Merge pull request #38654 from ROCmSoftwarePlatform:google_upstream_rocm_switch_to_rocm33
PiperOrigin-RevId: 308160300
Change-Id: Icd6e71c7bbfe22e0386ff7ebfdc926cf20b3b420
diff --git a/.github/ISSUE_TEMPLATE/60-tflite-converter-issue.md b/.github/ISSUE_TEMPLATE/60-tflite-converter-issue.md
index 32ebaff..6eab765e 100644
--- a/.github/ISSUE_TEMPLATE/60-tflite-converter-issue.md
+++ b/.github/ISSUE_TEMPLATE/60-tflite-converter-issue.md
@@ -38,6 +38,9 @@
- Producing correct results, but the model is slower than expected (model generated from old converter)
+**RNN conversion support**
+If converting TF RNN to TFLite fused RNN ops, please prefix [RNN] in the title.
+
**Any other info / logs**
Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.
diff --git a/.github/stale.yml b/.github/stale.yml
index ecc670a..e1184ce 100644
--- a/.github/stale.yml
+++ b/.github/stale.yml
@@ -25,7 +25,7 @@
daysUntilClose: 7
# Issues or Pull Requests with these labels will never be considered stale. Set to `[]` to disable
onlyLabels:
- - awaitingResponse
+ - stat:awaiting response
# Comment to post when marking as stale. Set to `false` to disable
markComment: >
This issue has been automatically marked as stale because it has not had
diff --git a/configure.py b/configure.py
index f051fab..ac9ed0c 100644
--- a/configure.py
+++ b/configure.py
@@ -1419,8 +1419,8 @@
environ_cp.get('LD_LIBRARY_PATH'))
if (environ_cp.get('TF_NEED_ROCM') == '1' and environ_cp.get('ROCM_PATH')):
- write_action_env_to_bazelrc('ROCM_PATH',environ_cp.get('ROCM_PATH'))
- write_action_env_to_bazelrc('ROCM_ROOT',environ_cp.get('ROCM_PATH'))
+ write_action_env_to_bazelrc('ROCM_PATH', environ_cp.get('ROCM_PATH'))
+ write_action_env_to_bazelrc('ROCM_ROOT', environ_cp.get('ROCM_PATH'))
environ_cp['TF_NEED_CUDA'] = str(
int(get_var(environ_cp, 'TF_NEED_CUDA', 'CUDA', False)))
diff --git a/tensorflow/BUILD b/tensorflow/BUILD
index f137e06..6ac4c81 100644
--- a/tensorflow/BUILD
+++ b/tensorflow/BUILD
@@ -523,6 +523,8 @@
],
)
+package_group(name = "ndarray_tensor_allow_list")
+
filegroup(
name = "intel_binary_blob",
data = if_mkl_ml(
diff --git a/tensorflow/api_template.__init__.py b/tensorflow/api_template.__init__.py
index d22eafa..f0f977a 100644
--- a/tensorflow/api_template.__init__.py
+++ b/tensorflow/api_template.__init__.py
@@ -116,7 +116,7 @@
# Get sitepackages directories for the python installation.
_site_packages_dirs = []
-_site_packages_dirs += [_site.USER_SITE]
+_site_packages_dirs += [] if _site.USER_SITE is None else [_site.USER_SITE]
_site_packages_dirs += [_p for _p in _sys.path if 'site-packages' in _p]
if 'getsitepackages' in dir(_site):
_site_packages_dirs += _site.getsitepackages()
diff --git a/tensorflow/api_template_v1.__init__.py b/tensorflow/api_template_v1.__init__.py
index f2856f8..dad91f2 100644
--- a/tensorflow/api_template_v1.__init__.py
+++ b/tensorflow/api_template_v1.__init__.py
@@ -126,7 +126,7 @@
# Get sitepackages directories for the python installation.
_site_packages_dirs = []
-_site_packages_dirs += [_site.USER_SITE]
+_site_packages_dirs += [] if _site.USER_SITE is None else [_site.USER_SITE]
_site_packages_dirs += [_p for _p in _sys.path if 'site-packages' in _p]
if 'getsitepackages' in dir(_site):
_site_packages_dirs += _site.getsitepackages()
diff --git a/tensorflow/c/BUILD b/tensorflow/c/BUILD
index 9bc96ff..c7ed2b6 100644
--- a/tensorflow/c/BUILD
+++ b/tensorflow/c/BUILD
@@ -118,6 +118,12 @@
visibility = ["//visibility:public"],
)
+cc_library(
+ name = "c_api_macros",
+ hdrs = ["c_api_macros.h"],
+ visibility = ["//visibility:public"],
+)
+
tf_cuda_library(
name = "c_api",
hdrs = [
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops.i b/tensorflow/c/c_api_macros.h
similarity index 60%
copy from tensorflow/lite/experimental/kernels/hashtable_ops.i
copy to tensorflow/c/c_api_macros.h
index fa2e6fa..85c9507 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops.i
+++ b/tensorflow/c/c_api_macros.h
@@ -13,8 +13,21 @@
limitations under the License.
==============================================================================*/
-%{
-#include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
-%}
+#ifndef TENSORFLOW_C_C_API_MACROS_H_
+#define TENSORFLOW_C_C_API_MACROS_H_
-%include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
+#ifdef SWIG
+#define TF_CAPI_EXPORT
+#else
+#if defined(_WIN32)
+#ifdef TF_COMPILE_LIBRARY
+#define TF_CAPI_EXPORT __declspec(dllexport)
+#else
+#define TF_CAPI_EXPORT __declspec(dllimport)
+#endif // TF_COMPILE_LIBRARY
+#else
+#define TF_CAPI_EXPORT __attribute__((visibility("default")))
+#endif // _WIN32
+#endif // SWIG
+
+#endif // TENSORFLOW_C_C_API_MACROS_H_
diff --git a/tensorflow/c/eager/BUILD b/tensorflow/c/eager/BUILD
index 3f9059b..e51a15f 100644
--- a/tensorflow/c/eager/BUILD
+++ b/tensorflow/c/eager/BUILD
@@ -41,12 +41,21 @@
":context_interface",
":operation_interface",
":tensor_handle_interface",
+ ":tfe_context_internal",
+ ":tfe_cancellation_manager_internal",
+ ":tfe_executor_internal",
+ ":tfe_monitoring_internal",
+ ":tfe_op_attrs_internal",
+ ":tfe_op_internal",
+ ":tfe_tensor_debug_info_internal",
+ ":tfe_tensorhandle_internal",
"@com_google_absl//absl/algorithm:container",
"@com_google_absl//absl/container:fixed_array",
"@com_google_absl//absl/types:span",
"@com_google_absl//absl/types:variant",
"//tensorflow/c:c_api",
"//tensorflow/c:c_api_internal",
+ "//tensorflow/c:tf_status_internal",
"//tensorflow/c:tf_tensor_internal",
"//tensorflow/core:core_cpu",
"//tensorflow/core/common_runtime/eager:attr_builder",
@@ -100,6 +109,14 @@
"dlpack.h",
"operation_interface.h",
"tensor_handle_interface.h",
+ "tfe_cancellation_manager_internal.h",
+ "tfe_context_internal.h",
+ "tfe_executor_internal.h",
+ "tfe_monitoring_internal.h",
+ "tfe_op_attrs_internal.h",
+ "tfe_op_internal.h",
+ "tfe_tensor_debug_info_internal.h",
+ "tfe_tensorhandle_internal.h",
],
visibility = [
"//tensorflow/core:__pkg__",
@@ -107,33 +124,27 @@
],
)
-tf_cuda_library(
+cc_library(
name = "c_api_internal",
- srcs = [
+ hdrs = [
"c_api_experimental.h",
- "c_api_unified_experimental.h",
+ "c_api_internal.h",
],
- hdrs = ["c_api_internal.h"],
visibility = [
"//learning/deepmind/courier:__subpackages__",
"//tensorflow:internal",
],
deps = [
":c_api",
- ":context_interface",
- ":operation_interface",
- ":tensor_handle_interface",
- "//tensorflow/c:c_api",
+ ":tfe_cancellation_manager_internal",
+ ":tfe_context_internal",
+ ":tfe_executor_internal",
+ ":tfe_monitoring_internal",
+ ":tfe_op_attrs_internal",
+ ":tfe_op_internal",
+ ":tfe_tensor_debug_info_internal",
+ ":tfe_tensorhandle_internal",
"//tensorflow/c:c_api_internal",
- "//tensorflow/core:core_cpu",
- "//tensorflow/core:core_cpu_lib",
- "//tensorflow/core:framework",
- "//tensorflow/core:framework_internal",
- "//tensorflow/core:framework_lite",
- "//tensorflow/core:lib",
- "//tensorflow/core:lib_internal",
- "//tensorflow/core/common_runtime/eager:attr_builder",
- "//tensorflow/core/common_runtime/eager:eager_executor",
],
)
@@ -184,6 +195,99 @@
],
)
+cc_library(
+ name = "tfe_context_internal",
+ hdrs = ["tfe_context_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ ":context_interface",
+ ],
+)
+
+cc_library(
+ name = "tfe_cancellation_manager_internal",
+ hdrs = ["tfe_cancellation_manager_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ "//tensorflow/core:framework",
+ ],
+)
+
+cc_library(
+ name = "tfe_executor_internal",
+ hdrs = ["tfe_executor_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ "//tensorflow/core/common_runtime/eager:eager_executor",
+ ],
+)
+
+cc_library(
+ name = "tfe_monitoring_internal",
+ hdrs = ["tfe_monitoring_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ "//tensorflow/core:lib",
+ "@com_google_absl//absl/memory",
+ ],
+)
+
+cc_library(
+ name = "tfe_op_attrs_internal",
+ hdrs = ["tfe_op_attrs_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ ":tfe_context_internal",
+ ":tfe_op_internal",
+ "//tensorflow/c:tf_status",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/core/common_runtime/eager:attr_builder",
+ ],
+)
+
+cc_library(
+ name = "tfe_op_internal",
+ hdrs = ["tfe_op_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ ":operation_interface",
+ ],
+)
+
+cc_library(
+ name = "tfe_tensor_debug_info_internal",
+ hdrs = ["tfe_tensor_debug_info_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "tfe_tensorhandle_internal",
+ hdrs = ["tfe_tensorhandle_internal.h"],
+ visibility = [
+ "//tensorflow:internal",
+ ],
+ deps = [
+ ":tensor_handle_interface",
+ ],
+)
+
tf_cuda_library(
name = "c_api_test_util",
testonly = 1,
diff --git a/tensorflow/c/eager/c_api.cc b/tensorflow/c/eager/c_api.cc
index 02a760d..658670a 100644
--- a/tensorflow/c/eager/c_api.cc
+++ b/tensorflow/c/eager/c_api.cc
@@ -611,13 +611,12 @@
}
}
- tensorflow::RemoteRendezvous* r =
- grpc_server->worker_env()->rendezvous_mgr->Find(context_id);
auto session_name = tensorflow::strings::StrCat("eager_", context_id);
- auto* device_mgr = grpc_server->worker_env()->device_mgr;
- std::shared_ptr<tensorflow::WorkerSession> worker_session;
-
if (reset_context) {
+ tensorflow::RemoteRendezvous* r =
+ grpc_server->worker_env()->rendezvous_mgr->Find(context_id);
+ auto* device_mgr = grpc_server->worker_env()->device_mgr;
+ std::shared_ptr<tensorflow::WorkerSession> worker_session;
TF_RETURN_IF_ERROR(grpc_server->worker_env()->session_mgr->CreateSession(
session_name, server_def, base_request.cluster_device_attributes(),
true));
@@ -647,10 +646,10 @@
LOG_AND_RETURN_IF_ERROR(
grpc_server->worker_env()->session_mgr->UpdateSession(
session_name, server_def, base_request.cluster_device_attributes(),
- true));
- LOG_AND_RETURN_IF_ERROR(context->UpdateRemoteMaster(
- grpc_server->worker_env(), std::move(remote_eager_workers),
- added_workers, removed_workers, context_id, r));
+ /*isolate_session_state=*/true));
+ LOG_AND_RETURN_IF_ERROR(
+ context->UpdateRemoteMaster(context_id, std::move(remote_eager_workers),
+ added_workers, removed_workers));
}
#undef LOG_AND_RETURN_IF_ERROR
@@ -1389,45 +1388,10 @@
return nullptr;
}
- tensorflow::TensorHandle* handle = nullptr;
- tensorflow::Device* device;
- tensorflow::EagerContext* context =
- tensorflow::ContextFromInterface(ctx->context);
- status->status = context->FindDeviceFromName(device_name, &device);
- if (!status->status.ok()) {
- tensorflow::CustomDevice* dev;
- status->status = context->FindCustomDeviceFromName(device_name, &dev);
- if (status->status.ok()) {
- status->status = dev->CopyTensorToDevice(
- tensorflow::TensorHandleFromInterface(h->handle), &handle);
- if (status->status.ok()) {
- return new TFE_TensorHandle{handle};
- }
- }
- return nullptr;
- }
- // Handle tensor handles currently in custom devices
- const char* handle_device_name = h->handle->DeviceName(&status->status);
- if (!status->status.ok()) {
- return nullptr;
- }
- tensorflow::CustomDevice* dev;
- status->status = context->FindCustomDeviceFromName(handle_device_name, &dev);
+ auto* result = ctx->context->CopyTensorHandleToDevice(h->handle, device_name,
+ &status->status);
if (status->status.ok()) {
- status->status = dev->CopyTensorFromDevice(
- tensorflow::TensorHandleFromInterface(h->handle), device_name, &handle);
- if (status->status.ok()) {
- return new TFE_TensorHandle{handle};
- }
- return nullptr;
- }
-
- // Handle regular case.
- status->status = tensorflow::EagerCopyToDevice(
- tensorflow::TensorHandleFromInterface(h->handle), context,
- &context->Executor(), device, false, &handle);
- if (status->status.ok()) {
- return new TFE_TensorHandle{handle};
+ return new TFE_TensorHandle{result};
}
return nullptr;
}
diff --git a/tensorflow/c/eager/c_api_debug.cc b/tensorflow/c/eager/c_api_debug.cc
index f5bf029..d3d1126 100644
--- a/tensorflow/c/eager/c_api_debug.cc
+++ b/tensorflow/c/eager/c_api_debug.cc
@@ -17,8 +17,11 @@
#include "tensorflow/c/c_api.h"
#include "tensorflow/c/eager/c_api.h"
-#include "tensorflow/c/eager/c_api_internal.h"
+#include "tensorflow/c/eager/tfe_tensor_debug_info_internal.h"
+#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
+#include "tensorflow/c/tf_status_internal.h"
#include "tensorflow/core/common_runtime/eager/tensor_handle.h"
+#include "tensorflow/core/platform/status.h"
#ifdef TENSORFLOW_EAGER_USE_XLA
#include "tensorflow/compiler/jit/xla_device.h"
#endif // TENSORFLOW_EAGER_USE_XLA
diff --git a/tensorflow/c/eager/c_api_internal.h b/tensorflow/c/eager/c_api_internal.h
index 00798c3..39b767d 100644
--- a/tensorflow/c/eager/c_api_internal.h
+++ b/tensorflow/c/eager/c_api_internal.h
@@ -15,39 +15,20 @@
#ifndef TENSORFLOW_C_EAGER_C_API_INTERNAL_H_
#define TENSORFLOW_C_EAGER_C_API_INTERNAL_H_
-#include <algorithm>
-#include <cstddef>
-#include <map>
-#include <memory>
-#include <queue>
-#include <string>
-#include <vector>
-
-#include "tensorflow/c/c_api.h"
#include "tensorflow/c/c_api_internal.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
-#include "tensorflow/c/eager/context_interface.h"
-#include "tensorflow/c/eager/operation_interface.h"
-#include "tensorflow/c/eager/tensor_handle_interface.h"
-#include "tensorflow/core/common_runtime/device_factory.h"
-#include "tensorflow/core/common_runtime/eager/attr_builder.h"
-#include "tensorflow/core/common_runtime/eager/eager_executor.h"
-#include "tensorflow/core/common_runtime/function.h"
-#include "tensorflow/core/common_runtime/rendezvous_mgr.h"
-#include "tensorflow/core/framework/cancellation.h"
-#include "tensorflow/core/framework/rendezvous.h"
-#include "tensorflow/core/lib/gtl/inlined_vector.h"
-#include "tensorflow/core/lib/gtl/map_util.h"
-#include "tensorflow/core/lib/monitoring/counter.h"
-#include "tensorflow/core/lib/monitoring/gauge.h"
-#include "tensorflow/core/lib/monitoring/sampler.h"
-#include "tensorflow/core/platform/errors.h"
-#include "tensorflow/core/platform/mutex.h"
-#include "tensorflow/core/platform/stringpiece.h"
-#include "tensorflow/core/platform/thread_annotations.h"
-#include "tensorflow/core/public/version.h"
+#include "tensorflow/c/eager/tfe_cancellation_manager_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_context_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_executor_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_monitoring_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_op_attrs_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_op_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_tensor_debug_info_internal.h" // IWYU pragma: export
+#include "tensorflow/c/eager/tfe_tensorhandle_internal.h" // IWYU pragma: export
+// TODO(b/154564140): Move this to its own header. This requires splitting
+// c_api_experimental.h
struct TFE_ContextOptions {
TF_SessionOptions session_options;
// true if async execution is enabled.
@@ -61,199 +42,4 @@
bool use_tfrt = false;
};
-// Wraps a pointer to a context implementation.
-//
-// WARNING: Since the underlying object could be ref-counted a user of this
-// interface cannot destruct the underlying context object. Instead, call
-// TFE_DeleteContext who calls Release() on the context pointer and deletes
-// the TFE_Context structure.
-struct TFE_Context {
- tensorflow::AbstractContextInterface* context;
-};
-
-// Wraps a pointer to a tensor handle implementation.
-//
-// WARNING: Since the underlying object could be ref-counted a user of this
-// interface cannot destruct the underlying handle object. Instead, call
-// TFE_DeleteTensorHandle who calls Release() on the handle pointer and deletes
-// the TFE_TensorHandle structure.
-struct TFE_TensorHandle {
- tensorflow::AbstractTensorHandleInterface* handle;
-};
-
-struct TFE_TensorDebugInfo {
- explicit TFE_TensorDebugInfo(const std::vector<tensorflow::int64>& dims)
- : dev_dims(dims) {}
-
- // Fully-padded, minor-to-major.
- std::vector<tensorflow::int64> dev_dims;
-};
-
-// Wraps a pointer to an operation implementation.
-//
-// WARNING: Since the underlying object could be ref-counted a user of this
-// interface cannot destruct the underlying operation object. Instead, call
-// TFE_DeleteOp who calls Release() on the operation pointer and deletes
-// the TFE_Op structure.
-struct TFE_Op {
- tensorflow::AbstractOperationInterface* operation;
-};
-
-struct TFE_MonitoringCounterCell {
- tensorflow::monitoring::CounterCell cell;
-};
-
-template <int NumLabels>
-struct TFE_MonitoringCounter {
- template <typename... LabelDesc>
- TFE_MonitoringCounter(const char* name, const char* description,
- LabelDesc&&... label) {
- counter = absl::WrapUnique(tensorflow::monitoring::Counter<NumLabels>::New(
- name, description, label...));
- }
-
- std::unique_ptr<tensorflow::monitoring::Counter<NumLabels>> counter;
-};
-
-struct TFE_MonitoringCounter0 : TFE_MonitoringCounter<0> {
- using TFE_MonitoringCounter::TFE_MonitoringCounter;
-};
-struct TFE_MonitoringCounter1 : TFE_MonitoringCounter<1> {
- using TFE_MonitoringCounter::TFE_MonitoringCounter;
-};
-struct TFE_MonitoringCounter2 : TFE_MonitoringCounter<2> {
- using TFE_MonitoringCounter::TFE_MonitoringCounter;
-};
-
-struct TFE_MonitoringIntGaugeCell {
- tensorflow::monitoring::GaugeCell<tensorflow::int64> cell;
-};
-struct TFE_MonitoringStringGaugeCell {
- tensorflow::monitoring::GaugeCell<tensorflow::string> cell;
-};
-struct TFE_MonitoringBoolGaugeCell {
- tensorflow::monitoring::GaugeCell<bool> cell;
-};
-
-template <typename ValueType, int NumLabels>
-struct TFE_MonitoringGauge {
- template <typename... LabelDesc>
- TFE_MonitoringGauge(const char* name, const char* description,
- LabelDesc&&... label) {
- gauge = absl::WrapUnique(
- tensorflow::monitoring::Gauge<ValueType, NumLabels>::New(
- name, description, label...));
- }
-
- std::unique_ptr<tensorflow::monitoring::Gauge<ValueType, NumLabels>> gauge;
-};
-
-struct TFE_MonitoringIntGauge0 : TFE_MonitoringGauge<tensorflow::int64, 0> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringIntGauge1 : TFE_MonitoringGauge<tensorflow::int64, 1> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringIntGauge2 : TFE_MonitoringGauge<tensorflow::int64, 2> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-
-struct TFE_MonitoringStringGauge0 : TFE_MonitoringGauge<tensorflow::string, 0> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringStringGauge1 : TFE_MonitoringGauge<tensorflow::string, 1> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringStringGauge2 : TFE_MonitoringGauge<tensorflow::string, 2> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-
-struct TFE_MonitoringBoolGauge0 : TFE_MonitoringGauge<bool, 0> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringBoolGauge1 : TFE_MonitoringGauge<bool, 1> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-struct TFE_MonitoringBoolGauge2 : TFE_MonitoringGauge<bool, 2> {
- using TFE_MonitoringGauge::TFE_MonitoringGauge;
-};
-
-struct TFE_MonitoringBuckets {
- explicit TFE_MonitoringBuckets(
- std::function<std::unique_ptr<tensorflow::monitoring::Buckets>(void)>
- fn) {
- create_buckets = fn;
- }
-
- std::function<std::unique_ptr<tensorflow::monitoring::Buckets>(void)>
- create_buckets;
-};
-
-struct TFE_MonitoringSamplerCell {
- tensorflow::monitoring::SamplerCell cell;
-};
-
-template <int NumLabels>
-struct TFE_MonitoringSampler {
- template <typename... LabelDesc>
- TFE_MonitoringSampler(
- const char* name,
- std::unique_ptr<tensorflow::monitoring::Buckets> buckets,
- const char* description, LabelDesc&&... label) {
- sampler = absl::WrapUnique(tensorflow::monitoring::Sampler<NumLabels>::New(
- {name, description, label...}, std::move(buckets)));
- }
-
- std::unique_ptr<tensorflow::monitoring::Sampler<NumLabels>> sampler;
-};
-
-struct TFE_MonitoringSampler0 : TFE_MonitoringSampler<0> {
- using TFE_MonitoringSampler::TFE_MonitoringSampler;
-};
-struct TFE_MonitoringSampler1 : TFE_MonitoringSampler<1> {
- using TFE_MonitoringSampler::TFE_MonitoringSampler;
-};
-struct TFE_MonitoringSampler2 : TFE_MonitoringSampler<2> {
- using TFE_MonitoringSampler::TFE_MonitoringSampler;
-};
-
-namespace tensorflow {
-// Set an AttrValue on the op. Doesn't handle the list types.
-void SetOpAttrValueScalar(TFE_Context* ctx, TFE_Op* op,
- const tensorflow::AttrValue& default_value,
- const char* attr_name, TF_Status* status);
-} // namespace tensorflow
-
-struct TFE_CancellationManager {
- tensorflow::CancellationManager cancellation_manager;
-};
-
-struct TFE_Executor {
- explicit TFE_Executor(bool async)
- : owned_executor(new tensorflow::EagerExecutor(async)) {}
-
- explicit TFE_Executor(tensorflow::EagerExecutor* executor)
- : owned_executor(nullptr), unowned_executor(executor) {}
-
- tensorflow::EagerExecutor* executor() {
- return owned_executor == nullptr ? unowned_executor : owned_executor.get();
- }
-
- std::unique_ptr<tensorflow::EagerExecutor> owned_executor;
- tensorflow::EagerExecutor* unowned_executor;
-};
-
-// An equivalent of a tensorflow::NameAttrList protocol buffer, but used in ways
-// that sometimes do not require serialization.
-struct TFE_OpAttrs {
- explicit TFE_OpAttrs() : name(nullptr), attributes(nullptr) {}
-
- explicit TFE_OpAttrs(const tensorflow::AttrBuilder* value,
- const char* op_name)
- : name(op_name), attributes(value) {}
-
- const char* name;
- const tensorflow::AttrBuilder* attributes;
-};
-
#endif // TENSORFLOW_C_EAGER_C_API_INTERNAL_H_
diff --git a/tensorflow/c/eager/context_interface.h b/tensorflow/c/eager/context_interface.h
index 157f10c..2d26eda 100644
--- a/tensorflow/c/eager/context_interface.h
+++ b/tensorflow/c/eager/context_interface.h
@@ -60,6 +60,10 @@
// Create a handle to wrap and manage a Tensor
virtual AbstractTensorHandleInterface* CreateLocalHandle(
AbstractTensorInterface* t) = 0;
+ // Copy the handle to another device.
+ virtual AbstractTensorHandleInterface* CopyTensorHandleToDevice(
+ AbstractTensorHandleInterface* handle, const char* device_name,
+ Status* status) = 0;
// Create an operation to perform op execution
virtual AbstractOperationInterface* CreateOperation() = 0;
diff --git a/tensorflow/c/eager/parallel_device/BUILD b/tensorflow/c/eager/parallel_device/BUILD
index 9d787d2..f4dbcc6 100644
--- a/tensorflow/c/eager/parallel_device/BUILD
+++ b/tensorflow/c/eager/parallel_device/BUILD
@@ -7,10 +7,26 @@
licenses = ["notice"], # Apache 2.0
)
+# Currently pybind extension shared objects must use only C API headers since
+# the C API has static initializers duplicated in the Python bindings. So we
+# need a second rule that omits .cc files, in
+# tensorflow/python:_pywrap_parallel_device.
+filegroup(
+ name = "headers",
+ srcs = ["parallel_device.h"],
+ visibility = ["//tensorflow/python:__pkg__"],
+)
+
+filegroup(
+ name = "sources",
+ srcs = ["parallel_device.cc"],
+ visibility = ["//tensorflow/python:__pkg__"],
+)
+
cc_library(
name = "parallel_device",
- srcs = ["parallel_device.cc"],
- hdrs = ["parallel_device.h"],
+ srcs = [":sources"],
+ hdrs = [":headers"],
deps = [
"//tensorflow/c:c_api",
"//tensorflow/c/eager:c_api",
diff --git a/tensorflow/c/eager/parallel_device/parallel_device.cc b/tensorflow/c/eager/parallel_device/parallel_device.cc
index bd5d8e7..e684680 100644
--- a/tensorflow/c/eager/parallel_device/parallel_device.cc
+++ b/tensorflow/c/eager/parallel_device/parallel_device.cc
@@ -574,23 +574,21 @@
} // namespace
-void RegisterParallelDevice(TFE_Context* context, const char* device_name,
- const char** underlying_devices,
- int num_underlying_devices, TF_Status* status) {
- TFE_CustomDevice custom_device;
- custom_device.copy_tensor_to_device = &CopyToParallelDevice;
- custom_device.copy_tensor_from_device = &CopyTensorFromParallelDevice;
- custom_device.delete_device = &DeleteParallelDevice;
- custom_device.execute = &ParallelDeviceExecute;
+void AllocateParallelDevice(const char* device_name,
+ const char* const* underlying_devices,
+ int num_underlying_devices,
+ TFE_CustomDevice* device, void** device_info) {
+ device->copy_tensor_to_device = &CopyToParallelDevice;
+ device->copy_tensor_from_device = &CopyTensorFromParallelDevice;
+ device->delete_device = &DeleteParallelDevice;
+ device->execute = &ParallelDeviceExecute;
std::vector<std::string> underlying_devices_vector;
underlying_devices_vector.reserve(num_underlying_devices);
for (int device_index = 0; device_index < num_underlying_devices;
++device_index) {
underlying_devices_vector.push_back(underlying_devices[device_index]);
}
- ParallelDevice* d =
- new ParallelDevice(device_name, underlying_devices_vector);
- TFE_RegisterCustomDevice(context, custom_device, device_name, d, status);
+ *device_info = new ParallelDevice(device_name, underlying_devices_vector);
}
} // namespace eager
diff --git a/tensorflow/c/eager/parallel_device/parallel_device.h b/tensorflow/c/eager/parallel_device/parallel_device.h
index b106524..f448a4c 100644
--- a/tensorflow/c/eager/parallel_device/parallel_device.h
+++ b/tensorflow/c/eager/parallel_device/parallel_device.h
@@ -16,12 +16,14 @@
#ifndef TENSORFLOW_C_EAGER_PARALLEL_DEVICE_PARALLEL_DEVICE_H_
#define TENSORFLOW_C_EAGER_PARALLEL_DEVICE_PARALLEL_DEVICE_H_
+#include "tensorflow/c/c_api.h"
#include "tensorflow/c/eager/c_api.h"
+#include "tensorflow/c/eager/c_api_experimental.h"
namespace tensorflow {
namespace eager {
-// Register a parallel device named `device_name` which forwards operations to
+// Allocate a parallel device named `device_name` which forwards operations to
// `underlying_devices`, maintaining "parallel tensors" with components placed
// on each underlying device.
//
@@ -50,11 +52,12 @@
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
-// `context` owns the parallel device. `underlying_devices` must stay valid
-// while the parallel device is in use.
-void RegisterParallelDevice(TFE_Context* context, const char* device_name,
- const char** underlying_devices,
- int num_underlying_devices, TF_Status* status);
+// The filled `device` struct and the allocated `device_info` struct may be
+// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
+void AllocateParallelDevice(const char* device_name,
+ const char* const* underlying_devices,
+ int num_underlying_devices,
+ TFE_CustomDevice* device, void** device_info);
} // namespace eager
} // namespace tensorflow
diff --git a/tensorflow/c/eager/parallel_device/parallel_device_test.cc b/tensorflow/c/eager/parallel_device/parallel_device_test.cc
index 41c7d64..9b0613b 100644
--- a/tensorflow/c/eager/parallel_device/parallel_device_test.cc
+++ b/tensorflow/c/eager/parallel_device/parallel_device_test.cc
@@ -288,6 +288,19 @@
*static_cast<float*>(TF_TensorData(value_zero.get())));
}
+template <std::size_t num_devices>
+void RegisterParallelDevice(
+ TFE_Context* context, const char* device_name,
+ const std::array<const char*, num_devices>& underlying_devices,
+ TF_Status* status) {
+ TFE_CustomDevice device;
+ void* device_info;
+ tensorflow::eager::AllocateParallelDevice(
+ device_name, underlying_devices.data(), underlying_devices.size(),
+ &device, &device_info);
+ TFE_RegisterCustomDevice(context, device, device_name, device_info, status);
+}
+
// Create and modify a variable placed on a parallel device which composes
// `first_device` and `second_device`.
void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
@@ -297,9 +310,8 @@
TF_NewStatus(), TF_DeleteStatus);
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
std::array<const char*, 2> underlying_devices{first_device, second_device};
- tensorflow::eager::RegisterParallelDevice(
- context, device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ RegisterParallelDevice(context, device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
// Create a variable handle (uninitialized to start) placed on the parallel
@@ -456,16 +468,14 @@
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> underlying_devices;
const char* first_device_name =
"/job:localhost/replica:0/task:0/device:CPU:0";
- underlying_devices.push_back(first_device_name);
const char* second_device_name =
"/job:localhost/replica:0/task:0/device:CPU:1";
- underlying_devices.push_back(second_device_name);
- tensorflow::eager::RegisterParallelDevice(
- context.get(), device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ std::array<const char*, 2> underlying_devices{first_device_name,
+ second_device_name};
+ RegisterParallelDevice(context.get(), device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
TensorHandlePtr cpu_value(FloatTensorHandle(3., status.get()));
@@ -524,12 +534,11 @@
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> underlying_devices;
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:0");
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:1");
- tensorflow::eager::RegisterParallelDevice(
- context.get(), device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ std::array<const char*, 2> underlying_devices{
+ "/job:localhost/replica:0/task:0/device:CPU:0",
+ "/job:localhost/replica:0/task:0/device:CPU:1"};
+ RegisterParallelDevice(context.get(), device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
// Create two vectors with different lengths
@@ -570,24 +579,22 @@
// Create a parallel device with two CPUs
const char* first_device_name =
"/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> first_underlying_devices{
+ std::array<const char*, 2> first_underlying_devices{
"/job:localhost/replica:0/task:0/device:CPU:0",
"/job:localhost/replica:0/task:0/device:CPU:1"};
- tensorflow::eager::RegisterParallelDevice(
- context.get(), first_device_name, first_underlying_devices.data(),
- first_underlying_devices.size(), status.get());
+ RegisterParallelDevice(context.get(), first_device_name,
+ first_underlying_devices, status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
// Create a second parallel device with the first parallel device and one
// additional CPU.
const char* second_device_name =
"/job:localhost/replica:0/task:0/device:CUSTOM:1";
- std::vector<const char*> second_underlying_devices{
+ std::array<const char*, 2> second_underlying_devices{
"/job:localhost/replica:0/task:0/device:CUSTOM:0",
"/job:localhost/replica:0/task:0/device:CPU:2"};
- tensorflow::eager::RegisterParallelDevice(
- context.get(), second_device_name, second_underlying_devices.data(),
- second_underlying_devices.size(), status.get());
+ RegisterParallelDevice(context.get(), second_device_name,
+ second_underlying_devices, status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
// Create a tensor on the first parallel device
@@ -656,11 +663,10 @@
std::unique_ptr<TFE_Context, decltype(&TFE_DeleteContext)> context(
TFE_NewContext(opts.get(), status.get()), TFE_DeleteContext);
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> underlying_devices;
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:0");
- tensorflow::eager::RegisterParallelDevice(
- context.get(), device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ std::array<const char*, 1> underlying_devices{
+ "/job:localhost/replica:0/task:0/device:CPU:0"};
+ RegisterParallelDevice(context.get(), device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
TensorHandlePtr value_one(FloatTensorHandle(1., status.get()));
@@ -775,12 +781,11 @@
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> underlying_devices;
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:0");
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:1");
- tensorflow::eager::RegisterParallelDevice(
- context.get(), device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ std::array<const char*, 2> underlying_devices{
+ "/job:localhost/replica:0/task:0/device:CPU:0",
+ "/job:localhost/replica:0/task:0/device:CPU:1"};
+ RegisterParallelDevice(context.get(), device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
// Create a tensor on the parallel device
@@ -867,12 +872,11 @@
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
- std::vector<const char*> underlying_devices;
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:0");
- underlying_devices.push_back("/job:localhost/replica:0/task:0/device:CPU:1");
- tensorflow::eager::RegisterParallelDevice(
- context.get(), device_name, underlying_devices.data(),
- underlying_devices.size(), status.get());
+ std::array<const char*, 2> underlying_devices{
+ "/job:localhost/replica:0/task:0/device:CPU:0",
+ "/job:localhost/replica:0/task:0/device:CPU:1"};
+ RegisterParallelDevice(context.get(), device_name, underlying_devices,
+ status.get());
ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
const char* function_name = "test_reduce_mul";
diff --git a/tensorflow/c/eager/tfe_cancellation_manager_internal.h b/tensorflow/c/eager/tfe_cancellation_manager_internal.h
new file mode 100644
index 0000000..7d500c8
--- /dev/null
+++ b/tensorflow/c/eager/tfe_cancellation_manager_internal.h
@@ -0,0 +1,24 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_CANCELLATION_MANAGER_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_CANCELLATION_MANAGER_INTERNAL_H_
+
+#include "tensorflow/core/framework/cancellation.h"
+
+struct TFE_CancellationManager {
+ tensorflow::CancellationManager cancellation_manager;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_CANCELLATION_MANAGER_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_context_internal.h b/tensorflow/c/eager/tfe_context_internal.h
new file mode 100644
index 0000000..4c2e650
--- /dev/null
+++ b/tensorflow/c/eager/tfe_context_internal.h
@@ -0,0 +1,30 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_CONTEXT_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_CONTEXT_INTERNAL_H_
+
+#include "tensorflow/c/eager/context_interface.h"
+
+// Wraps a pointer to a context implementation.
+//
+// WARNING: Since the underlying object could be ref-counted a user of this
+// interface cannot destruct the underlying context object. Instead, call
+// TFE_DeleteContext who calls Release() on the context pointer and deletes
+// the TFE_Context structure.
+struct TFE_Context {
+ tensorflow::AbstractContextInterface* context;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_CONTEXT_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_executor_internal.h b/tensorflow/c/eager/tfe_executor_internal.h
new file mode 100644
index 0000000..442103f
--- /dev/null
+++ b/tensorflow/c/eager/tfe_executor_internal.h
@@ -0,0 +1,37 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_EXECUTOR_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_EXECUTOR_INTERNAL_H_
+
+#include <memory>
+
+#include "tensorflow/core/common_runtime/eager/eager_executor.h"
+
+struct TFE_Executor {
+ explicit TFE_Executor(bool async)
+ : owned_executor(new tensorflow::EagerExecutor(async)) {}
+
+ explicit TFE_Executor(tensorflow::EagerExecutor* executor)
+ : owned_executor(nullptr), unowned_executor(executor) {}
+
+ tensorflow::EagerExecutor* executor() {
+ return owned_executor == nullptr ? unowned_executor : owned_executor.get();
+ }
+
+ std::unique_ptr<tensorflow::EagerExecutor> owned_executor;
+ tensorflow::EagerExecutor* unowned_executor;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_EXECUTOR_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_monitoring_internal.h b/tensorflow/c/eager/tfe_monitoring_internal.h
new file mode 100644
index 0000000..d822685
--- /dev/null
+++ b/tensorflow/c/eager/tfe_monitoring_internal.h
@@ -0,0 +1,146 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_MONITORING_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_MONITORING_INTERNAL_H_
+
+#include <functional>
+#include <memory>
+#include <string>
+
+#include "absl/memory/memory.h"
+#include "tensorflow/core/lib/monitoring/counter.h"
+#include "tensorflow/core/lib/monitoring/gauge.h"
+#include "tensorflow/core/lib/monitoring/sampler.h"
+#include "tensorflow/core/platform/types.h"
+
+struct TFE_MonitoringCounterCell {
+ tensorflow::monitoring::CounterCell cell;
+};
+
+template <int NumLabels>
+struct TFE_MonitoringCounter {
+ template <typename... LabelDesc>
+ TFE_MonitoringCounter(const char* name, const char* description,
+ LabelDesc&&... label) {
+ counter = absl::WrapUnique(tensorflow::monitoring::Counter<NumLabels>::New(
+ name, description, label...));
+ }
+
+ std::unique_ptr<tensorflow::monitoring::Counter<NumLabels>> counter;
+};
+
+struct TFE_MonitoringCounter0 : TFE_MonitoringCounter<0> {
+ using TFE_MonitoringCounter::TFE_MonitoringCounter;
+};
+struct TFE_MonitoringCounter1 : TFE_MonitoringCounter<1> {
+ using TFE_MonitoringCounter::TFE_MonitoringCounter;
+};
+struct TFE_MonitoringCounter2 : TFE_MonitoringCounter<2> {
+ using TFE_MonitoringCounter::TFE_MonitoringCounter;
+};
+
+struct TFE_MonitoringIntGaugeCell {
+ tensorflow::monitoring::GaugeCell<tensorflow::int64> cell;
+};
+struct TFE_MonitoringStringGaugeCell {
+ tensorflow::monitoring::GaugeCell<tensorflow::string> cell;
+};
+struct TFE_MonitoringBoolGaugeCell {
+ tensorflow::monitoring::GaugeCell<bool> cell;
+};
+
+template <typename ValueType, int NumLabels>
+struct TFE_MonitoringGauge {
+ template <typename... LabelDesc>
+ TFE_MonitoringGauge(const char* name, const char* description,
+ LabelDesc&&... label) {
+ gauge = absl::WrapUnique(
+ tensorflow::monitoring::Gauge<ValueType, NumLabels>::New(
+ name, description, label...));
+ }
+
+ std::unique_ptr<tensorflow::monitoring::Gauge<ValueType, NumLabels>> gauge;
+};
+
+struct TFE_MonitoringIntGauge0 : TFE_MonitoringGauge<tensorflow::int64, 0> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringIntGauge1 : TFE_MonitoringGauge<tensorflow::int64, 1> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringIntGauge2 : TFE_MonitoringGauge<tensorflow::int64, 2> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+
+struct TFE_MonitoringStringGauge0 : TFE_MonitoringGauge<tensorflow::string, 0> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringStringGauge1 : TFE_MonitoringGauge<tensorflow::string, 1> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringStringGauge2 : TFE_MonitoringGauge<tensorflow::string, 2> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+
+struct TFE_MonitoringBoolGauge0 : TFE_MonitoringGauge<bool, 0> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringBoolGauge1 : TFE_MonitoringGauge<bool, 1> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+struct TFE_MonitoringBoolGauge2 : TFE_MonitoringGauge<bool, 2> {
+ using TFE_MonitoringGauge::TFE_MonitoringGauge;
+};
+
+struct TFE_MonitoringBuckets {
+ explicit TFE_MonitoringBuckets(
+ std::function<std::unique_ptr<tensorflow::monitoring::Buckets>(void)>
+ fn) {
+ create_buckets = fn;
+ }
+
+ std::function<std::unique_ptr<tensorflow::monitoring::Buckets>(void)>
+ create_buckets;
+};
+
+struct TFE_MonitoringSamplerCell {
+ tensorflow::monitoring::SamplerCell cell;
+};
+
+template <int NumLabels>
+struct TFE_MonitoringSampler {
+ template <typename... LabelDesc>
+ TFE_MonitoringSampler(
+ const char* name,
+ std::unique_ptr<tensorflow::monitoring::Buckets> buckets,
+ const char* description, LabelDesc&&... label) {
+ sampler = absl::WrapUnique(tensorflow::monitoring::Sampler<NumLabels>::New(
+ {name, description, label...}, std::move(buckets)));
+ }
+
+ std::unique_ptr<tensorflow::monitoring::Sampler<NumLabels>> sampler;
+};
+
+struct TFE_MonitoringSampler0 : TFE_MonitoringSampler<0> {
+ using TFE_MonitoringSampler::TFE_MonitoringSampler;
+};
+struct TFE_MonitoringSampler1 : TFE_MonitoringSampler<1> {
+ using TFE_MonitoringSampler::TFE_MonitoringSampler;
+};
+struct TFE_MonitoringSampler2 : TFE_MonitoringSampler<2> {
+ using TFE_MonitoringSampler::TFE_MonitoringSampler;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_MONITORING_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_op_attrs_internal.h b/tensorflow/c/eager/tfe_op_attrs_internal.h
new file mode 100644
index 0000000..5351da5
--- /dev/null
+++ b/tensorflow/c/eager/tfe_op_attrs_internal.h
@@ -0,0 +1,52 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_OP_ATTRS_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_OP_ATTRS_INTERNAL_H_
+
+#include <algorithm>
+#include <cstddef>
+#include <map>
+#include <memory>
+#include <queue>
+#include <string>
+#include <vector>
+
+#include "tensorflow/c/eager/tfe_context_internal.h"
+#include "tensorflow/c/eager/tfe_op_internal.h"
+#include "tensorflow/c/tf_status.h"
+#include "tensorflow/core/common_runtime/eager/attr_builder.h"
+#include "tensorflow/core/framework/attr_value.pb.h"
+
+// An equivalent of a tensorflow::NameAttrList protocol buffer, but used in ways
+// that sometimes do not require serialization.
+struct TFE_OpAttrs {
+ explicit TFE_OpAttrs() : name(nullptr), attributes(nullptr) {}
+
+ explicit TFE_OpAttrs(const tensorflow::AttrBuilder* value,
+ const char* op_name)
+ : name(op_name), attributes(value) {}
+
+ const char* name;
+ const tensorflow::AttrBuilder* attributes;
+};
+
+namespace tensorflow {
+// Set an AttrValue on the op. Doesn't handle the list types.
+void SetOpAttrValueScalar(TFE_Context* ctx, TFE_Op* op,
+ const tensorflow::AttrValue& default_value,
+ const char* attr_name, TF_Status* status);
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EAGER_TFE_OP_ATTRS_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_op_internal.h b/tensorflow/c/eager/tfe_op_internal.h
new file mode 100644
index 0000000..b9292e2
--- /dev/null
+++ b/tensorflow/c/eager/tfe_op_internal.h
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_OP_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_OP_INTERNAL_H_
+
+#include "tensorflow/c/eager/operation_interface.h"
+
+// Wraps a pointer to an operation implementation.
+//
+// WARNING: Since the underlying object could be ref-counted a user of this
+// interface cannot destruct the underlying operation object. Instead, call
+// TFE_DeleteOp who calls Release() on the operation pointer and deletes
+// the TFE_Op structure.
+struct TFE_Op {
+ tensorflow::AbstractOperationInterface* operation;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_OP_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_tensor_debug_info_internal.h b/tensorflow/c/eager/tfe_tensor_debug_info_internal.h
new file mode 100644
index 0000000..a9cf12a
--- /dev/null
+++ b/tensorflow/c/eager/tfe_tensor_debug_info_internal.h
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_TENSOR_DEBUG_INFO_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_TENSOR_DEBUG_INFO_INTERNAL_H_
+
+#include <vector>
+
+#include "tensorflow/core/platform/types.h"
+
+struct TFE_TensorDebugInfo {
+ explicit TFE_TensorDebugInfo(const std::vector<tensorflow::int64>& dims)
+ : dev_dims(dims) {}
+
+ // Fully-padded, minor-to-major.
+ std::vector<tensorflow::int64> dev_dims;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_TENSOR_DEBUG_INFO_INTERNAL_H_
diff --git a/tensorflow/c/eager/tfe_tensorhandle_internal.h b/tensorflow/c/eager/tfe_tensorhandle_internal.h
new file mode 100644
index 0000000..39843c0
--- /dev/null
+++ b/tensorflow/c/eager/tfe_tensorhandle_internal.h
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_C_EAGER_TFE_TENSORHANDLE_INTERNAL_H_
+#define TENSORFLOW_C_EAGER_TFE_TENSORHANDLE_INTERNAL_H_
+
+#include "tensorflow/c/eager/tensor_handle_interface.h"
+
+// Wraps a pointer to a tensor handle implementation.
+//
+// WARNING: Since the underlying object could be ref-counted a user of this
+// interface cannot destruct the underlying handle object. Instead, call
+// TFE_DeleteTensorHandle who calls Release() on the handle pointer and deletes
+// the TFE_TensorHandle structure.
+struct TFE_TensorHandle {
+ tensorflow::AbstractTensorHandleInterface* handle;
+};
+
+#endif // TENSORFLOW_C_EAGER_TFE_TENSORHANDLE_INTERNAL_H_
diff --git a/tensorflow/c/experimental/saved_model/README.md b/tensorflow/c/experimental/saved_model/README.md
new file mode 100644
index 0000000..2fdb813
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/README.md
@@ -0,0 +1,66 @@
+# Tensorflow C SavedModel API
+
+## Overview
+
+These are the new experimental C SavedModel APIs for loading and running
+SavedModels in a TF2-idiomatic fashion. See
+[RFC 207](https://github.com/tensorflow/community/pull/207) for additional
+context.
+
+The directory structure is as follows:
+
+```none
+saved_model/
+
+ public/
+
+ internal/
+
+ core/
+
+```
+
+## saved_model/public
+
+`saved_model/public` is intended to house *only the public headers* of the
+SavedModel C API.
+
+These headers:
+
+1. declare opaque C types (like `TF_SavedModel`),
+
+2. declare the functions that operate on these types (like `TF_LoadSavedModel`).
+
+Once they leave experimental, these APIs should be considered stable for use
+by external clients.
+
+These headers are in a separate directory to make it obvious to clients which
+headers they should depend on, and which headers are implementation details.
+Separating these public headers by directory also allow future programmatic
+checks to ensure that TF public headers only `#include` other public TF headers.
+
+## saved_model/internal
+
+`saved_model/internal` is the "glue" between the C API and the internal C++
+implementation.
+
+Its role is to:
+
+1. implement the C API functions declared in `saved_model/public`
+
+2. define the C API types declared in `saved_model/public`
+
+The files fulfilling 1. are named `*.cc` (eg: `concrete_function.cc`), while
+the files fulfilling 2. are `*type.h` (eg: `concrete_function_type.h`).
+
+The headers exposing the internal implementation of the opaque C types are only
+visible to other implementors of the C API. This is similar to how other
+TF C API implementations use `tf_status_internal.h` (to extract the underlying
+`tensorflow::Status`). All other targets in this directory are private.
+
+## saved_model/core
+
+`saved_model/core` contains pure C++ "Classes" underlying the C API types
+in `saved_model/public/`. These are implementation
+details subject to change, and have limited visibility to implementors only.
+This is the bottom-most layer of the `C++ -> C -> C++` sandwich.
diff --git a/tensorflow/c/experimental/saved_model/core/BUILD b/tensorflow/c/experimental/saved_model/core/BUILD
new file mode 100644
index 0000000..68e46a4
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/core/BUILD
@@ -0,0 +1,46 @@
+# Experimental SavedModel C APIs for TensorFlow. See RFC
+# https://github.com/tensorflow/community/pull/207
+# Targets in this directory are pure C++ "Classes" underlying the C API types
+# under tf/c/experimental/saved_model/public/. They are subject to change and
+# have visibility limited to Tensorflow's implementation only.
+
+package(
+ default_visibility = [
+ "//tensorflow/c/experimental/saved_model/internal:__pkg__",
+ ],
+ licenses = ["notice"], # Apache 2.0
+)
+
+cc_library(
+ name = "concrete_function",
+ srcs = [
+ "concrete_function.cc",
+ ],
+ hdrs = [
+ "concrete_function.h",
+ ],
+ deps = [
+ ":function_metadata",
+ "//tensorflow/c/eager:operation_interface",
+ "//tensorflow/c/eager:tensor_handle_interface",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "function_metadata",
+ hdrs = [
+ "function_metadata.h",
+ ],
+)
+
+cc_library(
+ name = "saved_model_api",
+ hdrs = [
+ "saved_model_api.h",
+ ],
+ deps = [
+ ":concrete_function",
+ "//tensorflow/core:lib",
+ ],
+)
diff --git a/tensorflow/c/experimental/saved_model/core/concrete_function.cc b/tensorflow/c/experimental/saved_model/core/concrete_function.cc
new file mode 100644
index 0000000..d5da2ca
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/core/concrete_function.cc
@@ -0,0 +1,32 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+
+#include "tensorflow/c/eager/tensor_handle_interface.h"
+#include "tensorflow/c/experimental/saved_model/core/function_metadata.h"
+
+namespace tensorflow {
+
+const std::vector<tensorflow::AbstractTensorHandleInterface*>&
+ConcreteFunction::GetCaptures() const {
+ return captures_;
+}
+
+const FunctionMetadata& ConcreteFunction::GetFunctionMetadata() const {
+ return metadata_;
+}
+
+} // namespace tensorflow
diff --git a/tensorflow/c/experimental/saved_model/core/concrete_function.h b/tensorflow/c/experimental/saved_model/core/concrete_function.h
new file mode 100644
index 0000000..6f8a537
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/core/concrete_function.h
@@ -0,0 +1,55 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_CONCRETE_FUNCTION_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_CONCRETE_FUNCTION_H_
+
+#include <vector>
+
+#include "tensorflow/c/eager/operation_interface.h"
+#include "tensorflow/c/eager/tensor_handle_interface.h"
+#include "tensorflow/c/experimental/saved_model/core/function_metadata.h"
+#include "tensorflow/core/framework/function.pb.h"
+
+namespace tensorflow {
+
+// Note that ConcreteFunctions's lifetimes are effectively bound
+// to the SavedModel they are loaded from, since they retain pointers
+// to the TensorHandles owned by the SavedModel, and the FunctionDef
+// of the SavedModel.
+// Note(bmzhao): This class is only TEMPORARILY virtual, as a way to unblock
+// TFRT integration with TF Serving. Do not add more virtual implementations of
+// this class. Eventually we want to remove this virtual base class indirection
+// and have only a single implementation.
+class ConcreteFunction {
+ public:
+ virtual ~ConcreteFunction() = 0;
+
+ // This method returns the "Call" Op used to execute the function.
+ virtual AbstractOperationInterface* GetCallOp() = 0;
+
+ const std::vector<tensorflow::AbstractTensorHandleInterface*>& GetCaptures()
+ const;
+ const FunctionMetadata& GetFunctionMetadata() const;
+
+ private:
+ FunctionMetadata metadata_;
+ std::vector<tensorflow::AbstractTensorHandleInterface*> captures_;
+ FunctionDef* function_;
+};
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_CONCRETE_FUNCTION_H_
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops.i b/tensorflow/c/experimental/saved_model/core/function_metadata.h
similarity index 64%
copy from tensorflow/lite/experimental/kernels/hashtable_ops.i
copy to tensorflow/c/experimental/saved_model/core/function_metadata.h
index fa2e6fa..8499288 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops.i
+++ b/tensorflow/c/experimental/saved_model/core/function_metadata.h
@@ -13,8 +13,15 @@
limitations under the License.
==============================================================================*/
-%{
-#include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
-%}
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_FUNCTION_METADATA_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_FUNCTION_METADATA_H_
-%include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
+namespace tensorflow {
+
+class FunctionMetadata {
+ // TODO(bmzhao): Fill in with fields as necessary
+};
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_FUNCTION_METADATA_H_
diff --git a/tensorflow/c/experimental/saved_model/core/saved_model_api.h b/tensorflow/c/experimental/saved_model/core/saved_model_api.h
new file mode 100644
index 0000000..993ae93
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/core/saved_model_api.h
@@ -0,0 +1,55 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_SAVED_MODEL_API_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_SAVED_MODEL_API_H_
+
+#include <memory>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+#include "tensorflow/core/platform/status.h"
+
+namespace tensorflow {
+
+// Note(bmzhao): This class is only TEMPORARILY virtual, as a way to unblock
+// TFRT integration with TF Serving. Do not add more virtual implementations of
+// this class. Eventually we want to remove this virtual base class indirection
+// and have only a single implementation.
+class SavedModelAPI {
+ public:
+ // Retrieve a function from the TF2 SavedModel, using the "path" to a function
+ // in a TF2 savedmodel.
+ // Note: `function` is a double pointer, so that implementations are
+ // able to return a pointer to an internal member.
+ virtual Status GetFunction(const std::string& function_path,
+ ConcreteFunction** function) = 0;
+
+ // Retrieve a function from a SavedModel, using the key of the
+ // SignatureDef map:
+ // https://github.com/tensorflow/tensorflow/blob/69b08900b1e991d84bce31f3b404f5ed768f339f/tensorflow/core/protobuf/meta_graph.proto#L89
+ virtual Status GetSignatureDefFunction(const std::string& signature_def_key,
+ ConcreteFunction** function) = 0;
+
+ virtual const std::vector<ConcreteFunction*>& ListFunctions() = 0;
+
+ virtual ~SavedModelAPI() = default;
+};
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_CORE_SAVED_MODEL_API_H_
diff --git a/tensorflow/c/experimental/saved_model/internal/BUILD b/tensorflow/c/experimental/saved_model/internal/BUILD
new file mode 100644
index 0000000..1079804
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/BUILD
@@ -0,0 +1,157 @@
+# Experimental Implementation of SavedModel C APIs for TensorFlow. See RFC
+# https://github.com/tensorflow/community/pull/207
+# External clients should not worry about this directory; all contents are implementation details.
+# Code in this directory is intended to form the glue between the C API and the internal C++
+# implementation by
+# 1. mapping C API calls onto correponding methods of C++ objects
+# 2. mapping opaque C types onto C++ classes
+
+# Note(bmzhao): The *.cc files in this directory form the direct implementation of the
+# C API functions exposed in tf/c/experimental/saved_model/public/.
+
+# Note(bmzhao): All *type.h files in this directory are the internal definitions of
+# the opaque C types. These headers should only be visible to internal tensorflow
+# implementors.
+
+package(
+ licenses = ["notice"], # Apache 2.0
+)
+
+cc_library(
+ name = "conversion_macros",
+ hdrs = [
+ "conversion_macros.h",
+ ],
+)
+
+cc_library(
+ name = "concrete_function",
+ srcs = [
+ "concrete_function.cc",
+ ],
+ hdrs = [
+ "//tensorflow/c/experimental/saved_model/public:concrete_function.h",
+ ],
+ # TODO(bmzhao): Remove this as we refactor C API to granular targets,
+ # so that we can depend on c/eager/c_api_unified_experimental.h.
+ features = ["-layering_check"],
+ visibility = [
+ "//tensorflow/c/experimental/saved_model/public:__pkg__",
+ ],
+ deps = [
+ ":concrete_function_type",
+ ":function_metadata",
+ ":function_metadata_type",
+ "//tensorflow/c:c_api_macros",
+ "//tensorflow/c/eager:c_api",
+ "//tensorflow/c/eager:c_api_internal",
+ "//tensorflow/c/experimental/saved_model/core:concrete_function",
+ "//tensorflow/c/experimental/saved_model/core:function_metadata",
+ ],
+)
+
+cc_library(
+ name = "concrete_function_list",
+ srcs = [
+ "concrete_function_list.cc",
+ ],
+ hdrs = [
+ "//tensorflow/c/experimental/saved_model/public:concrete_function_list.h",
+ ],
+ visibility = [
+ "//tensorflow/c/experimental/saved_model/public:__pkg__",
+ ],
+ deps = [
+ ":concrete_function",
+ ":concrete_function_list_type",
+ ":concrete_function_type",
+ "//tensorflow/c:c_api_macros",
+ "//tensorflow/c/experimental/saved_model/core:concrete_function",
+ ],
+)
+
+cc_library(
+ name = "concrete_function_list_type",
+ hdrs = [
+ "concrete_function_list_type.h",
+ ],
+ deps = [
+ ":conversion_macros",
+ "//tensorflow/c/experimental/saved_model/core:concrete_function",
+ ],
+)
+
+cc_library(
+ name = "concrete_function_type",
+ hdrs = [
+ "concrete_function_type.h",
+ ],
+ deps = [
+ ":conversion_macros",
+ "//tensorflow/c/experimental/saved_model/core:concrete_function",
+ ],
+)
+
+cc_library(
+ name = "function_metadata",
+ srcs = [
+ "function_metadata.cc",
+ ],
+ hdrs = [
+ "//tensorflow/c/experimental/saved_model/public:function_metadata.h",
+ ],
+ visibility = [
+ "//tensorflow/c/experimental/saved_model/public:__pkg__",
+ ],
+ deps = [
+ ":function_metadata_type",
+ "//tensorflow/c:c_api_macros",
+ "//tensorflow/c/experimental/saved_model/core:function_metadata",
+ ],
+)
+
+cc_library(
+ name = "function_metadata_type",
+ hdrs = [
+ "function_metadata_type.h",
+ ],
+ deps = [
+ ":conversion_macros",
+ "//tensorflow/c/experimental/saved_model/core:function_metadata",
+ ],
+)
+
+cc_library(
+ name = "saved_model_api",
+ srcs = [
+ "saved_model_api.cc",
+ ],
+ hdrs = [
+ "//tensorflow/c/experimental/saved_model/public:saved_model_api.h",
+ ],
+ visibility = [
+ "//tensorflow/c/experimental/saved_model/public:__pkg__",
+ ],
+ deps = [
+ ":concrete_function",
+ ":concrete_function_list",
+ ":concrete_function_list_type",
+ ":concrete_function_type",
+ ":saved_model_api_type",
+ "//tensorflow/c:c_api_macros",
+ "//tensorflow/c:tf_status",
+ "//tensorflow/c:tf_status_internal",
+ "//tensorflow/c/experimental/saved_model/core:saved_model_api",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "saved_model_api_type",
+ hdrs = [
+ "saved_model_api_type.h",
+ ],
+ deps = [
+ "//tensorflow/c/experimental/saved_model/core:saved_model_api",
+ ],
+)
diff --git a/tensorflow/c/experimental/saved_model/internal/concrete_function.cc b/tensorflow/c/experimental/saved_model/internal/concrete_function.cc
new file mode 100644
index 0000000..da96490
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/concrete_function.cc
@@ -0,0 +1,40 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/c/experimental/saved_model/public/concrete_function.h"
+
+#include "tensorflow/c/eager/c_api_unified_experimental.h"
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/core/function_metadata.h"
+#include "tensorflow/c/experimental/saved_model/internal/concrete_function_type.h"
+#include "tensorflow/c/experimental/saved_model/internal/function_metadata_type.h"
+
+extern "C" {
+
+TF_FunctionMetadata* TF_ConcreteFunctionGetMetadata(TF_ConcreteFunction* func) {
+ return tensorflow::wrap(&tensorflow::unwrap(func)->GetFunctionMetadata());
+}
+
+TF_OutputList* TF_ConcreteFunctionGetCaptures(TF_ConcreteFunction* func) {
+ // TODO(bmzhao): Refactor TF_OutputList struct definition into a separate
+ // internal header, and implement this function.
+ return nullptr;
+}
+
+TFE_Op* TF_ConcreteFunctionGetCallOp(TF_ConcreteFunction* func) {
+ return new TFE_Op{tensorflow::unwrap(func)->GetCallOp()};
+}
+
+} // end extern "C"
diff --git a/tensorflow/c/experimental/saved_model/internal/concrete_function_list.cc b/tensorflow/c/experimental/saved_model/internal/concrete_function_list.cc
new file mode 100644
index 0000000..00ba314
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/concrete_function_list.cc
@@ -0,0 +1,33 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <stddef.h>
+
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/internal/concrete_function_list_type.h"
+#include "tensorflow/c/experimental/saved_model/internal/concrete_function_type.h"
+
+extern "C" {
+
+size_t TF_ConcreteFunctionListNumOutputs(TF_ConcreteFunctionList* list) {
+ return tensorflow::unwrap(list)->size();
+}
+
+TF_ConcreteFunction* TF_ConcreteFunctionListGet(TF_ConcreteFunctionList* list,
+ int i) {
+ return tensorflow::wrap((*tensorflow::unwrap(list))[i]);
+}
+
+} // end extern "C"
diff --git a/tensorflow/c/experimental/saved_model/internal/concrete_function_list_type.h b/tensorflow/c/experimental/saved_model/internal/concrete_function_list_type.h
new file mode 100644
index 0000000..5a89d7a
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/concrete_function_list_type.h
@@ -0,0 +1,36 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_LIST_TYPE_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_LIST_TYPE_H_
+
+#include <vector>
+
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/internal/conversion_macros.h"
+
+// Internal structures used by the SavedModel C API. These are likely to change
+// and should not be depended on.
+
+typedef struct TF_ConcreteFunctionList TF_ConcreteFunctionList;
+
+namespace tensorflow {
+
+DEFINE_CONVERSION_FUNCTIONS(std::vector<tensorflow::ConcreteFunction*>,
+ TF_ConcreteFunctionList)
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_LIST_TYPE_H_
diff --git a/tensorflow/c/experimental/saved_model/internal/concrete_function_type.h b/tensorflow/c/experimental/saved_model/internal/concrete_function_type.h
new file mode 100644
index 0000000..3797337
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/concrete_function_type.h
@@ -0,0 +1,36 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_TYPE_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_TYPE_H_
+
+#include "tensorflow/c/experimental/saved_model/core/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/internal/conversion_macros.h"
+
+// Internal structures used by the SavedModel C API. These are likely to change
+// and should not be depended on.
+
+// It doesn't make sense to wrap tensorflow::ConcreteFunction* in a separate
+// struct, since the lifetime of the struct and the raw pointer it wraps would
+// be different. Therefore TF_ConcreteFunction* = tensorflow::ConcreteFunction*.
+typedef struct TF_ConcreteFunction TF_ConcreteFunction;
+
+namespace tensorflow {
+
+DEFINE_CONVERSION_FUNCTIONS(tensorflow::ConcreteFunction, TF_ConcreteFunction)
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONCRETE_FUNCTION_TYPE_H_
diff --git a/tensorflow/c/experimental/saved_model/internal/conversion_macros.h b/tensorflow/c/experimental/saved_model/internal/conversion_macros.h
new file mode 100644
index 0000000..73875f0
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/conversion_macros.h
@@ -0,0 +1,28 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONVERSION_MACROS_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONVERSION_MACROS_H_
+
+#define DEFINE_CONVERSION_FUNCTIONS(cpp_impl, wrapper) \
+ inline cpp_impl *unwrap(wrapper *w) { \
+ return reinterpret_cast<cpp_impl *>(w); \
+ } \
+ \
+ inline wrapper *wrap(const cpp_impl *i) { \
+ return reinterpret_cast<wrapper *>(const_cast<cpp_impl *>(i)); \
+ }
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_CONVERSION_MACROS_H_
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops.i b/tensorflow/c/experimental/saved_model/internal/function_metadata.cc
similarity index 74%
rename from tensorflow/lite/experimental/kernels/hashtable_ops.i
rename to tensorflow/c/experimental/saved_model/internal/function_metadata.cc
index fa2e6fa..4cf31e1 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops.i
+++ b/tensorflow/c/experimental/saved_model/internal/function_metadata.cc
@@ -13,8 +13,8 @@
limitations under the License.
==============================================================================*/
-%{
-#include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
-%}
+#include "tensorflow/c/experimental/saved_model/public/function_metadata.h"
-%include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
+#include "tensorflow/c/experimental/saved_model/internal/function_metadata_type.h"
+
+// TODO(bmzhao): Add getter functions here as necessary.
diff --git a/tensorflow/c/experimental/saved_model/internal/function_metadata_type.h b/tensorflow/c/experimental/saved_model/internal/function_metadata_type.h
new file mode 100644
index 0000000..ab89cf2
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/function_metadata_type.h
@@ -0,0 +1,30 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_FUNCTION_METADATA_TYPE_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_FUNCTION_METADATA_TYPE_H_
+
+#include "tensorflow/c/experimental/saved_model/core/function_metadata.h"
+#include "tensorflow/c/experimental/saved_model/internal/conversion_macros.h"
+
+typedef struct TF_FunctionMetadata TF_FunctionMetadata;
+
+namespace tensorflow {
+
+DEFINE_CONVERSION_FUNCTIONS(tensorflow::FunctionMetadata, TF_FunctionMetadata)
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_FUNCTION_METADATA_TYPE_H_
diff --git a/tensorflow/c/experimental/saved_model/internal/saved_model_api.cc b/tensorflow/c/experimental/saved_model/internal/saved_model_api.cc
new file mode 100644
index 0000000..009d611
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/saved_model_api.cc
@@ -0,0 +1,67 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/c/experimental/saved_model/public/saved_model_api.h"
+
+#include "tensorflow/c/experimental/saved_model/core/saved_model_api.h"
+#include "tensorflow/c/experimental/saved_model/internal/concrete_function_list_type.h"
+#include "tensorflow/c/experimental/saved_model/internal/concrete_function_type.h"
+#include "tensorflow/c/experimental/saved_model/internal/saved_model_api_type.h"
+#include "tensorflow/c/tf_status.h"
+#include "tensorflow/c/tf_status_internal.h"
+#include "tensorflow/core/platform/status.h"
+
+extern "C" {
+
+TF_SavedModel* TF_LoadSavedModel(const char* dirname, TFE_Context* ctx,
+ const char* const* tags, int tags_len,
+ TF_Status* status) {
+ // TODO(bmzhao): Add a virtual "LoadSavedModel" method to
+ // AbstractContextInterface, and call it here.
+ return nullptr;
+}
+
+void TF_DeleteSavedModel(TF_SavedModel* model) { delete model; }
+
+TF_ConcreteFunction* TF_GetSavedModelConcreteFunction(TF_SavedModel* model,
+ char* function_path,
+ TF_Status* status) {
+ tensorflow::ConcreteFunction* result = nullptr;
+ tensorflow::Status get_function_status =
+ model->saved_model->GetFunction(function_path, &result);
+ status->status.Update(get_function_status);
+ if (!get_function_status.ok()) {
+ return nullptr;
+ }
+ return tensorflow::wrap(result);
+}
+
+TF_CAPI_EXPORT extern TF_ConcreteFunction* TF_GetSavedModelSignatureDefFunction(
+ TF_SavedModel* model, char* signature_def_key, TF_Status* status) {
+ tensorflow::ConcreteFunction* result = nullptr;
+ tensorflow::Status get_function_status =
+ model->saved_model->GetSignatureDefFunction(signature_def_key, &result);
+ status->status.Update(get_function_status);
+ if (!get_function_status.ok()) {
+ return nullptr;
+ }
+ return tensorflow::wrap(result);
+}
+
+TF_ConcreteFunctionList* TF_ListSavedModelFunctions(TF_SavedModel* model) {
+ return tensorflow::wrap(&model->saved_model->ListFunctions());
+}
+
+} // end extern "C"
diff --git a/tensorflow/c/experimental/saved_model/internal/saved_model_api_type.h b/tensorflow/c/experimental/saved_model/internal/saved_model_api_type.h
new file mode 100644
index 0000000..9e2d111
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/internal/saved_model_api_type.h
@@ -0,0 +1,30 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_SAVED_MODEL_API_TYPE_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_SAVED_MODEL_API_TYPE_H_
+
+#include <memory>
+
+#include "tensorflow/c/experimental/saved_model/core/saved_model_api.h"
+
+// Internal structures used by the SavedModel C API. These are likely to change
+// and should not be depended on.
+
+struct TF_SavedModel {
+ std::unique_ptr<tensorflow::SavedModelAPI> saved_model;
+};
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_INTERNAL_SAVED_MODEL_API_TYPE_H_
diff --git a/tensorflow/c/experimental/saved_model/public/BUILD b/tensorflow/c/experimental/saved_model/public/BUILD
new file mode 100644
index 0000000..af65e05
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/BUILD
@@ -0,0 +1,63 @@
+# Experimental SavedModel C APIs for TensorFlow.
+# See RFC https://github.com/tensorflow/community/pull/207
+# All headers are on the public surface of Tensorflow's C API.
+# Once moved out of experimental, these will be stable.
+# The idea behind a separate public/ directory is to make apparent
+# which headers are part of TF's public interface (and which headers)
+# are implementation details. This structure allows us to also perform future
+# programmatic checks that all "public" headers only include other "public"
+# headers.
+
+package(
+ # This is intentionally public
+ default_visibility = [
+ "//visibility:public",
+ ],
+ licenses = ["notice"], # Apache 2.0
+)
+
+# TODO(bmzhao): Remove these exports_files and rules, swap with cc_public_library instead.
+# cc_public_library would allows us to separate the header dep graph from header+srcs dep graph.
+exports_files(
+ [
+ "concrete_function.h",
+ "concrete_function_list.h",
+ "function_metadata.h",
+ "saved_model_api.h",
+ ],
+ visibility = ["//tensorflow/c/experimental/saved_model/internal:__pkg__"],
+)
+
+# The purpose of this header is to provide insulation against
+# future changes where we rename/move a public header, without
+# forcing all clients to change their "#includes".
+cc_library(
+ name = "c_saved_model_api",
+ hdrs = ["c_saved_model_api.h"],
+ deps = [
+ ":concrete_function",
+ ":concrete_function_list",
+ ":function_metadata",
+ ":saved_model_api",
+ ],
+)
+
+alias(
+ name = "concrete_function",
+ actual = "//tensorflow/c/experimental/saved_model/internal:concrete_function",
+)
+
+alias(
+ name = "concrete_function_list",
+ actual = "//tensorflow/c/experimental/saved_model/internal:concrete_function_list",
+)
+
+alias(
+ name = "function_metadata",
+ actual = "//tensorflow/c/experimental/saved_model/internal:function_metadata",
+)
+
+alias(
+ name = "saved_model_api",
+ actual = "//tensorflow/c/experimental/saved_model/internal:saved_model_api",
+)
diff --git a/tensorflow/c/experimental/saved_model/public/c_saved_model_api.h b/tensorflow/c/experimental/saved_model/public/c_saved_model_api.h
new file mode 100644
index 0000000..30f533f
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/c_saved_model_api.h
@@ -0,0 +1,26 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_C_SAVED_MODEL_API_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_C_SAVED_MODEL_API_H_
+
+// IWYU pragma: begin_exports
+#include "tensorflow/c/experimental/saved_model/public/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/public/concrete_function_list.h"
+#include "tensorflow/c/experimental/saved_model/public/function_metadata.h"
+#include "tensorflow/c/experimental/saved_model/public/saved_model_api.h"
+// IWYU pragma: end_exports
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_C_SAVED_MODEL_API_H_
diff --git a/tensorflow/c/experimental/saved_model/public/concrete_function.h b/tensorflow/c/experimental/saved_model/public/concrete_function.h
new file mode 100644
index 0000000..5baac90
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/concrete_function.h
@@ -0,0 +1,53 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_H_
+
+#include "tensorflow/c/c_api_macros.h"
+#include "tensorflow/c/eager/c_api_internal.h"
+#include "tensorflow/c/eager/c_api_unified_experimental.h"
+#include "tensorflow/c/experimental/saved_model/public/function_metadata.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif // __cplusplus
+
+// An opaque type that corresponds to a Function loaded from a SavedModel.
+// TODO(bmzhao): Work together w/srbs@ to make sure this composes w/the
+// C++ Unified Eager/Graph API's AbstractFunction
+typedef struct TF_ConcreteFunction TF_ConcreteFunction;
+
+// Returns FunctionMetadata associated with `func`. Metadata's lifetime is
+// bound to `func`, which is bound to the TF_SavedModel it was loaded from.
+TF_CAPI_EXPORT extern TF_FunctionMetadata* TF_ConcreteFunctionGetMetadata(
+ TF_ConcreteFunction* func);
+
+// Returns a list of TensorHandles implicitly captured by this function.
+TF_CAPI_EXPORT extern TF_OutputList* TF_ConcreteFunctionGetCaptures(
+ TF_ConcreteFunction* func);
+
+// Returns a TFE_Op suitable for executing this function.
+TF_CAPI_EXPORT extern TFE_Op* TF_ConcreteFunctionGetCallOp(
+ TF_ConcreteFunction* func);
+
+// Deletes `func`.
+TF_CAPI_EXPORT extern void TF_DeleteConcreteFunction(TF_ConcreteFunction* func);
+
+#ifdef __cplusplus
+} // end extern "C"
+#endif // __cplusplus
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_H_
diff --git a/tensorflow/c/experimental/saved_model/public/concrete_function_list.h b/tensorflow/c/experimental/saved_model/public/concrete_function_list.h
new file mode 100644
index 0000000..ab5ed35
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/concrete_function_list.h
@@ -0,0 +1,35 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_LIST_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_LIST_H_
+
+#include <stddef.h>
+
+#include "tensorflow/c/c_api_macros.h"
+#include "tensorflow/c/experimental/saved_model/public/concrete_function.h"
+
+// An opaque type that is acts like a list of TF_ConcreteFunction pointers.
+typedef struct TF_ConcreteFunctionList TF_ConcreteFunctionList;
+
+// Returns the size of `list`.
+TF_CAPI_EXPORT size_t
+TF_ConcreteFunctionListSize(TF_ConcreteFunctionList* list);
+
+// Returns the `i`th TF_ConcreteFunction in the list.
+TF_CAPI_EXPORT TF_ConcreteFunction* TF_ConcreteFunctionListGet(
+ TF_ConcreteFunctionList* list, int i);
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_CONCRETE_FUNCTION_LIST_H_
diff --git a/tensorflow/c/experimental/saved_model/public/function_metadata.h b/tensorflow/c/experimental/saved_model/public/function_metadata.h
new file mode 100644
index 0000000..83ca3c7
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/function_metadata.h
@@ -0,0 +1,35 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_FUNCTION_METADATA_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_FUNCTION_METADATA_H_
+
+#include "tensorflow/c/c_api_macros.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif // __cplusplus
+
+// An opaque type used to store any metadata associated with a function.
+typedef struct TF_FunctionMetadata TF_FunctionMetadata;
+
+// TODO(bmzhao): Add getters for fields as we determine what metadata
+// we want to expose.
+
+#ifdef __cplusplus
+} // end extern "C"
+#endif // __cplusplus
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_FUNCTION_METADATA_H_
diff --git a/tensorflow/c/experimental/saved_model/public/saved_model_api.h b/tensorflow/c/experimental/saved_model/public/saved_model_api.h
new file mode 100644
index 0000000..fa0976d
--- /dev/null
+++ b/tensorflow/c/experimental/saved_model/public/saved_model_api.h
@@ -0,0 +1,96 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_SAVED_MODEL_API_H_
+#define TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_SAVED_MODEL_API_H_
+
+#include "tensorflow/c/c_api_macros.h"
+#include "tensorflow/c/experimental/saved_model/public/concrete_function.h"
+#include "tensorflow/c/experimental/saved_model/public/concrete_function_list.h"
+#include "tensorflow/c/tf_status.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif // __cplusplus
+
+// An opaque type representing a Tensorflow "SavedModel"
+// (https://www.tensorflow.org/guide/saved_model) that we always pass by pointer
+// to achieve ABI stability.
+typedef struct TF_SavedModel TF_SavedModel;
+
+// Load a SavedModel from `dirname`.
+//
+// Params:
+// dirname - A directory filepath that the SavedModel is at.
+// ctx - A TFE_Context containing optional load/TF runtime options.
+// `ctx` must outlive the returned TF_SavedModel pointer.
+// tags - Pointer to char* array of SavedModel tags. Optional if the SavedModel
+// contains a single Metagraph, as for those exported from
+// `tf.saved_model.save`.
+// tags_len - number of elements in the `tags` array.
+// status - Set to OK on success and an appropriate error on failure.
+// Returns:
+// If status is not OK, returns nullptr. Otherwise, returns a newly created
+// TF_SavedModel instance. It must be deleted by calling TF_DeleteSavedModel.
+TF_CAPI_EXPORT extern TF_SavedModel* TF_LoadSavedModel(const char* dirname,
+ TFE_Context* ctx,
+ const char* const* tags,
+ int tags_len,
+ TF_Status* status);
+
+// Deletes a TF_SavedModel, and frees any resources owned by it.
+TF_CAPI_EXPORT extern void TF_DeleteSavedModel(TF_SavedModel* model);
+
+// Retrieve a function from the TF2 SavedModel via function path.
+//
+// Params:
+// model - The TF2 SavedModel to load a function from.
+// function_path - A string containing the path from the root saved python
+// object to a tf.function method.
+// TODO(bmzhao): Add a detailed example of this with a
+// python tf.module before moving this out of experimental.
+// status - Set to OK on success and an appropriate error on failure.
+// Returns:
+// If status is not OK, returns nullptr. Otherwise, returns a
+// TF_ConcreteFunction instance. The lifetime of this instance is
+// "conceptually" bound to `model`. Once `model` is deleted, all
+// `TF_ConcreteFunctions` retrieved from it are invalid, and have been deleted.
+TF_CAPI_EXPORT extern TF_ConcreteFunction* TF_GetSavedModelConcreteFunction(
+ TF_SavedModel* model, char* function_path, TF_Status* status);
+
+// Retrieve a function from the TF SavedModel via a SignatureDef key.
+//
+// Params:
+// model - The SavedModel to load a function from.
+// signature_def_key - The string key of the SignatureDef map of a SavedModel:
+// https://github.com/tensorflow/tensorflow/blob/69b08900b1e991d84bce31f3b404f5ed768f339f/tensorflow/core/protobuf/meta_graph.proto#L89
+// status - Set to OK on success and an appropriate error on failure.
+// Returns:
+// If status is not OK, returns nullptr. Otherwise, returns a
+// TF_ConcreteFunction instance. Once `model` is deleted, all
+// `TF_ConcreteFunctions` retrieved from it are invalid, and have been deleted.
+TF_CAPI_EXPORT extern TF_ConcreteFunction* TF_GetSavedModelSignatureDefFunction(
+ TF_SavedModel* model, char* signature_def_key, TF_Status* status);
+
+// Returns a list of all ConcreteFunctions stored in this SavedModel.
+// The lifetime of the returned list is bound to `model`.
+TF_CAPI_EXPORT extern TF_ConcreteFunctionList* TF_ListSavedModelFunctions(
+ TF_SavedModel* model);
+
+#ifdef __cplusplus
+} // end extern "C"
+#endif // __cplusplus
+
+#endif // TENSORFLOW_C_EXPERIMENTAL_SAVED_MODEL_PUBLIC_SAVED_MODEL_API_H_
diff --git a/tensorflow/compiler/aot/benchmark.h b/tensorflow/compiler/aot/benchmark.h
index 266b7fe..95bb766 100644
--- a/tensorflow/compiler/aot/benchmark.h
+++ b/tensorflow/compiler/aot/benchmark.h
@@ -38,7 +38,7 @@
struct Options {
// kDefaultMicros specifies the default time to run the benchmark, and is used
// if neither max_iters nor max_micros is set.
- static const int64 kDefaultMicros = 3000000;
+ static constexpr int64 kDefaultMicros = 3000000;
int64 max_iters = 0; // Maximum iterations to run, ignored if <= 0.
int64 max_micros = 0; // Maximum microseconds to run, ignored if <= 0.
diff --git a/tensorflow/compiler/aot/tfcompile.bzl b/tensorflow/compiler/aot/tfcompile.bzl
index 35a054a..abccefb 100644
--- a/tensorflow/compiler/aot/tfcompile.bzl
+++ b/tensorflow/compiler/aot/tfcompile.bzl
@@ -38,6 +38,7 @@
tfcompile_tool = "//tensorflow/compiler/aot:tfcompile",
include_standard_runtime_deps = True,
enable_xla_hlo_profiling = False,
+ enable_tracemes = False,
mlir_components = "None",
deps = None,
tags = []):
@@ -89,6 +90,9 @@
enable_xla_hlo_profiling: Enable XLA HLO profiling in the generated
program, and emit metadata that lets us pretty-print the gathered
profile counters.
+ enable_tracemes: Tell tfcompile to generate calls to
+ TraceMe::Activity{Start|End} around HLO instructions that can be used by
+ Xprof to construct profiler timelines.
mlir_components: When the value is "None", no components use MLIR. When
the value is "Bridge", use MLIR to translate GraphDef to HLO.
deps: a list of deps to include on the build rules for the generated
@@ -190,6 +194,11 @@
else:
profiling_flag = ""
+ if enable_tracemes:
+ traceme_flag = "--xla_cpu_enable_xprof_traceme=true"
+ else:
+ traceme_flag = "--xla_cpu_enable_xprof_traceme=false"
+
mlir_flag = "--mlir_components=" + mlir_components
srcs = [tfcompile_graph, config]
@@ -218,7 +227,7 @@
" --out_header=$(@D)/" + header_file +
" --out_metadata_object=$(@D)/" + metadata_object_file +
" --out_function_object=$(@D)/" + function_object_file +
- " " + flags + " " + profiling_flag + " " + mlir_flag
+ " " + flags + " " + profiling_flag + " " + mlir_flag + " " + traceme_flag
),
tools = [tfcompile_tool],
visibility = visibility,
diff --git a/tensorflow/compiler/jit/xla_launch_util.cc b/tensorflow/compiler/jit/xla_launch_util.cc
index 80bb9cd..e0ec990 100644
--- a/tensorflow/compiler/jit/xla_launch_util.cc
+++ b/tensorflow/compiler/jit/xla_launch_util.cc
@@ -479,12 +479,9 @@
input_output_alias, output_num, ctx, i, shape, &output,
definition_event, stream, use_multiple_streams_));
} else {
- auto program_shape =
- kernel->computation->GetProgramShape().ValueOrDie();
- if (program_shape.result().IsTuple() &&
- program_shape.result().tuple_shapes(output_num).IsTuple()) {
+ if (type == DT_VARIANT) {
return errors::Unimplemented(
- "Support for TensorList or Stack crossing the XLA/TF boundary "
+ "Support for TensorList crossing the XLA/TF boundary "
"is not implemented");
}
diff --git a/tensorflow/compiler/mlir/BUILD b/tensorflow/compiler/mlir/BUILD
index bc4094b..a904af6 100644
--- a/tensorflow/compiler/mlir/BUILD
+++ b/tensorflow/compiler/mlir/BUILD
@@ -48,7 +48,6 @@
"@llvm-project//mlir:MlirOptLib",
"@llvm-project//mlir:Pass",
"@llvm-project//mlir:Support",
- "@llvm-project//mlir/test:TestTransforms",
],
)
diff --git a/tensorflow/compiler/mlir/lite/BUILD b/tensorflow/compiler/mlir/lite/BUILD
index 6de717b..ce0b9c0 100644
--- a/tensorflow/compiler/mlir/lite/BUILD
+++ b/tensorflow/compiler/mlir/lite/BUILD
@@ -296,11 +296,9 @@
name = "tensorflow_lite_legalize_tf",
srcs = [
"transforms/dilated_conv.cc",
- "transforms/extract_ophint.cc",
"transforms/generated_legalize_tf.inc",
"transforms/generated_lower_static_tensor_list.inc",
"transforms/generated_prepare_tf.inc",
- "transforms/legalize_ophint_func_op.cc",
"transforms/legalize_tf.cc",
"transforms/legalize_tf_while.cc",
"transforms/lower_static_tensor_list.cc",
diff --git a/tensorflow/compiler/mlir/lite/experimental/estimators/arithmetic_count_util.h b/tensorflow/compiler/mlir/lite/experimental/estimators/arithmetic_count_util.h
index 0c50efb..2ca49e4 100644
--- a/tensorflow/compiler/mlir/lite/experimental/estimators/arithmetic_count_util.h
+++ b/tensorflow/compiler/mlir/lite/experimental/estimators/arithmetic_count_util.h
@@ -40,6 +40,37 @@
*count = total_count;
return true;
}
+
+ // For conv2d/depthwise_conv/fully_connected ops.
+ // This algorithm actually comes from TOCO tooling_util.cc
+ static bool GetArithmeticCountForConvAndFullyconnectedOp(Operation* op,
+ int64_t* count) {
+ auto weight = op->getOperand(1);
+ auto weight_type = weight.getType().dyn_cast_or_null<RankedTensorType>();
+ if (weight_type == nullptr || !weight_type.hasStaticShape()) return false;
+
+ auto output = op->getResult(0);
+ auto output_type = output.getType().dyn_cast_or_null<RankedTensorType>();
+ if (output_type == nullptr || !output_type.hasStaticShape()) return false;
+
+ int64_t cols = 1;
+ for (int i = 0; i < output_type.getRank() - 1; ++i) {
+ cols *= output_type.getDimSize(i);
+ }
+ const int64_t cost_per_col = 2 * weight_type.getNumElements();
+
+ *count = 2 * cost_per_col * cols;
+
+ auto bias = op->getOperand(2);
+ if (bias) {
+ auto bias_type = bias.getType().dyn_cast_or_null<RankedTensorType>();
+ if (bias_type && bias_type.hasStaticShape()) {
+ *count += bias_type.getNumElements();
+ }
+ }
+
+ return true;
+ }
};
#endif // TENSORFLOW_COMPILER_MLIR_LITE_EXPERIMENTAL_ESTIMATORS_ARITHMETIC_COUNT_UTIL_H_
diff --git a/tensorflow/compiler/mlir/lite/experimental/estimators/cpu_estimators.h b/tensorflow/compiler/mlir/lite/experimental/estimators/cpu_estimators.h
index 4a7415f..b47c08c 100644
--- a/tensorflow/compiler/mlir/lite/experimental/estimators/cpu_estimators.h
+++ b/tensorflow/compiler/mlir/lite/experimental/estimators/cpu_estimators.h
@@ -41,6 +41,68 @@
static bool IsSupported(mlir::Operation* op) { return true; }
};
+// tfl.concatenation
+template <>
+class TFLiteCostEstimator<ConcatenationOp, hardware::CPU> {
+ public:
+ static double GetCost(mlir::Operation* op) {
+ int64_t count;
+ if (ArithmeticCountUtilHelper::GetInputTensorTotalSize(op, &count))
+ return kCPUCopyUnitCost * count;
+ return kCPUDefaultFixedValuedCost;
+ }
+
+ static bool IsSupported(mlir::Operation* op) { return true; }
+};
+
+// tfl.conv_2d
+template <>
+class TFLiteCostEstimator<Conv2DOp, hardware::CPU> {
+ public:
+ static double GetCost(mlir::Operation* op) {
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kCPUArithmeticUnitCost;
+ }
+ return kCPUDefaultFixedValuedCost;
+ }
+
+ static bool IsSupported(mlir::Operation* op) { return true; }
+};
+
+// tfl.depthwise_conv_2d
+template <>
+class TFLiteCostEstimator<DepthwiseConv2DOp, hardware::CPU> {
+ public:
+ static double GetCost(mlir::Operation* op) {
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kCPUArithmeticUnitCost;
+ }
+ return kCPUDefaultFixedValuedCost;
+ }
+
+ static bool IsSupported(mlir::Operation* op) { return true; }
+};
+
+// tfl.fully_connected
+template <>
+class TFLiteCostEstimator<FullyConnectedOp, hardware::CPU> {
+ public:
+ static double GetCost(mlir::Operation* op) {
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kCPUArithmeticUnitCost;
+ }
+ return kCPUDefaultFixedValuedCost;
+ }
+
+ static bool IsSupported(mlir::Operation* op) { return true; }
+};
+
// tfl.mul
template <>
class TFLiteCostEstimator<MulOp, hardware::CPU> {
@@ -56,21 +118,6 @@
static bool IsSupported(mlir::Operation* op) { return true; }
};
-// tfl.concatenation
-template <>
-class TFLiteCostEstimator<ConcatenationOp, hardware::CPU> {
- public:
- static double GetCost(mlir::Operation* op) {
- int64_t count;
- if (ArithmeticCountUtilHelper::GetInputTensorTotalSize(op, &count))
- return kCPUCopyUnitCost * count;
- return kCPUDefaultFixedValuedCost;
- }
-
- // TODO(renjieliu): We probably need to check for dynamic weights.
- static bool IsSupported(mlir::Operation* op) { return true; }
-};
-
// tfl.pack
template <>
class TFLiteCostEstimator<PackOp, hardware::CPU> {
@@ -82,7 +129,6 @@
return kCPUDefaultFixedValuedCost;
}
- // TODO(renjieliu): We probably need to check for dynamic weights.
static bool IsSupported(mlir::Operation* op) { return true; }
};
diff --git a/tensorflow/compiler/mlir/lite/experimental/estimators/gpu_estimators.h b/tensorflow/compiler/mlir/lite/experimental/estimators/gpu_estimators.h
index 9d32bf1..5bb7b0b 100644
--- a/tensorflow/compiler/mlir/lite/experimental/estimators/gpu_estimators.h
+++ b/tensorflow/compiler/mlir/lite/experimental/estimators/gpu_estimators.h
@@ -87,9 +87,12 @@
class TFLiteCostEstimator<Conv2DOp, hardware::GPU> {
public:
static double GetCost(mlir::Operation* op) {
- llvm::errs() << "No defined cost function for op: "
- << op->getName().getStringRef().str();
- return 0.0;
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kGPUArithmeticUnitCost;
+ }
+ return kGPUDefaultFixedValuedCost;
}
// TODO(renjieliu): We probably need to check for dynamic weights.
@@ -114,9 +117,12 @@
class TFLiteCostEstimator<DepthwiseConv2DOp, hardware::GPU> {
public:
static double GetCost(mlir::Operation* op) {
- llvm::errs() << "No defined cost function for op: "
- << op->getName().getStringRef().str();
- return 0.0;
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kGPUArithmeticUnitCost;
+ }
+ return kGPUDefaultFixedValuedCost;
}
static bool IsSupported(mlir::Operation* op) { return true; }
@@ -153,9 +159,12 @@
class TFLiteCostEstimator<FullyConnectedOp, hardware::GPU> {
public:
static double GetCost(mlir::Operation* op) {
- llvm::errs() << "No defined cost function for op: "
- << op->getName().getStringRef().str();
- return 0.0;
+ int64_t arithmetic_count;
+ if (ArithmeticCountUtilHelper::GetArithmeticCountForConvAndFullyconnectedOp(
+ op, &arithmetic_count)) {
+ return arithmetic_count * kGPUArithmeticUnitCost;
+ }
+ return kGPUDefaultFixedValuedCost;
}
// TODO(renjieliu): we need to check for dynamic weights.
diff --git a/tensorflow/compiler/mlir/lite/ir/tfl_ops.cc b/tensorflow/compiler/mlir/lite/ir/tfl_ops.cc
index 47a7b32d..1b25ac6 100644
--- a/tensorflow/compiler/mlir/lite/ir/tfl_ops.cc
+++ b/tensorflow/compiler/mlir/lite/ir/tfl_ops.cc
@@ -1014,6 +1014,75 @@
return success();
}
+TFL::ConstOp NarrowDownInt64InputValuesForOp(Operation *input_op,
+ RankedTensorType value_type,
+ Location loc, OpBuilder *builder) {
+ if (input_op == nullptr) return nullptr;
+
+ mlir::DenseIntElementsAttr attr;
+ if (!matchPattern(input_op, m_Constant(&attr))) {
+ return nullptr;
+ }
+
+ auto value_shape_type = mlir::RankedTensorType::get(
+ value_type.getShape(), builder->getIntegerType(32));
+
+ SmallVector<int32_t, 4> value_i32;
+ value_i32.reserve(value_type.getRank());
+ for (const auto &size : attr) {
+ value_i32.push_back(static_cast<int32_t>(size.getSExtValue()));
+ }
+ auto new_value_i32_attr =
+ mlir::DenseIntElementsAttr::get(value_shape_type, value_i32);
+
+ return builder->create<TFL::ConstOp>(loc, new_value_i32_attr);
+}
+
+// This will cast donw int64 values for TFL slice op.
+// This will require the begin & size are constants.
+struct CastDonwInt64BeginEndToInt32 : public OpRewritePattern<TFL::SliceOp> {
+ using OpRewritePattern<TFL::SliceOp>::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(TFL::SliceOp slice_op,
+ PatternRewriter &rewriter) const override {
+ auto begin = slice_op.begin();
+ auto size = slice_op.size();
+ auto begin_type = begin.getType().dyn_cast_or_null<RankedTensorType>();
+ auto size_type = size.getType().dyn_cast_or_null<RankedTensorType>();
+ auto begin_op = begin.getDefiningOp();
+ auto size_op = size.getDefiningOp();
+
+ if (begin_op == nullptr && size_op == nullptr) return failure();
+
+ if (begin_type == nullptr && size_type == nullptr) return failure();
+
+ // Handle begin.
+ if (begin_op && begin_type && begin_type.getElementType().isInteger(64)) {
+ auto new_begin = NarrowDownInt64InputValuesForOp(
+ begin_op, begin_type, slice_op.getLoc(), &rewriter);
+ if (new_begin != nullptr) {
+ slice_op.setOperand(1, new_begin);
+ }
+ }
+
+ // Handle size.
+ if (size_op && size_type && size_type.getElementType().isInteger(64)) {
+ auto new_size = NarrowDownInt64InputValuesForOp(
+ size_op, size_type, slice_op.getLoc(), &rewriter);
+ if (new_size != nullptr) {
+ slice_op.setOperand(2, new_size);
+ }
+ }
+
+ return success();
+ }
+};
+
+void SliceOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
+ MLIRContext *context) {
+ results.insert<CastDonwInt64BeginEndToInt32>(context);
+}
+
//===----------------------------------------------------------------------===//
// SubOp
//===----------------------------------------------------------------------===//
diff --git a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
index 314f6f2..2298285 100644
--- a/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
+++ b/tensorflow/compiler/mlir/lite/ir/tfl_ops.td
@@ -103,8 +103,8 @@
// Derived shape attribute class.
//===----------------------------------------------------------------------===//
class DerivedShapeAttr<code body> : DerivedAttr<"ArrayRef<int64_t>", body>;
-class DerivedTFLiteTypeAttr<code body> :
- DerivedAttr<"tflite::TensorType", body>;
+class DerivedTFLiteTypeAttr<code body, code convert> :
+ DerivedAttr<"tflite::TensorType", body, convert>;
// TFL Runtime op trait predicate.
class TFL_RuntimePredOpTrait<string desc, Pred pred> :
@@ -237,12 +237,29 @@
Or<[CPred<"getElementTypeOrSelf($_op.getOperand(" # j # ")).isa<mlir::TF::Quint" # num # "Type>()">,
CPred<"getElementTypeOrSelf($_op.getOperand(" # j # ")).isUnsignedInteger(" # num # ")">]>]>;
-class TFL_OperandHasRankLessThan<int n, int m> :
- PredOpTrait<"operand " # n # " is maximum " # m # "-D",
+class TFL_OperandHasRankLessThanOrEqualTo<int n, int m> :
+ PredOpTrait<"operand " # n # " is at most " # m # "-D",
Or<[TFL_OperandIsUnrankedPred<n>,
CPred<"$_op.getOperand(" # n #
").getType().cast<ShapedType>().getRank() <= " # m>]>>;
+class TFL_OperandHasRankGreaterThanOrEqualTo<int n, int m> :
+ PredOpTrait<"operand " # n # " is at least " # m # "-D",
+ Or<[TFL_OperandIsUnrankedPred<n>,
+ CPred<"$_op.getOperand(" # n #
+ ").getType().cast<ShapedType>().getRank() >= " # m>]>>;
+
+class TFL_OperandHasRankRange<int n, int x, int y> :
+ PredOpTrait<"operand " # n # " has rank range [" # x # ", " # y # "]",
+ Or<[TFL_OperandIsUnrankedPred<n>,
+ CPred<"$_op.getOperand(" # n # ").getType().cast<ShapedType>().getRank() "
+ ">= " # x # " && $_op.getOperand(" # n # ").getType().cast<ShapedType>()."
+ "getRank() <= " # y>]>>;
+
+def TFL_FloatNonNegative : AttrConstraint<
+ CPred<"!$_self.cast<FloatAttr>().getValue().isNegative()">,
+ "whose value is non-negative">;
+
// This is a quantization-aware version of TCresVTEtIsSameAsOp
class TFL_TCresVTEtIsSameAsOp<int i, int j> : And<[
TCOpResIsShapedTypePred<i, j>,
@@ -461,6 +478,7 @@
TFL_1DTensorOf<[I32]>:$output_shape,
TFL_TensorOf<[F32, TFL_Uint8, QI8, QUI8]>:$weights,
TFL_TensorOf<[F32, TFL_Uint8, QI8, QUI8]>:$input,
+ TFL_TensorOfOrNone<[F32, QI32, QUI32]>:$bias,
TFL_PaddingAttr:$padding,
I32Attr:$stride_h,
I32Attr:$stride_w
@@ -474,7 +492,12 @@
}
def TFL_Convolution2DTransposeBiasOp :
- Op<TFL_Dialect, "convolution_2d_transpose_bias", [NoSideEffect]> {
+ Op<TFL_Dialect, "convolution_2d_transpose_bias", [
+ NoSideEffect,
+ TFL_OperandHasRank<0, 4>,
+ TFL_OperandHasRank<1, 4>,
+ TFL_OperandHasRankLessThanOrEqualTo<2, 1>
+ ]> {
let summary = " Transpose convolution with bias operator";
let description = [{
@@ -489,15 +512,15 @@
}];
let arguments = (
- ins AnyTensor:$input,
- AnyTensor:$filter,
- TFL_TensorOfOrNone<[AnyType]>:$bias,
+ ins TFL_FpTensor:$input,
+ TFL_FpTensor:$filter,
+ TFL_TensorOfOrNone<[F32]>:$bias,
TFL_PaddingAttr:$padding,
- I32Attr:$stride_h,
- I32Attr:$stride_w
+ Confined<I32Attr, [IntPositive]>:$stride_h,
+ Confined<I32Attr, [IntPositive]>:$stride_w
);
- let results = (outs AnyTensor:$output);
+ let results = (outs TFL_FpTensor:$output);
}
def TFL_AveragePool2DOp:
@@ -549,6 +572,8 @@
return getResult().getType().cast<TensorType>().getElementType().
cast<IntegerType>().getWidth() > 32 ? tflite::TensorType_INT64 :
tflite::TensorType_INT32;
+ }], [{
+ TypeAttr::get(getResult().getType().cast<TensorType>().getElementType())
}]>;
}
@@ -577,6 +602,8 @@
return getResult().getType().cast<TensorType>().getElementType().
cast<IntegerType>().getWidth() > 32 ? tflite::TensorType_INT64 :
tflite::TensorType_INT32;
+ }], [{
+ TypeAttr::get(getResult().getType().cast<TensorType>().getElementType())
}]>;
}
@@ -608,14 +635,14 @@
let arguments = (
ins TFL_VariadicTensorOf<
- [F32, I64, I32, I16, I8, QI8, QUI8, QI16, TFL_Uint8]>:$values,
+ [F32, I64, I32, I16, I8, QI8, QUI8, TFL_Uint8]>:$values,
I32Attr:$axis,
TFL_AFAttr:$fused_activation_function
);
let results = (outs
TFL_TensorOf<
- [F32, I64, I32, I16, I8, QI8, QUI8, QI16, TFL_Uint8]>:$output
+ [F32, I64, I32, I16, I8, QI8, QUI8, TFL_Uint8]>:$output
);
let hasOptions = 1;
@@ -1175,7 +1202,9 @@
def TFL_EmbeddingLookupOp: TFL_Op<"embedding_lookup",
[NoSideEffect,
PredOpTrait<"value and output must have same element type",
- TCresVTEtIsSameAsOp<0, 1>>
+ TFL_TCresVTEtIsSameAsOp<0, 1>>,
+ TFL_OperandHasRank<0, 1>,
+ TFL_OperandHasRankGreaterThanOrEqualTo<1, 2>
]> {
let summary = "Embedding lookup operator";
@@ -1193,6 +1222,8 @@
def TFL_EqualOp: TFL_Op<"equal", [Commutative, ResultsBroadcastableShape,
NoQuantizableResult,
+ ResultsBroadcastableShape,
+ TFL_BinaryOperandsHaveSameShapesOrBroadcastableShape<0, 1, 4>,
PredOpTrait<"Operands have same value type", TCopVTEtIsSameAs<0, 1>>]> {
let summary = "Equal operator";
@@ -1202,8 +1233,8 @@
let arguments = (
ins
- TFL_TensorOf<[I1, F32, I32, I64, I8, QI8, QUI8, TFL_Uint8]>:$x,
- TFL_TensorOf<[I1, F32, I32, I64, I8, QI8, QUI8, TFL_Uint8]>:$y
+ TFL_TensorOf<[I1, F32, I32, I64, QI8, QUI8, TFL_Uint8, TFL_Str]>:$x,
+ TFL_TensorOf<[I1, F32, I32, I64, QI8, QUI8, TFL_Uint8, TFL_Str]>:$y
);
let results = (outs TFL_BoolTensor:$output);
@@ -1808,6 +1839,8 @@
);
let verifier = [{ return Verify(*this); }];
+
+ let hasCanonicalizer = 1;
}
def TFL_SumOp: TFL_Op<"sum", [NoSideEffect]> {
@@ -2734,7 +2767,10 @@
NoSideEffect,
SameOperandsAndResultsScale,
PredOpTrait<"input and output must have same element type",
- TCresVTEtIsSameAsOp<0, 0>>
+ TFL_TCresVTEtIsSameAsOp<0, 0>>,
+ TFL_OperandHasRankRange<0, 3, 4>,
+ TFL_OperandHasRank<1, 1>,
+ TFL_OperandHasRank<2, 2>
]> {
let summary = "BatchToSpaceNd operator";
@@ -2743,13 +2779,13 @@
}];
let arguments = (ins
- TFL_TensorOf<[F32, I8, I32, I64, QI8, QUI8]>:$input,
+ TFL_TensorOf<[F32, I8, I32, I64, UI8, QI8, QUI8]>:$input,
TFL_TensorOf<[I32]>:$block_shape,
TFL_TensorOf<[I32]>:$indices
);
let results = (outs
- TFL_TensorOf<[F32, I16, I32, I64, QI8, QUI8]>:$output
+ TFL_TensorOf<[F32, I16, I32, I64, UI8, QI8, QUI8]>:$output
);
}
@@ -2808,7 +2844,8 @@
NoSideEffect,
SameOperandsAndResultsScale,
PredOpTrait<"input and output must have same element type",
- TFL_TCresVTEtIsSameAsOp<0, 0>>
+ TFL_TCresVTEtIsSameAsOp<0, 0>>,
+ TFL_OperandHasRankLessThanOrEqualTo<0, 4>
]> {
let summary = "DepthToSpace operator";
@@ -2822,12 +2859,12 @@
}];
let arguments = (ins
- TFL_TensorOf<[F32, I8, I32, I64, TFL_Uint8, TFL_Quint8, QUI8]>:$input,
- I32Attr:$block_size
+ TFL_TensorOf<[F32, I8, I32, I64, TFL_Quint8, TFL_Uint8, UI8, QI8, QUI8]>:$input,
+ Confined<I32Attr, [IntPositive]>:$block_size
);
let results = (outs
- TFL_TensorOf<[F32, I8, I32, I64, TFL_Uint8, TFL_Quint8, QUI8]>:$output
+ TFL_TensorOf<[F32, I8, I32, I64, TFL_Quint8, TFL_Uint8, UI8, QI8, QUI8]>:$output
);
let hasOptions = 1;
@@ -3080,6 +3117,8 @@
return getResult(1).getType().cast<TensorType>().getElementType().
cast<IntegerType>().getWidth() > 32 ? tflite::TensorType_INT64 :
tflite::TensorType_INT32;
+ }], [{
+ TypeAttr::get(getResult(1).getType().cast<TensorType>().getElementType())
}]>;
let hasOptions = 1;
@@ -3096,9 +3135,9 @@
quantization parameters.
}];
- let arguments = (ins AnyTensor:$input);
+ let arguments = (ins TFL_TensorOf<[QI8, QUI8, QI16, F16]>:$input);
- let results = (outs AnyTensor:$output);
+ let results = (outs TFL_FpTensor:$output);
}
def TFL_FakeQuantOp : TFL_Op<"fake_quant", [NoSideEffect]> {
@@ -3210,9 +3249,9 @@
Converts sparse tensor to dense format.
}];
- let arguments = (ins AnyTensor:$input);
+ let arguments = (ins TFL_TensorOf<[F32, I8]>:$input);
- let results = (outs AnyTensor:$output);
+ let results = (outs TFL_TensorOf<[F32, I8]>:$output);
}
//===----------------------------------------------------------------------===//
@@ -3275,16 +3314,16 @@
}];
let arguments = (
- ins TFL_TensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$data_input,
- TFL_TensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$prev_activ_input,
- TFL_TensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$weights_input,
- TFL_TensorOf<[F32, QI32, QUI32]>:$biases_input,
- TFL_TensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$prev_state_input,
+ ins TFL_TensorOf<[F32, QUI8]>:$data_input,
+ TFL_TensorOf<[F32, QUI8]>:$prev_activ_input,
+ TFL_TensorOf<[F32, QUI8]>:$weights_input,
+ TFL_TensorOf<[F32, QI32]>:$biases_input,
+ TFL_TensorOf<[F32, QI16]>:$prev_state_input,
// Attributes
DefaultValuedAttr<TFL_AFAttr, "TANH">:$fused_activation_function,
- DefaultValuedAttr<F32Attr, "0.0f">:$cell_clip,
- DefaultValuedAttr<F32Attr, "0.0f">:$proj_clip,
+ Confined<DefaultValuedAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$cell_clip,
+ Confined<DefaultValuedAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$proj_clip,
// Since this op is the BASIC kernel only, constrain it.
Confined<
DefaultValuedAttr<TFL_LSTMKernelTypeAttr, "BASIC">,
@@ -3293,10 +3332,10 @@
let hasOptions = 1;
- let results = (outs TFL_2DTensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$activ_output,
- TFL_2DTensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$state_output,
- TFL_2DTensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$concat_temp,
- TFL_2DTensorOf<[F32, I8, QI8, QUI8, QI16, QUI16]>:$activ_temp);
+ let results = (outs TFL_2DTensorOf<[F32, QUI8]>:$activ_output,
+ TFL_2DTensorOf<[F32, QUI16]>:$state_output,
+ TFL_2DTensorOf<[F32, QUI8]>:$concat_temp,
+ TFL_2DTensorOf<[F32, QUI16]>:$activ_temp);
}
// This is the FULL kernel type LSTM op.
@@ -3526,6 +3565,41 @@
BidiLstmOptionalPeepholeWeightConstraint,
BidiLstmProjectionWeightBiasConstraint,
LstmResultConstraint,
+ TFL_OperandHasRank<0, 3>, // input
+ TFL_OperandHasRank<1, 2>, // fw_input_to_input_weights
+ TFL_OperandHasRank<2, 2>, // fw_input_to_forget_weights
+ TFL_OperandHasRank<3, 2>, // fw_input_to_cell_weights
+ TFL_OperandHasRank<4, 2>, // fw_input_to_output_weights
+ TFL_OperandHasRank<5, 2>, // fw_recurrent_to_input_weights
+ TFL_OperandHasRank<6, 2>, // fw_recurrent_to_forget_weights
+ TFL_OperandHasRank<7, 2>, // fw_recurrent_to_cell_weights
+ TFL_OperandHasRank<8, 2>, // fw_recurrent_to_output_weights
+ TFL_OperandHasRank<9, 1>, // fw_cell_to_input_weights
+ TFL_OperandHasRank<10, 1>, // fw_cell_to_forget_weights
+ TFL_OperandHasRank<11, 1>, // fw_cell_to_output_weights
+ TFL_OperandHasRank<12, 1>, // fw_input_gate_bias
+ TFL_OperandHasRank<13, 1>, // fw_forget_gate_bias
+ TFL_OperandHasRank<14, 1>, // fw_cell_bias
+ TFL_OperandHasRank<15, 1>, // fw_output_gate_bias
+ TFL_OperandHasRank<16, 2>, // fw_projection_weights
+ TFL_OperandHasRank<17, 1>, // fw_projection_bias
+ TFL_OperandHasRank<18, 2>, // bw_input_to_input_weights
+ TFL_OperandHasRank<19, 2>, // bw_input_to_forget_weights
+ TFL_OperandHasRank<20, 2>, // bw_input_to_cell_weights
+ TFL_OperandHasRank<21, 2>, // bw_input_to_output_weights
+ TFL_OperandHasRank<22, 2>, // bw_recurrent_to_input_weights
+ TFL_OperandHasRank<23, 2>, // bw_recurrent_to_forget_weights
+ TFL_OperandHasRank<24, 2>, // bw_recurrent_to_cell_weights
+ TFL_OperandHasRank<25, 2>, // bw_recurrent_to_output_weights
+ TFL_OperandHasRank<26, 1>, // bw_cell_to_input_weights
+ TFL_OperandHasRank<27, 1>, // bw_cell_to_forget_weights
+ TFL_OperandHasRank<28, 1>, // bw_cell_to_output_weights
+ TFL_OperandHasRank<29, 1>, // bw_input_gate_bias
+ TFL_OperandHasRank<30, 1>, // bw_forget_gate_bias
+ TFL_OperandHasRank<31, 1>, // bw_cell_bias
+ TFL_OperandHasRank<32, 1>, // bw_output_gate_bias
+ TFL_OperandHasRank<33, 2>, // bw_projection_weights
+ TFL_OperandHasRank<34, 1>, // bw_projection_bias
TFL_StatefulOp]> {
let summary = "Bidirectional sequence lstm operator";
@@ -3619,8 +3693,8 @@
// Attributes
TFL_AFAttr:$fused_activation_function,
- DefaultValuedAttr<F32Attr, "0.0f">:$cell_clip,
- DefaultValuedAttr<F32Attr, "0.0f">:$proj_clip,
+ Confined<DefaultValuedAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$cell_clip,
+ Confined<DefaultValuedAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$proj_clip,
BoolAttr:$merge_outputs,
BoolAttr:$time_major
);
diff --git a/tensorflow/compiler/mlir/lite/quantization/device_target.cc b/tensorflow/compiler/mlir/lite/quantization/device_target.cc
index b1d0b10..48c0345 100644
--- a/tensorflow/compiler/mlir/lite/quantization/device_target.cc
+++ b/tensorflow/compiler/mlir/lite/quantization/device_target.cc
@@ -113,7 +113,7 @@
if (!in_spec || !w_spec || !b_spec || !o_spec) return failure();
double scale_product = in_spec.getScale() * w_spec.getScale();
- if (fabs(scale_product - b_spec.getScale()) < 1e-6) return failure();
+ if (fabs(scale_product - b_spec.getScale()) >= 1e-6) return failure();
// input multipliers
input_multipliers->append(3, kUnitQuantizedMultiplier);
diff --git a/tensorflow/compiler/mlir/lite/tests/canonicalize.mlir b/tensorflow/compiler/mlir/lite/tests/canonicalize.mlir
index c94eb1b..1f067aa 100644
--- a/tensorflow/compiler/mlir/lite/tests/canonicalize.mlir
+++ b/tensorflow/compiler/mlir/lite/tests/canonicalize.mlir
@@ -98,3 +98,16 @@
// CHECK-NOT: pack
// CHECK: return %arg0, %[[UNPACK]]#0 : tensor<2x5xf32>, tensor<5xf32>
}
+
+// -----
+
+func @Int64SliceBeginSize(%arg0: tensor<4x128x32xf32>) -> tensor<1x128x32xf32> {
+ %0 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi64>} : () -> tensor<3xi64>
+ %1 = "tfl.pseudo_const"() {value = dense<[1, 128, 32]> : tensor<3xi64>} : () -> tensor<3xi64>
+ %2 = "tfl.slice"(%arg0, %0, %1) : (tensor<4x128x32xf32>, tensor<3xi64>, tensor<3xi64>) -> tensor<1x128x32xf32>
+ return %2 : tensor<1x128x32xf32>
+
+// CHECK: [[VAL_1:%.*]] = constant dense<0> : tensor<3xi32>
+// CHECK: [[VAL_2:%.*]] = constant dense<[1, 128, 32]> : tensor<3xi32>
+// CHECK: [[VAL_3:%.*]] = "tfl.slice"(%arg0, [[VAL_1]], [[VAL_2]]) : (tensor<4x128x32xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x128x32xf32>
+}
diff --git a/tensorflow/compiler/mlir/lite/tests/end2end/BUILD b/tensorflow/compiler/mlir/lite/tests/end2end/BUILD
index 9d768fe..cf58498 100644
--- a/tensorflow/compiler/mlir/lite/tests/end2end/BUILD
+++ b/tensorflow/compiler/mlir/lite/tests/end2end/BUILD
@@ -12,7 +12,6 @@
"add.pbtxt": ["no_rocm"],
"conv_2d.pbtxt": ["no_rocm"],
"fake_quant_per_channel.pbtxt": ["no_rocm"],
- "ophint_lstm.pbtxt": ["no_rocm"],
},
test_file_exts = [
"pbtxt",
diff --git a/tensorflow/compiler/mlir/lite/tests/end2end/ophint_lstm.pbtxt b/tensorflow/compiler/mlir/lite/tests/end2end/ophint_lstm.pbtxt
deleted file mode 100644
index 1b42b60..0000000
--- a/tensorflow/compiler/mlir/lite/tests/end2end/ophint_lstm.pbtxt
+++ /dev/null
@@ -1,7822 +0,0 @@
-# RUN: tf_tfl_translate -tf-input-arrays=INPUT -tf-input-shapes=1,3,3 -tf-input-data-types=DT_FLOAT -tf-output-arrays=OUTPUT %s -o - --output-mlir | FileCheck %s
-
-node {
- name: "INPUT"
- op: "Placeholder"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "shape"
- value {
- shape {
- dim {
- size: -1
- }
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- }
- }
-}
-node {
- name: "unstack"
- op: "Unpack"
- input: "INPUT"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "axis"
- value {
- i: 1
- }
- }
- attr {
- key: "num"
- value {
- i: 3
- }
- }
-}
-node {
- name: "rnn/Shape"
- op: "Shape"
- input: "unstack"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "out_type"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/strided_slice/stack"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/strided_slice/stack_1"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/strided_slice/stack_2"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/strided_slice"
- op: "StridedSlice"
- input: "rnn/Shape"
- input: "rnn/strided_slice/stack"
- input: "rnn/strided_slice/stack_1"
- input: "rnn/strided_slice/stack_2"
- attr {
- key: "Index"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "begin_mask"
- value {
- i: 0
- }
- }
- attr {
- key: "ellipsis_mask"
- value {
- i: 0
- }
- }
- attr {
- key: "end_mask"
- value {
- i: 0
- }
- }
- attr {
- key: "new_axis_mask"
- value {
- i: 0
- }
- }
- attr {
- key: "shrink_axis_mask"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/ExpandDims/dim"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/ExpandDims"
- op: "ExpandDims"
- input: "rnn/strided_slice"
- input: "rnn/TFLiteLSTMCellZeroState/ExpandDims/dim"
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tdim"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 3
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/concat/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/concat"
- op: "ConcatV2"
- input: "rnn/TFLiteLSTMCellZeroState/ExpandDims"
- input: "rnn/TFLiteLSTMCellZeroState/Const"
- input: "rnn/TFLiteLSTMCellZeroState/concat/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/zeros/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- }
- float_val: 0.0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/zeros"
- op: "Fill"
- input: "rnn/TFLiteLSTMCellZeroState/concat"
- input: "rnn/TFLiteLSTMCellZeroState/zeros/Const"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "index_type"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/ExpandDims_2/dim"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/ExpandDims_2"
- op: "ExpandDims"
- input: "rnn/strided_slice"
- input: "rnn/TFLiteLSTMCellZeroState/ExpandDims_2/dim"
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tdim"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/Const_2"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 3
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/concat_1/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/concat_1"
- op: "ConcatV2"
- input: "rnn/TFLiteLSTMCellZeroState/ExpandDims_2"
- input: "rnn/TFLiteLSTMCellZeroState/Const_2"
- input: "rnn/TFLiteLSTMCellZeroState/concat_1/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/zeros_1/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- }
- float_val: 0.0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState/zeros_1"
- op: "Fill"
- input: "rnn/TFLiteLSTMCellZeroState/concat_1"
- input: "rnn/TFLiteLSTMCellZeroState/zeros_1/Const"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "index_type"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims/dim"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims"
- op: "ExpandDims"
- input: "rnn/strided_slice"
- input: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims/dim"
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tdim"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 3
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/concat/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/concat"
- op: "ConcatV2"
- input: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims"
- input: "rnn/TFLiteLSTMCellZeroState_1/Const"
- input: "rnn/TFLiteLSTMCellZeroState_1/concat/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/zeros/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- }
- float_val: 0.0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/zeros"
- op: "Fill"
- input: "rnn/TFLiteLSTMCellZeroState_1/concat"
- input: "rnn/TFLiteLSTMCellZeroState_1/zeros/Const"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "index_type"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims_2/dim"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims_2"
- op: "ExpandDims"
- input: "rnn/strided_slice"
- input: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims_2/dim"
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tdim"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/Const_2"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- dim {
- size: 1
- }
- }
- int_val: 3
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/concat_1/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/concat_1"
- op: "ConcatV2"
- input: "rnn/TFLiteLSTMCellZeroState_1/ExpandDims_2"
- input: "rnn/TFLiteLSTMCellZeroState_1/Const_2"
- input: "rnn/TFLiteLSTMCellZeroState_1/concat_1/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/zeros_1/Const"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- }
- float_val: 0.0
- }
- }
- }
-}
-node {
- name: "rnn/TFLiteLSTMCellZeroState_1/zeros_1"
- op: "Fill"
- input: "rnn/TFLiteLSTMCellZeroState_1/concat_1"
- input: "rnn/TFLiteLSTMCellZeroState_1/zeros_1/Const"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "index_type"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_input_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "p\217k>@\254:\276\270W\264\276\014\033N\277p\226a\276\220d+\277\330\277\216>\240VN\276\010\253 \277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_input_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/input_to_input_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/input_to_input_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_input_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/input_to_input_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-None-input_to_input_w"
- op: "Identity"
- input: "rnn/rnn1/input_to_input_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_forget_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "4X\003?\304g1\277\374H\014?@\341\205=\314\264\023?\324{w?\000.V<PG}>\370Y\242>"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_forget_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/input_to_forget_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/input_to_forget_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_forget_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/input_to_forget_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- op: "Identity"
- input: "rnn/rnn1/input_to_forget_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_cell_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "p\205\r\276@\321\336\2750_\n\276H\256r?\340\017_\277\220\326J\277\2001\013=T\021\n\277\250\000d?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_cell_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/input_to_cell_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/input_to_cell_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_cell_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/input_to_cell_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- op: "Identity"
- input: "rnn/rnn1/input_to_cell_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 3
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_output_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "`\222T\276l\273A\277 oZ\277\310\335\211\276\300\310?=H\303\264\276\000\367\217\275@\203\224=DXQ\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_output_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/input_to_output_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/input_to_output_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_to_output_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/input_to_output_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-4-None-input_to_output_w"
- op: "Identity"
- input: "rnn/rnn1/input_to_output_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 4
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_input_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\310\326\374\27609\310\276\250\036\263\276\200\231\256\274L\362\016?\230\337\003\277\350\023\333>\324;\036?p\026@\276"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_input_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/cell_to_input_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/cell_to_input_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_input_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/cell_to_input_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- op: "Identity"
- input: "rnn/rnn1/cell_to_input_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 5
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_forget_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\210\334b?\024,\033\277\230\r\347\276\030\257\246>\364\0071?\020\036-\277\000\023a>LD ?\024\374\030\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_forget_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/cell_to_forget_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/cell_to_forget_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_forget_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/cell_to_forget_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- op: "Identity"
- input: "rnn/rnn1/cell_to_forget_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 6
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_cell_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\010\341\314\276P6=?p\253N>\364\266-?H;\244>\214*s?\\\307N\277HP\010\277 \226\027>"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_cell_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/cell_to_cell_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/cell_to_cell_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_cell_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/cell_to_cell_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- op: "Identity"
- input: "rnn/rnn1/cell_to_cell_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 7
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_output_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\350\177\343>\300\212\010\276x\357V?\340\r\344>t[\022\277X\330\021?\330\025\356> s}\277L\352!\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_output_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/cell_to_output_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/cell_to_output_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_to_output_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/cell_to_output_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- op: "Identity"
- input: "rnn/rnn1/cell_to_output_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 8
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/input_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/input_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/input_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/input_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-12-None-input_bias"
- op: "Identity"
- input: "rnn/rnn1/input_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 12
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/forget_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\200?\000\000\200?\000\000\200?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/forget_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/forget_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/forget_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/forget_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-13-None-forget_bias"
- op: "Identity"
- input: "rnn/rnn1/forget_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 13
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/cell_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/cell_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/cell_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-14-None-cell_bias"
- op: "Identity"
- input: "rnn/rnn1/cell_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 14
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/output_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/output_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/output_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/output_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/output_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-15-None-output_bias"
- op: "Identity"
- input: "rnn/rnn1/output_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 15
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/w_f_diag"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\020o/> \030\035\276\364|\027?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_f_diag/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/w_f_diag"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/w_f_diag"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_f_diag/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/w_f_diag/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-10-None-w_f_diag"
- op: "Identity"
- input: "rnn/rnn1/w_f_diag/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 10
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/w_i_diag"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\324\331+\277h\331\322>\250z\017?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_i_diag/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/w_i_diag"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/w_i_diag"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_i_diag/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/w_i_diag/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-9-None-w_i_diag"
- op: "Identity"
- input: "rnn/rnn1/w_i_diag/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 9
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn1/w_o_diag"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\230\316\316>\210\316a\277\210\373d\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_o_diag/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn1/w_o_diag"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn1/w_o_diag"
- }
- }
- }
-}
-node {
- name: "rnn/rnn1/w_o_diag/Read/Identity"
- op: "Identity"
- input: "rnn/rnn1/w_o_diag/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-11-None-w_o_diag"
- op: "Identity"
- input: "rnn/rnn1/w_o_diag/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 11
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_input_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\220\305\000\2760;\245>HV\372>P\356\270>\324u{?\010\265\345\276\370bw?\300[D\2770\212\344>"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_input_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/input_to_input_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/input_to_input_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_input_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/input_to_input_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-None-input_to_input_w"
- op: "Identity"
- input: "rnn/rnn2/input_to_input_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_forget_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\354\037d?\000\254\216\276\374\210w?\020;J\277\200bm=P\270^>\234\2702\277$\300{\277\370\231U\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_forget_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/input_to_forget_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/input_to_forget_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_forget_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/input_to_forget_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- op: "Identity"
- input: "rnn/rnn2/input_to_forget_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_cell_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: ",AH?\200\3616\275,7Y?\024@\024\277p\305\320\276\350\200\342>\000\236\271;\3500\031?T>!?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_cell_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/input_to_cell_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/input_to_cell_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_cell_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/input_to_cell_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- op: "Identity"
- input: "rnn/rnn2/input_to_cell_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 3
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_output_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "HO[\2770\355L\277@\2007?\324Q\t?$\251\n?@\221\266\276\370mK\277\240\356\014>\300\2440?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_output_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/input_to_output_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/input_to_output_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_to_output_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/input_to_output_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-4-None-input_to_output_w"
- op: "Identity"
- input: "rnn/rnn2/input_to_output_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 4
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_input_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\274;\002\277\250\302\026\277`\234\361>\220\r\002\277\000\255\200\274\334\332M\277t\225z\277\000(\322:\024\201z\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_input_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/cell_to_input_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/cell_to_input_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_input_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/cell_to_input_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- op: "Identity"
- input: "rnn/rnn2/cell_to_input_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 5
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_forget_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "03:>\014\273\035?\020\333+\276\334\371;?HVu?0\310`\27782\275>\304\020x\277,\212a\277"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_forget_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/cell_to_forget_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/cell_to_forget_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_forget_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/cell_to_forget_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- op: "Identity"
- input: "rnn/rnn2/cell_to_forget_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 6
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_cell_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\244\251o\277\230xo\277\340\222\223>\2409y\276|\327 \277pA\364\276\200\325\003\277\300Lg\277\274=,?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_cell_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/cell_to_cell_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/cell_to_cell_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_cell_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/cell_to_cell_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- op: "Identity"
- input: "rnn/rnn2/cell_to_cell_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 7
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_output_w"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- dim {
- size: 3
- }
- }
- tensor_content: "\274\345\035\277`\202d?\364\333+?8\246W\2778X\267\276\024ER?4TJ?\254T6? g\215="
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_output_w/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/cell_to_output_w"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/cell_to_output_w"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_to_output_w/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/cell_to_output_w/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- op: "Identity"
- input: "rnn/rnn2/cell_to_output_w/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 8
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/input_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/input_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/input_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/input_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-12-None-input_bias"
- op: "Identity"
- input: "rnn/rnn2/input_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 12
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/forget_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\200?\000\000\200?\000\000\200?"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/forget_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/forget_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/forget_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/forget_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-13-None-forget_bias"
- op: "Identity"
- input: "rnn/rnn2/forget_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 13
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/cell_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/cell_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/cell_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-14-None-cell_bias"
- op: "Identity"
- input: "rnn/rnn2/cell_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 14
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/rnn2/output_bias"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_FLOAT
- tensor_shape {
- dim {
- size: 3
- }
- }
- tensor_content: "\000\000\000\000\000\000\000\000\000\000\000\000"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/output_bias/Read/ReadVariableOp"
- op: "Identity"
- input: "rnn/rnn2/output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_class"
- value {
- list {
- s: "loc:@rnn/rnn2/output_bias"
- }
- }
- }
-}
-node {
- name: "rnn/rnn2/output_bias/Read/Identity"
- op: "Identity"
- input: "rnn/rnn2/output_bias/Read/ReadVariableOp"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-15-None-output_bias"
- op: "Identity"
- input: "rnn/rnn2/output_bias/Read/Identity"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 15
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-0-input"
- op: "Identity"
- input: "unstack"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-0-c_prev"
- op: "Identity"
- input: "rnn/TFLiteLSTMCellZeroState/zeros"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-0-m_prev"
- op: "Identity"
- input: "rnn/TFLiteLSTMCellZeroState/zeros_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-0-input"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-0-m_prev"
- input: "rnn/stacked_rnn_cells/concat/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_1/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_1"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells/concat_1/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat"
- input: "rnn/stacked_rnn_cells/concat_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_2/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_2"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells/concat_2/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_1"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat"
- input: "rnn/stacked_rnn_cells/concat_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_1"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_1"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_3/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_3"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells/concat_3/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_2"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat"
- input: "rnn/stacked_rnn_cells/concat_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_2"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_4/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_4"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells/concat_4/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_3"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat"
- input: "rnn/stacked_rnn_cells/concat_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_3"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_3"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-10-None-w_f_diag"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-0-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/add"
- op: "Add"
- input: "rnn/stacked_rnn_cells/BiasAdd_1"
- input: "rnn/stacked_rnn_cells/mul"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/add"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_1"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-0-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_2"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-9-None-w_i_diag"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-0-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/add_1"
- op: "Add"
- input: "rnn/stacked_rnn_cells/BiasAdd"
- input: "rnn/stacked_rnn_cells/mul_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid_1"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/add_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Tanh"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells/BiasAdd_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_3"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid_1"
- input: "rnn/stacked_rnn_cells/Tanh"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/add_2"
- op: "Add"
- input: "rnn/stacked_rnn_cells/mul_1"
- input: "rnn/stacked_rnn_cells/mul_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_4"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-11-None-w_o_diag"
- input: "rnn/stacked_rnn_cells/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/add_3"
- op: "Add"
- input: "rnn/stacked_rnn_cells/BiasAdd_2"
- input: "rnn/stacked_rnn_cells/mul_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid_2"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/add_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Tanh_1"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_5"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid_2"
- input: "rnn/stacked_rnn_cells/Tanh_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-0-c"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "last"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-0-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/mul_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-0-input"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-0-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-0-c_prev"
- op: "Identity"
- input: "rnn/TFLiteLSTMCellZeroState_1/zeros"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-0-m_prev"
- op: "Identity"
- input: "rnn/TFLiteLSTMCellZeroState_1/zeros_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_5/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_5"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-0-input"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-0-m_prev"
- input: "rnn/stacked_rnn_cells/concat_5/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_6/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_6"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells/concat_6/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_4"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat_5"
- input: "rnn/stacked_rnn_cells/concat_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_4"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_4"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_7/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_7"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells/concat_7/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_5"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat_5"
- input: "rnn/stacked_rnn_cells/concat_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_5"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_5"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_8/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_8"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells/concat_8/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_6"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat_5"
- input: "rnn/stacked_rnn_cells/concat_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_6"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_6"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_9/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/concat_9"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells/concat_9/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/MatMul_7"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells/concat_5"
- input: "rnn/stacked_rnn_cells/concat_9"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/BiasAdd_7"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells/MatMul_7"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid_3"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/BiasAdd_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_6"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid_3"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-0-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid_4"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/BiasAdd_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Tanh_2"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells/BiasAdd_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_7"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid_4"
- input: "rnn/stacked_rnn_cells/Tanh_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/add_4"
- op: "Add"
- input: "rnn/stacked_rnn_cells/mul_6"
- input: "rnn/stacked_rnn_cells/mul_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Sigmoid_5"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells/BiasAdd_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/Tanh_3"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells/add_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/mul_8"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/Sigmoid_5"
- input: "rnn/stacked_rnn_cells/Tanh_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-0-c"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/add_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "last"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-0-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/mul_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-1-input"
- op: "Identity"
- input: "unstack:1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-1-c_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-0-c"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-1-m_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-0-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-1-input"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-1-m_prev"
- input: "rnn/stacked_rnn_cells_1/concat/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_1/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_1"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells_1/concat_1/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat"
- input: "rnn/stacked_rnn_cells_1/concat_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_2/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_2"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells_1/concat_2/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_1"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat"
- input: "rnn/stacked_rnn_cells_1/concat_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_1"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_1"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_3/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_3"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells_1/concat_3/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_2"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat"
- input: "rnn/stacked_rnn_cells_1/concat_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_2"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_4/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_4"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells_1/concat_4/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_3"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat"
- input: "rnn/stacked_rnn_cells_1/concat_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_3"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_3"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-10-None-w_f_diag"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-1-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/add"
- op: "Add"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_1"
- input: "rnn/stacked_rnn_cells_1/mul"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/add"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_1"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-1-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_2"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-9-None-w_i_diag"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-1-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/add_1"
- op: "Add"
- input: "rnn/stacked_rnn_cells_1/BiasAdd"
- input: "rnn/stacked_rnn_cells_1/mul_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid_1"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/add_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Tanh"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_3"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid_1"
- input: "rnn/stacked_rnn_cells_1/Tanh"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/add_2"
- op: "Add"
- input: "rnn/stacked_rnn_cells_1/mul_1"
- input: "rnn/stacked_rnn_cells_1/mul_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_4"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-11-None-w_o_diag"
- input: "rnn/stacked_rnn_cells_1/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/add_3"
- op: "Add"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_2"
- input: "rnn/stacked_rnn_cells_1/mul_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid_2"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/add_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Tanh_1"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_1/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_5"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid_2"
- input: "rnn/stacked_rnn_cells_1/Tanh_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-1-c"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "last"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-1-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/mul_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-1-input"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-1-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-1-c_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-0-c"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-1-m_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-0-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_5/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_5"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-1-input"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-1-m_prev"
- input: "rnn/stacked_rnn_cells_1/concat_5/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_6/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_6"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells_1/concat_6/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_4"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat_5"
- input: "rnn/stacked_rnn_cells_1/concat_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_4"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_4"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_7/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_7"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells_1/concat_7/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_5"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat_5"
- input: "rnn/stacked_rnn_cells_1/concat_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_5"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_5"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_8/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_8"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells_1/concat_8/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_6"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat_5"
- input: "rnn/stacked_rnn_cells_1/concat_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_6"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_6"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_9/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/concat_9"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells_1/concat_9/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/MatMul_7"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_1/concat_5"
- input: "rnn/stacked_rnn_cells_1/concat_9"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/BiasAdd_7"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_1/MatMul_7"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid_3"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_6"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid_3"
- input: "rnn/stacked_rnn_cells_1/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-1-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid_4"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Tanh_2"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_7"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid_4"
- input: "rnn/stacked_rnn_cells_1/Tanh_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/add_4"
- op: "Add"
- input: "rnn/stacked_rnn_cells_1/mul_6"
- input: "rnn/stacked_rnn_cells_1/mul_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Sigmoid_5"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_1/BiasAdd_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/Tanh_3"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_1/add_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/mul_8"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_1/Sigmoid_5"
- input: "rnn/stacked_rnn_cells_1/Tanh_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-1-c"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/add_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "last"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-1-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/mul_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 1
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-2-input"
- op: "Identity"
- input: "unstack:2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-2-c_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-1-c"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-2-m_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-1-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-0-2-input"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-18-2-m_prev"
- input: "rnn/stacked_rnn_cells_2/concat/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_1/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_1"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells_2/concat_1/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat"
- input: "rnn/stacked_rnn_cells_2/concat_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_2/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_2"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells_2/concat_2/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_1"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat"
- input: "rnn/stacked_rnn_cells_2/concat_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_1"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_1"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_3/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_3"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells_2/concat_3/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_2"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat"
- input: "rnn/stacked_rnn_cells_2/concat_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_2"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_4/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_4"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells_2/concat_4/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_3"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat"
- input: "rnn/stacked_rnn_cells_2/concat_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_3"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_3"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-10-None-w_f_diag"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-2-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/add"
- op: "Add"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_1"
- input: "rnn/stacked_rnn_cells_2/mul"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/add"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_1"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-2-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_2"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-9-None-w_i_diag"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-19-2-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/add_1"
- op: "Add"
- input: "rnn/stacked_rnn_cells_2/BiasAdd"
- input: "rnn/stacked_rnn_cells_2/mul_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid_1"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/add_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Tanh"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_3"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid_1"
- input: "rnn/stacked_rnn_cells_2/Tanh"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/add_2"
- op: "Add"
- input: "rnn/stacked_rnn_cells_2/mul_1"
- input: "rnn/stacked_rnn_cells_2/mul_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_4"
- op: "Mul"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-11-None-w_o_diag"
- input: "rnn/stacked_rnn_cells_2/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/add_3"
- op: "Add"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_2"
- input: "rnn/stacked_rnn_cells_2/mul_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid_2"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/add_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Tanh_1"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_2/add_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_5"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid_2"
- input: "rnn/stacked_rnn_cells_2/Tanh_1"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-2-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_2/mul_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae2de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-2-input"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_2/OutputHint-UnidirectionalSequenceLstm-47eb6ae2de2411e9a4834201c0a80701-2-2-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 0
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-2-c_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-1-c"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 19
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-2-m_prev"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_1/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-1-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "first"
- }
- }
- attr {
- key: "_tflite_function_input_index"
- value {
- i: 18
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_5/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_5"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-0-2-input"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-18-2-m_prev"
- input: "rnn/stacked_rnn_cells_2/concat_5/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_6/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_6"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-1-None-input_to_input_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-5-None-cell_to_input_w"
- input: "rnn/stacked_rnn_cells_2/concat_6/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_4"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat_5"
- input: "rnn/stacked_rnn_cells_2/concat_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_4"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_4"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-12-None-input_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_7/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_7"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-None-input_to_forget_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-6-None-cell_to_forget_w"
- input: "rnn/stacked_rnn_cells_2/concat_7/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_5"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat_5"
- input: "rnn/stacked_rnn_cells_2/concat_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_5"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_5"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-13-None-forget_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_8/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_8"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-4-None-input_to_output_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-8-None-cell_to_output_w"
- input: "rnn/stacked_rnn_cells_2/concat_8/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_6"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat_5"
- input: "rnn/stacked_rnn_cells_2/concat_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_6"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_6"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-15-None-output_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_9/axis"
- op: "Const"
- attr {
- key: "dtype"
- value {
- type: DT_INT32
- }
- }
- attr {
- key: "value"
- value {
- tensor {
- dtype: DT_INT32
- tensor_shape {
- }
- int_val: 1
- }
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/concat_9"
- op: "ConcatV2"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-3-None-input_to_cell_w"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-7-None-cell_to_cell_w"
- input: "rnn/stacked_rnn_cells_2/concat_9/axis"
- attr {
- key: "N"
- value {
- i: 2
- }
- }
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "Tidx"
- value {
- type: DT_INT32
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/MatMul_7"
- op: "MatMul"
- input: "rnn/stacked_rnn_cells_2/concat_5"
- input: "rnn/stacked_rnn_cells_2/concat_9"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "transpose_a"
- value {
- b: false
- }
- }
- attr {
- key: "transpose_b"
- value {
- b: true
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/BiasAdd_7"
- op: "BiasAdd"
- input: "rnn/stacked_rnn_cells_2/MatMul_7"
- input: "rnn/stacked_rnn_cells/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-14-None-cell_bias"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "data_format"
- value {
- s: "NHWC"
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid_3"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_5"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_6"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid_3"
- input: "rnn/stacked_rnn_cells_2/InputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-19-2-c_prev"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid_4"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Tanh_2"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_7"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid_4"
- input: "rnn/stacked_rnn_cells_2/Tanh_2"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/add_4"
- op: "Add"
- input: "rnn/stacked_rnn_cells_2/mul_6"
- input: "rnn/stacked_rnn_cells_2/mul_7"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Sigmoid_5"
- op: "Sigmoid"
- input: "rnn/stacked_rnn_cells_2/BiasAdd_6"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/Tanh_3"
- op: "Tanh"
- input: "rnn/stacked_rnn_cells_2/add_4"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/mul_8"
- op: "Mul"
- input: "rnn/stacked_rnn_cells_2/Sigmoid_5"
- input: "rnn/stacked_rnn_cells_2/Tanh_3"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-node {
- name: "rnn/stacked_rnn_cells_2/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-2-m"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_2/mul_8"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
- attr {
- key: "_tflite_function_aggregate"
- value {
- s: "stack"
- }
- }
- attr {
- key: "_tflite_function_name"
- value {
- s: "UnidirectionalSequenceLstm"
- }
- }
- attr {
- key: "_tflite_function_output_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_sort_index"
- value {
- i: 2
- }
- }
- attr {
- key: "_tflite_function_uuid"
- value {
- s: "47eb6ae3de2411e9a4834201c0a80701"
- }
- }
- attr {
- key: "_tflite_ophint_level"
- value {
- i: 1
- }
- }
-}
-node {
- name: "OUTPUT"
- op: "Identity"
- input: "rnn/stacked_rnn_cells_2/OutputHint-UnidirectionalSequenceLstm-47eb6ae3de2411e9a4834201c0a80701-2-2-m"
- attr {
- key: "T"
- value {
- type: DT_FLOAT
- }
- }
-}
-library {
-}
-
-# CHECK-LABEL: func @main
-# CHECK-SAME: (%[[ARG_0:[a-z0-9]+]]: tensor<1x3x3xf32>) -> tensor<1x3xf32>
-# CHECK-SAME: control_outputs = ""
-# CHECK-SAME: inputs = "INPUT"
-# CHECK-SAME: outputs = "OUTPUT"
-# CHECK: [[VAL_1:%.*]] = constant dense<0.000000e+00> : tensor<1x3xf32>
-# CHECK: [[VAL_2:%.*]] = constant dense<0.000000e+00> : tensor<3xf32>
-# CHECK: [[VAL_3:%.*]] = constant dense<{{\[\[}}-0.856678485, -0.800494194, 0.716800689], [0.536404848, 0.541643381, -0.35657692], [-0.794646739, 0.137629032, 0.690013885]]> : tensor<3x3xf32>
-# CHECK: [[VAL_4:%.*]] = constant dense<{{\[\[}}-0.125753641, 0.32271719, 0.488939524], [0.36119318, 0.982266664, -0.448646784], [0.966353893, -0.767024993, 0.446366787]]> : tensor<3x3xf32>
-# CHECK: [[VAL_5:%.*]] = constant dense<{{\[\[}}0.891112089, -2.786560e-01, 0.966933965], [-0.789963722, 0.057955265, 0.217499971], [-0.698129416, -0.983400583, -0.834380626]]> : tensor<3x3xf32>
-# CHECK: [[VAL_6:%.*]] = constant dense<{{\[\[}}0.782244444, -0.0446639061, 0.848498106], [-0.579102755, -0.407756329, 0.442389727], [0.00566458702, 0.5984025, 0.629857302]]> : tensor<3x3xf32>
-# CHECK: [[VAL_7:%.*]] = constant dense<1.000000e+00> : tensor<3xf32>
-# CHECK: [[VAL_8:%.*]] = constant dense<{{\[\[}}-0.616786718, 0.892614365, 0.671324968], [-0.842380046, -0.358094931, 0.821366549], [0.790347338, 0.71222949, 0.0690443515]]> : tensor<3x3xf32>
-# CHECK: [[VAL_9:%.*]] = constant dense<{{\[\[}}-5.087240e-01, -0.588907719, 0.471896172], [-0.508019447, -0.0157074928, -0.804120779], [-0.978842973, 0.00160336494, -0.978532075]]> : tensor<3x3xf32>
-# CHECK: [[VAL_10:%.*]] = constant dense<{{\[\[}}0.18183589, 0.616135359, -0.167827845], [0.734281301, 0.958347797, -0.878054618], [0.369523764, -0.969005823, -0.881014585]]> : tensor<3x3xf32>
-# CHECK: [[VAL_11:%.*]] = constant dense<{{\[\[}}-0.936182261, -0.935433864, 0.288229942], [-0.243383884, -0.628288031, -0.477061749], [-0.514976501, -0.903514862, 6.728170e-01]]> : tensor<3x3xf32>
-# CHECK: [[VAL_12:%.*]] = constant dense<{{\[}}0.403919935, -0.882057666, -0.894463062]> : tensor<3xf32>
-# CHECK: [[VAL_13:%.*]] = constant dense<{{\[}}-0.671292543, 0.411814928, 0.560465336]> : tensor<3xf32>
-# CHECK: [[VAL_14:%.*]] = constant dense<{{\[}}0.171322107, -0.153412342, 0.591750383]> : tensor<3xf32>
-# CHECK: [[VAL_15:%.*]] = constant dense<{{\[\[}}-0.207589626, -0.756766081, -0.853258133], [-0.269270182, 0.0468223095, -0.353052378], [-0.0702953338, 0.0725159645, -0.817753077]]> : tensor<3x3xf32>
-# CHECK: [[VAL_16:%.*]] = constant dense<{{\[\[}}0.230039358, -0.182297707, -0.352231741], [-0.805100203, -0.220300436, -0.669503212], [0.278807402, -0.201502323, -0.627609729]]> : tensor<3x3xf32>
-# CHECK: [[VAL_17:%.*]] = constant dense<{{\[\[}}0.513064623, -0.692989588, 0.547988653], [0.0653710365, 0.576977491, 0.966733217], [0.0130724907, 0.247342348, 0.317092657]]> : tensor<3x3xf32>
-# CHECK: [[VAL_18:%.*]] = constant dense<{{\[\[}}-0.138204336, -0.10879755, -0.135128736], [0.94797182, -8.713360e-01, -0.792336463], [0.0339827538, -0.539326906, 8.906350e-01]]> : tensor<3x3xf32>
-# CHECK: [[VAL_19:%.*]] = constant dense<{{\[\[}}0.444335222, -0.133341789, 0.839591503], [0.445418358, -0.571707964, 0.569707394], [0.465010405, -0.990037918, -0.632481337]]> : tensor<3x3xf32>
-# CHECK: [[VAL_20:%.*]] = constant dense<{{\[\[}}-0.493826151, -0.391061306, -0.349843264], [-0.0213134289, 0.558384657, -0.51513052], [0.427886248, 0.618100405, -0.187585592]]> : tensor<3x3xf32>
-# CHECK: [[VAL_21:%.*]] = constant dense<{{\[\[}}0.886177539, -0.606141329, -0.451275587], [0.325554609, 0.691527605, -0.676239967], [0.219799042, 0.626042128, -0.597596407]]> : tensor<3x3xf32>
-# CHECK: [[VAL_22:%.*]] = constant dense<{{\[\[}}-0.400154352, 0.739109992, 0.201825857], [0.678572893, 0.32076478, 0.949867963], [-0.807729483, -5.324750e-01, 0.148033619]]> : tensor<3x3xf32>
-# CHECK: [[VAL_23:%.*]] = constant unit
-# CHECK: [[UNPACK:%.*]]:3 = "tfl.unpack"(%arg0) {axis = 1 : i32, num = 3 : i32} : (tensor<1x3x3xf32>) -> (tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>)
-# CHECK: [[PACK:%.*]] = "tfl.pack"([[UNPACK]]#0, [[UNPACK]]#1, [[UNPACK]]#2) {axis = 0 : i32, values_count = 3 : i32} : (tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<3x1x3xf32>
-# CHECK: [[VAL_24:%.*]] = constant dense<0.000000e+00> : tensor<1x3xf32>
-# CHECK: [[UNIDIRECTIONAL_SEQUENCE_LSTM_1:%.*]] = "tfl.unidirectional_sequence_lstm"([[PACK]], [[VAL_16]], [[VAL_17]], [[VAL_18]], [[VAL_15]], [[VAL_20]], [[VAL_21]], [[VAL_22]], [[VAL_19]], [[VAL_13]], [[VAL_14]], [[VAL_12]], [[VAL_2]], [[VAL_7]], [[VAL_2]], [[VAL_2]], [[VAL_23]], [[VAL_23]], [[VAL_1]], [[VAL_24]], [[VAL_23]], [[VAL_23]], [[VAL_23]], [[VAL_23]]) {fused_activation_function = "TANH", time_major = true} : (tensor<3x1x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, none, none, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, none) -> tensor<3x1x3xf32>
-# CHECK: [[VAL_25:%.*]] = constant dense<0.000000e+00> : tensor<1x3xf32>
-# CHECK: [[VAL_26:%.*]] = constant dense<0.000000e+00> : tensor<1x3xf32>
-# CHECK: [[UNIDIRECTIONAL_SEQUENCE_LSTM_2:%.*]] = "tfl.unidirectional_sequence_lstm"([[UNIDIRECTIONAL_SEQUENCE_LSTM_1]], [[VAL_4]], [[VAL_5]], [[VAL_6]], [[VAL_3]], [[VAL_9]], [[VAL_10]], [[VAL_11]], [[VAL_8]], [[VAL_23]], [[VAL_23]], [[VAL_23]], [[VAL_2]], [[VAL_7]], [[VAL_2]], [[VAL_2]], [[VAL_23]], [[VAL_23]], [[VAL_25]], [[VAL_26]], [[VAL_23]], [[VAL_23]], [[VAL_23]], [[VAL_23]]) {fused_activation_function = "TANH", time_major = true} : (tensor<3x1x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<3x3xf32>, none, none, none, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, none, none, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, none) -> tensor<3x1x3xf32>
-# CHECK: [[RESULT:%.*]]:3 = "tfl.unpack"([[UNIDIRECTIONAL_SEQUENCE_LSTM_2]]) {axis = 0 : i32, num = 3 : i32} : (tensor<3x1x3xf32>) -> (tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>)
-# CHECK: return [[RESULT]]#2 : tensor<1x3xf32>
diff --git a/tensorflow/compiler/mlir/lite/tests/extract-ophint.mlir b/tensorflow/compiler/mlir/lite/tests/extract-ophint.mlir
deleted file mode 100644
index a18ba9c..0000000
--- a/tensorflow/compiler/mlir/lite/tests/extract-ophint.mlir
+++ /dev/null
@@ -1,201 +0,0 @@
-// RUN: tf-opt -tfl-extract-ophint %s -split-input-file -verify-diagnostics | FileCheck %s
-
-// CHECK-LABEL: extractSimpleOphint
-func @extractSimpleOphint() {
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @d4b1eb00b81211e99426dc4a3e957995(%0) : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-// CHECK: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[OP_HINT_CALL]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation", _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %3 = "tf.Mul"(%2, %1) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x16x1xf32>, tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- return
-}
-
-// CHECK: func @d4b1eb00b81211e99426dc4a3e957995(tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation"}
-
-// -----
-
-// CHECK-LABEL: extractPackedInputOphint
-func @extractPackedInputOphint() {
-// CHECK: %[[PACK:[0-9]*]] = "tfl.pack"(%0, %1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<2x1x16x1xf32>
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @"47393154b9af11e99426dc4a3e957995"(%[[PACK]]) : (tensor<2x1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[OP_HINT_CALL]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_stack", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "47393154b9af11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack-47393154b9af11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_stack", _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "47393154b9af11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_stack-47393154b9af11e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_stack", _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "47393154b9af11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_stack-47393154b9af11e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %5 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_1"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %6 = "tf.Mul"(%2, %5) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %7 = "tf.Identity"(%6) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_stack", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "47393154b9af11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack-47393154b9af11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @"47393154b9af11e99426dc4a3e957995"(tensor<2x1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation_stack"}
-
-// -----
-
-// CHECK-LABEL: extractFirstInputOphint
-func @extractFirstInputOphint() {
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @b703f0f4b9ec11e99426dc4a3e957995(%0) : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[OP_HINT_CALL]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_first", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "b703f0f4b9ec11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_first-b703f0f4b9ec11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "first", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_first", _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "b703f0f4b9ec11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_first-b703f0f4b9ec11e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "first", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_first", _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "b703f0f4b9ec11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_first-b703f0f4b9ec11e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %5 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_1"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %6 = "tf.Mul"(%2, %5) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %7 = "tf.Identity"(%6) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_first", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "b703f0f4b9ec11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_first-b703f0f4b9ec11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @b703f0f4b9ec11e99426dc4a3e957995(tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation_first"}
-
-// -----
-
-// CHECK-LABEL: extractLastInputOphint
-func @extractLastInputOphint() {
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @e31fcf90b9ed11e99426dc4a3e957995(%1) : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[OP_HINT_CALL]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_last", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "e31fcf90b9ed11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_last-e31fcf90b9ed11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "last", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_last", _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "e31fcf90b9ed11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_last-e31fcf90b9ed11e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "last", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_last", _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "e31fcf90b9ed11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_last-e31fcf90b9ed11e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %5 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_1"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %6 = "tf.Mul"(%2, %5) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %7 = "tf.Identity"(%6) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_last", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "e31fcf90b9ed11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_last-e31fcf90b9ed11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @e31fcf90b9ed11e99426dc4a3e957995(tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation_last"}
-
-// -----
-
-// CHECK-LABEL: extractPackOneInputOphint
-func @extractPackOneInputOphint() {
-// CHECK: %[[CST:.*]] = constant dense<[1, 1, 16, 1]> : tensor<4xi32>
-// CHECK: %[[RESHAPE:[0-9]*]] = "tfl.reshape"(%0, %[[CST]]) : (tensor<1x16x1xf32>, tensor<4xi32>) -> tensor<1x1x16x1xf32>
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @"33fab028b9ef11e99426dc4a3e957995"(%[[RESHAPE]]) : (tensor<1x1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[OP_HINT_CALL]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_pack_input_one", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "33fab028b9ef11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_pack_input_one-33fab028b9ef11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_pack_input_one", _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "33fab028b9ef11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_pack_input_one-33fab028b9ef11e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Identity"(%2) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_pack_input_one", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "33fab028b9ef11e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_pack_input_one-33fab028b9ef11e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @"33fab028b9ef11e99426dc4a3e957995"(tensor<1x1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation_pack_input_one"}
-
-// -----
-
-// CHECK-LABEL: extractStackInputOutputOphint
-func @extractStackInputOutputOphint() {
-// CHECK: %[[PACK:[0-9]*]] = "tfl.pack"(%0, %1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<2x1x16x1xf32>
-// CHECK: %[[OP_HINT_CALL:[0-9]*]] = call @b92ed354b9f011e99426dc4a3e957995(%[[PACK]]) : (tensor<2x1x16x1xf32>) -> tensor<2x1x16x1xf32>
-// CHECK: %[[UNPACK:[0-9]*]]:2 = "tfl.unpack"(%[[OP_HINT_CALL]]) {axis = 0 : i32, num = 2 : i32} : (tensor<2x1x16x1xf32>) -> (tensor<1x16x1xf32>, tensor<1x16x1xf32>)
-// CHECK-DAG: %[[OUTPUT:[0-9]*]] = "tf.Identity"(%[[UNPACK]]#1) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_output_index = 0 : i64, _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-// CHECK-DAG: %[[OUTPUT_1:[0-9]*]] = "tf.Identity"(%[[UNPACK]]#0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_output_index = 0 : i64, _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %5 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_1"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %6 = "tf.Mul"(%2, %5) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %7 = "tf.Identity"(%6) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_output_index = 0 : i64, _tflite_function_sort_index = 0 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-0-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %8 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_2"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %9 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_3"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %10 = "tf.Add"(%8, %9) {T = "tfdtype$DT_FLOAT", name = "add"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %11 = "tf.Identity"(%10) {T = "tfdtype$DT_FLOAT", _tflite_function_aggregate = "stack", _tflite_function_name = "cool_activation_stack_input_output", _tflite_function_output_index = 0 : i64, _tflite_function_sort_index = 1 : i64, _tflite_function_uuid = "b92ed354b9f011e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_stack_input_output-b92ed354b9f011e99426dc4a3e957995-0-1-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @b92ed354b9f011e99426dc4a3e957995(tensor<2x1x16x1xf32>) -> tensor<2x1x16x1xf32>
-// CHECK: attributes {_tflite_function_input_index = [0 : i32], _tflite_function_name = "cool_activation_stack_input_output"}
-
-// -----
-
-// CHECK-LABEL: extractMultipleInputsOutputsOphint
-func @extractMultipleInputsOutputsOphint() {
-// CHECK: %[[MULTI_INPUT_CALL:[0-9]*]]:2 = call @a6ca45beb9f411e99426dc4a3e957995(%0, %1) : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> (tensor<1x16x1xf32>, tensor<1x16x1xf32>)
-
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation_multiple_input_output", _tflite_function_uuid = "a6ca45beb9f411e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_multiple_input_output-a6ca45beb9f411e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %3 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 1 : i64, _tflite_function_name = "cool_activation_multiple_input_output", _tflite_function_uuid = "a6ca45beb9f411e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation_multiple_input_output-a6ca45beb9f411e99426dc4a3e957995-1-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %5 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_1"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %6 = "tf.Mul"(%2, %5) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %7 = "tf.Identity"(%6) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_multiple_input_output", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "a6ca45beb9f411e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_multiple_input_output-a6ca45beb9f411e99426dc4a3e957995-0-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %8 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_2"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %9 = "tf.Sigmoid"(%4) {T = "tfdtype$DT_FLOAT", name = "Sigmoid_3"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %10 = "tf.Add"(%8, %9) {T = "tfdtype$DT_FLOAT", name = "add"} : (tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- %11 = "tf.Identity"(%10) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation_multiple_input_output", _tflite_function_output_index = 1 : i64, _tflite_function_uuid = "a6ca45beb9f411e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation_multiple_input_output-a6ca45beb9f411e99426dc4a3e957995-1-None-None"} : (tensor<1x16x1xf32>) -> tensor<1x16x1xf32>
- return
-}
-
-// CHECK: func @a6ca45beb9f411e99426dc4a3e957995(tensor<1x16x1xf32>, tensor<1x16x1xf32>) -> (tensor<1x16x1xf32>, tensor<1x16x1xf32>)
-// CHECK: attributes {_tflite_function_input_index = [0 : i32, 1 : i32], _tflite_function_name = "cool_activation_multiple_input_output"}
-
-// -----
-
-// CHECK-LABEL: inputsAfterOutputs
-func @inputsAfterOutputs() {
-// CHECK: %[[PLACE_HOLDER:[0-9]*]] = "tf.Placeholder"() {device = "", dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 2 } dim { size: 2 }"} : () -> tensor<2x2xf32>
-// CHECK: %[[INPUT_PROCESS:[0-9]*]] = "tf.Sigmoid"(%[[PLACE_HOLDER]]) {T = "tfdtype$DT_FLOAT", device = "", name = "Sigmoid"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
-// CHECK: %[[OP_HINT_CALL:[0-9]*]]:2 = call @d6266124d2dd11e9b52cdc4a3e957995(%0, %1, %[[INPUT_PROCESS]]) : (tensor<2x2xf32>, tensor<f32>, tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>)
-
- %0 = "tf.Const"() {device = "", dtype = "tfdtype$DT_FLOAT", name = "Const", value = dense<0.000000e+00> : tensor<f32>} : () -> tensor<f32>
- %1 = "tf.Identity"(%0) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 1 : i64, _tflite_function_name = "CustomOp", _tflite_function_uuid = "d6266124d2dd11e9b52cdc4a3e957995", _tflite_ophint_level = 1 : i64, device = "", name = "InputHint-CustomOp-d6266124d2dd11e9b52cdc4a3e957995-1-None-None"} : (tensor<f32>) -> tensor<f32>
- %2 = "tf.Placeholder"() {device = "", dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 2 } dim { size: 2 }"} : () -> tensor<2x2xf32>
- %3 = "tf.Identity"(%2) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 0 : i64, _tflite_function_name = "CustomOp", _tflite_function_uuid = "d6266124d2dd11e9b52cdc4a3e957995", _tflite_ophint_level = 1 : i64, device = "", name = "InputHint-CustomOp-d6266124d2dd11e9b52cdc4a3e957995-0-None-None"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
- %4 = "tf.Add"(%3, %1) {T = "tfdtype$DT_FLOAT", device = "", name = "Add"} : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
- %5 = "tf.Identity"(%4) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "CustomOp", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "d6266124d2dd11e9b52cdc4a3e957995", _tflite_ophint_level = 1 : i64, device = "", name = "OutputHint-CustomOp-d6266124d2dd11e9b52cdc4a3e957995-0-None-None"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
- %6 = "tf.Placeholder"() {device = "", dtype = "tfdtype$DT_FLOAT", name = "Placeholder_1", shape = "tfshape$dim { size: 2 } dim { size: 2 }"} : () -> tensor<2x2xf32>
- %7 = "tf.Sigmoid"(%6) {T = "tfdtype$DT_FLOAT", device = "", name = "Sigmoid"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
- %8 = "tf.Identity"(%7) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 2 : i64, _tflite_function_name = "CustomOp", _tflite_function_uuid = "d6266124d2dd11e9b52cdc4a3e957995", _tflite_ophint_level = 1 : i64, device = "", name = "InputHint-CustomOp-d6266124d2dd11e9b52cdc4a3e957995-2-None-None"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
- %9 = "tf.Add"(%5, %8) {T = "tfdtype$DT_FLOAT", device = "", name = "Add_1"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
- %10 = "tf.Identity"(%9) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "CustomOp", _tflite_function_output_index = 1 : i64, _tflite_function_uuid = "d6266124d2dd11e9b52cdc4a3e957995", _tflite_ophint_level = 1 : i64, device = "", name = "OutputHint-CustomOp-d6266124d2dd11e9b52cdc4a3e957995-1-None-None"} : (tensor<2x2xf32>) -> tensor<2x2xf32>
- return
-}
-
-// CHECK: func @d6266124d2dd11e9b52cdc4a3e957995(tensor<2x2xf32>, tensor<f32>, tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>)
-// CHECK: attributes {_tflite_function_input_index = [0 : i32, 1 : i32, 2 : i32], _tflite_function_name = "CustomOp"}
-
-// -----
-
-module {
-func @extractOphintSame() {
- %0 = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x16x1xf32>
- %1 = call @AnotherFunc(%0) : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %2 = "tf.Sigmoid"(%1) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %3 = "tf.Mul"(%2, %1) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x16x1xf32>, tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- %4 = "tf.Identity"(%3) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- return
-
-// CHECK: [[VAL_0:%.*]] = "tf.Placeholder"() {dtype = "tfdtype$DT_FLOAT", name = "Placeholder", shape = "tfshape$dim { size: 1 } dim { size: 16 } dim { size: 16 } dim { size: 1 }"} : () -> tensor<1x16x16x1xf32>
-// CHECK: [[VAL_1:%.*]] = call @AnotherFunc([[VAL_0]]) : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-// CHECK: [[VAL_2:%.*]] = "tf.Sigmoid"([[VAL_1]]) {T = "tfdtype$DT_FLOAT", name = "Sigmoid"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-// CHECK: [[VAL_3:%.*]] = "tf.Mul"([[VAL_2]], [[VAL_1]]) {T = "tfdtype$DT_FLOAT", name = "mul"} : (tensor<1x16x16x1xf32>, tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-// CHECK: [[VAL_4:%.*]] = "tf.Identity"([[VAL_3]]) {T = "tfdtype$DT_FLOAT", _tflite_function_name = "cool_activation", _tflite_function_output_index = 0 : i64, _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "OutputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
-}
-
-func @AnotherFunc(%arg0: tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32> {
- %0 = "tf.Identity"(%arg0) {T = "tfdtype$DT_FLOAT", _tflite_function_input_index = 0 : i64, _tflite_function_name = "cool_activation", _tflite_function_uuid = "d4b1eb00b81211e99426dc4a3e957995", _tflite_ophint_level = 1 : i64, name = "InputHint-cool_activation-d4b1eb00b81211e99426dc4a3e957995-0-None-None"} : (tensor<1x16x16x1xf32>) -> tensor<1x16x16x1xf32>
- return %0 : tensor<1x16x16x1xf32>
-}
-}
diff --git a/tensorflow/compiler/mlir/lite/tests/legalize-ophint-func-op.mlir b/tensorflow/compiler/mlir/lite/tests/legalize-ophint-func-op.mlir
deleted file mode 100644
index 97bb6f2..0000000
--- a/tensorflow/compiler/mlir/lite/tests/legalize-ophint-func-op.mlir
+++ /dev/null
@@ -1,68 +0,0 @@
-// RUN: tf-opt -tfl-legalize-ophint-func-op %s -split-input-file | FileCheck %s
-
-module {
- // CHECK-LABEL: func @testConvertUnidirectionalSequenceRNN
- // CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<1x3xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<1x3xf32>)
- func @testConvertUnidirectionalSequenceRNN(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x4xf32> {
- // CHECK: %[[CST:.*]] = constant dense<0.000000e+00> : tensor<1x4xf32>
- // CHECK: %[[CST_0:.*]] = constant dense<0.000000e+00> : tensor<4xf32>
- // CHECK: %[[CST_1:.*]] = constant dense<0.000000e+00> : tensor<4x3xf32>
- // CHECK: %[[CST_2:.*]] = constant dense<0.000000e+00> : tensor<4x4xf32>
- // CHECK: %[[PACKED_INPUT:[a-z0-9]*]] = "tfl.pack"(%[[ARG_0]], %[[ARG_1]]) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x1x3xf32>
- // CHECK: %[[FUSED_OUTPUT:[a-z0-9]*]] = "tfl.unidirectional_sequence_rnn"(%[[PACKED_INPUT]], %[[CST_1]], %[[CST_2]], %[[CST_0]], %[[CST]]) {fused_activation_function = "TANH", time_major = true} : (tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<1x4xf32>) -> tensor<2x1x4xf32>
- // CHECK: %[[UNPACK:[0-9]*]]:2 = "tfl.unpack"(%[[FUSED_OUTPUT]]) {axis = 0 : i32, num = 2 : i32} : (tensor<2x1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xf32>)
-
- %cst = constant dense<0.000000e+00> : tensor<1x4xf32>
- %cst0 = constant dense<0.000000e+00> : tensor<4xf32>
- %cst1 = constant dense<0.000000e+00> : tensor<4x3xf32>
- %cst2 = constant dense<0.000000e+00> : tensor<4x4xf32>
- %2 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x1x3xf32>
- %3 = call @a9211722c23011e9875cdc4a3e957995(%2, %cst1, %cst2, %cst0, %cst) : (tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<1x4xf32>) -> tensor<2x1x4xf32>
- %4:2 = "tfl.unpack"(%3) {axis = 0 : i32, num = 2 : i32} : (tensor<2x1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xf32>)
- return %4#0 : tensor<1x4xf32>
- }
- func @a9211722c23011e9875cdc4a3e957995(tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<1x4xf32>) -> tensor<2x1x4xf32>
- attributes {_tflite_function_name = "UnidirectionalSequenceRnn"}
-}
-
-// -----
-
-module {
- // CHECK-LABEL: func @testConvertUnidirectionalSequenceLSTM
- // CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<1x3xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<1x3xf32>)
- func @testConvertUnidirectionalSequenceLSTM(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x4xf32> {
- // CHECK: %[[CST:.*]] = constant dense<0.000000e+00> : tensor<4x4xf32>
- // CHECK: %[[CST_0:.*]] = constant dense<0.000000e+00> : tensor<4x4xf32>
- // CHECK: %[[CST_1:.*]] = constant dense<0.000000e+00> : tensor<4x4xf32>
- // CHECK: %[[CST_2:.*]] = constant dense<0.000000e+00> : tensor<4x4xf32>
- // CHECK: %[[CST_3:.*]] = constant dense<1.000000e+00> : tensor<4xf32>
- // CHECK: %[[CST_4:.*]] = constant dense<0.000000e+00> : tensor<4x3xf32>
- // CHECK: %[[CST_5:.*]] = constant dense<0.000000e+00> : tensor<4x3xf32>
- // CHECK: %[[CST_6:.*]] = constant dense<0.000000e+00> : tensor<4x3xf32>
- // CHECK: %[[CST_7:.*]] = constant dense<0.000000e+00> : tensor<4x3xf32>
- // CHECK: %[[CST_8:.*]] = constant dense<0.000000e+00> : tensor<4xf32>
- // CHECK: %[[CST_9:.*]] = constant dense<0.000000e+00> : tensor<1x4xf32>
- // CHECK: %[[PACKED_INPUT:[a-z0-9]*]] = "tfl.pack"(%[[ARG_0]], %[[ARG_1]]) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x1x3xf32>
- // CHECK: %[[CST_10:.*]] = constant unit
- // CHECK: %[[FUSED_OUTPUT:[a-z0-9]*]] = "tfl.unidirectional_sequence_lstm"(%[[PACKED_INPUT]], %[[CST_6]], %[[CST_5]], %[[CST_4]], %[[CST_7]], %[[CST_1]], %[[CST_0]], %[[CST]], %[[CST_2]], %[[CST_10]], %[[CST_10]], %[[CST_10]], %[[CST_8]], %[[CST_3]], %[[CST_8]], %[[CST_8]], %[[CST_10]], %[[CST_10]], %[[CST_9]], %[[CST_9]], %[[CST_10]], %[[CST_10]], %[[CST_10]], %[[CST_10]]) {fused_activation_function = "TANH", time_major = true} : (tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, none, none, none, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, none, none, tensor<1x4xf32>, tensor<1x4xf32>, none, none, none, none) -> tensor<2x1x4xf32>
- // CHECK: %[[UNPACK:[0-9]*]]:2 = "tfl.unpack"(%[[FUSED_OUTPUT]]) {axis = 0 : i32, num = 2 : i32} : (tensor<2x1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xf32>)
-
- %cst = constant dense<0.000000e+00> : tensor<4x4xf32>
- %cst_0 = constant dense<0.000000e+00> : tensor<4x4xf32>
- %cst_1 = constant dense<0.000000e+00> : tensor<4x4xf32>
- %cst_2 = constant dense<0.000000e+00> : tensor<4x4xf32>
- %cst_3 = constant dense<1.000000e+00> : tensor<4xf32>
- %cst_4 = constant dense<0.000000e+00> : tensor<4x3xf32>
- %cst_5 = constant dense<0.000000e+00> : tensor<4x3xf32>
- %cst_6 = constant dense<0.000000e+00> : tensor<4x3xf32>
- %cst_7 = constant dense<0.000000e+00> : tensor<4x3xf32>
- %cst_8 = constant dense<0.000000e+00> : tensor<4xf32>
- %cst_9 = constant dense<0.000000e+00> : tensor<1x4xf32>
- %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x1x3xf32>
- %1:2 = call @a7addbdad08811e9b52cdc4a3e957995(%0, %cst_6, %cst_5, %cst_4, %cst_7, %cst_1, %cst_0, %cst, %cst_2, %cst_8, %cst_3, %cst_8, %cst_8, %cst_9, %cst_9) : (tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<1x4xf32>, tensor<1x4xf32>) -> (tensor<1x4xf32>, tensor<2x1x4xf32>)
- %2:2 = "tfl.unpack"(%1#1) {axis = 0 : i32, num = 2 : i32} : (tensor<2x1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xf32>)
- return %2#1 : tensor<1x4xf32>
- }
- func @a7addbdad08811e9b52cdc4a3e957995(tensor<2x1x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x3xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<1x4xf32>, tensor<1x4xf32>) -> (tensor<1x4xf32>, tensor<2x1x4xf32>)
- attributes {_tflite_function_input_index = [0 : i32, 1 : i32, 2 : i32, 3 : i32, 4 : i32, 5 : i32, 6 : i32, 7 : i32, 8 : i32, 12 : i32, 13 : i32, 14 : i32, 15 : i32, 18 : i32, 19 : i32], _tflite_function_name = "UnidirectionalSequenceLstm"}
-}
diff --git a/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir b/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir
index d6f2a83..91b38ab 100644
--- a/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir
+++ b/tensorflow/compiler/mlir/lite/tests/legalize-tf.mlir
@@ -1296,10 +1296,12 @@
// CHECK-LABEL: conv2d_backprop_input
// CHECK: %[[CST:.*]] = constant dense<[2, 0, 1, 3]> : tensor<4xi32>
// CHECK: %[[ARG0:.*]] = "tfl.transpose"(%arg1, %[[CST]]) : (tensor<3x3x1x32xf32>, tensor<4xi32>) -> tensor<1x3x3x32xf32>
- // CHECK: %[[ARG1:.*]] = "tfl.transpose_conv"(%arg0, %[[ARG0]], %arg2) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<1x3x3x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>
+ // CHECK: %[[CST_0:.*]] = constant unit
+ // CHECK: %[[ARG1:.*]] = "tfl.transpose_conv"(%arg0, %[[ARG0]], %arg2, %[[CST_0]]) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<1x3x3x32xf32>, tensor<15x14x14x32xf32>, none) -> tensor<15x28x28x1xf32>
// CHECK: %[[CST_1:.*]] = constant dense<[2, 0, 1, 3]> : tensor<4xi32>
// CHECK: %[[ARG2:.*]] = "tfl.transpose"(%arg1, %[[CST_1]]) : (tensor<3x3x1x32xf32>, tensor<4xi32>) -> tensor<1x3x3x32xf32>
- // CHECK: %[[ARG3:.*]] = "tfl.transpose_conv"(%arg0, %[[ARG2]], %arg2) {padding = "VALID", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<1x3x3x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>
+ // CHECK: %[[CST_2:.*]] = constant unit
+ // CHECK: %[[ARG3:.*]] = "tfl.transpose_conv"(%arg0, %[[ARG2]], %arg2, %[[CST_2]]) {padding = "VALID", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<1x3x3x32xf32>, tensor<15x14x14x32xf32>, none) -> tensor<15x28x28x1xf32>
// CHECK: %[[RESULT:.*]] = tfl.add %[[ARG1]], %[[ARG3]] {fused_activation_function = "NONE"} : tensor<15x28x28x1xf32>
// CHECK: return %[[RESULT]] : tensor<15x28x28x1xf32>
}
diff --git a/tensorflow/compiler/mlir/lite/tests/ops.mlir b/tensorflow/compiler/mlir/lite/tests/ops.mlir
index a85c7f2..697e93e 100644
--- a/tensorflow/compiler/mlir/lite/tests/ops.mlir
+++ b/tensorflow/compiler/mlir/lite/tests/ops.mlir
@@ -225,10 +225,10 @@
}
// CHECK-LABEL: testDequantize
-func @testDequantize(tensor<? x i32>) -> tensor<? x f32> {
-^bb0(%arg0: tensor<? x i32>):
- // CHECK: "tfl.dequantize"(%arg0) : (tensor<?xi32>) -> tensor<?xf32>
- %0 = "tfl.dequantize"(%arg0): (tensor<? x i32>) -> tensor<? x f32>
+func @testDequantize(tensor<? x !quant.uniform<i8:f32, 0.1>>) -> tensor<? x f32> {
+^bb0(%arg0: tensor<? x !quant.uniform<i8:f32, 0.1>>):
+ // CHECK: "tfl.dequantize"(%arg0) : (tensor<?x!quant.uniform<i8:f32, 1.000000e-01>>) -> tensor<?xf32>
+ %0 = "tfl.dequantize"(%arg0): (tensor<? x !quant.uniform<i8:f32, 0.1>>) -> tensor<? x f32>
return %0 : tensor<? x f32>
}
@@ -609,9 +609,9 @@
// -----
// CHECK-LABEL: testBidirectionalSequenceLstm
-func @testBidirectionalSequenceLstm(%arg0: tensor<? x f32>, %arg1: none, %arg2: tensor<? x f32>, %arg3: tensor<? x f32>, %arg4: tensor<? x f32>, %arg5: tensor<? x f32>, %arg6: tensor<? x f32>, %arg7: tensor<? x f32>, %arg8: tensor<? x f32>, %arg9: tensor<? x f32>, %arg10: tensor<? x f32>, %arg11: tensor<? x f32>, %arg12: tensor<? x f32>, %arg13: tensor<? x f32>, %arg14: tensor<? x f32>, %arg15: tensor<? x f32>, %arg16: tensor<? x f32>, %arg17: tensor<? x f32>, %arg18: tensor<? x f32>, %arg19: tensor<? x f32>, %arg20: tensor<? x f32>, %arg21: tensor<? x f32>, %arg22: tensor<? x f32>, %arg23: tensor<? x f32>, %arg24: tensor<? x f32>, %arg25: tensor<? x f32>, %arg26: tensor<? x f32>, %arg27: tensor<? x f32>, %arg28: tensor<? x f32>, %arg29: tensor<? x f32>, %arg30: tensor<? x f32>, %arg31: tensor<? x f32>, %arg32: tensor<? x f32>, %arg33: tensor<? x f32>, %arg34: tensor<? x f32>, %arg35: tensor<? x f32>, %arg36: tensor<? x f32>, %arg37: tensor<? x f32>, %arg38: tensor<? x f32>, %arg39: tensor<? x f32>, %arg40: tensor<? x f32>, %arg41: tensor<? x f32>, %arg42: tensor<? x f32>, %arg43: tensor<? x f32>, %arg44: tensor<? x f32>, %arg45: tensor<? x f32>, %arg46: tensor<? x f32>, %arg47: tensor<? x f32>) -> tensor<? x f32> {
- // CHECK: "tfl.bidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23, %arg24, %arg25, %arg26, %arg27, %arg28, %arg29, %arg30, %arg31, %arg32, %arg33, %arg34, %arg35, %arg36, %arg37, %arg38, %arg39, %arg40, %arg41, %arg42, %arg43, %arg44, %arg45, %arg46, %arg47) {cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", merge_outputs = true, time_major = false} : (tensor<?xf32>, none, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> (tensor<?xf32>, tensor<?xf32>)
- %0:2 = "tfl.bidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23, %arg24, %arg25, %arg26, %arg27, %arg28, %arg29, %arg30, %arg31, %arg32, %arg33, %arg34, %arg35, %arg36, %arg37, %arg38, %arg39, %arg40, %arg41, %arg42, %arg43, %arg44, %arg45, %arg46, %arg47) {cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", merge_outputs = true, time_major = false} : (tensor<?xf32>, none, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> (tensor<?xf32>, tensor<?xf32>)
+func @testBidirectionalSequenceLstm(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?xf32>, %arg2: tensor<?x?xf32>, %arg3: tensor<?x?xf32>, %arg4: tensor<?x?xf32>, %arg5: tensor<?x?xf32>, %arg6: tensor<?x?xf32>, %arg7: tensor<?x?xf32>, %arg8: tensor<?x?xf32>, %arg9: tensor<?xf32>, %arg10: tensor<?xf32>, %arg11: tensor<?xf32>, %arg12: tensor<?xf32>, %arg13: tensor<?xf32>, %arg14: tensor<?xf32>, %arg15: tensor<?xf32>, %arg16: tensor<?x?xf32>, %arg17: tensor<?xf32>, %arg18: tensor<?x?xf32>, %arg19: tensor<?x?xf32>, %arg20: tensor<?x?xf32>, %arg21: tensor<?x?xf32>, %arg22: tensor<?x?xf32>, %arg23: tensor<?x?xf32>, %arg24: tensor<?x?xf32>, %arg25: tensor<?x?xf32>, %arg26: tensor<?xf32>, %arg27: tensor<?xf32>, %arg28: tensor<?xf32>, %arg29: tensor<?xf32>, %arg30: tensor<?xf32>, %arg31: tensor<?xf32>, %arg32: tensor<?xf32>, %arg33: tensor<?x?xf32>, %arg34: tensor<?xf32>, %arg35: tensor<?xf32>, %arg36: tensor<?xf32>, %arg37: tensor<?xf32>, %arg38: tensor<?xf32>, %arg39: tensor<?xf32>, %arg40: tensor<?xf32>, %arg41: tensor<?xf32>, %arg42: tensor<?xf32>, %arg43: tensor<?xf32>, %arg44: tensor<?xf32>, %arg45: tensor<?xf32>, %arg46: tensor<?xf32>, %arg47: tensor<?xf32>) -> tensor<?xf32> {
+ // CHECK: "tfl.bidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23, %arg24, %arg25, %arg26, %arg27, %arg28, %arg29, %arg30, %arg31, %arg32, %arg33, %arg34, %arg35, %arg36, %arg37, %arg38, %arg39, %arg40, %arg41, %arg42, %arg43, %arg44, %arg45, %arg46, %arg47) {cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", merge_outputs = true, time_major = false} : (tensor<?x?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> (tensor<?xf32>, tensor<?xf32>)
+ %0:2 = "tfl.bidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23, %arg24, %arg25, %arg26, %arg27, %arg28, %arg29, %arg30, %arg31, %arg32, %arg33, %arg34, %arg35, %arg36, %arg37, %arg38, %arg39, %arg40, %arg41, %arg42, %arg43, %arg44, %arg45, %arg46, %arg47) {cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", merge_outputs = true, time_major = false} : (tensor<?x?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> (tensor<?xf32>, tensor<?xf32>)
return %0#0 : tensor<?xf32>
}
@@ -1444,16 +1444,16 @@
// -----
-func @testEmbeddingLookup(%arg0 : tensor<?xi32>, %arg1 : tensor<?xf32>) -> tensor<?xf32> {
- %0 = "tfl.embedding_lookup"(%arg0, %arg1) : (tensor<?xi32>,tensor<?xf32>) -> tensor<?xf32>
+func @testEmbeddingLookup(%arg0 : tensor<?xi32>, %arg1 : tensor<?x?xf32>) -> tensor<?xf32> {
+ %0 = "tfl.embedding_lookup"(%arg0, %arg1) : (tensor<?xi32>,tensor<?x?xf32>) -> tensor<?xf32>
return %0 : tensor<?xf32>
}
// -----
-func @testEmbeddingLookupValueAndResultElementTypeTraitFailed(%arg0 : tensor<?xi32>, %arg1 : tensor<?xi8>) -> tensor<?xf32> {
+func @testEmbeddingLookupValueAndResultElementTypeTraitFailed(%arg0 : tensor<?xi32>, %arg1 : tensor<?x?xi8>) -> tensor<?xf32> {
// expected-error @+1 {{'tfl.embedding_lookup' op failed to verify that value and output must have same element type}}
- %0 = "tfl.embedding_lookup"(%arg0, %arg1) : (tensor<?xi32>,tensor<?xi8>) -> tensor<?xf32>
+ %0 = "tfl.embedding_lookup"(%arg0, %arg1) : (tensor<?xi32>,tensor<?x?xi8>) -> tensor<?xf32>
return %0 : tensor<?xf32>
}
@@ -2032,7 +2032,8 @@
// -----
func @testTransposeConv(%arg0: tensor<4xi32>, %arg1: tensor<32x4x4x128xf32>, %arg2: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
- %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32>
+ %cst = constant unit
+ %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2, %cst) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x32xf32>
return %0 : tensor<1x64x84x32xf32>
}
@@ -2046,8 +2047,9 @@
// -----
func @testTransposeConvBadOutputRank(%arg0: tensor<4xi32>, %arg1: tensor<32x4x4x128xf32>, %arg2: tensor<1x32x42x128xf32>) -> tensor<64x84x32xf32> {
+ %cst = constant unit
// expected-error @+1 {{expect output type has rank = 4, got output type tensor<64x84x32xf32>}}
- %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>) -> tensor<64x84x32xf32>
+ %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2, %cst) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<64x84x32xf32>
return %0 : tensor<64x84x32xf32>
}
@@ -2055,8 +2057,9 @@
func @testTransposeConvBadOutputShape(%arg1: tensor<32x4x4x128xf32>, %arg2: tensor<1x32x42x128xf32>) -> tensor<1x64x84x31xf32> {
%cst = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
+ %cst_1 = constant unit
// expected-error @+1 {{expect output type tensor<1x64x84x32xf32>, got tensor<1x64x84x31xf32>}}
- %0 = "tfl.transpose_conv"(%cst, %arg1, %arg2) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>) -> tensor<1x64x84x31xf32>
+ %0 = "tfl.transpose_conv"(%cst, %arg1, %arg2, %cst_1) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x31xf32>
return %0 : tensor<1x64x84x31xf32>
}
diff --git a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
index 57f1571..e2354e6 100644
--- a/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
+++ b/tensorflow/compiler/mlir/lite/tf_tfl_passes.cc
@@ -80,26 +80,10 @@
}
// The conversion pipeline has to follow the following orders:
- // 1) Try to convert ophint nodes if present first like ophint lstm.
- // 2) Saved model related optimization like decompose resource ops
- // 3) Convert composite functions like lstm/rnns, along with proper function
+ // 1) Saved model related optimization like decompose resource ops
+ // 2) Convert composite functions like lstm/rnns, along with proper function
// inlining & dce.
- // 4) Lower static tensor list pass.
-
- // The ophint extractions happen before lots of other passes:
- // The assumption of ophint-extraction is each ophinted region is a black-box
- // and nodes within this black-box is NOT connected to the nodes OUTSIDE the
- // black-box.
- // Some passes may merge nodes together (such as const nodes), however, this
- // will break the ophint-extraction assumption. (The nodes within the black
- // box is not isolated anymore).
- // So ophint extraction and legalization needs to happen before
- // the canonicalization pass.
- if (pass_config.emit_builtin_tflite_ops) {
- pass_manager->addPass(mlir::TFL::CreateExtractOphintPass());
- // Convert composite op pass will happen after ophint extraction pass.
- pass_manager->addPass(mlir::TFL::CreateLegalizeOphintFuncOpPass());
- }
+ // 3) Lower static tensor list pass.
// This decomposes resource ops like ResourceGather into read-variable op
// followed by gather. This is used when the saved model import path is used
diff --git a/tensorflow/compiler/mlir/lite/transforms/extract_ophint.cc b/tensorflow/compiler/mlir/lite/transforms/extract_ophint.cc
deleted file mode 100644
index a783314..0000000
--- a/tensorflow/compiler/mlir/lite/transforms/extract_ophint.cc
+++ /dev/null
@@ -1,763 +0,0 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-#include <map>
-#include <queue>
-#include <vector>
-
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/STLExtras.h"
-#include "llvm/ADT/StringRef.h"
-#include "llvm/ADT/StringSwitch.h"
-#include "llvm/Support/Casting.h"
-#include "mlir/Analysis/LoopAnalysis.h" // from @llvm-project
-#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
-#include "mlir/IR/Attributes.h" // from @llvm-project
-#include "mlir/IR/Block.h" // from @llvm-project
-#include "mlir/IR/Builders.h" // from @llvm-project
-#include "mlir/IR/Function.h" // from @llvm-project
-#include "mlir/IR/MLIRContext.h" // from @llvm-project
-#include "mlir/IR/Module.h" // from @llvm-project
-#include "mlir/IR/Operation.h" // from @llvm-project
-#include "mlir/IR/OperationSupport.h" // from @llvm-project
-#include "mlir/IR/PatternMatch.h" // from @llvm-project
-#include "mlir/IR/StandardTypes.h" // from @llvm-project
-#include "mlir/IR/SymbolTable.h" // from @llvm-project
-#include "mlir/IR/Types.h" // from @llvm-project
-#include "mlir/IR/Value.h" // from @llvm-project
-#include "mlir/Pass/Pass.h" // from @llvm-project
-#include "mlir/Pass/PassRegistry.h" // from @llvm-project
-#include "mlir/Support/LLVM.h" // from @llvm-project
-#include "mlir/Support/LogicalResult.h" // from @llvm-project
-#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"
-#include "tensorflow/compiler/mlir/lite/transforms/passes.h"
-#include "tensorflow/compiler/mlir/lite/utils/attribute_utils.h"
-#include "tensorflow/compiler/mlir/lite/utils/validators.h"
-#include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h"
-#include "tensorflow/compiler/mlir/tensorflow/ir/tf_types.h"
-#include "tensorflow/core/platform/logging.h"
-
-namespace mlir {
-namespace TFL {
-namespace {
-
-constexpr char kTfLiteFunctionName[] = "_tflite_function_name";
-constexpr char kTfLiteFunctionUUID[] = "_tflite_function_uuid";
-constexpr char kTfLiteFunctionInputIndex[] = "_tflite_function_input_index";
-constexpr char kTfLiteFunctionOutputIndex[] = "_tflite_function_output_index";
-constexpr char kTfLiteFunctionSortIndex[] = "_tflite_function_sort_index";
-constexpr char kTfLiteFunctionAggregate[] = "_tflite_function_aggregate";
-
-constexpr char kStrategyNone[] = "None";
-constexpr char kStrategyStack[] = "stack";
-constexpr char kStrategyFirst[] = "first";
-constexpr char kStrategyLast[] = "last";
-
-// A Ophinted op typically looks like below"
-//
-// InputOp1 InputOp2 InputOp3
-// / \ | |
-// val1 val2 val3 val4
-// | | | |
-// identOp1 identOp2 identOp3 identOp4
-// \ | | /
-// \ | | /
-// .... a bunch of operations (needs to be fused) ...
-// / \
-// / \
-// identOp1 (output) identOp2 (output)
-// | |
-// Other ops Other ops
-//
-//
-// In this pass, we are trying to convert them into the following format:
-//
-// ||
-// ||
-// \ /
-//
-// InputOp1 InputOp2 InputOp3
-// / \ | /
-// val1 val2 val3 val4
-// \ | | /
-// PackOp | /
-// \ | | /
-// \ | | /
-// Call funcOp (fusedOp - name like 'UnidirectionalSequenceRNN')
-// (The funcOp will be inserted at the bottom of the module, also
-// . note every funcOp will be unique.)
-// |
-// UnpackOp
-// / \
-// / \
-// Other ops Other ops
-struct OphintCompositeOp {
- // OphintCompositeOp is a conceptually "composite op" which will be converted
- // to a "fused op" later.
- //
- // As a "composite op", it has "inputs" and "outputs", and all the inputs
- // and outputs are annotated by special-annotated identity ops.
- //
- // All inputs and outputs need to be processed based on different strategies,
- // See all the different strategies under
- // tensorflow/lite/python/op_hint.py
- //
- // For example, "stack" strategy means we need to pack the inputs together
- // or unpack the outputs.
- public:
- OphintCompositeOp(StringRef uuid, StringRef function_name)
- : uuid(uuid), function_name(function_name) {}
-
- void AddInput(int index, Operation* op, StringRef aggregation,
- int sort_index) {
- auto it = inputs.find(index);
- if (it == inputs.end()) {
- AggregatedOperand operand;
- operand.aggregation = aggregation;
- it = inputs.insert({index, operand}).first;
- }
- // TODO(renjieliu): check aggregation strategy stays the same.
- // Also needs to make sure if aggregation strategy is "None" we should not
- // have more than one op.
- it->second.ops[sort_index] = op;
- }
-
- void AddOutput(int index, Operation* op, llvm::StringRef aggregation,
- int sort_index) {
- auto it = outputs.find(index);
- if (it == outputs.end()) {
- AggregatedOperand operand;
- operand.aggregation = aggregation;
- it = outputs.insert({index, operand}).first;
- }
- // TODO(renjieliu): check aggregation strategy stays the same.
- // Also needs to make sure if aggregation strategy is "None" we should not
- // have more than one op.
- it->second.ops[sort_index] = op;
- }
-
- std::vector<Operation*> GetAllInputOps() {
- std::vector<Operation*> all_input_ops;
- for (const auto& kv : inputs) {
- if (kv.second.aggregation == kStrategyFirst) {
- all_input_ops.push_back(kv.second.ops.at(0));
- continue;
- }
- for (const auto& operandKv : kv.second.ops) {
- all_input_ops.push_back(operandKv.second);
- }
- }
- return all_input_ops;
- }
-
- std::vector<Operation*> GetAllOutputOps() {
- std::vector<Operation*> all_output_ops;
- for (const auto& kv : outputs) {
- for (const auto& operand_kv : kv.second.ops) {
- all_output_ops.push_back(operand_kv.second);
- }
- }
- return all_output_ops;
- }
-
- std::vector<Operation*> GetAllInUseOutputOps() {
- std::vector<Operation*> all_output_ops;
- for (const auto& kv : outputs) {
- auto& aggregated_operand = kv.second;
- if (aggregated_operand.aggregation != kStrategyStack) {
- continue;
- }
- for (const auto& operand_kv : aggregated_operand.ops) {
- all_output_ops.push_back(operand_kv.second);
- }
- }
- return all_output_ops;
- }
-
- // This function will process the aggregated inputs based on different
- // strategies like "first", "last", "stack".
- std::map<int, Value> GetAggregatedInputs(OpBuilder* builder) {
- std::map<int, Value> aggregated_inputs;
- for (const auto& kv : inputs) {
- Value op_input = nullptr;
- const AggregatedOperand& operand = kv.second;
- // Dealing with "stack" strategy:
- // This breaks into two parts:
- // 1) If the ops only has one element, we only add a reshape op to expand
- // the dim.
- // 2) If the ops contain more than one element, we need to append a
- // pack_op after the input ops.
- if (operand.aggregation == kStrategyStack) {
- if (operand.ops.size() == 1) {
- // If ops size is 1, it will be simply expanding dimensions at dim 0.
- Operation* current_identity_op = operand.ops.begin()->second;
- Value input = current_identity_op->getOperand(0);
- RankedTensorType input_type =
- input.getType().cast<RankedTensorType>();
- // The Reshape will be {1, (original_shape)}
- SmallVector<int64_t, 4> reshape_op_shape;
- reshape_op_shape.push_back(1);
- for (const auto& dim : input_type.getShape()) {
- reshape_op_shape.push_back(dim);
- }
-
- Operation* first_use = current_identity_op->getNextNode();
- builder->setInsertionPoint(first_use);
- Location loc = first_use->getLoc();
- auto shape_type = RankedTensorType::get({input_type.getRank() + 1},
- builder->getIntegerType(32));
- SmallVector<Attribute, 4> result_shape_data(reshape_op_shape.size());
- for (int i = 0; i < reshape_op_shape.size(); ++i) {
- result_shape_data[i] = builder->getI32IntegerAttr(
- static_cast<int32_t>(reshape_op_shape[i]));
- }
- auto shape_attr =
- DenseElementsAttr::get(shape_type, result_shape_data);
- auto shape = builder->create<ConstantOp>(loc, shape_type, shape_attr);
- auto reshape_output_type = RankedTensorType::get(
- reshape_op_shape, input_type.getElementType());
- Operation* reshape = builder->create<TFL::ReshapeOp>(
- first_use->getLoc(), reshape_output_type, input, shape);
- op_input = reshape->getResult(0);
-
- } else {
- // Insert a pack op to pack all the inputs together.
- std::vector<Value> pack_input_operands;
- std::vector<Value> packed_input_consumers;
- for (int i = 0, e = operand.ops.size(); i < e; ++i) {
- pack_input_operands.push_back(operand.ops.at(i)->getOperand(0));
- packed_input_consumers.push_back(operand.ops.at(i)->getResult(0));
- }
- // Find the first op that consumes the last value of the aggregated
- // inputs.
- Operation* first_use = *(packed_input_consumers.back().user_begin());
- // The pack reshape will be {N, (original_shape)}
- SmallVector<int64_t, 4> pack_shape;
- pack_shape.push_back(pack_input_operands.size());
- RankedTensorType type = operand.ops.at(0)
- ->getResult(0)
- .getType()
- .cast<RankedTensorType>();
- for (const auto& dim : type.getShape()) {
- pack_shape.push_back(dim);
- }
- auto pack_input_type =
- RankedTensorType::get(pack_shape, type.getElementType());
- builder->setInsertionPoint(first_use);
- Operation* pack_op = builder->create<TFL::PackOp>(
- first_use->getLoc(), pack_input_type, pack_input_operands,
- builder->getI32IntegerAttr(pack_input_operands.size()),
- builder->getI32IntegerAttr(0));
- op_input = pack_op->getResult(0);
- }
- } else if (operand.aggregation == kStrategyLast) {
- // This handle the strategy "last", if simply takes the last input.
- op_input = operand.ops.at(operand.ops.size() - 1)->getOperand(0);
- } else {
- // This handle the strategy "first" and default, if simply takes the
- // first input.
- op_input = operand.ops.at(0)->getOperand(0);
- }
- aggregated_inputs[kv.first] = op_input;
- }
- return aggregated_inputs;
- }
-
- // For now, we just return the first output's location which the fused op will
- // be inserted in.
- Operation* GetFirstOutputOp() { return outputs.begin()->second.ops.at(0); }
-
- // Since we have different aggregation strategies, e.g., "first", "last",
- // "stack". We don't somehow aggregated to get the outputs for the funcOp.
- // This function is simply compute the RankedTensorType (shape & element type)
- std::map<int, Type> GetAggregatedOutputTypes(OpBuilder* builder) {
- std::map<int, Type> aggregated_output_types;
- for (const auto& kv : outputs) {
- const AggregatedOperand& operand = kv.second;
- if (operand.aggregation == kStrategyStack) {
- const int output_numer = operand.ops.size();
- Value first_output = operand.ops.at(0)->getOperand(0);
- RankedTensorType first_output_type =
- first_output.getType().cast<RankedTensorType>();
- // The aggregated output shape will be {N, original_shape}.
- SmallVector<int64_t, 4> shape;
- shape.push_back(output_numer);
- for (const auto& dim : first_output_type.getShape()) {
- shape.push_back(dim);
- }
- aggregated_output_types[kv.first] =
- RankedTensorType::get(shape, first_output_type.getElementType());
- } else if (operand.aggregation == kStrategyLast) {
- Value last_output =
- operand.ops.at(operand.ops.size() - 1)->getOperand(0);
- aggregated_output_types[kv.first] = last_output.getType();
- } else {
- Value first_output = operand.ops.at(0)->getOperand(0);
- aggregated_output_types[kv.first] = first_output.getType();
- }
- }
- return aggregated_output_types;
- }
-
- void AggregateAndRewireOutputs(OpBuilder* builder, Operation* fused_op) {
- // TODO(renjieliu): Consider get rid of the ophinted identity nodes here
- // as well or just rely on the general path to get rid of the identity
- // nodes.
- int output_index = 0;
- for (const auto& kv : outputs) {
- const AggregatedOperand& operand = kv.second;
- // This handles the "stack" strategy. It push a unpack_op before all the
- // outputs and make all the outputs point to the unpack_op.
- if (operand.aggregation == kStrategyStack) {
- // TODO(renjieliu): Revisit here if we need to handle
- // operand.ops().size() == 1 case. Insert a unpack op to unpack the
- // outputs.
- const int output_number = operand.ops.size();
- // Find the first output.
- Operation* first_output = operand.ops.at(0);
- Location insert_loc = first_output->getLoc();
- SmallVector<Type, 4> unpack_output_types(
- output_number, first_output->getOperand(0).getType());
-
- builder->setInsertionPoint(first_output);
- Operation* unpack_op = builder->create<TFL::UnpackOp>(
- insert_loc, unpack_output_types, fused_op->getResult(output_index),
- builder->getI32IntegerAttr(output_number),
- builder->getI32IntegerAttr(0));
- // For every unpack output, make sure they point to the right ones.
- for (int i = 0; i < output_number; ++i) {
- Operation* to_be_replaced_op = operand.ops.at(i);
- to_be_replaced_op->replaceUsesOfWith(to_be_replaced_op->getOperand(0),
- unpack_op->getResult(i));
- }
- } else if (operand.aggregation == kStrategyLast) {
- // This handles the strategy "last", it simply takes the last output.
- Operation* op = operand.ops.at(operand.ops.size() - 1);
- op->replaceUsesOfWith(op->getOperand(0),
- fused_op->getResult(output_index));
- } else {
- // This handles the strategy "first" and default, it simply takes the
- // first output.
- Operation* op = operand.ops.at(0);
- op->replaceUsesOfWith(op->getOperand(0),
- fused_op->getResult(output_index));
- }
-
- output_index++;
- }
- }
-
- LogicalResult VerifyOphint() const {
- if (inputs.empty() || outputs.empty()) return failure();
- return success();
- }
-
- StringRef uuid;
- StringRef function_name;
-
- private:
- // The AggregatedOperand is used to hold one "aggregated operand".
- // For example, this can be
- // {
- // aggregation = "stack",
- // {0: ident_op1, 1: ident_op2, 2: ident_op3}
- // }
- struct AggregatedOperand {
- StringRef aggregation;
- std::map<int, Operation*> ops;
- };
-
- std::map<int, AggregatedOperand> inputs;
- std::map<int, AggregatedOperand> outputs;
-};
-
-// Preprocess the graph for topo sort. (each operation is a node, while
-// inputs/outputs indicate edges) Assume the graph is acyclic. The preprocess
-// does the following:
-// Compute each operations's in-degress (how many input nodes they're taken)
-// Get all consumer operations for every operations. (operation_to_outputs)
-// Get the init_queue (those operations will be processed first).
-void PreprocessTopoSortGraph(
- Block* block, std::queue<Operation*>* init_queue,
- llvm::DenseMap<Operation*, llvm::DenseSet<Operation*>>*
- operation_to_outputs,
- llvm::DenseMap<Operation*, int>* operation_to_in_degrees) {
- for (auto& op : *block) {
- if (&op == block->getTerminator()) continue;
- if (op.getNumOperands() == 0) {
- init_queue->push(&op);
- } else {
- // The operand of the ops is not a direct indication of the "edge" as we
- // can have a pack op after a unpack op (they have multiple edges), we
- // should only count as one.
- llvm::DenseSet<Operation*> input_ops;
- for (int i = 0; i < op.getNumOperands(); ++i) {
- Operation* input_op = op.getOperand(i).getDefiningOp();
- if (input_op) input_ops.insert(input_op);
- }
- if (input_ops.empty()) {
- init_queue->push(&op);
- continue;
- }
- operation_to_in_degrees->try_emplace(&op, input_ops.size());
- for (auto* input_op : input_ops) {
- auto preceding_op_it = operation_to_outputs->find(input_op);
- if (preceding_op_it == operation_to_outputs->end()) {
- auto result = operation_to_outputs->try_emplace(
- input_op, llvm::DenseSet<Operation*>());
- preceding_op_it = result.first;
- }
- preceding_op_it->second.insert(&op);
- }
- }
- }
-}
-
-bool IsSideEffectOp(Operation* op) {
- // TODO(riverriddle) Properly handle region side effects.
- if (MemoryEffectOpInterface::hasNoEffect(op) && op->getNumRegions() == 0)
- return false;
-
- // Identity op has no side effect.
- // Check the OperationName maybe more elegant here.
- auto tf_identity_op = dyn_cast_or_null<TF::IdentityOp>(op);
- if (tf_identity_op) return false;
- return true;
-}
-
-// It's possible other transformations can benefit from this util function, but
-// since currently there's none, so we only limit this function to the ophint
-// extraction pass. We may refactor this function to extend the usage in future.
-//
-// Assume the graph is disconnected from outside.
-// Also assume the block has no arguments.
-LogicalResult TopoSortOperations(OpBuilder* builder) {
- std::queue<Operation*> init_queue;
- llvm::DenseMap<Operation*, llvm::DenseSet<Operation*>> operation_to_outputs;
- llvm::DenseMap<Operation*, int> operation_to_in_degrees;
- std::vector<Operation*> sorted_ops;
-
- PreprocessTopoSortGraph(builder->getBlock(), &init_queue,
- &operation_to_outputs, &operation_to_in_degrees);
- while (!init_queue.empty()) {
- Operation* current_op = init_queue.front();
- init_queue.pop();
- sorted_ops.push_back(current_op);
-
- auto current_op_to_output_it = operation_to_outputs.find(current_op);
- if (current_op_to_output_it == operation_to_outputs.end()) {
- continue;
- }
- for (Operation* output_op : current_op_to_output_it->second) {
- auto output_op_it = operation_to_in_degrees.find(output_op);
- if (output_op_it == operation_to_in_degrees.end()) return failure();
-
- output_op_it->second -= 1;
- if (output_op_it->second == 0) {
- init_queue.push(output_op);
- operation_to_in_degrees.erase(output_op_it);
- }
- }
- operation_to_outputs.erase(current_op_to_output_it);
- }
-
- // Before we performs the sort. We need to make sure we didn't mess the
- // ordering of original side-effect operations.
- // It's possible those side-effect operations have no topological relations
- // at all!
- std::vector<Operation*> original_side_effect_ops;
- std::vector<Operation*> after_sort_side_effect_ops;
- for (auto& op : *builder->getBlock()) {
- if (IsSideEffectOp(&op) && (&op != builder->getBlock()->getTerminator()))
- original_side_effect_ops.push_back(&op);
- }
- for (auto* op : sorted_ops) {
- if (IsSideEffectOp(op)) after_sort_side_effect_ops.push_back(op);
- }
- if (original_side_effect_ops.size() != after_sort_side_effect_ops.size())
- return failure();
- for (int i = 0; i < original_side_effect_ops.size(); ++i) {
- if (original_side_effect_ops[i] != after_sort_side_effect_ops[i])
- return failure();
- }
-
- // Performs the sort.
- // Ideally it would be nice to just clear the block then write the sorted ops.
- // But unfortunately that's hard to do.
- for (int i = sorted_ops.size() - 1; i > 0; --i) {
- Operation* current_op = sorted_ops[i];
- for (int j = i - 1; j >= 0; --j) {
- Operation* prev_op = sorted_ops[j];
- prev_op->moveBefore(current_op);
- }
- }
-
- return success();
-}
-
-Operation* BuildFusedFuncOp(StringRef func_name, StringRef fused_func_type,
- Operation* insert_before_op,
- const std::map<int, Value>& inputs,
- const std::map<int, Type>& output_types,
- OpBuilder* builder, ModuleOp* module_op) {
- SmallVector<Type, 4> input_types;
- SmallVector<Value, 4> input_values;
- SmallVector<int, 4> input_indexes;
- for (const auto& kv : inputs) {
- Value input = kv.second;
- input_types.push_back(input.getType());
- input_values.push_back(input);
- input_indexes.push_back(kv.first);
- }
-
- SmallVector<Type, 4> func_output_types;
- for (const auto& kv : output_types) {
- func_output_types.push_back(kv.second);
- }
-
- FunctionType function_type =
- builder->getFunctionType(/*inputs=*/input_types,
- /*results=*/func_output_types);
-
- SmallVector<NamedAttribute, 4> attrs;
- attrs.push_back(builder->getNamedAttr(
- kTfLiteFunctionName, builder->getStringAttr(fused_func_type)));
- attrs.push_back(builder->getNamedAttr(
- kTfLiteFunctionInputIndex, builder->getI32ArrayAttr(input_indexes)));
- FuncOp func_op = FuncOp::create(insert_before_op->getLoc(), func_name,
- function_type, llvm::makeArrayRef(attrs));
- module_op->push_back(func_op);
- builder->setInsertionPoint(insert_before_op);
- return builder->create<CallOp>(insert_before_op->getLoc(), func_op,
- input_values);
-}
-
-llvm::StringMap<OphintCompositeOp> FindAllOphintNodes(Block* bb) {
- llvm::StringMap<OphintCompositeOp> ophint_composite_ops;
- for (auto& op : *bb) {
- auto nameAttr = op.getAttrOfType<StringAttr>(kTfLiteFunctionName);
- if (!nameAttr) continue;
- StringRef function_name = nameAttr.getValue();
- auto uuidAttr = op.getAttrOfType<StringAttr>(kTfLiteFunctionUUID);
- if (!uuidAttr) continue;
- StringRef uuid = uuidAttr.getValue();
- auto it = ophint_composite_ops.find(uuid);
- if (it == ophint_composite_ops.end()) {
- OphintCompositeOp ophint_composite_op(uuid, function_name);
- it = ophint_composite_ops.insert({uuid, ophint_composite_op}).first;
- }
-
- // The default aggregation strategy is "NONE".
- StringRef aggregation = kStrategyNone;
- auto aggregationAttr =
- op.getAttrOfType<StringAttr>(kTfLiteFunctionAggregate);
- if (aggregationAttr != nullptr) aggregation = aggregationAttr.getValue();
-
- // The default sort index is 0.
- int sortIndex = 0;
- auto sortIndexAttr =
- op.getAttrOfType<IntegerAttr>(kTfLiteFunctionSortIndex);
- if (sortIndexAttr != nullptr) sortIndex = sortIndexAttr.getInt();
-
- auto inputIndexAttr =
- op.getAttrOfType<IntegerAttr>(kTfLiteFunctionInputIndex);
- if (inputIndexAttr != nullptr) {
- it->second.AddInput(inputIndexAttr.getInt(), &op, aggregation, sortIndex);
- } else {
- auto outputIndexAttr =
- op.getAttrOfType<IntegerAttr>(kTfLiteFunctionOutputIndex);
- it->second.AddOutput(outputIndexAttr.getInt(), &op, aggregation,
- sortIndex);
- }
- }
-
- return ophint_composite_ops;
-}
-
-llvm::DenseSet<Operation*> BfsForReachableOps(ArrayRef<Operation*> input_ops) {
- llvm::DenseSet<Operation*> reachable_ops;
- std::queue<Operation*> ops_queue;
- for (auto& input_op : input_ops) {
- for (Value value : input_op->getOperands()) {
- Operation* op = value.getDefiningOp();
- if (op != nullptr) ops_queue.push(op);
- }
- }
-
- while (!ops_queue.empty()) {
- Operation* current_op = ops_queue.front();
- ops_queue.pop();
- reachable_ops.insert(current_op);
- for (Value value : current_op->getOperands()) {
- Operation* upstream_op = value.getDefiningOp();
- // Not visited, put it into the queue.
- if (upstream_op != nullptr &&
- !llvm::is_contained(reachable_ops, upstream_op)) {
- ops_queue.emplace(upstream_op);
- }
- }
- }
-
- return reachable_ops;
-}
-
-// Convert ophint to stub will remove all ops within the ophint region and
-// place a new fused op right before the first op.
-LogicalResult ConvertOphintToStub(StringRef stub_name,
- OphintCompositeOp ophint_composite_op,
- OpBuilder* builder, ModuleOp* module_op) {
- // Step 1, find all ops reachable by inputs.
- const llvm::DenseSet<Operation*>& reachable_by_inputs =
- BfsForReachableOps(ophint_composite_op.GetAllInputOps());
-
- // Step 2, find all ops reachable by outputs.
- const llvm::DenseSet<Operation*>& reachable_by_outputs =
- BfsForReachableOps(ophint_composite_op.GetAllOutputOps());
-
- // Step 3, deal with inputs aggregation strategies.
- const std::map<int, Value>& aggregated_inputs =
- ophint_composite_op.GetAggregatedInputs(builder);
-
- // Step 4, get aggregated output types.
- const std::map<int, Type>& aggregated_output_types =
- ophint_composite_op.GetAggregatedOutputTypes(builder);
-
- // Step 5, create & place the fused op and rewire the inputs.
- // Here we use a funcOp to represent the fused op. This "funcOp" will be
- // converted to other ops (like UnidirectionalSequenceRNNOp) in the
- // legalization phase.
- Operation* inserted_before_op = ophint_composite_op.GetFirstOutputOp();
- Operation* fused_op = BuildFusedFuncOp(
- stub_name, ophint_composite_op.function_name, inserted_before_op,
- aggregated_inputs, aggregated_output_types, builder, module_op);
-
- for (const auto& kv : aggregated_inputs) {
- Operation* op = kv.second.getDefiningOp();
- if (op == nullptr) return failure();
- op->moveBefore(fused_op);
- }
-
- // Step 6, deal with outputs aggregation strategies and rewire the outputs.
- ophint_composite_op.AggregateAndRewireOutputs(builder, fused_op);
-
- // Step 7, remove all the removable ops where
- // (reachable_by_outputs - reachable_by_inputs) as removable and the rest
- // ops are not removable.
- // We also need to make sure all the output identity nodes are there.
- llvm::DenseSet<Operation*> ophinted_identity_nodes;
- for (auto* output : ophint_composite_op.GetAllInUseOutputOps()) {
- ophinted_identity_nodes.insert(output);
- }
-
- auto removeRemovableOps = [&](Operation* op) {
- if (reachable_by_inputs.count(op) == 0 &&
- reachable_by_outputs.count(op) != 0 &&
- ophinted_identity_nodes.count(op) == 0) {
- op->dropAllDefinedValueUses();
- op->dropAllReferences();
- op->erase();
- }
- };
-
- builder->getBlock()->walk(removeRemovableOps);
-
- // Step 8: Topo sort to fix any invalid temporary IRs.
- if (failed(TopoSortOperations(builder))) return failure();
-
- return success();
-}
-
-struct ExtractOphintPass
- : public PassWrapper<ExtractOphintPass, OperationPass<ModuleOp>> {
- void runOnOperation() override;
- void Verify();
-
- private:
- int ophint_composite_ops_count = 0;
-};
-
-// TODO(renjieliu): Current ophint extraction does not support inputs/outputs
-// cross functions, we need to do that.
-void ExtractOphintPass::runOnOperation() {
- ModuleOp module = getOperation();
- for (auto function : module.getOps<FuncOp>()) {
- // Process block by block.
- for (auto& bb : function.getBody()) {
- // Find ophints.
- const llvm::StringMap<OphintCompositeOp>& ophint_composite_ops =
- FindAllOphintNodes(&bb);
- if (ophint_composite_ops.empty()) continue;
-
- // Verify: Make sure all ophint_composite_ops are valid.
- // If not valid, we just don't do anything.
- for (const auto& kv : ophint_composite_ops) {
- if (failed(kv.getValue().VerifyOphint())) {
- return;
- }
- }
-
- ophint_composite_ops_count = ophint_composite_ops.size();
-
- // Convert.
- OpBuilder builder = OpBuilder::atBlockEnd(&bb);
- for (const auto& kv : ophint_composite_ops) {
- if (failed(ConvertOphintToStub(kv.getKey(), kv.getValue(), &builder,
- &module))) {
- module.emitError()
- << "Convert ophint failed, malformed inputs or outputs.";
- return signalPassFailure();
- }
- }
- }
- }
-}
-
-void ExtractOphintPass::Verify() {
- ModuleOp module = getOperation();
- int ophint_func_op_count = 0;
- for (FuncOp func : getOperation().getOps<FuncOp>()) {
- for (const NamedAttribute attr : func.getAttrs()) {
- if (attr.first == kTfLiteFunctionName) {
- ophint_func_op_count++;
- if (func.getNumArguments() == 0) {
- module.emitError() << "Ophint function has no inputs.";
- return signalPassFailure();
- }
- if (func.getType().getNumResults() == 0) {
- module.emitError() << "Ophint function has no outputs.";
- return signalPassFailure();
- }
- }
- }
- }
- if (ophint_func_op_count != ophint_composite_ops_count) {
- module.emitError()
- << "Ophint converted functions do not match ophint regions founded.";
- return signalPassFailure();
- }
-}
-
-} // namespace
-
-/// Creates an instance of the TensorFlow Lite dialect ExtractOphintPass
-/// pass.
-std::unique_ptr<OperationPass<ModuleOp>> CreateExtractOphintPass() {
- return std::make_unique<ExtractOphintPass>();
-}
-
-static PassRegistration<ExtractOphintPass> pass(
- "tfl-extract-ophint", "Extract Ophint for TfLite dialect.");
-
-} // namespace TFL
-} // namespace mlir
diff --git a/tensorflow/compiler/mlir/lite/transforms/legalize_ophint_func_op.cc b/tensorflow/compiler/mlir/lite/transforms/legalize_ophint_func_op.cc
deleted file mode 100644
index 652d10a..0000000
--- a/tensorflow/compiler/mlir/lite/transforms/legalize_ophint_func_op.cc
+++ /dev/null
@@ -1,295 +0,0 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/StringMap.h"
-#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
-#include "mlir/IR/Attributes.h" // from @llvm-project
-#include "mlir/IR/Block.h" // from @llvm-project
-#include "mlir/IR/Builders.h" // from @llvm-project
-#include "mlir/IR/Function.h" // from @llvm-project
-#include "mlir/IR/MLIRContext.h" // from @llvm-project
-#include "mlir/IR/Module.h" // from @llvm-project
-#include "mlir/IR/Operation.h" // from @llvm-project
-#include "mlir/IR/OperationSupport.h" // from @llvm-project
-#include "mlir/IR/StandardTypes.h" // from @llvm-project
-#include "mlir/IR/SymbolTable.h" // from @llvm-project
-#include "mlir/IR/Types.h" // from @llvm-project
-#include "mlir/IR/Value.h" // from @llvm-project
-#include "mlir/Pass/Pass.h" // from @llvm-project
-#include "mlir/Pass/PassRegistry.h" // from @llvm-project
-#include "mlir/Support/LLVM.h" // from @llvm-project
-#include "mlir/Support/LogicalResult.h" // from @llvm-project
-#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"
-
-namespace mlir {
-namespace TFL {
-namespace {
-
-constexpr char kTfLiteFunctionName[] = "_tflite_function_name";
-constexpr char kTfLiteFunctionInputIndex[] = "_tflite_function_input_index";
-constexpr char kUnidirectionalSequenceRnn[] = "UnidirectionalSequenceRnn";
-constexpr char kUnidirectionalSequenceLstm[] = "UnidirectionalSequenceLstm";
-
-// This pass is used for converting to TFLite composite op like
-// UnidirectionalSequenceRNN, UnidirectionalSequenceLSTM or SVDF Op. Currently,
-// this pass is only for ophint converted function op only. See below diagram:
-//
-// InputOp1 InputOp2 ...
-// \ /
-// \ /
-// call funcOp (say UnidirectionalSequenceRNN)
-// |
-// |
-// OutputOp1
-//
-// funcOp() { '_tflite_function_name' = 'UnidirectionalSequenceRNN'}
-//
-// ||
-// ||
-// \ /
-//
-// InputOp1 InputOp2 ...
-// \ /
-// \ /
-// tfl.UnidirectionalSequenceRNN
-// |
-// |
-// OutputOp1
-struct LegalizeOphintFuncOpPass
- : public PassWrapper<LegalizeOphintFuncOpPass, OperationPass<ModuleOp>> {
- void runOnOperation() override;
-};
-
-llvm::StringMap<FuncOp> FindCompositeFuncOps(ModuleOp module) {
- llvm::StringMap<FuncOp> composite_func_ops;
- for (FuncOp func : module.getOps<FuncOp>()) {
- if (func.getAttr(kTfLiteFunctionName))
- composite_func_ops[func.getName()] = func;
- }
- return composite_func_ops;
-}
-
-LogicalResult BuildUnidirectionalSequenceRnnOp(FuncOp composite_func_op,
- CallOp call_op,
- OpBuilder* builder,
- Operation** fused_op) {
- // UnidirectionalSequenceRnn takes exactly 5 inputs.
- if (composite_func_op.getNumArguments() != 5) return failure();
- if (call_op.getNumOperands() != 5) return failure();
- // UnidirectionalSequenceRnn has exactly 1 input.
- if (call_op.getNumResults() != 1) return failure();
-
- // Inputs is indexed at 0.
- Value input = call_op.getOperand(0);
- // Input_weight is indexed at 1.
- Value weight = call_op.getOperand(1);
- // Recurrent_weight is indexed at 2.
- Value recurrent_weight = call_op.getOperand(2);
- // Bias is indexed at 3.
- Value bias = call_op.getOperand(3);
- // Hidden_state is indexed at 4.
- Value hidden_state = call_op.getOperand(4);
-
- // Build Output.
- auto output_type = call_op.getResult(0).getType();
-
- // Currently, ophinted RNN only supports time_major = True.
- const bool time_major = true;
- // Activation will always be TanH.
- StringAttr fused_activation_function = builder->getStringAttr("TANH");
-
- builder->setInsertionPoint(call_op.getOperation());
- *fused_op = builder->create<TFL::UnidirectionalSequenceRNNOp>(
- call_op.getLoc(), output_type, input, weight, recurrent_weight, bias,
- hidden_state, builder->getBoolAttr(time_major),
- fused_activation_function);
- return success();
-}
-
-LogicalResult BuildUnidirectionalSequenceLSTMOp(FuncOp composite_func_op,
- CallOp call_op,
- OpBuilder* builder,
- Operation** fused_op) {
- if (composite_func_op.getNumArguments() != call_op.getNumOperands())
- return failure();
- auto input_index_attr = composite_func_op.getAttr(kTfLiteFunctionInputIndex)
- .cast<ArrayAttr>()
- .getValue();
- llvm::DenseMap<int, Value> fused_ops_index_to_call_op_args;
-
- for (int i = 0; i < call_op.getNumOperands(); ++i) {
- int input_index = input_index_attr[i].cast<IntegerAttr>().getInt();
- fused_ops_index_to_call_op_args.try_emplace(input_index,
- call_op.getOperand(i));
- }
-
- constexpr int kUnidirectionalSequenceLSTMOpTotalIArgumentNum = 24;
-
- // We encounter some optional arguments not filled, so we need to create an
- // empty Value.
- Value none_value;
- if (call_op.getNumOperands() <
- kUnidirectionalSequenceLSTMOpTotalIArgumentNum) {
- builder->setInsertionPoint(call_op.getOperation());
- none_value = builder->create<mlir::ConstantOp>(
- call_op.getLoc(), builder->getNoneType(), builder->getUnitAttr());
- }
-
- // Prepare all operands for the UnidirectionalSequenceLSTMOp.
- SmallVector<Value, kUnidirectionalSequenceLSTMOpTotalIArgumentNum> operands;
- for (int i = 0; i < kUnidirectionalSequenceLSTMOpTotalIArgumentNum; ++i) {
- auto operand_it = fused_ops_index_to_call_op_args.find(i);
- if (operand_it == fused_ops_index_to_call_op_args.end()) {
- // Encounter optional arguments.
- operands.push_back(none_value);
- } else {
- operands.push_back(operand_it->second);
- }
- }
-
- // Prepare output types.
- SmallVector<Type, 4> output_types;
- // The output type set is somewhat adhoc here: The fused op only have exact
- // one output while the call_op can have more than one output. (but we only
- // take the last one).
- // And here we check the outputs are not used (except the last one) if the
- // call_op has more than one output.
- if (call_op.getNumResults() > 1) {
- for (int i = 0; i < call_op.getNumResults() - 1; ++i) {
- // This one should not be used.
- Value unused_output = call_op.getResult(i);
- if (!unused_output.use_empty()) return failure();
- }
- }
- output_types.push_back(
- call_op.getResult(call_op.getNumResults() - 1).getType());
-
- // Prepare attributes.
- SmallVector<NamedAttribute, 4> attributes;
- attributes.push_back(builder->getNamedAttr("fused_activation_function",
- builder->getStringAttr("TANH")));
- attributes.push_back(
- builder->getNamedAttr("time_major", builder->getBoolAttr(true)));
-
- builder->setInsertionPoint(call_op.getOperation());
-
- *fused_op = builder->create<TFL::UnidirectionalSequenceLSTMOp>(
- call_op.getLoc(), output_types, operands, attributes);
-
- return success();
-}
-
-LogicalResult ConvertTfLiteFusedOpIfAvailable(StringRef func_name,
- FuncOp composite_func_op,
- CallOp call_op,
- OpBuilder* builder) {
- Operation* fused_op = nullptr;
- if (func_name == kUnidirectionalSequenceRnn) {
- // TODO(renjieliu): Validate the func op inputs.
- LogicalResult build_fused_op_result = BuildUnidirectionalSequenceRnnOp(
- composite_func_op, call_op, builder, &fused_op);
- if (failed(build_fused_op_result)) return build_fused_op_result;
- call_op.replaceAllUsesWith(fused_op);
- } else if (func_name == kUnidirectionalSequenceLstm) {
- LogicalResult build_fused_op_result = BuildUnidirectionalSequenceLSTMOp(
- composite_func_op, call_op, builder, &fused_op);
- if (failed(build_fused_op_result)) return build_fused_op_result;
- Value call_output = call_op.getResult(call_op.getNumResults() - 1);
- if (call_output.getType() != fused_op->getResult(0).getType()) {
- return failure();
- }
- call_output.replaceAllUsesWith(fused_op->getResult(0));
- } else { // If we support more fused op, we should add the conversion here.
- return failure();
- }
-
- // Delete call op.
- Operation* call = call_op.getOperation();
- call->dropAllDefinedValueUses();
- call->dropAllReferences();
- call->erase();
- return success();
-}
-
-LogicalResult ConvertCallOps(llvm::StringMap<FuncOp>* composite_func_ops,
- ModuleOp* module) {
- for (auto func : module->getOps<FuncOp>()) {
- // Ideally it will be much simpler if we can just use walk, but we also
- // want to early return if encounter errors. :(
- OpBuilder builder(func.getBody());
- // The call_op replacement within this loop works like an in-place
- // replacement, so it should be safe to do so.
- for (auto call_op :
- llvm::make_early_inc_range(builder.getBlock()->getOps<CallOp>())) {
- auto it = composite_func_ops->find(call_op.getCallee());
- if (it == composite_func_ops->end()) return failure();
-
- // Replace the call op with TfLite fused op.
- // Currently it's only handled case by case, but ideally it would be
- // much better if we can do this automatically.
- FuncOp composite_func_op = it->second;
- StringRef func_name = composite_func_op.getAttr(kTfLiteFunctionName)
- .cast<StringAttr>()
- .getValue();
- if (failed(ConvertTfLiteFusedOpIfAvailable(func_name, composite_func_op,
- call_op, &builder)))
- return failure();
-
- composite_func_ops->erase(it);
- // Delete func op.
- Operation* func = composite_func_op.getOperation();
- func->erase();
- }
- }
- return success();
-}
-
-void LegalizeOphintFuncOpPass::runOnOperation() {
- ModuleOp module = getOperation();
-
- // Find all composite funcs, then for every call op inside every func op
- // within the module, we go ahead and replace the callop with the tflite
- // corresponding op and destroy the func op. This two-phase processing is
- // intended:
- //
- // Every func op is meant to be used exactly once.
- // Instead of finding the composite func then loop through the graph and
- // convert the call op immediately, we break finding & converting into two
- // phases. This changes the complexity from O(op_in_module *
- // function_in_module * attr_in_func) to O(op_in_module) * O(map_look_up) +
- // O(function_in_module * attr_in_func). O(op_in_module) is the dominant
- // factor here and map look up should be very cheap.
- llvm::StringMap<FuncOp> composite_func_ops = FindCompositeFuncOps(module);
- if (composite_func_ops.empty()) return;
- if (failed(ConvertCallOps(&composite_func_ops, &module))) {
- module.emitError() << "Legalize ophint: ConvertCallOps failed.";
- return signalPassFailure();
- }
-}
-
-} // namespace
-
-/// Creates an instance of the TensorFlow Lite dialect LegalizeOphintFuncOpPass
-/// pass.
-std::unique_ptr<OperationPass<ModuleOp>> CreateLegalizeOphintFuncOpPass() {
- return std::make_unique<LegalizeOphintFuncOpPass>();
-}
-
-static PassRegistration<LegalizeOphintFuncOpPass> pass(
- "tfl-legalize-ophint-func-op", "Convert composite op for TfLite dialect.");
-
-} // namespace TFL
-} // namespace mlir
diff --git a/tensorflow/compiler/mlir/lite/transforms/legalize_patterns.td b/tensorflow/compiler/mlir/lite/transforms/legalize_patterns.td
index 586ddf6..12796b8 100644
--- a/tensorflow/compiler/mlir/lite/transforms/legalize_patterns.td
+++ b/tensorflow/compiler/mlir/lite/transforms/legalize_patterns.td
@@ -58,6 +58,9 @@
def HasSameStaticShapes : Constraint<HasSameStaticShapesPred, "op must have static same input shapes">;
def HasNotSameStaticShapes : Constraint<Neg<HasSameStaticShapesPred>, "op must have not static same input shapes">;
+def CreateNoneValue : NativeCodeCall<
+ "$_builder.create<mlir::ConstantOp>($0.getLoc(), $_builder.getNoneType(), $_builder.getUnitAttr())">;
+
// Checks if the value has only one user.
// TODO(karimnosseir): Move to a common place?
def HasOneUse : Constraint<CPred<"$0.hasOneUse()">>;
@@ -343,6 +346,7 @@
(TF_TransposeOp $filter,
(ConstantOp ConstantAttr<I32VectorElementsAttr<4>, "{2, 0, 1, 3}">)),
$out_backprop,
+ /*bias=*/ (CreateNoneValue $input_sizes),
/*padding=*/ $padding,
/*stride_h=*/ ExtractI32At<1>:$strides,
/*stride_w=*/ ExtractI32At<2>:$strides)>;
diff --git a/tensorflow/compiler/mlir/lite/transforms/passes.h b/tensorflow/compiler/mlir/lite/transforms/passes.h
index a744a57..959c17e 100644
--- a/tensorflow/compiler/mlir/lite/transforms/passes.h
+++ b/tensorflow/compiler/mlir/lite/transforms/passes.h
@@ -67,13 +67,6 @@
// pass.
std::unique_ptr<OperationPass<ModuleOp>> CreatePrepareCompositeFunctionsPass();
-// Creates an instance of the TensorFlow Lite dialect ExtractOphint pass.
-std::unique_ptr<OperationPass<ModuleOp>> CreateExtractOphintPass();
-
-// Creates an instance of the TensorFlow Lite dialect LegalizeOphintFuncOpPass
-// pass. The composite op is created from the ophint extraction pass.
-std::unique_ptr<OperationPass<ModuleOp>> CreateLegalizeOphintFuncOpPass();
-
// Creates an instance of the TensorFlow Lite dialect SplitMergedOperandsPass.
std::unique_ptr<OperationPass<FuncOp>> CreateSplitMergedOperandsPass();
diff --git a/tensorflow/compiler/mlir/op_or_arg_name_mapper.cc b/tensorflow/compiler/mlir/op_or_arg_name_mapper.cc
index 272fab9..bce0ed4 100644
--- a/tensorflow/compiler/mlir/op_or_arg_name_mapper.cc
+++ b/tensorflow/compiler/mlir/op_or_arg_name_mapper.cc
@@ -55,8 +55,10 @@
// to be unique.
auto& val = prefix_it.first->second;
llvm::SmallString<64> probe_name(prefix);
+ probe_name.append(GetSuffixSeparator());
+ const int probe_prefix_size = probe_name.size();
while (true) {
- probe_name.resize(prefix.size());
+ probe_name.resize(probe_prefix_size);
// TODO(jpienaar): Subtract one so that the initial suffix is 0 instead
// of 1.
// TODO(jpienaar): Switch to radix 36 and update tests.
diff --git a/tensorflow/compiler/mlir/op_or_arg_name_mapper.h b/tensorflow/compiler/mlir/op_or_arg_name_mapper.h
index 108496e..6a52d13 100644
--- a/tensorflow/compiler/mlir/op_or_arg_name_mapper.h
+++ b/tensorflow/compiler/mlir/op_or_arg_name_mapper.h
@@ -64,6 +64,9 @@
return op_or_val_to_name_;
}
+ // Returns the separator used before uniqueing suffix.
+ virtual llvm::StringRef GetSuffixSeparator() { return ""; }
+
private:
// Returns name from the location of the operation or value.
virtual std::string GetName(OpOrVal op_or_val) = 0;
diff --git a/tensorflow/compiler/mlir/tensorflow/BUILD b/tensorflow/compiler/mlir/tensorflow/BUILD
index 4305d64..5a24ae3 100644
--- a/tensorflow/compiler/mlir/tensorflow/BUILD
+++ b/tensorflow/compiler/mlir/tensorflow/BUILD
@@ -10,6 +10,7 @@
name = "friends",
includes = ["//third_party/mlir:subpackages"],
packages = [
+ "//learning/brain/experimental/dtensor/...",
"//learning/brain/experimental/tfrt/...",
"//learning/pathways/data_parallel/tf2xla/...",
"//tensorflow/compiler/...",
@@ -34,6 +35,7 @@
"ir/tf_ops.td",
"@llvm-project//mlir:OpBaseTdFiles",
"@llvm-project//mlir:include/mlir/Interfaces/CallInterfaces.td",
+ "@llvm-project//mlir:include/mlir/Interfaces/InferTypeOpInterface.td",
"@llvm-project//mlir:include/mlir/Interfaces/SideEffects.td",
],
)
@@ -281,6 +283,7 @@
"@llvm-project//mlir:DerivedAttributeOpInterface",
"@llvm-project//mlir:Dialect",
"@llvm-project//mlir:IR",
+ "@llvm-project//mlir:InferTypeOpInterface",
"@llvm-project//mlir:Parser",
"@llvm-project//mlir:Pass",
"@llvm-project//mlir:SideEffects",
@@ -1014,6 +1017,7 @@
name = "derived_attr_populator_inc",
srcs = [
"@llvm-project//mlir:include/mlir/Interfaces/CallInterfaces.td",
+ "@llvm-project//mlir:include/mlir/Interfaces/InferTypeOpInterface.td",
"@llvm-project//mlir:include/mlir/Interfaces/SideEffects.td",
"@llvm-project//mlir:include/mlir/IR/OpBase.td",
"ir/tf_generated_ops.td",
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
index 3c52704..8be7c69 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td
@@ -2185,8 +2185,8 @@
}];
let arguments = (ins
- TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint8]>:$x,
- TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint8]>:$y,
+ TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$x,
+ TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$y,
DefaultValuedAttr<BoolAttr, "true">:$incompatible_shape_error
);
@@ -3279,22 +3279,6 @@
TF_DerivedResultTypeAttr Tout = TF_DerivedResultTypeAttr<0>;
}
-def TF_InfeedDequeueTupleOp : TF_Op<"InfeedDequeueTuple", []> {
- let summary = "Fetches multiple values from infeed as an XLA tuple.";
-
- let description = [{
- }];
-
- let arguments = (ins);
-
- let results = (outs
- Variadic<TF_Tensor>:$outputs
- );
-
- TF_DerivedResultShapeListAttr shapes = TF_DerivedResultShapeListAttr<0>;
- TF_DerivedResultTypeListAttr dtypes = TF_DerivedResultTypeListAttr<0>;
-}
-
def TF_InvOp : TF_Op<"Inv", [NoSideEffect, SameOperandsAndResultType]> {
let summary = "Computes the reciprocal of x element-wise.";
@@ -4730,12 +4714,12 @@
}];
let arguments = (ins
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$x,
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$y
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$x,
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$y
);
let results = (outs
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$z
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$z
);
TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -4778,12 +4762,12 @@
}];
let arguments = (ins
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$x,
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$y
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$x,
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$y
);
let results = (outs
- TensorOf<[BF16, F16, F32, F64, I16, I32, I64]>:$z
+ TensorOf<[BF16, F16, F32, F64, I16, I32, I64, TF_Uint8]>:$z
);
TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
@@ -5144,8 +5128,8 @@
}];
let arguments = (ins
- TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint8]>:$x,
- TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint8]>:$y,
+ TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$x,
+ TensorOf<[BF16, F16, F32, F64, I1, I16, I32, I64, I8, TF_Complex128, TF_Complex64, TF_Qint32, TF_Qint8, TF_Quint8, TF_Str, TF_Uint16, TF_Uint32, TF_Uint64, TF_Uint8]>:$y,
DefaultValuedAttr<BoolAttr, "true">:$incompatible_shape_error
);
@@ -5539,6 +5523,40 @@
TF_DerivedOperandTypeAttr Tidx = TF_DerivedOperandTypeAttr<1>;
}
+def TF_QrOp : TF_Op<"Qr", [NoSideEffect]> {
+ let summary = "Computes the QR decompositions of one or more matrices.";
+
+ let description = [{
+Computes the QR decomposition of each inner matrix in `tensor` such that
+`tensor[..., :, :] = q[..., :, :] * r[..., :,:])`
+
+```python
+# a is a tensor.
+# q is a tensor of orthonormal matrices.
+# r is a tensor of upper triangular matrices.
+q, r = qr(a)
+q_full, r_full = qr(a, full_matrices=True)
+```
+ }];
+
+ let arguments = (ins
+ TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$input,
+
+ DefaultValuedAttr<BoolAttr, "false">:$full_matrices
+ );
+
+ let results = (outs
+ TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$q,
+ TensorOf<[F16, F32, F64, TF_Complex128, TF_Complex64]>:$r
+ );
+
+ TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+
+ let verifier = [{
+ return Verify(*this);
+ }];
+}
+
def TF_QuantizeAndDequantizeOp : TF_Op<"QuantizeAndDequantize", [NoSideEffect, SameOperandsAndResultType]> {
let summary = "Use QuantizeAndDequantizeV2 instead.";
@@ -8856,6 +8874,32 @@
);
}
+def TF_TensorListScatterIntoExistingListOp : TF_Op<"TensorListScatterIntoExistingList", [NoSideEffect]> {
+ let summary = "Scatters tensor at indices in an input list.";
+
+ let description = [{
+Each member of the TensorList corresponds to one row of the input tensor,
+specified by the given index (see `tf.gather`).
+
+input_handle: The list to scatter into.
+tensor: The input tensor.
+indices: The indices used to index into the list.
+output_handle: The TensorList.
+ }];
+
+ let arguments = (ins
+ TF_VariantTensor:$input_handle,
+ TF_Tensor:$tensor,
+ I32Tensor:$indices
+ );
+
+ let results = (outs
+ TF_VariantTensor:$output_handle
+ );
+
+ TF_DerivedOperandTypeAttr element_dtype = TF_DerivedOperandTypeAttr<1>;
+}
+
def TF_TensorListSetItemOp : TF_Op<"TensorListSetItem", [NoSideEffect]> {
let summary = "";
@@ -9523,6 +9567,30 @@
let hasFolder = 1;
}
+def TF_VariableV2Op : TF_Op<"VariableV2", []> {
+ let summary = [{
+Holds state in the form of a tensor that persists across steps.
+ }];
+
+ let description = [{
+Outputs a ref to the tensor state so it may be read or modified.
+TODO(zhifengc/mrry): Adds a pointer to a more detail document
+about sharing states in tensorflow.
+ }];
+
+ let arguments = (ins
+ StrAttr:$shape,
+ StrAttr:$container,
+ StrAttr:$shared_name
+ );
+
+ let results = (outs
+ TF_Tensor:$ref
+ );
+
+ TF_DerivedResultTypeAttr dtype = TF_DerivedResultTypeAttr<0>;
+}
+
def TF_WhereOp : TF_Op<"Where", [NoSideEffect]> {
let summary = "Returns locations of nonzero / true values in a tensor.";
@@ -9889,25 +9957,6 @@
TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
}
-def TF_XlaShardingOp : TF_Op<"XlaSharding", [NoSideEffect]> {
- let summary = [{
-An op which shards the input based on the given sharding attribute.
- }];
-
- let description = [{
- }];
-
- let arguments = (ins
- TF_Tensor:$input
- );
-
- let results = (outs
- TF_Tensor:$output
- );
-
- TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
-}
-
def TF_XlaSvdOp : TF_Op<"XlaSvd", [NoSideEffect]> {
let summary = [{
Computes the eigen decomposition of a batch of self-adjoint matrices
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td
index 773025c..ec83fbf 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td
@@ -233,7 +233,8 @@
class TF_DerivedOperandSizeAttr<int idx> : DerivedAttr<
"size_t",
"auto range = getODSOperands(" # idx # ");\n"
- "return std::distance(range.begin(), range.end());">;
+ "return std::distance(range.begin(), range.end());",
+ [{ $_builder.getI64IntegerAttr($_self) }]>;
// A derived attribute that returns the element type of `idx`-th ODS-declared
// operand. If the `idx`-th operand is a variadic operand, then this attribute
@@ -251,7 +252,16 @@
"mlir::OperandElementTypeRange",
"auto values = getODSOperands(" # idx # ");\n"
"return {mlir::OperandElementTypeIterator(values.begin()), "
- "mlir::OperandElementTypeIterator(values.end())};"
+ "mlir::OperandElementTypeIterator(values.end())};",
+ [{
+ ArrayAttr::get(
+ [&]() {
+ llvm::SmallVector<Attribute, 4> ret;
+ for (auto t : $_self)
+ ret.push_back(TypeAttr::get(t));
+ return ret;
+ }(), $_ctx)
+ }]
>;
// A derived attribute that returns the shapes of the tensors in the actual
@@ -262,7 +272,9 @@
"mlir::TF::OperandShapeRange",
"auto values = getODSOperands(" # idx # ");\n"
"return {mlir::TF::OperandShapeIterator(values.begin()), "
- "mlir::TF::OperandShapeIterator(values.end())};"
+ "mlir::TF::OperandShapeIterator(values.end())};",
+ // TODO(jpienaar): Update post TensorShapeAttr landing.
+ [{ nullptr }]
>;
// A derived attribute that returns the size of `idx`-th ODS-declared variadic
@@ -270,7 +282,8 @@
class TF_DerivedResultSizeAttr<int idx> : DerivedAttr<
"size_t",
"auto range = getODSResults(" # idx # ");\n"
- "return std::distance(range.begin(), range.end());">;
+ "return std::distance(range.begin(), range.end());",
+ [{ $_builder.getI64IntegerAttr($_self) }]>;
// A derived attribute that returns the element type of `idx`-th ODS-declared
// result. If the `idx`-th result is a variadic result, then this attribute
@@ -288,7 +301,16 @@
"mlir::ResultElementTypeRange",
"auto values = getODSResults(" # idx # ");\n"
"return {mlir::ResultElementTypeIterator(values.begin()), "
- "mlir::ResultElementTypeIterator(values.end())};"
+ "mlir::ResultElementTypeIterator(values.end())};",
+ [{
+ ArrayAttr::get(
+ [&]() {
+ llvm::SmallVector<Attribute, 4> ret;
+ for (auto t : $_self)
+ ret.push_back(TypeAttr::get(t));
+ return ret;
+ }(), $_ctx)
+ }]
>;
// A derived attribute that returns the shapes of the tensors in the actual
@@ -299,12 +321,15 @@
"mlir::TF::ResultShapeRange",
"auto values = getODSResults(" # idx # ");\n"
"return {mlir::TF::ResultShapeIterator(values.begin()), "
- "mlir::TF::ResultShapeIterator(values.end())};"
+ "mlir::TF::ResultShapeIterator(values.end())};",
+ // TODO(jpienaar): Update post TensorShapeAttr landing.
+ [{ nullptr }]
>;
// A derived attribute that returns the shape of the first result type.
def TF_DerivedResultShapeAttr : DerivedAttr<"ShapedType",
- "return (*getOperation()->result_type_begin()).cast<ShapedType>();">;
+ "return (*getOperation()->result_type_begin()).cast<ShapedType>();",
+ [{ TypeAttr::get($_self) }]>;
// A derived attribute that returns the element type of the tensor held by a
// named resource-type operand or result.
@@ -315,7 +340,6 @@
"assert(!resource_type.getSubtypes().empty() && \"unknown type\");\n"
"return mlir::getElementTypeOrSelf(*resource_type.getSubtypes().begin());">;
-
// A derived attribute that returns the shape of the tensor held by a named
// resource-type operand or result.
class TF_DerivedOperandOrResultHandleShapeAttr<string name> : DerivedAttr<
@@ -324,7 +348,8 @@
" mlir::getElementTypeOrSelf(this->" # name # "())\n"
" .cast<TF::ResourceType>();\n"
"assert(!resource_type.getSubtypes().empty() && \"unknown shape\");\n"
- "return resource_type.getSubtypes().begin()->cast<ShapedType>();">;
+ "return resource_type.getSubtypes().begin()->cast<ShapedType>();",
+ [{ TypeAttr::get($_self) }]>;
def TF_IntTypeAttr : TypeAttrBase<"IntegerType", "integer type"> {
let returnType = "Type";
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc
index 874fd6b..1c4d207 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.cc
@@ -82,8 +82,7 @@
// Returns true if the given `value` is of ranked float tensor type with the
// given `rank`.
-static inline bool isOfRankedFloatTensorType(Value value, int rank) {
- RankedTensorType type = GetRankedTensorTypeForOperand(value);
+static inline bool IsOfRankedFloatTensorType(RankedTensorType type, int rank) {
return type && type.getRank() == rank &&
type.getElementType().isa<FloatType>();
}
@@ -155,6 +154,53 @@
return dim_or_rank == -1;
}
+bool BroadcastCompatible(ArrayRef<Type> lhs, ArrayRef<Type> rhs) {
+ if (lhs.size() != rhs.size()) return false;
+ for (auto types : llvm::zip(lhs, rhs)) {
+ auto lhs_type = std::get<0>(types);
+ auto rhs_type = std::get<1>(types);
+
+ // This should be true for all TF ops:
+ auto lhs_tt = lhs_type.dyn_cast<TensorType>();
+ auto rhs_tt = rhs_type.dyn_cast<TensorType>();
+ if (!lhs_tt || !rhs_tt) {
+ if (lhs_type != rhs_type) return false;
+ continue;
+ }
+
+ // Verify matching element types. These should be identical, except for
+ // variant type where unknown subtype is considered compatible with all
+ // subtypes.
+ auto lhs_et = lhs_tt.getElementType();
+ auto rhs_et = rhs_tt.getElementType();
+ if (lhs_et != rhs_et) {
+ // If either is not a variant type, then the element types don't match.
+ auto lhs_vt = lhs_et.dyn_cast<TF::VariantType>();
+ auto rhs_vt = rhs_et.dyn_cast<TF::VariantType>();
+ if (!lhs_vt || !rhs_vt) return false;
+
+ // Consider the subtype of variant types.
+ auto lhs_vt_st = lhs_vt.getSubtypes();
+ auto rhs_vt_st = rhs_vt.getSubtypes();
+ if (!lhs_vt_st.empty() && !rhs_vt_st.empty()) {
+ for (auto subtypes : llvm::zip(lhs_vt_st, rhs_vt_st)) {
+ if (!BroadcastCompatible(std::get<0>(subtypes),
+ std::get<1>(subtypes)))
+ return false;
+ }
+ }
+ }
+
+ auto lhs_rt = lhs_type.dyn_cast<RankedTensorType>();
+ auto rhs_rt = rhs_type.dyn_cast<RankedTensorType>();
+ if (!lhs_rt || !rhs_rt) return true;
+ SmallVector<int64_t, 4> shape;
+ return OpTrait::util::getBroadcastedShape(lhs_rt.getShape(),
+ rhs_rt.getShape(), shape);
+ }
+ return true;
+}
+
// Returns the tf.Equal/tf.NotEqual result type given `x` and `y` and inputs. If
// `incompatible_shape_error` is true, reports error if `x` and `y` has
// incompatible shapes. Otherwise, returns a tensor type with unknown rank.
@@ -892,10 +938,12 @@
// Builds a constant op with the specified attribute `value`. The result
// op's type is deduced from `value`; if `value` is of scalar type,
// wraps it up with a tensor type of empty shape.
+// TODO(jpienaar): This one differs from the autogenerated one as it takes an
+// attribute but always creates an ElementsAttr internally.
void ConstOp::build(Builder *builder, OperationState &result, Attribute value) {
ShapedType type;
- if (auto elemAttr = value.dyn_cast<ElementsAttr>()) {
- type = elemAttr.getType();
+ if (auto elem_attr = value.dyn_cast<ElementsAttr>()) {
+ return ConstOp::build(builder, result, elem_attr);
} else if (value.isa<BoolAttr>() || value.isa<FloatAttr>() ||
value.isa<IntegerAttr>()) {
// All TensorFlow types must be tensor types. In the build() method,
@@ -903,12 +951,10 @@
// types. But we need to wrap it up with ElementsAttr to construct
// valid TensorFlow constants.
type = RankedTensorType::get(/*shape=*/{}, value.getType());
- value = DenseElementsAttr::get(type, value);
+ return ConstOp::build(builder, result, DenseElementsAttr::get(type, value));
}
- // TODO: support other TensorFlow specific types.
- assert(type && "unsupported attribute type for building tf.Const");
- result.types.push_back(type);
- result.addAttribute("value", value);
+ // TODO(jpienaar): support other TensorFlow specific types.
+ llvm_unreachable("unsupported attribute type for building tf.Const");
}
void ConstOp::build(Builder *builder, OperationState &result, Type type,
@@ -925,6 +971,24 @@
assert(type == result.types[0] && "type mismatch in construction");
}
+LogicalResult ConstOp::inferReturnTypes(
+ MLIRContext *context, Optional<Location> location, ValueRange operands,
+ ArrayRef<NamedAttribute> attributes, RegionRange regions,
+ SmallVectorImpl<Type> &inferredReturnTypes) {
+ for (NamedAttribute named_attr : attributes) {
+ if (named_attr.first.strref() != "value") continue;
+ auto value = named_attr.second;
+ if (auto elem_attr = value.dyn_cast<ElementsAttr>()) {
+ inferredReturnTypes.assign({elem_attr.getType()});
+ return success();
+ }
+ return emitOptionalError(location,
+ "attribute 'value' failed to satisfy constraint: "
+ "constant vector/tensor");
+ }
+ return emitOptionalError(location, "missing attribute 'value'");
+}
+
//===----------------------------------------------------------------------===//
// Conv2DOp and Conv3DOp
//===----------------------------------------------------------------------===//
@@ -1461,10 +1525,12 @@
// FakeQuantWithMinMaxVarsOp
//===----------------------------------------------------------------------===//
static LogicalResult Verify(FakeQuantWithMinMaxVarsOp op) {
- if (!isOfRankedFloatTensorType(op.min(), 0))
+ auto min = GetRankedTensorTypeForOperand(op.min());
+ if (min && !IsOfRankedFloatTensorType(min, 0))
return op.emitOpError("requires min to be a 0d float tensor");
- if (!isOfRankedFloatTensorType(op.max(), 0))
+ auto max = GetRankedTensorTypeForOperand(op.max());
+ if (max && !IsOfRankedFloatTensorType(max, 0))
return op.emitOpError("requires max to be a 0d float tensor");
int64_t num_bits = op.num_bits().getSExtValue();
@@ -1479,30 +1545,33 @@
// FakeQuantWithMinMaxVarsPerChannelOp
//===----------------------------------------------------------------------===//
static LogicalResult Verify(FakeQuantWithMinMaxVarsPerChannelOp op) {
- if (!isOfRankedFloatTensorType(op.min(), 1))
+ auto min = GetRankedTensorTypeForOperand(op.min());
+ if (min && !IsOfRankedFloatTensorType(min, 1))
return op.emitOpError("requires min to be a 1d float tensor");
- if (!isOfRankedFloatTensorType(op.max(), 1))
+ auto max = GetRankedTensorTypeForOperand(op.max());
+ if (max && !IsOfRankedFloatTensorType(max, 1))
return op.emitOpError("requires max to be a 1d float tensor");
Value inputs = op.inputs();
- if (!HasRankAtLeast(inputs, 1) ||
- inputs.getType().isa<UnrankedTensorType>()) {
+ if (!HasRankAtLeast(inputs, 1))
return op.emitError("requires inputs to be at least 1d float tensor");
- }
- auto inputsType = inputs.getType().cast<ShapedType>();
- int depth = inputsType.getDimSize(inputsType.getRank() - 1);
- if (op.min().getType().cast<ShapedType>().getDimSize(0) != depth ||
- op.max().getType().cast<ShapedType>().getDimSize(0) != depth) {
- return op.emitOpError(
- "requires min and max to have same size as last dimension of inputs");
- }
int64_t num_bits = op.num_bits().getSExtValue();
if (num_bits < 2 || num_bits > 16) {
return op.emitOpError(
"requires num_bits to be between 2 and 16, inclusive");
}
+
+ auto inputs_type = inputs.getType().dyn_cast<RankedTensorType>();
+ if (!inputs_type) return success();
+ int depth = inputs_type.getDimSize(inputs_type.getRank() - 1);
+ if ((min && min.getDimSize(0) != depth) ||
+ (max && max.getDimSize(0) != depth)) {
+ return op.emitOpError(
+ "requires min and max to have same size as last dimension of inputs");
+ }
+
return success();
}
@@ -1573,19 +1642,24 @@
//===----------------------------------------------------------------------===//
static LogicalResult Verify(FusedBatchNormOp op) {
- if (!isOfRankedFloatTensorType(op.x(), 4))
+ auto x = GetRankedTensorTypeForOperand(op.x());
+ if (x && !IsOfRankedFloatTensorType(x, 4))
return op.emitOpError("requires x to be a 4D float tensor");
- if (!isOfRankedFloatTensorType(op.scale(), 1))
+ auto scale = GetRankedTensorTypeForOperand(op.scale());
+ if (scale && !IsOfRankedFloatTensorType(scale, 1))
return op.emitOpError("requires scale to be a 1D float tensor");
- if (!isOfRankedFloatTensorType(op.offset(), 1))
+ auto offset = GetRankedTensorTypeForOperand(op.offset());
+ if (offset && !IsOfRankedFloatTensorType(offset, 1))
return op.emitOpError("requires offset to be a 1D float tensor");
- if (!isOfRankedFloatTensorType(op.mean(), 1))
+ auto mean = GetRankedTensorTypeForOperand(op.mean());
+ if (mean && !IsOfRankedFloatTensorType(mean, 1))
return op.emitOpError("requires mean to be a 1D float tensor");
- if (!isOfRankedFloatTensorType(op.variance(), 1))
+ auto variance = GetRankedTensorTypeForOperand(op.variance());
+ if (variance && !IsOfRankedFloatTensorType(variance, 1))
return op.emitOpError("requires variance to be a 1D float tensor");
// TODO(antiagainst): check attributes
@@ -2195,6 +2269,28 @@
}
//===----------------------------------------------------------------------===//
+// QrOp
+//===----------------------------------------------------------------------===//
+
+// Verifies that,
+//
+// * Input type, if ranked, must have at least 2 dimensions and at most
+// INT32_MAX dimensions.
+//
+static LogicalResult Verify(QrOp op) {
+ auto ttype = op.input().getType().cast<TensorType>();
+ if (!ttype.hasRank()) return success();
+ if (!HasRankAtLeast(op.input(), 2))
+ return op.emitOpError(
+ "requires ranked input tensor to be of rank 2 or more");
+ if (!HasRankAtMost(op.input(), std::numeric_limits<int32_t>::max()))
+ return op.emitOpError(
+ "requires ranked input tensor to be of rank INT32_MAX or less");
+
+ return success();
+}
+
+//===----------------------------------------------------------------------===//
// ReciprocalOp
//===----------------------------------------------------------------------===//
@@ -2437,7 +2533,8 @@
variadic_idx < 0 ? "" : llvm::formatv(" #{0}", variadic_idx).str();
auto result_ranked_type = result_type.dyn_cast<RankedTensorType>();
- if (!result_ranked_type || result_ranked_type.getShape().size() != 1)
+ if (!result_ranked_type) return success();
+ if (result_ranked_type.getShape().size() != 1)
return op->emitOpError("requires 1D type for result") << variadic_idx_str;
auto operand_ranked_type = operand_type.dyn_cast_or_null<RankedTensorType>();
@@ -2857,6 +2954,12 @@
// StridedSliceOp
//===----------------------------------------------------------------------===//
+// TODO(b/154160827): Add a canonicalization pattern from tf.StridedSliceOp to
+// tf.SliceOp if both of the following are true:
+// - All strides have a known value equal to 1
+// - No masks are set (or masks can be applied by transforming the inputs to
+// Slice)
+
// Verifies that,
//
// - begin, end and strides operands are 1D and they have the same number of
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h
index 8dc8fb3..bd3d894 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h
@@ -30,6 +30,7 @@
#include "mlir/IR/TypeUtilities.h" // from @llvm-project
#include "mlir/Interfaces/CallInterfaces.h" // from @llvm-project
#include "mlir/Interfaces/DerivedAttributeOpInterface.h" // from @llvm-project
+#include "mlir/Interfaces/InferTypeOpInterface.h" // from @llvm-project
#include "mlir/Interfaces/SideEffects.h" // from @llvm-project
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_structs.h"
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_traits.h"
@@ -86,6 +87,9 @@
// both mutex.h and this header file.
#undef mutex_lock
+// Returns whether two arrays of Type are broadcast compatible.
+bool BroadcastCompatible(ArrayRef<Type> lhs, ArrayRef<Type> rhs);
+
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_op_interfaces.h.inc"
#define GET_OP_CLASSES
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h.inc"
diff --git a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
index fc60a76..7cc0d61 100644
--- a/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
+++ b/tensorflow/compiler/mlir/tensorflow/ir/tf_ops.td
@@ -28,7 +28,9 @@
#define TF_OPS
include "tensorflow/compiler/mlir/tensorflow/ir/tf_generated_ops.td"
+include "tensorflow/compiler/mlir/tensorflow/ir/tf_op_base.td"
include "mlir/Interfaces/CallInterfaces.td"
+include "mlir/Interfaces/InferTypeOpInterface.td"
include "mlir/IR/OpBase.td"
class TF_TensorListInitOp<string mnemonic> : TF_Op<mnemonic, [NoSideEffect]> {
@@ -64,7 +66,8 @@
// In MLIR, the TensorFlow tensor value is represented as an ElementsAttr, with
// its type encoding the tensor's shape and data type.
-def TF_ConstOp : TF_Op<"Const", [ConstantLike, NoSideEffect]> {
+def TF_ConstOp : TF_Op<"Const", [ConstantLike, NoSideEffect,
+ DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
let summary = "Constant tensor op";
let arguments = (ins
@@ -85,6 +88,12 @@
];
let hasFolder = 1;
+
+ let extraClassDeclaration = [{
+ static bool isCompatibleReturnTypes(ArrayRef<Type> l, ArrayRef<Type> r) {
+ return BroadcastCompatible(l, r);
+ }
+ }];
}
def TF_DataFormatVecPermuteOp : TF_Op<"DataFormatVecPermute", [NoSideEffect, SameOperandsAndResultType]> {
@@ -667,4 +676,43 @@
TF_DerivedOperandSizeAttr NN = TF_DerivedOperandSizeAttr<1>;
}
+def TF_XlaShardingOp : TF_Op<"XlaSharding", [NoSideEffect]> {
+ let summary = [{
+An op which shards the input based on the given sharding attribute.
+ }];
+
+ let description = [{
+ }];
+
+ let arguments = (ins
+ TF_Tensor:$input,
+
+ OptionalAttr<StrAttr>:$_XlaSharding
+ );
+
+ let results = (outs
+ TF_Tensor:$output
+ );
+
+ TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
+}
+
+def TF_InfeedDequeueTupleOp : TF_Op<"InfeedDequeueTuple", []> {
+ let summary = "Fetches multiple values from infeed as an XLA tuple.";
+
+ let description = [{
+ }];
+
+ let arguments = (ins
+ OptionalAttr<StrAttr>:$_XlaSharding
+ );
+
+ let results = (outs
+ Variadic<TF_Tensor>:$outputs
+ );
+
+ TF_DerivedResultShapeListAttr shapes = TF_DerivedResultShapeListAttr<0>;
+ TF_DerivedResultTypeListAttr dtypes = TF_DerivedResultTypeListAttr<0>;
+}
+
#endif // TF_OPS
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/graphdef2mlir/tensor-list.pbtxt b/tensorflow/compiler/mlir/tensorflow/tests/graphdef2mlir/tensor-list.pbtxt
index cc24caa..87d2221 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/graphdef2mlir/tensor-list.pbtxt
+++ b/tensorflow/compiler/mlir/tensorflow/tests/graphdef2mlir/tensor-list.pbtxt
@@ -59,7 +59,7 @@
key: "value"
value {
tensor {
- dtype: DT_INT32
+ dtype: DT_FLOAT
tensor_shape {
dim {
size: 2
@@ -68,10 +68,10 @@
size: 2
}
}
- int_val: 1
- int_val: 2
- int_val: 3
- int_val: 4
+ float_val: 1
+ float_val: 2
+ float_val: 3
+ float_val: 4
}
}
}
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-type-attr.mlir b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-type-attr.mlir
index 83ddf62..6de8b30 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-type-attr.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-type-attr.mlir
@@ -12,7 +12,7 @@
func @main() {
tf_executor.graph {
- %0:2 = tf_executor.island wraps "tf.VariableV2"() {dtype = "tfdtype$DT_INT32", value = dense<2> : tensor<i32>} : () -> tensor<!tf.int32ref> loc("Ref_Variable")
+ %0:2 = tf_executor.island wraps "tf.VariableV2"() {dtype = "tfdtype$DT_INT32", value = dense<2> : tensor<i32>, shape = "tfshape$dim {\0A size: 2\0A}\0A", container = "", shared_name = ""} : () -> tensor<!tf.int32ref> loc("Ref_Variable")
%1:2 = tf_executor.island wraps "tf.Identity"(%0#0) : (tensor<!tf.int32ref>) -> tensor<*x!tf.int32ref> loc("foo")
tf_executor.fetch
}
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-while-loop.mlir b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-while-loop.mlir
index 8b2d393..b053735 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-while-loop.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/ref-while-loop.mlir
@@ -9,7 +9,7 @@
// CHECK: op: "RefNextIteration"
tf_executor.graph {
%0:3 = tf_executor.NextIteration.Source : tensor<*x!tf.int32ref> {device = "", T = "tfdtype$DT_INT32"} loc("while/NextIteration")
- %1:2 = tf_executor.island wraps "tf.VariableV2"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<0> : tensor<i32>} : () -> tensor<!tf.int32ref> loc("Ref_Variable")
+ %1:2 = tf_executor.island wraps "tf.VariableV2"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<0> : tensor<i32>, shape = "tfshape$dim {\0A size: 0\0A}\0A", container = "", shared_name = ""} : () -> tensor<!tf.int32ref> loc("Ref_Variable")
%2:2 = tf_executor.Enter %1#0 frame "while/while_context" parallel_iterations 10 : (tensor<!tf.int32ref>) -> (tensor<*x!tf.int32ref>, !tf_executor.control) {device = "", T = "tfdtype$DT_INT32"} loc("while/Enter")
%3:3 = tf_executor.Merge %2#0, %0#0 : tensor<*x!tf.int32ref> {device = "", N = 2, T = "tfdtype$DT_INT32"} loc("while/Merge")
%4:2 = tf_executor.island(%3#2) wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<10> : tensor<i32>} : () -> tensor<i32> loc("while/Less/y")
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/while-loop.mlir b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/while-loop.mlir
index fb2eac8..b6d7079 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/while-loop.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/mlir2graphdef/while-loop.mlir
@@ -10,7 +10,7 @@
// CHECK-NEXT: input: "while/Add"
tf_executor.graph {
%0:3 = tf_executor.NextIteration.Source : tensor<*xi32> {device = "", T = "tfdtype$DT_INT32"} loc("while/NextIteration")
- %1:2 = tf_executor.island wraps "tf.VariableV2"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<0> : tensor<i32>} : () -> tensor<i32> loc("Ref_Variable")
+ %1:2 = tf_executor.island wraps "tf.VariableV2"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<0> : tensor<i32>, shape = "tfshape$dim {\0A size: 0\0A}\0A", container = "", shared_name = ""} : () -> tensor<i32> loc("Ref_Variable")
%2:2 = tf_executor.Enter %1#0 frame "while/while_context" parallel_iterations 10 : (tensor<i32>) -> (tensor<*xi32>, !tf_executor.control) {device = "", T = "tfdtype$DT_INT32"} loc("while/Enter")
%3:3 = tf_executor.Merge %2#0, %0#0 : tensor<*xi32> {device = "", N = 2, T = "tfdtype$DT_INT32"} loc("while/Merge")
%4:2 = tf_executor.island(%3#2) wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<10> : tensor<i32>} : () -> tensor<i32> loc("while/Less/y")
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir b/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir
index 757df9d..94f626c 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/shape_inference.mlir
@@ -1,4 +1,4 @@
-// RUN: tf-opt %s -tf-shape-inference -verify-diagnostics | FileCheck %s -dump-input=fail -color
+// RUN: tf-opt %s -tf-shape-inference -verify-diagnostics | FileCheck %s -dump-input=fail
module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 130 : i32}} {
// CHECK-LABEL: func @main(%arg0: tensor<1xi32>, %arg1: tensor<1xi32>) -> tensor<1xi32>
@@ -71,6 +71,15 @@
return %1 : tensor<?x?x?x?xf32>
}
+// Tests where tf.Const's value needs to be refined.
+
+ func @const_refine() -> tensor<*xi32> {
+ %0 = "tf.Const"() {value = dense<[3, 2]> : tensor<2xi32>} : () -> tensor<*xi32>
+ // CHECK: "tf.Const"
+ // CHECK-SAME: -> tensor<2xi32>
+ return %0 : tensor<*xi32>
+ }
+
// Tests the case where an op's shape function returns non-fully-defined shapes.
// CHECK-LABEL: func @op_non_fully_defined_shape_fn
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/stack_ops_decomposition.mlir b/tensorflow/compiler/mlir/tensorflow/tests/stack_ops_decomposition.mlir
index e8c5bb5..26801e5 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/stack_ops_decomposition.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/stack_ops_decomposition.mlir
@@ -185,7 +185,7 @@
}
// CHECK: func @callee(%[[AARG0:.*]]: tensor<!tf.resource>, %[[AARG1:.*]]: tensor<i1>) -> tensor<!tf.resource>
-func @callee(%arg0: tensor<!tf.resource>, %arg1: tensor<i1>) -> tensor<!tf.resource> {
+func @callee(%arg0: tensor<!tf.resource>, %arg1: tensor<i1>) -> tensor<!tf.resource> attributes {sym_visibility = "public"} {
%elem = "tf._SomeOp"(%arg1) : (tensor<i1>) -> tensor<f32>
// CHECK: tf.StackPushV2"
%push = "tf.StackPushV2"(%arg0, %elem) {swap_memory = false} : (tensor<!tf.resource>, tensor<f32>) -> tensor<f32>
@@ -201,6 +201,62 @@
// -----
+// Tests PartitionedCall/StatefulPartitionedCall with private callee function.
+
+// CHECK-LABEL: func @main
+func @main(%arg0: tensor<i1>) -> () {
+ %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
+ // CHECK-NOT: tf.Stack
+ %stack = "tf.StackV2"(%max_size) {elem_type = f32, stack_name = "s"} : (tensor<i32>) -> tensor<!tf.resource>
+ // CHECK: "tf.StatefulPartitionedCall"
+ // CHECK-SAME: f = @callee
+ %call = "tf.StatefulPartitionedCall"(%stack, %arg0) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.resource>, tensor<i1>) -> tensor<!tf.resource>
+ // CHECK: "tf.PartitionedCall"
+ // CHECK-SAME: f = @callee
+ %call2 = "tf.PartitionedCall"(%stack, %arg0) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.resource>, tensor<i1>) -> tensor<!tf.resource>
+ // CHECK: "tf.Slice"
+ %pop = "tf.StackPopV2"(%call) : (tensor<!tf.resource>) -> tensor<f32>
+ // CHECK-NOT: tf.Stack
+ "tf.StackCloseV2"(%stack) : (tensor<!tf.resource>) -> ()
+ // CHECK: return
+ return
+}
+
+// CHECK: func @callee(%[[ARG0:.*]]: tensor<!tf.resource<tensor<10xf32>>>, %[[ARG1:.*]]: tensor<i1>, %[[ARG2:.*]]: tensor<!tf.resource<tensor<1xi32>>>)
+func @callee(%arg0: tensor<!tf.resource>, %arg1: tensor<i1>) -> tensor<!tf.resource> attributes {sym_visibility = "private"} {
+ %elem = "tf._SomeOp"(%arg1) : (tensor<i1>) -> tensor<f32>
+ // CHECK-NOT: "tf.StackPushV2"
+ // CHECK: %[[UPDATE:.*]] = "tf.XlaDynamicUpdateSlice"
+ // CHECK: "tf.AssignVariableOp"(%[[TARG0:.*]], %[[UPDATE]])
+ // CHECK: "tf.AssignVariableOp"(%[[EARG1:.*]],
+ // CHECK-NOT: "tf.StackPushV2"
+ %push = "tf.StackPushV2"(%arg0, %elem) {swap_memory = false} : (tensor<!tf.resource>, tensor<f32>) -> tensor<f32>
+ return %arg0 : tensor<!tf.resource>
+}
+
+// -----
+
+// Tests PartitionedCall op with no signature change on callee.
+
+// CHECK-LABEL: func @main
+func @main() -> () {
+ "tf.PartitionedCall"() {f = @callee, config = "", config_proto = "", executor_type = ""} : () -> ()
+ return
+}
+// CHECK: func @callee()
+func @callee() -> () attributes {sym_visibility = "public"} {
+ %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
+ // CHECK-NOT: tf.Stack
+ %stack = "tf.StackV2"(%max_size) {elem_type = f32, stack_name = "s"} : (tensor<i32>) -> tensor<!tf.resource>
+ %elem = "tf._SomeOp"() : () -> tensor<f32>
+ %push = "tf.StackPushV2"(%stack, %elem) {swap_memory = false} : (tensor<!tf.resource>, tensor<f32>) -> tensor<f32>
+ return
+}
+
+// -----
+
// Tests that the pass reports error on unknown stack size.
func @main(%arg0: tensor<i32>) -> tensor<2xi32> {
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tensor_array_ops_decomposition.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tensor_array_ops_decomposition.mlir
index 1a13338..b76e2da 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tensor_array_ops_decomposition.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tensor_array_ops_decomposition.mlir
@@ -322,7 +322,7 @@
}
// CHECK-LABEL: func @callee
// CHECK-SAME: (%[[OCARG0:.*]]: tensor<!tf.resource>) -> tensor<!tf.resource>
-func @callee(%arg0: tensor<!tf.resource>) -> tensor<!tf.resource> {
+func @callee(%arg0: tensor<!tf.resource>) -> tensor<!tf.resource> attributes {sym_visibility = "public"} {
%const1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
%elem = "tf._SomeOp"() : () -> tensor<3xf32>
%flow = "tf.Const"() {value = dense<1.0> : tensor<f32>} : () -> tensor<f32>
@@ -343,6 +343,75 @@
// -----
+// Tests (Stateful)PartitionedCall op with private callee function.
+
+// CHECK-LABEL: func @main
+func @main() -> () {
+ // CHECK: %[[SIZE:.*]] = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
+ %size = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
+ %index = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
+ // CHECK: %[[VAR:.*]] = "tf.MlirLocalVarOp"() : () -> tensor<!tf.resource<tensor<5x3xf32>>>
+ %ta:2 = "tf.TensorArrayV3"(%size) {dtype = f32, element_shape = "tfshape$dim { size: 3 }", dynamic_size = false, clear_after_read = true, identical_element_shapes = true, tensor_array_name = "ta"} : (tensor<i32>) -> (tensor<!tf.resource>, tensor<f32>)
+ // CHECK: %[[COND:.*]] = "tf._SomeOp"() : () -> tensor<i1>
+ %cond = "tf._SomeOp"() : () -> tensor<i1>
+ // CHECK: %[[GVAR1:.*]] = "tf.MlirLocalVarOp"() : () -> tensor<!tf.resource<tensor<5x3xf32>>>
+ %grad:2 = "tf.TensorArrayGradV3"(%ta#0, %ta#1) {source = "a"} : (tensor<!tf.resource>, tensor<f32>) -> (tensor<!tf.resource>, tensor<f32>)
+ // CHECK: %[[GVAR2:.*]] = "tf.MlirLocalVarOp"() : () -> tensor<!tf.resource<tensor<5x3xf32>>>
+ // CHECK: "tf.StatefulPartitionedCall"(%[[VAR]], %[[GVAR1]], %[[GVAR2]])
+ // CHECK-SAME: f = @callee
+ %call = "tf.StatefulPartitionedCall"(%ta#0) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.resource>) -> tensor<!tf.resource>
+ // CHECK: "tf.PartitionedCall"(%[[VAR]], %[[GVAR1]], %[[GVAR2]])
+ // CHECK-SAME: f = @callee
+ %call2 = "tf.PartitionedCall"(%call) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.resource>) -> tensor<!tf.resource>
+ // CHECK: %[[READ:.*]] = "tf.ReadVariableOp"(%[[VAR]]) : (tensor<!tf.resource<tensor<5x3xf32>>>) -> tensor<5x3xf32>
+ // CHECK: "tf.Slice"(%[[READ]],
+ %read = "tf.TensorArrayReadV3"(%call2, %index, %ta#1) : (tensor<!tf.resource>, tensor<i32>, tensor<f32>) -> tensor<3xf32>
+ return
+}
+// CHECK: func @callee(%[[CARG0:.*]]: tensor<!tf.resource<tensor<5x3xf32>>>, %[[CARG1:.*]]: tensor<!tf.resource<tensor<5x3xf32>>>, %[[CARG2:.*]]: tensor<!tf.resource<tensor<5x3xf32>>>)
+func @callee(%arg0: tensor<!tf.resource>) -> tensor<!tf.resource> attributes {sym_visibility = "private"} {
+ // CHECK: %[[READ1:.*]] = "tf.ReadVariableOp"(%[[CARG1]]) : (tensor<!tf.resource<tensor<5x3xf32>>>) -> tensor<5x3xf32>
+ // CHECK: %[[UPDATE1:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ1]],
+ // CHECK: "tf.AssignVariableOp"(%[[CARG1]], %[[UPDATE1]])
+ // CHECK: %[[READ2:.*]] = "tf.ReadVariableOp"(%[[CARG2]]) : (tensor<!tf.resource<tensor<5x3xf32>>>) -> tensor<5x3xf32>
+ // CHECK: %[[UPDATE2:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ2]],
+ // CHECK: "tf.AssignVariableOp"(%[[CARG2]], %[[UPDATE2]])
+ %const1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
+ %elem = "tf._SomeOp"() : () -> tensor<3xf32>
+ %flow = "tf.Const"() {value = dense<1.0> : tensor<f32>} : () -> tensor<f32>
+ %grad:2 = "tf.TensorArrayGradV3"(%arg0, %flow) {source = "a"} : (tensor<!tf.resource>, tensor<f32>) -> (tensor<!tf.resource>, tensor<f32>)
+ %gwrite = "tf.TensorArrayWriteV3"(%grad#0, %const1, %elem, %grad#1) : (tensor<!tf.resource>, tensor<i32>, tensor<3xf32>, tensor<f32>) -> tensor<f32>
+ %grad2:2 = "tf.TensorArrayGradV3"(%arg0, %flow) {source = "b"} : (tensor<!tf.resource>, tensor<f32>) -> (tensor<!tf.resource>, tensor<f32>)
+ %gwrite2 = "tf.TensorArrayWriteV3"(%grad2#0, %const1, %elem, %grad2#1) : (tensor<!tf.resource>, tensor<i32>, tensor<3xf32>, tensor<f32>) -> tensor<f32>
+ // CHECK: return %[[CARG0]]
+ return %arg0 : tensor<!tf.resource>
+}
+
+// -----
+
+// Tests PartitionedCall op with no signature change on callee.
+
+// CHECK-LABEL: func @main
+func @main() -> () {
+ %call = "tf.PartitionedCall"() {f = @callee, config = "", config_proto = "", executor_type = ""} : () -> tensor<i32>
+ return
+}
+// CHECK: func @callee() -> tensor<i32>
+func @callee() -> tensor<i32> attributes {sym_visibility = "public"} {
+ %size = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
+ // CHECK: "tf.MlirLocalVarOp"() : () -> tensor<!tf.resource<tensor<5xf32>>>
+ // CHECK: "tf.AssignVariableOp"
+ %ta:2 = "tf.TensorArrayV3"(%size) {dtype = f32, element_shape = "tfshape$", dynamic_size = false, clear_after_read = true, identical_element_shapes = true, tensor_array_name = "ta"} : (tensor<i32>) -> (tensor<!tf.resource>, tensor<f32>)
+ // CHECK: %[[SIZE:.*]] = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
+ %size_out = "tf.TensorArraySizeV3"(%ta#0, %ta#1) : (tensor<!tf.resource>, tensor<f32>) -> tensor<i32>
+ // CHECK: return %[[SIZE]] : tensor<i32>
+ return %size_out : tensor<i32>
+}
+
+// -----
+
// Test the pass reports failure on unknown size.
func @main(%arg0: tensor<i32>) -> () {
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tensor_list_ops_decomposition.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tensor_list_ops_decomposition.mlir
index 682da38..7e9b85f 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tensor_list_ops_decomposition.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tensor_list_ops_decomposition.mlir
@@ -141,6 +141,25 @@
// -----
+// Test scatter into existing tensor list.
+
+// CHECK-LABEL: func @main
+// CHECK-SAME: (%[[ARG0:.*]]: tensor<10x8x9xf32>, %[[ARG1:.*]]: tensor<5xi32>, %[[ARG2:.*]]: tensor<5x8x9xf32>) -> tensor<10x8x9xf32>
+func @main(%arg0: tensor<10x8x9xf32>, %arg1: tensor<5xi32>, %arg2: tensor<5x8x9xf32>) -> tensor<10x8x9xf32> {
+ %elem_shape = "tf.Const"() {value = dense<[8, 9]> : tensor<2xi32>} : () -> tensor<2xi32>
+ // CHECK: %[[BUFFER:.*]] = "tf.Identity"(%[[ARG0]]) : (tensor<10x8x9xf32>) -> tensor<10x8x9xf32>
+ %tl = "tf.TensorListFromTensor"(%arg0, %elem_shape) : (tensor<10x8x9xf32>, tensor<2xi32>) -> tensor<!tf.variant<tensor<8x9xf32>>>
+ // CHECK: %[[IND_SHAPE:.*]] = "tf.Const"() {value = dense<[5, 1]> : tensor<2xi32>} : () -> tensor<2xi32>
+ // CHECK: %[[IND_RESHPE:.*]] = "tf.Reshape"(%[[ARG1]], %[[IND_SHAPE]]) : (tensor<5xi32>, tensor<2xi32>) -> tensor<5x1xi32>
+ // CHECK: %[[SC:.*]] = "tf.TensorScatterUpdate"(%[[BUFFER]], %[[IND_RESHPE]], %[[ARG2]]) : (tensor<10x8x9xf32>, tensor<5x1xi32>, tensor<5x8x9xf32>) -> tensor<10x8x9xf32>
+ %scatter = "tf.TensorListScatterIntoExistingList"(%tl, %arg2, %arg1) : (tensor<!tf.variant<tensor<8x9xf32>>>, tensor<5x8x9xf32>, tensor<5xi32>) -> tensor<!tf.variant<tensor<8x9xf32>>>
+ %stack = "tf.TensorListStack"(%scatter, %elem_shape) : (tensor<!tf.variant<tensor<8x9xf32>>>, tensor<2xi32>) -> tensor<10x8x9xf32>
+ // CHECK: return %[[SC]] : tensor<10x8x9xf32>
+ return %stack : tensor<10x8x9xf32>
+}
+
+// -----
+
// Tests while loop.
// CHECK-LABEL: func @main
@@ -255,7 +274,7 @@
}
// CHECK: func @callee(%[[AARG0:.*]]: tensor<!tf.variant<tensor<f32>>>, %[[AARG1:.*]]: tensor<i1>) -> tensor<!tf.variant<tensor<f32>>>
-func @callee(%arg0: tensor<!tf.variant<tensor<f32>>>, %arg1: tensor<i1>) -> tensor<!tf.variant<tensor<f32>>> {
+func @callee(%arg0: tensor<!tf.variant<tensor<f32>>>, %arg1: tensor<i1>) -> tensor<!tf.variant<tensor<f32>>> attributes {sym_visibility = "public"} {
%elem = "tf._SomeOp"(%arg1) : (tensor<i1>) -> tensor<f32>
// CHECK: "tf.TensorListPushBack"
%push = "tf.TensorListPushBack"(%arg0, %elem) : (tensor<!tf.variant<tensor<f32>>>, tensor<f32>) -> tensor<!tf.variant<tensor<f32>>>
@@ -272,6 +291,66 @@
// -----
+// Tests PartitionedCall/StatefulPartitionedCall with private callee function.
+
+// CHECK-LABEL: func @main
+func @main(%arg0: tensor<i1>) -> () {
+ %elem_shape = "tf.Const"() {value = dense<[]> : tensor<0xi32>} : () -> tensor<0xi32>
+ %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
+ // CHECK-NOT: tf.EmptyTensorList
+ // CHECK: %[[INIT:.*]] = "tf.BroadcastTo"
+ %tl = "tf.EmptyTensorList"(%elem_shape, %max_size) : (tensor<0xi32>, tensor<i32>) -> tensor<!tf.variant<tensor<f32>>>
+ // CHECK: "tf.StatefulPartitionedCall"(%[[INIT]],
+ // CHECK-SAME: f = @callee
+ %call = "tf.StatefulPartitionedCall"(%tl, %arg0) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.variant<tensor<f32>>>, tensor<i1>) -> tensor<!tf.variant<tensor<f32>>>
+ // CHECK: %[[CALL2:.*]]:2 = "tf.PartitionedCall"(%[[INIT]],
+ // CHECK-SAME: f = @callee
+ %call2 = "tf.PartitionedCall"(%tl, %arg0) {f = @callee, config = "", config_proto = "", executor_type = ""}
+ : (tensor<!tf.variant<tensor<f32>>>, tensor<i1>) -> tensor<!tf.variant<tensor<f32>>>
+ // CHECK: %[[COPY:.*]] = "tf.Identity"(%[[CALL2]]#0)
+ // CHECK: "tf.Slice"(%[[COPY]],
+ %pop:2 = "tf.TensorListPopBack"(%call2, %elem_shape) : (tensor<!tf.variant<tensor<f32>>>, tensor<0xi32>) -> (tensor<!tf.variant<tensor<f32>>>, tensor<f32>)
+ // CHECK-NOT: tf.TensorListPopBack
+ // CHECK: return
+ return
+}
+
+// CHECK: func @callee(%[[ARG0:.*]]: tensor<10xf32>, %[[ARG1:.*]]: tensor<i1>, %[[ARG2:.*]]: tensor<1xi32>) -> (tensor<10xf32>, tensor<1xi32>)
+func @callee(%arg0: tensor<!tf.variant<tensor<f32>>>, %arg1: tensor<i1>) -> tensor<!tf.variant<tensor<f32>>> attributes {sym_visibility = "private"} {
+ %elem = "tf._SomeOp"(%arg1) : (tensor<i1>) -> tensor<f32>
+
+ // CHECK-NOT: "tf.TensorListPushBack"
+ // CHECK: %[[UPDATE:.*]] = "tf.XlaDynamicUpdateSlice"
+ // CHECK: %[[CONST1:.*]] = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: %[[ADD:.*]] = "tf.AddV2"(%[[ARG2]], %[[CONST1]])
+ // CHECK-NOT: "tf.TensorListPushBack"
+ %push = "tf.TensorListPushBack"(%arg0, %elem) : (tensor<!tf.variant<tensor<f32>>>, tensor<f32>) -> tensor<!tf.variant<tensor<f32>>>
+ // CHECK: return %[[UPDATE]], %[[ADD]]
+ return %push : tensor<!tf.variant<tensor<f32>>>
+}
+
+// -----
+
+// Tests PartitionedCall op with no signature change on callee.
+
+// CHECK-LABEL: func @main
+func @main() -> () {
+ "tf.PartitionedCall"() {f = @callee, config = "", config_proto = "", executor_type = ""} : () -> ()
+ return
+}
+// CHECK: func @callee()
+func @callee() -> () attributes {sym_visibility = "public"} {
+ %elem_shape = "tf.Const"() {value = dense<[]> : tensor<0xi32>} : () -> tensor<0xi32>
+ %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
+ // CHECK-NOT: tf.EmptyTensorList
+ // CHECK: "tf.BroadcastTo"
+ %tl = "tf.EmptyTensorList"(%elem_shape, %max_size) : (tensor<0xi32>, tensor<i32>) -> tensor<!tf.variant<tensor<f32>>>
+ return
+}
+
+// -----
+
// Tests that the pass reports error on unknown maximum size.
func @main(%arg0: tensor<i32>) -> () {
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tf-ops.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tf-ops.mlir
index afe6367..9625439 100644
--- a/tensorflow/compiler/mlir/tensorflow/tests/tf-ops.mlir
+++ b/tensorflow/compiler/mlir/tensorflow/tests/tf-ops.mlir
@@ -725,10 +725,10 @@
// -----
// Test invalid tf.FusedBatchNorm
-func @testFusedBatchNormWrongVarianceType(tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>) -> tensor<8x8x8x8xf32> {
-^bb0(%arg0: tensor<8x8x8x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<*xf32>):
+func @testFusedBatchNormWrongVarianceType(tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<10x2xf32>) -> tensor<8x8x8x8xf32> {
+^bb0(%arg0: tensor<8x8x8x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<10x2xf32>):
// expected-error @+1 {{requires variance to be a 1D float tensor}}
- %0:5 = "tf.FusedBatchNorm"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>)
+ %0:5 = "tf.FusedBatchNorm"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<10x2xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<10x2xf32>)
return %0#0 : tensor<8x8x8x8xf32>
}
@@ -1297,11 +1297,11 @@
// -----
-func @testShapeWrongResultDim(tensor<1x32x32x16xf32>) -> tensor<*xi32> {
+func @testShapeWrongResultDim(tensor<1x32x32x16xf32>) -> tensor<3x2xi32> {
^bb0(%arg0: tensor<1x32x32x16xf32>):
// expected-error @+1 {{requires 1D type for result}}
- %0 = "tf.Shape"(%arg0) {T = "tfdtype$DT_FLOAT", output = "tfdtype$DT_INT32"} : (tensor<1x32x32x16xf32>) -> tensor<*xi32>
- return %0 : tensor<*xi32>
+ %0 = "tf.Shape"(%arg0) {T = "tfdtype$DT_FLOAT", output = "tfdtype$DT_INT32"} : (tensor<1x32x32x16xf32>) -> tensor<3x2xi32>
+ return %0 : tensor<3x2xi32>
}
// -----
@@ -1341,11 +1341,11 @@
// -----
-func @testShapeNWrongResultDim(tensor<1x32x32x16xf32>) -> tensor<*xi32> {
+func @testShapeNWrongResultDim(tensor<1x32x32x16xf32>) -> tensor<2x2xi32> {
^bb0(%arg0: tensor<1x32x32x16xf32>):
// expected-error @+1 {{requires 1D type for result #1}}
- %0:2 = "tf.ShapeN"(%arg0, %arg0) : (tensor<1x32x32x16xf32>, tensor<1x32x32x16xf32>) -> (tensor<4xi32>, tensor<*xi32>)
- return %0#1 : tensor<*xi32>
+ %0:2 = "tf.ShapeN"(%arg0, %arg0) : (tensor<1x32x32x16xf32>, tensor<1x32x32x16xf32>) -> (tensor<4xi32>, tensor<2x2xi32>)
+ return %0#1 : tensor<2x2xi32>
}
// -----
@@ -1402,10 +1402,10 @@
// -----
-func @testVariableShapeWrongResultDim(%arg0: tensor<*x!tf.resource<tensor<1x32x32x16xf32>>>) -> tensor<*xi32> {
+func @testVariableShapeWrongResultDim(%arg0: tensor<*x!tf.resource<tensor<1x32x32x16xf32>>>) -> tensor<2x3xi32> {
// expected-error @+1 {{requires 1D type for result}}
- %0 = "tf.VariableShape"(%arg0) {output = "tfdtype$DT_INT32"} : (tensor<*x!tf.resource<tensor<1x32x32x16xf32>>>) -> tensor<*xi32>
- return %0 : tensor<*xi32>
+ %0 = "tf.VariableShape"(%arg0) {output = "tfdtype$DT_INT32"} : (tensor<*x!tf.resource<tensor<1x32x32x16xf32>>>) -> tensor<2x3xi32>
+ return %0 : tensor<2x3xi32>
}
// -----
@@ -1768,7 +1768,7 @@
// -----
func @testOneHot(%indices: tensor<3xi32>, %on_value: tensor<f32>, %off_value: tensor<f32>) -> tensor<3x5xf32> {
- %depth = "tf.Const"() { value = dense<-5> : tensor<i64> } : () -> tensor<i32>
+ %depth = "tf.Const"() { value = dense<-5> : tensor<i32> } : () -> tensor<i32>
// expected-error @+1 {{depth must be non-negative}}
%result = "tf.OneHot"(%indices, %depth, %on_value, %off_value) {axis = -1 : i64} : (tensor<3xi32>, tensor<i32>, tensor<f32>, tensor<f32>) -> tensor<3x5xf32>
return %result : tensor<3x5xf32>
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc b/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc
index 02e1c99..31a80a4 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/mark_function_visibility.cc
@@ -97,6 +97,36 @@
MarkFunctionVisibilityUsingEntryFunctionSpecificationPass>();
}
+// Marks the main function with public visibility, while other functions are
+// marked with private visibility.
+LogicalResult MarkOnlyMainFunctionWithPublicVisibility(ModuleOp module) {
+ for (auto func : module.getOps<FuncOp>()) {
+ if (func.getName() == "main") {
+ func.setVisibility(FuncOp::Visibility::Public);
+ } else {
+ func.setVisibility(FuncOp::Visibility::Private);
+ }
+ }
+ return success();
+}
+
+namespace {
+struct MarkOnlyMainFunctionWithPublicVisibilityPass
+ : public PassWrapper<MarkOnlyMainFunctionWithPublicVisibilityPass,
+ OperationPass<ModuleOp>> {
+ void runOnOperation() override {
+ if (failed(MarkOnlyMainFunctionWithPublicVisibility(getOperation()))) {
+ signalPassFailure();
+ }
+ }
+};
+} // namespace
+
+std::unique_ptr<OperationPass<ModuleOp>>
+CreateMarkOnlyMainFunctionWithPublicVisibilityPass() {
+ return std::make_unique<MarkOnlyMainFunctionWithPublicVisibilityPass>();
+}
+
} // namespace TF
namespace tf_saved_model {
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h
index d6da961..d04a065 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/passes.h
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/passes.h
@@ -101,6 +101,11 @@
std::unique_ptr<OperationPass<ModuleOp>>
CreateMarkFunctionVisibilityUsingEntryFunctionSpecificationPass();
+// Creates a pass that marks the main function with public visibility, while
+// other functions are marked with private visibility.
+std::unique_ptr<OperationPass<ModuleOp>>
+CreateMarkOnlyMainFunctionWithPublicVisibilityPass();
+
// Creates a simple device assignment pass on TF dialect for CoreRT use case.
std::unique_ptr<OperationPass<FuncOp>> CreateSimpleTFDeviceAssignmentPass(
llvm::StringRef default_device);
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc b/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc
index 53e15b0..c2e21d2 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/shape_inference.cc
@@ -274,6 +274,15 @@
return changed;
}
+bool RefineTfConst(TF::ConstOp const_op) {
+ Type old_type = const_op.getType();
+ if (const_op.valueAttr().getType() == old_type) return false;
+ const_op.getResult().setType(const_op.valueAttr().getType());
+ AddCastBackForUnsupportedNonTFUses(const_op, const_op.getResult(),
+ const_op.getDialect(), old_type);
+ return true;
+}
+
} // namespace
bool InferShapeForSingleOperation(Operation* op, Dialect* tf_dialect,
@@ -622,6 +631,13 @@
return;
}
+ if (auto tf_const = dyn_cast<TF::ConstOp>(op)) {
+ changed |= RefineTfConst(tf_const);
+ // TODO(jpienaar): Debug why we can't just return here. We end up with
+ // additional constant due to the propagation of constant into attached
+ // function if we return already.
+ }
+
// Before attempting inference, just try to fold the operation.
if (succeeded(folder.tryToFold(op))) return;
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
index 55b22ad..28be174 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/stack_ops_decomposition.cc
@@ -154,14 +154,14 @@
LogicalResult DecomposeStackOpsInternal(
Block*, ModuleOp, llvm::SmallDenseMap<Value, Value>*,
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>*);
+ llvm::StringMap<PartitionedCallStackOpsInfo>*);
// Handles stack usage by a tf.While. It will convert the body and conditional
// function signatures, and performs stack ops decomposition on them.
LogicalResult HandleWhileOp(
TF::WhileOp while_op, ModuleOp module,
const llvm::SmallDenseMap<Value, Value>& data_var_to_size_var,
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>*
+ llvm::StringMap<PartitionedCallStackOpsInfo>*
decomposed_partitioned_call_callees) {
auto body = module.lookupSymbol<FuncOp>(while_op.body());
llvm::SmallDenseMap<Value, Value> body_map;
@@ -238,7 +238,7 @@
LogicalResult HandleIfOp(
TF::IfOp if_op, ModuleOp module,
const llvm::SmallDenseMap<Value, Value>& data_var_to_size_var,
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>*
+ llvm::StringMap<PartitionedCallStackOpsInfo>*
decomposed_partitioned_call_callees) {
auto then_branch = module.lookupSymbol<FuncOp>(if_op.then_branch());
auto else_branch = module.lookupSymbol<FuncOp>(if_op.else_branch());
@@ -295,11 +295,11 @@
LogicalResult HandlePartitionedCallOp(
CallOp call, FuncOp callee, ModuleOp module,
const llvm::SmallDenseMap<Value, Value>& data_var_to_size_var,
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>*
+ llvm::StringMap<PartitionedCallStackOpsInfo>*
decomposed_partitioned_call_callees) {
auto emplace_res = decomposed_partitioned_call_callees->try_emplace(
- callee, PartitionedCallStackOpsInfo());
- auto& info = emplace_res.first->getSecond();
+ callee.getName(), PartitionedCallStackOpsInfo());
+ auto& info = emplace_res.first->second;
// Recreate the call op with info.
auto recreate_caller = [&] {
auto new_operands = llvm::to_vector<8>(call.getOperands());
@@ -343,39 +343,38 @@
return recreate_caller();
}
llvm::SmallDenseMap<Value, Value> callee_map;
- auto callee_clone = callee.clone();
+ FuncOp lowered_callee = callee;
+ if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+ // Clone non-private callee in case of signature change.
+ lowered_callee = callee.clone();
+ lowered_callee.setVisibility(SymbolTable::Visibility::Private);
+ }
auto find_arg_stack_type = [&](int64_t index) -> llvm::Optional<Type> {
auto it = data_var_to_size_var.find(call.getOperand(index));
if (it == data_var_to_size_var.end()) return llvm::None;
return it->getFirst().getType();
};
- ModifyFunctionSignature(callee_clone, &callee_map, find_arg_stack_type);
- if (callee_map.empty()) {
+ ModifyFunctionSignature(lowered_callee, &callee_map, find_arg_stack_type);
+ info.signature_change = !callee_map.empty();
+ if (!info.signature_change) {
// Signature is not modified. We do not need the clone.
- info.signature_change = false;
- callee_clone.erase();
+ if (lowered_callee != callee) {
+ lowered_callee.erase();
+ }
} else {
- info.signature_change = true;
- info.decomposed_callee = callee_clone;
+ info.decomposed_callee = lowered_callee;
for (auto& entry : callee_map) {
info.stack_var_arg_to_size_arg
[entry.getFirst().cast<BlockArgument>().getArgNumber()] =
entry.getSecond().cast<BlockArgument>().getArgNumber();
}
- // Add the clone with a new name.
- auto name_base = llvm::join(
- std::vector<std::string>{callee.getName().str(), "stack_decomposed"},
- "_");
- auto name = name_base;
- {
- int64_t counter = 0;
- while (module.lookupSymbol(name)) {
- name = llvm::formatv("{0}_{1}", name_base, counter++).str();
- }
+ if (lowered_callee != callee) {
+ // Add the clone with a new name.
+ lowered_callee.setName(
+ llvm::formatv("{0}_stack_decomposed", callee.getName()).str());
+ SymbolTable(module).insert(lowered_callee);
+ callee = lowered_callee;
}
- callee_clone.setName(name);
- SymbolTable(module).insert(callee_clone);
- callee = callee_clone;
}
if (failed(DecomposeStackOpsInternal(&callee.front(), module, &callee_map,
decomposed_partitioned_call_callees))) {
@@ -487,7 +486,7 @@
LogicalResult DecomposeStackOpsInternal(
Block* block, ModuleOp module,
llvm::SmallDenseMap<Value, Value>* data_var_to_size_var,
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>*
+ llvm::StringMap<PartitionedCallStackOpsInfo>*
decomposed_partitioned_call_callees) {
for (auto& op : llvm::make_early_inc_range(block->getOperations())) {
if (llvm::isa<TF::IdentityOp>(&op) || llvm::isa<TF::IdentityNOp>(&op)) {
@@ -545,7 +544,7 @@
LogicalResult DecomposeStackOps(Block* block, ModuleOp module) {
llvm::SmallDenseMap<Value, Value> data_var_to_size_var;
- llvm::SmallDenseMap<FuncOp, PartitionedCallStackOpsInfo>
+ llvm::StringMap<PartitionedCallStackOpsInfo>
decomposed_partitioned_call_callees;
return DecomposeStackOpsInternal(block, module, &data_var_to_size_var,
&decomposed_partitioned_call_callees);
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
index 8e0c34a..89abef2 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_array_ops_decomposition.cc
@@ -531,13 +531,12 @@
LogicalResult DecomposeTensorArrayOps(
Block*, ModuleOp, llvm::SmallDenseMap<Value, TensorArrayStats>*,
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>*);
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>*);
-LogicalResult HandleWhileOp(
- TF::WhileOp while_op, ModuleOp module,
- llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>*
- decomposed_partitioned_call_callees) {
+LogicalResult HandleWhileOp(TF::WhileOp while_op, ModuleOp module,
+ llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>*
+ decomposed_partitioned_call_callees) {
auto body = module.lookupSymbol<FuncOp>(while_op.body());
auto cond = module.lookupSymbol<FuncOp>(while_op.cond());
auto grads = AccessedGradients({body, cond}, module);
@@ -619,11 +618,10 @@
return success();
}
-LogicalResult HandleIfOp(
- TF::IfOp if_op, ModuleOp module,
- llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>*
- decomposed_partitioned_call_callees) {
+LogicalResult HandleIfOp(TF::IfOp if_op, ModuleOp module,
+ llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>*
+ decomposed_partitioned_call_callees) {
auto then_branch = module.lookupSymbol<FuncOp>(if_op.then_branch());
auto else_branch = module.lookupSymbol<FuncOp>(if_op.else_branch());
auto grads = AccessedGradients({then_branch, else_branch}, module);
@@ -706,11 +704,11 @@
LogicalResult HandlePartitionedCallOp(
CallOp call, FuncOp callee, ModuleOp module,
llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>*
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>*
decomposed_partitioned_call_callees) {
auto emplace_res = decomposed_partitioned_call_callees->try_emplace(
- callee, PartitionedCallTensorArrayOpsInfo());
- auto& info = emplace_res.first->getSecond();
+ callee.getName(), PartitionedCallTensorArrayOpsInfo());
+ auto& info = emplace_res.first->second;
// Recreates the call op with info.
auto recreate_caller = [&]() -> LogicalResult {
auto new_operands = llvm::to_vector<8>(call.getOperands());
@@ -752,7 +750,7 @@
if (!info.signature_change) return success();
return recreate_caller();
}
- // Rewrite the callee on a cloned function.
+ // Rewrite the callee.
info.signature_change = false;
auto ta_arg_buffer_type = [&](int64_t index) -> Type {
auto it = stats->find(call.getOperand(index));
@@ -765,45 +763,46 @@
if (it == stats->end()) return false;
return it->getSecond().accumulate_on_write;
};
- auto callee_clone = callee.clone();
- callee_clone.setVisibility(SymbolTable::Visibility::Private);
- auto grads = AccessedGradients({callee_clone}, module);
- for (int64_t i = 0; i < callee_clone.getNumArguments(); ++i) {
+ FuncOp lowered_callee = callee;
+ if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+ // Clone non-private callee in case of signature change.
+ lowered_callee = callee.clone();
+ lowered_callee.setVisibility(SymbolTable::Visibility::Private);
+ }
+ auto grads = AccessedGradients({lowered_callee}, module);
+ for (int64_t i = 0; i < lowered_callee.getNumArguments(); ++i) {
auto it = grads.find(i);
if (it == grads.end()) continue;
info.arg_grads.emplace_back(i, it->getSecond());
}
llvm::SmallDenseMap<Value, TensorArrayStats> callee_stats;
- ChangeFunctionInputSignature(callee_clone, grads, ta_arg_buffer_type,
+ ChangeFunctionInputSignature(lowered_callee, grads, ta_arg_buffer_type,
ta_accumulate_on_write, &callee_stats);
- if (failed(DecomposeTensorArrayOps(&callee_clone.front(), module,
+ if (failed(DecomposeTensorArrayOps(&lowered_callee.front(), module,
&callee_stats,
decomposed_partitioned_call_callees))) {
return failure();
}
for (int64_t i = 0; i < call.getNumResults(); ++i) {
- auto ret = callee_clone.front().getTerminator()->getOperand(i);
+ auto ret = lowered_callee.front().getTerminator()->getOperand(i);
if (!getElementTypeOrSelf(ret.getType()).isa<TF::ResourceType>()) continue;
auto arg = ret.dyn_cast<BlockArgument>();
if (!arg) continue;
info.ret_forward_input.emplace_back(i, arg.getArgNumber());
}
- if (!info.signature_change) {
- // Signature is not modified. We do not need to keep two copies.
- info.signature_change = false;
- auto name = callee.getName();
- callee.erase();
- callee_clone.setName(name);
- SymbolTable(module).insert(callee_clone);
- } else {
- info.decomposed_callee = callee_clone;
- // Add the clone with a new name.
- auto name =
- llvm::formatv("{0}_{1}", callee.getName(), "tensorarray_decomposed")
- .str();
- callee_clone.setName(name);
- SymbolTable(module).insert(callee_clone);
+ info.decomposed_callee = lowered_callee;
+ if (lowered_callee != callee) {
+ if (!info.signature_change) {
+ // Signature is not modified. We do not need to keep two copies.
+ lowered_callee.setName(callee.getName());
+ callee.erase();
+ } else {
+ // Add the clone with a new name.
+ lowered_callee.setName(
+ llvm::formatv("{0}_tensorarray_decomposed", callee.getName()).str());
+ }
+ SymbolTable(module).insert(lowered_callee);
}
if (info.signature_change) return recreate_caller();
return success();
@@ -812,7 +811,7 @@
LogicalResult DecomposeTensorArrayOps(
Block* block, ModuleOp module,
llvm::SmallDenseMap<Value, TensorArrayStats>* stats,
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>*
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>*
decomposed_partitioned_call_callees) {
for (auto& op : llvm::make_early_inc_range(block->getOperations())) {
if (llvm::isa<TF::IdentityOp>(&op) || llvm::isa<TF::IdentityNOp>(&op)) {
@@ -880,7 +879,7 @@
auto main = module.lookupSymbol<FuncOp>("main");
if (!main) return;
llvm::SmallDenseMap<Value, TensorArrayStats> stats;
- llvm::SmallDenseMap<FuncOp, PartitionedCallTensorArrayOpsInfo>
+ llvm::StringMap<PartitionedCallTensorArrayOpsInfo>
decomposed_partitioned_call_callees;
if (failed(DecomposeTensorArrayOps(&main.front(), module, &stats,
&decomposed_partitioned_call_callees))) {
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
index d6b6216..1294248 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tensor_list_ops_decomposition.cc
@@ -16,6 +16,7 @@
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringMap.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/FormatVariadic.h"
#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
@@ -122,7 +123,7 @@
LogicalResult DecomposeTensorListOpsInternal(
Block*, ModuleOp, llvm::SmallDenseMap<Value, SizeInfo>*,
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>*);
+ llvm::StringMap<PartitionedCallDecompositionInfo>*);
// Adds the corresponding sizes of tensor list buffers in func's return values
// to the list of return values. Returns the mapping from the buffer indices to
@@ -151,7 +152,7 @@
LogicalResult HandleWhileOp(
TF::WhileOp while_op, ModuleOp module,
llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size,
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>*
+ llvm::StringMap<PartitionedCallDecompositionInfo>*
decomposed_partitioned_call_callees) {
// Rewrite body.
auto body = module.lookupSymbol<FuncOp>(while_op.body());
@@ -216,11 +217,10 @@
return success();
}
-LogicalResult HandleIfOp(
- TF::IfOp if_op, ModuleOp module,
- llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size,
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>*
- decomposed_partitioned_call_callees) {
+LogicalResult HandleIfOp(TF::IfOp if_op, ModuleOp module,
+ llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size,
+ llvm::StringMap<PartitionedCallDecompositionInfo>*
+ decomposed_partitioned_call_callees) {
// Rewrite the branches.
auto then_branch = module.lookupSymbol<FuncOp>(if_op.then_branch());
auto else_branch = module.lookupSymbol<FuncOp>(if_op.else_branch());
@@ -285,11 +285,11 @@
LogicalResult HandlePartitionedCallOp(
CallOp call, FuncOp callee, ModuleOp module,
llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size,
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>*
+ llvm::StringMap<PartitionedCallDecompositionInfo>*
decomposed_partitioned_call_callees) {
auto emplace_res = decomposed_partitioned_call_callees->try_emplace(
- callee, PartitionedCallDecompositionInfo());
- auto& info = emplace_res.first->getSecond();
+ callee.getName(), PartitionedCallDecompositionInfo());
+ auto& info = emplace_res.first->second;
// Recreates the call op with info.
auto recreate_caller = [&] {
auto new_operands = llvm::to_vector<8>(call.getOperands());
@@ -325,10 +325,14 @@
if (!info.signature_change) return success();
return recreate_caller();
}
- // Rewrite the callee on a cloned function.
+ // Rewrite the callee.
llvm::SmallDenseMap<Value, SizeInfo> callee_map;
- auto callee_clone = callee.clone();
- callee_clone.setVisibility(SymbolTable::Visibility::Private);
+ FuncOp lowered_callee = callee;
+ if (callee.getVisibility() != SymbolTable::Visibility::Private) {
+ // Clone non-private callee in case of signature change.
+ lowered_callee = callee.clone();
+ lowered_callee.setVisibility(SymbolTable::Visibility::Private);
+ }
auto find_arg_buffer_type = [&](int64_t index) -> llvm::Optional<Type> {
auto it = buffer_to_size->find(call.getOperand(index));
if (it == buffer_to_size->end()) return llvm::None;
@@ -337,41 +341,41 @@
auto arg_buffer_size_is_fixed = [&](int64_t index) {
return (*buffer_to_size)[call.getOperand(index)].fixed;
};
- ModifyFunctionSignature(callee_clone, cutil::GetSizeType(OpBuilder(call)),
+ ModifyFunctionSignature(lowered_callee, cutil::GetSizeType(OpBuilder(call)),
&callee_map, find_arg_buffer_type,
arg_buffer_size_is_fixed);
- const bool args_no_changed = callee.empty();
+ const bool args_no_changed = callee_map.empty();
if (failed(DecomposeTensorListOpsInternal(
- &callee_clone.front(), module, &callee_map,
+ &lowered_callee.front(), module, &callee_map,
decomposed_partitioned_call_callees))) {
return failure();
}
info.buffer_ret_to_size_ret =
- AddTensorListSizesToReturn(callee_clone, callee_map);
+ AddTensorListSizesToReturn(lowered_callee, callee_map);
+ info.decomposed_callee = lowered_callee;
if (args_no_changed && info.buffer_ret_to_size_ret.empty()) {
// Signature is not modified. We do not need to keep two copies.
info.signature_change = false;
- auto name = callee.getName();
- callee.erase();
- callee_clone.setName(name);
- SymbolTable(module).insert(callee_clone);
+ if (lowered_callee != callee) {
+ lowered_callee.setName(callee.getName());
+ callee.erase();
+ SymbolTable(module).insert(lowered_callee);
+ }
} else {
info.signature_change = true;
- info.decomposed_callee = callee_clone;
for (auto& entry : callee_map) {
auto buffer_arg = entry.getFirst().dyn_cast<BlockArgument>();
if (!buffer_arg) continue;
info.buffer_arg_to_size_arg[buffer_arg.getArgNumber()] =
entry.getSecond().size.cast<BlockArgument>().getArgNumber();
}
-
- // Add the clone with a new name.
- auto name = llvm::join(std::vector<std::string>{callee.getName().str(),
- "tensorlist_decomposed"},
- "_");
- callee_clone.setName(name);
- SymbolTable(module).insert(callee_clone);
- callee = callee_clone;
+ if (lowered_callee != callee) {
+ // Add the clone with a new name.
+ lowered_callee.setName(
+ llvm::formatv("{0}_tensorlist_decomposed", callee.getName()).str());
+ SymbolTable(module).insert(lowered_callee);
+ callee = lowered_callee;
+ }
}
if (info.signature_change) return recreate_caller();
return success();
@@ -608,10 +612,37 @@
return success();
}
+LogicalResult HandleTensorListScatterIntoExistingListOp(
+ TF::TensorListScatterIntoExistingListOp scatter,
+ llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size) {
+ auto it = buffer_to_size->find(scatter.input_handle());
+ if (it == buffer_to_size->end()) {
+ return scatter.emitOpError("unknown tensor list");
+ }
+ auto buffer = scatter.input_handle();
+ OpBuilder builder(scatter);
+ auto indices_type = scatter.indices().getType().cast<RankedTensorType>();
+ if (!indices_type) return scatter.emitOpError("unranked indices shape");
+ auto shape_type = RankedTensorType::get({2}, builder.getIntegerType(32));
+ auto shape = builder.create<TF::ConstOp>(
+ scatter.getLoc(),
+ DenseElementsAttr::get(
+ shape_type, {static_cast<int>(indices_type.getDimSize(0)), 1}));
+ auto indices =
+ builder.create<TF::ReshapeOp>(scatter.getLoc(), scatter.indices(), shape);
+ Value tensor_scatter_update = builder.create<TF::TensorScatterUpdateOp>(
+ scatter.getLoc(), buffer, indices, scatter.tensor());
+ scatter.output_handle().replaceAllUsesWith(tensor_scatter_update);
+ scatter.erase();
+ auto size = it->getSecond();
+ (*buffer_to_size)[tensor_scatter_update] = size;
+ return success();
+}
+
LogicalResult DecomposeTensorListOpsInternal(
Block* block, ModuleOp module,
llvm::SmallDenseMap<Value, SizeInfo>* buffer_to_size,
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>*
+ llvm::StringMap<PartitionedCallDecompositionInfo>*
decomposed_partitioned_call_callees) {
for (auto& op : llvm::make_early_inc_range(block->getOperations())) {
// TODO(yuanzx): Add a pass to remove identities in device computation.
@@ -662,6 +693,13 @@
if (failed(HandleTensorListGatherOp(gather, *buffer_to_size))) {
return failure();
}
+ } else if (auto scatter =
+ llvm::dyn_cast<TF::TensorListScatterIntoExistingListOp>(
+ &op)) {
+ if (failed(HandleTensorListScatterIntoExistingListOp(scatter,
+ buffer_to_size))) {
+ return failure();
+ }
} else if (auto addn = llvm::dyn_cast<TF::AddNOp>(&op)) {
auto it = buffer_to_size->find(addn.getOperand(0));
if (it != buffer_to_size->end()) {
@@ -710,7 +748,7 @@
LogicalResult DecomposeTensorListOps(Block* block, ModuleOp module) {
llvm::SmallDenseMap<Value, SizeInfo> buffer_to_size;
- llvm::SmallDenseMap<FuncOp, PartitionedCallDecompositionInfo>
+ llvm::StringMap<PartitionedCallDecompositionInfo>
decomposed_partitioned_call_callees;
return DecomposeTensorListOpsInternal(block, module, &buffer_to_size,
&decomposed_partitioned_call_callees);
diff --git a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
index eb47b8c..ce62773 100644
--- a/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
+++ b/tensorflow/compiler/mlir/tensorflow/transforms/tpu_sharding_identification_pass.cc
@@ -93,7 +93,7 @@
}
if (!parsed_sharding_op) return llvm::Optional<llvm::StringRef>();
- return tensorflow::ParseShardingAttribute(parsed_sharding_op->getOperation());
+ return parsed_sharding_op.getValue()._XlaSharding();
}
// Returns the provided sharding configuration if operand of return value
@@ -102,9 +102,8 @@
const int output_index,
const OpOperand& operand) {
if (auto sharding_op = llvm::dyn_cast_or_null<TF::XlaShardingOp>(
- operand.get().getDefiningOp())) {
- return tensorflow::ParseShardingAttribute(sharding_op.getOperation());
- }
+ operand.get().getDefiningOp()))
+ return sharding_op._XlaSharding();
return llvm::Optional<StringRef>();
}
diff --git a/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc b/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc
index 4cdce62..c22e86e 100644
--- a/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc
+++ b/tensorflow/compiler/mlir/tensorflow/utils/compile_mlir_util.cc
@@ -259,13 +259,10 @@
xla::XlaComputation* xla_computation, bool use_tuple_args,
bool return_tuple,
const XlaCompiler::ShapeRepresentationFn shape_representation_fn) {
- // Mark main function as public.
- mlir::FuncOp main_func = module_op.lookupSymbol<mlir::FuncOp>("main");
- if (main_func) {
- main_func.setVisibility(mlir::FuncOp::Visibility::Public);
- }
-
mlir::PassManager tf2xla(module_op.getContext());
+ // Mark main function as public, and other functions as private.
+ tf2xla.addPass(
+ mlir::TF::CreateMarkOnlyMainFunctionWithPublicVisibilityPass());
tf2xla.addNestedPass<mlir::FuncOp>(mlir::createCanonicalizerPass());
tf2xla.addPass(mlir::TF::CreateTensorListOpsDecompositionPass());
tf2xla.addPass(mlir::TF::CreateStackOpsDecompositionPass());
diff --git a/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.cc b/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.cc
index 1853183..aef3363 100644
--- a/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.cc
+++ b/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.cc
@@ -37,18 +37,9 @@
namespace tensorflow {
-const char* const kXlaShardingAttrName = "_XlaSharding";
const char* const kInputShardingAttr = "input_sharding_configuration";
const char* const kOutputShardingAttr = "output_sharding_configuration";
-llvm::Optional<mlir::StringRef> ParseShardingAttribute(
- mlir::Operation* operation) {
- const auto& sharding_attr =
- operation->getAttrOfType<mlir::StringAttr>(kXlaShardingAttrName);
- if (!sharding_attr) return llvm::Optional<mlir::StringRef>();
- return sharding_attr.getValue();
-}
-
namespace {
constexpr char kNumSplitAttr[] = "num_split";
diff --git a/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.h b/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.h
index 77bfd25..52a633d 100644
--- a/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.h
+++ b/tensorflow/compiler/mlir/tensorflow/utils/xla_sharding_util.h
@@ -29,14 +29,9 @@
namespace tensorflow {
-extern const char* const kXlaShardingAttrName;
extern const char* const kInputShardingAttr;
extern const char* const kOutputShardingAttr;
-// Parses "_XlaSharding" attribute from operation, if it exists.
-llvm::Optional<mlir::StringRef> ParseShardingAttribute(
- mlir::Operation* operation);
-
// Parses "input_sharding_configuration" attribute and returns a list where
// i-th element is a list of mlir::Value's which represent inputs for the
// TPU computation correponding to i-th logical device. If the attribute
diff --git a/tensorflow/compiler/mlir/xla/BUILD b/tensorflow/compiler/mlir/xla/BUILD
index 6b186d0..acdfd6d 100644
--- a/tensorflow/compiler/mlir/xla/BUILD
+++ b/tensorflow/compiler/mlir/xla/BUILD
@@ -11,8 +11,8 @@
includes = ["//third_party/mlir:subpackages"],
packages = [
"//babelfish/device/...",
+ "//learning/brain/experimental/dtensor/...",
"//learning/brain/experimental/mlir/...",
- "//learning/brain/experimental/swift_mlir/...",
"//learning/brain/google/xla/kernels/...",
"//learning/brain/swift/swift_mlir/...",
"//learning/pathways/data_parallel/tf2xla/...",
@@ -31,7 +31,7 @@
filegroup(
name = "hlo_ops_td_files",
srcs = [
- "ir/hlo_client_ops.td",
+ "ir/chlo_ops.td",
"ir/hlo_ops.td",
"ir/hlo_ops_base.td",
"ir/hlo_utils.td",
@@ -43,13 +43,13 @@
)
gentbl(
- name = "hlo_client_ops_inc_gen",
+ name = "chlo_ops_inc_gen",
tbl_outs = [
- ("-gen-op-decls", "ir/hlo_client_ops.h.inc"),
- ("-gen-op-defs", "ir/hlo_client_ops.cc.inc"),
+ ("-gen-op-decls", "ir/chlo_ops.h.inc"),
+ ("-gen-op-defs", "ir/chlo_ops.cc.inc"),
],
tblgen = "@llvm-project//mlir:mlir-tblgen",
- td_file = "ir/hlo_client_ops.td",
+ td_file = "ir/chlo_ops.td",
td_srcs = [
":hlo_ops_td_files",
],
@@ -138,6 +138,7 @@
"//tensorflow/compiler/mlir/tensorflow:lower_tf_lib",
"//tensorflow/compiler/xla:xla_data_proto_cc",
"//tensorflow/compiler/xla/client:padding",
+ "//tensorflow/compiler/xla/client:sharding_builder",
"//tensorflow/core:framework",
"//tensorflow/core/kernels:conv_grad_shape_utils",
"@llvm-project//llvm:support",
@@ -393,6 +394,29 @@
)
cc_library(
+ name = "xla_hlo_to_lhlo_with_xla",
+ srcs = [
+ "transforms/xla_hlo_to_lhlo_with_xla.cc",
+ ],
+ deps = [
+ ":hlo",
+ ":hlo_utils",
+ ":lhlo",
+ ":mlir_hlo_to_hlo",
+ ":xla_dialect_registration",
+ "//tensorflow/compiler/xla:statusor",
+ "//tensorflow/compiler/xla:util",
+ "//tensorflow/compiler/xla/service:buffer_assignment",
+ "//tensorflow/compiler/xla/service:hlo",
+ "@llvm-project//llvm:support",
+ "@llvm-project//mlir:IR",
+ "@llvm-project//mlir:Pass",
+ "@llvm-project//mlir:StandardOps",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
name = "xla_legalize_to_standard",
srcs = ["transforms/legalize_to_standard.cc"],
deps = [
@@ -470,17 +494,34 @@
)
cc_library(
+ name = "chlo_legalize_to_hlo",
+ srcs = [
+ "transforms/chlo_legalize_to_hlo.cc",
+ ],
+ deps = [
+ ":hlo",
+ "@llvm-project//llvm:support",
+ "@llvm-project//mlir:IR",
+ "@llvm-project//mlir:Shape",
+ "@llvm-project//mlir:Transforms",
+ ],
+)
+
+cc_library(
name = "xla_test_passes",
srcs = [
+ "transforms/chlo_legalize_to_hlo_pass.cc",
"transforms/materialize_broadcasts_pass.cc",
"transforms/unfuse_batch_norm_pass.cc",
],
deps = [
+ ":chlo_legalize_to_hlo", # build-cleaner: keep
":hlo",
- ":xla_materialize_broadcasts",
- ":xla_unfuse_batch_norm",
+ ":xla_materialize_broadcasts", # build-cleaner: keep
+ ":xla_unfuse_batch_norm", # build-cleaner: keep
"@llvm-project//mlir:IR",
"@llvm-project//mlir:Pass",
+ "@llvm-project//mlir:Shape",
"@llvm-project//mlir:StandardOps",
"@llvm-project//mlir:Transforms",
],
@@ -490,14 +531,14 @@
cc_library(
name = "hlo",
srcs = [
- "ir/hlo_client_ops.cc",
+ "ir/chlo_ops.cc",
"ir/hlo_ops.cc",
"ir/hlo_ops.cc.inc",
"ir/hlo_ops.h.inc",
"ir/hlo_utils.cc",
],
hdrs = [
- "ir/hlo_client_ops.h",
+ "ir/chlo_ops.h",
"ir/hlo_ops.h",
"ir/hlo_utils.h",
"transforms/passes.h",
@@ -505,8 +546,8 @@
],
includes = ["include"],
deps = [
+ ":chlo_ops_inc_gen",
":convert_op_folder",
- ":hlo_client_ops_inc_gen",
":hlo_ops_base_inc_gen",
":hlo_ops_inc_gen",
":xla_canonicalize_inc_gen",
@@ -804,6 +845,7 @@
deps = [
":buffer_assignment",
":buffer_assignment_test",
+ ":chlo_legalize_to_hlo",
":hlo",
":hlo_legalize_to_lhlo",
":lhlo",
@@ -813,6 +855,7 @@
":lhlo_legalize_to_gpu",
":lhlo_legalize_to_parallel_loops",
":xla_dialect_registration",
+ ":xla_hlo_to_lhlo_with_xla",
":xla_legalize_control_flow",
":xla_legalize_tf",
":xla_legalize_tf_with_tf2xla",
@@ -828,6 +871,8 @@
name = "xla-opt",
deps = [
":all_xla_passes_for_testing",
+ "//tensorflow/compiler/jit:xla_cpu_jit",
+ "//tensorflow/compiler/jit:xla_gpu_jit",
"//tensorflow/compiler/mlir:tf_mlir_opt_main",
],
)
diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.cc b/tensorflow/compiler/mlir/xla/ir/chlo_ops.cc
similarity index 83%
rename from tensorflow/compiler/mlir/xla/ir/hlo_client_ops.cc
rename to tensorflow/compiler/mlir/xla/ir/chlo_ops.cc
index 921c4f0..7864fa4 100644
--- a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.cc
+++ b/tensorflow/compiler/mlir/xla/ir/chlo_ops.cc
@@ -13,12 +13,12 @@
limitations under the License.
==============================================================================*/
-#include "tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h"
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.h"
#include "mlir/IR/TypeUtilities.h" // from @llvm-project
namespace mlir {
-namespace xla_hlo_client {
+namespace xla_chlo {
template <typename T>
static LogicalResult Verify(T op) {
@@ -90,38 +90,38 @@
broadcast_dimensions); \
}
-BINARY_BUILDER(AddOp);
-BINARY_BUILDER(AndOp);
-BINARY_BUILDER(Atan2Op);
-BINARY_BUILDER(DivOp);
-BINARY_BUILDER(MaxOp);
-BINARY_BUILDER(MinOp);
-BINARY_BUILDER(MulOp);
-BINARY_BUILDER(OrOp);
-BINARY_BUILDER(PowOp);
-BINARY_BUILDER(RemOp);
-BINARY_BUILDER(ShiftLeftOp);
-BINARY_BUILDER(ShiftRightArithmeticOp);
-BINARY_BUILDER(ShiftRightLogicalOp);
-BINARY_BUILDER(SubOp);
-BINARY_BUILDER(XorOp);
+BINARY_BUILDER(BroadcastAddOp);
+BINARY_BUILDER(BroadcastAndOp);
+BINARY_BUILDER(BroadcastAtan2Op);
+BINARY_BUILDER(BroadcastDivOp);
+BINARY_BUILDER(BroadcastMaxOp);
+BINARY_BUILDER(BroadcastMinOp);
+BINARY_BUILDER(BroadcastMulOp);
+BINARY_BUILDER(BroadcastOrOp);
+BINARY_BUILDER(BroadcastPowOp);
+BINARY_BUILDER(BroadcastRemOp);
+BINARY_BUILDER(BroadcastShiftLeftOp);
+BINARY_BUILDER(BroadcastShiftRightArithmeticOp);
+BINARY_BUILDER(BroadcastShiftRightLogicalOp);
+BINARY_BUILDER(BroadcastSubOp);
+BINARY_BUILDER(BroadcastXorOp);
#undef BINARY_BUILDER
#define GET_OP_CLASSES
-#include "tensorflow/compiler/mlir/xla/ir/hlo_client_ops.cc.inc"
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.cc.inc"
//===----------------------------------------------------------------------===//
-// xla_hlo_client Dialect Constructor
+// xla_chlo Dialect Constructor
//===----------------------------------------------------------------------===//
XlaHloClientDialect::XlaHloClientDialect(MLIRContext* context)
: Dialect(getDialectNamespace(), context) {
addOperations<
#define GET_OP_LIST
-#include "tensorflow/compiler/mlir/xla/ir/hlo_client_ops.cc.inc"
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.cc.inc"
>();
}
-} // namespace xla_hlo_client
+} // namespace xla_chlo
} // namespace mlir
diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h b/tensorflow/compiler/mlir/xla/ir/chlo_ops.h
similarity index 78%
rename from tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h
rename to tensorflow/compiler/mlir/xla/ir/chlo_ops.h
index 405b1ff..21cf463 100644
--- a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h
+++ b/tensorflow/compiler/mlir/xla/ir/chlo_ops.h
@@ -13,8 +13,8 @@
limitations under the License.
==============================================================================*/
-#ifndef TENSORFLOW_COMPILER_MLIR_XLA_IR_HLO_CLIENT_OPS_H_
-#define TENSORFLOW_COMPILER_MLIR_XLA_IR_HLO_CLIENT_OPS_H_
+#ifndef TENSORFLOW_COMPILER_MLIR_XLA_IR_CHLO_OPS_H_
+#define TENSORFLOW_COMPILER_MLIR_XLA_IR_CHLO_OPS_H_
#include "llvm/ADT/StringRef.h"
#include "mlir/IR/Dialect.h" // from @llvm-project
@@ -27,18 +27,18 @@
#include "mlir/Interfaces/SideEffects.h" // from @llvm-project
namespace mlir {
-namespace xla_hlo_client {
+namespace xla_chlo {
class XlaHloClientDialect : public Dialect {
public:
explicit XlaHloClientDialect(MLIRContext *context);
- static StringRef getDialectNamespace() { return "xla_hlo_client"; }
+ static StringRef getDialectNamespace() { return "xla_chlo"; }
};
#define GET_OP_CLASSES
-#include "tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h.inc"
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.h.inc"
-} // namespace xla_hlo_client
+} // namespace xla_chlo
} // namespace mlir
-#endif // TENSORFLOW_COMPILER_MLIR_XLA_IR_HLO_CLIENT_OPS_H_
+#endif // TENSORFLOW_COMPILER_MLIR_XLA_IR_CHLO_OPS_H_
diff --git a/tensorflow/compiler/mlir/xla/ir/chlo_ops.td b/tensorflow/compiler/mlir/xla/ir/chlo_ops.td
new file mode 100644
index 0000000..724e8f9
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/ir/chlo_ops.td
@@ -0,0 +1,313 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// Defines "client" aligned HLO ops.
+// These ops are not necessarily orthogonal or optimized for transformation but
+// for ease of expression in certain cases deemed important for client
+// libraries (i.e. implicit broadcasting, helper ops, etc).
+// This dialect is considered to exist in addition to augment the xla_hlo
+// dialect for ergonomic needs, not duplicate/replace it.
+//
+// The typical use of this dialect is for client libraries to be able to emit
+// less constrained ops and rely on the conversion framework to lower any
+// xla_chlo ops to canonical xla_hlo ops.
+//
+// See: https://www.tensorflow.org/xla/operation_semantics
+
+#ifndef CHLO_OPS
+#define CHLO_OPS
+
+include "mlir/IR/OpBase.td"
+include "mlir/Interfaces/SideEffects.td"
+include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td"
+
+def HLOClient_Dialect : Dialect {
+ let name = "xla_chlo";
+ let cppNamespace = "xla_chlo";
+ let summary = [{
+ XLA Client HLO Ops
+ }];
+
+ let description = [{
+ This dialect contains ops that align closely with the API surface area
+ of the XlaBuilder C++ API, where such ops have semantics that go beyond
+ what exists in the lower level dialects (such as `xla_hlo`). Essentially,
+ whenever the client library uses syntactic sugar or composition
+ of multiple ops for an API call, this dialect tries to model the API call
+ and provide conversion patterns to fully materialize into lower level
+ dialects.
+ }];
+}
+
+class HLOClient_Op<string mnemonic, list<OpTrait> traits> :
+ Op<HLOClient_Dialect, mnemonic, traits> {
+ // TODO(b/129012527) Much of this custom verification should be expressed as
+ // type constraints.
+ let verifier = [{ return Verify(*this); }];
+}
+
+//===----------------------------------------------------------------------===//
+// XLA binary elementwise op definitions.
+// From the client perspective, each of these support both explicit rank
+// broadcasting (via the broadcast_dimensions attribute) and implicit degenerate
+// shape broadcasting.
+//
+// These correspond to operations in the xla_hlo dialect without the
+// "broadcast_" prefix, except that those ops require same-shaped operands and
+// results.
+//
+// See:
+// https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations
+// https://www.tensorflow.org/xla/broadcasting
+//===----------------------------------------------------------------------===//
+
+class HLOClient_BroadcastBinaryElementwiseOp<
+ string mnemonic, list<OpTrait> traits> :
+ HLOClient_Op<mnemonic, traits> {
+ let arguments = (ins
+ HLO_Tensor:$lhs,
+ HLO_Tensor:$rhs,
+ // Explicit rank-broadcast dimension mappings. Defaults to "numpy" prefix
+ // padded rank-broadcast semantics if omitted.
+ OptionalAttr<BroadcastDimAttr>:$broadcast_dimensions
+ );
+
+ let builders = [OpBuilder<
+ "Builder *builder, OperationState &result, Value left, Value right, "
+ "DenseIntElementsAttr broadcast_dimensions"
+ >];
+
+ let results = (outs HLO_Tensor);
+
+ let assemblyFormat = [{
+ $lhs `,` $rhs attr-dict `:`
+ `(` type($lhs) `,` type($rhs) `)` `->` type(results)
+ }];
+}
+
+def HLOClient_BroadcastAddOp : HLOClient_BroadcastBinaryElementwiseOp<"broadcast_add",
+ [Commutative, NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Addition operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs + rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastAtan2Op : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_atan2",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Atan2 operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `atan2(lhs/rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastDivOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_divide",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Division operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs / rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastMaxOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_maximum",
+ [Commutative, NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Maximum operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `max(lhs, rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastMinOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_minimum",
+ [Commutative, NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Minimum operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `min(lhs, rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastMulOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_multiply",
+ [Commutative, NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Multiplication operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs * rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastPowOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_power",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Power operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs ^ rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastRemOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_remainder",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Remainder operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs % rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastShiftLeftOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_shift_left",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Shift left operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs << rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastShiftRightArithmeticOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_shift_right_arithmetic",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Shift right arithmetic operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs >> rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastShiftRightLogicalOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_shift_right_logical",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Shift right logical operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs >> rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastSubOp : HLOClient_BroadcastBinaryElementwiseOp<
+ "broadcast_subtract",
+ [NoSideEffect, SameOperandsAndResultElementType]> {
+ string summary = "Subtraction operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `lhs - rhs` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+//===----------------------------------------------------------------------===//
+// XLA binary elementwise op definitions.
+// The same description as the arithmetic binary elementwise ops applies.
+//===----------------------------------------------------------------------===//
+
+class HLOClient_BroadcastBinaryLogicalElementwiseOp<string mnemonic> :
+ HLOClient_BroadcastBinaryElementwiseOp<
+ mnemonic, [Commutative, NoSideEffect]> {
+ let arguments = (ins
+ HLO_PredOrIntTensor:$lhs,
+ HLO_PredOrIntTensor:$rhs,
+ // Explicit rank-broadcast dimension mappings. Defaults to "numpy" prefix
+ // padded rank-broadcast semantics if omitted.
+ OptionalAttr<BroadcastDimAttr>:$broadcast_dimensions
+ );
+
+ let assemblyFormat = [{
+ $lhs `,` $rhs attr-dict `:`
+ `(` type($lhs) `,` type($rhs) `)` `->` type(results)
+ }];
+}
+
+def HLOClient_BroadcastAndOp: HLOClient_BroadcastBinaryLogicalElementwiseOp<
+ "broadcast_and"> {
+ string summary = "Logical and operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `logical_and(lhs, rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastOrOp: HLOClient_BroadcastBinaryLogicalElementwiseOp<
+ "broadcast_or"> {
+ string summary = "Logical or operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `logical_or(lhs, rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+def HLOClient_BroadcastXorOp : HLOClient_BroadcastBinaryLogicalElementwiseOp<
+ "broadcast_xor"> {
+ string summary = "Logical xor operator (with optional broadcasting)";
+
+ string description = [{
+ Returns `logical_xor(lhs, rhs)` element-wise.
+
+ See
+ https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
+ }];
+}
+
+#endif // CHLO_OPS
diff --git a/tensorflow/compiler/mlir/xla/ir/dialect_registration.cc b/tensorflow/compiler/mlir/xla/ir/dialect_registration.cc
index bafbc1a..2d1bc8d 100644
--- a/tensorflow/compiler/mlir/xla/ir/dialect_registration.cc
+++ b/tensorflow/compiler/mlir/xla/ir/dialect_registration.cc
@@ -13,12 +13,12 @@
limitations under the License.
==============================================================================*/
-#include "tensorflow/compiler/mlir/xla/ir/hlo_client_ops.h"
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.h"
#include "tensorflow/compiler/mlir/xla/ir/hlo_ops.h"
#include "tensorflow/compiler/mlir/xla/ir/lhlo_ops.h"
// Static initialization for XLA dialect registration.
static mlir::DialectRegistration<mlir::xla_hlo::XlaHloDialect> xla_hlo_ops;
-static mlir::DialectRegistration<mlir::xla_hlo_client::XlaHloClientDialect>
- xla_hlo_client_ops;
+static mlir::DialectRegistration<mlir::xla_chlo::XlaHloClientDialect>
+ xla_chlo_ops;
static mlir::DialectRegistration<mlir::xla_lhlo::XlaLhloDialect> xla_lhlo_ops;
diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.td b/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.td
deleted file mode 100644
index 48b765f..0000000
--- a/tensorflow/compiler/mlir/xla/ir/hlo_client_ops.td
+++ /dev/null
@@ -1,134 +0,0 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// Defines "client" aligned HLO ops.
-// These ops are not necessarily orthogonal or optimized for transformation but
-// for ease of expression in certain cases deemed important for client
-// libraries (i.e. implicit broadcasting, helper ops, etc).
-// This dialect is considered to exist in addition to augment the xla_hlo
-// dialect for ergonomic needs, not duplicate/replace it.
-//
-// The typical use of this dialect is for client libraries to be able to emit
-// less constrained ops and rely on the conversion framework to lower any
-// xla_hlo_client ops to canonical xla_hlo ops.
-//
-// See: https://www.tensorflow.org/xla/operation_semantics
-
-#ifndef HLO_CLIENT_OPS
-#define HLO_CLIENT_OPS
-
-include "mlir/IR/OpBase.td"
-include "mlir/Interfaces/SideEffects.td"
-include "tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td"
-
-def HLOClient_Dialect : Dialect {
- let name = "xla_hlo_client";
- let cppNamespace = "xla_hlo_client";
-}
-
-class HLOClient_Op<string mnemonic, list<OpTrait> traits> :
- Op<HLOClient_Dialect, mnemonic, traits> {
- // TODO(b/129012527) Much of this custom verification should be expressed as
- // type constraints.
- let verifier = [{ return Verify(*this); }];
-}
-
-//===----------------------------------------------------------------------===//
-// XLA binary elementwise op definitions.
-// From the client perspective, each of these support both explicit rank
-// broadcasting (via the broadcast_dimensions attribute) and implicit degenerate
-// shape broadcasting.
-//
-// These have 1:1 correspondence with same-named ops in the xla_hlo dialect;
-// however, those operations do not support broadcasting.
-//
-// See:
-// https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations
-// https://www.tensorflow.org/xla/broadcasting
-//===----------------------------------------------------------------------===//
-
-class HLOClient_BinaryElementwiseOp<string mnemonic, list<OpTrait> traits> :
- HLOClient_Op<mnemonic, traits> {
- let arguments = (ins
- HLO_Tensor:$lhs,
- HLO_Tensor:$rhs,
- OptionalAttr<BroadcastDimAttr>:$broadcast_dimensions
- );
-
- let builders = [OpBuilder<
- "Builder *builder, OperationState &result, Value left, Value right, "
- "DenseIntElementsAttr broadcast_dimensions"
- >];
-
- let results = (outs HLO_Tensor);
- let parser = [{ return mlir::impl::parseOneResultSameOperandTypeOp(parser, result); }];
- let printer = [{ return mlir::impl::printOneResultOp(getOperation(), p); }];
-}
-
-def HLOClient_AddOp : HLOClient_BinaryElementwiseOp<"add",
- [Commutative, NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_AddOp;
-
-def HLOClient_Atan2Op : HLOClient_BinaryElementwiseOp<"atan2",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_Atan2Op;
-
-def HLOClient_DivOp : HLOClient_BinaryElementwiseOp<"divide",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_DivOp;
-
-def HLOClient_MaxOp : HLOClient_BinaryElementwiseOp<"maximum",
- [Commutative, NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_MaxOp;
-
-def HLOClient_MinOp : HLOClient_BinaryElementwiseOp<"minimum",
- [Commutative, NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_MinOp;
-
-def HLOClient_MulOp : HLOClient_BinaryElementwiseOp<"multiply",
- [Commutative, NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_MulOp;
-
-def HLOClient_PowOp : HLOClient_BinaryElementwiseOp<"pow",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_PowOp;
-
-def HLOClient_RemOp : HLOClient_BinaryElementwiseOp<"remainder",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_RemOp;
-
-def HLOClient_ShiftLeftOp : HLOClient_BinaryElementwiseOp<"shift_left",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_ShiftLeftOp;
-
-def HLOClient_ShiftRightArithmeticOp : HLOClient_BinaryElementwiseOp<"shift_right_arithmetic",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_ShiftRightArithmeticOp;
-
-def HLOClient_ShiftRightLogicalOp : HLOClient_BinaryElementwiseOp<"shift_right_logical",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_ShiftRightLogicalOp;
-
-def HLOClient_SubOp : HLOClient_BinaryElementwiseOp<"subtract",
- [NoSideEffect, SameOperandsAndResultElementType]>, BASE_HLO_SubOp;
-
-//===----------------------------------------------------------------------===//
-// XLA binary elementwise op definitions.
-// The same description as the arithmetic binary elementwise ops applies.
-//===----------------------------------------------------------------------===//
-
-class HLOClient_BinaryLogicalElementwiseOp<string mnemonic> :
- HLOClient_BinaryElementwiseOp<mnemonic, [Commutative, NoSideEffect]> {
- let arguments = (ins
- HLO_PredOrIntTensor:$lhs,
- HLO_PredOrIntTensor:$rhs,
- OptionalAttr<BroadcastDimAttr>:$broadcast_dimensions
- );
-}
-
-def HLOClient_AndOp: HLOClient_BinaryLogicalElementwiseOp<"and">, BASE_HLO_AndOp;
-def HLOClient_OrOp: HLOClient_BinaryLogicalElementwiseOp<"or">, BASE_HLO_OrOp;
-def HLOClient_XorOp : HLOClient_BinaryLogicalElementwiseOp<"xor">, BASE_HLO_XorOp;
-
-#endif // HLO_CLIENT_OPS
diff --git a/tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td b/tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td
index 7994026..287ad1b 100644
--- a/tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td
+++ b/tensorflow/compiler/mlir/xla/ir/hlo_ops_base.td
@@ -517,7 +517,7 @@
string summary = "Logical and";
string description = [{
- Returns `lhs /\ rhs` element-wise.
+ Returns `logical_and(lhs, rhs)` element-wise.
See
https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
@@ -528,7 +528,7 @@
string summary = "Logical or";
string description = [{
- Returns `lhs \/ rhs` element-wise.
+ Returns `logical_or(lhs, rhs)` element-wise.
See
https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
@@ -539,7 +539,7 @@
string summary = "Logical xor";
string description = [{
- Returns `lhs xor rhs` element-wise.
+ Returns `logical_xor(lhs, rhs)` element-wise.
See
https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations.
diff --git a/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td b/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
index 7613f1e..f0c84db 100644
--- a/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
+++ b/tensorflow/compiler/mlir/xla/ir/lhlo_ops.td
@@ -37,13 +37,12 @@
// Any floating-point tensor types
def LHLO_FpBuffer : MemRefOf<[AnyFloat]>;
-
def LHLO_PredBuffer : MemRefOf<[HLO_Pred]>;
// Any integer or floating-point tensor types
def LHLO_IntOrFpBuffer : MemRefOf<[HLO_Int, AnyFloat]>;
-def LHLO_Buffer : MemRefOf<[AnyFloat, AnySignlessInteger]>;
+def LHLO_Buffer : MemRefOf<[AnyFloat, AnySignlessInteger, AnyComplex]>;
def LHLO_TupleBuffer : NestedTupleOf<[LHLO_Buffer]>;
@@ -106,8 +105,30 @@
def LHLO_TanhOp: LHLO_UnaryElementwiseOp<"tanh">, BASE_HLO_TanhOp;
//===----------------------------------------------------------------------===//
+// XLA complex unary elementwise op definitions.
+//===----------------------------------------------------------------------===//
+// See https://www.tensorflow.org/xla/operation_semantics#element-wise_unary_functions
+
+def LHLO_ComplexOp: LHLO_Op<"complex", [SameOperandsShape]>, BASE_HLO_ComplexOp {
+ let arguments = (ins Arg<LHLO_Buffer, "", [MemRead]>:$lhs,
+ Arg<LHLO_Buffer, "", [MemRead]>:$rhs,
+ Arg<LHLO_Buffer, "", [MemWrite]>:$output);
+}
+
+def LHLO_ImagOp: LHLO_Op<"imag", [SameOperandsShape]>, BASE_HLO_ImagOp {
+ let arguments = (ins Arg<LHLO_Buffer, "", [MemRead]>:$input,
+ Arg<LHLO_Buffer, "", [MemWrite]>:$output);
+}
+
+def LHLO_RealOp: LHLO_Op<"real", [SameOperandsShape]>, BASE_HLO_RealOp {
+ let arguments = (ins Arg<LHLO_Buffer, "", [MemRead]>:$input,
+ Arg<LHLO_Buffer, "", [MemWrite]>:$output);
+}
+
+//===----------------------------------------------------------------------===//
// XLA binary elementwise op definitions.
//===----------------------------------------------------------------------===//
+// See https://www.tensorflow.org/xla/operation_semantics#element-wise_binary_arithmetic_operations
class LHLO_BinaryElementwiseOp<string mnemonic, list<OpTrait> traits> :
LHLO_Op<mnemonic, traits> {
diff --git a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
index 739f19e..cfa8c1b 100644
--- a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
+++ b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.cc
@@ -165,6 +165,90 @@
/*attributes=*/{});
}
+XlaOp MlirHloBuilder::CreateToken() {
+ return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::CreateTokenOp>(
+ loc_, mlir::xla_hlo::TokenType::get(builder_.getContext())));
+ });
+}
+
+StatusOr<XlaOp> MlirHloBuilder::InfeedWithTokenInternal(
+ const Shape& infeed_instruction_shape, XlaOp token, const string& config) {
+ TF_ASSIGN_OR_RETURN(mlir::Type result_type,
+ ConvertShapeToType<mlir::RankedTensorType>(
+ infeed_instruction_shape, builder_));
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::InfeedOp>(
+ loc_, result_type, GetValue(token),
+ /*infeed_config=*/config));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::OutfeedWithTokenInternal(
+ XlaOp operand, XlaOp token, const Shape& shape_with_layout,
+ const string& outfeed_config) {
+ auto token_type = mlir::xla_hlo::TokenType::get(builder_.getContext());
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::OutfeedOp>(
+ loc_, token_type, GetValue(operand), GetValue(token), outfeed_config));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::ConcatInDimInternal(
+ const Shape& shape, absl::Span<const XlaOp> operands, int64 dimension) {
+ TF_ASSIGN_OR_RETURN(
+ mlir::Type result_type,
+ ConvertShapeToType<mlir::RankedTensorType>(shape, builder_));
+ auto mlir_operands = GetValues(operands);
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::ConcatenateOp>(
+ loc_, result_type, mlir_operands, builder_.getI64IntegerAttr(dimension)));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::GetTupleElementInternal(const Shape& shape,
+ XlaOp tuple_data,
+ int64 index) {
+ TF_ASSIGN_OR_RETURN(
+ mlir::Type result_type,
+ ConvertShapeToType<mlir::RankedTensorType>(shape, builder_));
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::GetTupleElementOp>(
+ loc_, result_type, GetValue(tuple_data),
+ builder_.getI32IntegerAttr(index)));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::SliceInternal(
+ const Shape& shape, XlaOp operand, absl::Span<const int64> start_indices,
+ absl::Span<const int64> limit_indices, absl::Span<const int64> strides) {
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::SliceOp>(
+ loc_, GetValue(operand), GetI64ElementsAttr(start_indices, &builder_),
+ GetI64ElementsAttr(limit_indices, &builder_),
+ GetI64ElementsAttr(strides, &builder_)));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::PadInternal(
+ const Shape& shape, XlaOp operand, XlaOp padding_value,
+ const PaddingConfig& padding_config) {
+ TF_ASSIGN_OR_RETURN(
+ mlir::Type result_type,
+ ConvertShapeToType<mlir::RankedTensorType>(shape, builder_));
+ std::vector<int64> low;
+ std::vector<int64> high;
+ std::vector<int64> internal;
+ for (auto& dimension : padding_config.dimensions()) {
+ low.push_back(dimension.edge_padding_low());
+ high.push_back(dimension.edge_padding_high());
+ internal.push_back(dimension.interior_padding());
+ }
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::PadOp>(
+ loc_, result_type, GetValue(operand), GetValue(padding_value),
+ GetI64ElementsAttr(low, &builder_), GetI64ElementsAttr(high, &builder_),
+ GetI64ElementsAttr(internal, &builder_)));
+}
+
+StatusOr<XlaOp> MlirHloBuilder::TupleInternal(
+ const Shape& shape, absl::Span<const XlaOp> elements) {
+ mlir::SmallVector<mlir::Value, 4> operands;
+ for (auto& element : elements) {
+ operands.push_back(GetValue(element));
+ }
+ return MakeXlaOp(builder_.create<mlir::xla_hlo::TupleOp>(loc_, operands));
+}
+
StatusOr<XlaOp> MlirHloBuilder::CreateOp(
const std::string& op_name, const Shape& shape,
llvm::ArrayRef<XlaOp> operands,
diff --git a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
index 95dafbd..c0ef645 100644
--- a/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
+++ b/tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h
@@ -54,6 +54,9 @@
// TODO(hinsu): Add a constructor to build a new MLIR function from scratch
// and override Build methods.
+ MlirHloBuilder(std::string name, mlir::OpBuilder builder, mlir::Location loc)
+ : XlaBuilder(name), builder_(builder), loc_(loc) {}
+
MlirHloBuilder(const MlirHloBuilder&) = delete;
MlirHloBuilder& operator=(const MlirHloBuilder&) = delete;
@@ -75,6 +78,17 @@
return mlir::Value::getFromOpaquePointer(ptr);
}
+ // Returns MLIR values corresponding to the given XLA ops.
+ //
+ // Requires that the ops were created by this builder.
+ std::vector<mlir::Value> GetValues(absl::Span<const XlaOp> ops) {
+ std::vector<mlir::Value> values;
+ for (auto xla_op : ops) {
+ values.push_back(GetValue(xla_op));
+ }
+ return values;
+ }
+
// Sets location for newly built ops, until reset.
void SetLocation(mlir::Location loc) { loc_ = loc; }
@@ -120,6 +134,34 @@
StatusOr<XlaOp> AddOpWithShape(HloOpcode opcode, const Shape& shape,
absl::Span<const XlaOp> operands) override;
+ XlaOp CreateToken() override;
+
+ StatusOr<XlaOp> InfeedWithTokenInternal(const Shape& infeed_instruction_shape,
+ XlaOp token,
+ const string& config) override;
+ StatusOr<XlaOp> OutfeedWithTokenInternal(
+ XlaOp operand, XlaOp token, const Shape& shape_with_layout,
+ const string& outfeed_config) override;
+
+ StatusOr<XlaOp> ConcatInDimInternal(const Shape& shape,
+ absl::Span<const XlaOp> operands,
+ int64 dimension) override;
+
+ StatusOr<XlaOp> GetTupleElementInternal(const Shape& shape, XlaOp tuple_data,
+ int64 index) override;
+
+ StatusOr<XlaOp> SliceInternal(const Shape& shape, XlaOp operand,
+ absl::Span<const int64> start_indices,
+ absl::Span<const int64> limit_indices,
+ absl::Span<const int64> strides) override;
+
+ StatusOr<XlaOp> PadInternal(const Shape& shape, XlaOp operand,
+ XlaOp padding_value,
+ const PaddingConfig& padding_config) override;
+
+ StatusOr<XlaOp> TupleInternal(const Shape& shape,
+ absl::Span<const XlaOp> elements) override;
+
// Creates HLO dialect op and returns the result as an XlaOp.
StatusOr<XlaOp> CreateOp(const std::string& op_name, const Shape& shape,
llvm::ArrayRef<XlaOp> operands,
diff --git a/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.h b/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.h
index 1a341b0..8bfe4c7 100644
--- a/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.h
+++ b/tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.h
@@ -31,16 +31,16 @@
// are converted to a tuple even when there is only a single return value.
// Multiple return values are always converted to a tuple and returned as a
// single value.
-Status ConvertMlirHloToHlo(mlir::ModuleOp module, xla::HloProto* hlo_proto,
+Status ConvertMlirHloToHlo(mlir::ModuleOp module, ::xla::HloProto* hlo_proto,
bool use_tuple_args, bool return_tuple,
const tensorflow::XlaCompiler::ShapeRepresentationFn
shape_representation_fn = nullptr);
// Creates XlaOp equivalent of a given MLIR operation using the operand info
// from `value_lowering` map.
-llvm::Optional<xla::XlaOp> CreateXlaOperator(
+llvm::Optional<::xla::XlaOp> CreateXlaOperator(
mlir::Operation* op,
- llvm::DenseMap<mlir::Value, xla::XlaOp>* value_lowering);
+ llvm::DenseMap<mlir::Value, ::xla::XlaOp>* value_lowering);
} // namespace mlir
diff --git a/tensorflow/compiler/mlir/xla/tests/BUILD b/tensorflow/compiler/mlir/xla/tests/BUILD
index 989b846..ad69383 100644
--- a/tensorflow/compiler/mlir/xla/tests/BUILD
+++ b/tensorflow/compiler/mlir/xla/tests/BUILD
@@ -1,4 +1,5 @@
load("//tensorflow/compiler/mlir:glob_lit_test.bzl", "glob_lit_tests")
+load("//tensorflow:tensorflow.bzl", "tf_cc_test")
package(licenses = ["notice"])
@@ -18,3 +19,18 @@
"@llvm-project//llvm:FileCheck",
],
)
+
+tf_cc_test(
+ name = "mlir_hlo_builder_test",
+ srcs = ["mlir_hlo_builder_test.cc"],
+ deps = [
+ "//tensorflow/compiler/mlir/xla:hlo",
+ "//tensorflow/compiler/mlir/xla:mlir_hlo_builder",
+ "//tensorflow/compiler/xla:shape_util",
+ "//tensorflow/core:test",
+ "//tensorflow/core:test_main",
+ "//tensorflow/core:testlib",
+ "@llvm-project//llvm:support",
+ "@llvm-project//mlir:IR",
+ ],
+)
diff --git a/tensorflow/compiler/mlir/xla/tests/chlo_legalize_to_hlo_broadcasts.mlir b/tensorflow/compiler/mlir/xla/tests/chlo_legalize_to_hlo_broadcasts.mlir
new file mode 100644
index 0000000..895996e
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/tests/chlo_legalize_to_hlo_broadcasts.mlir
@@ -0,0 +1,177 @@
+// RUN: xla-opt -test-xla-chlo-legalize-to-hlo -split-input-file -verify-diagnostics %s -o - | FileCheck --dump-input=fail %s
+
+// Check the non-broadcast case for each registered op, then just check a
+// representative op for detailed broadcast semantics.
+// CHECK-LABEL: @addWithoutBroadcast
+func @addWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.add %arg0, %arg1
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @dynamicBroadcast
+// CHECK-SAME: %[[ARG0:.+]]: tensor<4xf32>
+// CHECK-SAME: %[[ARG1:.+]]: tensor<1x4xf32>
+func @dynamicBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<1x4xf32>) -> tensor<1x4xf32> {
+ // CHECK-DAG: %[[ARG0_S:.+]] = "shape.shape_of"(%[[ARG0]])
+ // CHECK-DAG: %[[ARG1_S:.+]] = "shape.shape_of"(%[[ARG1]])
+ // CHECK-DAG: %[[RESULT_S:.+]] = "shape.broadcast"(%[[ARG0_S]], %[[ARG1_S]])
+ // CHECK: %[[RESULT_EXTENTS:.+]] = "shape.to_extent_tensor"(%[[RESULT_S]])
+ // CHECK-DAG: %[[ARG0_B:.+]] = "xla_hlo.dynamic_broadcast_in_dim"(%[[ARG0]], %[[RESULT_EXTENTS]]) {broadcast_dimensions = dense<1> : tensor<1xi64>}
+ // CHECK-DAG: %[[ARG1_B:.+]] = "xla_hlo.dynamic_broadcast_in_dim"(%[[ARG1]], %[[RESULT_EXTENTS]]) {broadcast_dimensions = dense<[0, 1]> : tensor<2xi64>}
+ // CHECK-DAG: %[[RESULT:.+]] = xla_hlo.add %[[ARG0_B]], %[[ARG1_B]]
+ // CHECK: return %[[RESULT]] : tensor<1x4xf32>
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 : (tensor<4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
+ return %0 : tensor<1x4xf32>
+}
+
+// -----
+// Verifies that broadcast_dimensions validity checks are valid.
+// CHECK-LABEL: @dynamicNonScalarBroadcastDimensions
+func @dynamicNonScalarBroadcastDimensions(%arg0: tensor<1x4xf32>, %arg1: tensor<4xf32>) -> tensor<1x4xf32> {
+ // CHECK: xla_hlo.add
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 {broadcast_dimensions = dense<1> : tensor<1xi64>} : (tensor<1x4xf32>, tensor<4xf32>) -> tensor<1x4xf32>
+ return %0 : tensor<1x4xf32>
+}
+
+// -----
+// Verifies that broadcast_dimensions validity checks are valid.
+// CHECK-LABEL: @dynamicNonScalarByScalarBroadcastDimensions
+func @dynamicNonScalarByScalarBroadcastDimensions(%arg0: tensor<1x4xf32>, %arg1: tensor<f32>) -> tensor<1x4xf32> {
+ // CHECK: xla_hlo.add
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 {broadcast_dimensions = dense<[]> : tensor<0xi64>} : (tensor<1x4xf32>, tensor<f32>) -> tensor<1x4xf32>
+ return %0 : tensor<1x4xf32>
+}
+
+// -----
+// Verifies that invalid broadcast dimensions are rejected.
+func @dynamicNonScalarBroadcastDimensionsSizeMismatch(%arg0: tensor<1x4xf32>, %arg1: tensor<4xf32>) -> tensor<1x4xf32> {
+ // expected-warning @+2 {{unsupported non prefix-padded dynamic rank broadcast_dimensions}}
+ // expected-error @+1 {{failed to legalize operation}}
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 {broadcast_dimensions = dense<[1, 2]> : tensor<2xi64>} : (tensor<1x4xf32>, tensor<4xf32>) -> tensor<1x4xf32>
+ return %0 : tensor<1x4xf32>
+}
+
+// -----
+// Verifies that invalid broadcast dimensions are rejected.
+func @dynamicNonScalarBroadcastDimensionsMismatch(%arg0: tensor<1x4xf32>, %arg1: tensor<4xf32>) -> tensor<1x4xf32> {
+ // expected-warning @+2 {{unsupported non prefix-padded dynamic rank broadcast_dimensions}}
+ // expected-error @+1 {{failed to legalize operation}}
+ %0 = xla_chlo.broadcast_add %arg0, %arg1 {broadcast_dimensions = dense<2> : tensor<1xi64>} : (tensor<1x4xf32>, tensor<4xf32>) -> tensor<1x4xf32>
+ return %0 : tensor<1x4xf32>
+}
+
+// -----
+// Note that broadcast_add is used as a proxy for all of the template
+// expansions. Tests below merely verify that the op has an expansion.
+// CHECK-LABEL: @atan2WithoutBroadcast
+func @atan2WithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.atan2 %arg0, %arg1
+ %0 = xla_chlo.broadcast_atan2 %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @divideWithoutBroadcast
+func @divideWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.divide %arg0, %arg1
+ %0 = xla_chlo.broadcast_divide %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @maximumWithoutBroadcast
+func @maximumWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.maximum %arg0, %arg1
+ %0 = xla_chlo.broadcast_maximum %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @minimumWithoutBroadcast
+func @minimumWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.minimum %arg0, %arg1
+ %0 = xla_chlo.broadcast_minimum %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @multiplyWithoutBroadcast
+func @multiplyWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.multiply %arg0, %arg1
+ %0 = xla_chlo.broadcast_multiply %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @powerWithoutBroadcast
+func @powerWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.power %arg0, %arg1
+ %0 = xla_chlo.broadcast_power %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @remainderWithoutBroadcast
+func @remainderWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.remainder %arg0, %arg1
+ %0 = xla_chlo.broadcast_remainder %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @shift_leftWithoutBroadcast
+func @shift_leftWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.shift_left %arg0, %arg1
+ %0 = xla_chlo.broadcast_shift_left %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @shift_right_arithmeticWithoutBroadcast
+func @shift_right_arithmeticWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.shift_right_arithmetic %arg0, %arg1
+ %0 = xla_chlo.broadcast_shift_right_arithmetic %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @shift_right_logicalWithoutBroadcast
+func @shift_right_logicalWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.shift_right_logical %arg0, %arg1
+ %0 = xla_chlo.broadcast_shift_right_logical %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @subWithoutBroadcast
+func @subWithoutBroadcast(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
+ // CHECK: xla_hlo.subtract %arg0, %arg1
+ %0 = xla_chlo.broadcast_subtract %arg0, %arg1 : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+// CHECK-LABEL: @andWithoutBroadcast
+func @andWithoutBroadcast(%arg0: tensor<4xi1>, %arg1: tensor<4xi1>) -> tensor<4xi1> {
+ // CHECK: xla_hlo.and %arg0, %arg1
+ %0 = xla_chlo.broadcast_and %arg0, %arg1 : (tensor<4xi1>, tensor<4xi1>) -> tensor<4xi1>
+ return %0 : tensor<4xi1>
+}
+
+// -----
+// CHECK-LABEL: @orWithoutBroadcast
+func @orWithoutBroadcast(%arg0: tensor<4xi1>, %arg1: tensor<4xi1>) -> tensor<4xi1> {
+ // CHECK: xla_hlo.or %arg0, %arg1
+ %0 = xla_chlo.broadcast_or %arg0, %arg1 : (tensor<4xi1>, tensor<4xi1>) -> tensor<4xi1>
+ return %0 : tensor<4xi1>
+}
+
+// -----
+// CHECK-LABEL: @xorWithoutBroadcast
+func @xorWithoutBroadcast(%arg0: tensor<4xi1>, %arg1: tensor<4xi1>) -> tensor<4xi1> {
+ // CHECK: xla_hlo.xor %arg0, %arg1
+ %0 = xla_chlo.broadcast_xor %arg0, %arg1 : (tensor<4xi1>, tensor<4xi1>) -> tensor<4xi1>
+ return %0 : tensor<4xi1>
+}
diff --git a/tensorflow/compiler/mlir/xla/tests/hlo-legalize-to-lhlo.mlir b/tensorflow/compiler/mlir/xla/tests/hlo-legalize-to-lhlo.mlir
index c457f3d..262533b 100644
--- a/tensorflow/compiler/mlir/xla/tests/hlo-legalize-to-lhlo.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/hlo-legalize-to-lhlo.mlir
@@ -174,6 +174,45 @@
// -----
+// CHECK-LABEL: func @complex
+func @complex(%real: memref<2x2xf32>,
+ %imag: memref<2x2xf32>,
+ %result: memref<2x2xcomplex<f32>>) {
+ %tensor_real = tensor_load %real : memref<2x2xf32>
+ %tensor_imag = tensor_load %imag : memref<2x2xf32>
+ %tensor_result = "xla_hlo.complex"(%tensor_real, %tensor_imag)
+ : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xcomplex<f32>>
+ // CHECK: "xla_lhlo.complex"(%{{.*}}, %{{.*}})
+ tensor_store %tensor_result, %result : memref<2x2xcomplex<f32>>
+ return
+}
+
+// -----
+
+// CHECK-LABEL: func @real
+func @real(%operand: memref<2x2xcomplex<f32>>, %result: memref<2x2xf32>) {
+ %tensor_operand = tensor_load %operand : memref<2x2xcomplex<f32>>
+ %tensor_result = "xla_hlo.real"(%tensor_operand)
+ : (tensor<2x2xcomplex<f32>>) -> tensor<2x2xf32>
+ // CHECK: "xla_lhlo.real"(%{{.*}}, %{{.*}})
+ tensor_store %tensor_result, %result : memref<2x2xf32>
+ return
+}
+
+// -----
+
+// CHECK-LABEL: func @imag
+func @imag(%operand: memref<2x2xcomplex<f32>>, %result: memref<2x2xf32>) {
+ %tensor_operand = tensor_load %operand : memref<2x2xcomplex<f32>>
+ %tensor_result = "xla_hlo.imag"(%tensor_operand)
+ : (tensor<2x2xcomplex<f32>>) -> tensor<2x2xf32>
+ // CHECK: "xla_lhlo.imag"(%{{.*}}, %{{.*}})
+ tensor_store %tensor_result, %result : memref<2x2xf32>
+ return
+}
+
+// -----
+
// CHECK-LABEL: func @iota
func @iota(%result: memref<10xi32>) {
%tensor_result = "xla_hlo.iota"()
diff --git a/tensorflow/compiler/mlir/xla/tests/hlo_to_lhlo_with_xla/passthrough.mlir b/tensorflow/compiler/mlir/xla/tests/hlo_to_lhlo_with_xla/passthrough.mlir
new file mode 100644
index 0000000..cda1dc4
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/tests/hlo_to_lhlo_with_xla/passthrough.mlir
@@ -0,0 +1,15 @@
+// RUN: xla-opt -xla-hlo-to-lhlo-with-xla %s | FileCheck --enable-var-scope --dump-input=fail %s
+
+// Current allocation will lead to one buffer argument for the "value" and
+// another one for the output, an no returned values.
+// CHECK-LABEL: func @main
+// CHECK-SAME: %[[ARG0:.*]]: memref<2x2xf32> {xla_lhlo.params = 0 : index},
+// CHECK-SAME: %[[ARG1:.*]]: memref<16xi8> {xla_lhlo.alloc = 0 : index, xla_lhlo.liveout = true}
+// CHECK-SAME: ) {
+func @main(%value: tensor<2x2xf32>) -> tensor<2x2xf32> {
+ // The only expected instruction is a copy from the input into the output.
+ // CHECK: %[[OUTPUT:.*]] = std.view %[[ARG1]][][] : memref<16xi8> to memref<2x2xf32>
+ // CHECK: xla_lhlo.copy
+ // CHECK-SAME: %[[ARG0]], %[[OUTPUT]]
+ return %value : tensor<2x2xf32>
+}
diff --git a/tensorflow/compiler/mlir/xla/tests/legalize-tf.mlir b/tensorflow/compiler/mlir/xla/tests/legalize-tf.mlir
index cb27b57..0ae8594 100644
--- a/tensorflow/compiler/mlir/xla/tests/legalize-tf.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/legalize-tf.mlir
@@ -27,14 +27,20 @@
}
// CHECK-LABEL: fusedBatchNormV3_noTraining_mixedPrecision
-func @fusedBatchNormV3_noTraining_mixedPrecision(%arg0: tensor<8x8x8x8xbf16>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> (tensor<8x8x8x8xbf16>) {
- // CHECK: %[[RESULT0:.*]] = "xla_hlo.convert"(%arg0) : (tensor<8x8x8x8xbf16>) -> tensor<8x8x8x8xf32>
- // CHECK: %[[RESULT1:.*]] = "xla_hlo.batch_norm_inference"(%[[RESULT0]], %arg1, %arg2, %arg3, %arg4) {epsilon = 1.000000e-03 : f32, feature_index = 3 : i64} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> tensor<8x8x8x8xf32>
- %0:6 = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)
- // CHECK-NEXT: "xla_hlo.convert"(%[[RESULT1]]) : (tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xbf16>
- return %0#0 : tensor<8x8x8x8xbf16>
+// CHECK-SAME: ([[X:%.*]]: tensor<8x8x8x8xbf16>, [[SCALE:%.*]]: tensor<8xf32>, [[OFFSET:%.*]]: tensor<8xf32>, [[MEAN:%.*]]: tensor<8xf32>, [[VARIANCE:%.*]]: tensor<8xf32>)
+func @fusedBatchNormV3_noTraining_mixedPrecision(%arg0: tensor<8x8x8x8xbf16>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> (tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>) {
+ // CHECK: [[CONVERT_X:%.*]] = "xla_hlo.convert"([[X]]) : (tensor<8x8x8x8xbf16>) -> tensor<8x8x8x8xf32>
+ // CHECK: [[Y:%.*]] = "xla_hlo.batch_norm_inference"([[CONVERT_X]], [[SCALE]], [[OFFSET]], [[MEAN]], [[VARIANCE]]) {epsilon = 1.000000e-03 : f32, feature_index = 3 : i64}
+ %0:6 = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>)
+ // CHECK: [[Y_CONVERT:%.*]] = "xla_hlo.convert"([[Y]]) : (tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xbf16>
+ // CHECK: [[DUMMY:%.*]] = xla_hlo.constant {value = dense<0.000000e+00> : tensor<0xf32>} : tensor<*xf32>
+ // CHECK: return [[Y_CONVERT]], [[MEAN]], [[VARIANCE]], [[MEAN]], [[VARIANCE]], [[DUMMY]]
+ return %0#0, %0#1, %0#2, %0#3, %0#4, %0#5 : tensor<8x8x8x8xbf16>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<*xf32>
}
+
+
+
// CHECK-LABEL: fusedBatchNormV3_training
func @fusedBatchNormV3_training(%arg0: tensor<8x8x8x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> (tensor<8x8x8x8xf32>) {
// CHECK: %[[RESULT0:.*]] = "xla_hlo.batch_norm_training"({{.*}}, %arg1, %arg2) {epsilon = 1.000000e-03 : f32, feature_index = 3 : i64} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>) -> tuple<tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>>
@@ -1163,6 +1169,26 @@
return %0#0, %0#1 : tensor<3xi32>, tensor<4xf32>
}
+// The following op sharding is used:
+// Proto debug string:
+// type: TUPLE
+// tuple_shardings {
+// type: MAXIMAL
+// tile_assignment_dimensions: 1
+// tile_assignment_devices: 0
+// }
+// Serialized string:
+// "\08\02*\08\08\01\1A\01\01\22\01\00"
+
+// CHECK-LABEL: infeed_dequeue_tuple_sharding
+func @infeed_dequeue_tuple_sharding() -> tensor<8xi32> {
+ // CHECK: "xla_hlo.infeed"
+ // An additional sharding is added at the end to account for token result.
+ // CHECK-SAME: xla_hlo.sharding = "type: TUPLE\0Atuple_shardings {\0A type: MAXIMAL\0A tile_assignment_dimensions: 1\0A tile_assignment_devices: 0\0A}\0Atuple_shardings {\0A type: MAXIMAL\0A tile_assignment_dimensions: 1\0A tile_assignment_devices: 0\0A}\0A"
+ %0 = "tf.InfeedDequeueTuple"() {_XlaSharding = "\08\02*\08\08\01\1A\01\01\22\01\00"} : () -> tensor<8xi32>
+ return %0 : tensor<8xi32>
+}
+
//===----------------------------------------------------------------------===//
// Nullary op legalizations.
//===----------------------------------------------------------------------===//
@@ -1348,6 +1374,28 @@
return %0 : tensor<2x4x7x7xi32>
}
+// CHECK-LABEL: maxpool_3d_valid_padding
+// CHECK-SAME: %[[ARG:.*]]: tensor
+func @maxpool_3d_valid_padding(%arg0: tensor<2x8x12x20x7xf32>) -> tensor<2x8x3x5x7xf32> {
+ // CHECK: %[[INIT:.*]] = xla_hlo.constant dense<0xFF800000> : tensor<f32>
+ // CHECK: "xla_hlo.reduce_window"(%[[ARG]], %[[INIT]])
+ // CHECK: xla_hlo.maximum
+ // CHECK: xla_hlo.return
+ // CHECK: {window_dimensions = dense<[1, 1, 2, 2, 1]> : tensor<5xi64>, window_strides = dense<[1, 1, 4, 4, 1]> : tensor<5xi64>}
+
+ %0 = "tf.MaxPool3D"(%arg0) {data_format = "NDHWC", ksize = [1, 1, 2, 2, 1], padding = "VALID", strides = [1, 1, 4, 4, 1]} : (tensor<2x8x12x20x7xf32>) -> tensor<2x8x3x5x7xf32>
+ return %0 : tensor<2x8x3x5x7xf32>
+}
+
+// CHECK-LABEL: maxpool_3d_same_padding
+// CHECK-SAME: %[[ARG:.*]]: tensor
+func @maxpool_3d_same_padding(%arg0: tensor<2x8x13x25x7xf32>) -> tensor<2x8x4x7x7xf32> {
+ // CHECK: padding = dense<{{\[\[}}0, 0], [0, 0], [0, 1], [1, 1], [0, 0]]> : tensor<5x2xi64>
+
+ %0 = "tf.MaxPool3D"(%arg0) {data_format = "NDHWC", ksize = [1, 1, 2, 3, 1], padding = "SAME", strides = [1, 1, 4, 4, 1]} : (tensor<2x8x13x25x7xf32>) -> tensor<2x8x4x7x7xf32>
+ return %0 : tensor<2x8x4x7x7xf32>
+}
+
//===----------------------------------------------------------------------===//
// MaxPoolGrad op legalizations.
//===----------------------------------------------------------------------===//
@@ -1376,6 +1424,25 @@
return %result : tensor<10x24x24x64xf32>
}
+// CHECK-LABEL: @max_pool_3d_grad_valid
+// CHECK-SAME: %[[INPUT:.*]]: tensor<10x8x24x24x64xf32>, %arg1: tensor<10x8x12x12x64xf32>, %[[GRAD:.*]]: tensor<10x8x12x12x64xf32>
+func @max_pool_3d_grad_valid(%orig_input: tensor<10x8x24x24x64xf32>, %orig_output: tensor<10x8x12x12x64xf32>, %grad: tensor<10x8x12x12x64xf32>) -> tensor<10x8x24x24x64xf32> {
+ // CHECK: %[[ZERO:.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+ // CHECK: %[[RESULT:.*]] = "xla_hlo.select_and_scatter"(%[[INPUT]], %[[GRAD]], %[[ZERO]]) ( {
+ // CHECK: ^bb0(%[[VALUE_A:.*]]: tensor<f32>, %[[VALUE_B:.*]]: tensor<f32>):
+ // CHECK: %[[SELECT_RESULT:.*]] = "xla_hlo.compare"(%[[VALUE_A]], %[[VALUE_B]]) {comparison_direction = "GE"} : (tensor<f32>, tensor<f32>) -> tensor<i1>
+ // CHECK: "xla_hlo.return"(%[[SELECT_RESULT]]) : (tensor<i1>) -> ()
+ // CHECK: }, {
+ // CHECK: ^bb0(%[[VALUE_A:.*]]: tensor<f32>, %[[VALUE_B:.*]]: tensor<f32>):
+ // CHECK: %[[SELECT_RESULT:.*]] = xla_hlo.add %[[VALUE_A]], %[[VALUE_B]] : tensor<f32>
+ // CHECK: "xla_hlo.return"(%[[SELECT_RESULT]]) : (tensor<f32>) -> ()
+ // CHECK: }) {window_dimensions = dense<[1, 1, 2, 2, 1]> : tensor<5xi64>, window_strides = dense<[1, 1, 2, 2, 1]> : tensor<5xi64>} : (tensor<10x8x24x24x64xf32>, tensor<10x8x12x12x64xf32>, tensor<f32>) -> tensor<10x8x24x24x64xf32>
+ // CHECK: return %[[RESULT]] : tensor<10x8x24x24x64xf32>
+ // CHECK: }
+ %result = "tf.MaxPool3DGrad"(%orig_input, %orig_output, %grad) {data_format = "NDHWC", ksize = [1, 1, 2, 2, 1], padding = "VALID", strides = [1, 1, 2, 2, 1]} : (tensor<10x8x24x24x64xf32>, tensor<10x8x12x12x64xf32>, tensor<10x8x12x12x64xf32>) -> tensor<10x8x24x24x64xf32>
+ return %result : tensor<10x8x24x24x64xf32>
+}
+
// CHECK-LABEL: @max_pool_grad_same
func @max_pool_grad_same(%orig_input: tensor<2x13x25x7xf32>, %orig_output: tensor<2x4x7x7xf32>, %grad: tensor<2x4x7x7xf32>) -> tensor<2x13x25x7xf32> {
// CHECK: padding = dense<{{\[\[}}0, 0], [0, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>
@@ -1388,6 +1455,13 @@
return %result : tensor<2x13x25x7xf32>
}
+// CHECK-LABEL: @max_pool_3d_grad_same
+func @max_pool_3d_grad_same(%orig_input: tensor<2x8x13x25x7xf32>, %orig_output: tensor<2x8x4x7x7xf32>, %grad: tensor<2x8x4x7x7xf32>) -> tensor<2x8x13x25x7xf32> {
+ // CHECK: padding = dense<{{\[\[}}0, 0], [0, 0], [0, 1], [1, 1], [0, 0]]> : tensor<5x2xi64>
+ %result = "tf.MaxPool3DGrad"(%orig_input, %orig_output, %grad) {data_format = "NDHWC", ksize = [1, 1, 2, 3, 1], padding = "SAME", strides = [1, 1, 4, 4, 1]} : (tensor<2x8x13x25x7xf32>, tensor<2x8x4x7x7xf32>, tensor<2x8x4x7x7xf32>) -> tensor<2x8x13x25x7xf32>
+ return %result : tensor<2x8x13x25x7xf32>
+}
+
//===----------------------------------------------------------------------===//
// OneHot op legalizations.
//===----------------------------------------------------------------------===//
@@ -1400,7 +1474,7 @@
// CHECK: %[[OFF_VALUE:.*]] = "xla_hlo.broadcast"(%arg2) {broadcast_sizes = dense<[3, 5]> : tensor<2xi64>} : (tensor<f32>) -> tensor<3x5xf32>
// CHECK: %[[RESULT:.*]] = "xla_hlo.select"(%[[COMPARE]], %[[ON_VALUE]], %[[OFF_VALUE]]) : (tensor<3x5xi1>, tensor<3x5xf32>, tensor<3x5xf32>) -> tensor<3x5xf32>
// CHECK: return %[[RESULT]] : tensor<3x5xf32>
- %depth = "tf.Const"() { value = dense<5> : tensor<i64> } : () -> tensor<i32>
+ %depth = "tf.Const"() { value = dense<5> : tensor<i32> } : () -> tensor<i32>
%result = "tf.OneHot"(%indices, %depth, %on_value, %off_value) {axis = -1 : i64} : (tensor<3xi32>, tensor<i32>, tensor<f32>, tensor<f32>) -> tensor<3x5xf32>
return %result : tensor<3x5xf32>
}
@@ -2096,6 +2170,12 @@
return %0 : tensor<*xf32>
}
+// CHECK-LABEL: func @round
+func @round(%arg0: tensor<2xf32>) -> tensor<2xf32> {
+ // CHECK: "xla_hlo.round_nearest_afz"(%arg0) : (tensor<2xf32>) -> tensor<2xf32>
+ %0 = "tf.Round"(%arg0) : (tensor<2xf32>) -> tensor<2xf32>
+ return %0 : tensor<2xf32>
+}
// CHECK-LABEL: func @rsqrt
func @rsqrt(%arg0: tensor<2xf32>) -> tensor<2xf32> {
@@ -2434,8 +2514,8 @@
// Begin: 1, 4, -3
// End: 8, 65, 42
// Stride: 1, 4, -1
- // Begin mask: 1, 0, 0 (= 1)
- // End mask: 0, 0, 1 (= 4)
+ // Begin mask: 0, 0, 1 (= 1)
+ // End mask: 1, 0, 0 (= 4)
// So result shape:
// Dim #0: begin mask (1) -> begin = 0; end 8 canonicalized to 4: so 4
@@ -2582,6 +2662,142 @@
return %0 : tensor<2x16x2xf32>
}
+// CHECK-LABEL: strided_slice_nonconstant_begin_end
+func @strided_slice_nonconstant_begin_end(%arg0: tensor<i32>, %arg1: tensor<32x1x97xi32>) -> (tensor<1x97xi32>) {
+ // In this case, the `begin` and `end` inputs are unknown at compile time --
+ // so the StridedSlice needs to slice these vectors and use that as input to
+ // an HLO dynamic slice.
+ %begin = "tf.Pack"(%arg0) {N = 1 : i64, T = i32, axis = 0 : i64, device = ""} : (tensor<i32>) -> tensor<1xi32>
+ %0 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
+ %1 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ %2 = "tf.AddV2"(%arg0, %0) {T = i32, device = ""} : (tensor<i32>, tensor<i32>) -> tensor<i32>
+ %end = "tf.Pack"(%2) {N = 1 : i64, T = i32, axis = 0 : i64, device = ""} : (tensor<i32>) -> tensor<1xi32>
+ // CHECK: %[[A:.*]] = "xla_hlo.reshape"(%arg0) : (tensor<i32>) -> tensor<1xi32>
+ // CHECK-NEXT: %[[BEGIN:.*]] = "xla_hlo.concatenate"(%[[A]])
+ // CHECK-DAG-SAME: {dimension = 0 : i64} : (tensor<1xi32>) -> tensor<1xi32>
+ // CHECK: %[[ZERO:.*]] = xla_hlo.constant dense<0> : tensor<i32>
+ // CHECK-NEXT: %[[INDEX:.*]] = "xla_hlo.slice"(%[[BEGIN]])
+ // CHECK-DAG-SAME: {limit_indices = dense<1> : tensor<1xi64>,
+ // CHECK-DAG-SAME: start_indices = dense<0> : tensor<1xi64>,
+ // CHECK-DAG-SAME: strides = dense<1> : tensor<1xi64>} : (tensor<1xi32>) -> tensor<1xi32>
+ // CHECK-NEXT: %[[INDEX2:.*]] = "xla_hlo.reshape"(%[[INDEX]]) : (tensor<1xi32>) -> tensor<i32>
+ // CHECK-NEXT: %[[CMP:.*]] = "xla_hlo.compare"(%[[INDEX2]], %[[ZERO]])
+ // CHECK-DAG-SAME: {comparison_direction = "LT"} : (tensor<i32>, tensor<i32>) -> tensor<i1>
+ // CHECK-NEXT: %[[DIM:.*]] = xla_hlo.constant dense<32> : tensor<i32>
+ // CHECK-NEXT: %[[WRAP:.*]] = xla_hlo.add %[[DIM]], %[[INDEX2]] : tensor<i32>
+ // CHECK-NEXT: %[[INDEX3:.*]] = "xla_hlo.select"(%[[CMP]], %[[WRAP]], %[[INDEX2]]) :
+ // CHECK-DAG-SAME: (tensor<i1>, tensor<i32>, tensor<i32>) -> tensor<i32>
+ // CHECK-NEXT: %[[SLICED:.*]] = "xla_hlo.dynamic-slice"
+ // CHECK-DAG-SAME: (%arg1, %[[INDEX3]], %[[ZERO]], %[[ZERO]])
+ // CHECK-DAG-SAME: {slice_sizes = dense<[1, 1, 97]> : tensor<3xi64>} :
+ // CHECK-DAG-SAME: (tensor<32x1x97xi32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<1x97xi32>
+ // CHECK-NEXT: %[[FINAL:.*]] = "xla_hlo.reshape"(%[[SLICED]]) : (tensor<1x97xi32>) -> tensor<1x97xi32>
+ %result = "tf.StridedSlice"(%arg1, %begin, %end, %1) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ // CHECK-NEXT: return %[[FINAL]] : tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_stride_1
+func @strided_slice_nonconstant_begin_end_stride_1(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>, %strides: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // Dynamic stride: when `begin` and `end` inputs are unknown at compile time,
+ // `strides` must be known.
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 4 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_stride_2
+func @strided_slice_nonconstant_begin_end_stride_2(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // Invalid stride (not equal to 1): when `begin` and `end` inputs are unknown
+ // at compile time, `strides` must be known to have all 1 values.
+ %strides = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 4 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_invalid_elem_count
+func @strided_slice_nonconstant_begin_end_invalid_elem_count(%input: tensor<4x8xf32>, %begin: tensor<2xi64>, %end: tensor<2xi64>) -> tensor<6x10xf32> {
+ %strides = "tf.Const"() { value = dense<[1, 1]> : tensor<2xi64> } : () -> tensor<2xi64>
+ // When begin/end are dynamic, the number of output elements must be equal to
+ // the number of input elements sliced.
+ // CHECK: tf.StridedSlice
+ %0 = "tf.StridedSlice"(%input, %begin, %end, %strides) : (tensor<4x8xf32>, tensor<2xi64>, tensor<2xi64>, tensor<2xi64>) -> tensor<6x10xf32>
+ return %0 : tensor<6x10xf32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_begin_mask
+func @strided_slice_nonconstant_begin_end_and_begin_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // Begin mask: When `begin` and `end` inputs are unknown at compile time, we
+ // can't support a begin mask.
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 4 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_end_mask
+func @strided_slice_nonconstant_begin_end_and_end_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // End mask: When `begin` and `end` inputs are unknown at compile time, we
+ // can't support an end mask.
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 1 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_new_axis_mask
+func @strided_slice_nonconstant_begin_end_and_new_axis_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // New axis mask: When `begin` and `end` inputs are unknown at compile time,
+ // we can't support a new_axis mask.
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 15 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_ellipsis_mask
+func @strided_slice_nonconstant_begin_end_and_ellipsis_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // This ellipsis mask is not supported because it does not refer to the last
+ // dimension.
+ // [0, 1, 0] = 2
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 2 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_valid_ellipsis_mask
+func @strided_slice_nonconstant_begin_end_and_valid_ellipsis_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // This ellipsis mask is supported because it refers to the last dimension.
+ // [1, 0, 0] = 4
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: xla_hlo.dynamic-slice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 4 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 1 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_valid_shrink_axis_mask
+func @strided_slice_nonconstant_begin_end_and_valid_shrink_axis_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // This shrink_axis mask is supported because it refers to a major dimension.
+ // [1, 1, 1] = 7
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: xla_hlo.dynamic-slice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 7 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
+// CHECK-LABEL: strided_slice_nonconstant_begin_end_and_invalid_shrink_axis_mask
+func @strided_slice_nonconstant_begin_end_and_invalid_shrink_axis_mask(%input: tensor<32x1x97xi32>, %begin: tensor<1xi32>, %end: tensor<1xi32>) -> (tensor<1x97xi32>) {
+ // This shrink_axis mask is unsupported because it does not refer to a major
+ // dimension.
+ // [0, 1, 0] = 2
+ %strides = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
+ // CHECK: tf.StridedSlice
+ %result = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = i32, begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 0 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 2 : i64} : (tensor<32x1x97xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x97xi32>
+ return %result : tensor<1x97xi32>
+}
+
//===----------------------------------------------------------------------===//
// Reduction op legalizations.
@@ -3144,6 +3360,35 @@
return %result : tensor<100x28x28x1xf32>
}
+// CHECK-LABEL: @conv3d_backprop_filter
+func @conv3d_backprop_filter(%input: tensor<2x8x8x8x1xf32>, %out_backprop: tensor<2x8x8x8x6xf32>) -> tensor<2x8x8x8x1xf32> {
+ // CHECK: %[[RESULT:.*]] = "xla_hlo.convolution"(%arg0, %arg1)
+
+ // CHECK-DAG-SAME: batch_group_count = 1 : i64
+
+ // CHECK-DAG-SAME: dimension_numbers =
+ // CHECK-DAG-SAME: input_batch_dimension = 4 : i64
+ // CHECK-DAG-SAME: input_feature_dimension = 0 : i64
+ // CHECK-DAG-SAME: input_spatial_dimensions = dense<[1, 2, 3]> : tensor<3xi64>
+ // CHECK-DAG-SAME: kernel_input_feature_dimension = 0 : i64
+ // CHECK-DAG-SAME: kernel_output_feature_dimension = 4 : i64
+ // CHECK-DAG-SAME: kernel_spatial_dimensions = dense<[1, 2, 3]> : tensor<3xi64>
+ // CHECK-DAG-SAME: output_batch_dimension = 3 : i64
+ // CHECK-DAG-SAME: output_feature_dimension = 4 : i64
+ // CHECK-DAG-SAME: output_spatial_dimensions = dense<[0, 1, 2]> : tensor<3xi64>
+
+ // CHECK-DAG-SAME: feature_group_count = 1 : i64
+ // CHECK-DAG-SAME: lhs_dilation = dense<1> : tensor<3xi64>
+ // CHECK-DAG-SAME: padding = dense<1> : tensor<3x2xi64>
+ // CHECK-DAG-SAME: rhs_dilation = dense<1> : tensor<3xi64>
+ // CHECK-DAG-SAME: window_strides = dense<1> : tensor<3xi64>
+
+ // CHECK: return %[[RESULT]]
+ %filter_sizes = "tf.Const"() {value = dense<[3, 3, 3, 1, 6]> : tensor<5xi32>} : () -> tensor<5xi32>
+ %result = "tf.Conv3DBackpropFilterV2"(%input, %filter_sizes, %out_backprop) {data_format = "NDHWC", dilations = [1, 1, 1, 1, 1], padding = "SAME", strides = [1, 1, 1, 1, 1]} : (tensor<2x8x8x8x1xf32>, tensor<5xi32>, tensor<2x8x8x8x6xf32>) -> tensor<2x8x8x8x1xf32>
+ return %result : tensor<2x8x8x8x1xf32>
+}
+
// CHECK-LABEL: @cross_replica_sum
func @cross_replica_sum(%input: tensor<10xf32>) -> tensor<10xf32> {
%replica_groups = "tf.Const" () {
@@ -3954,3 +4199,180 @@
%0 = "tf.BatchMatMulV2"(%arg0, %arg1) {adj_x = true, adj_y = true, device = ""} : (tensor<5x2xcomplex<f32>>, tensor<2x4xcomplex<f32>>) -> tensor<5x4xcomplex<f32>>
return %0 : tensor<5x4xcomplex<f32>>
}
+
+// CHECK: func @qr([[VAL_0:%.*]]: tensor<500x100x75xf32>) -> (tensor<500x100x75xf32>, tensor<500x75x75xf32>)
+func @qr(%arg0: tensor<500x100x75xf32>) -> (tensor<500x100x75xf32>, tensor<500x75x75xf32>) {
+// CHECK: [[VAL_1:%.*]] = "xla_hlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<100x100xi32>
+// CHECK: [[VAL_2:%.*]] = "xla_hlo.iota"() {iota_dimension = 1 : i64} : () -> tensor<100x100xi32>
+// CHECK: [[VAL_3:%.*]] = "xla_hlo.compare"([[VAL_1]], [[VAL_2]]) {comparison_direction = "EQ"} : (tensor<100x100xi32>, tensor<100x100xi32>) -> tensor<100x100xi1>
+// CHECK: [[VAL_4:%.*]] = "xla_hlo.convert"([[VAL_3]]) : (tensor<100x100xi1>) -> tensor<100x100xf32>
+// CHECK: [[VAL_5:%.*]] = "xla_hlo.broadcast"([[VAL_4]]) {broadcast_sizes = dense<500> : tensor<1xi64>} : (tensor<100x100xf32>) -> tensor<500x100x100xf32>
+// CHECK: [[VAL_6:%.*]] = "xla_hlo.slice"([[VAL_0]]) {limit_indices = dense<[500, 100, 75]> : tensor<3xi64>, start_indices = dense<0> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_7:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_8:%.*]] = "xla_hlo.broadcast"([[VAL_7]]) {broadcast_sizes = dense<[500, 100, 75]> : tensor<3xi64>} : (tensor<f32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_9:%.*]] = "xla_hlo.broadcast"([[VAL_7]]) {broadcast_sizes = dense<[500, 75]> : tensor<2xi64>} : (tensor<f32>) -> tensor<500x75xf32>
+// CHECK: [[VAL_10:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_11:%.*]] = "xla_hlo.tuple"([[VAL_10]], [[VAL_6]], [[VAL_8]], [[VAL_9]]) : (tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: [[VAL_12:%.*]] = "xla_hlo.while"([[VAL_11]]) ( {
+// CHECK: ^bb0([[VAL_13:%.*]]: tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>):
+// CHECK: [[VAL_14:%.*]] = "xla_hlo.get_tuple_element"([[VAL_13]]) {index = 0 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<i32>
+// CHECK: [[VAL_15:%.*]] = xla_hlo.constant dense<75> : tensor<i32>
+// CHECK: [[VAL_16:%.*]] = "xla_hlo.compare"([[VAL_14]], [[VAL_15]]) {comparison_direction = "LT"} : (tensor<i32>, tensor<i32>) -> tensor<i1>
+// CHECK: "xla_hlo.return"([[VAL_16]]) : (tensor<i1>) -> ()
+// CHECK: }, {
+// CHECK: ^bb0([[VAL_17:%.*]]: tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>):
+// CHECK: [[VAL_18:%.*]] = "xla_hlo.get_tuple_element"([[VAL_17]]) {index = 0 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<i32>
+// CHECK: [[VAL_19:%.*]] = "xla_hlo.get_tuple_element"([[VAL_17]]) {index = 1 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_20:%.*]] = "xla_hlo.get_tuple_element"([[VAL_17]]) {index = 2 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_21:%.*]] = "xla_hlo.get_tuple_element"([[VAL_17]]) {index = 3 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x75xf32>
+// CHECK: [[VAL_22:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_23:%.*]] = "xla_hlo.dynamic-slice"([[VAL_19]], [[VAL_22]], [[VAL_22]], [[VAL_18]]) {slice_sizes = dense<[500, 100, 1]> : tensor<3xi64>} : (tensor<500x100x75xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_24:%.*]] = "xla_hlo.reshape"([[VAL_23]]) : (tensor<500x100x1xf32>) -> tensor<500x100xf32>
+// CHECK: [[VAL_25:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_26:%.*]] = xla_hlo.constant dense<1.000000e+00> : tensor<f32>
+// CHECK: [[VAL_27:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_28:%.*]] = "xla_hlo.dynamic-slice"([[VAL_24]], [[VAL_27]], [[VAL_18]]) {slice_sizes = dense<[500, 1]> : tensor<2xi64>} : (tensor<500x100xf32>, tensor<i32>, tensor<i32>) -> tensor<500x1xf32>
+// CHECK: [[VAL_29:%.*]] = "xla_hlo.reshape"([[VAL_28]]) : (tensor<500x1xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_30:%.*]] = "xla_hlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<100xi32>
+// CHECK: [[VAL_31:%.*]] = "xla_hlo.compare"([[VAL_30]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "GT"} : (tensor<100xi32>, tensor<i32>) -> tensor<100xi1>
+// CHECK: [[VAL_32:%.*]] = "xla_hlo.convert"([[VAL_31]]) : (tensor<100xi1>) -> tensor<100xf32>
+// CHECK: [[VAL_33:%.*]] = "xla_hlo.multiply"([[VAL_24]], [[VAL_32]]) {broadcast_dimensions = dense<1> : tensor<1xi64>} : (tensor<500x100xf32>, tensor<100xf32>) -> tensor<500x100xf32>
+// CHECK: [[VAL_34:%.*]] = xla_hlo.multiply [[VAL_33]], [[VAL_33]] : tensor<500x100xf32>
+// CHECK: [[VAL_35:%.*]] = "xla_hlo.reduce"([[VAL_34]], [[VAL_25]]) ( {
+// CHECK: ^bb0([[VAL_36:%.*]]: tensor<f32>, [[VAL_37:%.*]]: tensor<f32>):
+// CHECK: [[VAL_38:%.*]] = xla_hlo.add [[VAL_36]], [[VAL_37]] : tensor<f32>
+// CHECK: "xla_hlo.return"([[VAL_38]]) : (tensor<f32>) -> ()
+// CHECK: }) {dimensions = dense<1> : tensor<1xi64>} : (tensor<500x100xf32>, tensor<f32>) -> tensor<500xf32>
+// CHECK: [[VAL_39:%.*]] = xla_hlo.multiply [[VAL_29]], [[VAL_29]] : tensor<500xf32>
+// CHECK: [[VAL_40:%.*]] = xla_hlo.add [[VAL_39]], [[VAL_41:%.*]] : tensor<500xf32>
+// CHECK: [[VAL_42:%.*]] = "xla_hlo.sqrt"([[VAL_40]]) : (tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_43:%.*]] = "xla_hlo.compare"([[VAL_41]], [[VAL_25]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "EQ"} : (tensor<500xf32>, tensor<f32>) -> tensor<500xi1>
+// CHECK: [[VAL_44:%.*]] = "xla_hlo.compare"([[VAL_29]], [[VAL_25]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "LT"} : (tensor<500xf32>, tensor<f32>) -> tensor<500xi1>
+// CHECK: [[VAL_45:%.*]] = "xla_hlo.broadcast"([[VAL_26]]) {broadcast_sizes = dense<500> : tensor<1xi64>} : (tensor<f32>) -> tensor<500xf32>
+// CHECK: [[VAL_46:%.*]] = "xla_hlo.negate"([[VAL_45]]) : (tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_47:%.*]] = "xla_hlo.select"([[VAL_44]], [[VAL_45]], [[VAL_46]]) : (tensor<500xi1>, tensor<500xf32>, tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_48:%.*]] = xla_hlo.multiply [[VAL_47]], [[VAL_42]] {broadcast_dimensions = dense<[]> : tensor<0xi64>} : tensor<500xf32>
+// CHECK: [[VAL_49:%.*]] = "xla_hlo.select"([[VAL_43]], [[VAL_29]], [[VAL_48]]) : (tensor<500xi1>, tensor<500xf32>, tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_50:%.*]] = xla_hlo.subtract [[VAL_49]], [[VAL_29]] : tensor<500xf32>
+// CHECK: [[VAL_51:%.*]] = xla_hlo.divide [[VAL_50]], [[VAL_49]] : tensor<500xf32>
+// CHECK: [[VAL_52:%.*]] = "xla_hlo.broadcast"([[VAL_25]]) {broadcast_sizes = dense<500> : tensor<1xi64>} : (tensor<f32>) -> tensor<500xf32>
+// CHECK: [[VAL_53:%.*]] = "xla_hlo.select"([[VAL_43]], [[VAL_52]], [[VAL_51]]) : (tensor<500xi1>, tensor<500xf32>, tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_54:%.*]] = xla_hlo.subtract [[VAL_29]], [[VAL_49]] : tensor<500xf32>
+// CHECK: [[VAL_55:%.*]] = "xla_hlo.select"([[VAL_43]], [[VAL_45]], [[VAL_54]]) : (tensor<500xi1>, tensor<500xf32>, tensor<500xf32>) -> tensor<500xf32>
+// CHECK: [[VAL_56:%.*]] = "xla_hlo.compare"([[VAL_30]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "EQ"} : (tensor<100xi32>, tensor<i32>) -> tensor<100xi1>
+// CHECK: [[VAL_57:%.*]] = "xla_hlo.convert"([[VAL_56]]) : (tensor<100xi1>) -> tensor<100xf32>
+// CHECK: [[VAL_58:%.*]] = "xla_hlo.broadcast"([[VAL_57]]) {broadcast_sizes = dense<1> : tensor<1xi64>} : (tensor<100xf32>) -> tensor<1x100xf32>
+// CHECK: [[VAL_59:%.*]] = "xla_hlo.divide"([[VAL_33]], [[VAL_55]]) {broadcast_dimensions = dense<0> : tensor<1xi64>} : (tensor<500x100xf32>, tensor<500xf32>) -> tensor<500x100xf32>
+// CHECK: [[VAL_60:%.*]] = "xla_hlo.add"([[VAL_58]], [[VAL_59]]) : (tensor<1x100xf32>, tensor<500x100xf32>) -> tensor<500x100xf32>
+// CHECK: [[VAL_61:%.*]] = "xla_hlo.reshape"([[VAL_60]]) : (tensor<500x100xf32>) -> tensor<500x1x100xf32>
+// CHECK: [[VAL_62:%.*]] = "xla_hlo.dot_general"([[VAL_61]], [[VAL_19]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<2> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x1x100xf32>, tensor<500x100x75xf32>) -> tensor<500x1x75xf32>
+// CHECK: [[VAL_63:%.*]] = "xla_hlo.dot_general"([[VAL_61]], [[VAL_62]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<1> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x1x100xf32>, tensor<500x1x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_64:%.*]] = "xla_hlo.multiply"([[VAL_53]], [[VAL_63]]) {broadcast_dimensions = dense<0> : tensor<1xi64>} : (tensor<500xf32>, tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_65:%.*]] = xla_hlo.subtract [[VAL_19]], [[VAL_64]] : tensor<500x100x75xf32>
+// CHECK: [[VAL_66:%.*]] = "xla_hlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<100x1xi32>
+// CHECK: [[VAL_67:%.*]] = "xla_hlo.compare"([[VAL_66]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "LT"} : (tensor<100x1xi32>, tensor<i32>) -> tensor<100x1xi1>
+// CHECK: [[VAL_68:%.*]] = "xla_hlo.convert"([[VAL_67]]) : (tensor<100x1xi1>) -> tensor<100x1xf32>
+// CHECK: [[VAL_69:%.*]] = "xla_hlo.compare"([[VAL_66]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "EQ"} : (tensor<100x1xi32>, tensor<i32>) -> tensor<100x1xi1>
+// CHECK: [[VAL_70:%.*]] = "xla_hlo.convert"([[VAL_69]]) : (tensor<100x1xi1>) -> tensor<100x1xf32>
+// CHECK: [[VAL_71:%.*]] = "xla_hlo.broadcast"([[VAL_70]]) {broadcast_sizes = dense<1> : tensor<1xi64>} : (tensor<100x1xf32>) -> tensor<1x100x1xf32>
+// CHECK: [[VAL_72:%.*]] = "xla_hlo.multiply"([[VAL_23]], [[VAL_68]]) {broadcast_dimensions = dense<[1, 2]> : tensor<2xi64>} : (tensor<500x100x1xf32>, tensor<100x1xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_73:%.*]] = "xla_hlo.multiply"([[VAL_49]], [[VAL_71]]) {broadcast_dimensions = dense<0> : tensor<1xi64>} : (tensor<500xf32>, tensor<1x100x1xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_74:%.*]] = xla_hlo.add [[VAL_72]], [[VAL_73]] : tensor<500x100x1xf32>
+// CHECK: [[VAL_75:%.*]] = "xla_hlo.broadcast_in_dim"([[VAL_74]]) {broadcast_dimensions = dense<[0, 1, 2]> : tensor<3xi64>} : (tensor<500x100x1xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_76:%.*]] = "xla_hlo.iota"() {iota_dimension = 2 : i64} : () -> tensor<500x100x75xi32>
+// CHECK: [[VAL_77:%.*]] = "xla_hlo.compare"([[VAL_76]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "EQ"} : (tensor<500x100x75xi32>, tensor<i32>) -> tensor<500x100x75xi1>
+// CHECK: [[VAL_78:%.*]] = "xla_hlo.select"([[VAL_77]], [[VAL_75]], [[VAL_65]]) : (tensor<500x100x75xi1>, tensor<500x100x75xf32>, tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_79:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_80:%.*]] = "xla_hlo.broadcast"([[VAL_79]]) {broadcast_sizes = dense<[500, 100, 75]> : tensor<3xi64>} : (tensor<f32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_81:%.*]] = "xla_hlo.add"([[VAL_80]], [[VAL_60]]) {broadcast_dimensions = dense<[0, 1]> : tensor<2xi64>} : (tensor<500x100x75xf32>, tensor<500x100xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_82:%.*]] = "xla_hlo.select"([[VAL_77]], [[VAL_81]], [[VAL_80]]) : (tensor<500x100x75xi1>, tensor<500x100x75xf32>, tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_83:%.*]] = xla_hlo.add [[VAL_20]], [[VAL_82]] : tensor<500x100x75xf32>
+// CHECK: [[VAL_84:%.*]] = "xla_hlo.iota"() {iota_dimension = 1 : i64} : () -> tensor<500x75xi32>
+// CHECK: [[VAL_85:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_86:%.*]] = "xla_hlo.broadcast"([[VAL_85]]) {broadcast_sizes = dense<[500, 75]> : tensor<2xi64>} : (tensor<f32>) -> tensor<500x75xf32>
+// CHECK: [[VAL_87:%.*]] = "xla_hlo.compare"([[VAL_84]], [[VAL_18]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "EQ"} : (tensor<500x75xi32>, tensor<i32>) -> tensor<500x75xi1>
+// CHECK: [[VAL_88:%.*]] = "xla_hlo.add"([[VAL_86]], [[VAL_53]]) {broadcast_dimensions = dense<0> : tensor<1xi64>} : (tensor<500x75xf32>, tensor<500xf32>) -> tensor<500x75xf32>
+// CHECK: [[VAL_89:%.*]] = "xla_hlo.select"([[VAL_87]], [[VAL_88]], [[VAL_86]]) : (tensor<500x75xi1>, tensor<500x75xf32>, tensor<500x75xf32>) -> tensor<500x75xf32>
+// CHECK: [[VAL_90:%.*]] = xla_hlo.add [[VAL_21]], [[VAL_89]] : tensor<500x75xf32>
+// CHECK: [[VAL_91:%.*]] = xla_hlo.constant dense<1> : tensor<i32>
+// CHECK: [[VAL_92:%.*]] = xla_hlo.add [[VAL_18]], [[VAL_91]] {broadcast_dimensions = dense<[]> : tensor<0xi64>} : tensor<i32>
+// CHECK: [[VAL_93:%.*]] = "xla_hlo.tuple"([[VAL_92]], [[VAL_78]], [[VAL_83]], [[VAL_90]]) : (tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: "xla_hlo.return"([[VAL_93]]) : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> ()
+// CHECK: }) : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: [[VAL_94:%.*]] = "xla_hlo.get_tuple_element"([[VAL_95:%.*]]) {index = 1 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_96:%.*]] = "xla_hlo.get_tuple_element"([[VAL_95]]) {index = 2 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_97:%.*]] = "xla_hlo.get_tuple_element"([[VAL_95]]) {index = 3 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x75xf32>
+// CHECK: [[VAL_98:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_99:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_100:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_101:%.*]] = "xla_hlo.dynamic-update-slice"([[VAL_0]], [[VAL_94]], [[VAL_100]], [[VAL_98]], [[VAL_99]]) : (tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_102:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_103:%.*]] = "xla_hlo.broadcast"([[VAL_102]]) {broadcast_sizes = dense<[500, 100, 75]> : tensor<3xi64>} : (tensor<f32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_104:%.*]] = "xla_hlo.slice"([[VAL_96]]) {limit_indices = dense<[500, 100, 1]> : tensor<3xi64>, start_indices = dense<0> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x75xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_105:%.*]] = "xla_hlo.slice"([[VAL_97]]) {limit_indices = dense<[500, 1]> : tensor<2xi64>, start_indices = dense<0> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} : (tensor<500x75xf32>) -> tensor<500x1xf32>
+// CHECK: [[VAL_106:%.*]] = "xla_hlo.negate"([[VAL_105]]) : (tensor<500x1xf32>) -> tensor<500x1xf32>
+// CHECK: [[VAL_107:%.*]] = "xla_hlo.multiply"([[VAL_106]], [[VAL_104]]) {broadcast_dimensions = dense<[0, 2]> : tensor<2xi64>} : (tensor<500x1xf32>, tensor<500x100x1xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_108:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_109:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_110:%.*]] = "xla_hlo.dynamic-update-slice"([[VAL_103]], [[VAL_107]], [[VAL_109]], [[VAL_109]], [[VAL_108]]) : (tensor<500x100x75xf32>, tensor<500x100x1xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_111:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_112:%.*]] = "xla_hlo.tuple"([[VAL_111]], [[VAL_110]], [[VAL_96]], [[VAL_97]]) : (tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: [[VAL_113:%.*]] = "xla_hlo.while"([[VAL_112]]) ( {
+// CHECK: ^bb0([[VAL_114:%.*]]: tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>):
+// CHECK: [[VAL_115:%.*]] = "xla_hlo.get_tuple_element"([[VAL_114]]) {index = 0 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<i32>
+// CHECK: [[VAL_116:%.*]] = xla_hlo.constant dense<74> : tensor<i32>
+// CHECK: [[VAL_117:%.*]] = "xla_hlo.compare"([[VAL_115]], [[VAL_116]]) {comparison_direction = "LT"} : (tensor<i32>, tensor<i32>) -> tensor<i1>
+// CHECK: "xla_hlo.return"([[VAL_117]]) : (tensor<i1>) -> ()
+// CHECK: }, {
+// CHECK: ^bb0([[VAL_118:%.*]]: tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>):
+// CHECK: [[VAL_119:%.*]] = "xla_hlo.get_tuple_element"([[VAL_118]]) {index = 0 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<i32>
+// CHECK: [[VAL_120:%.*]] = "xla_hlo.get_tuple_element"([[VAL_118]]) {index = 1 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_121:%.*]] = "xla_hlo.get_tuple_element"([[VAL_118]]) {index = 2 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_122:%.*]] = "xla_hlo.get_tuple_element"([[VAL_118]]) {index = 3 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x75xf32>
+// CHECK: [[VAL_123:%.*]] = xla_hlo.constant dense<1> : tensor<i32>
+// CHECK: [[VAL_124:%.*]] = xla_hlo.add [[VAL_119]], [[VAL_123]] : tensor<i32>
+// CHECK: [[VAL_125:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_126:%.*]] = "xla_hlo.dynamic-slice"([[VAL_121]], [[VAL_125]], [[VAL_125]], [[VAL_124]]) {slice_sizes = dense<[500, 100, 1]> : tensor<3xi64>} : (tensor<500x100x75xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_127:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_128:%.*]] = "xla_hlo.dynamic-slice"([[VAL_122]], [[VAL_127]], [[VAL_124]]) {slice_sizes = dense<[500, 1]> : tensor<2xi64>} : (tensor<500x75xf32>, tensor<i32>, tensor<i32>) -> tensor<500x1xf32>
+// CHECK: [[VAL_129:%.*]] = "xla_hlo.iota"() {iota_dimension = 2 : i64} : () -> tensor<500x100x75xi32>
+// CHECK: [[VAL_130:%.*]] = xla_hlo.constant dense<0.000000e+00> : tensor<f32>
+// CHECK: [[VAL_131:%.*]] = "xla_hlo.broadcast"([[VAL_130]]) {broadcast_sizes = dense<[500, 100, 75]> : tensor<3xi64>} : (tensor<f32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_132:%.*]] = "xla_hlo.compare"([[VAL_129]], [[VAL_124]]) {broadcast_dimensions = dense<[]> : tensor<0xi64>, comparison_direction = "GE"} : (tensor<500x100x75xi32>, tensor<i32>) -> tensor<500x100x75xi1>
+// CHECK: [[VAL_133:%.*]] = "xla_hlo.select"([[VAL_132]], [[VAL_131]], [[VAL_121]]) : (tensor<500x100x75xi1>, tensor<500x100x75xf32>, tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_134:%.*]] = "xla_hlo.dot_general"([[VAL_133]], [[VAL_126]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<1> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x75xf32>, tensor<500x100x1xf32>) -> tensor<500x75x1xf32>
+// CHECK: [[VAL_135:%.*]] = "xla_hlo.dot_general"([[VAL_120]], [[VAL_134]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<2> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x75xf32>, tensor<500x75x1xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_136:%.*]] = "xla_hlo.negate"([[VAL_128]]) : (tensor<500x1xf32>) -> tensor<500x1xf32>
+// CHECK: [[VAL_137:%.*]] = xla_hlo.add [[VAL_126]], [[VAL_135]] : tensor<500x100x1xf32>
+// CHECK: [[VAL_138:%.*]] = "xla_hlo.multiply"([[VAL_136]], [[VAL_137]]) {broadcast_dimensions = dense<[0, 2]> : tensor<2xi64>} : (tensor<500x1xf32>, tensor<500x100x1xf32>) -> tensor<500x100x1xf32>
+// CHECK: [[VAL_139:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_140:%.*]] = "xla_hlo.dynamic-update-slice"([[VAL_120]], [[VAL_138]], [[VAL_139]], [[VAL_139]], [[VAL_124]]) : (tensor<500x100x75xf32>, tensor<500x100x1xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_141:%.*]] = xla_hlo.constant dense<1> : tensor<i32>
+// CHECK: [[VAL_142:%.*]] = xla_hlo.add [[VAL_119]], [[VAL_141]] {broadcast_dimensions = dense<[]> : tensor<0xi64>} : tensor<i32>
+// CHECK: [[VAL_143:%.*]] = "xla_hlo.tuple"([[VAL_142]], [[VAL_140]], [[VAL_121]], [[VAL_122]]) : (tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: "xla_hlo.return"([[VAL_143]]) : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> ()
+// CHECK: }) : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>
+// CHECK: [[VAL_144:%.*]] = "xla_hlo.get_tuple_element"([[VAL_145:%.*]]) {index = 1 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_146:%.*]] = "xla_hlo.get_tuple_element"([[VAL_145]]) {index = 2 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_147:%.*]] = "xla_hlo.get_tuple_element"([[VAL_145]]) {index = 3 : i32} : (tuple<tensor<i32>, tensor<500x100x75xf32>, tensor<500x100x75xf32>, tensor<500x75xf32>>) -> tensor<500x75xf32>
+// CHECK: [[VAL_148:%.*]] = "xla_hlo.slice"([[VAL_101]]) {limit_indices = dense<[500, 100, 75]> : tensor<3xi64>, start_indices = dense<[0, 0, 75]> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x75xf32>) -> tensor<500x100x0xf32>
+// CHECK: [[VAL_149:%.*]] = "xla_hlo.dot_general"([[VAL_144]], [[VAL_148]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<1> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x75xf32>, tensor<500x100x0xf32>) -> tensor<500x75x0xf32>
+// CHECK: [[VAL_150:%.*]] = "xla_hlo.dot_general"([[VAL_96]], [[VAL_149]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<2> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x75xf32>, tensor<500x75x0xf32>) -> tensor<500x100x0xf32>
+// CHECK: [[VAL_151:%.*]] = xla_hlo.add [[VAL_148]], [[VAL_150]] : tensor<500x100x0xf32>
+// CHECK: [[VAL_152:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_153:%.*]] = xla_hlo.constant dense<75> : tensor<i32>
+// CHECK: [[VAL_154:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_155:%.*]] = "xla_hlo.dynamic-update-slice"([[VAL_101]], [[VAL_151]], [[VAL_154]], [[VAL_152]], [[VAL_153]]) : (tensor<500x100x75xf32>, tensor<500x100x0xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_156:%.*]] = "xla_hlo.slice"([[VAL_5]]) {limit_indices = dense<[500, 100, 100]> : tensor<3xi64>, start_indices = dense<0> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x100xf32>) -> tensor<500x100x100xf32>
+// CHECK: [[VAL_157:%.*]] = "xla_hlo.dot_general"([[VAL_156]], [[VAL_144]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<2> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<1> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x100xf32>, tensor<500x100x75xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_158:%.*]] = "xla_hlo.dot_general"([[VAL_157]], [[VAL_96]]) {dot_dimension_numbers = {lhs_batching_dimensions = dense<0> : tensor<1xi64>, lhs_contracting_dimensions = dense<2> : tensor<1xi64>, rhs_batching_dimensions = dense<0> : tensor<1xi64>, rhs_contracting_dimensions = dense<2> : tensor<1xi64>}, precision_config = ["HIGHEST", "HIGHEST"]} : (tensor<500x100x75xf32>, tensor<500x100x75xf32>) -> tensor<500x100x100xf32>
+// CHECK: [[VAL_159:%.*]] = xla_hlo.add [[VAL_156]], [[VAL_158]] : tensor<500x100x100xf32>
+// CHECK: [[VAL_160:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_161:%.*]] = xla_hlo.constant dense<0> : tensor<i32>
+// CHECK: [[VAL_162:%.*]] = "xla_hlo.dynamic-update-slice"([[VAL_5]], [[VAL_159]], [[VAL_161]], [[VAL_161]], [[VAL_160]]) : (tensor<500x100x100xf32>, tensor<500x100x100xf32>, tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<500x100x100xf32>
+// CHECK: [[VAL_163:%.*]] = "xla_hlo.slice"([[VAL_162]]) {limit_indices = dense<[500, 100, 75]> : tensor<3xi64>, start_indices = dense<0> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x100xf32>) -> tensor<500x100x75xf32>
+// CHECK: [[VAL_164:%.*]] = "xla_hlo.slice"([[VAL_155]]) {limit_indices = dense<[500, 75, 75]> : tensor<3xi64>, start_indices = dense<0> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>} : (tensor<500x100x75xf32>) -> tensor<500x75x75xf32>
+// CHECK: return [[VAL_163]], [[VAL_164]] : tensor<500x100x75xf32>, tensor<500x75x75xf32>
+ %0:2 = "tf.Qr"(%arg0) {full_matrices = false} : (tensor<500x100x75xf32>) -> (tensor<500x100x75xf32>, tensor<500x75x75xf32>)
+ return %0#0, %0#1 : tensor<500x100x75xf32>, tensor<500x75x75xf32>
+}
diff --git a/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-select-and-scatter.mlir b/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-select-and-scatter.mlir
new file mode 100644
index 0000000..5b763cd
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-select-and-scatter.mlir
@@ -0,0 +1,199 @@
+// GenericAtomicRMWOp should contain only ops with no side effects.
+// Unfortunately, the legalization pattern for SelectAndScatterOp has to adapt
+// to XLA LHLO dialect using allocs/deallocs inside of GenericAtomicRMWOp body.
+// Lowering to STD dialect and store forwarding pass would be required to get
+// rid of them. This is exactly what is done in the real MLIR GPU pipeline, but
+// here we disable verification with `verify-each=0` to check the output IR.
+// RUN: xla-opt %s -lhlo-legalize-to-parallel-loops -canonicalize --verify-each=0 | FileCheck %s --dump-input-on-failure
+
+func @select_and_scatter(%arg: memref<112x112xf32>,
+ %src: memref<56x56xf32>,
+ %init: memref<f32>,
+ %result: memref<112x112xf32>) {
+ "xla_lhlo.select_and_scatter"(%arg, %src, %init, %result) ( {
+ // select
+ ^bb0(%lhs: memref<f32>, %rhs: memref<f32>, %pred: memref<i1>):
+ "xla_lhlo.compare"(%lhs, %rhs, %pred) {comparison_direction = "GE"} :
+ (memref<f32>, memref<f32>, memref<i1>) -> ()
+ "xla_lhlo.terminator"() : () -> ()
+ }, {
+ // scatter
+ ^bb0(%lhs: memref<f32>, %rhs: memref<f32>, %out: memref<f32>):
+ "xla_lhlo.add"(%lhs, %rhs, %out) :
+ (memref<f32>, memref<f32>, memref<f32>) -> ()
+ "xla_lhlo.terminator"() : () -> ()
+ }) {
+ padding = dense<[[0, 1], [0, 1]]> : tensor<2x2xi64>,
+ window_dimensions = dense<[3, 3]> : tensor<2xi64>,
+ window_strides = dense<[2, 2]> : tensor<2xi64>
+ } : (memref<112x112xf32>,
+ memref<56x56xf32>,
+ memref<f32>, memref<112x112xf32>) -> ()
+ "xla_lhlo.terminator"() : () -> ()
+}
+// CHECK-LABEL: func @select_and_scatter(
+// CHECK-SAME: [[ARG_BUF:%.*]]: memref<112x112xf32>,
+// CHECK-SAME: [[SRC_BUF:%.*]]: memref<56x56xf32>,
+// CHECK-SAME: [[INIT_BUF:%.*]]: memref<f32>,
+// CHECK-SAME: [[RESULT_BUF:%.*]]: memref<112x112xf32>) {
+
+// Constants.
+// CHECK: [[C56:%.*]] = constant 56 : index
+// CHECK: [[C1:%.*]] = constant 1 : index
+// CHECK: [[C0_F32:%.*]] = constant 0.000000e+00 : f32
+// CHECK: [[CFALSE:%.*]] = constant 0 : i1
+// CHECK: [[C3:%.*]] = constant 3 : index
+// CHECK: [[C2:%.*]] = constant 2 : index
+// CHECK: [[C0:%.*]] = constant 0 : index
+// CHECK: [[C112:%.*]] = constant 112 : index
+// CHECK: [[CTRUE:%.*]] = constant 1 : i1
+
+// Parallel loop to initialize the output buffer.
+// CHECK: [[INIT:%.*]] = load [[INIT_BUF]][] : memref<f32>
+// CHECK: loop.parallel ([[I:%.*]], [[J:%.*]]) = ([[C0]], [[C0]])
+// CHECK-SAME: to ([[C112]], [[C112]]) step ([[C1]], [[C1]]) {
+// CHECK: store [[INIT]], [[RESULT_BUF]]{{\[}}[[I]], [[J]]]
+// CHECK: loop.yield
+// CHECK: }
+
+// Parallel loop over source buffer to compute scattered values.
+// CHECK: loop.parallel ([[II:%.*]], [[JJ:%.*]]) = ([[C0]], [[C0]])
+// CHECK-SAME: to ([[C56]], [[C56]]) step ([[C1]], [[C1]]) {
+
+// Window loop w.r.t. first dim.
+// CHECK: [[SEL_RES_I:%.*]]:4
+// CHECK-SAME: = loop.for [[WIN_I:%.*]] = [[C0]] to [[C3]] step [[C1]]
+// CHECK-SAME: iter_args(
+// CHECK-SAME: [[SEL_I_0:%.*]] = [[C0]], [[SEL_J_0:%.*]] = [[C0]],
+// CHECK-SAME: [[SEL_VAL_0:%.*]] = [[C0_F32]],
+// CHECK-SAME: [[SEL_INIT_0:%.*]] = [[CFALSE]]
+// CHECK-SAME: ) -> (index, index, f32, i1) {
+
+// Window loop w.r.t. second dim.
+// CHECK: [[SEL_RES_J:%.*]]:4
+// CHECK-SAME: = loop.for [[WIN_J:%.*]] = [[C0]] to [[C3]] step [[C1]]
+// CHECK-SAME: iter_args(
+// CHECK-SAME: [[SEL_I:%.*]] = [[SEL_I_0]], [[SEL_J:%.*]] = [[SEL_J_0]],
+// CHECK-SAME: [[SEL_VAL:%.*]] = [[SEL_VAL_0]],
+// CHECK-SAME: [[SEL_INIT:%.*]] = [[SEL_INIT_0]]
+// CHECK-SAME: ) -> (index, index, f32, i1) {
+
+// Compute index I of the ARG buffer and check whether it is in padding area.
+// CHECK: [[START_I:%.*]] = muli [[II]], [[C2]] : index
+// CHECK: [[OFFSET_I:%.*]] = subi [[WIN_I]], [[C0]] : index
+// CHECK: [[ARG_I:%.*]] = addi [[START_I]], [[OFFSET_I]] : index
+// CHECK: [[ARG_I_FITS:%.*]] = cmpi "ult", [[ARG_I]], [[C112]] : index
+
+// Update `INBOUNDS`, i.e. whether or not ARG indices are inside the boundaries
+// of the buffer or they are in the padding area.
+// CHECK: [[INBOUNDS_0:%.*]] = and [[ARG_I_FITS]], [[CTRUE]] : i1
+
+// Compute index J of the ARG buffer and check whether it is in padding area.
+// CHECK: [[START_J:%.*]] = muli [[JJ]], [[C2]] : index
+// CHECK: [[OFFSET_J:%.*]] = subi [[WIN_J]], [[C0]] : index
+// CHECK: [[ARG_J:%.*]] = addi [[START_J]], [[OFFSET_J]] : index
+// CHECK: [[ARG_J_FITS:%.*]] = cmpi "ult", [[ARG_J]], [[C112]] : index
+
+// Update `INBOUNDS`, i.e. whether or not ARG indices are inside the boundaries
+// of the buffer or they are in the padding area.
+// CHECK: [[INBOUNDS_1:%.*]] = and [[INBOUNDS_0]], [[ARG_J_FITS]] : i1
+
+// If ARG ivs are in the padding area, then 'select' function does not have to
+// be applied, current selected ivs (SEL_I, SEL_J) and value (SEL_VAL) are
+// returned in that case.
+// CHECK: [[IF_INBOUNDS_RES:%.*]]:4
+// CHECK-SAME: = loop.if [[INBOUNDS_1]] -> (index, index, f32, i1) {
+
+
+ // INBOUNDS-THEN-BODY, i.e. if INBOUNDS == true
+
+ // CHECK: [[ARG_ELEM:%.*]] = load [[ARG_BUF]]{{\[}}[[ARG_I]], [[ARG_J]]]
+ // CHECK: [[IF_INIT_RES:%.*]]:4
+ // CHECK-SAME: = loop.if [[SEL_INIT]] -> (index, index, f32, i1) {
+
+ // INIT-THEN-BODY, i.e. INBOUNDS == true and INIT = true
+
+ // The LHLO IR of the select block of the lhlo.select_and_scatter is applied
+ // to the current selected value (SEL_VAL) and the element of the ARG buffer
+ // to compute boolean PRED, whether the new value and ivs should replace the
+ // current ones.
+
+ // Allocate buffers for ARG element, current selected value to adapt LHLO
+ // code.
+ // CHECK: [[ARG_ELEM_BUF:%.*]] = alloc() : memref<f32>
+ // CHECK: [[SEL_VAL_BUF:%.*]] = alloc() : memref<f32>
+ // CHECK: [[PRED_BUF:%.*]] = alloc() : memref<i1>
+ // CHECK: store [[ARG_ELEM]], [[ARG_ELEM_BUF]][] : memref<f32>
+ // CHECK: store [[SEL_VAL]], [[SEL_VAL_BUF]][] : memref<f32>
+
+ // Compute PRED.
+ // CHECK: "xla_lhlo.compare"(
+ // CHECK-SAME: [[ARG_ELEM_BUF]], [[SEL_VAL_BUF]], [[PRED_BUF]])
+ // CHECK: [[PRED:%.*]] = load [[PRED_BUF]][] : memref<i1>
+
+
+ // Depending on PRED, return ARG ivs & elem or current select ivs and value.
+ // CHECK: [[IF_PRED_RES:%.*]]:4 = loop.if [[PRED]]
+ // CHECK: loop.yield [[ARG_I]], [[ARG_J]], [[ARG_ELEM]], [[CTRUE]]
+ // CHECK: } else {
+ // CHECK: loop.yield [[SEL_I]], [[SEL_J]], [[SEL_VAL]], [[SEL_INIT]]
+ // CHECK: }
+
+ // INIT-THEN-BODY yield.
+ // CHECK: loop.yield [[IF_PRED_RES]]#0, [[IF_PRED_RES]]#1,
+ // CHECK-SAME: [[IF_PRED_RES]]#2, [[IF_PRED_RES]]#3
+
+ // INIT-ELSE-BODY, i.e. if INBOUNDS == TRUE and INIT == FALSE, returns ARG
+ // ivs and element without computing Select function.
+ // CHECK: loop.yield [[ARG_I]], [[ARG_J]], [[ARG_ELEM]],
+ // CHECK-SAME: [[CTRUE]] : index, index, f32, i1
+ // CHECK: }
+
+ // INBOUNDS-THEN-BODY yield.
+ // CHECK: loop.yield [[IF_INIT_RES]]#0, [[IF_INIT_RES]]#1, [[IF_INIT_RES]]#2,
+ // CHECK-SAME: [[IF_INIT_RES]]#3 : index, index, f32, i1
+ // CHECK: }
+
+ // INBOUNDS-ELSE-REGION, i.e. if INBOUNDS == FALSE
+ // We are in the pad area, return current iter_args.
+ // CHECK: loop.yield [[SEL_I]], [[SEL_J]], [[SEL_VAL]],
+ // CHECK-SAME: [[SEL_INIT]] : index, index, f32, i1
+ // CHECK: }
+
+// Window loop w.r.t. second dim yield.
+// CHECK: loop.yield [[IF_INBOUNDS_RES]]#0, [[IF_INBOUNDS_RES]]#1,
+// CHECK-SAME: [[IF_INBOUNDS_RES]]#2, [[IF_INBOUNDS_RES]]#3
+// CHECK: }
+
+// Window loop w.r.t. first dim yield.
+// CHECK: loop.yield [[SEL_RES_J]]#0, [[SEL_RES_J]]#1, [[SEL_RES_J]]#2,
+// CHECK-SAME: [[SEL_RES_J]]#3 : index, index, f32, i1
+// CHECK: }
+
+// Use selected ivs to load element from the SRC buffer.
+// CHECK: [[SRC_ELEM:%.*]] = load [[SRC_BUF]]{{\[}}[[II]], [[JJ]]]
+
+// Update of RESULT[SELECTED_I, SELECTED_J] should be done atomically, because
+// it may happen that several other threads select the same IVs if the windows
+// overlap.
+// CHECK: generic_atomic_rmw [[RESULT_BUF]]{{\[}}[[SEL_RES_I]]#0,
+// CHECK-SAME: [[SEL_RES_I]]#1] : memref<112x112xf32>
+// CHECK: ^bb0([[CUR_RES:%.*]]: f32):
+
+// Allocate buffers for ARG element, current selected value to adapt LHLO code.
+// CHECK: [[SRC_ELEM_BUF:%.*]] = alloc() : memref<f32>
+// CHECK: [[CUR_RES_BUF:%.*]] = alloc() : memref<f32>
+// CHECK: [[RES_BUF:%.*]] = alloc() : memref<f32>
+// CHECK: store [[SRC_ELEM]], [[SRC_ELEM_BUF]][] : memref<f32>
+// CHECK: store [[CUR_RES]], [[CUR_RES_BUF]][] : memref<f32>
+
+// Compute scatter value.
+// CHECK: "xla_lhlo.add"([[SRC_ELEM_BUF]], [[CUR_RES_BUF]], [[RES_BUF]]) :
+// CHECK-SAME: (memref<f32>, memref<f32>, memref<f32>) -> ()
+// CHECK: [[RES:%.*]] = load [[RES_BUF]][] : memref<f32>
+
+// Atomic RMW terminator that returns updated value.
+// CHECK: atomic_yield [[RES]] : f32
+
+// Parallel loop over source buffer yield
+// CHECK: loop.yield
diff --git a/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-to-parallel-loops.mlir b/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-to-parallel-loops.mlir
index 17b7d69f..cb169e0 100644
--- a/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-to-parallel-loops.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/lhlo-legalize-to-parallel-loops.mlir
@@ -203,192 +203,3 @@
// CHECK: }
// CHECK: return
// CHECK: }
-
-// -----
-
-func @select_and_scatter(%arg: memref<112x112xf32>,
- %src: memref<56x56xf32>,
- %init: memref<f32>,
- %result: memref<112x112xf32>) {
- "xla_lhlo.select_and_scatter"(%arg, %src, %init, %result) ( {
- // select
- ^bb0(%lhs: memref<f32>, %rhs: memref<f32>, %pred: memref<i1>):
- "xla_lhlo.compare"(%lhs, %rhs, %pred) {comparison_direction = "GE"} :
- (memref<f32>, memref<f32>, memref<i1>) -> ()
- "xla_lhlo.terminator"() : () -> ()
- }, {
- // scatter
- ^bb0(%lhs: memref<f32>, %rhs: memref<f32>, %out: memref<f32>):
- "xla_lhlo.add"(%lhs, %rhs, %out) :
- (memref<f32>, memref<f32>, memref<f32>) -> ()
- "xla_lhlo.terminator"() : () -> ()
- }) {
- padding = dense<[[0, 1], [0, 1]]> : tensor<2x2xi64>,
- window_dimensions = dense<[3, 3]> : tensor<2xi64>,
- window_strides = dense<[2, 2]> : tensor<2xi64>
- } : (memref<112x112xf32>,
- memref<56x56xf32>,
- memref<f32>, memref<112x112xf32>) -> ()
- "xla_lhlo.terminator"() : () -> ()
-}
-// CHECK-LABEL: func @select_and_scatter(
-// CHECK-SAME: [[ARG_BUF:%.*]]: memref<112x112xf32>,
-// CHECK-SAME: [[SRC_BUF:%.*]]: memref<56x56xf32>,
-// CHECK-SAME: [[INIT_BUF:%.*]]: memref<f32>,
-// CHECK-SAME: [[RESULT_BUF:%.*]]: memref<112x112xf32>) {
-
-// Constants.
-// CHECK: [[C56:%.*]] = constant 56 : index
-// CHECK: [[C1:%.*]] = constant 1 : index
-// CHECK: [[C0_F32:%.*]] = constant 0.000000e+00 : f32
-// CHECK: [[CFALSE:%.*]] = constant 0 : i1
-// CHECK: [[C3:%.*]] = constant 3 : index
-// CHECK: [[C2:%.*]] = constant 2 : index
-// CHECK: [[C0:%.*]] = constant 0 : index
-// CHECK: [[C112:%.*]] = constant 112 : index
-// CHECK: [[CTRUE:%.*]] = constant 1 : i1
-
-// Parallel loop to initialize the output buffer.
-// CHECK: [[INIT:%.*]] = load [[INIT_BUF]][] : memref<f32>
-// CHECK: loop.parallel ([[I:%.*]], [[J:%.*]]) = ([[C0]], [[C0]])
-// CHECK-SAME: to ([[C112]], [[C112]]) step ([[C1]], [[C1]]) {
-// CHECK: store [[INIT]], [[RESULT_BUF]]{{\[}}[[I]], [[J]]]
-// CHECK: loop.yield
-// CHECK: }
-
-// Parallel loop over source buffer to compute scattered values.
-// CHECK: loop.parallel ([[II:%.*]], [[JJ:%.*]]) = ([[C0]], [[C0]])
-// CHECK-SAME: to ([[C56]], [[C56]]) step ([[C1]], [[C1]]) {
-
-// Window loop w.r.t. first dim.
-// CHECK: [[SEL_RES_I:%.*]]:4
-// CHECK-SAME: = loop.for [[WIN_I:%.*]] = [[C0]] to [[C3]] step [[C1]]
-// CHECK-SAME: iter_args(
-// CHECK-SAME: [[SEL_I_0:%.*]] = [[C0]], [[SEL_J_0:%.*]] = [[C0]],
-// CHECK-SAME: [[SEL_VAL_0:%.*]] = [[C0_F32]],
-// CHECK-SAME: [[SEL_INIT_0:%.*]] = [[CFALSE]]
-// CHECK-SAME: ) -> (index, index, f32, i1) {
-
-// Window loop w.r.t. second dim.
-// CHECK: [[SEL_RES_J:%.*]]:4
-// CHECK-SAME: = loop.for [[WIN_J:%.*]] = [[C0]] to [[C3]] step [[C1]]
-// CHECK-SAME: iter_args(
-// CHECK-SAME: [[SEL_I:%.*]] = [[SEL_I_0]], [[SEL_J:%.*]] = [[SEL_J_0]],
-// CHECK-SAME: [[SEL_VAL:%.*]] = [[SEL_VAL_0]],
-// CHECK-SAME: [[SEL_INIT:%.*]] = [[SEL_INIT_0]]
-// CHECK-SAME: ) -> (index, index, f32, i1) {
-
-// Compute index I of the ARG buffer and check whether it is in padding area.
-// CHECK: [[START_I:%.*]] = muli [[II]], [[C2]] : index
-// CHECK: [[OFFSET_I:%.*]] = subi [[WIN_I]], [[C0]] : index
-// CHECK: [[ARG_I:%.*]] = addi [[START_I]], [[OFFSET_I]] : index
-// CHECK: [[ARG_I_FITS:%.*]] = cmpi "ult", [[ARG_I]], [[C112]] : index
-
-// Update `INBOUNDS`, i.e. whether or not ARG indices are inside the boundaries
-// of the buffer or they are in the padding area.
-// CHECK: [[INBOUNDS_0:%.*]] = and [[ARG_I_FITS]], [[CTRUE]] : i1
-
-// Compute index J of the ARG buffer and check whether it is in padding area.
-// CHECK: [[START_J:%.*]] = muli [[JJ]], [[C2]] : index
-// CHECK: [[OFFSET_J:%.*]] = subi [[WIN_J]], [[C0]] : index
-// CHECK: [[ARG_J:%.*]] = addi [[START_J]], [[OFFSET_J]] : index
-// CHECK: [[ARG_J_FITS:%.*]] = cmpi "ult", [[ARG_J]], [[C112]] : index
-
-// Update `INBOUNDS`, i.e. whether or not ARG indices are inside the boundaries
-// of the buffer or they are in the padding area.
-// CHECK: [[INBOUNDS_1:%.*]] = and [[INBOUNDS_0]], [[ARG_J_FITS]] : i1
-
-// If ARG ivs are in the padding area, then 'select' function does not have to
-// be applied, current selected ivs (SEL_I, SEL_J) and value (SEL_VAL) are
-// returned in that case.
-// CHECK: [[IF_INBOUNDS_RES:%.*]]:4
-// CHECK-SAME: = loop.if [[INBOUNDS_1]] -> (index, index, f32, i1) {
-
-
- // INBOUNDS-THEN-BODY, i.e. if INBOUNDS == true
-
- // CHECK: [[ARG_ELEM:%.*]] = load [[ARG_BUF]]{{\[}}[[ARG_I]], [[ARG_J]]]
- // CHECK: [[IF_INIT_RES:%.*]]:4
- // CHECK-SAME: = loop.if [[SEL_INIT]] -> (index, index, f32, i1) {
-
- // INIT-THEN-BODY, i.e. INBOUNDS == true and INIT = true
-
- // The LHLO IR of the select block of the lhlo.select_and_scatter is applied
- // to the current selected value (SEL_VAL) and the element of the ARG buffer
- // to compute boolean PRED, whether the new value and ivs should replace the
- // current ones.
-
- // Allocate buffers for ARG element, current selected value to adapt LHLO
- // code.
- // CHECK: [[ARG_ELEM_BUF:%.*]] = alloc() : memref<f32>
- // CHECK: [[SEL_VAL_BUF:%.*]] = alloc() : memref<f32>
- // CHECK: [[PRED_BUF:%.*]] = alloc() : memref<i1>
- // CHECK: store [[ARG_ELEM]], [[ARG_ELEM_BUF]][] : memref<f32>
- // CHECK: store [[SEL_VAL]], [[SEL_VAL_BUF]][] : memref<f32>
-
- // Compute PRED.
- // CHECK: "xla_lhlo.compare"(
- // CHECK-SAME: [[ARG_ELEM_BUF]], [[SEL_VAL_BUF]], [[PRED_BUF]])
- // CHECK: [[PRED:%.*]] = load [[PRED_BUF]][] : memref<i1>
-
-
- // Depending on PRED, return ARG ivs & elem or current select ivs and value.
- // CHECK: [[IF_PRED_RES:%.*]]:4 = loop.if [[PRED]]
- // CHECK: loop.yield [[ARG_I]], [[ARG_J]], [[ARG_ELEM]], [[CTRUE]]
- // CHECK: } else {
- // CHECK: loop.yield [[SEL_I]], [[SEL_J]], [[SEL_VAL]], [[SEL_INIT]]
- // CHECK: }
-
- // INIT-THEN-BODY yield.
- // CHECK: loop.yield [[IF_PRED_RES]]#0, [[IF_PRED_RES]]#1,
- // CHECK-SAME: [[IF_PRED_RES]]#2, [[IF_PRED_RES]]#3
-
- // INIT-ELSE-BODY, i.e. if INBOUNDS == TRUE and INIT == FALSE, returns ARG
- // ivs and element without computing Select function.
- // CHECK: loop.yield [[ARG_I]], [[ARG_J]], [[ARG_ELEM]],
- // CHECK-SAME: [[CTRUE]] : index, index, f32, i1
- // CHECK: }
-
- // INBOUNDS-THEN-BODY yield.
- // CHECK: loop.yield [[IF_INIT_RES]]#0, [[IF_INIT_RES]]#1, [[IF_INIT_RES]]#2,
- // CHECK-SAME: [[IF_INIT_RES]]#3 : index, index, f32, i1
- // CHECK: }
-
- // INBOUNDS-ELSE-REGION, i.e. if INBOUNDS == FALSE
- // We are in the pad area, return current iter_args.
- // CHECK: loop.yield [[SEL_I]], [[SEL_J]], [[SEL_VAL]],
- // CHECK-SAME: [[SEL_INIT]] : index, index, f32, i1
- // CHECK: }
-
-// Window loop w.r.t. second dim yield.
-// CHECK: loop.yield [[IF_INBOUNDS_RES]]#0, [[IF_INBOUNDS_RES]]#1,
-// CHECK-SAME: [[IF_INBOUNDS_RES]]#2, [[IF_INBOUNDS_RES]]#3
-// CHECK: }
-
-// Window loop w.r.t. first dim yield.
-// CHECK: loop.yield [[SEL_RES_J]]#0, [[SEL_RES_J]]#1, [[SEL_RES_J]]#2,
-// CHECK-SAME: [[SEL_RES_J]]#3 : index, index, f32, i1
-// CHECK: }
-
-// Use selected ivs to load element from the SRC buffer.
-// CHECK: [[CUR_RES:%.*]] = load [[RESULT_BUF]]{{\[}}[[SEL_RES_I:%.*]]#0,
-// CHECK-SAME: [[SEL_RES_I]]#1] : memref<112x112xf32>
-// CHECK: [[SRC_ELEM:%.*]] = load [[SRC_BUF]]{{\[}}[[II]], [[JJ]]]
-
-// Allocate buffers for ARG element, current selected value to adapt LHLO code.
-// CHECK: [[SRC_ELEM_BUF:%.*]] = alloc() : memref<f32>
-// CHECK: [[CUR_RES_BUF:%.*]] = alloc() : memref<f32>
-// CHECK: [[RES_BUF:%.*]] = alloc() : memref<f32>
-// CHECK: store [[SRC_ELEM]], [[SRC_ELEM_BUF]][] : memref<f32>
-// CHECK: store [[CUR_RES]], [[CUR_RES_BUF]][] : memref<f32>
-
-// Compute scatter value.
-// CHECK: "xla_lhlo.add"([[SRC_ELEM_BUF]], [[CUR_RES_BUF]], [[RES_BUF]]) :
-// CHECK-SAME: (memref<f32>, memref<f32>, memref<f32>) -> ()
-// CHECK: [[RES:%.*]] = load [[RES_BUF]][] : memref<f32>
-
-// Update RESULT[SELECTED_I, SELECTED_J] with RES.
-// CHECK: store [[RES]], [[RESULT_BUF]]{{\[}}[[SEL_RES_I]]#0, [[SEL_RES_I]]#1]
-
-// Parallel loop over source buffer yield
-// CHECK: loop.yield
diff --git a/tensorflow/compiler/mlir/xla/tests/materialize-broadcasts.mlir b/tensorflow/compiler/mlir/xla/tests/materialize-broadcasts.mlir
index fde5c12..c77f685 100644
--- a/tensorflow/compiler/mlir/xla/tests/materialize-broadcasts.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/materialize-broadcasts.mlir
@@ -227,6 +227,18 @@
// -----
+// CHECK-LABEL: @clampBroadcast
+// CHECK-SAME: (%[[MIN:.+]]: tensor<f32>, %[[VAL:.+]]: tensor<4xf32>, %[[MAX:.+]]: tensor<f32>)
+func @clampBroadcast(%min: tensor<f32>, %value: tensor<4xf32>, %max: tensor<f32>) -> tensor<4xf32> {
+ // CHECK-DAG: %[[MIN_BC:.+]] = "xla_hlo.broadcast"(%[[MIN]]) {broadcast_sizes = dense<4> : tensor<1xi64>} : (tensor<f32>) -> tensor<4xf32>
+ // CHECK-DAG: %[[MAX_BC:.+]] = "xla_hlo.broadcast"(%[[MAX]]) {broadcast_sizes = dense<4> : tensor<1xi64>} : (tensor<f32>) -> tensor<4xf32>
+ // CHECK: "xla_hlo.clamp"(%[[MIN_BC]], %[[VAL]], %[[MAX_BC]]) : (tensor<4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
+ %0 = "xla_hlo.clamp"(%min, %value, %max) : (tensor<f32>, tensor<4xf32>, tensor<f32>) -> tensor<4xf32>
+ return %0 : tensor<4xf32>
+}
+
+// -----
+
// CHECK-LABEL: @compareBroadcastRhs
func @compareBroadcastRhs(%arg0: tensor<1x4xf32>, %arg1: tensor<4xf32>) -> tensor<1x4xi1> {
// CHECK-NEXT: %[[BROADCAST0:.*]] = "xla_hlo.broadcast_in_dim"(%arg0) {broadcast_dimensions = dense<[0, 1]> : tensor<2xi64>} : (tensor<1x4xf32>) -> tensor<1x4xf32>
diff --git a/tensorflow/compiler/mlir/xla/tests/mlir_hlo_builder_test.cc b/tensorflow/compiler/mlir/xla/tests/mlir_hlo_builder_test.cc
new file mode 100644
index 0000000..54791e1
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/tests/mlir_hlo_builder_test.cc
@@ -0,0 +1,179 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/compiler/mlir/xla/ir/mlir_hlo_builder.h"
+
+#include <string>
+
+#include "llvm/Support/raw_ostream.h"
+#include "mlir/IR/Attributes.h" // from @llvm-project
+#include "mlir/IR/Builders.h" // from @llvm-project
+#include "mlir/IR/Dialect.h" // from @llvm-project
+#include "mlir/IR/Location.h" // from @llvm-project
+#include "mlir/IR/MLIRContext.h" // from @llvm-project
+#include "mlir/IR/Module.h" // from @llvm-project
+#include "tensorflow/compiler/mlir/xla/ir/hlo_ops.h"
+#include "tensorflow/compiler/xla/shape_util.h"
+#include "tensorflow/core/lib/core/status_test_util.h"
+#include "tensorflow/core/platform/test.h"
+
+namespace xla {
+
+namespace {
+
+static void ExpectHasSubstr(absl::string_view s, absl::string_view expected) {
+ EXPECT_TRUE(absl::StrContains(s, expected))
+ << s << " does not contain " << expected;
+}
+
+class XlaBuilderTest : public ::testing::Test {
+ protected:
+ XlaBuilderTest()
+ : name_(SetupTest()),
+ context_(),
+ module_(mlir::ModuleOp::create(mlir::UnknownLoc::get(&context_))),
+ builder_(&module_->getBodyRegion()),
+ xla_builder_(name_, builder_, module_->getLoc()) {}
+
+ string SetupTest() {
+ mlir::registerDialect<mlir::xla_hlo::XlaHloDialect>();
+ return ::testing::UnitTest::GetInstance()->current_test_info()->name();
+ }
+
+ // Retuns the MLIR op string representation of the given XlaOp.
+ string GetMlirOpString(XlaOp xla_op) {
+ string str;
+ llvm::raw_string_ostream ostream{str};
+ xla_builder_.GetValue(xla_op).print(ostream);
+ ostream.flush();
+ return str;
+ }
+
+ string name_;
+ mlir::MLIRContext context_;
+ mlir::OwningModuleRef module_;
+ mlir::OpBuilder builder_;
+ MlirHloBuilder xla_builder_;
+};
+
+TEST_F(XlaBuilderTest, CreateToken) {
+ auto token = CreateToken(&xla_builder_);
+ auto str = GetMlirOpString(token);
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+
+ ExpectHasSubstr(GetMlirOpString(token),
+ R"("xla_hlo.create_token"() : () -> !xla_hlo.token)");
+}
+
+TEST_F(XlaBuilderTest, Infeed) {
+ auto token = CreateToken(&xla_builder_);
+ auto infeed = InfeedWithToken(token, ShapeUtil::MakeShape(F32, {4, 8}), "");
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(infeed),
+ R"("xla_hlo.infeed"(%0) {infeed_config = ""} : (!xla_hlo.token) -> tuple<tensor<4x8xf32>, !xla_hlo.token>)");
+}
+
+TEST_F(XlaBuilderTest, Outfeed) {
+ auto outfeed_shape = ShapeUtil::MakeShape(F32, {4, 8});
+ auto data = ConstantLiteral(
+ &xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, outfeed_shape.dimensions()));
+ auto token = CreateToken(&xla_builder_);
+ auto outfeed = OutfeedWithToken(data, token, outfeed_shape, "");
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(outfeed),
+ R"("xla_hlo.outfeed"(%0, %1) {outfeed_config = ""} : (tensor<4x8xf32>, !xla_hlo.token) -> !xla_hlo.token)");
+}
+
+TEST_F(XlaBuilderTest, ConcatInDim) {
+ auto data0 = ConstantLiteral(
+ &xla_builder_, LiteralUtil::CreateFromDimensions(F32, {2, 4, 5}));
+ auto data1 = ConstantLiteral(
+ &xla_builder_, LiteralUtil::CreateFromDimensions(F32, {2, 6, 5}));
+ auto concat = ConcatInDim(&xla_builder_, {data0, data1}, 1);
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(concat),
+ R"("xla_hlo.concatenate"(%0, %1) {dimension = 1 : i64} : (tensor<2x4x5xf32>, tensor<2x6x5xf32>) -> tensor<2x10x5xf32>)");
+}
+
+TEST_F(XlaBuilderTest, Tuple) {
+ auto data0 = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {3, 7}));
+ auto data1 = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {}));
+ auto tuple = Tuple(&xla_builder_, {data0, data1});
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(tuple),
+ R"("xla_hlo.tuple"(%0, %1) : (tensor<3x7xf32>, tensor<f32>) -> tuple<tensor<3x7xf32>, tensor<f32>>)");
+}
+
+TEST_F(XlaBuilderTest, GetTupleElement) {
+ auto data0 = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {3, 7}));
+ auto data1 = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {}));
+ auto tuple_data = Tuple(&xla_builder_, {data0, data1});
+ auto gte = GetTupleElement(tuple_data, 1);
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(gte),
+ R"("xla_hlo.get_tuple_element"(%2) {index = 1 : i32} : (tuple<tensor<3x7xf32>, tensor<f32>>) -> tensor<f32>)");
+}
+
+TEST_F(XlaBuilderTest, Slice) {
+ auto data = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {3, 7}));
+ auto slice = Slice(data, {0, 1}, {2, 5}, {1, 1});
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(slice),
+ R"("xla_hlo.slice"(%0) {limit_indices = dense<[2, 5]> : tensor<2xi64>, start_indices = dense<[0, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} : (tensor<3x7xf32>) -> tensor<2x4xf32>)");
+}
+
+TEST_F(XlaBuilderTest, Pad) {
+ auto data = ConstantLiteral(&xla_builder_,
+ LiteralUtil::CreateFromDimensions(F32, {3, 7}));
+ auto zero = ConstantLiteral(&xla_builder_, LiteralUtil::Zero(F32));
+
+ PaddingConfig padding_config;
+ auto* dims0 = padding_config.add_dimensions();
+ dims0->set_edge_padding_low(1);
+ dims0->set_interior_padding(0);
+ dims0->set_edge_padding_high(2);
+ auto* dims1 = padding_config.add_dimensions();
+ dims1->set_edge_padding_low(3);
+ dims1->set_interior_padding(1);
+ dims1->set_edge_padding_high(0);
+ auto pad = Pad(data, zero, padding_config);
+
+ TF_ASSERT_OK(xla_builder_.GetCurrentStatus());
+ ExpectHasSubstr(
+ GetMlirOpString(pad),
+ R"("xla_hlo.pad"(%0, %1) {edge_padding_high = dense<[2, 0]> : tensor<2xi64>, edge_padding_low = dense<[1, 3]> : tensor<2xi64>, interior_padding = dense<[0, 1]> : tensor<2xi64>} : (tensor<3x7xf32>, tensor<f32>) -> tensor<6x16xf32>)");
+}
+
+} // namespace
+} // namespace xla
diff --git a/tensorflow/compiler/mlir/xla/tests/translate/export.mlir b/tensorflow/compiler/mlir/xla/tests/translate/export.mlir
index b65381b..f27c1be 100644
--- a/tensorflow/compiler/mlir/xla/tests/translate/export.mlir
+++ b/tensorflow/compiler/mlir/xla/tests/translate/export.mlir
@@ -460,14 +460,18 @@
// -----
// CHECK: HloModule
-func @main(%arg: tensor<4x2xf32>) -> tensor<i32> {
- %0 = "xla_hlo.get_dimension_size"(%arg) {dimension = 1 : i32} : (tensor<4x2xf32>) -> tensor<i32>
- return %0 : tensor<i32>
+func @main(%arg: tensor<4x2xf32>, %size: tensor<i32>) -> tensor<i32> {
+ %0 = "xla_hlo.set_dimension_size"(%arg, %size) {dimension = 1 : i32} : (tensor<4x2xf32>, tensor<i32>) -> tensor<4x2xf32>
+ %1 = "xla_hlo.get_dimension_size"(%0) {dimension = 1 : i32} : (tensor<4x2xf32>) -> tensor<i32>
+ return %1 : tensor<i32>
}
// CHECK: ENTRY
// CHECK: [[ARG:%.*]] = f32[4,2] parameter(0)
-// CHECK: s32[] get-dimension-size(f32[4,2] [[ARG]]), dimensions={1}
+// CHECK: [[SIZE:%.*]] = s32[] parameter(1)
+// CHECK: [[DYNAMIC:%.*]] = f32[4,<=2] set-dimension-size(f32[4,2] [[ARG]], s32[] [[SIZE]]), dimensions={1}
+// CHECK: ROOT %[[RESULT:.*]] = s32[] get-dimension-size(f32[4,<=2] [[DYNAMIC]]), dimensions={1}
+
// -----
diff --git a/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo.cc b/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo.cc
new file mode 100644
index 0000000..8ef2ed4
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo.cc
@@ -0,0 +1,208 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "mlir/Dialect/Shape/IR/Shape.h" // from @llvm-project
+#include "mlir/IR/Attributes.h" // from @llvm-project
+#include "mlir/IR/MLIRContext.h" // from @llvm-project
+#include "mlir/IR/OperationSupport.h" // from @llvm-project
+#include "mlir/IR/PatternMatch.h" // from @llvm-project
+#include "mlir/IR/StandardTypes.h" // from @llvm-project
+#include "mlir/Transforms/DialectConversion.h" // from @llvm-project
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.h"
+#include "tensorflow/compiler/mlir/xla/ir/hlo_ops.h"
+#include "tensorflow/compiler/mlir/xla/transforms/rewriters.h"
+
+namespace mlir {
+namespace xla_chlo {
+
+namespace {
+
+// Converts binary ops that statically are determined to not broadcast directly
+// to the corresponding xla_hlo non-broadcasting op.
+template <typename ChloOpTy, typename HloOpTy>
+struct ConvertTrivialNonBroadcastBinaryOp
+ : public OpConversionPattern<ChloOpTy> {
+ using OpConversionPattern<ChloOpTy>::OpConversionPattern;
+ LogicalResult matchAndRewrite(
+ ChloOpTy op, ArrayRef<Value> operands,
+ ConversionPatternRewriter &rewriter) const override {
+ // Only rewrite for statically determinable non-broadcasting cases.
+ auto lhs = operands[0].getType().dyn_cast<RankedTensorType>();
+ auto rhs = operands[1].getType().dyn_cast<RankedTensorType>();
+ if (!lhs || !rhs) return failure();
+
+ // Requires rank broadcast.
+ if (lhs.getRank() != rhs.getRank()) return failure();
+ // Any dynamic dimension may require broadcasting and requires more
+ // analysis.
+ if (!lhs.hasStaticShape() || !rhs.hasStaticShape()) return failure();
+
+ for (auto extents : llvm::zip(lhs.getShape(), rhs.getShape())) {
+ auto lhs_extent = std::get<0>(extents);
+ auto rhs_extent = std::get<1>(extents);
+ if (lhs_extent != rhs_extent) {
+ return failure();
+ }
+ }
+
+ rewriter.replaceOp(op, rewriter.createOrFold<HloOpTy>(
+ op.getLoc(), operands[0], operands[1],
+ /*broadcast_dimensions=*/nullptr));
+ return success();
+ }
+};
+
+// Checks whether the given operand types and broadcast_dims attr represent a
+// legal combination for "numpy" style broadcasting (where 1-dims are prepended
+// to the smaller ranked operand until it is of the same rank as the larger).
+bool IsLegalNumpyRankedBroadcast(RankedTensorType lhs_type,
+ RankedTensorType rhs_type,
+ DenseIntElementsAttr broadcast_dims) {
+ if (lhs_type.getRank() == rhs_type.getRank()) return true;
+
+ // Otherwise, verify that broadcast_dims strictly performs left-padding.
+ auto smaller_rank = std::min(lhs_type.getRank(), rhs_type.getRank());
+ auto larger_rank = std::max(lhs_type.getRank(), rhs_type.getRank());
+
+ auto expected_extents = llvm::to_vector<4>(
+ llvm::seq<int64_t>(larger_rank - smaller_rank, larger_rank));
+ if (expected_extents.size() != broadcast_dims.getNumElements()) {
+ return false;
+ }
+ return std::equal(expected_extents.begin(), expected_extents.end(),
+ broadcast_dims.getIntValues().begin());
+}
+
+// Converts a binary op with ranked broadcasting operands to explicitly
+// broadcast and invoke the corresponding xla_hlo non-broadcasting op.
+// Note that dynamic broadcasting supported by this pattern is only valid for
+// "numpy" broadcasting semantics as defined here:
+// https://docs.scipy.org/doc/numpy/reference/ufuncs.html
+// Specifically, this includes the following cases:
+// - Same rank broadcast (operands have the same static rank).
+// - Different-rank broadcast, either without a broadcast_dims attribte or
+// with the broadcast_dims attribute set to map to a prefix padding.
+// - Legal combinations of degenerate (1-dim) implicit broadcasting.
+// The restriction on broadcast_dims derives from the definition of the
+// `shape.broadcast` op, which only supports prefix-padding.
+//
+// It may be possible to expand this pattern to operate on unranked tensors in
+// the future by emitting more code to dynamically differentiate based on rank.
+// Whether that is of any practical benefit remains to be seen.
+template <typename ChloOpTy, typename HloOpTy>
+struct ConvertRankedDynamicBroadcastBinaryOp
+ : public OpConversionPattern<ChloOpTy> {
+ using OpConversionPattern<ChloOpTy>::OpConversionPattern;
+ LogicalResult matchAndRewrite(
+ ChloOpTy op, ArrayRef<Value> operands,
+ ConversionPatternRewriter &rewriter) const override {
+ // Only support ranked operands.
+ Value lhs = operands[0];
+ Value rhs = operands[1];
+ auto lhs_type = lhs.getType().dyn_cast<RankedTensorType>();
+ auto rhs_type = rhs.getType().dyn_cast<RankedTensorType>();
+ auto result_type = op.getResult().getType();
+ if (!lhs_type || !rhs_type) return failure();
+
+ // Check for "numpy"-style rank broadcast.
+ auto broadcast_dimensions = op.broadcast_dimensions();
+ if (broadcast_dimensions &&
+ !IsLegalNumpyRankedBroadcast(lhs_type, rhs_type,
+ *op.broadcast_dimensions())) {
+ // Note: It is unclear whether the general specification of explicit
+ // broadcast_dimensions on binary ops is a feature we want to carry
+ // forward. While it can technically be implemented for ranked-dynamic,
+ // it is incompatible with unranked inputs. If this warning is emitted
+ // in real programs, it is an indication that the feature should be
+ // implemented versus just falling back on the more standard definition
+ // of numpy-like prefix-padding.
+ op.emitWarning() << "unsupported non prefix-padded dynamic rank "
+ << "broadcast_dimensions = " << *broadcast_dimensions;
+ return failure();
+ }
+
+ // Compute result shape.
+ auto loc = op.getLoc();
+ int64_t result_rank = std::max(lhs_type.getRank(), rhs_type.getRank());
+ auto shape_type = shape::ShapeType::get(rewriter.getContext());
+ Value lhs_shape_v =
+ rewriter.createOrFold<shape::ShapeOfOp>(loc, shape_type, lhs);
+ Value rhs_shape_v =
+ rewriter.createOrFold<shape::ShapeOfOp>(loc, shape_type, rhs);
+ Value result_shape_v = rewriter.createOrFold<shape::BroadcastOp>(
+ loc, shape_type, lhs_shape_v, rhs_shape_v, nullptr /* error */);
+ Value result_extents = rewriter.createOrFold<shape::ToExtentTensorOp>(
+ loc, RankedTensorType::get({result_rank}, rewriter.getIndexType()),
+ result_shape_v);
+
+ // Note that we unconditionally emit DynamicBroadcastInDim ops and let
+ // downstream canonicalizations fold them away if possible. This is
+ // because, in the dynamic case, there are many corner cases regarding
+ // when it is safe to omit, and some of them require analysis to prove
+ // properly.
+ auto lhs_broadcast_dimensions = llvm::to_vector<4>(
+ llvm::seq<int64_t>(result_rank - lhs_type.getRank(), result_rank));
+ Value broadcasted_lhs = rewriter.create<xla_hlo::DynamicBroadcastInDimOp>(
+ loc, result_type, lhs, result_extents,
+ rewriter.getI64TensorAttr(lhs_broadcast_dimensions));
+ auto rhs_broadcast_dimensions = llvm::to_vector<4>(
+ llvm::seq<int64_t>(result_rank - rhs_type.getRank(), result_rank));
+ Value broadcasted_rhs = rewriter.create<xla_hlo::DynamicBroadcastInDimOp>(
+ loc, result_type, rhs, result_extents,
+ rewriter.getI64TensorAttr(rhs_broadcast_dimensions));
+
+ // And generate the final non-broadcasted binary op.
+ rewriter.replaceOpWithNewOp<HloOpTy>(op, result_type, broadcasted_lhs,
+ broadcasted_rhs,
+ /*broadcast_dimensions=*/nullptr);
+ return success();
+ }
+};
+
+template <typename ChloOpTy, typename HloOpTy>
+void PopulateForBinaryOp(MLIRContext *context,
+ OwningRewritePatternList *patterns) {
+ patterns->insert<ConvertTrivialNonBroadcastBinaryOp<ChloOpTy, HloOpTy>>(
+ context, 10);
+ patterns->insert<ConvertRankedDynamicBroadcastBinaryOp<ChloOpTy, HloOpTy>>(
+ context, 5);
+}
+
+} // namespace
+
+void PopulateLegalizeChloToHloPatterns(MLIRContext *context,
+ OwningRewritePatternList *patterns) {
+#define POPULATE_BCAST(ChloOp, HloOp) \
+ PopulateForBinaryOp<ChloOp, xla_hlo::HloOp>(context, patterns);
+
+ POPULATE_BCAST(BroadcastAddOp, AddOp);
+ POPULATE_BCAST(BroadcastAndOp, AndOp);
+ POPULATE_BCAST(BroadcastAtan2Op, Atan2Op);
+ POPULATE_BCAST(BroadcastDivOp, DivOp);
+ POPULATE_BCAST(BroadcastMaxOp, MaxOp);
+ POPULATE_BCAST(BroadcastMinOp, MinOp);
+ POPULATE_BCAST(BroadcastMulOp, MulOp);
+ POPULATE_BCAST(BroadcastOrOp, OrOp);
+ POPULATE_BCAST(BroadcastPowOp, PowOp);
+ POPULATE_BCAST(BroadcastRemOp, RemOp);
+ POPULATE_BCAST(BroadcastShiftLeftOp, ShiftLeftOp);
+ POPULATE_BCAST(BroadcastShiftRightArithmeticOp, ShiftRightArithmeticOp);
+ POPULATE_BCAST(BroadcastShiftRightLogicalOp, ShiftRightLogicalOp);
+ POPULATE_BCAST(BroadcastSubOp, SubOp);
+ POPULATE_BCAST(BroadcastXorOp, XorOp);
+}
+
+} // namespace xla_chlo
+} // namespace mlir
diff --git a/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo_pass.cc b/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo_pass.cc
new file mode 100644
index 0000000..a4d0918
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/transforms/chlo_legalize_to_hlo_pass.cc
@@ -0,0 +1,57 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "mlir/Dialect/Shape/IR/Shape.h" // from @llvm-project
+#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
+#include "mlir/Pass/Pass.h" // from @llvm-project
+#include "tensorflow/compiler/mlir/xla/ir/chlo_ops.h"
+#include "tensorflow/compiler/mlir/xla/ir/hlo_ops.h"
+#include "tensorflow/compiler/mlir/xla/transforms/rewriters.h"
+
+namespace mlir {
+namespace xla_chlo {
+
+namespace {
+
+struct TestChloLegalizeToHloPass
+ : public PassWrapper<TestChloLegalizeToHloPass, FunctionPass> {
+ void runOnFunction() override {
+ ConversionTarget conversionTarget(getContext());
+ OwningRewritePatternList conversionPatterns;
+
+ conversionTarget.addIllegalDialect<XlaHloClientDialect>();
+ // Consider the xla_hlo dialect legal for tests.
+ conversionTarget.addLegalDialect<xla_hlo::XlaHloDialect>();
+ // The conversion uses helpers from the Standard dialect.
+ conversionTarget.addLegalDialect<mlir::StandardOpsDialect>();
+ conversionTarget.addLegalDialect<mlir::shape::ShapeDialect>();
+
+ PopulateLegalizeChloToHloPatterns(&getContext(), &conversionPatterns);
+
+ if (failed(applyPartialConversion(getFunction(), conversionTarget,
+ conversionPatterns))) {
+ return signalPassFailure();
+ }
+ }
+};
+
+} // namespace
+
+} // namespace xla_chlo
+} // namespace mlir
+
+static mlir::PassRegistration<mlir::xla_chlo::TestChloLegalizeToHloPass> pass(
+ "test-xla-chlo-legalize-to-hlo",
+ "Test pass for applying chlo -> hlo legalization patterns");
diff --git a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
index d3fb832..aa29241 100644
--- a/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc
@@ -447,18 +447,21 @@
HloToLhloOpConverter<xla_hlo::BroadcastInDimOp>,
HloToLhloOpConverter<xla_hlo::CeilOp>,
HloToLhloOpConverter<xla_hlo::CompareOp>,
+ HloToLhloOpConverter<xla_hlo::ComplexOp>,
HloToLhloOpConverter<xla_hlo::ConstOp>,
HloToLhloOpConverter<xla_hlo::ConvertOp>,
HloToLhloOpConverter<xla_hlo::CopyOp>,
HloToLhloOpConverter<xla_hlo::CosOp>,
HloToLhloOpConverter<xla_hlo::DivOp>,
HloToLhloOpConverter<xla_hlo::ExpOp>,
+ HloToLhloOpConverter<xla_hlo::ImagOp>,
HloToLhloOpConverter<xla_hlo::IotaOp>,
HloToLhloOpConverter<xla_hlo::LogOp>,
HloToLhloOpConverter<xla_hlo::MaxOp>,
HloToLhloOpConverter<xla_hlo::MinOp>,
HloToLhloOpConverter<xla_hlo::MulOp>,
HloToLhloOpConverter<xla_hlo::NegOp>,
+ HloToLhloOpConverter<xla_hlo::RealOp>,
HloToLhloOpConverter<xla_hlo::RemOp>,
HloToLhloOpConverter<xla_hlo::RsqrtOp>,
HloToLhloOpConverter<xla_hlo::SelectOp>,
diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc b/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
index 7dafa4d..3853821 100644
--- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf.cc
@@ -46,6 +46,7 @@
#include "tensorflow/compiler/mlir/xla/ir/hlo_utils.h"
#include "tensorflow/compiler/mlir/xla/transforms/passes.h"
#include "tensorflow/compiler/xla/client/padding.h"
+#include "tensorflow/compiler/xla/client/sharding_builder.h"
#include "tensorflow/compiler/xla/xla_data.pb.h"
#include "tensorflow/core/framework/kernel_shape_util.h"
#include "tensorflow/core/kernels/conv_grad_shape_utils.h"
@@ -56,6 +57,8 @@
namespace xla_hlo {
namespace {
+constexpr char kShardingAttr[] = "xla_hlo.sharding";
+
class LegalizeTF : public PassWrapper<LegalizeTF, FunctionPass> {
public:
LegalizeTF() = default;
@@ -115,6 +118,28 @@
return DenseIntElementsAttr::get(ty, values);
}
+// Returns a 1-d i64 elements attribute populated with numbers from start to
+// end, excluding.
+static DenseIntElementsAttr GetI64ElementsAttrForSeq(int start, int end,
+ Builder *builder) {
+ int size = end - start;
+
+ SmallVector<int64_t, 4> vals;
+ vals.resize(size);
+ std::iota(vals.begin(), vals.end(), start);
+
+ TensorType ty = RankedTensorType::get({size}, builder->getIntegerType(64));
+ return DenseIntElementsAttr::get(ty, vals);
+}
+
+// Returns a 1-d i64 elements attribute populated with `val` repeated `size`
+// times.
+static DenseIntElementsAttr GetI64ElementsAttrForValue(int size, int64_t val,
+ Builder *builder) {
+ TensorType ty = RankedTensorType::get({size}, builder->getIntegerType(64));
+ return DenseIntElementsAttr::get(ty, val);
+}
+
// Returns the corresponding type that should be used for performing sum
// accumulation over the given input type.
Type GetSumAccumulationType(Type input_type) {
@@ -194,10 +219,17 @@
return ranked_ty.getDimSize(index);
}
-template <typename T>
+template <typename T, int num_dims>
tensorflow::TensorShape ToTensorShape(llvm::ArrayRef<T> sizes) {
- return tensorflow::TensorShape(
- llvm::SmallVector<tensorflow::int64, 4>(sizes.begin(), sizes.end()));
+ return tensorflow::TensorShape(llvm::SmallVector<tensorflow::int64, num_dims>(
+ sizes.begin(), sizes.end()));
+}
+
+template <typename T, int num_dims>
+tensorflow::TensorShape ToTensorShape(
+ llvm::iterator_range<DenseElementsAttr::ElementIterator<T>> sizes) {
+ return tensorflow::TensorShape(llvm::SmallVector<tensorflow::int64, num_dims>(
+ sizes.begin(), sizes.end()));
}
// Returns minimal value for the given int or float element type.
@@ -239,8 +271,122 @@
// Returns int or float scalar DenseElementsAttr attribute with the given
// element type and the value.
static ConstOp GetScalarConstOfType(Type ty, Location loc, int64_t raw_value,
- PatternRewriter *rewriter) {
- return rewriter->create<ConstOp>(loc, xla::GetScalarOfType(ty, raw_value));
+ OpBuilder *builder) {
+ return builder->create<ConstOp>(loc, xla::GetScalarOfType(ty, raw_value));
+}
+
+// Creates an xla_hlo::SliceOp where the major dimensions have full size, and
+// the minor dimensions have the provided offsets and sizes.
+static Value SliceInMinorDims(Location loc, Value v,
+ ArrayRef<int64_t> minor_starts,
+ ArrayRef<int64_t> minor_limits,
+ OpBuilder *builder) {
+ auto type = v.getType().cast<RankedTensorType>();
+ llvm::SmallVector<int64_t, 4> slice_starts(type.getRank(), 0);
+ int64_t major_dims = type.getRank() - minor_starts.size();
+ std::copy(minor_starts.begin(), minor_starts.end(),
+ slice_starts.begin() + major_dims);
+ auto slice_limits = llvm::to_vector<4>(type.getShape());
+ std::copy(minor_limits.begin(), minor_limits.end(),
+ slice_limits.begin() + major_dims);
+ llvm::SmallVector<int64_t, 4> slice_strides(type.getRank(), 1);
+ return builder->create<SliceOp>(loc, v,
+ GetI64ElementsAttr(slice_starts, builder),
+ GetI64ElementsAttr(slice_limits, builder),
+ GetI64ElementsAttr(slice_strides, builder));
+}
+
+// Creates a vector of index values:
+// [0, 0, ..., minor_indices[0], minor_indices[1], ... minor_indices[-1]]
+// with length `rank`.
+static llvm::SmallVector<Value, 4> CreateFullIndexVectorFromMinorIndices(
+ Location loc, ArrayRef<Value> minor_indices, int64_t rank,
+ OpBuilder *builder) {
+ auto zero =
+ GetScalarConstOfType(getElementTypeOrSelf(minor_indices[0].getType()),
+ loc, 0, builder)
+ .output();
+ llvm::SmallVector<Value, 4> indices(rank, zero);
+ std::copy(minor_indices.begin(), minor_indices.end(),
+ indices.begin() + (rank - minor_indices.size()));
+ return indices;
+}
+
+// Creates an xla_hlo::DynamicSliceOp where the major dimensions have full size,
+// and the minor dimensions have the provided offsets and sizes.
+static Value DynamicSliceInMinorDims(Location loc, Value v,
+ ArrayRef<Value> minor_starts,
+ ArrayRef<int64_t> minor_sizes,
+ OpBuilder *builder) {
+ if (minor_starts.empty()) return v;
+ auto type = v.getType().cast<RankedTensorType>();
+ auto slice_starts = CreateFullIndexVectorFromMinorIndices(
+ loc, minor_starts, type.getRank(), builder);
+ int64_t major_dims = type.getRank() - minor_starts.size();
+ auto slice_sizes = llvm::to_vector<4>(type.getShape());
+ std::copy(minor_sizes.begin(), minor_sizes.end(),
+ slice_sizes.begin() + major_dims);
+ auto slice_type = RankedTensorType::get(slice_sizes, type.getElementType());
+ return builder->create<xla_hlo::DynamicSliceOp>(
+ loc, slice_type, v, slice_starts,
+ GetI64ElementsAttr(slice_sizes, builder));
+}
+
+// Creates an xla_hlo::DynamicUpdateSliceOp where the major dimensions have zero
+// offsets, and the minor dimensions have the provided offsets.
+static Value DynamicUpdateSliceInMinorDims(Location loc, Value v, Value update,
+ ArrayRef<Value> minor_starts,
+ OpBuilder *builder) {
+ if (minor_starts.empty()) return v;
+ auto type = v.getType().cast<RankedTensorType>();
+ auto dus_starts = CreateFullIndexVectorFromMinorIndices(
+ loc, minor_starts, type.getRank(), builder);
+ return builder->create<DynamicUpdateSliceOp>(loc, type, v, update,
+ llvm::makeArrayRef(dus_starts));
+}
+
+// Creates an xla_hlo::DynamicUpdateSliceOp where the major dimensions have zero
+// offsets, and the minor dimensions have the provided static offsets.
+static Value UpdateSliceInMinorDims(Location loc, Value v, Value update,
+ ArrayRef<int64_t> minor_starts,
+ OpBuilder *builder) {
+ llvm::SmallVector<Value, 4> dus_starts(minor_starts.size());
+ for (int64_t i = 0; i < minor_starts.size(); ++i) {
+ dus_starts[i] = GetScalarConstOfType(builder->getIntegerType(32), loc,
+ minor_starts[i], builder);
+ }
+ return DynamicUpdateSliceInMinorDims(loc, v, update, dus_starts, builder);
+}
+
+// Creates a batch dot using xla_hlo::DotGeneralOp.
+Value BatchDot(Location loc, Value lhs, bool transpose_lhs, Value rhs,
+ bool transpose_rhs, int64_t num_batch_dims,
+ ArrayAttr precision_config, OpBuilder *builder) {
+ auto batch_dimensions = GetI64ElementsAttr(
+ llvm::to_vector<4>(llvm::seq<int64_t>(0, num_batch_dims)), builder);
+ auto lhs_contracting_dimensions = GetI64ElementsAttr(
+ llvm::makeArrayRef({transpose_lhs ? num_batch_dims : num_batch_dims + 1}),
+ builder);
+ auto rhs_contracting_dimensions = GetI64ElementsAttr(
+ llvm::makeArrayRef({transpose_rhs ? num_batch_dims + 1 : num_batch_dims}),
+ builder);
+ auto dimension_numbers = DotDimensionNumbers::get(
+ /*lhs_batching_dimensions=*/batch_dimensions,
+ /*rhs_batching_dimensions=*/batch_dimensions,
+ /*lhs_contracting_dimensions=*/lhs_contracting_dimensions,
+ /*rhs_contracting_dimensions=*/rhs_contracting_dimensions,
+ builder->getContext());
+ auto lhs_shape = lhs.getType().cast<RankedTensorType>().getShape();
+ auto rhs_shape = rhs.getType().cast<RankedTensorType>().getShape();
+ auto shape = llvm::to_vector<4>(lhs_shape);
+ shape[shape.size() - 2] =
+ transpose_lhs ? lhs_shape.back() : lhs_shape[lhs_shape.size() - 2];
+ shape[shape.size() - 1] =
+ transpose_rhs ? rhs_shape[rhs_shape.size() - 2] : rhs_shape.back();
+ Type element_type = getElementTypeOrSelf(lhs.getType());
+ return builder->create<DotGeneralOp>(
+ loc, RankedTensorType::get(shape, element_type), lhs, rhs,
+ dimension_numbers, precision_config);
}
// Builds body for reduce op by using the using the template binary op as the
@@ -567,20 +713,6 @@
// Softmax op utilities.
//===----------------------------------------------------------------------===//
-// Returns a 1-d i64 elements attribute populated with numbers from start to
-// end, excluding.
-static DenseIntElementsAttr GetI64ElementsAttrForSeq(int start, int end,
- Builder *builder) {
- int size = end - start;
-
- SmallVector<int64_t, 4> vals;
- vals.resize(size);
- std::iota(vals.begin(), vals.end(), start);
-
- TensorType ty = RankedTensorType::get({size}, builder->getIntegerType(64));
- return DenseIntElementsAttr::get(ty, vals);
-}
-
// Returns the type to use for accumulating the given type.
static Type GetAccumulationType(Type ty) {
// Upcast 16 bit sum reductions to 32 bit to reduce the precision loss from
@@ -1290,11 +1422,25 @@
// The mean, variance, and reserved space outputs of the batch norm op are
// not used for inference. It doesn't matter what values we provide for
- // the last 5 results.
- rewriter.replaceOp(
- op, {/*y=*/y_out, /*batch_mean=*/op.x(),
- /*batch_variance=*/op.x(), /*reserve_space_1=*/op.x(),
- /*reserve_space_2=*/op.x(), /*reserve_space_3=*/op.x()});
+ // the last 5 results as long as they are of the same type. Forward
+ // input mean and variance to output mean, variance, reserved_space_1 and
+ // reserver_space_2. Create a constant tensor to forward to last
+ // reserve_space_3 output.
+ auto reserve_space_3_type = op.getResult(5).getType().cast<TensorType>();
+ int num_elements = reserve_space_3_type.hasStaticShape()
+ ? reserve_space_3_type.getNumElements()
+ : 0;
+ auto const_attr_type = RankedTensorType::get(
+ {num_elements}, getElementTypeOrSelf(reserve_space_3_type));
+ auto dummy_const = rewriter.create<ConstOp>(
+ op.getLoc(), reserve_space_3_type,
+ DenseElementsAttr::get<float>(const_attr_type, 0.0));
+ rewriter.replaceOp(op, {/*y=*/y_out,
+ /*batch_mean=*/op.mean(),
+ /*batch_variance=*/op.variance(),
+ /*reserve_space_1=*/op.mean(),
+ /*reserve_space_2=*/op.variance(),
+ /*reserve_space_3=*/dummy_const.getResult()});
}
return success();
}
@@ -1304,13 +1450,15 @@
//
// Requires padding to be either 'SAME' or 'VALID' and the number of input
// dimensions to be equal to the size of window dimensions and window strides.
+template <int num_dims>
static DenseIntElementsAttr GetReduceWindowPadding(
llvm::ArrayRef<int64_t> input_dims, ArrayAttr window_dims,
ArrayAttr window_strides, StringRef padding, Builder *builder) {
if (padding == "VALID") return {};
DCHECK_EQ(padding.str(), "SAME");
- llvm::SmallVector<tensorflow::int64, 4> input_shape, window_shape, strides;
+ llvm::SmallVector<tensorflow::int64, num_dims> input_shape, window_shape,
+ strides;
input_shape.reserve(input_dims.size());
window_shape.reserve(window_shape.size());
strides.reserve(window_strides.size());
@@ -1325,7 +1473,7 @@
::xla::MakePadding(input_shape, window_shape, strides,
::xla::Padding::kSame);
int64_t rank = paddings.size();
- llvm::SmallVector<int64_t, 8> flatten_paddings(rank * 2);
+ llvm::SmallVector<int64_t, num_dims * 2> flatten_paddings(rank * 2);
for (int i = 0; i < rank; i++) {
flatten_paddings[2 * i] = paddings[i].first;
flatten_paddings[2 * i + 1] = paddings[i].second;
@@ -1335,7 +1483,7 @@
flatten_paddings);
}
-// Converts MaxPool op to HLO ReduceWindow op by setting appropriate window
+// Converts AvgPool op to HLO ReduceWindow op by setting appropriate window
// dimensions with add as the reduction function. The reduction result is
// then divided by the number of elements in the window.
class ConvertAvgPoolOp : public OpRewritePattern<TF::AvgPoolOp> {
@@ -1375,8 +1523,8 @@
Value init =
GetScalarConstOfType(sum_element_type, op.getLoc(), 0, &rewriter);
DenseIntElementsAttr paddings_attr =
- GetReduceWindowPadding(input_type.getShape(), op.ksize(), op.strides(),
- op.padding(), &rewriter);
+ GetReduceWindowPadding<4>(input_type.getShape(), op.ksize(),
+ op.strides(), op.padding(), &rewriter);
auto reduce = rewriter.create<ReduceWindowOp>(
op.getLoc(), result_type, input_value, init,
GetI64ElementsAttr(op.ksize()), GetI64ElementsAttr(op.strides()),
@@ -1418,21 +1566,22 @@
// %max_pool = "xla_hlo.reduce"(%inp, %init) ["xla_hlo.maximum"]
// {window_dimensions = ..., window_strides = ... }
//
-class ConvertMaxPoolOp : public OpRewritePattern<TF::MaxPoolOp> {
+template <typename OpTy, int num_dims>
+class ConvertMaxPoolOp : public OpRewritePattern<OpTy> {
public:
- using OpRewritePattern::OpRewritePattern;
+ using OpRewritePattern<OpTy>::OpRewritePattern;
- LogicalResult matchAndRewrite(TF::MaxPoolOp op,
+ LogicalResult matchAndRewrite(OpTy op,
PatternRewriter &rewriter) const override {
Type element_type =
- op.input().getType().cast<TensorType>().getElementType();
+ op.input().getType().template cast<TensorType>().getElementType();
if (!element_type.isSignlessIntOrFloat()) return failure();
Location loc = op.getLoc();
ConstOp init = GetMinValueForType(element_type, loc, &rewriter);
- auto input_ty = op.input().getType().dyn_cast<RankedTensorType>();
+ auto input_ty = op.input().getType().template dyn_cast<RankedTensorType>();
if (!input_ty) return failure();
- DenseIntElementsAttr paddings_attr = GetReduceWindowPadding(
+ DenseIntElementsAttr paddings_attr = GetReduceWindowPadding<num_dims>(
input_ty.getShape(), op.ksize(), op.strides(), op.padding(), &rewriter);
auto reduce = rewriter.create<ReduceWindowOp>(
loc, op.getType(), op.input(), init.getResult(),
@@ -1446,6 +1595,9 @@
}
};
+using ConvertMaxPool2DOp = ConvertMaxPoolOp<TF::MaxPoolOp, /*num_dims=*/4>;
+using ConvertMaxPool3DOp = ConvertMaxPoolOp<TF::MaxPool3DOp, /*num_dims=*/5>;
+
// Converts SelectV2 to HLO Select op and necessary BroadcastInDim ops on
// operands.
//
@@ -1995,11 +2147,16 @@
// negative strides and Reshape op to update the output shape. Indices and
// strides operands are converted to attributes with non-negative indexing.
//
+// If the begin input is not a compile time constant, the begin input needs to
+// be sliced and the slice needs to be lowered to xla_hlo.DynamicSlice. In this
+// case, strides must have a known value of 1 (otherwise we have insufficient
+// information to conform to XLA's op semantics).
+//
// For example with an op like following,
// tf.StridedSlice(%input, %begin, %end, %strides) {shrink_axis_mask = 1}
// : tensor<AxBxf32> -> tensor<Pxf32>
//
-// Output would be:
+// If the %begin input is constant, output would be:
// %reversed = "xla_hlo.Reverse" (%input) {dimensions = ...}
// %sliced = "xla_hlo.Slice" (%input)
// {start_indices = ..., limit_indices = ..., strides = ...}
@@ -2009,31 +2166,16 @@
public:
using OpRewritePattern::OpRewritePattern;
- LogicalResult matchAndRewrite(TF::StridedSliceOp op,
- PatternRewriter &rewriter) const override {
- // Input shape needs to be static to convert negative indices in TensorFlow
- // to absolute indices required by HLO.
- //
- // TODO(hinsu): Relax this constraint for ops without negative indices and
- // strides.
- auto input_ty = op.input().getType().dyn_cast<RankedTensorType>();
- if (!input_ty || !input_ty.hasStaticShape()) return failure();
- ArrayRef<int64_t> input_shape = input_ty.getShape();
-
- // Output shape needs to be static to apply 'new_axis_mask' or
- // 'shrink_axis_mask' by reshaping tensor after slice.
- //
- // TODO(hinsu): Relax this constraint for ops without the above masks.
- auto result_ty = op.getType().dyn_cast<RankedTensorType>();
- if (!result_ty || !result_ty.hasStaticShape()) return failure();
-
- SmallVector<int64_t, 4> begin_indices, end_indices, strides;
- if (!op.GetSlicedBoundRanges(&begin_indices, &end_indices, &strides))
- return failure();
-
+ LogicalResult rewriteWithConstantBegin(TF::StridedSliceOp op,
+ ArrayRef<int64_t> begin_indices,
+ ArrayRef<int64_t> end_indices,
+ ArrayRef<int64_t> strides,
+ RankedTensorType input_ty,
+ PatternRewriter &rewriter) const {
SmallVector<int64_t, 4> hlo_begin_indices, hlo_end_indices, hlo_strides,
dims_to_reverse;
int64_t input_rank = input_ty.getRank();
+ ArrayRef<int64_t> input_shape = input_ty.getShape();
hlo_begin_indices.reserve(input_rank);
hlo_end_indices.reserve(input_rank);
hlo_strides.reserve(input_rank);
@@ -2085,6 +2227,170 @@
rewriter.replaceOpWithNewOp<ReshapeOp>(op, op.getType(), sliced);
return success();
}
+
+ LogicalResult rewriteWithUnknownBegin(TF::StridedSliceOp op,
+ RankedTensorType input_ty,
+ RankedTensorType result_ty,
+ PatternRewriter &rewriter) const {
+ // If begin and end values are dynamic, we can only support this lowering
+ // if strides are a known value of 1.
+ DenseIntElementsAttr sparse_strides_attr;
+ if (!matchPattern(op.strides(), m_Constant(&sparse_strides_attr))) {
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that strides are known when begin/end values are dynamic");
+ }
+ SmallVector<int64_t, 4> strides;
+ int64_t stride_value;
+ for (const APInt &stride : sparse_strides_attr) {
+ if ((stride_value = stride.getSExtValue()) != 1) {
+ return rewriter.notifyMatchFailure(op,
+ "requires that strides are all 1 "
+ "when begin/end values are dynamic");
+ }
+ strides.push_back(stride_value);
+ }
+
+ ArrayRef<int64_t> input_shape = input_ty.getShape();
+ int last_dim = std::max(static_cast<int>(input_shape.size()) - 1, 0);
+
+ // When begin/end values are dynamic, we can only support shrinking a major
+ // axis. For instance, if there are 4 dims, we can support a
+ // shrink_axis_mask of 0001 (1), 0011 (3), 0111 (7), or 1111 (15), but no
+ // other.
+ bool shrink_axis_mask_ok = op.shrink_axis_mask().isMask();
+ if (!shrink_axis_mask_ok)
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that shrink_axis_mask, if set, refer to a major axis "
+ "dimension (when begin/end values are dynamic)");
+
+ // When begin/end values are dynamic, the ellipsis mask, if set, must refer
+ // to the last dimension.
+ int ellipsis_mask = op.ellipsis_mask().getZExtValue();
+ if (!(ellipsis_mask == 0 || ellipsis_mask == (1 << last_dim)))
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that ellipsis_mask, if set, refer to the last dimension of "
+ "input (when begin/end values are dynamic)");
+
+ APInt begin_mask = op.begin_mask();
+ if (!begin_mask.isNullValue())
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that begin_mask is either set to 0 or not set when "
+ "begin/end values are dynamic");
+ APInt end_mask = op.end_mask();
+ if (!end_mask.isNullValue())
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that end_mask is either set to 0 or not set when begin/end "
+ "values are dynamic");
+ APInt new_axis_mask = op.new_axis_mask();
+ if (!new_axis_mask.isNullValue())
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires that new_axis_mask is either set to 0 or not set when "
+ "begin/end values are dynamic");
+
+ // In this case where the begin and end values are dynamic, the number of
+ // output elements has to be equal to the number of input elements that
+ // are sliced.
+ int output_elements = result_ty.getNumElements();
+ int input_elements_sliced = 1;
+
+ // Begin must be a ranked, 1-dimensional tensor: This is checked by the
+ // verifier.
+ int64_t slicing_dim_size =
+ op.begin().getType().cast<RankedTensorType>().getShape()[0];
+ auto input_rank = input_shape.size();
+ for (int d = slicing_dim_size; d < input_rank; ++d) {
+ // We only support slicing major dimensions, so minor dimensions after
+ // slicing dimensions are all sliced with their full sizes.
+ input_elements_sliced *= input_shape[d];
+ }
+ if (input_elements_sliced != output_elements) {
+ return rewriter.notifyMatchFailure(
+ op,
+ "requires the number of output elements to be equal to the number of "
+ "input elements sliced (when begin/end values are dynamic)");
+ }
+
+ SmallVector<Value, 4> slice_begin_indices;
+ // For the dimensions that are to be sliced, all have slice sizes of 1.
+ SmallVector<int64_t, 4> slice_sizes(slicing_dim_size, 1);
+ auto input_element_ty = input_ty.getElementType();
+ // Scalar tensor type.
+ TensorType type = RankedTensorType::get(/*shape=*/{}, input_element_ty);
+ Location loc = op.getLoc();
+ auto zero = GetScalarConstOfType(input_element_ty, loc, 0, &rewriter);
+ for (int d = 0; d < slicing_dim_size; ++d) {
+ auto index = rewriter.create<SliceOp>(
+ loc, op.begin(), GetI64ElementsAttr({d}, &rewriter),
+ GetI64ElementsAttr({d + 1}, &rewriter),
+ GetI64ElementsAttr({1}, &rewriter));
+ // Convert index to scalar.
+ auto reshaped_index = rewriter.create<ReshapeOp>(loc, type, index);
+ // If the index is negative, wrap it around with dimension size.
+ auto index_negative =
+ rewriter.create<TF::LessOp>(loc, reshaped_index, zero);
+ auto input_val = GetScalarConstOfType(input_element_ty, loc,
+ input_shape[d], &rewriter);
+ auto wrapped_index =
+ rewriter.create<TF::AddOp>(loc, input_val, reshaped_index);
+ auto final_index = rewriter.create<SelectOp>(
+ loc, type, index_negative, wrapped_index, reshaped_index);
+ slice_begin_indices.push_back(final_index);
+ }
+
+ // For non-slice dims, get the full slice of that dimension.
+ for (int d = slicing_dim_size; d < input_shape.size(); ++d) {
+ slice_sizes.push_back(input_shape[d]);
+ slice_begin_indices.push_back(zero);
+ }
+
+ auto slice_sizes_attr = GetI64ElementsAttr(slice_sizes, &rewriter);
+ // This must be an xla DynamicSlice op due to the inputs that aren't
+ // constant.
+ auto sliced = rewriter.create<DynamicSliceOp>(
+ loc, op.getType(), op.input(), slice_begin_indices, slice_sizes_attr);
+
+ // Reshape slice result so that the shape is updated depending on
+ // 'new_axis_mask' or 'shrink_axis_mask' attributes.
+ rewriter.replaceOpWithNewOp<ReshapeOp>(op, op.getType(), sliced);
+ return success();
+ }
+
+ LogicalResult matchAndRewrite(TF::StridedSliceOp op,
+ PatternRewriter &rewriter) const override {
+ // Input shape needs to be static to convert negative indices in TensorFlow
+ // to absolute indices required by HLO.
+ //
+ // TODO(hinsu): Relax this constraint for ops without negative indices and
+ // strides.
+ auto input_ty = op.input().getType().dyn_cast<RankedTensorType>();
+ if (!input_ty || !input_ty.hasStaticShape()) return failure();
+
+ // Output shape needs to be static to apply 'new_axis_mask' or
+ // 'shrink_axis_mask' by reshaping tensor after slice.
+ //
+ // TODO(hinsu): Relax this constraint for ops without the above masks.
+ auto result_ty = op.getType().dyn_cast<RankedTensorType>();
+ if (!result_ty || !result_ty.hasStaticShape()) return failure();
+
+ DenseIntElementsAttr sparse_begin_attr, sparse_end_attr;
+ if (!matchPattern(op.begin(), m_Constant(&sparse_begin_attr)) ||
+ !matchPattern(op.end(), m_Constant(&sparse_end_attr))) {
+ return rewriteWithUnknownBegin(op, input_ty, result_ty, rewriter);
+ }
+
+ SmallVector<int64_t, 4> begin_indices, end_indices, strides;
+ if (!op.GetSlicedBoundRanges(&begin_indices, &end_indices, &strides)) {
+ return failure();
+ }
+ return rewriteWithConstantBegin(op, begin_indices, end_indices, strides,
+ input_ty, rewriter);
+ }
};
// Converts tf.StridedSliceGrad to HLO reshape, reverse and padding ops.
@@ -2684,23 +2990,25 @@
}
};
-class ConvertMaxPoolGradOp : public OpRewritePattern<TF::MaxPoolGradOp> {
+template <typename OpTy, int num_dims>
+class ConvertMaxPoolGradOp : public OpRewritePattern<OpTy> {
public:
- using OpRewritePattern::OpRewritePattern;
+ using OpRewritePattern<OpTy>::OpRewritePattern;
- LogicalResult matchAndRewrite(TF::MaxPoolGradOp op,
+ LogicalResult matchAndRewrite(OpTy op,
PatternRewriter &rewriter) const override {
Location loc = op.getLoc();
Type element_type =
- op.orig_input().getType().cast<TensorType>().getElementType();
+ op.orig_input().getType().template cast<TensorType>().getElementType();
// Compute paddings using the original input and kernel shape and strides.
// Here, ReduceWindow op as used as the MaxPool op is lowered to the
// ReduceWindow op.
- auto input_ty = op.orig_input().getType().dyn_cast<RankedTensorType>();
+ auto input_ty =
+ op.orig_input().getType().template dyn_cast<RankedTensorType>();
if (!input_ty) return failure();
- DenseIntElementsAttr paddings_attr = GetReduceWindowPadding(
+ DenseIntElementsAttr paddings_attr = GetReduceWindowPadding<num_dims>(
input_ty.getShape(), op.ksize(), op.strides(), op.padding(), &rewriter);
auto result = rewriter.create<SelectAndScatterOp>(
@@ -2731,6 +3039,11 @@
}
};
+using ConvertMaxPool2DGradOp =
+ ConvertMaxPoolGradOp<TF::MaxPoolGradOp, /*num_dims=*/4>;
+using ConvertMaxPool3DGradOp =
+ ConvertMaxPoolGradOp<TF::MaxPool3DGradOp, /*num_dims=*/5>;
+
// Converts tf.Conv?DBackpropInputOp into:
// %rev_filter = "xla_hlo.reverse"(%filter)
// %result = "xla_hlo.convolution"(%out_backprop, %rev_filter)
@@ -2763,8 +3076,7 @@
input_shape_attr.getType().getRank() != 1)
return failure();
- auto input_shape = llvm::to_vector<num_spatial_dims>(
- input_shape_attr.getValues<int32_t>());
+ auto input_shape = input_shape_attr.getValues<int32_t>();
auto dilations_attr = GetI64ElementsAttr(op.dilations());
std::vector<int> dilations{
@@ -2788,31 +3100,30 @@
explicit_padding.cast<IntegerAttr>().getInt());
}
+ constexpr int num_dims = num_spatial_dims + 2;
ArrayRef<int64_t> filter_shape = filter_ty.getShape();
+
// Reuse dimension computation logic from conv_grad_shape_utils.cc.
tensorflow::ConvBackpropDimensions dims;
if (!tensorflow::ConvBackpropComputeDimensionsV2(
/*label=*/"", num_spatial_dims,
- ToTensorShape<int32_t>(input_shape),
- ToTensorShape<int64_t>(filter_shape),
- ToTensorShape<int64_t>(out_backprop_ty.getShape()), dilations,
- strides, padding, explicit_paddings, data_format, &dims)
+ ToTensorShape<int32_t, num_dims>(input_shape),
+ ToTensorShape<int64_t, num_dims>(filter_shape),
+ ToTensorShape<int64_t, num_dims>(out_backprop_ty.getShape()),
+ dilations, strides, padding, explicit_paddings, data_format, &dims)
.ok()) {
return failure();
}
// Compute ConvDimensionNumbers, dilation, and padding.
SmallVector<int64_t, num_spatial_dims> spatial_dims;
- SmallVector<int64_t, num_spatial_dims> kernel_spatial_dims;
SmallVector<int64_t, num_spatial_dims> lhs_dilation;
SmallVector<int64_t, num_spatial_dims> rhs_dilation;
SmallVector<int64_t, num_spatial_dims * 2> paddings;
- const int num_dims = num_spatial_dims + 2;
for (int i : llvm::seq<int>(0, num_spatial_dims)) {
const int64_t dim = GetTensorSpatialDimIndex(num_dims, data_format, i);
spatial_dims.push_back(dim);
- kernel_spatial_dims.push_back(i);
const auto &spatial_dim_i = dims.spatial_dims[i];
lhs_dilation.push_back(spatial_dim_i.stride);
rhs_dilation.push_back(dilations[dim]);
@@ -2830,7 +3141,7 @@
const int feature_dim =
tensorflow::GetTensorFeatureDimIndex(num_dims, data_format);
- const int64_t in_depth = input_shape[feature_dim];
+ const int64_t in_depth = *(input_shape.begin() + feature_dim);
const int64_t filter_in_depth = filter_shape[num_spatial_dims];
const int64_t feature_group_count = in_depth / filter_in_depth;
@@ -2843,12 +3154,12 @@
return failure();
}
- // Mirror the filter in the spatial dimensions.
- filter = rewriter.create<ReverseOp>(
- op.getLoc(), filter,
- GetI64ElementsAttr(kernel_spatial_dims, &rewriter));
+ auto kernel_spatial_dims_attr =
+ GetI64ElementsAttrForSeq(0, num_spatial_dims, &rewriter);
- SmallVector<int64_t, num_spatial_dims> ones(num_spatial_dims, 1);
+ // Mirror the filter in the spatial dimensions.
+ filter = rewriter.create<ReverseOp>(op.getLoc(), filter,
+ kernel_spatial_dims_attr);
const int batch_dim =
tensorflow::GetTensorBatchDimIndex(num_dims, data_format);
@@ -2859,7 +3170,9 @@
// = gradients (with padding and dilation) <conv> mirrored_weights
Value result = rewriter.create<ConvOp>(
op.getLoc(), op.getType(), op.out_backprop(), filter,
- /*window_strides=*/GetI64ElementsAttr(ones, &rewriter),
+ /*window_strides=*/
+ GetI64ElementsAttrForValue(/*size=*/num_spatial_dims, /*val=*/1,
+ &rewriter),
/*padding=*/paddings_attr, GetI64ElementsAttr(lhs_dilation, &rewriter),
GetI64ElementsAttr(rhs_dilation, &rewriter),
ConvDimensionNumbers::get(
@@ -2873,8 +3186,7 @@
rewriter.getI64IntegerAttr(num_spatial_dims + 1),
/*kernel_output_feature_dimension=*/
rewriter.getI64IntegerAttr(num_spatial_dims),
- /*kernel_spatial_dimensions=*/
- GetI64ElementsAttr(kernel_spatial_dims, &rewriter),
+ /*kernel_spatial_dimensions=*/kernel_spatial_dims_attr,
/*output_batch_dimension=*/batch_dim_attr,
/*output_feature_dimension=*/feature_dim_attr,
/*output_spatial_dimensions=*/spatial_dims_attr,
@@ -2896,67 +3208,72 @@
ConvertConvBackpropInputOp<TF::Conv3DBackpropInputV2Op,
/*num_spatial_dims=*/3>;
-// Converts tf.Conv2DBackpropFilterOp into:
+// Converts tf.Conv?DBackpropFilterOp into:
// %result = "xla_hlo.convolution"(%input, %out_backprop)
-class ConvertConv2DBackpropFilterOp
- : public OpRewritePattern<TF::Conv2DBackpropFilterOp> {
+template <typename OpTy, int num_spatial_dims>
+class ConvertConvBackpropFilterOp : public OpRewritePattern<OpTy> {
public:
- using OpRewritePattern::OpRewritePattern;
+ using OpRewritePattern<OpTy>::OpRewritePattern;
- LogicalResult matchAndRewrite(TF::Conv2DBackpropFilterOp op,
+ LogicalResult matchAndRewrite(OpTy op,
PatternRewriter &rewriter) const override {
// Unpack all of the attributes.
tensorflow::TensorFormat data_format;
- if (!FormatFromString(op.data_format().str(), &data_format)) {
+ if (!FormatFromString(op.data_format().str(), &data_format))
return failure();
- }
+
tensorflow::Padding padding;
if (!GetPaddingFromString(op.padding().str(), &padding).ok())
return failure();
auto out_backprop_ty =
- op.out_backprop().getType().dyn_cast<RankedTensorType>();
- if (!out_backprop_ty || !out_backprop_ty.hasStaticShape()) return failure();
+ op.out_backprop().getType().template dyn_cast<RankedTensorType>();
+ auto input_ty = op.input().getType().template dyn_cast<RankedTensorType>();
+
+ for (RankedTensorType ty : {out_backprop_ty, input_ty})
+ if (!ty || !ty.hasStaticShape()) return failure();
+
ArrayRef<int64_t> out_backprop_shape = out_backprop_ty.getShape();
- auto input_ty = op.input().getType().dyn_cast<RankedTensorType>();
- if (!input_ty || !input_ty.hasStaticShape()) return failure();
ArrayRef<int64_t> input_shape = input_ty.getShape();
DenseIntElementsAttr filter_shape_attr;
if (!matchPattern(op.filter_sizes(), m_Constant(&filter_shape_attr)) ||
- filter_shape_attr.getType().getRank() != 1) {
+ filter_shape_attr.getType().getRank() != 1)
return failure();
- }
+ auto dilations_attr = GetI64ElementsAttr(op.dilations());
+ std::vector<int> dilations{
+ dilations_attr.template getValues<int64_t>().begin(),
+ dilations_attr.template getValues<int64_t>().end()};
auto strides_attr = GetI64ElementsAttr(op.strides());
std::vector<tensorflow::int32> strides{
- strides_attr.getValues<int64_t>().begin(),
- strides_attr.getValues<int64_t>().end()};
- auto dilations_attr = GetI64ElementsAttr(op.dilations());
- SmallVector<int, 4> dilations{dilations_attr.getValues<int64_t>().begin(),
- dilations_attr.getValues<int64_t>().end()};
- auto explicit_paddings_attr = GetI64ElementsAttr(op.explicit_paddings());
- SmallVector<tensorflow::int64, 4> explicit_paddings{
- explicit_paddings_attr.getValues<int64_t>().begin(),
- explicit_paddings_attr.getValues<int64_t>().end()};
+ strides_attr.template getValues<int64_t>().begin(),
+ strides_attr.template getValues<int64_t>().end()};
- int num_spatial_dims = 2;
- int num_dims = num_spatial_dims + 2;
- int batch_dim = tensorflow::GetTensorBatchDimIndex(num_dims, data_format);
- int feature_dim =
- tensorflow::GetTensorFeatureDimIndex(num_dims, data_format);
+ std::vector<tensorflow::int64> explicit_paddings;
+ if (padding == tensorflow::Padding::EXPLICIT) {
+ // EXPLICIT padding mode and the associated attribute is limited to
+ // Conv2DBackpropFilter. So, fetch attribute by identifier instead of the
+ // op.explicit_paddings() attribute getter.
+ ArrayRef<Attribute> explicit_paddings_attr =
+ op.template getAttrOfType<ArrayAttr>("explicit_paddings").getValue();
+ explicit_paddings.reserve(explicit_paddings_attr.size());
+ for (Attribute explicit_padding : explicit_paddings_attr)
+ explicit_paddings.push_back(
+ explicit_padding.cast<IntegerAttr>().getInt());
+ }
- auto filter_shape =
- llvm::to_vector<4>(filter_shape_attr.getValues<int32_t>());
- if (filter_shape.size() != num_dims) return failure();
+ constexpr int num_dims = num_spatial_dims + 2;
+ auto filter_shape = filter_shape_attr.getValues<int32_t>();
// Reuse dimension computation logic from conv_grad_shape_utils.cc.
tensorflow::ConvBackpropDimensions dims;
if (!tensorflow::ConvBackpropComputeDimensionsV2(
- "", num_spatial_dims, ToTensorShape<int64_t>(input_shape),
- ToTensorShape<int>(filter_shape),
- ToTensorShape<int64_t>(out_backprop_shape), dilations, strides,
- padding, explicit_paddings, data_format, &dims)
+ /*label=*/"", num_spatial_dims,
+ ToTensorShape<int64_t, num_dims>(input_shape),
+ ToTensorShape<int32_t, num_dims>(filter_shape),
+ ToTensorShape<int64_t, num_dims>(out_backprop_shape), dilations,
+ strides, padding, explicit_paddings, data_format, &dims)
.ok()) {
return failure();
}
@@ -2967,9 +3284,12 @@
// 1. In the case of group convolution, move the num_groups dimension before
// the batch dimension
// 2. Swap the roles of the batch and feature dimensions.
- int64_t in_depth = input_shape[feature_dim];
- int64_t filter_in_depth = filter_shape[num_spatial_dims];
- int64_t feature_group_count = in_depth / filter_in_depth;
+ const int feature_dim =
+ tensorflow::GetTensorFeatureDimIndex(num_dims, data_format);
+ const int64_t in_depth = input_shape[feature_dim];
+ const int64_t filter_in_depth = *(filter_shape.begin() + num_spatial_dims);
+ const int64_t feature_group_count = in_depth / filter_in_depth;
+
if (feature_group_count != 1) {
/*
// TODO(parkers): translate this code to mlir.
@@ -2981,21 +3301,20 @@
}
// Compute ConvDimensionNumbers, dilation, and padding.
- SmallVector<int64_t, 8> conv_padding(num_spatial_dims * 2);
- SmallVector<int64_t, 4> rhs_dilation(num_spatial_dims);
- SmallVector<int64_t, 4> window_strides(num_spatial_dims);
- SmallVector<int64_t, 4> lhs_dilation(num_spatial_dims, 1);
- SmallVector<int64_t, 4> spatial_dims(num_spatial_dims);
- SmallVector<int64_t, 4> kernel_spatial_dims(num_spatial_dims);
+ SmallVector<int64_t, num_spatial_dims> spatial_dims;
+ SmallVector<int64_t, num_spatial_dims> kernel_spatial_dims;
+ SmallVector<int64_t, num_spatial_dims> rhs_dilation;
+ SmallVector<int64_t, num_spatial_dims * 2> paddings;
+ SmallVector<int64_t, num_spatial_dims> window_strides;
// The filter gradients are computed by a convolution of the input
// activations and the output gradients, with some appropriate padding.
// See the comment at the top of conv_grad_ops.h for details.
- for (int64_t i = 0; i < num_spatial_dims; ++i) {
- int64_t dim =
+ for (int i : llvm::seq<int>(0, num_spatial_dims)) {
+ const int64_t dim =
tensorflow::GetTensorSpatialDimIndex(num_dims, data_format, i);
- kernel_spatial_dims[i] = dim;
+ kernel_spatial_dims.push_back(dim);
// Besides padding the input, we will also expand output_rows to
// expanded_out_rows = (output_rows - 1) * stride + 1
// with zeros in between:
@@ -3004,8 +3323,9 @@
//
// This is done by specifying the window dilation factors in the
// convolution HLO below.
- rhs_dilation[i] = dims.spatial_dims[i].stride;
- window_strides[i] = dilations[dim];
+ const auto &spatial_dim_i = dims.spatial_dims[i];
+ rhs_dilation.push_back(spatial_dim_i.stride);
+ window_strides.push_back(dilations[dim]);
// We will also need to pad the input with zeros such that after the
// convolution, we get the right size for the filter.
@@ -3013,8 +3333,8 @@
// expanded_out_rows as a filter, we should get filter_rows back.
const int64_t padded_in_size =
- dims.spatial_dims[i].expanded_output_size +
- (dims.spatial_dims[i].filter_size - 1) * dilations[dim];
+ spatial_dim_i.expanded_output_size +
+ (spatial_dim_i.filter_size - 1) * dilations[dim];
// However it can be smaller than input_rows: in this
// case it means some of the inputs are not used.
@@ -3030,8 +3350,7 @@
// and input "C" is not used at all.
//
// We apply negative padding in this case.
- const int64_t pad_total =
- padded_in_size - dims.spatial_dims[i].input_size;
+ const int64_t pad_total = padded_in_size - spatial_dim_i.input_size;
// + For the EXPLICIT padding, we pad the top/left side with the explicit
// padding and pad the bottom/right side with the remaining space.
@@ -3048,26 +3367,27 @@
: padding == tensorflow::Padding::SAME
? std::max<int64_t>(pad_total / 2, 0)
: 0;
- conv_padding[i * 2] = pad_before;
- conv_padding[i * 2 + 1] = pad_total - pad_before;
+ paddings.push_back(pad_before);
+ paddings.push_back(pad_total - pad_before);
}
RankedTensorType paddings_ty = RankedTensorType::get(
{num_spatial_dims, 2}, rewriter.getIntegerType(64));
- auto paddings_attr = DenseIntElementsAttr::get(paddings_ty, conv_padding);
- auto out_spatial_dims_attr =
- GetI64ElementsAttrForSeq(0, num_spatial_dims, &rewriter);
+ auto paddings_attr = DenseIntElementsAttr::get(paddings_ty, paddings);
auto kernel_spatial_dims_attr =
GetI64ElementsAttr(kernel_spatial_dims, &rewriter);
+ const int batch_dim =
+ tensorflow::GetTensorBatchDimIndex(num_dims, data_format);
auto batch_dim_attr = rewriter.getI64IntegerAttr(batch_dim);
auto feature_dim_attr = rewriter.getI64IntegerAttr(feature_dim);
- Location loc = op.getLoc();
Value result = rewriter.create<ConvOp>(
- loc, op.getType(), op.input(), op.out_backprop(),
+ op.getLoc(), op.getType(), op.input(), op.out_backprop(),
/*window_strides=*/GetI64ElementsAttr(window_strides, &rewriter),
- /*padding=*/paddings_attr, GetI64ElementsAttr(lhs_dilation, &rewriter),
+ /*padding=*/paddings_attr, /*lhs_dilation=*/
+ GetI64ElementsAttrForValue(/*size=*/num_spatial_dims, /*val=*/1,
+ &rewriter),
GetI64ElementsAttr(rhs_dilation, &rewriter),
ConvDimensionNumbers::get(
// Swap batch_dim and feature_dim in the activations.
@@ -3085,7 +3405,8 @@
rewriter.getI64IntegerAttr(num_spatial_dims),
/*output_feature_dimension=*/
rewriter.getI64IntegerAttr(num_spatial_dims + 1),
- /*output_spatial_dimensions=*/out_spatial_dims_attr,
+ /*output_spatial_dimensions=*/
+ GetI64ElementsAttrForSeq(0, num_spatial_dims, &rewriter),
rewriter.getContext()),
rewriter.getI64IntegerAttr(feature_group_count),
/*batch_group_count=*/rewriter.getI64IntegerAttr(1),
@@ -3097,6 +3418,13 @@
}
};
+using ConvertConv2DBackpropFilterOp =
+ ConvertConvBackpropFilterOp<TF::Conv2DBackpropFilterOp,
+ /*num_spatial_dims=*/2>;
+using ConvertConv3DBackpropFilterOp =
+ ConvertConvBackpropFilterOp<TF::Conv3DBackpropFilterV2Op,
+ /*num_spatial_dims=*/3>;
+
class ConvertOneHotOp : public OpRewritePattern<TF::OneHotOp> {
public:
using OpRewritePattern::OpRewritePattern;
@@ -3198,6 +3526,27 @@
auto data_and_token =
rewriter.create<InfeedOp>(op.getLoc(), data_and_token_type, token,
/*infeed_config=*/rewriter.getStringAttr(""));
+ if (op._XlaSharding().hasValue()) {
+ // _XlaSharding attribute in TF is a serialized string of the OpSharding
+ // proto, so convert to a text form here.
+ ::xla::OpSharding sharding_proto;
+ if (!sharding_proto.ParseFromString(op._XlaSharding().getValue().str()))
+ return failure();
+
+ // Token is a control signal and not a real data, so arbitrarily assign
+ // the token to device 0.
+ if (sharding_proto.type() == ::xla::OpSharding::TUPLE)
+ *sharding_proto.add_tuple_shardings() =
+ ::xla::sharding_builder::AssignDevice(0);
+
+ std::string sharding_str;
+ if (!::tensorflow::protobuf::TextFormat::PrintToString(sharding_proto,
+ &sharding_str))
+ return failure();
+
+ data_and_token.setAttr(kShardingAttr,
+ rewriter.getStringAttr(sharding_str));
+ }
// The infeed instruction produces a tuple of the infeed data and a token
// type. Emit get_tuple_element to get infeed data tuple.
@@ -3737,30 +4086,23 @@
PatternRewriter &rewriter) const override {
// TODO(b/148313088): define sharding attribute struct in MLIR intead of
// using a string.
- auto sharding = op.getAttrOfType<StringAttr>("_XlaSharding");
- if (!sharding) {
- return failure();
- }
+ if (!op._XlaSharding().hasValue()) return failure();
// _XlaSharding attribute in TF is a serialized string of the OpSharding
// proto, so convert to a text form here.
::xla::OpSharding sharding_proto;
std::string sharding_str;
- if (!sharding_proto.ParseFromString(sharding.getValue().str())) {
+ if (!sharding_proto.ParseFromString(op._XlaSharding().getValue().str()) ||
+ !::tensorflow::protobuf::TextFormat::PrintToString(sharding_proto,
+ &sharding_str))
return failure();
- }
- if (!::tensorflow::protobuf::TextFormat::PrintToString(sharding_proto,
- &sharding_str)) {
- return failure();
- }
auto custom_call = rewriter.create<xla_hlo::CustomCallOp>(
op.getLoc(), op.getType(), op.input(),
/*call_target_name=*/rewriter.getStringAttr("Sharding"),
/*has_side_effect=*/rewriter.getBoolAttr(false),
/*backend_config=*/rewriter.getStringAttr(""));
- custom_call.setAttr("xla_hlo.sharding",
- rewriter.getStringAttr(sharding_str));
+ custom_call.setAttr(kShardingAttr, rewriter.getStringAttr(sharding_str));
rewriter.replaceOp(op, custom_call.getResult());
return success();
@@ -3866,6 +4208,511 @@
}
};
+// Converts a TF QR op to HLO.
+class ConvertQrOp : public OpRewritePattern<TF::QrOp> {
+ public:
+ using OpRewritePattern::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(TF::QrOp op,
+ PatternRewriter &rewriter) const override {
+ // Block Householder QR Factorization. Algorithm 5.2.2 of Golub and van
+ // Loan. def qr_blocked(a, block_size):
+ // m = a.shape[0]
+ // n = a.shape[1]
+ // q = np.eye(m)
+ // for i in xrange(0, min(m, n), block_size):
+ // k = min(block_size, min(m, n) - s)
+ // (a, vs, taus) = qr(a[i:, i:i+k])
+ // y = vs
+ // w = ComputeWYRepresentation(vs, taus, m-i, k)
+ // a[i:, i+r:] += np.dot(y, np.dot(w.T, a[i:, i+k:]))
+ // q[:, i:] += np.dot(q[:, i:], np.dot(w, y.T))
+ // return (q, a)
+ auto type = op.input().getType().dyn_cast<RankedTensorType>();
+ if (!type || !type.hasStaticShape()) return failure();
+ // The block size is chosen to match old bridge lowering.
+ constexpr int64_t kBlockSize = 128;
+ Value a = op.input();
+ int64_t m = type.getDimSize(type.getRank() - 2);
+ int64_t n = type.getDimSize(type.getRank() - 1);
+ int64_t p = std::min(m, n);
+ auto batch_dims = type.getShape().drop_back(2);
+ auto iota_type = RankedTensorType::get({m, m}, rewriter.getIntegerType(32));
+ auto iota0 = rewriter.create<IotaOp>(op.getLoc(), iota_type,
+ rewriter.getI64IntegerAttr(0));
+ auto iota1 = rewriter.create<IotaOp>(op.getLoc(), iota_type,
+ rewriter.getI64IntegerAttr(1));
+ Value compare = rewriter.create<CompareOp>(
+ op.getLoc(), iota0, iota1,
+ /*broadcast_dimensions=*/nullptr,
+ StringAttr::get("EQ", rewriter.getContext()));
+ Value identity_matrix =
+ rewriter.create<ConvertOp>(op.getLoc(), compare, type.getElementType());
+ auto q_shape = llvm::to_vector<4>(type.getShape());
+ q_shape.back() = m;
+ Value q = rewriter.create<BroadcastOp>(
+ op.getLoc(), RankedTensorType::get(q_shape, type.getElementType()),
+ identity_matrix, GetI64ElementsAttr(batch_dims, &rewriter));
+ auto precision_config = rewriter.getStrArrayAttr({"HIGHEST", "HIGHEST"});
+ for (int64_t i = 0; i < p; i += kBlockSize) {
+ int64_t k = std::min(kBlockSize, p - i);
+ auto a_block =
+ SliceInMinorDims(op.getLoc(), a, {i, i}, {m, i + k}, &rewriter);
+ Value r_block;
+ Value taus;
+ Value vs;
+ QRBlock(op.getLoc(), a_block, &r_block, &taus, &vs, &rewriter);
+ a = UpdateSliceInMinorDims(op.getLoc(), a, r_block, {i, i}, &rewriter);
+
+ // Compute the I-WY block representation of a product of Householder
+ // matrices.
+ Value w =
+ ComputeWYRepresentation(op.getLoc(), type.getElementType(),
+ batch_dims, vs, taus, m - i, k, &rewriter);
+ auto y = vs;
+
+ // a[i:, i+k:] += np.dot(Y, np.dot(W.T, a[i:, i+k:]))
+ Value a_panel =
+ SliceInMinorDims(op.getLoc(), a, {i, i + k}, {m, n}, &rewriter);
+ auto a_update = BatchDot(op.getLoc(), w, true, a_panel, false,
+ batch_dims.size(), precision_config, &rewriter);
+ a_update = BatchDot(op.getLoc(), y, false, a_update, false,
+ batch_dims.size(), precision_config, &rewriter);
+ a_panel = rewriter.create<AddOp>(op.getLoc(), a_panel, a_update,
+ /*broadcast_dimensions=*/nullptr);
+ a = UpdateSliceInMinorDims(op.getLoc(), a, a_panel, {i, i + k},
+ &rewriter);
+
+ // q[:, i:] += np.dot(np.dot(q[:, i:], W), Y.T))
+ Value q_panel =
+ SliceInMinorDims(op.getLoc(), q, {0, i}, {m, m}, &rewriter);
+ Value q_update = BatchDot(op.getLoc(), q_panel, false, w, false,
+ batch_dims.size(), precision_config, &rewriter);
+ q_update = BatchDot(op.getLoc(), q_update, false, y, true,
+ batch_dims.size(), precision_config, &rewriter);
+ q_panel = rewriter.create<AddOp>(op.getLoc(), q_panel, q_update,
+ /*broadcast_dimensions=*/nullptr);
+ q = UpdateSliceInMinorDims(op.getLoc(), q, q_panel, {i}, &rewriter);
+ }
+ // full_matrices is false when only a partial result in needed. Slice to the
+ // needed dimensions here.
+ if (!op.full_matrices()) {
+ q = SliceInMinorDims(op.getLoc(), q, {0, 0}, {m, p}, &rewriter);
+ a = SliceInMinorDims(op.getLoc(), a, {0, 0}, {p, n}, &rewriter);
+ }
+ rewriter.replaceOp(op, {q, a});
+ return success();
+ }
+
+ private:
+ // Computes a Householder reflection of the form:
+ // H = I - tau v v.T.
+ // such that
+ // H . ( x1 ) = ( x1 )
+ // ( x2 ) = ( x2 )
+ // ( ... ) = ( ... )
+ // ( xk ) = ( beta )
+ // ( ... ) ( 0 )
+ // ( ... ) ( 0 )
+ // Unlike the usual formulation, we allow the caller to supply 'k' rather than
+ // only providing the relevant part of 'x' to maintain XLA's static shape
+ // invariant. In addition, the implementation supports batching.
+ // Pseudo-code, without batching:
+ // alpha = x[k]
+ // x_copy = np.copy(x)
+ // x_copy[:k+1] = 0
+ // xnorm = norm2(x_copy)
+ // if xnorm == 0:
+ // beta = alpha
+ // tau = 0
+ // v = np.zeros_like(x)
+ // else:
+ // beta = - np.sign(alpha) * dlapy2(alpha, xnorm)
+ // tau = (beta - alpha) / beta
+ // v = x / (alpha - beta)
+ // v[k] = 1
+ // return (v, tau, beta)
+ void House(Location loc, Value x, Value k, ArrayRef<int64_t> batch_dims,
+ const int64_t m, OpBuilder *builder, Value *v, Value *tau,
+ Value *beta) const {
+ auto x_type = x.getType().cast<RankedTensorType>();
+
+ llvm::SmallVector<int64_t, 4> batch_dim_ids(batch_dims.size());
+ std::iota(batch_dim_ids.begin(), batch_dim_ids.end(), 0);
+ const int64_t minor_dim = batch_dims.size();
+
+ Value zero = GetScalarConstOfType(x_type.getElementType(), loc, 0, builder);
+ Value one = GetScalarConstOfType(x_type.getElementType(), loc, 1, builder);
+
+ // alpha = x[k]
+ Value alpha = DynamicSliceInMinorDims(loc, x, {k}, {1}, builder);
+ alpha = builder->create<ReshapeOp>(
+ loc, RankedTensorType::get(batch_dims, x_type.getElementType()), alpha);
+
+ // Compute x[k+1:] (padded with zeros in elements 0..k)
+ Value iota = builder->create<IotaOp>(
+ loc, RankedTensorType::get({m}, builder->getIntegerType(32)),
+ builder->getI64IntegerAttr(0));
+ Value gtk = builder->create<CompareOp>(
+ loc, iota, k, GetI64ElementsAttr({}, builder),
+ StringAttr::get("GT", builder->getContext()));
+ gtk = builder->create<ConvertOp>(loc, gtk, x_type.getElementType());
+ Value x_after_k = builder->create<MulOp>(
+ loc, x, gtk, GetI64ElementsAttr({minor_dim}, builder));
+ Value x_after_k_sq = builder->create<MulOp>(
+ loc, x_after_k, x_after_k, /*broadcast_dimensions=*/nullptr);
+ // sigma = np.dot(x[k+1:], x[k+1:])
+ auto sigma = builder->create<ReduceOp>(
+ loc, x_after_k_sq, zero, GetI64ElementsAttr({minor_dim}, builder));
+ BuildReduceBody<AddOp>(x_type.getElementType(), &sigma.body(), builder);
+ // mu = np.sqrt(x[k]*x[k] + sigma)
+ Value alpha_sq = builder->create<MulOp>(loc, alpha, alpha,
+ /*broadcast_dimensions=*/nullptr);
+ Value mu = builder->create<SqrtOp>(
+ loc, builder->create<AddOp>(loc, alpha_sq, sigma.getResult(0),
+ /*broadcast_dimensions=*/nullptr));
+
+ Value sigma_is_zero = builder->create<CompareOp>(
+ loc, sigma.getResult(0), zero, GetI64ElementsAttr({}, builder),
+ StringAttr::get("EQ", builder->getContext()));
+ Value alpha_is_negative = builder->create<CompareOp>(
+ loc, alpha, zero, GetI64ElementsAttr({}, builder),
+ StringAttr::get("LT", builder->getContext()));
+ auto batch_size_one = builder->create<BroadcastOp>(
+ loc, alpha.getType(), one, GetI64ElementsAttr(batch_dims, builder));
+ Value signed_mu = builder->create<MulOp>(
+ loc,
+ builder->create<SelectOp>(loc, mu.getType(), alpha_is_negative,
+ batch_size_one,
+ builder->create<NegOp>(loc, batch_size_one)),
+ mu, GetI64ElementsAttr({}, builder));
+ *beta = builder->create<SelectOp>(loc, alpha.getType(), sigma_is_zero,
+ alpha, signed_mu);
+ *tau = builder->create<DivOp>(
+ loc,
+ builder->create<SubOp>(loc, *beta, alpha,
+ /*broadcast_dimensions=*/nullptr),
+ *beta,
+ /*broadcast_dimensions=*/nullptr);
+ Value zero_tau = builder->create<BroadcastOp>(
+ loc, alpha.getType(), zero, GetI64ElementsAttr(batch_dims, builder));
+ *tau = builder->create<SelectOp>(loc, alpha.getType(), sigma_is_zero,
+ zero_tau, *tau);
+ Value divisor = builder->create<SubOp>(loc, alpha, *beta,
+ /*broadcast_dimensions=*/nullptr);
+ divisor = builder->create<SelectOp>(loc, divisor.getType(), sigma_is_zero,
+ batch_size_one, divisor);
+
+ Value eqk = builder->create<CompareOp>(
+ loc, iota, k, GetI64ElementsAttr({}, builder),
+ StringAttr::get("EQ", builder->getContext()));
+ eqk = builder->create<ConvertOp>(loc, eqk, x_type.getElementType());
+ llvm::SmallVector<int64_t, 4> e_k_shape(batch_dims.size(), 1);
+ e_k_shape.push_back(m);
+ auto e_k = builder->create<BroadcastOp>(
+ loc, RankedTensorType::get(e_k_shape, x_type.getElementType()), eqk,
+ GetI64ElementsAttr(llvm::SmallVector<int64_t, 4>(batch_dims.size(), 1),
+ builder));
+
+ // Form v as [0, 0, ..., 1] ++ x[k+1:] / divisor
+ // If sigma is zero, x[k+1:] is zero, so use any non-zero divisor.
+ *v = builder->create<AddOp>(
+ loc, e_k,
+ builder->create<DivOp>(loc, x_after_k, divisor,
+ GetI64ElementsAttr(batch_dim_ids, builder)),
+ /*broadcast_dimensions=*/nullptr);
+ }
+
+ // Householder QR decomposition. Algorithm 5.2.1 from Golub and Van
+ // Loan "Matrix Computations", 4th Edition. This is an unblocked
+ // implementation used as an inner routine of the blocked implementation.
+ // Algorithm is adapted slightly so the shapes inside the loop are static, at
+ // the cost of some redundant computation. Since this is used as an inner
+ // block kernel, accumulates the Householder transformations (vs, taus) rather
+ // than the matrix q. Equivalent Python code, without batching: def qr(a):
+ // m = a.shape[0]
+ // n = a.shape[1]
+ // vs = np.zeros([m, n])
+ // taus = np.zeros([n])
+ // for j in xrange(min(m, n)):
+ // v, tau, beta = house(a[:, j], j)
+ // # Unusually, we apply the Householder transformation to the entirety of
+ // # a, wasting FLOPs to maintain the static shape invariant that XLA
+ // # requires. For columns that precede j this has no effect.
+ // a[:, :] -= tau * np.dot(v[:, np.newaxis],
+ // np.dot(v[np.newaxis, :], a[:, :]))
+ // # Form column j explicitly rather than relying on the precision of the
+ // # Householder update.
+ // a[j, j] = beta
+ // a[j+1:, j] = np.zeros([m - j - 1], dtype=a.dtype)
+ // vs[:, j] = v
+ // taus[j] = tau
+ // return (q, vs, taus)
+ void QRBlock(Location loc, Value a, Value *r, Value *taus, Value *vs,
+ PatternRewriter *rewriter) const {
+ auto a_type = a.getType().cast<RankedTensorType>();
+ const int num_dims = a_type.getRank();
+ assert(num_dims >= 2 && "Argument to QR must have rank >= 2");
+
+ const int64_t m = a_type.getDimSize(a_type.getRank() - 2);
+ const int64_t n = a_type.getDimSize(a_type.getRank() - 1);
+
+ const int64_t num_batch_dims = num_dims - 2;
+ auto batch_dims = a_type.getShape().take_front(num_batch_dims);
+ llvm::SmallVector<int64_t, 4> batch_dim_indices(batch_dims.size());
+ std::iota(batch_dim_indices.begin(), batch_dim_indices.end(), 0);
+
+ auto qr_body_fn = [&](Location loc, Value j, ArrayRef<Value> old_values,
+ SmallVectorImpl<Value> *new_values,
+ OpBuilder *builder) {
+ auto a = old_values[0];
+ auto vs = old_values[1];
+ auto taus = old_values[2];
+
+ // v, beta = house(a[:, j], j)
+ auto x = DynamicSliceInMinorDims(loc, a, {j}, {1}, builder);
+ auto x_collapsed_shape = llvm::to_vector<4>(batch_dims);
+ x_collapsed_shape.push_back(m);
+ auto x_collapsed = builder->create<ReshapeOp>(
+ loc,
+ RankedTensorType::get(x_collapsed_shape,
+ getElementTypeOrSelf(x.getType())),
+ x);
+ Value v, tau, beta;
+ House(loc, x_collapsed, j, batch_dims, m, builder, &v, &tau, &beta);
+
+ auto shape = llvm::to_vector<4>(batch_dims);
+ shape.append({1, m});
+ auto v_broadcast = builder->create<ReshapeOp>(
+ loc, RankedTensorType::get(shape, getElementTypeOrSelf(v.getType())),
+ v);
+ // a[:, :] -= tau * np.dot(v[:, np.newaxis],
+ // np.dot(v[np.newaxis, :], a[:, :]))
+ auto precision = builder->getStrArrayAttr({"HIGHEST", "HIGHEST"});
+ auto vva = BatchDot(loc, v_broadcast, false, a, false, num_batch_dims,
+ precision, builder);
+ vva = BatchDot(loc, v_broadcast, true, vva, false, num_batch_dims,
+ precision, builder);
+ auto tau_x_vva = builder->create<MulOp>(
+ loc, tau, vva, GetI64ElementsAttr(batch_dim_indices, builder));
+ a = builder->create<SubOp>(loc, a, tau_x_vva,
+ /*broadcast_dimensions=*/nullptr);
+
+ // It is more precise to populate column 'k' explicitly, rather than
+ // computing it implicitly by applying the Householder transformation.
+ // a[k,k] = beta
+ // a[k+1:,k] = np.zeros([m-k-1], dtype=a.dtype)
+ auto iota = builder->create<IotaOp>(
+ loc, RankedTensorType::get({m, 1}, builder->getIntegerType(32)),
+ builder->getI64IntegerAttr(0));
+ Value predecessor_mask = builder->create<CompareOp>(
+ loc, iota, j, GetI64ElementsAttr({}, builder),
+ StringAttr::get("LT", builder->getContext()));
+ predecessor_mask = builder->create<ConvertOp>(loc, predecessor_mask,
+ a_type.getElementType());
+ Value mask = builder->create<CompareOp>(
+ loc, iota, j, GetI64ElementsAttr({}, builder),
+ StringAttr::get("EQ", builder->getContext()));
+ mask = builder->create<ConvertOp>(loc, mask, a_type.getElementType());
+ llvm::SmallVector<int64_t, 4> broadcast_mask_shape(a_type.getRank(), 1);
+ broadcast_mask_shape[a_type.getRank() - 2] = m;
+ mask = builder->create<BroadcastOp>(
+ loc,
+ RankedTensorType::get(broadcast_mask_shape, a_type.getElementType()),
+ mask,
+ GetI64ElementsAttr(llvm::SmallVector<int64_t, 4>(num_batch_dims, 1),
+ builder));
+ Value predecessor_masked_x = builder->create<MulOp>(
+ loc, x, predecessor_mask,
+ GetI64ElementsAttr({num_dims - 2, num_dims - 1}, builder));
+ Value masked_beta = builder->create<MulOp>(
+ loc, beta, mask, GetI64ElementsAttr(batch_dim_indices, builder));
+ Value new_x =
+ builder->create<AddOp>(loc, predecessor_masked_x, masked_beta,
+ /*broadcast_dimensions=*/nullptr);
+ // Update a[:,j]
+ llvm::SmallVector<int64_t, 4> dim_ids(num_dims);
+ std::iota(dim_ids.begin(), dim_ids.end(), 0);
+ new_x = builder->create<BroadcastInDimOp>(
+ loc, a_type, new_x, GetI64ElementsAttr(dim_ids, builder));
+ const int64_t minor_dim = num_batch_dims;
+ auto iota_mn = builder->create<IotaOp>(
+ loc,
+ RankedTensorType::get(a_type.getShape(), builder->getIntegerType(32)),
+ builder->getI64IntegerAttr(minor_dim + 1));
+ Value xa_mask = builder->create<CompareOp>(
+ loc, iota_mn, j, GetI64ElementsAttr({}, builder),
+ StringAttr::get("EQ", builder->getContext()));
+ a = builder->create<SelectOp>(loc, a_type, xa_mask, new_x, a);
+
+ // vs[:, j] = v
+ llvm::SmallVector<int64_t, 4> vs_broadcast_dims(num_batch_dims + 1);
+ std::iota(vs_broadcast_dims.begin(), vs_broadcast_dims.end(), 0);
+ Value vs_zeros =
+ GetScalarConstOfType(a_type.getElementType(), loc, 0, builder);
+ vs_zeros = builder->create<BroadcastOp>(
+ loc, vs.getType(), vs_zeros,
+ GetI64ElementsAttr(vs.getType().cast<RankedTensorType>().getShape(),
+ builder));
+ auto vs_update = builder->create<SelectOp>(
+ loc, vs.getType(), xa_mask,
+ builder->create<AddOp>(
+ loc, vs_zeros, v, GetI64ElementsAttr(vs_broadcast_dims, builder)),
+ vs_zeros);
+ vs = builder->create<AddOp>(loc, vs, vs_update,
+ /*broadcast_dimensions=*/nullptr);
+
+ // taus[j] = tau
+ llvm::SmallVector<int64_t, 4> tau_broadcast_dims(batch_dims.size());
+ std::iota(tau_broadcast_dims.begin(), tau_broadcast_dims.end(), 0);
+
+ auto iota_shape = llvm::to_vector<4>(batch_dims);
+ iota_shape.push_back(n);
+ auto iota_n = builder->create<IotaOp>(
+ loc, RankedTensorType::get(iota_shape, builder->getIntegerType(32)),
+ builder->getI64IntegerAttr(minor_dim));
+ Value taus_zeros =
+ GetScalarConstOfType(a_type.getElementType(), loc, 0, builder);
+ taus_zeros = builder->create<BroadcastOp>(
+ loc, taus.getType(), taus_zeros,
+ GetI64ElementsAttr(taus.getType().cast<RankedTensorType>().getShape(),
+ builder));
+ Value taus_mask = builder->create<CompareOp>(
+ loc, iota_n, j, GetI64ElementsAttr({}, builder),
+ StringAttr::get("EQ", builder->getContext()));
+ auto taus_update = builder->create<SelectOp>(
+ loc, taus.getType(), taus_mask,
+ builder->create<AddOp>(
+ loc, taus_zeros, tau,
+ GetI64ElementsAttr(tau_broadcast_dims, builder)),
+ taus_zeros);
+ taus = builder->create<AddOp>(loc, taus, taus_update,
+ /*broadcast_dimensions=*/nullptr);
+ new_values->assign({a, vs, taus});
+ };
+
+ Value zero =
+ GetScalarConstOfType(a_type.getElementType(), loc, 0, rewriter);
+ *vs = rewriter->create<BroadcastOp>(
+ loc, a_type, zero, GetI64ElementsAttr(a_type.getShape(), rewriter));
+ auto taus_shape = llvm::to_vector<4>(batch_dims);
+ taus_shape.push_back(n);
+ *taus = rewriter->create<BroadcastOp>(
+ loc, RankedTensorType::get(taus_shape, a_type.getElementType()), zero,
+ GetI64ElementsAttr(taus_shape, rewriter));
+
+ SmallVector<Value, 4> while_output;
+ CreateWhile32(loc, std::min(m, n), qr_body_fn, {a, *vs, *taus},
+ &while_output, rewriter);
+ *r = while_output[0];
+ *vs = while_output[1];
+ *taus = while_output[2];
+ }
+
+ // Computes W and Y such that I-WY is equivalent to the sequence of
+ // Householder
+ // transformations given by vs and taus.
+ // Golub and van Loan, "Matrix Computations", algorithm 5.1.2.
+ // Y = np.zeros([m, n])
+ // W = np.zeros([m, n])
+ // Y[:, 0] = vs[:, 0]
+ // W[:, 0] = -taus[0] * vs[:, 0]
+ // for j in xrange(1, n):
+ // v = vs[:, j]
+ // z = -taus[j] * v - taus[j] * np.dot(W, np.dot(Y.T, v))
+ // W[:, j] = z
+ // Y[:, j] = v
+ // return W
+ // There is no need to return Y since at termination of the loop it is equal
+ // to vs.
+ Value ComputeWYRepresentation(Location loc, Type data_type,
+ ArrayRef<int64_t> batch_dims, Value vs,
+ Value taus, int64_t m, int64_t n,
+ PatternRewriter *rewriter) const {
+ int64_t n_index = batch_dims.size() + 1;
+ llvm::SmallVector<int64_t, 4> batch_dim_indices(batch_dims.size());
+ std::iota(batch_dim_indices.begin(), batch_dim_indices.end(), 0);
+
+ auto body_fn = [&](Location loc, Value j, ArrayRef<Value> old_values,
+ SmallVectorImpl<Value> *new_values, OpBuilder *builder) {
+ // w has shape [..., m, n]
+ auto w = old_values[0];
+ const auto vs = old_values[1];
+ const auto taus = old_values[2];
+
+ // Want j values in range [1, ... n).
+ j = builder->create<AddOp>(
+ loc, j,
+ GetScalarConstOfType(getElementTypeOrSelf(j.getType()), loc, 1,
+ builder),
+ /*broadcast_dimensions=*/nullptr);
+ // vs has shape [..., m, 1]
+ auto v = DynamicSliceInMinorDims(loc, vs, {j}, {1}, builder);
+ // beta has shape [..., 1]
+ auto beta = DynamicSliceInMinorDims(loc, taus, {j}, {1}, builder);
+
+ auto iota_shape = llvm::to_vector<4>(batch_dims);
+ iota_shape.append({m, n});
+ auto iota_mn = builder->create<IotaOp>(
+ loc, RankedTensorType::get(iota_shape, builder->getIntegerType(32)),
+ builder->getI64IntegerAttr(n_index));
+
+ // y has shape [..., m, n]
+ Value zero = GetScalarConstOfType(getElementTypeOrSelf(vs.getType()), loc,
+ 0, builder);
+ zero = builder->create<BroadcastOp>(
+ loc, vs.getType(), zero,
+ GetI64ElementsAttr(vs.getType().cast<RankedTensorType>().getShape(),
+ builder));
+ auto compare = builder->create<CompareOp>(
+ loc, iota_mn, j, GetI64ElementsAttr({}, builder),
+ StringAttr::get("GE", builder->getContext()));
+ auto y = builder->create<SelectOp>(loc, vs.getType(), compare, zero, vs);
+
+ // yv has shape [..., n, 1]
+ auto precision = builder->getStrArrayAttr({"HIGHEST", "HIGHEST"});
+ auto yv = BatchDot(loc, y, true, v, false, batch_dims.size(), precision,
+ builder);
+ // wyv has shape [..., m, 1]
+ auto wyv = BatchDot(loc, w, false, yv, false, batch_dims.size(),
+ precision, builder);
+
+ // z = -beta * (v + wyv)
+ auto neg_beta = builder->create<NegOp>(loc, beta);
+ auto v_wyv = builder->create<AddOp>(loc, v, wyv,
+ /*broadcast_dimensions=*/nullptr);
+ auto beta_broadcast_dims = llvm::to_vector<4>(batch_dim_indices);
+ beta_broadcast_dims.push_back(n_index);
+ auto z = builder->create<MulOp>(
+ loc, neg_beta, v_wyv,
+ GetI64ElementsAttr(beta_broadcast_dims, builder));
+
+ w = DynamicUpdateSliceInMinorDims(loc, w, z, {j}, builder);
+ new_values->assign({w, vs, taus});
+ };
+
+ Value w =
+ GetScalarConstOfType(getElementTypeOrSelf(data_type), loc, 0, rewriter);
+ auto w_shape = llvm::to_vector<4>(batch_dims);
+ w_shape.append({m, n});
+ w = rewriter->create<BroadcastOp>(loc,
+ RankedTensorType::get(w_shape, data_type),
+ w, GetI64ElementsAttr(w_shape, rewriter));
+ auto v = SliceInMinorDims(loc, vs, {0}, {1}, rewriter);
+ auto beta = SliceInMinorDims(loc, taus, {0}, {1}, rewriter);
+ auto neg_beta = rewriter->create<NegOp>(loc, beta);
+ auto beta_broadcast_dims = llvm::to_vector<4>(batch_dim_indices);
+ beta_broadcast_dims.push_back(n_index);
+ auto bv = rewriter->create<MulOp>(
+ loc, neg_beta, v, GetI64ElementsAttr(beta_broadcast_dims, rewriter));
+ w = UpdateSliceInMinorDims(loc, w, bv, {0}, rewriter);
+
+ SmallVector<Value, 4> while_output;
+ CreateWhile32(loc, n - 1, body_fn, {w, vs, taus}, &while_output, rewriter);
+ return while_output[0];
+ }
+};
+
#include "tensorflow/compiler/mlir/xla/transforms/generated_legalize_tf.inc"
LogicalResult legalizeTF(Operation *op, bool allow_partial_conversion) {
@@ -3883,15 +4730,16 @@
ConvertAllOp, ConvertAnyOp, ConvertArgMaxOp, ConvertBatchMatMulV2Op,
ConvertBroadcastToOp, ConvertBF16FloorDivOp, ConvertConv2DOp,
ConvertConv3DOp, ConvertDepthConv2DOp, ConvertConv2DBackpropFilterOp,
- ConvertConv2DBackpropInputOp, ConvertConv3DBackpropInputOp,
- ConvertCumsumOp, ConvertDiagPartOp, ConvertEinsumOp,
- ConvertFusedBatchNormGradOp, ConvertFusedBatchNormGradV2Op,
- ConvertFusedBatchNormGradV3Op, ConvertFusedBatchNormV3Op,
- ConvertInfeedDequeueTupleOp, ConvertLinSpaceOp, ConvertMaxOp,
- ConvertMinOp, ConvertAvgPoolOp, ConvertMaxPoolOp, ConvertMaxPoolGradOp,
+ ConvertConv3DBackpropFilterOp, ConvertConv2DBackpropInputOp,
+ ConvertConv3DBackpropInputOp, ConvertCumsumOp, ConvertDiagPartOp,
+ ConvertEinsumOp, ConvertFusedBatchNormGradOp,
+ ConvertFusedBatchNormGradV2Op, ConvertFusedBatchNormGradV3Op,
+ ConvertFusedBatchNormV3Op, ConvertInfeedDequeueTupleOp, ConvertLinSpaceOp,
+ ConvertMaxOp, ConvertMinOp, ConvertAvgPoolOp, ConvertMaxPool2DOp,
+ ConvertMaxPool3DOp, ConvertMaxPool2DGradOp, ConvertMaxPool3DGradOp,
ConvertMeanOp, ConvertOneHotOp, ConvertOutfeedEnqueueTupleOp,
- ConvertProdOp, ConvertRangeOp, ConvertSelectV2Op, ConvertSigmoidOp,
- ConvertSizeOp, ConvertSoftmaxOp<TF::LogSoftmaxOp, true>,
+ ConvertProdOp, ConvertQrOp, ConvertRangeOp, ConvertSelectV2Op,
+ ConvertSigmoidOp, ConvertSizeOp, ConvertSoftmaxOp<TF::LogSoftmaxOp, true>,
ConvertSoftmaxOp<TF::SoftmaxOp, false>, ConvertSplitOp, ConvertSplitVOp,
ConvertStridedSliceOp, ConvertStridedSliceGradOp, ConvertSumOp,
ConvertTensorScatterUpdateOp, ConvertTileOp, ConvertTopKV2Op,
diff --git a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_patterns.td b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_patterns.td
index 6a36f3e..3e89890 100644
--- a/tensorflow/compiler/mlir/xla/transforms/legalize_tf_patterns.td
+++ b/tensorflow/compiler/mlir/xla/transforms/legalize_tf_patterns.td
@@ -541,6 +541,7 @@
[TF_LogicalNotOp, HLO_NotOp],
[TF_NegOp, HLO_NegOp],
[TF_RealOp, HLO_RealOp],
+ [TF_RoundOp, HLO_RoundOp],
[TF_RsqrtOp, HLO_RsqrtOp],
[TF_SinOp, HLO_SinOp],
[TF_SqrtOp, HLO_SqrtOp],
diff --git a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
index b177827..d3c3311 100644
--- a/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/lhlo_legalize_to_parallel_loops.cc
@@ -488,23 +488,27 @@
LogicalResult matchAndRewrite(
xla_lhlo::SelectAndScatterOp s_and_s_op, ArrayRef<Value> /*args*/,
ConversionPatternRewriter& rewriter) const final {
+ auto loc = s_and_s_op.getLoc();
InitializeOutput(s_and_s_op, &rewriter);
loop::ParallelOp loop_over_src =
- MakeLoopOverShape(s_and_s_op.getLoc(), s_and_s_op.source(), &rewriter);
+ MakeLoopOverShape(loc, s_and_s_op.source(), &rewriter);
rewriter.setInsertionPointToStart(loop_over_src.getBody());
// Compute indices of the selected element in the window.
auto selected_ivs = SelectIvs(s_and_s_op, loop_over_src, &rewriter);
- // Compute `acc_result` = scatter(out[selected_ivs], src_element)`.
- Value acc_result =
- Scatter(s_and_s_op, loop_over_src, selected_ivs, &rewriter);
- // Updates `out[selected_ivs]`.
- //
- // TODO(pifon): This has to become AtomicRMWOp that updates an element of
- // s_and_s_op.out().
- rewriter.create<StoreOp>(s_and_s_op.getLoc(), acc_result, s_and_s_op.out(),
- selected_ivs);
+ // Load `source[selected_ivs]`.
+ auto src_elem = rewriter.create<LoadOp>(loc, s_and_s_op.source(),
+ loop_over_src.getInductionVars());
+
+ // Compute `out[selected_ivs]` = scatter(out[selected_ivs], src_element)`.
+ auto rmw = rewriter.create<GenericAtomicRMWOp>(loc, s_and_s_op.out(),
+ selected_ivs);
+ OpBuilder rmw_builder = rmw.getBodyBuilder();
+ auto acc_result =
+ ApplySingleResultLhloCode(loc, {src_elem, rmw.getCurrentValue()},
+ &s_and_s_op.scatter().front(), &rmw_builder);
+ rmw_builder.create<AtomicYieldOp>(loc, acc_result);
rewriter.replaceOp(s_and_s_op, llvm::None);
return success();
@@ -685,19 +689,6 @@
}
return if_init.getResults();
}
-
- Value Scatter(xla_lhlo::SelectAndScatterOp s_and_s_op,
- loop::ParallelOp loop_over_src, ValueRange selected_ivs,
- OpBuilder* b) const {
- auto loc = s_and_s_op.getLoc();
-
- auto acc_current = b->create<LoadOp>(loc, s_and_s_op.out(), selected_ivs);
- auto src_elem = b->create<LoadOp>(loc, s_and_s_op.source(),
- loop_over_src.getInductionVars());
-
- return ApplySingleResultLhloCode(loc, {src_elem, acc_current},
- &s_and_s_op.scatter().front(), b);
- }
};
struct LhloLegalizeToParallelLoops
diff --git a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
index 9d04e82..6178434 100644
--- a/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
+++ b/tensorflow/compiler/mlir/xla/transforms/map_hlo_to_lhlo_op.h
@@ -44,17 +44,20 @@
MAP_HLO_TO_LHLO(CeilOp);
MAP_HLO_TO_LHLO(ConstOp);
MAP_HLO_TO_LHLO(CompareOp);
+MAP_HLO_TO_LHLO(ComplexOp);
MAP_HLO_TO_LHLO(ConvertOp);
MAP_HLO_TO_LHLO(CopyOp);
MAP_HLO_TO_LHLO(CosOp);
MAP_HLO_TO_LHLO(DivOp);
MAP_HLO_TO_LHLO(ExpOp);
+MAP_HLO_TO_LHLO(ImagOp);
MAP_HLO_TO_LHLO(IotaOp);
MAP_HLO_TO_LHLO(LogOp);
MAP_HLO_TO_LHLO(MaxOp);
MAP_HLO_TO_LHLO(MinOp);
MAP_HLO_TO_LHLO(MulOp);
MAP_HLO_TO_LHLO(NegOp);
+MAP_HLO_TO_LHLO(RealOp);
MAP_HLO_TO_LHLO(ReduceOp);
MAP_HLO_TO_LHLO(RemOp);
MAP_HLO_TO_LHLO(RsqrtOp);
diff --git a/tensorflow/compiler/mlir/xla/transforms/materialize_broadcasts.cc b/tensorflow/compiler/mlir/xla/transforms/materialize_broadcasts.cc
index 237cac6..421fafb 100644
--- a/tensorflow/compiler/mlir/xla/transforms/materialize_broadcasts.cc
+++ b/tensorflow/compiler/mlir/xla/transforms/materialize_broadcasts.cc
@@ -287,6 +287,47 @@
}
};
+// Converts ClampOp with broadcast semantics. ClampOp requires "all three arrays
+// must be the same shape. Alternatively, as a restricted form of broadcasting,
+// min and/or max can be a scalar of type T."
+struct ClampWithBroadcastConvert : public OpRewritePattern<ClampOp> {
+ explicit ClampWithBroadcastConvert(MLIRContext *context)
+ : OpRewritePattern<ClampOp>(context) {}
+
+ LogicalResult matchAndRewrite(ClampOp op,
+ PatternRewriter &rewriter) const override {
+ auto operand_type = op.operand().getType().dyn_cast<RankedTensorType>();
+ auto max_type = op.max().getType().dyn_cast<RankedTensorType>();
+ auto min_type = op.min().getType().dyn_cast<RankedTensorType>();
+ // Unrancked types are not supported.
+ if (!operand_type || !max_type || !min_type) return failure();
+ // Does not support operand with dynamic dimensions for now.
+ if (!operand_type.hasStaticShape()) return failure();
+
+ ArrayRef<int64_t> operand_shape = operand_type.getShape();
+
+ Value max_value = op.max();
+ if (max_type != operand_type) {
+ assert(max_type.getRank() == 0);
+ max_value = rewriter.createOrFold<BroadcastOp>(
+ op.getLoc(), operand_type, max_value,
+ rewriter.getI64TensorAttr(operand_shape));
+ }
+
+ Value min_value = op.min();
+ if (min_type != operand_type) {
+ assert(min_type.getRank() == 0);
+ min_value = rewriter.createOrFold<BroadcastOp>(
+ op.getLoc(), operand_type, min_value,
+ rewriter.getI64TensorAttr(operand_shape));
+ }
+
+ rewriter.replaceOpWithNewOp<ClampOp>(op, op.getType(), min_value,
+ op.operand(), max_value);
+ return success();
+ }
+};
+
// Specialized class for CompareOp, as it has an additional builder argument.
struct CompareWithBroadcastConvert : public OpRewritePattern<CompareOp> {
explicit CompareWithBroadcastConvert(MLIRContext *context)
@@ -337,6 +378,11 @@
ADD_DYNAMICALLY_LEGAL_OP_WITH_BROADCAST(CompareOp);
#undef ADD_DYNAMICALLY_LEGAL_OP_WITH_BROADCAST
+
+ conversionTarget->addDynamicallyLegalOp<ClampOp>([](ClampOp op) {
+ return op.max().getType() == op.operand().getType() &&
+ op.min().getType() == op.operand().getType();
+ });
}
void PopulateMaterializeBroadcastsPatterns(MLIRContext *context,
@@ -361,6 +407,8 @@
patterns->insert<BinaryOpWithBroadcastConvert<OrOp>>(context);
patterns->insert<BinaryOpWithBroadcastConvert<XorOp>>(context);
+ // ClampOp. It can have a restricted form of broadcasting.
+ patterns->insert<ClampWithBroadcastConvert>(context);
// CompareOp. Note the specialized class instead of using the template.
patterns->insert<CompareWithBroadcastConvert>(context);
}
diff --git a/tensorflow/compiler/mlir/xla/transforms/rewriters.h b/tensorflow/compiler/mlir/xla/transforms/rewriters.h
index 7656c89..ad81cda 100644
--- a/tensorflow/compiler/mlir/xla/transforms/rewriters.h
+++ b/tensorflow/compiler/mlir/xla/transforms/rewriters.h
@@ -61,6 +61,16 @@
OwningRewritePatternList *patterns);
} // namespace xla_hlo
+
+namespace xla_chlo {
+
+// Populates a collection of conversion patterns for legalizing client-HLO to
+// HLO.
+void PopulateLegalizeChloToHloPatterns(MLIRContext *context,
+ OwningRewritePatternList *patterns);
+
+} // namespace xla_chlo
+
} // namespace mlir
#endif // TENSORFLOW_COMPILER_MLIR_XLA_TRANSFORMS_REWRITERS_H_
diff --git a/tensorflow/compiler/mlir/xla/transforms/xla_hlo_to_lhlo_with_xla.cc b/tensorflow/compiler/mlir/xla/transforms/xla_hlo_to_lhlo_with_xla.cc
new file mode 100644
index 0000000..aba8188
--- /dev/null
+++ b/tensorflow/compiler/mlir/xla/transforms/xla_hlo_to_lhlo_with_xla.cc
@@ -0,0 +1,379 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <memory>
+#include <tuple>
+
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallVector.h"
+#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
+#include "mlir/IR/AffineExpr.h" // from @llvm-project
+#include "mlir/IR/AffineMap.h" // from @llvm-project
+#include "mlir/IR/Attributes.h" // from @llvm-project
+#include "mlir/IR/Builders.h" // from @llvm-project
+#include "mlir/IR/Location.h" // from @llvm-project
+#include "mlir/IR/MLIRContext.h" // from @llvm-project
+#include "mlir/IR/Module.h" // from @llvm-project
+#include "mlir/IR/Operation.h" // from @llvm-project
+#include "mlir/IR/StandardTypes.h" // from @llvm-project
+#include "mlir/IR/SymbolTable.h" // from @llvm-project
+#include "mlir/Pass/Pass.h" // from @llvm-project
+#include "mlir/Pass/PassOptions.h" // from @llvm-project
+#include "tensorflow/compiler/mlir/xla/hlo_utils.h"
+#include "tensorflow/compiler/mlir/xla/ir/lhlo_ops.h"
+#include "tensorflow/compiler/mlir/xla/mlir_hlo_to_hlo.h"
+#include "tensorflow/compiler/xla/service/buffer_assignment.h"
+#include "tensorflow/compiler/xla/service/hlo_computation.h"
+#include "tensorflow/compiler/xla/service/hlo_instruction.h"
+#include "tensorflow/compiler/xla/service/hlo_module.h"
+#include "tensorflow/compiler/xla/statusor.h"
+#include "tensorflow/compiler/xla/util.h"
+
+using xla::BufferAllocation;
+using xla::BufferAssignment;
+using xla::HloComputation;
+using xla::HloInstruction;
+using xla::HloModule;
+using xla::HloModuleProto;
+using xla::HloProto;
+using xla::Shape;
+using xla::StatusOr;
+
+namespace mlir {
+namespace {
+
+absl::string_view StringRefToView(llvm::StringRef ref) {
+ return {ref.data(), ref.size()};
+}
+
+StatusOr<std::unique_ptr<HloModule>> HloModuleFromProto(
+ const HloProto& hlo_proto) {
+ const HloModuleProto& module_proto = hlo_proto.hlo_module();
+ TF_ASSIGN_OR_RETURN(const ::xla::HloModuleConfig module_config,
+ HloModule::CreateModuleConfigFromProto(
+ module_proto, ::xla::GetDebugOptionsFromFlags()));
+ return HloModule::CreateFromProto(module_proto, module_config);
+}
+
+// This class will process an HloModule with the supplied BufferAssignment and
+// populate the MLIR ModuleOp with the computation converted in the LHLO
+// dialect.
+class LhloDialectEmitter : public ::xla::DfsHloVisitorWithDefault {
+ public:
+ // Populate the MLIR `module` with the computation from the `hlo_module` using
+ // the provided buffer `assignment`. The returned `Status` indicates success
+ // or failure in the conversion.
+ static Status EmitModule(const BufferAssignment& assignment,
+ const HloModule& hlo_module, ModuleOp module) {
+ return LhloDialectEmitter(assignment, hlo_module, module).Run();
+ }
+
+ private:
+ // Main entry point of the processing: after this call the MLIR ModuleOp is
+ // populated with the computation from the HloModule. The returned `Status`
+ // indicates success or failure in the conversion.
+ Status Run();
+
+ LhloDialectEmitter(const BufferAssignment& assignment,
+ const HloModule& hlo_module, ModuleOp module)
+ : assignment_(std::move(assignment)),
+ hlo_module_(hlo_module),
+ module_(module),
+ builder_(module.getContext()),
+ i8_type_(builder_.getIntegerType(8)) {}
+
+ Status DefaultAction(HloInstruction* hlo) final {
+ return ::xla::Unimplemented("unsupported HLO %s", hlo->name());
+ }
+
+ // Computation parameters don't need any specific handling when they are
+ // visited, they are already processed when we enter a new computation.
+ Status HandleParameter(HloInstruction* instr) final { return Status::OK(); }
+
+ // HLO Copy is translated 1:1 to an lhlo.copy operation.
+ Status HandleCopy(HloInstruction* instr) final {
+ TF_ASSIGN_OR_RETURN(Value source, GetOrCreateView(instr->operand(0)));
+ TF_ASSIGN_OR_RETURN(Value dest, GetOrCreateView(instr));
+ if (source != dest)
+ builder_.create<xla_lhlo::CopyOp>(getLocation(instr),
+ llvm::ArrayRef<Type>{}, source, dest);
+ return Status::OK();
+ }
+
+ // Helper function to create view in a buffer for a given slice. The view is
+ // cached in the `slices_` map.
+ Value GetOrCreateView(const BufferAllocation::Slice& slice);
+
+ // Helper function to create view in a buffer for a given instruction result.
+ StatusOr<Value> GetOrCreateView(const HloInstruction* instr);
+
+ // Return an MLIR location for an HLO instruction.
+ Location getLocation(HloInstruction* inst) {
+ return NameLoc::get(builder_.getIdentifier(inst->name()),
+ builder_.getContext());
+ }
+
+ // This map provides access to MLIR buffers for each HLO buffer allocation.
+ // The MLIR buffers are all `memref<{size}xi8>` and correspond to function
+ // parameters. It is populated at the beginning of the processing with all the
+ // buffer allocations and is unchanged afterward. Every HLOInstruction is
+ // using a "slice" of the buffer allocation and providing shape, layout, and
+ // Dtype. An MLIR view is used separately to model slices into the allocations
+ // (see below).
+ llvm::DenseMap<const BufferAllocation*, Value> allocations_;
+
+ // This map provides access to MLIR buffers for each HLO buffer slice. A slice
+ // is contained in a BufferAllocation, and has an offset and a size.
+ // The MLIR buffers are all `memref<{size}xi8>`. If the slice is the entire
+ // BufferAllocation then the MLIR buffer corresponds to function
+ // parameter for the allocation, otherwise it will map to a ViewOp in the
+ // allocation. It is populated lazily in the `GetOrCreateView()` helper as we
+ // process every instruction.
+ using SliceKey = std::tuple<const BufferAllocation*, int64_t, int64_t>;
+ llvm::DenseMap<SliceKey, Value> slices_;
+
+ // The BufferAssignment computed by XLA ahead of time.
+ const BufferAssignment& assignment_;
+
+ // The HLO module that will be converted.
+ const HloModule& hlo_module_;
+
+ // This is the MLIR module in which a function will be created for every HLO
+ // computation.
+ ModuleOp module_;
+
+ // The builder keeps track of the current insertion point in the MLIR module.
+ OpBuilder builder_;
+ // Convenient "cached" access to this widely used MLIR type (i8).
+ Type i8_type_;
+};
+
+Value LhloDialectEmitter::GetOrCreateView(
+ const BufferAllocation::Slice& slice) {
+ // Check if we already have a view for this slice, otherwise we need to create
+ // a new one.
+ SliceKey slice_key(slice.allocation(), slice.offset(), slice.size());
+ auto slice_view_it = slices_.find(slice_key);
+ if (slice_view_it != slices_.end()) return slice_view_it->second;
+
+ // Check if we can just use the entire allocation before creating a view.
+ Value alloc_buffer = allocations_[slice.allocation()];
+ if (slice.offset() == 0 && slice.size() == slice.allocation()->size()) {
+ slices_.insert({slice_key, alloc_buffer});
+ return alloc_buffer;
+ }
+
+ // Create the view for this slice size, possible with an affine map to model
+ // the offset. The result is cached in the slices_ map.
+ SmallVector<AffineMap, 1> offset_map;
+ if (slice.offset()) {
+ offset_map.push_back(AffineMap::get(
+ /*dimCount=*/1, /*symbolCount=*/0,
+ {getAffineDimExpr(0, builder_.getContext()) + slice.offset()},
+ builder_.getContext()));
+ }
+ auto slice_type = MemRefType::get({slice.size()}, i8_type_, offset_map);
+
+ auto slice_view = builder_.create<ViewOp>(
+ alloc_buffer.getLoc(), slice_type, alloc_buffer, /*operands=*/llvm::None);
+ slices_.insert({slice_key, slice_view});
+ return slice_view;
+}
+
+// Returns a view for the result of an instruction.
+// We first get a view for the slice in the allocation, and then may need to
+// create another view to adjust the slice for the shape of the instruction.
+StatusOr<Value> LhloDialectEmitter::GetOrCreateView(
+ const HloInstruction* instr) {
+ const Shape& target_shape = instr->shape();
+ TF_ASSIGN_OR_RETURN(const BufferAllocation::Slice out_slice,
+ assignment_.GetUniqueTopLevelSlice(instr));
+ Value slice_view = GetOrCreateView(out_slice);
+ TF_ASSIGN_OR_RETURN(Type out_type, ::xla::ConvertShapeToType<MemRefType>(
+ target_shape, builder_));
+ if (slice_view.getType() != out_type)
+ slice_view = builder_.create<ViewOp>(builder_.getUnknownLoc(), out_type,
+ slice_view, llvm::None);
+ return slice_view;
+}
+
+Status LhloDialectEmitter::Run() {
+ HloComputation* computation = hlo_module_.entry_computation();
+ std::string function_name =
+ computation->name().empty() ? "__compute" : computation->name();
+
+ // Create the function as () -> (), we'll compute the arguments from the
+ // buffer allocation and update the type then.
+ auto func_op = FuncOp::create(builder_.getUnknownLoc(), function_name,
+ builder_.getFunctionType({}, {}));
+ Block* block = func_op.addEntryBlock();
+
+ // The function signature will be composed of:
+ // - one memref for each of the parameters.
+ // - one memref for each other buffer allocation.
+ llvm::SmallVector<NamedAttributeList, 8> args_attrs;
+ for (const HloInstruction* param : computation->parameter_instructions()) {
+ TF_ASSIGN_OR_RETURN(auto arg_type, ::xla::ConvertShapeToType<MemRefType>(
+ param->shape(), builder_));
+ // First map parameters to memrefs on the operation.
+ block->addArgument(arg_type);
+ TF_ASSIGN_OR_RETURN(const BufferAllocation::Slice slice,
+ assignment_.GetUniqueTopLevelSlice(param));
+ allocations_[slice.allocation()] = block->getArguments().back();
+ args_attrs.emplace_back();
+ args_attrs.back().set(builder_.getIdentifier("xla_lhlo.params"),
+ builder_.getIndexAttr(param->parameter_number()));
+ }
+
+ for (const BufferAllocation& alloc : assignment_.Allocations()) {
+ if (alloc.is_entry_computation_parameter()) continue;
+ block->addArgument(MemRefType::get({alloc.size()}, i8_type_));
+ allocations_[&alloc] = block->getArguments().back();
+ args_attrs.emplace_back();
+ args_attrs.back().set(builder_.getIdentifier("xla_lhlo.alloc"),
+ builder_.getIndexAttr(alloc.index()));
+ if (alloc.maybe_live_out())
+ args_attrs.back().set(builder_.getIdentifier("xla_lhlo.liveout"),
+ builder_.getBoolAttr(true));
+ }
+
+ FunctionType function_type = builder_.getFunctionType(
+ llvm::to_vector<8>(block->getArgumentTypes()), {});
+ func_op.setType(function_type);
+ func_op.setAllArgAttrs(args_attrs);
+
+ SymbolTable symbol_table(module_);
+ symbol_table.insert(func_op);
+ builder_.setInsertionPointToEnd(block);
+
+ const ::xla::HloInstructionSequence* schedule =
+ assignment_.hlo_ordering().SequentialOrder(*computation);
+ if (!schedule)
+ return ::xla::Unimplemented("Missing sequential order for the computation");
+
+ const std::vector<HloInstruction*>& ordering = schedule->instructions();
+ TF_RETURN_IF_ERROR(computation->AcceptOrdered(this, ordering));
+ builder_.create<ReturnOp>(builder_.getUnknownLoc());
+ return Status::OK();
+}
+
+// Convert the MLIR `module` from HLO dialect to LHLO dialect using XLA for the
+// given platform.
+Status ConvertModule(ModuleOp module, StringRef platform_name) {
+ SymbolTable symbol_table(module);
+ if (!symbol_table.lookup("main")) {
+ return ::xla::InvalidArgument(
+ "conversion to HLO module failed: missing main()");
+ }
+ HloProto hlo_proto;
+ TF_RETURN_WITH_CONTEXT_IF_ERROR(
+ ConvertMlirHloToHlo(module, &hlo_proto,
+ /*use_tuple_args=*/false,
+ /*return_tuple=*/false,
+ /*shape_representation_fn=*/nullptr),
+ "conversion to XLA HLO proto failed");
+
+ auto statusOrHloModule = HloModuleFromProto(hlo_proto);
+ TF_RETURN_WITH_CONTEXT_IF_ERROR(statusOrHloModule.status(),
+ "parsing HLO proto to HLO module failed");
+ std::unique_ptr<HloModule> hlo_module =
+ std::move(statusOrHloModule.ValueOrDie());
+
+ auto platform = ::xla::se::MultiPlatformManager::PlatformWithName(
+ StringRefToView(platform_name));
+ if (!platform.ok()) {
+ std::string error_msg;
+ llvm::raw_string_ostream os(error_msg);
+ os << "failed to get platform: " << platform.status().ToString()
+ << " (available Platform: ";
+ std::vector<std::string> available_platforms;
+ (void)::xla::se::MultiPlatformManager::PlatformsWithFilter(
+ [&](const stream_executor::Platform* p) {
+ available_platforms.push_back(p->Name());
+ return false;
+ });
+ llvm::interleaveComma(available_platforms, os);
+ os << ")";
+ return ::xla::InvalidArgument("%s", os.str().c_str());
+ }
+
+ ::xla::BackendOptions backend_options;
+ backend_options.set_platform(platform.ValueOrDie());
+ auto backend_or_err = ::xla::Backend::CreateBackend(backend_options);
+ TF_RETURN_WITH_CONTEXT_IF_ERROR(backend_or_err.status(),
+ "failed to create XLA Backend ");
+ auto backend = std::move(backend_or_err.ValueOrDie());
+
+ // Run all HLO passes to produce an optimized module.
+ auto result_or = backend->compiler()->RunHloPassesAndBufferAssignement(
+ std::move(hlo_module), backend->default_stream_executor(),
+ backend->memory_allocator());
+ TF_RETURN_WITH_CONTEXT_IF_ERROR(result_or.status(),
+ "running XLA pass pipeline");
+ std::unique_ptr<HloModule> optimized_hlo_module =
+ std::move(std::get<0>(result_or.ValueOrDie()));
+ std::unique_ptr<BufferAssignment> assignment =
+ std::move(std::get<1>(result_or.ValueOrDie()));
+
+ // Clear the module before populating it back with the result of the
+ // conversion.
+ module.getBody()->clear();
+ OpBuilder builder(module);
+ module.ensureTerminator(module.getBodyRegion(), builder, module.getLoc());
+
+ TF_RETURN_WITH_CONTEXT_IF_ERROR(
+ LhloDialectEmitter::EmitModule(*assignment, *optimized_hlo_module,
+ module),
+ "converting HLO to LHLO");
+
+ return Status::OK();
+}
+
+// This pass take a MLIR HLO module, convert it to XLA to perform the HLO
+// optimization pipeline for the required platform, and then convert back to
+// MLIR LHLO.
+class XlaHloToLhloPass
+ : public PassWrapper<XlaHloToLhloPass, OperationPass<ModuleOp>> {
+ public:
+ XlaHloToLhloPass() = default;
+ XlaHloToLhloPass(const XlaHloToLhloPass&) {}
+
+ private:
+ void runOnOperation() final {
+ ModuleOp module = getOperation();
+ Status status = ConvertModule(module, platform_);
+ if (!status.ok()) {
+ module.emitError() << status.ToString();
+ return signalPassFailure();
+ }
+ }
+
+ Option<std::string> platform_{
+ *this, "platform",
+ llvm::cl::desc("The platform to use for the XLA optimization pipeline."),
+ llvm::cl::init("Host")};
+};
+
+} // namespace
+
+std::unique_ptr<OperationPass<ModuleOp>> createXlaHloToLhloWithXlaPass() {
+ return std::make_unique<XlaHloToLhloPass>();
+}
+
+static PassRegistration<XlaHloToLhloPass> registration(
+ "xla-hlo-to-lhlo-with-xla",
+ "Emit LHLO from HLO using the existing XLA implementation");
+
+} // namespace mlir
diff --git a/tensorflow/compiler/tf2tensorrt/BUILD b/tensorflow/compiler/tf2tensorrt/BUILD
index af1877a..337d198 100644
--- a/tensorflow/compiler/tf2tensorrt/BUILD
+++ b/tensorflow/compiler/tf2tensorrt/BUILD
@@ -487,6 +487,8 @@
copts = tf_copts(),
deps = [
"//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
"//tensorflow/core:lib_proto_parsing",
"//tensorflow/core:protos_all_cc",
"@com_google_absl//absl/strings",
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
index bb70581..a729df7 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
@@ -2027,6 +2027,24 @@
return Status::OK();
}
+namespace {
+// Extracts the spatial dimensions from `output_sizes` and returns them as a
+// vector of size 2.
+std::vector<int64_t> GetSpatialDimsFromOutputSizes(
+ const TRT_TensorOrWeights& output_sizes, const int h_index,
+ const int w_index) {
+ // We use h_index and w_index instead of 1 and 2 because we haven't
+ // transposed output_sizes along with the input.
+ const TRT_ShapedWeights& weights = output_sizes.weights();
+ const int output_sizes_length = weights.count();
+ auto output_sizes_values = static_cast<int*>(weights.GetValues());
+ // The length of output_sizes can be 2 or 4. When the length is 4,
+ // output_sizes represents <height,width>.
+ return {output_sizes_values[output_sizes_length == 4 ? h_index : 0],
+ output_sizes_values[output_sizes_length == 4 ? w_index : 1]};
+}
+} // namespace
+
Status ConvertConv2DHelper(OpConverterParams* params, int group,
bool is_conv2d_backprop_input) {
const auto& inputs = params->inputs;
@@ -2125,11 +2143,8 @@
// For backprop, calculate padding based on "input_sizes" input, which
// actually corresponds to output size. ("input_sizes" makes sense in the
// context of Conv2DBackpropInput).
- // We use h_index and w_index instead of 1 and 2 because we havent
- // transposed backprop_output_size along with the input.
- auto output_size_weights =
- static_cast<int*>(backprop_output_size.weights().GetValues());
- input_dims = {output_size_weights[h_index], output_size_weights[w_index]};
+ input_dims =
+ GetSpatialDimsFromOutputSizes(backprop_output_size, h_index, w_index);
} else {
// Use 1 and 2 because tensor_dim has the dimensions of the transposed
// input.
@@ -2189,22 +2204,24 @@
// argument output_shape and thus the TRT output shape could be wrong
// in case of strides>1.
if (is_conv2d_backprop_input) {
- auto tf_output_shape =
- static_cast<int*>(backprop_output_size.weights().GetValues());
+ std::vector<int64_t> output_spatial_dims =
+ GetSpatialDimsFromOutputSizes(backprop_output_size, h_index, w_index);
+ const int output_height = output_spatial_dims[0];
+ const int output_width = output_spatial_dims[1];
nvinfer1::Dims trt_output_shape = output_tensor->getDimensions();
// What determines the padding size is the difference between the given
// input_sizes (tf_output_shape) and TRT computed size.
- const int height_diff = tf_output_shape[h_index] - trt_output_shape.d[1];
- const int width_diff = tf_output_shape[w_index] - trt_output_shape.d[2];
+ const int height_diff = output_height - trt_output_shape.d[1];
+ const int width_diff = output_width - trt_output_shape.d[2];
if ((height_diff < 0) || (width_diff < 0)) {
return errors::InvalidArgument(
"input_sizes argument of Conv2DBackprop (i.e. output_shape argument "
"of conv2d_transpose) ",
"is too small for the given out_backprop argument of Conv2DBackprop "
"(i.e. input argument of conv2d_transpose). Expect: ",
- "(", tf_output_shape[h_index], ", ", tf_output_shape[w_index],
- ") >= ", "(", trt_output_shape.d[1], ", ", trt_output_shape.d[2],
- ") for op ", node_def.name());
+ "(", output_height, ", ", output_width, ") >= ", "(",
+ trt_output_shape.d[1], ", ", trt_output_shape.d[2], ") for op ",
+ node_def.name());
}
// Only add a padding layer if padding sizes are larger than 0
if ((height_diff > 0) || (width_diff > 0)) {
@@ -2283,112 +2300,70 @@
Status ConvertReshape(OpConverterParams* params) {
const auto& inputs = params->inputs;
- const auto& node_def = params->node_def;
TF_RETURN_IF_ERROR(
CheckInputsWeights(*params, {{"tensor", false}, {"shape", true}}));
TF_RETURN_IF_ERROR(AllowDataTypes(
*params, {DataType::DT_FLOAT, DataType::DT_HALF, DataType::DT_INT32}));
const TRT_TensorOrWeights& input_tensor = inputs.at(0);
+
+ // TODO(bixia): we can't use inputs.at(1).weights().ToVector<int>() for two
+ // reasons: (1) When weights.count()==0, TRT_ShapedWeights::tensor_ dtype is
+ // not properly set to INT32. (2) I tried a fix for the first problem, I got
+ // shared pointer related error in convert_nodes_test. We should fix the
+ // problems and switch to use inputs.at(1).weights().ToVector<int>(), a type
+ // safe method to access the content of the tensor.
TRT_ShapedWeights weights = inputs.at(1).weights();
if (weights.count() == 0) {
return errors::Unimplemented("Reshape to shape=[] is not supported, at ",
- node_def.name());
+ params->node_def.name());
}
- const int* weights_ptr = static_cast<int*>(weights.GetValues());
-
- // Check that it doesn't change the batch dimension. This check is
- // conservative, for example, when the first dim of the shape is -1 and input
- // tensor shape is not fixed, it is still possible that the reshape doesn't
- // change the batch dim, but as long as there is a possibility that it could
- // change the batch dim, it reject the conversion. The parameters are:
- //
- // * reshape_batch_dim: the value of the first dim of the input shape constant
- // * reshape_dims: all other dims of the input shape constant
- // * input_batch_dim: the value of the first dim of the input tensor to
- // reshape
- // * input_dims: all other dims of the input tensor to reshape
- //
- // The validation logic is:
- //
- // if input_batch_dim is fixed:
- // if reshape_batch_dim == input_batch_dim:
- // ok
- // elif reshape_batch_dim == -1 (meaning reshape_dims are fixed) and
- // input_dims are fixed and
- // prod(input_dims) == prod(reshape_dims)
- // ok
- // else:
- // not ok
- // elif input_dims are fixed:
- // if reshape_dims are fixed and
- // prod(input_dims) == prod(reshape_dims):
- // ok
- // else:
- // not ok
- // else:
- // not ok
- //
- // Note that the following is ok no matter whether reshape_batch_dim is fixed
- // or not:
- //
- // ```
- // input_batch_dim is not fixed &&
- // reshape_dims are fixed &&
- // prod(input_dims) == prod(reshape_dims),
- // ```
- //
- // because the non-batch dims of the new and old shapes match, and TF runtime
- // should make sure the batch dim is not changed.
+ const int* output_shape_dims = static_cast<int*>(weights.GetValues());
+ size_t output_shape_dims_count = weights.count();
const int input_batch_dim = input_tensor.batch_size();
- const int reshape_batch_dim = weights_ptr[0];
- const nvinfer1::Dims input_dims = input_tensor.GetTrtDims();
+ const int output_batch_dim = output_shape_dims[0];
- nvinfer1::Dims reshape_dims;
- reshape_dims.nbDims = weights.count() - 1;
- for (int i = 1; i < weights.count(); i++) {
- reshape_dims.d[i - 1] = weights_ptr[i];
+ const nvinfer1::Dims input_nonbatch_dims = input_tensor.GetTrtDims();
+ nvinfer1::Dims output_nonbatch_dims;
+ output_nonbatch_dims.nbDims = output_shape_dims_count - 1;
+ for (int i = 1; i < output_shape_dims_count; i++) {
+ output_nonbatch_dims.d[i - 1] = output_shape_dims[i];
}
- // Check that it doesn't change the batch dimension according to the logic
- // mentioned above.
- bool reshape_may_change_batch_dim = false;
- if (input_batch_dim > 0) { // Batch size is fixed.
- if (reshape_batch_dim == -1) { // Other dims of the shape must be fixed.
- if (!AreDimsStaticWithSameSize(input_dims, reshape_dims,
- /*is_tensor=*/true)) {
- reshape_may_change_batch_dim = true;
- }
- } else if (reshape_batch_dim != input_batch_dim) {
- reshape_may_change_batch_dim = true;
- } else {
- // This means (input_batch_dim>0 && input_batch_dim==reshape_batch_dim),
- // and TF runtime should make sure non-batch dims are matched.
- }
- } else if (!AreDimsStaticWithSameSize(input_dims, reshape_dims,
- /*is_tensor=*/true)) {
- reshape_may_change_batch_dim = true;
- }
VLOG(1) << "input_batch_dim=" << input_batch_dim
- << ", input_dims=" << DebugString(input_dims)
- << "\nreshape_batch_dim=" << reshape_batch_dim
- << ", reshape_dims=" << DebugString(reshape_dims);
+ << ", input_nonbatch_dims=" << DebugString(input_nonbatch_dims)
+ << "\nresult_batch_dim=" << output_batch_dim
+ << ", result_nonbatch_dims=" << DebugString(output_nonbatch_dims);
+
+ // Check whether input_batch_dim and output_batch_dim will have the same
+ // static value.
+ bool reshape_may_change_batch_dim = false;
+ if (input_batch_dim != -1 && output_batch_dim != -1) {
+ reshape_may_change_batch_dim = (input_batch_dim != output_batch_dim);
+ } else {
+ reshape_may_change_batch_dim =
+ !AreDimsStaticWithSameSize(input_nonbatch_dims, output_nonbatch_dims,
+ /*is_tensor=*/true);
+ }
if (reshape_may_change_batch_dim) {
- const string msg = StrCat(
- "Reshape on batch dimension is not supported, at ", node_def.name(),
- ". input_batch_dim=", input_batch_dim, ", ", DebugString(input_dims),
- "; reshape_batch_dim=", reshape_batch_dim, ", ",
- DebugString(reshape_dims));
+ const string msg =
+ StrCat("Reshape on batch dimension is not supported, at ",
+ params->node_def.name(), ". input_batch_dim=", input_batch_dim,
+ ", ", DebugString(input_nonbatch_dims),
+ "; output_batch_dim=", output_batch_dim, ", ",
+ DebugString(output_nonbatch_dims));
return errors::Unimplemented(msg);
}
- // Start conversion.
+ // Perform the conversion.
nvinfer1::ITensor* output_tensor = nullptr;
TF_RETURN_IF_ERROR(params->converter->PrepareTensorForShape(
- input_tensor, reshape_dims, params->validation_only, &output_tensor));
+ input_tensor, output_nonbatch_dims, params->validation_only,
+ &output_tensor));
if (params->validation_only) return Status::OK();
+ // Record the conversion result.
params->outputs->push_back(TRT_TensorOrWeights(output_tensor));
return Status::OK();
}
diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
index 64d82d1..3e9c5db 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
@@ -3698,28 +3698,16 @@
// Get nodedef for Conv2D layer.
auto get_conv2d_nodedef =
[](std::vector<int> strides = {1, 1, 1, 1}, string padding = "SAME",
- string data_format = "NCHW", std::vector<int> dilations = {1, 1, 1, 1},
- bool is_conv2d_backprop_input = false) -> NodeDef {
+ string data_format = "NCHW",
+ std::vector<int> dilations = {1, 1, 1, 1}) -> NodeDef {
Scope s = Scope::NewRootScope();
auto input = ops::Placeholder(s.WithOpName("input"), DT_FLOAT);
auto filter = ops::Placeholder(s.WithOpName("weights"), DT_FLOAT);
- if (is_conv2d_backprop_input) {
- auto input_sizes =
- ops::Placeholder(s.WithOpName("input_sizes"), DT_INT32);
- ops::Conv2DBackpropInput::Attrs attrs = ops::Conv2DBackpropInput::Attrs()
- .DataFormat(data_format)
- .Dilations(dilations);
- auto conv2d =
- ops::Conv2DBackpropInput(s.WithOpName("my_conv2d"), input_sizes,
- filter, input, strides, padding, attrs);
- return conv2d.operation.node()->def();
- } else {
- ops::Conv2D::Attrs attrs =
- ops::Conv2D::Attrs().DataFormat(data_format).Dilations(dilations);
- auto conv2d = ops::Conv2D(s.WithOpName("my_conv2d"), input, filter,
- strides, padding, attrs);
- return conv2d.operation.node()->def();
- }
+ ops::Conv2D::Attrs attrs =
+ ops::Conv2D::Attrs().DataFormat(data_format).Dilations(dilations);
+ auto conv2d = ops::Conv2D(s.WithOpName("my_conv2d"), input, filter, strides,
+ padding, attrs);
+ return conv2d.operation.node()->def();
};
{
@@ -3786,19 +3774,6 @@
"dimensions, at my_conv2d");
}
{
- // Dilation + Conv2DBackpropInput, should fail.
- Reset();
- NodeDef node_def =
- get_conv2d_nodedef({1, 1, 1, 1}, "SAME", "NHWC", {1, 1, 2, 1}, true);
- AddTestTensor("input", {2, 3, 1});
- AddTestWeights<float>("weights", {3, 3, 1, 1}, {1, 2, 3, 4, 5, 6, 7, 8, 9});
- AddTestWeights<int>("input_sizes", {4}, {1, 2, 3, 1});
- RunValidationAndConversion(node_def, error::UNIMPLEMENTED,
- "Dilation with Conv2DBackpropInput "
- "(conv2d_transpose) is not supported, "
- "at my_conv2d");
- }
- {
// Strides is not 4D, should fail.
Reset();
NodeDef node_def =
@@ -3830,7 +3805,6 @@
string padding;
string data_format;
std::vector<int> dilations;
- bool is_conv2d_backprop_input;
std::vector<int> expected_output_dims;
std::vector<float> expected_output;
};
@@ -3846,7 +3820,6 @@
/*padding=*/"VALID",
/*data_format=*/"NCHW",
/*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{1, 2, 2},
/*expected_output=*/{1, 1, 0, 1}},
// SAME padding (Asymmetric)
@@ -3858,7 +3831,6 @@
/*padding=*/"SAME",
/*data_format=*/"NCHW",
/*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{1, 2, 3},
/*expected_output=*/{1, 1, -2, 0, 1, -4}},
// SAME padding (Symmetric)
@@ -3870,7 +3842,6 @@
/*padding=*/"SAME",
/*data_format=*/"NCHW",
/*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{1, 2, 3},
/*expected_output=*/{1, 2, -1, 3, 1, -3}},
// NHWC
@@ -3882,7 +3853,6 @@
/*padding=*/"VALID",
/*data_format=*/"NHWC",
/*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{2, 2, 1},
/*expected_output=*/{1, 1, 0, 1}},
// Dilated
@@ -3894,7 +3864,6 @@
/*padding=*/"VALID",
/*data_format=*/"NCHW",
/*dilations=*/{1, 1, 1, 2},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{1, 2, 1},
/*expected_output=*/{2, 1}},
// Strided
@@ -3906,62 +3875,18 @@
/*padding=*/"VALID",
/*data_format=*/"NCHW",
/*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/false,
/*expected_output_dims=*/{1, 2, 2},
/*expected_output=*/{1, 0, 1, 3}},
- // Transpose Strided
- TestParams{/*input_dims=*/{1, 2, 2},
- /*input=*/{0, 1, 2, 3},
- /*filter_dims=*/{1, 2, 1, 1},
- /*filter=*/{-1, 1},
- /*strides=*/{1, 1, 1, 2},
- /*padding=*/"SAME",
- /*data_format=*/"NCHW",
- /*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/true,
- /*expected_output_dims=*/{1, 2, 4},
- /*expected_output=*/{0, 0, -1, 1, -2, 2, -3, 3}},
- // Transpose Strided NHWC
- TestParams{/*input_dims=*/{2, 2, 1},
- /*input=*/{0, 1, 2, 3},
- /*filter_dims=*/{1, 2, 1, 1},
- /*filter=*/{-1, 1},
- /*strides=*/{1, 1, 2, 1},
- /*padding=*/"SAME",
- /*data_format=*/"NHWC",
- /*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/true,
- /*expected_output_dims=*/{2, 4, 1},
- /*expected_output=*/{0, 0, -1, 1, -2, 2, -3, 3}},
- // Transpose Strided NHWC with VALID padding
- TestParams{/*input_dims=*/{3, 1, 1},
- /*input=*/{0, 1, 2},
- /*filter_dims=*/{2, 1, 1, 1},
- /*filter=*/{-1, 1},
- /*strides=*/{1, 2, 1, 1},
- /*padding=*/"VALID",
- /*data_format=*/"NHWC",
- /*dilations=*/{1, 1, 1, 1},
- /*is_conv2d_backprop_input=*/true,
- /*expected_output_dims=*/{7, 1, 1},
- /*expected_output=*/{0, 0, -1, 1, -2, 2, 0}},
-
};
for (int i = 0; i < ok_params.size(); i++) {
Reset();
- NodeDef node_def = get_conv2d_nodedef(
- ok_params[i].strides, ok_params[i].padding, ok_params[i].data_format,
- ok_params[i].dilations, ok_params[i].is_conv2d_backprop_input);
+ NodeDef node_def =
+ get_conv2d_nodedef(ok_params[i].strides, ok_params[i].padding,
+ ok_params[i].data_format, ok_params[i].dilations);
AddTestTensor("input", ok_params[i].input_dims);
AddTestWeights<float>("weights", ok_params[i].filter_dims,
ok_params[i].filter);
- if (ok_params[i].is_conv2d_backprop_input) {
- std::vector<int> tf_input_sizes = ok_params[i].expected_output_dims;
- tf_input_sizes.insert(tf_input_sizes.begin(), 1); // Add batch dimension.
- QCHECK_EQ(4, tf_input_sizes.size());
- AddTestWeights<int>("input_sizes", {4}, tf_input_sizes);
- }
RunValidationAndConversion(node_def);
TRT_TensorOrWeights output;
TF_EXPECT_OK(GetTensorOrWeights("my_conv2d", &output));
@@ -3979,6 +3904,134 @@
}
}
+TEST_F(OpConverterTest, ConvertConv2DBackpropInput) {
+ // Get nodedef for Conv2D layer.
+ auto get_conv2d_backprop_input_nodedef =
+ [](std::vector<int> strides = {1, 1, 1, 1}, string padding = "SAME",
+ string data_format = "NCHW",
+ std::vector<int> dilations = {1, 1, 1, 1}) -> NodeDef {
+ Scope s = Scope::NewRootScope();
+ auto input = ops::Placeholder(s.WithOpName("input"), DT_FLOAT);
+ auto filter = ops::Placeholder(s.WithOpName("weights"), DT_FLOAT);
+ auto input_sizes = ops::Placeholder(s.WithOpName("input_sizes"), DT_INT32);
+ ops::Conv2DBackpropInput::Attrs attrs = ops::Conv2DBackpropInput::Attrs()
+ .DataFormat(data_format)
+ .Dilations(dilations);
+ auto conv2d = ops::Conv2DBackpropInput(
+ s.WithOpName("my_conv2d_backprop_input"), input_sizes, filter, input,
+ strides, padding, attrs);
+ return conv2d.operation.node()->def();
+ };
+
+ {
+ // Dilation + Conv2DBackpropInput, should fail.
+ Reset();
+ NodeDef node_def = get_conv2d_backprop_input_nodedef({1, 1, 1, 1}, "SAME",
+ "NHWC", {1, 1, 2, 1});
+ AddTestTensor("input", {2, 3, 1});
+ AddTestWeights<float>("weights", {3, 3, 1, 1}, {1, 2, 3, 4, 5, 6, 7, 8, 9});
+ AddTestWeights<int>("input_sizes", {4}, {1, 2, 3, 1});
+ RunValidationAndConversion(node_def, error::UNIMPLEMENTED,
+ "Dilation with Conv2DBackpropInput "
+ "(conv2d_transpose) is not supported, "
+ "at my_conv2d_backprop_input");
+ }
+
+ struct TestParams {
+ std::vector<int> input_dims;
+ std::vector<float> input;
+ std::vector<int> filter_dims;
+ std::vector<float> filter;
+ std::vector<int> strides;
+ string padding;
+ string data_format;
+ std::vector<int> dilations;
+ std::vector<int> expected_output_dims;
+ std::vector<float> expected_output;
+ };
+
+ // Ok.
+ std::vector<TestParams> ok_params = {
+ // Transpose Strided
+ TestParams{/*input_dims=*/{1, 2, 2},
+ /*input=*/{0, 1, 2, 3},
+ /*filter_dims=*/{1, 2, 1, 1},
+ /*filter=*/{-1, 1},
+ /*strides=*/{1, 1, 1, 2},
+ /*padding=*/"SAME",
+ /*data_format=*/"NCHW",
+ /*dilations=*/{1, 1, 1, 1},
+ /*expected_output_dims=*/{1, 2, 4},
+ /*expected_output=*/{0, 0, -1, 1, -2, 2, -3, 3}},
+ // Transpose Strided NHWC
+ TestParams{/*input_dims=*/{2, 2, 1},
+ /*input=*/{0, 1, 2, 3},
+ /*filter_dims=*/{1, 2, 1, 1},
+ /*filter=*/{-1, 1},
+ /*strides=*/{1, 1, 2, 1},
+ /*padding=*/"SAME",
+ /*data_format=*/"NHWC",
+ /*dilations=*/{1, 1, 1, 1},
+ /*expected_output_dims=*/{2, 4, 1},
+ /*expected_output=*/{0, 0, -1, 1, -2, 2, -3, 3}},
+ // Transpose Strided NHWC with VALID padding
+ TestParams{/*input_dims=*/{3, 1, 1},
+ /*input=*/{0, 1, 2},
+ /*filter_dims=*/{2, 1, 1, 1},
+ /*filter=*/{-1, 1},
+ /*strides=*/{1, 2, 1, 1},
+ /*padding=*/"VALID",
+ /*data_format=*/"NHWC",
+ /*dilations=*/{1, 1, 1, 1},
+ /*expected_output_dims=*/{7, 1, 1},
+ /*expected_output=*/{0, 0, -1, 1, -2, 2, 0}},
+ };
+
+ for (int i = 0; i < ok_params.size(); i++) {
+ for (int input_sizes_length : {2, 4}) {
+ Reset();
+ NodeDef node_def = get_conv2d_backprop_input_nodedef(
+ ok_params[i].strides, ok_params[i].padding, ok_params[i].data_format,
+ ok_params[i].dilations);
+ AddTestTensor("input", ok_params[i].input_dims);
+ AddTestWeights<float>("weights", ok_params[i].filter_dims,
+ ok_params[i].filter);
+
+ std::vector<int> tf_input_sizes = ok_params[i].expected_output_dims;
+ if (input_sizes_length == 4) {
+ tf_input_sizes.insert(tf_input_sizes.begin(),
+ 1); // Add batch dimension.
+ QCHECK_EQ(4, tf_input_sizes.size());
+ AddTestWeights<int>("input_sizes", {4}, tf_input_sizes);
+ } else {
+ // Remove the channel dimension.
+ if (ok_params[i].data_format == "NHWC") {
+ tf_input_sizes.pop_back();
+ } else {
+ tf_input_sizes.erase(tf_input_sizes.begin());
+ }
+ QCHECK_EQ(2, tf_input_sizes.size());
+ AddTestWeights<int>("input_sizes", {2}, tf_input_sizes);
+ }
+
+ RunValidationAndConversion(node_def);
+ TRT_TensorOrWeights output;
+ TF_EXPECT_OK(GetTensorOrWeights("my_conv2d_backprop_input", &output));
+ ASSERT_TRUE(output.is_tensor());
+ ExpectTrtDimsEqualsArray(ok_params[i].expected_output_dims,
+ output.tensor()->getDimensions());
+
+ const DataVec input_data{{"input", AsTensor<float>(ok_params[i].input)}};
+ DataVec output_data{
+ {"my_conv2d_backprop_input",
+ ConstructTensor<float>(ok_params[i].expected_output.size())}};
+ BuildAndRun(input_data, &output_data);
+ EXPECT_THAT(GetSpanForData<float>(output_data[0]),
+ ElementsAreArray(ok_params[i].expected_output));
+ }
+ }
+}
+
#if IS_TRT_VERSION_GE(6, 0, 0, 0)
TEST_F(OpConverterTest, ConvertConv3D) {
// Get nodedef for Conv3D layer.
diff --git a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
index 66a1a96..269d71d 100644
--- a/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
+++ b/tensorflow/compiler/tf2tensorrt/kernels/trt_engine_op.cc
@@ -569,7 +569,15 @@
input_concrete_shapes.push_back(ctx->input(i).shape());
}
- OP_REQUIRES_OK_ASYNC(ctx, VerifyInputShapes(input_concrete_shapes), *helper);
+ Status verify_input_shape_status = VerifyInputShapes(input_concrete_shapes);
+ // TODO(bixia): Fix the segmentation.
+ if (!verify_input_shape_status.ok()) {
+ LOG_FIRST_N(WARNING, 5) << "Running native segment for" << name()
+ << " due to failure in verifying input shapes: "
+ << verify_input_shape_status.error_message();
+ ExecuteNativeSegment(ctx, helper);
+ return;
+ }
if (!use_implicit_batch_) {
if (profile_generation_mode_) {
diff --git a/tensorflow/compiler/tf2tensorrt/segment/segment.cc b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
index 4d9dd42..9b15137 100644
--- a/tensorflow/compiler/tf2tensorrt/segment/segment.cc
+++ b/tensorflow/compiler/tf2tensorrt/segment/segment.cc
@@ -27,8 +27,11 @@
#include "tensorflow/core/graph/graph_constructor.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/status.h"
+#include "tensorflow/core/lib/gtl/flatset.h"
+#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/types.h"
+#include "tensorflow/core/util/env_var.h"
#if GOOGLE_CUDA
#if GOOGLE_TENSORRT
@@ -422,6 +425,19 @@
// for TRT.
std::unordered_set<string> unsupported_ops;
int num_unsupported_ops = 0;
+
+ // Getting the operations blacklisted for conversion
+ string tftrt_op_blacklist_str;
+ TF_CHECK_OK(
+ ReadStringFromEnvVar("TF_TRT_OP_BLACKLIST", "", &tftrt_op_blacklist_str));
+
+ auto tftrt_op_blacklist = gtl::FlatSet<string>{}; // non-absl ok
+
+ for (const auto& x : str_util::Split(tftrt_op_blacklist_str, ",")) {
+ tftrt_op_blacklist.insert(x);
+ }
+
+ // Parsing each node of the graph
std::vector<UnionFind<SimpleNode*>> node_segments;
for (int i = 0; i < graph->num_node_ids(); ++i) {
SimpleNode* node = graph->FindNodeId(i);
@@ -443,6 +459,16 @@
unsupported_ops.emplace(node->tf_node()->type_string());
num_unsupported_ops++;
node = nullptr;
+ } else if (tftrt_op_blacklist.count(node->tf_node()->type_string())) {
+ // WARNING verbosity since the user explicitly requests this behavior.
+ LOG(WARNING)
+ << "Blacklisted as TF-TRT candidate, "
+ << "(Op type: " << node->tf_node()->type_string() << "), "
+ << "(Op name: " << node->name() << "), "
+ << "(Reason: Blacklisted with the env var TF_TRT_OP_BLACKLIST)";
+ unsupported_ops.emplace(node->tf_node()->type_string());
+ num_unsupported_ops++;
+ node = nullptr;
} else {
VLOG(2) << "Accepted as a TF-TRT candidate, "
<< "(Op type: " << node->tf_node()->type_string() << "), "
diff --git a/tensorflow/compiler/tf2xla/kernels/shape_op.cc b/tensorflow/compiler/tf2xla/kernels/shape_op.cc
index 5d2b08f..85917af 100644
--- a/tensorflow/compiler/tf2xla/kernels/shape_op.cc
+++ b/tensorflow/compiler/tf2xla/kernels/shape_op.cc
@@ -274,10 +274,23 @@
auto list_shape_or = ctx->builder()->GetShape(list);
OP_REQUIRES_OK(ctx, list_shape_or.status());
+ const xla::Shape& list_shape = list_shape_or.ValueOrDie();
+ std::vector<std::vector<xla::XlaOp>> list_dynamic_dims;
+ list_dynamic_dims.reserve(list_shape.tuple_shapes_size() - 1);
+ for (int64 i = 0; i < list_shape.tuple_shapes_size() - 1; ++i) {
+ // Set dynamic dimension size to 0 for initialization value.
+ std::vector<xla::XlaOp> dynamic_dims;
+ const xla::Shape& shape = list_shape.tuple_shapes(i);
+ auto sub_element = xla::GetTupleElement(list, i);
+ for (int64 dim = 0; dim < shape.dimensions_size(); ++dim) {
+ dynamic_dims.push_back(xla::GetDimensionSize(sub_element, dim));
+ }
+ list_dynamic_dims.push_back(dynamic_dims);
+ }
xla::XlaOp new_list;
OP_REQUIRES_OK(
- ctx, CreateZerosTensorListWithShape(
- ctx->builder(), list_shape_or.ValueOrDie(), &new_list));
+ ctx, CreateZerosTensorListWithShape(ctx->builder(), list_shape,
+ list_dynamic_dims, &new_list));
xla::XlaOp push_index;
OP_REQUIRES_OK(ctx, GetTensorListPushIndex(list, &push_index));
@@ -287,10 +300,20 @@
SetTensorListPushIndex(new_list, push_index, &result));
ctx->SetTensorListOutput(0, result);
} else {
- const TensorShape input_shape = ctx->InputShape(0);
-
auto zero = XlaHelpers::Zero(ctx->builder(), input_type(0));
- ctx->SetOutput(0, xla::Broadcast(zero, input_shape.dim_sizes()));
+ xla::XlaOp input = ctx->Input(0);
+ auto input_shape = ctx->InputXlaShape(0).ValueOrDie();
+ auto result = xla::Broadcast(zero, input_shape.dimensions());
+
+ // Setting up dynamic dimensions of the broadcast.
+ for (int64 i = 0; i < input_shape.dimensions_size(); ++i) {
+ if (input_shape.is_dynamic_dimension(i)) {
+ xla::XlaOp input_dynamic_dim = xla::GetDimensionSize(input, i);
+ result = xla::SetDimensionSize(result, input_dynamic_dim, i);
+ }
+ }
+
+ ctx->SetOutput(0, result);
}
}
};
diff --git a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc
index 9093175..2684c98 100644
--- a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc
+++ b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc
@@ -202,7 +202,7 @@
ctx, output_elements == input_elements_sliced,
errors::InvalidArgument(
"The number of output elements ", output_elements,
- " has to equal to number of input elements that are sliced ",
+ " has to equal to number of input elements that are sliced ",
input_elements_sliced, " when input indices are not constant."));
for (int64 i = 0; i < ctx->InputShape("begin").dims(); ++i) {
diff --git a/tensorflow/compiler/tf2xla/kernels/tensor_list_ops.cc b/tensorflow/compiler/tf2xla/kernels/tensor_list_ops.cc
index 4af3d42..fa5a96c 100644
--- a/tensorflow/compiler/tf2xla/kernels/tensor_list_ops.cc
+++ b/tensorflow/compiler/tf2xla/kernels/tensor_list_ops.cc
@@ -44,6 +44,36 @@
namespace {
+// GetTensorListDynamicDims collects the dynamic dimensions that a tensorlist
+// may carry and returns them in a 2D vector: int64[ElementSize][DimSize]. If a
+// dimension is static, a constant dimension is returned.
+xla::StatusOr<std::vector<std::vector<xla::XlaOp>>> GetTensorListDynamicDims(
+ XlaOpKernelContext* ctx, const xla::Shape& element_shape,
+ const xla::Shape& list_shape, int64 num_elements) {
+ std::vector<int64> dynamic_sizes;
+ ctx->set_dynamic_dimension_is_minus_one(true);
+ // The multiplier can be a dynamic value.
+ TF_RETURN_IF_ERROR(ctx->ConstantInputAsIntVector(0, &dynamic_sizes));
+ std::vector<std::vector<xla::XlaOp>> list_dynamic_dims;
+ // Set dynamic dimension size to 0 for initialization value.
+ std::vector<xla::XlaOp> dynamic_dims;
+ // Leading dim is a static dimension.
+ dynamic_dims.push_back(xla::ConstantR0<int32>(ctx->builder(), num_elements));
+ for (int64 dim = 0; dim < element_shape.dimensions_size(); ++dim) {
+ if (ctx->is_dynamic_dimension(dynamic_sizes[dim])) {
+ auto dynamic_dim_size = xla::Slice(ctx->Input(0), {dim}, {dim + 1}, {1});
+ dynamic_dim_size = xla::Reshape(dynamic_dim_size, {});
+ dynamic_dim_size = xla::ConvertElementType(dynamic_dim_size, xla::S32);
+ dynamic_dims.push_back(dynamic_dim_size);
+ } else {
+ dynamic_dims.push_back(
+ xla::ConstantR0<int32>(ctx->builder(), dynamic_sizes[dim]));
+ }
+ }
+ list_dynamic_dims.push_back(dynamic_dims);
+ return list_dynamic_dims;
+}
+
class TensorListLengthOp : public XlaOpKernel {
public:
explicit TensorListLengthOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {}
@@ -124,10 +154,14 @@
xla::Shape list_shape;
OP_REQUIRES_OK(ctx, GetTensorListShapeFromElementShape(
element_shape, num_elements, &list_shape));
-
+ // Set up dynamic dimension sizes to create the zero tensor.
+ auto list_dynamic_dims_or = GetTensorListDynamicDims(
+ ctx, element_shape, list_shape, num_elements);
+ OP_REQUIRES_OK(ctx, list_dynamic_dims_or.status());
xla::XlaOp new_list;
OP_REQUIRES_OK(ctx, CreateZerosTensorListWithShape(
- ctx->builder(), list_shape, &new_list));
+ ctx->builder(), list_shape,
+ list_dynamic_dims_or.ValueOrDie(), &new_list));
xla::XlaOp result;
OP_REQUIRES_OK(
ctx,
@@ -185,10 +219,16 @@
xla::Shape list_shape;
OP_REQUIRES_OK(ctx, GetTensorListShapeFromElementShape(
element_shape, max_num_elements, &list_shape));
+ // Set up dynamic dimension sizes to create the zero tensor.
+ auto list_dynamic_dims_or = GetTensorListDynamicDims(
+ ctx, element_shape, list_shape, max_num_elements);
+ OP_REQUIRES_OK(ctx, list_dynamic_dims_or.status());
xla::XlaOp result;
OP_REQUIRES_OK(ctx, CreateZerosTensorListWithShape(
- ctx->builder(), list_shape, &result));
+ ctx->builder(), list_shape,
+ list_dynamic_dims_or.ValueOrDie(), &result));
+
ctx->SetTensorListOutput(0, result);
return;
}
diff --git a/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.cc b/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.cc
index 6020b00..aa71e4d 100644
--- a/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.cc
+++ b/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.cc
@@ -16,6 +16,7 @@
#include "tensorflow/compiler/tf2xla/kernels/tensor_list_utils.h"
#include "tensorflow/compiler/tf2xla/shape_util.h"
+#include "tensorflow/compiler/xla/client/xla_builder.h"
#include "tensorflow/compiler/xla/literal_util.h"
#include "tensorflow/compiler/xla/shape.h"
#include "tensorflow/compiler/xla/shape_util.h"
@@ -247,19 +248,29 @@
return Status::OK();
}
-Status CreateZerosTensorListWithShape(xla::XlaBuilder* b,
- const xla::Shape& list_shape,
- xla::XlaOp* list) {
+Status CreateZerosTensorListWithShape(
+ xla::XlaBuilder* b, const xla::Shape& list_shape,
+ const std::vector<std::vector<xla::XlaOp>>& dynamic_dims,
+ xla::XlaOp* list) {
int tuple_size = xla::ShapeUtil::TupleElementCount(list_shape);
std::vector<xla::XlaOp> elements;
- for (int i = 0; i < tuple_size; i++) {
+ TF_RET_CHECK(dynamic_dims.size() == tuple_size - 1);
+ for (int i = 0; i < tuple_size - 1; i++) {
const xla::Shape& shape =
xla::ShapeUtil::GetTupleElementShape(list_shape, i);
xla::XlaOp zero =
xla::ConstantLiteral(b, xla::LiteralUtil::Zero(shape.element_type()));
xla::XlaOp zeros = xla::Broadcast(zero, shape.dimensions());
+ TF_RET_CHECK(dynamic_dims[i].size() == shape.dimensions_size());
+ for (int64 dim = 0; dim < shape.dimensions_size(); ++dim) {
+ zeros = xla::SetDimensionSize(zeros, dynamic_dims[i][dim], dim);
+ }
elements.push_back(zeros);
}
+ // List size (last item) has to be S32.
+ TF_RET_CHECK(xla::ShapeUtil::GetTupleElementShape(list_shape, tuple_size - 1)
+ .element_type() == xla::S32);
+ elements.push_back(xla::ConstantLiteral(b, xla::LiteralUtil::Zero(xla::S32)));
*list = xla::Tuple(b, elements);
return Status::OK();
}
@@ -272,12 +283,12 @@
xla::XlaBuilder* b = list.builder();
xla::Shape list_shape;
+ TF_ASSIGN_OR_RETURN(xla::Shape element_shape, b->GetShape(element));
+
if (element_is_tensor_list) {
- TF_ASSIGN_OR_RETURN(xla::Shape element_shape, b->GetShape(element));
TF_RETURN_IF_ERROR(GetTensorListShapeFromElementTensorListShape(
element_shape, leading_dim, &list_shape));
} else {
- TF_ASSIGN_OR_RETURN(xla::Shape element_shape, b->GetShape(element));
TF_RETURN_IF_ERROR(GetTensorListShapeFromElementShape(
element_shape, leading_dim, &list_shape));
}
@@ -295,7 +306,27 @@
*initialized_list = list;
return Status::OK();
} else {
- return CreateZerosTensorListWithShape(b, list_shape, initialized_list);
+ // Prepare dynamic dimension dimensions for zero tensor list. The dynamic
+ // sizes are created by reading the dynamic dimension size of sub-elements.
+ std::vector<std::vector<xla::XlaOp>> list_dynamic_dims;
+ for (int64 i = 0; i < list_shape.tuple_shapes_size() - 1; ++i) {
+ std::vector<xla::XlaOp> dynamic_dims;
+ const xla::Shape& shape = list_shape.tuple_shapes(i);
+ // Leading dim is a static dimension.
+ dynamic_dims.push_back(xla::ConstantR0<int32>(b, leading_dim));
+ xla::XlaOp sub_element;
+ if (element_is_tensor_list) {
+ sub_element = xla::GetTupleElement(element, i);
+ } else {
+ sub_element = element;
+ }
+ for (int64 dim = 0; dim < shape.dimensions_size() - 1; ++dim) {
+ dynamic_dims.push_back(xla::GetDimensionSize(sub_element, dim));
+ }
+ list_dynamic_dims.push_back(dynamic_dims);
+ }
+ return CreateZerosTensorListWithShape(b, list_shape, list_dynamic_dims,
+ initialized_list);
}
}
@@ -473,7 +504,13 @@
xla::XlaOp list_part = xla::GetTupleElement(list, 0);
xla::XlaOp read = xla::DynamicSlice(list_part, start_indices, slice_shape);
-
+ for (int64 i = 0; i < buffer_shape.dimensions_size(); ++i) {
+ if (buffer_shape.is_dynamic_dimension(i)) {
+ auto buffer = xla::GetTupleElement(list, 0);
+ auto gds = xla::GetDimensionSize(buffer, i);
+ read = xla::SetDimensionSize(read, gds, i);
+ }
+ }
slice_shape.erase(slice_shape.begin());
*result = xla::Reshape(read, slice_shape);
return Status::OK();
diff --git a/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.h b/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.h
index 7fac2d9..ef3c8ba 100644
--- a/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.h
+++ b/tensorflow/compiler/tf2xla/kernels/tensor_list_utils.h
@@ -74,9 +74,9 @@
xla::Shape* tensor_list_shape);
// Returns a TensorList filled by zeros with the given shape.
-Status CreateZerosTensorListWithShape(xla::XlaBuilder* b,
- const xla::Shape& list_shape,
- xla::XlaOp* list);
+Status CreateZerosTensorListWithShape(
+ xla::XlaBuilder* b, const xla::Shape& list_shape,
+ const std::vector<std::vector<xla::XlaOp>>& dynamic_dims, xla::XlaOp* list);
// If the TensorList is initialized, check that its shape matches element shape;
// If the TensorList is uninitialized, initialize it with the element shape.
diff --git a/tensorflow/compiler/tf2xla/kernels/while_op.cc b/tensorflow/compiler/tf2xla/kernels/while_op.cc
index 21568a1..fe7a589 100644
--- a/tensorflow/compiler/tf2xla/kernels/while_op.cc
+++ b/tensorflow/compiler/tf2xla/kernels/while_op.cc
@@ -510,8 +510,25 @@
// first compilation and the body/cond was recompiled with the updated
// shape/datatype of the list.
if (input_shape != list_shape) {
- OP_REQUIRES_OK(ctx, CreateZerosTensorListWithShape(
- ctx->builder(), list_shape, &inputs[i]));
+ // Prepare dynamic dimensions for element shapes.
+ std::vector<std::vector<xla::XlaOp>> list_dynamic_dims;
+ for (int64 i = 0; i < list_shape.tuple_shapes_size() - 1; ++i) {
+ // Set dynamic dimension size to 0 for initilization value.
+ std::vector<xla::XlaOp> dynamic_dims;
+ const xla::Shape& shape = list_shape.tuple_shapes(i);
+ for (int64 dim = 0; dim < shape.dimensions_size(); ++dim) {
+ int32 dim_size = shape.dimensions(dim);
+ if (shape.is_dynamic_dimension(dim)) {
+ dim_size = 0;
+ }
+ dynamic_dims.push_back(
+ xla::ConstantR0<int32>(ctx->builder(), dim_size));
+ }
+ list_dynamic_dims.push_back(dynamic_dims);
+ }
+ OP_REQUIRES_OK(
+ ctx, CreateZerosTensorListWithShape(ctx->builder(), list_shape,
+ list_dynamic_dims, &inputs[i]));
} else {
inputs[i] = ctx->Input(input_num);
}
diff --git a/tensorflow/compiler/tf2xla/xla_compiler.cc b/tensorflow/compiler/tf2xla/xla_compiler.cc
index 85f2d5c..368cbe0 100644
--- a/tensorflow/compiler/tf2xla/xla_compiler.cc
+++ b/tensorflow/compiler/tf2xla/xla_compiler.cc
@@ -621,6 +621,7 @@
graph_optimizer_options.cf_consider_fn = cf_consider_fn;
graph_optimizer_options.inline_multi_device_functions = true;
graph_optimizer_options.inline_impl_selection_group_functions = true;
+ graph_optimizer_options.inline_with_single_device_body_placer = true;
optimizer.Optimize(flib_runtime_, flib_runtime_->env(),
/*device=*/nullptr, &graph, graph_optimizer_options);
diff --git a/tensorflow/compiler/tf2xla/xla_op_kernel.h b/tensorflow/compiler/tf2xla/xla_op_kernel.h
index 8a38439..6987b6f 100644
--- a/tensorflow/compiler/tf2xla/xla_op_kernel.h
+++ b/tensorflow/compiler/tf2xla/xla_op_kernel.h
@@ -217,6 +217,8 @@
return dynamic_dimension_is_minus_one_;
}
+ bool is_dynamic_dimension(int64 dim_size) { return dim_size == -1; }
+
// Reads the current value of the resource variable referred to by input
// `index`. If `shape` is not nullptr, sets `*shape` to the shape of the
// variable. Returns an error if the variable has not been initialized, or if
diff --git a/tensorflow/compiler/xla/client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_builder.cc
index 2a69023..c72361c 100644
--- a/tensorflow/compiler/xla/client/xla_builder.cc
+++ b/tensorflow/compiler/xla/client/xla_builder.cc
@@ -822,23 +822,29 @@
absl::Span<const int64> limit_indices,
absl::Span<const int64> strides) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand));
TF_ASSIGN_OR_RETURN(Shape shape, ShapeInference::InferSliceShape(
*operand_shape, start_indices,
limit_indices, strides));
- *instr.mutable_shape() = shape.ToProto();
- for (int i = 0; i < start_indices.size(); i++) {
- auto* slice_config = instr.add_slice_dimensions();
- slice_config->set_start(start_indices[i]);
- slice_config->set_limit(limit_indices[i]);
- slice_config->set_stride(strides[i]);
- }
-
- return AddInstruction(std::move(instr), HloOpcode::kSlice, {operand});
+ return SliceInternal(shape, operand, start_indices, limit_indices, strides);
});
}
+StatusOr<XlaOp> XlaBuilder::SliceInternal(const Shape& shape, XlaOp operand,
+ absl::Span<const int64> start_indices,
+ absl::Span<const int64> limit_indices,
+ absl::Span<const int64> strides) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = shape.ToProto();
+ for (int i = 0; i < start_indices.size(); i++) {
+ auto* slice_config = instr.add_slice_dimensions();
+ slice_config->set_start(start_indices[i]);
+ slice_config->set_limit(limit_indices[i]);
+ slice_config->set_stride(strides[i]);
+ }
+ return AddInstruction(std::move(instr), HloOpcode::kSlice, {operand});
+}
+
XlaOp XlaBuilder::SliceInDim(XlaOp operand, int64 start_index,
int64 limit_index, int64 stride, int64 dimno) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
@@ -952,41 +958,49 @@
XlaOp XlaBuilder::ConcatInDim(absl::Span<const XlaOp> operands,
int64 dimension) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
-
std::vector<const Shape*> operand_shape_ptrs;
TF_ASSIGN_OR_RETURN(const auto& operand_shapes, GetOperandShapes(operands));
absl::c_transform(operand_shapes, std::back_inserter(operand_shape_ptrs),
[](const Shape& shape) { return &shape; });
TF_ASSIGN_OR_RETURN(Shape shape, ShapeInference::InferConcatOpShape(
operand_shape_ptrs, dimension));
- *instr.mutable_shape() = shape.ToProto();
-
- instr.add_dimensions(dimension);
-
- return AddInstruction(std::move(instr), HloOpcode::kConcatenate, operands);
+ return ConcatInDimInternal(shape, operands, dimension);
});
}
+StatusOr<XlaOp> XlaBuilder::ConcatInDimInternal(
+ const Shape& shape, absl::Span<const XlaOp> operands, int64 dimension) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = shape.ToProto();
+
+ instr.add_dimensions(dimension);
+
+ return AddInstruction(std::move(instr), HloOpcode::kConcatenate, operands);
+}
+
XlaOp XlaBuilder::Pad(XlaOp operand, XlaOp padding_value,
const PaddingConfig& padding_config) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
-
TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand));
TF_ASSIGN_OR_RETURN(const Shape* padding_value_shape,
GetShapePtr(padding_value));
TF_ASSIGN_OR_RETURN(
Shape shape, ShapeInference::InferPadShape(
*operand_shape, *padding_value_shape, padding_config));
- *instr.mutable_shape() = shape.ToProto();
- *instr.mutable_padding_config() = padding_config;
-
- return AddInstruction(std::move(instr), HloOpcode::kPad,
- {operand, padding_value});
+ return PadInternal(shape, operand, padding_value, padding_config);
});
}
+StatusOr<XlaOp> XlaBuilder::PadInternal(const Shape& shape, XlaOp operand,
+ XlaOp padding_value,
+ const PaddingConfig& padding_config) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = shape.ToProto();
+ *instr.mutable_padding_config() = padding_config;
+ return AddInstruction(std::move(instr), HloOpcode::kPad,
+ {operand, padding_value});
+}
+
XlaOp XlaBuilder::Reshape(XlaOp operand, absl::Span<const int64> dimensions,
absl::Span<const int64> new_sizes,
int64 inferred_dimension) {
@@ -1080,7 +1094,6 @@
XlaOp XlaBuilder::Tuple(absl::Span<const XlaOp> elements) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
std::vector<const Shape*> operand_shape_ptrs;
TF_ASSIGN_OR_RETURN(const auto& operand_shapes, GetOperandShapes(elements));
absl::c_transform(operand_shapes, std::back_inserter(operand_shape_ptrs),
@@ -1088,14 +1101,19 @@
TF_ASSIGN_OR_RETURN(const Shape shape,
ShapeInference::InferVariadicOpShape(
HloOpcode::kTuple, operand_shape_ptrs));
- *instr.mutable_shape() = shape.ToProto();
- return AddInstruction(std::move(instr), HloOpcode::kTuple, elements);
+ return TupleInternal(shape, elements);
});
}
+StatusOr<XlaOp> XlaBuilder::TupleInternal(const Shape& shape,
+ absl::Span<const XlaOp> elements) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = shape.ToProto();
+ return AddInstruction(std::move(instr), HloOpcode::kTuple, elements);
+}
+
XlaOp XlaBuilder::GetTupleElement(XlaOp tuple_data, int64 index) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
TF_ASSIGN_OR_RETURN(const Shape* tuple_shape, GetShapePtr(tuple_data));
if (!tuple_shape->IsTuple()) {
return InvalidArgument(
@@ -1107,16 +1125,22 @@
"GetTupleElement() index (%d) out of range for tuple shape %s", index,
ShapeUtil::HumanString(*tuple_shape));
}
- *instr.mutable_shape() =
- ShapeUtil::GetTupleElementShape(*tuple_shape, index).ToProto();
-
- instr.set_tuple_index(index);
-
- return AddInstruction(std::move(instr), HloOpcode::kGetTupleElement,
- {tuple_data});
+ return GetTupleElementInternal(
+ ShapeUtil::GetTupleElementShape(*tuple_shape, index), tuple_data,
+ index);
});
}
+StatusOr<XlaOp> XlaBuilder::GetTupleElementInternal(const Shape& shape,
+ XlaOp tuple_data,
+ int64 index) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = shape.ToProto();
+ instr.set_tuple_index(index);
+ return AddInstruction(std::move(instr), HloOpcode::kGetTupleElement,
+ {tuple_data});
+}
+
XlaOp XlaBuilder::Dot(XlaOp lhs, XlaOp rhs,
const PrecisionConfig* precision_config) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
@@ -1407,14 +1431,11 @@
XlaOp XlaBuilder::InfeedWithToken(XlaOp token, const Shape& shape,
const string& config) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
if (!LayoutUtil::HasLayout(shape)) {
return InvalidArgument("Given shape to Infeed must have a layout");
}
const Shape infeed_instruction_shape =
ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeTokenShape()});
- *instr.mutable_shape() = infeed_instruction_shape.ToProto();
- instr.set_infeed_config(config);
if (shape.IsArray() && sharding() &&
sharding()->type() == OpSharding::OTHER) {
@@ -1427,11 +1448,18 @@
return InvalidArgument(
"Replicated sharding is not yet supported for infeeds");
}
-
- return AddInstruction(std::move(instr), HloOpcode::kInfeed, {token});
+ return InfeedWithTokenInternal(infeed_instruction_shape, token, config);
});
}
+StatusOr<XlaOp> XlaBuilder::InfeedWithTokenInternal(
+ const Shape& infeed_instruction_shape, XlaOp token, const string& config) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = infeed_instruction_shape.ToProto();
+ instr.set_infeed_config(config);
+ return AddInstruction(std::move(instr), HloOpcode::kInfeed, {token});
+}
+
void XlaBuilder::Outfeed(XlaOp operand, const Shape& shape_with_layout,
const string& outfeed_config) {
ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
@@ -1488,10 +1516,6 @@
const Shape& shape_with_layout,
const string& outfeed_config) {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
- HloInstructionProto instr;
-
- *instr.mutable_shape() = ShapeUtil::MakeTokenShape().ToProto();
-
// Check and set outfeed shape.
if (!LayoutUtil::HasLayout(shape_with_layout)) {
return InvalidArgument("Given shape to Outfeed must have a layout");
@@ -1503,15 +1527,22 @@
ShapeUtil::HumanStringWithLayout(shape_with_layout),
ShapeUtil::HumanStringWithLayout(*operand_shape));
}
- *instr.mutable_outfeed_shape() = shape_with_layout.ToProto();
-
- instr.set_outfeed_config(outfeed_config);
-
- return AddInstruction(std::move(instr), HloOpcode::kOutfeed,
- {operand, token});
+ return OutfeedWithTokenInternal(operand, token, shape_with_layout,
+ outfeed_config);
});
}
+StatusOr<XlaOp> XlaBuilder::OutfeedWithTokenInternal(
+ XlaOp operand, XlaOp token, const Shape& shape_with_layout,
+ const string& outfeed_config) {
+ HloInstructionProto instr;
+ *instr.mutable_shape() = ShapeUtil::MakeTokenShape().ToProto();
+ *instr.mutable_outfeed_shape() = shape_with_layout.ToProto();
+ instr.set_outfeed_config(outfeed_config);
+ return AddInstruction(std::move(instr), HloOpcode::kOutfeed,
+ {operand, token});
+}
+
XlaOp XlaBuilder::CreateToken() {
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
HloInstructionProto instr;
@@ -2615,6 +2646,11 @@
TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand));
TF_ASSIGN_OR_RETURN(Shape shape, ShapeInference::InferGetDimensionSizeShape(
*operand_shape, dimension));
+ // Calling GetDimensionSize on a static dimension returns a constant
+ // instruction.
+ if (!operand_shape->is_dynamic_dimension(dimension)) {
+ return ConstantR0<int32>(this, operand_shape->dimensions(dimension));
+ }
*instr.mutable_shape() = shape.ToProto();
instr.add_dimensions(dimension);
return AddInstruction(std::move(instr), HloOpcode::kGetDimensionSize,
@@ -2626,8 +2662,20 @@
return ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
HloInstructionProto instr;
TF_ASSIGN_OR_RETURN(const Shape* operand_shape, GetShapePtr(operand));
+
TF_ASSIGN_OR_RETURN(Shape shape, ShapeInference::InferSetDimensionSizeShape(
*operand_shape, dimension));
+ // Setting an op's dynamic dimension to the static size is a noop.
+ TF_ASSIGN_OR_RETURN(const HloInstructionProto* val_proto,
+ LookUpInstruction(val));
+ if (StringToHloOpcode(val_proto->opcode()).ValueOrDie() ==
+ HloOpcode::kConstant) {
+ TF_ASSIGN_OR_RETURN(auto literal,
+ Literal::CreateFromProto(val_proto->literal(), true));
+ if (literal.Get<int32>({}) == shape.dimensions(dimension)) {
+ return operand;
+ }
+ }
*instr.mutable_shape() = shape.ToProto();
instr.add_dimensions(dimension);
return AddInstruction(std::move(instr), HloOpcode::kSetDimensionSize,
diff --git a/tensorflow/compiler/xla/client/xla_builder.h b/tensorflow/compiler/xla/client/xla_builder.h
index f320fee..4eba598 100644
--- a/tensorflow/compiler/xla/client/xla_builder.h
+++ b/tensorflow/compiler/xla/client/xla_builder.h
@@ -364,6 +364,10 @@
Status SetInstructionFrontendAttribute(XlaOp op, string attribute,
string value);
+ // Returns shapes for the operands.
+ StatusOr<std::vector<Shape>> GetOperandShapes(
+ absl::Span<const XlaOp> operands) const;
+
private:
// Build helper which takes the id of the root operation..
StatusOr<XlaComputation> Build(int64 root_id, bool remove_dynamic_dimensions);
@@ -391,6 +395,10 @@
XlaOp Pad(XlaOp operand, XlaOp padding_value,
const PaddingConfig& padding_config);
+ virtual StatusOr<XlaOp> PadInternal(const Shape& shape, XlaOp operand,
+ XlaOp padding_value,
+ const PaddingConfig& padding_config);
+
XlaOp Reshape(XlaOp operand, absl::Span<const int64> dimensions,
absl::Span<const int64> new_sizes,
int64 inferred_dimension = -1);
@@ -406,9 +414,12 @@
XlaOp Slice(XlaOp operand, absl::Span<const int64> start_indices,
absl::Span<const int64> limit_indices,
absl::Span<const int64> strides);
-
- XlaOp SliceInDim(XlaOp operand, int64 start_index, int64 limit_index,
- int64 stride, int64 dimno);
+ virtual StatusOr<XlaOp> SliceInternal(const Shape& shape, XlaOp operand,
+ absl::Span<const int64> start_indices,
+ absl::Span<const int64> limit_indices,
+ absl::Span<const int64> strides);
+ virtual XlaOp SliceInDim(XlaOp operand, int64 start_index, int64 limit_index,
+ int64 stride, int64 dimno);
ABSL_DEPRECATED("Use span-of-indices form instead")
XlaOp DynamicSlice(XlaOp operand, XlaOp start_indices,
@@ -422,14 +433,22 @@
absl::Span<const XlaOp> start_indices);
XlaOp ConcatInDim(absl::Span<const XlaOp> operands, int64 dimension);
+ virtual StatusOr<XlaOp> ConcatInDimInternal(const Shape& shape,
+ absl::Span<const XlaOp> operands,
+ int64 dimension);
void Trace(const string& tag, XlaOp operand);
XlaOp Select(XlaOp pred, XlaOp on_true, XlaOp on_false);
XlaOp Tuple(absl::Span<const XlaOp> elements);
+ virtual StatusOr<XlaOp> TupleInternal(const Shape& shape,
+ absl::Span<const XlaOp> elements);
XlaOp GetTupleElement(XlaOp tuple_data, int64 index);
+ virtual StatusOr<XlaOp> GetTupleElementInternal(const Shape& shape,
+ XlaOp tuple_data,
+ int64 index);
XlaOp Dot(XlaOp lhs, XlaOp rhs,
const PrecisionConfig* precision_config = nullptr);
@@ -476,15 +495,18 @@
absl::Span<const int64> fft_length);
XlaOp Infeed(const Shape& shape, const string& config = "");
- XlaOp InfeedWithToken(XlaOp token, const Shape& shape,
- const string& config = "");
+ XlaOp InfeedWithToken(XlaOp token, const Shape& shape, const string& config);
+ virtual StatusOr<XlaOp> InfeedWithTokenInternal(
+ const Shape& infeed_instruction_shape, XlaOp token, const string& config);
void Outfeed(XlaOp operand, const Shape& shape_with_layout,
const string& outfeed_config);
XlaOp OutfeedWithToken(XlaOp operand, XlaOp token,
const Shape& shape_with_layout,
const string& outfeed_config);
-
+ virtual StatusOr<XlaOp> OutfeedWithTokenInternal(
+ XlaOp operand, XlaOp token, const Shape& shape_with_layout,
+ const string& outfeed_config);
XlaOp Call(const XlaComputation& computation,
absl::Span<const XlaOp> operands);
@@ -624,7 +646,7 @@
XlaOp RecvFromHost(XlaOp token, const Shape& shape,
const ChannelHandle& handle);
- XlaOp CreateToken();
+ virtual XlaOp CreateToken();
XlaOp AfterAll(absl::Span<const XlaOp> tokens);
@@ -701,10 +723,6 @@
// Returns the (inferred) result for the program shape using the given root.
StatusOr<ProgramShape> GetProgramShape(int64 root_id) const;
- // Returns shapes for the operands.
- StatusOr<std::vector<Shape>> GetOperandShapes(
- absl::Span<const XlaOp> operands) const;
-
// A visitor which checks whether an operation is a compile-time constant,
// meaning that it doesn't depend on any parameters, or on any stateful
// operation such as `RngNormal` or `Infeed`. The visitor walks the
diff --git a/tensorflow/compiler/xla/client/xla_builder_test.cc b/tensorflow/compiler/xla/client/xla_builder_test.cc
index 115a822..1fa839b 100644
--- a/tensorflow/compiler/xla/client/xla_builder_test.cc
+++ b/tensorflow/compiler/xla/client/xla_builder_test.cc
@@ -407,13 +407,25 @@
TEST_F(XlaBuilderTest, GetDimensionSize) {
XlaBuilder b(TestName());
- auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}), "x");
+ auto x =
+ Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}, {false, true}), "x");
GetDimensionSize(x, 1);
TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b));
auto root = module->entry_computation()->root_instruction();
EXPECT_EQ(root->opcode(), HloOpcode::kGetDimensionSize);
}
+TEST_F(XlaBuilderTest, GetDimensionSizeConstant) {
+ XlaBuilder b(TestName());
+ auto x =
+ Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}, {false, true}), "x");
+ // Get dimension size from a contant dimension gives us a constant.
+ GetDimensionSize(x, 0);
+ TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b));
+ auto root = module->entry_computation()->root_instruction();
+ EXPECT_EQ(root->opcode(), HloOpcode::kConstant);
+}
+
TEST_F(XlaBuilderTest, ReportError) {
XlaBuilder b(TestName());
auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}), "x");
diff --git a/tensorflow/compiler/xla/debug_options_flags.cc b/tensorflow/compiler/xla/debug_options_flags.cc
index 8604531..e6d60e5 100644
--- a/tensorflow/compiler/xla/debug_options_flags.cc
+++ b/tensorflow/compiler/xla/debug_options_flags.cc
@@ -63,6 +63,8 @@
opts.set_xla_allow_excess_precision(true);
opts.set_xla_force_host_platform_device_count(1);
opts.set_xla_gpu_deterministic_reductions(false);
+ opts.set_xla_cpu_enable_xprof_traceme(true);
+
return opts;
}
@@ -529,7 +531,6 @@
flag_values->xla_gpu_algorithm_blacklist_path(),
"An AlgorithmBlacklist text proto file as a blacklist "
"of convolutions to avoid to use."),
-
tensorflow::Flag(
"xla_gpu_deterministic_reductions",
bool_setter_for(&DebugOptions::set_xla_gpu_deterministic_reductions),
@@ -545,6 +546,12 @@
bool_setter_for(&DebugOptions::set_xla_tpu_detect_inf),
flag_values->xla_tpu_detect_inf(),
"Trigger error on execution on TPU if a INF value is detected"),
+ tensorflow::Flag(
+ "xla_cpu_enable_xprof_traceme",
+ bool_setter_for(&DebugOptions::set_xla_cpu_enable_xprof_traceme),
+ flag_values->xla_cpu_enable_xprof_traceme(),
+ "If true, XLA CPU generates code to call "
+ "TraceMe::Activity{Start|End} around HLO operations."),
});
ParseFlagsFromEnvAndDieIfUnknown("XLA_FLAGS", *flag_objects);
}
diff --git a/tensorflow/compiler/xla/executable_run_options.h b/tensorflow/compiler/xla/executable_run_options.h
index 6981b35..43ee0fd 100644
--- a/tensorflow/compiler/xla/executable_run_options.h
+++ b/tensorflow/compiler/xla/executable_run_options.h
@@ -127,6 +127,13 @@
ExecutableRunOptions& set_rng_seed(int rng_seed);
int rng_seed() const;
+ ExecutableRunOptions& set_launch_id(int32 launch_id) {
+ launch_id_ = launch_id;
+ return *this;
+ }
+
+ int32 launch_id() const { return launch_id_; }
+
ExecutableRunOptions& set_run_id(RunId id);
RunId run_id() const;
@@ -153,6 +160,7 @@
const Eigen::ThreadPoolDevice* intra_op_thread_pool_ = nullptr;
ExecutionProfile* execution_profile_ = nullptr;
int rng_seed_ = 0;
+ int32 launch_id_ = 0;
stream_executor::Stream* host_to_device_stream_ = nullptr;
ThenExecuteFunction* then_execute_function_ = nullptr;
RunId run_id_;
diff --git a/tensorflow/compiler/xla/python/BUILD b/tensorflow/compiler/xla/python/BUILD
index 7c11091..0aed818 100644
--- a/tensorflow/compiler/xla/python/BUILD
+++ b/tensorflow/compiler/xla/python/BUILD
@@ -25,9 +25,25 @@
srcs = ["custom_call_for_test.pyx"],
)
-py_test(
+py_library(
name = "xla_client_test",
+ testonly = 1,
srcs = ["xla_client_test.py"],
+ srcs_version = "PY3",
+ deps = [
+ ":custom_call_for_test",
+ ":xla_client",
+ ":xla_extension",
+ "@absl_py//absl/flags",
+ "@absl_py//absl/testing:absltest",
+ "@absl_py//absl/testing:parameterized",
+ ],
+)
+
+py_test(
+ name = "xla_client_test_cpu",
+ srcs = ["xla_client_test.py"],
+ args = ["--backend=cpu"],
main = "xla_client_test.py",
python_version = "PY3",
srcs_version = "PY3",
@@ -36,6 +52,27 @@
":custom_call_for_test",
":xla_client",
":xla_extension",
+ "@absl_py//absl/flags",
+ "@absl_py//absl/testing:absltest",
+ "@absl_py//absl/testing:parameterized",
+ ] + xla_py_test_deps(),
+)
+
+py_test(
+ name = "xla_client_test_gpu",
+ srcs = ["xla_client_test.py"],
+ args = ["--backend=gpu"],
+ main = "xla_client_test.py",
+ python_version = "PY3",
+ srcs_version = "PY3",
+ tags = [
+ "no_oss",
+ "requires-gpu-nvidia",
+ ], # TODO(phawkins): This test passes, but requires --config=monolithic.
+ deps = [
+ ":xla_client",
+ ":xla_extension",
+ "@absl_py//absl/flags",
"@absl_py//absl/testing:absltest",
"@absl_py//absl/testing:parameterized",
] + xla_py_test_deps(),
diff --git a/tensorflow/compiler/xla/python/local_client.cc b/tensorflow/compiler/xla/python/local_client.cc
index f19d0f85..74669d0 100644
--- a/tensorflow/compiler/xla/python/local_client.cc
+++ b/tensorflow/compiler/xla/python/local_client.cc
@@ -774,10 +774,10 @@
}
}
if (block_stream != nullptr) {
+ se::Stream* block_stream_ptr = block_stream.release();
local_device_state->ThenExecuteOnCallbackThread(
- block_stream.get(),
- [device_buffer, block_stream_ptr{block_stream.release()},
- local_device_state]() {
+ block_stream_ptr,
+ [device_buffer, block_stream_ptr, local_device_state]() {
local_device_state->ReturnStreamToPool(
std::unique_ptr<se::Stream>(block_stream_ptr));
});
diff --git a/tensorflow/compiler/xla/python/tpu_driver/client/tpu_client.py b/tensorflow/compiler/xla/python/tpu_driver/client/tpu_client.py
index 8933893..e99ba05 100644
--- a/tensorflow/compiler/xla/python/tpu_driver/client/tpu_client.py
+++ b/tensorflow/compiler/xla/python/tpu_driver/client/tpu_client.py
@@ -86,7 +86,8 @@
device = self.client.local_devices()[0]
return _tpu_client.PyTpuBuffer.from_python(pyval, self.client, device)
- def compile(self, c_computation, compile_options):
+ def compile(self, c_computation, compile_options=None):
+ compile_options = compile_options or xla_client.CompileOptions()
options = _xla.ExecutableBuildOptions()
options.num_replicas = compile_options.num_replicas
options.num_partitions = compile_options.num_partitions
diff --git a/tensorflow/compiler/xla/python/xla.cc b/tensorflow/compiler/xla/python/xla.cc
index 9916948..3d6d78c 100644
--- a/tensorflow/compiler/xla/python/xla.cc
+++ b/tensorflow/compiler/xla/python/xla.cc
@@ -310,6 +310,13 @@
// XlaBuilder.
py::module ops = m->def_submodule("ops", "XLA operations");
+ py::enum_<TriangularSolveOptions::Transpose>(
+ ops, "TriangularSolveOptions_Transpose")
+ .value("TRANSPOSE_INVALID", TriangularSolveOptions::TRANSPOSE_INVALID)
+ .value("NO_TRANSPOSE", TriangularSolveOptions::NO_TRANSPOSE)
+ .value("TRANSPOSE", TriangularSolveOptions::TRANSPOSE)
+ .value("ADJOINT", TriangularSolveOptions::ADJOINT);
+
ops.def("AfterAll", &AfterAll, py::arg("builder"), py::arg("tokens"));
ops.def(
"AllReduce",
@@ -317,19 +324,20 @@
XlaOp, const XlaComputation&, absl::Span<const ReplicaGroup>,
const absl::optional<ChannelHandle>&, const absl::optional<Shape>&)>(
&AllReduce),
- py::arg("operand"), py::arg("computation"), py::arg("replica_groups"),
+ py::arg("operand"), py::arg("computation"),
+ py::arg("replica_groups") = py::list(),
py::arg("channel_id") = absl::nullopt,
py::arg("shape_with_layout") = absl::nullopt);
ops.def("AllToAll", &AllToAll, py::arg("operand"), py::arg("split_dimension"),
py::arg("concat_dimension"), py::arg("split_count"),
- py::arg("replica_groups"));
+ py::arg("replica_groups") = py::list());
ops.def("CollectivePermute", &CollectivePermute, py::arg("operand"),
py::arg("source_target_pairs"));
ops.def("CreateToken", &CreateToken, py::arg("builder"));
ops.def("CrossReplicaSum",
static_cast<XlaOp (*)(XlaOp, absl::Span<const ReplicaGroup>)>(
&CrossReplicaSum),
- py::arg("operand"), py::arg("replica_groups"));
+ py::arg("operand"), py::arg("replica_groups") = py::list());
ops.def("BitcastConvertType", &BitcastConvertType, py::arg("operand"),
py::arg("new_element_type"));
ops.def("Broadcast", &Broadcast, py::arg("operand"), py::arg("sizes"));
@@ -353,6 +361,7 @@
py::arg("predicate"), py::arg("true_operand"),
py::arg("true_computation"), py::arg("false_operand"),
py::arg("false_computation"));
+ ops.def("Constant", &ConstantLiteral, py::arg("builder"), py::arg("literal"));
ops.def("ConstantLiteral", &ConstantLiteral, py::arg("builder"),
py::arg("literal"));
ops.def("ConvGeneralDilated", &ConvGeneralDilated, py::arg("lhs"),
@@ -363,13 +372,28 @@
py::arg("precision_config") = nullptr);
ops.def("ConvertElementType", &ConvertElementType, py::arg("operand"),
py::arg("new_element_type"));
- ops.def("CustomCall", &CustomCall, py::arg("builder"),
- py::arg("call_target_name"), py::arg("operands"), py::arg("shape"),
- py::arg("opaque"));
- ops.def("CustomCallWithLayout", &CustomCallWithLayout, py::arg("builder"),
- py::arg("call_target_name"), py::arg("operands"),
- py::arg("shape_with_layout"), py::arg("operand_shapes_with_layout"),
- py::arg("opaque"));
+ ops.def(
+ "CustomCall",
+ [](XlaBuilder* builder, const py::bytes& call_target_name,
+ absl::Span<const XlaOp> operands, const Shape& shape,
+ const py::bytes& opaque) -> XlaOp {
+ return CustomCall(builder, call_target_name, operands, shape, opaque);
+ },
+ py::arg("builder"), py::arg("call_target_name"), py::arg("operands"),
+ py::arg("shape"), py::arg("opaque") = py::bytes(""));
+ ops.def(
+ "CustomCallWithLayout",
+ [](XlaBuilder* builder, const py::bytes& call_target_name,
+ absl::Span<const XlaOp> operands, const Shape& shape_with_layout,
+ absl::Span<const Shape> operand_shapes_with_layout,
+ const py::bytes& opaque) -> XlaOp {
+ return CustomCallWithLayout(builder, call_target_name, operands,
+ shape_with_layout,
+ operand_shapes_with_layout, opaque);
+ },
+ py::arg("builder"), py::arg("call_target_name"), py::arg("operands"),
+ py::arg("shape_with_layout"), py::arg("operand_shapes_with_layout"),
+ py::arg("opaque") = py::bytes(""));
ops.def("Dot", &Dot, py::arg("lhs"), py::arg("rhs"),
py::arg("precision_config") = nullptr);
ops.def("DotGeneral", &DotGeneral, py::arg("lhs"), py::arg("rhs"),
@@ -388,7 +412,7 @@
ops.def("Gather", &Gather, py::arg("a"), py::arg("start_indices"),
py::arg("dimension_numbers"), py::arg("slice_sizes"),
- py::arg("indices_are_sorted"));
+ py::arg("indices_are_sorted") = false);
ops.def("GetTupleElement", &GetTupleElement, py::arg("tuple_data"),
py::arg("index"));
ops.def("InfeedWithToken", &InfeedWithToken, py::arg("token"),
@@ -401,7 +425,7 @@
py::arg("builder"), py::arg("type"), py::arg("size"));
ops.def("Map", &Map, py::arg("builder"), py::arg("operands"),
py::arg("computation"), py::arg("dimensions"),
- py::arg("static_operands"));
+ py::arg("static_operands") = py::list());
ops.def("NextAfter", &NextAfter, py::arg("from"), py::arg("to"));
ops.def("OutfeedWithToken", &OutfeedWithToken, py::arg("operand"),
py::arg("token"), py::arg("shape_with_layout"),
@@ -410,15 +434,11 @@
py::arg("padding_config"));
ops.def("Parameter",
static_cast<XlaOp (*)(XlaBuilder*, int64, const Shape&,
- const std::string&)>(&Parameter),
- py::arg("builder"), py::arg("parameter_number"), py::arg("shape"),
- py::arg("name"));
- ops.def("Parameter",
- static_cast<XlaOp (*)(XlaBuilder*, int64, const Shape&,
const std::string&, const std::vector<bool>&)>(
&Parameter),
py::arg("builder"), py::arg("parameter_number"), py::arg("shape"),
- py::arg("name"), py::arg("replicated_at_leaf_buffers"));
+ py::arg("name") = "",
+ py::arg("replicated_at_leaf_buffers") = std::vector<bool>());
ops.def(
"QR",
[](XlaOp a, bool full_matrices) -> StatusOr<std::pair<XlaOp, XlaOp>> {
@@ -486,8 +506,9 @@
py::arg("limit_index"), py::arg("stride"), py::arg("dimno"));
ops.def(
"Sort",
- [](XlaBuilder* builder, absl::Span<const XlaOp> operands, int64 dimension,
- absl::optional<const XlaComputation*> comparator) -> XlaOp {
+ [](XlaBuilder* builder, absl::Span<const XlaOp> operands,
+ absl::optional<const XlaComputation*> comparator, int64 dimension,
+ bool is_stable) -> XlaOp {
return builder->ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
std::vector<PrimitiveType> operand_types;
for (const auto& operand : operands) {
@@ -496,16 +517,17 @@
}
if (comparator) {
- return Sort(operands, **comparator, dimension);
+ return Sort(operands, **comparator, dimension, is_stable);
} else {
return Sort(operands,
CreateScalarLtComputation(operand_types, builder),
- dimension);
+ dimension, is_stable);
}
});
},
- py::arg("builder"), py::arg("operands"), py::arg("dimension") = -1,
- py::arg("comparator") = absl::nullopt);
+ py::arg("builder"), py::arg("operands"),
+ py::arg("comparator") = absl::nullopt, py::arg("dimension") = -1,
+ py::arg("is_stable") = false);
ops.def("TopK", &TopK, py::arg("input"), py::arg("k"));
ops.def("Transpose", &Transpose, py::arg("operand"), py::arg("permutation"));
ops.def("TriangularSolve", &TriangularSolve, py::arg("a"), py::arg("b"),
@@ -1399,13 +1421,6 @@
return WrapWithClient(std::move(client), std::move(buffer));
});
- py::enum_<TriangularSolveOptions::Transpose>(
- m, "TriangularSolveOptions_Transpose")
- .value("TRANSPOSE_INVALID", TriangularSolveOptions::TRANSPOSE_INVALID)
- .value("NO_TRANSPOSE", TriangularSolveOptions::NO_TRANSPOSE)
- .value("TRANSPOSE", TriangularSolveOptions::TRANSPOSE)
- .value("ADJOINT", TriangularSolveOptions::ADJOINT);
-
py::enum_<PrecisionConfig::Precision>(m, "PrecisionConfig_Precision")
.value("DEFAULT", PrecisionConfig::DEFAULT)
.value("HIGH", PrecisionConfig::HIGH)
diff --git a/tensorflow/compiler/xla/python/xla_client.py b/tensorflow/compiler/xla/python/xla_client.py
index fa887ec..be19a09 100644
--- a/tensorflow/compiler/xla/python/xla_client.py
+++ b/tensorflow/compiler/xla/python/xla_client.py
@@ -25,6 +25,7 @@
import inspect
import itertools
import os
+from typing import List, Sequence, Tuple, Union
from absl import logging
import numpy as np
@@ -42,6 +43,9 @@
# consistency with XLA.
# pylint: disable=invalid-name
+# Pylint has false positives for type annotations.
+# pylint: disable=invalid-sequence-index
+
profiler = _xla.profiler
@@ -77,7 +81,7 @@
"""Allocates a fresh buffer and populates it with `pyval`."""
@abc.abstractmethod
- def compile(self, computation, compile_options):
+ def compile(self, computation, compile_options=None):
"""Compiles a computation. Returns an executable."""
@abc.abstractmethod
@@ -133,7 +137,8 @@
return _xla.PyLocalBuffer.from_python(pyval, self.client, device,
force_copy)
- def compile(self, c_computation, compile_options):
+ def compile(self, c_computation, compile_options=None):
+ compile_options = compile_options or CompileOptions()
options = _xla.ExecutableBuildOptions()
options.num_replicas = compile_options.num_replicas
options.num_partitions = compile_options.num_partitions
@@ -654,8 +659,8 @@
SAME = 2
-def _convert_padding_type_to_pad_values(padding_type, lhs_dims, rhs_dims,
- window_strides):
+def window_padding_type_to_pad_values(padding_type, lhs_dims, rhs_dims,
+ window_strides):
"""Maps PaddingType or string to pad values (list of pairs of ints)."""
if not isinstance(padding_type, (str, PaddingType)):
msg = 'padding_type must be str or PaddingType, got {}.'
@@ -685,6 +690,9 @@
raise ValueError(msg.format(padding_type))
+XlaBuilder = _xla.XlaBuilder
+
+
class ComputationBuilder(object):
"""XLA computation builder.
@@ -987,9 +995,7 @@
Returns:
An XlaOp representing the added Pad op.
"""
- if isinstance(padding_config, tuple) or isinstance(padding_config, list):
- padding_config = GetPaddingConfigFromTriples(padding_config)
- return ops.Pad(operand, padding_value, padding_config)
+ return ops.Pad(operand, padding_value, make_padding_config(padding_config))
def Reshape(self, operand, dimensions, new_sizes):
"""Enqueues a reshape op onto the computation.
@@ -1022,7 +1028,7 @@
Returns:
An XlaOp that represents the all-reduced result.
"""
- replica_groups_protos = _get_replica_groups_protos(replica_groups)
+ replica_groups_protos = make_replica_groups(replica_groups)
return ops.AllReduce(operand, computation.computation,
replica_groups_protos, None, None)
@@ -1046,7 +1052,7 @@
Returns:
An XlaOp that represents the all-to-all concatenation.
"""
- replica_groups_protos = _get_replica_groups_protos(replica_groups)
+ replica_groups_protos = make_replica_groups(replica_groups)
if not replica_groups:
split_count = 1
else:
@@ -1069,7 +1075,7 @@
Returns:
An XlaOp that represents on each replica the sum of its group's values.
"""
- replica_groups_protos = _get_replica_groups_protos(replica_groups)
+ replica_groups_protos = make_replica_groups(replica_groups)
return ops.CrossReplicaSum(operand, replica_groups_protos)
def Trans(self, operand):
@@ -1100,7 +1106,7 @@
Returns:
An XlaOp representing the added SelectAndScatter op.
"""
- pads = _convert_padding_type_to_pad_values(
+ pads = window_padding_type_to_pad_values(
padding,
self.GetShape(operand).dimensions(), window_dimensions, window_strides)
return ops.SelectAndScatterWithGeneralPadding(operand, select.computation,
@@ -1218,12 +1224,16 @@
An XlaOp representing the added custom call op.
"""
opaque = opaque or b''
- return ops.CustomCallWithLayout(
- self._builder, call_target_name, list(operands), shape_with_layout,
- list(operand_shapes_with_layout), opaque)
+ return ops.CustomCallWithLayout(self._builder, call_target_name,
+ list(operands), shape_with_layout,
+ list(operand_shapes_with_layout), opaque)
- def CustomCall(self, call_target_name, operands, shape,
- operand_shapes_with_layout=None, opaque=None):
+ def CustomCall(self,
+ call_target_name,
+ operands,
+ shape,
+ operand_shapes_with_layout=None,
+ opaque=None):
"""Enqueues a custom call operation onto the computation.
Args:
@@ -1244,9 +1254,9 @@
return ops.CustomCall(self._builder, call_target_name, list(operands),
shape, opaque)
else:
- return ops.CustomCallWithLayout(
- self._builder, call_target_name, list(operands), shape,
- list(operand_shapes_with_layout), opaque)
+ return ops.CustomCallWithLayout(self._builder, call_target_name,
+ list(operands), shape,
+ list(operand_shapes_with_layout), opaque)
def Map(self, operands, computation_to_apply, dimensions):
"""Enqueues a map operation onto the computation.
@@ -1292,7 +1302,7 @@
Returns:
An XlaOp representing the added ReduceWindow op.
"""
- pads = _convert_padding_type_to_pad_values(
+ pads = window_padding_type_to_pad_values(
padding,
self.GetShape(operand).dimensions(), window_dimensions, window_strides)
return ops.ReduceWindowWithGeneralPadding(operand, init_value,
@@ -1413,8 +1423,7 @@
and batch dimensions on each input operand.
Returns: a XlaOp representing the DotGeneral operation.
"""
- if isinstance(dimension_numbers, tuple):
- dimension_numbers = GetDotDimensionsFromLists(dimension_numbers)
+ dimension_numbers = make_dot_dimension_numbers(dimension_numbers)
return ops.DotGeneral(
lhs, rhs, dimension_numbers, precision_config=precision_config)
@@ -1437,7 +1446,7 @@
batch_group_count: number of batch groups for grouped convolution.
Returns: a XlaOp representing the Conv operation.
"""
- pads = _convert_padding_type_to_pad_values(
+ pads = window_padding_type_to_pad_values(
padding,
self.GetShape(lhs).dimensions()[2:],
self.GetShape(rhs).dimensions()[2:], window_strides)
@@ -1488,21 +1497,6 @@
batch_group_count=batch_group_count,
precision_config=precision_config)
- def _GetConvDimensionNumbers(self, num_spatial_dims):
- """Create ConvolutionDimensionNumbers proto for convolutions."""
- nd = num_spatial_dims
- dimension_numbers = ConvolutionDimensionNumbers()
- dimension_numbers.input_batch_dimension = 0
- dimension_numbers.input_feature_dimension = 1
- dimension_numbers.output_batch_dimension = 0
- dimension_numbers.output_feature_dimension = 1
- dimension_numbers.kernel_output_feature_dimension = 0
- dimension_numbers.kernel_input_feature_dimension = 1
- dimension_numbers.input_spatial_dimensions.extend(range(2, 2 + nd))
- dimension_numbers.kernel_spatial_dimensions.extend(range(2, 2 + nd))
- dimension_numbers.output_spatial_dimensions.extend(range(2, 2 + nd))
- return dimension_numbers
-
def ConvGeneralDilated(self,
lhs,
rhs,
@@ -1546,27 +1540,6 @@
batch_group_count: number of batch groups for grouped convolution.
Returns: a XlaOp representing the ConvGeneralDilated operation.
"""
- if dimension_numbers is None:
- dimension_numbers = self._GetConvDimensionNumbers(len(window_strides))
- elif isinstance(dimension_numbers, tuple):
- lhs_spec, rhs_spec, out_spec = dimension_numbers
- dimension_numbers = ConvolutionDimensionNumbers()
-
- dimension_numbers.input_batch_dimension = lhs_spec.index('N')
- dimension_numbers.input_feature_dimension = lhs_spec.index('C')
- dimension_numbers.output_batch_dimension = out_spec.index('N')
- dimension_numbers.output_feature_dimension = out_spec.index('C')
- dimension_numbers.kernel_output_feature_dimension = rhs_spec.index('O')
- dimension_numbers.kernel_input_feature_dimension = rhs_spec.index('I')
-
- dimension_numbers.kernel_spatial_dimensions.extend(
- i for i, c in enumerate(rhs_spec) if c not in {'I', 'O'})
- dimension_numbers.input_spatial_dimensions.extend(
- sorted((i for i, c in enumerate(lhs_spec) if c not in {'N', 'C'}),
- key=lambda i: rhs_spec.index(lhs_spec[i])))
- dimension_numbers.output_spatial_dimensions.extend(
- sorted((i for i, c in enumerate(out_spec) if c not in {'N', 'C'}),
- key=lambda i: rhs_spec.index(out_spec[i])))
return ops.ConvGeneralDilated(
lhs,
rhs,
@@ -1574,12 +1547,13 @@
padding,
lhs_dilation,
rhs_dilation,
- dimension_numbers,
+ make_convolution_dimension_numbers(dimension_numbers,
+ len(window_strides)),
feature_group_count,
batch_group_count,
precision_config=precision_config)
- def Sort(self, operands, dimension=-1, comparator=None):
+ def Sort(self, operands, dimension=-1, comparator=None, is_stable=False):
"""Enqueues a sort operation onto the computation.
Args:
@@ -1596,15 +1570,16 @@
operands = (
list(operands)
if isinstance(operands, collections.abc.Sequence) else [operands])
- return ops.Sort(self._builder, operands, dimension,
- comparator.computation if comparator else None)
+ return ops.Sort(self._builder, operands,
+ comparator.computation if comparator else None, dimension,
+ is_stable)
- def SortKeyVal(self, keys, values, dimension=-1):
+ def SortKeyVal(self, keys, values, dimension=-1, is_stable=False):
"""Enqueues a key-value sort operation onto the computation.
Deprecated. Use `Sort` instead.
"""
- return ops.Sort(self._builder, [keys, values], dimension)
+ return ops.Sort(self._builder, [keys, values], None, dimension, is_stable)
def QR(self, a, full_matrices=True):
"""Enqueues a QR decomposition onto the computation."""
@@ -1620,13 +1595,13 @@
unit_diagonal=False):
"""Enqueues a triangular-solve operation onto the computation."""
if not transpose_a:
- transpose = _xla.TriangularSolveOptions_Transpose.NO_TRANSPOSE
+ transpose = ops.TriangularSolveOptions_Transpose.NO_TRANSPOSE
if conjugate_a:
a = self.Conj(a)
else:
transpose = (
- _xla.TriangularSolveOptions_Transpose.ADJOINT
- if conjugate_a else _xla.TriangularSolveOptions_Transpose.TRANSPOSE)
+ ops.TriangularSolveOptions_Transpose.ADJOINT
+ if conjugate_a else ops.TriangularSolveOptions_Transpose.TRANSPOSE)
return ops.TriangularSolve(a, b, left_side, lower, unit_diagonal, transpose)
def Eigh(self, a, full_matrices=True):
@@ -1817,15 +1792,28 @@
self.dimensions = []
-def GetPaddingConfigFromTriples(triples):
- """Create PaddingConfig proto from list of triples of integers."""
- padding_config = PaddingConfig()
- for lo, hi, interior in triples:
- dimension = PaddingConfigDimension()
- dimension.edge_padding_low = lo
- dimension.edge_padding_high = hi
- dimension.interior_padding = interior
- padding_config.dimensions.append(dimension)
+def make_padding_config(
+ padding_config: Union[PaddingConfig, Sequence[Tuple[int, int, int]]]
+) -> PaddingConfig:
+ """Create PaddingConfig proto from list of triples of integers.
+
+ Args:
+ padding_config: either a PaddingConfig or a list of integer triples
+ (edge_padding_low, edge_padding_high, interior_padding) representing the
+ configuration of the padding operation.
+
+ Returns:
+ A `PaddingConfig` object.
+ """
+ if isinstance(padding_config, tuple) or isinstance(padding_config, list):
+ triples = padding_config
+ padding_config = PaddingConfig()
+ for lo, hi, interior in triples:
+ dimension = PaddingConfigDimension()
+ dimension.edge_padding_low = lo
+ dimension.edge_padding_high = hi
+ dimension.interior_padding = interior
+ padding_config.dimensions.append(dimension)
return padding_config
@@ -1841,14 +1829,32 @@
self.rhs_batch_dimensions = []
-def GetDotDimensionsFromLists(dimension_numbers):
- (lhs_contract, rhs_contract), (lhs_batch, rhs_batch) = dimension_numbers
- dot_dims_proto = DotDimensionNumbers()
- dot_dims_proto.lhs_contracting_dimensions.extend(lhs_contract)
- dot_dims_proto.rhs_contracting_dimensions.extend(rhs_contract)
- dot_dims_proto.lhs_batch_dimensions.extend(lhs_batch)
- dot_dims_proto.rhs_batch_dimensions.extend(rhs_batch)
- return dot_dims_proto
+def make_dot_dimension_numbers(
+ dimension_numbers: Union[DotDimensionNumbers,
+ Tuple[Tuple[List[int], List[int]],
+ Tuple[List[int], List[int]]]]
+) -> DotDimensionNumbers:
+ """Builds a DotDimensionNumbers object from a specification.
+
+ Args:
+ dimension_numbers: either a `DotDimensionNumbers` or a nested tuple
+ `((lhs_contract, rhs_contract), (lhs_batch, rhs_batch))` of lists of
+ integers representing the dimensions to treat as contracting dimensions
+ and batch dimensions on each input operand.
+
+ Returns:
+ A `DotDimensionNumbers` object.
+ """
+ if isinstance(dimension_numbers, (list, tuple)):
+ (lhs_contract, rhs_contract), (lhs_batch, rhs_batch) = dimension_numbers
+ dot_dims_proto = DotDimensionNumbers()
+ dot_dims_proto.lhs_contracting_dimensions.extend(lhs_contract)
+ dot_dims_proto.rhs_contracting_dimensions.extend(rhs_contract)
+ dot_dims_proto.lhs_batch_dimensions.extend(lhs_batch)
+ dot_dims_proto.rhs_batch_dimensions.extend(rhs_batch)
+ return dot_dims_proto
+ else:
+ return dimension_numbers
class ConvolutionDimensionNumbers(object):
@@ -1871,6 +1877,70 @@
self.output_spatial_dimensions = []
+def make_convolution_dimension_numbers(
+ dimension_numbers: Union[None, ConvolutionDimensionNumbers, Tuple[str, str,
+ str]],
+ num_spatial_dimensions: int) -> ConvolutionDimensionNumbers:
+ """Builds a ConvolutionDimensionNumbers object from a specification.
+
+ Args:
+ dimension_numbers: optional, either a ConvolutionDimensionNumbers object or
+ a tuple (lhs_spec, rhs_spec, out_spec). Each element is a string of
+ length N+2 identifying by position: (1) batch dimensions in lhs, rhs, and
+ the output with the character 'N', (2) feature dimensions in lhs and the
+ output with the character 'C', (3) input and output feature dimensions
+ in rhs with the characters 'I' and 'O' respectively, and (4) spatial
+ dimension correspondences between lhs, rhs, and the output using any
+ distinct characters. For example, to indicate dimension numbers
+ consistent with the Conv operation with two spatial dimensions, one
+ could use ('NCHW', 'OIHW', 'NCHW'). As another example, to indicate
+ dimension numbers consistent with the TensorFlow Conv2D operation, one
+ could use ('NHWC', 'HWIO', 'NHWC'). When using the latter form of
+ convolution dimension specification, window strides are associated with
+ spatial dimension character labels according to the order in which the
+ labels appear in the rhs_spec string, so that window_strides[0] is
+ matched with the dimension corresponding to the first character
+ appearing in rhs_spec that is not 'I' or 'O'. By default, use the same
+ dimension numbering as Conv and ConvWithGeneralPadding.
+ num_spatial_dimensions: the number of spatial dimensions.
+
+ Returns:
+ A `ConvolutionDimensionNumbers` object.
+ """
+ if dimension_numbers is None:
+ nd = num_spatial_dimensions
+ dimension_numbers = ConvolutionDimensionNumbers()
+ dimension_numbers.input_batch_dimension = 0
+ dimension_numbers.input_feature_dimension = 1
+ dimension_numbers.output_batch_dimension = 0
+ dimension_numbers.output_feature_dimension = 1
+ dimension_numbers.kernel_output_feature_dimension = 0
+ dimension_numbers.kernel_input_feature_dimension = 1
+ dimension_numbers.input_spatial_dimensions.extend(range(2, 2 + nd))
+ dimension_numbers.kernel_spatial_dimensions.extend(range(2, 2 + nd))
+ dimension_numbers.output_spatial_dimensions.extend(range(2, 2 + nd))
+ elif isinstance(dimension_numbers, tuple):
+ lhs_spec, rhs_spec, out_spec = dimension_numbers
+ dimension_numbers = ConvolutionDimensionNumbers()
+
+ dimension_numbers.input_batch_dimension = lhs_spec.index('N')
+ dimension_numbers.input_feature_dimension = lhs_spec.index('C')
+ dimension_numbers.output_batch_dimension = out_spec.index('N')
+ dimension_numbers.output_feature_dimension = out_spec.index('C')
+ dimension_numbers.kernel_output_feature_dimension = rhs_spec.index('O')
+ dimension_numbers.kernel_input_feature_dimension = rhs_spec.index('I')
+
+ dimension_numbers.kernel_spatial_dimensions.extend(
+ i for i, c in enumerate(rhs_spec) if c not in {'I', 'O'})
+ dimension_numbers.input_spatial_dimensions.extend(
+ sorted((i for i, c in enumerate(lhs_spec) if c not in {'N', 'C'}),
+ key=lambda i: rhs_spec.index(lhs_spec[i])))
+ dimension_numbers.output_spatial_dimensions.extend(
+ sorted((i for i, c in enumerate(out_spec) if c not in {'N', 'C'}),
+ key=lambda i: rhs_spec.index(out_spec[i])))
+ return dimension_numbers
+
+
class OpSharding(object):
"""Python representation of a xla.OpSharding protobuf."""
__slots__ = ('type', 'tile_assignment_dimensions', 'tile_assignment_devices',
@@ -1933,7 +2003,7 @@
return replica_group_proto
-def _get_replica_groups_protos(replica_groups):
+def make_replica_groups(replica_groups):
if replica_groups is None:
replica_groups_protos = [] # special value for XLA API
else:
diff --git a/tensorflow/compiler/xla/python/xla_client_test.py b/tensorflow/compiler/xla/python/xla_client_test.py
index e40006e..0e7dfb6 100644
--- a/tensorflow/compiler/xla/python/xla_client_test.py
+++ b/tensorflow/compiler/xla/python/xla_client_test.py
@@ -24,2136 +24,2025 @@
import threading
import unittest
+from absl import flags
from absl.testing import absltest
from absl.testing import parameterized
import numpy as np
-from tensorflow.compiler.xla.python import custom_call_for_test
from tensorflow.compiler.xla.python import xla_client
# pylint: disable=g-import-not-at-top
try:
+ from tensorflow.compiler.xla.python import custom_call_for_test
+except ImportError:
+ custom_call_for_test = None
+
+try:
import portpicker
except ImportError:
portpicker = None
# pylint: enable=g-import-not-at-top
bfloat16 = xla_client.bfloat16
+ops = xla_client.ops
+FLAGS = flags.FLAGS
-class ComputationTest(absltest.TestCase):
- """Base class for running an XLA Computation through the local client."""
+# We choose to ignore pylint's complaints about complex comprehensions, which we
+# use widely for parameterizing tests.
+# pylint: disable=g-complex-comprehension
- def _NewComputation(self, name=None):
- if name is None:
- name = self.id()
- return xla_client.ComputationBuilder(name)
- def _Execute(self, c, arguments):
- compiled_c = c.Build().Compile()
- return xla_client.execute_with_python_values(compiled_c, arguments)
+def TestFactory(xla_backend, cloud_tpu=False):
+ tests = []
- def _ExecuteAndAssertWith(self, assert_func, c, arguments, expected):
- assert expected is not None
- results = self._Execute(c, arguments)
- self.assertLen(results, len(expected))
- for result, e in zip(results, expected):
- # Numpy's comparison methods are a bit too lenient by treating inputs as
- # "array-like", meaning that scalar 4 will be happily compared equal to
- # [[4]]. We'd like to be more strict so assert shapes as well.
- self.assertEqual(np.asanyarray(result).shape, np.asanyarray(e).shape)
- assert_func(result, e)
+ if not cloud_tpu:
+ int_dtypes = [np.int32, np.int64, np.uint32, np.uint64]
+ # TODO(phawkins): test np.float16, where supported.
+ float_dtypes = [bfloat16, np.float32, np.float64]
+ complex_dtypes = [np.complex64, np.complex128]
+ standard_dtypes = int_dtypes + float_dtypes + complex_dtypes + [np.bool_]
+ else:
+ int_dtypes = [np.int32, np.uint32]
+ float_dtypes = [np.float32]
+ complex_dtypes = [np.complex64]
+ standard_dtypes = int_dtypes + float_dtypes + complex_dtypes + [np.bool_]
+ dlpack_dtypes = int_dtypes + float_dtypes
- def _ExecuteAndCompareExact(self, c, arguments=(), expected=None):
- self._ExecuteAndAssertWith(np.testing.assert_equal, c, arguments, expected)
+ class ComputationTest(parameterized.TestCase):
+ """Base class for running an XLA Computation through the local client."""
- def _ExecuteAndCompareClose(self,
- c,
- arguments=(),
- expected=None,
- rtol=1e-7,
- atol=0):
- self._ExecuteAndAssertWith(
- functools.partial(np.testing.assert_allclose, rtol=rtol, atol=atol), c,
- arguments, expected)
+ def setUp(self):
+ super(ComputationTest, self).setUp()
+ self.backend = xla_backend()
+ def _NewComputation(self, name=None):
+ if name is None:
+ name = self.id()
+ return xla_client.XlaBuilder(name)
-def NumpyArrayF32(*args, **kwargs):
- """Convenience wrapper to create Numpy arrays with a np.float32 dtype."""
- return np.array(*args, dtype=np.float32, **kwargs)
+ def _Execute(self, c, arguments):
+ compiled_c = self.backend.compile(c.Build())
+ return xla_client.execute_with_python_values(
+ compiled_c, arguments, backend=self.backend)
+ def _ExecuteAndAssertWith(self, assert_func, c, arguments, expected):
+ assert expected is not None
+ results = self._Execute(c, arguments)
+ self.assertLen(results, len(expected))
+ for result, e in zip(results, expected):
+ # Numpy's comparison methods are a bit too lenient by treating inputs as
+ # "array-like", meaning that scalar 4 will be happily compared equal to
+ # [[4]]. We'd like to be more strict so assert shapes as well.
+ self.assertEqual(np.asanyarray(result).shape, np.asanyarray(e).shape)
+ assert_func(result, e)
-def NumpyArrayF64(*args, **kwargs):
- """Convenience wrapper to create Numpy arrays with a np.float64 dtype."""
- return np.array(*args, dtype=np.float64, **kwargs)
+ def _ExecuteAndCompareExact(self, c, arguments=(), expected=None):
+ self._ExecuteAndAssertWith(np.testing.assert_equal, c, arguments,
+ expected)
+ def _ExecuteAndCompareClose(self,
+ c,
+ arguments=(),
+ expected=None,
+ rtol=1e-7,
+ atol=0):
+ self._ExecuteAndAssertWith(
+ functools.partial(np.testing.assert_allclose, rtol=rtol, atol=atol),
+ c, arguments, expected)
-def NumpyArrayS32(*args, **kwargs):
- """Convenience wrapper to create Numpy arrays with a np.int32 dtype."""
- return np.array(*args, dtype=np.int32, **kwargs)
+ def NumpyArrayF32(*args, **kwargs):
+ """Convenience wrapper to create Numpy arrays with a np.float32 dtype."""
+ return np.array(*args, dtype=np.float32, **kwargs)
+ def NumpyArrayS32(*args, **kwargs):
+ """Convenience wrapper to create Numpy arrays with a np.int32 dtype."""
+ return np.array(*args, dtype=np.int32, **kwargs)
-def NumpyArrayS64(*args, **kwargs):
- """Convenience wrapper to create Numpy arrays with a np.int64 dtype."""
- return np.array(*args, dtype=np.int64, **kwargs)
+ def NumpyArrayBool(*args, **kwargs):
+ """Convenience wrapper to create Numpy arrays with a np.bool dtype."""
+ return np.array(*args, dtype=np.bool, **kwargs)
+ class ComputationPrinting(absltest.TestCase):
-def NumpyArrayBool(*args, **kwargs):
- """Convenience wrapper to create Numpy arrays with a np.bool dtype."""
- return np.array(*args, dtype=np.bool, **kwargs)
+ def setUp(self):
+ super(ComputationPrinting, self).setUp()
+ self.backend = xla_backend()
+ def ExampleComputation(self):
+ builder = xla_client.XlaBuilder("acomputation")
+ p0 = ops.Parameter(builder, 0, xla_client.shape_from_pyval(np.float32(0)))
+ p1 = ops.Parameter(
+ builder, 1, xla_client.shape_from_pyval(np.zeros((4,), np.float32)))
+ x = ops.Mul(p0, p1)
+ ops.Add(x, x)
+ return builder.Build()
-class ComputationPrinting(absltest.TestCase):
+ def testComputationToHloText(self):
+ computation = self.ExampleComputation()
+ hlo_text = computation.GetHloText()
+ self.assertTrue(hlo_text.startswith("HloModule acomputation"))
- def ExampleComputation(self):
- builder = xla_client.ComputationBuilder("acomputation")
- p0 = builder.ParameterFromNumpy(np.float32(0))
- p1 = builder.ParameterFromNumpy(np.zeros((4,), np.float32))
- x = builder.Mul(p0, p1)
- builder.Add(x, x)
- return builder.Build()
+ def testComputationToHloGraph(self):
+ computation = self.ExampleComputation()
+ hlo_dot_graph = computation.GetHloDotGraph()
+ self.assertTrue(hlo_dot_graph.startswith("digraph "))
- def testComputationToHloText(self):
- computation = self.ExampleComputation()
- hlo_text = computation.GetHloText()
- self.assertTrue(hlo_text.startswith("HloModule acomputation"))
+ def testHloModuleToHloText(self):
+ computation = self.ExampleComputation()
+ hlo_text = computation.get_hlo_module().to_string()
+ self.assertTrue(hlo_text.startswith("HloModule acomputation"))
- def testComputationToHloGraph(self):
- computation = self.ExampleComputation()
- hlo_dot_graph = computation.GetHloDotGraph()
- self.assertTrue(hlo_dot_graph.startswith("digraph "))
+ def testHloModuleToHloGraph(self):
+ computation = self.ExampleComputation()
+ hlo_dot_graph = xla_client._xla.hlo_module_to_dot_graph(
+ computation.get_hlo_module())
+ self.assertTrue(hlo_dot_graph.startswith("digraph "))
- def testHloModuleToHloText(self):
- computation = self.ExampleComputation()
- hlo_text = computation.computation.get_hlo_module().to_string()
- self.assertTrue(hlo_text.startswith("HloModule acomputation"))
+ @unittest.skipIf(cloud_tpu, "not implemented")
+ def testCompiledHloModuleToHloText(self):
+ computation = self.ExampleComputation()
+ executable = self.backend.compile(computation)
+ hlo_modules = executable.get_hlo_modules()
+ self.assertLen(hlo_modules, 1)
+ hlo_text = hlo_modules[0].to_string()
+ self.assertTrue(hlo_text.startswith("HloModule acomputation"))
+ self.assertIn("fusion", hlo_text)
- def testHloModuleToHloGraph(self):
- computation = self.ExampleComputation()
- hlo_dot_graph = xla_client._xla.hlo_module_to_dot_graph(
- computation.computation.get_hlo_module())
- self.assertTrue(hlo_dot_graph.startswith("digraph "))
+ tests.append(ComputationPrinting)
- def testCompiledHloModuleToHloText(self):
- computation = self.ExampleComputation()
- executable = computation.Compile()
- hlo_modules = executable.get_hlo_modules()
- self.assertLen(hlo_modules, 1)
- hlo_text = hlo_modules[0].to_string()
- self.assertTrue(hlo_text.startswith("HloModule acomputation"))
- self.assertIn("fusion", hlo_text)
+ class ComputationHashTest(absltest.TestCase):
+ def testHash(self):
+ builder0 = xla_client.XlaBuilder("computation0")
+ p0 = ops.Parameter(builder0, 0,
+ xla_client.shape_from_pyval(np.float32(0)))
+ p1 = ops.Parameter(
+ builder0, 1, xla_client.shape_from_pyval(np.zeros((4,), np.float32)))
+ ops.Mul(p0, p1)
+ computation0 = builder0.Build()
-class ComputationHashTest(absltest.TestCase):
+ builder1 = xla_client.XlaBuilder("computation1")
+ p0 = ops.Parameter(builder1, 0,
+ xla_client.shape_from_pyval(np.float32(0)))
+ p1 = ops.Parameter(
+ builder1, 1, xla_client.shape_from_pyval(np.zeros((4,), np.float32)))
+ ops.Mul(p0, p1)
+ computation1 = builder1.Build()
- def testHash(self):
- builder0 = xla_client.ComputationBuilder("computation0")
- p0 = builder0.ParameterFromNumpy(np.float32(0))
- p1 = builder0.ParameterFromNumpy(np.zeros((4,), np.float32))
- builder0.Mul(p0, p1)
- computation0 = builder0.Build()
+ self.assertEqual(computation0.Hash(), computation1.Hash())
- builder1 = xla_client.ComputationBuilder("computation1")
- p0 = builder1.ParameterFromNumpy(np.float32(0))
- p1 = builder1.ParameterFromNumpy(np.zeros((4,), np.float32))
- builder1.Mul(p0, p1)
- computation1 = builder1.Build()
+ tests.append(ComputationHashTest)
- self.assertEqual(computation0.Hash(), computation1.Hash())
+ class ComputationsWithConstantsTest(ComputationTest):
+ """Tests focusing on Constant ops."""
-
-class ComputationsWithConstantsTest(ComputationTest):
- """Tests focusing on Constant ops."""
-
- def testConstantScalarSumS8(self):
- c = self._NewComputation()
- c.Add(c.Constant(np.int8(1)), c.Constant(np.int8(2)))
- self._ExecuteAndCompareExact(c, expected=[np.int8(3)])
-
- def testConstantScalarSumBF16(self):
- c = self._NewComputation()
- c.Add(c.Constant(bfloat16(1.11)), c.Constant(bfloat16(3.14)))
- self._ExecuteAndCompareClose(c, expected=[bfloat16(4.25)])
-
- def testConstantScalarSumF32(self):
- c = self._NewComputation()
- c.Add(c.ConstantF32Scalar(1.11), c.ConstantF32Scalar(3.14))
- self._ExecuteAndCompareClose(c, expected=[4.25])
-
- def testConstantScalarSumF64(self):
- c = self._NewComputation()
- c.Add(c.ConstantF64Scalar(1.11), c.ConstantF64Scalar(3.14))
- self._ExecuteAndCompareClose(c, expected=[4.25])
-
- def testConstantScalarSumS32(self):
- c = self._NewComputation()
- c.Add(c.ConstantS32Scalar(1), c.ConstantS32Scalar(2))
- self._ExecuteAndCompareClose(c, expected=[3])
-
- def testConstantScalarSumS64(self):
- c = self._NewComputation()
- c.Add(c.ConstantS64Scalar(1), c.ConstantS64Scalar(2))
- self._ExecuteAndCompareClose(c, expected=[3])
-
- def testConstantVectorMulF16(self):
- c = self._NewComputation()
- c.Mul(
- c.Constant(np.array([2.5, 3.3, -1.2, 0.7], np.float16)),
- c.Constant(np.array([-1.2, 2, -2, -3], np.float16)))
- self._ExecuteAndCompareClose(
- c, expected=[np.array([-3, 6.6, 2.4, -2.1], np.float16)], rtol=2e-3)
-
- def testConstantVectorMulF32(self):
- c = self._NewComputation()
- c.Mul(
- c.Constant(NumpyArrayF32([2.5, 3.3, -1.2, 0.7])),
- c.Constant(NumpyArrayF32([-1.2, 2, -2, -3])))
- self._ExecuteAndCompareClose(c, expected=[[-3, 6.6, 2.4, -2.1]])
-
- def testConstantVectorMulF64(self):
- c = self._NewComputation()
- c.Mul(
- c.Constant(NumpyArrayF64([2.5, 3.3, -1.2, 0.7])),
- c.Constant(NumpyArrayF64([-1.2, 2, -2, -3])))
- self._ExecuteAndCompareClose(c, expected=[[-3, 6.6, 2.4, -2.1]])
-
- def testConstantVectorScalarDivF32(self):
- c = self._NewComputation()
- c.Div(
- c.Constant(NumpyArrayF32([1.5, 2.5, 3.0, -10.8])),
- c.ConstantF32Scalar(2.0))
- self._ExecuteAndCompareClose(c, expected=[[0.75, 1.25, 1.5, -5.4]])
-
- def testConstantVectorScalarDivF64(self):
- c = self._NewComputation()
- c.Div(
- c.Constant(NumpyArrayF64([1.5, 2.5, 3.0, -10.8])),
- c.ConstantF64Scalar(2.0))
- self._ExecuteAndCompareClose(c, expected=[[0.75, 1.25, 1.5, -5.4]])
-
- def testConstantVectorScalarPowF32(self):
- c = self._NewComputation()
- c.Pow(c.Constant(NumpyArrayF32([1.5, 2.5, 3.0])), c.ConstantF32Scalar(2.))
- self._ExecuteAndCompareClose(c, expected=[[2.25, 6.25, 9.]])
-
- def testConstantVectorScalarPowF64(self):
- c = self._NewComputation()
- c.Pow(c.Constant(NumpyArrayF64([1.5, 2.5, 3.0])), c.ConstantF64Scalar(2.))
- self._ExecuteAndCompareClose(c, expected=[[2.25, 6.25, 9.]])
-
- def testIota(self):
- c = self._NewComputation()
- c.Iota(np.float32, 10)
- self._ExecuteAndCompareExact(c, expected=[np.arange(10, dtype=np.float32)])
-
- def testBroadcastedIota(self):
- c = self._NewComputation()
- c.BroadcastedIota(np.int64, (2, 3), 1)
- expected = np.array([[0, 1, 2], [0, 1, 2]], dtype=np.int64)
- self._ExecuteAndCompareExact(c, expected=[expected])
-
- def testBooleanAnd(self):
- c = self._NewComputation()
- c.And(
- c.Constant(NumpyArrayBool([True, False, True, False])),
- c.Constant(NumpyArrayBool([True, True, False, False])))
- self._ExecuteAndCompareExact(c, expected=[[True, False, False, False]])
-
- def testBooleanOr(self):
- c = self._NewComputation()
- c.Or(
- c.Constant(NumpyArrayBool([True, False, True, False])),
- c.Constant(NumpyArrayBool([True, True, False, False])))
- self._ExecuteAndCompareExact(c, expected=[[True, True, True, False]])
-
- def testBooleanXor(self):
- c = self._NewComputation()
- c.Xor(
- c.Constant(NumpyArrayBool([True, False, True, False])),
- c.Constant(NumpyArrayBool([True, True, False, False])))
- self._ExecuteAndCompareExact(c, expected=[[False, True, True, False]])
-
- def testSum2DF32(self):
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF32([[1, 2, 3], [4, 5, 6]])),
- c.Constant(NumpyArrayF32([[1, -1, 1], [-1, 1, -1]])))
- self._ExecuteAndCompareClose(c, expected=[[[2, 1, 4], [3, 6, 5]]])
-
- def testShiftLeft(self):
- c = self._NewComputation()
- c.ShiftLeft(c.Constant(NumpyArrayS32([3])), c.Constant(NumpyArrayS32([2])))
- self._ExecuteAndCompareClose(c, expected=[[12]])
-
- def testShiftRightArithmetic(self):
- c = self._NewComputation()
- c.ShiftRightArithmetic(
- c.Constant(NumpyArrayS32([-2])), c.Constant(NumpyArrayS32([1])))
- self._ExecuteAndCompareClose(c, expected=[[-1]])
-
- def testShiftRightLogical(self):
- c = self._NewComputation()
- c.ShiftRightLogical(
- c.Constant(NumpyArrayS32([-1])), c.Constant(NumpyArrayS32([1])))
- self._ExecuteAndCompareClose(c, expected=[[2**31 - 1]])
-
- def testSum2DF64(self):
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF64([[1, 2, 3], [4, 5, 6]])),
- c.Constant(NumpyArrayF64([[1, -1, 1], [-1, 1, -1]])))
- self._ExecuteAndCompareClose(c, expected=[[[2, 1, 4], [3, 6, 5]]])
-
- def testSum2DWith1DBroadcastDim0F32(self):
- # sum of a 2D array with a 1D array where the latter is replicated across
- # dimension 0 to match the former's shape.
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF32([10, 20, 30])),
- broadcast_dimensions=(0,))
- self._ExecuteAndCompareClose(
- c, expected=[[[11, 12, 13], [24, 25, 26], [37, 38, 39]]])
-
- def testSum2DWith1DBroadcastDim0F64(self):
- # sum of a 2D array with a 1D array where the latter is replicated across
- # dimension 0 to match the former's shape.
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF64([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF64([10, 20, 30])),
- broadcast_dimensions=(0,))
- self._ExecuteAndCompareClose(
- c, expected=[[[11, 12, 13], [24, 25, 26], [37, 38, 39]]])
-
- def testSum2DWith1DBroadcastDim1F32(self):
- # sum of a 2D array with a 1D array where the latter is replicated across
- # dimension 1 to match the former's shape.
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF32([10, 20, 30])),
- broadcast_dimensions=(1,))
- self._ExecuteAndCompareClose(
- c, expected=[[[11, 22, 33], [14, 25, 36], [17, 28, 39]]])
-
- def testSum2DWith1DBroadcastDim1F64(self):
- # sum of a 2D array with a 1D array where the latter is replicated across
- # dimension 1 to match the former's shape.
- c = self._NewComputation()
- c.Add(
- c.Constant(NumpyArrayF64([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF64([10, 20, 30])),
- broadcast_dimensions=(1,))
- self._ExecuteAndCompareClose(
- c, expected=[[[11, 22, 33], [14, 25, 36], [17, 28, 39]]])
-
- def testConstantAxpyF32(self):
- c = self._NewComputation()
- c.Add(
- c.Mul(
- c.ConstantF32Scalar(2),
- c.Constant(NumpyArrayF32([2.2, 3.3, 4.4, 5.5]))),
- c.Constant(NumpyArrayF32([100, -100, 200, -200])))
- self._ExecuteAndCompareClose(c, expected=[[104.4, -93.4, 208.8, -189]])
-
- def testConstantAxpyF64(self):
- c = self._NewComputation()
- c.Add(
- c.Mul(
- c.ConstantF64Scalar(2),
- c.Constant(NumpyArrayF64([2.2, 3.3, 4.4, 5.5]))),
- c.Constant(NumpyArrayF64([100, -100, 200, -200])))
- self._ExecuteAndCompareClose(c, expected=[[104.4, -93.4, 208.8, -189]])
-
- def testCustomCall(self):
- c = self._NewComputation()
- for name, fn in custom_call_for_test.cpu_custom_call_targets.items():
- xla_client.register_custom_call_target(name, fn, platform="cpu")
- c.CustomCall(
- b"test_subtract_f32",
- operands=(c.ConstantF32Scalar(1.25), c.ConstantF32Scalar(0.5)),
- shape=xla_client.Shape.array_shape(np.dtype(np.float32), (), ()),
- operand_shapes_with_layout=(
- xla_client.Shape.array_shape(np.dtype(np.float32), (), ()),
- xla_client.Shape.array_shape(np.dtype(np.float32), (), ()),
- ))
- self._ExecuteAndCompareClose(c, expected=[0.75])
-
-
-class ComputationFromProtoTest(absltest.TestCase):
- """Test computation execution from HLO proto."""
-
- def testExecuteFromProto(self):
- # Build the HLO proto
- b = xla_client.ComputationBuilder("computation")
- b.Add(b.Constant(np.int8(1)), b.Constant(np.int8(2)))
- serialized_proto = b.Build().GetSerializedProto()
-
- # Load and execute the proto
- c = xla_client.Computation(xla_client._xla.XlaComputation(serialized_proto))
- ans, = xla_client.execute_with_python_values(c.Compile())
- np.testing.assert_equal(ans, np.int8(3))
-
-
-class ParametersTest(ComputationTest):
- """Tests focusing on Parameter ops and argument-passing."""
-
- def setUp(self):
- self.f32_scalar_2 = NumpyArrayF32(2.0)
- self.f32_4vector = NumpyArrayF32([-2.3, 3.3, -4.3, 5.3])
- self.f64_scalar_2 = NumpyArrayF64(2.0)
- self.f64_4vector = NumpyArrayF64([-2.3, 3.3, -4.3, 5.3])
- self.s32_scalar_3 = NumpyArrayS32(3)
- self.s32_4vector = NumpyArrayS32([10, 15, -2, 7])
- self.s64_scalar_3 = NumpyArrayS64(3)
- self.s64_4vector = NumpyArrayS64([10, 15, -2, 7])
-
- def testScalarTimesVectorAutonumberF32(self):
- c = self._NewComputation()
- p0 = c.ParameterFromNumpy(self.f32_scalar_2)
- p1 = c.ParameterFromNumpy(self.f32_4vector)
- c.Mul(p0, p1)
- self._ExecuteAndCompareClose(
- c,
- arguments=[self.f32_scalar_2, self.f32_4vector],
- expected=[[-4.6, 6.6, -8.6, 10.6]])
-
- def testScalarTimesVectorAutonumberF64(self):
- c = self._NewComputation()
- p0 = c.ParameterFromNumpy(self.f64_scalar_2)
- p1 = c.ParameterFromNumpy(self.f64_4vector)
- c.Mul(p0, p1)
- self._ExecuteAndCompareClose(
- c,
- arguments=[self.f64_scalar_2, self.f64_4vector],
- expected=[[-4.6, 6.6, -8.6, 10.6]])
-
- def testScalarTimesVectorS32(self):
- c = self._NewComputation()
- p0 = c.ParameterFromNumpy(self.s32_scalar_3)
- p1 = c.ParameterFromNumpy(self.s32_4vector)
- c.Mul(p0, p1)
- self._ExecuteAndCompareExact(
- c,
- arguments=[self.s32_scalar_3, self.s32_4vector],
- expected=[[30, 45, -6, 21]])
-
- def testScalarTimesVectorS64(self):
- c = self._NewComputation()
- p0 = c.ParameterFromNumpy(self.s64_scalar_3)
- p1 = c.ParameterFromNumpy(self.s64_4vector)
- c.Mul(p0, p1)
- self._ExecuteAndCompareExact(
- c,
- arguments=[self.s64_scalar_3, self.s64_4vector],
- expected=[[30, 45, -6, 21]])
-
- def testScalarMinusVectorExplicitNumberingF32(self):
- # Use explicit numbering and pass parameter_num first. Sub is used since
- # it's not commutative and can help catch parameter reversal within the
- # computation.
- c = self._NewComputation()
- p1 = c.ParameterFromNumpy(self.f32_4vector, parameter_num=1)
- p0 = c.ParameterFromNumpy(self.f32_scalar_2, parameter_num=0)
- c.Sub(p1, p0)
- self._ExecuteAndCompareClose(
- c,
- arguments=[self.f32_scalar_2, self.f32_4vector],
- expected=[[-4.3, 1.3, -6.3, 3.3]])
-
- def testScalarMinusVectorExplicitNumberingF64(self):
- # Use explicit numbering and pass parameter_num first. Sub is used since
- # it's not commutative and can help catch parameter reversal within the
- # computation.
- c = self._NewComputation()
- p1 = c.ParameterFromNumpy(self.f64_4vector, parameter_num=1)
- p0 = c.ParameterFromNumpy(self.f64_scalar_2, parameter_num=0)
- c.Sub(p1, p0)
- self._ExecuteAndCompareClose(
- c,
- arguments=[self.f64_scalar_2, self.f64_4vector],
- expected=[[-4.3, 1.3, -6.3, 3.3]])
-
-
-class BufferTest(ComputationTest):
- """Tests focusing on execution with Buffers."""
-
- def testConstantSum(self):
- c = self._NewComputation()
- c.Add(c.ConstantF32Scalar(1.11), c.ConstantF32Scalar(3.14))
- self._ExecuteAndCompareClose(c, expected=[4.25])
-
- def testOneParameterSum(self):
- c = self._NewComputation()
- c.Add(c.ParameterFromNumpy(NumpyArrayF32(0.)), c.ConstantF32Scalar(3.14))
- self._ExecuteAndCompareClose(
- c, arguments=[NumpyArrayF32(1.11)], expected=[4.25])
-
- def testTwoParameterSum(self):
- c = self._NewComputation()
- c.Add(
- c.ParameterFromNumpy(NumpyArrayF32(0.)),
- c.ParameterFromNumpy(NumpyArrayF32(0.)))
- self._ExecuteAndCompareClose(
- c,
- arguments=[NumpyArrayF32(1.11),
- NumpyArrayF32(3.14)],
- expected=[4.25])
-
- def testCannotCallWithDeletedBuffers(self):
- c = self._NewComputation()
- c.Add(c.ParameterFromNumpy(NumpyArrayF32(0.)), c.ConstantF32Scalar(3.14))
- arg = NumpyArrayF32(1.11)
- compiled_c = c.Build().Compile()
- arg_buffer = xla_client.Buffer.from_pyval(arg)
- arg_buffer.delete()
- with self.assertRaises(RuntimeError):
- compiled_c.Execute([arg_buffer])
-
- def testShape(self):
- pyval = np.array([[1., 2.]], np.float32)
- local_buffer = xla_client.Buffer.from_pyval(pyval)
- xla_shape = local_buffer.shape()
- self.assertEqual(xla_shape.dimensions(), (1, 2))
- self.assertEqual(np.dtype(xla_shape.element_type()), np.dtype(np.float32))
-
- def testBlockHostUntilReadyWorks(self):
- arg = np.array([[1., 2.]], np.float32)
- arg_buffer = xla_client.Buffer.from_pyval(arg)
- arg_buffer.block_host_until_ready()
- # This test merely checks that nothing goes awry when we call
- # block_host_until_ready(); it's difficult to test anything else.
-
- def testCopyToHost(self):
- arg0 = np.array([[1., 2.]], np.float32)
- arg1 = np.array([[3., 4.]], np.float32)
- arg0_buffer = xla_client.Buffer.from_pyval(arg0)
- arg1_buffer = xla_client.Buffer.from_pyval(arg1)
- # Prefetch two buffers using copy_to_host_async, and then retrieve their
- # values using to_py.
- arg0_buffer.copy_to_host_async()
- arg0_buffer.copy_to_host_async() # Duplicate calls don't do anything.
- arg1_buffer.copy_to_host_async()
- np.testing.assert_equal(arg0, arg0_buffer.to_py())
- np.testing.assert_equal(arg1, arg1_buffer.to_py())
- # copy_to_host_async does nothing after to_py is called.
- arg0_buffer.copy_to_host_async()
- np.testing.assert_equal(arg0, arg0_buffer.to_py())
-
- def testDevice(self):
- x = np.arange(8)
- for device in xla_client.get_local_backend().local_devices():
- buf = xla_client.Buffer.from_pyval(x, device=device)
- self.assertEqual(buf.device(), device)
- np.testing.assert_equal(x, buf.to_py())
-
-
-class SingleOpTest(ComputationTest):
- """Tests for single ops.
-
- The goal here is smoke testing - to exercise the most basic functionality of
- single XLA ops. As minimal as possible number of additional ops are added
- around the op being tested.
- """
-
- def testConcatenateF32(self):
- c = self._NewComputation()
- args = (
- c.Constant(NumpyArrayF32([1.0, 2.0, 3.0])),
- c.Constant(NumpyArrayF32([4.0, 5.0, 6.0])),
- )
- c.Concatenate(args, dimension=0)
- self._ExecuteAndCompareClose(c, expected=[[1.0, 2.0, 3.0, 4.0, 5.0, 6.0]])
-
- def testConcatenateF64(self):
- c = self._NewComputation()
- args = (
- c.Constant(NumpyArrayF64([1.0, 2.0, 3.0])),
- c.Constant(NumpyArrayF64([4.0, 5.0, 6.0])),
- )
- c.Concatenate(args, dimension=0)
- self._ExecuteAndCompareClose(c, expected=[[1.0, 2.0, 3.0, 4.0, 5.0, 6.0]])
-
- def testConvertElementType(self):
- xla_types = {
- np.bool: xla_client.PrimitiveType.PRED,
- np.int32: xla_client.PrimitiveType.S32,
- np.int64: xla_client.PrimitiveType.S64,
- np.float32: xla_client.PrimitiveType.F32,
- np.float64: xla_client.PrimitiveType.F64,
- }
-
- def _ConvertAndTest(template, src_dtype, dst_dtype):
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in int_dtypes + float_dtypes)
+ def testConstantScalarSum(self, dtype):
+ if dtype == np.int8 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support int8")
c = self._NewComputation()
- x = c.Constant(np.array(template, dtype=src_dtype))
- c.ConvertElementType(x, xla_types[dst_dtype])
+ ops.Add(ops.Constant(c, dtype(1.11)), ops.Constant(c, dtype(3.14)))
+ self._ExecuteAndCompareClose(c, expected=[dtype(1.11) + dtype(3.14)])
- result = xla_client.execute_with_python_values(c.Build().Compile())
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testConstantVectorMul(self, dtype):
+ c = self._NewComputation()
+ ops.Mul(
+ ops.Constant(c, np.array([2.5, 3.3, -1.2, 0.7], dtype)),
+ ops.Constant(c, np.array([-1.2, 2, -2, -3], dtype)))
+ self._ExecuteAndCompareClose(
+ c, expected=[[-3, 6.6, 2.4, -2.1]], rtol=3e-3)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testConstantVectorScalarDiv(self, dtype):
+ c = self._NewComputation()
+ ops.Div(
+ ops.Constant(c, np.array([1.5, 2.5, 3.0, -10.8], dtype=dtype)),
+ ops.Constant(c, dtype(2.0)))
+ self._ExecuteAndCompareClose(
+ c, expected=[[0.75, 1.25, 1.5, -5.4]], rtol=2e-3)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testConstantVectorScalarPow(self, dtype):
+ c = self._NewComputation()
+ ops.Pow(
+ ops.Constant(c, np.array([1.5, 2.5, 3.0], dtype=dtype)),
+ ops.Constant(c, dtype(2.)))
+ self._ExecuteAndCompareClose(c, expected=[[2.25, 6.25, 9.]])
+
+ def testIota(self):
+ c = self._NewComputation()
+ ops.Iota(c, xla_client.PrimitiveType.F32, 10)
+ self._ExecuteAndCompareExact(
+ c, expected=[np.arange(10, dtype=np.float32)])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in int_dtypes)
+ def testBroadcastedIota(self, dtype):
+ c = self._NewComputation()
+ shape = xla_client.Shape.array_shape(
+ xla_client.dtype_to_etype(dtype), (2, 3))
+ ops.Iota(c, shape, 1)
+ expected = np.array([[0, 1, 2], [0, 1, 2]], dtype=dtype)
+ self._ExecuteAndCompareExact(c, expected=[expected])
+
+ def testBooleanAnd(self):
+ c = self._NewComputation()
+ ops.And(
+ ops.Constant(c, NumpyArrayBool([True, False, True, False])),
+ ops.Constant(c, NumpyArrayBool([True, True, False, False])))
+ self._ExecuteAndCompareExact(c, expected=[[True, False, False, False]])
+
+ def testBooleanOr(self):
+ c = self._NewComputation()
+ ops.Or(
+ ops.Constant(c, NumpyArrayBool([True, False, True, False])),
+ ops.Constant(c, NumpyArrayBool([True, True, False, False])))
+ self._ExecuteAndCompareExact(c, expected=[[True, True, True, False]])
+
+ def testBooleanXor(self):
+ c = self._NewComputation()
+ ops.Xor(
+ ops.Constant(c, NumpyArrayBool([True, False, True, False])),
+ ops.Constant(c, NumpyArrayBool([True, True, False, False])))
+ self._ExecuteAndCompareExact(c, expected=[[False, True, True, False]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testSum2D(self, dtype):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Constant(c, np.array([[1, 2, 3], [4, 5, 6]], dtype=dtype)),
+ ops.Constant(c, np.array([[1, -1, 1], [-1, 1, -1]], dtype=dtype)))
+ self._ExecuteAndCompareClose(c, expected=[[[2, 1, 4], [3, 6, 5]]])
+
+ def testShiftLeft(self):
+ c = self._NewComputation()
+ ops.ShiftLeft(
+ ops.Constant(c, NumpyArrayS32([3])),
+ ops.Constant(c, NumpyArrayS32([2])))
+ self._ExecuteAndCompareClose(c, expected=[[12]])
+
+ def testShiftRightArithmetic(self):
+ c = self._NewComputation()
+ ops.ShiftRightArithmetic(
+ ops.Constant(c, NumpyArrayS32([-2])),
+ ops.Constant(c, NumpyArrayS32([1])))
+ self._ExecuteAndCompareClose(c, expected=[[-1]])
+
+ def testShiftRightLogical(self):
+ c = self._NewComputation()
+ ops.ShiftRightLogical(
+ ops.Constant(c, NumpyArrayS32([-1])),
+ ops.Constant(c, NumpyArrayS32([1])))
+ self._ExecuteAndCompareClose(c, expected=[[2**31 - 1]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testSum2DWith1DBroadcastDim0(self, dtype):
+ # sum of a 2D array with a 1D array where the latter is replicated across
+ # dimension 0 to match the former's shape.
+ c = self._NewComputation()
+ ops.Add(
+ ops.Constant(c,
+ np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]],
+ dtype=dtype)),
+ ops.Constant(c, np.array([10, 20, 30], dtype=dtype)),
+ broadcast_dimensions=(0,))
+ self._ExecuteAndCompareClose(
+ c, expected=[[[11, 12, 13], [24, 25, 26], [37, 38, 39]]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testSum2DWith1DBroadcastDim1(self, dtype):
+ # sum of a 2D array with a 1D array where the latter is replicated across
+ # dimension 1 to match the former's shape.
+ c = self._NewComputation()
+ ops.Add(
+ ops.Constant(c,
+ np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]],
+ dtype=dtype)),
+ ops.Constant(c, np.array([10, 20, 30], dtype=dtype)),
+ broadcast_dimensions=(1,))
+ self._ExecuteAndCompareClose(
+ c, expected=[[[11, 22, 33], [14, 25, 36], [17, 28, 39]]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testConstantAxpy(self, dtype):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Mul(
+ ops.Constant(c, dtype(2)),
+ ops.Constant(c, np.array([2.2, 3.3, 4.4, 5.5], dtype=dtype))),
+ ops.Constant(c, np.array([100, -100, 200, -200], dtype)))
+ self._ExecuteAndCompareClose(
+ c, expected=[[104.4, -93.4, 208.8, -189]], rtol=2e-3)
+
+ def testCustomCall(self):
+ if self.backend.platform != "cpu":
+ self.skipTest("Test requires cpu platform")
+ c = self._NewComputation()
+ for name, fn in custom_call_for_test.cpu_custom_call_targets.items():
+ xla_client.register_custom_call_target(name, fn, platform="cpu")
+ ops.CustomCallWithLayout(
+ c,
+ b"test_subtract_f32",
+ operands=[
+ ops.Constant(c, np.float32(1.25)),
+ ops.Constant(c, np.float32(0.5))
+ ],
+ shape_with_layout=xla_client.Shape.array_shape(
+ np.dtype(np.float32), (), ()),
+ operand_shapes_with_layout=[
+ xla_client.Shape.array_shape(np.dtype(np.float32), (), ()),
+ xla_client.Shape.array_shape(np.dtype(np.float32), (), ()),
+ ])
+ self._ExecuteAndCompareClose(c, expected=[0.75])
+
+ tests.append(ComputationsWithConstantsTest)
+
+ class ComputationFromProtoTest(absltest.TestCase):
+ """Test computation execution from HLO proto."""
+
+ def setUp(self):
+ super(ComputationFromProtoTest, self).setUp()
+ self.backend = xla_backend()
+
+ def testExecuteFromProto(self):
+ # Build the HLO proto
+ b = xla_client.XlaBuilder("computation")
+ ops.Add(ops.Constant(b, np.int8(1)), ops.Constant(b, np.int8(2)))
+ serialized_proto = b.Build().GetSerializedProto()
+
+ # Load and execute the proto
+ c = xla_client.Computation(
+ xla_client._xla.XlaComputation(serialized_proto))
+ ans, = xla_client.execute_with_python_values(
+ c.Compile(), backend=self.backend)
+ np.testing.assert_equal(ans, np.int8(3))
+
+ tests.append(ComputationFromProtoTest)
+
+ class ParametersTest(ComputationTest):
+ """Tests focusing on Parameter ops and argument-passing."""
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in int_dtypes)
+ def testScalarTimesVector(self, dtype):
+ c = self._NewComputation()
+ arg0 = np.array(3, dtype=dtype)
+ arg1 = np.array([10, 15, -2, 7], dtype=dtype)
+ p0 = ops.Parameter(c, 0, xla_client.shape_from_pyval(arg0))
+ p1 = ops.Parameter(c, 1, xla_client.shape_from_pyval(arg1))
+ ops.Mul(p0, p1)
+ self._ExecuteAndCompareExact(
+ c, arguments=[arg0, arg1], expected=[arg0 * arg1])
+
+ # TODO(phawkins): test comparison harness doesn't support bfloat16
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes if dtype != bfloat16)
+ def testScalarMinusVectorExplicitNumbering(self, dtype):
+ # Use explicit numbering and pass parameter_num first. Sub is used since
+ # it's not commutative and can help catch parameter reversal within the
+ # computation.
+ c = self._NewComputation()
+ arg0 = np.array(2.0, dtype=dtype)
+ arg1 = np.array([-2.3, 3.3, -4.3, 5.3], dtype=dtype)
+ p1 = ops.Parameter(c, 1, xla_client.shape_from_pyval(arg1))
+ p0 = ops.Parameter(c, 0, xla_client.shape_from_pyval(arg0))
+ ops.Sub(p1, p0)
+ self._ExecuteAndCompareClose(
+ c, arguments=[arg0, arg1], expected=[arg1 - arg0])
+
+ tests.append(ParametersTest)
+
+ class BufferTest(ComputationTest):
+ """Tests focusing on execution with Buffers."""
+
+ def testConstantSum(self):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Constant(c, np.float32(1.11)), ops.Constant(c, np.float32(3.14)))
+ self._ExecuteAndCompareClose(c, expected=[4.25])
+
+ def testOneParameterSum(self):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(0.))),
+ ops.Constant(c, np.float32(3.14)))
+ self._ExecuteAndCompareClose(
+ c, arguments=[NumpyArrayF32(1.11)], expected=[4.25])
+
+ def testTwoParameterSum(self):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(0.))),
+ ops.Parameter(c, 1, xla_client.shape_from_pyval(NumpyArrayF32(0.))))
+ self._ExecuteAndCompareClose(
+ c,
+ arguments=[NumpyArrayF32(1.11),
+ NumpyArrayF32(3.14)],
+ expected=[4.25])
+
+ @unittest.skipIf(cloud_tpu, "not implemented")
+ def testCannotCallWithDeletedBuffers(self):
+ c = self._NewComputation()
+ ops.Add(
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(0.))),
+ ops.Constant(c, np.float32(3.14)))
+ arg = NumpyArrayF32(1.11)
+ compiled_c = self.backend.compile(c.Build())
+ arg_buffer = xla_client.Buffer.from_pyval(arg, backend=self.backend)
+ arg_buffer.delete()
+ with self.assertRaises(RuntimeError):
+ compiled_c.Execute([arg_buffer])
+
+ def testShape(self):
+ pyval = np.array([[1., 2.]], np.float32)
+ local_buffer = xla_client.Buffer.from_pyval(pyval)
+ xla_shape = local_buffer.shape()
+ self.assertEqual(xla_shape.dimensions(), (1, 2))
+ self.assertEqual(np.dtype(xla_shape.element_type()), np.dtype(np.float32))
+
+ def testBlockHostUntilReadyWorks(self):
+ arg = np.array([[1., 2.]], np.float32)
+ arg_buffer = xla_client.Buffer.from_pyval(arg)
+ arg_buffer.block_host_until_ready()
+ # This test merely checks that nothing goes awry when we call
+ # block_host_until_ready(); it's difficult to test anything else.
+
+ def testCopyToHost(self):
+ arg0 = np.array([[1., 2.]], np.float32)
+ arg1 = np.array([[3., 4.]], np.float32)
+ arg0_buffer = xla_client.Buffer.from_pyval(arg0)
+ arg1_buffer = xla_client.Buffer.from_pyval(arg1)
+ # Prefetch two buffers using copy_to_host_async, and then retrieve their
+ # values using to_py.
+ arg0_buffer.copy_to_host_async()
+ arg0_buffer.copy_to_host_async() # Duplicate calls don't do anything.
+ arg1_buffer.copy_to_host_async()
+ np.testing.assert_equal(arg0, arg0_buffer.to_py())
+ np.testing.assert_equal(arg1, arg1_buffer.to_py())
+ # copy_to_host_async does nothing after to_py is called.
+ arg0_buffer.copy_to_host_async()
+ np.testing.assert_equal(arg0, arg0_buffer.to_py())
+
+ def testDevice(self):
+ x = np.arange(8, dtype=np.int32)
+ for device in self.backend.local_devices():
+ buf = xla_client.Buffer.from_pyval(
+ x, device=device, backend=self.backend)
+ self.assertEqual(buf.device(), device)
+ np.testing.assert_equal(x, buf.to_py())
+
+ tests.append(BufferTest)
+
+ class SingleOpTest(ComputationTest):
+ """Tests for single ops.
+
+ The goal here is smoke testing - to exercise the most basic functionality of
+ single XLA ops. As minimal as possible number of additional ops are added
+ around the op being tested.
+ """
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testConcatenate(self, dtype):
+ c = self._NewComputation()
+ args = (
+ ops.Constant(c, np.array([1.0, 2.0, 3.0], dtype=dtype)),
+ ops.Constant(c, np.array([4.0, 5.0, 6.0], dtype=dtype)),
+ )
+ ops.ConcatInDim(c, args, dimension=0)
+ self._ExecuteAndCompareExact(
+ c, expected=[np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], dtype=dtype)])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}_{}".format(src_dtype.__name__,
+ dst_dtype.__name__),
+ "src_dtype": src_dtype,
+ "dst_dtype": dst_dtype,
+ } for src_dtype, dst_dtype in itertools.permutations(
+ [np.bool, np.int32, np.int64, np.float32, np.float64], 2))
+ def testConvertElementType(self, src_dtype, dst_dtype):
+ if ((src_dtype in [np.int64, np.float64] or
+ dst_dtype in [np.int64, np.float64]) and
+ self.backend.platform == "tpu"):
+ self.skipTest("TPU doesn't support float64")
+ c = self._NewComputation()
+ x = np.array([0, 1, 0, 0, 1], dtype=src_dtype)
+ ops.ConvertElementType(
+ ops.Constant(c, x), xla_client.dtype_to_etype(dst_dtype))
+
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
self.assertLen(result, 1)
- expected = np.array(template, dtype=dst_dtype)
+ expected = np.array(x, dtype=dst_dtype)
self.assertEqual(result[0].shape, expected.shape)
self.assertEqual(result[0].dtype, expected.dtype)
np.testing.assert_equal(result[0], expected)
- x = [0, 1, 0, 0, 1]
- for src_dtype, dst_dtype in itertools.product(xla_types, xla_types):
- _ConvertAndTest(x, src_dtype, dst_dtype)
-
- def testBitcastConvertType(self):
- xla_x32_types = {
- np.int32: xla_client.PrimitiveType.S32,
- np.float32: xla_client.PrimitiveType.F32,
- }
-
- xla_x64_types = {
- np.int64: xla_client.PrimitiveType.S64,
- np.float64: xla_client.PrimitiveType.F64,
- }
-
- def _ConvertAndTest(template, src_dtype, dst_dtype, dst_etype):
+ @parameterized.named_parameters(
+ {
+ "testcase_name": "_{}_{}".format(src_dtype.__name__,
+ dst_dtype.__name__),
+ "src_dtype": src_dtype,
+ "dst_dtype": dst_dtype,
+ }
+ for dtypes in [[np.int32, np.float32], [np.int64, np.float64]]
+ for src_dtype, dst_dtype in itertools.permutations(dtypes, 2))
+ def testBitcastConvertType(self, src_dtype, dst_dtype):
+ if (np.float64 in (src_dtype, dst_dtype) and
+ self.backend.platform == "tpu"):
+ self.skipTest("TPU doesn't support float64")
c = self._NewComputation()
- x = c.Constant(np.array(template, dtype=src_dtype))
- c.BitcastConvertType(x, dst_etype)
+ x = np.array([0, 1, 0, 0, 1], dtype=src_dtype)
+ ops.BitcastConvertType(
+ ops.Constant(c, x), xla_client.dtype_to_etype(dst_dtype))
- result = xla_client.execute_with_python_values(c.Build().Compile())
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
self.assertLen(result, 1)
- expected = np.array(template, src_dtype).view(dst_dtype)
+ expected = x.view(dst_dtype)
self.assertEqual(result[0].shape, expected.shape)
self.assertEqual(result[0].dtype, expected.dtype)
np.testing.assert_equal(result[0], expected)
- x = [0, 1, 0, 0, 1]
- for xla_types in [xla_x32_types, xla_x64_types]:
- for src_dtype, dst_dtype in itertools.product(xla_types, xla_types):
- _ConvertAndTest(x, src_dtype, dst_dtype, xla_types[dst_dtype])
+ # TODO(b/123523486) implement AllToAll on CPU
+ def DISABLED_testAllToAllOneReplica(self):
+ samples = [
+ NumpyArrayF32([97.0]),
+ NumpyArrayF32([64.0, 117.0]),
+ NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
+ ]
+ for lhs in samples[:1]:
+ c = self._NewComputation()
+ ops.AllToAll(ops.Constant(c, lhs), 0, 0)
+ self._ExecuteAndCompareExact(c, expected=[lhs])
- # TODO(b/123523486) implement AllToAll on CPU
- def DISABLED_testAllToAllOneReplica(self):
- samples = [
- NumpyArrayF32([97.0]),
- NumpyArrayF32([64.0, 117.0]),
- NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
- ]
- for lhs in samples[:1]:
+ def testCrossReplicaSumOneReplica(self):
+ samples = [
+ NumpyArrayF32(42.0),
+ NumpyArrayF32([97.0]),
+ NumpyArrayF32([64.0, 117.0]),
+ NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
+ ]
+ for lhs in samples:
+ c = self._NewComputation()
+ ops.CrossReplicaSum(ops.Constant(c, lhs))
+ self._ExecuteAndCompareExact(c, expected=[lhs])
+
+ def testReplicaId(self):
c = self._NewComputation()
- c.AllToAll(c.Constant(lhs), 0, 0)
- self._ExecuteAndCompareExact(c, expected=[lhs])
+ _ = ops.ReplicaId(c)
+ self._ExecuteAndCompareExact(c, expected=[0])
- def testCrossReplicaSumOneReplica(self):
- samples = [
- NumpyArrayF32(42.0),
- NumpyArrayF32([97.0]),
- NumpyArrayF32([64.0, 117.0]),
- NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
- ]
- for lhs in samples:
+ def testCrossReplicaSumOneReplicaWithSingletonGroup(self):
+ samples = [
+ NumpyArrayF32(42.0),
+ NumpyArrayF32([97.0]),
+ NumpyArrayF32([64.0, 117.0]),
+ NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
+ ]
+ for lhs in samples:
+ c = self._NewComputation()
+ ops.CrossReplicaSum(
+ ops.Constant(c, lhs), xla_client.make_replica_groups([[0]]))
+ self._ExecuteAndCompareExact(c, expected=[lhs])
+
+ # TODO(phawkins): np.dot implementation doesn't support bfloat16
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes if dtype != bfloat16)
+ def testDotMatrixVector(self, dtype):
c = self._NewComputation()
- c.CrossReplicaSum(c.Constant(lhs))
- self._ExecuteAndCompareExact(c, expected=[lhs])
+ lhs = np.array([[2.0, 3.0], [4.0, 5.0]], dtype=dtype)
+ rhs = np.array([[10.0], [20.0]], dtype=dtype)
+ ops.Dot(ops.Constant(c, lhs), ops.Constant(c, rhs))
+ self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
- def testReplicaId(self):
- c = self._NewComputation()
- _ = c.ReplicaId()
- self._ExecuteAndCompareExact(c, expected=[0])
-
- def testCrossReplicaSumOneReplicaWithSingletonGroup(self):
- samples = [
- NumpyArrayF32(42.0),
- NumpyArrayF32([97.0]),
- NumpyArrayF32([64.0, 117.0]),
- NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]]),
- ]
- for lhs in samples:
+ # TODO(phawkins): np.dot implementation doesn't support bfloat16
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes if dtype != bfloat16)
+ def testDotMatrixMatrix(self, dtype):
c = self._NewComputation()
- c.CrossReplicaSum(c.Constant(lhs), [[0]])
- self._ExecuteAndCompareExact(c, expected=[lhs])
+ lhs = np.array([[2.0, 3.0], [4.0, 5.0]], dtype=dtype)
+ rhs = np.array([[10.0, 20.0], [100.0, 200.0]], dtype=dtype)
+ ops.Dot(ops.Constant(c, lhs), ops.Constant(c, rhs))
+ self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
- def testDotMatrixVectorF32(self):
- c = self._NewComputation()
- lhs = NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]])
- rhs = NumpyArrayF32([[10.0], [20.0]])
- c.Dot(c.Constant(lhs), c.Constant(rhs))
- self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
-
- def testDotMatrixVectorF64(self):
- c = self._NewComputation()
- lhs = NumpyArrayF64([[2.0, 3.0], [4.0, 5.0]])
- rhs = NumpyArrayF64([[10.0], [20.0]])
- c.Dot(c.Constant(lhs), c.Constant(rhs))
- self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
-
- def testDotMatrixMatrixF32(self):
- c = self._NewComputation()
- lhs = NumpyArrayF32([[2.0, 3.0], [4.0, 5.0]])
- rhs = NumpyArrayF32([[10.0, 20.0], [100.0, 200.0]])
- c.Dot(c.Constant(lhs), c.Constant(rhs))
- self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
-
- def testDotMatrixMatrixF64(self):
- c = self._NewComputation()
- lhs = NumpyArrayF64([[2.0, 3.0], [4.0, 5.0]])
- rhs = NumpyArrayF64([[10.0, 20.0], [100.0, 200.0]])
- c.Dot(c.Constant(lhs), c.Constant(rhs))
- self._ExecuteAndCompareClose(c, expected=[np.dot(lhs, rhs)])
-
- def testDotGeneral(self):
- c = self._NewComputation()
- rng = np.random.RandomState(0)
- lhs = NumpyArrayF32(rng.randn(10, 3, 4))
- rhs = NumpyArrayF32(rng.randn(10, 4, 5))
- dimension_numbers = (([2], [1]), ([0], [0]))
- c.DotGeneral(c.Constant(lhs), c.Constant(rhs), dimension_numbers)
- self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
-
- def testDotGeneralWithDotDimensionNumbersProto(self):
- c = self._NewComputation()
- rng = np.random.RandomState(0)
- lhs = NumpyArrayF32(rng.randn(10, 3, 4))
- rhs = NumpyArrayF32(rng.randn(10, 4, 5))
-
- dimension_numbers = xla_client.DotDimensionNumbers()
- dimension_numbers.lhs_contracting_dimensions.append(2)
- dimension_numbers.rhs_contracting_dimensions.append(1)
- dimension_numbers.lhs_batch_dimensions.append(0)
- dimension_numbers.rhs_batch_dimensions.append(0)
-
- c.DotGeneral(c.Constant(lhs), c.Constant(rhs), dimension_numbers)
- self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
-
- def testDotGeneralWithPrecisionConfig(self):
- c = self._NewComputation()
- rng = np.random.RandomState(0)
- lhs = NumpyArrayF32(rng.randn(10, 3, 4))
- rhs = NumpyArrayF32(rng.randn(10, 4, 5))
- dimension_numbers = (([2], [1]), ([0], [0]))
- config = xla_client.PrecisionConfig()
- config.operand_precision.append(config.Precision.HIGH)
- config.operand_precision.append(config.Precision.HIGHEST)
- c.DotGeneral(
- c.Constant(lhs),
- c.Constant(rhs),
- dimension_numbers,
- precision_config=config)
- self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
-
- def testConvF32Same(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 2, 3, 4)
- rhs = a(1, 2, 1, 2) * 10
- c.Conv(
- c.Constant(lhs), c.Constant(rhs), [1, 1], xla_client.PaddingType.SAME)
- result = np.array([[[
- [640., 700., 760., 300.],
- [880., 940., 1000., 380.],
- [1120., 1180., 1240., 460.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testConvF32Valid(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 2, 3, 4)
- rhs = a(1, 2, 1, 2) * 10
- c.Conv(
- c.Constant(lhs), c.Constant(rhs), [2, 1], xla_client.PaddingType.VALID)
- result = np.array([[[
- [640., 700., 760.],
- [1120., 1180., 1240.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testConvWithGeneralPaddingF32(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 1, 2, 3)
- rhs = a(1, 1, 1, 2) * 10
- strides = [1, 1]
- pads = [(1, 0), (0, 1)]
- lhs_dilation = (2, 1)
- rhs_dilation = (1, 1)
- c.ConvWithGeneralPadding(
- c.Constant(lhs), c.Constant(rhs), strides, pads, lhs_dilation,
- rhs_dilation)
- result = np.array([[[
- [0., 0., 0.],
- [10., 20., 0.],
- [0., 0., 0.],
- [40., 50., 0.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testConvGeneralDilatedF32(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 1, 2, 3)
- rhs = a(1, 1, 1, 2) * 10
- strides = [1, 1]
- pads = [(1, 0), (0, 1)]
- lhs_dilation = (2, 1)
- rhs_dilation = (1, 1)
- dimension_numbers = ("NCHW", "OIHW", "NCHW")
- c.ConvGeneralDilated(
- c.Constant(lhs), c.Constant(rhs), strides, pads, lhs_dilation,
- rhs_dilation, dimension_numbers)
- result = np.array([[[
- [0., 0., 0.],
- [10., 20., 0.],
- [0., 0., 0.],
- [40., 50., 0.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testConvGeneralDilatedF32WithPrecisionConfig(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 1, 2, 3)
- rhs = a(1, 1, 1, 2) * 10
- strides = [1, 1]
- pads = [(1, 0), (0, 1)]
- lhs_dilation = (2, 1)
- rhs_dilation = (1, 1)
- dimension_numbers = ("NCHW", "OIHW", "NCHW")
- config = xla_client.PrecisionConfig()
- config.operand_precision.append(config.Precision.HIGHEST)
- config.operand_precision.append(config.Precision.DEFAULT)
- c.ConvGeneralDilated(
- c.Constant(lhs),
- c.Constant(rhs),
- strides,
- pads,
- lhs_dilation,
- rhs_dilation,
- dimension_numbers,
- precision_config=config)
- result = np.array([[[
- [0., 0., 0.],
- [10., 20., 0.],
- [0., 0., 0.],
- [40., 50., 0.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testConvGeneralDilatedPermutedF32(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 1, 2, 3)
- rhs = a(1, 1, 1, 2) * 10
- strides = [1, 1]
- pads = [(1, 0), (0, 1)]
- lhs_dilation = (2, 1)
- rhs_dilation = (1, 1)
-
- dimension_numbers = ("NHWC", "OIHW", "CWNH")
- c.ConvGeneralDilated(
- c.Constant(np.transpose(lhs, (0, 2, 3, 1))), c.Constant(rhs), strides,
- pads, lhs_dilation, rhs_dilation, dimension_numbers)
- result = np.array([[[[0., 0., 0.], [10., 20., 0.], [0., 0., 0.],
- [40., 50., 0.]]]])
- self._ExecuteAndCompareClose(
- c, expected=[np.transpose(result, (1, 3, 0, 2))])
-
- def testConvGeneralDilatedGroupedConvolutionF32(self):
- c = self._NewComputation()
- a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
- lhs = a(1, 2, 2, 3)
- rhs = a(2, 1, 1, 2) * 10
- strides = [1, 1]
- pads = [(1, 0), (0, 1)]
- lhs_dilation = (2, 1)
- rhs_dilation = (1, 1)
- dimension_numbers = ("NCHW", "OIHW", "NCHW")
- feature_group_count = 2
- c.ConvGeneralDilated(
- c.Constant(lhs), c.Constant(rhs), strides, pads, lhs_dilation,
- rhs_dilation, dimension_numbers, feature_group_count)
- result = np.array([[[
- [0., 0., 0.],
- [10., 20., 0.],
- [0., 0., 0.],
- [40., 50., 0.],
- ], [
- [0., 0., 0.],
- [330., 380., 160.],
- [0., 0., 0.],
- [480., 530., 220.],
- ]]])
- self._ExecuteAndCompareClose(c, expected=[result])
-
- def testBooleanNot(self):
- c = self._NewComputation()
- arr = NumpyArrayBool([True, False, True])
- c.Not(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[~arr])
-
- def testPopulationCount(self):
- c = self._NewComputation()
- arr = NumpyArrayS32([3, 0, 1])
- c.PopulationCount(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.array([2, 0, 1])])
-
- def testCountLeadingZeros(self):
- c = self._NewComputation()
- arr = NumpyArrayS32([0x7FFF, 0x12345678])
- c.Clz(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[[17, 3]])
-
- def testExp(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Exp(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.exp(arr)])
-
- def testExpm1(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Expm1(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.expm1(arr)])
-
- def testRound(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Round(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.round(arr)])
-
- def testLog(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Log(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.log(arr)])
-
- def testLog1p(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Log1p(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.log1p(arr)])
-
- def testNeg(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Neg(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[-arr])
-
- def testFloor(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Floor(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.floor(arr)])
-
- def testCeil(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Ceil(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.ceil(arr)])
-
- def testAbs(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, -12.1, 2.4, -1.])
- c.Abs(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.abs(arr)])
-
- def testTanh(self):
- c = self._NewComputation()
- arr = NumpyArrayF32([3.3, 12.1])
- c.Tanh(c.Constant(arr))
- self._ExecuteAndCompareClose(c, expected=[np.tanh(arr)])
-
- def testTrans(self):
-
- def _TransposeAndTest(array):
+ def testDotGeneral(self):
c = self._NewComputation()
- c.Trans(c.Constant(array))
- self._ExecuteAndCompareClose(c, expected=[array.T])
+ rng = np.random.RandomState(0)
+ lhs = NumpyArrayF32(rng.randn(10, 3, 4))
+ rhs = NumpyArrayF32(rng.randn(10, 4, 5))
+ dimension_numbers = xla_client.make_dot_dimension_numbers(
+ (([2], [1]), ([0], [0])))
+ ops.DotGeneral(
+ ops.Constant(c, lhs), ops.Constant(c, rhs), dimension_numbers)
+ self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
- # Test square and non-square matrices in both default (C) and F orders.
- for array_fun in [NumpyArrayF32, NumpyArrayF64]:
- _TransposeAndTest(array_fun([[1, 2, 3], [4, 5, 6]]))
- _TransposeAndTest(array_fun([[1, 2, 3], [4, 5, 6]], order="F"))
- _TransposeAndTest(array_fun([[1, 2], [4, 5]]))
- _TransposeAndTest(array_fun([[1, 2], [4, 5]], order="F"))
-
- def testTranspose(self):
-
- def _TransposeAndTest(array, permutation):
+ def testDotGeneralWithDotDimensionNumbersProto(self):
c = self._NewComputation()
- c.Transpose(c.Constant(array), permutation)
- expected = np.transpose(array, permutation)
+ rng = np.random.RandomState(0)
+ lhs = NumpyArrayF32(rng.randn(10, 3, 4))
+ rhs = NumpyArrayF32(rng.randn(10, 4, 5))
+
+ dimension_numbers = xla_client.DotDimensionNumbers()
+ dimension_numbers.lhs_contracting_dimensions.append(2)
+ dimension_numbers.rhs_contracting_dimensions.append(1)
+ dimension_numbers.lhs_batch_dimensions.append(0)
+ dimension_numbers.rhs_batch_dimensions.append(0)
+
+ ops.DotGeneral(
+ ops.Constant(c, lhs), ops.Constant(c, rhs), dimension_numbers)
+ self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
+
+ def testDotGeneralWithPrecisionConfig(self):
+ c = self._NewComputation()
+ rng = np.random.RandomState(0)
+ lhs = NumpyArrayF32(rng.randn(10, 3, 4))
+ rhs = NumpyArrayF32(rng.randn(10, 4, 5))
+ dimension_numbers = xla_client.make_dot_dimension_numbers(
+ (([2], [1]), ([0], [0])))
+ config = xla_client.PrecisionConfig()
+ config.operand_precision.append(config.Precision.HIGH)
+ config.operand_precision.append(config.Precision.HIGHEST)
+ ops.DotGeneral(
+ ops.Constant(c, lhs),
+ ops.Constant(c, rhs),
+ dimension_numbers,
+ precision_config=config)
+ self._ExecuteAndCompareClose(c, expected=[np.matmul(lhs, rhs)], rtol=1e-6)
+
+ def testConvGeneralDilatedF32(self):
+ c = self._NewComputation()
+ a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
+ lhs = a(1, 1, 2, 3)
+ rhs = a(1, 1, 1, 2) * 10
+ strides = [1, 1]
+ pads = [(1, 0), (0, 1)]
+ lhs_dilation = (2, 1)
+ rhs_dilation = (1, 1)
+ dimension_numbers = xla_client.make_convolution_dimension_numbers(
+ ("NCHW", "OIHW", "NCHW"), 2)
+ ops.ConvGeneralDilated(
+ ops.Constant(c, lhs), ops.Constant(c, rhs), strides, pads,
+ lhs_dilation, rhs_dilation, dimension_numbers)
+ result = np.array([[[
+ [0., 0., 0.],
+ [10., 20., 0.],
+ [0., 0., 0.],
+ [40., 50., 0.],
+ ]]])
+ self._ExecuteAndCompareClose(c, expected=[result])
+
+ def testConvGeneralDilatedF32WithPrecisionConfig(self):
+ c = self._NewComputation()
+ a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
+ lhs = a(1, 1, 2, 3)
+ rhs = a(1, 1, 1, 2) * 10
+ strides = [1, 1]
+ pads = [(1, 0), (0, 1)]
+ lhs_dilation = (2, 1)
+ rhs_dilation = (1, 1)
+ dimension_numbers = xla_client.make_convolution_dimension_numbers(
+ ("NCHW", "OIHW", "NCHW"), 2)
+ config = xla_client.PrecisionConfig()
+ config.operand_precision.append(config.Precision.HIGHEST)
+ config.operand_precision.append(config.Precision.DEFAULT)
+ ops.ConvGeneralDilated(
+ ops.Constant(c, lhs),
+ ops.Constant(c, rhs),
+ strides,
+ pads,
+ lhs_dilation,
+ rhs_dilation,
+ dimension_numbers,
+ precision_config=config)
+ result = np.array([[[
+ [0., 0., 0.],
+ [10., 20., 0.],
+ [0., 0., 0.],
+ [40., 50., 0.],
+ ]]])
+ self._ExecuteAndCompareClose(c, expected=[result])
+
+ def testConvGeneralDilatedPermutedF32(self):
+ c = self._NewComputation()
+ a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
+ lhs = a(1, 1, 2, 3)
+ rhs = a(1, 1, 1, 2) * 10
+ strides = [1, 1]
+ pads = [(1, 0), (0, 1)]
+ lhs_dilation = (2, 1)
+ rhs_dilation = (1, 1)
+
+ dimension_numbers = xla_client.make_convolution_dimension_numbers(
+ ("NHWC", "OIHW", "CWNH"), 2)
+ ops.ConvGeneralDilated(
+ ops.Constant(c, np.transpose(lhs,
+ (0, 2, 3, 1))), ops.Constant(c, rhs),
+ strides, pads, lhs_dilation, rhs_dilation, dimension_numbers)
+ result = np.array([[[[0., 0., 0.], [10., 20., 0.], [0., 0., 0.],
+ [40., 50., 0.]]]])
+ self._ExecuteAndCompareClose(
+ c, expected=[np.transpose(result, (1, 3, 0, 2))])
+
+ def testConvGeneralDilatedGroupedConvolutionF32(self):
+ c = self._NewComputation()
+ a = lambda *dims: np.arange(np.prod(dims)).reshape(dims).astype("float32")
+ lhs = a(1, 2, 2, 3)
+ rhs = a(2, 1, 1, 2) * 10
+ strides = [1, 1]
+ pads = [(1, 0), (0, 1)]
+ lhs_dilation = (2, 1)
+ rhs_dilation = (1, 1)
+ dimension_numbers = xla_client.make_convolution_dimension_numbers(
+ ("NCHW", "OIHW", "NCHW"), 2)
+ feature_group_count = 2
+ ops.ConvGeneralDilated(
+ ops.Constant(c, lhs), ops.Constant(c, rhs), strides, pads,
+ lhs_dilation, rhs_dilation, dimension_numbers, feature_group_count)
+ result = np.array([[[
+ [0., 0., 0.],
+ [10., 20., 0.],
+ [0., 0., 0.],
+ [40., 50., 0.],
+ ], [
+ [0., 0., 0.],
+ [330., 380., 160.],
+ [0., 0., 0.],
+ [480., 530., 220.],
+ ]]])
+ self._ExecuteAndCompareClose(c, expected=[result])
+
+ def testBooleanNot(self):
+ c = self._NewComputation()
+ arr = NumpyArrayBool([True, False, True])
+ ops.Not(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[~arr])
+
+ def testPopulationCount(self):
+ c = self._NewComputation()
+ arr = NumpyArrayS32([3, 0, 1])
+ ops.PopulationCount(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.array([2, 0, 1])])
+
+ def testCountLeadingZeros(self):
+ c = self._NewComputation()
+ arr = NumpyArrayS32([0x7FFF, 0x12345678])
+ ops.Clz(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[[17, 3]])
+
+ def testExp(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Exp(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.exp(arr)])
+
+ def testExpm1(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Expm1(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.expm1(arr)])
+
+ def testRound(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Round(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.round(arr)])
+
+ def testLog(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Log(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.log(arr)])
+
+ def testLog1p(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Log1p(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.log1p(arr)])
+
+ def testNeg(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Neg(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[-arr])
+
+ def testFloor(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Floor(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.floor(arr)])
+
+ def testCeil(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Ceil(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.ceil(arr)])
+
+ def testAbs(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, -12.1, 2.4, -1.])
+ ops.Abs(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.abs(arr)])
+
+ def testTanh(self):
+ c = self._NewComputation()
+ arr = NumpyArrayF32([3.3, 12.1])
+ ops.Tanh(ops.Constant(c, arr))
+ self._ExecuteAndCompareClose(c, expected=[np.tanh(arr)])
+
+ def testTranspose(self):
+
+ def _TransposeAndTest(array, permutation):
+ c = self._NewComputation()
+ ops.Transpose(ops.Constant(c, array), permutation)
+ expected = np.transpose(array, permutation)
+ self._ExecuteAndCompareClose(c, expected=[expected])
+
+ _TransposeAndTest(NumpyArrayF32([[1, 2, 3], [4, 5, 6]]), [0, 1])
+ _TransposeAndTest(NumpyArrayF32([[1, 2, 3], [4, 5, 6]]), [1, 0])
+ _TransposeAndTest(NumpyArrayF32([[1, 2], [4, 5]]), [0, 1])
+ _TransposeAndTest(NumpyArrayF32([[1, 2], [4, 5]]), [1, 0])
+
+ arr = np.random.RandomState(0).randn(2, 3, 4).astype(np.float32)
+ for permutation in itertools.permutations(range(arr.ndim)):
+ _TransposeAndTest(arr, permutation)
+ _TransposeAndTest(np.asfortranarray(arr), permutation)
+
+ def testEq(self):
+ c = self._NewComputation()
+ ops.Eq(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4])),
+ ops.Constant(c, NumpyArrayS32([4, 2, 3, 1])))
+ self._ExecuteAndCompareExact(c, expected=[[False, True, True, False]])
+
+ def testNe(self):
+ c = self._NewComputation()
+ ops.Ne(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4])),
+ ops.Constant(c, NumpyArrayS32([4, 2, 3, 1])))
+ self._ExecuteAndCompareExact(c, expected=[[True, False, False, True]])
+
+ ops.Ne(
+ ops.Constant(c, NumpyArrayF32([-2.0, 0.0,
+ float("nan"),
+ float("nan")])),
+ ops.Constant(c, NumpyArrayF32([2.0, -0.0, 1.0,
+ float("nan")])))
+ self._ExecuteAndAssertWith(
+ np.testing.assert_allclose,
+ c, (),
+ expected=[[True, False, True, True]])
+
+ def testGt(self):
+ c = self._NewComputation()
+ ops.Gt(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4, 9])),
+ ops.Constant(c, NumpyArrayS32([1, 0, 2, 7, 12])))
+ self._ExecuteAndCompareExact(
+ c, expected=[[False, True, True, False, False]])
+
+ def testGe(self):
+ c = self._NewComputation()
+ ops.Ge(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4, 9])),
+ ops.Constant(c, NumpyArrayS32([1, 0, 2, 7, 12])))
+ self._ExecuteAndCompareExact(
+ c, expected=[[True, True, True, False, False]])
+
+ def testLt(self):
+ c = self._NewComputation()
+ ops.Lt(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4, 9])),
+ ops.Constant(c, NumpyArrayS32([1, 0, 2, 7, 12])))
+ self._ExecuteAndCompareExact(
+ c, expected=[[False, False, False, True, True]])
+
+ def testLe(self):
+ c = self._NewComputation()
+ ops.Le(
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4, 9])),
+ ops.Constant(c, NumpyArrayS32([1, 0, 2, 7, 12])))
+ self._ExecuteAndCompareExact(
+ c, expected=[[True, False, False, True, True]])
+
+ def testMax(self):
+ c = self._NewComputation()
+ ops.Max(
+ ops.Constant(c, NumpyArrayF32([1.0, 2.0, 3.0, 4.0, 9.0])),
+ ops.Constant(c, NumpyArrayF32([1.0, 0.0, 2.0, 7.0, 12.0])))
+ self._ExecuteAndCompareExact(c, expected=[[1.0, 2.0, 3.0, 7.0, 12.0]])
+
+ def testMaxExplicitBroadcastDim0(self):
+ c = self._NewComputation()
+ ops.Max(
+ ops.Constant(c, NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ ops.Constant(c, NumpyArrayF32([3, 4, 5])),
+ broadcast_dimensions=(0,))
+ self._ExecuteAndCompareExact(
+ c, expected=[[[3, 3, 3], [4, 5, 6], [7, 8, 9]]])
+
+ def testMaxExplicitBroadcastDim1(self):
+ c = self._NewComputation()
+ ops.Max(
+ ops.Constant(c, NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ ops.Constant(c, NumpyArrayF32([3, 4, 5])),
+ broadcast_dimensions=(1,))
+ self._ExecuteAndCompareExact(
+ c, expected=[[[3, 4, 5], [4, 5, 6], [7, 8, 9]]])
+
+ def testMin(self):
+ c = self._NewComputation()
+ ops.Min(
+ ops.Constant(c, NumpyArrayF32([1.0, 2.0, 3.0, 4.0, 9.0])),
+ ops.Constant(c, NumpyArrayF32([1.0, 0.0, 2.0, 7.0, 12.0])))
+ self._ExecuteAndCompareExact(c, expected=[[1.0, 0.0, 2.0, 4.0, 9.0]])
+
+ def testPad(self):
+ c = self._NewComputation()
+ ops.Pad(
+ ops.Constant(c, NumpyArrayF32([[1.0, 2.0], [3.0, 4.0]])),
+ ops.Constant(c, NumpyArrayF32(0.0)),
+ xla_client.make_padding_config([(1, 2, 1), (0, 1, 0)]))
+ self._ExecuteAndCompareClose(
+ c,
+ expected=[[[0.0, 0.0, 0.0], [1.0, 2.0, 0.0], [0.0, 0.0, 0.0],
+ [3.0, 4.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
+
+ def testPadWithPaddingConfig(self):
+ c = self._NewComputation()
+ padding_config = xla_client.PaddingConfig()
+ for lo, hi, interior in [(1, 2, 1), (0, 1, 0)]:
+ dimension = xla_client.PaddingConfigDimension()
+ dimension.edge_padding_low = lo
+ dimension.edge_padding_high = hi
+ dimension.interior_padding = interior
+ padding_config.dimensions.append(dimension)
+ ops.Pad(
+ ops.Constant(c, NumpyArrayF32([[1.0, 2.0], [3.0, 4.0]])),
+ ops.Constant(c, NumpyArrayF32(0.0)), padding_config)
+ self._ExecuteAndCompareClose(
+ c,
+ expected=[[[0.0, 0.0, 0.0], [1.0, 2.0, 0.0], [0.0, 0.0, 0.0],
+ [3.0, 4.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
+
+ def testReshape(self):
+ c = self._NewComputation()
+ ops.Reshape(
+ ops.Constant(c, NumpyArrayS32([[1, 2], [3, 4], [5, 6]])),
+ dimensions=[0, 1],
+ new_sizes=[2, 3])
+ self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3], [4, 5, 6]]])
+
+ def testCollapse(self):
+ c = self._NewComputation()
+ ops.Collapse(
+ ops.Constant(c, NumpyArrayS32([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])),
+ dimensions=[1, 2])
+ self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3, 4], [5, 6, 7, 8]]])
+
+ def testRev(self):
+ c = self._NewComputation()
+ ops.Rev(
+ ops.Constant(c, NumpyArrayS32([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])),
+ dimensions=[0, 2])
+ self._ExecuteAndCompareExact(
+ c, expected=[[[[6, 5], [8, 7]], [[2, 1], [4, 3]]]])
+
+ def testReducePrecision(self):
+ c = self._NewComputation()
+ ops.ReducePrecision(
+ ops.Constant(c, NumpyArrayF32([float.fromhex("0x1.32fffep-3")])),
+ exponent_bits=8,
+ mantissa_bits=7)
+ self._ExecuteAndCompareClose(c, expected=[[float.fromhex("0x1.32p-3")]])
+
+ def testClampF32(self):
+ c = self._NewComputation()
+ ops.Clamp(
+ ops.Constant(c, NumpyArrayF32(-1)),
+ ops.Constant(c, NumpyArrayF32([-2, -1, 0, 1, 2, 3])),
+ ops.Constant(c, NumpyArrayF32(2)))
+ self._ExecuteAndCompareExact(c, expected=[[-1, -1, 0, 1, 2, 2]])
+
+ def testClampS32(self):
+ c = self._NewComputation()
+ ops.Clamp(
+ ops.Constant(c, NumpyArrayS32(-1)),
+ ops.Constant(c, NumpyArrayS32([-2, -1, 0, 1, 2, 3])),
+ ops.Constant(c, NumpyArrayS32(2)))
+ self._ExecuteAndCompareExact(c, expected=[[-1, -1, 0, 1, 2, 2]])
+
+ def testSelect(self):
+ c = self._NewComputation()
+ ops.Select(
+ ops.Constant(c, NumpyArrayBool([True, False, False, True, False])),
+ ops.Constant(c, NumpyArrayS32([1, 2, 3, 4, 5])),
+ ops.Constant(c, NumpyArrayS32([-1, -2, -3, -4, -5])))
+ self._ExecuteAndCompareExact(c, expected=[[1, -2, -3, 4, -5]])
+
+ def testSlice(self):
+ c = self._NewComputation()
+ ops.Slice(
+ ops.Constant(c, NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ [1, 0], [3, 2], [1, 1])
+ self._ExecuteAndCompareExact(c, expected=[[[4, 5], [7, 8]]])
+
+ def testSliceInDim(self):
+ c = self._NewComputation()
+ ops.SliceInDim(
+ ops.Constant(c, NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ start_index=1,
+ limit_index=2,
+ stride=1,
+ dimno=1)
+ self._ExecuteAndCompareExact(c, expected=[[[2], [5], [8]]])
+ ops.SliceInDim(
+ ops.Constant(c, NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ start_index=0,
+ limit_index=3,
+ stride=2,
+ dimno=0)
+ self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3], [7, 8, 9]]])
+
+ def testDynamicSlice(self):
+ c = self._NewComputation()
+ ops.DynamicSlice(
+ ops.Constant(c, NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ [ops.Constant(c, NumpyArrayS32([1, 0]))], [2, 2])
+ self._ExecuteAndCompareExact(c, expected=[[[4, 5], [7, 8]]])
+
+ def testDynamicUpdateSlice(self):
+ c = self._NewComputation()
+ ops.DynamicUpdateSlice(
+ ops.Constant(c, NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
+ ops.Constant(c, NumpyArrayS32([[1, 2], [3, 4]])),
+ [ops.Constant(c, NumpyArrayS32([1, 1]))])
+ self._ExecuteAndCompareExact(
+ c, expected=[[[1, 2, 3], [4, 1, 2], [7, 3, 4]]])
+
+ def testTuple(self):
+ c = self._NewComputation()
+ ops.Tuple(c, [
+ ops.Constant(c, np.int32(42)),
+ ops.Constant(c, NumpyArrayF32([1.0, 2.0])),
+ ops.Constant(c, NumpyArrayBool([True, False, False, True]))
+ ])
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
+ self.assertLen(result, 3)
+ np.testing.assert_equal(result[0], 42)
+ np.testing.assert_allclose(result[1], [1.0, 2.0])
+ np.testing.assert_equal(result[2], [True, False, False, True])
+
+ def testGetTupleElement(self):
+ c = self._NewComputation()
+ ops.GetTupleElement(
+ ops.Tuple(c, [
+ ops.Constant(c, np.int32(42)),
+ ops.Constant(c, NumpyArrayF32([1.0, 2.0])),
+ ops.Constant(c, NumpyArrayBool([True, False, False, True]))
+ ]), 1)
+ self._ExecuteAndCompareClose(c, expected=[[1.0, 2.0]])
+
+ def testBroadcast(self):
+ c = self._NewComputation()
+ ops.Broadcast(
+ ops.Constant(c, NumpyArrayS32([10, 20, 30, 40])), sizes=(3,))
+ self._ExecuteAndCompareExact(
+ c, expected=[[[10, 20, 30, 40], [10, 20, 30, 40], [10, 20, 30, 40]]])
+
+ def testBroadcastInDim(self):
+ c = self._NewComputation()
+ ops.BroadcastInDim(ops.Constant(c, NumpyArrayS32([1, 2])), [2, 2], [0])
+ self._ExecuteAndCompareExact(c, expected=[[[1, 1], [2, 2]]])
+ ops.BroadcastInDim(ops.Constant(c, NumpyArrayS32([1, 2])), [2, 2], [1])
+ self._ExecuteAndCompareExact(c, expected=[[[1, 2], [1, 2]]])
+
+ def testRngNormal(self):
+ shape = (2, 3)
+ c = self._NewComputation()
+ ops.RngNormal(
+ ops.Constant(c, NumpyArrayF32(0.)),
+ ops.Constant(c, NumpyArrayF32(1.)),
+ shape=xla_client.Shape.array_shape(xla_client.PrimitiveType.F32,
+ shape))
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
+ # since the result is random, we just check shape and uniqueness
+ self.assertLen(result, 1)
+ self.assertEqual(result[0].shape, shape)
+ self.assertLen(np.unique(result[0]), np.prod(shape))
+
+ def testRngUniformF32(self):
+ lo, hi = 2., 4.
+ shape = (2, 3)
+ c = self._NewComputation()
+ ops.RngUniform(
+ ops.Constant(c, NumpyArrayF32(lo)),
+ ops.Constant(c, NumpyArrayF32(hi)),
+ shape=xla_client.Shape.array_shape(xla_client.PrimitiveType.F32,
+ shape))
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
+ # since the result is random, we just check shape, uniqueness, and range
+ self.assertLen(result, 1)
+ self.assertEqual(result[0].shape, shape)
+ self.assertLen(np.unique(result[0]), np.prod(shape))
+ self.assertTrue(np.all(lo <= result[0]))
+ self.assertTrue(np.all(result[0] < hi))
+
+ def testRngUniformS32(self):
+ lo, hi = 2, 4
+ shape = (2, 3)
+ c = self._NewComputation()
+ ops.RngUniform(
+ ops.Constant(c, NumpyArrayS32(lo)),
+ ops.Constant(c, NumpyArrayS32(hi)),
+ shape=xla_client.Shape.array_shape(xla_client.PrimitiveType.S32,
+ shape))
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
+ # since the result is random, we just check shape, integrality, and range
+ self.assertLen(result, 1)
+ self.assertEqual(result[0].shape, shape)
+ self.assertEqual(result[0].dtype, np.int32)
+ self.assertTrue(np.all(lo <= result[0]))
+ self.assertTrue(np.all(result[0] < hi))
+
+ def testCholesky(self):
+ l = np.array([[4, 0, 0, 0], [6, 5, 0, 0], [2, 14, 16, 0], [3, 6, 1, 4]],
+ dtype=np.float32)
+ c = self._NewComputation()
+ ops.Cholesky(ops.Constant(c, np.tril(np.dot(l, l.T))))
+ self._ExecuteAndCompareClose(c, expected=[l], rtol=1e-4)
+
+ def testSort(self):
+ keys = np.array([[2, 4, 1, 3], [3, 1, 4, 2]], dtype=np.float32)
+ c = self._NewComputation()
+ ops.Sort(c, [ops.Constant(c, keys)], is_stable=True)
+ self._ExecuteAndCompareClose(
+ c,
+ expected=[np.array([[1, 2, 3, 4], [1, 2, 3, 4]], dtype=np.float32)])
+
+ def testSortKeyVal(self):
+ keys = np.array([[2, 4, 1, 3], [3, 1, 4, 2]], dtype=np.float32)
+ values = np.array([[0, 1, 2, 3], [4, 5, 6, 7]], dtype=np.int32)
+ c = self._NewComputation()
+ ops.Sort(c, (ops.Constant(c, keys), ops.Constant(c, values)), dimension=0)
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), backend=self.backend)
+ self.assertLen(result, 2)
+ np.testing.assert_allclose(result[0], [[2, 1, 1, 2], [3, 4, 4, 3]])
+ np.testing.assert_equal(result[1], [[0, 5, 2, 7], [4, 1, 6, 3]])
+
+ def testSortCustomComparator(self):
+ b = self._NewComputation("comparator")
+ p0 = ops.Parameter(b, 0, xla_client.shape_from_pyval(NumpyArrayF32(0)))
+ q0 = ops.Parameter(b, 1, xla_client.shape_from_pyval(NumpyArrayF32(0)))
+ p1 = ops.Parameter(b, 2, xla_client.shape_from_pyval(NumpyArrayS32(0)))
+ q1 = ops.Parameter(b, 3, xla_client.shape_from_pyval(NumpyArrayS32(0)))
+ ops.Or(ops.Lt(p0, q0), ops.And(ops.Eq(p0, q0), ops.Gt(p1, q1)))
+ comparator = b.Build()
+
+ keys = np.array([[2, 3, 1, 3], [3, 1, 2, 2]], dtype=np.float32)
+ values = np.array([[0, 1, 2, 3], [4, 5, 6, 7]], dtype=np.int32)
+ c = self._NewComputation()
+ ops.Sort(
+ c, (ops.Constant(c, keys), ops.Constant(c, values)),
+ dimension=1,
+ comparator=comparator)
+ result = xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()))
+ self.assertLen(result, 2)
+ np.testing.assert_allclose(result[0], [[1, 2, 3, 3], [1, 2, 2, 3]])
+ np.testing.assert_equal(result[1], [[2, 0, 3, 1], [5, 7, 6, 4]])
+
+ def testQR(self):
+ a = np.array([[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166],
+ [10, 63, 166, 310]],
+ dtype=np.float32)
+ c = self._NewComputation()
+ ops.Tuple(c, ops.QR(ops.Constant(c, a), full_matrices=True))
+ q, r = self._Execute(c, ())
+ np.testing.assert_allclose(np.dot(q, r), a, rtol=1e-4)
+
+ def testEigh(self):
+ a = np.array([[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166],
+ [10, 63, 166, 310]],
+ dtype=np.float32)
+ a = (a + a.T) / 2
+
+ c = self._NewComputation()
+ ops.Tuple(c, ops.Eigh(ops.Constant(c, a), lower=True))
+ # TODO(b/129396575): Turn this test back on when it passes without
+ # fastmath.
+ # v, w = self._Execute(c, ())
+ # self.assertLess(np.linalg.norm(np.dot(a, v) - w * v), 1e-3)
+
+ def testSVD(self):
+ a = np.array([[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166],
+ [10, 63, 166, 310]],
+ dtype=np.float32)
+ c = self._NewComputation()
+ ops.Tuple(c, ops.SVD(ops.Constant(c, a)))
+ u, d, v = self._Execute(c, ())
+ self.assertLess(np.linalg.norm(a - np.matmul(u * d, v.T)), 1e-3)
+
+ def testTriangularSolve(self):
+ a_vals = np.array(
+ [[2, 0, 0, 0], [3, 6, 0, 0], [4, 7, 9, 0], [5, 8, 10, 11]],
+ dtype=np.float32)
+ b_vals = np.array([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]],
+ dtype=np.float32)
+
+ c = self._NewComputation()
+ ops.TriangularSolve(
+ ops.Constant(c, a_vals),
+ ops.Constant(c, b_vals),
+ left_side=False,
+ lower=True,
+ transpose_a=ops.TriangularSolveOptions_Transpose.TRANSPOSE,
+ unit_diagonal=False)
+ self._ExecuteAndCompareClose(
+ c,
+ expected=[
+ np.array([
+ [0.5, 0.08333334, 0.04629629, 0.03367003],
+ [2.5, -0.25, -0.1388889, -0.1010101],
+ [4.5, -0.58333331, -0.32407406, -0.23569024],
+ ],
+ dtype=np.float32)
+ ],
+ rtol=1e-4)
+
+ def testIsConstant(self):
+ c = self._NewComputation()
+ a = ops.Constant(c, np.int32(3))
+ b = ops.Constant(c, np.int32(1))
+ x = ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayS32(0)))
+ const_expr = ops.Sub(b, a)
+ non_const_expr = ops.Mul(const_expr, x)
+ self.assertTrue(c.IsConstant(const_expr))
+ self.assertFalse(c.IsConstant(non_const_expr))
+
+ def testGather(self):
+ a = np.arange(9).astype(np.int32).reshape((3, 3))
+ indices = np.array([[[0, 2], [2, 1]], [[1, 2], [2, 0]]], dtype=np.int32)
+ dnums = xla_client.GatherDimensionNumbers()
+ dnums.offset_dims.append(1)
+ dnums.offset_dims.append(2)
+ dnums.start_index_map.append(0)
+ dnums.start_index_map.append(1)
+ dnums.index_vector_dim = 2
+ c = self._NewComputation()
+ ops.Gather(
+ ops.Constant(c, a),
+ ops.Constant(c, indices),
+ dnums,
+ slice_sizes=[1, 1])
+ g, = self._Execute(c, ())
+ expected = np.array([[[[2, 7]]], [[[5, 6]]]], dtype=np.int32)
+ np.testing.assert_allclose(g, expected, rtol=1e-4)
+
+ def testFft(self):
+ if self.backend.platform == "tpu":
+ self.skipTest("TPU only supports 1D FFT")
+ shape = [2, 3, 4, 5]
+ rng = np.random.RandomState(0)
+ a = rng.randn(*shape) + 1.0j * rng.randn(*shape)
+ a = a.astype(np.complex64)
+ # FFT
+ c = self._NewComputation()
+ ops.Fft(ops.Constant(c, a), xla_client.FftType.FFT, shape[-3:])
+ self._ExecuteAndCompareClose(
+ c, expected=[np.fft.fftn(a, axes=(1, 2, 3))], rtol=1e-4)
+ # IFFT
+ c = self._NewComputation()
+ ops.Fft(ops.Constant(c, a), xla_client.FftType.IFFT, shape[-3:])
+ self._ExecuteAndCompareClose(
+ c, expected=[np.fft.ifftn(a, axes=(1, 2, 3))], rtol=1e-4)
+ # RFFT
+ b = rng.randn(*shape).astype(np.float32)
+ c = self._NewComputation()
+ ops.Fft(ops.Constant(c, b), xla_client.FftType.RFFT, shape[-3:])
+ self._ExecuteAndCompareClose(
+ c, expected=[np.fft.rfftn(b, axes=(1, 2, 3))], rtol=1e-4)
+ # IRFFT
+ c = self._NewComputation()
+ ops.Fft(ops.Constant(c, a), xla_client.FftType.IRFFT, [3, 4, 8])
+ self._ExecuteAndCompareClose(
+ c, expected=[np.fft.irfftn(a, axes=(1, 2, 3))], rtol=1e-4)
+
+ def testNextAfter(self):
+ c = self._NewComputation()
+ ops.NextAfter(
+ ops.Constant(c, np.array([1, 2], dtype=np.float32)),
+ ops.Constant(c, np.array([2, 1], dtype=np.float32)))
+ out, = self._Execute(c, ())
+ eps = np.finfo(np.float32).eps
+ np.testing.assert_equal(
+ np.array([eps + 1, 2 - eps], dtype=np.float32), out)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testRegularizedIncompleteBeta(self, dtype):
+ x = np.array([0.53787335, 0.24015466, 0.47494545, 0.13567594, 0.95114538],
+ dtype=dtype)
+ a = np.array([0.00753073, 0.34813385, 0.30485708, 1.29298632, 0.51472606],
+ dtype=dtype)
+ b = np.array([0.55688389, 0.59794214, 0.42661022, 1.59748339, 0.95047677],
+ dtype=dtype)
+ c = self._NewComputation()
+ ops.RegularizedIncompleteBeta(
+ ops.Constant(c, a), ops.Constant(c, b), ops.Constant(c, x))
+ expected = np.array(
+ [0.98923271, 0.48575411, 0.57952568, 0.12579775, 0.96989155])
+ self._ExecuteAndCompareClose(c, expected=[expected], rtol=2e-2)
+
+ tests.append(SingleOpTest)
+
+ class EmbeddedComputationsTest(ComputationTest):
+ """Tests for XLA graphs with embedded computations (such as maps)."""
+
+ def _CreateConstantComputation(self, in_dtype, out_dtype):
+ """Computation (A) -> B that returns a constant 1 for any input."""
+ c = self._NewComputation("constant_{}_{}_one".format(
+ in_dtype.__name__, out_dtype.__name__))
+ ops.Parameter(c, 0,
+ xla_client.shape_from_pyval(np.array(0, dtype=in_dtype)))
+ ops.Constant(c, out_dtype(1))
+ return c.Build()
+
+ def _CreateMulBy2Computation(self, dtype):
+ """Computation (dtype) -> dtype that multiplies its parameter by 2."""
+ c = self._NewComputation("mul_f32_by2")
+ ops.Mul(
+ ops.Parameter(
+ c, 0,
+ xla_client.shape_from_pyval(np.array(
+ 0, dtype=dtype)).with_major_to_minor_layout_if_absent()),
+ ops.Constant(c, dtype(2.0)))
+ return c.Build()
+
+ def _CreateMulF32ByParamComputation(self):
+ """Computation (f32) -> f32 that multiplies one parameter by the other."""
+ c = self._NewComputation("mul_f32_by_param")
+ ops.Mul(
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(0))),
+ ops.Parameter(c, 1, xla_client.shape_from_pyval(NumpyArrayF32(0))))
+ return c.Build()
+
+ def _CreateBinaryAddComputation(self, dtype):
+ """Computation (dtype, dtype) -> dtype that adds its two parameters."""
+ c = self._NewComputation("add_param0_by_param1")
+ shape = xla_client.shape_from_pyval(np.array(0, dtype=dtype))
+ shape = shape.with_major_to_minor_layout_if_absent()
+ ops.Add(ops.Parameter(c, 0, shape), ops.Parameter(c, 1, shape))
+ return c.Build()
+
+ def _CreateBinaryGeComputation(self, dtype):
+ """Computation (dtype, dtype) -> bool that tests param0 >= param1."""
+ c = self._NewComputation("param0_lt_param1")
+ shape = xla_client.shape_from_pyval(np.array(0, dtype=dtype))
+ shape = shape.with_major_to_minor_layout_if_absent()
+ ops.Ge(ops.Parameter(c, 0, shape), ops.Parameter(c, 1, shape))
+ return c.Build()
+
+ def _MakeSample3DArray(self, dtype):
+ return np.array([[[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]],
+ [[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]]],
+ dtype=dtype)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testCall(self, dtype):
+ c = self._NewComputation()
+ ops.Call(
+ c,
+ self._CreateMulBy2Computation(dtype),
+ operands=(ops.Constant(c, dtype(5.0)),))
+ self._ExecuteAndCompareClose(c, expected=[10.0])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}_{}".format(in_dtype.__name__, out_dtype.__name__),
+ "in_dtype": in_dtype,
+ "out_dtype": out_dtype,
+ } for in_dtype, out_dtype in [[np.float32, np.int32]])
+ def testMapEachElementToConstant(self, in_dtype, out_dtype):
+ c = self._NewComputation()
+ ops.Map(c,
+ [ops.Constant(c, np.array([1.0, 2.0, 3.0, 4.0], dtype=in_dtype))],
+ self._CreateConstantComputation(in_dtype, out_dtype), [0])
+ self._ExecuteAndCompareExact(c, expected=[[1, 1, 1, 1]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testMapMulBy2(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ c = self._NewComputation()
+ ops.Map(c, [ops.Constant(c, np.array([1.0, 2.0, 3.0, 4.0], dtype=dtype))],
+ self._CreateMulBy2Computation(dtype), [0])
+ self._ExecuteAndCompareClose(c, expected=[[2.0, 4.0, 6.0, 8.0]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testSimpleMapChain(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ # Chains a map of constant-out with a map of mul-by-2
+ c = self._NewComputation()
+ const = ops.Map(
+ c, [ops.Constant(c, np.array([1.0, 2.0, 3.0, 4.0], dtype=dtype))],
+ self._CreateConstantComputation(dtype, dtype), [0])
+ ops.Map(c, [const], self._CreateMulBy2Computation(dtype), [0])
+ self._ExecuteAndCompareClose(c, expected=[[2.0, 2.0, 2.0, 2.0]])
+
+ # TODO(b/154752816): bfloat16 crashes in evaluator.
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes if dtype != bfloat16)
+ def testDivVectorsWithMap(self, dtype):
+
+ def DivComputation():
+ c = self._NewComputation("div_param0_by_param1")
+ shape = xla_client.shape_from_pyval(np.array(0, dtype=dtype))
+ ops.Div(ops.Parameter(c, 0, shape), ops.Parameter(c, 1, shape))
+ return c.Build()
+
+ c = self._NewComputation()
+ ops.Map(c, (ops.Constant(c, np.array([1.0, 2.0, 3.0, 4.0], dtype=dtype)),
+ ops.Constant(c, np.array([5.0, 5.0, 4.0, 4.0], dtype=dtype))),
+ DivComputation(), [0])
+ self._ExecuteAndCompareClose(
+ c, expected=[[0.2, 0.4, 0.75, 1.0]], rtol=1e-3)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testSelectAndScatter(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ c = self._NewComputation()
+ operand = ops.Constant(
+ c, np.array([[1., 2., 6.], [4., 5., 3.]], dtype=dtype))
+ window_dimensions = (2, 1)
+ window_strides = (1, 2)
+ padding = xla_client.window_padding_type_to_pad_values(
+ xla_client.PaddingType.VALID,
+ c.GetShape(operand).dimensions(), window_dimensions, window_strides)
+ ops.SelectAndScatterWithGeneralPadding(
+ operand,
+ select=self._CreateBinaryGeComputation(dtype),
+ window_dimensions=window_dimensions,
+ window_strides=window_strides,
+ padding=padding,
+ source=ops.Constant(c, np.array([[0.1, 0.2]], dtype=dtype)),
+ init_value=ops.Constant(c, np.array(1, dtype=dtype)),
+ scatter=self._CreateBinaryAddComputation(dtype))
+ self._ExecuteAndCompareClose(
+ c, expected=[[[1., 1., 1.2], [1.1, 1., 1.]]], rtol=5e-3)
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testReduce1DtoScalar(self, dtype):
+ c = self._NewComputation()
+ ops.Reduce(
+ c,
+ operands=[
+ ops.Constant(c, np.array([1.0, 2.0, 3.0, 4.0], dtype=dtype))
+ ],
+ init_values=[ops.Constant(c, dtype(0))],
+ computation=self._CreateBinaryAddComputation(dtype),
+ dimensions_to_reduce=[0])
+ self._ExecuteAndCompareClose(c, expected=[10])
+
+ # TODO(phawkins): test comparison harness doesn't support bfloat16
+ @parameterized.named_parameters({
+ "testcase_name": "_{}_dim{}".format(dtype.__name__, dim),
+ "dtype": dtype,
+ "dim": dim,
+ } for dtype in float_dtypes if dtype != bfloat16 for dim in range(2))
+ def testReduce2DTo1D(self, dtype, dim):
+ input_array = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=dtype)
+ c = self._NewComputation()
+ ops.Reduce(
+ c,
+ operands=[ops.Constant(c, input_array)],
+ init_values=[ops.Constant(c, dtype(0))],
+ computation=self._CreateBinaryAddComputation(dtype),
+ dimensions_to_reduce=[dim])
+ self._ExecuteAndCompareClose(c, expected=[np.sum(input_array, axis=dim)])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}_dims[{}]".format(dtype.__name__, dims),
+ "dtype": dtype,
+ "dims": tuple(dims)
+ } for dtype in float_dtypes for dims in itertools.permutations(range(3)))
+ def testReduce3DAllPossibleWaysF32(self, dtype, dims):
+ input_array = self._MakeSample3DArray(dtype)
+ c = self._NewComputation()
+ ops.Reduce(
+ c,
+ operands=[ops.Constant(c, input_array)],
+ init_values=[ops.Constant(c, dtype(0))],
+ computation=self._CreateBinaryAddComputation(dtype),
+ dimensions_to_reduce=dims)
+ self._ExecuteAndCompareClose(c, expected=[np.sum(input_array, axis=dims)])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testReduceWindowValidUnitStrides(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ input_array = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=dtype)
+ c = self._NewComputation()
+ window_dimensions = (2, 1)
+ window_strides = (1, 1)
+ padding = xla_client.window_padding_type_to_pad_values(
+ xla_client.PaddingType.VALID, input_array.shape, window_dimensions,
+ window_strides)
+ ops.ReduceWindowWithGeneralPadding(
+ operand=ops.Constant(c, input_array),
+ init_value=ops.Constant(c, dtype(0)),
+ computation=self._CreateBinaryAddComputation(dtype),
+ window_dimensions=window_dimensions,
+ window_strides=window_strides,
+ base_dilations=[],
+ window_dilations=[],
+ padding=padding)
+ self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.]]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testReduceWindowSameUnitStrides(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ input_array = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=dtype)
+ c = self._NewComputation()
+ window_dimensions = (2, 1)
+ window_strides = (1, 1)
+ padding = xla_client.window_padding_type_to_pad_values(
+ xla_client.PaddingType.SAME, input_array.shape, window_dimensions,
+ window_strides)
+ ops.ReduceWindowWithGeneralPadding(
+ operand=ops.Constant(c, input_array),
+ init_value=ops.Constant(c, dtype(0)),
+ computation=self._CreateBinaryAddComputation(dtype),
+ window_dimensions=window_dimensions,
+ window_strides=window_strides,
+ base_dilations=[],
+ window_dilations=[],
+ padding=padding)
+ self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.], [4., 5., 6.]]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testReduceWindowValidGeneralStrides(self, dtype):
+ if dtype == np.float64 and self.backend.platform == "tpu":
+ self.skipTest("TPU doesn't support float64")
+ input_array = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=dtype)
+ c = self._NewComputation()
+ window_dimensions = (2, 1)
+ window_strides = (1, 2)
+ padding = xla_client.window_padding_type_to_pad_values(
+ xla_client.PaddingType.VALID, input_array.shape, window_dimensions,
+ window_strides)
+ ops.ReduceWindowWithGeneralPadding(
+ operand=ops.Constant(c, input_array),
+ init_value=ops.Constant(c, dtype(0)),
+ computation=self._CreateBinaryAddComputation(dtype),
+ window_dimensions=window_dimensions,
+ window_strides=window_strides,
+ base_dilations=[],
+ window_dilations=[],
+ padding=padding)
+ self._ExecuteAndCompareClose(c, expected=[[[5., 9.]]])
+
+ @parameterized.named_parameters({
+ "testcase_name": "_{}".format(dtype.__name__),
+ "dtype": dtype,
+ } for dtype in float_dtypes)
+ def testWhile(self, dtype):
+
+ def LessThan10Cond():
+ c = self._NewComputation("test_lt_10")
+ shape = xla_client.shape_from_pyval(np.array(0, dtype=dtype))
+ ops.Lt(ops.Parameter(c, 0, shape), ops.Constant(c, dtype(10.)))
+ return c.Build()
+
+ cond = LessThan10Cond()
+ body = self._CreateMulBy2Computation(dtype)
+ c = self._NewComputation()
+ init = ops.Constant(c, dtype(1.))
+ ops.While(cond, body, init)
+ self._ExecuteAndCompareClose(c, expected=[16.])
+
+ def testConditionalTrue(self):
+ c = self._NewComputation()
+ pred = ops.Constant(c, np.bool_(True))
+ true_operand = ops.Constant(c, np.float32(3.))
+ true_computation = self._CreateMulBy2Computation(np.float32)
+ false_operand = ops.Constant(c, np.float32(2.))
+ false_computation = self._CreateConstantComputation(
+ np.float32, np.float32)
+ ops.Conditional(pred, true_operand, true_computation, false_operand,
+ false_computation)
+ self._ExecuteAndCompareClose(c, expected=[6.])
+
+ def testConditionalFalse(self):
+ c = self._NewComputation()
+ pred = ops.Constant(c, np.bool_(False))
+ true_operand = ops.Constant(c, np.float32(3.))
+ true_computation = self._CreateMulBy2Computation(np.float32)
+ false_operand = ops.Constant(c, np.float32(2.))
+ false_computation = self._CreateConstantComputation(
+ np.float32, np.float32)
+ ops.Conditional(pred, true_operand, true_computation, false_operand,
+ false_computation)
+ self._ExecuteAndCompareClose(c, expected=[1.])
+
+ @unittest.skipIf(cloud_tpu, "not implemented")
+ def testInfeedS32Values(self):
+ to_infeed = NumpyArrayS32([1, 2, 3, 4])
+ c = self._NewComputation()
+ ops.GetTupleElement(
+ ops.InfeedWithToken(
+ ops.CreateToken(c),
+ xla_client.shape_from_pyval(
+ to_infeed[0]).with_major_to_minor_layout_if_absent()), 0)
+ compiled_c = self.backend.compile(c.Build())
+ device = self.backend.local_devices()[0]
+ for item in to_infeed:
+ xla_client.transfer_to_infeed(item, device=device)
+
+ for item in to_infeed:
+ result, = xla_client.execute_with_python_values(
+ compiled_c, backend=self.backend)
+ self.assertEqual(result, item)
+
+ @unittest.skipIf(cloud_tpu, "not implemented")
+ def testInfeedTuple(self):
+ to_infeed = (NumpyArrayS32([1, 2, 3, 4]), NumpyArrayS32([[7], [8]]))
+ c = self._NewComputation()
+ ops.GetTupleElement(
+ ops.InfeedWithToken(
+ ops.CreateToken(c),
+ xla_client.shape_from_pyval(
+ to_infeed).with_major_to_minor_layout_if_absent()), 0)
+ compiled_c = self.backend.compile(c.Build())
+ device = self.backend.local_devices()[0]
+ xla_client.transfer_to_infeed(to_infeed, device=device)
+
+ result = xla_client.execute_with_python_values(
+ compiled_c, backend=self.backend)
+ self.assertLen(result, 2)
+ np.testing.assert_equal(result[0], to_infeed[0])
+ np.testing.assert_equal(result[1], to_infeed[1])
+
+ @unittest.skipIf(cloud_tpu, "not implemented")
+ def testInfeedThenOutfeedS32(self):
+ to_round_trip = NumpyArrayS32([1, 2, 3, 4])
+ c = self._NewComputation()
+ x_and_token = ops.InfeedWithToken(
+ ops.CreateToken(c),
+ xla_client.shape_from_pyval(
+ to_round_trip[0]).with_major_to_minor_layout_if_absent())
+ x = ops.GetTupleElement(x_and_token, 0)
+ token = ops.GetTupleElement(x_and_token, 1)
+ outfeed_shape = xla_client.shape_from_pyval(
+ to_round_trip[0]).with_major_to_minor_layout_if_absent()
+ ops.OutfeedWithToken(x, token, outfeed_shape)
+
+ compiled_c = self.backend.compile(c.Build())
+ device = self.backend.local_devices()[0]
+
+ for want in to_round_trip:
+ execution = threading.Thread(target=lambda: compiled_c.Execute([]))
+ execution.start()
+ xla_client.transfer_to_infeed(want, device=device)
+ got = xla_client.transfer_from_outfeed(outfeed_shape, device=device)
+ execution.join()
+ self.assertEqual(want, got)
+
+ def testScatter(self):
+ a = np.arange(9).astype(np.int32).reshape((3, 3))
+ scatter_indices = np.array([0, 2], dtype=np.int32)
+ updates = np.array([[10, 20, 30], [70, 80, 90]], dtype=np.int32)
+
+ dnums = xla_client.ScatterDimensionNumbers()
+ dnums.update_window_dims.append(1)
+ dnums.inserted_window_dims.append(0)
+ dnums.scatter_dims_to_operand_dims.append(0)
+ dnums.index_vector_dim = 1
+
+ c = self._NewComputation()
+ ops.Scatter(
+ ops.Constant(c, a), ops.Constant(c, scatter_indices),
+ ops.Constant(c, updates), self._CreateBinaryAddComputation(np.int32),
+ dnums)
+ expected = np.array([[10, 21, 32], [3, 4, 5], [76, 87, 98]],
+ dtype=np.int32)
self._ExecuteAndCompareClose(c, expected=[expected])
- _TransposeAndTest(NumpyArrayF32([[1, 2, 3], [4, 5, 6]]), [0, 1])
- _TransposeAndTest(NumpyArrayF32([[1, 2, 3], [4, 5, 6]]), [1, 0])
- _TransposeAndTest(NumpyArrayF32([[1, 2], [4, 5]]), [0, 1])
- _TransposeAndTest(NumpyArrayF32([[1, 2], [4, 5]]), [1, 0])
+ class ErrorTest(ComputationTest):
- arr = np.random.RandomState(0).randn(2, 3, 4).astype(np.float32)
- for permutation in itertools.permutations(range(arr.ndim)):
- _TransposeAndTest(arr, permutation)
- _TransposeAndTest(np.asfortranarray(arr), permutation)
+ def setUp(self):
+ super(ErrorTest, self).setUp()
+ self.f32_scalar_2 = NumpyArrayF32(2.0)
+ self.s32_scalar_2 = NumpyArrayS32(2)
- def testEq(self):
- c = self._NewComputation()
- c.Eq(
- c.Constant(NumpyArrayS32([1, 2, 3, 4])),
- c.Constant(NumpyArrayS32([4, 2, 3, 1])))
- self._ExecuteAndCompareExact(c, expected=[[False, True, True, False]])
-
- def testNe(self):
- c = self._NewComputation()
- c.Ne(
- c.Constant(NumpyArrayS32([1, 2, 3, 4])),
- c.Constant(NumpyArrayS32([4, 2, 3, 1])))
- self._ExecuteAndCompareExact(c, expected=[[True, False, False, True]])
-
- c.Ne(
- c.Constant(NumpyArrayF32([-2.0, 0.0,
- float("nan"),
- float("nan")])),
- c.Constant(NumpyArrayF32([2.0, -0.0, 1.0, float("nan")])))
- self._ExecuteAndAssertWith(
- np.testing.assert_allclose, c, (), expected=[[True, False, True, True]])
-
- def testGt(self):
- c = self._NewComputation()
- c.Gt(
- c.Constant(NumpyArrayS32([1, 2, 3, 4, 9])),
- c.Constant(NumpyArrayS32([1, 0, 2, 7, 12])))
- self._ExecuteAndCompareExact(
- c, expected=[[False, True, True, False, False]])
-
- def testGe(self):
- c = self._NewComputation()
- c.Ge(
- c.Constant(NumpyArrayS32([1, 2, 3, 4, 9])),
- c.Constant(NumpyArrayS32([1, 0, 2, 7, 12])))
- self._ExecuteAndCompareExact(c, expected=[[True, True, True, False, False]])
-
- def testLt(self):
- c = self._NewComputation()
- c.Lt(
- c.Constant(NumpyArrayS32([1, 2, 3, 4, 9])),
- c.Constant(NumpyArrayS32([1, 0, 2, 7, 12])))
- self._ExecuteAndCompareExact(
- c, expected=[[False, False, False, True, True]])
-
- def testLe(self):
- c = self._NewComputation()
- c.Le(
- c.Constant(NumpyArrayS32([1, 2, 3, 4, 9])),
- c.Constant(NumpyArrayS32([1, 0, 2, 7, 12])))
- self._ExecuteAndCompareExact(c, expected=[[True, False, False, True, True]])
-
- def testMax(self):
- c = self._NewComputation()
- c.Max(
- c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0, 9.0])),
- c.Constant(NumpyArrayF32([1.0, 0.0, 2.0, 7.0, 12.0])))
- self._ExecuteAndCompareExact(c, expected=[[1.0, 2.0, 3.0, 7.0, 12.0]])
-
- def testMaxExplicitBroadcastDim0(self):
- c = self._NewComputation()
- c.Max(
- c.Constant(NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF32([3, 4, 5])),
- broadcast_dimensions=(0,))
- self._ExecuteAndCompareExact(
- c, expected=[[[3, 3, 3], [4, 5, 6], [7, 8, 9]]])
-
- def testMaxExplicitBroadcastDim1(self):
- c = self._NewComputation()
- c.Max(
- c.Constant(NumpyArrayF32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayF32([3, 4, 5])),
- broadcast_dimensions=(1,))
- self._ExecuteAndCompareExact(
- c, expected=[[[3, 4, 5], [4, 5, 6], [7, 8, 9]]])
-
- def testMin(self):
- c = self._NewComputation()
- c.Min(
- c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0, 9.0])),
- c.Constant(NumpyArrayF32([1.0, 0.0, 2.0, 7.0, 12.0])))
- self._ExecuteAndCompareExact(c, expected=[[1.0, 0.0, 2.0, 4.0, 9.0]])
-
- def testPad(self):
- c = self._NewComputation()
- c.Pad(
- c.Constant(NumpyArrayF32([[1.0, 2.0], [3.0, 4.0]])),
- c.Constant(NumpyArrayF32(0.0)), [(1, 2, 1), (0, 1, 0)])
- self._ExecuteAndCompareClose(
- c,
- expected=[[[0.0, 0.0, 0.0], [1.0, 2.0, 0.0], [0.0, 0.0, 0.0],
- [3.0, 4.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
-
- def testPadWithPaddingConfig(self):
- c = self._NewComputation()
- padding_config = xla_client.PaddingConfig()
- for lo, hi, interior in [(1, 2, 1), (0, 1, 0)]:
- dimension = xla_client.PaddingConfigDimension()
- dimension.edge_padding_low = lo
- dimension.edge_padding_high = hi
- dimension.interior_padding = interior
- padding_config.dimensions.append(dimension)
- c.Pad(
- c.Constant(NumpyArrayF32([[1.0, 2.0], [3.0, 4.0]])),
- c.Constant(NumpyArrayF32(0.0)), padding_config)
- self._ExecuteAndCompareClose(
- c,
- expected=[[[0.0, 0.0, 0.0], [1.0, 2.0, 0.0], [0.0, 0.0, 0.0],
- [3.0, 4.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
-
- def testReshape(self):
- c = self._NewComputation()
- c.Reshape(
- c.Constant(NumpyArrayS32([[1, 2], [3, 4], [5, 6]])),
- dimensions=[0, 1],
- new_sizes=[2, 3])
- self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3], [4, 5, 6]]])
-
- def testCollapse(self):
- c = self._NewComputation()
- c.Collapse(
- c.Constant(NumpyArrayS32([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])),
- dimensions=[1, 2])
- self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3, 4], [5, 6, 7, 8]]])
-
- def testRev(self):
- c = self._NewComputation()
- c.Rev(
- c.Constant(NumpyArrayS32([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])),
- dimensions=[0, 2])
- self._ExecuteAndCompareExact(
- c, expected=[[[[6, 5], [8, 7]], [[2, 1], [4, 3]]]])
-
- def testReducePrecision(self):
- c = self._NewComputation()
- c.ReducePrecision(
- c.Constant(NumpyArrayF32([float.fromhex("0x1.32fffep-3")])),
- exponent_bits=8,
- mantissa_bits=7)
- self._ExecuteAndCompareClose(c, expected=[[float.fromhex("0x1.32p-3")]])
-
- def testClampF32(self):
- c = self._NewComputation()
- c.Clamp(
- c.Constant(NumpyArrayF32(-1)),
- c.Constant(NumpyArrayF32([-2, -1, 0, 1, 2, 3])),
- c.Constant(NumpyArrayF32(2)))
- self._ExecuteAndCompareExact(c, expected=[[-1, -1, 0, 1, 2, 2]])
-
- def testClampS32(self):
- c = self._NewComputation()
- c.Clamp(
- c.Constant(NumpyArrayS32(-1)),
- c.Constant(NumpyArrayS32([-2, -1, 0, 1, 2, 3])),
- c.Constant(NumpyArrayS32(2)))
- self._ExecuteAndCompareExact(c, expected=[[-1, -1, 0, 1, 2, 2]])
-
- def testSelect(self):
- c = self._NewComputation()
- c.Select(
- c.Constant(NumpyArrayBool([True, False, False, True, False])),
- c.Constant(NumpyArrayS32([1, 2, 3, 4, 5])),
- c.Constant(NumpyArrayS32([-1, -2, -3, -4, -5])))
- self._ExecuteAndCompareExact(c, expected=[[1, -2, -3, 4, -5]])
-
- def testSlice(self):
- c = self._NewComputation()
- c.Slice(
- c.Constant(NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])), [1, 0],
- [3, 2])
- self._ExecuteAndCompareExact(c, expected=[[[4, 5], [7, 8]]])
-
- def testSliceInDim(self):
- c = self._NewComputation()
- c.SliceInDim(
- c.Constant(NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- start_index=1,
- limit_index=2,
- stride=1,
- dimno=1)
- self._ExecuteAndCompareExact(c, expected=[[[2], [5], [8]]])
- c.SliceInDim(
- c.Constant(NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- start_index=0,
- limit_index=3,
- stride=2,
- dimno=0)
- self._ExecuteAndCompareExact(c, expected=[[[1, 2, 3], [7, 8, 9]]])
-
- def testDynamicSlice(self):
- c = self._NewComputation()
- c.DynamicSlice(
- c.Constant(NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayS32([1, 0])), [2, 2])
- self._ExecuteAndCompareExact(c, expected=[[[4, 5], [7, 8]]])
-
- def testDynamicUpdateSlice(self):
- c = self._NewComputation()
- c.DynamicUpdateSlice(
- c.Constant(NumpyArrayS32([[1, 2, 3], [4, 5, 6], [7, 8, 9]])),
- c.Constant(NumpyArrayS32([[1, 2], [3, 4]])),
- c.Constant(NumpyArrayS32([1, 1])))
- self._ExecuteAndCompareExact(
- c, expected=[[[1, 2, 3], [4, 1, 2], [7, 3, 4]]])
-
- def testTuple(self):
- c = self._NewComputation()
- c.Tuple(
- c.ConstantS32Scalar(42), c.Constant(NumpyArrayF32([1.0, 2.0])),
- c.Constant(NumpyArrayBool([True, False, False, True])))
- result = xla_client.execute_with_python_values(c.Build().Compile())
- self.assertLen(result, 3)
- np.testing.assert_equal(result[0], 42)
- np.testing.assert_allclose(result[1], [1.0, 2.0])
- np.testing.assert_equal(result[2], [True, False, False, True])
-
- def testGetTupleElement(self):
- c = self._NewComputation()
- c.GetTupleElement(
- c.Tuple(
- c.ConstantS32Scalar(42), c.Constant(NumpyArrayF32([1.0, 2.0])),
- c.Constant(NumpyArrayBool([True, False, False, True]))), 1)
- self._ExecuteAndCompareClose(c, expected=[[1.0, 2.0]])
-
- def testBroadcast(self):
- c = self._NewComputation()
- c.Broadcast(c.Constant(NumpyArrayS32([10, 20, 30, 40])), sizes=(3,))
- self._ExecuteAndCompareExact(
- c, expected=[[[10, 20, 30, 40], [10, 20, 30, 40], [10, 20, 30, 40]]])
-
- def testBroadcastInDim(self):
- c = self._NewComputation()
- c.BroadcastInDim(c.Constant(NumpyArrayS32([1, 2])), [2, 2], [0])
- self._ExecuteAndCompareExact(c, expected=[[[1, 1], [2, 2]]])
- c.BroadcastInDim(c.Constant(NumpyArrayS32([1, 2])), [2, 2], [1])
- self._ExecuteAndCompareExact(c, expected=[[[1, 2], [1, 2]]])
-
- def testRngNormal(self):
- shape = (2, 3)
- c = self._NewComputation()
- c.RngNormal(
- c.Constant(NumpyArrayF32(0.)),
- c.Constant(NumpyArrayF32(1.)),
- dims=shape)
- result = xla_client.execute_with_python_values(c.Build().Compile())
- # since the result is random, we just check shape and uniqueness
- self.assertLen(result, 1)
- self.assertEqual(result[0].shape, shape)
- self.assertLen(np.unique(result[0]), np.prod(shape))
-
- def testRngUniformF32(self):
- lo, hi = 2., 4.
- shape = (2, 3)
- c = self._NewComputation()
- c.RngUniform(
- c.Constant(NumpyArrayF32(lo)),
- c.Constant(NumpyArrayF32(hi)),
- dims=shape)
- result = xla_client.execute_with_python_values(c.Build().Compile())
- # since the result is random, we just check shape, uniqueness, and range
- self.assertLen(result, 1)
- self.assertEqual(result[0].shape, shape)
- self.assertLen(np.unique(result[0]), np.prod(shape))
- self.assertTrue(np.all(lo <= result[0]))
- self.assertTrue(np.all(result[0] < hi))
-
- def testRngUniformS32(self):
- lo, hi = 2, 4
- shape = (2, 3)
- c = self._NewComputation()
- c.RngUniform(
- c.Constant(NumpyArrayS32(lo)),
- c.Constant(NumpyArrayS32(hi)),
- dims=shape)
- result = xla_client.execute_with_python_values(c.Build().Compile())
- # since the result is random, we just check shape, integrality, and range
- self.assertLen(result, 1)
- self.assertEqual(result[0].shape, shape)
- self.assertEqual(result[0].dtype, np.int32)
- self.assertTrue(np.all(lo <= result[0]))
- self.assertTrue(np.all(result[0] < hi))
-
- def testCholesky(self):
- l = np.array([[4, 0, 0, 0], [6, 5, 0, 0], [2, 14, 16, 0], [3, 6, 1, 4]],
- dtype=np.float32)
- c = self._NewComputation()
- c.Cholesky(c.Constant(np.dot(l, l.T)))
- self._ExecuteAndCompareClose(c, expected=[l], rtol=1e-4)
-
- def testSort(self):
- keys = np.array([[2, 4, 1, 3], [3, 1, 4, 2]], dtype=np.float32)
- c = self._NewComputation()
- c.Sort(c.Constant(keys))
- self._ExecuteAndCompareClose(
- c, expected=[np.array([[1, 2, 3, 4], [1, 2, 3, 4]], dtype=np.float32)])
-
- def testSortKeyVal(self):
- keys = np.array([[2, 4, 1, 3], [3, 1, 4, 2]], dtype=np.float32)
- values = np.array([[0, 1, 2, 3], [4, 5, 6, 7]], dtype=np.int32)
- c = self._NewComputation()
- c.Sort((c.Constant(keys), c.Constant(values)), dimension=0)
- result = xla_client.execute_with_python_values(c.Build().Compile())
- self.assertLen(result, 2)
- np.testing.assert_allclose(result[0], [[2, 1, 1, 2], [3, 4, 4, 3]])
- np.testing.assert_equal(result[1], [[0, 5, 2, 7], [4, 1, 6, 3]])
-
- def testSortCustomComparator(self):
- b = self._NewComputation("comparator")
- p0 = b.ParameterFromNumpy(NumpyArrayF32(0))
- q0 = b.ParameterFromNumpy(NumpyArrayF32(0))
- p1 = b.ParameterFromNumpy(NumpyArrayS32(0))
- q1 = b.ParameterFromNumpy(NumpyArrayS32(0))
- b.Or(b.Lt(p0, q0), b.And(b.Eq(p0, q0), b.Gt(p1, q1)))
- comparator = b.Build()
-
- keys = np.array([[2, 3, 1, 3], [3, 1, 2, 2]], dtype=np.float32)
- values = np.array([[0, 1, 2, 3], [4, 5, 6, 7]], dtype=np.int32)
- c = self._NewComputation()
- c.Sort((c.Constant(keys), c.Constant(values)),
- dimension=1,
- comparator=comparator)
- result = xla_client.execute_with_python_values(c.Build().Compile())
- self.assertLen(result, 2)
- np.testing.assert_allclose(result[0], [[1, 2, 3, 3], [1, 2, 2, 3]])
- np.testing.assert_equal(result[1], [[2, 0, 3, 1], [5, 7, 6, 4]])
-
- def testQR(self):
- a = np.array(
- [[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166], [10, 63, 166, 310]],
- dtype=np.float32)
- c = self._NewComputation()
- c.QR(c.Constant(a), full_matrices=True)
- q, r = self._Execute(c, ())
- np.testing.assert_allclose(np.dot(q, r), a, rtol=1e-4)
-
- def testEigh(self):
- a = np.array(
- [[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166], [10, 63, 166, 310]],
- dtype=np.float32)
- a = (a + a.T) / 2
-
- c = self._NewComputation()
- c.Eigh(c.Constant(a), full_matrices=True)
- # TODO(b/129396575): Turn this test back on when it passes without fastmath.
- # v, w = self._Execute(c, ())
- # self.assertLess(np.linalg.norm(np.dot(a, v) - w * v), 1e-3)
-
- def testSVD(self):
- a = np.array(
- [[4, 6, 8, 10], [6, 45, 54, 63], [8, 54, 146, 166], [10, 63, 166, 310]],
- dtype=np.float32)
- c = self._NewComputation()
- c.SVD(c.Constant(a))
- u, d, v = self._Execute(c, ())
- self.assertLess(np.linalg.norm(a - np.matmul(u * d, v.T)), 1e-3)
-
- def testTriangularSolve(self):
- a_vals = np.array(
- [[2, 0, 0, 0], [3, 6, 0, 0], [4, 7, 9, 0], [5, 8, 10, 11]],
- dtype=np.float32)
- b_vals = np.array([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]],
- dtype=np.float32)
-
- c = self._NewComputation()
- c.TriangularSolve(
- c.Constant(a_vals),
- c.Constant(b_vals),
- left_side=False,
- lower=True,
- transpose_a=True)
- self._ExecuteAndCompareClose(
- c,
- expected=[
- np.array([
- [0.5, 0.08333334, 0.04629629, 0.03367003],
- [2.5, -0.25, -0.1388889, -0.1010101],
- [4.5, -0.58333331, -0.32407406, -0.23569024],
- ],
- dtype=np.float32)
- ],
- rtol=1e-4)
-
- def testIsConstant(self):
- c = self._NewComputation()
- a = c.ConstantS32Scalar(3)
- b = c.ConstantS32Scalar(1)
- x = c.ParameterFromNumpy(NumpyArrayS32(0))
- const_expr = c.Sub(b, a)
- non_const_expr = c.Mul(const_expr, x)
- self.assertTrue(c.IsConstant(const_expr))
- self.assertFalse(c.IsConstant(non_const_expr))
- # self.assertTrue(c.IsConstant(c.Sub(c.Add(x, a), x))) # TODO(b/77245564)
-
- def testGather(self):
- a = np.arange(9).astype(np.int32).reshape((3, 3))
- indices = np.array([[[0, 2], [2, 1]], [[1, 2], [2, 0]]], dtype=np.int32)
- dnums = xla_client.GatherDimensionNumbers()
- dnums.offset_dims.append(1)
- dnums.offset_dims.append(2)
- dnums.start_index_map.append(0)
- dnums.start_index_map.append(1)
- dnums.index_vector_dim = 2
- c = self._NewComputation()
- c.Gather(c.Constant(a), c.Constant(indices), dnums, slice_sizes=[1, 1])
- g, = self._Execute(c, ())
- expected = np.array([[[[2, 7]]], [[[5, 6]]]], dtype=np.int32)
- np.testing.assert_allclose(g, expected, rtol=1e-4)
-
- def testFft(self):
- shape = [2, 3, 4, 5]
- rng = np.random.RandomState(0)
- a = rng.randn(*shape) + 1.0j * rng.randn(*shape)
- a = a.astype(np.complex64)
- # FFT
- c = self._NewComputation()
- c.Fft(c.Constant(a), xla_client.FftType.FFT, shape[-3:])
- self._ExecuteAndCompareClose(
- c, expected=[np.fft.fftn(a, axes=(1, 2, 3))], rtol=1e-4)
- # IFFT
- c = self._NewComputation()
- c.Fft(c.Constant(a), xla_client.FftType.IFFT, shape[-3:])
- self._ExecuteAndCompareClose(
- c, expected=[np.fft.ifftn(a, axes=(1, 2, 3))], rtol=1e-4)
- # RFFT
- b = rng.randn(*shape).astype(np.float32)
- c = self._NewComputation()
- c.Fft(c.Constant(b), xla_client.FftType.RFFT, shape[-3:])
- self._ExecuteAndCompareClose(
- c, expected=[np.fft.rfftn(b, axes=(1, 2, 3))], rtol=1e-4)
- # IRFFT
- c = self._NewComputation()
- c.Fft(c.Constant(a), xla_client.FftType.IRFFT, [3, 4, 8])
- self._ExecuteAndCompareClose(
- c, expected=[np.fft.irfftn(a, axes=(1, 2, 3))], rtol=1e-4)
-
- def testNextAfter(self):
- c = self._NewComputation()
- c.NextAfter(
- c.Constant(np.array([1, 2], dtype=np.float32)),
- c.Constant(np.array([2, 1], dtype=np.float32)))
- out, = self._Execute(c, ())
- eps = np.finfo(np.float32).eps
- np.testing.assert_equal(np.array([eps + 1, 2 - eps], dtype=np.float32), out)
-
- def testRegularizedIncompleteBeta(self):
- x = np.array([0.53787335, 0.24015466, 0.47494545, 0.13567594, 0.95114538])
- a = np.array([0.00753073, 0.34813385, 0.30485708, 1.29298632, 0.51472606])
- b = np.array([0.55688389, 0.59794214, 0.42661022, 1.59748339, 0.95047677])
- c = self._NewComputation()
- c.RegularizedIncompleteBeta(c.Constant(a), c.Constant(b), c.Constant(x))
- expected = np.array(
- [0.98923271, 0.48575411, 0.57952568, 0.12579775, 0.96989155])
- self._ExecuteAndCompareClose(c, expected=[expected], rtol=1e-4)
-
-
-class EmbeddedComputationsTest(ComputationTest):
- """Tests for XLA graphs with embedded computations (such as maps)."""
-
- def _CreateConstantS32Computation(self):
- """Computation (f32) -> s32 that returns a constant 1 for any input."""
- c = self._NewComputation("constant_s32_one")
- # TODO(eliben): consider adding a nicer way to create new parameters without
- # having to create dummy Numpy arrays or populating Shape messages. Perhaps
- # we need our own (Python-client-own) way to represent Shapes conveniently.
- c.ParameterFromNumpy(NumpyArrayF32(0))
- c.ConstantS32Scalar(1)
- return c.Build()
-
- def _CreateConstantS64Computation(self):
- """Computation (f64) -> s64 that returns a constant 1 for any input."""
- c = self._NewComputation("constant_s64_one")
- # TODO(eliben): consider adding a nicer way to create new parameters without
- # having to create dummy Numpy arrays or populating Shape messages. Perhaps
- # we need our own (Python-client-own) way to represent Shapes conveniently.
- c.ParameterFromNumpy(NumpyArrayF64(0))
- c.ConstantS64Scalar(1)
- return c.Build()
-
- def _CreateConstantF32Computation(self):
- """Computation (f32) -> f32 that returns a constant 1.0 for any input."""
- c = self._NewComputation("constant_f32_one")
- c.ParameterFromNumpy(NumpyArrayF32(0))
- c.ConstantF32Scalar(1.0)
- return c.Build()
-
- def _CreateConstantF64Computation(self):
- """Computation (f64) -> f64 that returns a constant 1.0 for any input."""
- c = self._NewComputation("constant_f64_one")
- c.ParameterFromNumpy(NumpyArrayF64(0))
- c.ConstantF64Scalar(1.0)
- return c.Build()
-
- def _CreateMulF32By2Computation(self):
- """Computation (f32) -> f32 that multiplies its parameter by 2."""
- c = self._NewComputation("mul_f32_by2")
- c.Mul(c.ParameterFromNumpy(NumpyArrayF32(0)), c.ConstantF32Scalar(2.0))
- return c.Build()
-
- def _CreateMulF32ByParamComputation(self):
- """Computation (f32) -> f32 that multiplies one parameter by the other."""
- c = self._NewComputation("mul_f32_by_param")
- c.Mul(
- c.ParameterFromNumpy(NumpyArrayF32(0)),
- c.ParameterFromNumpy(NumpyArrayF32(0)))
- return c.Build()
-
- def _CreateMulF64By2Computation(self):
- """Computation (f64) -> f64 that multiplies its parameter by 2."""
- c = self._NewComputation("mul_f64_by2")
- c.Mul(c.ParameterFromNumpy(NumpyArrayF64(0)), c.ConstantF64Scalar(2.0))
- return c.Build()
-
- def _CreateBinaryAddS32Computation(self):
- """Computation (s32, s32) -> s32 that adds its two parameters."""
- c = self._NewComputation("add_param0_by_param1")
- c.Add(
- c.ParameterFromNumpy(NumpyArrayS32(0)),
- c.ParameterFromNumpy(NumpyArrayS32(0)))
- return c.Build()
-
- def _CreateBinaryAddF32Computation(self):
- """Computation (f32, f32) -> f32 that adds its two parameters."""
- c = self._NewComputation("add_param0_by_param1")
- c.Add(
- c.ParameterFromNumpy(NumpyArrayF32(0)),
- c.ParameterFromNumpy(NumpyArrayF32(0)))
- return c.Build()
-
- def _CreateBinaryAddF64Computation(self):
- """Computation (f64, f64) -> f64 that adds its two parameters."""
- c = self._NewComputation("add_param0_by_param1")
- c.Add(
- c.ParameterFromNumpy(NumpyArrayF64(0)),
- c.ParameterFromNumpy(NumpyArrayF64(0)))
- return c.Build()
-
- def _CreateBinaryDivF32Computation(self):
- """Computation (f32, f32) -> f32 that divides its two parameters."""
- c = self._NewComputation("div_param0_by_param1")
- c.Div(
- c.ParameterFromNumpy(NumpyArrayF32(0)),
- c.ParameterFromNumpy(NumpyArrayF32(0)))
- return c.Build()
-
- def _CreateBinaryDivF64Computation(self):
- """Computation (f64, f64) -> f64 that divides its two parameters."""
- c = self._NewComputation("div_param0_by_param1")
- c.Div(
- c.ParameterFromNumpy(NumpyArrayF64(0)),
- c.ParameterFromNumpy(NumpyArrayF64(0)))
- return c.Build()
-
- def _CreateTestF32Lt10Computation(self):
- """Computation (f32) -> bool that tests if its parameter is less than 10."""
- c = self._NewComputation("test_f32_lt_10")
- c.Lt(c.ParameterFromNumpy(NumpyArrayF32(0)), c.ConstantF32Scalar(10.))
- return c.Build()
-
- def _CreateTestF64Lt10Computation(self):
- """Computation (f64) -> bool that tests if its parameter is less than 10."""
- c = self._NewComputation("test_f64_lt_10")
- c.Lt(c.ParameterFromNumpy(NumpyArrayF64(0)), c.ConstantF64Scalar(10.))
- return c.Build()
-
- def _CreateBinaryGeF32Computation(self):
- """Computation (f32, f32) -> bool that tests first_param >= second_param."""
- c = self._NewComputation("param0_lt_param1")
- c.Ge(
- c.ParameterFromNumpy(NumpyArrayF32(0)),
- c.ParameterFromNumpy(NumpyArrayF32(0)))
- return c.Build()
-
- def _CreateBinaryGeF64Computation(self):
- """Computation (f64, f64) -> bool that tests first_param >= second_param."""
- c = self._NewComputation("param0_lt_param1")
- c.Ge(
- c.ParameterFromNumpy(NumpyArrayF64(0)),
- c.ParameterFromNumpy(NumpyArrayF64(0)))
- return c.Build()
-
- def _MakeSample3DArrayF32(self):
- return NumpyArrayF32([[[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]],
- [[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]]])
-
- def _MakeSample3DArrayF64(self):
- return NumpyArrayF64([[[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]],
- [[1, 2, 3], [4, 5, 6]], [[1, 2, 3], [4, 5, 6]]])
-
- def testCallF32(self):
- c = self._NewComputation()
- c.Call(
- self._CreateMulF32By2Computation(),
- operands=(c.ConstantF32Scalar(5.0),))
- self._ExecuteAndCompareClose(c, expected=[10.0])
-
- def testCallF64(self):
- c = self._NewComputation()
- c.Call(
- self._CreateMulF64By2Computation(),
- operands=(c.ConstantF64Scalar(5.0),))
- self._ExecuteAndCompareClose(c, expected=[10.0])
-
- def testMapEachElementToS32Constant(self):
- c = self._NewComputation()
- c.Map([c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0]))],
- self._CreateConstantS32Computation(), [0])
- self._ExecuteAndCompareExact(c, expected=[[1, 1, 1, 1]])
-
- def testMapEachElementToS64Constant(self):
- c = self._NewComputation()
- c.Map([c.Constant(NumpyArrayF64([1.0, 2.0, 3.0, 4.0]))],
- self._CreateConstantS64Computation(), [0])
- self._ExecuteAndCompareExact(c, expected=[[1, 1, 1, 1]])
-
- def testMapMulBy2F32(self):
- c = self._NewComputation()
- c.Map([c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0]))],
- self._CreateMulF32By2Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[2.0, 4.0, 6.0, 8.0]])
-
- def testMapMulBy2F64(self):
- c = self._NewComputation()
- c.Map([c.Constant(NumpyArrayF64([1.0, 2.0, 3.0, 4.0]))],
- self._CreateMulF64By2Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[2.0, 4.0, 6.0, 8.0]])
-
- def testSimpleMapChainF32(self):
- # Chains a map of constant-f32 with a map of mul-by-2
- c = self._NewComputation()
- const_f32 = c.Map([c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0]))],
- self._CreateConstantF32Computation(), [0])
- c.Map([const_f32], self._CreateMulF32By2Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[2.0, 2.0, 2.0, 2.0]])
-
- def testSimpleMapChainF64(self):
- # Chains a map of constant-f64 with a map of mul-by-2
- c = self._NewComputation()
- const_f64 = c.Map([c.Constant(NumpyArrayF64([1.0, 2.0, 3.0, 4.0]))],
- self._CreateConstantF64Computation(), [0])
- c.Map([const_f64], self._CreateMulF64By2Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[2.0, 2.0, 2.0, 2.0]])
-
- def testDivVectorsWithMapF32(self):
- c = self._NewComputation()
- c.Map((c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0])),
- c.Constant(NumpyArrayF32([5.0, 5.0, 4.0, 4.0]))),
- self._CreateBinaryDivF32Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[0.2, 0.4, 0.75, 1.0]])
-
- def testDivVectorsWithMapF64(self):
- c = self._NewComputation()
- c.Map((c.Constant(NumpyArrayF64([1.0, 2.0, 3.0, 4.0])),
- c.Constant(NumpyArrayF64([5.0, 5.0, 4.0, 4.0]))),
- self._CreateBinaryDivF64Computation(), [0])
- self._ExecuteAndCompareClose(c, expected=[[0.2, 0.4, 0.75, 1.0]])
-
- def testSelectAndScatterF32(self):
- c = self._NewComputation()
- c.SelectAndScatter(
- c.Constant(NumpyArrayF32([[1., 2., 6.], [4., 5., 3.]])),
- select=self._CreateBinaryGeF32Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 2),
- padding=xla_client.PaddingType.VALID,
- source=c.Constant(NumpyArrayF32([[0.1, 0.2]])),
- init_value=c.Constant(NumpyArrayF32(1)),
- scatter=self._CreateBinaryAddF32Computation())
- self._ExecuteAndCompareClose(c, expected=[[[1., 1., 1.2], [1.1, 1., 1.]]])
-
- def testSelectAndScatterF64(self):
- c = self._NewComputation()
- c.SelectAndScatter(
- c.Constant(NumpyArrayF64([[1., 2., 6.], [4., 5., 3.]])),
- select=self._CreateBinaryGeF64Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 2),
- padding=xla_client.PaddingType.VALID,
- source=c.Constant(NumpyArrayF64([[0.1, 0.2]])),
- init_value=c.Constant(NumpyArrayF64(1)),
- scatter=self._CreateBinaryAddF64Computation())
- self._ExecuteAndCompareClose(c, expected=[[[1., 1., 1.2], [1.1, 1., 1.]]])
-
- def testReduce1DtoScalarF32(self):
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0])),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- dimensions=[0])
- self._ExecuteAndCompareClose(c, expected=[10])
-
- def testReduce1DtoScalarF64(self):
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(NumpyArrayF64([1.0, 2.0, 3.0, 4.0])),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- dimensions=[0])
- self._ExecuteAndCompareClose(c, expected=[10])
-
- def testReduce2DTo1DDim0F32(self):
- input_array = NumpyArrayF32([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- dimensions=[0])
- self._ExecuteAndCompareClose(c, expected=[[5, 7, 9]])
-
- def testReduce2DTo1DDim0F64(self):
- input_array = NumpyArrayF64([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- dimensions=[0])
- self._ExecuteAndCompareClose(c, expected=[[5, 7, 9]])
-
- def testReduce2DTo1DDim1F32(self):
- input_array = NumpyArrayF32([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- dimensions=[1])
- self._ExecuteAndCompareClose(c, expected=[[6, 15]])
-
- def testReduce2DTo1DDim1F64(self):
- input_array = NumpyArrayF64([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- dimensions=[1])
- self._ExecuteAndCompareClose(c, expected=[[6, 15]])
-
- def testReduce3DAllPossibleWaysF32(self):
- input_array = self._MakeSample3DArrayF32()
-
- def _ReduceAndTest(*dims):
+ def testCompileWithWrongElementTypeInLayout(self):
c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- dimensions=dims)
- self._ExecuteAndCompareClose(
- c, expected=[np.sum(input_array, axis=tuple(dims))])
+ c.SetOpMetadata(xla_client.CurrentSourceInfoMetadata())
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(self.s32_scalar_2))
+ c.ClearOpMetadata()
- _ReduceAndTest(0)
- _ReduceAndTest(0, 1)
- _ReduceAndTest(0, 2)
- _ReduceAndTest(1, 2)
- _ReduceAndTest(0, 1, 2)
+ options = xla_client.CompileOptions()
+ options.argument_layouts = [
+ xla_client.Shape.array_shape(np.dtype(np.float32), [])
+ ]
- def testReduce3DAllPossibleWaysF64(self):
- input_array = self._MakeSample3DArrayF64()
+ def TestFun():
+ return self.backend.compile(c.Build(), compile_options=options)
- def _ReduceAndTest(*dims):
+ self.assertRaisesRegex(
+ RuntimeError, r".*Invalid argument shape.*"
+ r"expected s32\[\], got f32\[\].*", TestFun)
+
+ def testInvokeWithWrongElementType(self):
c = self._NewComputation()
- c.Reduce(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- dimensions=dims)
- self._ExecuteAndCompareClose(
- c, expected=[np.sum(input_array, axis=tuple(dims))])
+ c.SetOpMetadata(xla_client.CurrentSourceInfoMetadata())
+ ops.Parameter(c, 0, xla_client.shape_from_pyval(self.s32_scalar_2))
+ c.ClearOpMetadata()
- _ReduceAndTest(0)
- _ReduceAndTest(0)
- _ReduceAndTest(0, 1)
- _ReduceAndTest(0, 2)
- _ReduceAndTest(1, 2)
- _ReduceAndTest(0, 1, 2)
+ def TestFun():
+ return xla_client.execute_with_python_values(
+ self.backend.compile(c.Build()), [self.f32_scalar_2])
- def testReduceWindowValidUnitStridesF32(self):
- input_array = NumpyArrayF32([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 1),
- padding=xla_client.PaddingType.VALID)
- self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.]]])
+ self.assertRaisesRegex(
+ RuntimeError, r"Invalid argument: Argument does not match.*"
+ r"want s32\[\], got f32\[\].*", TestFun)
- def testReduceWindowSameUnitStridesF32(self):
- input_array = NumpyArrayF32([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 1),
- padding=xla_client.PaddingType.SAME)
- self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.], [4., 5., 6.]]])
+ tests.append(EmbeddedComputationsTest)
- def testReduceWindowValidGeneralStridesF32(self):
- input_array = NumpyArrayF32([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF32Scalar(0),
- computation_to_apply=self._CreateBinaryAddF32Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 2),
- padding=xla_client.PaddingType.VALID)
- self._ExecuteAndCompareClose(c, expected=[[[5., 9.]]])
+ class ComputationRootTest(ComputationTest):
+ """Tests related to setting the root of the computation."""
- def testReduceWindowValidUnitStridesF64(self):
- input_array = NumpyArrayF64([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 1),
- padding=xla_client.PaddingType.VALID)
- self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.]]])
+ def testComputationRootDifferentFromLastOp(self):
+ c = self._NewComputation()
+ x = ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(2.0)))
+ result = ops.Add(x, ops.Constant(c, np.float32(3.14)))
+ ops.Add(result, ops.Constant(c, np.float32(1.618)))
- def testReduceWindowSameUnitStridesF64(self):
- input_array = NumpyArrayF64([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 1),
- padding=xla_client.PaddingType.SAME)
- self._ExecuteAndCompareClose(c, expected=[[[5., 7., 9.], [4., 5., 6.]]])
+ arg = NumpyArrayF32(1.0)
+ compiled_c = self.backend.compile(c.Build(result))
+ ans, = xla_client.execute_with_python_values(
+ compiled_c, [arg], backend=self.backend)
+ np.testing.assert_allclose(ans, 4.14)
- def testReduceWindowValidGeneralStridesF64(self):
- input_array = NumpyArrayF64([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
- c = self._NewComputation()
- c.ReduceWindow(
- operand=c.Constant(input_array),
- init_value=c.ConstantF64Scalar(0),
- computation_to_apply=self._CreateBinaryAddF64Computation(),
- window_dimensions=(2, 1),
- window_strides=(1, 2),
- padding=xla_client.PaddingType.VALID)
- self._ExecuteAndCompareClose(c, expected=[[[5., 9.]]])
+ tests.append(ComputationRootTest)
- def testWhileF32(self):
- cond = self._CreateTestF32Lt10Computation()
- body = self._CreateMulF32By2Computation()
- c = self._NewComputation()
- init = c.ConstantF32Scalar(1.)
- c.While(cond, body, init)
- self._ExecuteAndCompareClose(c, expected=[16.])
+ class SetShardingTest(ComputationTest):
+ """Tests related to set OpSharding."""
- def testWhileF64(self):
- cond = self._CreateTestF64Lt10Computation()
- body = self._CreateMulF64By2Computation()
- c = self._NewComputation()
- init = c.ConstantF64Scalar(1.)
- c.While(cond, body, init)
- self._ExecuteAndCompareClose(c, expected=[16.])
+ def testSetSharding(self):
+ c = self._NewComputation()
+ sharding = xla_client.OpSharding()
+ sharding.type = sharding.type.REPLICATED
+ sharding.tile_assignment_dimensions.extend([1])
+ sharding.tile_assignment_devices.extend([0])
+ # Set Sharding.
+ c.SetSharding(sharding)
+ x = ops.Parameter(c, 0, xla_client.shape_from_pyval(NumpyArrayF32(2.0)))
+ # Clear Sharding.
+ c.ClearSharding()
- def testConditionalTrue(self):
- c = self._NewComputation()
- pred = c.ConstantPredScalar(True)
- true_operand = c.ConstantF32Scalar(3.)
- true_computation = self._CreateMulF32By2Computation()
- false_operand = c.ConstantF32Scalar(2.)
- false_computation = self._CreateConstantF32Computation()
- c.Conditional(pred, true_operand, true_computation, false_operand,
- false_computation)
- self._ExecuteAndCompareClose(c, expected=[6.])
+ result = ops.Add(x, ops.Constant(c, np.float32(3.14)))
+ ops.Add(result, ops.Constant(c, np.float32(1.618)))
+ arg = NumpyArrayF32(1.0)
+ compiled_c = self.backend.compile(c.Build(result))
+ ans, = xla_client.execute_with_python_values(
+ compiled_c, [arg], backend=self.backend)
+ np.testing.assert_allclose(ans, 4.14)
- def testConditionalFalse(self):
- c = self._NewComputation()
- pred = c.ConstantPredScalar(False)
- true_operand = c.ConstantF32Scalar(3.)
- true_computation = self._CreateMulF32By2Computation()
- false_operand = c.ConstantF32Scalar(2.)
- false_computation = self._CreateConstantF32Computation()
- c.Conditional(pred, true_operand, true_computation, false_operand,
- false_computation)
- self._ExecuteAndCompareClose(c, expected=[1.])
+ tests.append(SetShardingTest)
- def testInfeedS32Values(self):
- to_infeed = NumpyArrayS32([1, 2, 3, 4])
- c = self._NewComputation()
- c.GetTupleElement(c.Infeed(xla_client.shape_from_pyval(to_infeed[0])), 0)
- compiled_c = c.Build().Compile()
- for item in to_infeed:
- xla_client.transfer_to_infeed(item)
+ class AliasTest(ComputationTest):
- for item in to_infeed:
- result, = xla_client.execute_with_python_values(compiled_c)
- self.assertEqual(result, item)
+ def testSetUpAlias(self):
+ c = self._NewComputation()
+ p1 = ops.Parameter(
+ c, 0,
+ xla_client.shape_from_pyval(
+ NumpyArrayF32(1.0)).with_major_to_minor_layout_if_absent())
+ p2 = ops.Parameter(
+ c, 1,
+ xla_client.shape_from_pyval(
+ NumpyArrayF32(1.0)).with_major_to_minor_layout_if_absent())
+ out = ops.Add(p1, p2)
+ c.SetUpAlias([], 0, [])
+ c = c.Build(out)
+ if self.backend.platform != "tpu":
+ with self.assertRaisesRegex(
+ RuntimeError, "Buffer aliasing is not supported "
+ "by XLA for non-TPU backends"):
+ self.backend.compile(c)
- def testInfeedTuple(self):
- to_infeed = (NumpyArrayS32([1, 2, 3, 4]), NumpyArrayS32([[7], [8]]))
- c = self._NewComputation()
- c.GetTupleElement(c.Infeed(xla_client.shape_from_pyval(to_infeed)), 0)
- compiled_c = c.Build().Compile()
- xla_client.transfer_to_infeed(to_infeed)
+ tests.append(AliasTest)
- result = xla_client.execute_with_python_values(compiled_c)
- self.assertLen(result, 2)
- np.testing.assert_equal(result[0], to_infeed[0])
- np.testing.assert_equal(result[1], to_infeed[1])
+ testcase_shapes = [
+ (),
+ (1,),
+ (2, 3),
+ (2, 0),
+ (0, 7),
+ (4, 1, 2),
+ (2, 1, 3),
+ (2, 4, 1),
+ (3, 1),
+ (1, 3),
+ ]
- def testInfeedThenOutfeedS32(self):
- to_round_trip = NumpyArrayS32([1, 2, 3, 4])
- c = self._NewComputation()
- x_and_token = c.Infeed(xla_client.shape_from_pyval(to_round_trip[0]))
- x = c.GetTupleElement(x_and_token, 0)
- token = c.GetTupleElement(x_and_token, 1)
- c.Outfeed(x, token)
+ def FormatShapeAndDtype(shape, dtype):
+ return "_{}[{}]".format(np.dtype(dtype).name, ",".join(map(str, shape)))
- compiled_c = c.Build().Compile()
+ class DLPackTest(parameterized.TestCase):
- for want in to_round_trip:
- execution = threading.Thread(target=lambda: compiled_c.Execute([]))
- execution.start()
- xla_client.transfer_to_infeed(want)
- got = xla_client.transfer_from_outfeed(
- xla_client.shape_from_pyval(to_round_trip[0]))
- execution.join()
- self.assertEqual(want, got)
+ def setUp(self):
+ super(DLPackTest, self).setUp()
+ self.backend = xla_backend()
+ if self.backend.platform not in ("cpu", "gpu"):
+ self.skipTest("DLPack requires CPU or GPU")
- def testScatter(self):
- a = np.arange(9).astype(np.int32).reshape((3, 3))
- scatter_indices = np.array([0, 2], dtype=np.int32)
- updates = np.array([[10, 20, 30], [70, 80, 90]], dtype=np.int32)
+ # pylint: disable=g-complex-comprehension
+ @parameterized.named_parameters({
+ "testcase_name": FormatShapeAndDtype(shape, dtype),
+ "dtype": dtype,
+ "shape": shape
+ } for dtype in dlpack_dtypes for shape in testcase_shapes)
+ def testRoundTrip(self, dtype, shape):
+ x = np.array(np.random.rand(*shape) * 100, dtype=dtype)
+ buffer = xla_client.Buffer.from_pyval(x, backend=self.backend)
+ dlt = xla_client._xla.BufferToDLPackManagedTensor(buffer)
+ del buffer # Free "buffer" to make sure dlt retains ownership.
+ self.assertEqual(type(dlt).__name__, "PyCapsule")
+ y = xla_client._xla.DLPackManagedTensorToBuffer(dlt, self.backend.client)
+ np.testing.assert_array_equal(x, y.to_py())
- dnums = xla_client.ScatterDimensionNumbers()
- dnums.update_window_dims.append(1)
- dnums.inserted_window_dims.append(0)
- dnums.scatter_dims_to_operand_dims.append(0)
- dnums.index_vector_dim = 1
+ def testTensorsCanBeConsumedOnceOnly(self):
+ x = np.array(np.random.rand(3, 4, 5, 6), dtype=np.float32)
+ buffer = xla_client.Buffer.from_pyval(x, backend=self.backend)
+ dlt = xla_client._xla.BufferToDLPackManagedTensor(buffer)
- c = self._NewComputation()
- c.Scatter(
- c.Constant(a), c.Constant(scatter_indices), c.Constant(updates),
- self._CreateBinaryAddS32Computation(), dnums)
- expected = np.array([[10, 21, 32], [3, 4, 5], [76, 87, 98]], dtype=np.int32)
- self._ExecuteAndCompareClose(c, expected=[expected])
+ def ConsumeDLPackTensor():
+ _ = xla_client._xla.DLPackManagedTensorToBuffer(dlt,
+ self.backend.client)
+
+ ConsumeDLPackTensor()
+ self.assertRaisesRegex(
+ RuntimeError, ".*a DLPack tensor may be consumed at most once.*",
+ ConsumeDLPackTensor)
+
+ tests.append(DLPackTest)
+
+ class BufferProtocolTest(parameterized.TestCase):
+
+ def setUp(self):
+ super(BufferProtocolTest, self).setUp()
+ self.backend = xla_backend()
+ if self.backend.platform != "cpu":
+ self.skipTest("Test requires CPU")
+
+ # pylint: disable=g-complex-comprehension
+ @parameterized.named_parameters({
+ "testcase_name": FormatShapeAndDtype(shape, dtype),
+ "dtype": dtype,
+ "shape": shape
+ } for dtype in standard_dtypes if dtype != bfloat16
+ for shape in testcase_shapes)
+ def testRoundTrip(self, dtype, shape):
+ x = np.array(np.random.rand(*shape) * 100, dtype=dtype)
+ x_ptr = x.__array_interface__["data"][0]
+ buffer = xla_client.Buffer.from_pyval(x, backend=self.backend)
+ y = np.array(buffer, copy=False)
+ y_ptr = y.__array_interface__["data"][0]
+ np.testing.assert_array_equal(x, y)
+ # If the input was sufficiently aligned, the input and output should
+ # alias.
+ self.assertTrue((x_ptr & 63) != 0 or x_ptr == y_ptr)
+ self.assertEqual(y_ptr, buffer.unsafe_buffer_pointer())
+
+ buffer2 = xla_client.Buffer.from_pyval(
+ x, backend=self.backend, force_copy=True)
+ z = np.array(buffer2, copy=False)
+ self.assertNotEqual(x.__array_interface__["data"][0],
+ z.__array_interface__["data"][0])
+
+ def testDeleteWithActiveView(self):
+ x = np.random.randn(20, 10)
+ buffer = xla_client.Buffer.from_pyval(x, backend=self.backend)
+ buffer_ptr = buffer.unsafe_buffer_pointer()
+ y = np.array(buffer, copy=False)
+ buffer.delete()
+ # It is still legal to access `y`; the array view must keep it alive.
+ np.testing.assert_array_equal(x, y)
+ self.assertEqual(y.__array_interface__["data"][0], buffer_ptr)
+
+ tests.append(BufferProtocolTest)
+
+ class ProfilerTest(absltest.TestCase):
+
+ def testTraceMe(self):
+ # TODO(phawkins): These tests just check that the TraceMe context manager
+ # acts like a context manager and doesn't explode. Ideally we'd check that
+ # the profiler saw the traceme too.
+ with xla_client.profiler.TraceMe("test1"):
+ pass
+ with xla_client.profiler.TraceMe("test2", foo=123):
+ pass
+ with self.assertRaises(ValueError):
+ with xla_client.profiler.TraceMe("test3"):
+ raise ValueError("test")
+
+ @unittest.skipIf(portpicker is None, "Test requires portpicker")
+ def testStartServer(self):
+ port = portpicker.pick_unused_port()
+ server = xla_client.profiler.start_server(port)
+ del server
+
+ tests.append(ProfilerTest)
+ return tests
-class ErrorTest(ComputationTest):
-
- def setUp(self):
- self.f32_scalar_2 = NumpyArrayF32(2.0)
- self.s32_scalar_2 = NumpyArrayS32(2)
-
- def testCompileWithWrongElementTypeInLayout(self):
- c = self._NewComputation()
- c.SetOpMetadata(xla_client.CurrentSourceInfoMetadata())
- c.ParameterFromNumpy(self.s32_scalar_2)
- c.ClearOpMetadata()
-
- options = xla_client.CompileOptions()
- options.argument_layouts = [
- xla_client.Shape.array_shape(np.dtype(np.float32), [])
- ]
-
- def TestFun():
- return c.Build().Compile(compile_options=options)
-
- self.assertRaisesRegex(
- RuntimeError, r".*Invalid argument shape.*"
- r"expected s32\[\], got f32\[\].*", TestFun)
-
- def testInvokeWithWrongElementType(self):
- c = self._NewComputation()
- c.SetOpMetadata(xla_client.CurrentSourceInfoMetadata())
- c.ParameterFromNumpy(self.s32_scalar_2)
- c.ClearOpMetadata()
-
- def TestFun():
- return xla_client.execute_with_python_values(c.Build().Compile(),
- [self.f32_scalar_2])
-
- self.assertRaisesRegex(
- RuntimeError, r"Invalid argument: Argument does not match.*"
- r"want s32\[\], got f32\[\].*", TestFun)
-
-
-class ComputationRootTest(ComputationTest):
- """Tests related to setting the root of the computation."""
-
- def testComputationRootDifferentFromLastOp(self):
- c = self._NewComputation()
- x = c.ParameterFromNumpy(NumpyArrayF32(2.0))
- result = c.Add(x, c.ConstantF32Scalar(3.14))
- extra = c.Add(result, c.ConstantF32Scalar(1.618)) # pylint: disable=unused-variable
-
- arg = NumpyArrayF32(1.0)
- compiled_c = c.Build(result).Compile()
- ans, = xla_client.execute_with_python_values(compiled_c, [arg])
- np.testing.assert_allclose(ans, 4.14)
-
-
-class SetShardingTest(ComputationTest):
- """Tests related to set OpSharding."""
-
- def testSetSharding(self):
- c = self._NewComputation()
- sharding = xla_client.OpSharding()
- sharding.type = sharding.type.REPLICATED
- sharding.tile_assignment_dimensions.extend([1])
- sharding.tile_assignment_devices.extend([0])
- # Set Sharding.
- c.SetSharding(sharding)
- x = c.ParameterFromNumpy(NumpyArrayF32(2.0))
- # Clear Sharding.
- c.ClearSharding()
-
- result = c.Add(x, c.ConstantF32Scalar(3.14))
- extra = c.Add(result, c.ConstantF32Scalar(1.618)) # pylint: disable=unused-variable
- arg = NumpyArrayF32(1.0)
- compiled_c = c.Build(result).Compile()
- ans, = xla_client.execute_with_python_values(compiled_c, [arg])
- np.testing.assert_allclose(ans, 4.14)
-
-
-class AliasTest(ComputationTest):
-
- def testSetUpAlias(self):
- c = self._NewComputation()
- p1 = c.ParameterFromNumpy(NumpyArrayF32(1.0))
- p2 = c.ParameterFromNumpy(NumpyArrayF32(1.0))
- out = c.Add(p1, p2)
- c.SetUpAlias([], 0, [])
- c = c.Build(out)
- with self.assertRaisesRegex(RuntimeError,
- "Buffer aliasing is not supported "
- "by XLA for non-TPU backends"):
- c.Compile()
-
-
-int_dtypes = [
- np.int8, np.int16, np.int32, np.int64, np.uint8, np.uint16, np.uint32,
- np.uint64
-]
-float_dtypes = [np.float16, np.float32, np.float64]
-complex_dtypes = [np.complex64, np.complex128]
-dlpack_dtypes = int_dtypes + float_dtypes + [bfloat16]
-standard_dtypes = int_dtypes + float_dtypes + complex_dtypes + [np.bool_]
-
-testcase_shapes = [
- (),
- (1,),
- (2, 3),
- (2, 0),
- (0, 7),
- (4, 1, 2),
- (2, 1, 3),
- (2, 4, 1),
- (3, 1),
- (1, 3),
-]
-
-
-def FormatShapeAndDtype(shape, dtype):
- return "_{}[{}]".format(np.dtype(dtype).name, ",".join(map(str, shape)))
-
-
-class DLPackTest(parameterized.TestCase):
-
- # pylint: disable=g-complex-comprehension
- @parameterized.named_parameters({
- "testcase_name": FormatShapeAndDtype(shape, dtype),
- "dtype": dtype,
- "shape": shape
- } for dtype in dlpack_dtypes for shape in testcase_shapes)
- def testRoundTrip(self, dtype, shape):
- x = np.array(np.random.rand(*shape) * 100, dtype=dtype)
- backend = xla_client.get_local_backend()
- buffer = xla_client.Buffer.from_pyval(x, backend=backend)
- dlt = xla_client._xla.BufferToDLPackManagedTensor(buffer)
- del buffer # Free "buffer" to make sure dlt retains ownership.
- self.assertEqual(type(dlt).__name__, "PyCapsule")
- y = xla_client._xla.DLPackManagedTensorToBuffer(dlt, backend.client)
- np.testing.assert_array_equal(x, y.to_py())
-
- def testTensorsCanBeConsumedOnceOnly(self):
- x = np.array(np.random.rand(3, 4, 5, 6), dtype=np.float32)
- backend = xla_client.get_local_backend()
- buffer = xla_client.Buffer.from_pyval(x, backend=backend)
- dlt = xla_client._xla.BufferToDLPackManagedTensor(buffer)
-
- def ConsumeDLPackTensor():
- _ = xla_client._xla.DLPackManagedTensorToBuffer(dlt, backend.client)
-
- ConsumeDLPackTensor()
- self.assertRaisesRegex(RuntimeError,
- ".*a DLPack tensor may be consumed at most once.*",
- ConsumeDLPackTensor)
-
-
-class BufferProtocolTest(parameterized.TestCase):
-
- # pylint: disable=g-complex-comprehension
- @parameterized.named_parameters({
- "testcase_name": FormatShapeAndDtype(shape, dtype),
- "dtype": dtype,
- "shape": shape
- } for dtype in standard_dtypes for shape in testcase_shapes)
- def testRoundTrip(self, dtype, shape):
- x = np.array(np.random.rand(*shape) * 100, dtype=dtype)
- x_ptr = x.__array_interface__["data"][0]
- backend = xla_client.get_local_backend("cpu")
- buffer = xla_client.Buffer.from_pyval(x, backend=backend)
- y = np.array(buffer, copy=False)
- y_ptr = y.__array_interface__["data"][0]
- np.testing.assert_array_equal(x, y)
- # If the input was sufficiently aligned, the input and output should alias.
- self.assertTrue((x_ptr & 63) != 0 or x_ptr == y_ptr)
- self.assertEqual(y_ptr, buffer.unsafe_buffer_pointer())
-
- buffer2 = xla_client.Buffer.from_pyval(x, backend=backend, force_copy=True)
- z = np.array(buffer2, copy=False)
- self.assertNotEqual(x.__array_interface__["data"][0],
- z.__array_interface__["data"][0])
-
- def testDeleteWithActiveView(self):
- x = np.random.randn(20, 10)
- backend = xla_client.get_local_backend("cpu")
- buffer = xla_client.Buffer.from_pyval(x, backend=backend)
- buffer_ptr = buffer.unsafe_buffer_pointer()
- y = np.array(buffer, copy=False)
- buffer.delete()
- # It is still legal to access `y`; the array view must keep it alive.
- np.testing.assert_array_equal(x, y)
- self.assertEqual(y.__array_interface__["data"][0], buffer_ptr)
-
-
-class ProfilerTest(absltest.TestCase):
-
- def testTraceMe(self):
- # TODO(phawkins): These tests just check that the TraceMe context manager
- # acts like a context manager and doesn't explode. Ideally we'd check that
- # the profiler saw the traceme too.
- with xla_client.profiler.TraceMe("test1"):
- pass
- with xla_client.profiler.TraceMe("test2", foo=123):
- pass
- with self.assertRaises(ValueError):
- with xla_client.profiler.TraceMe("test3"):
- raise ValueError("test")
-
- @unittest.skipIf(portpicker is None, "Test requires portpicker")
- def testStartServer(self):
- port = portpicker.pick_unused_port()
- server = xla_client.profiler.start_server(port)
- del server
+def InstantiateTests(globals_dict, backend, test_prefix="", **kw):
+ for klass in TestFactory(backend, **kw):
+ test = type(test_prefix + klass.__name__, (klass,), {})
+ # Clean up the qualified names of the tests to not include the test factory.
+ test.__qualname__ = test.__name__
+ globals_dict[test.__name__] = test
if __name__ == "__main__":
+ flags.DEFINE_string("backend", "cpu", "Target backend.")
+ InstantiateTests(globals(),
+ lambda: xla_client.get_local_backend(FLAGS.backend))
absltest.main()
diff --git a/tensorflow/compiler/xla/service/BUILD b/tensorflow/compiler/xla/service/BUILD
index 484e967..aef215e 100644
--- a/tensorflow/compiler/xla/service/BUILD
+++ b/tensorflow/compiler/xla/service/BUILD
@@ -977,7 +977,7 @@
"//tensorflow/compiler/xla/service/gpu:gpu_transfer_manager",
"//tensorflow/core:stream_executor_no_cuda",
] + if_cuda_is_configured([
- "//tensorflow/compiler/xla/service/mlir_gpu:mlir_compiler",
+ "//tensorflow/compiler/xla/service/mlir_gpu:mlir_compiler_impl",
]),
)
@@ -1078,6 +1078,7 @@
srcs = ["compiler.cc"],
hdrs = ["compiler.h"],
deps = [
+ ":buffer_assignment",
":buffer_value",
":computation_placer",
":executable",
@@ -4118,6 +4119,28 @@
)
cc_library(
+ name = "root_instruction_sinker",
+ srcs = ["root_instruction_sinker.cc"],
+ hdrs = ["root_instruction_sinker.h"],
+ deps = [
+ ":hlo",
+ ":hlo_pass",
+ ":tuple_util",
+ ],
+)
+
+tf_cc_test(
+ name = "root_instruction_sinker_test",
+ srcs = ["root_instruction_sinker_test.cc"],
+ deps = [
+ ":hlo_matchers",
+ ":root_instruction_sinker",
+ "//tensorflow/compiler/xla/tests:hlo_test_base",
+ "//tensorflow/compiler/xla/tests:xla_internal_test_main",
+ ],
+)
+
+cc_library(
name = "while_util",
srcs = ["while_util.cc"],
hdrs = ["while_util.h"],
diff --git a/tensorflow/compiler/xla/service/algebraic_simplifier.cc b/tensorflow/compiler/xla/service/algebraic_simplifier.cc
index fcf5c9c..1fbb486 100644
--- a/tensorflow/compiler/xla/service/algebraic_simplifier.cc
+++ b/tensorflow/compiler/xla/service/algebraic_simplifier.cc
@@ -2969,26 +2969,6 @@
MakeScalarLike(lhs, 1), lhs));
}
- VLOG(10) << "trying transform [pow(pow(A, X), Y) => pow(A, X*Y)]: "
- << power->ToString();
-
- // Don't perform this optimization if either of the exponents is complex; this
- // identity is true only for real-valued exponents. In addition, we cowardly
- // refuse to do this transformation if the two exponents have different
- // element types.
- if (lhs->opcode() == HloOpcode::kPower &&
- !ShapeUtil::ElementIsComplex(lhs->operand(1)->shape()) &&
- !ShapeUtil::ElementIsComplex(rhs->shape()) &&
- ShapeUtil::SameElementType(lhs->operand(1)->shape(), rhs->shape())) {
- auto exponent_product =
- computation_->AddInstruction(HloInstruction::CreateBinary(
- rhs->shape(), HloOpcode::kMultiply, lhs->mutable_operand(1), rhs));
- return ReplaceWithNewInstruction(
- power, HloInstruction::CreateBinary(power->shape(), HloOpcode::kPower,
- lhs->mutable_operand(0),
- exponent_product));
- }
-
return Status::OK();
}
@@ -3714,7 +3694,7 @@
auto bcast_width = ShapeUtil::GetDimension(updated_shape, dim);
padding_config_dim->set_edge_padding_low(beg);
padding_config_dim->set_edge_padding_high(
- std::max(bcast_width - (beg + update_width), 0LL));
+ std::max(bcast_width - (beg + update_width), int64{0}));
// dynamic_update_slice does not specify a stride
padding_config_dim->set_interior_padding(0);
}
diff --git a/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc b/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc
index dde1bcb..5604146 100644
--- a/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc
+++ b/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc
@@ -1011,13 +1011,8 @@
builder.AddInstruction(HloInstruction::CreateBinary(r1f32, HloOpcode::kPower,
inner_power, exp2));
- auto computation = m->AddEntryComputation(builder.Build());
AlgebraicSimplifier simplifier(default_options_);
- ASSERT_TRUE(simplifier.Run(m.get()).ValueOrDie());
- EXPECT_THAT(
- computation->root_instruction(),
- GmockMatch(m::Power(m::Op().Is(base),
- m::Multiply(m::Op().Is(exp1), m::Op().Is(exp2)))));
+ ASSERT_FALSE(simplifier.Run(m.get()).ValueOrDie());
}
// Don't simplify pow(pow(A, X), Y) => pow(A, X*Y) if X and Y are complex
diff --git a/tensorflow/compiler/xla/service/compiler.cc b/tensorflow/compiler/xla/service/compiler.cc
index 653f455..f03b27c 100644
--- a/tensorflow/compiler/xla/service/compiler.cc
+++ b/tensorflow/compiler/xla/service/compiler.cc
@@ -28,6 +28,14 @@
/* static */ tensorflow::mutex Compiler::platform_compiler_mutex_(
tensorflow::LINKER_INITIALIZED);
+StatusOr<
+ std::tuple<std::unique_ptr<HloModule>, std::unique_ptr<BufferAssignment>>>
+Compiler::RunHloPassesAndBufferAssignement(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* executor,
+ se::DeviceMemoryAllocator* device_allocator) {
+ return Unimplemented("This compiler does not support this method");
+}
+
std::vector<std::unique_ptr<tensorflow::protobuf::Message>>
Compiler::ComputeBackendConfigs(const HloInstruction& hlo,
se::StreamExecutor* executor) const {
diff --git a/tensorflow/compiler/xla/service/compiler.h b/tensorflow/compiler/xla/service/compiler.h
index b2e1231..cf64615 100644
--- a/tensorflow/compiler/xla/service/compiler.h
+++ b/tensorflow/compiler/xla/service/compiler.h
@@ -27,6 +27,7 @@
#include <vector>
#include "absl/types/span.h"
+#include "tensorflow/compiler/xla/service/buffer_assignment.h"
#include "tensorflow/compiler/xla/service/buffer_value.h"
#include "tensorflow/compiler/xla/service/computation_placer.h"
#include "tensorflow/compiler/xla/service/executable.h"
@@ -172,6 +173,21 @@
std::unique_ptr<HloModule> module, se::StreamExecutor* executor,
se::DeviceMemoryAllocator* device_allocator) = 0;
+ // Runs HLO passes to optimize the given HloModule, perform scheduling and
+ // buffer assignment, returns the optimized module and the buffer assignments.
+ // This interface is intentionally narrow.
+ //
+ // If device_allocator is not null, the compiler may use it to allocate temp
+ // space on the device for use during compilation. For example, the compiler
+ // may allocate buffers on the device and then run variants of a given
+ // algorithm over those buffers, to see which variant is fastest. Any space
+ // allocated should be deallocated before this function returns.
+ virtual StatusOr<
+ std::tuple<std::unique_ptr<HloModule>, std::unique_ptr<BufferAssignment>>>
+ RunHloPassesAndBufferAssignement(std::unique_ptr<HloModule> module,
+ se::StreamExecutor* executor,
+ se::DeviceMemoryAllocator* device_allocator);
+
// Compiles the HLO module for execution on a device given by the executor,
// and returns an executable object or an error status. No HLO passes are
// applied to module. Generally a module should be passed through RunHloPasses
diff --git a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc
index 53d0d14..8f56aa8 100644
--- a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc
+++ b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc
@@ -521,6 +521,33 @@
return std::move(module);
}
+StatusOr<
+ std::tuple<std::unique_ptr<HloModule>, std::unique_ptr<BufferAssignment>>>
+CpuCompiler::RunHloPassesAndBufferAssignement(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* executor,
+ se::DeviceMemoryAllocator* device_allocator) {
+ TF_ASSIGN_OR_RETURN(
+ module, RunHloPasses(std::move(module), executor, device_allocator));
+
+ // Select an order for emitting the HLO instructions for each computation.
+ // Using this sequence enables tighter buffer liveness analysis and reduced
+ // memory usage (as compared to using DependencyHloOrdering).
+ TF_ASSIGN_OR_RETURN(HloSchedule schedule,
+ ScheduleModule(module.get(), BufferSizeBytesFunction(),
+ ComputationSchedulerToModuleScheduler(
+ DFSMemoryScheduler)));
+
+ // Run buffer allocation on the HLO graph.
+ TF_ASSIGN_OR_RETURN(
+ std::unique_ptr<BufferAssignment> assignment,
+ BufferAssigner::Run(module.get(),
+ absl::make_unique<SequentialHloOrdering>(schedule),
+ BufferSizeBytesFunction(), memory_alignment,
+ /*allocate_buffers_for_constants=*/true));
+
+ return std::make_tuple(std::move(module), std::move(assignment));
+}
+
namespace {
// Post-compilation callback functor for use by SimpleOrcJIT.
diff --git a/tensorflow/compiler/xla/service/cpu/cpu_compiler.h b/tensorflow/compiler/xla/service/cpu/cpu_compiler.h
index 537bf8b..d28ccd9 100644
--- a/tensorflow/compiler/xla/service/cpu/cpu_compiler.h
+++ b/tensorflow/compiler/xla/service/cpu/cpu_compiler.h
@@ -136,6 +136,12 @@
std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
se::DeviceMemoryAllocator* device_allocator) override;
+ StatusOr<
+ std::tuple<std::unique_ptr<HloModule>, std::unique_ptr<BufferAssignment>>>
+ RunHloPassesAndBufferAssignement(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* executor,
+ se::DeviceMemoryAllocator* device_allocator) override;
+
StatusOr<std::unique_ptr<Executable>> RunBackend(
std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
se::DeviceMemoryAllocator* device_allocator) override;
diff --git a/tensorflow/compiler/xla/service/cpu/ir_emitter.cc b/tensorflow/compiler/xla/service/cpu/ir_emitter.cc
index f4549ac..c19fa77 100644
--- a/tensorflow/compiler/xla/service/cpu/ir_emitter.cc
+++ b/tensorflow/compiler/xla/service/cpu/ir_emitter.cc
@@ -182,11 +182,8 @@
arch_type_ == llvm::Triple::ArchType::x86_64;
profiling_state_ = ProfilingState(use_rdtscp);
- bool emit_tracing =
- hlo_module_config_.hlo_profiling_enabled() &&
- hlo_module_config_.debug_options().xla_backend_extra_options().count(
- "xla_hlo_trace");
- tracing_state_.set_enabled(emit_tracing);
+ tracing_state_.set_enabled(
+ computation->parent()->config().cpu_traceme_enabled());
TF_RETURN_IF_ERROR(computation->AcceptOrdered(this, instruction_order));
llvm::Function* ir_function = compute_function_->function();
@@ -3126,7 +3123,8 @@
}
llvm::Type* int8_ptr_type = b->getInt8Ty()->getPointerTo();
- llvm::Type* void_ptr_type = b->getVoidTy()->getPointerTo();
+ llvm::Type* void_ptr_type =
+ int8_ptr_type; // LLVM does not have a void*, we use an int8* instead.
llvm::FunctionType* fn_type =
llvm::FunctionType::get(b->getInt64Ty(), {void_ptr_type, int8_ptr_type},
/*isVarArg=*/false);
@@ -3156,7 +3154,9 @@
return;
}
- llvm::Type* void_ptr_type = b->getVoidTy()->getPointerTo();
+ llvm::Type* void_ptr_type =
+ b->getInt8Ty()->getPointerTo(); // LLVM does not have a void*, we use an
+ // int8* instead.
llvm::FunctionType* fn_type =
llvm::FunctionType::get(b->getVoidTy(), {void_ptr_type, b->getInt64Ty()},
/*isVarArg=*/false);
diff --git a/tensorflow/compiler/xla/service/cpu/llvm_ir_runtime.cc b/tensorflow/compiler/xla/service/cpu/llvm_ir_runtime.cc
index 8a52301..f62769c 100644
--- a/tensorflow/compiler/xla/service/cpu/llvm_ir_runtime.cc
+++ b/tensorflow/compiler/xla/service/cpu/llvm_ir_runtime.cc
@@ -121,7 +121,8 @@
}
// Generate the vectorized code.
- CHECK_EQ(vector_width, input->getType()->getVectorNumElements());
+ CHECK_EQ(vector_width,
+ llvm::cast<llvm::VectorType>(input->getType())->getNumElements());
llvm::Value* result = fn_body_generator(&b, input, vector_width);
// Downcast result to scalar type if necessary.
diff --git a/tensorflow/compiler/xla/service/dynamic_dimension_inference.cc b/tensorflow/compiler/xla/service/dynamic_dimension_inference.cc
index a103b55..be26f9a 100644
--- a/tensorflow/compiler/xla/service/dynamic_dimension_inference.cc
+++ b/tensorflow/compiler/xla/service/dynamic_dimension_inference.cc
@@ -1369,77 +1369,27 @@
}
Status DynamicDimensionInferenceVisitor::HandleWhile(HloInstruction* hlo) {
- // While loop is handled by passing dynamic size hlos as parameters into the
- // hlo while loop. This is done by replacing the original while with a new
- // one.
- //
- // Before:
- //
- // op1 = ...
- // op2 = ...
- // op1_x = ... // dynamic dimension size of op1
- // while = while(op1, op2)
- //
- //
- // After:
- //
- // op1 = ...
- // op2 = ...
- // op1_x = ... // dynamic dimension size of op1
- // while = while(op1, op2, op1_x)
- //
- // In the above graph, op_x is the bound of the dynamic dimension size of op1
- // and is wired into the while loop as new parameter.
- //
- // TODO(b/119843103): Once we implement dynamic bounds in XLA backend, dynamic
- // bound can be propagated through native xla values instead of relying on
- // additional parameter.
-
- // dynamic_size_to_operand_id_index_map keeps track of dynamic size operations
- // to their operand ids in the new while loop.
- absl::flat_hash_map<HloInstruction*, int64>
- dynamic_size_to_operand_id_index_map;
-
- // operands_to_add collects dynamic sizes that need to be added to the while
- // loop as parameters. Note that a dynamic size is ignored if it is already
- // part of the parameter. i.e.:
- //
- // We don't do:
- //
- // op1 = ...
- // op2 = ...
- // op_x = ... // dynamic dimension size of both op1 and op2
- // while = while(op1, op2, op_x, op_x) // 4 parameters
- //
- // But we do:
- //
- // op1 = ...
- // op2 = ...
- // op_x = ... // dynamic dimension size of both op1 and op2
- // while = while(op1, op2, op_x)
- //
- // An alternative is to do this in a while loop CSE pass.
- //
+ // If the output of the conditional contains dynamic dimension. We send
+ // dynamic dimension size out by adding additional root element. A mapping
+ // from the root instruction's dynamic dimension index (represented by a shape
+ // index as output index and a int64 dimension number) to output index
+ // (represented by an int64) is tracked for the conditional instruction (all
+ // branches should have the same mapping).
+ ShapeTree<absl::flat_hash_map<int64, int64>> dynamic_output_mapping(
+ hlo->shape());
std::vector<HloInstruction*> operands_to_add;
- int64 operand_count = hlo->shape().tuple_shapes_size();
+ const int64 original_tuple_count = hlo->shape().tuple_shapes_size();
+ int64 operand_count = original_tuple_count;
TF_RETURN_IF_ERROR(ForEachOperandDynamicDimension(
- hlo, [&](HloInstruction*, ShapeIndex, int64, int64,
+ hlo, [&](HloInstruction*, ShapeIndex index, int64 dim, int64,
HloInstruction* dynamic_size, DimensionConstraint constraint) {
- const HloInstruction* tuple_operand = hlo->operand(0);
- for (int64 i = 0; i < tuple_operand->operand_count(); ++i) {
- if (dynamic_size == tuple_operand->operand(i)) {
- dynamic_size_to_operand_id_index_map[dynamic_size] = i;
- return Status::OK();
- }
- }
- auto iter = dynamic_size_to_operand_id_index_map.find(dynamic_size);
- if (iter == dynamic_size_to_operand_id_index_map.end()) {
- operands_to_add.push_back(dynamic_size);
- dynamic_size_to_operand_id_index_map[dynamic_size] = operand_count++;
- }
+ operands_to_add.push_back(dynamic_size);
+ dynamic_output_mapping.mutable_element(index)->emplace(dim,
+ operand_count++);
return Status::OK();
}));
+ DynamicParameterBinding binding_for_while;
if (!operands_to_add.empty()) {
// Only replace the while loop if there are new parameters to add.
HloInstruction* old_tuple_operand = hlo->mutable_operand(0);
@@ -1453,37 +1403,78 @@
parent_->CopyMapping(/*from=*/old_tuple_operand,
/*to=*/new_tuple_operand);
hlo = result.new_while_instr;
+ // We have replaced the while loop, now set the dynamic dimensions for the
+ // newly created while loop so that the hlos that consumes the while loop
+ // can see the dynamic dimensions. Also sets the dynamic parameter binding
+ // for running inference in the while loop.
+ TF_RETURN_IF_ERROR(ForEachOperandDynamicDimension(
+ hlo,
+ [&](HloInstruction*, ShapeIndex index, int64 dimension,
+ int64 operand_index, HloInstruction* dynamic_size,
+ DimensionConstraint constraint) -> Status {
+ TF_RET_CHECK(!operands_to_add.empty());
+ const int64 output_dynamic_size_index =
+ dynamic_output_mapping.element(index).at(dimension);
+ DynamicParameterBinding::DynamicParameter dynamic_parameter{
+ operand_index, {output_dynamic_size_index}};
+ DynamicParameterBinding::DynamicDimension dynamic_dimension{
+ operand_index, index, dimension};
+ TF_RETURN_IF_ERROR(
+ binding_for_while.Bind(dynamic_parameter, dynamic_dimension));
+ // This is the updated output dynamic size coming out of hlo while
+ // loop.
+ HloInstruction* output_dynamic_size = hlo->parent()->AddInstruction(
+ HloInstruction::CreateGetTupleElement(
+ ShapeUtil::MakeScalarShape(S32), hlo,
+ output_dynamic_size_index));
+ parent_->SetDynamicSize(result.replacement_instr, index, dimension,
+ output_dynamic_size, constraint);
+ return Status::OK();
+ }));
+ // Set the replacement instruction as visited to avoid visiting it again.
+ SetVisited(*result.replacement_instr);
}
- // We have replaced the while loop, now set the dynamic dimensions for the
- // newly created while loop so that the hlos that consumes the while loop can
- // see the dynamic dimensions. Also sets the dynamic parameter binding for
- // running inference in the while loop.
- DynamicParameterBinding binding_for_while;
- TF_RETURN_IF_ERROR(ForEachOperandDynamicDimension(
- hlo, [&](HloInstruction*, ShapeIndex index, int64 dimension,
- int64 operand_index, HloInstruction* dynamic_size,
- DimensionConstraint constraint) {
- DynamicParameterBinding::DynamicParameter dynamic_parameter{
- operand_index,
- {dynamic_size_to_operand_id_index_map[dynamic_size]}};
- DynamicParameterBinding::DynamicDimension dynamic_dimension{
- operand_index, index, dimension};
- TF_RETURN_IF_ERROR(
- binding_for_while.Bind(dynamic_parameter, dynamic_dimension));
- parent_->SetDynamicSize(hlo, index, dimension, dynamic_size,
- constraint);
- return Status::OK();
- }));
-
// Run inference in while body and condition.
TF_RETURN_IF_ERROR(DynamicDimensionInferenceVisitor::Run(
hlo->while_body(), binding_for_while, parent_));
TF_RETURN_IF_ERROR(DynamicDimensionInferenceVisitor::Run(
hlo->while_condition(), binding_for_while, parent_));
- // Set the replacement while loop as visited to avoid visiting it again.
- SetVisited(*hlo);
+ if (operands_to_add.empty()) {
+ // No dynamic dimension in the inputs and outputs.
+ return Status::OK();
+ }
+
+ // The dynamic dimension size could have been changed in the loop body (e.g, A
+ // loop that inserts items in a stack, the stack size increases with each
+ // iteration). Rewrite the dynamic dimension size at the root.
+ HloInstruction* body_root = hlo->while_body()->root_instruction();
+ std::vector<HloInstruction*> new_root_operands(body_root->operand_count(),
+ nullptr);
+
+ // Original non-dynamic-dim operands of root are pass-through.
+ for (int64 i = 0; i < original_tuple_count; ++i) {
+ new_root_operands[i] =
+ hlo->while_body()->AddInstruction(HloInstruction::CreateGetTupleElement(
+ body_root->shape().tuple_shapes(i), body_root, i));
+ }
+ // Add dynamic dimension size as new parameters.
+ TF_RETURN_IF_ERROR(ForEachDynamicDimension(
+ hlo->while_body()->root_instruction(),
+ [&](ShapeIndex index, int64 dim, HloInstruction* dynamic_size,
+ DimensionConstraint) -> Status {
+ const int64 output_index =
+ dynamic_output_mapping.element(index).at(dim);
+ new_root_operands[output_index] = dynamic_size;
+ return Status::OK();
+ }));
+ for (auto operand : new_root_operands) {
+ TF_RET_CHECK(operand != nullptr);
+ }
+ HloInstruction* new_body_root = hlo->while_body()->AddInstruction(
+ HloInstruction::CreateTuple(new_root_operands));
+ hlo->while_body()->set_root_instruction(new_body_root);
return Status::OK();
}
diff --git a/tensorflow/compiler/xla/service/dynamic_dimension_inference_test.cc b/tensorflow/compiler/xla/service/dynamic_dimension_inference_test.cc
index dc29566..b5a1761 100644
--- a/tensorflow/compiler/xla/service/dynamic_dimension_inference_test.cc
+++ b/tensorflow/compiler/xla/service/dynamic_dimension_inference_test.cc
@@ -767,7 +767,7 @@
// While
auto* a_param = builder.AddInstruction(HloInstruction::CreateParameter(
/*parameter_number=*/0, tuple_shape, "A"));
- auto* size_param = builder.AddInstruction(HloInstruction::CreateParameter(
+ builder.AddInstruction(HloInstruction::CreateParameter(
/*parameter_number=*/1, scalar_shape_, "size_param"));
builder.AddInstruction(
HloInstruction::CreateWhile(tuple_shape, condition, body, a_param));
@@ -782,37 +782,32 @@
DynamicParameterBinding::DynamicParameter{1, {}},
DynamicParameterBinding::DynamicDimension{0, {1}, 0}));
- // Test that dynamic dimension inference does the right thing. A lambda is
- // used here since we want to test twice by running inference again
- // (idempotency).
- auto test_dynamic_dimension = [&]() {
- HloInstruction* while_hlo = nullptr;
- // The while hlo has been replaced, find the new one.
- for (HloInstruction* inst : module_->entry_computation()->instructions()) {
- if (inst->opcode() == HloOpcode::kWhile) {
- while_hlo = inst;
- }
- }
- ASSERT_NE(while_hlo, nullptr);
- // The original while shape has 2 parameters. With dynamic size passed in
- // as an extra parameter, the tuple should have 3 elements.
- EXPECT_EQ(while_hlo->shape().tuple_shapes_size(), 3);
- HloInstruction* add = nullptr;
- for (HloInstruction* inst : while_hlo->while_body()->instructions()) {
- if (inst->opcode() == HloOpcode::kAdd) {
- add = inst;
- }
- }
- EXPECT_NE(add, nullptr);
- EXPECT_NE(inference_->GetDynamicSize(add, {}, 0), nullptr);
- EXPECT_EQ(inference_->GetDynamicSize(while_hlo, {0}, 0), size_param);
- EXPECT_EQ(inference_->GetDynamicSize(while_hlo, {1}, 0), size_param);
- };
-
TF_ASSERT_OK(RunInference());
- test_dynamic_dimension();
- TF_ASSERT_OK(RunInference());
- test_dynamic_dimension();
+ HloInstruction* while_hlo = nullptr;
+ // The while hlo has been replaced, find the new one.
+ for (HloInstruction* inst : module_->entry_computation()->instructions()) {
+ if (inst->opcode() == HloOpcode::kWhile) {
+ while_hlo = inst;
+ }
+ }
+ ASSERT_NE(while_hlo, nullptr);
+ // The original while shape has 2 parameters. With dynamic size, the tuple
+ // should have 4 elements (We don't deduplicate the arguments).
+ EXPECT_EQ(while_hlo->shape().tuple_shapes_size(), 4);
+ HloInstruction* add_inst = nullptr;
+ for (HloInstruction* inst : while_hlo->while_body()->instructions()) {
+ if (inst->opcode() == HloOpcode::kAdd) {
+ add_inst = inst;
+ }
+ }
+ EXPECT_NE(add_inst, nullptr);
+ EXPECT_NE(inference_->GetDynamicSize(add_inst, {}, 0), nullptr);
+ EXPECT_NE(inference_->GetDynamicSize(
+ module_->entry_computation()->root_instruction(), {0}, 0),
+ nullptr);
+ EXPECT_NE(inference_->GetDynamicSize(
+ module_->entry_computation()->root_instruction(), {1}, 0),
+ nullptr);
}
TEST_F(DynamicDimensionInferenceTest, ConditionalInputTest) {
diff --git a/tensorflow/compiler/xla/service/dynamic_padder.cc b/tensorflow/compiler/xla/service/dynamic_padder.cc
index e0fe9c0..09b1578 100644
--- a/tensorflow/compiler/xla/service/dynamic_padder.cc
+++ b/tensorflow/compiler/xla/service/dynamic_padder.cc
@@ -244,8 +244,9 @@
Status RewriteDynamicReshapeSplitInput(
HloInstruction* reshape, int64 input_dim,
absl::Span<const int64> output_dims,
+ absl::Span<HloInstruction*> output_dynamic_dims,
DynamicDimensionInference* dynamic_dimension_inference) {
- VLOG(1) << "Reshaping input dim " << input_dim << "to "
+ VLOG(2) << "Reshaping input dim " << input_dim << "to "
<< VectorString(output_dims);
const Shape operand_shape = reshape->operand(0)->shape();
TF_RET_CHECK(output_dims.size() > 1);
@@ -280,8 +281,7 @@
// dimension.
for (int64 i = 1; i < output_dims.size(); ++i) {
const int64 output_dim = output_dims[i];
- HloInstruction* dynamic_size =
- dynamic_dimension_inference->GetDynamicSize(reshape, {}, output_dim);
+ HloInstruction* dynamic_size = output_dynamic_dims[output_dim];
if (dynamic_size == nullptr) {
continue;
}
@@ -331,10 +331,7 @@
mask_input_shape, HloOpcode::kSubtract, cumsum, broadcast_ones));
GatherDimensionNumbers gather_dim_numbers;
- // We use gather to rearrange the input dim dimension. However the current
- // semantic of gather doesn't allow us to collapse dimension in this case so
- // we keep it, which make the gather from shape [..., input_dim, ...] to
- // [..., 1, input_dim, ...]
+ // Use gather to rearrange the input dim dimension.
for (int64 i = 0; i < operand_shape.dimensions_size(); ++i) {
// Offset dim is every dimension including newly added size 1 dim, except
// for input_dim, which acts as a batch_dim.
@@ -396,177 +393,255 @@
return Status::OK();
}
+// RewriteDynamicReshapeCombineInput is similar to
+// RewriteDynamicReshapeSplitInput, in a reshape if multiple dimensions are
+// combined into one dimension, we need to rewrite the output.
+//
+// The reason for this is that a continuous input may not be evenly reshaped
+// into output. Image we have [2, <=3] where second dimension has size 2 and
+// padding(P) data has size 1:
+// [[a,b,P]
+// [c,d,P]]
+//
+// And we have a reshape that combines this two input dimensions.
+//
+// [2, <=3]
+// |
+// Reshape
+// |
+// [6]
+//
+// This should produce the same result as if the data has no padding:
+//
+// [2, 2] // [[a, b], [c, d]]
+// |
+// Reshape
+// |
+// [4] // [a,b,c,d]
+//
+// Without rewriting, the result would be:
+//
+// [a,b,P,c,d,P], which is incorrect.
+//
+// We need to rewrite the reshape such that it produces:
+// [a,b,c,d,P,P]
+//
+// The way we do this is by a 5-steps sort-gather algorithm:
+//
+// 1.First we use the input shape to generate a binary 0-1 masking, which masks
+// out the padded area of the output:
+// [[0,0,1]
+// [0,0,1]]
+//
+// 2.Then we do an reshape to reshape the mask from input shape to output
+// shape [2,3]->[6]:
+// [0,0,1,0,0,1]
+//
+// 3.We then generate an iota mask using the output shape:
+// [0,1,2,3,4,5]
+//
+// 4.Stable sort the iota mask using the binary mask as key:
+// key [0,0,1,0,0,1]
+// value[0,1,2,3,4,5]
+// | Sort by key
+// v
+// key [0,0,0,0,1,1]
+// value[0,1,3,4,2,5]
+//
+// 5.Gather the original output [a,b,P,c,d,P] using the sorted iota mask:
+// original output gather indices
+// [a,b,P,c,d,P] [0,1,3,4,2,5]
+// | |
+// Gather ----------------+
+// |
+// [a,b,c,d,P,P]
+//
Status RewriteDynamicReshapeCombineInput(
- HloInstruction* reshape, int64 input_dim, int64 output_dim,
- HloInstruction* dynamic_size,
+ HloInstruction* reshape, absl::Span<const int64> input_dims,
+ int64 output_dim, absl::Span<HloInstruction*> input_dynamic_dims,
DynamicDimensionInference* dynamic_dimension_inference) {
// Rewrite dynamic reshape into reshape followed by a sort, all padded
// data will be moved to the end.
- const HloInstruction* operand = reshape->operand(0);
HloComputation* comp = reshape->parent();
HloInstruction* zero = comp->AddInstruction(
HloInstruction::CreateConstant(LiteralUtil::Zero(S32)));
HloInstruction* one = comp->AddInstruction(
HloInstruction::CreateConstant(LiteralUtil::One(S32)));
- const Shape mask_shape =
- ShapeUtil::ChangeElementType(operand->shape(), xla::S32);
- const Shape mask_reshaped_shape =
- ShapeUtil::ChangeElementType(reshape->shape(), xla::S32);
- HloInstruction* broadcasted_zero = comp->AddInstruction(
- HloInstruction::CreateBroadcast(mask_shape, zero, {}));
- // Pad masking area with 1s, rest with 0s.
- HloInstruction* padding_mask =
- PadWithScalar(broadcasted_zero, input_dim, dynamic_size, one);
- HloInstruction* mask_reshaped = comp->AddInstruction(
- HloInstruction::CreateReshape(mask_reshaped_shape, padding_mask));
+ const Shape output_shape = reshape->shape();
+ const Shape input_shape = reshape->operand(0)->shape();
+ const Shape mask_output_shape =
+ ShapeUtil::MakeShape(xla::S32, {output_shape.dimensions(output_dim)});
+ std::vector<int64> input_dim_sizes;
+ for (int64 input_dim : input_dims) {
+ input_dim_sizes.push_back(input_shape.dimensions(input_dim));
+ }
- // Build computation for reshape, key is the mask shape, value is reshape's
- // original data.
+ const Shape mask_input_shape =
+ ShapeUtil::MakeShape(xla::S32, input_dim_sizes);
+
+ // Step 1 -- generate binary mask.
+ // Mask starts with all zero, each dynamic dimension sets that dimension of
+ // the mask to partially ones in the end.
+ HloInstruction* binary_mask = comp->AddInstruction(
+ HloInstruction::CreateBroadcast(mask_input_shape, zero, {}));
+
+ bool need_rewrite = false;
+
+ // Pad the effective dimension with 1.
+ //
+ // Index starts from 1 since there is no need to rewrite a major output
+ // dimension.
+ for (int64 i = 1; i < input_dims.size(); ++i) {
+ const int64 input_dim = input_dims[i];
+ HloInstruction* dynamic_size = input_dynamic_dims[input_dim];
+ if (dynamic_size == nullptr) {
+ continue;
+ }
+ // If there is a dynamic dimension in the input, need to rewrite the output.
+ need_rewrite = true;
+
+ binary_mask = PadWithScalar(binary_mask, i, dynamic_size, one);
+ }
+ if (!need_rewrite) {
+ VLOG(2) << "No need to rewrite";
+ return Status::OK();
+ }
+
+ // Step 2.
+ // Do a reshape to flatten the binary mask into output_shape
+ HloInstruction* output_shape_binary_mask = comp->AddInstruction(
+ HloInstruction::CreateReshape(mask_output_shape, binary_mask));
+
+ // Step 3.
+ // Generate an iota with output shape.
+ HloInstruction* iota =
+ comp->AddInstruction(HloInstruction::CreateIota(mask_output_shape, 0));
+
+ // Step 4.
+ // Stable sort the iota mask using the binary mask as key and iota as value:
+
+ // Build computation for sort, key is the mask, value is the iota.
HloComputation::Builder comp_builder("compare");
HloInstruction* lhs_key =
comp_builder.AddInstruction(HloInstruction::CreateParameter(
- 0, ShapeUtil::MakeShape(S32, {}), "lhs_key"));
+ 0, ShapeUtil::MakeScalarShape(S32), "lhs_key"));
HloInstruction* rhs_key =
comp_builder.AddInstruction(HloInstruction::CreateParameter(
- 1, ShapeUtil::MakeShape(S32, {}), "rhs_key"));
+ 1, ShapeUtil::MakeScalarShape(S32), "rhs_key"));
// Values for lhs and rhs
comp_builder.AddInstruction(HloInstruction::CreateParameter(
- 2, ShapeUtil::MakeShape(operand->shape().element_type(), {}),
- "lhs_value"));
+ 2, ShapeUtil::MakeScalarShape(S32), "lhs_value"));
comp_builder.AddInstruction(HloInstruction::CreateParameter(
- 3, ShapeUtil::MakeShape(operand->shape().element_type(), {}),
- "rhs_value"));
+ 3, ShapeUtil::MakeScalarShape(S32), "rhs_value"));
comp_builder.AddInstruction(
HloInstruction::CreateCompare(ShapeUtil::MakeShape(PRED, {}), lhs_key,
rhs_key, ComparisonDirection::kLt));
HloComputation* compare =
comp->parent()->AddEmbeddedComputation(comp_builder.Build());
+ // Use mask_reshaped as key, sort reshaped data as value.
+ HloInstruction* sort = comp->AddInstruction(HloInstruction::CreateSort(
+ ShapeUtil::MakeTupleShape({mask_output_shape, mask_output_shape}), 0,
+ {output_shape_binary_mask, iota}, compare,
+ /*is_stable=*/true));
+
+ HloInstruction* gather_indices = comp->AddInstruction(
+ HloInstruction::CreateGetTupleElement(mask_output_shape, sort, 1));
+
+ // Step 5.Gather the original output using the sorted iota mask:
+
+ GatherDimensionNumbers gather_dim_numbers;
+ // Use gather to rearrange the output dim dimension.
+ for (int64 i = 0; i < output_shape.dimensions_size(); ++i) {
+ // Offset dim is every dimension including newly added size 1 dim, except
+ // for input_dim, which acts as a batch_dim.
+ if (i != output_dim) {
+ gather_dim_numbers.add_offset_dims(i);
+ }
+ }
+ // The dimension to rewrite is the index dim.
+ gather_dim_numbers.add_start_index_map(output_dim);
+ gather_dim_numbers.set_index_vector_dim(1);
+ gather_dim_numbers.add_collapsed_slice_dims(output_dim);
+
HloInstruction* static_dim_size = comp->AddInstruction(
HloInstruction::CreateConstant(LiteralUtil::CreateR0<int32>(
reshape->shape().dimensions(output_dim))));
// Temporarily removes dynamic dimension of the reshape before we send it to
- // the sort -- we want padded area to also participate in the sort.
+ // the sort -- we want padded area to also participate in the gather.
HloInstruction* reshape_static =
comp->AddInstruction(HloInstruction::CreateSetDimensionSize(
reshape->shape(), reshape, static_dim_size, output_dim));
+ std::vector<int64> gather_slice_sizes(output_shape.dimensions().begin(),
+ output_shape.dimensions().end());
+ gather_slice_sizes[output_dim] = 1;
+ HloInstruction* gather = comp->AddInstruction(HloInstruction::CreateGather(
+ output_shape, reshape_static, gather_indices, gather_dim_numbers,
+ gather_slice_sizes, true));
- // Use mask_reshaped as key, sort reshaped data as value.
- HloInstruction* sort = comp->AddInstruction(HloInstruction::CreateSort(
- ShapeUtil::MakeTupleShape({mask_reshaped_shape, reshape->shape()}),
- output_dim, {mask_reshaped, reshape_static}, compare,
- /*is_stable=*/true));
- HloInstruction* dynamic_reshape = comp->AddInstruction(
- HloInstruction::CreateGetTupleElement(reshape->shape(), sort, 1));
- // Forward dynamic size to the newly created reshape.
+ // Forward dynamic size to the newly created gather.
HloInstruction* output_dynamic_size =
dynamic_dimension_inference->GetDynamicSize(reshape, {}, output_dim);
TF_RET_CHECK(output_dynamic_size != nullptr);
- dynamic_reshape = comp->AddInstruction(HloInstruction::CreateSetDimensionSize(
- dynamic_reshape->shape(), dynamic_reshape, output_dynamic_size,
- output_dim));
+ gather = comp->AddInstruction(HloInstruction::CreateSetDimensionSize(
+ gather->shape(), gather, output_dynamic_size, output_dim));
auto users = reshape->users();
for (auto* user : users) {
// Avoid cycles by not replacing the staic reshape and get_dimension_size.
if (user != reshape_static && user != output_dynamic_size) {
- TF_RETURN_IF_ERROR(reshape->ReplaceUseWith(user, dynamic_reshape));
+ TF_RETURN_IF_ERROR(reshape->ReplaceUseWith(user, gather));
}
}
if (reshape == comp->root_instruction()) {
- comp->set_root_instruction(dynamic_reshape);
+ comp->set_root_instruction(gather);
}
- TF_RETURN_IF_ERROR(dynamic_dimension_inference->ForwardDynamicSize(
- reshape, dynamic_reshape, {}));
+ TF_RETURN_IF_ERROR(
+ dynamic_dimension_inference->ForwardDynamicSize(reshape, gather, {}));
return Status::OK();
}
-Status RewriteDynamicReshapeSingleDim(
- HloInstruction* reshape, int64 input_dim, HloInstruction* dynamic_size,
+Status RewriteDynamicReshapeSingleGroup(
+ HloInstruction* reshape, absl::Span<const int64> input_dims,
+ absl::Span<const int64> output_dims,
+ absl::Span<HloInstruction*> input_dynamic_dims,
+ absl::Span<HloInstruction*> output_dynamic_dims,
DynamicDimensionInference* dynamic_dimension_inference) {
VLOG(2) << "Rewriting dynamic reshape " << reshape->ToString()
- << " input dim: " << input_dim;
+ << " input dims: " << VectorString(input_dims)
+ << " output dims: " << VectorString(output_dims);
+
const Shape operand_shape = reshape->operand(0)->shape();
const Shape output_shape = reshape->shape();
- const int64 static_input_dim_size = operand_shape.dimensions()[input_dim];
-
- // Don't need to rewrite size 1 input dims.
- if (static_input_dim_size == 1) {
- return Status::OK();
- }
-
- auto common_factors =
- CommonFactors(operand_shape.dimensions(), output_shape.dimensions());
- // If there are multiple input dims combining into one output dim,
- // input_dim_start and input_dim_end represent the input dimension range.
- int64 input_dim_start = -1;
- int64 input_dim_end = -1;
- // Similarly when one input dim is splitted into multiple outputs, we use
- // output_dim_start and output_dim_start to represent the output dimension
- // range.
- int64 output_dim_start = -1;
- int64 output_dim_end = -1;
- // Find common_factors that the input belong to.
- for (int64 i = 0; i < common_factors.size() - 1; ++i) {
- auto start = common_factors[i];
- auto end = common_factors[i + 1];
- if (input_dim >= start.first && input_dim < end.first) {
- // Found the common_factor group that the input_dim belongs to.
- input_dim_start = start.first;
- input_dim_end = end.first;
- output_dim_start = start.second;
- output_dim_end = end.second;
+ if (input_dims.size() == 1) {
+ int64 input_dim = input_dims[0];
+ // Size 1 dimension doesn't need a rewrite.
+ if (operand_shape.dimensions()[input_dim] == 1) {
+ return Status::OK();
}
- }
-
- TF_RET_CHECK(output_dim_end - output_dim_start > 0);
-
- std::vector<int64> output_dims;
- for (int64 i = output_dim_start; i < output_dim_end; ++i) {
- output_dims.push_back(i);
- }
-
- const int64 first_output_dim = output_dims[0];
-
- if (reshape->shape().dimensions(first_output_dim) < static_input_dim_size) {
// One input dimension is splitted into multiple output dimensions.
return RewriteDynamicReshapeSplitInput(reshape, input_dim, output_dims,
+ output_dynamic_dims,
dynamic_dimension_inference);
}
- if (reshape->shape().dimensions(first_output_dim) == static_input_dim_size) {
- // Unchanged dynamic dimension doesn't need a rewrite.
- return Status::OK();
- }
-
- // Multiple dimensions got combined into one output.
- if (input_dim != input_dim_start) {
- // If 'input_dim' is not the first dimension that got combined into the
- // output. A reshape rewrite on the output is needed:
- //
- // Need a write (d is dynamic):
- // 1, 2, d
- // |
- // Reshape
- // |
- // 2d
- //
- // Don't need rewrite:
- // d, 2
- // |
- // Reshape
- // |
- // 2d
- //
- return RewriteDynamicReshapeCombineInput(reshape, input_dim,
- first_output_dim, dynamic_size,
+ if (output_dims.size() == 1) {
+ int64 output_dim = output_dims[0];
+ if (output_shape.dimensions()[output_dim] == 1) {
+ return Status::OK();
+ }
+ // One input dimension is splitted into multiple output dimensions.
+ return RewriteDynamicReshapeCombineInput(reshape, input_dims, output_dim,
+ input_dynamic_dims,
dynamic_dimension_inference);
}
+ // Shouldn't get here;
+ TF_RET_CHECK(false);
return Status::OK();
}
@@ -718,23 +793,85 @@
DynamicDimensionInference* dynamic_dimension_inference) {
bool changed = false;
HloInstruction* operand = reshape->mutable_operand(0);
+ std::vector<HloInstruction*> input_dynamic_dims;
+ for (int64 dim = 0; dim < operand->shape().dimensions_size(); ++dim) {
+ input_dynamic_dims.push_back(
+ dynamic_dimension_inference->GetDynamicSize(operand, {}, dim));
+ }
- // We append sort instructions after reshape if there is a dynamic input, and
- // the order of sort matters. Rewrite minor dimensions first in case multiple
- // inputs have dynamic dimensions to ensure correct order of sort.
- for (int64 input_dim = operand->shape().rank() - 1; input_dim >= 0;
- --input_dim) {
- HloInstruction* operand_dynamic_size =
- dynamic_dimension_inference->GetDynamicSize(operand, {}, input_dim);
+ std::vector<HloInstruction*> output_dynamic_dims;
+ for (int64 dim = 0; dim < reshape->shape().dimensions_size(); ++dim) {
+ output_dynamic_dims.push_back(
+ dynamic_dimension_inference->GetDynamicSize(reshape, {}, dim));
+ }
- if (operand_dynamic_size == nullptr) {
+ auto common_factors = CommonFactors(operand->shape().dimensions(),
+ reshape->shape().dimensions());
+ // Find common_factors that the input belongs to.
+ for (int64 i = 0; i < common_factors.size() - 1; ++i) {
+ auto start = common_factors[i];
+ auto end = common_factors[i + 1];
+ std::vector<int64> input_dims;
+ std::vector<int64> output_dims;
+ for (int64 dim = start.first; dim < end.first; ++dim) {
+ input_dims.push_back(dim);
+ }
+ for (int64 dim = start.second; dim < end.second; ++dim) {
+ output_dims.push_back(dim);
+ }
+
+ VLOG(2) << "input_dims: " << VectorString(input_dims);
+ VLOG(2) << "output_dims: " << VectorString(output_dims);
+
+ if (input_dims.empty() || output_dims.empty()) {
continue;
}
- TF_RETURN_IF_ERROR(RewriteDynamicReshapeSingleDim(
- reshape, input_dim, operand_dynamic_size, dynamic_dimension_inference));
+ bool has_dynamic_dimension = absl::c_any_of(output_dims, [&](int64 dim) {
+ HloInstruction* operand_dynamic_size =
+ dynamic_dimension_inference->GetDynamicSize(reshape, {}, dim);
- changed = true;
+ return operand_dynamic_size != nullptr ||
+ reshape->shape().is_dynamic_dimension(dim);
+ });
+
+ if (!has_dynamic_dimension) {
+ // Don't need to rewrite any group without dynamic dimensions.
+ VLOG(2) << "All dimensions are static in this common factor group";
+ continue;
+ }
+
+ if (input_dims.size() == 1 && output_dims.size() == 1) {
+ // The dimension is unchanged. No rewrite needed.
+ continue;
+ }
+ if (input_dims.size() > 1 && output_dims.size() > 1) {
+ // We don't support the case when a dynamic dimension is both combined
+ // with and splitted into other dimensions:
+ //
+ // [x, yz]
+ // | Reshape
+ // [xy, z]
+ //
+ // TODO(yunxing): This can be supported by canonicalizing
+ // the offending reshape into two reshapes:
+ //
+ // [x,yz]
+ // | Reshape
+ // [x, y, z]
+ // | Reshape
+ // [xy, z]
+ //
+ return Unimplemented(
+ "Dynamic input dimension to reshape that is both splitted and "
+ "combined is not supported %s",
+ reshape->ToString());
+ }
+
+ TF_RETURN_IF_ERROR(RewriteDynamicReshapeSingleGroup(
+ reshape, input_dims, output_dims, absl::MakeSpan(input_dynamic_dims),
+ absl::MakeSpan(output_dynamic_dims), dynamic_dimension_inference));
}
+
return changed;
}
diff --git a/tensorflow/compiler/xla/service/dynamic_padder_test.cc b/tensorflow/compiler/xla/service/dynamic_padder_test.cc
index c937bf2..31ae1ab 100644
--- a/tensorflow/compiler/xla/service/dynamic_padder_test.cc
+++ b/tensorflow/compiler/xla/service/dynamic_padder_test.cc
@@ -844,6 +844,149 @@
EXPECT_EQ(result, expected);
}
+XLA_TEST_F(ExecutionTest, ReshapeSplitCombineSameTime) {
+ // [<=4, 2, <=2]
+ // |
+ // Reshape
+ // |
+ // [2, <=2, <=4]
+ //
+ // Split one input dynamic dim to multiple output dims while combining two
+ // dimensions together.
+ //
+ const string hlo_text = R"(
+HloModule TensorFlowScatterV1
+
+update_s32 (lhs: s32[], rhs: s32[]) -> s32[] {
+ lhs = s32[] parameter(0)
+ rhs = s32[] parameter(1)
+ ROOT add = s32[] add(lhs, rhs)
+}
+
+ENTRY main {
+ param = s32[4, 2, 2] parameter(0)
+ two = s32[] constant(2)
+ one = s32[] constant(1)
+ param_padded_partial = s32[<=4, 2, 2] set-dimension-size(param, two),
+ dimensions={0}
+
+ param_padded_dynamic = s32[<=4, 2, <=2] set-dimension-size(param_padded_partial,
+ one),
+ dimensions={2}
+ reshaped = s32[2, <=2, <=4] reshape(param_padded_dynamic),
+ inferred_dimension=1
+ init = s32[] constant(0)
+ ROOT reduce = s32[] reduce(reshaped, init),
+ dimensions={0, 1, 2},
+ to_apply=update_s32
+}
+)";
+
+ // First and last dims are dynamic. Padded data are expressed as -1.
+ Literal operand = LiteralUtil::CreateR3<int32>({{{0, -1}, {1, -1}},
+ {{2, -1}, {3, -1}},
+ {{-1, -1}, {-1, -1}},
+ {{-1, -1}, {-1, -1}}});
+ auto module = GetHloModule(hlo_text);
+
+ Literal result = PadAndExecute(std::move(module), {&operand});
+
+ // Reshaping (with correct reshape rewriting) produces:
+ // [[[0, 1, -1, -1], [-1, -1, -1, -1]], [[2, 3, -1, -1], [-1, -1, -1, -1]]]
+ //
+ // Dynamic padder auto pads -1 with 0.
+ //
+ // Reducing it produces 0 + 1 + 2 + 3 = 6
+
+ Literal expected = LiteralUtil::CreateR0<int32>(6);
+
+ EXPECT_EQ(result, expected);
+}
+
+XLA_TEST_F(ExecutionTest, WhileLoopStack) {
+ // Push into a dynamic sized stack with iteration number:
+ // init:
+ // [[P, P],
+ // [P, P],
+ // [P, P],
+ // [P, P]]
+ // First iteration i = 0:
+ // [[0, 0],
+ // [P, P],
+ // [P, P],
+ // [P, P]]
+ // Second iteration i = 1:
+ // [[0, 0],
+ // [1, 1],
+ // [P, P],
+ // [P, P]]
+ // Third iteration i = 2:
+ // [[0, 0],
+ // [1, 1],
+ // [2, 2],
+ // [P, P]]
+
+ const string hlo_text = R"(
+HloModule module
+
+update_s32 (lhs: s32[], rhs: s32[]) -> s32[] {
+ lhs = s32[] parameter(0)
+ rhs = s32[] parameter(1)
+ ROOT add = s32[] add(lhs, rhs)
+}
+
+body {
+ stack = (s32[<=4,2]) parameter(0)
+ stack_buffer = s32[<=4, 2] get-tuple-element(stack), index=0
+ stack_size = s32[] get-dimension-size(stack_buffer), dimensions={0}
+ zero = s32[] constant(0)
+ one = s32[] constant(1)
+ // content of the stack is the stack index broadcasted.
+ new_data = s32[1, 2] broadcast(s32[] stack_size), dimensions={}
+ new_stack_buffer = s32[<=4, 2] dynamic-update-slice(stack_buffer, new_data, stack_size, zero)
+ new_stack_size = s32[] add(stack_size, one)
+ new_stack_buffer_dynamic = s32[<=4, 2]set-dimension-size(new_stack_buffer, new_stack_size), dimensions={0}
+ ROOT new_stack = (s32[<=4,2]) tuple(new_stack_buffer_dynamic)
+}
+
+condition {
+ stack = (s32[<=4,2]) parameter(0)
+ stack_buffer = s32[<=4, 2] get-tuple-element(stack), index=0
+ stack_size = s32[] get-dimension-size(stack_buffer), dimensions={0}
+ three = s32[] constant(3)
+ ROOT less-than = pred[] compare(s32[] stack_size, s32[] three), direction=LT
+}
+
+ENTRY entry {
+ zero = s32[] constant(0)
+ pad = s32[] constant(-1)
+ stack_buffer_input = s32[4, 2] broadcast(s32[] pad), dimensions={}
+ stack_buffer_input_dynamic = s32[<=4, 2] set-dimension-size(stack_buffer_input, zero), dimensions={0}
+ input_tuple = (s32[<=4 ,2]) tuple(stack_buffer_input_dynamic)
+ while = (s32[<=4, 2]) while(input_tuple), body=body, condition=condition
+ stack_buffer = s32[<=4, 2] get-tuple-element(while), index=0
+ ROOT reduce = s32[2] reduce(stack_buffer, zero),
+ dimensions={0},
+ to_apply=update_s32
+}
+)";
+
+ auto module = GetHloModule(hlo_text);
+
+ Literal result = PadAndExecute(std::move(module), {});
+
+ // Stack has three valid items in it:
+ // [[0, 0],
+ // [1, 1],
+ // [2, 2],
+ // [P, P]]
+ //
+ // Reducing along major dimension gives us [3, 3]
+ Literal expected = LiteralUtil::CreateR1<int32>({{3, 3}});
+
+ EXPECT_EQ(result, expected);
+}
+
XLA_TEST_F(ExecutionTest, DoubleDynamicDimension) {
const string hlo_text = R"(
HloModule TensorFlowScatterV1
diff --git a/tensorflow/compiler/xla/service/gpu/horizontal_fusion.cc b/tensorflow/compiler/xla/service/gpu/horizontal_fusion.cc
index 5e7593a..6d663c6 100644
--- a/tensorflow/compiler/xla/service/gpu/horizontal_fusion.cc
+++ b/tensorflow/compiler/xla/service/gpu/horizontal_fusion.cc
@@ -192,6 +192,14 @@
return false;
}
+ // We can emit DUS in-place, horizontally fusing it makes the emitter no
+ // longer recognize that it can be done in-place. This creates much slower
+ // code. This restriction could be lifted if buffer assignment would recognize
+ // that the DUS can be done in-place even inside of a horizontal fusion.
+ if (root->opcode() == HloOpcode::kDynamicUpdateSlice) {
+ return false;
+ }
+
return true;
}
diff --git a/tensorflow/compiler/xla/service/gpu/horizontal_fusion_test.cc b/tensorflow/compiler/xla/service/gpu/horizontal_fusion_test.cc
index e1024f6..bad5899 100644
--- a/tensorflow/compiler/xla/service/gpu/horizontal_fusion_test.cc
+++ b/tensorflow/compiler/xla/service/gpu/horizontal_fusion_test.cc
@@ -364,6 +364,45 @@
EXPECT_TRUE(RunAndCompare(std::move(module), ErrorSpec{1.0e-5, 1.0e-5}));
}
+TEST_F(HorizontalFusionTest, NegativeTestForDynamicUpdateSlice) {
+ auto module = ParseAndReturnVerifiedModule(R"(
+ HloModule NegativeTestForDynamicUpdateSlice
+
+ fusion.1 {
+ p.0 = f16[5,9,10]{2,1,0} parameter(0)
+ p.1 = s32[1]{0} parameter(1)
+ p.2 = f16[1,9,10]{2,1,0} parameter(2)
+ c.0 = s32[] constant(0)
+ pad = s32[3]{0} pad(p.1, c.0), padding=0_2
+ ROOT %dynamic-update-slice = f16[5,9,10]{2,1,0} dynamic-update-slice(p.0, p.2, pad)
+ }
+
+ fusion.2 {
+ p.0 = f16[5,9,10]{2,1,0} parameter(0)
+ p.1 = s32[1]{0} parameter(1)
+ p.2 = f16[1,9,10]{2,1,0} parameter(2)
+ c.0 = s32[] constant(0)
+ pad = s32[3]{0} pad(p.1, c.0), padding=0_2
+ ROOT %dynamic-update-slice = f16[5,9,10]{2,1,0} dynamic-update-slice(p.0, p.2, pad)
+ }
+
+ ENTRY entry {
+ p.00 = f16[5,9,10]{2,1,0} parameter(0)
+ p.01 = f16[5,9,10]{2,1,0} parameter(1)
+ p.10 = s32[1]{0} parameter(2)
+ p.11 = s32[1]{0} parameter(3)
+ p.20 = f16[1,9,10]{2,1,0} parameter(4)
+ p.21 = f16[1,9,10]{2,1,0} parameter(5)
+
+ f1 = f16[5,9,10] fusion(p.00, p.10, p.20), kind=kLoop, calls=fusion.1
+ f2 = f16[5,9,10] fusion(p.01, p.11, p.21), kind=kLoop, calls=fusion.2
+ ROOT tuple = (f16[5,9,10],f16[5,9,10]) tuple(f1, f2)
+ })")
+ .ValueOrDie();
+
+ EXPECT_FALSE(GpuHorizontalFusion().Run(module.get()).ValueOrDie());
+}
+
} // namespace
} // namespace gpu
} // namespace xla
diff --git a/tensorflow/compiler/xla/service/hlo_evaluator_test.cc b/tensorflow/compiler/xla/service/hlo_evaluator_test.cc
index db651d3..b04635d 100644
--- a/tensorflow/compiler/xla/service/hlo_evaluator_test.cc
+++ b/tensorflow/compiler/xla/service/hlo_evaluator_test.cc
@@ -4442,5 +4442,27 @@
EXPECT_TRUE(LiteralTestUtil::Equal(expected, result));
}
+TEST_F(HloEvaluatorTest, MapBF16) {
+ const absl::string_view hlo_text = R"(
+ HloModule test
+
+ map_computation {
+ p = bf16[] parameter(0)
+ add = bf16[] add(p, p)
+ ROOT conv = f32[] convert(add)
+ }
+
+ ENTRY CopyStartCopyDone {
+ c = bf16[3] constant({1, 2, 3})
+ ROOT map = f32[3] map(c), to_apply=map_computation
+ }
+ )";
+ TF_ASSERT_OK_AND_ASSIGN(m_, ParseAndReturnVerifiedModule(hlo_text));
+ Literal expected = LiteralUtil::CreateR1<float>({2.f, 4.f, 6.f});
+ TF_ASSERT_OK_AND_ASSIGN(
+ Literal result, HloEvaluator().Evaluate(*m_->entry_computation(), {}));
+ EXPECT_TRUE(LiteralTestUtil::Equal(expected, result));
+}
+
} // namespace
} // namespace xla
diff --git a/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h b/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h
index 6fa3f9f..e105ea8 100644
--- a/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h
+++ b/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h
@@ -1680,6 +1680,10 @@
MapImpl<Eigen::half>(map));
break;
}
+ case BF16: {
+ TF_ASSIGN_OR_RETURN(parent_->evaluated_[map], MapImpl<bfloat16>(map));
+ break;
+ }
case F32: {
TF_ASSIGN_OR_RETURN(parent_->evaluated_[map], MapImpl<float>(map));
break;
diff --git a/tensorflow/compiler/xla/service/hlo_module_config.h b/tensorflow/compiler/xla/service/hlo_module_config.h
index d90a148..b31a9ae 100644
--- a/tensorflow/compiler/xla/service/hlo_module_config.h
+++ b/tensorflow/compiler/xla/service/hlo_module_config.h
@@ -104,10 +104,20 @@
return debug_options_.xla_hlo_profile();
}
+ bool cpu_traceme_enabled() const {
+ return debug_options_.xla_cpu_enable_xprof_traceme();
+ }
+
// Sets/returns the module seed set during execution.
void set_seed(uint64 seed) { seed_ = seed; }
uint64 seed() const { return seed_; }
+ // Set the launch id of the program. Launch id identifies a set of programs
+ // that should be launched together.
+ void set_launch_id(uint64 launch_id) { launch_id_ = launch_id; }
+
+ int32 launch_id() const { return launch_id_; }
+
void set_replica_count(int64 replica_count) {
replica_count_ = replica_count;
}
@@ -197,6 +207,9 @@
// Module/graph-level seed handle.
uint64 seed_ = 0;
+ // Program id that identifies a set of program to be launched together.
+ int32 launch_id_ = 0;
+
// The number of replicas (data parallelism) to compile this binary for.
int64 replica_count_ = 1;
diff --git a/tensorflow/compiler/xla/service/hlo_reachability.h b/tensorflow/compiler/xla/service/hlo_reachability.h
index 0b68cc2..1d08933 100644
--- a/tensorflow/compiler/xla/service/hlo_reachability.h
+++ b/tensorflow/compiler/xla/service/hlo_reachability.h
@@ -148,7 +148,7 @@
private:
using Word = uint64;
- static const size_t kBits = 64;
+ static constexpr size_t kBits = 64;
// Number of bits in the bitvector.
size_t size_;
diff --git a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc
index 7fbd01e..0371ce7 100644
--- a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc
+++ b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc
@@ -22,6 +22,7 @@
#include "absl/container/flat_hash_set.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Value.h"
+#include "tensorflow/compiler/xla/map_util.h"
#include "tensorflow/compiler/xla/service/elemental_ir_emitter.h"
#include "tensorflow/compiler/xla/service/hlo_computation.h"
#include "tensorflow/compiler/xla/service/hlo_instruction.h"
@@ -43,9 +44,8 @@
Status FusedIrEmitter::DefaultAction(const HloInstruction* hlo) {
indexed_generators_[hlo] =
[=](const IrArray::Index& index) -> StatusOr<llvm::Value*> {
- if (generated_value_cache_[hlo].contains(index.multidim())) {
- llvm::Value* generated_value =
- generated_value_cache_[hlo][index.multidim()];
+ if (llvm::Value* generated_value = FindOrDefault(
+ generated_value_cache_[hlo], index.multidim(), nullptr)) {
llvm::BasicBlock* generated_value_bb = nullptr;
if (auto* generated_instruction =
llvm::dyn_cast<llvm::Instruction>(generated_value)) {
@@ -71,10 +71,11 @@
<< b_->GetInsertBlock()->getName().str() << ").";
}
- TF_ASSIGN_OR_RETURN(generated_value_cache_[hlo][index.multidim()],
+ TF_ASSIGN_OR_RETURN(llvm::Value* const generated_value,
elemental_emitter_->MakeElementGenerator(
hlo, indexed_generators_)(index));
- return generated_value_cache_[hlo][index.multidim()];
+ generated_value_cache_[hlo][index.multidim()] = generated_value;
+ return generated_value;
};
return Status::OK();
}
diff --git a/tensorflow/compiler/xla/service/memory_space_assignment.cc b/tensorflow/compiler/xla/service/memory_space_assignment.cc
index e5b1756..564a60f 100644
--- a/tensorflow/compiler/xla/service/memory_space_assignment.cc
+++ b/tensorflow/compiler/xla/service/memory_space_assignment.cc
@@ -504,12 +504,12 @@
AddInputAndOutputRequiredAssignments();
- if (VLOG_IS_ON(4)) {
- VLOG(4) << "Flattened instruction sequence:";
+ if (VLOG_IS_ON(3)) {
+ VLOG(3) << "Flattened instruction sequence:";
const auto& instruction_sequence =
hlo_live_range_.flattened_instruction_sequence().instructions();
for (int i = 0; i < instruction_sequence.size(); ++i) {
- VLOG(4) << " " << i << ": " << instruction_sequence[i]->parent()->name()
+ VLOG(3) << " " << i << ": " << instruction_sequence[i]->parent()->name()
<< " " << instruction_sequence[i]->name();
}
}
@@ -545,7 +545,7 @@
}
if (AreIntervalsReservedInAlternateMemory(colocated_intervals)) {
- VLOG(4) << "Interval " << interval.buffer->ToShortString()
+ VLOG(3) << "Interval " << interval.buffer->ToShortString()
<< " is reserved in the alternate memory. Total reserved bytes = "
<< reserved_in_bytes_;
for (const BufferInterval* colocated_interval : colocated_intervals) {
@@ -554,7 +554,7 @@
// alternate memory allocations will not have an entry in preset
// allocations that is normally used for coloring.
for (auto& position : value->positions()) {
- VLOG(3) << "Coloring " << position.ToString();
+ VLOG(4) << "Coloring " << position.ToString();
Shape* shape = ShapeUtil::GetMutableSubshape(
position.instruction->mutable_shape(), position.index);
CHECK(shape->IsArray()) << "Coloring a shape that is not an array: "
@@ -672,6 +672,12 @@
// interval (5-6) can be allocated separately and this buffer
// doesn't waste alternate memory space within the while loop body.
HloComputation* while_body = use.instruction->while_body();
+ // We require while body ROOTs to be the last in the schedule.
+ CHECK_EQ(
+ instruction_schedule.at(while_body->root_instruction()) + 1,
+ instruction_schedule.at(use.instruction))
+ << "While body ROOTs need to be the last in the schedule! "
+ "Please run RootInstructionSinker.";
// Replace the use time with the parameter time so that we can
// decide on alternate memory allocations within the while loop body
// when we look at uses within the while loop body.
@@ -994,7 +1000,7 @@
for (const auto& interval_and_chunk : pending_chunks_) {
const BufferInterval& interval = interval_and_chunk.first;
const Chunk& chunk = interval_and_chunk.second.chunk;
- VLOG(4) << "Uncommitting: (" << interval.start << ", " << interval.end
+ VLOG(3) << "Uncommitting: (" << interval.start << ", " << interval.end
<< ") off = " << chunk.offset << " size = " << chunk.size;
interval_tree_.Remove(interval.start, interval.end, chunk);
}
@@ -1139,7 +1145,7 @@
// If the buffer must be in default memory at the end_time, don't prefetch.
if (required_memory_space_at_end == MemorySpace::kDefault) {
- VLOG(4)
+ VLOG(3)
<< "Not trying to prefetch because use requires buffer in default mem.";
(*prev_allocation_in_default_mem_it)->Extend(request.end_time);
(*prev_allocation_in_default_mem_it)->AddUse(request.use);
@@ -1267,7 +1273,7 @@
preferred_offset = request.preferred_offset;
}
- VLOG(4) << "We can eliminate copy to alternate memory. Preferred offset = "
+ VLOG(3) << "We can eliminate copy to alternate memory. Preferred offset = "
<< (preferred_offset ? *preferred_offset : -1);
// In case there are additional uses after this use, we rely on the last use
// time to try to reserve a chunk in the heap simulator. This is to prevent
@@ -1347,7 +1353,7 @@
eviction_mem_interval.end =
std::min(preferred_eviction_end_time, global_max_time_);
int64 preferred_offset = prev_allocation->chunk().offset;
- VLOG(4) << "Eviction (" << eviction_start_time << ", " << eviction_end_time
+ VLOG(3) << "Eviction (" << eviction_start_time << ", " << eviction_end_time
<< ") preferred end time = " << eviction_mem_interval.end;
for (; eviction_mem_interval.end > eviction_end_time;
@@ -1387,7 +1393,7 @@
// this interval.
bool eviction_scheduled = false;
for (int64 time = eviction_start_time; time < eviction_end_time; ++time) {
- VLOG(3) << "Try evicting (" << time << ", " << time + 1 << ")";
+ VLOG(4) << "Try evicting (" << time << ", " << time + 1 << ")";
if (!ViolatesMaximumOutstandingAsyncCopies(time, time + 1)) {
VLOG(3) << "Eviction successful.";
AddAsyncCopy(*prev_allocation, MemorySpace::kDefault,
@@ -1433,7 +1439,7 @@
options_.prefetch_interval_picker->Begin(
request.use, prev_allocation_in_default_mem.earliest_available_time(),
request.latest_prefetch_time);
- VLOG(4) << "Trying prefetch picker = "
+ VLOG(3) << "Trying prefetch picker = "
<< options_.prefetch_interval_picker->ToDebugString();
// Create an alternate memory interval that starts at the earliest
@@ -1701,9 +1707,9 @@
const HloAliasAnalysis& alias_analysis,
const Options& options) {
CHECK(module->has_schedule());
- VLOG(4) << "Module before memory space assignment: ";
- XLA_VLOG_LINES(4, module->ToString());
- VLOG(4) << "Schedule: " << module->schedule().ToString();
+ VLOG(3) << "Module before memory space assignment: ";
+ XLA_VLOG_LINES(3, module->ToString());
+ VLOG(3) << "Schedule: " << module->schedule().ToString();
MemorySpaceAssignment memory_space_assignment(module, options,
hlo_live_range);
@@ -1713,9 +1719,10 @@
memory_space_assignment.ScheduleAsynchronousCopies();
TF_RETURN_IF_ERROR(memory_space_assignment.SimplifyGraph());
TF_RETURN_IF_ERROR(memory_space_assignment.FixSchedule());
+ TF_RETURN_IF_ERROR(memory_space_assignment.ExportAndColorBuffers());
- VLOG(4) << "Module after memory space assignment: ";
- XLA_VLOG_LINES(4, module->ToString());
+ VLOG(3) << "Module after memory space assignment: ";
+ XLA_VLOG_LINES(3, module->ToString());
TF_CHECK_OK(module->schedule().Verify());
VLOG(1) << "Maximum number of outstanding async copies: "
<< CountMaximumOutstandingAsyncCopies(*module);
@@ -1870,6 +1877,18 @@
return producing_instruction;
}
+std::string MemorySpaceAssignment::Allocation::ToString() const {
+ return absl::StrCat("Allocation in ",
+ memory_space_ == MemorySpace::kDefault ? "def" : "alt",
+ " defined at ", defining_position_.ToString());
+}
+
+std::string MemorySpaceAssignment::CopyAllocation::ToString() const {
+ return absl::StrCat("Copy Allocation in ",
+ memory_space_ == MemorySpace::kDefault ? "def" : "alt",
+ " from ", prev_allocation_.ToString());
+}
+
Status MemorySpaceAssignment::CopyAllocation::Process(
MemorySpaceAssignment* memory_space_assignment) {
// Copy allocations need to insert asynchronous copy nodes.
@@ -1914,25 +1933,29 @@
}
Status MemorySpaceAssignment::Process() {
+ VLOG(1) << "Processing assigned buffers...";
// Insert CopyStart/CopyDone pairs.
- int64 alternate_memory_size = 0;
- std::vector<std::pair<HloPosition, Chunk>> position_and_chunks;
for (auto& allocation : allocations_) {
+ VLOG(3) << "Processing: " << allocation->ToString();
TF_RETURN_IF_ERROR(allocation->Process(this));
// Add the offset and size of the allocation in the alternate memory to
// the output map.
if (allocation->memory_space() == MemorySpace::kAlternate) {
- position_and_chunks.emplace_back(allocation->defining_position(),
- allocation->chunk());
- alternate_memory_size =
- std::max(alternate_memory_size, allocation->chunk().chunk_end());
+ alternate_memory_assignments_.emplace_back(
+ allocation->defining_position(), allocation->chunk());
+ alternate_memory_size_ =
+ std::max(alternate_memory_size_, allocation->chunk().chunk_end());
}
}
+ return Status::OK();
+}
+Status MemorySpaceAssignment::ExportAndColorBuffers() {
+ VLOG(1) << "Exporting buffers...";
TF_ASSIGN_OR_RETURN(auto alias_analysis, HloAliasAnalysis::Run(module_));
absl::flat_hash_map<int64, int64> seen_buffer_offsets;
VLOG(3) << "Exported alternate memory allocations:";
- for (const auto& position_and_chunk : position_and_chunks) {
+ for (const auto& position_and_chunk : alternate_memory_assignments_) {
const HloPosition& defining_position = position_and_chunk.first;
const Chunk& chunk = position_and_chunk.second;
const HloBuffer& buffer = alias_analysis->GetUniqueBufferAt(
@@ -1954,7 +1977,7 @@
if (!preset_assignments_->chunks().empty()) {
preset_assignments_
->assignment_information_for_space(options_.alternate_memory_space)
- ->size = alternate_memory_size;
+ ->size = alternate_memory_size_;
}
VLOG(3) << "Exported alternate memory sizes:";
@@ -1962,6 +1985,7 @@
VLOG(3) << " space: " << pair.first << ", size: " << pair.second.size;
}
+ VLOG(1) << "Coloring buffers...";
// Color the pending positions and all of their aliased buffers.
for (const auto& defining_position_and_chunk :
preset_assignments_->chunks()) {
@@ -1970,7 +1994,7 @@
defining_position.instruction, defining_position.index)) {
for (auto& value : buffer->values()) {
for (auto& position : value->positions()) {
- VLOG(3) << "Coloring " << position.ToString();
+ VLOG(4) << "Coloring " << position.ToString();
Shape* shape = ShapeUtil::GetMutableSubshape(
position.instruction->mutable_shape(), position.index);
CHECK(shape->IsArray()) << "Coloring a shape that is not an array: "
@@ -1981,25 +2005,25 @@
}
}
}
-
return Status::OK();
}
-void PresetAssignments::RemoveAssignmentForInstruction(
+void MemorySpaceAssignment::RemoveAssignmentForInstruction(
const HloInstruction* instruction) {
- for (auto& position_and_chunk : chunks_) {
+ for (auto& position_and_chunk : alternate_memory_assignments_) {
const HloPosition& position = position_and_chunk.first;
if (position.instruction == instruction) {
- VLOG(3) << "Removing instruction from preset assignments.";
+ VLOG(3) << "Removing instruction from alternate memory assignments.";
// Swap the removed position and chunk with the back and pop back.
- position_and_chunk = chunks_.back();
- chunks_.pop_back();
+ position_and_chunk = alternate_memory_assignments_.back();
+ alternate_memory_assignments_.pop_back();
break;
}
}
}
Status MemorySpaceAssignment::SimplifyGraph() {
+ VLOG(1) << "Simplifying graph...";
for (HloComputation* computation : module_->MakeNonfusionComputations()) {
// Parallel computations aren't in the schedule and don't need to be
// modified.
@@ -2034,9 +2058,9 @@
instruction->opcode() != HloOpcode::kCopyStart &&
instruction->opcode() != HloOpcode::kCopyDone) {
VLOG(4) << "Instruction removed: " << instruction->ToString();
- // Ensure the exported preset assignments don't contain a reference to
- // the removed instruction.
- preset_assignments_->RemoveAssignmentForInstruction(instruction);
+ // Ensure the alternate memory assignments don't contain a reference
+ // to the removed instruction.
+ RemoveAssignmentForInstruction(instruction);
// Instead of deleting the instruction from the schedule, replace it
// with a nullptr. This is needed because FixSchedule relies on the
// logical time that is the index into flattened_instructions_ for
@@ -2122,6 +2146,7 @@
}
void MemorySpaceAssignment::ScheduleAsynchronousCopies() {
+ VLOG(1) << "Scheduling asynchronous copies...";
for (MemorySpace memory_space :
{MemorySpace::kDefault, MemorySpace::kAlternate}) {
std::vector<CopyAllocation*> copy_allocations;
@@ -2170,6 +2195,7 @@
}
Status MemorySpaceAssignment::FixSchedule() {
+ VLOG(1) << "Fixing schedule...";
CHECK(module_->has_schedule());
HloSchedule& schedule = module_->schedule();
for (const HloComputation* computation :
@@ -2243,7 +2269,7 @@
}
Status MemorySpaceAssignment::VerifyAndExportHeapSimulatorTrace() {
- VLOG(3) << "Verifying:";
+ VLOG(1) << "Verifying...";
TF_ASSIGN_OR_RETURN(std::unique_ptr<HloAliasAnalysis> alias_analysis,
HloAliasAnalysis::Run(module_));
TF_ASSIGN_OR_RETURN(std::unique_ptr<HloLiveRange> hlo_live_range,
@@ -2348,7 +2374,7 @@
memory_usage -= chunk.size;
}
max_memory_usage = std::max(max_memory_usage, memory_usage);
- VLOG(3) << "Memory usage: " << memory_usage << " at time: " << time;
+ VLOG(4) << "Memory usage: " << memory_usage << " at time: " << time;
}
VLOG(1) << "Max memory usage ignoring fragmentation: " << max_memory_usage;
diff --git a/tensorflow/compiler/xla/service/memory_space_assignment.h b/tensorflow/compiler/xla/service/memory_space_assignment.h
index fcb325f..5572aca 100644
--- a/tensorflow/compiler/xla/service/memory_space_assignment.h
+++ b/tensorflow/compiler/xla/service/memory_space_assignment.h
@@ -63,9 +63,6 @@
return assignment_info_;
}
- // Remove the chunks_ entry that corresponds to instruction.
- void RemoveAssignmentForInstruction(const HloInstruction* instruction);
-
private:
std::vector<std::pair<HloPosition, HeapSimulator::Chunk>> chunks_;
std::vector<std::pair<int64, AssignmentInformation>> assignment_info_;
@@ -398,6 +395,8 @@
int64 start_time() const { return start_time_; }
int64 end_time() const { return end_time_; }
+ virtual std::string ToString() const;
+
protected:
// Descend to the shape_index element of the tuple and replace that with
// new_instruction.
@@ -467,6 +466,8 @@
copy_start_schedule_after_ = copy_start_schedule_after;
}
+ std::string ToString() const override;
+
private:
const Allocation& prev_allocation_;
// These variables define the scheduling boundaries where CopyStart and
@@ -635,6 +636,10 @@
// FixSchedule inserts asynchronous copies in the schedule.
Status FixSchedule();
+ // Export the alternate memory assignments to the PresetAssignments and color
+ // the HLO graph with the determined memory spaces.
+ Status ExportAndColorBuffers();
+
// Insert an instruction to the schedule, and make sure its dependencies
// (operands) are already in the schedule. If not, insert these operands
// before the instruction.
@@ -646,12 +651,18 @@
// corresponding CopyDones follow the same order.
void ScheduleAsynchronousCopies();
+ // Remove the positions and chunks associated with the instruction from
+ // alternate_memory_assignments_.
+ void RemoveAssignmentForInstruction(const HloInstruction* instruction);
+
HloModule* module_;
Options options_;
std::vector<HloInstruction*> flattened_instructions_;
absl::flat_hash_set<const HloComputation*> computations_in_schedule_;
AllocationSequence allocations_;
std::unique_ptr<PresetAssignments> preset_assignments_;
+ std::vector<std::pair<HloPosition, Chunk>> alternate_memory_assignments_;
+ int64 alternate_memory_size_ = 0;
// These maps hold vectors of new instructions that need to be scheduled after
// (or before) the instruction index in the key. FixSchedule uses these maps
diff --git a/tensorflow/compiler/xla/service/memory_space_assignment_test.cc b/tensorflow/compiler/xla/service/memory_space_assignment_test.cc
index 2788dcf..b2125d3 100644
--- a/tensorflow/compiler/xla/service/memory_space_assignment_test.cc
+++ b/tensorflow/compiler/xla/service/memory_space_assignment_test.cc
@@ -737,16 +737,17 @@
// refer to unique positions.
HloComputation::Builder builder(TestName());
Shape shape = ShapeUtil::MakeShape(F32, {2, 3});
+ Shape param_shape = ShapeUtil::MakeShape(F32, {6});
HloInstruction* p0 =
builder.AddInstruction(HloInstruction::CreateParameter(0, shape, "p0"));
- HloInstruction* p1 =
- builder.AddInstruction(HloInstruction::CreateParameter(1, shape, "p1"));
+ HloInstruction* p1 = builder.AddInstruction(
+ HloInstruction::CreateParameter(1, param_shape, "p1"));
HloInstruction* negate = builder.AddInstruction(
HloInstruction::CreateUnary(shape, HloOpcode::kNegate, p0));
- HloInstruction* bitcast =
- builder.AddInstruction(HloInstruction::CreateBitcast(shape, negate));
+ HloInstruction* bitcast = builder.AddInstruction(
+ HloInstruction::CreateBitcast(param_shape, negate));
HloInstruction* add = builder.AddInstruction(
- HloInstruction::CreateBinary(shape, HloOpcode::kAdd, bitcast, p1));
+ HloInstruction::CreateBinary(param_shape, HloOpcode::kAdd, bitcast, p1));
auto module = CreateNewVerifiedModule();
HloComputation* computation = module->AddEntryComputation(builder.Build());
@@ -757,6 +758,8 @@
AssignMemorySpace(module.get());
+ bitcast = add->mutable_operand(0);
+ EXPECT_EQ(bitcast->opcode(), HloOpcode::kBitcast);
EXPECT_EQ(bitcast->shape().layout().memory_space(), kAlternateMemorySpace);
}
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/BUILD b/tensorflow/compiler/xla/service/mlir_gpu/BUILD
index ab58230..cd679f7 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/BUILD
+++ b/tensorflow/compiler/xla/service/mlir_gpu/BUILD
@@ -59,11 +59,26 @@
cc_library(
name = "mlir_compiler",
- srcs = if_cuda_is_configured(["mlir_compiler.cc"]),
- hdrs = if_cuda_is_configured(["mlir_compiler.h"]),
- deps = if_cuda_is_configured([
+ srcs = ["mlir_compiler.cc"],
+ hdrs = ["mlir_compiler.h"],
+ deps = [
":emission_context",
+ "//tensorflow/compiler/xla/service:compiler",
+ "//tensorflow/compiler/xla/service/gpu:target_constants",
+ "//tensorflow/core:stream_executor_no_cuda",
+ "@llvm-project//llvm:core",
+ "@llvm-project//mlir:IR",
+ "@llvm-project//mlir:LLVMDialect",
+ ],
+)
+
+cc_library(
+ name = "mlir_compiler_impl",
+ srcs = if_cuda_is_configured(["mlir_compiler_impl.cc"]),
+ deps = if_cuda_is_configured([
+ ":mlir_compiler",
":failover_compiler",
+ ":emission_context",
":kernel_lowering",
":lhlo_dialect_emitter",
"@com_google_absl//absl/container:flat_hash_map",
@@ -77,7 +92,6 @@
"@llvm-project//mlir:TargetNVVMIR",
"//tensorflow/compiler/xla:util",
"//tensorflow/compiler/xla/service:buffer_assignment",
- "//tensorflow/compiler/xla/service:compiler",
"//tensorflow/compiler/xla/service:dump",
"//tensorflow/compiler/xla/service:hlo",
"//tensorflow/compiler/xla/service/gpu:gpu_constants",
@@ -93,7 +107,6 @@
"//tensorflow/compiler/xla/service/gpu/llvm_gpu_backend",
"//tensorflow/core:cuda_libdevice_path",
"//tensorflow/core:lib",
- "//tensorflow/stream_executor:stream_executor_headers",
"//tensorflow/stream_executor/gpu:asm_compiler",
]),
alwayslink = True, # Contains compiler registration
@@ -186,8 +199,8 @@
cc_library(
name = "xla_gpu_opt_lib",
testonly = True,
- srcs = if_cuda_is_configured(["xla_gpu_opt.cc"]),
- hdrs = if_cuda_is_configured(["xla_gpu_opt.h"]),
+ srcs = ["xla_gpu_opt.cc"],
+ hdrs = ["xla_gpu_opt.h"],
tags = ["no_pip"],
deps = [
":failover_compiler",
@@ -212,7 +225,7 @@
tf_cc_binary(
name = "xla-gpu-opt",
testonly = True,
- srcs = if_cuda_is_configured(["xla_gpu_opt_main.cc"]),
+ srcs = ["xla_gpu_opt_main.cc"],
tags = ["no_pip"],
deps = [
":mlir_compiler",
@@ -222,6 +235,7 @@
"//tensorflow/compiler/xla/service:gpu_plugin_mlir",
"//tensorflow/core:lib",
"@llvm-project//llvm:support",
+ "@llvm-project//mlir:Pass",
"@llvm-project//mlir:Support",
],
)
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/hlo_dialect_emitter.cc b/tensorflow/compiler/xla/service/mlir_gpu/hlo_dialect_emitter.cc
index 0a2c15b..3355027 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/hlo_dialect_emitter.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/hlo_dialect_emitter.cc
@@ -58,6 +58,8 @@
return {func_builder.create<hlo::AndOp>(loc, rets, args, attrs)};
case HloOpcode::kCeil:
return {func_builder.create<hlo::CeilOp>(loc, rets, args, attrs)};
+ case HloOpcode::kComplex:
+ return {func_builder.create<hlo::ComplexOp>(loc, rets, args, attrs)};
case HloOpcode::kCopy:
return {func_builder.create<hlo::CopyOp>(loc, rets, args, attrs)};
case HloOpcode::kCos:
@@ -66,6 +68,8 @@
return {func_builder.create<hlo::DivOp>(loc, rets, args, attrs)};
case HloOpcode::kExp:
return {func_builder.create<hlo::ExpOp>(loc, rets, args, attrs)};
+ case HloOpcode::kImag:
+ return {func_builder.create<hlo::ImagOp>(loc, rets, args, attrs)};
case HloOpcode::kLog:
return {func_builder.create<hlo::LogOp>(loc, rets, args, attrs)};
case HloOpcode::kMaximum:
@@ -76,6 +80,8 @@
return {func_builder.create<hlo::MulOp>(loc, rets, args, attrs)};
case HloOpcode::kNegate:
return {func_builder.create<hlo::NegOp>(loc, rets, args, attrs)};
+ case HloOpcode::kReal:
+ return {func_builder.create<hlo::RealOp>(loc, rets, args, attrs)};
case HloOpcode::kRemainder:
return {func_builder.create<hlo::RemOp>(loc, rets, args, attrs)};
case HloOpcode::kRsqrt:
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
index 056a6bf..33d3690 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.cc
@@ -312,11 +312,8 @@
mlir::FuncOp func = getFunction();
mlir::ModuleOp module = func.getParentOfType<mlir::ModuleOp>();
getFunction().walk([&](mlir::gpu::LaunchFuncOp launchOp) {
- mlir::gpu::GPUModuleOp gpu_module =
- module.lookupSymbol<mlir::gpu::GPUModuleOp>(
- launchOp.getKernelModuleName());
mlir::gpu::GPUFuncOp kernel =
- gpu_module.lookupSymbol<mlir::gpu::GPUFuncOp>(launchOp.kernel());
+ module.lookupSymbol<mlir::gpu::GPUFuncOp>(launchOp.kernel());
// Compute a map from function arguments to kernel function operands.
mlir::BlockAndValueMapping func_to_kernel;
for (mlir::BlockArgument arg : func.getArguments()) {
@@ -331,6 +328,7 @@
// Create a new kernel function with modified signature. We know that it
// will have the same signature as the original function, so just reuse it
// here.
+ auto gpu_module = kernel.getParentOfType<mlir::gpu::GPUModuleOp>();
mlir::OpBuilder kernel_builder(gpu_module.body());
auto new_kernel = kernel_builder.create<mlir::gpu::GPUFuncOp>(
kernel.getLoc(), kernel.getName(), func.getType());
@@ -372,17 +370,6 @@
}
};
-void EnableIRPrinting(mlir::PassManager* passManager) {
- auto enable_if_vlog_is_on = [](mlir::Pass* pass, mlir::Operation* op) {
- return VLOG_IS_ON(1);
- };
- passManager->enableIRPrinting(/*shouldPrintBeforePass=*/enable_if_vlog_is_on,
- /*shouldPrintAfterPass=*/{},
- /*printModuleScope=*/false,
- /*printAfterOnlyOnChange=*/true, llvm::dbgs());
- passManager->disableMultithreading();
-}
-
// Extract_element(xla_hlo_scalars_to_dimension_tensor(v_i), i) -> v_i
//
// We need to direct fusion to the inner loops. This cannot be done with
@@ -432,7 +419,7 @@
llvm::ArrayRef<unsigned> unroll_factors,
bool collapseParallelLoops) {
mlir::PassManager pm(module.getContext());
- EnableIRPrinting(&pm);
+ applyPassManagerCLOptions(pm);
// We have to anticipate later unrolling in tiling to make sure that we get
// the requested tiling after unrolling. Compute the new tiling here if
@@ -547,7 +534,7 @@
Status LowerKernelBodiesToNVVM(mlir::ModuleOp module) {
// We cannot verify as the signature of the kernel is rewritten.
::mlir::PassManager pm(module.getContext(), /*verifyPasses=*/false);
- EnableIRPrinting(&pm);
+ applyPassManagerCLOptions(pm);
// Rewrite kernel functions to LLVM IR.
auto& kernelPm = pm.nest<::mlir::gpu::GPUModuleOp>();
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.cc b/tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.cc
index 3c90d27..6e26d85 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.cc
@@ -77,6 +77,9 @@
case HloOpcode::kCeil:
func_builder.create<lhlo::CeilOp>(loc, rets, args, attrs);
break;
+ case HloOpcode::kComplex:
+ func_builder.create<lhlo::ComplexOp>(loc, rets, args, attrs);
+ break;
case HloOpcode::kCopy:
func_builder.create<lhlo::CopyOp>(loc, rets, args, attrs);
break;
@@ -89,6 +92,9 @@
case HloOpcode::kExp:
func_builder.create<lhlo::ExpOp>(loc, rets, args, attrs);
break;
+ case HloOpcode::kImag:
+ func_builder.create<lhlo::ImagOp>(loc, rets, args, attrs);
+ break;
case HloOpcode::kLog:
func_builder.create<lhlo::LogOp>(loc, rets, args, attrs);
break;
@@ -104,6 +110,9 @@
case HloOpcode::kNegate:
func_builder.create<lhlo::NegOp>(loc, rets, args, attrs);
break;
+ case HloOpcode::kReal:
+ func_builder.create<lhlo::RealOp>(loc, rets, args, attrs);
+ break;
case HloOpcode::kRemainder:
func_builder.create<lhlo::RemOp>(loc, rets, args, attrs);
break;
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.cc b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.cc
index dc33be5..458522f 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.cc
@@ -17,69 +17,18 @@
#include <memory>
-#include "absl/container/flat_hash_map.h"
-#include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVM.h" // from @llvm-project
-#include "mlir/Dialect/GPU/GPUDialect.h" // from @llvm-project
+#include "llvm/IR/Module.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h" // from @llvm-project
-#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
-#include "mlir/IR/Attributes.h" // from @llvm-project
-#include "mlir/IR/Function.h" // from @llvm-project
-#include "mlir/IR/Location.h" // from @llvm-project
#include "mlir/IR/MLIRContext.h" // from @llvm-project
-#include "mlir/IR/Module.h" // from @llvm-project
-#include "mlir/IR/OperationSupport.h" // from @llvm-project
-#include "mlir/IR/StandardTypes.h" // from @llvm-project
-#include "mlir/IR/Value.h" // from @llvm-project
-#include "mlir/Support/LLVM.h" // from @llvm-project
-#include "mlir/Target/NVVMIR.h" // from @llvm-project
-#include "tensorflow/compiler/xla/service/buffer_assignment.h"
-#include "tensorflow/compiler/xla/service/dump.h"
-#include "tensorflow/compiler/xla/service/gpu/gpu_constants.h"
-#include "tensorflow/compiler/xla/service/gpu/gpu_executable.h"
-#include "tensorflow/compiler/xla/service/gpu/gpu_hlo_schedule.h"
-#include "tensorflow/compiler/xla/service/gpu/gpu_types.h"
-#include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h"
-#include "tensorflow/compiler/xla/service/gpu/kernel_thunk.h"
-#include "tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h"
-#include "tensorflow/compiler/xla/service/gpu/nvptx_compiler.h"
-#include "tensorflow/compiler/xla/service/gpu/partition_assignment.h"
-#include "tensorflow/compiler/xla/service/gpu/stream_assignment.h"
-#include "tensorflow/compiler/xla/service/gpu/stream_executor_util.h"
#include "tensorflow/compiler/xla/service/gpu/target_constants.h"
-#include "tensorflow/compiler/xla/service/gpu/thunk_schedule.h"
-#include "tensorflow/compiler/xla/service/hlo_instruction.h"
-#include "tensorflow/compiler/xla/service/hlo_opcode.h"
-#include "tensorflow/compiler/xla/service/mlir_gpu/emission_context.h"
-#include "tensorflow/compiler/xla/service/mlir_gpu/failover_compiler.h"
-#include "tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h"
-#include "tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.h"
-#include "tensorflow/compiler/xla/util.h"
-#include "tensorflow/core/lib/core/errors.h"
-#include "tensorflow/core/lib/io/path.h"
-#include "tensorflow/core/platform/cuda_libdevice_path.h"
-#include "tensorflow/stream_executor/gpu/asm_compiler.h"
+#include "tensorflow/core/platform/stream_executor_no_cuda.h"
namespace xla {
namespace mlir_gpu {
namespace {
-using ::mlir::BlockArgument;
-using ::mlir::dyn_cast;
-using ::mlir::FuncOp;
using ::mlir::MLIRContext;
-using ::mlir::ModuleOp;
-using ::mlir::OwningModuleRef;
-using ::mlir::UnknownLoc;
-using ::mlir::Value;
-using ::mlir::gpu::LaunchFuncOp;
using ::mlir::LLVM::LLVMDialect;
-using ::mlir::LLVM::LLVMFuncOp;
-using ::mlir::LLVM::LLVMType;
-using ::xla::gpu::GpuExecutable;
-using ::xla::gpu::GpuHloSchedule;
-using ::xla::gpu::GpuVersion;
-using ::xla::gpu::StreamAssignment;
-using ::xla::gpu::ThunkSchedule;
int64 ConfigureLLVMModuleAndGetPointerSize(MLIRContext* context) {
LLVMDialect* dialect = context->getRegisteredDialect<LLVMDialect>();
@@ -89,49 +38,6 @@
return module.getDataLayout().getPointerSize();
}
-// TODO(b/137624192) Share with NVPTX compiler
-static std::vector<std::string> CandidateCudaRoots(
- const HloModuleConfig& config) {
- return tensorflow::CandidateCudaRoots(
- config.debug_options().xla_gpu_cuda_data_dir());
-}
-
-void PrintCantFindCudaMessage(absl::string_view msg,
- const HloModuleConfig& hlo_module_config) {
- LOG(WARNING) << msg;
- LOG(WARNING) << "Searched for CUDA in the following directories:";
-
- for (const auto& dir : CandidateCudaRoots(hlo_module_config)) {
- LOG(WARNING) << " " << dir;
- }
- LOG(WARNING)
- << "You can choose the search directory by setting xla_gpu_cuda_data_dir "
- "in HloModule's DebugOptions. For most apps, setting the environment "
- "variable XLA_FLAGS=--xla_gpu_cuda_data_dir=/path/to/cuda will work.";
-}
-
-// Returns the directory containing nvvm libdevice files.
-string GetLibdeviceDir(const HloModuleConfig& hlo_module_config) {
- for (const string& cuda_root : CandidateCudaRoots(hlo_module_config)) {
- const string libdevice_dir =
- tensorflow::io::JoinPath(cuda_root, "nvvm", "libdevice");
- VLOG(2) << "Looking for libdevice at " << libdevice_dir;
- if (tensorflow::Env::Default()->IsDirectory(libdevice_dir).ok()) {
- VLOG(2) << "Found libdevice dir " << libdevice_dir;
- return libdevice_dir;
- }
- }
- PrintCantFindCudaMessage(
- "Can't find libdevice directory ${CUDA_DIR}/nvvm/libdevice. This may "
- "result in compilation or runtime failures, if the program we try to run "
- "uses routines from libdevice.",
- hlo_module_config);
-
- // GetCudaRootCandidates always includes ".", but if everything fails, we
- // return it anyway. Better than returning the empty string.
- return ".";
-}
-
} // namespace
MlirCompiler::MlirCompiler()
@@ -141,428 +47,6 @@
return stream_executor::cuda::kCudaPlatformId;
}
-StatusOr<std::unique_ptr<HloModule>> MlirCompiler::RunHloPasses(
- std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
- se::DeviceMemoryAllocator* device_allocator) {
- // Until we find a reason to do something different, run the same passes
- // that the normal GPU backend runs.
- gpu::NVPTXCompiler xla_compiler;
- TF_RETURN_IF_ERROR(xla_compiler.OptimizeHloModule(module.get(), stream_exec,
- device_allocator));
- TF_RETURN_IF_ERROR(xla_compiler.PrepareHloModuleForIrEmitting(module.get()));
-
- return std::move(module);
-}
-
-namespace {
-
-// TODO(b/137624192): Move this to custom call handling and share.
-absl::optional<bool> CanShareBufferHint(const HloInstruction* user,
- const HloInstruction* operand,
- const ShapeIndex& user_index) {
- if (user->opcode() == HloOpcode::kCustomCall) {
- // Share the bias buffer with the parent instruction.
- if (user->custom_call_target() == xla::gpu::kGemmCallTarget) {
- if (user->operand_count() == 3 && user->operand(2) == operand) {
- return true;
- }
- }
- // The operand of cholesky can be shared with the first output.
- if (user->custom_call_target() == xla::gpu::kCusolverCholeskyCallTarget) {
- return user_index.size() == 1 && user_index[0] == 0;
- }
- }
- return absl::nullopt;
-}
-
-// TODO(b/137624192): Share this with nvptx backend.
-GpuVersion GetGpuVersion(se::StreamExecutor* stream_exec) {
- int cc_major, cc_minor;
- const auto& device_description = stream_exec->GetDeviceDescription();
- if (!device_description.cuda_compute_capability(&cc_major, &cc_minor)) {
- LOG(WARNING)
- << "Couldn't get compute capability for device; assuming sm_20.";
- cc_major = 2;
- cc_minor = 0;
- }
- return std::make_pair(cc_major, cc_minor);
-}
-
-// Return the constant launch bound along the "x" dimension in "dim" if all the
-// other dimensions are 1. Return nullopt otherwise or when any of the bounds
-// is not constant.
-static absl::optional<int64> getLaunchBound(const mlir::gpu::KernelDim3& dim) {
- auto get_constant = [](mlir::Operation* op,
- mlir::StringRef name) -> absl::optional<int64> {
- if (auto constant = llvm::dyn_cast_or_null<mlir::ConstantOp>(op)) {
- return constant.value().cast<mlir::IntegerAttr>().getInt();
- }
- op->emitError() << "bound " << name << " is not constant";
- return absl::nullopt;
- };
- auto y_op = dim.y.getDefiningOp();
- auto dim_y = get_constant(y_op, "y");
- if (!dim_y.has_value() || dim_y.value() != 1) {
- y_op->emitError() << "bound 'y' is not constant 1";
- return absl::nullopt;
- }
- auto z_op = dim.z.getDefiningOp();
- auto dim_z = get_constant(z_op, "z");
- if (!dim_z.has_value() || dim_z.value() != 1) {
- z_op->emitError() << "bound 'z' is not constant 1";
- return absl::nullopt;
- }
- return get_constant(dim.x.getDefiningOp(), "x");
-}
-
-namespace {
-
-// Indexes of a range of arguments in a GPU function. This is used to keep the
-// range of arguments that correspond to a lowered kernel argument of
-// (previously) memref type.
-struct LaunchFuncArgument {
- int kernel_argument_begin;
- int kernel_argument_size;
-};
-
-} // end namespace
-
-using OperandToValueMap =
- absl::flat_hash_map<const HloInstruction*, std::vector<LaunchFuncArgument>>;
-
-static StatusOr<std::vector<const HloInstruction*>> ComputeOperandToValueMap(
- OperandToValueMap* operand_to_value_map, const HloInstruction* instr,
- LaunchFuncOp launchOp, LLVMFuncOp kernel) {
- auto operands = instr->operands();
- std::vector<const HloInstruction*> ordered_operands;
- bool has_failed = false;
- // A memref will expand into multiple kernel operands, accumulate their number
- // in order to find them later.
- int cur_operand_position = 0;
-
- for (int kernel_index = 0; kernel_index < launchOp.getNumKernelOperands();
- ++kernel_index) {
- auto launchop_operand =
- launchOp.getKernelOperand(kernel_index).dyn_cast<BlockArgument>();
- if (!launchop_operand) {
- launchOp.emitError("argument to kernel is not a function input");
- has_failed = true;
- continue;
- }
- auto memref_type =
- launchop_operand.getType().dyn_cast<::mlir::MemRefType>();
- if (!memref_type) {
- launchOp.emitError("only memref-typed arguments are supported");
- has_failed = true;
- break;
- }
- // host_index is the argument position to the surrounding function that
- // contains the launch. This index corresponds to HLO operand indices
- // by construction.
- auto host_index = launchop_operand.getArgNumber();
- // The trailing argument to the outer function are the results.
- auto operand =
- (host_index < operands.size()) ? operands[host_index] : instr;
- if (!operand_to_value_map->count(operand)) {
- ordered_operands.push_back(operand);
- }
- // Associate the HLO operand with the argument values of the kernel
- // function.
- int num_unpacked =
- mlir::MemRefDescriptor::getNumUnpackedValues(memref_type);
- (*operand_to_value_map)[operand].push_back(
- {cur_operand_position, num_unpacked});
- cur_operand_position += num_unpacked;
- }
- if (has_failed) {
- return InternalError("Mapping operands to kernel arguments has failed.");
- }
- return ordered_operands;
-}
-
-Status InsertBufferLoadPreduleIntoKernel(
- LLVMFuncOp kernel, const OperandToValueMap& operand_to_value_map,
- const std::vector<const HloInstruction*>& ordered_operands,
- BufferAssignment* assignment,
- const std::vector<const BufferAllocation*>& buffers) {
- mlir::OpBuilder builder(kernel.getBody());
- auto llvm_dialect = kernel.getContext()->getRegisteredDialect<LLVMDialect>();
- auto offset_type = LLVMType::getInt64Ty(llvm_dialect);
- auto ptr_type = LLVMType::getInt8PtrTy(llvm_dialect);
- auto void_type = LLVMType::getVoidTy(llvm_dialect);
- auto loc = kernel.getLoc();
-
- auto num_original_args = kernel.getNumArguments();
- std::vector<LLVMType> new_arg_types(buffers.size(), ptr_type);
- kernel.setAttr(kernel.getTypeAttrName(),
- mlir::TypeAttr::get(LLVMType::getFunctionTy(
- void_type, new_arg_types, /*isVarArg=*/false)));
- std::vector<Value> original_args(kernel.args_begin(), kernel.args_end());
-
- std::vector<mlir::Type> as_mlir_types(new_arg_types.begin(),
- new_arg_types.end());
- auto new_args = kernel.front().addArguments(as_mlir_types);
- std::vector<Value> buffer_args(new_args.begin(), new_args.end());
-
- for (auto operand : ordered_operands) {
- TF_ASSIGN_OR_RETURN(auto slice,
- assignment->GetUniqueTopLevelSlice(operand));
- auto buffer = std::find(buffers.begin(), buffers.end(), slice.allocation());
- auto index = buffer - buffers.begin();
- auto offset = builder.create<mlir::LLVM::ConstantOp>(
- loc, offset_type, builder.getI64IntegerAttr(slice.offset()));
- auto ptr = buffer_args[index];
-
- // Replace uses of function arguments pertaining to memref descriptors with
- // values derived from HLO buffers. The instructions inserting these values
- // into memref descriptors were already introduced during the lowering phase
- // as per MLIR calling convention.
- for (auto arg : operand_to_value_map.at(operand)) {
- mlir::MemRefDescriptorView original(
- mlir::ValueRange(original_args)
- .slice(arg.kernel_argument_begin, arg.kernel_argument_size));
-
- // Allocated and aligned pointers are the same.
- auto casted = builder.create<mlir::LLVM::BitcastOp>(
- loc, original.alignedPtr().getType().cast<LLVMType>(),
- mlir::ValueRange(ptr));
- original.alignedPtr().replaceAllUsesWith(casted);
- original.allocatedPtr().replaceAllUsesWith(casted);
-
- // Use the offset of the HLO buffer instead of the one expected in the
- // function call.
- original.offset().replaceAllUsesWith(offset);
-
- // Fill the shape.
- auto shape = operand->shape();
- // Unless the operand is a scalar pointer, also fill shape and strides.
- if (shape.dimensions().empty()) {
- continue;
- }
-
- // TODO(b/137624192) Pass in the descriptor to allow for dynamic shapes.
- assert(shape.IsArray() && shape.is_static());
- for (auto extent : llvm::enumerate(shape.dimensions())) {
- auto shape = builder.create<mlir::LLVM::ConstantOp>(
- loc, original.size(extent.index()).getType(),
- builder.getI64IntegerAttr(extent.value()));
- original.size(extent.index()).replaceAllUsesWith(shape);
- }
- // Finally, fill the strides.
- // TODO(b/137624192): Take assigned layout into account.
- uint64_t accumulator = 0;
- for (int64_t idx = shape.rank() - 1; idx >= 0; --idx) {
- if (accumulator == 0) {
- accumulator = 1;
- } else {
- accumulator *= shape.dimensions(idx + 1);
- }
- auto stride = builder.create<mlir::LLVM::ConstantOp>(
- loc, original.stride(idx).getType(),
- builder.getI64IntegerAttr(accumulator));
- original.stride(idx).replaceAllUsesWith(stride);
- }
- }
- }
-
- // Now we can remove the original arguments, as they should have no more
- // users.
- for (int i = 0; i < num_original_args; ++i) {
- kernel.front().eraseArgument(0);
- }
-
- return Status::OK();
-}
-
-StatusOr<std::unique_ptr<gpu::KernelThunk>> TransformKernelToXlaThunk(
- FuncOp func, const HloInstruction* const instr, ModuleOp kernel_module,
- BufferAssignment* assignment) {
- // Find the single LaunchFuncOp and compute a mapping from operands of
- // the hlo instruction to the corresponding values of the kernel
- // function in the target module;
- LaunchFuncOp launchOp;
- auto walkResult = func.walk([&launchOp](LaunchFuncOp op) {
- if (launchOp) {
- op.emitError("multiple kernels for single top-level HLO");
- return mlir::WalkResult::interrupt();
- }
- launchOp = op;
- return mlir::WalkResult::advance();
- });
- if (walkResult.wasInterrupted()) {
- return InternalError("Multiple kernels for single top-level HLO");
- }
- if (!launchOp) {
- // If there was no launchOp, then no kernel was generated, so the lowering
- // from the LHLO ops to the GPU dialect is not implemented yet.
- return Unimplemented("No kernel was generated.");
- }
-
- auto kernel = kernel_module.lookupSymbol<LLVMFuncOp>(launchOp.kernel());
-
- // Store the assignment of operands to block arguments. Note that an operand
- // might be used in multiple argument positions, hence the vector.
- OperandToValueMap operand_to_value_map;
- TF_ASSIGN_OR_RETURN(
- auto ordered_operands,
- ComputeOperandToValueMap(&operand_to_value_map, instr, launchOp, kernel));
-
- // Get the required buffers to support the inputs. Use a set and vector here
- // to keep the order fixed. This is mostly useful for testing.
- std::unordered_set<const BufferAllocation*> buffers_needed;
- std::vector<const BufferAllocation*> buffers;
- // TODO(b/137624192) Add support for tuples.
- for (auto operand : ordered_operands) {
- TF_ASSIGN_OR_RETURN(auto buffer,
- assignment->GetUniqueTopLevelSlice(operand));
- if (buffers_needed.insert(buffer.allocation()).second) {
- buffers.push_back(buffer.allocation());
- }
- }
-
- // TODO(b/137624192) Add support for temp buffer.
- // TODO(b/137624192) Add support for constant buffers.
-
- // Change the signature to match what the XLA runtime expects from the
- // kernel.
- TF_RETURN_IF_ERROR(InsertBufferLoadPreduleIntoKernel(
- kernel, operand_to_value_map, ordered_operands, assignment, buffers));
-
- // Finally, create the thunk and set the launch dimensions.
- auto thunk = absl::make_unique<gpu::KernelThunk>(
- buffers, kernel.getName().str(), instr,
- /*unroll_factor=*/1);
-
- // Set launch bounds.
- mlir::gpu::KernelDim3 block = launchOp.getBlockSizeOperandValues();
- mlir::gpu::KernelDim3 grid = launchOp.getGridSizeOperandValues();
- absl::optional<int64> num_threads = getLaunchBound(block);
- absl::optional<int64> num_blocks = getLaunchBound(grid);
- if (!num_threads || !num_blocks) {
- return Unimplemented("Unsupported launch bounds");
- }
- thunk->SetLaunchDimensions(gpu::LaunchDimensions(*num_blocks, *num_threads));
- return std::move(thunk);
-}
-
-} // namespace
-
-StatusOr<std::unique_ptr<Executable>> MlirCompiler::RunBackend(
- std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
- se::DeviceMemoryAllocator* device_allocator) {
- // Determine the HLO schedule, which is an ordering of HLO instructions. This
- // is used by buffer assignment to enable buffer reuse, and the same ordering
- // must also be used to determine the thunk launch schedule.
- std::unique_ptr<StreamAssignment> stream_assignment =
- xla::gpu::AssignStreams(*module);
- TF_ASSIGN_OR_RETURN(
- std::unique_ptr<GpuHloSchedule> hlo_schedule,
- GpuHloSchedule::Build(*module, *stream_assignment, pointer_size_));
-
- // Run buffer analysis on the HLO graph. This analysis figures out which
- // temporary buffers are required to run the computation.
- TF_ASSIGN_OR_RETURN(std::unique_ptr<BufferAssignment> buffer_assignment,
- BufferAssigner::Run(
- module.get(), hlo_schedule->ConsumeHloOrdering(),
- BufferSizeBytesFunction(),
- /*color_alignment=*/
- [](LogicalBuffer::Color) {
- return xla::gpu::kXlaAllocatedBufferAlignBytes;
- },
- /*allocate_buffers_for_constants=*/true,
- /*colorer=*/BufferAssigner::DefaultColorer(),
- /*must_not_live_out=*/{}, &CanShareBufferHint));
- DumpHloModuleIfEnabled(*module, *buffer_assignment, "after_optimizations");
-
- EmissionContext emission_context(std::move(module));
- if (error_handler_) {
- emission_context.setErrorHandler(error_handler_);
- }
-
- OwningModuleRef mlir_module =
- ModuleOp::create(UnknownLoc::get(emission_context.getContext()));
- LhloDialectEmitter lhlo_emitter(&emission_context, *buffer_assignment,
- stream_exec->platform(), *mlir_module);
-
- TF_RETURN_IF_ERROR(lhlo_emitter.EmitComputation(
- *emission_context.getHloModule()->entry_computation()));
-
- TF_RETURN_IF_ERROR(
- module_hook_.invoke(IRHook::LoweringStage::LHLO, *mlir_module));
-
- TF_RETURN_IF_ERROR(LowerLHLOToGPU(*mlir_module));
-
- TF_RETURN_IF_ERROR(
- module_hook_.invoke(IRHook::LoweringStage::GPU, *mlir_module));
-
- TF_RETURN_IF_ERROR(LowerKernelBodiesToNVVM(*mlir_module));
-
- TF_RETURN_IF_ERROR(
- module_hook_.invoke(IRHook::LoweringStage::LLVM, *mlir_module));
-
- TF_ASSIGN_OR_RETURN(OwningModuleRef kernel_module,
- ExtractKernelModule(*mlir_module));
-
- auto thunk_sequence = lhlo_emitter.ConsumeThunkSequence();
- for (auto entry : lhlo_emitter.InstructionToFunctionMap()) {
- TF_ASSIGN_OR_RETURN(
- auto thunk,
- TransformKernelToXlaThunk(entry.second, entry.first, *kernel_module,
- buffer_assignment.get()));
- thunk_sequence->push_back(std::move(thunk));
- }
-
- TF_RETURN_IF_ERROR(
- module_hook_.invoke(IRHook::LoweringStage::KERNEL, *kernel_module));
-
- auto llvmModule = mlir::translateModuleToNVVMIR(*kernel_module);
-
- if (!llvmModule) {
- return InternalError("Translation to LLVM failed");
- }
-
- llvmModule->setModuleIdentifier(emission_context.getHloModule()->name());
- // TODO(herhut): Why is this needed and does not come from the template?
- llvmModule->setDataLayout(gpu::nvptx::kDataLayout);
-
- const auto& config = emission_context.getHloModule()->config();
- TF_ASSIGN_OR_RETURN(
- auto ptx, xla::gpu::nvptx::CompileToPtx(llvmModule.get(),
- GetGpuVersion(stream_exec),
- config, GetLibdeviceDir(config)));
- TF_ASSIGN_OR_RETURN(
- auto cubin, se::CompileGpuAsm(stream_exec->device_ordinal(), ptx.c_str(),
- gpu::PtxOptsFromConfig(config)));
-
- auto thunk_schedule = absl::make_unique<ThunkSchedule>(
- std::move(thunk_sequence), std::move(stream_assignment),
- hlo_schedule->ThunkLaunchOrder());
-
- if (DumpingEnabledForHloModule(*emission_context.getHloModule())) {
- DumpToFileInDirOrStdout(*emission_context.getHloModule(), "",
- "thunk_schedule", thunk_schedule->ToString());
- }
-
- // TODO(b/137624192): Add profiling support.
- return {absl::make_unique<GpuExecutable>(
- ptx, cubin, GetGpuVersion(stream_exec), std::move(thunk_schedule),
- emission_context.releaseHloModule(), std::move(buffer_assignment),
- nullptr, nullptr)};
-}
-
-StatusOr<std::vector<std::unique_ptr<Executable>>> MlirCompiler::Compile(
- std::unique_ptr<HloModuleGroup> module_group,
- std::vector<std::vector<se::StreamExecutor*>> stream_execs,
- se::DeviceMemoryAllocator* device_allocator) {
- return Unimplemented("Not yet implemented in MLIR compiler");
-}
-
-StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
-MlirCompiler::CompileAheadOfTime(std::unique_ptr<HloModuleGroup> module_group,
- const AotCompilationOptions& options) {
- return Unimplemented("Not yet implemented in MLIR compiler");
-}
-
void MlirCompiler::SetModuleHook(IRHook module_hook) {
module_hook_ = module_hook;
}
@@ -579,14 +63,3 @@
} // namespace mlir_gpu
} // namespace xla
-
-static bool InitModule() {
- xla::Compiler::RegisterCompilerFactory(
- stream_executor::cuda::kCudaPlatformId, []() {
- return absl::make_unique<xla::FailoverCompiler>(
- absl::make_unique<xla::mlir_gpu::MlirCompiler>(),
- absl::make_unique<xla::gpu::NVPTXCompiler>());
- });
- return true;
-}
-static bool module_initialized = InitModule();
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h
index 9aeef12..a7b2f94 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h
+++ b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h
@@ -16,7 +16,6 @@
#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_MLIR_GPU_MLIR_COMPILER_H_
#define TENSORFLOW_COMPILER_XLA_SERVICE_MLIR_GPU_MLIR_COMPILER_H_
-#include "absl/container/flat_hash_map.h"
#include "mlir/IR/MLIRContext.h" // from @llvm-project
#include "mlir/IR/Module.h" // from @llvm-project
#include "tensorflow/compiler/xla/service/compiler.h"
@@ -27,7 +26,8 @@
// A Compiler implementation that converts XLAs IR to a matching MLIR dialect,
// performs all lowering on the MLIR IR and finally converts MLIR to LLVMIR for
-// generation of a think suitable for XLAs runtime.
+// generation of a thunk suitable for XLAs runtime. MlirCompilerImpl contains
+// the implementation.
class MlirCompiler : public Compiler {
using ErrorHandler =
std::function<void(const EmissionContext::ErrorMap&, HloModule*)>;
@@ -37,30 +37,6 @@
se::Platform::Id PlatformId() const override;
- StatusOr<std::unique_ptr<HloModule>> RunHloPasses(
- std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
- se::DeviceMemoryAllocator* device_allocator) override;
-
- StatusOr<std::unique_ptr<Executable>> RunBackend(
- std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
- se::DeviceMemoryAllocator* device_allocator) override;
-
- StatusOr<std::vector<std::unique_ptr<Executable>>> Compile(
- std::unique_ptr<HloModuleGroup> module_group,
- std::vector<std::vector<se::StreamExecutor*>> stream_execs,
- se::DeviceMemoryAllocator* device_allocator) override;
-
- StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
- CompileAheadOfTime(std::unique_ptr<HloModuleGroup> module_group,
- const AotCompilationOptions& options) override;
-
- HloCostAnalysis::ShapeSizeFunction ShapeSizeBytesFunction() const override {
- int64 pointer_size = pointer_size_;
- return [pointer_size](const Shape& shape) {
- return ShapeUtil::ByteSizeOf(shape, pointer_size);
- };
- }
-
struct IRHook {
enum class LoweringStage { LHLO, GPU, LLVM, KERNEL };
@@ -80,7 +56,7 @@
void SetErrorHandler(ErrorHandler error_handler);
void RemoveErrorHandler();
- private:
+ protected:
::mlir::MLIRContext context_;
int64 pointer_size_;
IRHook module_hook_;
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler_impl.cc b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler_impl.cc
new file mode 100644
index 0000000..35ac3b2
--- /dev/null
+++ b/tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler_impl.cc
@@ -0,0 +1,585 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "absl/container/flat_hash_map.h"
+#include "mlir/Conversion/StandardToLLVM/ConvertStandardToLLVM.h" // from @llvm-project
+#include "mlir/Dialect/GPU/GPUDialect.h" // from @llvm-project
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h" // from @llvm-project
+#include "mlir/Dialect/StandardOps/IR/Ops.h" // from @llvm-project
+#include "mlir/IR/Attributes.h" // from @llvm-project
+#include "mlir/IR/Function.h" // from @llvm-project
+#include "mlir/IR/Location.h" // from @llvm-project
+#include "mlir/IR/Module.h" // from @llvm-project
+#include "mlir/IR/OperationSupport.h" // from @llvm-project
+#include "mlir/IR/StandardTypes.h" // from @llvm-project
+#include "mlir/IR/Value.h" // from @llvm-project
+#include "mlir/Support/LLVM.h" // from @llvm-project
+#include "mlir/Target/NVVMIR.h" // from @llvm-project
+#include "tensorflow/compiler/xla/service/buffer_assignment.h"
+#include "tensorflow/compiler/xla/service/dump.h"
+#include "tensorflow/compiler/xla/service/gpu/gpu_constants.h"
+#include "tensorflow/compiler/xla/service/gpu/gpu_executable.h"
+#include "tensorflow/compiler/xla/service/gpu/gpu_hlo_schedule.h"
+#include "tensorflow/compiler/xla/service/gpu/gpu_types.h"
+#include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h"
+#include "tensorflow/compiler/xla/service/gpu/kernel_thunk.h"
+#include "tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h"
+#include "tensorflow/compiler/xla/service/gpu/nvptx_compiler.h"
+#include "tensorflow/compiler/xla/service/gpu/partition_assignment.h"
+#include "tensorflow/compiler/xla/service/gpu/stream_assignment.h"
+#include "tensorflow/compiler/xla/service/gpu/stream_executor_util.h"
+#include "tensorflow/compiler/xla/service/gpu/target_constants.h"
+#include "tensorflow/compiler/xla/service/gpu/thunk_schedule.h"
+#include "tensorflow/compiler/xla/service/hlo_instruction.h"
+#include "tensorflow/compiler/xla/service/hlo_opcode.h"
+#include "tensorflow/compiler/xla/service/mlir_gpu/emission_context.h"
+#include "tensorflow/compiler/xla/service/mlir_gpu/failover_compiler.h"
+#include "tensorflow/compiler/xla/service/mlir_gpu/kernel_lowering.h"
+#include "tensorflow/compiler/xla/service/mlir_gpu/lhlo_dialect_emitter.h"
+#include "tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h"
+#include "tensorflow/compiler/xla/util.h"
+#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/lib/io/path.h"
+#include "tensorflow/core/platform/cuda_libdevice_path.h"
+#include "tensorflow/stream_executor/gpu/asm_compiler.h"
+
+namespace xla {
+namespace mlir_gpu {
+namespace {
+
+using ::mlir::BlockArgument;
+using ::mlir::dyn_cast;
+using ::mlir::FuncOp;
+using ::mlir::ModuleOp;
+using ::mlir::OwningModuleRef;
+using ::mlir::UnknownLoc;
+using ::mlir::Value;
+using ::mlir::gpu::LaunchFuncOp;
+using ::mlir::LLVM::LLVMDialect;
+using ::mlir::LLVM::LLVMFuncOp;
+using ::mlir::LLVM::LLVMType;
+using ::xla::gpu::GpuExecutable;
+using ::xla::gpu::GpuHloSchedule;
+using ::xla::gpu::GpuVersion;
+using ::xla::gpu::StreamAssignment;
+using ::xla::gpu::ThunkSchedule;
+
+// A Compiler implementation that converts XLAs IR to a matching MLIR dialect,
+// performs all lowering on the MLIR IR and finally converts MLIR to LLVMIR for
+// generation of a thunk suitable for XLAs runtime.
+class MlirCompilerImpl : public MlirCompiler {
+ public:
+ StatusOr<std::unique_ptr<HloModule>> RunHloPasses(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
+ se::DeviceMemoryAllocator* device_allocator) override;
+
+ StatusOr<std::unique_ptr<Executable>> RunBackend(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
+ se::DeviceMemoryAllocator* device_allocator) override;
+
+ StatusOr<std::vector<std::unique_ptr<Executable>>> Compile(
+ std::unique_ptr<HloModuleGroup> module_group,
+ std::vector<std::vector<se::StreamExecutor*>> stream_execs,
+ se::DeviceMemoryAllocator* device_allocator) override;
+
+ StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
+ CompileAheadOfTime(std::unique_ptr<HloModuleGroup> module_group,
+ const AotCompilationOptions& options) override;
+
+ HloCostAnalysis::ShapeSizeFunction ShapeSizeBytesFunction() const override {
+ int64 pointer_size = pointer_size_;
+ return [pointer_size](const Shape& shape) {
+ return ShapeUtil::ByteSizeOf(shape, pointer_size);
+ };
+ }
+};
+
+// TODO(b/137624192) Share with NVPTX compiler
+static std::vector<std::string> CandidateCudaRoots(
+ const HloModuleConfig& config) {
+ return tensorflow::CandidateCudaRoots(
+ config.debug_options().xla_gpu_cuda_data_dir());
+}
+
+void PrintCantFindCudaMessage(absl::string_view msg,
+ const HloModuleConfig& hlo_module_config) {
+ LOG(WARNING) << msg;
+ LOG(WARNING) << "Searched for CUDA in the following directories:";
+
+ for (const auto& dir : CandidateCudaRoots(hlo_module_config)) {
+ LOG(WARNING) << " " << dir;
+ }
+ LOG(WARNING)
+ << "You can choose the search directory by setting xla_gpu_cuda_data_dir "
+ "in HloModule's DebugOptions. For most apps, setting the environment "
+ "variable XLA_FLAGS=--xla_gpu_cuda_data_dir=/path/to/cuda will work.";
+}
+
+// Returns the directory containing nvvm libdevice files.
+std::string GetLibdeviceDir(const HloModuleConfig& hlo_module_config) {
+ for (const string& cuda_root : CandidateCudaRoots(hlo_module_config)) {
+ const std::string libdevice_dir =
+ tensorflow::io::JoinPath(cuda_root, "nvvm", "libdevice");
+ VLOG(2) << "Looking for libdevice at " << libdevice_dir;
+ if (tensorflow::Env::Default()->IsDirectory(libdevice_dir).ok()) {
+ VLOG(2) << "Found libdevice dir " << libdevice_dir;
+ return libdevice_dir;
+ }
+ }
+ PrintCantFindCudaMessage(
+ "Can't find libdevice directory ${CUDA_DIR}/nvvm/libdevice. This may "
+ "result in compilation or runtime failures, if the program we try to run "
+ "uses routines from libdevice.",
+ hlo_module_config);
+
+ // GetCudaRootCandidates always includes ".", but if everything fails, we
+ // return it anyway. Better than returning the empty string.
+ return ".";
+}
+
+StatusOr<std::unique_ptr<HloModule>> MlirCompilerImpl::RunHloPasses(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
+ se::DeviceMemoryAllocator* device_allocator) {
+ // Until we find a reason to do something different, run the same passes
+ // that the normal GPU backend runs.
+ gpu::NVPTXCompiler xla_compiler;
+ TF_RETURN_IF_ERROR(xla_compiler.OptimizeHloModule(module.get(), stream_exec,
+ device_allocator));
+ TF_RETURN_IF_ERROR(xla_compiler.PrepareHloModuleForIrEmitting(module.get()));
+
+ return std::move(module);
+}
+
+// TODO(b/137624192): Move this to custom call handling and share.
+absl::optional<bool> CanShareBufferHint(const HloInstruction* user,
+ const HloInstruction* operand,
+ const ShapeIndex& user_index) {
+ if (user->opcode() == HloOpcode::kCustomCall) {
+ // Share the bias buffer with the parent instruction.
+ if (user->custom_call_target() == xla::gpu::kGemmCallTarget) {
+ if (user->operand_count() == 3 && user->operand(2) == operand) {
+ return true;
+ }
+ }
+ // The operand of cholesky can be shared with the first output.
+ if (user->custom_call_target() == xla::gpu::kCusolverCholeskyCallTarget) {
+ return user_index.size() == 1 && user_index[0] == 0;
+ }
+ }
+ return absl::nullopt;
+}
+
+// TODO(b/137624192): Share this with nvptx backend.
+GpuVersion GetGpuVersion(se::StreamExecutor* stream_exec) {
+ int cc_major, cc_minor;
+ const auto& device_description = stream_exec->GetDeviceDescription();
+ if (!device_description.cuda_compute_capability(&cc_major, &cc_minor)) {
+ LOG(WARNING)
+ << "Couldn't get compute capability for device; assuming sm_20.";
+ cc_major = 2;
+ cc_minor = 0;
+ }
+ return std::make_pair(cc_major, cc_minor);
+}
+
+// Return the constant launch bound along the "x" dimension in "dim" if all the
+// other dimensions are 1. Return nullopt otherwise or when any of the bounds
+// is not constant.
+static absl::optional<int64> getLaunchBound(const mlir::gpu::KernelDim3& dim) {
+ auto get_constant = [](mlir::Operation* op,
+ mlir::StringRef name) -> absl::optional<int64> {
+ if (auto constant = llvm::dyn_cast_or_null<mlir::ConstantOp>(op)) {
+ return constant.value().cast<mlir::IntegerAttr>().getInt();
+ }
+ op->emitError() << "bound " << name << " is not constant";
+ return absl::nullopt;
+ };
+ auto y_op = dim.y.getDefiningOp();
+ auto dim_y = get_constant(y_op, "y");
+ if (!dim_y.has_value() || dim_y.value() != 1) {
+ y_op->emitError() << "bound 'y' is not constant 1";
+ return absl::nullopt;
+ }
+ auto z_op = dim.z.getDefiningOp();
+ auto dim_z = get_constant(z_op, "z");
+ if (!dim_z.has_value() || dim_z.value() != 1) {
+ z_op->emitError() << "bound 'z' is not constant 1";
+ return absl::nullopt;
+ }
+ return get_constant(dim.x.getDefiningOp(), "x");
+}
+
+// Indexes of a range of arguments in a GPU function. This is used to keep the
+// range of arguments that correspond to a lowered kernel argument of
+// (previously) memref type.
+struct LaunchFuncArgument {
+ int kernel_argument_begin;
+ int kernel_argument_size;
+};
+
+using OperandToValueMap =
+ absl::flat_hash_map<const HloInstruction*, std::vector<LaunchFuncArgument>>;
+
+static StatusOr<std::vector<const HloInstruction*>> ComputeOperandToValueMap(
+ OperandToValueMap* operand_to_value_map, const HloInstruction* instr,
+ LaunchFuncOp launchOp, LLVMFuncOp kernel) {
+ auto operands = instr->operands();
+ std::vector<const HloInstruction*> ordered_operands;
+ bool has_failed = false;
+ // A memref will expand into multiple kernel operands, accumulate their number
+ // in order to find them later.
+ int cur_operand_position = 0;
+
+ for (int kernel_index = 0; kernel_index < launchOp.getNumKernelOperands();
+ ++kernel_index) {
+ auto launchop_operand =
+ launchOp.getKernelOperand(kernel_index).dyn_cast<BlockArgument>();
+ if (!launchop_operand) {
+ launchOp.emitError("argument to kernel is not a function input");
+ has_failed = true;
+ continue;
+ }
+ auto memref_type =
+ launchop_operand.getType().dyn_cast<::mlir::MemRefType>();
+ if (!memref_type) {
+ launchOp.emitError("only memref-typed arguments are supported");
+ has_failed = true;
+ break;
+ }
+ // host_index is the argument position to the surrounding function that
+ // contains the launch. This index corresponds to HLO operand indices
+ // by construction.
+ auto host_index = launchop_operand.getArgNumber();
+ // The trailing argument to the outer function are the results.
+ auto operand =
+ (host_index < operands.size()) ? operands[host_index] : instr;
+ if (!operand_to_value_map->count(operand)) {
+ ordered_operands.push_back(operand);
+ }
+ // Associate the HLO operand with the argument values of the kernel
+ // function.
+ int num_unpacked =
+ mlir::MemRefDescriptor::getNumUnpackedValues(memref_type);
+ (*operand_to_value_map)[operand].push_back(
+ {cur_operand_position, num_unpacked});
+ cur_operand_position += num_unpacked;
+ }
+ if (has_failed) {
+ return InternalError("Mapping operands to kernel arguments has failed.");
+ }
+ return ordered_operands;
+}
+
+Status InsertBufferLoadPreduleIntoKernel(
+ LLVMFuncOp kernel, const OperandToValueMap& operand_to_value_map,
+ const std::vector<const HloInstruction*>& ordered_operands,
+ BufferAssignment* assignment,
+ const std::vector<const BufferAllocation*>& buffers) {
+ mlir::OpBuilder builder(kernel.getBody());
+ auto llvm_dialect = kernel.getContext()->getRegisteredDialect<LLVMDialect>();
+ auto offset_type = LLVMType::getInt64Ty(llvm_dialect);
+ auto ptr_type = LLVMType::getInt8PtrTy(llvm_dialect);
+ auto void_type = LLVMType::getVoidTy(llvm_dialect);
+ auto loc = kernel.getLoc();
+
+ auto num_original_args = kernel.getNumArguments();
+ std::vector<LLVMType> new_arg_types(buffers.size(), ptr_type);
+ kernel.setAttr(kernel.getTypeAttrName(),
+ mlir::TypeAttr::get(LLVMType::getFunctionTy(
+ void_type, new_arg_types, /*isVarArg=*/false)));
+ std::vector<Value> original_args(kernel.args_begin(), kernel.args_end());
+
+ std::vector<mlir::Type> as_mlir_types(new_arg_types.begin(),
+ new_arg_types.end());
+ auto new_args = kernel.front().addArguments(as_mlir_types);
+ std::vector<Value> buffer_args(new_args.begin(), new_args.end());
+
+ for (auto operand : ordered_operands) {
+ TF_ASSIGN_OR_RETURN(auto slice,
+ assignment->GetUniqueTopLevelSlice(operand));
+ auto buffer = std::find(buffers.begin(), buffers.end(), slice.allocation());
+ auto index = buffer - buffers.begin();
+ auto offset = builder.create<mlir::LLVM::ConstantOp>(
+ loc, offset_type, builder.getI64IntegerAttr(slice.offset()));
+ auto ptr = buffer_args[index];
+
+ // Replace uses of function arguments pertaining to memref descriptors with
+ // values derived from HLO buffers. The instructions inserting these values
+ // into memref descriptors were already introduced during the lowering phase
+ // as per MLIR calling convention.
+ for (auto arg : operand_to_value_map.at(operand)) {
+ mlir::MemRefDescriptorView original(
+ mlir::ValueRange(original_args)
+ .slice(arg.kernel_argument_begin, arg.kernel_argument_size));
+
+ // Allocated and aligned pointers are the same.
+ auto casted = builder.create<mlir::LLVM::BitcastOp>(
+ loc, original.alignedPtr().getType().cast<LLVMType>(),
+ mlir::ValueRange(ptr));
+ original.alignedPtr().replaceAllUsesWith(casted);
+ original.allocatedPtr().replaceAllUsesWith(casted);
+
+ // Use the offset of the HLO buffer instead of the one expected in the
+ // function call.
+ original.offset().replaceAllUsesWith(offset);
+
+ // Fill the shape.
+ auto shape = operand->shape();
+ // Unless the operand is a scalar pointer, also fill shape and strides.
+ if (shape.dimensions().empty()) {
+ continue;
+ }
+
+ // TODO(b/137624192) Pass in the descriptor to allow for dynamic shapes.
+ assert(shape.IsArray() && shape.is_static());
+ for (auto extent : llvm::enumerate(shape.dimensions())) {
+ auto shape = builder.create<mlir::LLVM::ConstantOp>(
+ loc, original.size(extent.index()).getType(),
+ builder.getI64IntegerAttr(extent.value()));
+ original.size(extent.index()).replaceAllUsesWith(shape);
+ }
+ // Finally, fill the strides.
+ // TODO(b/137624192): Take assigned layout into account.
+ uint64_t accumulator = 0;
+ for (int64_t idx = shape.rank() - 1; idx >= 0; --idx) {
+ if (accumulator == 0) {
+ accumulator = 1;
+ } else {
+ accumulator *= shape.dimensions(idx + 1);
+ }
+ auto stride = builder.create<mlir::LLVM::ConstantOp>(
+ loc, original.stride(idx).getType(),
+ builder.getI64IntegerAttr(accumulator));
+ original.stride(idx).replaceAllUsesWith(stride);
+ }
+ }
+ }
+
+ // Now we can remove the original arguments, as they should have no more
+ // users.
+ for (int i = 0; i < num_original_args; ++i) {
+ kernel.front().eraseArgument(0);
+ }
+
+ return Status::OK();
+}
+
+StatusOr<std::unique_ptr<gpu::KernelThunk>> TransformKernelToXlaThunk(
+ FuncOp func, const HloInstruction* const instr, ModuleOp kernel_module,
+ BufferAssignment* assignment) {
+ // Find the single LaunchFuncOp and compute a mapping from operands of
+ // the hlo instruction to the corresponding values of the kernel
+ // function in the target module;
+ LaunchFuncOp launchOp;
+ auto walkResult = func.walk([&launchOp](LaunchFuncOp op) {
+ if (launchOp) {
+ op.emitError("multiple kernels for single top-level HLO");
+ return mlir::WalkResult::interrupt();
+ }
+ launchOp = op;
+ return mlir::WalkResult::advance();
+ });
+ if (walkResult.wasInterrupted()) {
+ return InternalError("Multiple kernels for single top-level HLO");
+ }
+ if (!launchOp) {
+ // If there was no launchOp, then no kernel was generated, so the lowering
+ // from the LHLO ops to the GPU dialect is not implemented yet.
+ return Unimplemented("No kernel was generated.");
+ }
+
+ auto kernel =
+ kernel_module.lookupSymbol<LLVMFuncOp>(launchOp.getKernelName());
+
+ // Store the assignment of operands to block arguments. Note that an operand
+ // might be used in multiple argument positions, hence the vector.
+ OperandToValueMap operand_to_value_map;
+ TF_ASSIGN_OR_RETURN(
+ auto ordered_operands,
+ ComputeOperandToValueMap(&operand_to_value_map, instr, launchOp, kernel));
+
+ // Get the required buffers to support the inputs. Use a set and vector here
+ // to keep the order fixed. This is mostly useful for testing.
+ std::unordered_set<const BufferAllocation*> buffers_needed;
+ std::vector<const BufferAllocation*> buffers;
+ // TODO(b/137624192) Add support for tuples.
+ for (auto operand : ordered_operands) {
+ TF_ASSIGN_OR_RETURN(auto buffer,
+ assignment->GetUniqueTopLevelSlice(operand));
+ if (buffers_needed.insert(buffer.allocation()).second) {
+ buffers.push_back(buffer.allocation());
+ }
+ }
+
+ // TODO(b/137624192) Add support for temp buffer.
+ // TODO(b/137624192) Add support for constant buffers.
+
+ // Change the signature to match what the XLA runtime expects from the
+ // kernel.
+ TF_RETURN_IF_ERROR(InsertBufferLoadPreduleIntoKernel(
+ kernel, operand_to_value_map, ordered_operands, assignment, buffers));
+
+ // Finally, create the thunk and set the launch dimensions.
+ auto thunk = absl::make_unique<gpu::KernelThunk>(
+ buffers, kernel.getName().str(), instr,
+ /*unroll_factor=*/1);
+
+ // Set launch bounds.
+ mlir::gpu::KernelDim3 block = launchOp.getBlockSizeOperandValues();
+ mlir::gpu::KernelDim3 grid = launchOp.getGridSizeOperandValues();
+ absl::optional<int64> num_threads = getLaunchBound(block);
+ absl::optional<int64> num_blocks = getLaunchBound(grid);
+ if (!num_threads || !num_blocks) {
+ return Unimplemented("Unsupported launch bounds");
+ }
+ thunk->SetLaunchDimensions(gpu::LaunchDimensions(*num_blocks, *num_threads));
+ return std::move(thunk);
+}
+
+StatusOr<std::unique_ptr<Executable>> MlirCompilerImpl::RunBackend(
+ std::unique_ptr<HloModule> module, se::StreamExecutor* stream_exec,
+ se::DeviceMemoryAllocator* device_allocator) {
+ // Determine the HLO schedule, which is an ordering of HLO instructions. This
+ // is used by buffer assignment to enable buffer reuse, and the same ordering
+ // must also be used to determine the thunk launch schedule.
+ std::unique_ptr<StreamAssignment> stream_assignment =
+ xla::gpu::AssignStreams(*module);
+ TF_ASSIGN_OR_RETURN(
+ std::unique_ptr<GpuHloSchedule> hlo_schedule,
+ GpuHloSchedule::Build(*module, *stream_assignment, pointer_size_));
+
+ // Run buffer analysis on the HLO graph. This analysis figures out which
+ // temporary buffers are required to run the computation.
+ TF_ASSIGN_OR_RETURN(std::unique_ptr<BufferAssignment> buffer_assignment,
+ BufferAssigner::Run(
+ module.get(), hlo_schedule->ConsumeHloOrdering(),
+ BufferSizeBytesFunction(),
+ /*color_alignment=*/
+ [](LogicalBuffer::Color) {
+ return xla::gpu::kXlaAllocatedBufferAlignBytes;
+ },
+ /*allocate_buffers_for_constants=*/true,
+ /*colorer=*/BufferAssigner::DefaultColorer(),
+ /*must_not_live_out=*/{}, &CanShareBufferHint));
+ DumpHloModuleIfEnabled(*module, *buffer_assignment, "after_optimizations");
+
+ EmissionContext emission_context(std::move(module));
+ if (error_handler_) {
+ emission_context.setErrorHandler(error_handler_);
+ }
+
+ OwningModuleRef mlir_module =
+ ModuleOp::create(UnknownLoc::get(emission_context.getContext()));
+ LhloDialectEmitter lhlo_emitter(&emission_context, *buffer_assignment,
+ stream_exec->platform(), *mlir_module);
+
+ TF_RETURN_IF_ERROR(lhlo_emitter.EmitComputation(
+ *emission_context.getHloModule()->entry_computation()));
+
+ TF_RETURN_IF_ERROR(
+ module_hook_.invoke(IRHook::LoweringStage::LHLO, *mlir_module));
+
+ TF_RETURN_IF_ERROR(LowerLHLOToGPU(*mlir_module));
+
+ TF_RETURN_IF_ERROR(
+ module_hook_.invoke(IRHook::LoweringStage::GPU, *mlir_module));
+
+ TF_RETURN_IF_ERROR(LowerKernelBodiesToNVVM(*mlir_module));
+
+ TF_RETURN_IF_ERROR(
+ module_hook_.invoke(IRHook::LoweringStage::LLVM, *mlir_module));
+
+ TF_ASSIGN_OR_RETURN(OwningModuleRef kernel_module,
+ ExtractKernelModule(*mlir_module));
+
+ auto thunk_sequence = lhlo_emitter.ConsumeThunkSequence();
+ for (auto entry : lhlo_emitter.InstructionToFunctionMap()) {
+ TF_ASSIGN_OR_RETURN(
+ auto thunk,
+ TransformKernelToXlaThunk(entry.second, entry.first, *kernel_module,
+ buffer_assignment.get()));
+ thunk_sequence->push_back(std::move(thunk));
+ }
+
+ TF_RETURN_IF_ERROR(
+ module_hook_.invoke(IRHook::LoweringStage::KERNEL, *kernel_module));
+
+ auto llvmModule = mlir::translateModuleToNVVMIR(*kernel_module);
+
+ if (!llvmModule) {
+ return InternalError("Translation to LLVM failed");
+ }
+
+ llvmModule->setModuleIdentifier(emission_context.getHloModule()->name());
+ // TODO(herhut): Why is this needed and does not come from the template?
+ llvmModule->setDataLayout(gpu::nvptx::kDataLayout);
+
+ const auto& config = emission_context.getHloModule()->config();
+ TF_ASSIGN_OR_RETURN(
+ auto ptx, xla::gpu::nvptx::CompileToPtx(llvmModule.get(),
+ GetGpuVersion(stream_exec),
+ config, GetLibdeviceDir(config)));
+ TF_ASSIGN_OR_RETURN(
+ auto cubin, se::CompileGpuAsm(stream_exec->device_ordinal(), ptx.c_str(),
+ gpu::PtxOptsFromConfig(config)));
+
+ auto thunk_schedule = absl::make_unique<ThunkSchedule>(
+ std::move(thunk_sequence), std::move(stream_assignment),
+ hlo_schedule->ThunkLaunchOrder());
+
+ if (DumpingEnabledForHloModule(*emission_context.getHloModule())) {
+ DumpToFileInDirOrStdout(*emission_context.getHloModule(), "",
+ "thunk_schedule", thunk_schedule->ToString());
+ }
+
+ // TODO(b/137624192): Add profiling support.
+ return {absl::make_unique<GpuExecutable>(
+ ptx, cubin, GetGpuVersion(stream_exec), std::move(thunk_schedule),
+ emission_context.releaseHloModule(), std::move(buffer_assignment),
+ nullptr, nullptr)};
+}
+
+StatusOr<std::vector<std::unique_ptr<Executable>>> MlirCompilerImpl::Compile(
+ std::unique_ptr<HloModuleGroup> module_group,
+ std::vector<std::vector<se::StreamExecutor*>> stream_execs,
+ se::DeviceMemoryAllocator* device_allocator) {
+ return Unimplemented("Not yet implemented in MLIR compiler");
+}
+
+StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
+MlirCompilerImpl::CompileAheadOfTime(
+ std::unique_ptr<HloModuleGroup> /*module_group*/,
+ const AotCompilationOptions& /*options*/) {
+ return Unimplemented("Not yet implemented in MLIR compiler");
+}
+
+} // namespace
+} // namespace mlir_gpu
+} // namespace xla
+
+static bool InitModule() {
+ xla::Compiler::RegisterCompilerFactory(
+ stream_executor::cuda::kCudaPlatformId, []() {
+ return absl::make_unique<xla::FailoverCompiler>(
+ absl::make_unique<xla::mlir_gpu::MlirCompilerImpl>(),
+ absl::make_unique<xla::gpu::NVPTXCompiler>());
+ });
+ return true;
+}
+static bool module_initialized = InitModule();
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/BUILD b/tensorflow/compiler/xla/service/mlir_gpu/tests/BUILD
index ddf5a48..014b26c 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/tests/BUILD
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/BUILD
@@ -25,8 +25,6 @@
],
driver = "@llvm-project//mlir:run_lit.sh",
exclude = [
- # TODO(b/149302060) Reenable once fusion is fixed.
- "iota_add_multiply.hlo",
# TODO(b/137624192): Reenable once we can fuse reductions.
"fused_reduce.hlo",
],
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/complex.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/complex.hlo
new file mode 100644
index 0000000..974eb4e
--- /dev/null
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/complex.hlo
@@ -0,0 +1,12 @@
+// RUN: xla-gpu-opt %s | FileCheck %s -dump-input-on-failure
+HloModule Complex
+
+ENTRY %Complex (real: f32[2,2]{0,1}, imag: f32[2,2]{0,1}) -> c64[2,2] {
+ %real = f32[2,2]{0,1} parameter(0)
+ %imag = f32[2,2]{0,1} parameter(1)
+ ROOT %compl = c64[2,2]{0,1} complex(%real, %imag)
+}
+
+// CHECK: func @complex(%[[REAL:.*]]: [[BUF_F32:.*]], %[[IMAG:.*]]: [[BUF_F32]], %[[OUT:.*]]: [[BUF_C64:.*]]) {
+// CHECK: "xla_lhlo.complex"(%[[REAL]], %[[IMAG]], %[[OUT]]) : ([[BUF_F32]], [[BUF_F32]], [[BUF_C64]]) -> ()
+// CHECK: }
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/imag.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/imag.hlo
new file mode 100644
index 0000000..ca79c84
--- /dev/null
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/imag.hlo
@@ -0,0 +1,11 @@
+// RUN: xla-gpu-opt %s | FileCheck %s -dump-input-on-failure
+HloModule Imag
+
+ENTRY %Imag (x: c64[2,2]{0,1}) -> f32[2,2] {
+ %x = c64[2,2]{0,1} parameter(0)
+ ROOT %imag = f32[2,2]{0,1} imag(%x)
+}
+
+// CHECK: func @imag(%[[IN:.*]]: [[BUF_C64:.*]], %[[OUT:.*]]: [[BUF_F32:.*]]) {
+// CHECK: "xla_lhlo.imag"(%[[IN]], %[[OUT]]) : ([[BUF_C64]], [[BUF_F32]]) -> ()
+// CHECK: }
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_multiply.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_multiply.hlo
deleted file mode 100644
index 1c52d43..0000000
--- a/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_multiply.hlo
+++ /dev/null
@@ -1,16 +0,0 @@
-// RUN: xla-gpu-opt -lowering-stage=GPU %s | FileCheck %s -dump-input-on-failure
-HloModule AddMultiply
-
-ENTRY %AddMultiply (x: s32[2,2], y: s32[2,2]) -> s32[2,2] {
- %x = s32[2,2]{1,0} parameter(0)
- %y = s32[2,2]{1,0} parameter(1)
-
- %add = s32[2,2]{1,0} add(s32[2,2]{1,0} %x, s32[2,2]{1,0} %y)
- %iota = s32[2, 2]{1,0} iota(), iota_dimension=0
-
- ROOT %mul = s32[2,2]{1,0} multiply(s32[2,2]{1,0} %add, s32[2,2]{1,0} %iota)
-}
-
-// CHECK-NOT: store
-// CHECK: %[[RESULT:.*]] = muli %{{.*}}, %{{.*}}
-// CHECK: store %[[RESULT]]
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_subtract.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_subtract.hlo
new file mode 100644
index 0000000..f42a7cf
--- /dev/null
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/iota_add_subtract.hlo
@@ -0,0 +1,16 @@
+// RUN: xla-gpu-opt -lowering-stage=GPU %s | FileCheck %s -dump-input-on-failure
+HloModule AddSubtract
+
+ENTRY %AddSubtract (x: s32[2,2], y: s32[2,2]) -> s32[2,2] {
+ %x = s32[2,2]{1,0} parameter(0)
+ %y = s32[2,2]{1,0} parameter(1)
+
+ %add = s32[2,2]{1,0} add(s32[2,2]{1,0} %x, s32[2,2]{1,0} %y)
+ %iota = s32[2, 2]{1,0} iota(), iota_dimension=0
+
+ ROOT %sub = s32[2,2]{1,0} subtract(s32[2,2]{1,0} %add, s32[2,2]{1,0} %iota)
+}
+
+// CHECK-NOT: store
+// CHECK: [[RESULT:%.*]] = subi %{{.*}}, %{{.*}}
+// CHECK: store [[RESULT]]
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/tests/real.hlo b/tensorflow/compiler/xla/service/mlir_gpu/tests/real.hlo
new file mode 100644
index 0000000..cb19c39
--- /dev/null
+++ b/tensorflow/compiler/xla/service/mlir_gpu/tests/real.hlo
@@ -0,0 +1,11 @@
+// RUN: xla-gpu-opt %s | FileCheck %s -dump-input-on-failure
+HloModule Real
+
+ENTRY %Real (x: c64[2,2]{0,1}) -> f32[2,2] {
+ %x = c64[2,2]{0,1} parameter(0)
+ ROOT %real = f32[2,2]{0,1} real(%x)
+}
+
+// CHECK: func @real(%[[IN:.*]]: [[BUF_C64:.*]], %[[OUT:.*]]: [[BUF_F32:.*]]) {
+// CHECK: "xla_lhlo.real"(%[[IN]], %[[OUT]]) : ([[BUF_C64]], [[BUF_F32]]) -> ()
+// CHECK: }
diff --git a/tensorflow/compiler/xla/service/mlir_gpu/xla_gpu_opt_main.cc b/tensorflow/compiler/xla/service/mlir_gpu/xla_gpu_opt_main.cc
index d5e789a..f60eea6 100644
--- a/tensorflow/compiler/xla/service/mlir_gpu/xla_gpu_opt_main.cc
+++ b/tensorflow/compiler/xla/service/mlir_gpu/xla_gpu_opt_main.cc
@@ -17,6 +17,7 @@
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ToolOutputFile.h"
+#include "mlir/Pass/PassManager.h" // from @llvm-project
#include "mlir/Support/FileUtilities.h" // from @llvm-project
#include "tensorflow/compiler/mlir/init_mlir.h"
#include "tensorflow/compiler/xla/service/mlir_gpu/mlir_compiler.h"
@@ -61,6 +62,7 @@
int main(int argc, char **argv) {
tensorflow::InitMlir y(&argc, &argv);
+ mlir::registerPassManagerCLOptions();
llvm::cl::ParseCommandLineOptions(argc, argv,
"XLA GPU modular optimizer driver\n");
diff --git a/tensorflow/compiler/xla/service/root_instruction_sinker.cc b/tensorflow/compiler/xla/service/root_instruction_sinker.cc
new file mode 100644
index 0000000..bee703b
--- /dev/null
+++ b/tensorflow/compiler/xla/service/root_instruction_sinker.cc
@@ -0,0 +1,73 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/compiler/xla/service/root_instruction_sinker.h"
+
+#include "tensorflow/compiler/xla/service/tuple_util.h"
+namespace xla {
+
+namespace {
+
+// Sinks the root of the given computation for tuple root types.
+void SinkTupleRoot(HloComputation* computation) {
+ HloInstruction* root = computation->root_instruction();
+ CHECK(root->shape().IsTuple());
+ HloInstruction* new_root = TupleUtil::Duplicate(root);
+ // Add the new instructions to the schedule.
+ HloInstructionSequence& sequence =
+ computation->parent()->schedule().GetOrCreateSequence(computation);
+ for (HloInstruction* operand : new_root->operands()) {
+ sequence.push_back(operand);
+ }
+ sequence.push_back(new_root);
+ computation->set_root_instruction(new_root);
+}
+
+// Sinks the root of the given computation for not-tuple root types.
+void SinkNontupleRoot(HloComputation* computation) {
+ HloInstruction* root = computation->root_instruction();
+ CHECK(!root->shape().IsTuple());
+ HloInstruction* new_root = computation->AddInstruction(
+ HloInstruction::CreateBitcast(root->shape(), root));
+ HloInstructionSequence& sequence =
+ computation->parent()->schedule().GetOrCreateSequence(computation);
+ sequence.push_back(new_root);
+ computation->set_root_instruction(new_root);
+}
+
+} // namespace
+
+StatusOr<bool> RootInstructionSinker::Run(HloModule* module) {
+ TF_RET_CHECK(module->has_schedule());
+
+ bool modified = false;
+ for (HloComputation* computation : module->MakeNonfusionComputations()) {
+ HloInstructionSequence& sequence =
+ module->schedule().GetOrCreateSequence(computation);
+ if (computation->root_instruction() ==
+ sequence.instructions().at(sequence.size() - 1)) {
+ continue;
+ }
+ if (computation->root_instruction()->shape().IsTuple()) {
+ SinkTupleRoot(computation);
+ } else {
+ SinkNontupleRoot(computation);
+ }
+ modified = true;
+ }
+ return modified;
+}
+
+} // namespace xla
diff --git a/tensorflow/compiler/xla/service/root_instruction_sinker.h b/tensorflow/compiler/xla/service/root_instruction_sinker.h
new file mode 100644
index 0000000..d4d0887
--- /dev/null
+++ b/tensorflow/compiler/xla/service/root_instruction_sinker.h
@@ -0,0 +1,41 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
+#define TENSORFLOW_COMPILER_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
+
+#include "tensorflow/compiler/xla/service/hlo_module.h"
+#include "tensorflow/compiler/xla/service/hlo_pass_interface.h"
+
+namespace xla {
+
+// Given a scheduled HLO module, this pass sinks the ROOT of the instruction to
+// the bottom of the non-fusion computations. To avoid dependency violations of
+// moving the ROOT instruction, it creates a new ROOT instruction that looks
+// like the following:
+// - For tuple ROOT type:
+// new_root = tuple(gte(old_root), gte(old_root), ...)
+// - For non-tuple ROOT type:
+// new_root = bitcast(old_root)
+class RootInstructionSinker : public HloModulePass {
+ public:
+ ~RootInstructionSinker() override = default;
+ absl::string_view name() const override { return "root-instruction-sinker"; }
+ StatusOr<bool> Run(HloModule* module) override;
+};
+
+} // namespace xla
+
+#endif // TENSORFLOW_COMPILER_XLA_SERVICE_ROOT_INSTRUCTION_SINKER_H_
diff --git a/tensorflow/compiler/xla/service/root_instruction_sinker_test.cc b/tensorflow/compiler/xla/service/root_instruction_sinker_test.cc
new file mode 100644
index 0000000..8a03a92
--- /dev/null
+++ b/tensorflow/compiler/xla/service/root_instruction_sinker_test.cc
@@ -0,0 +1,170 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/compiler/xla/service/root_instruction_sinker.h"
+
+#include "tensorflow/compiler/xla/service/hlo_matchers.h"
+#include "tensorflow/compiler/xla/tests/hlo_test_base.h"
+
+namespace xla {
+namespace {
+
+namespace op = xla::testing::opcode_matchers;
+
+using RootInstructionSinkerTest = HloTestBase;
+
+TEST_F(RootInstructionSinkerTest, TupleNoChange) {
+ // ROOTS are already sunk, no change performed to the module.
+ absl::string_view hlo_string = R"(
+ HloModule While, is_scheduled=true
+ While.body {
+ loop_var.1 = (s32[], s32[3]{0}) parameter(0)
+ get-tuple-element.1 = s32[] get-tuple-element(loop_var.1), index=0
+ constant.1 = s32[] constant(1)
+ add = s32[] add(get-tuple-element.1, constant.1)
+ get-tuple-element.2 = s32[3]{0} get-tuple-element(loop_var.1), index=1
+ multiply = s32[3]{0} multiply(get-tuple-element.2, get-tuple-element.2)
+ ROOT tuple = (s32[], s32[3]{0}) tuple(add, multiply)
+ }
+ While.condition {
+ loop_var.2 = (s32[], s32[3]{0}) parameter(0)
+ get-tuple-element.3 = s32[] get-tuple-element(loop_var.2), index=0
+ constant.2 = s32[] constant(100)
+ ROOT less-than = pred[] compare(get-tuple-element.3, constant.2), direction=LT
+ }
+ ENTRY While {
+ constant.3 = s32[] constant(42)
+ constant.4 = s32[3]{0} constant({0, 1, 2})
+ tuple.1 = (s32[], s32[3]{0}) tuple(constant.3, constant.4)
+ ROOT while = (s32[], s32[3]{0}) while(tuple.1), condition=
+ While.condition, body=While.body
+ }
+ )";
+ TF_ASSERT_OK_AND_ASSIGN(auto module,
+ ParseAndReturnVerifiedModule(hlo_string));
+ auto while_body =
+ module->entry_computation()->root_instruction()->while_body();
+ int num_body_instructions = while_body->instruction_count();
+ RootInstructionSinker sinker;
+ EXPECT_FALSE(sinker.Run(module.get()).ValueOrDie());
+ EXPECT_EQ(module->entry_computation()
+ ->root_instruction()
+ ->while_body()
+ ->instruction_count(),
+ num_body_instructions);
+}
+
+TEST_F(RootInstructionSinkerTest, Tuple) {
+ // Sink tuple return type.
+ absl::string_view hlo_string = R"(
+ HloModule While, is_scheduled=true
+ While.body {
+ loop_var.1 = (s32[], s32[3]{0}) parameter(0)
+ get-tuple-element.1 = s32[] get-tuple-element(loop_var.1), index=0
+ constant.1 = s32[] constant(1)
+ add = s32[] add(get-tuple-element.1, constant.1)
+ get-tuple-element.2 = s32[3]{0} get-tuple-element(loop_var.1), index=1
+ multiply = s32[3]{0} multiply(get-tuple-element.2, get-tuple-element.2)
+ ROOT tuple = (s32[], s32[3]{0}) tuple(add, multiply)
+ after-all = token[] after-all()
+ send = (s32[3]{0}, u32[], token[]) send(multiply, after-all), channel_id=1
+ send-done = token[] send-done(send), channel_id=1
+ }
+ While.condition {
+ loop_var.2 = (s32[], s32[3]{0}) parameter(0)
+ get-tuple-element.3 = s32[] get-tuple-element(loop_var.2), index=0
+ constant.2 = s32[] constant(100)
+ ROOT less-than = pred[] compare(get-tuple-element.3, constant.2), direction=LT
+ }
+ ENTRY While {
+ constant.3 = s32[] constant(42)
+ constant.4 = s32[3]{0} constant({0, 1, 2})
+ tuple.1 = (s32[], s32[3]{0}) tuple(constant.3, constant.4)
+ ROOT while = (s32[], s32[3]{0}) while(tuple.1), condition=
+ While.condition, body=While.body
+ }
+ )";
+ TF_ASSERT_OK_AND_ASSIGN(auto module,
+ ParseAndReturnVerifiedModule(hlo_string));
+ RootInstructionSinker sinker;
+ EXPECT_TRUE(sinker.Run(module.get()).ValueOrDie());
+ auto while_body =
+ module->entry_computation()->root_instruction()->while_body();
+ const auto& sequence = module->schedule().sequence(while_body);
+ EXPECT_EQ(sequence.instructions().at(sequence.size() - 1),
+ while_body->root_instruction());
+ EXPECT_THAT(while_body->root_instruction(),
+ op::Tuple(op::GetTupleElement(op::Tuple()),
+ op::GetTupleElement(op::Tuple())));
+}
+
+TEST_F(RootInstructionSinkerTest, NontupleNoChange) {
+ // ROOTS are already sunk, no change performed to the module.
+ absl::string_view hlo_string = R"(
+ HloModule Call, is_scheduled=true
+ Call {
+ param = s32[3]{0} parameter(0)
+ ROOT multiply = s32[3]{0} multiply(param, param)
+ }
+ ENTRY While {
+ constant.4 = s32[3]{0} constant({0, 1, 2})
+ ROOT call = s32[3]{0} call(constant.4), to_apply=Call
+ }
+ )";
+ TF_ASSERT_OK_AND_ASSIGN(auto module,
+ ParseAndReturnVerifiedModule(hlo_string));
+ auto called_computation =
+ module->entry_computation()->root_instruction()->called_computations()[0];
+ int num_instructions = called_computation->instruction_count();
+ RootInstructionSinker sinker;
+ EXPECT_FALSE(sinker.Run(module.get()).ValueOrDie());
+ EXPECT_EQ(module->entry_computation()
+ ->root_instruction()
+ ->called_computations()[0]
+ ->instruction_count(),
+ num_instructions);
+}
+
+TEST_F(RootInstructionSinkerTest, Nontuple) {
+ // Sink a non-tuple return type.
+ absl::string_view hlo_string = R"(
+ HloModule Call, is_scheduled=true
+ Call {
+ param = s32[3]{0} parameter(0)
+ ROOT multiply = s32[3]{0} multiply(param, param)
+ after-all = token[] after-all()
+ send = (s32[3]{0}, u32[], token[]) send(multiply, after-all), channel_id=1
+ send-done = token[] send-done(send), channel_id=1
+ }
+ ENTRY While {
+ constant.4 = s32[3]{0} constant({0, 1, 2})
+ ROOT call = s32[3]{0} call(constant.4), to_apply=Call
+ }
+ )";
+ TF_ASSERT_OK_AND_ASSIGN(auto module,
+ ParseAndReturnVerifiedModule(hlo_string));
+ RootInstructionSinker sinker;
+ EXPECT_TRUE(sinker.Run(module.get()).ValueOrDie());
+ auto called_computation =
+ module->entry_computation()->root_instruction()->called_computations()[0];
+ const auto& sequence = module->schedule().sequence(called_computation);
+ EXPECT_EQ(sequence.instructions().at(sequence.size() - 1),
+ called_computation->root_instruction());
+ EXPECT_THAT(called_computation->root_instruction(),
+ op::Bitcast(op::Multiply()));
+}
+
+} // namespace
+} // namespace xla
diff --git a/tensorflow/compiler/xla/service/service.cc b/tensorflow/compiler/xla/service/service.cc
index e12e157..ab71c30 100644
--- a/tensorflow/compiler/xla/service/service.cc
+++ b/tensorflow/compiler/xla/service/service.cc
@@ -314,6 +314,7 @@
config->set_num_partitions(execution_options->num_partitions());
}
config->set_seed(execution_options->seed());
+ config->set_launch_id(execution_options->launch_id());
config->set_debug_options(execution_options->debug_options());
} else {
config->set_replica_count(options_.number_of_replicas());
diff --git a/tensorflow/compiler/xla/service/shape_inference.cc b/tensorflow/compiler/xla/service/shape_inference.cc
index 3b8c2f4..d2cbddd 100644
--- a/tensorflow/compiler/xla/service/shape_inference.cc
+++ b/tensorflow/compiler/xla/service/shape_inference.cc
@@ -2596,7 +2596,18 @@
VLOG(2) << StrFormat("update_sizes[%d] = %d", dim, update_dim_size);
}
- return operand_shape;
+ auto result_shape = operand_shape;
+
+ // If any of the operand shape and update shape is dynamic, update the result
+ // dimension to dynamic.
+ for (int64 i = 0; i < update_shape.rank(); ++i) {
+ if (update_shape.is_dynamic_dimension(i) ||
+ operand_shape.is_dynamic_dimension(i)) {
+ result_shape.set_dynamic_dimension(i, true);
+ }
+ }
+
+ return result_shape;
}
/*static */ StatusOr<Shape> ShapeInference::InferReverseShape(
diff --git a/tensorflow/compiler/xla/service/tuple_util.h b/tensorflow/compiler/xla/service/tuple_util.h
index bc5aac0..ee7b8be 100644
--- a/tensorflow/compiler/xla/service/tuple_util.h
+++ b/tensorflow/compiler/xla/service/tuple_util.h
@@ -39,6 +39,13 @@
static HloInstruction* AppendSuffix(
HloInstruction* input_tuple,
absl::Span<HloInstruction* const> trailing_values);
+
+ // Generates HLO instructions that duplicates the tuple by inserting
+ // get-tuple-elements and a new tuple instruction. Returns the root of the
+ // graph of instructions generated.
+ static HloInstruction* Duplicate(HloInstruction* input_tuple) {
+ return ExtractPrefix(input_tuple, input_tuple->shape().tuple_shapes_size());
+ }
};
} // namespace xla
diff --git a/tensorflow/compiler/xla/service/while_util.cc b/tensorflow/compiler/xla/service/while_util.cc
index e5d64b2..f2c4f7f 100644
--- a/tensorflow/compiler/xla/service/while_util.cc
+++ b/tensorflow/compiler/xla/service/while_util.cc
@@ -125,8 +125,9 @@
// We want to get rid of the old while instruction even if it has side
// effecting operations so we do a manual HloComputation::RemoveInstruction
// instead of relying on HloComputation::ReplaceInstruction.
- TF_RETURN_IF_ERROR(while_instr->ReplaceAllUsesWith(TupleUtil::ExtractPrefix(
- new_while, while_instr->shape().tuple_shapes_size())));
+ HloInstruction* replacement_instr = TupleUtil::ExtractPrefix(
+ new_while, while_instr->shape().tuple_shapes_size());
+ TF_RETURN_IF_ERROR(while_instr->ReplaceAllUsesWith(replacement_instr));
TF_RETURN_IF_ERROR(containing_computation->RemoveInstruction(while_instr));
HloInstruction* while_body_param = new_while_body->parameter_instruction(0);
@@ -142,6 +143,7 @@
WhileUtil::MakeInstructionsLiveInResult result;
result.new_while_instr = new_while;
+ result.replacement_instr = replacement_instr;
result.while_body_live_in_values = std::move(live_in_instructions);
result.while_body_instruction_map = std::move(inlined_instructions_map);
diff --git a/tensorflow/compiler/xla/service/while_util.h b/tensorflow/compiler/xla/service/while_util.h
index cba41cc..b4b9d29 100644
--- a/tensorflow/compiler/xla/service/while_util.h
+++ b/tensorflow/compiler/xla/service/while_util.h
@@ -29,6 +29,10 @@
// The new while operation that has the requested values live in.
HloInstruction* new_while_instr;
+ // The new tuple instruction that replaced the original while instruction
+ // with the same shape.
+ HloInstruction* replacement_instr;
+
// The i'th element of `while_body_live_in_values` is an instruction in the
// while body that holds the i'th *newly added* live in value at runtime.
std::vector<HloInstruction*> while_body_live_in_values;
diff --git a/tensorflow/compiler/xla/tests/half_test.cc b/tensorflow/compiler/xla/tests/half_test.cc
index 74333d6..566f655 100644
--- a/tensorflow/compiler/xla/tests/half_test.cc
+++ b/tensorflow/compiler/xla/tests/half_test.cc
@@ -34,7 +34,7 @@
protected:
const ErrorSpec error_spec_{0.001, 0.001};
// Number of elements in the input buffers.
- static const int kNumElements = 4;
+ static constexpr int kNumElements = 4;
};
using UnaryBuildFuncTy = std::function<void(const xla::XlaOp& src)>;
diff --git a/tensorflow/compiler/xla/xla.proto b/tensorflow/compiler/xla/xla.proto
index f8bd7a0..826876e 100644
--- a/tensorflow/compiler/xla/xla.proto
+++ b/tensorflow/compiler/xla/xla.proto
@@ -269,7 +269,10 @@
bool xla_tpu_detect_nan = 135;
bool xla_tpu_detect_inf = 136;
- // Next id: 137
+ // True if TraceMe annotations are enabled for XLA:CPU.
+ bool xla_cpu_enable_xprof_traceme = 137;
+
+ // Next id: 138
// Extra options to pass to the compilation backend (e.g. LLVM); specific
// interpretation of these values is left to the backend.
@@ -319,6 +322,9 @@
// Number of partitions of the computation to run (model parallelism).
// If zero, uses the default number of partitions for the XLA service.
int32 num_partitions = 9;
+
+ // Used to identify a set of programs that should be launch together.
+ int32 launch_id = 10;
}
message GetDeviceHandlesRequest {
diff --git a/tensorflow/compiler/xrt/xrt_state.cc b/tensorflow/compiler/xrt/xrt_state.cc
index a0daa5c..c2f9a1c 100644
--- a/tensorflow/compiler/xrt/xrt_state.cc
+++ b/tensorflow/compiler/xrt/xrt_state.cc
@@ -588,7 +588,8 @@
allocator_->platform(), device_ordinal_);
for (const auto& index_buffer : buffers_) {
if (index_buffer.second == nullptr ||
- index_buffer.second->allocation().is_null()) {
+ (index_buffer.second->allocation().is_null() &&
+ index_buffer.second->allocation().size() > 0)) {
return errors::InvalidArgument("Literal buffer at index ",
index_buffer.first.ToString(),
" has been released");
@@ -652,7 +653,8 @@
xla::ExecutionInput result(on_device_shape());
for (const auto& index_buffer : buffers_) {
if (index_buffer.second == nullptr ||
- index_buffer.second->allocation().is_null()) {
+ (index_buffer.second->allocation().is_null() &&
+ index_buffer.second->allocation().size() > 0)) {
return errors::InvalidArgument("Literal buffer at index ",
index_buffer.first.ToString(),
" has been released");
diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index d2d8996..d6b2e5f 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -2065,7 +2065,13 @@
"//tensorflow/core/platform/default:logging.h",
],
copts = tf_copts(),
- linkopts = ["-ldl"],
+ linkopts = select({
+ "//tensorflow:freebsd": [],
+ "//tensorflow:windows": [],
+ "//conditions:default": [
+ "-ldl",
+ ],
+ }),
visibility = ["//visibility:public"],
deps = [
":platform_base",
diff --git a/tensorflow/core/api_def/base_api/api_def_DummyMemoryCache.pbtxt b/tensorflow/core/api_def/base_api/api_def_DummyMemoryCache.pbtxt
new file mode 100644
index 0000000..3b940d4
--- /dev/null
+++ b/tensorflow/core/api_def/base_api/api_def_DummyMemoryCache.pbtxt
@@ -0,0 +1,4 @@
+op {
+ graph_op_name: "DummyMemoryCache"
+ visibility: HIDDEN
+}
diff --git a/tensorflow/core/api_def/base_api/api_def_DummySeedGenerator.pbtxt b/tensorflow/core/api_def/base_api/api_def_DummySeedGenerator.pbtxt
new file mode 100644
index 0000000..3e771fd
--- /dev/null
+++ b/tensorflow/core/api_def/base_api/api_def_DummySeedGenerator.pbtxt
@@ -0,0 +1,4 @@
+op {
+ graph_op_name: "DummySeedGenerator"
+ visibility: HIDDEN
+}
diff --git a/tensorflow/core/api_def/base_api/api_def_ShuffleDatasetV3.pbtxt b/tensorflow/core/api_def/base_api/api_def_ShuffleDatasetV3.pbtxt
new file mode 100644
index 0000000..0cb628b
--- /dev/null
+++ b/tensorflow/core/api_def/base_api/api_def_ShuffleDatasetV3.pbtxt
@@ -0,0 +1,4 @@
+op {
+ graph_op_name: "ShuffleDatasetV3"
+ visibility: HIDDEN
+}
diff --git a/tensorflow/core/common_runtime/BUILD b/tensorflow/core/common_runtime/BUILD
index bf5b1cf..0fc7389 100644
--- a/tensorflow/core/common_runtime/BUILD
+++ b/tensorflow/core/common_runtime/BUILD
@@ -53,6 +53,7 @@
tf_cuda_library(
name = "core_cpu",
hdrs = [
+ "composite_device.h",
"device.h",
"device_factory.h",
"function.h",
@@ -145,6 +146,7 @@
filegroup(
name = "core_cpu_base_headers",
srcs = [
+ "composite_device.h",
"device.h",
"device_factory.h",
"device_mgr.h",
@@ -167,8 +169,6 @@
srcs = [
"eval_const_tensor.cc",
"graph_optimizer.h",
- "scoped_allocator.cc",
- "scoped_allocator_mgr.cc",
"shape_refiner.cc",
"//tensorflow/core/graph:core_cpu_base_no_ops_srcs",
"//tensorflow/core/public:session_options.h",
@@ -180,6 +180,7 @@
],
copts = tf_copts(),
deps = [
+ ":scoped_allocator",
"//tensorflow/core:graph",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
@@ -217,16 +218,14 @@
"debugger_state_interface.h",
"device_resolver_local.h",
"dma_helper.h",
- "entry.h",
"executor.h",
"executor_factory.h",
"function_optimization_registry.h",
"graph_optimizer.h",
- "graph_view.h",
- "immutable_executor_state.h",
"input_colocation_exemption_registry.h",
"isolate_placer_inspection_required_ops_pass.h",
"local_device.h",
+ "local_executor_params.h",
"lower_function_call_op.h",
"lower_if_op.h",
"lower_case_op.h",
@@ -235,14 +234,11 @@
"memory_types.h",
"mkl_cpu_allocator.h",
"optimization_registry.h",
- "pending_counts.h",
"partitioning_utils.h",
"placer.h",
"process_util.h",
"inspecting_placer.h",
"profile_handler.h",
- "propagator_debug_utils.h",
- "propagator_state.h",
"renamed_device.h",
"rendezvous_mgr.h",
"rendezvous_util.h",
@@ -251,7 +247,6 @@
"ring_alg.h",
"ring_gatherer.h",
"session_factory.h",
- "simple_propagator_state.h",
"single_threaded_cpu_device.h",
"stats_publisher_interface.h",
"step_stats_collector.h",
@@ -262,92 +257,852 @@
] + if_mkl(["//tensorflow/core/graph:mkl_graph_util_header"]),
)
-tf_cuda_library(
- name = "core_cpu_impl",
+cc_library(
+ name = "accumulate_n_optimizer",
+ srcs = ["accumulate_n_optimizer.cc"],
+ copts = tf_copts(),
+ deps = [
+ ":optimization_registry",
+ "//tensorflow/core:graph",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "base_collective_executor",
+ srcs = ["base_collective_executor.cc"],
+ hdrs = ["base_collective_executor.h"],
+ copts = tf_copts(),
+ deps = [
+ ":buf_rendezvous",
+ ":copy_tensor",
+ ":device_mgr",
+ ":dma_helper",
+ ":process_util",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+)
+
+cc_library(
+ name = "buf_rendezvous",
+ srcs = ["buf_rendezvous.cc"],
+ hdrs = ["buf_rendezvous.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":device_mgr",
+ ":process_util",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "@com_google_absl//absl/strings",
+ ],
+)
+
+cc_library(
+ name = "build_graph_options",
+ srcs = ["build_graph_options.cc"],
+ hdrs = ["build_graph_options.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "collective_executor_mgr",
+ srcs = ["collective_executor_mgr.cc"],
+ hdrs = ["collective_executor_mgr.h"],
+ copts = tf_copts(),
+ deps = [
+ ":base_collective_executor",
+ ":build_graph_options",
+ ":collective_rma_local",
+ ":device_mgr",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "collective_util",
+ srcs = ["collective_util.cc"],
+ hdrs = ["collective_util.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":device_mgr",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "copy_tensor",
+ srcs = ["copy_tensor.cc"],
+ hdrs = ["copy_tensor.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":dma_helper",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core/profiler/lib:scoped_annotation",
+ ],
+)
+
+cc_library(
+ name = "collective_param_resolver_local",
+ srcs = ["collective_param_resolver_local.cc"],
+ hdrs = ["collective_param_resolver_local.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":device_mgr",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "collective_rma_local",
+ srcs = ["collective_rma_local.cc"],
+ hdrs = ["collective_rma_local.h"],
+ copts = tf_copts(),
+ deps = [
+ ":buf_rendezvous",
+ ":copy_tensor",
+ ":device",
+ ":device_mgr",
+ ":dma_helper",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "costmodel_manager",
+ srcs = ["costmodel_manager.cc"],
+ hdrs = ["costmodel_manager.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "debugger_state_interface",
+ srcs = ["debugger_state_interface.cc"],
+ hdrs = ["debugger_state_interface.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "device",
+ srcs = ["device.cc"],
+ hdrs = ["device.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework_internal",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "device_factory",
+ srcs = ["device_factory.cc"],
+ hdrs = ["device_factory.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":session_options",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "device_mgr",
srcs = [
- "accumulate_n_optimizer.cc",
- "base_collective_executor.cc",
- "buf_rendezvous.cc",
- "build_graph_options.cc",
- "collective_executor_mgr.cc",
- "collective_param_resolver_local.cc",
- "collective_rma_local.cc",
- "collective_util.cc",
- "colocation_graph.cc",
- "constant_folding.cc",
- "copy_tensor.cc",
- "costmodel_manager.cc",
- "debugger_state_interface.cc",
- "device.cc",
- "device_factory.cc",
"device_mgr.cc",
- "device_resolver_local.cc",
- "device_set.cc",
"dynamic_device_mgr.cc",
- "executor.cc",
- "executor_factory.cc",
+ ],
+ hdrs = ["device_mgr.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":local_device",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "device_resolver_local",
+ srcs = ["device_resolver_local.cc"],
+ hdrs = ["device_resolver_local.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device_mgr",
+ "//tensorflow/core:framework",
+ ],
+)
+
+cc_library(
+ name = "entry",
+ hdrs = ["entry.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "executor",
+ srcs = ["executor.cc"],
+ hdrs = ["executor.h"],
+ copts = tf_copts(),
+ deps = [
+ ":costmodel_manager",
+ ":device",
+ ":entry",
+ ":executor_factory",
+ ":graph_view",
+ ":immutable_executor_state",
+ ":local_executor_params",
+ ":pending_counts",
+ ":propagator_state",
+ ":renamed_device",
+ ":simple_propagator_state",
+ ":step_stats_collector",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:framework_internal",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/core/profiler/lib:annotated_traceme",
+ "//tensorflow/core/profiler/lib:scoped_annotation",
+ "//tensorflow/core/profiler/lib:traceme",
+ "@com_google_absl//absl/memory",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "executor_factory",
+ srcs = ["executor_factory.cc"],
+ hdrs = ["executor_factory.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "graph_view",
+ srcs = ["graph_view.cc"],
+ hdrs = ["graph_view.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "device_set",
+ srcs = ["device_set.cc"],
+ hdrs = ["device_set.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":device_factory",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "dma_helper",
+ hdrs = ["dma_helper.h"],
+ copts = tf_copts(),
+ deps = ["//tensorflow/core:framework"],
+)
+
+cc_library(
+ name = "hierarchical_tree_broadcaster",
+ srcs = ["hierarchical_tree_broadcaster.cc"],
+ hdrs = ["hierarchical_tree_broadcaster.h"],
+ copts = tf_copts(),
+ deps = [
+ ":base_collective_executor",
+ ":collective_rma_local",
+ ":collective_util",
+ ":device_mgr",
+ ":dma_helper",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "immutable_executor_state",
+ srcs = ["immutable_executor_state.cc"],
+ hdrs = ["immutable_executor_state.h"],
+ copts = tf_copts(),
+ deps = [
+ ":graph_view",
+ ":local_executor_params",
+ ":pending_counts",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:framework_internal",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "@com_google_absl//absl/memory",
+ ],
+)
+
+cc_library(
+ name = "input_colocation_exemption_registry",
+ srcs = ["input_colocation_exemption_registry.cc"],
+ hdrs = ["input_colocation_exemption_registry.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "local_device",
+ srcs = ["local_device.cc"],
+ hdrs = ["local_device.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ ":process_state",
+ ":process_util",
+ ":session_options",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//third_party/eigen3",
+ ],
+)
+
+cc_library(
+ name = "local_executor_params",
+ hdrs = ["local_executor_params.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "memory_types",
+ srcs = ["memory_types.cc"],
+ hdrs = ["memory_types.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "mkl_cpu_allocator",
+ srcs = ["mkl_cpu_allocator.cc"],
+ hdrs = ["mkl_cpu_allocator.h"],
+ copts = tf_copts(),
+ deps = [
+ ":bfc_allocator",
+ ":pool_allocator",
+ "//tensorflow/core:lib",
+ ] + mkl_deps(),
+)
+
+cc_library(
+ name = "optimization_registry",
+ srcs = ["optimization_registry.cc"],
+ hdrs = ["optimization_registry.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device_set",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:framework_internal",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "parallel_concat_optimizer",
+ srcs = ["parallel_concat_optimizer.cc"],
+ copts = tf_copts(),
+ deps = [
+ ":optimization_registry",
+ "//tensorflow/core:graph",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "partitioning_utils",
+ srcs = ["partitioning_utils.cc"],
+ hdrs = ["partitioning_utils.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device_set",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "pending_counts",
+ hdrs = ["pending_counts.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "pool_allocator",
+ srcs = ["pool_allocator.cc"],
+ hdrs = ["pool_allocator.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "process_state",
+ srcs = ["process_state.cc"],
+ hdrs = ["process_state.h"],
+ copts = tf_copts(),
+ deps = [
+ ":bfc_allocator",
+ ":pool_allocator",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/core/util:env_var",
+ "@com_google_absl//absl/base",
+ ],
+)
+
+cc_library(
+ name = "process_util",
+ srcs = ["process_util.cc"],
+ hdrs = ["process_util.h"],
+ copts = tf_copts() + tf_openmp_copts(),
+ deps = [
+ ":session_options",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ ],
+)
+
+cc_library(
+ name = "profile_handler",
+ hdrs = ["profile_handler.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "propagator_debug_utils",
+ srcs = ["propagator_debug_utils.cc"],
+ hdrs = ["propagator_debug_utils.h"],
+ copts = tf_copts(),
+ deps = [
+ ":entry",
+ ":graph_view",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "propagator_state",
+ srcs = ["propagator_state.cc"],
+ hdrs = ["propagator_state.h"],
+ copts = tf_copts(),
+ deps = [
+ ":entry",
+ ":graph_view",
+ ":immutable_executor_state",
+ ":pending_counts",
+ ":propagator_debug_utils",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+)
+
+cc_library(
+ name = "renamed_device",
+ srcs = ["renamed_device.cc"],
+ hdrs = ["renamed_device.h"],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//third_party/eigen3",
+ "@com_google_absl//absl/memory",
+ ],
+)
+
+cc_library(
+ name = "rendezvous_mgr",
+ srcs = ["rendezvous_mgr.cc"],
+ hdrs = ["rendezvous_mgr.h"],
+ copts = tf_copts(),
+ deps = [
+ ":copy_tensor",
+ ":device",
+ ":device_mgr",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "ring_alg",
+ srcs = ["ring_alg.cc"],
+ hdrs = ["ring_alg.h"],
+ copts = tf_copts(),
+ deps = [
+ ":base_collective_executor",
+ ":collective_rma_local",
+ ":collective_util",
+ ":copy_tensor",
+ ":device",
+ ":device_mgr",
+ ":dma_helper",
+ ":process_util",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "ring_gatherer",
+ srcs = ["ring_gatherer.cc"],
+ hdrs = ["ring_gatherer.h"],
+ copts = tf_copts(),
+ deps = [
+ ":base_collective_executor",
+ ":collective_rma_local",
+ ":collective_util",
+ ":copy_tensor",
+ ":device",
+ ":device_mgr",
+ ":dma_helper",
+ ":process_util",
+ ":ring_alg",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "ring_reducer",
+ srcs = ["ring_reducer.cc"],
+ hdrs = ["ring_reducer.h"],
+ copts = tf_copts(),
+ deps = [
+ ":base_collective_executor",
+ ":collective_rma_local",
+ ":collective_util",
+ ":copy_tensor",
+ ":device",
+ ":device_mgr",
+ ":dma_helper",
+ ":process_util",
+ ":ring_alg",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "rendezvous_util",
+ srcs = ["rendezvous_util.cc"],
+ hdrs = ["rendezvous_util.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "replicate_per_replica_nodes",
+ srcs = ["replicate_per_replica_nodes.cc"],
+ hdrs = ["replicate_per_replica_nodes.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "scoped_allocator",
+ srcs = [
+ "scoped_allocator.cc",
+ "scoped_allocator_mgr.cc",
+ ],
+ hdrs = [
+ "scoped_allocator.h",
+ "scoped_allocator_mgr.h",
+ ],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "session",
+ srcs = ["session.cc"],
+ hdrs = ["//tensorflow/core/public:session.h"],
+ copts = tf_copts(),
+ deps = [
+ ":session_factory",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "session_factory",
+ srcs = ["session_factory.cc"],
+ hdrs = ["session_factory.h"],
+ copts = tf_copts(),
+ deps = [
+ ":session_options",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "session_options",
+ srcs = ["session_options.cc"],
+ hdrs = [
+ "//tensorflow/core/public:session_options.h",
+ ],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "simple_propagator_state",
+ srcs = ["simple_propagator_state.cc"],
+ hdrs = ["simple_propagator_state.h"],
+ copts = tf_copts(),
+ deps = [
+ ":entry",
+ ":graph_view",
+ ":immutable_executor_state",
+ ":pending_counts",
+ ":propagator_debug_utils",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core/profiler/lib:traceme",
+ ],
+)
+
+cc_library(
+ name = "single_threaded_cpu_device",
+ srcs = ["single_threaded_cpu_device.cc"],
+ hdrs = [
+ "single_threaded_cpu_device.h",
+ ],
+ copts = tf_copts(),
+ deps = [
+ ":device",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:protos_all_cc",
+ "//third_party/eigen3",
+ ],
+)
+
+cc_library(
+ name = "session_state",
+ srcs = ["session_state.cc"],
+ hdrs = ["//tensorflow/core/framework:session_state.h"],
+ copts = tf_copts(),
+ deps = [
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ ],
+)
+
+cc_library(
+ name = "stats_publisher_interface",
+ srcs = ["stats_publisher_interface.cc"],
+ hdrs = ["stats_publisher_interface.h"],
+ copts = tf_copts(),
+ deps = [
+ ":build_graph_options",
+ ":profile_handler",
+ ":session_options",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "step_stats_collector",
+ srcs = ["step_stats_collector.cc"],
+ hdrs = ["step_stats_collector.h"],
+ copts = tf_copts(),
+ deps = [
+ ":costmodel_manager",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "threadpool_device",
+ srcs = ["threadpool_device.cc"],
+ hdrs = ["threadpool_device.h"],
+ copts = tf_copts() + tf_openmp_copts(),
+ deps = [
+ ":device_factory",
+ ":local_device",
+ ":scoped_allocator",
+ ":session_options",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:graph",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core:protos_all_cc",
+ ],
+)
+
+cc_library(
+ name = "threadpool_device_factory",
+ srcs = ["threadpool_device_factory.cc"],
+ copts = tf_copts(),
+ deps = [
+ ":device_factory",
+ ":process_state",
+ ":session_options",
+ ":threadpool_device",
+ "//tensorflow/core:framework",
+ "//tensorflow/core:lib",
+ "@com_google_absl//absl/memory",
+ ],
+ alwayslink = 1,
+)
+
+tf_cuda_library(
+ name = "core_cpu_rump_impl",
+ srcs = [
+ "colocation_graph.cc",
+ "composite_device.cc",
+ "constant_folding.cc",
"function.cc",
"function_optimization_registry.cc",
"graph_optimizer.cc",
"graph_runner.cc",
- "graph_view.cc",
- "hierarchical_tree_broadcaster.cc",
- "immutable_executor_state.cc",
- "input_colocation_exemption_registry.cc",
"inspecting_placer.cc",
"isolate_placer_inspection_required_ops_pass.cc",
- "local_device.cc",
"lower_case_op.cc",
"lower_function_call_op.cc",
"lower_functional_ops.cc",
"lower_if_op.cc",
"lower_while_op.cc",
- "memory_types.cc",
- "mkl_cpu_allocator.cc",
- "optimization_registry.cc",
- "parallel_concat_optimizer.cc",
- "partitioning_utils.cc",
"placer.cc",
"placer_inspection_required_ops_utils.cc",
"placer_inspection_required_ops_utils.h",
- "pool_allocator.cc",
"process_function_library_runtime.cc",
- "process_state.cc",
- "process_util.cc",
- "propagator_debug_utils.cc",
- "propagator_state.cc",
- "renamed_device.cc",
- "rendezvous_mgr.cc",
- "rendezvous_util.cc",
- "replicate_per_replica_nodes.cc",
- "ring_alg.cc",
- "ring_gatherer.cc",
- "ring_reducer.cc",
- "session.cc",
- "session_factory.cc",
- "session_options.cc",
- "session_state.cc",
- "simple_propagator_state.cc",
- "single_threaded_cpu_device.cc",
- "stats_publisher_interface.cc",
- "step_stats_collector.cc",
- "threadpool_device.cc",
- "threadpool_device_factory.cc",
"//tensorflow/core/graph:core_cpu_impl_srcs",
- "//tensorflow/core/public:session.h",
- "//tensorflow/core/public:session_options.h",
],
hdrs = [":core_cpu_lib_headers"],
- copts = tf_copts() + tf_openmp_copts(),
+ copts = tf_copts(),
deps = [
- ":bfc_allocator",
+ ":device",
+ ":entry",
+ ":executor",
+ ":executor_factory",
+ ":graph_view",
+ ":local_executor_params",
+ ":immutable_executor_state",
+ ":input_colocation_exemption_registry",
+ ":pending_counts",
+ ":propagator_debug_utils",
+ ":propagator_state",
+ ":session_options",
+ ":simple_propagator_state",
+ ":single_threaded_cpu_device",
"//tensorflow/core:graph",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
"//tensorflow/core:protos_all_cc",
- "@com_google_absl//absl/base",
"@com_google_absl//absl/algorithm:container",
"@com_google_absl//absl/container:flat_hash_map",
"@com_google_absl//absl/container:flat_hash_set",
@@ -355,7 +1110,6 @@
"@com_google_absl//absl/strings",
"@com_google_absl//absl/types:optional",
"@com_google_absl//absl/types:variant",
- "//third_party/eigen3",
"//tensorflow/core/public:version",
"//tensorflow/core/grappler/utils:functions",
"//tensorflow/core/profiler/lib:annotated_traceme",
@@ -366,6 +1120,63 @@
)
tf_cuda_library(
+ name = "core_cpu_impl",
+ hdrs = [":core_cpu_lib_headers"],
+ copts = tf_copts(),
+ deps = [
+ ":accumulate_n_optimizer",
+ ":base_collective_executor",
+ ":bfc_allocator",
+ ":buf_rendezvous",
+ ":build_graph_options",
+ ":collective_executor_mgr",
+ ":collective_param_resolver_local",
+ ":collective_rma_local",
+ ":collective_util",
+ ":copy_tensor",
+ ":core_cpu_rump_impl",
+ ":costmodel_manager",
+ ":debugger_state_interface",
+ ":device",
+ ":device_factory",
+ ":device_mgr",
+ ":device_resolver_local",
+ ":device_set",
+ ":entry",
+ ":graph_view",
+ ":hierarchical_tree_broadcaster",
+ ":input_colocation_exemption_registry",
+ ":local_device",
+ ":memory_types",
+ ":mkl_cpu_allocator",
+ ":optimization_registry",
+ ":parallel_concat_optimizer",
+ ":partitioning_utils",
+ ":pending_counts",
+ ":pool_allocator",
+ ":process_state",
+ ":process_util",
+ ":profile_handler",
+ ":renamed_device",
+ ":rendezvous_mgr",
+ ":rendezvous_util",
+ ":replicate_per_replica_nodes",
+ ":ring_alg",
+ ":ring_gatherer",
+ ":ring_reducer",
+ ":session",
+ ":session_factory",
+ ":session_options",
+ ":session_state",
+ ":single_threaded_cpu_device",
+ ":stats_publisher_interface",
+ ":step_stats_collector",
+ ":threadpool_device",
+ ":threadpool_device_factory",
+ ],
+)
+
+tf_cuda_library(
name = "core_cpu_lib",
hdrs = [":core_cpu_lib_headers"],
deps = [
@@ -529,6 +1340,22 @@
],
)
+tf_cc_test(
+ name = "composite_device_test",
+ size = "small",
+ srcs = [
+ "composite_device_test.cc",
+ ],
+ linkstatic = tf_kernel_tests_linkstatic(),
+ deps = [
+ ":core_cpu",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/core:test",
+ "//tensorflow/core:test_main",
+ "//tensorflow/core:testlib",
+ ],
+)
+
tf_cc_tests(
name = "core_higher_level_tests",
size = "small",
@@ -562,6 +1389,7 @@
":core_cpu",
":core_cpu_internal",
":direct_session_internal",
+ ":pending_counts",
"//tensorflow/cc:cc_ops",
"//tensorflow/cc:cc_ops_internal",
"//tensorflow/cc:function_ops",
@@ -636,7 +1464,6 @@
":core",
":core_cpu",
":core_cpu_internal",
- ":direct_session_internal",
"//tensorflow/core:all_kernels",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
@@ -665,7 +1492,6 @@
":core",
":core_cpu",
":core_cpu_internal",
- ":direct_session_internal",
"//tensorflow/core:all_kernels",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
@@ -694,7 +1520,6 @@
":core",
":core_cpu",
":core_cpu_internal",
- ":direct_session_internal",
"//tensorflow/core:all_kernels",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
@@ -739,9 +1564,7 @@
linkstatic = tf_kernel_tests_linkstatic(),
tags = tf_cuda_tests_tags(),
deps = [
- ":core",
- ":core_cpu",
- ":core_cpu_internal",
+ ":memory_types",
"//tensorflow/cc:cc_ops",
"//tensorflow/core:framework",
"//tensorflow/core:framework_internal",
@@ -854,7 +1677,7 @@
srcs = ["process_util_test.cc"],
linkstatic = tf_kernel_tests_linkstatic(),
deps = [
- ":core_cpu_internal",
+ ":process_util",
"//tensorflow/core:test",
"//tensorflow/core:test_main",
],
@@ -866,7 +1689,7 @@
srcs = ["rendezvous_util_test.cc"],
linkstatic = tf_kernel_tests_linkstatic(),
deps = [
- ":core_cpu_internal",
+ ":rendezvous_util",
"//tensorflow/core:lib",
"//tensorflow/core:test",
"//tensorflow/core:test_main",
@@ -879,7 +1702,7 @@
srcs = ["replicate_per_replica_nodes_test.cc"],
linkstatic = tf_kernel_tests_linkstatic(),
deps = [
- ":core_cpu_internal",
+ ":replicate_per_replica_nodes",
"//tensorflow/cc:cc_ops",
"//tensorflow/cc:cc_ops_internal",
"//tensorflow/cc:function_ops",
@@ -1164,8 +1987,8 @@
srcs = ["scoped_allocator_mgr_test.cc"],
linkstatic = tf_kernel_tests_linkstatic(),
deps = [
- ":core_cpu",
- ":core_cpu_internal",
+ ":dma_helper",
+ ":scoped_allocator",
"//tensorflow/core:framework",
"//tensorflow/core:lib",
"//tensorflow/core:test",
@@ -1178,8 +2001,7 @@
size = "small",
srcs = ["input_colocation_exemption_registry_test.cc"],
deps = [
- ":core_cpu",
- ":core_cpu_internal",
+ ":input_colocation_exemption_registry",
"//tensorflow/core:test",
"//tensorflow/core:test_main",
"//tensorflow/core:testlib",
diff --git a/tensorflow/core/common_runtime/base_collective_executor.cc b/tensorflow/core/common_runtime/base_collective_executor.cc
index de2dc28..7a614a8 100644
--- a/tensorflow/core/common_runtime/base_collective_executor.cc
+++ b/tensorflow/core/common_runtime/base_collective_executor.cc
@@ -21,9 +21,7 @@
#include "tensorflow/core/common_runtime/copy_tensor.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/dma_helper.h"
-#include "tensorflow/core/common_runtime/hierarchical_tree_broadcaster.h"
#include "tensorflow/core/common_runtime/process_util.h"
-#include "tensorflow/core/common_runtime/ring_reducer.h"
#include "tensorflow/core/framework/allocator.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor.h"
diff --git a/tensorflow/core/common_runtime/bfc_allocator.cc b/tensorflow/core/common_runtime/bfc_allocator.cc
index f65cfcf..f9c0ba1 100644
--- a/tensorflow/core/common_runtime/bfc_allocator.cc
+++ b/tensorflow/core/common_runtime/bfc_allocator.cc
@@ -441,8 +441,30 @@
return nullptr;
}
+int64 BFCAllocator::LargestFreeChunk() {
+ for (int i = kNumBins - 1; i >= 0; i--) {
+ if (!BinFromIndex(i)->free_chunks.empty()) {
+ return ChunkFromHandle(*BinFromIndex(i)->free_chunks.rbegin())->size;
+ }
+ }
+ return 0;
+}
+
+double BFCAllocator::GetFragmentation() {
+ int64 bytes_available = total_region_allocated_bytes_ - stats_.bytes_in_use;
+ DCHECK_GT(bytes_available, 0);
+ return static_cast<double>(bytes_available - LargestFreeChunk()) /
+ bytes_available;
+}
+
+void BFCAllocator::AddTraceMe(absl::string_view traceme_name, const void* ptr) {
+ BFCAllocator::Chunk* chunk = ChunkFromHandle(region_manager_.get_handle(ptr));
+ AddTraceMe(traceme_name, chunk->ptr, chunk->requested_size, chunk->size);
+}
+
void BFCAllocator::AddTraceMe(absl::string_view traceme_name,
- const void* chunk_ptr) {
+ const void* chunk_ptr, int64 req_bytes,
+ int64 alloc_bytes) {
// Internal users will see the memory profile with default trace level.
auto traceme_level = profiler::TraceMeLevel::kVerbose;
#ifdef PLATFORM_GOOGLE
@@ -454,21 +476,19 @@
AllocatorStats stats = stats_;
int64 bytes_available =
memory_limit_ - stats.bytes_reserved - stats.bytes_in_use;
- BFCAllocator::Chunk* chunk =
- ChunkFromHandle(region_manager_.get_handle(chunk_ptr));
const auto& annotation =
ScopedMemoryDebugAnnotation::CurrentAnnotation();
std::string tensor_shape = annotation.pending_shape
? annotation.pending_shape->DebugString()
: "";
-
return absl::StrCat(traceme_name, "#allocator_name=", name_,
",bytes_reserved=", stats.bytes_reserved,
",bytes_allocated=", stats.bytes_in_use,
",bytes_available=", bytes_available,
+ ",fragmentation=", GetFragmentation(),
",peak_bytes_in_use=", stats.peak_bytes_in_use,
- ",requested_bytes=", chunk->requested_size,
- ",allocation_bytes=", chunk->size,
+ ",requested_bytes=", req_bytes,
+ ",allocation_bytes=", alloc_bytes,
",addr=", reinterpret_cast<uint64>(chunk_ptr),
",tf_op=", annotation.pending_op_name,
",id=", annotation.pending_step_id,
@@ -613,11 +633,13 @@
// Find the chunk from the ptr.
BFCAllocator::ChunkHandle h = region_manager_.get_handle(ptr);
CHECK(h != kInvalidChunkHandle);
+ // Record chunk information before it's freed.
+ Chunk* chunk = ChunkFromHandle(h);
+ void* chunk_ptr = chunk->ptr;
+ int64 req_bytes = chunk->requested_size;
+ int64 alloc_bytes = chunk->size;
MarkFree(h);
- // TraceMe needs to be added after MarkFree and before InsertFreeChunkIntoBin
- // for correct memory stats.
- AddTraceMe("MemoryDeallocation", ptr);
// Consider coalescing it.
if (timing_counter_) {
@@ -627,6 +649,10 @@
InsertFreeChunkIntoBin(TryToCoalesce(h, false));
}
+ // TraceMe needs to be added after MarkFree and InsertFreeChunkIntoBin for
+ // correct aggregation stats (bytes_in_use, fragmentation).
+ AddTraceMe("MemoryDeallocation", chunk_ptr, req_bytes, alloc_bytes);
+
if (VLOG_IS_ON(4)) {
LOG(INFO) << "F: " << RenderOccupancy();
}
@@ -1115,31 +1141,6 @@
return md;
}
-double BFCAllocator::GetFragmentation() {
- int64 largest_free_chunk = 0;
- int64 free_bytes = 0;
- for (const auto& region : region_manager_.regions()) {
- ChunkHandle chunk_handle = region_manager_.get_handle(region.ptr());
- while (chunk_handle != kInvalidChunkHandle) {
- const Chunk* chunk = ChunkFromHandle(chunk_handle);
- if (!chunk->in_use()) {
- free_bytes += chunk->size;
- if (chunk->size > largest_free_chunk) {
- largest_free_chunk = chunk->size;
- }
- }
- chunk_handle = chunk->next;
- }
- }
- double frag_metric = 0.0;
- if (free_bytes > 0) {
- frag_metric =
- (free_bytes - largest_free_chunk) / static_cast<double>(free_bytes);
- }
-
- return frag_metric;
-}
-
absl::optional<AllocatorStats> BFCAllocator::GetStats() {
mutex_lock l(lock_);
return stats_;
diff --git a/tensorflow/core/common_runtime/bfc_allocator.h b/tensorflow/core/common_runtime/bfc_allocator.h
index 94506bb..509fa9e 100644
--- a/tensorflow/core/common_runtime/bfc_allocator.h
+++ b/tensorflow/core/common_runtime/bfc_allocator.h
@@ -115,21 +115,30 @@
bool MergeTimestampedChunks(size_t required_bytes)
TF_EXCLUSIVE_LOCKS_REQUIRED(lock_);
+ // Return the largest free chunk bytes from the largest bin in constant time.
+ // The free chunks are sorted by size (and then address) in a bin.
+ int64 LargestFreeChunk() TF_EXCLUSIVE_LOCKS_REQUIRED(lock_);
+
// Add TraceMe (in memory allocation and deallocation) for memory stats
// profiling. The chunk_ptr is passed to get information such as address,
// chunk size and requested_size.
- void AddTraceMe(absl::string_view traceme_name, const void* chunk_ptr)
+ void AddTraceMe(absl::string_view traceme_name, const void* ptr)
+ TF_EXCLUSIVE_LOCKS_REQUIRED(lock_);
+
+ // Overloaded AddTraceMe function with chunk information.
+ void AddTraceMe(absl::string_view traceme_name, const void* chunk_ptr,
+ int64 req_bytes, int64 alloc_bytes)
TF_EXCLUSIVE_LOCKS_REQUIRED(lock_);
// A ChunkHandle is an index into the chunks_ vector in BFCAllocator
// kInvalidChunkHandle means an invalid chunk
typedef size_t ChunkHandle;
- static const int kInvalidChunkHandle = -1;
+ static constexpr int kInvalidChunkHandle = -1;
typedef int BinNum;
- static const int kInvalidBinNum = -1;
+ static constexpr int kInvalidBinNum = -1;
// The following means that the largest bin'd chunk size is 256 << 21 = 512MB.
- static const int kNumBins = 21;
+ static constexpr int kNumBins = 21;
// A Chunk points to a piece of memory that's either entirely free or entirely
// in use by one user memory allocation.
@@ -243,8 +252,8 @@
: bin_size(bs), free_chunks(ChunkComparator(allocator)) {}
};
- static const size_t kMinAllocationBits = 8;
- static const size_t kMinAllocationSize = 1 << kMinAllocationBits;
+ static constexpr size_t kMinAllocationBits = 8;
+ static constexpr size_t kMinAllocationSize = 1 << kMinAllocationBits;
// BFCAllocator allocates memory into a collection of disjoint
// AllocationRegions. Each AllocationRegion corresponds to one call to
diff --git a/tensorflow/core/common_runtime/buf_rendezvous.cc b/tensorflow/core/common_runtime/buf_rendezvous.cc
index 6733a2e..49cc9fd 100644
--- a/tensorflow/core/common_runtime/buf_rendezvous.cc
+++ b/tensorflow/core/common_runtime/buf_rendezvous.cc
@@ -18,6 +18,7 @@
#include "absl/strings/str_cat.h"
#include "absl/strings/string_view.h"
#include "tensorflow/core/common_runtime/device.h"
+#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/notification.h"
diff --git a/tensorflow/core/common_runtime/buf_rendezvous.h b/tensorflow/core/common_runtime/buf_rendezvous.h
index 527d0e2..74857e4 100644
--- a/tensorflow/core/common_runtime/buf_rendezvous.h
+++ b/tensorflow/core/common_runtime/buf_rendezvous.h
@@ -20,7 +20,6 @@
#include "absl/container/flat_hash_map.h"
#include "absl/strings/string_view.h"
-#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/framework/allocator.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/mutex.h"
@@ -28,6 +27,7 @@
namespace tensorflow {
class Device;
class DeviceContext;
+class DeviceMgr;
class Tensor;
// EXPERIMENTAL: RDMA oriented producer/consumer rendezvous on a local
diff --git a/tensorflow/core/common_runtime/buf_rendezvous_test.cc b/tensorflow/core/common_runtime/buf_rendezvous_test.cc
index c9a27b5..2701651 100644
--- a/tensorflow/core/common_runtime/buf_rendezvous_test.cc
+++ b/tensorflow/core/common_runtime/buf_rendezvous_test.cc
@@ -14,6 +14,8 @@
==============================================================================*/
#include "tensorflow/core/common_runtime/buf_rendezvous.h"
+#include "tensorflow/core/common_runtime/device.h"
+#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/core/notification.h"
diff --git a/tensorflow/core/common_runtime/build_graph_options.h b/tensorflow/core/common_runtime/build_graph_options.h
index 24b71cc..4bc953d 100644
--- a/tensorflow/core/common_runtime/build_graph_options.h
+++ b/tensorflow/core/common_runtime/build_graph_options.h
@@ -32,7 +32,7 @@
// TODO(mrry): Remove this when the distributed runtime supports Arg/Retval.
bool use_function_convention = false;
- static const int64 kNoCollectiveGraphKey = 0;
+ static constexpr int64 kNoCollectiveGraphKey = 0;
int64 collective_graph_key = kNoCollectiveGraphKey;
// If not `kNone`, order all CollectiveReduce operations statically and
diff --git a/tensorflow/core/common_runtime/collective_param_resolver_local.cc b/tensorflow/core/common_runtime/collective_param_resolver_local.cc
index ad67ce5..f3dea5c 100644
--- a/tensorflow/core/common_runtime/collective_param_resolver_local.cc
+++ b/tensorflow/core/common_runtime/collective_param_resolver_local.cc
@@ -31,6 +31,7 @@
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/types.h"
+#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/util/device_name_utils.h"
namespace tensorflow {
diff --git a/tensorflow/core/common_runtime/collective_param_resolver_local.h b/tensorflow/core/common_runtime/collective_param_resolver_local.h
index f1bdf63..c724ed9 100644
--- a/tensorflow/core/common_runtime/collective_param_resolver_local.h
+++ b/tensorflow/core/common_runtime/collective_param_resolver_local.h
@@ -23,13 +23,13 @@
#include "tensorflow/core/framework/collective.h"
#include "tensorflow/core/lib/gtl/flatmap.h"
-#include "tensorflow/core/protobuf/config.pb.h"
namespace tensorflow {
class CompleteGroupRequest;
class CompleteGroupResponse;
class CompleteInstanceRequest;
class CompleteInstanceResponse;
+class ConfigProto;
class DeviceMgr;
// Implements ParamResolverInterface for a single-task context.
diff --git a/tensorflow/core/common_runtime/colocation_graph.cc b/tensorflow/core/common_runtime/colocation_graph.cc
index 41058ae..f731edc 100644
--- a/tensorflow/core/common_runtime/colocation_graph.cc
+++ b/tensorflow/core/common_runtime/colocation_graph.cc
@@ -26,6 +26,7 @@
#include "absl/container/flat_hash_set.h"
#include "absl/strings/str_join.h"
#include "absl/types/optional.h"
+#include "tensorflow/core/common_runtime/composite_device.h"
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_set.h"
#include "tensorflow/core/common_runtime/function.h"
@@ -136,6 +137,10 @@
device_type == "TPU");
}
+bool IsCompositeDevice(absl::string_view device_type) {
+ return device_type == kCompositeDeviceType;
+}
+
} // namespace
Status Member::SetParentAndSupportedDevices(
@@ -220,6 +225,26 @@
return Status::OK();
}
+bool Member::IsEdgeFromCompositeDeviceToPhysicalDevice(
+ const Member& src_root) const {
+ auto compatible_edge_from_composite_device_to_physical_device =
+ [](const DeviceNameUtils::ParsedName& src_device,
+ const DeviceNameUtils::ParsedName& dst_device) -> bool {
+ return src_device.has_type && dst_device.has_type &&
+ IsCompositeDevice(src_device.type) &&
+ !IsCompositeDevice(dst_device.type);
+ };
+ if (compatible_edge_from_composite_device_to_physical_device(
+ src_root.assigned_device_name_, assigned_device_name_) ||
+ compatible_edge_from_composite_device_to_physical_device(
+ src_root.resource_device_name_, resource_device_name_) ||
+ compatible_edge_from_composite_device_to_physical_device(
+ src_root.requested_device_name_, requested_device_name_)) {
+ return true;
+ }
+ return false;
+}
+
Status Member::EnsureCompatibilityAcrossResourceEdge(
const Node& src, const Member& src_root,
const Node& dst, /*dst_root is this*/
@@ -484,7 +509,10 @@
void Member::MaybeExcludeXlaDevices() {
for (const auto& parsed_name :
{requested_device_name_, assigned_device_name_, resource_device_name_}) {
- if (parsed_name.has_type && IsXlaDevice(parsed_name.type)) {
+ // Don't exculde XLA devices from supported devices if member is explicitly
+ // assigned to a CompositeDevice.
+ if (parsed_name.has_type && (IsXlaDevice(parsed_name.type) ||
+ IsCompositeDevice(parsed_name.type))) {
return;
}
}
@@ -664,6 +692,12 @@
auto& src_root = members_[src_root_id];
auto& dst_root = members_[dst_root_id];
+ if (dst_root.IsEdgeFromCompositeDeviceToPhysicalDevice(src_root)) {
+ // If the src root is assigned to a composite device and the dst root is
+ // assigned to a physical device, don't colocate the dst root with the src
+ // root.
+ return Status::OK();
+ }
TF_RETURN_IF_ERROR(dst_root.EnsureCompatibilityAcrossResourceEdge(
*src, src_root, *dst, log_device_placement_));
Status status = ColocateNodes(*src, src_root_id, *dst, dst_root_id);
@@ -890,6 +924,15 @@
return Status::OK();
}
+// Returns whether the device_type in `device_attributes` is supported.
+bool IsSupportedDeviceType(const DeviceAttributes& device_attributes,
+ const DeviceType& supported_type) {
+ if (DeviceType(device_attributes.device_type()) == supported_type) {
+ return true;
+ }
+ return IsCompositeDevice(device_attributes.device_type());
+}
+
} // namespace
Status ColocationGraph::ApplyIOColocationGroups(
@@ -1364,7 +1407,7 @@
}
for (const auto& d : member->supported_device_types()) {
- if (DeviceType(assigned_device->attributes().device_type()) == d.first) {
+ if (IsSupportedDeviceType(assigned_device->attributes(), d.first)) {
return Status::OK();
}
}
@@ -1434,8 +1477,8 @@
PrioritizedDeviceVector prioritized_filtered_devices;
for (const auto& supported_device_type : supported_device_types) {
for (Device* device : devices) {
- if (DeviceType(device->attributes().device_type()) ==
- supported_device_type.first) {
+ if (IsSupportedDeviceType(device->attributes(),
+ supported_device_type.first)) {
if (default_local_device &&
(device == default_local_device ||
// TODO(nareshmodi, fishx): At times the device pointer in the
diff --git a/tensorflow/core/common_runtime/colocation_graph.h b/tensorflow/core/common_runtime/colocation_graph.h
index d0714d5..c9f48a1 100644
--- a/tensorflow/core/common_runtime/colocation_graph.h
+++ b/tensorflow/core/common_runtime/colocation_graph.h
@@ -51,6 +51,10 @@
Status FillPossibleDevices(PossibleDevices* possible_device) const;
+ // Returns whether `src_root` is assigned to a CompositeDevice and `this` is
+ // assigned to a physical device.
+ bool IsEdgeFromCompositeDeviceToPhysicalDevice(const Member& src_root) const;
+
Status EnsureCompatibilityAcrossResourceEdge(
const Node& src, const Member& src_root,
const Node& dst, /*dst_root is this*/
diff --git a/tensorflow/core/common_runtime/composite_device.cc b/tensorflow/core/common_runtime/composite_device.cc
new file mode 100644
index 0000000..b726515
--- /dev/null
+++ b/tensorflow/core/common_runtime/composite_device.cc
@@ -0,0 +1,69 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/common_runtime/composite_device.h"
+
+#include "absl/strings/str_join.h"
+#include "tensorflow/core/util/device_name_utils.h"
+
+namespace tensorflow {
+
+const char* const kCompositeDeviceType = "COMPOSITE";
+
+std::unique_ptr<CompositeDevice> CompositeDevice::MakeDevice(
+ const std::vector<string>& underlying_devices, const int unique_device_id,
+ Status* status) {
+ if (underlying_devices.empty()) {
+ status->Update(
+ errors::InvalidArgument("underlying_devices should not be empty."));
+ return nullptr;
+ }
+ DeviceNameUtils::ParsedName parsed_name;
+ if (!DeviceNameUtils::ParseFullName(underlying_devices.at(0), &parsed_name)) {
+ status->Update(tensorflow::errors::InvalidArgument(
+ "Cannot parse device name ", underlying_devices.at(0),
+ " when creating CompositeDevice."));
+ return nullptr;
+ }
+ const string& underlying_type = parsed_name.type;
+ for (int i = 1; i < underlying_devices.size(); ++i) {
+ DeviceNameUtils::ParsedName name;
+ if (!DeviceNameUtils::ParseFullName(underlying_devices.at(i), &name)) {
+ status->Update(tensorflow::errors::InvalidArgument(
+ "Cannot parse device name ", underlying_devices.at(i),
+ " when creating CompositeDevice."));
+ return nullptr;
+ }
+ if (name.type != underlying_type) {
+ status->Update(tensorflow::errors::InvalidArgument(
+ "Expect device type ", parsed_name.type, "; but got type ", name.type,
+ " from device: ", underlying_devices.at(i),
+ " when creating CompositeDevice."));
+ return nullptr;
+ }
+ }
+ DeviceAttributes device_attributes;
+ parsed_name.type = kCompositeDeviceType;
+ device_attributes.set_device_type(parsed_name.type);
+ parsed_name.id = unique_device_id;
+ const string composite_name =
+ DeviceNameUtils::ParsedNameToString(parsed_name);
+ device_attributes.set_name(composite_name);
+
+ return absl::WrapUnique(
+ new CompositeDevice(device_attributes, underlying_devices));
+}
+
+} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/composite_device.h b/tensorflow/core/common_runtime/composite_device.h
new file mode 100644
index 0000000..127e5b8
--- /dev/null
+++ b/tensorflow/core/common_runtime/composite_device.h
@@ -0,0 +1,63 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_COMMON_RUNTIME_COMPOSITE_DEVICE_H_
+#define TENSORFLOW_CORE_COMMON_RUNTIME_COMPOSITE_DEVICE_H_
+
+#include "absl/strings/string_view.h"
+#include "tensorflow/core/common_runtime/device.h"
+#include "tensorflow/core/framework/allocator.h"
+#include "tensorflow/core/framework/device_attributes.pb.h"
+#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/lib/core/status.h"
+
+namespace tensorflow {
+
+extern const char* const kCompositeDeviceType;
+
+// A virtual device which represents a set of devices. We don't execute any
+// op on this virtial device.
+class CompositeDevice : public Device {
+ public:
+ Status Sync() override {
+ return errors::Internal(
+ "Sync() should never been invoked on CompositeDevice.");
+ }
+
+ Allocator* GetAllocator(AllocatorAttributes) override { return nullptr; }
+
+ const std::vector<string>* underlying_devices() const {
+ return &underlying_devices_;
+ }
+
+ // Helper for creating a CompositeDevice.
+ static std::unique_ptr<CompositeDevice> MakeDevice(
+ const std::vector<string>& underlying_devices, const int unique_device_id,
+ Status* status);
+
+ private:
+ CompositeDevice(const DeviceAttributes& device_attributes,
+ const std::vector<string>& underlying_devices)
+ : Device(/*env=*/nullptr, device_attributes),
+ underlying_devices_(underlying_devices) {}
+
+ const std::vector<string> underlying_devices_;
+
+ TF_DISALLOW_COPY_AND_ASSIGN(CompositeDevice);
+};
+
+} // namespace tensorflow
+
+#endif // TENSORFLOW_CORE_COMMON_RUNTIME_COMPOSITE_DEVICE_H_
diff --git a/tensorflow/core/common_runtime/composite_device_test.cc b/tensorflow/core/common_runtime/composite_device_test.cc
new file mode 100644
index 0000000..73c87b5
--- /dev/null
+++ b/tensorflow/core/common_runtime/composite_device_test.cc
@@ -0,0 +1,65 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/common_runtime/composite_device.h"
+
+#include "tensorflow/core/lib/core/status_test_util.h"
+
+namespace tensorflow {
+
+TEST(CompositeDeviceTest, Basic) {
+ std::vector<string> underlying_devices;
+ {
+ Status status;
+ std::unique_ptr<CompositeDevice> composite_device =
+ CompositeDevice::MakeDevice(underlying_devices, /*unique_device_id=*/0,
+ &status);
+ EXPECT_EQ(composite_device, nullptr);
+ EXPECT_EQ(error::INVALID_ARGUMENT, status.code());
+ EXPECT_TRUE(absl::StrContains(status.error_message(),
+ "underlying_devices should not be empty"))
+ << status.ToString();
+ }
+
+ {
+ Status status;
+ underlying_devices.push_back(
+ "/job:localhost/replica:0/task:0/device:CPU:0");
+ underlying_devices.push_back(
+ "/job:localhost/replica:0/task:0/device:CPU:1");
+ std::unique_ptr<CompositeDevice> composite_device =
+ CompositeDevice::MakeDevice(underlying_devices, /*unique_device_id=*/0,
+ &status);
+ TF_ASSERT_OK(status);
+ EXPECT_EQ(composite_device->device_type(), kCompositeDeviceType);
+ EXPECT_EQ(underlying_devices, *composite_device->underlying_devices());
+ }
+
+ {
+ Status status;
+ underlying_devices.push_back(
+ "/job:localhost/replica:0/task:0/device:GPU:0");
+ std::unique_ptr<CompositeDevice> composite_device =
+ CompositeDevice::MakeDevice(underlying_devices, /*unique_device_id=*/1,
+ &status);
+ EXPECT_EQ(composite_device, nullptr);
+ EXPECT_EQ(error::INVALID_ARGUMENT, status.code());
+ EXPECT_TRUE(absl::StrContains(status.error_message(),
+ "Expect device type CPU; but got type GPU"))
+ << status.ToString();
+ }
+}
+
+} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/device_set.cc b/tensorflow/core/common_runtime/device_set.cc
index 112769f..b062529 100644
--- a/tensorflow/core/common_runtime/device_set.cc
+++ b/tensorflow/core/common_runtime/device_set.cc
@@ -21,7 +21,6 @@
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_factory.h"
-#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/core/stringpiece.h"
#include "tensorflow/core/lib/gtl/map_util.h"
diff --git a/tensorflow/core/common_runtime/device_set.h b/tensorflow/core/common_runtime/device_set.h
index fdb7453..608705c 100644
--- a/tensorflow/core/common_runtime/device_set.h
+++ b/tensorflow/core/common_runtime/device_set.h
@@ -21,7 +21,6 @@
#include <vector>
#include "tensorflow/core/common_runtime/device.h"
-#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/util/device_name_utils.h"
diff --git a/tensorflow/core/common_runtime/eager/attr_builder.h b/tensorflow/core/common_runtime/eager/attr_builder.h
index 1a871b0..c53286d 100644
--- a/tensorflow/core/common_runtime/eager/attr_builder.h
+++ b/tensorflow/core/common_runtime/eager/attr_builder.h
@@ -111,6 +111,12 @@
return *this;
}
+ AttrBuilder& Set(StringPiece attr_name, const AttrValue& value) {
+ AddAttrIfNotPresent(attr_name, value);
+ cached_cache_key_ = absl::nullopt;
+ return *this;
+ }
+
// Retrieves the attribute value.
// Note that Get() can involve a linear scan of all attributes with the same
// value type in this Node. This is not an issue, because Get is used rarely
diff --git a/tensorflow/core/common_runtime/eager/context.cc b/tensorflow/core/common_runtime/eager/context.cc
index b74829f..85c8c2b 100644
--- a/tensorflow/core/common_runtime/eager/context.cc
+++ b/tensorflow/core/common_runtime/eager/context.cc
@@ -110,6 +110,7 @@
#if !defined(IS_MOBILE_PLATFORM)
context_id_ = kInvalidContextId;
+ context_view_id_ = 0;
#endif // IS_MOBILE_PLATFORM
std::unique_ptr<DeviceResolverInterface> drl(
@@ -1095,11 +1096,10 @@
}
Status EagerContext::UpdateRemoteMaster(
- WorkerEnv* worker_env,
+ uint64 context_id,
std::unique_ptr<eager::EagerClientCache> remote_eager_workers,
const std::vector<string>& add_remote_contexts,
- const std::vector<string>& remove_remote_contexts, uint64 context_id,
- Rendezvous* r) {
+ const std::vector<string>& remove_remote_contexts) {
{
tf_shared_lock l(remote_state_mu_);
if (context_id != context_id_) {
@@ -1135,9 +1135,6 @@
mutex_lock l(remote_state_mu_);
context_view_id_++;
- worker_env_ = worker_env;
- if (rendezvous_ != nullptr) rendezvous_->Unref();
- rendezvous_ = r;
remote_eager_workers_ = std::move(remote_eager_workers);
pflr_->InitializeDeviceSet();
InitPrioritizedDeviceTypeList();
@@ -1337,11 +1334,8 @@
}
Status EagerContext::UpdateRemoteWorker(
- const DeviceMgr* worker_session_device_mgr,
std::unique_ptr<eager::EagerClientCache> remote_eager_workers,
- DynamicDeviceMgr* remote_device_mgr,
- const std::vector<string>& remote_contexts, uint64 context_id,
- DistributedFunctionLibraryRuntime* cluster_flr) {
+ const std::vector<string>& remote_contexts, uint64 context_id) {
{
mutex_lock l(remote_state_mu_);
if (context_id != context_id_) {
@@ -1351,15 +1345,20 @@
" but current id = ", context_id_);
}
context_view_id_++;
+
+ remote_contexts_ = remote_contexts;
+ remote_eager_workers_ = std::move(remote_eager_workers);
+ InitPrioritizedDeviceTypeList();
+ pflr_->InitializeDeviceSet();
}
- remote_contexts_ = remote_contexts;
-
- remote_eager_workers_ = std::move(remote_eager_workers);
- ResetClusterFLR(cluster_flr);
-
- remote_device_manager_.Reset(remote_device_mgr);
- InitPrioritizedDeviceTypeList();
+ // No need to update remote_device_manager_ since it's not owned for remote
+ // worker context (owned by the corresponding worker session).
+ if (remote_device_manager_.Owned()) {
+ return errors::FailedPrecondition(
+ "EagerContext::UpdateRemoteWorker failed because the context was "
+ "initialized as a master context.");
+ }
ClearCachesAndThreadExecutors();
default_executor_.ClearError();
@@ -1369,13 +1368,6 @@
entry.second->ClearError();
}
}
-
- SessionOptions options = SessionOptions();
- const auto* config = pflr_->config();
- ResetPFLR(worker_session_device_mgr, options.env, config,
- TF_GRAPH_DEF_VERSION, FuncLibDef(),
- config->graph_options().optimizer_options(), thread_pool_.get(),
- cluster_flr_.Get(), custom_kernel_creator_);
return Status::OK();
}
#endif // !IS_MOBILE_PLATFORM
diff --git a/tensorflow/core/common_runtime/eager/context.h b/tensorflow/core/common_runtime/eager/context.h
index 1670345..6c751f0 100644
--- a/tensorflow/core/common_runtime/eager/context.h
+++ b/tensorflow/core/common_runtime/eager/context.h
@@ -132,7 +132,7 @@
class EagerContext : public AbstractContextInterface, public core::RefCounted {
public:
- static const uint64 kInvalidContextId = 0;
+ static constexpr uint64 kInvalidContextId = 0;
static uint64 NewContextId() {
uint64 context_id = random::New64();
@@ -170,6 +170,9 @@
AbstractTensorHandleInterface* CreateLocalHandle(
AbstractTensorInterface* t) override;
+ AbstractTensorHandleInterface* CopyTensorHandleToDevice(
+ AbstractTensorHandleInterface* handle, const char* device_name,
+ Status* status) override;
AbstractOperationInterface* CreateOperation() override;
void ListDevices(std::vector<DeviceAttributes>* devices) override;
@@ -382,11 +385,10 @@
// can still be accessed, and will automatically register existing functions
// if there are newly added hosts.
Status UpdateRemoteMaster(
- WorkerEnv* worker_env,
+ uint64 context_id,
std::unique_ptr<eager::EagerClientCache> remote_eager_workers,
const std::vector<string>& add_remote_contexts,
- const std::vector<string>& remove_remote_contexts, uint64 context_id,
- Rendezvous* r);
+ const std::vector<string>& remove_remote_contexts);
// Similar with InitializeRemoteMaster but this context will not kill remote
// contexts in shutdown.
@@ -404,11 +406,8 @@
// Similar with InitializeRemoteWorker but will reuse existing context and
// increment context_view_id.
Status UpdateRemoteWorker(
- const DeviceMgr* worker_session_device_mgr,
std::unique_ptr<eager::EagerClientCache> remote_eager_workers,
- DynamicDeviceMgr* remote_device_mgr,
- const std::vector<string>& remote_contexts, uint64 context_id,
- DistributedFunctionLibraryRuntime* cluster_flr);
+ const std::vector<string>& remote_contexts, uint64 context_id);
Status StoreCollectiveOpsServer(
std::unique_ptr<ServerInterface> new_server, DeviceMgr* device_mgr,
diff --git a/tensorflow/core/common_runtime/eager/core.cc b/tensorflow/core/common_runtime/eager/core.cc
index cfb188b..c0d1b93 100644
--- a/tensorflow/core/common_runtime/eager/core.cc
+++ b/tensorflow/core/common_runtime/eager/core.cc
@@ -98,6 +98,48 @@
}
}
+AbstractTensorHandleInterface* EagerContext::CopyTensorHandleToDevice(
+ AbstractTensorHandleInterface* handle, const char* device_name,
+ Status* status) {
+ TensorHandle* input = TensorHandleFromInterface(handle);
+ TensorHandle* result = nullptr;
+ Device* device;
+ *status = this->FindDeviceFromName(device_name, &device);
+ if (!status->ok()) {
+ tensorflow::CustomDevice* dev;
+ *status = this->FindCustomDeviceFromName(device_name, &dev);
+ if (status->ok()) {
+ *status = dev->CopyTensorToDevice(input, &result);
+ if (status->ok()) {
+ return result;
+ }
+ }
+ return nullptr;
+ }
+ // Handle tensor handles currently in custom devices
+ const char* handle_device_name = input->DeviceName(status);
+ if (!status->ok()) {
+ return nullptr;
+ }
+ tensorflow::CustomDevice* dev;
+ *status = this->FindCustomDeviceFromName(handle_device_name, &dev);
+ if (status->ok()) {
+ *status = dev->CopyTensorFromDevice(input, device_name, &result);
+ if (status->ok()) {
+ return result;
+ }
+ return nullptr;
+ }
+
+ // Handle regular case.
+ *status =
+ EagerCopyToDevice(input, this, &this->Executor(), device, false, &result);
+ if (status->ok()) {
+ return result;
+ }
+ return nullptr;
+}
+
// TODO(b/152902651): We unfortunately need to put this EagerContext function
// here to a circular BUILD dep issue. If we move this to context.cc, then we
// will have the circular dependency of:
diff --git a/tensorflow/core/common_runtime/eager/eager_operation.cc b/tensorflow/core/common_runtime/eager/eager_operation.cc
index 44d2fe4..090bfef 100644
--- a/tensorflow/core/common_runtime/eager/eager_operation.cc
+++ b/tensorflow/core/common_runtime/eager/eager_operation.cc
@@ -36,6 +36,12 @@
ClearInferenceState();
}
+Status EagerOperation::SetAttrValue(const char* attr_name,
+ const AttrValue& value) {
+ MutableAttrs()->Set(attr_name, value);
+ return Status::OK();
+}
+
Status EagerOperation::SetAttrString(const char* attr_name, const char* data,
size_t length) {
MutableAttrs()->Set(attr_name, StringPiece(data, length));
diff --git a/tensorflow/core/common_runtime/eager/eager_operation.h b/tensorflow/core/common_runtime/eager/eager_operation.h
index b92a144..14268ef 100644
--- a/tensorflow/core/common_runtime/eager/eager_operation.h
+++ b/tensorflow/core/common_runtime/eager/eager_operation.h
@@ -74,6 +74,8 @@
last_set_device_name_ = "\177"; // DEL (an invalid value)
}
+ Status SetAttrValue(const char* attr_name, const AttrValue& value);
+
Status AddInput(AbstractTensorHandleInterface* input) override;
Status AddInputList(
absl::Span<AbstractTensorHandleInterface*> inputs) override;
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.cc b/tensorflow/core/common_runtime/eager/tensor_handle.cc
index 2b59588..c870d2c 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.cc
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.cc
@@ -626,10 +626,14 @@
DCHECK(IsRemote()) << "SetRemoteShape is only called on remote handles.";
auto& data = absl::get<RemoteTensorHandleData>(data_);
- if (data.context_view_id() != context_view_id) {
- return errors::Internal("Attempted to set remote shape for an old handle.");
- }
-
+ // context_view_id is currently used to validate mirrors. The shape of
+ // RemoteTensorHandleData should be set without checking context_view_id.
+ // The reason behind it is that for the primary copy of data, if the remote
+ // worker / device is removed, the consumer should report a connection error
+ // indicating the remote tensor is no longer available.
+ // For mirrors, this is not the case because they colocate with the data
+ // consuming op/function device, and we (for now) have to aggressively
+ // invalidate those copies to avoid any false positives during cluster update.
return data.SetShape(shape);
}
diff --git a/tensorflow/core/common_runtime/eager/tensor_handle.h b/tensorflow/core/common_runtime/eager/tensor_handle.h
index 0b39161..cf63f9d 100644
--- a/tensorflow/core/common_runtime/eager/tensor_handle.h
+++ b/tensorflow/core/common_runtime/eager/tensor_handle.h
@@ -184,6 +184,7 @@
// tensor for a specific device.
void Poison(Status status, const Device* d);
+ // TODO(b/154282629): Consider moving it to EagerContext.
Status CopyToDevice(const EagerContext& ctx, tensorflow::Device* d,
tensorflow::Tensor* output);
diff --git a/tensorflow/core/common_runtime/executor.cc b/tensorflow/core/common_runtime/executor.cc
index ebed956..1f2a364 100644
--- a/tensorflow/core/common_runtime/executor.cc
+++ b/tensorflow/core/common_runtime/executor.cc
@@ -200,9 +200,9 @@
// Initial time (in CPU cycles) we expect an operation to take. Used to
// determine whether an operation should be place in a threadpool.
// Operations start out "expensive".
- static const uint64 kInitialCostEstimateCycles = 100 * 1000 * 1000;
- static const uint64 kOpIsExpensiveThresholdCycles = 5000;
- static const uint64 kCostDecay = 10;
+ static constexpr uint64 kInitialCostEstimateCycles = 100 * 1000 * 1000;
+ static constexpr uint64 kOpIsExpensiveThresholdCycles = 5000;
+ static constexpr uint64 kCostDecay = 10;
std::unique_ptr<std::atomic<bool>[]> is_expensive_;
std::unique_ptr<std::atomic_uint_fast64_t[]> cost_estimates_;
diff --git a/tensorflow/core/common_runtime/executor.h b/tensorflow/core/common_runtime/executor.h
index b2a01f3..d590ae0 100644
--- a/tensorflow/core/common_runtime/executor.h
+++ b/tensorflow/core/common_runtime/executor.h
@@ -17,7 +17,7 @@
#define TENSORFLOW_CORE_COMMON_RUNTIME_EXECUTOR_H_
#include "tensorflow/core/common_runtime/device.h"
-#include "tensorflow/core/common_runtime/rendezvous_mgr.h"
+#include "tensorflow/core/common_runtime/local_executor_params.h"
#include "tensorflow/core/framework/rendezvous.h"
#include "tensorflow/core/framework/session_state.h"
#include "tensorflow/core/framework/tensor.h"
@@ -131,23 +131,6 @@
//
// "params" provides a set of context for the executor. We expect that
// different context would provide different implementations.
-struct LocalExecutorParams {
- Device* device;
-
- const SessionMetadata* session_metadata = nullptr;
-
- // The library runtime support.
- FunctionLibraryRuntime* function_library = nullptr;
-
- // create_kernel returns an instance of op kernel based on NodeDef.
- // delete_kernel is called for every kernel used by the executor
- // when the executor is deleted.
- std::function<Status(const std::shared_ptr<const NodeProperties>&,
- OpKernel**)>
- create_kernel;
- std::function<void(OpKernel*)> delete_kernel;
-};
-
::tensorflow::Status NewLocalExecutor(const LocalExecutorParams& params,
const Graph& graph, Executor** executor);
diff --git a/tensorflow/core/common_runtime/gpu/gpu_device.cc b/tensorflow/core/common_runtime/gpu/gpu_device.cc
index 00e237a..cf2e704 100644
--- a/tensorflow/core/common_runtime/gpu/gpu_device.cc
+++ b/tensorflow/core/common_runtime/gpu/gpu_device.cc
@@ -1444,7 +1444,7 @@
};
std::vector<CudaVersion> supported_cuda_compute_capabilities = {
- TF_CUDA_CAPABILITIES,};
+ CudaVersion("3.5"), CudaVersion("5.2")};
std::vector<CudaVersion> GetSupportedCudaComputeCapabilities() {
auto cuda_caps = supported_cuda_compute_capabilities;
diff --git a/tensorflow/core/common_runtime/gpu/gpu_event_mgr_test.cc b/tensorflow/core/common_runtime/gpu/gpu_event_mgr_test.cc
index 0df7a84..c89bf54 100644
--- a/tensorflow/core/common_runtime/gpu/gpu_event_mgr_test.cc
+++ b/tensorflow/core/common_runtime/gpu/gpu_event_mgr_test.cc
@@ -191,7 +191,7 @@
public:
// Length of tensors. TODO(tucker): make this a variable parameter.
- static const int kTDim = 1024;
+ static constexpr int kTDim = 1024;
int num_ops() const { return add_kernels_.size(); }
size_t tensor_size() const {
diff --git a/tensorflow/core/common_runtime/graph_optimizer.cc b/tensorflow/core/common_runtime/graph_optimizer.cc
index 774a506..410c4c2 100644
--- a/tensorflow/core/common_runtime/graph_optimizer.cc
+++ b/tensorflow/core/common_runtime/graph_optimizer.cc
@@ -40,7 +40,8 @@
shape_map,
const NodePredicate& cse_consider_fn, const NodePredicate& cf_consider_fn,
bool inline_multi_device_functions,
- bool inline_impl_selection_group_functions) {
+ bool inline_impl_selection_group_functions,
+ bool inline_with_single_device_body_placer) {
Graph* g = graph->get();
DumpGraph("Initial", g);
@@ -92,6 +93,13 @@
ExpandInlineFunctionsOptions expand_inline_opts;
expand_inline_opts.native_options.inlined_function_body_placer =
InlinedFunctionBodyPlacer::SingleDevice();
+
+ // Force single device placement strategy for multi-device function body.
+ if (inline_with_single_device_body_placer) {
+ expand_inline_opts.multi_device_options.inlined_function_body_placer =
+ InlinedFunctionBodyPlacer::SingleDevice();
+ }
+
if (!inline_multi_device_functions) {
// GraphOptimizer is running:
// (1) After partitioning when executing with a Session API.
@@ -132,7 +140,8 @@
Optimize(runtime, env, device, graph, options.shape_map,
options.cse_consider_fn, options.cf_consider_fn,
options.inline_multi_device_functions,
- options.inline_impl_selection_group_functions);
+ options.inline_impl_selection_group_functions,
+ options.inline_with_single_device_body_placer);
}
} // end namespace tensorflow
diff --git a/tensorflow/core/common_runtime/graph_optimizer.h b/tensorflow/core/common_runtime/graph_optimizer.h
index df8b37d..77c9d62 100644
--- a/tensorflow/core/common_runtime/graph_optimizer.h
+++ b/tensorflow/core/common_runtime/graph_optimizer.h
@@ -54,6 +54,10 @@
// If true, functions in implementation selection group will be inlined if
// opts_.do_function_inlining() is true.
bool inline_impl_selection_group_functions = false;
+
+ // If true all functions will be inlined with a single device function
+ // body placer strategy.
+ bool inline_with_single_device_body_placer = false;
};
explicit GraphOptimizer(const OptimizerOptions& opts);
@@ -76,7 +80,8 @@
const NodePredicate& cse_consider_fn = nullptr,
const NodePredicate& cf_consider_fn = nullptr,
bool inline_multi_device_functions = false,
- bool inline_impl_selection_group_functions = false);
+ bool inline_impl_selection_group_functions = false,
+ bool inline_with_single_device_body_placer = false);
const OptimizerOptions& options() { return opts_; }
diff --git a/tensorflow/core/common_runtime/hierarchical_tree_broadcaster.cc b/tensorflow/core/common_runtime/hierarchical_tree_broadcaster.cc
index e509792..d4cb79e 100644
--- a/tensorflow/core/common_runtime/hierarchical_tree_broadcaster.cc
+++ b/tensorflow/core/common_runtime/hierarchical_tree_broadcaster.cc
@@ -31,7 +31,6 @@
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/env.h"
-#include "tensorflow/core/platform/tracing.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/profiler/lib/traceme.h"
@@ -448,6 +447,8 @@
col_ctx_->device_locality, 0 /*stream_index*/, done);
}
+namespace {
REGISTER_COLLECTIVE(HierarchicalTreeBroadcast, HierarchicalTreeBroadcaster);
+} // namespace
} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/hierarchical_tree_broadcaster_test.cc b/tensorflow/core/common_runtime/hierarchical_tree_broadcaster_test.cc
index 63fdf4c..2006947 100644
--- a/tensorflow/core/common_runtime/hierarchical_tree_broadcaster_test.cc
+++ b/tensorflow/core/common_runtime/hierarchical_tree_broadcaster_test.cc
@@ -21,7 +21,6 @@
#include "tensorflow/core/common_runtime/collective_rma_local.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/device_resolver_local.h"
-#include "tensorflow/core/common_runtime/dma_helper.h"
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/common_runtime/test_collective_executor_mgr.h"
#include "tensorflow/core/common_runtime/threadpool_device.h"
diff --git a/tensorflow/core/common_runtime/immutable_executor_state.cc b/tensorflow/core/common_runtime/immutable_executor_state.cc
index 2f6d985..a98d9f0 100644
--- a/tensorflow/core/common_runtime/immutable_executor_state.cc
+++ b/tensorflow/core/common_runtime/immutable_executor_state.cc
@@ -16,8 +16,8 @@
#include "tensorflow/core/common_runtime/immutable_executor_state.h"
#include "absl/memory/memory.h"
-#include "tensorflow/core/common_runtime/metrics.h"
#include "tensorflow/core/framework/function.h"
+#include "tensorflow/core/framework/metrics.h"
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/graph/edgeset.h"
#include "tensorflow/core/graph/graph.h"
diff --git a/tensorflow/core/common_runtime/immutable_executor_state.h b/tensorflow/core/common_runtime/immutable_executor_state.h
index 9a2987c..50c9893 100644
--- a/tensorflow/core/common_runtime/immutable_executor_state.h
+++ b/tensorflow/core/common_runtime/immutable_executor_state.h
@@ -20,8 +20,8 @@
#include <memory>
#include <vector>
-#include "tensorflow/core/common_runtime/executor.h"
#include "tensorflow/core/common_runtime/graph_view.h"
+#include "tensorflow/core/common_runtime/local_executor_params.h"
#include "tensorflow/core/common_runtime/pending_counts.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/lib/core/status.h"
diff --git a/tensorflow/core/common_runtime/local_executor_params.h b/tensorflow/core/common_runtime/local_executor_params.h
new file mode 100644
index 0000000..caa9b68
--- /dev/null
+++ b/tensorflow/core/common_runtime/local_executor_params.h
@@ -0,0 +1,54 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_COMMON_RUNTIME_LOCAL_EXECUTOR_PARAMS_H_
+#define TENSORFLOW_CORE_COMMON_RUNTIME_LOCAL_EXECUTOR_PARAMS_H_
+
+#include <functional>
+#include <memory>
+
+namespace tensorflow {
+
+class Device;
+class StepStatsCollector;
+class SessionMetadata;
+class FunctionLibraryRuntime;
+class NodeProperties;
+class OpKernel;
+class Status;
+
+// LocalExecutorParams provides arguments that will be shared by all invocations
+// of an executor. We expect that different contexts would provide different
+// implementations (e.g. local versus distributed).
+struct LocalExecutorParams {
+ Device* device;
+
+ const SessionMetadata* session_metadata = nullptr;
+
+ // The library runtime support.
+ FunctionLibraryRuntime* function_library = nullptr;
+
+ // create_kernel returns an instance of op kernel based on NodeDef.
+ // delete_kernel is called for every kernel used by the executor
+ // when the executor is deleted.
+ std::function<Status(const std::shared_ptr<const NodeProperties>&,
+ OpKernel**)>
+ create_kernel;
+ std::function<void(OpKernel*)> delete_kernel;
+};
+
+} // end namespace tensorflow
+
+#endif // TENSORFLOW_CORE_COMMON_RUNTIME_LOCAL_EXECUTOR_PARAMS_H_
diff --git a/tensorflow/core/common_runtime/optimization_registry.cc b/tensorflow/core/common_runtime/optimization_registry.cc
index 03d0e7b..cfaeb05 100644
--- a/tensorflow/core/common_runtime/optimization_registry.cc
+++ b/tensorflow/core/common_runtime/optimization_registry.cc
@@ -15,7 +15,7 @@
#include "tensorflow/core/common_runtime/optimization_registry.h"
-#include "tensorflow/core/common_runtime/metrics.h"
+#include "tensorflow/core/framework/metrics.h"
#include "tensorflow/core/util/dump_graph.h"
namespace tensorflow {
diff --git a/tensorflow/core/common_runtime/parallel_concat_optimizer.cc b/tensorflow/core/common_runtime/parallel_concat_optimizer.cc
index 2bcd6fb..5aa53d5 100644
--- a/tensorflow/core/common_runtime/parallel_concat_optimizer.cc
+++ b/tensorflow/core/common_runtime/parallel_concat_optimizer.cc
@@ -13,10 +13,6 @@
limitations under the License.
==============================================================================*/
-#include "tensorflow/core/common_runtime/graph_optimizer.h"
-
-#include "tensorflow/core/common_runtime/constant_folding.h"
-#include "tensorflow/core/common_runtime/function.h"
#include "tensorflow/core/common_runtime/optimization_registry.h"
#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/graph/node_builder.h"
diff --git a/tensorflow/core/common_runtime/pending_counts.h b/tensorflow/core/common_runtime/pending_counts.h
index b4338af..46b6509 100644
--- a/tensorflow/core/common_runtime/pending_counts.h
+++ b/tensorflow/core/common_runtime/pending_counts.h
@@ -261,7 +261,7 @@
// Each frame in this subgraph has its own PendingCounts.
// We use 3 bits each for dead_count and pending.
- static const int kMaxCountForPackedCounts = 7;
+ static constexpr int kMaxCountForPackedCounts = 7;
// Most counts are small, so we pack a pending count and a dead
// count into 3 bits each, use 1 bit to indicate that the node has
diff --git a/tensorflow/core/common_runtime/placer_test.cc b/tensorflow/core/common_runtime/placer_test.cc
index f1fd257..2b3152a 100644
--- a/tensorflow/core/common_runtime/placer_test.cc
+++ b/tensorflow/core/common_runtime/placer_test.cc
@@ -96,7 +96,7 @@
const string& device_type) {
DeviceAttributes device_attributes;
device_attributes.set_name(name);
- device_attributes.set_device_type(DeviceType(device_type).type());
+ device_attributes.set_device_type(device_type);
return std::unique_ptr<Device>(new FakeDevice(device_attributes));
}
@@ -233,6 +233,9 @@
local_devices_.emplace_back(FakeDevice::MakeDevice(
"/job:a/replica:0/task:0/device:XLA_CPU:0", "XLA_CPU"));
devices_.AddDevice(local_devices_.back().get());
+ local_devices_.emplace_back(FakeDevice::MakeDevice(
+ "/job:a/replica:0/task:0/device:COMPOSITE:0", "COMPOSITE"));
+ devices_.AddDevice(local_devices_.back().get());
}
// Builds the given graph, and (if successful) indexes the node
@@ -1175,6 +1178,40 @@
EXPECT_DEVICE_TYPE(g, "assign", "FakeCPU");
}
+TEST_F(PlacerTest, TestResourceHandleOnCompositeDevice) {
+ auto build_graph = [this](Graph* g) -> Status {
+ GraphDefBuilder b(GraphDefBuilder::kFailImmediately);
+ Node* input = ops::SourceOp("TestInput", b.opts().WithName("in"));
+ // Build ten variable-and-assignment pairs.
+ Node* var = ops::SourceOp("HandleVariableCPU", b.opts().WithName("var"));
+ ops::BinaryOp("TestHandleAssign", var, input, b.opts().WithName("assign"));
+ TF_RETURN_IF_ERROR(BuildGraph(b, g));
+ // `var` is assigned to COMPOSITE.
+ GetNodeByName(*g, "var")->set_assigned_device_name(
+ "/job:a/replica:0/task:0/device:COMPOSITE:0");
+ return Status::OK();
+ };
+
+ {
+ // `assign` is not assigned to any device.
+ Graph g(OpRegistry::Global());
+ TF_ASSERT_OK(build_graph(&g));
+ TF_ASSERT_OK(Place(&g));
+ EXPECT_DEVICE_TYPE(g, "var", "COMPOSITE");
+ EXPECT_DEVICE_TYPE(g, "assign", "COMPOSITE");
+ }
+ {
+ // `assign` is assigned to FakeCPU.
+ Graph g(OpRegistry::Global());
+ TF_ASSERT_OK(build_graph(&g));
+ GetNodeByName(g, "assign")
+ ->set_assigned_device_name("/job:a/replica:0/task:0/device:FakeCPU:0");
+ TF_ASSERT_OK(Place(&g));
+ EXPECT_DEVICE_TYPE(g, "var", "COMPOSITE");
+ EXPECT_DEVICE_TYPE(g, "assign", "FakeCPU");
+ }
+}
+
TEST_F(PlacerTest, TestColocationGroup) {
Graph g(OpRegistry::Global());
{ // Scope for temporary variables used to construct g.
@@ -1282,6 +1319,9 @@
Node* input = ops::SourceOp("TestInput", b.opts().WithName("in"));
Node* var1 = ops::SourceOp("VariableCPU", b.opts().WithName("var1"));
Node* var2 = ops::SourceOp("VariableCPU", b.opts().WithName("var2"));
+ Node* var3 = ops::SourceOp(
+ "VariableCPU",
+ b.opts().WithName("var3").WithDevice("/device:COMPOSITE:0"));
// Two assigns (reference connections) with two different
// colocation groups. Because their colocation groups all map to the
@@ -1292,14 +1332,20 @@
ops::BinaryOp(
"TestAssign", var2, input,
b.opts().WithName("assign2").WithAttr("_class", {"loc:@var2"}));
+ ops::BinaryOp(
+ "TestAssign", var3, input,
+ b.opts().WithName("assign3").WithAttr("_class", {"loc:@var3"}));
TF_EXPECT_OK(BuildGraph(b, &g));
}
TF_EXPECT_OK(Place(&g));
+ EXPECT_DEVICE_TYPE(g, "in", "FakeCPU");
EXPECT_COLOCATED(g, "in", "var1");
EXPECT_COLOCATED(g, "in", "var2");
EXPECT_COLOCATED(g, "var1", "assign2");
EXPECT_COLOCATED(g, "var2", "assign1");
+ EXPECT_DEVICE_TYPE(g, "var3", "COMPOSITE");
+ EXPECT_COLOCATED(g, "var3", "assign3");
}
TEST_P(SoftPlacementPlacerTest,
diff --git a/tensorflow/core/common_runtime/process_state.h b/tensorflow/core/common_runtime/process_state.h
index a106a93..a833c22 100644
--- a/tensorflow/core/common_runtime/process_state.h
+++ b/tensorflow/core/common_runtime/process_state.h
@@ -84,8 +84,8 @@
// If these flags need to be runtime configurable consider adding
// them to ConfigProto.
- static const bool FLAGS_brain_mem_reg_gpu_dma = true;
- static const bool FLAGS_brain_gpu_record_mem_types = false;
+ static constexpr bool FLAGS_brain_mem_reg_gpu_dma = true;
+ static constexpr bool FLAGS_brain_gpu_record_mem_types = false;
// Helper method for unit tests to reset the ProcessState singleton by
// cleaning up everything. Never use in production.
diff --git a/tensorflow/core/common_runtime/process_util.cc b/tensorflow/core/common_runtime/process_util.cc
index 2c7f66b..8f87873 100644
--- a/tensorflow/core/common_runtime/process_util.cc
+++ b/tensorflow/core/common_runtime/process_util.cc
@@ -101,8 +101,7 @@
const char* val = std::getenv("TF_NUM_INTRAOP_THREADS");
return (val && strings::safe_strto32(val, &num)) ? num : 0;
}
-
-#ifdef INTEL_MKL
+#if !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
int32 OMPThreadsFromEnvironment() {
// 1) std::getenv is thread-safe (as long as no other function modifies the
// host env) from C++11 onward. 2) Most of TF code (except tests and
@@ -122,14 +121,14 @@
// Default to the maximum parallelism for the current process.
return port::MaxParallelism();
}
-#endif // INTEL_MKL
+#endif // !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
int32 NumInterOpThreadsFromSessionOptions(const SessionOptions& options) {
const int32 inter_op = options.config.inter_op_parallelism_threads();
if (inter_op > 0) return inter_op;
const int32 env_inter_op = GetEnvNumInterOpThreads();
if (env_inter_op > 0) return env_inter_op;
-#ifdef INTEL_MKL
+#if !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
if (!DisableMKL()) {
// MKL library executes ops in parallel using OMP threads.
// Setting inter_op conservatively to avoid thread oversubscription that
@@ -150,7 +149,7 @@
<< ". Tune using inter_op_parallelism_threads for best performance.";
return mkl_inter_op;
}
-#endif // INTEL_MKL
+#endif // !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
return DefaultNumInterOpThreads();
}
diff --git a/tensorflow/core/common_runtime/propagator_debug_utils.cc b/tensorflow/core/common_runtime/propagator_debug_utils.cc
index 27f9da7..1bf75fd 100644
--- a/tensorflow/core/common_runtime/propagator_debug_utils.cc
+++ b/tensorflow/core/common_runtime/propagator_debug_utils.cc
@@ -14,11 +14,12 @@
==============================================================================*/
#include "tensorflow/core/common_runtime/propagator_debug_utils.h"
-#include <vector>
-
#include "tensorflow/core/common_runtime/entry.h"
-#include "tensorflow/core/common_runtime/immutable_executor_state.h"
+#include "tensorflow/core/common_runtime/graph_view.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/platform/logging.h"
+#include "tensorflow/core/platform/strcat.h"
namespace tensorflow {
@@ -39,10 +40,8 @@
}
}
-void DumpPendingNodeState(const ImmutableExecutorState& immutable_state,
- const int node_id, const Entry* input_vector,
+void DumpPendingNodeState(const NodeItem& node_item, const Entry* input_vector,
const bool show_nodes_with_no_ready_inputs) {
- const NodeItem& node_item = immutable_state.graph_view().node_ref(node_id);
const int input_base = node_item.input_start;
if (!show_nodes_with_no_ready_inputs) {
bool has_ready_input = false;
@@ -73,9 +72,7 @@
}
}
-void DumpActiveNodeState(const ImmutableExecutorState& immutable_state,
- const int node_id, const Entry* input_vector) {
- const NodeItem& node_item = immutable_state.graph_view().node_ref(node_id);
+void DumpActiveNodeState(const NodeItem& node_item, const Entry* input_vector) {
LOG(WARNING) << " Active Node: " << node_item.DebugString();
const int input_base = node_item.input_start;
for (int i = 0; i < node_item.num_inputs; ++i) {
diff --git a/tensorflow/core/common_runtime/propagator_debug_utils.h b/tensorflow/core/common_runtime/propagator_debug_utils.h
index 8f12049..2e83710 100644
--- a/tensorflow/core/common_runtime/propagator_debug_utils.h
+++ b/tensorflow/core/common_runtime/propagator_debug_utils.h
@@ -18,22 +18,20 @@
namespace tensorflow {
struct Entry;
-class ImmutableExecutorState;
+struct NodeItem;
class Tensor;
// Returns a pointer to the tensor in `input` if one exists, or `nullptr`.
const Tensor* GetTensorValueForDump(const Entry& input);
-// Writes a LOG(WARNING) message describing the state of the pending node
-// `node_id` in the graph described by `immutable_state`.
-void DumpPendingNodeState(const ImmutableExecutorState& immutable_state,
- const int node_id, const Entry* input_vector,
+// Writes a LOG(WARNING) message describing the state of the given pending node
+// in the graph described by `immutable_state`.
+void DumpPendingNodeState(const NodeItem& node_item, const Entry* input_vector,
const bool show_nodes_with_no_ready_inputs);
-// Writes a LOG(WARNING) message describing the state of the active node
-// `node_id` in the graph described by `immutable_state`.
-void DumpActiveNodeState(const ImmutableExecutorState& immutable_state,
- const int node_id, const Entry* input_vector);
+// Writes a LOG(WARNING) message describing the state of the given active node
+// in the graph described by `immutable_state`.
+void DumpActiveNodeState(const NodeItem& node_item, const Entry* input_vector);
} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/propagator_state.cc b/tensorflow/core/common_runtime/propagator_state.cc
index a4e311c..6d714d2 100644
--- a/tensorflow/core/common_runtime/propagator_state.cc
+++ b/tensorflow/core/common_runtime/propagator_state.cc
@@ -17,6 +17,7 @@
#include "tensorflow/core/common_runtime/graph_view.h"
#include "tensorflow/core/common_runtime/propagator_debug_utils.h"
+#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/lib/hash/hash.h"
#include "tensorflow/core/profiler/lib/traceme.h"
@@ -177,8 +178,7 @@
immutable_state_.pending_ids()[node->node_id];
if (iteration->node_state(pending_id) == PendingCounts::PENDING_NOTREADY ||
iteration->node_state(pending_id) == PendingCounts::PENDING_READY) {
- DumpPendingNodeState(immutable_state_, node->node_id,
- iteration->input_tensors, false);
+ DumpPendingNodeState(*node, iteration->input_tensors, false);
}
}
// Then the active nodes.
@@ -186,8 +186,7 @@
PendingCounts::Handle pending_id =
immutable_state_.pending_ids()[node->node_id];
if (iteration->node_state(pending_id) == PendingCounts::STARTED) {
- DumpActiveNodeState(immutable_state_, node->node_id,
- iteration->input_tensors);
+ DumpActiveNodeState(*node, iteration->input_tensors);
}
}
// Show all input tensors in use.
diff --git a/tensorflow/core/common_runtime/propagator_state.h b/tensorflow/core/common_runtime/propagator_state.h
index 6d5abd0..13aadde 100644
--- a/tensorflow/core/common_runtime/propagator_state.h
+++ b/tensorflow/core/common_runtime/propagator_state.h
@@ -33,7 +33,6 @@
namespace tensorflow {
-typedef gtl::InlinedVector<TensorValue, 4> TensorValueVec;
typedef gtl::InlinedVector<AllocatorAttributes, 4> AllocatorAttributeVec;
// Represents the ephemeral "edge state" associated with one invocation of
diff --git a/tensorflow/core/common_runtime/ring_gatherer.cc b/tensorflow/core/common_runtime/ring_gatherer.cc
index db096ba..ecffd4a 100644
--- a/tensorflow/core/common_runtime/ring_gatherer.cc
+++ b/tensorflow/core/common_runtime/ring_gatherer.cc
@@ -38,7 +38,6 @@
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/env.h"
-#include "tensorflow/core/platform/tracing.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/profiler/lib/traceme.h"
@@ -273,6 +272,8 @@
return !aborted;
}
+namespace {
REGISTER_COLLECTIVE(RingGather, RingGatherer);
+} // namespace
} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/ring_gatherer_test.cc b/tensorflow/core/common_runtime/ring_gatherer_test.cc
index 0f05178..3af4890 100644
--- a/tensorflow/core/common_runtime/ring_gatherer_test.cc
+++ b/tensorflow/core/common_runtime/ring_gatherer_test.cc
@@ -22,7 +22,6 @@
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/device_resolver_local.h"
-#include "tensorflow/core/common_runtime/dma_helper.h"
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/common_runtime/test_collective_executor_mgr.h"
#include "tensorflow/core/common_runtime/threadpool_device.h"
diff --git a/tensorflow/core/common_runtime/ring_reducer.cc b/tensorflow/core/common_runtime/ring_reducer.cc
index cc950df..ab4542d 100644
--- a/tensorflow/core/common_runtime/ring_reducer.cc
+++ b/tensorflow/core/common_runtime/ring_reducer.cc
@@ -38,7 +38,6 @@
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/env.h"
-#include "tensorflow/core/platform/tracing.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/profiler/lib/traceme.h"
@@ -350,6 +349,8 @@
return !aborted;
}
+namespace {
REGISTER_COLLECTIVE(RingReduce, RingReducer);
+} // namespace
} // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/ring_reducer_test.cc b/tensorflow/core/common_runtime/ring_reducer_test.cc
index 3502b46..318d6e9 100644
--- a/tensorflow/core/common_runtime/ring_reducer_test.cc
+++ b/tensorflow/core/common_runtime/ring_reducer_test.cc
@@ -22,7 +22,6 @@
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_mgr.h"
#include "tensorflow/core/common_runtime/device_resolver_local.h"
-#include "tensorflow/core/common_runtime/dma_helper.h"
#include "tensorflow/core/common_runtime/process_util.h"
#include "tensorflow/core/common_runtime/test_collective_executor_mgr.h"
#include "tensorflow/core/common_runtime/threadpool_device.h"
diff --git a/tensorflow/core/common_runtime/scoped_allocator.h b/tensorflow/core/common_runtime/scoped_allocator.h
index 8ab5ee0..f25bee4 100644
--- a/tensorflow/core/common_runtime/scoped_allocator.h
+++ b/tensorflow/core/common_runtime/scoped_allocator.h
@@ -27,8 +27,8 @@
// Manages a single backing tensor and a collection of aliases.
class ScopedAllocator {
public:
- static const int32 kInvalidId = 0;
- static const size_t kMaxAlignment = 64;
+ static constexpr int32 kInvalidId = 0;
+ static constexpr size_t kMaxAlignment = 64;
// A subrange of the TensorBuffer associated with this object that
// will be the backing memory for one aliased tensor.
@@ -39,7 +39,7 @@
size_t bytes_allocated;
};
// Field index that refers to backing tensor, not any aliased field.
- static const int32 kBackingIndex = -1;
+ static constexpr int32 kBackingIndex = -1;
// backing_tensor is expected to be newly allocated by a ScopedAllocatorOp
// instance. It must be large enough to back all of the specified
diff --git a/tensorflow/core/common_runtime/shape_refiner.h b/tensorflow/core/common_runtime/shape_refiner.h
index 33ebba0..c83bd81 100644
--- a/tensorflow/core/common_runtime/shape_refiner.h
+++ b/tensorflow/core/common_runtime/shape_refiner.h
@@ -17,6 +17,7 @@
#include <vector>
+#include "absl/container/flat_hash_map.h"
#include "tensorflow/core/common_runtime/graph_runner.h"
#include "tensorflow/core/framework/function.pb.h"
#include "tensorflow/core/framework/shape_inference.h"
@@ -236,7 +237,8 @@
GraphRunner graph_runner_;
// Stores a map from a node to its ExtendedInferenceContext.
- std::unordered_map<const Node*, std::unique_ptr<ExtendedInferenceContext>>
+ absl::flat_hash_map<const Node*, std::unique_ptr<ExtendedInferenceContext>,
+ hash<const Node*>>
node_to_context_;
// Holds a cache from 'tensor name' to the tensor that is
@@ -257,9 +259,10 @@
// shape inference.
const tensorflow::FunctionLibraryDefinition* function_library_ = nullptr;
- // Cache the graph corresponding to each functin definition for which shapes
+ // Cache the graph corresponding to each function definition for which shapes
// are refined.
- std::unordered_map<const FunctionDef*, std::unique_ptr<const Graph>>
+ absl::flat_hash_map<const FunctionDef*, std::unique_ptr<const Graph>,
+ hash<const FunctionDef*>>
functions_;
TF_DISALLOW_COPY_AND_ASSIGN(ShapeRefiner);
diff --git a/tensorflow/core/common_runtime/simple_propagator_state.cc b/tensorflow/core/common_runtime/simple_propagator_state.cc
index bf6172b..48fac96 100644
--- a/tensorflow/core/common_runtime/simple_propagator_state.cc
+++ b/tensorflow/core/common_runtime/simple_propagator_state.cc
@@ -17,6 +17,7 @@
#include <atomic>
#include "tensorflow/core/common_runtime/propagator_debug_utils.h"
+#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/profiler/lib/traceme.h"
namespace tensorflow {
@@ -107,15 +108,13 @@
// Dump any waiting nodes that are holding on to tensors.
for (const NodeItem* node : *nodes_) {
if (pending_[node->node_id]) {
- DumpPendingNodeState(immutable_state_, node->node_id,
- input_tensors_.data(), false);
+ DumpPendingNodeState(*node, input_tensors_.data(), false);
}
}
// Then the active nodes.
for (const NodeItem* node : *nodes_) {
if ((*active_)[node->node_id]) {
- DumpActiveNodeState(immutable_state_, node->node_id,
- input_tensors_.data());
+ DumpActiveNodeState(*node, input_tensors_.data());
}
}
// Show all input tensors in use.
diff --git a/tensorflow/core/common_runtime/step_stats_collector.h b/tensorflow/core/common_runtime/step_stats_collector.h
index f2789c7..ce6fa97 100644
--- a/tensorflow/core/common_runtime/step_stats_collector.h
+++ b/tensorflow/core/common_runtime/step_stats_collector.h
@@ -19,7 +19,6 @@
#include <unordered_map>
#include <vector>
#include "tensorflow/core/framework/step_stats.pb.h"
-#include "tensorflow/core/framework/tensor_reference.h"
#include "tensorflow/core/lib/gtl/inlined_vector.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/mutex.h"
@@ -188,7 +187,7 @@
private:
// TODO(suharshs): Make this configurable if its not possible to find a value
// that works for all cases.
- static const uint64 kMaxCollectedNodes = 1 << 20;
+ static constexpr uint64 kMaxCollectedNodes = 1 << 20;
typedef std::vector<std::unique_ptr<NodeExecStatsWrapper>> NodeStatsVector;
typedef std::unordered_map<uint32, string> ThreadNamesMap;
diff --git a/tensorflow/core/common_runtime/threadpool_device.cc b/tensorflow/core/common_runtime/threadpool_device.cc
index 68fcc9a..44fa5bf 100644
--- a/tensorflow/core/common_runtime/threadpool_device.cc
+++ b/tensorflow/core/common_runtime/threadpool_device.cc
@@ -50,7 +50,7 @@
name, DEVICE_CPU, memory_limit, locality)),
allocator_(allocator),
scoped_allocator_mgr_(new ScopedAllocatorMgr(name)) {
-#ifdef INTEL_MKL
+#if !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
// Early return when MKL is disabled
if (DisableMKL()) return;
#ifdef _OPENMP
@@ -69,7 +69,7 @@
}
}
#endif // _OPENMP
-#endif // INTEL_MKL
+#endif // !defined(ENABLE_MKLDNN_THREADPOOL) && defined(INTEL_MKL)
}
ThreadPoolDevice::~ThreadPoolDevice() {}
diff --git a/tensorflow/core/data/service/BUILD b/tensorflow/core/data/service/BUILD
index ef270f2..8fd70fb 100644
--- a/tensorflow/core/data/service/BUILD
+++ b/tensorflow/core/data/service/BUILD
@@ -278,6 +278,7 @@
deps = [
":credentials_factory",
":grpc_master_impl",
+ ":grpc_util",
":grpc_worker_impl",
":local_credentials_factory",
"//tensorflow/core:lib",
diff --git a/tensorflow/core/data/service/grpc_master_impl.cc b/tensorflow/core/data/service/grpc_master_impl.cc
index 9d00b8c..4e5e9f4 100644
--- a/tensorflow/core/data/service/grpc_master_impl.cc
+++ b/tensorflow/core/data/service/grpc_master_impl.cc
@@ -39,6 +39,7 @@
return ToGrpcStatus(impl_.method(request, response)); \
}
HANDLER(RegisterWorker);
+HANDLER(WorkerUpdate);
HANDLER(GetOrRegisterDataset);
HANDLER(CreateJob);
HANDLER(GetTasks);
diff --git a/tensorflow/core/data/service/grpc_master_impl.h b/tensorflow/core/data/service/grpc_master_impl.h
index c4f12be..2f775f8 100644
--- a/tensorflow/core/data/service/grpc_master_impl.h
+++ b/tensorflow/core/data/service/grpc_master_impl.h
@@ -38,17 +38,18 @@
const std::string& protocol);
~GrpcMasterImpl() override {}
- private:
#define HANDLER(method) \
grpc::Status method(grpc::ServerContext* context, \
const method##Request* request, \
method##Response* response) override;
HANDLER(RegisterWorker);
+ HANDLER(WorkerUpdate);
HANDLER(GetOrRegisterDataset);
HANDLER(CreateJob);
HANDLER(GetTasks);
#undef HANDLER
+ private:
DataServiceMasterImpl impl_;
TF_DISALLOW_COPY_AND_ASSIGN(GrpcMasterImpl);
diff --git a/tensorflow/core/data/service/master.proto b/tensorflow/core/data/service/master.proto
index 3a8a514..9361b7b 100644
--- a/tensorflow/core/data/service/master.proto
+++ b/tensorflow/core/data/service/master.proto
@@ -16,6 +16,21 @@
repeated TaskDef tasks = 2;
}
+message TaskProgress {
+ // The task that this message is about.
+ int64 task_id = 1;
+ // Whether the task has completed.
+ bool completed = 2;
+}
+
+message WorkerUpdateRequest {
+ // The worker id that the update is for.
+ int64 worker_id = 1;
+ repeated TaskProgress updates = 2;
+}
+
+message WorkerUpdateResponse {}
+
message GetOrRegisterDatasetRequest {
// The dataset to register.
DatasetDef dataset = 1;
@@ -60,12 +75,19 @@
message GetTasksResponse {
// A list of all tasks for a job.
repeated TaskInfo task_info = 1;
+ // Whether the job has finished. An empty `task_info` list could either mean
+ // that no tasks have been started yet, or that all tasks have finished. This
+ // field gives us a way to tell the difference.
+ bool job_finished = 2;
}
service MasterService {
// Registers a worker with the master.
rpc RegisterWorker(RegisterWorkerRequest) returns (RegisterWorkerResponse);
+ // Updates the master with information about the worker's state.
+ rpc WorkerUpdate(WorkerUpdateRequest) returns (WorkerUpdateResponse);
+
// Registers a dataset with the server, or returns its id if it is already
// registered.
//
diff --git a/tensorflow/core/data/service/master_impl.cc b/tensorflow/core/data/service/master_impl.cc
index 75bbf00..4141b26 100644
--- a/tensorflow/core/data/service/master_impl.cc
+++ b/tensorflow/core/data/service/master_impl.cc
@@ -15,6 +15,10 @@
#include "tensorflow/core/data/service/master_impl.h"
+#include <memory>
+#include <tuple>
+#include <utility>
+
#include "grpcpp/create_channel.h"
#include "grpcpp/impl/codegen/server_context.h"
#include "grpcpp/security/credentials.h"
@@ -27,6 +31,7 @@
#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/kernels/data/dataset_utils.h"
#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/public/session_options.h"
namespace tensorflow {
@@ -55,34 +60,49 @@
VLOG(3) << "Received register worker request";
mutex_lock l(mu_);
int64 worker_id = next_worker_id_++;
- workers_.emplace_back();
- workers_.back().address = request->worker_address();
- workers_.back().id = worker_id;
+ workers_.emplace_back(worker_id, request->worker_address());
response->set_worker_id(worker_id);
// Allocate tasks to the worker.
for (auto& entry : jobs_) {
Job& job = entry.second;
- int64 task_id = next_task_id_++;
- DCHECK(!tasks_.contains(task_id));
- Task& task = tasks_[task_id];
- task.id = task_id;
- task.dataset_id = job.dataset_id;
- task.worker_address = request->worker_address();
- job.task_ids.push_back(task_id);
+ if (job.finished()) {
+ continue;
+ }
+ int64 task_id = CreateTask(&job, request->worker_address());
TaskDef* task_def = response->add_tasks();
*task_def->mutable_dataset() =
- datasets_by_id_[task.dataset_id]->dataset_def;
- task_def->set_dataset_id(task.dataset_id);
- task_def->set_job_id(job.id);
- task_def->set_task_id(task.id);
+ datasets_by_id_[job.dataset_id()]->dataset_def();
+ task_def->set_dataset_id(job.dataset_id());
+ task_def->set_job_id(job.job_id());
+ task_def->set_task_id(task_id);
}
VLOG(1) << "Registered worker " << workers_.back().DebugString();
return Status::OK();
}
+Status DataServiceMasterImpl::WorkerUpdate(const WorkerUpdateRequest* request,
+ WorkerUpdateResponse* response) {
+ mutex_lock l(mu_);
+ int64 worker_id = request->worker_id();
+ for (auto& update : request->updates()) {
+ int64 task_id = update.task_id();
+ if (!tasks_.contains(task_id)) {
+ return errors::NotFound("WorkerUpdate called for worker ", worker_id,
+ " with unknown task id ", task_id);
+ }
+ if (update.completed()) {
+ int64 job_id = tasks_.at(task_id).job_id();
+ DCHECK(jobs_.contains(job_id));
+ jobs_.at(job_id).task_finished(task_id);
+ VLOG(3) << "Task " << task_id << " from job " << job_id << " completed";
+ }
+ }
+ return Status::OK();
+}
+
Status DataServiceMasterImpl::GetOrRegisterDataset(
const GetOrRegisterDatasetRequest* request,
GetOrRegisterDatasetResponse* response) {
@@ -92,7 +112,7 @@
VLOG(3) << "Registering dataset graph: "
<< request->dataset().graph().DebugString();
if (datasets_by_fingerprint_.contains(fingerprint)) {
- int64 id = datasets_by_fingerprint_[fingerprint]->id;
+ int64 id = datasets_by_fingerprint_[fingerprint]->dataset_id();
VLOG(3) << "Received duplicate RegisterDataset request with fingerprint "
<< fingerprint << ". Returning id " << id;
response->set_dataset_id(id);
@@ -108,11 +128,9 @@
int64 DataServiceMasterImpl::RegisterDataset(uint64 fingerprint,
const DatasetDef& dataset)
EXCLUSIVE_LOCKS_REQUIRED(mu_) {
- auto new_dataset = std::make_shared<Dataset>();
int64 dataset_id = next_dataset_id_++;
- new_dataset->id = dataset_id;
- new_dataset->fingerprint = fingerprint;
- new_dataset->dataset_def = dataset;
+ auto new_dataset =
+ std::make_shared<Dataset>(dataset_id, fingerprint, dataset);
DCHECK(!datasets_by_id_.contains(dataset_id));
datasets_by_id_[dataset_id] = new_dataset;
@@ -144,24 +162,18 @@
int64 job_id = next_job_id_++;
DCHECK(!jobs_.contains(job_id));
- Job& job = jobs_[job_id];
- job.id = job_id;
- job.dataset_id = request->dataset_id();
+ auto result =
+ jobs_.emplace(std::piecewise_construct, std::forward_as_tuple(job_id),
+ std::forward_as_tuple(job_id, request->dataset_id()));
+ DCHECK(result.second);
+ Job& job = result.first->second;
response->set_job_id(job_id);
for (auto& worker : workers_) {
- int64 task_id = next_task_id_++;
- DCHECK(!tasks_.contains(task_id));
- Task& task = tasks_[task_id];
- task.id = task_id;
- task.dataset_id = request->dataset_id();
- task.worker_address = worker.address;
- job.task_ids.push_back(task_id);
+ int64 task_id = CreateTask(&job, worker.address());
- std::unique_ptr<WorkerService::Stub> stub;
- TF_RETURN_IF_ERROR(CreateWorkerStub(worker.address, protocol_, &stub));
// TODO(aaudibert): perform these calls asynchronously.
- TF_RETURN_IF_ERROR(AllocateTaskToWorker(task, &worker));
+ TF_RETURN_IF_ERROR(AllocateTaskToWorker(tasks_.at(task_id), &worker));
}
VLOG(3) << "Beginning job " << job_id << " for dataset "
@@ -169,25 +181,45 @@
return Status::OK();
}
-Status DataServiceMasterImpl::AllocateTaskToWorker(const Task& task,
- WorkerInfo* worker)
+int64 DataServiceMasterImpl::CreateTask(Job* job,
+ const std::string& worker_address)
EXCLUSIVE_LOCKS_REQUIRED(mu_) {
- if (!worker->stub) {
- TF_RETURN_IF_ERROR(
- CreateWorkerStub(worker->address, protocol_, &worker->stub));
+ int64 task_id = next_task_id_++;
+ DCHECK(!tasks_.contains(task_id));
+ auto result =
+ tasks_.emplace(std::piecewise_construct, std::forward_as_tuple(task_id),
+ std::forward_as_tuple(task_id, job->job_id(),
+ job->dataset_id(), worker_address));
+ job->add_task_id(task_id);
+ DCHECK(result.second);
+ return task_id;
+}
+
+Status DataServiceMasterImpl::EnsureWorkerStubInitialized(Worker* worker) {
+ if (!worker->stub()) {
+ std::unique_ptr<WorkerService::Stub> stub;
+ TF_RETURN_IF_ERROR(CreateWorkerStub(worker->address(), protocol_, &stub));
+ worker->set_stub(std::move(stub));
}
+ return Status::OK();
+}
+
+Status DataServiceMasterImpl::AllocateTaskToWorker(const Task& task,
+ Worker* worker)
+ EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+ TF_RETURN_IF_ERROR(EnsureWorkerStubInitialized(worker));
grpc::ClientContext client_ctx;
ProcessTaskRequest req;
- req.mutable_task()->set_dataset_id(task.dataset_id);
- DCHECK(datasets_by_id_.contains(task.dataset_id));
+ req.mutable_task()->set_dataset_id(task.dataset_id());
+ DCHECK(datasets_by_id_.contains(task.dataset_id()));
*req.mutable_task()->mutable_dataset() =
- datasets_by_id_[task.dataset_id]->dataset_def;
- req.mutable_task()->set_task_id(task.id);
+ datasets_by_id_.at(task.dataset_id())->dataset_def();
+ req.mutable_task()->set_task_id(task.task_id());
ProcessTaskResponse resp;
- grpc::Status s = worker->stub->ProcessTask(&client_ctx, req, &resp);
+ grpc::Status s = worker->stub()->ProcessTask(&client_ctx, req, &resp);
if (!s.ok()) {
return grpc_util::WrapError(
- absl::StrCat("Failed to submit task to worker ", worker->address), s);
+ absl::StrCat("Failed to submit task to worker ", worker->address()), s);
}
return Status::OK();
}
@@ -202,14 +234,15 @@
"> not found.");
}
Job& job = it->second;
- for (const auto& task_id : job.task_ids) {
+ for (const auto& task_id : job.task_ids()) {
auto task_iter = tasks_.find(task_id);
DCHECK(task_iter != tasks_.end());
Task& task = task_iter->second;
TaskInfo* task_info = response->mutable_task_info()->Add();
- task_info->set_worker_address(task.worker_address);
- task_info->set_id(task.id);
+ task_info->set_worker_address(task.worker_address());
+ task_info->set_id(task.task_id());
}
+ response->set_job_finished(job.finished());
VLOG(3) << "Found " << response->task_info_size() << " tasks for job id "
<< request->job_id();
return Status::OK();
diff --git a/tensorflow/core/data/service/master_impl.h b/tensorflow/core/data/service/master_impl.h
index 8286bb8..b7cfc49 100644
--- a/tensorflow/core/data/service/master_impl.h
+++ b/tensorflow/core/data/service/master_impl.h
@@ -48,6 +48,8 @@
/// Worker-facing API.
Status RegisterWorker(const RegisterWorkerRequest* request,
RegisterWorkerResponse* response);
+ Status WorkerUpdate(const WorkerUpdateRequest* request,
+ WorkerUpdateResponse* response);
/// Client-facing API.
Status GetOrRegisterDataset(const GetOrRegisterDatasetRequest* request,
@@ -57,38 +59,99 @@
Status GetTasks(const GetTasksRequest* request, GetTasksResponse* response);
private:
- typedef struct WorkerInfo {
- std::string address;
- int64 id;
- std::unique_ptr<WorkerService::Stub> stub;
+ class Worker {
+ public:
+ Worker(int64 worker_id, const std::string address)
+ : worker_id_(worker_id), address_(address) {}
+
+ int64 worker_id() { return worker_id_; }
+ std::string address() { return address_; }
+ WorkerService::Stub* stub() { return stub_.get(); }
+ void set_stub(std::unique_ptr<WorkerService::Stub> stub) {
+ stub_ = std::move(stub);
+ }
std::string DebugString() {
- return absl::StrCat("id: ", id, "address: ", address);
+ return absl::StrCat("id: ", worker_id_, "address: ", address_);
}
- } WorkerInfo;
- typedef struct Dataset {
- int64 id;
- int64 fingerprint;
- DatasetDef dataset_def;
- } Dataset;
+ private:
+ const int64 worker_id_;
+ const std::string address_;
+ std::unique_ptr<WorkerService::Stub> stub_;
+ };
- typedef struct Job {
- int64 id;
- int64 dataset_id;
- std::vector<int64> task_ids;
- } Job;
+ class Dataset {
+ public:
+ Dataset(int64 dataset_id, int64 fingerprint, const DatasetDef& dataset_def)
+ : dataset_id_(dataset_id),
+ fingerprint_(fingerprint),
+ dataset_def_(dataset_def) {}
- typedef struct Task {
- int64 id;
- int64 dataset_id;
- std::string worker_address;
- } Task;
+ int64 dataset_id() const { return dataset_id_; }
+ int64 fingerprint() const { return fingerprint_; }
+ const DatasetDef& dataset_def() { return dataset_def_; }
+
+ private:
+ const int64 dataset_id_;
+ const int64 fingerprint_;
+ const DatasetDef dataset_def_;
+ };
+
+ class Job {
+ public:
+ Job(int64 job_id, int64 dataset_id)
+ : job_id_(job_id), dataset_id_(dataset_id) {}
+
+ int64 job_id() const { return job_id_; }
+ int64 dataset_id() const { return dataset_id_; }
+ const std::vector<int64>& task_ids() const { return task_ids_; }
+ void add_task_id(int64 task_id) { task_ids_.push_back(task_id); }
+ void task_finished(int64 task_id) {
+ finished_tasks_.push_back(task_id);
+ if (finished_tasks_.size() == task_ids_.size()) {
+ finished_ = true;
+ }
+ }
+ bool finished() const { return finished_; }
+
+ private:
+ const int64 job_id_;
+ const int64 dataset_id_;
+ std::vector<int64> task_ids_;
+ std::vector<int64> finished_tasks_;
+ bool finished_ = false;
+ };
+
+ class Task {
+ public:
+ Task(int64 task_id, int64 job_id, int64 dataset_id,
+ const std::string& worker_address)
+ : task_id_(task_id),
+ job_id_(job_id),
+ dataset_id_(dataset_id),
+ worker_address_(worker_address) {}
+
+ int64 task_id() const { return task_id_; }
+ int64 job_id() const { return job_id_; }
+ int64 dataset_id() const { return dataset_id_; }
+ std::string worker_address() const { return worker_address_; }
+
+ private:
+ const int64 task_id_;
+ const int64 job_id_;
+ const int64 dataset_id_;
+ const std::string worker_address_;
+ };
// Registers a dataset with the given fingerprint, returning a new dataset id.
int64 RegisterDataset(uint64 fingerprint, const DatasetDef& dataset);
+ // Initializes a workers stub, if it hasn't been initialized already.
+ Status EnsureWorkerStubInitialized(Worker* worker);
// Instructs a worker to begin processing a task.
- Status AllocateTaskToWorker(const Task& task_id, WorkerInfo* worker);
+ Status AllocateTaskToWorker(const Task& task_id, Worker* worker);
+ // Creates a new task for a job, returning the new task's id.
+ int64 CreateTask(Job* job, const std::string& worker_address);
// Protocol to use for communicating with workers.
const std::string protocol_;
@@ -101,7 +164,7 @@
int64 next_task_id_ TF_GUARDED_BY(mu_) = 0;
// Registered workers.
- std::vector<WorkerInfo> workers_ TF_GUARDED_BY(mu_);
+ std::vector<Worker> workers_ TF_GUARDED_BY(mu_);
// Registered datasets, keyed by dataset ids.
absl::flat_hash_map<int64, std::shared_ptr<Dataset>> datasets_by_id_
TF_GUARDED_BY(mu_);
diff --git a/tensorflow/core/data/service/server_lib.cc b/tensorflow/core/data/service/server_lib.cc
index d4ec8dd..40139b3 100644
--- a/tensorflow/core/data/service/server_lib.cc
+++ b/tensorflow/core/data/service/server_lib.cc
@@ -17,20 +17,16 @@
#include "tensorflow/core/data/service/credentials_factory.h"
#include "tensorflow/core/data/service/grpc_master_impl.h"
+#include "tensorflow/core/data/service/grpc_util.h"
#include "tensorflow/core/data/service/grpc_worker_impl.h"
namespace tensorflow {
namespace data {
-GrpcDataServer::GrpcDataServer(int port, const std::string& protocol,
- bool is_master,
- const std::string& master_address)
- : requested_port_(port),
- protocol_(protocol),
- is_master_(is_master),
- master_address_(master_address) {}
+GrpcDataServerBase::GrpcDataServerBase(int port, const std::string& protocol)
+ : requested_port_(port), protocol_(protocol) {}
-Status GrpcDataServer::Start() {
+Status GrpcDataServerBase::Start() {
::grpc::ServerBuilder builder;
std::shared_ptr<::grpc::ServerCredentials> credentials;
TF_RETURN_IF_ERROR(
@@ -39,48 +35,78 @@
credentials, &bound_port_);
builder.SetMaxReceiveMessageSize(-1);
- if (is_master_) {
- service_ = absl::make_unique<GrpcMasterImpl>(&builder, protocol_);
- } else {
- service_ =
- absl::make_unique<GrpcWorkerImpl>(&builder, master_address_, protocol_);
- }
-
+ AddServiceToBuilder(&builder);
server_ = builder.BuildAndStart();
if (!server_) {
return errors::Internal("Could not start gRPC server");
}
- if (!is_master_) {
- static_cast<GrpcWorkerImpl*>(service_.get())
- ->Start(strings::StrCat("localhost:", bound_port_));
- }
+ TF_RETURN_IF_ERROR(StartServiceInternal());
- LOG(INFO) << "Started data service " << (is_master_ ? "master" : "worker")
- << " running at " << Target();
+ LOG(INFO) << "Started data service running at " << Target();
return Status::OK();
}
-void GrpcDataServer::Stop() { server_->Shutdown(); }
+void GrpcDataServerBase::Stop() { server_->Shutdown(); }
-void GrpcDataServer::Join() { server_->Wait(); }
+void GrpcDataServerBase::Join() { server_->Wait(); }
-std::string GrpcDataServer::Target() {
+std::string GrpcDataServerBase::Target() {
return strings::StrCat(protocol_, "://localhost:", bound_port_);
}
+MasterGrpcDataServer::MasterGrpcDataServer(int port,
+ const std::string& protocol)
+ : GrpcDataServerBase(port, protocol) {}
+
+MasterGrpcDataServer::~MasterGrpcDataServer() { delete service_; }
+
+void MasterGrpcDataServer::AddServiceToBuilder(grpc::ServerBuilder* builder) {
+ auto service = absl::make_unique<GrpcMasterImpl>(builder, protocol_);
+ service_ = service.release();
+}
+
+Status MasterGrpcDataServer::NumTasks(int* num_tasks) {
+ GetTasksRequest req;
+ GetTasksResponse resp;
+ grpc::ServerContext ctx;
+ grpc::Status s = service_->GetTasks(&ctx, &req, &resp);
+ if (!s.ok()) {
+ return grpc_util::WrapError("Failed to get num tasks", s);
+ }
+ *num_tasks = resp.task_info_size();
+ return Status::OK();
+}
+
+WorkerGrpcDataServer::WorkerGrpcDataServer(int port,
+ const std::string& protocol,
+ const std::string& master_address)
+ : GrpcDataServerBase(port, protocol), master_address_(master_address) {}
+
+WorkerGrpcDataServer::~WorkerGrpcDataServer() { delete service_; }
+
+void WorkerGrpcDataServer::AddServiceToBuilder(grpc::ServerBuilder* builder) {
+ auto service =
+ absl::make_unique<GrpcWorkerImpl>(builder, master_address_, protocol_);
+ service_ = service.release();
+}
+
+Status WorkerGrpcDataServer::StartServiceInternal() {
+ service_->Start(strings::StrCat("localhost:", bound_port()));
+ return Status::OK();
+}
+
Status NewMasterServer(int port, const std::string& protocol,
- std::unique_ptr<GrpcDataServer>* out_server) {
- *out_server = absl::make_unique<GrpcDataServer>(
- port, protocol, /*is_master=*/true, /*master_address=*/"");
+ std::unique_ptr<MasterGrpcDataServer>* out_server) {
+ *out_server = absl::make_unique<MasterGrpcDataServer>(port, protocol);
return Status::OK();
}
Status NewWorkerServer(int port, const std::string& protocol,
const std::string& master_address,
- std::unique_ptr<GrpcDataServer>* out_server) {
- *out_server = absl::make_unique<GrpcDataServer>(
- port, protocol, /*is_master=*/false, master_address);
+ std::unique_ptr<WorkerGrpcDataServer>* out_server) {
+ *out_server =
+ absl::make_unique<WorkerGrpcDataServer>(port, protocol, master_address);
return Status::OK();
}
diff --git a/tensorflow/core/data/service/server_lib.h b/tensorflow/core/data/service/server_lib.h
index 753dd5d..4f35444 100644
--- a/tensorflow/core/data/service/server_lib.h
+++ b/tensorflow/core/data/service/server_lib.h
@@ -16,23 +16,28 @@
#ifndef TENSORFLOW_CORE_DATA_SERVICE_SERVER_LIB_H_
#define TENSORFLOW_CORE_DATA_SERVICE_SERVER_LIB_H_
-#include "grpcpp/impl/codegen/service_type.h"
#include "grpcpp/server.h"
+#include "grpcpp/server_builder.h"
#include "tensorflow/core/lib/core/status.h"
namespace tensorflow {
namespace data {
+// Forward declared because transitively depending on .grpc.pb.h files causes
+// issues in the pywrap build.
+class GrpcMasterImpl;
+class GrpcWorkerImpl;
+
// A grpc server for the dataset service.
-class GrpcDataServer {
+class GrpcDataServerBase {
public:
// Constructs a dataset server with the specified port. If the port is 0, the
// server will find an available port in `Start()`. The chosen port can be
// found in the output of `Target()`.
//
// master_address is only needed for worker data servers.
- explicit GrpcDataServer(int requested_port, const std::string& protocol,
- bool is_master, const std::string& master_address);
+ GrpcDataServerBase(int requested_port, const std::string& protocol);
+ virtual ~GrpcDataServerBase() {}
// Starts the server running asynchronously.
Status Start();
@@ -46,26 +51,64 @@
// Returns the target string for the server. Only valid after calling Start().
std::string Target();
- private:
+ protected:
+ virtual void AddServiceToBuilder(::grpc::ServerBuilder* builder) = 0;
+ // Starts the service. This will be called after building the service, so
+ // bound_port() will return the actual bound port.
+ virtual Status StartServiceInternal() = 0;
+
+ int bound_port() { return bound_port_; }
+
const int requested_port_;
const std::string protocol_;
- const bool is_master_;
- const std::string master_address_;
+ private:
int bound_port_;
- std::unique_ptr<grpc::Service> service_;
std::unique_ptr<grpc::Server> server_;
};
+class MasterGrpcDataServer : public GrpcDataServerBase {
+ public:
+ MasterGrpcDataServer(int requested_port, const std::string& protocol);
+ ~MasterGrpcDataServer() override;
+
+ // Returns the number of tasks created by the master.
+ Status NumTasks(int* num_tasks);
+
+ protected:
+ void AddServiceToBuilder(grpc::ServerBuilder* builder) override;
+ Status StartServiceInternal() override { return Status::OK(); }
+
+ private:
+ // Owned. We use a raw pointer because GrpcMasterImpl is forward-declared.
+ GrpcMasterImpl* service_;
+};
+
+class WorkerGrpcDataServer : public GrpcDataServerBase {
+ public:
+ WorkerGrpcDataServer(int requested_port, const std::string& protocol,
+ const std::string& master_address);
+ ~WorkerGrpcDataServer() override;
+
+ protected:
+ void AddServiceToBuilder(grpc::ServerBuilder* builder) override;
+ Status StartServiceInternal() override;
+
+ private:
+ const std::string master_address_;
+ // Owned. We use a raw pointer because GrpcWorkerImpl is forward-declared.
+ GrpcWorkerImpl* service_;
+};
+
// Creates a master dataset server and stores it in `*out_server`.
Status NewMasterServer(int port, const std::string& protocol,
- std::unique_ptr<GrpcDataServer>* out_server);
+ std::unique_ptr<MasterGrpcDataServer>* out_server);
// Creates a worker dataset server and stores it in `*out_server`.
Status NewWorkerServer(int port, const std::string& protocol,
const std::string& master_address,
- std::unique_ptr<GrpcDataServer>* out_server);
+ std::unique_ptr<WorkerGrpcDataServer>* out_server);
} // namespace data
} // namespace tensorflow
diff --git a/tensorflow/core/data/service/test_cluster.cc b/tensorflow/core/data/service/test_cluster.cc
index bfa337b..058eec7 100644
--- a/tensorflow/core/data/service/test_cluster.cc
+++ b/tensorflow/core/data/service/test_cluster.cc
@@ -57,7 +57,7 @@
}
Status TestCluster::AddWorker() {
- std::unique_ptr<GrpcDataServer> worker;
+ std::unique_ptr<WorkerGrpcDataServer> worker;
TF_RETURN_IF_ERROR(
NewWorkerServer(/*port=*/0, kProtocol, master_address_, &worker));
TF_RETURN_IF_ERROR(worker->Start());
diff --git a/tensorflow/core/data/service/test_cluster.h b/tensorflow/core/data/service/test_cluster.h
index 6aa75f4..c4b05ad 100644
--- a/tensorflow/core/data/service/test_cluster.h
+++ b/tensorflow/core/data/service/test_cluster.h
@@ -42,9 +42,9 @@
private:
bool initialized_ = false;
int num_workers_;
- std::unique_ptr<GrpcDataServer> master_;
+ std::unique_ptr<MasterGrpcDataServer> master_;
std::string master_address_;
- std::vector<std::unique_ptr<GrpcDataServer>> workers_;
+ std::vector<std::unique_ptr<WorkerGrpcDataServer>> workers_;
std::vector<std::string> worker_addresses_;
};
diff --git a/tensorflow/core/data/service/worker_impl.cc b/tensorflow/core/data/service/worker_impl.cc
index dde51ab..b0d3275 100644
--- a/tensorflow/core/data/service/worker_impl.cc
+++ b/tensorflow/core/data/service/worker_impl.cc
@@ -50,11 +50,20 @@
tf_data_service_created->GetCell()->Set(true);
}
+DataServiceWorkerImpl::~DataServiceWorkerImpl() {
+ mutex_lock l(mu_);
+ cancelled_ = true;
+ heartbeat_cv_.notify_one();
+}
+
void DataServiceWorkerImpl::Start(const std::string& worker_address) {
VLOG(3) << "Starting tf.data service worker at address " << worker_address;
mutex_lock l(mu_);
worker_address_ = worker_address;
+ Thread* thread = Env::Default()->StartThread(
+ {}, "data-service-worker-heartbeat", [this]() { HeartbeatThread(); });
+ heartbeat_thread_.reset(thread);
Status s = Register();
while (!s.ok()) {
LOG(WARNING) << "Failed to register with master at " << master_address_
@@ -64,32 +73,6 @@
}
}
-Status DataServiceWorkerImpl::Register() EXCLUSIVE_LOCKS_REQUIRED(mu_) {
- VLOG(3) << "Registering with master at " << master_address_;
- if (!master_stub_) {
- ::grpc::ChannelArguments args;
- std::shared_ptr<::grpc::ChannelCredentials> credentials;
- TF_RETURN_IF_ERROR(
- CredentialsFactory::CreateClientCredentials(protocol_, &credentials));
- auto channel =
- ::grpc::CreateCustomChannel(master_address_, credentials, args);
- master_stub_ = MasterService::NewStub(channel);
- }
- RegisterWorkerRequest req;
- req.set_worker_address(worker_address_);
- RegisterWorkerResponse resp;
-
- grpc::ClientContext ctx;
- grpc::Status s = master_stub_->RegisterWorker(&ctx, req, &resp);
- if (!s.ok()) {
- return grpc_util::WrapError("Failed to register worker", s);
- }
- for (const TaskDef& task : resp.tasks()) {
- TF_RETURN_IF_ERROR(ProcessTaskInternal(task));
- }
- VLOG(3) << "Registered worker with id " << resp.worker_id();
- return Status::OK();
-}
Status DataServiceWorkerImpl::ProcessTask(const ProcessTaskRequest* request,
ProcessTaskResponse* response) {
@@ -141,6 +124,8 @@
if (end_of_sequence) {
// Release iterator memory and leave a null entry as a tombstone.
iter.reset();
+ pending_completed_tasks_.push_back(request->task_id());
+ heartbeat_cv_.notify_one();
}
}
@@ -153,5 +138,79 @@
return Status::OK();
}
+Status DataServiceWorkerImpl::EnsureMasterStubInitialized()
+ EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+ if (!master_stub_) {
+ ::grpc::ChannelArguments args;
+ std::shared_ptr<::grpc::ChannelCredentials> credentials;
+ TF_RETURN_IF_ERROR(
+ CredentialsFactory::CreateClientCredentials(protocol_, &credentials));
+ auto channel =
+ ::grpc::CreateCustomChannel(master_address_, credentials, args);
+ master_stub_ = MasterService::NewStub(channel);
+ }
+ return Status::OK();
+}
+
+Status DataServiceWorkerImpl::Register() EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+ VLOG(3) << "Registering with master at " << master_address_;
+ TF_RETURN_IF_ERROR(EnsureMasterStubInitialized());
+ RegisterWorkerRequest req;
+ req.set_worker_address(worker_address_);
+ RegisterWorkerResponse resp;
+
+ grpc::ClientContext ctx;
+ grpc::Status s = master_stub_->RegisterWorker(&ctx, req, &resp);
+ if (!s.ok()) {
+ return grpc_util::WrapError("Failed to register worker", s);
+ }
+ for (const TaskDef& task : resp.tasks()) {
+ TF_RETURN_IF_ERROR(ProcessTaskInternal(task));
+ }
+ worker_id_ = resp.worker_id();
+ VLOG(3) << "Registered worker with id " << resp.worker_id();
+ return Status::OK();
+}
+
+Status DataServiceWorkerImpl::SendTaskUpdate() EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+ VLOG(3) << "Sending " << pending_completed_tasks_.size()
+ << " task updates to master";
+ TF_RETURN_IF_ERROR(EnsureMasterStubInitialized());
+ WorkerUpdateRequest req;
+ req.set_worker_id(worker_id_);
+ for (int task_id : pending_completed_tasks_) {
+ TaskProgress* update = req.add_updates();
+ update->set_task_id(task_id);
+ update->set_completed(true);
+ }
+
+ WorkerUpdateResponse resp;
+ grpc::ClientContext ctx;
+ grpc::Status s = master_stub_->WorkerUpdate(&ctx, req, &resp);
+ if (!s.ok()) {
+ return grpc_util::WrapError("Failed to send task updates", s);
+ }
+ pending_completed_tasks_.clear();
+ VLOG(3) << "Sent " << req.updates().size() << " task updates ";
+ return Status::OK();
+}
+
+void DataServiceWorkerImpl::HeartbeatThread() {
+ while (true) {
+ mutex_lock l(mu_);
+ while (!cancelled_ && pending_completed_tasks_.empty()) {
+ heartbeat_cv_.wait(l);
+ }
+ if (cancelled_) {
+ VLOG(3) << "Heartbeat thread shutting down";
+ return;
+ }
+ Status s = SendTaskUpdate();
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to send task updates to master: " << s;
+ }
+ }
+}
+
} // namespace data
} // namespace tensorflow
diff --git a/tensorflow/core/data/service/worker_impl.h b/tensorflow/core/data/service/worker_impl.h
index 9595702..8c5fc2e 100644
--- a/tensorflow/core/data/service/worker_impl.h
+++ b/tensorflow/core/data/service/worker_impl.h
@@ -31,7 +31,7 @@
public:
explicit DataServiceWorkerImpl(const std::string& master_address,
const std::string& protocol);
- virtual ~DataServiceWorkerImpl() {}
+ ~DataServiceWorkerImpl();
// Starts the worker. The worker needs to know its own address so that it can
// register with the master.
@@ -48,10 +48,16 @@
GetElementResponse* response);
private:
+ // Sets master_stub_ if it isn't already set.
+ Status EnsureMasterStubInitialized();
// Registers the worker with the master.
Status Register();
+ // Sends task status to the master.
+ Status SendTaskUpdate();
// Creates an iterator to process a task.
Status ProcessTaskInternal(const TaskDef& task);
+ // A thread for updating the master with worker status.
+ void HeartbeatThread();
typedef struct Task {
int64 id;
@@ -68,9 +74,16 @@
std::string worker_address_;
mutex mu_;
+ int64 worker_id_ TF_GUARDED_BY(mu_);
std::unique_ptr<MasterService::Stub> master_stub_ TF_GUARDED_BY(mu_);
// Information about tasks, keyed by task ids.
absl::flat_hash_map<int64, Task> tasks_ TF_GUARDED_BY(mu_);
+ // List of completed tasks which haven't yet been communicated to the master.
+ std::vector<int64> pending_completed_tasks_ TF_GUARDED_BY(mu_);
+ bool cancelled_ TF_GUARDED_BY(mu_) = false;
+ // Condition variable for notifying the heartbeat thread.
+ condition_variable heartbeat_cv_ TF_GUARDED_BY(mu_);
+ std::unique_ptr<Thread> heartbeat_thread_;
TF_DISALLOW_COPY_AND_ASSIGN(DataServiceWorkerImpl);
};
diff --git a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
index 169ace9..ab511e4 100644
--- a/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
+++ b/tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
@@ -286,14 +286,9 @@
TF_RETURN_IF_ERROR(worker_session->worker_cache()->GetEagerClientCache(
&remote_eager_workers));
- DistributedFunctionLibraryRuntime* cluster_flr =
- eager::CreateClusterFLR(request->context_id(), ctx, worker_session.get());
-
ctx->ClearCachesAndThreadExecutors();
- Status s = ctx->UpdateRemoteWorker(
- device_mgr, std::move(remote_eager_workers),
- worker_session->remote_device_mgr(), remote_workers,
- request->context_id(), cluster_flr);
+ Status s = ctx->UpdateRemoteWorker(std::move(remote_eager_workers),
+ remote_workers, request->context_id());
if (!s.ok()) {
VLOG(1) << "EagerContext::UpdateRemoteWorker failed with " << s.ToString();
return s;
diff --git a/tensorflow/core/distributed_runtime/eager/eager_service_impl.h b/tensorflow/core/distributed_runtime/eager/eager_service_impl.h
index f1c02f0..2386998 100644
--- a/tensorflow/core/distributed_runtime/eager/eager_service_impl.h
+++ b/tensorflow/core/distributed_runtime/eager/eager_service_impl.h
@@ -87,7 +87,7 @@
Status CreateMasterContext(const tensorflow::uint64 context_id,
EagerContext* context);
- static const uint64 kInvalidStreamId = 0;
+ static constexpr uint64 kInvalidStreamId = 0;
// Used by both Enqueue and StreamingEnqueue RPCs.
Status Enqueue(const EnqueueRequest* request, EnqueueResponse* response,
diff --git a/tensorflow/core/distributed_runtime/eager/remote_execute_node.cc b/tensorflow/core/distributed_runtime/eager/remote_execute_node.cc
index 3eab62b..067e26a 100644
--- a/tensorflow/core/distributed_runtime/eager/remote_execute_node.cc
+++ b/tensorflow/core/distributed_runtime/eager/remote_execute_node.cc
@@ -76,7 +76,8 @@
if (!s.ok()) {
LOG(ERROR) << "Ignoring an error encountered when setting "
"remote shape of tensor handle: "
- << retvals[i] << " with status: " << status.ToString()
+ << retvals[i]
+ << " with execute status: " << status.ToString()
<< " and SetRemoteShape status: " << s.ToString()
<< "\nThis should never happen. "
"Please file an issue with the TensorFlow Team.";
diff --git a/tensorflow/core/distributed_runtime/eager/remote_mgr_test.cc b/tensorflow/core/distributed_runtime/eager/remote_mgr_test.cc
index eb2f2ae..84eee59 100644
--- a/tensorflow/core/distributed_runtime/eager/remote_mgr_test.cc
+++ b/tensorflow/core/distributed_runtime/eager/remote_mgr_test.cc
@@ -102,6 +102,52 @@
handle->Unref();
}
+TEST_F(RemoteMgrTest, InvalidateRemoteMirrorWithClusterUpdate) {
+ RemoteMgr remote_mgr(false, ctx_);
+ Tensor t(DT_FLOAT, TensorShape({0}));
+
+ TensorHandle* handle = TensorHandle::CreateLocalHandle(
+ std::move(t), local_device_, local_device_, ctx_);
+ const uint64 op_id = 2;
+ const int output_num = 3;
+ TF_ASSERT_OK(handle->AddUnshapedRemoteMirror(remote_device_, op_id,
+ output_num, "", ctx_));
+ EXPECT_TRUE(
+ handle->HasRemoteMirror(remote_device_, ctx_->GetContextViewId()));
+
+ // When updating cluster, remote mirror should be invalidated.
+ ctx_->IncrementContextViewId();
+ EXPECT_FALSE(
+ handle->HasRemoteMirror(remote_device_, ctx_->GetContextViewId()));
+ // Setting remote shape should still be OK
+ TF_ASSERT_OK(handle->SetRemoteShape(TensorShape({0}), remote_device_,
+ ctx_->GetContextViewId()));
+ handle->Unref();
+}
+
+TEST_F(RemoteMgrTest, SetRemoteShapeWithClusterUpdate) {
+ RemoteMgr remote_mgr(false, ctx_);
+
+ const uint64 op_id = 3;
+ const int output_num = 1;
+ TensorHandle* handle = TensorHandle::CreateUnshapedRemoteHandle(
+ op_id, output_num,
+ /*remote_task=*/"", DT_FLOAT, remote_device_, ctx_);
+ TF_ASSERT_OK(handle->SetRemoteShape(TensorShape({0}), remote_device_,
+ ctx_->GetContextViewId()));
+ handle->Unref();
+
+ // Setting remote shape on primary (non-mirror) remote handle works after
+ // cluster being updated
+ handle = TensorHandle::CreateUnshapedRemoteHandle(
+ op_id, output_num,
+ /*remote_task=*/"", DT_FLOAT, remote_device_, ctx_);
+ ctx_->IncrementContextViewId();
+ TF_ASSERT_OK(handle->SetRemoteShape(TensorShape({0}), remote_device_,
+ ctx_->GetContextViewId()));
+ handle->Unref();
+}
+
} // namespace
} // namespace eager
} // namespace tensorflow
diff --git a/tensorflow/core/distributed_runtime/eager/remote_tensor_handle_data.cc b/tensorflow/core/distributed_runtime/eager/remote_tensor_handle_data.cc
index 1f0f5a4..8e6b97f 100644
--- a/tensorflow/core/distributed_runtime/eager/remote_tensor_handle_data.cc
+++ b/tensorflow/core/distributed_runtime/eager/remote_tensor_handle_data.cc
@@ -173,6 +173,10 @@
}
Status RemoteTensorHandleData::SetShape(const TensorShape& shape) {
+ // If `is_ready_` is set previously due to poisoning, return the original
+ // error that poisoned this tensor.
+ TF_RETURN_IF_ERROR(IsPoisoned());
+
mutex_lock l(mu_);
if (is_ready_) {
return errors::Internal("SetShape is only called on non-ready handles.");
diff --git a/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc b/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
index 37e88ba..5bb61eb 100644
--- a/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
+++ b/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
@@ -197,7 +197,7 @@
}
private:
- static const int kMaxObjects = 1000;
+ static constexpr int kMaxObjects = 1000;
mutex mu_;
std::vector<RpcRecvTensorCall*> objects_ TF_GUARDED_BY(mu_);
diff --git a/tensorflow/core/distributed_runtime/worker_session.cc b/tensorflow/core/distributed_runtime/worker_session.cc
index 8a75870..ecdbfec 100644
--- a/tensorflow/core/distributed_runtime/worker_session.cc
+++ b/tensorflow/core/distributed_runtime/worker_session.cc
@@ -134,9 +134,6 @@
TF_RETURN_IF_ERROR(remote_device_mgr_->RemoveDevices(removed_remote_devices));
TF_RETURN_IF_ERROR(
remote_device_mgr_->AddDevices(std::move(added_remote_devices)));
- cluster_flr_ = std::unique_ptr<ClusterFunctionLibraryRuntime>(
- new ClusterFunctionLibraryRuntime(this, !session_name_.empty(),
- remote_device_mgr()));
return Status::OK();
}
diff --git a/tensorflow/core/framework/model.cc b/tensorflow/core/framework/model.cc
index 85eeeb6..322e301 100644
--- a/tensorflow/core/framework/model.cc
+++ b/tensorflow/core/framework/model.cc
@@ -636,6 +636,8 @@
} // namespace
+thread_local int64 Node::work_start_;
+
std::shared_ptr<Parameter> MakeParameter(const string& name,
std::shared_ptr<SharedState> state,
double min, double max) {
diff --git a/tensorflow/core/framework/model.h b/tensorflow/core/framework/model.h
index c5d477e..1cea4d5 100644
--- a/tensorflow/core/framework/model.h
+++ b/tensorflow/core/framework/model.h
@@ -236,18 +236,15 @@
// Records that a node thread has started executing.
void record_start(int64 time_nanos) TF_LOCKS_EXCLUDED(mu_) {
- mutex_lock l(mu_);
- work_start_[std::this_thread::get_id()] = time_nanos;
+ DCHECK_EQ(work_start_, 0);
+ work_start_ = time_nanos;
}
// Records that a node thread has stopped executing.
void record_stop(int64 time_nanos) TF_LOCKS_EXCLUDED(mu_) {
- mutex_lock l(mu_);
- std::thread::id tid = std::this_thread::get_id();
- auto iter = work_start_.find(tid);
- if (iter != work_start_.end()) {
- processing_time_ += time_nanos - iter->second;
- work_start_.erase(iter);
+ if (work_start_ != 0) {
+ processing_time_ += time_nanos - work_start_;
+ work_start_ = 0;
} else {
VLOG(1) << "Encountered a stop event without a matching start event.";
}
@@ -436,6 +433,17 @@
void TotalMaximumBufferedBytesHelper(
absl::flat_hash_map<string, double>* total_bytes) const;
+ // Stores the time passed to the last call to `Node::record_start()` on the
+ // current thread.
+ //
+ // NOTE: This thread-local variable is shared between all instances of `Node`
+ // on which the same thread calls `record_start()` or `record_stop()`. It
+ // relies on the invariant that at most one `Node` can be "active" on a
+ // particular thread at any time. Therefore if `n->record_start()` is called
+ // on thread `t`, then `n->record_stop()` must be called before another call
+ // to `Node::record_start()` (for any node).
+ static thread_local int64 work_start_; // Will be initialized to zero.
+
mutable mutex mu_;
const int64 id_;
const string name_;
@@ -454,7 +462,6 @@
Metrics metrics_;
absl::flat_hash_map<string, std::shared_ptr<Parameter>> parameters_
TF_GUARDED_BY(mu_);
- absl::flat_hash_map<std::thread::id, int64> work_start_ TF_GUARDED_BY(mu_);
// Statistic of inputs processing time history.
double input_processing_time_sum_ = 0.0L;
diff --git a/tensorflow/core/framework/node_def_util.cc b/tensorflow/core/framework/node_def_util.cc
index 2873431..86911c7 100644
--- a/tensorflow/core/framework/node_def_util.cc
+++ b/tensorflow/core/framework/node_def_util.cc
@@ -487,6 +487,11 @@
if (arg_def.is_ref()) {
// For all types that were added by this function call, make them refs.
for (size_t i = original_size; i < sig->size(); ++i) {
+ if (IsRefType((*sig)[i])) {
+ return errors::InvalidArgument(
+ "Requested reference to a reference type: ",
+ arg_def.ShortDebugString());
+ }
(*sig)[i] = MakeRefType((*sig)[i]);
}
}
diff --git a/tensorflow/core/framework/op_kernel.h b/tensorflow/core/framework/op_kernel.h
index 2f14031..f31effb 100644
--- a/tensorflow/core/framework/op_kernel.h
+++ b/tensorflow/core/framework/op_kernel.h
@@ -694,8 +694,8 @@
checkpoint::TensorSliceReaderCacheWrapper* slice_reader_cache = nullptr;
// Support for forwarding reservations (used by ScopedAllocator).
- static const int kNeverForward = -2;
- static const int kNoReservation = -1;
+ static constexpr int kNeverForward = -2;
+ static constexpr int kNoReservation = -1;
// Values in [0,...) represent reservations for the indexed output.
const int* forward_from_array = nullptr;
diff --git a/tensorflow/core/framework/shape_inference_test.cc b/tensorflow/core/framework/shape_inference_test.cc
index d413882..45cfb23 100644
--- a/tensorflow/core/framework/shape_inference_test.cc
+++ b/tensorflow/core/framework/shape_inference_test.cc
@@ -74,7 +74,7 @@
void TestMergeHandles(bool input_not_output);
void TestRelaxHandles(bool input_not_output);
- static const int kVersion = 0; // used for graph-def version.
+ static constexpr int kVersion = 0; // used for graph-def version.
};
TEST_F(ShapeInferenceTest, InputOutputByName) {
diff --git a/tensorflow/core/framework/tensor_shape.h b/tensorflow/core/framework/tensor_shape.h
index ac1bef1..b0d4944 100644
--- a/tensorflow/core/framework/tensor_shape.h
+++ b/tensorflow/core/framework/tensor_shape.h
@@ -103,10 +103,10 @@
// We use the max value of uint16 or uint32 to represent unknown shapes, so
// the maximum representable valid shape in these representations is one less.
- static const int64 kMaxRep16 = std::numeric_limits<uint16>::max() - 1;
- static const int64 kMaxRep32 = std::numeric_limits<uint32>::max() - 1;
- static const uint16 kUnknownRep16 = std::numeric_limits<uint16>::max();
- static const uint32 kUnknownRep32 = std::numeric_limits<uint32>::max();
+ static constexpr int64 kMaxRep16 = std::numeric_limits<uint16>::max() - 1;
+ static constexpr int64 kMaxRep32 = std::numeric_limits<uint32>::max() - 1;
+ static constexpr uint16 kUnknownRep16 = std::numeric_limits<uint16>::max();
+ static constexpr uint32 kUnknownRep32 = std::numeric_limits<uint32>::max();
Rep16* as16() { return reinterpret_cast<Rep16*>(buf()); }
Rep32* as32() { return reinterpret_cast<Rep32*>(buf()); }
@@ -134,7 +134,7 @@
// We store the number of dimensions in byte 14, and the RepTag in byte 15.
// Bytes [0..13] vary depending on the representation.
// A value of 255 indicates unknown rank in the PartialTensorShape case.
- static const uint8 kUnknownRank = 255;
+ static constexpr uint8 kUnknownRank = 255;
uint8 ndims_byte() const { return buf()[14]; }
void set_ndims_byte(uint8 nd) { buf()[14] = nd; }
diff --git a/tensorflow/core/framework/tensor_testutil.h b/tensorflow/core/framework/tensor_testutil.h
index 1d476ba..80dddfb 100644
--- a/tensorflow/core/framework/tensor_testutil.h
+++ b/tensorflow/core/framework/tensor_testutil.h
@@ -116,11 +116,11 @@
template <typename T>
struct is_floating_point_type {
- static const bool value = std::is_same<T, Eigen::half>::value ||
- std::is_same<T, float>::value ||
- std::is_same<T, double>::value ||
- std::is_same<T, std::complex<float>>::value ||
- std::is_same<T, std::complex<double>>::value;
+ static constexpr bool value = std::is_same<T, Eigen::half>::value ||
+ std::is_same<T, float>::value ||
+ std::is_same<T, double>::value ||
+ std::is_same<T, std::complex<float>>::value ||
+ std::is_same<T, std::complex<double>>::value;
};
template <typename T>
diff --git a/tensorflow/core/framework/type_traits.h b/tensorflow/core/framework/type_traits.h
index 96fbf92..a7826a6 100644
--- a/tensorflow/core/framework/type_traits.h
+++ b/tensorflow/core/framework/type_traits.h
@@ -26,10 +26,10 @@
// Functions to define quantization attribute of types.
struct true_type {
- static const bool value = true;
+ static constexpr bool value = true;
};
struct false_type {
- static const bool value = false;
+ static constexpr bool value = false;
};
// Default is_quantized is false.
diff --git a/tensorflow/core/graph/graph_partition.h b/tensorflow/core/graph/graph_partition.h
index 8020c2d..04ea0ac 100644
--- a/tensorflow/core/graph/graph_partition.h
+++ b/tensorflow/core/graph/graph_partition.h
@@ -42,7 +42,7 @@
// A function that returns the incarnation of a device given the
// device's fullname. If not found, GetIncarnationFunc should return
// kIllegalIncarnation.
- static const uint64 kIllegalIncarnation = 0;
+ static constexpr uint64 kIllegalIncarnation = 0;
typedef std::function<uint64(const string&)> GetIncarnationFunc;
GetIncarnationFunc get_incarnation = nullptr;
diff --git a/tensorflow/core/graph/mkl_layout_pass.cc b/tensorflow/core/graph/mkl_layout_pass.cc
index c27c7aa..b1d692d 100644
--- a/tensorflow/core/graph/mkl_layout_pass.cc
+++ b/tensorflow/core/graph/mkl_layout_pass.cc
@@ -273,6 +273,7 @@
csinfo_.fused_batch_norm_v3 = "FusedBatchNormV3";
csinfo_.fused_batch_norm_grad_v3 = "FusedBatchNormGradV3";
csinfo_.fused_conv2d = "_FusedConv2D";
+ csinfo_.fused_depthwise_conv2d = "_FusedDepthwiseConv2dNative";
csinfo_.fused_matmul = "_FusedMatMul";
csinfo_.identity = "Identity";
csinfo_.leakyrelu = "LeakyRelu";
@@ -295,6 +296,7 @@
csinfo_.mkl_depthwise_conv2d_grad_filter =
"_MklDepthwiseConv2dNativeBackpropFilter";
csinfo_.mkl_fused_conv2d = "_MklFusedConv2D";
+ csinfo_.mkl_fused_depthwise_conv2d = "_MklFusedDepthwiseConv2dNative";
csinfo_.mkl_fused_matmul = "_MklFusedMatMul";
csinfo_.mkl_pad_with_conv2d = "_MklPadWithConv2D";
csinfo_.mkl_pad_with_fused_conv2d = "_MklPadWithFusedConv2D";
@@ -479,6 +481,10 @@
rinfo_.push_back({csinfo_.fused_conv2d, csinfo_.mkl_fused_conv2d,
CopyAttrsFusedConv2D, FusedConv2DRewrite,
kRewriteForLayoutPropagation});
+ rinfo_.push_back({csinfo_.fused_depthwise_conv2d,
+ csinfo_.mkl_fused_depthwise_conv2d, CopyAttrsFusedConv2D,
+ FusedDepthwiseConv2DRewrite,
+ kRewriteForLayoutPropagation});
rinfo_.push_back({csinfo_.fused_matmul, csinfo_.mkl_fused_matmul,
CopyAttrsAllCheckConstFilter, FusedMatMulRewrite});
@@ -925,6 +931,7 @@
string fused_batch_norm_v3;
string fused_batch_norm_grad_v3;
string fused_conv2d;
+ string fused_depthwise_conv2d;
string fused_matmul;
string identity;
string leakyrelu;
@@ -945,6 +952,7 @@
string mkl_depthwise_conv2d_grad_input;
string mkl_depthwise_conv2d_grad_filter;
string mkl_fused_conv2d;
+ string mkl_fused_depthwise_conv2d;
string mkl_fused_matmul;
string mkl_pad_with_conv2d;
string mkl_pad_with_fused_conv2d;
@@ -1675,6 +1683,25 @@
fused_ops == std::vector<string>{"BiasAdd", "Add", "Relu"});
}
+ static bool FusedDepthwiseConv2DRewrite(const Node* n) {
+ // MKL DNN currently doesn't support all fusions that grappler fuses
+ // together with DepthwiseConv2D (ex. batchnorm). We rewrite
+ // _FusedDepthwiseConv2DNative only if it includes those we support.
+ DataType T;
+ if (!TryGetNodeAttr(n->def(), "T", &T) ||
+ !mkl_op_registry::IsMklLayoutDependentOp(
+ csinfo_.mkl_fused_depthwise_conv2d, T)) {
+ return false;
+ }
+
+ std::vector<string> fused_ops;
+ TF_CHECK_OK(GetNodeAttr(n->def(), "fused_ops", &fused_ops));
+ return (fused_ops == std::vector<string>{"BiasAdd"} ||
+ fused_ops == std::vector<string>{"BiasAdd", "Relu"} ||
+ fused_ops == std::vector<string>{"BiasAdd", "Relu6"} ||
+ fused_ops == std::vector<string>{"BiasAdd", "Elu"});
+ }
+
// Rewrites input node to a new node specified by its matching rewrite info.
//
// Method first searches matching rewrite info for input node and then
@@ -3703,6 +3730,7 @@
n->type_string() != csinfo_.pad_with_fused_conv2d &&
n->type_string() != csinfo_.conv2d_grad_filter_with_bias &&
n->type_string() != csinfo_.fused_conv2d &&
+ n->type_string() != csinfo_.fused_depthwise_conv2d &&
n->type_string() != csinfo_.fused_matmul &&
!mkl_op_registry::IsMklOp(mkl_op_registry::GetMklOpName(n->type_string()),
T)) {
diff --git a/tensorflow/core/graph/mkl_layout_pass_test.cc b/tensorflow/core/graph/mkl_layout_pass_test.cc
index 6fe969a..f66b3fc 100644
--- a/tensorflow/core/graph/mkl_layout_pass_test.cc
+++ b/tensorflow/core/graph/mkl_layout_pass_test.cc
@@ -1789,6 +1789,56 @@
REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedConv2D_Positive7);
#undef REGISTER_TEST
+// Rewrite test for _FusedDepthwiseConv2dNative Op fusion
+#define REGISTER_TEST(NAME, T, INPUT) \
+ TEST_F(MklLayoutPassTest, NAME##_##T) { \
+ InitGraph( \
+ "node { name: 'A' op: '" #INPUT "'}" \
+ "node { name: 'B' op: '" #INPUT "'}" \
+ "node { name: 'C' op: '" #INPUT "'}" \
+ "node { name: 'D' op: '_FusedDepthwiseConv2dNative'" \
+ " attr { key: 'T' value { type: " #T " } }" \
+ " attr { key: 'num_args' value { i: 1 } }" \
+ " attr { key: 'data_format' value { s: 'NCHW' } }" \
+ " attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'padding' value { s: 'SAME' } }" \
+ " attr { key: 'dilations' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'fused_ops' value { list: " FUSED_OPS " } }" \
+ " attr { key: 'epsilon' value { f: 0.001 }}" \
+ " input: ['A', 'B', 'C']}" \
+ "node { name: 'E' op: 'Zeta'" \
+ "attr { key: 'T' value { type: " #T " } }" \
+ " input: ['D', 'C'] }"); \
+ EXPECT_EQ(DoMklLayoutOptimizationPass(), \
+ "A(" #INPUT ");B(" #INPUT ");C(" #INPUT ");" \
+ "D(_MklFusedDepthwiseConv2dNative);" \
+ "DMT/_0(Const);DMT/_1(Const);DMT/_2(Const);E(Zeta)|A->D;" \
+ "A:control->DMT/_0:control;A:control->DMT/_1:control;" \
+ "A:control->DMT/_2:control;B->D:1;C->D:2;C->E:1;D->E;" \
+ "DMT/_0->D:3;DMT/_1->D:4;DMT/_2->D:5"); \
+ }
+
+// BiasAdd fusion
+#define FUSED_OPS "{s: 'BiasAdd'}"
+REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedDepthwiseConv2dNative_Positive1);
+
+// BiasAdd + Relu fusion
+#define FUSED_OPS "{s: 'BiasAdd', s: 'Relu'}"
+REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedDepthwiseConv2dNative_Positive2);
+
+// BiasAdd + Relu6 fusion
+#define FUSED_OPS "{s: 'BiasAdd', s: 'Relu6'}"
+REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedDepthwiseConv2dNative_Positive3);
+
+// BiasAdd + Elu fusion
+#define FUSED_OPS "{s: 'BiasAdd', s: 'Elu'}"
+REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedDepthwiseConv2dNative_Positive4);
+
+#undef FUSED_OPS
+#undef REGISTER_TEST
+
// Rewrite test for _FusedConv2D Op with unsupported fusion
#define REGISTER_TEST(NAME, T, INPUT) \
TEST_F(MklLayoutPassTest, NAME##_##T) { \
@@ -1818,6 +1868,36 @@
REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedConv2D_Negative1);
#undef REGISTER_TEST
+// Rewrite test for _FusedDepthwiseConv2dNative with unsupported fusion
+#define REGISTER_TEST(NAME, T, INPUT) \
+ TEST_F(MklLayoutPassTest, NAME##_##T) { \
+ InitGraph( \
+ "node { name: 'A' op: '" #INPUT "'}" \
+ "node { name: 'B' op: '" #INPUT "'}" \
+ "node { name: 'C' op: '" #INPUT "'}" \
+ "node { name: 'D' op: '_FusedDepthwiseConv2dNative'" \
+ " attr { key: 'T' value { type: " #T " } }" \
+ " attr { key: 'num_args' value { i: 1 } }" \
+ " attr { key: 'data_format' value { s: 'NCHW' } }" \
+ " attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'padding' value { s: 'SAME' } }" \
+ " attr { key: 'dilations' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'fused_ops' value { list: {s: 'Unsupported'} } }" \
+ " attr { key: 'epsilon' value { f: 0.001 }}" \
+ " input: ['A', 'B', 'C']}" \
+ "node { name: 'E' op: 'Zeta'" \
+ "attr { key: 'T' value { type: " #T " } }" \
+ " input: ['D', 'C'] }"); \
+ EXPECT_EQ(DoMklLayoutOptimizationPass(), \
+ "A(" #INPUT ");B(" #INPUT ");C(" #INPUT ");" \
+ "D(_FusedDepthwiseConv2dNative);" \
+ "E(Zeta)|A->D;B->D:1;C->D:2;C->E:1;D->E"); \
+ }
+REGISTER_TEST_ALL_TYPES(NodeRewrite_FusedDepthwiseConv2dNative_Negative1);
+#undef REGISTER_TEST
+
// Rewrite test for _FusedConv2D Op with unsupported type
#define REGISTER_TEST(NAME, T, INPUT) \
TEST_F(MklLayoutPassTest, NAME##_##T) { \
@@ -1847,6 +1927,37 @@
REGISTER_TEST(NodeRewrite_FusedConv2D_Negative2, DT_DOUBLE, DoubleInput);
#undef REGISTER_TEST
+// Rewrite test for _FusedDepthwiseConv2dNativeOp with unsupported type
+#define REGISTER_TEST(NAME, T, INPUT) \
+ TEST_F(MklLayoutPassTest, NAME##_##T) { \
+ InitGraph( \
+ "node { name: 'A' op: '" #INPUT "'}" \
+ "node { name: 'B' op: '" #INPUT "'}" \
+ "node { name: 'C' op: '" #INPUT "'}" \
+ "node { name: 'D' op: '_FusedDepthwiseConv2dNative'" \
+ " attr { key: 'T' value { type:" #T "} }" \
+ " attr { key: 'num_args' value { i: 1 } }" \
+ " attr { key: 'data_format' value { s: 'NCHW' } }" \
+ " attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'padding' value { s: 'SAME' } }" \
+ " attr { key: 'dilations' value { list: {i: 1, i:1, i:1, i:1} " \
+ "} }" \
+ " attr { key: 'fused_ops' value { list: {s: 'BiasAdd'} } }" \
+ " attr { key: 'epsilon' value { f: 0.001 }}" \
+ " input: ['A', 'B', 'C']}" \
+ "node { name: 'E' op: 'Zeta'" \
+ "attr { key: 'T' value { type: " #T "} }" \
+ " input: ['D', 'C'] }"); \
+ EXPECT_EQ(DoMklLayoutOptimizationPass(), \
+ "A(" #INPUT ");B(" #INPUT ");C(" #INPUT ");" \
+ "D(_FusedDepthwiseConv2dNative);" \
+ "E(Zeta)|A->D;B->D:1;C->D:2;C->E:1;D->E"); \
+}
+REGISTER_TEST(NodeRewrite_FusedDepthwiseConv2dNative_Negative2,
+ DT_DOUBLE, DoubleInput);
+#undef REGISTER_TEST
+
// Test set: _FusedMatMul -> MklFusedMatMul rewrite tests
#define REGISTER_TEST(NAME, T, INPUT) \
TEST_F(MklLayoutPassTest, NAME##_##T) { \
@@ -4240,6 +4351,33 @@
"_MklFusedConv2D"));
}
+// _FusedDepthwiseConv2dNative + BiasAdd fusion where filter is a constant.
+TEST_F(MklLayoutPassTest,
+ FusedDepthwiseConv2dNativeWithBias_FilterCaching_Positive) {
+ InitGraph(
+ "node { name: 'A' op: 'Input'}"
+ "node { name: 'B' op: 'Const'" // Filter
+ " attr { key: 'dtype' value { type: DT_FLOAT } }"
+ " attr { key: 'value' value { "
+ " tensor { dtype: DT_FLOAT tensor_shape { dim { size: 1 } } "
+ " int_val: 0 } } } }"
+ "node { name: 'C' op: 'Input'}"
+ "node { name: 'D' op: '_FusedDepthwiseConv2dNative'"
+ " attr { key: 'T' value { type: DT_FLOAT } }"
+ " attr { key: 'num_args' value { i: 1 } }"
+ " attr { key: 'data_format' value { s: 'NCHW' } }"
+ " attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} } }"
+ " attr { key: 'padding' value { s: 'SAME' } }"
+ " attr { key: 'dilations' value { list: {i: 1, i:1, i:1, i:1} } }"
+ " attr { key: 'fused_ops' value { list: {s: 'BiasAdd'} } }"
+ " attr { key: 'epsilon' value { f: 0.001 }}"
+ " input: ['A', 'B', 'C']}"
+ "node { name: 'E' op: 'Zeta' attr { key: 'T' value { type: DT_FLOAT } }"
+ " input: ['D', 'C'] }");
+ EXPECT_TRUE(DoMklLayoutOptimizationPassGetAttrVal<bool>(
+ "is_filter_const", "_MklFusedDepthwiseConv2dNative"));
+}
+
// _FusedConv2D + BiasAdd fusion where filter is NOT a constant.
TEST_F(MklLayoutPassTest, FusedConv2DWithBias_FilterCaching_Negative) {
InitGraph(
@@ -4262,6 +4400,28 @@
"_MklFusedConv2D"));
}
+// _FusedDepthwiseConv2dNative + BiasAdd fusion where filter is NOT a constant.
+TEST_F(MklLayoutPassTest,
+ FusedDepthwiseConv2dNativeWithBias_FilterCaching_Negative) {
+ InitGraph(
+ "node { name: 'A' op: 'Input'}"
+ "node { name: 'B' op: 'Input'}" // Filter
+ "node { name: 'C' op: 'Input'}"
+ "node { name: 'D' op: '_FusedDepthwiseConv2dNative'"
+ " attr { key: 'T' value { type: DT_FLOAT } }"
+ " attr { key: 'num_args' value { i: 1 } }"
+ " attr { key: 'data_format' value { s: 'NCHW' } }"
+ " attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} } }"
+ " attr { key: 'padding' value { s: 'SAME' } }"
+ " attr { key: 'dilations' value { list: {i: 1, i:1, i:1, i:1} } }"
+ " attr { key: 'fused_ops' value { list: {s: 'BiasAdd'} } }"
+ " attr { key: 'epsilon' value { f: 0.001 }}"
+ " input: ['A', 'B', 'C']}"
+ "node { name: 'E' op: 'Zeta' attr { key: 'T' value { type: DT_FLOAT } }"
+ " input: ['D', 'C'] }");
+ EXPECT_FALSE(DoMklLayoutOptimizationPassGetAttrVal<bool>(
+ "is_filter_const", "_MklFusedDepthwiseConv2dNative"));
+}
// Depthwise Conv2D op where filter is a constant.
TEST_F(MklLayoutPassTest, DepthwiseConv2dNative_FilterCaching_Positive) {
InitGraph(
diff --git a/tensorflow/core/grappler/costs/BUILD b/tensorflow/core/grappler/costs/BUILD
index 443b191..1c7493c 100644
--- a/tensorflow/core/grappler/costs/BUILD
+++ b/tensorflow/core/grappler/costs/BUILD
@@ -270,6 +270,7 @@
"//tensorflow/core/grappler:op_types",
"//tensorflow/core/grappler:utils",
"//tensorflow/core/grappler/clusters:utils",
+ "//tensorflow/core/grappler/utils:transitive_fanin",
"@com_google_absl//absl/strings",
"@com_google_absl//absl/strings:str_format",
],
diff --git a/tensorflow/core/grappler/costs/virtual_scheduler.cc b/tensorflow/core/grappler/costs/virtual_scheduler.cc
index 7cb61b7..ffdec02 100644
--- a/tensorflow/core/grappler/costs/virtual_scheduler.cc
+++ b/tensorflow/core/grappler/costs/virtual_scheduler.cc
@@ -27,6 +27,7 @@
#include "tensorflow/core/grappler/costs/utils.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/utils.h"
+#include "tensorflow/core/grappler/utils/transitive_fanin.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/strings/numbers.h"
#include "tensorflow/core/platform/logging.h"
@@ -405,14 +406,10 @@
}
// Get the nodes that would run to output fetch_nodes.
- bool ill_formed = false;
std::unordered_map<string, const NodeDef*> name_to_node;
- const std::vector<const NodeDef*> fetch_fanin_nodes =
- ComputeTransitiveFanin(graph, fetch_nodes, &name_to_node, &ill_formed);
- if (ill_formed) {
- return errors::InvalidArgument(
- "Ill formed graph or invalid set of fetch nodes specified");
- }
+ std::vector<const NodeDef*> fetch_fanin_nodes;
+ TF_RETURN_IF_ERROR(ComputeTransitiveFanin(graph, fetch_nodes, &name_to_node,
+ &fetch_fanin_nodes));
// Once ComputeTransitiveFanin is complete, only the nodes that can be reached
// from the fetch nodes are scheduled. So the scheduled nodes should be
diff --git a/tensorflow/core/grappler/graph_analyzer/BUILD b/tensorflow/core/grappler/graph_analyzer/BUILD
index da27c10..d252a95 100644
--- a/tensorflow/core/grappler/graph_analyzer/BUILD
+++ b/tensorflow/core/grappler/graph_analyzer/BUILD
@@ -44,7 +44,7 @@
"//tensorflow/core:lib",
"//tensorflow/core:protos_all_cc",
"//tensorflow/core:tensorflow",
- "//tensorflow/core/grappler:grappler_item",
+ "//tensorflow/core/grappler/utils:transitive_fanin",
"@com_google_absl//absl/strings",
],
)
diff --git a/tensorflow/core/grappler/graph_analyzer/graph_analyzer_tool.cc b/tensorflow/core/grappler/graph_analyzer/graph_analyzer_tool.cc
index 924ca11..f5fd66f 100644
--- a/tensorflow/core/grappler/graph_analyzer/graph_analyzer_tool.cc
+++ b/tensorflow/core/grappler/graph_analyzer/graph_analyzer_tool.cc
@@ -16,7 +16,7 @@
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/grappler/graph_analyzer/graph_analyzer.h"
-#include "tensorflow/core/grappler/grappler_item.h"
+#include "tensorflow/core/grappler/utils/transitive_fanin.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/init_main.h"
@@ -54,9 +54,9 @@
if (fetch_nodes.empty()) {
*graph = metagraph.graph_def();
} else {
- std::vector<const tensorflow::NodeDef*> fanin_nodes =
- tensorflow::grappler::ComputeTransitiveFanin(metagraph.graph_def(),
- fetch_nodes);
+ std::vector<const tensorflow::NodeDef*> fanin_nodes;
+ TF_CHECK_OK(tensorflow::grappler::ComputeTransitiveFanin(
+ metagraph.graph_def(), fetch_nodes, &fanin_nodes));
for (const tensorflow::NodeDef* node : fanin_nodes) {
*(graph->add_node()) = *node;
}
diff --git a/tensorflow/core/grappler/grappler_item.cc b/tensorflow/core/grappler/grappler_item.cc
index 9c3fef4..4b58456 100644
--- a/tensorflow/core/grappler/grappler_item.cc
+++ b/tensorflow/core/grappler/grappler_item.cc
@@ -50,7 +50,9 @@
}
std::vector<const NodeDef*> GrapplerItem::MainOpsFanin() const {
- return ComputeTransitiveFanin(graph, fetch);
+ std::vector<const NodeDef*> fanin_nodes;
+ TF_CHECK_OK(ComputeTransitiveFanin(graph, fetch, &fanin_nodes));
+ return fanin_nodes;
}
std::vector<const NodeDef*> GrapplerItem::EnqueueOpsFanin() const {
@@ -60,15 +62,20 @@
enqueue_ops.push_back(enqueue_op);
}
}
- return ComputeTransitiveFanin(graph, enqueue_ops);
+ std::vector<const NodeDef*> fanin_nodes;
+ TF_CHECK_OK(ComputeTransitiveFanin(graph, fetch, &fanin_nodes));
+ return fanin_nodes;
}
std::vector<const NodeDef*> GrapplerItem::InitOpsFanin() const {
- return ComputeTransitiveFanin(graph, init_ops);
+ std::vector<const NodeDef*> fanin_nodes;
+ TF_CHECK_OK(ComputeTransitiveFanin(graph, init_ops, &fanin_nodes));
+ return fanin_nodes;
}
std::vector<const NodeDef*> GrapplerItem::MainVariables() const {
- std::vector<const NodeDef*> fanin = ComputeTransitiveFanin(graph, init_ops);
+ std::vector<const NodeDef*> fanin;
+ TF_CHECK_OK(ComputeTransitiveFanin(graph, init_ops, &fanin));
std::vector<const NodeDef*> vars;
for (const NodeDef* node : fanin) {
if (IsVariable(*node)) {
@@ -200,22 +207,5 @@
return optimization_options_;
}
-std::vector<const NodeDef*> ComputeTransitiveFanin(
- const GraphDef& graph, const std::vector<string>& terminal_nodes) {
- bool ill_formed = false;
- std::vector<const NodeDef*> result =
- ComputeTransitiveFanin(graph, terminal_nodes, &ill_formed);
- CHECK(!ill_formed);
- return result;
-}
-
-std::vector<const NodeDef*> ComputeTransitiveFanin(
- const GraphDef& graph, const std::vector<string>& terminal_nodes,
- bool* ill_formed) {
- std::unordered_map<string, const NodeDef*> name_to_fanin_node;
- return ComputeTransitiveFanin(graph, terminal_nodes, &name_to_fanin_node,
- ill_formed);
-}
-
} // end namespace grappler
} // end namespace tensorflow
diff --git a/tensorflow/core/grappler/grappler_item.h b/tensorflow/core/grappler/grappler_item.h
index ed97eec..99d6d2c4 100644
--- a/tensorflow/core/grappler/grappler_item.h
+++ b/tensorflow/core/grappler/grappler_item.h
@@ -133,25 +133,6 @@
OptimizationOptions optimization_options_;
};
-// Return the transitive fanin of a set of terminal nodes.
-std::vector<const NodeDef*> ComputeTransitiveFanin(
- const GraphDef& graph, const std::vector<string>& terminal_nodes);
-
-// Return the transitive fanin of a set of terminal nodes. Sets 'ill_formed' to
-// true if one of the node is missing in the graph, or some node inputs don't
-// exist.
-std::vector<const NodeDef*> ComputeTransitiveFanin(
- const GraphDef& graph, const std::vector<string>& terminal_nodes,
- bool* ill_formed);
-
-// Return the transitive fanin of a set of terminal nodes. Sets 'ill_formed' to
-// true if one of the node is missing in the graph, or some node inputs don't
-// exist. Sets name_to_fanin_node for name to fanin nodes map.
-std::vector<const NodeDef*> ComputeTransitiveFanin(
- const GraphDef& graph, const std::vector<string>& terminal_nodes,
- std::unordered_map<string, const NodeDef*>* name_to_fanin_node,
- bool* ill_formed);
-
} // end namespace grappler
} // end namespace tensorflow
diff --git a/tensorflow/core/grappler/optimizers/BUILD b/tensorflow/core/grappler/optimizers/BUILD
index ae854ad..0b8846f 100644
--- a/tensorflow/core/grappler/optimizers/BUILD
+++ b/tensorflow/core/grappler/optimizers/BUILD
@@ -75,6 +75,7 @@
"//tensorflow/core/grappler:op_types",
"//tensorflow/core/grappler:utils",
"//tensorflow/core/grappler/clusters:cluster",
+ "//tensorflow/core/grappler/utils:transitive_fanin",
],
)
diff --git a/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc b/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc
index 55f83eb..fa6ca31 100644
--- a/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc
+++ b/tensorflow/core/grappler/optimizers/auto_mixed_precision.cc
@@ -58,7 +58,7 @@
// node. It handles regular type attributes, list type attributes (where
// type_index is set to the index in the type list), and fixed types.
struct TypeAttrId {
- static const int kSingleType = -1;
+ static constexpr int kSingleType = -1;
explicit TypeAttrId(const string& _attr_name, int _type_index = kSingleType)
: attr_name(_attr_name),
diff --git a/tensorflow/core/grappler/optimizers/auto_parallel.cc b/tensorflow/core/grappler/optimizers/auto_parallel.cc
index 3f58a2a..a537fa2 100644
--- a/tensorflow/core/grappler/optimizers/auto_parallel.cc
+++ b/tensorflow/core/grappler/optimizers/auto_parallel.cc
@@ -25,6 +25,7 @@
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/utils.h"
+#include "tensorflow/core/grappler/utils/transitive_fanin.h"
#include "tensorflow/core/lib/strings/strcat.h"
namespace tensorflow {
@@ -147,7 +148,8 @@
}
LOG(INFO) << "Graph size after adding div nodes: " << all_nodes_.size();
- auto train_nodes = ComputeTransitiveFanin(graph_, item.fetch);
+ std::vector<const NodeDef*> train_nodes;
+ TF_RETURN_IF_ERROR(ComputeTransitiveFanin(graph_, item.fetch, &train_nodes));
LOG(INFO) << "Number of training nodes: " << train_nodes.size();
const NodeDef* dequeue_node;
@@ -161,7 +163,8 @@
std::vector<const NodeDef*> input_nodes;
if (dequeue_node) {
LOG(INFO) << "Dequeue node: " << dequeue_node->name();
- input_nodes = ComputeTransitiveFanin(graph_, {dequeue_node->name()});
+ TF_RETURN_IF_ERROR(ComputeTransitiveFanin(graph_, {dequeue_node->name()},
+ {}, &input_nodes));
}
LOG(INFO) << "Number of input nodes: " << input_nodes.size();
diff --git a/tensorflow/core/grappler/optimizers/constant_folding_test.cc b/tensorflow/core/grappler/optimizers/constant_folding_test.cc
index 6c829bb..7fae405 100644
--- a/tensorflow/core/grappler/optimizers/constant_folding_test.cc
+++ b/tensorflow/core/grappler/optimizers/constant_folding_test.cc
@@ -548,6 +548,8 @@
}
}
+// This test fails on ROCm platform (see commit message for details)
+#ifndef TENSORFLOW_USE_ROCM
TEST_F(ConstantFoldingTest, MulConvPushDownTest_Conv2D_ScalarConst) {
for (string data_format : {
"NHWC",
@@ -565,7 +567,10 @@
/*expect_folded=*/true);
}
}
+#endif
+// This test fails on ROCm platform (see commit message for details)
+#ifndef TENSORFLOW_USE_ROCM
TEST_F(ConstantFoldingTest, MulConvPushDownTest_Conv2D_SingletonConst) {
for (string data_format : {
"NHWC",
@@ -585,6 +590,7 @@
}
}
}
+#endif
TEST_F(ConstantFoldingTest,
MulConvPushDownTest_Conv2D_SingletonConst_ShapeMismatch) {
@@ -668,8 +674,7 @@
}
}
-// This test fails on ROCm platform with two vaue miscompare
-// TODO(rocm) : analysze and fix the cause of the failure and re-enable test
+// This test fails on ROCm platform (see commit message for details)
#ifndef TENSORFLOW_USE_ROCM
TEST_F(ConstantFoldingTest, MulConvPushDownTest_Conv3D_NDHWC_1x1x3Const) {
MulConvPushDownTest(
diff --git a/tensorflow/core/grappler/optimizers/data/BUILD b/tensorflow/core/grappler/optimizers/data/BUILD
index 0b88f2f..bab28d4 100644
--- a/tensorflow/core/grappler/optimizers/data/BUILD
+++ b/tensorflow/core/grappler/optimizers/data/BUILD
@@ -20,7 +20,6 @@
":inject_prefetch",
":latency_all_edges",
":make_sloppy",
- ":make_stateless",
":map_and_batch_fusion",
":map_and_filter_fusion",
":map_fusion",
@@ -43,6 +42,7 @@
":optimizer_base",
"@com_google_absl//absl/container:flat_hash_map",
"@com_google_absl//absl/container:flat_hash_set",
+ "@com_google_absl//absl/strings",
"//tensorflow/core/grappler:grappler_item",
"//tensorflow/core/grappler:mutable_graph_view",
"//tensorflow/core/grappler:op_types",
@@ -391,37 +391,6 @@
)
cc_library(
- name = "make_stateless",
- srcs = ["make_stateless.cc"],
- hdrs = ["make_stateless.h"],
- deps = [
- ":graph_utils",
- ":optimizer_base",
- "//tensorflow/core/grappler:grappler_item",
- "//tensorflow/core/grappler:mutable_graph_view",
- "//tensorflow/core/grappler:op_types",
- "//tensorflow/core/grappler/clusters:cluster",
- "//tensorflow/core/grappler/optimizers:custom_graph_optimizer_registry",
- ] + tf_protos_all(),
- alwayslink = 1,
-)
-
-tf_cc_test(
- name = "make_stateless_test",
- srcs = ["make_stateless_test.cc"],
- deps = [
- ":graph_test_utils",
- ":graph_utils",
- ":make_stateless",
- "//tensorflow/core:framework",
- "//tensorflow/core:test",
- "//tensorflow/core:test_main",
- "//tensorflow/core:testlib",
- "//tensorflow/core/grappler:grappler_item",
- ],
-)
-
-cc_library(
name = "map_and_batch_fusion",
srcs = ["map_and_batch_fusion.cc"],
hdrs = [
diff --git a/tensorflow/core/grappler/optimizers/data/auto_shard.cc b/tensorflow/core/grappler/optimizers/data/auto_shard.cc
index e3578b9..feabd7b 100644
--- a/tensorflow/core/grappler/optimizers/data/auto_shard.cc
+++ b/tensorflow/core/grappler/optimizers/data/auto_shard.cc
@@ -17,6 +17,7 @@
#include "absl/container/flat_hash_map.h"
#include "absl/container/flat_hash_set.h"
+#include "absl/strings/match.h"
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/function.pb.h"
@@ -38,10 +39,14 @@
constexpr char kShardDatasetOpName[] = "ShardDataset";
constexpr char kShuffleDatasetOpName[] = "ShuffleDataset";
constexpr char kShuffleDatasetV2OpName[] = "ShuffleDatasetV2";
+constexpr char kShuffleDatasetV3OpName[] = "ShuffleDatasetV3";
constexpr char kNumWorkersAttrName[] = "num_workers";
constexpr char kIndexAttrName[] = "index";
constexpr char kAutoShardPolicyAttrName[] = "auto_shard_policy";
+constexpr char kReshuffleEachIteration[] = "reshuffle_each_iteration";
+constexpr char kOutputShapes[] = "output_shapes";
+constexpr char kOutputTypes[] = "output_types";
constexpr std::array<const char*, 6> kReaderDatasetOps = {
"FixedLengthRecordDataset",
@@ -57,7 +62,7 @@
"ZipDataset"
};
-constexpr std::array<const char*, 30> kPassThroughOps = {
+constexpr std::array<const char*, 31> kPassThroughOps = {
"_Retval",
"AssertCardinalityDataset",
"AssertNextDataset",
@@ -85,6 +90,7 @@
"ShuffleAndRepeatDataset",
"ShuffleDataset",
"ShuffleDatasetV2",
+ "ShuffleDatasetV3",
"SkipDataset",
"TakeDataset",
"WindowDataset",
@@ -146,28 +152,27 @@
// Add shapes and other attributes
NodeDef* add_after = graph->GetNode(add_before.input(0));
- if (absl::EndsWith(add_after->op(), "Dataset") ||
- absl::EndsWith(add_after->op(), "DatasetV2")) {
+ if (absl::StrContains(add_after->op(), "Dataset")) {
// We still may or may not have the right attributes because Datasets like
// TFRecordDataset doesn't have a output type or shape, and by default we
// set them to DT_STRING and an unknown shape.
- if (add_after->attr().count("output_shapes") > 0) {
- graph_utils::CopyAttribute("output_shapes", *add_after, &new_node);
+ if (add_after->attr().count(kOutputShapes) > 0) {
+ graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
} else {
tensorflow::TensorShapeProto* shape =
- (*(new_node.mutable_attr()))["output_shapes"]
+ (*(new_node.mutable_attr()))[kOutputShapes]
.mutable_list()
->add_shape();
shape->set_unknown_rank(true);
}
- if (add_after->attr().count("output_types") > 0) {
- graph_utils::CopyAttribute("output_types", *add_after, &new_node);
+ if (add_after->attr().count(kOutputTypes) > 0) {
+ graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);
} else if (add_after->attr().count("Toutput_types") > 0) {
- (*(new_node.mutable_attr()))["output_types"] =
+ (*(new_node.mutable_attr()))[kOutputTypes] =
add_after->attr().at("Toutput_types");
} else {
- (*(new_node.mutable_attr()))["output_types"].mutable_list()->add_type(
+ (*(new_node.mutable_attr()))[kOutputTypes].mutable_list()->add_type(
tensorflow::DataType::DT_STRING);
}
} else {
@@ -189,9 +194,10 @@
return Status::OK();
}
-Status AddShuffleNode(MutableGraphView* graph, const NodeDef& add_before,
- const string& buffer_size_node, const string& seed_node,
- const string& seed2_node, bool reshuffle_each_iteration) {
+Status AddShuffleDataset(MutableGraphView* graph, const NodeDef& add_before,
+ const string& buffer_size_node,
+ const string& seed_node, const string& seed2_node,
+ bool reshuffle_each_iteration) {
NodeDef* add_after = graph->GetNode(add_before.input(0));
NodeDef new_node;
new_node.set_op(kShuffleDatasetOpName);
@@ -203,12 +209,12 @@
new_node.add_input(seed_node);
new_node.add_input(seed2_node);
- graph_utils::CopyAttribute("output_shapes", *add_after, &new_node);
- graph_utils::CopyAttribute("output_types", *add_after, &new_node);
+ graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
+ graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);
AttrValue reshuffle_attr;
reshuffle_attr.set_b(reshuffle_each_iteration);
- (*new_node.mutable_attr())["reshuffle_each_iteration"] = reshuffle_attr;
+ (*new_node.mutable_attr())[kReshuffleEachIteration] = reshuffle_attr;
NodeDef* new_node_graph = graph->AddNode(std::move(new_node));
@@ -217,9 +223,9 @@
return Status::OK();
}
-Status AddShuffleV2Node(MutableGraphView* graph, const NodeDef& add_before,
- const string& buffer_size_node,
- const string& seed_generator_node) {
+Status AddShuffleDatasetV2(MutableGraphView* graph, const NodeDef& add_before,
+ const string& buffer_size_node,
+ const string& seed_generator_node) {
NodeDef* add_after = graph->GetNode(add_before.input(0));
NodeDef new_node;
new_node.set_op(kShuffleDatasetV2OpName);
@@ -230,8 +236,39 @@
new_node.add_input(buffer_size_node);
new_node.add_input(seed_generator_node);
- graph_utils::CopyAttribute("output_shapes", *add_after, &new_node);
- graph_utils::CopyAttribute("output_types", *add_after, &new_node);
+ graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
+ graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);
+
+ NodeDef* new_node_graph = graph->AddNode(std::move(new_node));
+
+ TF_RETURN_IF_ERROR(
+ graph->UpdateFanouts(add_after->name(), new_node_graph->name()));
+ return Status::OK();
+}
+
+Status AddShuffleDatasetV3(MutableGraphView* graph, const NodeDef& add_before,
+ const string& buffer_size_node,
+ const string& seed_node, const string& seed2_node,
+ const string& seed_generator_node,
+ bool reshuffle_each_iteration) {
+ NodeDef* add_after = graph->GetNode(add_before.input(0));
+ NodeDef new_node;
+ new_node.set_op(kShuffleDatasetV3OpName);
+ graph_utils::SetUniqueGraphNodeName(kShuffleDatasetV3OpName, graph->graph(),
+ &new_node);
+
+ new_node.add_input(add_before.input(0));
+ new_node.add_input(buffer_size_node);
+ new_node.add_input(seed_node);
+ new_node.add_input(seed2_node);
+ new_node.add_input(seed_generator_node);
+
+ graph_utils::CopyAttribute(kOutputShapes, *add_after, &new_node);
+ graph_utils::CopyAttribute(kOutputTypes, *add_after, &new_node);
+
+ AttrValue reshuffle_attr;
+ reshuffle_attr.set_b(reshuffle_each_iteration);
+ (*new_node.mutable_attr())[kReshuffleEachIteration] = reshuffle_attr;
NodeDef* new_node_graph = graph->AddNode(std::move(new_node));
@@ -268,7 +305,7 @@
*buffer_size_node = node.input(1);
*seed_node = node.input(2);
*seed2_node = node.input(3);
- *reshuffle_each_iteration = node.attr().at("reshuffle_each_iteration").b();
+ *reshuffle_each_iteration = node.attr().at(kReshuffleEachIteration).b();
TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
nodes_to_delete->insert(node.name());
}
@@ -305,6 +342,33 @@
return Status::OK();
}
+Status RemoveShuffleDatasetV3(MutableGraphView* graph, const NodeDef& node,
+ absl::flat_hash_set<string>* nodes_to_delete,
+ string* op_name, string* buffer_size_node,
+ string* seed_node, string* seed2_node,
+ string* seed_generator_node,
+ bool* reshuffle_each_iteration) {
+ if (node.op() == kShuffleDatasetV3OpName) {
+ *op_name = node.op();
+ *buffer_size_node = node.input(1);
+ *seed_node = node.input(2);
+ *seed2_node = node.input(3);
+ *seed_generator_node = node.input(4);
+ *reshuffle_each_iteration = node.attr().at(kReshuffleEachIteration).b();
+ TF_RETURN_IF_ERROR(graph->UpdateFanouts(node.name(), node.input(0)));
+ nodes_to_delete->insert(node.name());
+ }
+
+ for (const auto& fanin : graph->GetFanins(node, true)) {
+ TF_RETURN_IF_ERROR(RemoveShuffleDatasetV3(
+ graph, *fanin.node, nodes_to_delete, op_name, buffer_size_node,
+ seed_node, seed2_node, seed_generator_node, reshuffle_each_iteration));
+ }
+
+ // TODO(frankchn): Traverse functions too.
+ return Status::OK();
+}
+
Status ProcessDatasetSourceNode(MutableGraphView* graph, const NodeDef& node,
absl::flat_hash_set<string>* nodes_to_delete,
int64 num_workers, int64 index) {
@@ -324,13 +388,24 @@
RemoveShuffleDatasetV2(graph, node, nodes_to_delete, &shuffle_op_name,
&buffer_size_node, &seed_generator_node));
}
+ if (shuffle_op_name.empty()) {
+ TF_RETURN_IF_ERROR(RemoveShuffleDatasetV3(
+ graph, node, nodes_to_delete, &shuffle_op_name, &buffer_size_node,
+ &seed_node, &seed2_node, &seed_generator_node,
+ &reshuffle_each_iteration));
+ }
if (shuffle_op_name == kShuffleDatasetOpName) {
- TF_RETURN_IF_ERROR(AddShuffleNode(graph, node, buffer_size_node, seed_node,
- seed2_node, reshuffle_each_iteration));
+ TF_RETURN_IF_ERROR(AddShuffleDataset(graph, node, buffer_size_node,
+ seed_node, seed2_node,
+ reshuffle_each_iteration));
} else if (shuffle_op_name == kShuffleDatasetV2OpName) {
- TF_RETURN_IF_ERROR(
- AddShuffleV2Node(graph, node, buffer_size_node, seed_generator_node));
+ TF_RETURN_IF_ERROR(AddShuffleDatasetV2(graph, node, buffer_size_node,
+ seed_generator_node));
+ } else if (shuffle_op_name == kShuffleDatasetV3OpName) {
+ TF_RETURN_IF_ERROR(AddShuffleDatasetV3(
+ graph, node, buffer_size_node, seed_node, seed2_node,
+ seed_generator_node, reshuffle_each_iteration));
}
return Status::OK();
diff --git a/tensorflow/core/grappler/optimizers/data/make_stateless.cc b/tensorflow/core/grappler/optimizers/data/make_stateless.cc
deleted file mode 100644
index a18ca58..0000000
--- a/tensorflow/core/grappler/optimizers/data/make_stateless.cc
+++ /dev/null
@@ -1,71 +0,0 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#include "tensorflow/core/grappler/optimizers/data/make_stateless.h"
-
-#include "tensorflow/core/framework/node_def.pb.h"
-#include "tensorflow/core/grappler/clusters/cluster.h"
-#include "tensorflow/core/grappler/grappler_item.h"
-#include "tensorflow/core/grappler/mutable_graph_view.h"
-#include "tensorflow/core/grappler/op_types.h"
-#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
-#include "tensorflow/core/grappler/optimizers/data/graph_utils.h"
-
-namespace tensorflow {
-namespace grappler {
-namespace {
-
-constexpr char kCacheDataset[] = "CacheDataset";
-constexpr char kCacheDatasetV2[] = "CacheDatasetV2";
-constexpr char kReshuffleEachIteration[] = "reshuffle_each_iteration";
-constexpr char kShuffleDataset[] = "ShuffleDataset";
-constexpr char kShuffleDatasetV2[] = "ShuffleDatasetV2";
-
-} // namespace
-
-Status MakeStateless::OptimizeAndCollectStats(Cluster* cluster,
- const GrapplerItem& item,
- GraphDef* output,
- OptimizationStats* stats) {
- *output = item.graph;
- MutableGraphView graph(output);
-
- NodeDef* zero_node = graph_utils::AddScalarConstNode<int64>(0, &graph);
-
- for (NodeDef& node : *output->mutable_node()) {
- if (node.op() == kShuffleDatasetV2) {
- *node.mutable_op() = kShuffleDataset;
- // remove `seed_generator` input
- node.mutable_input()->RemoveLast();
- // add `seed` input
- node.add_input(zero_node->name());
- // add `seed2` input
- node.add_input(zero_node->name());
- // set `reshuffle_each_iteration` attr
- (*node.mutable_attr())[kReshuffleEachIteration].set_b(true);
- } else if (node.op() == kCacheDatasetV2) {
- *node.mutable_op() = kCacheDataset;
- // remove `cache` input
- node.mutable_input()->RemoveLast();
- }
- }
-
- return Status::OK();
-}
-
-REGISTER_GRAPH_OPTIMIZER_AS(MakeStateless, "make_stateless");
-
-} // namespace grappler
-} // namespace tensorflow
diff --git a/tensorflow/core/grappler/optimizers/data/make_stateless.h b/tensorflow/core/grappler/optimizers/data/make_stateless.h
deleted file mode 100644
index cd95c23..0000000
--- a/tensorflow/core/grappler/optimizers/data/make_stateless.h
+++ /dev/null
@@ -1,56 +0,0 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_MAKE_STATELESS_H_
-#define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_MAKE_STATELESS_H_
-
-#include "tensorflow/core/grappler/optimizers/data/optimizer_base.h"
-
-namespace tensorflow {
-namespace grappler {
-
-// This rewrite replaces transformations that depend on external state (such as
-// `ShuffleDatasetV2`) with a stateless alternative so that the input pipeline
-// graph can be cloned.
-//
-// Note that this rewrites may change observable behavior of the input pipeline
-// (e.g. `reshuffle_each_iteration` will not work) and is a stop gap solution
-// to enable cloning until a better mechanism exists.
-class MakeStateless : public TFDataOptimizerBase {
- public:
- MakeStateless() = default;
- ~MakeStateless() override = default;
-
- string name() const override { return "make_stateless"; }
-
- bool UsesFunctionLibrary() const override { return false; }
-
- Status Init(
- const tensorflow::RewriterConfig_CustomGraphOptimizer* config) override {
- return Status::OK();
- }
-
- Status OptimizeAndCollectStats(Cluster* cluster, const GrapplerItem& item,
- GraphDef* output,
- OptimizationStats* stats) override;
-
- void Feedback(Cluster* cluster, const GrapplerItem& item,
- const GraphDef& optimize_output, double result) override {}
-};
-
-} // namespace grappler
-} // namespace tensorflow
-
-#endif // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_MAKE_STATELESS_H_
diff --git a/tensorflow/core/grappler/optimizers/data/make_stateless_test.cc b/tensorflow/core/grappler/optimizers/data/make_stateless_test.cc
deleted file mode 100644
index a30b7c6..0000000
--- a/tensorflow/core/grappler/optimizers/data/make_stateless_test.cc
+++ /dev/null
@@ -1,79 +0,0 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#include "tensorflow/core/grappler/optimizers/data/make_stateless.h"
-
-#include "tensorflow/core/framework/attr_value_util.h"
-#include "tensorflow/core/framework/function_testlib.h"
-#include "tensorflow/core/framework/tensor_testutil.h"
-#include "tensorflow/core/grappler/grappler_item.h"
-#include "tensorflow/core/grappler/optimizers/data/graph_test_utils.h"
-#include "tensorflow/core/grappler/optimizers/data/graph_utils.h"
-#include "tensorflow/core/lib/core/status_test_util.h"
-#include "tensorflow/core/platform/test.h"
-
-namespace tensorflow {
-namespace grappler {
-namespace {
-
-TEST(MakeStateless, Cache) {
- using test::function::NDef;
- GrapplerItem item;
- item.graph = test::function::GDef(
- {NDef("start", "Const", {}, {{"value", 0}, {"dtype", DT_INT32}}),
- NDef("stop", "Const", {}, {{"value", 10}, {"dtype", DT_INT32}}),
- NDef("step", "Const", {}, {{"value", 1}, {"dtype", DT_INT32}}),
- NDef("range", "RangeDataset", {"start", "stop", "step"}, {}),
- NDef("filename", "Const", {}, {{"value", ""}, {"dtype", DT_INT64}}),
- NDef("handle", "Const", {}, {{"value", 1}, {"dtype", DT_RESOURCE}}),
- graph_tests_utils::MakeCacheV2Node("cache", "range", "filename",
- "handle")},
- {});
-
- MakeStateless optimizer;
- GraphDef output;
- TF_ASSERT_OK(optimizer.Optimize(nullptr, item, &output));
- EXPECT_TRUE(graph_utils::ContainsGraphNodeWithName("cache", output));
- int index = graph_utils::FindGraphNodeWithName("cache", output);
- EXPECT_EQ(output.node(index).op(), "CacheDataset");
- EXPECT_EQ(output.node(index).input_size(), 2);
-}
-
-TEST(MakeStateless, Shuffle) {
- using test::function::NDef;
- GrapplerItem item;
- item.graph = test::function::GDef(
- {NDef("start", "Const", {}, {{"value", 0}, {"dtype", DT_INT32}}),
- NDef("stop", "Const", {}, {{"value", 10}, {"dtype", DT_INT32}}),
- NDef("step", "Const", {}, {{"value", 1}, {"dtype", DT_INT32}}),
- NDef("range", "RangeDataset", {"start", "stop", "step"}, {}),
- NDef("buffer_size", "Const", {}, {{"value", 1}, {"dtype", DT_INT64}}),
- NDef("handle", "Const", {}, {{"value", 1}, {"dtype", DT_RESOURCE}}),
- graph_tests_utils::MakeShuffleV2Node("shuffle", "range", "buffer_size",
- "handle")},
- {});
-
- MakeStateless optimizer;
- GraphDef output;
- TF_ASSERT_OK(optimizer.Optimize(nullptr, item, &output));
- EXPECT_TRUE(graph_utils::ContainsGraphNodeWithName("shuffle", output));
- int index = graph_utils::FindGraphNodeWithName("shuffle", output);
- EXPECT_EQ(output.node(index).op(), "ShuffleDataset");
- EXPECT_EQ(output.node(index).input_size(), 4);
-}
-
-} // namespace
-} // namespace grappler
-} // namespace tensorflow
diff --git a/tensorflow/core/grappler/optimizers/data/meta_optimizer.cc b/tensorflow/core/grappler/optimizers/data/meta_optimizer.cc
index b8b630c..3591cd5 100644
--- a/tensorflow/core/grappler/optimizers/data/meta_optimizer.cc
+++ b/tensorflow/core/grappler/optimizers/data/meta_optimizer.cc
@@ -43,8 +43,7 @@
std::map<string, tensorflow::RewriterConfig_CustomGraphOptimizer>;
// tf.data optimizations, in the order we want to perform them.
-constexpr std::array<const char*, 16> kTFDataOptimizations = {
- "make_stateless",
+constexpr std::array<const char*, 15> kTFDataOptimizations = {
"noop_elimination",
"shuffle_and_repeat_fusion",
"map_fusion",
diff --git a/tensorflow/core/grappler/optimizers/data/slack.cc b/tensorflow/core/grappler/optimizers/data/slack.cc
index 6d1aab0..27915e2 100644
--- a/tensorflow/core/grappler/optimizers/data/slack.cc
+++ b/tensorflow/core/grappler/optimizers/data/slack.cc
@@ -51,7 +51,7 @@
constexpr std::array<const char*, 2> kMultipleInputsDatasetOps = {
"ZipDataset", "ConcatenateDataset"};
-constexpr std::array<const char*, 21> kPassThroughOps = {
+constexpr std::array<const char*, 22> kPassThroughOps = {
"CacheDataset",
"CacheDatasetV2",
"ExperimentalMaxIntraOpParallelismDataset",
@@ -70,6 +70,7 @@
"ShuffleAndRepeatDataset",
"ShuffleDataset",
"ShuffleDatasetV2",
+ "ShuffleDatasetV3",
"SkipDataset",
"TakeDataset",
"WindowDataset",
diff --git a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_test.cc b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_test.cc
index 4ceb026..c85d85e 100644
--- a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_test.cc
+++ b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_test.cc
@@ -77,17 +77,21 @@
Output SimpleConv2DBackpropInput(tensorflow::Scope* s, int input_size,
int filter_size, const string& padding,
- bool dilated) {
+ bool dilated, const int input_sizes_length) {
int batch_size = 128;
int input_height = input_size;
int input_width = input_size;
int input_depth = 3;
int filter_count = 2;
int stride = 1;
- TensorShape input_sizes_shape({4});
+ TensorShape input_sizes_shape({input_sizes_length});
Tensor input_data(DT_INT32, input_sizes_shape);
- test::FillValues<int>(&input_data,
- {batch_size, input_height, input_width, input_depth});
+ if (input_sizes_length == 4) {
+ test::FillValues<int>(&input_data,
+ {batch_size, input_height, input_width, input_depth});
+ } else {
+ test::FillValues<int>(&input_data, {input_height, input_width});
+ }
Output input_sizes =
ops::Const(s->WithOpName("InputSizes"), Input::Initializer(input_data));
@@ -353,7 +357,8 @@
GTEST_SKIP() << "Neither CUDA nor ROCm is enabled";
#endif // !(GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
Scope s = Scope::NewRootScope();
- auto conv = SimpleConv2DBackpropInput(&s, 7, 2, "SAME", /*dilated=*/false);
+ auto conv = SimpleConv2DBackpropInput(&s, 7, 2, "SAME", /*dilated=*/false,
+ /*input_sizes_length=*/4);
Output fetch = ops::Identity(s.WithOpName("Fetch"), {conv});
GrapplerItem item;
TF_ASSERT_OK(s.ToGraphDef(&item.graph));
@@ -380,6 +385,30 @@
VerifyRegularFaninMatch(input_sizes_node, 0, "InputSizesIdentity", 0);
}
+TEST_F(GenericLayoutOptimizerTest, Conv2DBackpropInput2DInputSizes) {
+#if !(GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
+ GTEST_SKIP() << "Neither CUDA nor ROCm is enabled";
+#endif // !(GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
+ Scope s = Scope::NewRootScope();
+ auto conv = SimpleConv2DBackpropInput(&s, 7, 2, "SAME", /*dilated=*/false,
+ /*input_sizes_length=*/2);
+ Output fetch = ops::Identity(s.WithOpName("Fetch"), {conv});
+ GrapplerItem item;
+ TF_ASSERT_OK(s.ToGraphDef(&item.graph));
+
+ GenericLayoutOptimizer optimizer;
+ GraphDef output;
+ TF_ASSERT_OK(optimizer.Optimize(virtual_cluster_.get(), item, &output));
+
+ Status status;
+ utils::GraphView graph_view(&output, &status);
+ TF_ASSERT_OK(status);
+ auto* conv2d_backprop_node = graph_view.GetNode("Conv2DBackpropInput");
+ ASSERT_NE(conv2d_backprop_node, nullptr);
+ ASSERT_EQ(conv2d_backprop_node->NumRegularFanins(), 3);
+ VerifyRegularFaninMatch(conv2d_backprop_node, 0, "InputSizesIdentity", 0);
+}
+
TEST_F(GenericLayoutOptimizerTest, Conv2DDataFormatVecPermuteCollapse) {
#if !(GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
GTEST_SKIP() << "Neither CUDA nor ROCm is enabled";
diff --git a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc
index ca28faf..e9691a1 100644
--- a/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc
+++ b/tensorflow/core/grappler/optimizers/generic_layout_optimizer_transposer.cc
@@ -725,12 +725,42 @@
if (!ShouldProcess(*context, *node) || !IsFanoutPortRankN(*node, 0, 4)) {
return Status::OK();
}
+
+ const auto& fanin = node->GetRegularFanin(0);
+ auto* fanin_node = fanin.node_view();
+ const auto* output_shape_attr = fanin_node->GetAttr(kAttrOutputShape);
+ if (output_shape_attr == nullptr) {
+ VLOG(3) << "Cannot compute the shape of " << fanin_node->GetName()
+ << " because it is missing attribute " << kAttrOutputShape;
+ return Status::OK();
+ }
+ TensorShapeProto fanin_shape = output_shape_attr->list().shape(fanin.index());
+ if (fanin_shape.dim_size() != 1) {
+ VLOG(3) << fanin_node->GetName() << " is not a vector.";
+ return Status::OK();
+ }
+ int vector_size = fanin_shape.dim(0).size();
+ if (vector_size == -1) {
+ VLOG(3) << "The number of elements in " << fanin_node->GetName()
+ << " is unknown.";
+ return Status::OK();
+ }
+ if (vector_size != 2 && vector_size != 4) {
+ return errors::InvalidArgument(
+ fanin_node->GetName(), " must be a vector of size 2 or 4, but found ",
+ vector_size);
+ }
+
VLOG(3) << "GenericLayoutOptimizer: transforming node '" << node->GetName()
<< "' with op '" << node->GetOp() << "' from data format '"
<< context->src_format << "' to '" << context->dst_format << "'";
TF_RETURN_IF_ERROR(UpdateNode(context, node));
- TF_RETURN_IF_ERROR(
- UpdateFaninEdgesWithOp(context, {0}, node, kOpDataFormatVecPermute));
+ // Do not permute a input_sizes of size 2 because it represents HW regardless
+ // of whether NCHW or NHWC.
+ if (vector_size != 2) {
+ TF_RETURN_IF_ERROR(
+ UpdateFaninEdgesWithOp(context, {0}, node, kOpDataFormatVecPermute));
+ }
TF_RETURN_IF_ERROR(UpdateFaninEdgesWithOp(context, {2}, node, kOpTranspose));
TF_RETURN_IF_ERROR(UpdateFanoutEdgesWithOp(context, {0}, node, kOpTranspose));
return context->graph_view->GetMutationBuilder()->Apply();
diff --git a/tensorflow/core/grappler/optimizers/mkl_remapper_test.cc b/tensorflow/core/grappler/optimizers/mkl_remapper_test.cc
index 120f252..7a6b490 100644
--- a/tensorflow/core/grappler/optimizers/mkl_remapper_test.cc
+++ b/tensorflow/core/grappler/optimizers/mkl_remapper_test.cc
@@ -206,6 +206,94 @@
#undef CREATE_CONV2DFUSION_ADD_BCAST_TEST
#undef CREATE_CONV2DFUSION_TEST
+TEST_F(MklRemapperTest, FuseDepthwiseConv2DWithBiasAndActivation) {
+ using ::tensorflow::ops::Placeholder;
+
+ for (const string& activation : {"Relu", "Relu6", "Elu", "None"}) {
+ tensorflow::Scope s = tensorflow::Scope::NewRootScope();
+
+ auto input_shape = Placeholder::Shape({8, 32, 32, 3});
+ auto filter_shape = Placeholder::Shape({1, 1, 3, 1});
+ auto bias_shape = Placeholder::Shape({3});
+
+ auto input = Placeholder(s.WithOpName("input"), DT_FLOAT, input_shape);
+ auto filter = Placeholder(s.WithOpName("filter"), DT_FLOAT, filter_shape);
+ auto bias = Placeholder(s.WithOpName("bias"), DT_FLOAT, bias_shape);
+
+ std::vector<int> strides = {1, 1, 1, 1};
+ auto conv = ops::DepthwiseConv2dNative(s.WithOpName("depthwise_conv"),
+ input, filter, strides, "SAME");
+ auto bias_add = ops::BiasAdd(s.WithOpName("bias_add"), conv, bias);
+
+ ops::Identity fetch = [&]() -> ops::Identity {
+ auto activate = s.WithOpName("activation");
+ auto fetch = s.WithOpName("fetch");
+
+ if (activation == "Relu") {
+ return ops::Identity(fetch, ops::Relu(activate, bias_add));
+ } else if (activation == "Relu6") {
+ return ops::Identity(fetch, ops::Relu6(activate, bias_add));
+ } else if (activation == "Elu") {
+ return ops::Identity(fetch, ops::Elu(activate, bias_add));
+ }
+
+ DCHECK(activation == "None");
+ return ops::Identity(fetch, bias_add);
+ }();
+
+ auto input_t = GenerateRandomTensor<DT_FLOAT>({8, 32, 32, 3});
+ auto filter_t = GenerateRandomTensor<DT_FLOAT>({1, 1, 3, 1});
+ auto bias_t = GenerateRandomTensor<DT_FLOAT>({3});
+
+ GrapplerItem item;
+ item.fetch = {"fetch"};
+ item.feed = {{"input", input_t}, {"filter", filter_t}, {"bias", bias_t}};
+ TF_CHECK_OK(s.ToGraphDef(&item.graph));
+
+ // Place all nodes on CPU.
+ for (int i = 0; i < item.graph.node_size(); ++i) {
+ item.graph.mutable_node(i)->set_device("/device:CPU:0");
+ }
+
+ Remapper optimizer(RewriterConfig::ON);
+ GraphDef output;
+ TF_CHECK_OK(optimizer.Optimize(nullptr, item, &output));
+
+ int found = 0;
+ for (const NodeDef& node : output.node()) {
+ if (node.name() != "bias_add" && node.name() != "activation") continue;
+
+ EXPECT_EQ(node.op(), "_FusedDepthwiseConv2dNative");
+ ASSERT_EQ(node.input_size(), 3);
+ EXPECT_EQ(node.input(0), "input");
+ EXPECT_EQ(node.input(1), "filter");
+
+ EXPECT_EQ(node.attr().at("num_args").i(), 1);
+ EXPECT_EQ(node.input(2), "bias");
+
+ const auto fused_ops = node.attr().at("fused_ops").list().s();
+ if (node.name() == "bias_add") {
+ ASSERT_EQ(fused_ops.size(), 1);
+ EXPECT_EQ(fused_ops[0], "BiasAdd");
+ found++;
+ }
+ if (node.name() == "activation") {
+ ASSERT_EQ(fused_ops.size(), 2);
+ EXPECT_EQ(fused_ops[0], "BiasAdd");
+ EXPECT_EQ(fused_ops[1], activation);
+ found++;
+ }
+ }
+ EXPECT_EQ(found, 1);
+
+ auto tensors_expected = EvaluateNodes(item.graph, item.fetch, item.feed);
+ ASSERT_EQ(tensors_expected.size(), 1);
+ auto tensors = EvaluateNodes(output, item.fetch, item.feed);
+ ASSERT_EQ(tensors.size(), 1);
+ test::ExpectTensorNear<float>(tensors[0], tensors_expected[0], 1e-6);
+ }
+}
+
} // namespace grappler
} // namespace tensorflow
#endif // INTEL_MKL
diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc
index 4386935..c9cce4d 100644
--- a/tensorflow/core/grappler/optimizers/remapper.cc
+++ b/tensorflow/core/grappler/optimizers/remapper.cc
@@ -57,6 +57,7 @@
constexpr char kFusedConv2D[] = "_FusedConv2D";
constexpr char kFusedMatMul[] = "_FusedMatMul";
+constexpr char kFusedDepthwiseConv2dNative[] = "_FusedDepthwiseConv2dNative";
constexpr char kFusedBatchNormEx[] = "_FusedBatchNormEx";
constexpr char kDataFormat[] = "data_format";
@@ -279,12 +280,24 @@
return NodeIsOnCpu(matmul) && IsCpuCompatibleDataType(matmul);
}
+bool IsCpuCompatibleDepthwiseConv2dNative(const NodeDef* dw_conv2d) {
+ DCHECK(IsDepthwiseConv2dNative(*dw_conv2d))
+ << "Expected DepthwiseConv2dNative op";
+ return NodeIsOnCpu(dw_conv2d) && IsCpuCompatibleDataType(dw_conv2d);
+}
+
// Checks if we can rewrite a pattern to the `_Fused{Conv2D,MatMul}` on CPU.
template <typename Pattern>
bool IsCpuCompatible(const RemapperContext& ctx, const Pattern& matched) {
const NodeDef& node = ctx.graph_view.graph()->node(matched.contraction);
if (IsConv2D(node)) {
return IsCpuCompatibleConv2D(&node);
+ } else if (IsDepthwiseConv2dNative(node)) {
+#ifdef INTEL_MKL
+ return IsCpuCompatibleDepthwiseConv2dNative(&node);
+#else
+ return false;
+#endif // INTEL_MKL
} else if (IsMatMul(node)) {
return IsCpuCompatibleMatMul(&node);
} else {
@@ -381,11 +394,12 @@
const auto* contraction_node_view = regular_fanin_0.node_view();
const auto* contraction_node_def = contraction_node_view->node();
- bool is_conv2d_or_matmul =
- IsConv2D(*contraction_node_def) || IsMatMul(*contraction_node_def);
+ // Conv2D, MatMul or DepthwiseConv2D
+ bool is_contraction = IsConv2D(*contraction_node_def) ||
+ IsMatMul(*contraction_node_def) ||
+ IsDepthwiseConv2dNative(*contraction_node_def);
- if (!is_conv2d_or_matmul ||
- !HaveSameDataType(node_def, contraction_node_def) ||
+ if (!is_contraction || !HaveSameDataType(node_def, contraction_node_def) ||
HasControlFaninOrFanout(*contraction_node_view) ||
!HasAtMostOneFanoutAtPort0(*contraction_node_view) ||
IsInPreserveSet(ctx, contraction_node_def))
@@ -902,6 +916,21 @@
(*attr)["use_cudnn_on_gpu"] = src_attr.at("use_cudnn_on_gpu");
}
+void CopyDepthwiseConv2dNativeAttributes(const NodeDef& dw_conv2d,
+ NodeDef* fused_dw_conv2d) {
+ DCHECK(IsDepthwiseConv2dNative(dw_conv2d))
+ << "Input node must be a DepthwiseConv2dNative";
+
+ auto* attr = fused_dw_conv2d->mutable_attr();
+ auto& src_attr = dw_conv2d.attr();
+
+ (*attr)["T"] = src_attr.at("T");
+ (*attr)["strides"] = src_attr.at("strides");
+ (*attr)["padding"] = src_attr.at("padding");
+ (*attr)["dilations"] = src_attr.at("dilations");
+ (*attr)["data_format"] = src_attr.at("data_format");
+}
+
void CopyFusedBatchNormAttributes(const NodeDef& fused_batch_norm,
NodeDef* fused_batch_norm_ex) {
DCHECK(IsFusedBatchNorm(fused_batch_norm))
@@ -966,6 +995,9 @@
if (IsConv2D(contraction)) {
fused_op.set_op(kFusedConv2D);
CopyConv2DAttributes(contraction, &fused_op);
+ } else if (IsDepthwiseConv2dNative(contraction)) {
+ fused_op.set_op(kFusedDepthwiseConv2dNative);
+ CopyDepthwiseConv2dNativeAttributes(contraction, &fused_op);
} else if (IsMatMul(contraction)) {
fused_op.set_op(kFusedMatMul);
CopyMatMulAttributes(contraction, &fused_op);
@@ -1010,6 +1042,9 @@
if (IsConv2D(contraction)) {
fused_op.set_op(kFusedConv2D);
CopyConv2DAttributes(contraction, &fused_op);
+ } else if (IsDepthwiseConv2dNative(contraction)) {
+ fused_op.set_op(kFusedDepthwiseConv2dNative);
+ CopyDepthwiseConv2dNativeAttributes(contraction, &fused_op);
} else if (IsMatMul(contraction)) {
fused_op.set_op(kFusedMatMul);
CopyMatMulAttributes(contraction, &fused_op);
@@ -1660,7 +1695,8 @@
}
#endif //! INTEL_MKL
- // Remap {Conv2D,MatMul}+BiasAdd into the _Fused{Conv2D,MatMul}
+ // Remap {Conv2D,DepthwiseConv2D,MatMul}+BiasAdd into the
+ // _Fused{Conv2D,DepthwiseConv2dNative,MatMul}
ContractionWithBiasAdd contract_with_bias;
if (allow_non_differentiable_rewrites &&
FindContractionWithBias(ctx, i, &contract_with_bias)) {
@@ -1669,7 +1705,8 @@
continue;
}
- // Remap {Conv2D,MatMul}+BiasAdd+Activation into the _Fused{Conv2D,MatMul}.
+ // Remap {Conv2D,DepthwiseConv2D,MatMul}+BiasAdd+Activation into the
+ // _Fused{Conv2D,DepthwiseConv2dNative,MatMul}.
ContractionWithBiasAddAndActivation contract_with_bias_and_activation;
if (allow_non_differentiable_rewrites &&
FindContractionWithBiasAndActivation(
diff --git a/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc b/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc
index 358cc79..464a2c1 100644
--- a/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc
+++ b/tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc
@@ -117,7 +117,11 @@
*type = props.dtype();
} else if (*type != props.dtype()) {
return errors::Internal("Group ops don't all have same type");
- } else if (!TensorShape::IsValid(props.shape())) {
+ } else if (!TensorShape::IsValid(props.shape()) ||
+ props.shape().unknown_rank()) {
+ // TensorShape::IsValid may return true if unknown_rank is True, i.e.
+ // number of dimensions is unknown. But for ScopedAllocatorOptimizer we
+ // need to know the shape fully.
return errors::Internal("Complete shape not known for ", n->name());
}
if (*type != dtype) {
diff --git a/tensorflow/core/grappler/utils/transitive_fanin.cc b/tensorflow/core/grappler/utils/transitive_fanin.cc
index 92de380..6622807 100644
--- a/tensorflow/core/grappler/utils/transitive_fanin.cc
+++ b/tensorflow/core/grappler/utils/transitive_fanin.cc
@@ -20,15 +20,15 @@
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/grappler/utils.h"
+#include "tensorflow/core/platform/errors.h"
namespace tensorflow {
namespace grappler {
-std::vector<const NodeDef*> ComputeTransitiveFanin(
+Status ComputeTransitiveFanin(
const GraphDef& graph, const std::vector<string>& terminal_nodes,
std::unordered_map<string, const NodeDef*>* name_to_fanin_node,
- bool* ill_formed) {
- *ill_formed = false;
+ std::vector<const NodeDef*>* fanin_nodes) {
std::unordered_map<string, const NodeDef*> name_to_node;
std::unordered_map<string, const NodeDef*> name_to_send;
for (const auto& node : graph.node()) {
@@ -43,14 +43,12 @@
for (const string& root : terminal_nodes) {
const NodeDef* node = name_to_node[NodeName(root)];
if (!node) {
- *ill_formed = true;
- VLOG(2) << "ComputeTransitiveFanin: problem with root node: " << root;
- return {};
+ return errors::InvalidArgument("Graph does not contain terminal node ",
+ root, ".");
}
queue.push_back(node);
}
- std::vector<const NodeDef*> result;
std::unordered_set<const NodeDef*> visited;
while (!queue.empty()) {
@@ -60,15 +58,17 @@
// The node has already been visited.
continue;
}
- result.push_back(node);
- name_to_fanin_node->insert(
- std::pair<string, const NodeDef*>(node->name(), node));
+ fanin_nodes->push_back(node);
+ if (name_to_fanin_node) {
+ name_to_fanin_node->insert(
+ std::pair<string, const NodeDef*>(node->name(), node));
+ }
for (const string& input : node->input()) {
const NodeDef* in = name_to_node[NodeName(input)];
if (!in) {
- VLOG(2) << "ComputeTransitiveFanin: problem with node: " << input;
- *ill_formed = true;
- return {};
+ return errors::InvalidArgument("Graph does not contain input ",
+ NodeName(input), " of node ",
+ node->name(), ".");
}
queue.push_back(in);
}
@@ -82,7 +82,13 @@
// So, we do not set ill_formed for missing _Send.
}
}
- return result;
+ return Status::OK();
+}
+
+Status ComputeTransitiveFanin(const GraphDef& graph,
+ const std::vector<string>& terminal_nodes,
+ std::vector<const NodeDef*>* fanin_nodes) {
+ return ComputeTransitiveFanin(graph, terminal_nodes, nullptr, fanin_nodes);
}
Status SetTransitiveFaninGraph(const GraphDef& input_graph,
@@ -90,15 +96,9 @@
const std::vector<string>& terminal_nodes) {
// Determines transitive fanin nodes from terminal nodes and add them to the
// output graph.
- bool ill_formed = false;
- std::unordered_map<string, const NodeDef*> name_to_fanin_node;
- std::vector<const NodeDef*> keep = ComputeTransitiveFanin(
- input_graph, terminal_nodes, &name_to_fanin_node, &ill_formed);
- if (ill_formed) {
- // Some graph edges are invalid, or some of the feeds/fetch don't exist:
- // let's be conservative and preserve the graph as is.
- return errors::InvalidArgument("Invalid input graph.");
- }
+ std::vector<const NodeDef*> keep;
+ TF_RETURN_IF_ERROR(
+ ComputeTransitiveFanin(input_graph, terminal_nodes, &keep));
// Try to keep the nodes ordered somewhat topologically since this helps
// further optimizations perform better.
output_graph->mutable_node()->Reserve(keep.size());
diff --git a/tensorflow/core/grappler/utils/transitive_fanin.h b/tensorflow/core/grappler/utils/transitive_fanin.h
index 1b72eb4..11dccfc 100644
--- a/tensorflow/core/grappler/utils/transitive_fanin.h
+++ b/tensorflow/core/grappler/utils/transitive_fanin.h
@@ -25,13 +25,17 @@
namespace grappler {
// Computes the transitive fanin of the graph based on reachability from the
-// specified terminal nodes. ill_formed will be set to true if the graph is
-// deemed structurally invalid. Returns the set of nodes comprising the
-// transitive fanin.
-std::vector<const NodeDef*> ComputeTransitiveFanin(
+// specified terminal nodes. Returns the set of nodes comprising the
+// transitive fanin into fanin_nodes. Optionally returns a map of name->node
+// for that graph into name_to_fanin_node if that is not set to nullptr.
+Status ComputeTransitiveFanin(
const GraphDef& graph, const std::vector<string>& terminal_nodes,
std::unordered_map<string, const NodeDef*>* name_to_fanin_node,
- bool* ill_formed);
+ std::vector<const NodeDef*>* fanin_nodes);
+
+Status ComputeTransitiveFanin(const GraphDef& graph,
+ const std::vector<string>& terminal_nodes,
+ std::vector<const NodeDef*>* fanin_nodes);
// Creates output_graph from input_graph using the transitive fanin from the
// specified terminal nodes. Returns error if the input_graph is deemed
diff --git a/tensorflow/core/grappler/utils/transitive_fanin_test.cc b/tensorflow/core/grappler/utils/transitive_fanin_test.cc
index 94d98b9..3233dfd 100644
--- a/tensorflow/core/grappler/utils/transitive_fanin_test.cc
+++ b/tensorflow/core/grappler/utils/transitive_fanin_test.cc
@@ -117,7 +117,7 @@
ASSERT_FALSE(node_map.NodeExists("6"));
}
-TEST_F(TransitiveFaninTest, InvalidGraph) {
+TEST_F(TransitiveFaninTest, InvalidGraphOrTerminalNodes) {
GraphDef graph = CreateGraph({
{"1", {"2"}}, //
{"2", {"3"}}, //
@@ -131,7 +131,11 @@
const std::vector<string> terminal_nodes = {"1", "5"};
auto s = SetTransitiveFaninGraph(graph, &output_graph, terminal_nodes);
EXPECT_FALSE(s.ok());
- EXPECT_EQ(s.error_message(), "Invalid input graph.");
+ EXPECT_EQ(s.error_message(), "Graph does not contain input 6 of node 5.");
+ const std::vector<string> invalid_terminal_nodes = {"0", "1", "5"};
+ s = SetTransitiveFaninGraph(graph, &output_graph, invalid_terminal_nodes);
+ EXPECT_FALSE(s.ok());
+ EXPECT_EQ(s.error_message(), "Graph does not contain terminal node 0.");
}
} // namespace
diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
index 2632728..2016612 100644
--- a/tensorflow/core/kernels/BUILD
+++ b/tensorflow/core/kernels/BUILD
@@ -1371,7 +1371,9 @@
tf_kernel_library(
name = "unique_op",
prefix = "unique_op",
- deps = ARRAY_DEPS,
+ deps = ARRAY_DEPS + [
+ "@com_google_absl//absl/container:flat_hash_map",
+ ],
)
tf_kernel_library(
@@ -2335,6 +2337,7 @@
"//tensorflow/core:test",
"//tensorflow/core:test_main",
"//tensorflow/core:testlib",
+ "//tensorflow/core/kernels/data:single_threaded_executor",
],
)
@@ -6991,6 +6994,11 @@
# rule above. Seems to have only have worked before because of
# hdrs_check loose.
"stateful_random_ops_cpu_gpu.h",
+ # Allows conv_3d ops for android but excluded from *_3d* rule above.
+ "conv_3d.h",
+ "conv_ops_3d.h",
+ "conv_ops_3d.cc",
+ "conv_ops_gpu.h",
],
visibility = ["//visibility:public"],
)
diff --git a/tensorflow/core/kernels/check_numerics_op.cc b/tensorflow/core/kernels/check_numerics_op.cc
index 0feef16..6922158 100644
--- a/tensorflow/core/kernels/check_numerics_op.cc
+++ b/tensorflow/core/kernels/check_numerics_op.cc
@@ -353,7 +353,7 @@
}
}
- static const int abnormal_detected_size = 3;
+ static constexpr int abnormal_detected_size = 3;
};
#endif // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
diff --git a/tensorflow/core/kernels/cwise_op_equal_to_1.cc b/tensorflow/core/kernels/cwise_op_equal_to_1.cc
index ac66e55..64cd784 100644
--- a/tensorflow/core/kernels/cwise_op_equal_to_1.cc
+++ b/tensorflow/core/kernels/cwise_op_equal_to_1.cc
@@ -18,6 +18,7 @@
namespace tensorflow {
REGISTER7(BinaryOp, CPU, "Equal", functor::equal_to, float, Eigen::half, double,
uint8, int8, int16, bfloat16);
+REGISTER3(BinaryOp, CPU, "Equal", functor::equal_to, uint16, uint32, uint64);
REGISTER_KERNEL_BUILDER(
Name("ApproximateEqual").Device(DEVICE_CPU).TypeConstraint<float>("T"),
ApproximateEqualOp<CPUDevice, float>);
diff --git a/tensorflow/core/kernels/cwise_op_gpu_maximum.cu.cc b/tensorflow/core/kernels/cwise_op_gpu_maximum.cu.cc
index a68791d..53530f9 100644
--- a/tensorflow/core/kernels/cwise_op_gpu_maximum.cu.cc
+++ b/tensorflow/core/kernels/cwise_op_gpu_maximum.cu.cc
@@ -19,7 +19,7 @@
namespace tensorflow {
namespace functor {
-DEFINE_BINARY5(maximum, Eigen::half, float, double, int16, int64);
+DEFINE_BINARY6(maximum, Eigen::half, float, double, uint8, int16, int64);
} // namespace functor
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/cwise_op_gpu_minimum.cu.cc b/tensorflow/core/kernels/cwise_op_gpu_minimum.cu.cc
index df0f431..beab671 100644
--- a/tensorflow/core/kernels/cwise_op_gpu_minimum.cu.cc
+++ b/tensorflow/core/kernels/cwise_op_gpu_minimum.cu.cc
@@ -19,7 +19,7 @@
namespace tensorflow {
namespace functor {
-DEFINE_BINARY5(minimum, Eigen::half, float, double, int16, int64);
+DEFINE_BINARY6(minimum, Eigen::half, float, double, uint8, int16, int64);
} // namespace functor
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/cwise_op_maximum.cc b/tensorflow/core/kernels/cwise_op_maximum.cc
index a7b58d8..5ebfa74 100644
--- a/tensorflow/core/kernels/cwise_op_maximum.cc
+++ b/tensorflow/core/kernels/cwise_op_maximum.cc
@@ -16,11 +16,11 @@
#include "tensorflow/core/kernels/cwise_ops_common.h"
namespace tensorflow {
-REGISTER7(BinaryOp, CPU, "Maximum", functor::maximum, float, Eigen::half,
- bfloat16, double, int16, int32, int64);
+REGISTER8(BinaryOp, CPU, "Maximum", functor::maximum, float, Eigen::half,
+ bfloat16, double, uint8, int16, int32, int64);
#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
-REGISTER5(BinaryOp, GPU, "Maximum", functor::maximum, float, Eigen::half,
- double, int16, int64);
+REGISTER6(BinaryOp, GPU, "Maximum", functor::maximum, float, Eigen::half,
+ double, uint8, int16, int64);
// A special GPU kernel for int32.
// TODO(b/25387198): Also enable int32 in device memory. This kernel
diff --git a/tensorflow/core/kernels/cwise_op_minimum.cc b/tensorflow/core/kernels/cwise_op_minimum.cc
index 25cbbc8..8b301e8 100644
--- a/tensorflow/core/kernels/cwise_op_minimum.cc
+++ b/tensorflow/core/kernels/cwise_op_minimum.cc
@@ -16,11 +16,11 @@
#include "tensorflow/core/kernels/cwise_ops_common.h"
namespace tensorflow {
-REGISTER7(BinaryOp, CPU, "Minimum", functor::minimum, float, Eigen::half,
- bfloat16, double, int16, int32, int64);
+REGISTER8(BinaryOp, CPU, "Minimum", functor::minimum, float, Eigen::half,
+ bfloat16, double, uint8, int16, int32, int64);
#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
-REGISTER5(BinaryOp, GPU, "Minimum", functor::minimum, float, Eigen::half,
- double, int16, int64);
+REGISTER6(BinaryOp, GPU, "Minimum", functor::minimum, float, Eigen::half,
+ double, uint8, int16, int64);
// A special GPU kernel for int32.
// TODO(b/25387198): Also enable int32 in device memory. This kernel
diff --git a/tensorflow/core/kernels/cwise_op_not_equal_to_1.cc b/tensorflow/core/kernels/cwise_op_not_equal_to_1.cc
index f207158..4de69ed 100644
--- a/tensorflow/core/kernels/cwise_op_not_equal_to_1.cc
+++ b/tensorflow/core/kernels/cwise_op_not_equal_to_1.cc
@@ -18,6 +18,8 @@
namespace tensorflow {
REGISTER7(BinaryOp, CPU, "NotEqual", functor::not_equal_to, float, Eigen::half,
double, uint8, int8, int16, bfloat16);
+REGISTER3(BinaryOp, CPU, "NotEqual", functor::not_equal_to, uint16, uint32,
+ uint64);
#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
REGISTER4(BinaryOp, GPU, "NotEqual", functor::not_equal_to, float, Eigen::half,
double, uint8);
diff --git a/tensorflow/core/kernels/cwise_ops.h b/tensorflow/core/kernels/cwise_ops.h
index a64f59a..b8bf19c 100644
--- a/tensorflow/core/kernels/cwise_ops.h
+++ b/tensorflow/core/kernels/cwise_ops.h
@@ -793,7 +793,7 @@
// operation. Each functor for which this is enabled increases the
// code size, so by default this is disabled for binary functors and
// is enabled on a per-op basis as needed.
- static const bool use_bcast_optimization = false;
+ static constexpr bool use_bcast_optimization = false;
// operator() has the signature:
// out_type operator()(in_type in0, in_type in1 ...)
@@ -811,24 +811,24 @@
// Whether the functor can error out. Currently applies only to integer
// div and mod.
- static const bool has_errors = false;
+ static constexpr bool has_errors = false;
};
// For now, we only apply certain speed optimization for
// float/double's broadcast binary op.
template <typename T>
struct use_bcast_optimization {
- static const bool value = false;
+ static constexpr bool value = false;
};
template <>
struct use_bcast_optimization<float> {
- static const bool value = true;
+ static constexpr bool value = true;
};
template <>
struct use_bcast_optimization<double> {
- static const bool value = true;
+ static constexpr bool value = true;
};
////////////////////////////////////////////////////////////////////////////////
@@ -1007,17 +1007,17 @@
template <typename T>
struct add : base<T, Eigen::internal::scalar_sum_op<T>> {
- static const bool use_bcast_optimization = true;
+ static constexpr bool use_bcast_optimization = true;
};
template <typename T>
struct sub : base<T, Eigen::internal::scalar_difference_op<T>> {
- static const bool use_bcast_optimization = true;
+ static constexpr bool use_bcast_optimization = true;
};
template <typename T>
struct mul : base<T, Eigen::internal::scalar_product_op<T>> {
- static const bool use_bcast_optimization = true;
+ static constexpr bool use_bcast_optimization = true;
};
template <typename T>
@@ -1029,7 +1029,7 @@
template <typename T>
struct safe_div : base<T, Eigen::internal::safe_div_or_mod_op<
T, Eigen::internal::scalar_quotient_op<T>>> {
- static const bool has_errors = true;
+ static constexpr bool has_errors = true;
};
template <typename T>
@@ -1044,7 +1044,7 @@
template <typename T>
struct safe_mod : base<T, Eigen::internal::safe_div_or_mod_op<
T, Eigen::internal::scalar_mod2_op<T>>> {
- static const bool has_errors = true;
+ static constexpr bool has_errors = true;
};
template <typename T>
@@ -1053,7 +1053,7 @@
template <typename T>
struct safe_floor_mod : base<T, Eigen::internal::safe_div_or_mod_op<
T, Eigen::internal::google_floor_mod<T>>> {
- static const bool has_errors = true;
+ static constexpr bool has_errors = true;
};
template <typename T>
@@ -1062,7 +1062,7 @@
template <typename T>
struct safe_floor_div : base<T, Eigen::internal::safe_div_or_mod_op<
T, Eigen::internal::google_floor_div<T>>> {
- static const bool has_errors = true;
+ static constexpr bool has_errors = true;
};
template <typename T>
@@ -1073,7 +1073,7 @@
template <typename T>
struct safe_pow : base<T, Eigen::internal::safe_scalar_binary_pow_op<T, T>> {
- static const bool has_errors = true;
+ static constexpr bool has_errors = true;
};
template <typename T>
diff --git a/tensorflow/core/kernels/data/BUILD b/tensorflow/core/kernels/data/BUILD
index 823a800..65881da 100644
--- a/tensorflow/core/kernels/data/BUILD
+++ b/tensorflow/core/kernels/data/BUILD
@@ -170,6 +170,7 @@
"//tensorflow/core:core_cpu",
"//tensorflow/core:core_cpu_internal",
"//tensorflow/core:lib",
+ "//tensorflow/core/common_runtime:local_executor_params",
],
alwayslink = 1,
)
diff --git a/tensorflow/core/kernels/data/cache_dataset_ops.cc b/tensorflow/core/kernels/data/cache_dataset_ops.cc
index 707800b..9a1a4ee 100644
--- a/tensorflow/core/kernels/data/cache_dataset_ops.cc
+++ b/tensorflow/core/kernels/data/cache_dataset_ops.cc
@@ -48,7 +48,6 @@
constexpr char kCreatedAt[] = "Created at";
constexpr char kMemoryDatasetPrefix[] = "Memory";
constexpr char kMemoryCache[] = "MemoryCache";
-constexpr char kTFData[] = "tf_data";
constexpr char kCacheClaimed[] = "cache_claimed";
constexpr char kCacheSize[] = "cache_size";
constexpr char kCache[] = "cache";
@@ -58,10 +57,10 @@
constexpr char kImpl[] = "Impl";
constexpr char kCacheDataset[] = "CacheDataset";
-class CacheDatasetOp::FileDataset : public DatasetBase {
+class CacheDatasetOp::FileDatasetBase : public DatasetBase {
public:
- explicit FileDataset(OpKernelContext* ctx, const DatasetBase* input,
- string filename, Env* env)
+ FileDatasetBase(OpKernelContext* ctx, const DatasetBase* input,
+ string filename, Env* env)
: DatasetBase(DatasetContext(ctx)),
input_(input),
filename_(std::move(filename)),
@@ -76,7 +75,7 @@
DCHECK_EQ(item_index_padding_size_, 7);
}
- ~FileDataset() override { input_->Unref(); }
+ ~FileDatasetBase() override { input_->Unref(); }
std::unique_ptr<IteratorBase> MakeIteratorInternal(
const string& prefix) const override {
@@ -107,17 +106,6 @@
}
protected:
- Status AsGraphDefInternal(SerializationContext* ctx,
- DatasetGraphDefBuilder* b,
- Node** output) const override {
- Node* input_graph = nullptr;
- TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph));
- Node* filename = nullptr;
- TF_RETURN_IF_ERROR(b->AddScalar(filename_, &filename));
- TF_RETURN_IF_ERROR(b->AddDataset(this, {input_graph, filename}, output));
- return Status::OK();
- }
-
const DatasetBase* const input_;
const tstring filename_;
@@ -131,10 +119,10 @@
tensor_index);
}
- class FileIterator : public DatasetIterator<FileDataset> {
+ class FileIterator : public DatasetIterator<FileDatasetBase> {
public:
explicit FileIterator(const Params& params)
- : DatasetIterator<FileDataset>(params) {
+ : DatasetIterator<FileDatasetBase>(params) {
if (params.dataset->env_
->FileExists(MetaFilename(params.dataset->filename_))
.ok()) {
@@ -199,7 +187,7 @@
private:
// FileWriterIterator passes through and caches items from the input
- // FileDataset.
+ // FileDatasetBase.
//
// This iterator is used when the cache directory is not found on disk. It
// creates the cache directory, and passes on the underlying iterator's
@@ -214,10 +202,10 @@
// partial cache gets flushed to disk in files with prefix
// <filename>_<shard_id> where shard_id is unique for each checkpoint.
// When all elements have been produced, these shards get coalesced.
- class FileWriterIterator : public DatasetIterator<FileDataset> {
+ class FileWriterIterator : public DatasetIterator<FileDatasetBase> {
public:
explicit FileWriterIterator(const Params& params)
- : DatasetIterator<FileDataset>(params),
+ : DatasetIterator<FileDatasetBase>(params),
cur_index_(0),
shard_id_(0),
filename_(
@@ -483,10 +471,10 @@
bool iteration_completed_ TF_GUARDED_BY(mu_);
}; // FileWriterIterator
- class FileReaderIterator : public DatasetIterator<FileDataset> {
+ class FileReaderIterator : public DatasetIterator<FileDatasetBase> {
public:
explicit FileReaderIterator(const Params& params)
- : DatasetIterator<FileDataset>(params),
+ : DatasetIterator<FileDatasetBase>(params),
cur_index_(0),
reader_(dataset()->env_, dataset()->filename_),
iterator_restored_(false) {}
@@ -603,17 +591,34 @@
Env* const env_;
const size_t num_tensors_;
const size_t tensor_index_padding_size_;
- static const size_t kMaxItems = 10000000; // 10 million
+ static constexpr size_t kMaxItems = 10000000; // 10 million
const size_t item_index_padding_size_;
const string tensor_format_string_;
-}; // FileDataset
+}; // FileDatasetBase
-class CacheDatasetOp::FileDatasetV2 : public CacheDatasetOp::FileDataset {
+class CacheDatasetOp::FileDataset : public CacheDatasetOp::FileDatasetBase {
+ public:
+ using FileDatasetBase::FileDatasetBase;
+
+ protected:
+ Status AsGraphDefInternal(SerializationContext* ctx,
+ DatasetGraphDefBuilder* b,
+ Node** output) const override {
+ Node* input_graph = nullptr;
+ TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph));
+ Node* filename = nullptr;
+ TF_RETURN_IF_ERROR(b->AddScalar(filename_, &filename));
+ TF_RETURN_IF_ERROR(b->AddDataset(this, {input_graph, filename}, output));
+ return Status::OK();
+ }
+};
+
+class CacheDatasetOp::FileDatasetV2 : public CacheDatasetOp::FileDatasetBase {
public:
explicit FileDatasetV2(OpKernelContext* ctx, const DatasetBase* input,
string filename, Env* env,
const Tensor& resource_handle)
- : FileDataset(ctx, input, filename, env),
+ : FileDatasetBase(ctx, input, filename, env),
resource_handle_(resource_handle) {}
protected:
@@ -686,20 +691,17 @@
} // namespace
-class CacheDatasetOp::MemoryDataset : public DatasetBase {
+class CacheDatasetOp::MemoryDatasetBase : public DatasetBase {
public:
- explicit MemoryDataset(OpKernelContext* ctx, const DatasetBase* input,
- MemoryCache* cache)
- : DatasetBase(DatasetContext(ctx)), input_(input), cache_(cache) {
+ explicit MemoryDatasetBase(OpKernelContext* ctx, const DatasetBase* input,
+ std::shared_ptr<MemoryCache> cache)
+ : DatasetBase(DatasetContext(ctx)),
+ input_(input),
+ cache_(std::move(cache)) {
input_->Ref();
}
- ~MemoryDataset() override {
- input_->Unref();
- if (cache_) {
- cache_->Unref();
- }
- }
+ ~MemoryDatasetBase() override { input_->Unref(); }
std::unique_ptr<IteratorBase> MakeIteratorInternal(
const string& prefix) const override {
@@ -708,7 +710,7 @@
return absl::make_unique<MemoryIterator>(
MemoryIterator::Params{
this, name_utils::IteratorPrefix(kDatasetType, prefix, params)},
- cache_);
+ cache_.get());
}
const DataTypeVector& output_dtypes() const override {
@@ -732,44 +734,13 @@
}
protected:
- Status AsGraphDefInternal(SerializationContext* ctx,
- DatasetGraphDefBuilder* b,
- Node** output) const override {
- Node* input_node = nullptr;
- TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_node));
- Node* filename_node = nullptr;
- TF_RETURN_IF_ERROR(b->AddScalar(tstring(""), &filename_node));
- TF_RETURN_IF_ERROR(
- b->AddDataset(this, {input_node, filename_node}, output));
- return Status::OK();
- }
-
- class MemoryIterator : public DatasetIterator<MemoryDataset> {
+ class MemoryIterator : public DatasetIterator<MemoryDatasetBase> {
public:
explicit MemoryIterator(const Params& params, MemoryCache* cache)
- : DatasetIterator<MemoryDataset>(params), cache_(cache) {}
-
- ~MemoryIterator() override {
- if (dataset()->cache_ == nullptr) {
- cache_->Unref();
- }
- }
+ : DatasetIterator<MemoryDatasetBase>(params), cache_(cache) {}
Status Initialize(IteratorContext* ctx) override {
mutex_lock l(mu_);
- if (cache_ == nullptr) {
- // Use the resource manager in the iterator context to get / create
- // a cache.
- ResourceMgr* mgr = ctx->resource_mgr();
- const string name = strings::StrCat(
- prefix(), name_utils::kDelimiter, dataset()->node_name(),
- name_utils::kDelimiter, kMemoryCache);
- TF_RETURN_IF_ERROR(mgr->LookupOrCreate<MemoryCache>(
- kTFData, name, &cache_, [](MemoryCache** cache) {
- *cache = new MemoryCache();
- return Status::OK();
- }));
- }
InitializeIterator();
return iterator_->Initialize(ctx);
}
@@ -778,7 +749,14 @@
std::vector<Tensor>* out_tensors,
bool* end_of_sequence) override {
mutex_lock l(mu_);
- return iterator_->GetNext(ctx, out_tensors, end_of_sequence);
+ // TODO(b/154341936): Explicitly stopping and starting this iterator
+ // should not be necessary, but the `kImpl` added to the prefix passed
+ // to `iterator_` when it was created prevents the model from identifying
+ // this iterator as the output of `iterator_`.
+ RecordStop(ctx);
+ Status s = iterator_->GetNext(ctx, out_tensors, end_of_sequence);
+ RecordStart(ctx);
+ return s;
}
protected:
@@ -817,10 +795,10 @@
}
private:
- class MemoryWriterIterator : public DatasetIterator<MemoryDataset> {
+ class MemoryWriterIterator : public DatasetIterator<MemoryDatasetBase> {
public:
explicit MemoryWriterIterator(const Params& params, MemoryCache* cache)
- : DatasetIterator<MemoryDataset>(params), cache_(cache) {}
+ : DatasetIterator<MemoryDatasetBase>(params), cache_(cache) {}
~MemoryWriterIterator() override {
mutex_lock l(mu_);
@@ -900,12 +878,12 @@
std::vector<std::vector<Tensor>> temp_cache_ TF_GUARDED_BY(mu_);
}; // MemoryWriterIterator
- class MemoryReaderIterator : public DatasetIterator<MemoryDataset> {
+ class MemoryReaderIterator : public DatasetIterator<MemoryDatasetBase> {
public:
explicit MemoryReaderIterator(const Params& params, MemoryCache* cache)
- : DatasetIterator<MemoryDataset>(params), cache_(cache), index_(0) {
- CHECK(cache);
- }
+ : DatasetIterator<MemoryDatasetBase>(params),
+ cache_(cache),
+ index_(0) {}
Status Initialize(IteratorContext* ctx) override {
// The memory allocated for the cache is owned by the parent
@@ -988,19 +966,73 @@
}; // MemoryIterator
const DatasetBase* const input_;
- MemoryCache* cache_ = nullptr;
-}; // MemoryDataset
+ const std::shared_ptr<MemoryCache> cache_;
+}; // MemoryDatasetBase
-class CacheDatasetOp::MemoryDatasetV2 : public CacheDatasetOp::MemoryDataset {
+// This version of memory dataset has an exclusive ownership of the memory cache
+// resource. It supports sharing of the cache across different iterations of the
+// `repeat` transformation but not across different iterators.
+class CacheDatasetOp::MemoryDataset : public CacheDatasetOp::MemoryDatasetBase {
public:
- explicit MemoryDatasetV2(OpKernelContext* ctx, const DatasetBase* input,
- MemoryCache* cache,
- std::unique_ptr<OwnedResourceHandle> handle)
- : MemoryDataset(ctx, input, cache), handle_(std::move(handle)) {}
+ MemoryDataset(OpKernelContext* ctx, const DatasetBase* input,
+ MemoryCacheManager* manager, ResourceHandle&& resource_handle)
+ : MemoryDatasetBase(ctx, input, manager->get()),
+ manager_(manager),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()) {}
- Status CheckExternalState() const override {
- return errors::FailedPrecondition(DebugString(),
- " depends on memory cache resource.");
+ ~MemoryDataset() override {
+ manager_->Unref();
+ Status s = resource_mgr_->Delete<MemoryCacheManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete cache resource: " << s.ToString();
+ }
+ }
+
+ protected:
+ Status AsGraphDefInternal(SerializationContext* ctx,
+ DatasetGraphDefBuilder* b,
+ Node** output) const override {
+ Node* input_node = nullptr;
+ TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_node));
+ Node* filename_node = nullptr;
+ TF_RETURN_IF_ERROR(b->AddScalar(tstring(""), &filename_node));
+ TF_RETURN_IF_ERROR(
+ b->AddDataset(this, {input_node, filename_node}, output));
+ return Status::OK();
+ }
+
+ private:
+ MemoryCacheManager* const manager_; // Owned.
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
+};
+
+// This version of memory dataset has a shared ownership of the memory cache
+// resource. It supports sharing of the cache across different iterations of
+// the `repeat` transformation and also across different iterators.
+class CacheDatasetOp::MemoryDatasetV2
+ : public CacheDatasetOp::MemoryDatasetBase {
+ public:
+ MemoryDatasetV2(OpKernelContext* ctx, const DatasetBase* input,
+ MemoryCacheManager* manager, ResourceHandle&& resource_handle,
+ bool owns_resource)
+ : MemoryDatasetBase(ctx, input, manager->get()),
+ manager_(manager),
+ owns_resource_(owns_resource),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()) {}
+
+ ~MemoryDatasetV2() override {
+ manager_->Unref();
+ if (owns_resource_) {
+ Status s = resource_mgr_->Delete<MemoryCacheManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete cache resource: " << s.ToString();
+ }
+ }
}
protected:
@@ -1013,7 +1045,7 @@
TF_RETURN_IF_ERROR(b->AddScalar(tstring(""), &filename_node));
Node* resource_handle_node = nullptr;
Tensor handle(DT_RESOURCE, TensorShape({}));
- handle.scalar<ResourceHandle>()() = handle_->handle();
+ handle.scalar<ResourceHandle>()() = resource_handle_;
TF_RETURN_IF_ERROR(b->AddTensor(handle, &resource_handle_node));
TF_RETURN_IF_ERROR(b->AddDataset(
this, {input_node, filename_node, resource_handle_node}, output));
@@ -1021,7 +1053,10 @@
}
private:
- std::unique_ptr<OwnedResourceHandle> handle_;
+ MemoryCacheManager* const manager_; // Owned.
+ const bool owns_resource_;
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
};
CacheDatasetOp::CacheDatasetOp(OpKernelConstruction* ctx)
@@ -1033,22 +1068,45 @@
// Parse out the filenames tensor.
tstring filename;
OP_REQUIRES_OK(ctx, ParseScalarArgument<tstring>(ctx, kFileName, &filename));
-
if (filename.empty()) {
+ static std::atomic<int64> resource_id_counter(0);
+ const string& container = ctx->resource_manager()->default_container();
+ auto name = strings::StrCat(ctx->op_kernel().name(), "/", kMemoryCache, "_",
+ resource_id_counter.fetch_add(1));
if (op_version_ == 2) {
- MemoryCache* cache = nullptr;
- OP_REQUIRES_OK(ctx, LookupResource(ctx, HandleFromInput(ctx, 2), &cache));
-
- // Create a fresh handle for the resource because the input handle can
- // become invalid after this op executes.
- std::unique_ptr<OwnedResourceHandle> handle;
- OP_REQUIRES_OK(
- ctx, OwnedResourceHandle::Create(ctx, cache, kMemoryCache, &handle));
-
- // Ownership of cache is transferred onto `MemoryDatasetV2`.
- *output = new MemoryDatasetV2(ctx, input, cache, std::move(handle));
+ bool owns_resource = false;
+ MemoryCacheManager* manager = nullptr;
+ auto handle = HandleFromInput(ctx, 2);
+ Status s = ctx->resource_manager()->Lookup<MemoryCacheManager>(
+ handle.container(), handle.name(), &manager);
+ if (errors::IsNotFound(s)) {
+ owns_resource = true;
+ OP_REQUIRES_OK(
+ ctx,
+ ctx->resource_manager()->LookupOrCreate<MemoryCacheManager>(
+ container, name, &manager, [](MemoryCacheManager** manager) {
+ *manager = new MemoryCacheManager();
+ return Status::OK();
+ }));
+ handle = MakeResourceHandle<MemoryCacheManager>(ctx, container, name);
+ } else {
+ OP_REQUIRES_OK(ctx, s);
+ }
+ // Ownership of manager is transferred onto `MemoryDatasetV2`.
+ *output = new MemoryDatasetV2(ctx, input, manager, std::move(handle),
+ owns_resource);
} else {
- *output = new MemoryDataset(ctx, input, /*cache=*/nullptr);
+ MemoryCacheManager* manager;
+ OP_REQUIRES_OK(
+ ctx, ctx->resource_manager()->LookupOrCreate<MemoryCacheManager>(
+ container, name, &manager, [](MemoryCacheManager** manager) {
+ *manager = new MemoryCacheManager();
+ return Status::OK();
+ }));
+ auto handle =
+ MakeResourceHandle<MemoryCacheManager>(ctx, container, name);
+ // Ownership of manager is transferred onto `MemoryDataset`.
+ *output = new MemoryDataset(ctx, input, manager, std::move(handle));
}
} else {
if (op_version_ == 2) {
diff --git a/tensorflow/core/kernels/data/cache_dataset_ops.h b/tensorflow/core/kernels/data/cache_dataset_ops.h
index 484d048..e0ceee2 100644
--- a/tensorflow/core/kernels/data/cache_dataset_ops.h
+++ b/tensorflow/core/kernels/data/cache_dataset_ops.h
@@ -22,8 +22,8 @@
class CacheDatasetOp : public UnaryDatasetOpKernel {
public:
- class FileDataset;
- class MemoryDataset;
+ class FileDatasetBase;
+ class MemoryDatasetBase;
static constexpr const char* const kDatasetType = "Cache";
static constexpr const char* const kInputDataset = "input_dataset";
@@ -38,10 +38,12 @@
DatasetBase** output) override;
private:
+ class FileDataset;
class FileDatasetV2;
+ class MemoryDataset;
class MemoryDatasetV2;
- int op_version_;
+ const int op_version_;
};
} // namespace data
diff --git a/tensorflow/core/kernels/data/cache_ops.cc b/tensorflow/core/kernels/data/cache_ops.cc
index 371a2ae..90c2e90 100644
--- a/tensorflow/core/kernels/data/cache_ops.cc
+++ b/tensorflow/core/kernels/data/cache_ops.cc
@@ -18,6 +18,7 @@
#include "tensorflow/core/framework/partial_tensor_shape.h"
#include "tensorflow/core/framework/resource_mgr.h"
#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/kernels/data/dataset_utils.h"
#include "tensorflow/core/lib/random/philox_random.h"
#include "tensorflow/core/lib/random/random.h"
#include "tensorflow/core/lib/random/random_distributions.h"
@@ -26,11 +27,11 @@
namespace data {
namespace {
-const char kMemoryCache[] = "MemoryCache";
+constexpr char kMemoryCache[] = "MemoryCache";
} // namespace
-string MemoryCache::DebugString() const { return kMemoryCache; }
+string MemoryCacheManager::DebugString() const { return kMemoryCache; }
void MemoryCache::Complete(std::vector<std::vector<Tensor>>&& cache) {
mutex_lock l(mu_);
@@ -64,28 +65,25 @@
AnonymousMemoryCacheHandleOp::AnonymousMemoryCacheHandleOp(
OpKernelConstruction* ctx)
- : AnonymousResourceOp<MemoryCache>(ctx) {}
-
-void AnonymousMemoryCacheHandleOp::Compute(OpKernelContext* ctx) {
- AnonymousResourceOp<MemoryCache>::Compute(ctx);
-}
+ : AnonymousResourceOp<MemoryCacheManager>(ctx) {}
string AnonymousMemoryCacheHandleOp::name() { return kMemoryCache; }
Status AnonymousMemoryCacheHandleOp::CreateResource(
OpKernelContext* ctx, std::unique_ptr<FunctionLibraryDefinition> flib_def,
std::unique_ptr<ProcessFunctionLibraryRuntime> pflr,
- FunctionLibraryRuntime* lib, MemoryCache** resource) {
- *resource = new MemoryCache();
+ FunctionLibraryRuntime* lib, MemoryCacheManager** manager) {
+ *manager = new MemoryCacheManager();
return Status::OK();
}
void DeleteMemoryCacheOp::Compute(OpKernelContext* ctx) {
const ResourceHandle& handle = ctx->input(0).flat<ResourceHandle>()(0);
- // The resource is guaranteed to exist because the variant tensor wrapping the
- // deleter is provided as an unused input to this op, which guarantees that it
- // has not run yet.
- OP_REQUIRES_OK(ctx, ctx->resource_manager()->Delete(handle));
+ // The resource might have been already deleted by the dataset.
+ Status s = ctx->resource_manager()->Delete(handle);
+ if (!errors::IsNotFound(s)) {
+ OP_REQUIRES_OK(ctx, s);
+ }
}
namespace {
@@ -96,6 +94,9 @@
REGISTER_KERNEL_BUILDER(Name("DeleteMemoryCache").Device(DEVICE_CPU),
DeleteMemoryCacheOp);
+REGISTER_KERNEL_BUILDER(Name("DummyMemoryCache").Device(DEVICE_CPU),
+ DummyResourceOp<MemoryCache>);
+
} // namespace
} // namespace data
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/data/cache_ops.h b/tensorflow/core/kernels/data/cache_ops.h
index d21679b..c670d6f 100644
--- a/tensorflow/core/kernels/data/cache_ops.h
+++ b/tensorflow/core/kernels/data/cache_ops.h
@@ -27,12 +27,10 @@
// The expected use is that a single `MemoryWriterIterator` populates the
// cache with dataset elements. Once all elements are cached, the cache can
// be used by one or more `MemoryReaderIterator`s.
-class MemoryCache : public ResourceBase {
+class MemoryCache {
public:
MemoryCache() = default;
- string DebugString() const override;
-
// Marks the cache as completed.
void Complete(std::vector<std::vector<Tensor>>&& cache);
@@ -55,11 +53,24 @@
std::vector<std::vector<Tensor>> cache_ TF_GUARDED_BY(mu_);
};
+// A resource wrapping a shared instance of a memory cache.
+class MemoryCacheManager : public ResourceBase {
+ public:
+ MemoryCacheManager() : cache_(std::make_shared<MemoryCache>()) {}
+
+ string DebugString() const override;
+
+ std::shared_ptr<MemoryCache> get() { return cache_; }
+
+ private:
+ std::shared_ptr<MemoryCache> cache_;
+};
+
// Creates an instance of cache resource and transfers ownership to the caller.
-class AnonymousMemoryCacheHandleOp : public AnonymousResourceOp<MemoryCache> {
+class AnonymousMemoryCacheHandleOp
+ : public AnonymousResourceOp<MemoryCacheManager> {
public:
explicit AnonymousMemoryCacheHandleOp(OpKernelConstruction* ctx);
- void Compute(OpKernelContext* ctx) override;
private:
string name() override;
@@ -67,7 +78,7 @@
std::unique_ptr<FunctionLibraryDefinition> flib_def,
std::unique_ptr<ProcessFunctionLibraryRuntime> pflr,
FunctionLibraryRuntime* lib,
- MemoryCache** resource) override;
+ MemoryCacheManager** manager) override;
};
// Deletes an instance of cache resource.
diff --git a/tensorflow/core/kernels/data/dataset_utils.h b/tensorflow/core/kernels/data/dataset_utils.h
index bedd5fa..d8ae719 100644
--- a/tensorflow/core/kernels/data/dataset_utils.h
+++ b/tensorflow/core/kernels/data/dataset_utils.h
@@ -39,40 +39,6 @@
return Status::OK();
}
-// A wrapper class that manages the lifetime of a resource handle from its
-// creation to its deletion from the resource manager.
-class OwnedResourceHandle {
- public:
- template <typename T>
- static Status Create(OpKernelContext* ctx, T* resource, const string& name,
- std::unique_ptr<OwnedResourceHandle>* result) {
- ResourceHandle handle;
- TF_RETURN_IF_ERROR(CreateHandle<T>(ctx, resource, name, &handle));
- // We need to increase the refcount to match the decrease that occurs when
- // the resource associate.
- resource->Ref();
- *result = absl::make_unique<OwnedResourceHandle>(ctx, std::move(handle));
- return Status::OK();
- }
-
- OwnedResourceHandle(OpKernelContext* ctx, ResourceHandle&& handle)
- : mgr_(ctx->resource_manager()), handle_(handle) {}
-
- ~OwnedResourceHandle() {
- Status s = mgr_->Delete(handle_);
- if (!s.ok()) {
- VLOG(2) << s.ToString();
- }
- }
-
- // Returns the wrapped `ResourceHandle` object.
- const ResourceHandle& handle() const { return handle_; }
-
- private:
- ResourceMgr* mgr_; // not owned
- const ResourceHandle handle_;
-};
-
template <typename T>
class AnonymousResourceOp : public OpKernel {
public:
@@ -286,6 +252,28 @@
std::function<void(std::function<void()>)> RunnerWithMaxParallelism(
std::function<void(std::function<void()>)> runner, int max_parallelism);
+// Op for creating a typed dummy resource.
+//
+// This op is used to provide a resource "placeholder" for ops such as
+// `CacheDatasetV2` or `ShuffleDatasetV2` that expects a resource input.
+// Originally, the lifetime of the resources passed into these ops was managed
+// externally. After the implementation changed to manage the lifetime of the
+// resources (including creation) by the ops themselves, the resource input is
+// only needed to pass a resource handle through graph rewrites. When they are
+// invoked from user code, the implementation passes in a dummy resource.
+template <typename ResourceType>
+class DummyResourceOp : public OpKernel {
+ public:
+ explicit DummyResourceOp(OpKernelConstruction* ctx) : OpKernel(ctx) {}
+
+ void Compute(OpKernelContext* ctx) override {
+ Tensor* tensor;
+ OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({}), &tensor));
+ tensor->scalar<ResourceHandle>()() = MakeResourceHandle<ResourceType>(
+ ctx, /*container=*/"", /*name=*/"dummy_resource");
+ }
+};
+
} // namespace data
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/data/experimental/BUILD b/tensorflow/core/kernels/data/experimental/BUILD
index 16f496a..d6c9f3c 100644
--- a/tensorflow/core/kernels/data/experimental/BUILD
+++ b/tensorflow/core/kernels/data/experimental/BUILD
@@ -526,6 +526,7 @@
"//tensorflow/core:lib",
"//tensorflow/core:lib_internal",
"//tensorflow/core:protos_all_cc",
+ "//tensorflow/core/kernels/data:name_utils",
"//tensorflow/core/platform:coding",
"//tensorflow/core/platform:random",
"//tensorflow/core/profiler/lib:traceme",
diff --git a/tensorflow/core/kernels/data/experimental/data_service_dataset_op.cc b/tensorflow/core/kernels/data/experimental/data_service_dataset_op.cc
index f0174bf..d32383a 100644
--- a/tensorflow/core/kernels/data/experimental/data_service_dataset_op.cc
+++ b/tensorflow/core/kernels/data/experimental/data_service_dataset_op.cc
@@ -60,8 +60,8 @@
// the current attempt to complete and perform no more retries.
const int64 kRetryTimeoutMicros = 1000LL * 1000 * 60 * 60; // 60 minutes.
-// How often to refresh the task list.
-const int64 kRefreshTasksIntervalMicros = 1000LL * 1000 * 60; // 60 seconds.
+// Default interval between task list refreshes.
+const int64 kDefaultTaskRefreshIntervalMs = 1000; // 1 second.
// Dataset for reading data from the tf.data service non-deterministically.
//
@@ -71,13 +71,14 @@
class DataServiceDatasetOp::Dataset : public DatasetBase {
public:
Dataset(OpKernelContext* ctx, const std::string& address,
- const std::string& protocol, const int64 max_outstanding_requests,
- const DataTypeVector& output_types,
+ const std::string& protocol, int64 max_outstanding_requests,
+ int64 task_refresh_interval_ms, const DataTypeVector& output_types,
const std::vector<PartialTensorShape>& output_shapes)
: DatasetBase(DatasetContext(ctx)),
address_(address),
protocol_(protocol),
max_outstanding_requests_(max_outstanding_requests),
+ task_refresh_interval_ms_(task_refresh_interval_ms),
output_types_(output_types),
output_shapes_(output_shapes) {}
@@ -117,8 +118,15 @@
TF_RETURN_IF_ERROR(
b->AddScalar(max_outstanding_requests_, &max_outstanding_requests));
- TF_RETURN_IF_ERROR(b->AddDataset(
- this, {address, protocol, max_outstanding_requests}, {}, output));
+ AttrValue task_refresh_interval_hint_ms;
+ b->BuildAttrValue(task_refresh_interval_ms_,
+ &task_refresh_interval_hint_ms);
+
+ TF_RETURN_IF_ERROR(
+ b->AddDataset(this, {address, protocol, max_outstanding_requests},
+ {std::make_pair(kTaskRefreshIntervalHintMs,
+ task_refresh_interval_hint_ms)},
+ output));
return Status::OK();
}
@@ -158,14 +166,13 @@
VLOG(3) << "Calling GetNext in data service dataset op";
mutex_lock l(mu_);
if (!task_thread_manager_ && !cancelled_) {
- task_thread_manager_ = ctx->StartThread(
- "task-thread-manager", [this, ctx]() { TaskThreadManager(ctx); });
+ task_thread_manager_ =
+ ctx->StartThread("task-thread-manager", [this, ctx]() {
+ TaskThreadManager(absl::make_unique<IteratorContext>(*ctx));
+ });
}
- // tasks_.empty() indicates that we haven't yet received tasks from the
- // master, so we should wait.
- while (results_.empty() &&
- (tasks_.empty() || num_unfinished_tasks_ > 0) && !cancelled_) {
+ while (results_.empty() && !job_finished_ && !cancelled_) {
cv_.wait(l);
}
if (cancelled_) {
@@ -208,13 +215,15 @@
std::unique_ptr<WorkerService::Stub> worker_stub;
std::unique_ptr<Thread> thread;
bool end_of_sequence = false;
+ // Indicates that the thread has finished running.
+ bool finished = false;
} TaskThread;
// Periodically refresh the task list.
// Maintain one thread fetching elements for each task.
// TODO(aaudibert): Instead of polling, have master send updates when
// the list of tasks changes.
- void TaskThreadManager(IteratorContext* ctx) {
+ void TaskThreadManager(std::unique_ptr<IteratorContext> ctx) {
VLOG(3) << "Starting task handler manager";
auto channel = ::grpc::CreateChannel(dataset()->address_, credentials_);
std::unique_ptr<MasterService::Stub> master_stub =
@@ -231,11 +240,13 @@
cv_.wait_for(l, std::chrono::microseconds(remaining_time));
}
if (cancelled_) {
+ VLOG(3) << "Task thread manager finished";
return;
}
}
- UpdateTaskThreads(master_stub.get(), ctx);
- next_check = Env::Default()->NowMicros() + kRefreshTasksIntervalMicros;
+ UpdateTaskThreads(master_stub.get(), ctx.get());
+ next_check = Env::Default()->NowMicros() +
+ dataset()->task_refresh_interval_ms_ * 1000;
}
}
@@ -254,24 +265,36 @@
}
absl::flat_hash_set<int64> task_ids;
mutex_lock l(mu_);
+ job_finished_ = resp.job_finished();
for (auto& task : resp.task_info()) {
task_ids.insert(task.id());
if (task_threads_.contains(task.id())) {
continue;
}
- tasks_[task.id()] = task;
task_threads_[task.id()] = absl::make_unique<TaskThread>();
TaskThread* task_handler = task_threads_[task.id()].get();
task_handler->task_id = task.id();
+ task_handler->address = task.worker_address();
num_unfinished_tasks_++;
- task_handler->thread = ctx->StartThread(
- "tf-data-service-task_handler",
- [this, task_handler]() { RunTaskThread(task_handler); });
+ outstanding_requests_++;
+ auto done = [this, task_handler]() {
+ mutex_lock l(mu_);
+ num_unfinished_tasks_--;
+ outstanding_requests_--;
+ cv_.notify_all();
+ task_handler->finished = true;
+ VLOG(3) << "Task thread " << task_handler->task_id << " finished";
+ };
+ task_handler->thread =
+ ctx->StartThread("tf-data-service-task_handler",
+ [this, task_handler, done = std::move(done)]() {
+ RunTaskThread(task_handler, std::move(done));
+ });
}
// Mark deleted tasks and clean up finished task threads.
for (auto it = task_threads_.begin(); it != task_threads_.end();) {
TaskThread* task_thread = it->second.get();
- if (task_thread->end_of_sequence) {
+ if (task_thread->finished) {
task_threads_.erase(it++);
continue;
}
@@ -286,18 +309,8 @@
}
}
- void RunTaskThread(TaskThread* task_handler) {
- auto cleanup = gtl::MakeCleanup([this]() {
- mutex_lock l(mu_);
- outstanding_requests_--;
- num_unfinished_tasks_--;
- cv_.notify_all();
- });
- {
- mutex_lock l(mu_);
- outstanding_requests_++;
- task_handler->address = tasks_[task_handler->task_id].worker_address();
- }
+ void RunTaskThread(TaskThread* task_handler, std::function<void()> done) {
+ auto cleanup = gtl::MakeCleanup([done = std::move(done)]() { done(); });
VLOG(3) << "Starting task handler thread for task "
<< task_handler->task_id << " with worker address "
<< task_handler->address;
@@ -339,73 +352,32 @@
}
}
+ // Fetches an element from a task and adds the element to `results_`.
+ //
+ // If the task reaches end_of_sequence or is cancelled (e.g. due to a
+ // worker dying), FetchElement returns Status::OK() without adding to
+ // `results_`.
Status FetchElement(TaskThread* task_handler, int64 deadline_micros) {
- VLOG(3) << "Fetchng an element for task id " << task_handler->task_id;
+ VLOG(3) << "Fetching an element for task id " << task_handler->task_id;
GetElementResponse resp;
- TF_RETURN_IF_ERROR(
- GetElementWithDeadline(task_handler, &resp, deadline_micros));
- std::vector<Tensor> element;
- if (!resp.end_of_sequence()) {
- TF_RETURN_IF_ERROR(
- service_util::Uncompress(resp.compressed_element(), &element));
- }
- mutex_lock l(mu_);
- if (resp.end_of_sequence()) {
- task_handler->end_of_sequence = true;
- return Status::OK();
- }
- results_.push(std::move(element));
- cv_.notify_all();
- VLOG(3) << "Fetched an element for task id " << task_handler->task_id;
- return Status::OK();
- }
-
- Status CreateWorkerStub(const std::string& worker_address,
- std::unique_ptr<WorkerService::Stub>* stub) {
- ::grpc::ChannelArguments args;
- args.SetMaxReceiveMessageSize(-1);
- std::shared_ptr<::grpc::ChannelCredentials> credentials;
- TF_RETURN_IF_ERROR(CredentialsFactory::CreateClientCredentials(
- dataset()->protocol_, &credentials));
- auto channel =
- ::grpc::CreateCustomChannel(worker_address, credentials, args);
- *stub = WorkerService::NewStub(channel);
- return Status::OK();
- }
-
- Status GetElementWithDeadline(TaskThread* task_handler,
- GetElementResponse* resp,
- int64 deadline_micros) {
- return RetryWithDeadline(
- [task_handler, resp] {
- GetElementRequest req;
- req.set_task_id(task_handler->task_id);
- grpc::ClientContext client_ctx;
- grpc::Status s =
- task_handler->worker_stub->GetElement(&client_ctx, req, resp);
- if (s.ok()) {
- return Status::OK();
- }
- return grpc_util::WrapError("Failed to fetch an element", s);
- },
- deadline_micros);
- }
-
- static bool ShouldRetryError(error::Code error_code) {
- // Retry all errors that could indicate preemption.
- return error_code == error::Code::UNAVAILABLE ||
- error_code == error::Code::CANCELLED ||
- error_code == error::Code::ABORTED;
- }
-
- static Status RetryWithDeadline(const std::function<Status()>& call,
- int64 deadline_micros) {
- Status s;
for (int num_retries = 0;; ++num_retries) {
- s = call();
- if (s.ok() || !ShouldRetryError(s.code())) {
+ Status s = RequestElement(task_handler, &resp);
+ if (s.ok()) {
+ break;
+ }
+ // Retry all errors that could indicate preemption.
+ if (!errors::IsUnavailable(s) && !errors::IsCancelled(s) &&
+ !errors::IsAborted(s)) {
return s;
}
+ {
+ mutex_lock l(mu_);
+ // If `UpdateTaskThreads` finds that the task has been cancelled, it
+ // will set end_of_sequence to `true`.
+ if (task_handler->end_of_sequence || cancelled_) {
+ return Status::OK();
+ }
+ }
const int64 now_micros = EnvTime::NowMicros();
if (now_micros > deadline_micros) {
return s;
@@ -421,6 +393,46 @@
: deadline_micros;
Env::Default()->SleepForMicroseconds(backoff_until - now_micros);
}
+
+ std::vector<Tensor> element;
+ if (!resp.end_of_sequence()) {
+ TF_RETURN_IF_ERROR(
+ service_util::Uncompress(resp.compressed_element(), &element));
+ }
+ mutex_lock l(mu_);
+ if (resp.end_of_sequence()) {
+ task_handler->end_of_sequence = true;
+ return Status::OK();
+ }
+ results_.push(std::move(element));
+ cv_.notify_all();
+ VLOG(3) << "Fetched an element for task id " << task_handler->task_id;
+ return Status::OK();
+ }
+
+ Status RequestElement(TaskThread* task_handler, GetElementResponse* resp) {
+ GetElementRequest req;
+ req.set_task_id(task_handler->task_id);
+ grpc::ClientContext client_ctx;
+ grpc::Status s =
+ task_handler->worker_stub->GetElement(&client_ctx, req, resp);
+ if (s.ok()) {
+ return Status::OK();
+ }
+ return grpc_util::WrapError("Failed to request an element", s);
+ }
+
+ Status CreateWorkerStub(const std::string& worker_address,
+ std::unique_ptr<WorkerService::Stub>* stub) {
+ ::grpc::ChannelArguments args;
+ args.SetMaxReceiveMessageSize(-1);
+ std::shared_ptr<::grpc::ChannelCredentials> credentials;
+ TF_RETURN_IF_ERROR(CredentialsFactory::CreateClientCredentials(
+ dataset()->protocol_, &credentials));
+ auto channel =
+ ::grpc::CreateCustomChannel(worker_address, credentials, args);
+ *stub = WorkerService::NewStub(channel);
+ return Status::OK();
}
mutex mu_;
@@ -440,9 +452,8 @@
int64 job_id_;
std::shared_ptr<::grpc::ChannelCredentials> credentials_;
int64 num_unfinished_tasks_ TF_GUARDED_BY(mu_) = 0;
- // Map from task id to task info.
- absl::flat_hash_map<int64, TaskInfo> tasks_ TF_GUARDED_BY(mu_);
+ bool job_finished_ = false;
// Must come second to last so that task threads are joined before
// destroying other fields.
absl::flat_hash_map<int64, std::unique_ptr<TaskThread>> task_threads_
@@ -455,12 +466,18 @@
const tstring address_;
const tstring protocol_;
const int64 max_outstanding_requests_;
+ const int64 task_refresh_interval_ms_;
const DataTypeVector output_types_;
const std::vector<PartialTensorShape> output_shapes_;
};
DataServiceDatasetOp::DataServiceDatasetOp(OpKernelConstruction* ctx)
: DatasetOpKernel(ctx) {
+ OP_REQUIRES_OK(ctx, ctx->GetAttr(kTaskRefreshIntervalHintMs,
+ &task_refresh_interval_hint_ms_));
+ if (task_refresh_interval_hint_ms_ == model::kAutotune) {
+ task_refresh_interval_hint_ms_ = kDefaultTaskRefreshIntervalMs;
+ }
OP_REQUIRES_OK(ctx, ctx->GetAttr(kOutputTypes, &output_types_));
OP_REQUIRES_OK(ctx, ctx->GetAttr(kOutputShapes, &output_shapes_));
}
@@ -488,7 +505,8 @@
model::kAutotune));
*output = new Dataset(ctx, address, protocol, max_outstanding_requests,
- output_types_, output_shapes_);
+ task_refresh_interval_hint_ms_, output_types_,
+ output_shapes_);
}
REGISTER_KERNEL_BUILDER(Name("DataServiceDataset").Device(DEVICE_CPU),
diff --git a/tensorflow/core/kernels/data/experimental/data_service_dataset_op.h b/tensorflow/core/kernels/data/experimental/data_service_dataset_op.h
index 1154741..d51cb8c 100644
--- a/tensorflow/core/kernels/data/experimental/data_service_dataset_op.h
+++ b/tensorflow/core/kernels/data/experimental/data_service_dataset_op.h
@@ -28,6 +28,8 @@
static constexpr const char* const kProtocol = "protocol";
static constexpr const char* const kMaxOutstandingRequests =
"max_outstanding_requests";
+ static constexpr const char* const kTaskRefreshIntervalHintMs =
+ "task_refresh_interval_hint_ms";
static constexpr const char* const kOutputTypes = "output_types";
static constexpr const char* const kOutputShapes = "output_shapes";
@@ -39,6 +41,7 @@
private:
class Dataset;
+ int64 task_refresh_interval_hint_ms_;
DataTypeVector output_types_;
std::vector<PartialTensorShape> output_shapes_;
};
diff --git a/tensorflow/core/kernels/data/experimental/snapshot_util.cc b/tensorflow/core/kernels/data/experimental/snapshot_util.cc
index 6ff4476..3ad1345 100644
--- a/tensorflow/core/kernels/data/experimental/snapshot_util.cc
+++ b/tensorflow/core/kernels/data/experimental/snapshot_util.cc
@@ -15,10 +15,14 @@
#include "tensorflow/core/kernels/data/experimental/snapshot_util.h"
+#include <queue>
+
#include "absl/memory/memory.h"
#include "tensorflow/core/common_runtime/dma_helper.h"
+#include "tensorflow/core/framework/dataset.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/tensor.pb.h"
+#include "tensorflow/core/kernels/data/name_utils.h"
#include "tensorflow/core/lib/io/buffered_inputstream.h"
#include "tensorflow/core/lib/io/random_inputstream.h"
#include "tensorflow/core/lib/io/snappy/snappy_inputbuffer.h"
@@ -219,6 +223,197 @@
return (*out_reader)->Initialize(env);
}
+class Reader::Dataset : public DatasetBase {
+ public:
+ explicit Dataset(const std::string& filename, const std::string& compression,
+ const int64 version, const DataTypeVector& dtypes,
+ const std::vector<PartialTensorShape>& shapes,
+ DatasetContext::Params params)
+ : DatasetBase(DatasetContext(std::move(params))),
+ filename_(filename),
+ compression_(compression),
+ version_(version),
+ dtypes_(dtypes),
+ shapes_(shapes) {}
+
+ const DataTypeVector& output_dtypes() const override { return dtypes_; }
+
+ const std::vector<PartialTensorShape>& output_shapes() const override {
+ return shapes_;
+ }
+
+ std::string DebugString() const override {
+ return "snapshot_util::Reader::Dataset";
+ }
+
+ Status CheckExternalState() const override { return Status::OK(); }
+
+ protected:
+ Status AsGraphDefInternal(SerializationContext* ctx,
+ DatasetGraphDefBuilder* b,
+ Node** node) const override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ std::unique_ptr<IteratorBase> MakeIteratorInternal(
+ const string& prefix) const override {
+ return absl::make_unique<Iterator>(Iterator::Params{
+ this, name_utils::IteratorPrefix(node_name(), prefix)});
+ }
+
+ private:
+ std::string filename_;
+ std::string compression_;
+ int64 version_;
+ DataTypeVector dtypes_;
+ std::vector<PartialTensorShape> shapes_;
+
+ class Iterator : public DatasetIterator<Dataset> {
+ public:
+ explicit Iterator(const Params& params)
+ : DatasetIterator<Dataset>(params) {}
+
+ Status Initialize(IteratorContext* ctx) override {
+ return Reader::Create(ctx->env(), dataset()->filename_,
+ dataset()->compression_, dataset()->version_,
+ dataset()->dtypes_, &reader_);
+ }
+
+ protected:
+ Status GetNextInternal(IteratorContext* ctx,
+ std::vector<Tensor>* out_tensors,
+ bool* end_of_sequence) override {
+ *end_of_sequence = false;
+ Status s = reader_->ReadTensors(out_tensors);
+ if (errors::IsOutOfRange(s)) {
+ *end_of_sequence = true;
+ return Status::OK();
+ }
+ return s;
+ }
+
+ Status SaveInternal(SerializationContext* ctx,
+ IteratorStateWriter* writer) override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ Status RestoreInternal(IteratorContext* ctx,
+ IteratorStateReader* reader) override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ private:
+ std::unique_ptr<Reader> reader_;
+ };
+};
+
+class Reader::NestedDataset : public DatasetBase {
+ public:
+ explicit NestedDataset(std::vector<DatasetBase*> datasets,
+ DatasetContext::Params params)
+ : DatasetBase(DatasetContext(std::move(params))), datasets_(datasets) {
+ dtypes_.push_back(DT_VARIANT);
+ gtl::InlinedVector<int64, 1> element_dim_sizes;
+ element_dim_sizes.push_back(1);
+ partial_shapes_.emplace_back(element_dim_sizes);
+ }
+
+ const DataTypeVector& output_dtypes() const override { return dtypes_; }
+
+ const std::vector<PartialTensorShape>& output_shapes() const override {
+ return partial_shapes_;
+ }
+
+ std::string DebugString() const override {
+ return "snapshot_util::Reader::NestedDataset";
+ }
+
+ Status CheckExternalState() const override { return Status::OK(); }
+
+ protected:
+ Status AsGraphDefInternal(SerializationContext* ctx,
+ DatasetGraphDefBuilder* b,
+ Node** node) const override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ std::unique_ptr<IteratorBase> MakeIteratorInternal(
+ const string& prefix) const override {
+ return absl::make_unique<Iterator>(Iterator::Params{
+ this, name_utils::IteratorPrefix(node_name(), prefix)});
+ }
+
+ private:
+ std::vector<DatasetBase*> datasets_;
+ DataTypeVector dtypes_;
+ std::vector<PartialTensorShape> partial_shapes_;
+
+ class Iterator : public DatasetIterator<NestedDataset> {
+ public:
+ explicit Iterator(const Params& params)
+ : DatasetIterator<NestedDataset>(params), index_(0) {}
+
+ protected:
+ Status GetNextInternal(IteratorContext* ctx,
+ std::vector<Tensor>* out_tensors,
+ bool* end_of_sequence) override {
+ *end_of_sequence = dataset()->datasets_.size() == index_;
+ if (!*end_of_sequence) {
+ Tensor tensor(DT_VARIANT, TensorShape({}));
+
+ TF_RETURN_IF_ERROR(
+ StoreDatasetInVariantTensor(dataset()->datasets_[index_], &tensor));
+ out_tensors->clear();
+ out_tensors->push_back(std::move(tensor));
+
+ index_++;
+ }
+ return Status::OK();
+ }
+
+ Status SaveInternal(SerializationContext* ctx,
+ IteratorStateWriter* writer) override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ Status RestoreInternal(IteratorContext* ctx,
+ IteratorStateReader* reader) override {
+ // TODO(frankchn): Implement for serialization and checkpointing.
+ return Status::OK();
+ }
+
+ private:
+ int64 index_;
+ };
+};
+
+Status Reader::MakeNestedDataset(Env* env,
+ const std::vector<std::string>& filenames,
+ const string& compression_type, int version,
+ const DataTypeVector& dtypes,
+ const std::vector<PartialTensorShape>& shapes,
+ DatasetBase** output) {
+ std::vector<DatasetBase*> datasets;
+
+ datasets.reserve(filenames.size());
+ for (const auto& filename : filenames) {
+ datasets.push_back(
+ new Dataset(filename, compression_type, version, dtypes, shapes,
+ DatasetContext::Params({"snapshot_util::Reader::Dataset",
+ "snapshot_util_reader_Dataset"})));
+ }
+
+ *output = new NestedDataset(
+ datasets, DatasetContext::Params({"snapshot_util::Reader::NestedDataset",
+ "snapshot_util_reader_NestedDataset"}));
+ return Status::OK();
+}
+
Reader::Reader(const std::string& filename, const string& compression_type,
int version, const DataTypeVector& dtypes)
: filename_(filename),
diff --git a/tensorflow/core/kernels/data/experimental/snapshot_util.h b/tensorflow/core/kernels/data/experimental/snapshot_util.h
index 3816525..dd15c59 100644
--- a/tensorflow/core/kernels/data/experimental/snapshot_util.h
+++ b/tensorflow/core/kernels/data/experimental/snapshot_util.h
@@ -16,6 +16,7 @@
#ifndef TENSORFLOW_CORE_KERNELS_DATA_EXPERIMENTAL_SNAPSHOT_UTIL_H_
#define TENSORFLOW_CORE_KERNELS_DATA_EXPERIMENTAL_SNAPSHOT_UTIL_H_
+#include "tensorflow/core/framework/dataset.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/io/compression.h"
@@ -119,6 +120,18 @@
const DataTypeVector& dtypes,
std::unique_ptr<Reader>* out_reader);
+ // Returns a nested dataset for a set of given snapshot file names.
+ //
+ // This function takes a vector of snapshot files, and returns a nested
+ // dataset. Each element within the nested dataset is itself a dataset, and
+ // contains all the elements written out to each individual snapshot file.
+ static Status MakeNestedDataset(Env* env,
+ const std::vector<std::string>& filenames,
+ const string& compression_type, int version,
+ const DataTypeVector& dtypes,
+ const std::vector<PartialTensorShape>& shapes,
+ DatasetBase** output);
+
Status ReadTensors(std::vector<Tensor>* read_tensors);
private:
@@ -150,6 +163,9 @@
int num_simple_ = 0;
int num_complex_ = 0;
std::vector<bool> simple_tensor_mask_; // true for simple, false for complex.
+
+ class Dataset;
+ class NestedDataset;
};
Status WriteMetadataFile(const string& hash_dir,
diff --git a/tensorflow/core/kernels/data/random_seed_ops.cc b/tensorflow/core/kernels/data/random_seed_ops.cc
index ea40308..3b66f9e 100644
--- a/tensorflow/core/kernels/data/random_seed_ops.cc
+++ b/tensorflow/core/kernels/data/random_seed_ops.cc
@@ -21,6 +21,7 @@
#include "tensorflow/core/lib/random/philox_random.h"
#include "tensorflow/core/lib/random/random.h"
#include "tensorflow/core/lib/random/random_distributions.h"
+#include "tensorflow/core/platform/errors.h"
namespace tensorflow {
namespace data {
@@ -28,8 +29,6 @@
const char kAnonymousRandomSeedGenerator[] = "AnonymousRandomSeedGenerator";
const char kNumRandomSamples[] = "num_random_samples";
-const char kFixedSeedGenerator[] = "FixedSeedGenerator";
-const char kRandomSeedGenerator[] = "RandomSeedGenerator";
const char kSeedGenerator[] = "SeedGenerator";
const char kSeed[] = "seed";
const char kSeed2[] = "seed2";
@@ -37,27 +36,15 @@
} // namespace
-int64 SeedGenerator::num_random_samples() {
- tf_shared_lock l(mu_);
- return num_random_samples_;
-}
-
-void SeedGenerator::set_num_random_samples(int64 num_random_samples) {
- mutex_lock l(mu_);
- num_random_samples_ = num_random_samples;
-}
-
-string FixedSeedGenerator::DebugString() const { return kFixedSeedGenerator; }
+string SeedGeneratorManager::DebugString() const { return kSeedGenerator; }
void FixedSeedGenerator::GenerateSeeds(int64* seed1, int64* seed2) {
mutex_lock l(mu_);
num_random_samples_++;
- *seed1 = seed_;
- *seed2 = seed2_;
+ *seed1 = seeds_.seed();
+ *seed2 = seeds_.seed2();
}
-string RandomSeedGenerator::DebugString() const { return kRandomSeedGenerator; }
-
void RandomSeedGenerator::GenerateSeeds(int64* seed1, int64* seed2) {
mutex_lock l(mu_);
num_random_samples_++;
@@ -69,7 +56,7 @@
void RandomSeedGenerator::Reset() {
mutex_lock l(mu_);
// Reset the generators based on the current seeds.
- parent_generator_ = random::PhiloxRandom(seed_, seed2_);
+ parent_generator_ = random::PhiloxRandom(seeds_.seed(), seeds_.seed2());
generator_ =
random::SingleSampleAdapter<random::PhiloxRandom>(&parent_generator_);
generator_.Skip(num_random_samples_);
@@ -77,29 +64,17 @@
AnonymousSeedGeneratorHandleOp::AnonymousSeedGeneratorHandleOp(
OpKernelConstruction* ctx)
- : AnonymousResourceOp<SeedGenerator>(ctx) {}
+ : AnonymousResourceOp<SeedGeneratorManager>(ctx) {}
void AnonymousSeedGeneratorHandleOp::Compute(OpKernelContext* ctx) {
int64 seed;
OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed, &seed));
int64 seed2;
OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed2, &seed2));
- if (seed == 0 && seed2 == 0) {
- seed = random::New64();
- seed2 = random::New64();
- }
- seed_ = seed;
- seed2_ = seed2;
-
- // TODO(b/151115950): Remove this case when the forward compatibility window
- // expires.
- if (ctx->op_kernel().def().op() == kAnonymousRandomSeedGenerator) {
- reshuffle_ = true;
- } else {
- OP_REQUIRES_OK(ctx,
- ParseScalarArgument<bool>(ctx, kReshuffle, &reshuffle_));
- }
- AnonymousResourceOp<SeedGenerator>::Compute(ctx);
+ // Seeds will be consumed by `CreateResource`, which is called via `Compute`.
+ seeds_ = absl::make_unique<RandomSeeds>(seed, seed2);
+ OP_REQUIRES_OK(ctx, ParseScalarArgument<bool>(ctx, kReshuffle, &reshuffle_));
+ AnonymousResourceOp<SeedGeneratorManager>::Compute(ctx);
}
std::string AnonymousSeedGeneratorHandleOp::name() { return kSeedGenerator; }
@@ -107,12 +82,13 @@
Status AnonymousSeedGeneratorHandleOp::CreateResource(
OpKernelContext* ctx, std::unique_ptr<FunctionLibraryDefinition> flib_def,
std::unique_ptr<ProcessFunctionLibraryRuntime> pflr,
- FunctionLibraryRuntime* lib, SeedGenerator** resource) {
+ FunctionLibraryRuntime* lib, SeedGeneratorManager** manager) {
if (reshuffle_) {
- *resource = new RandomSeedGenerator(seed_, seed2_);
+ *manager = new SeedGeneratorManager(new RandomSeedGenerator(*seeds_));
} else {
- *resource = new FixedSeedGenerator(seed_, seed2_);
+ *manager = new SeedGeneratorManager(new FixedSeedGenerator(*seeds_));
}
+ seeds_ = nullptr;
return Status::OK();
}
@@ -137,6 +113,9 @@
REGISTER_KERNEL_BUILDER(Name("DeleteRandomSeedGenerator").Device(DEVICE_CPU),
DeleteSeedGeneratorOp);
+REGISTER_KERNEL_BUILDER(Name("DummySeedGenerator").Device(DEVICE_CPU),
+ DummyResourceOp<SeedGenerator>);
+
} // namespace
} // namespace data
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/data/random_seed_ops.h b/tensorflow/core/kernels/data/random_seed_ops.h
index 54332c7..e013cc0 100644
--- a/tensorflow/core/kernels/data/random_seed_ops.h
+++ b/tensorflow/core/kernels/data/random_seed_ops.h
@@ -25,51 +25,102 @@
namespace tensorflow {
namespace data {
+// Represents a pair of random seeds. By TensorFlow convention, if both seeds
+// are 0, then pseudo-random values are used instead.
+class RandomSeeds {
+ public:
+ RandomSeeds(int64 seed, int64 seed2)
+ : input_seed_(seed),
+ input_seed2_(seed2),
+ seed_((seed | seed2) == 0 ? random::New64() : seed),
+ seed2_((seed | seed2) == 0 ? random::New64() : seed2) {}
+
+ int64 input_seed() const { return input_seed_; }
+ int64 input_seed2() const { return input_seed2_; }
+ int64 seed() const { return seed_; }
+ int64 seed2() const { return seed2_; }
+
+ private:
+ const int64 input_seed_;
+ const int64 input_seed2_;
+ const int64 seed_;
+ const int64 seed2_;
+};
+
// Base class for seed generator resources. Subclasses customize how seeds are
// generated.
-class SeedGenerator : public ResourceBase {
+class SeedGenerator {
public:
+ virtual ~SeedGenerator() {}
+
+ virtual int64 seed() const = 0;
+ virtual int64 seed2() const = 0;
+ virtual bool reshuffle_each_iteration() const = 0;
+
virtual void GenerateSeeds(int64* seed1, int64* seed2) = 0;
virtual void Reset() = 0;
- virtual int64 num_random_samples();
- virtual void set_num_random_samples(int64 num_random_samples);
+ virtual int64 num_random_samples() const {
+ tf_shared_lock l(mu_);
+ return num_random_samples_;
+ }
+ virtual void set_num_random_samples(int64 num_random_samples) {
+ mutex_lock l(mu_);
+ num_random_samples_ = num_random_samples;
+ }
protected:
- mutex mu_;
+ mutable mutex mu_;
int64 num_random_samples_ TF_GUARDED_BY(mu_) = 0;
};
+// A resource wrapping a shared instance of a seed generator.
+class SeedGeneratorManager : public ResourceBase {
+ public:
+ explicit SeedGeneratorManager(SeedGenerator* seed_generator)
+ : seed_generator_(seed_generator) {}
+
+ std::string DebugString() const override;
+
+ std::shared_ptr<SeedGenerator> get() { return seed_generator_; }
+
+ private:
+ std::shared_ptr<SeedGenerator> seed_generator_;
+};
+
// Always generates the specified seed values.
class FixedSeedGenerator : public SeedGenerator {
public:
- FixedSeedGenerator(int64 seed, int64 seed2) : seed_(seed), seed2_(seed2) {}
+ explicit FixedSeedGenerator(RandomSeeds seeds) : seeds_(std::move(seeds)) {}
- std::string DebugString() const override;
+ int64 seed() const override { return seeds_.seed(); }
+ int64 seed2() const override { return seeds_.seed(); }
+ bool reshuffle_each_iteration() const override { return false; }
+
void GenerateSeeds(int64* seed1, int64* seed2) override;
void Reset() override {}
private:
- const int64 seed_;
- const int64 seed2_;
+ const RandomSeeds seeds_;
};
// Generates different (but deterministically chosen) seed values.
class RandomSeedGenerator : public SeedGenerator {
public:
- RandomSeedGenerator(int64 seed, int64 seed2)
- : seed_(seed),
- seed2_(seed2),
- parent_generator_(seed, seed2),
+ explicit RandomSeedGenerator(RandomSeeds seeds)
+ : seeds_(std::move(seeds)),
+ parent_generator_(seeds_.seed(), seeds_.seed2()),
generator_(&parent_generator_) {}
- std::string DebugString() const override;
+ int64 seed() const override { return seeds_.seed(); }
+ int64 seed2() const override { return seeds_.seed2(); }
+ bool reshuffle_each_iteration() const override { return true; }
+
void GenerateSeeds(int64* seed1, int64* seed2) override;
void Reset() override;
private:
- const int64 seed_;
- const int64 seed2_;
+ const RandomSeeds seeds_;
random::PhiloxRandom parent_generator_ TF_GUARDED_BY(mu_);
random::SingleSampleAdapter<random::PhiloxRandom> generator_
TF_GUARDED_BY(mu_);
@@ -78,7 +129,7 @@
// Creates an instance of seed generator resource and transfers ownership
// to the caller.
class AnonymousSeedGeneratorHandleOp
- : public AnonymousResourceOp<SeedGenerator> {
+ : public AnonymousResourceOp<SeedGeneratorManager> {
public:
explicit AnonymousSeedGeneratorHandleOp(OpKernelConstruction* ctx);
void Compute(OpKernelContext* ctx) override;
@@ -89,10 +140,9 @@
std::unique_ptr<FunctionLibraryDefinition> flib_def,
std::unique_ptr<ProcessFunctionLibraryRuntime> pflr,
FunctionLibraryRuntime* lib,
- SeedGenerator** resource) override;
+ SeedGeneratorManager** manager) override;
- int64 seed_;
- int64 seed2_;
+ std::unique_ptr<RandomSeeds> seeds_ = nullptr;
bool reshuffle_;
};
diff --git a/tensorflow/core/kernels/data/shuffle_dataset_op.cc b/tensorflow/core/kernels/data/shuffle_dataset_op.cc
index e0312b5..852ba23 100644
--- a/tensorflow/core/kernels/data/shuffle_dataset_op.cc
+++ b/tensorflow/core/kernels/data/shuffle_dataset_op.cc
@@ -68,34 +68,10 @@
constexpr char kSize[] = "size";
constexpr char kSeedGenerator[] = "SeedGenerator";
constexpr char kTFData[] = "tf_data";
-constexpr char kDSNumRandomSamples[] = "ds_num_random_samples";
-constexpr char kFixedSeedDatasetPrefix[] = "FixedSeed";
-constexpr char kDatasetPrefix[] = "Dataset";
-constexpr char kDatasetV2Prefix[] = "DatasetV2";
-constexpr char kShuffleDataset[] = "ShuffleDataset";
-
-namespace {
-class Seeds {
- public:
- Seeds(int64 seed, int64 seed2) {
- input_seed_ = seed;
- input_seed2_ = seed2;
- seed_ = seed;
- seed2_ = seed2;
- // By TensorFlow convention, if both seeds are 0, then shuffling should be
- // seeded non-deterministically.
- if (seed == 0 && seed2 == 0) {
- seed_ = random::New64();
- seed2_ = random::New64();
- }
- }
-
- int64 input_seed_;
- int64 input_seed2_;
- int64 seed_;
- int64 seed2_;
-};
-} // namespace
+constexpr char kEpochNumRandomSamples[] = "epoch_num_random_samples";
+constexpr char kShuffleDatasetV1[] = "ShuffleDataset";
+constexpr char kShuffleDatasetV2[] = "ShuffleDatasetV2";
+constexpr char kShuffleDatasetV3[] = "ShuffleDatasetV3";
ShuffleDatasetOpBase::ShuffleDatasetOpBase(OpKernelConstruction* ctx)
: UnaryDatasetOpKernel(ctx) {}
@@ -104,10 +80,12 @@
class ShuffleDatasetOpBase::ShuffleDatasetBase : public DatasetBase {
public:
ShuffleDatasetBase(OpKernelContext* ctx, const DatasetBase* input,
- int64 buffer_size, int64 count)
+ int64 buffer_size,
+ std::shared_ptr<SeedGenerator> seed_generator, int64 count)
: DatasetBase(DatasetContext(ctx)),
input_(input),
buffer_size_(buffer_size),
+ seed_generator_(std::move(seed_generator)),
count_(count),
traceme_metadata_(
{{"buffer_size",
@@ -117,6 +95,8 @@
~ShuffleDatasetBase() override { input_->Unref(); }
+ virtual string op_type() const = 0;
+
const DataTypeVector& output_dtypes() const override {
return input_->output_dtypes();
}
@@ -139,37 +119,40 @@
return input_->CheckExternalState();
}
- protected:
- // Adds the seeds to the given graphdef builder. `preserve_random_seeds`
- // controls whether to add the input seeds or the resolved seeds.
- Status AddSeeds(Seeds seeds, bool preserve_random_seeds,
- DatasetGraphDefBuilder* b, Node** seed, Node** seed2) const {
- int64 seed_to_add = preserve_random_seeds ? seeds.input_seed_ : seeds.seed_;
- int64 seed2_to_add =
- preserve_random_seeds ? seeds.input_seed2_ : seeds.seed2_;
- TF_RETURN_IF_ERROR(b->AddScalar(seed_to_add, seed));
- TF_RETURN_IF_ERROR(b->AddScalar(seed2_to_add, seed2));
- return Status::OK();
+ string DebugString() const override {
+ name_utils::DatasetDebugStringParams params;
+ params.set_args(buffer_size_, seed_generator_->seed(),
+ seed_generator_->seed2(), count_);
+ return name_utils::DatasetDebugString(op_type(), params);
}
- template <class T>
- class Iterator : public DatasetIterator<T> {
+ std::unique_ptr<IteratorBase> MakeIteratorInternal(
+ const string& prefix) const override {
+ return absl::make_unique<Iterator>(
+ Iterator::Params{this, name_utils::IteratorPrefix(op_type(), prefix)},
+ seed_generator_.get());
+ }
+
+ protected:
+ class Iterator : public DatasetIterator<ShuffleDatasetBase> {
public:
- explicit Iterator(const typename DatasetIterator<T>::Params& params,
- int64 seed, int64 seed2)
- : DatasetIterator<T>(params),
- seed_(seed),
- seed2_(seed2),
- input_impl_(nullptr),
- epoch_(0),
- num_elements_(0),
- parent_generator_(seed, seed2),
+ explicit Iterator(const Params& params, SeedGenerator* seed_generator)
+ : DatasetIterator<ShuffleDatasetBase>(params),
+ seed_generator_(seed_generator),
+ parent_generator_(seed_generator->seed(), seed_generator->seed2()),
generator_(&parent_generator_) {
buffer_ = absl::make_unique<std::vector<Tensor>[]>(
params.dataset->buffer_size_);
slices_.push_back(absl::make_unique<Slice>(0, 0));
}
+ Status Initialize(IteratorContext* ctx) override {
+ mutex_lock l(mu_);
+ seed_generator_->GenerateSeeds(&seed_, &seed2_);
+ ResetRngs();
+ return Status::OK();
+ }
+
Status GetNextInternal(IteratorContext* ctx,
std::vector<Tensor>* out_tensors,
bool* end_of_sequence) override {
@@ -283,6 +266,9 @@
IteratorStateWriter* writer) override {
mutex_lock l(mu_);
// Save state needed to restore the random number generators.
+ TF_RETURN_IF_ERROR(
+ writer->WriteScalar(full_name(kEpochNumRandomSamples),
+ seed_generator_->num_random_samples()));
TF_RETURN_IF_ERROR(writer->WriteScalar(this->full_name(kNumRandomSamples),
num_random_samples_));
TF_RETURN_IF_ERROR(writer->WriteScalar(this->full_name(kSeed), seed_));
@@ -337,6 +323,11 @@
IteratorStateReader* reader) override {
mutex_lock l(mu_);
// Restore the random number generators.
+ int64 num_random_samples;
+ TF_RETURN_IF_ERROR(reader->ReadScalar(full_name(kEpochNumRandomSamples),
+ &num_random_samples));
+ seed_generator_->set_num_random_samples(num_random_samples);
+ seed_generator_->Reset();
TF_RETURN_IF_ERROR(reader->ReadScalar(this->full_name(kNumRandomSamples),
&num_random_samples_));
TF_RETURN_IF_ERROR(reader->ReadScalar(this->full_name(kSeed), &seed_));
@@ -402,10 +393,6 @@
return this->dataset()->traceme_metadata_;
}
- mutex mu_;
- int64 seed_ TF_GUARDED_BY(mu_);
- int64 seed2_ TF_GUARDED_BY(mu_);
-
private:
// Used to represent slices of `buffer_` that belong to different epochs.
// The invariant maintained by the implementation is: `start` <= `end`.
@@ -426,10 +413,14 @@
return out;
}
+ mutex mu_;
+ SeedGenerator* const seed_generator_ TF_GUARDED_BY(mu_); // Not owned.
std::unique_ptr<std::vector<Tensor>[]> buffer_ TF_GUARDED_BY(mu_);
- std::unique_ptr<IteratorBase> input_impl_ TF_GUARDED_BY(mu_);
- int64 epoch_ TF_GUARDED_BY(mu_);
- int64 num_elements_ TF_GUARDED_BY(mu_);
+ std::unique_ptr<IteratorBase> input_impl_ TF_GUARDED_BY(mu_) = nullptr;
+ int64 epoch_ TF_GUARDED_BY(mu_) = 0;
+ int64 num_elements_ TF_GUARDED_BY(mu_) = 0;
+ int64 seed_ TF_GUARDED_BY(mu_) = 0;
+ int64 seed2_ TF_GUARDED_BY(mu_) = 0;
// Indices into `buffer_` indicating which data belongs to which epoch.
// The slice at the front of the deque references data from the earliest
// buffered epoch. It is an invariant that all slices reference
@@ -444,135 +435,59 @@
const DatasetBase* const input_;
const int64 buffer_size_;
+ const std::shared_ptr<SeedGenerator> seed_generator_;
// The number of epochs to run for. Normally this is just 1, but sometimes we
// fuse shuffle and repeat together, and make the shuffle dataset op
// responsible for repeating as well.
const int64 count_;
const TraceMeMetadata traceme_metadata_;
-};
+}; // ShuffleDatasetBase
+// This version of memory dataset has an exclusive ownership of the seed
+// generator resource. It supports sharing of the seed generator across
+// different iterations of the `repeat` transformation but not across different
+// iterators.
class ShuffleDatasetOp::Dataset : public ShuffleDatasetBase {
public:
Dataset(OpKernelContext* ctx, const DatasetBase* input, int64 buffer_size,
- Seeds seeds, int64 count, bool reshuffle_each_iteration)
- : ShuffleDatasetBase(ctx, input, buffer_size, count),
- seeds_(seeds),
- reshuffle_each_iteration_(reshuffle_each_iteration) {}
+ int64 count, RandomSeeds&& seeds, SeedGeneratorManager* manager,
+ ResourceHandle&& resource_handle)
+ : ShuffleDatasetBase(ctx, input, buffer_size, manager->get(), count),
+ manager_(manager),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()),
+ seeds_(std::move(seeds)) {}
- string DebugString() const override {
- name_utils::DatasetDebugStringParams params;
- params.dataset_prefix = kDatasetPrefix;
- params.set_args(buffer_size_, seeds_.seed_, seeds_.seed2_);
- return name_utils::DatasetDebugString(kDatasetType, params);
+ ~Dataset() override {
+ manager_->Unref();
+ Status s = resource_mgr_->Delete<SeedGeneratorManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete RNG resource: " << s.ToString();
+ }
}
- std::unique_ptr<IteratorBase> MakeIteratorInternal(
- const string& prefix) const override {
- return absl::make_unique<Iterator>(
- Iterator::Params{this,
- name_utils::IteratorPrefix(kDatasetType, prefix)},
- seeds_.seed_, seeds_.seed2_);
- }
+ string op_type() const override { return kDatasetType; }
protected:
- class Iterator : public ShuffleDatasetBase::Iterator<Dataset> {
- public:
- Iterator(const Params& params, int64 seed, int64 seed2)
- : ShuffleDatasetBase::Iterator<Dataset>(params, seed, seed2) {}
-
- ~Iterator() override { seed_generator_->Unref(); }
-
- Status Initialize(IteratorContext* ctx) override {
- // Firstly, lookup or create a seed generator from the IteratorResource
- // resource_mgr.
- ResourceMgr* mgr = ctx->resource_mgr();
- SeedGenerator* seed_generator;
- const string name = strings::StrCat(
- prefix(), name_utils::kDelimiter, dataset()->type_string(),
- name_utils::kDelimiter, kSeedGenerator);
-
- int64 dataset_seed, dataset_seed2;
- {
- tf_shared_lock l(mu_);
- // Ideally we'd like to hold this lock in the LookupOrCreate method,
- // but that trips up our Deadlock detection code.
- dataset_seed = seed_;
- dataset_seed2 = seed2_;
- }
- TF_RETURN_IF_ERROR(mgr->LookupOrCreate<SeedGenerator>(
- kTFData, name, &seed_generator,
- [this, dataset_seed, dataset_seed2](SeedGenerator** seed_generator) {
- // On the first iterator creation, use the original seeds from the
- // dataset to seed a `SeedGenerator` that will provide seeds
- // for subsequent repetitions of the same dataset.
- if (dataset()->reshuffle_each_iteration_) {
- *seed_generator =
- new RandomSeedGenerator(dataset_seed, dataset_seed2);
- } else {
- *seed_generator =
- new FixedSeedGenerator(dataset_seed, dataset_seed2);
- }
- return Status::OK();
- }));
- seed_generator_ = seed_generator;
- seed_generator_->GenerateSeeds(&seed_, &seed2_);
- mutex_lock l(mu_);
- ResetRngs();
- return Status::OK();
- }
-
- protected:
- std::shared_ptr<model::Node> CreateNode(
- IteratorContext* ctx, model::Node::Args args) const override {
- return model::MakeKnownRatioNode(std::move(args),
- /*ratio=*/1);
- }
-
- Status SaveInternal(SerializationContext* ctx,
- IteratorStateWriter* writer) override {
- // Save RNG state of Dataset.
- TF_RETURN_IF_ERROR(
- writer->WriteScalar(full_name(kDSNumRandomSamples),
- seed_generator_->num_random_samples()));
-
- // Save the Iterator.
- return ShuffleDatasetBase::Iterator<Dataset>::SaveInternal(ctx, writer);
- }
-
- Status RestoreInternal(IteratorContext* ctx,
- IteratorStateReader* reader) override {
- // Restore RNG state of Dataset.
- int64 num_random_samples;
- TF_RETURN_IF_ERROR(reader->ReadScalar(full_name(kDSNumRandomSamples),
- &num_random_samples));
- seed_generator_->set_num_random_samples(num_random_samples);
- seed_generator_->Reset();
-
- // Restore the Iterator.
- return ShuffleDatasetBase::Iterator<Dataset>::RestoreInternal(ctx,
- reader);
- }
-
- private:
- SeedGenerator* seed_generator_;
- };
-
Status AsGraphDefInternal(SerializationContext* ctx,
DatasetGraphDefBuilder* b,
Node** output) const override {
Node* input_graph_node = nullptr;
TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph_node));
- Node* buffer_size = nullptr;
- Node* seed = nullptr;
- Node* seed2 = nullptr;
+ Node* buffer_size_node = nullptr;
+ Node* seed_node = nullptr;
+ Node* seed2_node = nullptr;
AttrValue reshuffle_each_iteration;
- TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size));
- TF_RETURN_IF_ERROR(
- AddSeeds(seeds_, /*preserve_random_seeds=*/true, b, &seed, &seed2));
- b->BuildAttrValue(reshuffle_each_iteration_, &reshuffle_each_iteration);
+ TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size_node));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed(), &seed_node));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed2(), &seed2_node));
+ b->BuildAttrValue(seed_generator_->reshuffle_each_iteration(),
+ &reshuffle_each_iteration);
TF_RETURN_IF_ERROR(b->AddDataset(
- this, {input_graph_node, buffer_size, seed, seed2}, // Inputs
+ this,
+ {input_graph_node, buffer_size_node, seed_node, seed2_node}, // Inputs
{std::make_pair(kReshuffleEachIteration,
reshuffle_each_iteration)}, // Attrs
output));
@@ -580,92 +495,41 @@
}
private:
- const Seeds seeds_;
- const bool reshuffle_each_iteration_;
+ SeedGeneratorManager* const manager_; // Owned.
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
+ const RandomSeeds seeds_;
};
-// A shuffle dataset that uses an external seed generator resource to choose the
-// shuffle seeds for each iteration.
+// This version of shuffle dataset has a shared ownership of the seed generator
+// resource. It supports sharing of the generator state across different
+// iterations of the `repeat` transformation and also across different
+// iterators.
class ShuffleDatasetOp::DatasetV2 : public ShuffleDatasetBase {
public:
DatasetV2(OpKernelContext* ctx, const DatasetBase* input, int64 buffer_size,
- int64 count, SeedGenerator* seed_generator,
- std::unique_ptr<OwnedResourceHandle> handle)
- : ShuffleDatasetBase(ctx, input, buffer_size, count),
- seed_generator_(seed_generator),
- handle_(std::move(handle)) {}
+ int64 count, SeedGeneratorManager* manager,
+ ResourceHandle&& resource_handle, bool owns_resource)
+ : ShuffleDatasetBase(ctx, input, buffer_size, manager->get(), count),
+ manager_(manager),
+ owns_resource_(owns_resource),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()) {}
- ~DatasetV2() override { seed_generator_->Unref(); }
-
- string DebugString() const override {
- name_utils::DatasetDebugStringParams params;
- params.dataset_prefix = kDatasetV2Prefix;
- params.set_args(buffer_size_);
- return name_utils::DatasetDebugString(kDatasetType, params);
+ ~DatasetV2() override {
+ manager_->Unref();
+ if (owns_resource_) {
+ Status s = resource_mgr_->Delete<SeedGeneratorManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete RNG resource: " << s.ToString();
+ }
+ }
}
- Status CheckExternalState() const override {
- return errors::FailedPrecondition(
- DebugString(), " depends on random seed generator resource.");
- }
-
- std::unique_ptr<IteratorBase> MakeIteratorInternal(
- const string& prefix) const override {
- return absl::make_unique<Iterator>(
- Iterator::Params{this,
- name_utils::IteratorPrefix(kDatasetType, prefix)},
- seed_generator_);
- }
+ string op_type() const override { return kDatasetType; }
protected:
- class Iterator : public ShuffleDatasetBase::Iterator<DatasetV2> {
- public:
- Iterator(const Params& params, SeedGenerator* seed_generator)
- : ShuffleDatasetBase::Iterator<DatasetV2>(params, 0, 0),
- seed_generator_(seed_generator) {}
-
- Status Initialize(IteratorContext* ctx) override {
- mutex_lock l(mu_);
- seed_generator_->GenerateSeeds(&seed_, &seed2_);
- ResetRngs();
- return Status::OK();
- }
-
- protected:
- std::shared_ptr<model::Node> CreateNode(
- IteratorContext* ctx, model::Node::Args args) const override {
- return model::MakeKnownRatioNode(std::move(args), /*ratio=*/1);
- }
-
- Status SaveInternal(SerializationContext* ctx,
- IteratorStateWriter* writer) override {
- // Save state of the seed generator.
- TF_RETURN_IF_ERROR(
- writer->WriteScalar(full_name(kDSNumRandomSamples),
- seed_generator_->num_random_samples()));
-
- // Save the tterator state.
- return ShuffleDatasetBase::Iterator<DatasetV2>::SaveInternal(ctx, writer);
- }
-
- Status RestoreInternal(IteratorContext* ctx,
- IteratorStateReader* reader) override {
- // Restore state of the seed generator.
- int64 num_random_samples;
- TF_RETURN_IF_ERROR(reader->ReadScalar(full_name(kDSNumRandomSamples),
- &num_random_samples));
- seed_generator_->set_num_random_samples(num_random_samples);
- seed_generator_->Reset();
-
- // Restore the iterator state.
- return ShuffleDatasetBase::Iterator<DatasetV2>::RestoreInternal(ctx,
- reader);
- }
-
- private:
- SeedGenerator* seed_generator_;
- };
-
Status AsGraphDefInternal(SerializationContext* ctx,
DatasetGraphDefBuilder* b,
Node** output) const override {
@@ -675,7 +539,7 @@
TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size_node));
Node* resource_handle_node = nullptr;
Tensor handle(DT_RESOURCE, TensorShape({}));
- handle.scalar<ResourceHandle>()() = handle_->handle();
+ handle.scalar<ResourceHandle>()() = resource_handle_;
TF_RETURN_IF_ERROR(b->AddTensor(handle, &resource_handle_node));
TF_RETURN_IF_ERROR(b->AddDataset(
this,
@@ -686,33 +550,39 @@
}
private:
- SeedGenerator* seed_generator_ = nullptr;
- std::unique_ptr<OwnedResourceHandle> handle_;
+ SeedGeneratorManager* const manager_; // Owned.
+ const bool owns_resource_;
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
};
-// A dataset that uses the same fixed seed for all iterators created from it.
-// Used when `reshuffle_each_iteration` is false.
-// TODO(b/151115950): delete this class.
-class ShuffleDatasetOp::FixedSeedDataset : public ShuffleDatasetBase {
+// This version of shuffle dataset extends the functionality of DatasetV2 with
+// the ability to preserve seed generator configuration (i.e. initial seeds and
+// whether to reshuffle each iteration) across serialization of the dataset.
+class ShuffleDatasetOp::DatasetV3 : public ShuffleDatasetBase {
public:
- FixedSeedDataset(OpKernelContext* ctx, const DatasetBase* input,
- int64 buffer_size, Seeds seeds, int64 count)
- : ShuffleDatasetBase(ctx, input, buffer_size, count), seeds_(seeds) {}
+ DatasetV3(OpKernelContext* ctx, const DatasetBase* input, int64 buffer_size,
+ int64 count, RandomSeeds&& seeds, SeedGeneratorManager* manager,
+ ResourceHandle&& resource_handle, bool owns_resource)
+ : ShuffleDatasetBase(ctx, input, buffer_size, manager->get(), count),
+ manager_(manager),
+ owns_resource_(owns_resource),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()),
+ seeds_(std::move(seeds)) {}
- string DebugString() const override {
- name_utils::DatasetDebugStringParams params;
- params.dataset_prefix = kFixedSeedDatasetPrefix;
- params.set_args(buffer_size_, seeds_.seed_, seeds_.seed2_);
- return name_utils::DatasetDebugString(kDatasetType, params);
+ ~DatasetV3() override {
+ manager_->Unref();
+ if (owns_resource_) {
+ Status s = resource_mgr_->Delete<SeedGeneratorManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete RNG resource: " << s.ToString();
+ }
+ }
}
- std::unique_ptr<IteratorBase> MakeIteratorInternal(
- const string& prefix) const override {
- return absl::make_unique<ShuffleDatasetBase::Iterator<ShuffleDatasetBase>>(
- ShuffleDatasetBase::Iterator<ShuffleDatasetBase>::Params{
- this, name_utils::IteratorPrefix(kDatasetType, prefix)},
- seeds_.seed_, seeds_.seed2_);
- }
+ string op_type() const override { return kDatasetType; }
protected:
Status AsGraphDefInternal(SerializationContext* ctx,
@@ -720,30 +590,47 @@
Node** output) const override {
Node* input_graph_node = nullptr;
TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph_node));
- Node* buffer_size = nullptr;
- Node* seed = nullptr;
- Node* seed2 = nullptr;
+ Node* buffer_size_node = nullptr;
+ Node* seed_node = nullptr;
+ Node* seed2_node = nullptr;
+ TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size_node));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed(), &seed_node));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed2(), &seed2_node));
+ Node* resource_handle_node = nullptr;
+ Tensor handle(DT_RESOURCE, TensorShape({}));
+ handle.scalar<ResourceHandle>()() = resource_handle_;
+ TF_RETURN_IF_ERROR(b->AddTensor(handle, &resource_handle_node));
AttrValue reshuffle_each_iteration;
-
- TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size));
+ b->BuildAttrValue(seed_generator_->reshuffle_each_iteration(),
+ &reshuffle_each_iteration);
TF_RETURN_IF_ERROR(
- AddSeeds(seeds_, ctx->preserve_random_seeds(), b, &seed, &seed2));
- b->BuildAttrValue(false, &reshuffle_each_iteration);
- TF_RETURN_IF_ERROR(b->AddDataset(
- this, {input_graph_node, buffer_size, seed, seed2}, // Inputs
- {std::make_pair(kReshuffleEachIteration,
- reshuffle_each_iteration)}, // Attrs
- output));
+ b->AddDataset(this,
+ {input_graph_node, buffer_size_node, seed_node,
+ seed2_node, resource_handle_node}, // Inputs
+ {std::make_pair(kReshuffleEachIteration,
+ reshuffle_each_iteration)}, // Attrs
+ output));
return Status::OK();
}
private:
- const Seeds seeds_;
+ SeedGeneratorManager* const manager_; // Owned
+ const bool owns_resource_;
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
+ const RandomSeeds seeds_;
};
ShuffleDatasetOp::ShuffleDatasetOp(OpKernelConstruction* ctx)
- : ShuffleDatasetOpBase(ctx),
- op_version_(ctx->def().op() == kShuffleDataset ? 1 : 2) {
+ : ShuffleDatasetOpBase(ctx) {
+ auto& op_name = ctx->def().op();
+ if (op_name == kShuffleDatasetV3) {
+ op_version_ = 3;
+ } else if (op_name == kShuffleDatasetV2) {
+ op_version_ = 2;
+ } else if (op_name == kShuffleDatasetV1) {
+ op_version_ = 1;
+ }
if (ctx->HasAttr(kReshuffleEachIteration)) {
OP_REQUIRES_OK(
ctx, ctx->GetAttr(kReshuffleEachIteration, &reshuffle_each_iteration_));
@@ -760,71 +647,133 @@
errors::InvalidArgument("buffer_size must be greater than zero."));
int64 count = 1;
- if (op_version_ == 2) {
- SeedGenerator* seed_generator = nullptr;
- Status s = LookupResource(ctx, HandleFromInput(ctx, 2), &seed_generator);
+ static std::atomic<int64> resource_id_counter(0);
+ const string& container = ctx->resource_manager()->default_container();
+ auto name = strings::StrCat(ctx->op_kernel().name(), "/", kSeedGenerator, "_",
+ resource_id_counter.fetch_add(1));
+ if (op_version_ == 3) {
+ auto handle = HandleFromInput(ctx, 4);
+ SeedGeneratorManager* manager = nullptr;
+ Status s = ctx->resource_manager()->Lookup<SeedGeneratorManager>(
+ handle.container(), handle.name(), &manager);
+ int64 seed;
+ OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed, &seed));
+ int64 seed2;
+ OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed2, &seed2));
+ RandomSeeds seeds(seed, seed2);
+ bool owns_resource = false;
+ if (errors::IsNotFound(s)) {
+ OP_REQUIRES_OK(
+ ctx,
+ ctx->resource_manager()->LookupOrCreate<SeedGeneratorManager>(
+ container, name, &manager,
+ [reshuffle = reshuffle_each_iteration_,
+ &seeds](SeedGeneratorManager** manager) {
+ if (reshuffle) {
+ *manager =
+ new SeedGeneratorManager(new RandomSeedGenerator(seeds));
+ } else {
+ *manager =
+ new SeedGeneratorManager(new FixedSeedGenerator(seeds));
+ }
+ return Status::OK();
+ }));
+ handle = MakeResourceHandle<SeedGenerator>(ctx, container, name);
+ owns_resource = true;
+ } else {
+ OP_REQUIRES_OK(ctx, s);
+ }
+
+ // Ownership of manager is transferred onto `DatasetV3`.
+ *output = new ShuffleDatasetOp::DatasetV3(ctx, input, buffer_size, count,
+ std::move(seeds), manager,
+ std::move(handle), owns_resource);
+ } else if (op_version_ == 2) {
+ auto handle = HandleFromInput(ctx, 2);
+ SeedGeneratorManager* manager = nullptr;
+ Status s = ctx->resource_manager()->Lookup<SeedGeneratorManager>(
+ handle.container(), handle.name(), &manager);
+ bool owns_resource = false;
if (errors::IsNotFound(s)) {
LOG(WARNING) << "Failed to find seed generator resource. Falling back to "
- "using a non-deterministically-seeded seed generator.";
- *output =
- new ShuffleDatasetOp::Dataset(ctx, input, buffer_size, Seeds(0, 0),
- count, reshuffle_each_iteration_);
- return;
+ "using a non-deterministically seeded generator and "
+ "reshuffling each iteration.";
+ RandomSeeds seeds(0, 0);
+ OP_REQUIRES_OK(
+ ctx, ctx->resource_manager()->LookupOrCreate<SeedGeneratorManager>(
+ container, name, &manager,
+ [&seeds](SeedGeneratorManager** manager) {
+ *manager = new SeedGeneratorManager(
+ new RandomSeedGenerator(seeds));
+ return Status::OK();
+ }));
+ handle = MakeResourceHandle<SeedGeneratorManager>(ctx, container, name);
+ owns_resource = true;
+ } else {
+ OP_REQUIRES_OK(ctx, s);
}
- OP_REQUIRES_OK(ctx, s);
- // Create a fresh handle for the resource because the input handle can
- // become invalid after this op executes.
- std::unique_ptr<OwnedResourceHandle> handle;
- OP_REQUIRES_OK(
- ctx, OwnedResourceHandle::Create(
- ctx, seed_generator, seed_generator->DebugString(), &handle));
-
- // Ownership of seed generator is transferred onto `DatasetV2`.
- *output = new ShuffleDatasetOp::DatasetV2(
- ctx, input, buffer_size, count, seed_generator, std::move(handle));
- return;
- }
-
- int64 seed;
- OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed, &seed));
-
- int64 seed2;
- OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed2, &seed2));
-
- if (!reshuffle_each_iteration_) {
- // This dataset is only needed to support old clients running v2 eager with
- // reshuffle_each_iteration_=false. We can't tell here whether we are in v2
- // eager, so we conservatively always use FixedSeedDataset when
- // reshuffle_each_iteration=false.
- *output = new FixedSeedDataset(ctx, input, buffer_size, Seeds(seed, seed2),
- count);
+ // Ownership of manager is transferred onto `DatasetV2`.
+ *output =
+ new ShuffleDatasetOp::DatasetV2(ctx, input, buffer_size, count, manager,
+ std::move(handle), owns_resource);
} else {
- *output = new ShuffleDatasetOp::Dataset(ctx, input, buffer_size,
- Seeds(seed, seed2), count,
- reshuffle_each_iteration_);
+ if (op_version_ != 1) {
+ LOG(WARNING) << "Unsupported version of shuffle dataset op: "
+ << op_version_ << ". Defaulting to version 1.";
+ }
+ int64 seed;
+ OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed, &seed));
+ int64 seed2;
+ OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kSeed2, &seed2));
+ RandomSeeds seeds(seed, seed2);
+ SeedGeneratorManager* manager;
+ OP_REQUIRES_OK(
+ ctx,
+ ctx->resource_manager()->LookupOrCreate<SeedGeneratorManager>(
+ container, name, &manager,
+ [reshuffle = reshuffle_each_iteration_,
+ &seeds](SeedGeneratorManager** manager) {
+ if (reshuffle) {
+ *manager =
+ new SeedGeneratorManager(new RandomSeedGenerator(seeds));
+ } else {
+ *manager =
+ new SeedGeneratorManager(new FixedSeedGenerator(seeds));
+ }
+ return Status::OK();
+ }));
+ auto handle =
+ MakeResourceHandle<SeedGeneratorManager>(ctx, container, name);
+
+ // Ownership of manager is transferred onto `Dataset`.
+ *output = new ShuffleDatasetOp::Dataset(ctx, input, buffer_size, count,
+ std::move(seeds), manager,
+ std::move(handle));
}
}
class ShuffleAndRepeatDatasetOp::Dataset : public ShuffleDatasetBase {
public:
Dataset(OpKernelContext* ctx, const DatasetBase* input, int64 buffer_size,
- Seeds seeds, int64 count)
- : ShuffleDatasetBase(ctx, input, buffer_size, count), seeds_(seeds) {}
+ RandomSeeds&& seeds, SeedGeneratorManager* manager, int64 count,
+ ResourceHandle&& resource_handle)
+ : ShuffleDatasetBase(ctx, input, buffer_size, manager->get(), count),
+ manager_(manager),
+ resource_handle_(std::move(resource_handle)),
+ resource_mgr_(ctx->resource_manager()),
+ seeds_(std::move(seeds)) {}
- string DebugString() const override {
- name_utils::DatasetDebugStringParams params;
- params.set_args(buffer_size_, seeds_.seed_, seeds_.seed2_);
- return name_utils::DatasetDebugString(kDatasetType, params);
+ ~Dataset() override {
+ manager_->Unref();
+ Status s = resource_mgr_->Delete<SeedGeneratorManager>(
+ resource_handle_.container(), resource_handle_.name());
+ if (!s.ok()) {
+ LOG(WARNING) << "Failed to delete RNG resource: " << s.ToString();
+ }
}
- std::unique_ptr<IteratorBase> MakeIteratorInternal(
- const string& prefix) const override {
- return absl::make_unique<ShuffleDatasetBase::Iterator<ShuffleDatasetBase>>(
- ShuffleDatasetBase::Iterator<ShuffleDatasetBase>::Params{
- this, name_utils::IteratorPrefix(kDatasetType, prefix)},
- seeds_.seed_, seeds_.seed2_);
- }
+ string op_type() const override { return kDatasetType; }
protected:
Status AsGraphDefInternal(SerializationContext* ctx,
@@ -838,8 +787,8 @@
Node* count = nullptr;
TF_RETURN_IF_ERROR(b->AddScalar(buffer_size_, &buffer_size));
- TF_RETURN_IF_ERROR(
- AddSeeds(seeds_, /*preserve_random_seeds=*/true, b, &seed, &seed2));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed(), &seed));
+ TF_RETURN_IF_ERROR(b->AddScalar(seeds_.input_seed2(), &seed2));
TF_RETURN_IF_ERROR(b->AddScalar(count_, &count));
TF_RETURN_IF_ERROR(b->AddDataset(
this, {input_graph_node, buffer_size, seed, seed2, count}, // Inputs
@@ -849,7 +798,10 @@
}
private:
- const Seeds seeds_;
+ SeedGeneratorManager* const manager_; // Owned.
+ const ResourceHandle resource_handle_;
+ ResourceMgr* const resource_mgr_; // Not owned.
+ const RandomSeeds seeds_;
};
ShuffleAndRepeatDatasetOp::ShuffleAndRepeatDatasetOp(OpKernelConstruction* ctx)
@@ -874,11 +826,29 @@
int64 count;
OP_REQUIRES_OK(ctx, ParseScalarArgument<int64>(ctx, kCount, &count));
+ RandomSeeds seeds(seed, seed2);
+
OP_REQUIRES(ctx, count > 0 || count == -1,
errors::InvalidArgument(
"count must be greater than zero or equal to -1."));
- *output = new Dataset(ctx, input, buffer_size, Seeds(seed, seed2), count);
+ static std::atomic<int64> resource_id_counter(0);
+ const string& container = ctx->resource_manager()->default_container();
+ auto name = strings::StrCat(ctx->op_kernel().name(), "/", kSeedGenerator, "_",
+ resource_id_counter.fetch_add(1));
+ SeedGeneratorManager* manager;
+ OP_REQUIRES_OK(
+ ctx,
+ ctx->resource_manager()->LookupOrCreate<SeedGeneratorManager>(
+ container, name, &manager, [&seeds](SeedGeneratorManager** manager) {
+ *manager = new SeedGeneratorManager(new RandomSeedGenerator(seeds));
+ return Status::OK();
+ }));
+ auto handle = MakeResourceHandle<SeedGeneratorManager>(ctx, container, name);
+
+ // Ownership of manager is transferred onto `Dataset`.
+ *output = new Dataset(ctx, input, buffer_size, std::move(seeds), manager,
+ count, std::move(handle));
}
namespace {
@@ -888,6 +858,9 @@
REGISTER_KERNEL_BUILDER(Name("ShuffleDatasetV2").Device(DEVICE_CPU),
ShuffleDatasetOp);
+REGISTER_KERNEL_BUILDER(Name("ShuffleDatasetV3").Device(DEVICE_CPU),
+ ShuffleDatasetOp);
+
REGISTER_KERNEL_BUILDER(Name("ShuffleAndRepeatDataset").Device(DEVICE_CPU),
ShuffleAndRepeatDatasetOp);
} // namespace
diff --git a/tensorflow/core/kernels/data/shuffle_dataset_op.h b/tensorflow/core/kernels/data/shuffle_dataset_op.h
index 165a1db..7aa3c0e 100644
--- a/tensorflow/core/kernels/data/shuffle_dataset_op.h
+++ b/tensorflow/core/kernels/data/shuffle_dataset_op.h
@@ -50,8 +50,8 @@
private:
class Dataset;
class DatasetV2;
- class FixedSeedDataset;
- int op_version_;
+ class DatasetV3;
+ int op_version_ = 0;
bool reshuffle_each_iteration_;
};
diff --git a/tensorflow/core/kernels/data/shuffle_dataset_op_test.cc b/tensorflow/core/kernels/data/shuffle_dataset_op_test.cc
index ca9afce..6d16d76 100644
--- a/tensorflow/core/kernels/data/shuffle_dataset_op_test.cc
+++ b/tensorflow/core/kernels/data/shuffle_dataset_op_test.cc
@@ -297,23 +297,23 @@
{/*dataset_params=*/ShuffleDatasetParams7(),
/*expected_shuffle_outputs=*/
CreateTensors<int64>(TensorShape({}),
- {{9}, {0}, {8}, {6}, {1}, {3}, {7}, {2}, {4}, {5},
- {4}, {3}, {0}, {5}, {8}, {2}, {6}, {9}, {7}, {1}}),
+ {{2}, {6}, {1}, {3}, {9}, {5}, {0}, {8}, {7}, {4},
+ {0}, {5}, {1}, {7}, {2}, {9}, {8}, {4}, {6}, {3}}),
/*expected_reshuffle_outputs=*/
- CreateTensors<int64>(TensorShape({}), {{9}, {0}, {8}, {6}, {1}, {3}, {7},
- {2}, {4}, {5}, {4}, {3}, {0}, {5},
- {8}, {2}, {6}, {9}, {7}, {1}})},
+ CreateTensors<int64>(TensorShape({}), {{1}, {6}, {0}, {5}, {2}, {7}, {4},
+ {3}, {9}, {8}, {6}, {5}, {0}, {9},
+ {4}, {7}, {2}, {8}, {1}, {3}})},
{/*dataset_params=*/ShuffleDatasetParams8(),
/*expected_shuffle_outputs=*/
CreateTensors<int64>(
TensorShape({}),
- {{2}, {0}, {1}, {2}, {0}, {1}, {1}, {2}, {0}, {1}, {0},
- {2}, {2}, {0}, {1}, {1}, {0}, {2}, {2}, {1}, {0}}),
+ {{1}, {2}, {0}, {1}, {2}, {0}, {1}, {0}, {2}, {1}, {0},
+ {2}, {0}, {2}, {1}, {0}, {1}, {2}, {1}, {2}, {0}}),
/*expected_reshuffle_outputs=*/
CreateTensors<int64>(
TensorShape({}),
- {{2}, {0}, {1}, {2}, {0}, {1}, {1}, {2}, {0}, {1}, {0},
- {2}, {2}, {0}, {1}, {1}, {0}, {2}, {2}, {1}, {0}})}};
+ {{1}, {0}, {2}, {0}, {1}, {2}, {2}, {1}, {0}, {0}, {1},
+ {2}, {0}, {2}, {1}, {0}, {1}, {2}, {1}, {0}, {2}})}};
}
class ParameterizedGetNextTest : public ShuffleDatasetOpTest,
@@ -496,16 +496,16 @@
{/*dataset_params=*/ShuffleDatasetParams7(),
/*breakpoints=*/{0, 5, 22},
/*expected_shuffle_outputs=*/
- CreateTensors<int64>(TensorShape({}), {{9}, {0}, {8}, {6}, {1}, {3}, {7},
- {2}, {4}, {5}, {4}, {3}, {0}, {5},
- {8}, {2}, {6}, {9}, {7}, {1}})},
+ CreateTensors<int64>(TensorShape({}), {{2}, {6}, {1}, {3}, {9}, {5}, {0},
+ {8}, {7}, {4}, {0}, {5}, {1}, {7},
+ {2}, {9}, {8}, {4}, {6}, {3}})},
{/*dataset_params=*/ShuffleDatasetParams8(),
/*breakpoints=*/{0, 5, 20},
/*expected_shuffle_outputs=*/
CreateTensors<int64>(
TensorShape({}),
- {{2}, {0}, {1}, {2}, {0}, {1}, {1}, {2}, {0}, {1}, {0},
- {2}, {2}, {0}, {1}, {1}, {0}, {2}, {2}, {1}, {0}})}};
+ {{1}, {2}, {0}, {1}, {2}, {0}, {1}, {0}, {2}, {1}, {0},
+ {2}, {0}, {2}, {1}, {0}, {1}, {2}, {1}, {2}, {0}})}};
}
class ParameterizedIteratorSaveAndRestoreTest
diff --git a/tensorflow/core/kernels/deep_conv2d.cc b/tensorflow/core/kernels/deep_conv2d.cc
index 750c031..7a2bc1e 100644
--- a/tensorflow/core/kernels/deep_conv2d.cc
+++ b/tensorflow/core/kernels/deep_conv2d.cc
@@ -185,7 +185,7 @@
template <typename T>
struct ComputeFilterRangeTransform {
typedef typename Eigen::internal::packet_traits<T>::type Packet;
- static const int64 kPacketSize = (sizeof(Packet) / sizeof(T));
+ static constexpr int64 kPacketSize = (sizeof(Packet) / sizeof(T));
typedef Eigen::Map<
Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>
diff --git a/tensorflow/core/kernels/eigen_contraction_kernel.h b/tensorflow/core/kernels/eigen_contraction_kernel.h
index 56ac04c..7233020 100644
--- a/tensorflow/core/kernels/eigen_contraction_kernel.h
+++ b/tensorflow/core/kernels/eigen_contraction_kernel.h
@@ -270,10 +270,10 @@
static constexpr float kScaleN = 1.0;
// Mkldnn Avx/Avx2/Avx512 unroll factors are: 8/16/48.
- static const StorageIndex kUnrollM = 48;
+ static constexpr StorageIndex kUnrollM = 48;
// Mkldnn Avx/Avx2/Avx512 unroll factors are: 6/6/8.
- static const StorageIndex kUnrollN = 24;
+ static constexpr StorageIndex kUnrollN = 24;
public:
TensorContractionBlocking(StorageIndex k, StorageIndex m, StorageIndex n,
diff --git a/tensorflow/core/kernels/eigen_convolution_helpers.h b/tensorflow/core/kernels/eigen_convolution_helpers.h
index b6587da..965a283 100644
--- a/tensorflow/core/kernels/eigen_convolution_helpers.h
+++ b/tensorflow/core/kernels/eigen_convolution_helpers.h
@@ -63,7 +63,7 @@
functionExistsSfinae<TensorEvaluatorType, PacketType, IndexType>(
nullptr)) status;
- static const bool value = status::value;
+ static constexpr bool value = status::value;
};
// Compute a mask for loading/storing coefficients in/from a packet in a
diff --git a/tensorflow/core/kernels/eigen_pooling.h b/tensorflow/core/kernels/eigen_pooling.h
index 2b36bad..7db4a69 100644
--- a/tensorflow/core/kernels/eigen_pooling.h
+++ b/tensorflow/core/kernels/eigen_pooling.h
@@ -277,11 +277,11 @@
#if (EIGEN_ARCH_i386 || EIGEN_ARCH_x86_64) && !defined(__CUDACC__) && \
!defined(__HIPCC__)
// We only support packet access for floats.
- static const bool PacketAccess = internal::is_same<T, float>::value;
+ static constexpr bool PacketAccess = internal::is_same<T, float>::value;
#else
static const bool PacketAccess = false;
#endif
- static const bool IsStateful = true;
+ static constexpr bool IsStateful = true;
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE AvgPoolMeanReducer() : scalarCount_(0) {
typedef typename packet_traits<T>::type Packet;
diff --git a/tensorflow/core/kernels/eigen_volume_patch.h b/tensorflow/core/kernels/eigen_volume_patch.h
index bb71919..7afc05d 100644
--- a/tensorflow/core/kernels/eigen_volume_patch.h
+++ b/tensorflow/core/kernels/eigen_volume_patch.h
@@ -28,15 +28,15 @@
struct CustomTensorEvaluator {
typedef TensorVolumePatchOp<Planes, Rows, Cols, ArgType> XprType;
typedef typename XprType::Index Index;
- static const int NumInputDims = internal::array_size<
+ static constexpr int NumInputDims = internal::array_size<
typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
- static const int NumDims = NumInputDims + 1;
+ static constexpr int NumDims = NumInputDims + 1;
typedef DSizes<Index, NumDims> Dimensions;
typedef
typename internal::remove_const<typename XprType::Scalar>::type Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
- static const Index PacketSize =
+ static constexpr Index PacketSize =
internal::unpacket_traits<PacketReturnType>::size;
enum {
diff --git a/tensorflow/core/kernels/fuzzing/string_split_v2_fuzz.cc b/tensorflow/core/kernels/fuzzing/string_split_v2_fuzz.cc
index 367759d..cc14899 100644
--- a/tensorflow/core/kernels/fuzzing/string_split_v2_fuzz.cc
+++ b/tensorflow/core/kernels/fuzzing/string_split_v2_fuzz.cc
@@ -56,7 +56,7 @@
}
private:
- static const size_t kMaxSepSize = 4;
+ static constexpr size_t kMaxSepSize = 4;
};
STANDARD_TF_FUZZ_FUNCTION(FuzzStringSplitV2);
diff --git a/tensorflow/core/kernels/mkl_conv_ops.cc b/tensorflow/core/kernels/mkl_conv_ops.cc
index 80a53ad..59de322 100644
--- a/tensorflow/core/kernels/mkl_conv_ops.cc
+++ b/tensorflow/core/kernels/mkl_conv_ops.cc
@@ -51,12 +51,10 @@
using mkldnn::convolution_forward;
using mkldnn::prop_kind;
using mkldnn::stream;
-
using ConvFwdPd = mkldnn::convolution_forward::primitive_desc;
using ReorderPd = mkldnn::reorder::primitive_desc;
namespace tensorflow {
-
// This structure aggregates multiple inputs to Conv2DFwd* methods.
struct MklConvFwdParams {
memory::dims src_dims;
@@ -96,14 +94,12 @@
class MklConvFwdPrimitive : public MklPrimitive {
public:
explicit MklConvFwdPrimitive(const MklConvFwdParams& convFwdDims)
- : cpu_engine_(ENGINE_CPU, 0) {
- context_.fwd_stream.reset(new CPU_STREAM(cpu_engine_));
+ : MklPrimitive(engine(ENGINE_CPU, 0)) {
// Create convolution primitive
if (context_.conv_fwd == nullptr) {
Setup(convFwdDims);
}
}
-
~MklConvFwdPrimitive() {}
// Convolution forward execute with bias
@@ -112,7 +108,8 @@
// bias_data: input data buffer of bias
// dst_data: output data buffer of dst
void Execute(const Tinput* src_data, const Tfilter* filter_data,
- const Tbias* bias_data, const Toutput* dst_data) {
+ const Tbias* bias_data, const Toutput* dst_data,
+ std::shared_ptr<stream> fwd_stream) {
context_.src_mem->set_data_handle(
static_cast<void*>(const_cast<Tinput*>(src_data)));
context_.filter_mem->set_data_handle(
@@ -127,11 +124,11 @@
DCHECK_EQ(context_.fwd_primitives.size(),
context_.fwd_primitives_args.size());
for (size_t i = 0; i < context_.fwd_primitives.size(); ++i) {
- context_.fwd_primitives.at(i).execute(*context_.fwd_stream,
+ context_.fwd_primitives.at(i).execute(*fwd_stream,
context_.fwd_primitives_args.at(i));
}
#else
- context_.fwd_stream->submit(context_.fwd_primitives);
+ fwd_stream->submit(context_.fwd_primitives);
#endif // ENABLE_MKLDNN_V1
// After execution, set data handle back
@@ -148,8 +145,8 @@
// filter_data: input data buffer of filter (weights)
// dst_data: output data buffer of dst
void Execute(const Tinput* src_data, const Tfilter* filter_data,
- const Toutput* dst_data) {
- Execute(src_data, filter_data, nullptr, dst_data);
+ const Toutput* dst_data, std::shared_ptr<stream> fwd_stream) {
+ Execute(src_data, filter_data, nullptr, dst_data, fwd_stream);
}
#ifndef ENABLE_MKLDNN_V1
@@ -191,7 +188,6 @@
std::shared_ptr<ConvFwdPd> fwd_pd;
std::shared_ptr<mkldnn::primitive> conv_fwd;
- std::shared_ptr<mkldnn::stream> fwd_stream;
std::vector<mkldnn::primitive> fwd_primitives;
#ifdef ENABLE_MKLDNN_V1
@@ -213,8 +209,7 @@
filter_md(nullptr),
bias_md(nullptr),
fwd_pd(nullptr),
- conv_fwd(nullptr),
- fwd_stream(nullptr) {
+ conv_fwd(nullptr) {
}
};
@@ -346,7 +341,6 @@
}
struct ConvFwdContext context_;
- engine cpu_engine_;
};
// TODO(nhasabni): We should not require passing a type to MklPrimitiveFactory.
@@ -678,11 +672,9 @@
// TODO(mdfaijul): Extend the basic parameters for data types and fusions
this->ExtendConvFwdParams(context, convFwdDims);
-
conv_fwd =
MklConvFwdPrimitiveFactory<Tinput, Tfilter, Tbias, Ttemp_output>::Get(
convFwdDims, do_not_cache);
-
// Allocate output tensors `output_tensor` and `filter_out_tensor`
MklDnnShape output_mkl_shape;
std::shared_ptr<ConvFwdPd> conv_fwd_pd = conv_fwd->GetPrimitiveDesc();
@@ -703,8 +695,10 @@
Tinput* src_data = nullptr;
if (IS_SRC_REORDER_NEEDED(src_md, conv_fwd_pd, conv_fwd)) {
src.SetUsrMem(src_md, &src_tensor);
- src.CheckReorderToOpMem(MEMORY_PD_WITHOUT_DATA(
- GET_SRC_DESC_FROM_OP_PD(conv_fwd_pd), cpu_engine_));
+ src.CheckReorderToOpMem(
+ MEMORY_PD_WITHOUT_DATA(GET_SRC_DESC_FROM_OP_PD(conv_fwd_pd),
+ cpu_engine_),
+ context);
src_data = static_cast<Tinput*>(src.GetOpMem().get_data_handle());
} else {
src_data = static_cast<Tinput*>(
@@ -735,13 +729,16 @@
if (!is_filter_cached) {
filter.SetUsrMem(filter_md, &filter_tensor);
if (filter_out_tensor == nullptr) {
- filter.CheckReorderToOpMem(MEMORY_PD_WITHOUT_DATA(
- GET_WEIGHTS_DESC_FROM_OP_PD(conv_fwd_pd), cpu_engine_));
+ filter.CheckReorderToOpMem(
+ MEMORY_PD_WITHOUT_DATA(GET_WEIGHTS_DESC_FROM_OP_PD(conv_fwd_pd),
+ cpu_engine_),
+ context);
} else {
filter.CheckReorderToOpMem(
GET_WEIGHTS_DESC_FROM_OP_PD(conv_fwd_pd),
DATA_WITH_ENGINE(filter.GetTensorBuffer(filter_out_tensor),
- cpu_engine_));
+ cpu_engine_),
+ context);
}
filter_data =
static_cast<Tfilter*>(filter.GetOpMem().get_data_handle());
@@ -752,20 +749,23 @@
}
// Execute convolution
+ std::shared_ptr<stream> fwd_cpu_stream;
+ fwd_cpu_stream.reset(CreateStream(context, conv_fwd->GetEngine()));
if (fuse_biasadd_) {
const Tensor& bias_tensor = MklGetInput(context, kInputIndex_Bias);
Tbias* bias_data =
this->GetBiasHandle(context, conv_fwd_pd, bias_tensor);
- conv_fwd->Execute(src_data, filter_data, bias_data, dst_data);
+ conv_fwd->Execute(src_data, filter_data, bias_data, dst_data,
+ fwd_cpu_stream);
} else {
if (!eager_mode) {
- conv_fwd->Execute(src_data, filter_data, dst_data);
+ conv_fwd->Execute(src_data, filter_data, dst_data, fwd_cpu_stream);
} else {
// In eager mode we first write the output to temporary
// buffer in MKL format. Then we convert the data to TF format.
Ttemp_output* tmp_data = reinterpret_cast<Ttemp_output*>(
tmp_tensor.flat<Toutput>().data());
- conv_fwd->Execute(src_data, filter_data, tmp_data);
+ conv_fwd->Execute(src_data, filter_data, tmp_data, fwd_cpu_stream);
// Now we need to convert the output to TF format.
auto output_tf_md = output_mkl_shape.GetTfLayout();
@@ -780,12 +780,13 @@
memory* dst_data_mem =
new MEMORY_CONSTRUCTOR(OUTPUT_TF_MD, cpu_engine_, dst_data);
CreateAndExecuteReorder(reorder_pd, *tmp_data_mem, *dst_data_mem,
- cpu_engine_);
+ cpu_engine_, context);
}
}
// Delete primitive since it is not cached.
if (do_not_cache) delete conv_fwd;
+
} catch (mkldnn::error& e) {
string error_msg = tensorflow::strings::StrCat(
"Status: ", e.status, ", message: ", string(e.message), ", in file ",
@@ -970,8 +971,9 @@
new MEMORY_CONSTRUCTOR(DST_MD, this->cpu_engine_, dst_buf));
auto reorder_desc =
REORDER_PD_CONSTRUCTOR(ADD_MD, DST_MD, this->cpu_engine_);
+
CreateAndExecuteReorder(reorder_desc, *fuse_add_src_, *fuse_add_dst_,
- this->cpu_engine_);
+ this->cpu_engine_, context);
}
} else {
AllocateOutputSetMklShape(context, kOutputIndex_Dst, output_tensor,
@@ -1097,6 +1099,7 @@
filter_tf_shape, filter_mkl_shape);
}
+ // TODO(intel-mkl): This function does not seem to be called. Remove it.
// Prepare and execute net - checks for input and output reorders.
void PrepareAndExecuteNet(const ConvFwdPd& conv_prim_desc,
MklDnnData<Tinput>* src,
@@ -1185,7 +1188,7 @@
// Otherwise, cache filter
filter.SetUsrMem(filter_md, &filter_tensor);
filter.CheckReorderToOpMem(conv_fwd_pd.get()->weights_desc(),
- this->cpu_engine_);
+ this->cpu_engine_, context);
filter_data = static_cast<Tfilter*>(filter.GetOpMem().get_data_handle());
Tensor* filter_tensor_ptr = nullptr;
@@ -1251,9 +1254,9 @@
const Tensor& cached_filter_md =
*cached_filter_md_ptensor_.AccessTensor(context);
- // Check if the memory descriptor of the cached weights is same as
- // filter_md. If so, we can use the cached weights; otherwise
- // return nullptr.
+// Check if the memory descriptor of the cached weights is same as
+// filter_md. If so, we can use the cached weights; otherwise
+// return nullptr.
#ifdef ENABLE_MKLDNN_V1
if (filter_md == *static_cast<memory::desc*>(cached_filter_md.data())) {
#else
@@ -1363,6 +1366,58 @@
virtual ~MklFusedConvOp() {}
};
+template <typename Device, typename Tinput, typename Tfilter, typename Tbias,
+ typename Toutput, typename Ttemp_output, typename Tpadding,
+ bool pad_enabled, bool bias_enabled, bool is_depthwise>
+class MklFusedDepthwiseConvOp
+ : public MklConvOp<Device, Tinput, Tfilter, Tbias, Toutput, Ttemp_output,
+ Tpadding, bias_enabled, false, is_depthwise, false> {
+ public:
+ explicit MklFusedDepthwiseConvOp(OpKernelConstruction* context)
+ : MklConvOp<Device, Tinput, Tfilter, Tbias, Toutput, Ttemp_output,
+ Tpadding, bias_enabled, false, is_depthwise, false>(context) {
+ // Since we came here through the registration of
+ // _MklFusedDepthwiseConv2dNative, get all
+ // information from 'fused_ops' and 'num_args'
+ std::vector<string> fused_ops;
+ OP_REQUIRES_OK(context, context->GetAttr("fused_ops", &fused_ops));
+
+ int num_args;
+ OP_REQUIRES_OK(context, context->GetAttr("num_args", &num_args));
+ OP_REQUIRES(context, !fused_ops.empty(),
+ errors::InvalidArgument(
+ "Fused DepthwiseConv2D must have at least one fused op."));
+
+ if (fused_ops == std::vector<string>{"BiasAdd"}) {
+ this->set_fuse_biasadd(true);
+ } else if (fused_ops == std::vector<string>{"BiasAdd", "Relu"}) {
+ this->set_fuse_biasadd(true);
+ this->set_fuse_activation(true, ALGORITHM::eltwise_relu);
+ } else if (fused_ops == std::vector<string>{"BiasAdd", "Relu6"}) {
+ this->set_fuse_biasadd(true);
+ this->set_fuse_activation(true, ALGORITHM::eltwise_bounded_relu, 6.0);
+ } else if (fused_ops == std::vector<string>{"BiasAdd", "Elu"}) {
+ this->set_fuse_biasadd(true);
+ this->set_fuse_activation(true, ALGORITHM::eltwise_elu, 1.0);
+ } else {
+ OP_REQUIRES(context, false,
+ errors::Unimplemented("Fusion is not implemented: [",
+ absl::StrJoin(fused_ops, ","), "]"));
+ }
+
+ OP_REQUIRES(
+ context, num_args == 1,
+ errors::InvalidArgument(
+ "Fused DepthwiseConv2D must have one extra argument: bias."));
+
+ if (pad_enabled) {
+ this->set_fuse_pad(true);
+ }
+ }
+
+ virtual ~MklFusedDepthwiseConvOp() {}
+};
+
// We create new class for each version of Quantized Convolution and inherit
// from the FP32 version of the base class
template <typename Device, typename Tinput, typename Tbias, typename Toutput,
@@ -1600,7 +1655,7 @@
input_bias_->GET_DESC, scaled_bias_->GET_DESC, this->cpu_engine_,
bias_attr);
CreateAndExecuteReorder(reorder_desc, *input_bias_, *scaled_bias_,
- this->cpu_engine_);
+ this->cpu_engine_, context);
Tbias* bias_data =
reinterpret_cast<Tbias*>(scaled_bias_->get_data_handle());
@@ -1856,7 +1911,8 @@
auto reorder_desc = REORDER_PD_CONSTRUCTOR_WITH_ATTR(
SUMMAND_MD, conv_prim_desc.PRIMITIVE_DESC_DST, this->cpu_engine_,
reorder_attr);
- CreateAndExecuteReorder(reorder_desc, *summand_, *dst_, this->cpu_engine_);
+ CreateAndExecuteReorder(reorder_desc, *summand_, *dst_, this->cpu_engine_,
+ context);
}
std::shared_ptr<mkldnn::memory> summand_;
@@ -2253,6 +2309,11 @@
.TypeConstraint<quint8>("out_type"),
NoOp);
+REGISTER_KERNEL_BUILDER(Name("_FusedDepthwiseConv2dNative")
+ .Device(DEVICE_CPU)
+ .TypeConstraint<float>("T"),
+ NoOp);
+
// Register templatized MKL kernels for non-fused and fused-versions of
// QuantizedDepthwiseConv2D.
REGISTER_KERNEL_BUILDER(Name("_MklQuantizedDepthwiseConv2D")
@@ -2306,6 +2367,14 @@
MklQuantizedConv2DReluOp<CPUDevice, quint8, qint32, quint8, quint8, true,
true>);
+REGISTER_KERNEL_BUILDER(
+ Name("_MklFusedDepthwiseConv2dNative")
+ .Device(DEVICE_CPU)
+ .TypeConstraint<float>("T")
+ .Label(mkl_op_registry::kMklLayoutDependentOpLabel),
+ MklFusedDepthwiseConvOp<CPUDevice, float, float, float, float, float, int32,
+ false, true, true>);
+
// Register 2D operations
#define REGISTER_MKL_CPU_2D(T) \
REGISTER_KERNEL_BUILDER( \
diff --git a/tensorflow/core/kernels/mkl_fused_ops_test.cc b/tensorflow/core/kernels/mkl_fused_ops_test.cc
index ff4f678..edd1201 100644
--- a/tensorflow/core/kernels/mkl_fused_ops_test.cc
+++ b/tensorflow/core/kernels/mkl_fused_ops_test.cc
@@ -134,6 +134,7 @@
static void VerifyFusedTensorsClose(int depth, int image_width,
int image_height, int image_batch_count,
int filter_size, int filter_count,
+ int bias_size,
const std::vector<string>& fused_ops,
const FusedGraphRunner& run_default,
const FusedGraphRunner& run_fused) {
@@ -145,7 +146,6 @@
Tensor filter(dtype, {filter_size, filter_size, depth, filter_count});
filter.flat<T>() = filter.flat<T>().template setRandom<random_gen_>();
- const int bias_size = filter_count;
Tensor bias(dtype, {bias_size});
bias.flat<T>() = bias.flat<T>().template setRandom<random_gen_>();
@@ -321,9 +321,10 @@
out);
};
+ const int bias_size = filter_count;
CommonTestUtilities<T>::VerifyFusedTensorsClose(
depth, image_width, image_height, image_batch_count, filter_size,
- filter_count, fused_ops, run_default, run_fused);
+ filter_count, bias_size, fused_ops, run_default, run_fused);
}
};
@@ -449,6 +450,223 @@
using MklFusedBiasAddDataTypes = ::testing::Types<float>;
INSTANTIATE_TYPED_TEST_CASE_P(Test, MklFusedConv2DWithBiasOpTest,
MklFusedBiasAddDataTypes);
+
+// Testing MKL's fused depthwise convolution ops
+template <typename T>
+class MklFusedDepthwiseConv2DOpTest : public OpsTestBase {
+ protected:
+ static constexpr int kDepth = 3;
+ static constexpr int kImageWidth = 32;
+ static constexpr int kImageHeight = 32;
+ static constexpr int kImageBatchCount = 8;
+
+ void RunDepthwiseConv2DUnfused(const Tensor& input_data,
+ const Tensor& filter_data,
+ const Tensor& bias_data,
+ const std::vector<string>& fused_ops,
+ Tensor* output, int stride = 1) {
+ auto root = tensorflow::Scope::NewRootScope();
+ auto input_data_op =
+ ops::Const(root.WithOpName("input"), Input::Initializer(input_data));
+ Output next_op = ops::DepthwiseConv2dNative(
+ root.WithOpName("depthwise_conv"), input_data_op,
+ ops::Const(root.WithOpName("filter"), Input::Initializer(filter_data)),
+ {1, stride, stride, 1}, "SAME");
+
+ string last_op = "";
+ if (std::find(fused_ops.begin(), fused_ops.end(), "BiasAdd") !=
+ fused_ops.end()) {
+ last_op = "with_bias";
+ next_op = ops::BiasAdd(
+ root.WithOpName(last_op), next_op,
+ ops::Const(root.WithOpName("bias"), Input::Initializer(bias_data)));
+ }
+
+ if (std::find(fused_ops.begin(), fused_ops.end(), "Relu") !=
+ fused_ops.end()) {
+ last_op = "with_relu";
+ next_op = ops::Relu(root.WithOpName(last_op), next_op);
+ }
+
+ if (std::find(fused_ops.begin(), fused_ops.end(), "Relu6") !=
+ fused_ops.end()) {
+ last_op = "with_relu6";
+ next_op = ops::Relu6(root.WithOpName(last_op), next_op);
+ }
+
+ if (std::find(fused_ops.begin(), fused_ops.end(), "Elu") !=
+ fused_ops.end()) {
+ last_op = "with_elu";
+ next_op = ops::Elu(root.WithOpName(last_op), next_op);
+ }
+
+ CommonTestUtilities<T>::RunAndFetch(root, last_op, output);
+ }
+
+ void RunMklFusedDepthwiseConv2DOp(const Tensor& image, const Tensor& filter,
+ const std::vector<Tensor>& args,
+ const std::vector<string>& fused_ops,
+ Tensor* output, int stride = 1) {
+ DataType dtype = DataTypeToEnum<T>::v();
+ int num_args = static_cast<int>(args.size());
+
+ TF_EXPECT_OK(NodeDefBuilder("fused_depthwise_conv_op",
+ "_MklFusedDepthwiseConv2dNative")
+ .Input(FakeInput(dtype))
+ .Input(FakeInput(dtype))
+ .Input(FakeInput(num_args, dtype))
+ .Input(FakeInput(DT_UINT8))
+ .Input(FakeInput(DT_UINT8))
+ .Input(FakeInput(num_args, DT_UINT8))
+ .Attr("T", dtype)
+ .Attr("num_args", num_args)
+ .Attr("strides", {1, stride, stride, 1})
+ .Attr("padding", "SAME")
+ .Attr("fused_ops", fused_ops)
+ .Attr("_kernel", "MklLayoutDependentOp")
+ .Finalize(node_def()));
+
+ TF_EXPECT_OK(InitOp());
+
+ AddInputFromArray<T>(image.shape(), image.flat<T>());
+ AddInputFromArray<T>(filter.shape(), filter.flat<T>());
+ for (const Tensor& arg : args)
+ AddInputFromArray<T>(arg.shape(), arg.flat<T>());
+ AddInputFromArray<uint8>(dummy_shape, dummy_tensor);
+ AddInputFromArray<uint8>(dummy_shape, dummy_tensor);
+ for (const Tensor& arg : args)
+ AddInputFromArray<uint8>(dummy_shape, dummy_tensor);
+ TF_ASSERT_OK(RunOpKernel());
+
+ // Compare output to expected results
+ const Tensor& output_tensor = *GetOutput(0);
+ // Index 2 will need to be changed if the number of outputs produced
+ // by MklDepthwiseConv2D change.
+ const Tensor& output_meta_tensor = *GetOutput(2);
+ CommonTestUtilities<T> test_util;
+ test_util.PerformConversion(dtype, output_tensor, output_meta_tensor,
+ output);
+ }
+
+ // Verifies computing unfused ops in a graph is identical to
+ // FusedDepthwiseConv2D.
+ void VerifyFusedDepthwiseConv2D(int filter_size, int filter_count,
+ int bias_size,
+ const std::vector<string>& fused_ops,
+ int depth = kDepth,
+ int image_width = kImageWidth,
+ int image_height = kImageHeight,
+ int image_batch_count = kImageBatchCount) {
+ const FusedGraphRunner run_default =
+ [this](const Tensor& input_data, const Tensor& filter_data,
+ const Tensor& bias_data, const std::vector<string>& fused_ops,
+ Tensor* out) {
+ RunDepthwiseConv2DUnfused(input_data, filter_data, bias_data,
+ fused_ops, out);
+ };
+
+ const FusedGraphRunner run_fused =
+ [this](const Tensor& input_data, const Tensor& filter_data,
+ const Tensor& bias_data, const std::vector<string>& fused_ops,
+ Tensor* out) {
+ std::vector<Tensor> fused_input = {bias_data};
+ RunMklFusedDepthwiseConv2DOp(input_data, filter_data, fused_input,
+ fused_ops, out);
+ };
+
+ CommonTestUtilities<T>::VerifyFusedTensorsClose(
+ depth, image_width, image_height, image_batch_count, filter_size,
+ filter_count, bias_size, fused_ops, run_default, run_fused);
+ }
+};
+
+template <typename T>
+class MklFusedDepthwiseConv2DWithBiasOpTest
+ : public MklFusedDepthwiseConv2DOpTest<T> {};
+
+TYPED_TEST_SUITE_P(MklFusedDepthwiseConv2DWithBiasOpTest);
+
+// -------------------------------------------------------------------------- //
+// DepthwiseConv2D + BiasAdd + {Activation} //
+// -------------------------------------------------------------------------- //
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest, OneByOneConvolution) {
+ const int kFilterSize = 1;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest, SpatialConvolution) {
+ const int kFilterSize = 3;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest,
+ OneByOneConvolutionAndRelu) {
+ const int kFilterSize = 1;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Relu"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest, SpatialConvolutionAndRelu) {
+ const int kFilterSize = 3;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Relu"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest,
+ OneByOneConvolutionAndRelu6) {
+ const int kFilterSize = 1;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Relu6"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest,
+ SpatialConvolutionAndRelu6) {
+ const int kFilterSize = 3;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Relu6"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest, OneByOneConvolutionAndElu) {
+ const int kFilterSize = 1;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Elu"});
+}
+
+TYPED_TEST_P(MklFusedDepthwiseConv2DWithBiasOpTest, SpatialConvolutionAndElu) {
+ const int kFilterSize = 3;
+ const int kFilterCount = 1;
+ const int kBiasSize = 3;
+ this->VerifyFusedDepthwiseConv2D(kFilterSize, kFilterCount, kBiasSize,
+ {"BiasAdd", "Elu"});
+}
+
+REGISTER_TYPED_TEST_SUITE_P(
+ MklFusedDepthwiseConv2DWithBiasOpTest, OneByOneConvolution,
+ SpatialConvolution, OneByOneConvolutionAndRelu, SpatialConvolutionAndRelu,
+ OneByOneConvolutionAndRelu6, SpatialConvolutionAndRelu6,
+ OneByOneConvolutionAndElu, SpatialConvolutionAndElu);
+
+using MklFusedBiasAddDataTypes = ::testing::Types<float>;
+INSTANTIATE_TYPED_TEST_SUITE_P(Test, MklFusedDepthwiseConv2DWithBiasOpTest,
+ MklFusedBiasAddDataTypes);
+
// Testing fusion of pad and convolution
class FusedPadConvOpTest : public OpsTestBase {
diff --git a/tensorflow/core/kernels/mkl_input_conversion_op.cc b/tensorflow/core/kernels/mkl_input_conversion_op.cc
index e69fddd..f7866cb 100644
--- a/tensorflow/core/kernels/mkl_input_conversion_op.cc
+++ b/tensorflow/core/kernels/mkl_input_conversion_op.cc
@@ -165,7 +165,7 @@
input1_md, tensor_out, net, net_args, cpu_engine)),
errors::Internal(
"MklInputConversionOp: Failed to create reorder for input0"));
- ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine);
+ ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine, context);
// Input1 will be passed through
ForwardMklTensorInToOut(context, kInputIndex_1, kInputIndex_1);
return;
@@ -273,7 +273,7 @@
errors::Internal("MklInputConversionOp: Failed to forward "
"input tensor to output"));
} else {
- ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine);
+ ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine, context);
}
// -- The tensor in MKL format passes through --
diff --git a/tensorflow/core/kernels/mkl_reshape_op.cc b/tensorflow/core/kernels/mkl_reshape_op.cc
index c31e67b..bda3fad 100644
--- a/tensorflow/core/kernels/mkl_reshape_op.cc
+++ b/tensorflow/core/kernels/mkl_reshape_op.cc
@@ -172,7 +172,8 @@
// shape_from != shape_to), then we just copy input tensor to
// output tensor with target shape (we cannot forward Mkl layout
// in such case because shape has changed.)
- if (dnn_data_input.CheckReorderToOpMem(OUTPUT_TF_MD, output_tensor)) {
+ if (dnn_data_input.CheckReorderToOpMem(OUTPUT_TF_MD, output_tensor,
+ context)) {
} else {
OP_REQUIRES(context,
output_tensor->CopyFrom(input_tensor, shape_to),
diff --git a/tensorflow/core/kernels/mkl_tfconv_op.h b/tensorflow/core/kernels/mkl_tfconv_op.h
index 2e48961..f7aa4d2 100644
--- a/tensorflow/core/kernels/mkl_tfconv_op.h
+++ b/tensorflow/core/kernels/mkl_tfconv_op.h
@@ -111,7 +111,8 @@
if (input.IsReorderNeeded(OUTPUT_TF_MD)) {
// Insert reorder between MKL layout and TensorFlow layout
OP_REQUIRES(
- context, input.CheckReorderToOpMem(OUTPUT_TF_MD, output_tensor),
+ context,
+ input.CheckReorderToOpMem(OUTPUT_TF_MD, output_tensor, context),
errors::Internal("MklToTfOp: Failed to create input reorder"));
} else {
// If not, just forward input tensor to output tensor.
diff --git a/tensorflow/core/kernels/parameterized_truncated_normal_op.cc b/tensorflow/core/kernels/parameterized_truncated_normal_op.cc
index 09dc3ff..4eab905 100644
--- a/tensorflow/core/kernels/parameterized_truncated_normal_op.cc
+++ b/tensorflow/core/kernels/parameterized_truncated_normal_op.cc
@@ -317,7 +317,7 @@
template <typename Device, typename T>
class ParameterizedTruncatedNormalOp : public OpKernel {
// Reshape batches so each batch is this size if possible.
- static const int32 kDesiredBatchSize = 100;
+ static constexpr int32 kDesiredBatchSize = 100;
public:
explicit ParameterizedTruncatedNormalOp(OpKernelConstruction* context)
diff --git a/tensorflow/core/kernels/queue_base.h b/tensorflow/core/kernels/queue_base.h
index d1a9403..ea60403 100644
--- a/tensorflow/core/kernels/queue_base.h
+++ b/tensorflow/core/kernels/queue_base.h
@@ -36,7 +36,7 @@
class QueueBase : public QueueInterface {
public:
// As a possible value of 'capacity'.
- static const int32 kUnbounded = INT_MAX;
+ static constexpr int32 kUnbounded = INT_MAX;
// Args:
// component_dtypes: The types of each component in a queue-element tuple.
diff --git a/tensorflow/core/kernels/random_binomial_op.cc b/tensorflow/core/kernels/random_binomial_op.cc
index ea42239..4647457 100644
--- a/tensorflow/core/kernels/random_binomial_op.cc
+++ b/tensorflow/core/kernels/random_binomial_op.cc
@@ -326,7 +326,7 @@
template <typename Device, typename T, typename U>
class RandomBinomialOp : public OpKernel {
// Reshape batches so each batch is this size if possible.
- static const int32 kDesiredBatchSize = 100;
+ static constexpr int32 kDesiredBatchSize = 100;
public:
explicit RandomBinomialOp(OpKernelConstruction* context)
@@ -439,7 +439,7 @@
template <typename Device, typename T, typename U>
class StatelessRandomBinomialOp : public OpKernel {
// Reshape batches so each batch is this size if possible.
- static const int32 kDesiredBatchSize = 100;
+ static constexpr int32 kDesiredBatchSize = 100;
public:
explicit StatelessRandomBinomialOp(OpKernelConstruction* context)
diff --git a/tensorflow/core/kernels/random_op.cc b/tensorflow/core/kernels/random_op.cc
index 7b7f515..152ab5f 100644
--- a/tensorflow/core/kernels/random_op.cc
+++ b/tensorflow/core/kernels/random_op.cc
@@ -208,6 +208,7 @@
alpha_flat](int start_output, int limit_output) {
using Eigen::numext::exp;
using Eigen::numext::log;
+ using Eigen::numext::log1p;
using Eigen::numext::pow;
// Capturing "rng" by-value would only make a copy for the _shared_
@@ -241,7 +242,7 @@
gen.Skip(kReservedSamplesPerOutput * output_idx);
int16 uniform_remaining = 0;
UNIFORM(u);
- const double res = -log(1.0 - u);
+ const double res = -log1p(-u);
samples_alpha_offset[sample_idx * num_alphas] = static_cast<T>(res);
} // for (sample_idx)
} else { // if alpha != 1.0
diff --git a/tensorflow/core/kernels/random_op_cpu.h b/tensorflow/core/kernels/random_op_cpu.h
index 0e24bc3..eac1fae 100644
--- a/tensorflow/core/kernels/random_op_cpu.h
+++ b/tensorflow/core/kernels/random_op_cpu.h
@@ -111,7 +111,7 @@
template <class Distribution>
struct FillPhiloxRandomTask<Distribution, true> {
typedef typename Distribution::ResultElementType T;
- static const int64 kReservedSamplesPerOutput = 256;
+ static constexpr int64 kReservedSamplesPerOutput = 256;
static void Run(random::PhiloxRandom base_gen, T* data, int64 size,
int64 start_group, int64 limit_group, Distribution dist) {
diff --git a/tensorflow/core/kernels/resize_bicubic_op_test.cc b/tensorflow/core/kernels/resize_bicubic_op_test.cc
index 9af5626..9646461 100644
--- a/tensorflow/core/kernels/resize_bicubic_op_test.cc
+++ b/tensorflow/core/kernels/resize_bicubic_op_test.cc
@@ -58,7 +58,7 @@
}
private:
- static const int64 kTableSize = (1 << 10);
+ static constexpr int64 kTableSize = (1 << 10);
const float* InitCoeffsTable() {
// Allocate and initialize coefficients table using Bicubic
diff --git a/tensorflow/core/kernels/scan_ops.h b/tensorflow/core/kernels/scan_ops.h
index 1fd98f6..8afcac8 100644
--- a/tensorflow/core/kernels/scan_ops.h
+++ b/tensorflow/core/kernels/scan_ops.h
@@ -24,6 +24,7 @@
typedef Eigen::Index Index;
+// TODO(b/154339590): Needs to be vectorized.
template <typename Device, typename Reducer, typename T>
struct Scan {
void operator()(const Device& d, typename TTypes<T, 3>::ConstTensor in,
@@ -44,18 +45,33 @@
struct LogSumExp {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& a,
const T& b) const {
- Eigen::internal::scalar_sum_op<T> sum_op;
- Eigen::internal::scalar_exp_op<T> exp_op;
- Eigen::internal::scalar_log_op<T> log_op;
- Eigen::internal::scalar_max_op<T> max_op;
- Eigen::internal::scalar_min_op<T> min_op;
- Eigen::internal::scalar_log1p_op<T> log1p_op;
- Eigen::internal::scalar_difference_op<T> diff_op;
+ auto mi = Eigen::internal::scalar_min_op<T>()(a, b);
+ auto ma = Eigen::internal::scalar_max_op<T>()(a, b);
- auto mi = min_op(a, b);
- auto ma = max_op(a, b);
+ auto sub = Eigen::internal::scalar_difference_op<T>();
+ auto add = Eigen::internal::scalar_sum_op<T>();
+ auto exp = Eigen::internal::scalar_exp_op<T>();
+ auto log1p = Eigen::internal::scalar_log1p_op<T>();
+ auto cmp_lt =
+ Eigen::internal::scalar_cmp_op<T, T, Eigen::internal::cmp_LT>();
- return sum_op(log1p_op(exp_op(diff_op(mi, ma))), ma);
+ auto logsumexp = add(log1p(exp(sub(mi, ma))), ma);
+ return cmp_lt(ma, Eigen::NumTraits<T>::lowest()) ? ma : logsumexp;
+ }
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const T& a,
+ const T& b) const {
+ auto mi = Eigen::internal::pmin(a, b);
+ auto ma = Eigen::internal::pmax(a, b);
+ using Eigen::internal::padd;
+ using Eigen::internal::pcmp_lt;
+ using Eigen::internal::pexp;
+ using Eigen::internal::plog1p;
+ using Eigen::internal::pset1;
+ using Eigen::internal::psub;
+
+ auto logsumexp = padd(plog1p(pexp(psub(mi, ma))), ma);
+ return pselect(pcmp_lt(ma, pset1(Eigen::NumTraits<T>::lowest())), ma,
+ logsumexp);
}
};
@@ -66,13 +82,58 @@
*accum = logsumexp(*accum, t);
}
+ template <typename Packet>
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p,
+ Packet* accum) const {
+ LogSumExp<T> logsumexp;
+ *accum = logsumexp.packetOp(*accum, p);
+ }
+
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const {
- return Eigen::NumTraits<T>::lowest();
+ return -Eigen::NumTraits<T>::infinity();
+ }
+
+ template <typename Packet>
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet initializePacket() const {
+ return Eigen::internal::pset1(initialize());
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const {
return accum;
}
+
+ template <typename Packet>
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
+ finalizePacket(const Packet& vaccum) const {
+ return vaccum;
+ }
+
+ template <typename Packet>
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T
+ finalizeBoth(const T saccum, const Packet& vaccum) const {
+ auto max_reducer = Eigen::internal::MaxReducer<T>();
+ auto sum_reducer = Eigen::internal::SumReducer<T>();
+ auto exp = Eigen::internal::scalar_exp_op<T>();
+ auto cmp_lt =
+ Eigen::internal::scalar_cmp_op<T, T, Eigen::internal::cmp_LT>();
+ auto log = Eigen::internal::scalar_log_op<T>();
+ auto add = Eigen::internal::scalar_sum_op<T>();
+
+ using Eigen::internal::pexp;
+ using Eigen::internal::psub;
+
+ // `ma = max(x1, ..., xn)`
+ // If the max of all of the `xi` is `-infinity` then the result is
+ // -infinity. If the max is larger than `-infinity` then it's safe to use
+ // for normalization even if the other elements are `-infinity`.
+ //
+ // `logsumexp(x1, ..., xn) = ma + log (exp(x1 - ma) + ... + exp(xn - ma))`
+ auto ma = max_reducer.finalizeBoth(saccum, vaccum);
+ auto logsumexp = add(log(sum_reducer.finalizeBoth(
+ exp(saccum - ma), pexp(psub(vaccum, pset1(ma))))),
+ ma);
+ return cmp_lt(ma, Eigen::NumTraits<T>::lowest()) ? initialize() : logsumexp;
+ }
};
} // namespace functor
diff --git a/tensorflow/core/kernels/segment_reduction_ops_impl.h b/tensorflow/core/kernels/segment_reduction_ops_impl.h
index ba75150..8954dcd 100644
--- a/tensorflow/core/kernels/segment_reduction_ops_impl.h
+++ b/tensorflow/core/kernels/segment_reduction_ops_impl.h
@@ -425,7 +425,13 @@
// Same as SegmentReductionOp but takes as input a "sparse" tensor, represented
// by two dense tensors, one containing the data, and the other containing
// indices into the data.
-template <typename Device, class T>
+//
+// The template parameters are:
+// * Device: An Eigen device object, on which the kernel will execute.
+// * T: The value type.
+// * Index: The element type of the indices tensor (int32 or int64).
+// * SegmentId: The element type of the segment_ids tensor (int32 or int64).
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionOpBase : public OpKernel {
public:
explicit SparseSegmentReductionOpBase(OpKernelConstruction* context,
@@ -468,11 +474,10 @@
auto input_flat = input.flat_outer_dims<T>();
const int64 num_col = input_flat.dimension(1);
const auto indices_vec = indices.vec<Index>();
- typedef int32 OutputRow;
- const auto segment_vec = segment_ids.vec<OutputRow>();
+ const auto segment_vec = segment_ids.vec<SegmentId>();
// Note that the current implementation assumes that segment_vec values are
// sorted.
- const OutputRow last_segment_id_plus_one =
+ const SegmentId last_segment_id_plus_one =
num_indices > 0
? internal::SubtleMustCopy(segment_vec(num_indices - 1)) + 1
: 0;
@@ -505,14 +510,14 @@
int64 start = 0, end = 1;
// Index from which the output is not initialized.
- OutputRow uninitialized_index = 0;
- OutputRow out_index = internal::SubtleMustCopy(segment_vec(start));
+ SegmentId uninitialized_index = 0;
+ SegmentId out_index = internal::SubtleMustCopy(segment_vec(start));
while (true) {
// We initialize next_index to 0 to avoid "warning: 'next_index' may be
// used uninitialized in this function" in the Mac build (since the
// compiler isn't smart enough to realize the code is safe).
- OutputRow next_index = 0;
+ SegmentId next_index = 0;
if (end < num_indices) {
next_index = internal::SubtleMustCopy(segment_vec(end));
if (out_index == next_index) {
@@ -567,8 +572,6 @@
}
private:
- typedef int32 Index;
-
int64 Reduce(const typename TTypes<T>::ConstMatrix& input_flat,
const typename TTypes<Index>::ConstVec& indices_vec, int64 start,
int64 num,
@@ -702,70 +705,78 @@
const T default_value_;
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionMeanOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionMeanOp(OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, true /*is_mean*/, false /*is_sqrtn*/,
false /* has_num_segments */, T(0) /* default_value */) {}
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionMeanWithNumSegmentsOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionMeanWithNumSegmentsOp(
OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, true /*is_mean*/, false /*is_sqrtn*/,
true /* has_num_segments */, T(0) /* default_value */) {}
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionSqrtNOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionSqrtNOp(OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, false /*is_mean*/, true /*is_sqrtn*/,
false /* has_num_segments */, T(0) /* default_value */) {}
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionSqrtNWithNumSegmentsOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionSqrtNWithNumSegmentsOp(
OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, false /*is_mean*/, true /*is_sqrtn*/,
true /* has_num_segments */, T(0) /* default_value */) {}
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionSumOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionSumOp(OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, false /*is_mean*/, false /*is_sqrtn*/,
false /* has_num_segments */, T(0) /* default_value */) {}
};
-template <typename Device, class T>
+template <typename Device, class T, typename Index, typename SegmentId>
class SparseSegmentReductionSumWithNumSegmentsOp
- : public SparseSegmentReductionOpBase<Device, T> {
+ : public SparseSegmentReductionOpBase<Device, T, Index, SegmentId> {
public:
explicit SparseSegmentReductionSumWithNumSegmentsOp(
OpKernelConstruction* context)
- : SparseSegmentReductionOpBase<Device, T>(
+ : SparseSegmentReductionOpBase<Device, T, Index, SegmentId>(
context, false /*is_mean*/, false /*is_sqrtn*/,
true /* has_num_segments */, T(0) /* default_value */) {}
};
-template <class T>
+// Implements the common logic for the gradients of SparseSegmentReduction
+// kernels.
+//
+// The template parameters are:
+// * Device: An Eigen device object, on which the kernel will execute.
+// * T: The value type.
+// * Index: The element type of the indices tensor (int32 or int64).
+// * SegmentId: The element type of the segment_ids tensor (int32 or int64).
+template <class T, typename Index, typename SegmentId>
class SparseSegmentGradOpBase : public OpKernel {
public:
explicit SparseSegmentGradOpBase(OpKernelConstruction* context, bool is_sqrtn)
@@ -788,12 +799,9 @@
OP_REQUIRES(context, N == segment_ids.NumElements(),
errors::InvalidArgument(
"segment_ids and indices should have same size."));
- typedef int32 SegmentId;
- const SegmentId M =
- internal::SubtleMustCopy(output_dim0.scalar<SegmentId>()());
+ const SegmentId M = internal::SubtleMustCopy(output_dim0.scalar<int32>()());
auto input_flat = input.flat_outer_dims<T>();
- typedef int32 Index;
const auto indices_vec = indices.vec<Index>();
const auto segment_vec = segment_ids.vec<SegmentId>();
@@ -871,18 +879,22 @@
const bool is_sqrtn_;
};
-template <class T>
-class SparseSegmentMeanGradOp : public SparseSegmentGradOpBase<T> {
+template <class T, typename Index, typename SegmentId>
+class SparseSegmentMeanGradOp
+ : public SparseSegmentGradOpBase<T, Index, SegmentId> {
public:
explicit SparseSegmentMeanGradOp(OpKernelConstruction* context)
- : SparseSegmentGradOpBase<T>(context, false /*is_sqrtn*/) {}
+ : SparseSegmentGradOpBase<T, Index, SegmentId>(context,
+ false /*is_sqrtn*/) {}
};
-template <class T>
-class SparseSegmentSqrtNGradOp : public SparseSegmentGradOpBase<T> {
+template <class T, typename Index, typename SegmentId>
+class SparseSegmentSqrtNGradOp
+ : public SparseSegmentGradOpBase<T, Index, SegmentId> {
public:
explicit SparseSegmentSqrtNGradOp(OpKernelConstruction* context)
- : SparseSegmentGradOpBase<T>(context, true /*is_sqrtn*/) {}
+ : SparseSegmentGradOpBase<T, Index, SegmentId>(context,
+ true /*is_sqrtn*/) {}
};
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/segment_reduction_ops_impl_5.cc b/tensorflow/core/kernels/segment_reduction_ops_impl_5.cc
index 638c698..fee0f81 100644
--- a/tensorflow/core/kernels/segment_reduction_ops_impl_5.cc
+++ b/tensorflow/core/kernels/segment_reduction_ops_impl_5.cc
@@ -18,71 +18,100 @@
namespace tensorflow {
-#define REGISTER_CPU_SPARSE_KERNELS(type) \
- REGISTER_KERNEL_BUILDER(Name("SparseSegmentSum") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionSumOp<CPUDevice, type>); \
- REGISTER_KERNEL_BUILDER( \
- Name("SparseSegmentSumWithNumSegments") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionSumWithNumSegmentsOp<CPUDevice, type>);
-TF_CALL_REAL_NUMBER_TYPES(REGISTER_CPU_SPARSE_KERNELS);
+#define REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_SEGMENT_ID_TYPE(type, index_type) \
+ REGISTER_CPU_SPARSE_KERNELS(type, index_type, int32) \
+ REGISTER_CPU_SPARSE_KERNELS(type, index_type, int64)
+#define REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(type) \
+ REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_SEGMENT_ID_TYPE(type, int32) \
+ REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_SEGMENT_ID_TYPE(type, int64)
+
+#define REGISTER_CPU_SPARSE_KERNELS(type, index_type, segment_ids_type) \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentSum") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionSumOp<CPUDevice, type, index_type, \
+ segment_ids_type>); \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentSumWithNumSegments") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionSumWithNumSegmentsOp<CPUDevice, type, index_type, \
+ segment_ids_type>);
+TF_CALL_REAL_NUMBER_TYPES(REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE);
#undef REGISTER_CPU_SPARSE_KERNELS
-#define REGISTER_CPU_SPARSE_KERNELS(type) \
- REGISTER_KERNEL_BUILDER(Name("SparseSegmentMean") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionMeanOp<CPUDevice, type>); \
- REGISTER_KERNEL_BUILDER( \
- Name("SparseSegmentMeanWithNumSegments") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionMeanWithNumSegmentsOp<CPUDevice, type>);
-REGISTER_CPU_SPARSE_KERNELS(float);
-REGISTER_CPU_SPARSE_KERNELS(double);
+#define REGISTER_CPU_SPARSE_KERNELS(type, index_type, segment_ids_type) \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentMean") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionMeanOp<CPUDevice, type, index_type, \
+ segment_ids_type>); \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentMeanWithNumSegments") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionMeanWithNumSegmentsOp<CPUDevice, type, index_type, \
+ segment_ids_type>);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(float);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(double);
#undef REGISTER_CPU_SPARSE_KERNELS
-#define REGISTER_CPU_SPARSE_KERNELS(type) \
- REGISTER_KERNEL_BUILDER(Name("SparseSegmentSqrtN") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionSqrtNOp<CPUDevice, type>); \
- REGISTER_KERNEL_BUILDER( \
- Name("SparseSegmentSqrtNWithNumSegments") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentReductionSqrtNWithNumSegmentsOp<CPUDevice, type>);
-REGISTER_CPU_SPARSE_KERNELS(float);
-REGISTER_CPU_SPARSE_KERNELS(double);
+#define REGISTER_CPU_SPARSE_KERNELS(type, index_type, segment_ids_type) \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentSqrtN") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionSqrtNOp<CPUDevice, type, index_type, \
+ segment_ids_type>); \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentSqrtNWithNumSegments") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentReductionSqrtNWithNumSegmentsOp< \
+ CPUDevice, type, index_type, segment_ids_type>);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(float);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(double);
#undef REGISTER_CPU_SPARSE_KERNELS
-#define REGISTER_CPU_SPARSE_KERNELS(type) \
- REGISTER_KERNEL_BUILDER(Name("SparseSegmentMeanGrad") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentMeanGradOp<type>);
-REGISTER_CPU_SPARSE_KERNELS(float);
-REGISTER_CPU_SPARSE_KERNELS(double);
+#define REGISTER_CPU_SPARSE_KERNELS(type, index_type, segment_ids_type) \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentMeanGrad") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentMeanGradOp<type, index_type, segment_ids_type>);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(float);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(double);
#undef REGISTER_CPU_SPARSE_KERNELS
-#define REGISTER_CPU_SPARSE_KERNELS(type) \
- REGISTER_KERNEL_BUILDER(Name("SparseSegmentSqrtNGrad") \
- .Device(DEVICE_CPU) \
- .TypeConstraint<type>("T") \
- .TypeConstraint<int32>("Tidx"), \
- SparseSegmentSqrtNGradOp<type>);
-REGISTER_CPU_SPARSE_KERNELS(float);
-REGISTER_CPU_SPARSE_KERNELS(double);
+#define REGISTER_CPU_SPARSE_KERNELS(type, index_type, segment_ids_type) \
+ REGISTER_KERNEL_BUILDER( \
+ Name("SparseSegmentSqrtNGrad") \
+ .Device(DEVICE_CPU) \
+ .TypeConstraint<type>("T") \
+ .TypeConstraint<index_type>("Tidx") \
+ .TypeConstraint<segment_ids_type>("Tsegmentids"), \
+ SparseSegmentSqrtNGradOp<type, index_type, segment_ids_type>);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(float);
+REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE(double);
#undef REGISTER_CPU_SPARSE_KERNELS
+#undef REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_INDEX_TYPE
+#undef REGISTER_CPU_SPARSE_KERNELS_FOR_EACH_SEGMENT_ID_TYPE
+
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/sparse_matmul_op_test.cc b/tensorflow/core/kernels/sparse_matmul_op_test.cc
index ebc6d8f..1dc51cd 100644
--- a/tensorflow/core/kernels/sparse_matmul_op_test.cc
+++ b/tensorflow/core/kernels/sparse_matmul_op_test.cc
@@ -311,7 +311,7 @@
#elif defined EIGEN_VECTORIZE_AVX || defined EIGEN_VECTORIZE_AVX2
static const int kMaxPacketSize = 8;
#else
- static const int kMaxPacketSize = 4;
+ static constexpr int kMaxPacketSize = 4;
#endif
typedef typename Eigen::internal::packet_traits<float>::type Packet;
const int PacketSize;
diff --git a/tensorflow/core/kernels/sparse_tensor_dense_matmul_op.cc b/tensorflow/core/kernels/sparse_tensor_dense_matmul_op.cc
index d2dce3f..9baaa6e 100644
--- a/tensorflow/core/kernels/sparse_tensor_dense_matmul_op.cc
+++ b/tensorflow/core/kernels/sparse_tensor_dense_matmul_op.cc
@@ -244,7 +244,7 @@
template <typename T, typename Tindices, bool ADJ_A, bool ADJ_B>
struct SparseTensorDenseMatMulFunctor<CPUDevice, T, Tindices, ADJ_A, ADJ_B> {
// Vectorize certain operations above this size.
- static const std::size_t kNumVectorize = 32;
+ static constexpr std::size_t kNumVectorize = 32;
static Status Compute(const CPUDevice& d, typename TTypes<T>::Matrix out,
typename TTypes<Tindices>::ConstMatrix a_indices,
diff --git a/tensorflow/core/kernels/sparse_utils.cc b/tensorflow/core/kernels/sparse_utils.cc
index 1988629..678c94a 100644
--- a/tensorflow/core/kernels/sparse_utils.cc
+++ b/tensorflow/core/kernels/sparse_utils.cc
@@ -79,11 +79,14 @@
std::vector<Tindices> segment_indices;
const Tindices num_entries_in_sparse_tensor = indices_mat.dimension(0);
const Tindices num_dense_rows_in_sparse_tensor =
- 1 + indices_mat(num_entries_in_sparse_tensor - 1, 0) - indices_mat(0, 0);
+ 1 + indices_mat(num_entries_in_sparse_tensor - 1, 0);
// Reserve an extra slot for the 0 we store in the first entry by convention.
segment_indices.reserve(1 + num_dense_rows_in_sparse_tensor);
segment_indices.push_back(0);
- *contains_empty_rows = false;
+ for (Tindices i = 0; i < indices_mat(0, 0); ++i) {
+ segment_indices.push_back(0);
+ }
+ *contains_empty_rows = indices_mat(0, 0) > 0;
while (true) {
const Tindices start_sparse_index_of_next_dense_row =
FindNextDenseRowStartIndex<Tindices>(
@@ -127,9 +130,9 @@
template <typename Tindices>
bool ContainsEmptyRows(const std::vector<Tindices>& row_start_indices) {
- // Skip checking the lengths of the first and last dense rows since those are
+ // Skip checking the length of the last dense row since it is
// always non-empty.
- for (size_t i = 2; i < row_start_indices.size() - 1; ++i) {
+ for (size_t i = 1; i < row_start_indices.size() - 1; ++i) {
if (row_start_indices.at(i) - row_start_indices.at(i - 1) == 0) {
return true;
}
diff --git a/tensorflow/core/kernels/sparse_utils.h b/tensorflow/core/kernels/sparse_utils.h
index 9e3c41a..d43b2e3 100644
--- a/tensorflow/core/kernels/sparse_utils.h
+++ b/tensorflow/core/kernels/sparse_utils.h
@@ -44,7 +44,7 @@
// v.front() = 0, v.back() = indices_mat.dimension(0), and for i > 0,
// v[i] - v[i-1] is the length of the ith dense row in indices_mat.
// *contains_empty_rows = true if and only if indices_mat contains empty rows
-// (rows without values) between its first and last row.
+// (rows without values) between row 0 and the last row.
template <typename Tindices>
std::vector<Tindices> GetStartIndicesOfEachDenseRow(
const typename TTypes<Tindices>::ConstMatrix& indices_mat,
diff --git a/tensorflow/core/kernels/sparse_utils_test.cc b/tensorflow/core/kernels/sparse_utils_test.cc
index 5d0adff..5b0e521 100644
--- a/tensorflow/core/kernels/sparse_utils_test.cc
+++ b/tensorflow/core/kernels/sparse_utils_test.cc
@@ -66,8 +66,8 @@
bool contains_empty_rows;
EXPECT_TRUE(GetStartIndicesOfEachDenseRow<int64>(indices_mat,
&contains_empty_rows) ==
- std::vector<int64>({0, 1}));
- EXPECT_FALSE(contains_empty_rows);
+ std::vector<int64>({0, 0, 0, 0, 1}));
+ EXPECT_TRUE(contains_empty_rows);
}
{
uint32 data[] = {3, 0, 3, 0};
@@ -75,8 +75,8 @@
bool contains_empty_rows;
EXPECT_TRUE(GetStartIndicesOfEachDenseRow<uint32>(indices_mat,
&contains_empty_rows) ==
- std::vector<uint32>({0, 2}));
- EXPECT_FALSE(contains_empty_rows);
+ std::vector<uint32>({0, 0, 0, 0, 2}));
+ EXPECT_TRUE(contains_empty_rows);
}
{
uint16 data[] = {0, 0, 0, 0, 0, 0, 1, 0};
@@ -165,7 +165,7 @@
const auto segment_indices =
GetStartIndicesOfEachDenseRow<int32>(indices_mat, &contains_empty_rows);
// indices_list = {1, 1, 2, 2, 2, 3};
- EXPECT_FALSE(ContainsEmptyRows(segment_indices));
+ EXPECT_TRUE(ContainsEmptyRows(segment_indices));
}
{
uint16 data[] = {1, 0, 1, 1, 2, 0, 2, 1, 2, 2, 3, 4};
@@ -174,7 +174,7 @@
const auto segment_indices = GetStartIndicesOfEachDenseRow<uint16>(
indices_mat, &contains_empty_rows);
// indices_list = {1, 1, 2, 2, 2, 3};
- EXPECT_FALSE(ContainsEmptyRows(segment_indices));
+ EXPECT_TRUE(ContainsEmptyRows(segment_indices));
}
{
int32 data[] = {0, 0, 1, 0, 1, 1, 2, 0, 2, 1, 2, 2, 3, 4};
diff --git a/tensorflow/core/kernels/stateless_random_ops.cc b/tensorflow/core/kernels/stateless_random_ops.cc
index 167daf2..6738a34 100644
--- a/tensorflow/core/kernels/stateless_random_ops.cc
+++ b/tensorflow/core/kernels/stateless_random_ops.cc
@@ -259,6 +259,7 @@
using Eigen::numext::exp;
using Eigen::numext::log;
+ using Eigen::numext::log1p;
using Eigen::numext::pow;
Normal normal;
@@ -288,7 +289,7 @@
gen.Skip(kReservedSamplesPerOutput * output_idx);
int16 uniform_remaining = 0;
UNIFORM(u);
- const double res = -log(1.0 - u);
+ const double res = -log1p(-u);
samples_alpha_offset[sample_idx * num_alphas] = static_cast<T>(res);
} // for (sample_idx)
} else { // if alpha != 1.0
diff --git a/tensorflow/core/kernels/unary_ops_composition.cc b/tensorflow/core/kernels/unary_ops_composition.cc
index 0b76a60..dc8c9b7 100644
--- a/tensorflow/core/kernels/unary_ops_composition.cc
+++ b/tensorflow/core/kernels/unary_ops_composition.cc
@@ -137,7 +137,8 @@
}
private:
- static const int kPacketSize = Eigen::internal::unpacket_traits<Packet>::size;
+ static constexpr int kPacketSize =
+ Eigen::internal::unpacket_traits<Packet>::size;
static inline int64 AlignBlockSize(int64 block_size) {
// Align block size to packet size and account for unrolling in run above.
diff --git a/tensorflow/core/kernels/unique_op.cc b/tensorflow/core/kernels/unique_op.cc
index 8a9965f..8316018 100644
--- a/tensorflow/core/kernels/unique_op.cc
+++ b/tensorflow/core/kernels/unique_op.cc
@@ -17,18 +17,62 @@
#include <unordered_map>
#include <utility>
+#include "absl/container/flat_hash_map.h"
#include "tensorflow/core/framework/bounds_check.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
+#include "tensorflow/core/lib/bfloat16/bfloat16.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/lib/hash/hash.h"
namespace tensorflow {
+namespace {
typedef Eigen::ThreadPoolDevice CPUDevice;
+// `UniqueOpHashMap` defines the map type that is used when elements of type
+// `T` are to be uniquified. By default, we use `absl::flat_hash_map<T, TIndex>`
+// as the map type. Subsequent specializations are provided for
+// performance and/or correctness.
+template <typename T, typename TIndex>
+struct UniqueOpHashMap {
+ using map_type = absl::flat_hash_map<T, TIndex>;
+};
+
+// NOTE(mrry): For `tstring` elements, we use an `absl::string_view` key to
+// avoid copying the input strings into the map.
+template <typename TIndex>
+struct UniqueOpHashMap<tstring, TIndex> {
+ using map_type = absl::flat_hash_map<absl::string_view, TIndex>;
+};
+
+// NOTE(mrry): `absl::flat_hash_map<float, ...>` does not allow `NaN` as a key,
+// because `NaN != NaN`, so we fall back to `std::unordered_map<>` for
+// floating-point types.
+template <typename TIndex>
+struct UniqueOpHashMap<float, TIndex> {
+ using map_type = std::unordered_map<float, TIndex>;
+};
+template <typename TIndex>
+struct UniqueOpHashMap<double, TIndex> {
+ using map_type = std::unordered_map<double, TIndex>;
+};
+template <typename TIndex>
+struct UniqueOpHashMap<Eigen::half, TIndex> {
+ using map_type = std::unordered_map<Eigen::half, TIndex>;
+};
+template <typename TIndex>
+struct UniqueOpHashMap<bfloat16, TIndex> {
+ using map_type = std::unordered_map<bfloat16, TIndex>;
+};
+
+// `UniqueOp` computes the unique elements in the input tensor.
+//
+// * `T` is the element type.
+// * `TIndex` is the type used to represent indices in the output, either
+// `int32` or `int64`.
template <typename T, typename TIndex>
class UniqueOp : public OpKernel {
public:
@@ -106,10 +150,10 @@
auto Tin = input.flat<T>();
const int64 N = static_cast<int64>(Tin.size());
- std::unordered_map<T, TIndex> uniq;
+ typename UniqueOpHashMap<T, TIndex>::map_type uniq;
uniq.reserve(2 * N);
for (Eigen::Index i = 0, j = 0; i < N; ++i) {
- auto it = uniq.insert(std::make_pair(Tin(i), j));
+ auto it = uniq.emplace(Tin(i), j);
idx_vec(i) = it.first->second;
if (it.second) {
++j;
@@ -153,13 +197,14 @@
return true;
};
- std::unordered_map<int64, int64, decltype(hash_fn), decltype(equal_to_fn)>
+ absl::flat_hash_map<int64, int64, decltype(hash_fn),
+ decltype(equal_to_fn)>
uniq(0, hash_fn, equal_to_fn);
uniq.reserve(2 * Tin.dimension(1));
for (int64 i = 0, j = 0; i < Tin.dimension(1); ++i) {
- auto it = uniq.insert(std::make_pair(i, j));
+ auto it = uniq.emplace(i, j);
idx_vec(i) = it.first->second;
if (it.second) {
++j;
@@ -311,4 +356,6 @@
.HostMemory("idx"),
UniqueOp<int64, int64>);
#endif // TENSORFLOW_USE_SYCL
+
+} // namespace
} // namespace tensorflow
diff --git a/tensorflow/core/kernels/unique_op_test.cc b/tensorflow/core/kernels/unique_op_test.cc
index 4861a45..a0249d9 100644
--- a/tensorflow/core/kernels/unique_op_test.cc
+++ b/tensorflow/core/kernels/unique_op_test.cc
@@ -22,6 +22,7 @@
#include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/framework/types.pb.h"
+#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/graph/node_builder.h"
#include "tensorflow/core/graph/testlib.h"
#include "tensorflow/core/kernels/ops_testutil.h"
@@ -75,11 +76,14 @@
.Input(test::graph::Constant(g, input))
.Attr("T", DT_INT32)
.Finalize(g, &node));
+ FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(int32));
testing::UseRealTime();
testing::StartTiming();
- test::Benchmark("cpu", g).Run(iters);
+ test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
+ "SINGLE_THREADED_EXECUTOR")
+ .Run(iters);
}
static void BM_Unique_INT32_Repeat(int iters, int dim, int max_int) {
@@ -95,12 +99,15 @@
.Input(test::graph::Constant(g, input))
.Attr("T", DT_INT32)
.Finalize(g, &node));
+ FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * 200 *
sizeof(int32));
testing::UseRealTime();
testing::StartTiming();
- test::Benchmark("cpu", g).Run(iters);
+ test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
+ "SINGLE_THREADED_EXECUTOR")
+ .Run(iters);
}
TensorProto GetRandomStringsTensorProto(int dim, int max_str_len) {
@@ -132,11 +139,14 @@
.Input(test::graph::Constant(g, input))
.Attr("T", DT_STRING)
.Finalize(g, &node));
+ FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(tstring));
testing::UseRealTime();
testing::StartTiming();
- test::Benchmark("cpu", g).Run(iters);
+ test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
+ "SINGLE_THREADED_EXECUTOR")
+ .Run(iters);
}
BENCHMARK(BM_Unique_INT32)
diff --git a/tensorflow/core/lib/bfloat16/bfloat16.h b/tensorflow/core/lib/bfloat16/bfloat16.h
index e3bbff6..4c38738 100644
--- a/tensorflow/core/lib/bfloat16/bfloat16.h
+++ b/tensorflow/core/lib/bfloat16/bfloat16.h
@@ -392,11 +392,11 @@
uint16_t value;
// A value that represents "not a number".
- static const uint16_t NAN_VALUE = 0x7FC0;
+ static constexpr uint16_t NAN_VALUE = 0x7FC0;
private:
// A value that represents "zero".
- static const uint16_t ZERO_VALUE = 0;
+ static constexpr uint16_t ZERO_VALUE = 0;
B16_DEVICE_FUNC static bool float_isnan(const float& x) {
#ifdef __CUDA_ARCH__
diff --git a/tensorflow/core/lib/core/arena.h b/tensorflow/core/lib/core/arena.h
index 624ee77..4b791af 100644
--- a/tensorflow/core/lib/core/arena.h
+++ b/tensorflow/core/lib/core/arena.h
@@ -57,7 +57,7 @@
#ifdef __i386__
static const int kDefaultAlignment = 4;
#else
- static const int kDefaultAlignment = 8;
+ static constexpr int kDefaultAlignment = 8;
#endif
protected:
diff --git a/tensorflow/core/lib/core/bitmap.h b/tensorflow/core/lib/core/bitmap.h
index 8ff1e66..726c647 100644
--- a/tensorflow/core/lib/core/bitmap.h
+++ b/tensorflow/core/lib/core/bitmap.h
@@ -63,7 +63,7 @@
private:
typedef uint32 Word;
- static const size_t kBits = 32;
+ static constexpr size_t kBits = 32;
// Return the number of words needed to store n bits.
static size_t NumWords(size_t n) { return (n + kBits - 1) / kBits; }
diff --git a/tensorflow/core/lib/gtl/flatrep.h b/tensorflow/core/lib/gtl/flatrep.h
index 65a076b..d9d9622 100644
--- a/tensorflow/core/lib/gtl/flatrep.h
+++ b/tensorflow/core/lib/gtl/flatrep.h
@@ -45,8 +45,8 @@
class FlatRep {
public:
// kWidth is the number of entries stored in a bucket.
- static const uint32 kBase = 3;
- static const uint32 kWidth = (1 << kBase);
+ static constexpr uint32 kBase = 3;
+ static constexpr uint32 kWidth = (1 << kBase);
FlatRep(size_t N, const Hash& hf, const Eq& eq) : hash_(hf), equal_(eq) {
Init(N);
diff --git a/tensorflow/core/lib/io/cache_test.cc b/tensorflow/core/lib/io/cache_test.cc
index 38552d4..002ab0b 100644
--- a/tensorflow/core/lib/io/cache_test.cc
+++ b/tensorflow/core/lib/io/cache_test.cc
@@ -44,7 +44,7 @@
current_->deleted_values_.push_back(DecodeValue(v));
}
- static const int kCacheSize = 1000;
+ static constexpr int kCacheSize = 1000;
std::vector<int> deleted_keys_;
std::vector<int> deleted_values_;
Cache* cache_;
diff --git a/tensorflow/core/lib/io/record_reader.h b/tensorflow/core/lib/io/record_reader.h
index d1453e7..dd7def7 100644
--- a/tensorflow/core/lib/io/record_reader.h
+++ b/tensorflow/core/lib/io/record_reader.h
@@ -63,8 +63,8 @@
// uint32 masked crc of length
// byte data[length]
// uint32 masked crc of data
- static const size_t kHeaderSize = sizeof(uint64) + sizeof(uint32);
- static const size_t kFooterSize = sizeof(uint32);
+ static constexpr size_t kHeaderSize = sizeof(uint64) + sizeof(uint32);
+ static constexpr size_t kFooterSize = sizeof(uint32);
// Statistics (sizes are in units of bytes)
struct Stats {
diff --git a/tensorflow/core/lib/io/record_writer.h b/tensorflow/core/lib/io/record_writer.h
index dba4d75..012c2fb 100644
--- a/tensorflow/core/lib/io/record_writer.h
+++ b/tensorflow/core/lib/io/record_writer.h
@@ -55,8 +55,8 @@
// uint32 masked crc of length
// byte data[length]
// uint32 masked crc of data
- static const size_t kHeaderSize = sizeof(uint64) + sizeof(uint32);
- static const size_t kFooterSize = sizeof(uint32);
+ static constexpr size_t kHeaderSize = sizeof(uint64) + sizeof(uint32);
+ static constexpr size_t kFooterSize = sizeof(uint32);
// Create a writer that will append data to "*dest".
// "*dest" must be initially empty.
diff --git a/tensorflow/core/lib/random/philox_random.h b/tensorflow/core/lib/random/philox_random.h
index ee32e40..1d419f1 100644
--- a/tensorflow/core/lib/random/philox_random.h
+++ b/tensorflow/core/lib/random/philox_random.h
@@ -49,7 +49,7 @@
template <typename T, int ElementCount>
class Array {
public:
- static const int kElementCount = ElementCount;
+ static constexpr int kElementCount = ElementCount;
PHILOX_DEVICE_INLINE Array() {
for (int i = 0; i < ElementCount; ++i) {
data_[i] = T(0);
@@ -105,9 +105,9 @@
using ResultType = Array<uint32, 4>;
using ResultElementType = uint32;
// The number of elements that will be returned.
- static const int kResultElementCount = 4;
+ static constexpr int kResultElementCount = 4;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 10;
+ static constexpr int kElementCost = 10;
// The type for the 64-bit key stored in the form of two 32-bit uint
// that are used in the diffusion process.
using Key = Array<uint32, 2>;
@@ -192,10 +192,10 @@
private:
// We use the same constants as recommended by the original paper.
- static const uint32 kPhiloxW32A = 0x9E3779B9;
- static const uint32 kPhiloxW32B = 0xBB67AE85;
- static const uint32 kPhiloxM4x32A = 0xD2511F53;
- static const uint32 kPhiloxM4x32B = 0xCD9E8D57;
+ static constexpr uint32 kPhiloxW32A = 0x9E3779B9;
+ static constexpr uint32 kPhiloxW32B = 0xBB67AE85;
+ static constexpr uint32 kPhiloxM4x32A = 0xD2511F53;
+ static constexpr uint32 kPhiloxM4x32B = 0xCD9E8D57;
// Helper function to skip the next sample of 128-bits in the current stream.
PHILOX_DEVICE_INLINE void SkipOne() {
diff --git a/tensorflow/core/lib/random/random_distributions.h b/tensorflow/core/lib/random/random_distributions.h
index 2da5e29..386f133 100644
--- a/tensorflow/core/lib/random/random_distributions.h
+++ b/tensorflow/core/lib/random/random_distributions.h
@@ -71,12 +71,12 @@
class UniformDistribution<Generator, Eigen::half> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<Eigen::half, kResultElementCount> ResultType;
typedef Eigen::half ResultElementType;
@@ -95,12 +95,12 @@
class UniformDistribution<Generator, bfloat16> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<bfloat16, kResultElementCount> ResultType;
typedef bfloat16 ResultElementType;
@@ -119,12 +119,12 @@
class UniformDistribution<Generator, float> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<float, kResultElementCount> ResultType;
typedef float ResultElementType;
@@ -143,12 +143,12 @@
class UniformDistribution<Generator, double> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount / 2;
+ static constexpr int kResultElementCount = Generator::kResultElementCount / 2;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<double, kResultElementCount> ResultType;
typedef double ResultElementType;
@@ -167,12 +167,12 @@
class UniformDistribution<Generator, int32> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<int32, kResultElementCount> ResultType;
typedef int32 ResultElementType;
@@ -202,12 +202,12 @@
class UniformDistribution<Generator, int64> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount / 2;
+ static constexpr int kResultElementCount = Generator::kResultElementCount / 2;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<int64, kResultElementCount> ResultType;
typedef int64 ResultElementType;
@@ -244,12 +244,12 @@
class UniformFullIntDistribution32 {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<IntType, kResultElementCount> ResultType;
typedef IntType ResultElementType;
@@ -268,12 +268,12 @@
class UniformFullIntDistribution64 {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount / 2;
+ static constexpr int kResultElementCount = Generator::kResultElementCount / 2;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 3;
+ static constexpr int kElementCost = 3;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<IntType, kResultElementCount> ResultType;
typedef IntType ResultElementType;
@@ -307,9 +307,9 @@
class SingleSampleAdapter {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = 1;
+ static constexpr int kResultElementCount = 1;
// The number of elements that will be returned by the underlying generator.
- static const int kNativeElementCount = Generator::kResultElementCount;
+ static constexpr int kNativeElementCount = Generator::kResultElementCount;
typedef typename Generator::ResultElementType ResultType;
typedef typename Generator::ResultElementType ResultElementType;
@@ -391,12 +391,12 @@
class NormalDistribution<Generator, Eigen::half> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 70;
+ static constexpr int kElementCost = 70;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<Eigen::half, kResultElementCount> ResultType;
typedef Eigen::half ResultElementType;
@@ -418,12 +418,12 @@
class NormalDistribution<Generator, bfloat16> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 70;
+ static constexpr int kElementCost = 70;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<bfloat16, kResultElementCount> ResultType;
typedef bfloat16 ResultElementType;
@@ -448,12 +448,12 @@
class NormalDistribution<Generator, float> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount;
+ static constexpr int kResultElementCount = Generator::kResultElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 70;
+ static constexpr int kElementCost = 70;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<float, kResultElementCount> ResultType;
typedef float ResultElementType;
@@ -472,12 +472,12 @@
class NormalDistribution<Generator, double> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount = Generator::kResultElementCount / 2;
+ static constexpr int kResultElementCount = Generator::kResultElementCount / 2;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 70;
+ static constexpr int kElementCost = 70;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = false;
+ static constexpr bool kVariableSamplesPerOutput = false;
typedef Array<double, kResultElementCount> ResultType;
typedef double ResultElementType;
@@ -515,13 +515,13 @@
class TruncatedNormalDistribution<SingleSampleGenerator, Eigen::half> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount =
+ static constexpr int kResultElementCount =
SingleSampleGenerator::kNativeElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 90;
+ static constexpr int kElementCost = 90;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = true;
+ static constexpr bool kVariableSamplesPerOutput = true;
// The threshold where the normal distribution is truncated.
const float kTruncateValue = 2.0f;
@@ -561,13 +561,13 @@
class TruncatedNormalDistribution<SingleSampleGenerator, bfloat16> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount =
+ static constexpr int kResultElementCount =
SingleSampleGenerator::kNativeElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 90;
+ static constexpr int kElementCost = 90;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = true;
+ static constexpr bool kVariableSamplesPerOutput = true;
// The threshold where the normal distribution is truncated.
const float kTruncateValue = 2.0f;
@@ -608,13 +608,13 @@
class TruncatedNormalDistribution<SingleSampleGenerator, float> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount =
+ static constexpr int kResultElementCount =
SingleSampleGenerator::kNativeElementCount;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 90;
+ static constexpr int kElementCost = 90;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = true;
+ static constexpr bool kVariableSamplesPerOutput = true;
// The threshold where the normal distribution is truncated.
const float kTruncateValue = 2.0f;
@@ -655,15 +655,15 @@
class TruncatedNormalDistribution<SingleSampleGenerator, double> {
public:
// The number of elements that will be returned.
- static const int kResultElementCount =
+ static constexpr int kResultElementCount =
(SingleSampleGenerator::kNativeElementCount > 1)
? SingleSampleGenerator::kNativeElementCount / 2
: 1;
// Cost of generation of a single element (in cycles).
- static const int kElementCost = 90;
+ static constexpr int kElementCost = 90;
// Indicate that this distribution may take variable number of samples
// during the runtime.
- static const bool kVariableSamplesPerOutput = true;
+ static constexpr bool kVariableSamplesPerOutput = true;
typedef Array<double, kResultElementCount> ResultType;
typedef double ResultElementType;
const double kTruncateValue = 2.0;
diff --git a/tensorflow/core/lib/random/random_distributions_test.cc b/tensorflow/core/lib/random/random_distributions_test.cc
index a497316..13dfca7 100644
--- a/tensorflow/core/lib/random/random_distributions_test.cc
+++ b/tensorflow/core/lib/random/random_distributions_test.cc
@@ -307,7 +307,7 @@
explicit MockGenerator(uint64 seed) : counter_(seed) {}
using ResultType = std::vector<uint32>;
using ResultElementType = uint32;
- static const int kResultElementCount = 1;
+ static constexpr int kResultElementCount = 1;
ResultType operator()() {
ResultType result;
result.push_back(counter_++);
diff --git a/tensorflow/core/ops/compat/ops_history_v2/DataServiceDataset.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/DataServiceDataset.pbtxt
index 5fc666e..89db2e0 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/DataServiceDataset.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/DataServiceDataset.pbtxt
@@ -30,3 +30,42 @@
}
is_stateful: true
}
+op {
+ name: "DataServiceDataset"
+ input_arg {
+ name: "address"
+ type: DT_STRING
+ }
+ input_arg {
+ name: "protocol"
+ type: DT_STRING
+ }
+ input_arg {
+ name: "max_outstanding_requests"
+ type: DT_INT64
+ }
+ output_arg {
+ name: "handle"
+ type: DT_VARIANT
+ }
+ attr {
+ name: "task_refresh_interval_hint_ms"
+ type: "int"
+ default_value {
+ i: -1
+ }
+ }
+ attr {
+ name: "output_types"
+ type: "list(type)"
+ has_minimum: true
+ minimum: 1
+ }
+ attr {
+ name: "output_shapes"
+ type: "list(shape)"
+ has_minimum: true
+ minimum: 1
+ }
+ is_stateful: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/DummyMemoryCache.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/DummyMemoryCache.pbtxt
new file mode 100644
index 0000000..63901e2
--- /dev/null
+++ b/tensorflow/core/ops/compat/ops_history_v2/DummyMemoryCache.pbtxt
@@ -0,0 +1,8 @@
+op {
+ name: "DummyMemoryCache"
+ output_arg {
+ name: "handle"
+ type: DT_RESOURCE
+ }
+ is_stateful: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/DummySeedGenerator.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/DummySeedGenerator.pbtxt
new file mode 100644
index 0000000..585bc7c
--- /dev/null
+++ b/tensorflow/core/ops/compat/ops_history_v2/DummySeedGenerator.pbtxt
@@ -0,0 +1,8 @@
+op {
+ name: "DummySeedGenerator"
+ output_arg {
+ name: "handle"
+ type: DT_RESOURCE
+ }
+ is_stateful: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/Equal.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Equal.pbtxt
index 2ebe636..f84733a 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/Equal.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/Equal.pbtxt
@@ -164,3 +164,53 @@
}
is_commutative: true
}
+op {
+ name: "Equal"
+ input_arg {
+ name: "x"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "y"
+ type_attr: "T"
+ }
+ output_arg {
+ name: "z"
+ type: DT_BOOL
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_BFLOAT16
+ type: DT_HALF
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_UINT8
+ type: DT_INT8
+ type: DT_INT16
+ type: DT_INT32
+ type: DT_INT64
+ type: DT_UINT16
+ type: DT_UINT32
+ type: DT_UINT64
+ type: DT_COMPLEX64
+ type: DT_QUINT8
+ type: DT_QINT8
+ type: DT_QINT32
+ type: DT_STRING
+ type: DT_BOOL
+ type: DT_COMPLEX128
+ }
+ }
+ }
+ attr {
+ name: "incompatible_shape_error"
+ type: "bool"
+ default_value {
+ b: true
+ }
+ }
+ is_commutative: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/Maximum.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Maximum.pbtxt
index 47625b4..f0ac23b 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/Maximum.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/Maximum.pbtxt
@@ -146,3 +146,34 @@
}
}
}
+op {
+ name: "Maximum"
+ input_arg {
+ name: "x"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "y"
+ type_attr: "T"
+ }
+ output_arg {
+ name: "z"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_BFLOAT16
+ type: DT_HALF
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_UINT8
+ type: DT_INT16
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/Minimum.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/Minimum.pbtxt
index 7b33bf7..5920b63 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/Minimum.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/Minimum.pbtxt
@@ -146,3 +146,34 @@
}
}
}
+op {
+ name: "Minimum"
+ input_arg {
+ name: "x"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "y"
+ type_attr: "T"
+ }
+ output_arg {
+ name: "z"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_BFLOAT16
+ type: DT_HALF
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_UINT8
+ type: DT_INT16
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/NotEqual.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/NotEqual.pbtxt
index 3c9fcbc..0039b5e 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/NotEqual.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/NotEqual.pbtxt
@@ -164,3 +164,53 @@
}
is_commutative: true
}
+op {
+ name: "NotEqual"
+ input_arg {
+ name: "x"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "y"
+ type_attr: "T"
+ }
+ output_arg {
+ name: "z"
+ type: DT_BOOL
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_BFLOAT16
+ type: DT_HALF
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_UINT8
+ type: DT_INT8
+ type: DT_INT16
+ type: DT_INT32
+ type: DT_INT64
+ type: DT_UINT16
+ type: DT_UINT32
+ type: DT_UINT64
+ type: DT_COMPLEX64
+ type: DT_QUINT8
+ type: DT_QINT8
+ type: DT_QINT32
+ type: DT_STRING
+ type: DT_BOOL
+ type: DT_COMPLEX128
+ }
+ }
+ }
+ attr {
+ name: "incompatible_shape_error"
+ type: "bool"
+ default_value {
+ b: true
+ }
+ }
+ is_commutative: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/ShuffleDatasetV3.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/ShuffleDatasetV3.pbtxt
new file mode 100644
index 0000000..792a1b4
--- /dev/null
+++ b/tensorflow/core/ops/compat/ops_history_v2/ShuffleDatasetV3.pbtxt
@@ -0,0 +1,47 @@
+op {
+ name: "ShuffleDatasetV3"
+ input_arg {
+ name: "input_dataset"
+ type: DT_VARIANT
+ }
+ input_arg {
+ name: "buffer_size"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed2"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed_generator"
+ type: DT_RESOURCE
+ }
+ output_arg {
+ name: "handle"
+ type: DT_VARIANT
+ }
+ attr {
+ name: "reshuffle_each_iteration"
+ type: "bool"
+ default_value {
+ b: true
+ }
+ }
+ attr {
+ name: "output_types"
+ type: "list(type)"
+ has_minimum: true
+ minimum: 1
+ }
+ attr {
+ name: "output_shapes"
+ type: "list(shape)"
+ has_minimum: true
+ minimum: 1
+ }
+ is_stateful: true
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMean.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMean.pbtxt
index 0447e6f..a3fde86 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMean.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMean.pbtxt
@@ -40,3 +40,58 @@
}
}
}
+op {
+ name: "SparseSegmentMean"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanGrad.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanGrad.pbtxt
index c31439f..092e04a 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanGrad.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanGrad.pbtxt
@@ -44,3 +44,62 @@
}
}
}
+op {
+ name: "SparseSegmentMeanGrad"
+ input_arg {
+ name: "grad"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ input_arg {
+ name: "output_dim0"
+ type: DT_INT32
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanWithNumSegments.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanWithNumSegments.pbtxt
index ed3693a..2d1d816 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanWithNumSegments.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentMeanWithNumSegments.pbtxt
@@ -57,3 +57,75 @@
}
}
}
+op {
+ name: "SparseSegmentMeanWithNumSegments"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ input_arg {
+ name: "num_segments"
+ type_attr: "Tnumsegments"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tnumsegments"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtN.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtN.pbtxt
index f856480..6ab44de 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtN.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtN.pbtxt
@@ -40,3 +40,58 @@
}
}
}
+op {
+ name: "SparseSegmentSqrtN"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNGrad.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNGrad.pbtxt
index 569b5b8..7520a44 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNGrad.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNGrad.pbtxt
@@ -44,3 +44,62 @@
}
}
}
+op {
+ name: "SparseSegmentSqrtNGrad"
+ input_arg {
+ name: "grad"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ input_arg {
+ name: "output_dim0"
+ type: DT_INT32
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNWithNumSegments.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNWithNumSegments.pbtxt
index 753cfe4..038a5a2 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNWithNumSegments.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSqrtNWithNumSegments.pbtxt
@@ -57,3 +57,75 @@
}
}
}
+op {
+ name: "SparseSegmentSqrtNWithNumSegments"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ input_arg {
+ name: "num_segments"
+ type_attr: "Tnumsegments"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tnumsegments"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSum.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSum.pbtxt
index 9ecc207..c8a078d 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSum.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSum.pbtxt
@@ -202,3 +202,68 @@
}
}
}
+op {
+ name: "SparseSegmentSum"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_INT32
+ type: DT_UINT8
+ type: DT_INT16
+ type: DT_INT8
+ type: DT_INT64
+ type: DT_BFLOAT16
+ type: DT_UINT16
+ type: DT_HALF
+ type: DT_UINT32
+ type: DT_UINT64
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSumWithNumSegments.pbtxt b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSumWithNumSegments.pbtxt
index 0608745..067eef8 100644
--- a/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSumWithNumSegments.pbtxt
+++ b/tensorflow/core/ops/compat/ops_history_v2/SparseSegmentSumWithNumSegments.pbtxt
@@ -136,3 +136,85 @@
}
}
}
+op {
+ name: "SparseSegmentSumWithNumSegments"
+ input_arg {
+ name: "data"
+ type_attr: "T"
+ }
+ input_arg {
+ name: "indices"
+ type_attr: "Tidx"
+ }
+ input_arg {
+ name: "segment_ids"
+ type_attr: "Tsegmentids"
+ }
+ input_arg {
+ name: "num_segments"
+ type_attr: "Tnumsegments"
+ }
+ output_arg {
+ name: "output"
+ type_attr: "T"
+ }
+ attr {
+ name: "T"
+ type: "type"
+ allowed_values {
+ list {
+ type: DT_FLOAT
+ type: DT_DOUBLE
+ type: DT_INT32
+ type: DT_UINT8
+ type: DT_INT16
+ type: DT_INT8
+ type: DT_INT64
+ type: DT_BFLOAT16
+ type: DT_UINT16
+ type: DT_HALF
+ type: DT_UINT32
+ type: DT_UINT64
+ }
+ }
+ }
+ attr {
+ name: "Tidx"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tnumsegments"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
+}
diff --git a/tensorflow/core/ops/dataset_ops.cc b/tensorflow/core/ops/dataset_ops.cc
index 74e0d5b..6dc2280 100644
--- a/tensorflow/core/ops/dataset_ops.cc
+++ b/tensorflow/core/ops/dataset_ops.cc
@@ -438,6 +438,13 @@
.Input("deleter: variant")
.SetShapeFn(shape_inference::NoOutputs);
+REGISTER_OP("DummySeedGenerator")
+ .Output("handle: resource")
+ .SetShapeFn([](shape_inference::InferenceContext* c) {
+ c->set_output(0, c->Scalar());
+ return Status::OK();
+ });
+
REGISTER_OP("ShuffleDataset")
.Input("input_dataset: variant")
.Input("buffer_size: int64")
@@ -465,12 +472,32 @@
.Attr("output_shapes: list(shape) >= 1")
.SetShapeFn([](shape_inference::InferenceContext* c) {
shape_inference::ShapeHandle unused;
- // buffer_size, seed, and seed2 should be scalars.
+ // buffer_size and seed_generator should be scalars.
TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 0, &unused));
TF_RETURN_IF_ERROR(c->WithRank(c->input(2), 0, &unused));
return shape_inference::ScalarShape(c);
});
+REGISTER_OP("ShuffleDatasetV3")
+ .Input("input_dataset: variant")
+ .Input("buffer_size: int64")
+ .Input("seed: int64")
+ .Input("seed2: int64")
+ .Input("seed_generator: resource")
+ .Output("handle: variant")
+ .Attr("reshuffle_each_iteration: bool = true")
+ .Attr("output_types: list(type) >= 1")
+ .Attr("output_shapes: list(shape) >= 1")
+ .SetShapeFn([](shape_inference::InferenceContext* c) {
+ shape_inference::ShapeHandle unused;
+ // buffer_size, seed, seed2, and seed_generator should be scalars.
+ TF_RETURN_IF_ERROR(c->WithRank(c->input(1), 0, &unused));
+ TF_RETURN_IF_ERROR(c->WithRank(c->input(2), 0, &unused));
+ TF_RETURN_IF_ERROR(c->WithRank(c->input(3), 0, &unused));
+ TF_RETURN_IF_ERROR(c->WithRank(c->input(4), 0, &unused));
+ return shape_inference::ScalarShape(c);
+ });
+
REGISTER_OP("ShuffleAndRepeatDataset")
.Input("input_dataset: variant")
.Input("buffer_size: int64")
@@ -504,6 +531,13 @@
.Input("deleter: variant")
.SetShapeFn(shape_inference::NoOutputs);
+REGISTER_OP("DummyMemoryCache")
+ .Output("handle: resource")
+ .SetShapeFn([](shape_inference::InferenceContext* c) {
+ c->set_output(0, c->Scalar());
+ return Status::OK();
+ });
+
REGISTER_OP("CacheDataset")
.Input("input_dataset: variant")
.Input("filename: string")
diff --git a/tensorflow/core/ops/experimental_dataset_ops.cc b/tensorflow/core/ops/experimental_dataset_ops.cc
index f17c94a..a4c72b0 100644
--- a/tensorflow/core/ops/experimental_dataset_ops.cc
+++ b/tensorflow/core/ops/experimental_dataset_ops.cc
@@ -1042,6 +1042,7 @@
.Input("protocol: string")
.Input("max_outstanding_requests: int64")
.Output("handle: variant")
+ .Attr("task_refresh_interval_hint_ms: int = -1")
.Attr("output_types: list(type) >= 1")
.Attr("output_shapes: list(shape) >= 1")
.SetIsStateful()
diff --git a/tensorflow/core/ops/math_ops.cc b/tensorflow/core/ops/math_ops.cc
index 4252044..6fe3447 100644
--- a/tensorflow/core/ops/math_ops.cc
+++ b/tensorflow/core/ops/math_ops.cc
@@ -549,7 +549,7 @@
.Input("x: T")
.Input("y: T")
.Output("z: T")
- .Attr("T: {bfloat16, half, float, double, int16, int32, int64}")
+ .Attr("T: {bfloat16, half, float, double, uint8, int16, int32, int64}")
.SetShapeFn(shape_inference::BroadcastBinaryOpShapeFn);
// Note: This op is not commutative w.r.t. to all its inputs.
@@ -573,7 +573,7 @@
.Input("x: T")
.Input("y: T")
.Output("z: T")
- .Attr("T: {bfloat16, half, float, double, int16, int32, int64}")
+ .Attr("T: {bfloat16, half, float, double, uint8, int16, int32, int64}")
.SetShapeFn(shape_inference::BroadcastBinaryOpShapeFn);
REGISTER_OP("Mod")
@@ -717,8 +717,8 @@
.SetIsCommutative() \
.Attr( \
"T: {bfloat16, half, float, double, uint8, int8, int16, int32, " \
- "int64, complex64, quint8, qint8, qint32, string, bool, " \
- "complex128}") \
+ "int64, uint16, uint32, uint64, complex64, " \
+ "quint8, qint8, qint32, string, bool, complex128}") \
.Attr("incompatible_shape_error: bool = true") \
.SetShapeFn([](InferenceContext* c) { \
ShapeHandle x = c->input(0); \
@@ -1313,81 +1313,89 @@
REGISTER_OP("SparseSegmentSum")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Output("output: T")
.Attr("T: realnumbertype")
.Attr("Tidx: {int32, int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionShapeFn);
REGISTER_OP("SparseSegmentSumWithNumSegments")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Input("num_segments: Tnumsegments")
.Output("output: T")
.Attr("T: realnumbertype")
.Attr("Tidx: {int32, int64} = DT_INT32")
.Attr("Tnumsegments: {int32,int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionWithNumSegmentsShapeFn);
REGISTER_OP("SparseSegmentMean")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionShapeFn);
REGISTER_OP("SparseSegmentMeanWithNumSegments")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Input("num_segments: Tnumsegments")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
.Attr("Tnumsegments: {int32,int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionWithNumSegmentsShapeFn);
REGISTER_OP("SparseSegmentMeanGrad")
.Input("grad: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Input("output_dim0: int32")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionGradShapeFn);
REGISTER_OP("SparseSegmentSqrtN")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionShapeFn);
REGISTER_OP("SparseSegmentSqrtNWithNumSegments")
.Input("data: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Input("num_segments: Tnumsegments")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
.Attr("Tnumsegments: {int32,int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionWithNumSegmentsShapeFn);
REGISTER_OP("SparseSegmentSqrtNGrad")
.Input("grad: T")
.Input("indices: Tidx")
- .Input("segment_ids: int32")
+ .Input("segment_ids: Tsegmentids")
.Input("output_dim0: int32")
.Output("output: T")
.Attr("T: {float, double}")
.Attr("Tidx: {int32, int64} = DT_INT32")
+ .Attr("Tsegmentids: {int32, int64} = DT_INT32")
.SetShapeFn(SparseSegmentReductionGradShapeFn);
REGISTER_OP("All")
diff --git a/tensorflow/core/ops/mkl_nn_ops.cc b/tensorflow/core/ops/mkl_nn_ops.cc
index 47b3745..a625fb6 100644
--- a/tensorflow/core/ops/mkl_nn_ops.cc
+++ b/tensorflow/core/ops/mkl_nn_ops.cc
@@ -61,6 +61,30 @@
is expected to create these operators.
)doc");
+REGISTER_OP("_MklFusedDepthwiseConv2dNative")
+ .Input("input: T")
+ .Input("filter: T")
+ .Input("args: num_args * T")
+ .Input("mkl_input: uint8")
+ .Input("mkl_filter: uint8")
+ .Input("mkl_args: num_args * uint8")
+ .Output("output: T")
+ .Output("filter_output: T")
+ .Output("mkl_output: uint8")
+ .Output("mkl_filter_output: uint8")
+ .Attr("T: {bfloat16, float}")
+ .Attr("num_args: int >= 0")
+ .Attr("strides: list(int)")
+ .Attr("is_filter_const: bool = false")
+ .Attr(GetPaddingAttrString())
+ .Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
+ .Attr("fused_ops: list(string) = []")
+ // Attributes for the FusedBatchNorm ------------------------------------ //
+ .Attr("epsilon: float = 0.0001")
+ // ---------------------------------------------------------------------- //
+ .SetShapeFn(shape_inference::DepthwiseConv2DNativeShape);
+
REGISTER_OP("_MklFusedMatMul")
.Input("a: T")
.Input("b: T")
diff --git a/tensorflow/core/ops/nn_ops.cc b/tensorflow/core/ops/nn_ops.cc
index 83260bf..9200547 100644
--- a/tensorflow/core/ops/nn_ops.cc
+++ b/tensorflow/core/ops/nn_ops.cc
@@ -596,6 +596,23 @@
return Status::OK();
});
+REGISTER_OP("_FusedDepthwiseConv2dNative")
+ .Input("input: T")
+ .Input("filter: T")
+ .Input("args: num_args * T")
+ .Output("output: T")
+ .Attr("T: {half, bfloat16, float, double}")
+ .Attr("num_args: int >= 0")
+ .Attr("strides: list(int)")
+ .Attr(GetPaddingAttrString())
+ .Attr(GetConvnetDataFormatAttrString())
+ .Attr("dilations: list(int) = [1, 1, 1, 1]")
+ .Attr("fused_ops: list(string) = []")
+ // Attributes for the FusedBatchNorm ------------------------------------ //
+ .Attr("epsilon: float = 0.0001")
+ // ---------------------------------------------------------------------- //
+ .SetShapeFn(shape_inference::DepthwiseConv2DNativeShape);
+
// --------------------------------------------------------------------------
REGISTER_OP("Conv3D")
.Input("input: T")
diff --git a/tensorflow/core/ops/ops.pbtxt b/tensorflow/core/ops/ops.pbtxt
index f876b45..95cd4e2 100644
--- a/tensorflow/core/ops/ops.pbtxt
+++ b/tensorflow/core/ops/ops.pbtxt
@@ -10517,6 +10517,13 @@
type: DT_VARIANT
}
attr {
+ name: "task_refresh_interval_hint_ms"
+ type: "int"
+ default_value {
+ i: -1
+ }
+ }
+ attr {
name: "output_types"
type: "list(type)"
has_minimum: true
@@ -12499,6 +12506,22 @@
}
}
op {
+ name: "DummyMemoryCache"
+ output_arg {
+ name: "handle"
+ type: DT_RESOURCE
+ }
+ is_stateful: true
+}
+op {
+ name: "DummySeedGenerator"
+ output_arg {
+ name: "handle"
+ type: DT_RESOURCE
+ }
+ is_stateful: true
+}
+op {
name: "DynamicPartition"
input_arg {
name: "data"
@@ -13379,6 +13402,9 @@
type: DT_INT16
type: DT_INT32
type: DT_INT64
+ type: DT_UINT16
+ type: DT_UINT32
+ type: DT_UINT64
type: DT_COMPLEX64
type: DT_QUINT8
type: DT_QINT8
@@ -23713,6 +23739,7 @@
type: DT_HALF
type: DT_FLOAT
type: DT_DOUBLE
+ type: DT_UINT8
type: DT_INT16
type: DT_INT32
type: DT_INT64
@@ -23969,6 +23996,7 @@
type: DT_HALF
type: DT_FLOAT
type: DT_DOUBLE
+ type: DT_UINT8
type: DT_INT16
type: DT_INT32
type: DT_INT64
@@ -25368,6 +25396,9 @@
type: DT_INT16
type: DT_INT32
type: DT_INT64
+ type: DT_UINT16
+ type: DT_UINT32
+ type: DT_UINT64
type: DT_COMPLEX64
type: DT_QUINT8
type: DT_QINT8
@@ -42320,6 +42351,53 @@
is_stateful: true
}
op {
+ name: "ShuffleDatasetV3"
+ input_arg {
+ name: "input_dataset"
+ type: DT_VARIANT
+ }
+ input_arg {
+ name: "buffer_size"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed2"
+ type: DT_INT64
+ }
+ input_arg {
+ name: "seed_generator"
+ type: DT_RESOURCE
+ }
+ output_arg {
+ name: "handle"
+ type: DT_VARIANT
+ }
+ attr {
+ name: "reshuffle_each_iteration"
+ type: "bool"
+ default_value {
+ b: true
+ }
+ }
+ attr {
+ name: "output_types"
+ type: "list(type)"
+ has_minimum: true
+ minimum: 1
+ }
+ attr {
+ name: "output_shapes"
+ type: "list(shape)"
+ has_minimum: true
+ minimum: 1
+ }
+ is_stateful: true
+}
+op {
name: "ShutdownDistributedTPU"
is_stateful: true
}
@@ -45395,7 +45473,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
output_arg {
name: "output"
@@ -45424,6 +45502,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentMeanGrad"
@@ -45437,7 +45528,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
input_arg {
name: "output_dim0"
@@ -45470,6 +45561,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentMeanWithNumSegments"
@@ -45483,7 +45587,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
input_arg {
name: "num_segments"
@@ -45529,6 +45633,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentSqrtN"
@@ -45542,7 +45659,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
output_arg {
name: "output"
@@ -45571,6 +45688,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentSqrtNGrad"
@@ -45584,7 +45714,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
input_arg {
name: "output_dim0"
@@ -45617,6 +45747,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentSqrtNWithNumSegments"
@@ -45630,7 +45773,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
input_arg {
name: "num_segments"
@@ -45676,6 +45819,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentSum"
@@ -45689,7 +45845,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
output_arg {
name: "output"
@@ -45728,6 +45884,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSegmentSumWithNumSegments"
@@ -45741,7 +45910,7 @@
}
input_arg {
name: "segment_ids"
- type: DT_INT32
+ type_attr: "Tsegmentids"
}
input_arg {
name: "num_segments"
@@ -45797,6 +45966,19 @@
}
}
}
+ attr {
+ name: "Tsegmentids"
+ type: "type"
+ default_value {
+ type: DT_INT32
+ }
+ allowed_values {
+ list {
+ type: DT_INT32
+ type: DT_INT64
+ }
+ }
+ }
}
op {
name: "SparseSlice"
diff --git a/tensorflow/core/platform/default/distribute.bzl b/tensorflow/core/platform/default/distribute.bzl
index 35e716b..ffd4b1c 100644
--- a/tensorflow/core/platform/default/distribute.bzl
+++ b/tensorflow/core/platform/default/distribute.bzl
@@ -1,5 +1,9 @@
"""Build rules for tf.distribute testing."""
+load(
+ "//tensorflow/core/platform:build_config_root.bzl",
+ "register_extension_info",
+)
load("//tensorflow/python/tpu:tpu.bzl", _tpu_py_test = "tpu_py_test")
load("//tensorflow:tensorflow.bzl", "cuda_py_test")
@@ -69,3 +73,8 @@
disable_v2 = disable_v2,
disable_v3 = disable_v3,
)
+
+register_extension_info(
+ extension_name = "distribute_py_test",
+ label_regex_for_dep = "{extension_name}",
+)
diff --git a/tensorflow/core/platform/default/subprocess.h b/tensorflow/core/platform/default/subprocess.h
index 31b0ef3..b066274 100644
--- a/tensorflow/core/platform/default/subprocess.h
+++ b/tensorflow/core/platform/default/subprocess.h
@@ -101,7 +101,7 @@
string* stderr_output);
private:
- static const int kNFds = 3;
+ static constexpr int kNFds = 3;
static bool chan_valid(int chan) { return ((chan >= 0) && (chan < kNFds)); }
static bool retry(int e) {
return ((e == EINTR) || (e == EAGAIN) || (e == EWOULDBLOCK));
diff --git a/tensorflow/core/platform/env.cc b/tensorflow/core/platform/env.cc
index b98fd3c..74da5b9 100644
--- a/tensorflow/core/platform/env.cc
+++ b/tensorflow/core/platform/env.cc
@@ -522,7 +522,7 @@
}
private:
- static const int kBufSize = 512 << 10;
+ static constexpr int kBufSize = 512 << 10;
RandomAccessFile* file_;
int64 pos_;
diff --git a/tensorflow/core/platform/protobuf.h b/tensorflow/core/platform/protobuf.h
index d05095d..e16d89d 100644
--- a/tensorflow/core/platform/protobuf.h
+++ b/tensorflow/core/platform/protobuf.h
@@ -114,7 +114,7 @@
int64_t ByteCount() const override;
private:
- static const int kMinimumSize = 16;
+ static constexpr int kMinimumSize = 16;
tstring* target_;
};
diff --git a/tensorflow/core/profiler/convert/op_metrics_db_combiner.cc b/tensorflow/core/profiler/convert/op_metrics_db_combiner.cc
index b4c856e..3f601bb 100644
--- a/tensorflow/core/profiler/convert/op_metrics_db_combiner.cc
+++ b/tensorflow/core/profiler/convert/op_metrics_db_combiner.cc
@@ -28,6 +28,7 @@
DCHECK_EQ(src.name(), dst->name());
dst->set_category(src.category());
dst->set_provenance(src.provenance());
+ dst->set_is_eager(dst->is_eager() || src.is_eager());
dst->set_deduplicated_name(src.deduplicated_name());
if (!dst->has_layout() && src.has_layout()) {
*dst->mutable_layout() = src.layout();
diff --git a/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.cc b/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.cc
index 5b2e076..ca2a6c2 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.cc
+++ b/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.cc
@@ -84,6 +84,10 @@
const char* kAllOthersPythonExplanation =
" % of the total step time sampled is spent on 'All Others' time. "
"This could be due to Python execution overhead.";
+// Explanation for "Kernel Launch" time due to CPU contention with tf.data.
+const char* kKernelLaunchTfDataContention =
+ " It could be due to CPU contention with tf.data. In this case, you may "
+ "try to set the environment variable TF_GPU_THREAD_MODE=gpu_private.";
template <class Collection>
double GetTimeInMs(const Collection& type_ps, EventType event_type) {
@@ -357,7 +361,7 @@
return 0.0;
}
-void KernelLaunchAnalysis(double kernel_launch_percent,
+void KernelLaunchAnalysis(bool tfdata_used, double kernel_launch_percent,
string* kernel_launch_classification,
string* kernel_launch_statement) {
string percent_str = absl::StrFormat("%.1lf", kernel_launch_percent);
@@ -366,12 +370,18 @@
*kernel_launch_statement = absl::StrCat(
percent_str,
" % of the total step time sampled is spent on 'Kernel Launch'.");
+ if (tfdata_used) {
+ absl::StrAppend(kernel_launch_statement, kKernelLaunchTfDataContention);
+ }
} else if (kernel_launch_percent >=
kModeratelyKernelLaunchBoundThresholdInPercent) {
*kernel_launch_classification = "moderate";
*kernel_launch_statement = absl::StrCat(
percent_str,
" % of the total step time sampled is spent on 'Kernel Launch'.");
+ if (tfdata_used) {
+ absl::StrAppend(kernel_launch_statement, kKernelLaunchTfDataContention);
+ }
} else {
*kernel_launch_classification = "no";
*kernel_launch_statement = "";
@@ -566,8 +576,8 @@
GenerateHostResult(op_stats.host_op_metrics_db(), &result);
InputPipelineAnalysisRecommendation recommendation = GenerateRecommendation();
- BottleneckAnalysis bottleneck_analysis =
- ComputeBottleneckAnalysis(result.step_details());
+ BottleneckAnalysis bottleneck_analysis = ComputeBottleneckAnalysis(
+ result.input_time_breakdown(), result.step_details());
recommendation.mutable_bottleneck_analysis()->PackFrom(bottleneck_analysis);
*recommendation.mutable_summary_next_step() =
GetSummaryNextStep(bottleneck_analysis.input_classification(),
@@ -602,7 +612,7 @@
*input_classification = "both";
string all_other_percent_str = absl::StrFormat("%.1lf", all_other_percent);
*input_statement = absl::StrCat(
- "Your program in POTENTIALLY input-bound because ",
+ "Your program is POTENTIALLY input-bound because ",
all_other_percent_str,
"% of the total step time sampled is spent on 'All Others' time (which "
"could be due to I/O or Python execution or both).");
@@ -646,6 +656,7 @@
}
BottleneckAnalysis ComputeBottleneckAnalysis(
+ const InputTimeBreakdown& input_time_breakdown,
const ::tensorflow::protobuf::RepeatedPtrField<::google::protobuf::Any>&
any_step_details) {
double total_step_time_ms = 0;
@@ -700,8 +711,8 @@
string kernel_launch_classification;
string kernel_launch_statement;
- KernelLaunchAnalysis(kernel_launch_percent, &kernel_launch_classification,
- &kernel_launch_statement);
+ KernelLaunchAnalysis(TfDataInUse(input_time_breakdown), kernel_launch_percent,
+ &kernel_launch_classification, &kernel_launch_statement);
string all_other_classification;
string all_other_statement;
diff --git a/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.h b/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.h
index 511298c..738daea 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.h
+++ b/tensorflow/core/profiler/convert/op_stats_to_input_pipeline_analysis.h
@@ -40,6 +40,7 @@
// Returns the performance bottleneck of the program executed.
BottleneckAnalysis ComputeBottleneckAnalysis(
+ const InputTimeBreakdown& input_time_breakdown,
const ::tensorflow::protobuf::RepeatedPtrField<::google::protobuf::Any>&
any_step_details);
diff --git a/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc b/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
index 00c4af9..e19690a 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
+++ b/tensorflow/core/profiler/convert/op_stats_to_overview_page.cc
@@ -253,8 +253,9 @@
*overview_page.mutable_analysis() = ComputeAnalysisResult(op_stats);
*overview_page.mutable_input_analysis() =
ConvertOpStatsToInputPipelineAnalysis(op_stats, hardware_type);
- BottleneckAnalysis bottleneck =
- ComputeBottleneckAnalysis(overview_page.input_analysis().step_details());
+ BottleneckAnalysis bottleneck = ComputeBottleneckAnalysis(
+ overview_page.input_analysis().input_time_breakdown(),
+ overview_page.input_analysis().step_details());
*overview_page.mutable_recommendation() = ComputeGenericRecommendation(
bottleneck, op_stats.device_op_metrics_db().precision_stats());
SetCommonRecommendation(bottleneck.input_classification(),
diff --git a/tensorflow/core/profiler/convert/op_stats_to_tf_stats.cc b/tensorflow/core/profiler/convert/op_stats_to_tf_stats.cc
index 2356116..da409f8 100644
--- a/tensorflow/core/profiler/convert/op_stats_to_tf_stats.cc
+++ b/tensorflow/core/profiler/convert/op_stats_to_tf_stats.cc
@@ -33,6 +33,7 @@
double ridge_point_operational_intensity) {
TfStatsRecord record;
record.set_host_or_device(on_device ? "Device" : "Host");
+ record.set_is_eager(metrics.is_eager());
record.set_op_type(metrics.category());
record.set_op_name(metrics.name());
SetExecutionTimes(metrics, &record);
diff --git a/tensorflow/core/profiler/internal/gpu/BUILD b/tensorflow/core/profiler/internal/gpu/BUILD
index 2456809..c6fe4d7 100644
--- a/tensorflow/core/profiler/internal/gpu/BUILD
+++ b/tensorflow/core/profiler/internal/gpu/BUILD
@@ -91,8 +91,9 @@
deps = [
"//tensorflow/core:lib",
"//tensorflow/core:platform_base",
+ ] + if_cuda_is_configured_compat([
"//tensorflow/stream_executor/cuda:cupti_stub",
- ],
+ ]),
)
tf_cuda_library(
@@ -103,8 +104,9 @@
visibility = ["//visibility:public"],
deps = [
":cupti_interface",
+ ] + if_cuda_is_configured_compat([
"//tensorflow/stream_executor/cuda:cupti_stub",
- ],
+ ]),
)
tf_cuda_library(
diff --git a/tensorflow/core/profiler/protobuf/BUILD b/tensorflow/core/profiler/protobuf/BUILD
index ce5bc9b..03f77ee 100644
--- a/tensorflow/core/profiler/protobuf/BUILD
+++ b/tensorflow/core/profiler/protobuf/BUILD
@@ -112,3 +112,10 @@
cc_api_version = 2,
visibility = [":friends"],
)
+
+tf_proto_library(
+ name = "tfstreamz_proto",
+ srcs = ["tfstreamz.proto"],
+ cc_api_version = 2,
+ visibility = [":friends"],
+)
diff --git a/tensorflow/core/profiler/protobuf/tf_stats.proto b/tensorflow/core/profiler/protobuf/tf_stats.proto
index 0b4cbb7..2dae623 100644
--- a/tensorflow/core/profiler/protobuf/tf_stats.proto
+++ b/tensorflow/core/profiler/protobuf/tf_stats.proto
@@ -69,4 +69,6 @@
// Whether this operation is "Compute" or "Memory" bound,
// according to the Roofline Model.
string bound_by = 17;
+ // Whether this TF-op is eagerly executed.
+ bool is_eager = 18;
}
diff --git a/tensorflow/core/profiler/protobuf/tfstreamz.proto b/tensorflow/core/profiler/protobuf/tfstreamz.proto
new file mode 100644
index 0000000..4fe5c16
--- /dev/null
+++ b/tensorflow/core/profiler/protobuf/tfstreamz.proto
@@ -0,0 +1,33 @@
+// This proto describes the format of the output profile file from
+// the TF-stats tool.
+syntax = "proto3";
+
+package tensorflow.profiler.tfstreamz;
+
+// A proxy proto to serialize tensorflow::monitoring::Percentiles
+
+enum UnitOfMeasure {
+ NUMBER = 0;
+ TIME = 1;
+ BYTES = 2;
+}
+
+message PercentilePoint {
+ // In the [0, 100] range.
+ double percentile = 1;
+ double value = 2;
+}
+
+message Percentiles {
+ UnitOfMeasure unit_of_measure = 1;
+ uint64 start_nstime = 2;
+ uint64 end_nstime = 3;
+ double min_value = 4;
+ double max_value = 5;
+ double mean = 6;
+ double stddev = 7;
+ uint64 num_samples = 8;
+ uint64 total_samples = 9;
+ double accumulator = 10;
+ repeated PercentilePoint points = 11;
+}
diff --git a/tensorflow/core/profiler/utils/BUILD b/tensorflow/core/profiler/utils/BUILD
index 6bfb7ad..189b39d 100644
--- a/tensorflow/core/profiler/utils/BUILD
+++ b/tensorflow/core/profiler/utils/BUILD
@@ -325,3 +325,19 @@
"@com_google_absl//absl/strings",
],
)
+
+cc_library(
+ name = "tfstreamz_utils",
+ srcs = ["tfstreamz_utils.cc"],
+ hdrs = ["tfstreamz_utils.h"],
+ deps = [
+ ":xplane_builder",
+ "//tensorflow/core:lib",
+ "//tensorflow/core:lib_internal",
+ "//tensorflow/core/profiler/protobuf:tfstreamz_proto_cc",
+ "//tensorflow/core/profiler/protobuf:xplane_proto_cc",
+ "@com_google_absl//absl/memory",
+ "@com_google_absl//absl/strings",
+ "@com_google_absl//absl/strings:str_format",
+ ],
+)
diff --git a/tensorflow/core/profiler/utils/group_events.cc b/tensorflow/core/profiler/utils/group_events.cc
index e68cb52..f19aab4 100644
--- a/tensorflow/core/profiler/utils/group_events.cc
+++ b/tensorflow/core/profiler/utils/group_events.cc
@@ -32,17 +32,18 @@
static const int64 kFunctionalOpEventTypes[] = {
HostEventType::kCallOp,
- HostEventType::kParallelForOp,
- HostEventType::kForeverOp,
HostEventType::kNumericalGradientOpEvalRight,
HostEventType::kNumericalGradientOpEvalLeft,
HostEventType::kSymbolicGradientOp,
HostEventType::kRemoteCallOp,
HostEventType::kIfOp,
HostEventType::kCaseOp,
- HostEventType::kWhileOpEvalCond,
- HostEventType::kWhileOpStartBody,
- HostEventType::kForOp,
+ // TODO(b/154510598): Fix handling of the loop ops.
+ // HostEventType::kWhileOpEvalCond,
+ // HostEventType::kWhileOpStartBody,
+ // HostEventType::kForOp,
+ // HostEventType::kParallelForOp,
+ // HostEventType::kForeverOp,
HostEventType::kPartitionedCallOp,
};
@@ -193,9 +194,11 @@
}
bool EventNode::IsEager() {
- // It is eagerly executed if its trace context does not include the TF
- // executor.
- return FindParent(HostEventType::kExecutorStateProcess) == nullptr;
+ // It is eagerly executed if its trace context includes the EagerKernelExecute
+ // event (which may execute an op eagerly or through the TF executor) but not
+ // the TF executor event.
+ return FindParent(HostEventType::kExecutorStateProcess) == nullptr &&
+ FindParent(HostEventType::kEagerKernelExecute) != nullptr;
}
bool EventNode::IsNestedIn(EventNode* parent) {
diff --git a/tensorflow/core/profiler/utils/group_events_test.cc b/tensorflow/core/profiler/utils/group_events_test.cc
index 76d9405..f3f58bf 100644
--- a/tensorflow/core/profiler/utils/group_events_test.cc
+++ b/tensorflow/core/profiler/utils/group_events_test.cc
@@ -152,9 +152,13 @@
auto main_thread = host_plane_builder.GetOrCreateLine(0);
// Eagerly scheduled GPU kernel.
+ CreateXEvent(&host_plane_builder, &main_thread,
+ HostEventType::kEagerKernelExecute, 10, 100, {});
CreateXEvent(&host_plane_builder, &main_thread, "matmul", 10, 100,
{{StatType::kCorrelationId, 100}});
// Eagerly executed CPU TF op.
+ CreateXEvent(&host_plane_builder, &main_thread,
+ HostEventType::kEagerKernelExecute, 120, 80, {});
CreateXEvent(&host_plane_builder, &main_thread, "add:Add", 120, 80, {});
XPlane* device_plane = space.add_planes();
@@ -168,7 +172,7 @@
GroupTfEvents(&space, /*event_group_name_map=*/nullptr);
XPlaneVisitor host_plane_visitor = CreateTfXPlaneVisitor(host_plane);
- const XEvent& eager_cpu_tf_op = host_plane->lines(0).events(1);
+ const XEvent& eager_cpu_tf_op = host_plane->lines(0).events(3);
EXPECT_EQ(eager_cpu_tf_op.stats_size(), 1);
EXPECT_EQ(host_plane_visitor.GetStatType(eager_cpu_tf_op.stats(0)),
StatType::kIsEager);
@@ -191,6 +195,8 @@
auto main_thread = host_plane_builder.GetOrCreateLine(0);
CreateXEvent(&host_plane_builder, &main_thread, HostEventType::kTraceContext,
0, 100, {{StatType::kStepNum, 123}});
+ CreateXEvent(&host_plane_builder, &main_thread,
+ HostEventType::kEagerKernelExecute, 10, 90, {});
CreateXEvent(&host_plane_builder, &main_thread, HostEventType::kFunctionRun,
10, 90, {{StatType::kStepId, 0}});
diff --git a/tensorflow/core/profiler/utils/op_utils.cc b/tensorflow/core/profiler/utils/op_utils.cc
index cbc16f9..74ce13d 100644
--- a/tensorflow/core/profiler/utils/op_utils.cc
+++ b/tensorflow/core/profiler/utils/op_utils.cc
@@ -41,7 +41,7 @@
OpMetrics* op_metrics = LookupOrInsertNewOpMetrics(/*hlo_module_id=*/0, name);
if (op_metrics->category().empty())
op_metrics->set_category(category.data(), category.size());
- op_metrics->set_is_eager(is_eager);
+ op_metrics->set_is_eager(op_metrics->is_eager() || is_eager);
op_metrics->set_occurrences(op_metrics->occurrences() + 1);
op_metrics->set_time_ps(op_metrics->time_ps() + time_ps);
op_metrics->set_self_time_ps(op_metrics->self_time_ps() + self_time_ps);
@@ -72,7 +72,7 @@
: string(category));
if (op_metrics->provenance().empty())
op_metrics->set_provenance(string(provenance));
- op_metrics->set_is_eager(is_eager);
+ op_metrics->set_is_eager(op_metrics->is_eager() || is_eager);
op_metrics->set_occurrences(op_metrics->occurrences() + occurrences);
op_metrics->set_time_ps(op_metrics->time_ps() + time_ps);
op_metrics->set_self_time_ps(op_metrics->self_time_ps() + self_time_ps);
diff --git a/tensorflow/core/profiler/utils/tfstreamz_utils.cc b/tensorflow/core/profiler/utils/tfstreamz_utils.cc
new file mode 100644
index 0000000..b531c69
--- /dev/null
+++ b/tensorflow/core/profiler/utils/tfstreamz_utils.cc
@@ -0,0 +1,123 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include "tensorflow/core/profiler/utils/tfstreamz_utils.h"
+
+#include <memory>
+
+#include "absl/memory/memory.h"
+#include "absl/strings/str_cat.h"
+#include "absl/strings/str_format.h"
+#include "absl/strings/str_join.h"
+#include "absl/strings/substitute.h"
+#include "tensorflow/core/lib/core/errors.h"
+#include "tensorflow/core/lib/gtl/map_util.h"
+#include "tensorflow/core/lib/monitoring/collected_metrics.h"
+#include "tensorflow/core/lib/monitoring/collection_registry.h"
+#include "tensorflow/core/platform/errors.h"
+#include "tensorflow/core/profiler/protobuf/tfstreamz.pb.h"
+
+namespace tensorflow {
+namespace profiler {
+
+namespace {
+string ConstructXStatName(const string& name, const monitoring::Point& point) {
+ if (point.labels.empty()) {
+ return name;
+ }
+ return absl::Substitute(
+ "$0{$1}", name,
+ absl::StrJoin(point.labels, ", ",
+ [](string* out, const monitoring::Point::Label& label) {
+ absl::StrAppend(out, label.name, "=", label.value);
+ }));
+}
+
+string SerializePercentile(const monitoring::Percentiles& percentiles) {
+ tfstreamz::Percentiles output;
+ output.set_unit_of_measure(
+ static_cast<tfstreamz::UnitOfMeasure>(percentiles.unit_of_measure));
+ output.set_start_nstime(percentiles.start_nstime);
+ output.set_end_nstime(percentiles.end_nstime);
+ output.set_min_value(percentiles.min_value);
+ output.set_max_value(percentiles.max_value);
+ output.set_mean(percentiles.mean);
+ output.set_stddev(percentiles.stddev);
+ output.set_num_samples(percentiles.num_samples);
+ output.set_total_samples(percentiles.total_samples);
+ output.set_accumulator(percentiles.accumulator);
+ for (const auto& pp : percentiles.points) {
+ auto* percentile_point = output.add_points();
+ percentile_point->set_percentile(pp.percentile);
+ percentile_point->set_value(pp.value);
+ }
+ return output.SerializeAsString();
+}
+
+} // namespace
+
+Status SerializeToXPlane(const std::vector<TfStreamzSnapshot>& snapshots,
+ XPlane* plane, uint64 line_start_time_ns) {
+ XPlaneBuilder xplane(plane);
+ XLineBuilder line = xplane.GetOrCreateLine(0); // This plane has single line.
+ line.SetTimestampNs(line_start_time_ns);
+
+ // For each snapshot, create a virtual event.
+ for (const auto& snapshot : snapshots) {
+ XEventMetadata* event_metadata =
+ xplane.GetOrCreateEventMetadata("TFStreamz Snapshot");
+ XEventBuilder xevent = line.AddEvent(*event_metadata);
+ xevent.SetTimestampNs(snapshot.start_time_ns);
+ xevent.SetEndTimestampNs(snapshot.end_time_ns);
+ auto& metric_descriptor_map = snapshot.metrics->metric_descriptor_map;
+ for (const auto& point_set : snapshot.metrics->point_set_map) {
+ const string& metric_name = point_set.first;
+ // Each metrics have multiple points corresponding to different labels.
+ for (const auto& point : point_set.second->points) {
+ // Generates one KPI metric for each point.
+ string stat_name = ConstructXStatName(metric_name, *point);
+ auto* metadata = xplane.GetOrCreateStatMetadata(stat_name);
+ auto it = metric_descriptor_map.find(metric_name);
+ if (it != metric_descriptor_map.end()) {
+ metadata->set_description(it->second->description);
+ }
+ switch (point->value_type) {
+ case monitoring::ValueType::kInt64:
+ xevent.AddStatValue(*metadata, point->int64_value);
+ break;
+ case monitoring::ValueType::kBool:
+ xevent.AddStatValue(*metadata, point->bool_value);
+ break;
+ case monitoring::ValueType::kString:
+ xevent.AddStatValue(*metadata, point->string_value);
+ break;
+ case monitoring::ValueType::kHistogram:
+ xevent.AddStatValue(*metadata,
+ point->histogram_value.SerializeAsString(),
+ /*is_bytes=*/true);
+ break;
+ case monitoring::ValueType::kPercentiles:
+ xevent.AddStatValue(*metadata,
+ SerializePercentile(point->percentiles_value),
+ /*is_bytes=*/true);
+ break;
+ }
+ }
+ }
+ }
+ return Status::OK();
+}
+
+} // namespace profiler
+} // namespace tensorflow
diff --git a/tensorflow/core/profiler/utils/tfstreamz_utils.h b/tensorflow/core/profiler/utils/tfstreamz_utils.h
new file mode 100644
index 0000000..ae8e407
--- /dev/null
+++ b/tensorflow/core/profiler/utils/tfstreamz_utils.h
@@ -0,0 +1,39 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_CORE_PROFILER_UTILS_TFSTREAMZ_UTILS_H_
+#define TENSORFLOW_CORE_PROFILER_UTILS_TFSTREAMZ_UTILS_H_
+
+#include "tensorflow/core/lib/monitoring/collected_metrics.h"
+#include "tensorflow/core/lib/monitoring/collection_registry.h"
+#include "tensorflow/core/platform/status.h"
+#include "tensorflow/core/profiler/protobuf/xplane.pb.h"
+#include "tensorflow/core/profiler/utils/xplane_builder.h"
+
+namespace tensorflow {
+namespace profiler {
+
+struct TfStreamzSnapshot {
+ std::unique_ptr<monitoring::CollectedMetrics> metrics;
+ uint64 start_time_ns; // time before collection.
+ uint64 end_time_ns; // time after collection.
+};
+
+Status SerializeToXPlane(const std::vector<TfStreamzSnapshot>& snapshots,
+ XPlane* plane, uint64 line_start_time_ns);
+
+} // namespace profiler
+} // namespace tensorflow
+
+#endif // TENSORFLOW_CORE_PROFILER_UTILS_TFSTREAMZ_UTILS_H_
diff --git a/tensorflow/core/profiler/utils/xplane_schema.cc b/tensorflow/core/profiler/utils/xplane_schema.cc
index a040f93..402b755 100644
--- a/tensorflow/core/profiler/utils/xplane_schema.cc
+++ b/tensorflow/core/profiler/utils/xplane_schema.cc
@@ -26,11 +26,16 @@
const absl::string_view kGpuPlanePrefix = "/device:GPU:";
const absl::string_view kCuptiDriverApiPlaneName = "/host:CUPTI";
const absl::string_view kMetadataPlane = "/host:metadata";
+const absl::string_view kTFStreamzPlane = "/host:tfstreamz";
const int32 kHostPlaneId = 49;
const int32 kGpuPlaneBaseId = 0;
const int32 kCuptiDriverApiPlaneId = 50;
-const int32 kMetadataPlaneId = 51;
+const int32 kMetadataPlaneId = 99;
+const int32 kTFStreamzPlaneId = 98;
+
+const int32 kThreadGroupMinPlaneId = kCuptiDriverApiPlaneId + 1;
+const int32 kThreadGroupMaxPlaneId = kTFStreamzPlaneId - 1;
namespace {
diff --git a/tensorflow/core/profiler/utils/xplane_schema.h b/tensorflow/core/profiler/utils/xplane_schema.h
index 3a74171..6ffb4f9 100644
--- a/tensorflow/core/profiler/utils/xplane_schema.h
+++ b/tensorflow/core/profiler/utils/xplane_schema.h
@@ -33,6 +33,8 @@
ABSL_CONST_INIT extern const absl::string_view kCuptiDriverApiPlaneName;
// Name of XPlane that contains profile metadata such as XLA debug info.
ABSL_CONST_INIT extern const absl::string_view kMetadataPlane;
+// Name of XPlane that contains kpi related metrics.
+ABSL_CONST_INIT extern const absl::string_view kTFStreamzPlane;
// Id of XPlane that contains TraceMe events.
ABSL_CONST_INIT extern const int32 kHostPlaneId;
@@ -43,6 +45,11 @@
ABSL_CONST_INIT extern const int32 kCuptiDriverApiPlaneId;
// Id of XPlane that contains profile metadata such as XLA debug info.
ABSL_CONST_INIT extern const int32 kMetadataPlaneId;
+// Id of XPlane that contains kpi related metrics.
+ABSL_CONST_INIT extern const int32 kTFStreamzPlaneId;
+
+ABSL_CONST_INIT extern const int32 kThreadGroupMinPlaneId;
+ABSL_CONST_INIT extern const int32 kThreadGroupMaxPlaneId;
// Interesting event types (i.e., TraceMe names).
enum HostEventType {
diff --git a/tensorflow/core/public/version.h b/tensorflow/core/public/version.h
index 9c2792b..094bb19 100644
--- a/tensorflow/core/public/version.h
+++ b/tensorflow/core/public/version.h
@@ -108,7 +108,7 @@
#define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
#define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
-#define TF_GRAPH_DEF_VERSION 377 // Updated: 2020/4/20
+#define TF_GRAPH_DEF_VERSION 380 // Updated: 2020/4/23
// Checkpoint compatibility versions (the versions field in SavedSliceMeta).
//
diff --git a/tensorflow/core/util/BUILD b/tensorflow/core/util/BUILD
index b8c2b3b..de2dce9c 100644
--- a/tensorflow/core/util/BUILD
+++ b/tensorflow/core/util/BUILD
@@ -144,6 +144,7 @@
"matmul_autotune.h",
"matmul_bcast.h",
"mirror_pad_mode.h",
+ "mkl_threadpool.h",
"mkl_types.h",
"mkl_util.h",
"overflow.h",
@@ -273,6 +274,7 @@
filegroup(
name = "mkl_util_hdrs",
srcs = [
+ "mkl_threadpool.h",
"mkl_util.h",
],
visibility = ["//tensorflow/core:__pkg__"],
diff --git a/tensorflow/core/util/mkl_threadpool.h b/tensorflow/core/util/mkl_threadpool.h
new file mode 100644
index 0000000..8c9db0a
--- /dev/null
+++ b/tensorflow/core/util/mkl_threadpool.h
@@ -0,0 +1,138 @@
+
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_UTIL_MKL_THREADPOOL_H_
+#define TENSORFLOW_CORE_UTIL_MKL_THREADPOOL_H_
+#ifdef INTEL_MKL
+
+#include <list>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "mkldnn.hpp"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/platform/threadpool.h"
+#define EIGEN_USE_THREADS
+#ifdef ENABLE_MKLDNN_THREADPOOL
+using dnnl::stream_attr;
+using dnnl::threadpool_iface;
+
+namespace tensorflow {
+
+// Divide 'n' units of work equally among 'teams' threads. If 'n' is not
+// divisible by 'teams' and has a remainder 'r', the first 'r' teams have one
+// unit of work more than the rest. Returns the range of work that belongs to
+// the team 'tid'.
+// Parameters
+// n Total number of jobs.
+// team Number of workers.
+// tid Current thread_id.
+// n_start start of range operated by the thread.
+// n_end end of the range operated by the thread.
+
+template <typename T, typename U>
+inline void balance211(T n, U team, U tid, T* n_start, T* n_end) {
+ if (team <= 1 || n == 0) {
+ *n_start = 0;
+ *n_end = n;
+ return;
+ }
+ T min_per_team = n / team;
+ T remainder = n - min_per_team * team; // i.e., n % teams.
+ *n_start = tid * min_per_team + std::min(tid, remainder);
+ *n_end = *n_start + min_per_team + (tid < remainder);
+}
+
+struct MklDnnThreadPool : public dnnl::threadpool_iface {
+ MklDnnThreadPool() = default;
+
+ MklDnnThreadPool(OpKernelContext* ctx)
+ : eigen_interface_(ctx->device()
+ ->tensorflow_cpu_worker_threads()
+ ->workers->AsEigenThreadPool()) {}
+ virtual int get_num_threads() override {
+ return eigen_interface_->NumThreads();
+ }
+ virtual bool get_in_parallel() override {
+ return (eigen_interface_->CurrentThreadId() != -1) ? true : false;
+ }
+ virtual uint64_t get_flags() override { return ASYNCHRONOUS; }
+ virtual void parallel_for(int n,
+ const std::function<void(int, int)>& fn) override {
+ // Should never happen (handled by DNNL)
+ if (n == 0) return;
+
+ // Should never happen (handled by DNNL)
+ if (n == 1) {
+ fn(0, 1);
+ return;
+ }
+
+ int nthr = get_num_threads();
+ int njobs = std::min(n, nthr);
+ for (int i = 0; i < njobs; i++) {
+ eigen_interface_->ScheduleWithHint(
+ [i, n, njobs, fn]() {
+ int start, end;
+ balance211(n, njobs, i, &start, &end);
+ for (int j = start; j < end; j++) fn(j, n);
+ },
+ i, i + 1);
+ }
+ }
+ ~MklDnnThreadPool() {}
+
+ private:
+ Eigen::ThreadPoolInterface* eigen_interface_ = nullptr;
+};
+
+class MklDnnThreadPoolWrapper {
+ public:
+ static MklDnnThreadPoolWrapper& GetInstance() {
+ static MklDnnThreadPoolWrapper instance_;
+ return instance_;
+ }
+ MklDnnThreadPool* CreateThreadPoolPtr(OpKernelContext* ctx) {
+ if (threadpool_map_.empty() ||
+ threadpool_map_.find(ctx->device()) == threadpool_map_.end()) {
+ auto tp_iface = new MklDnnThreadPool(ctx);
+ threadpool_map_.emplace(std::make_pair(ctx->device(), tp_iface));
+ return tp_iface;
+ } else {
+ auto entry = threadpool_map_.find(ctx->device());
+ return entry->second;
+ }
+ }
+
+ private:
+ std::unordered_map<DeviceBase*, MklDnnThreadPool*> threadpool_map_;
+ MklDnnThreadPoolWrapper() {}
+ MklDnnThreadPoolWrapper(const MklDnnThreadPoolWrapper&) = delete;
+ MklDnnThreadPoolWrapper& operator=(const MklDnnThreadPoolWrapper&) = delete;
+ ~MklDnnThreadPoolWrapper() {
+ for (auto& tp : threadpool_map_) {
+ delete tp.second;
+ }
+ }
+};
+
+} // namespace tensorflow
+#endif // ENABLE_MKLDNN_THREADPOOL
+#endif // INTEL_MKL
+#endif // TENSORFLOW_CORE_UTIL_MKL_THREADPOOL_H_
diff --git a/tensorflow/core/util/mkl_util.h b/tensorflow/core/util/mkl_util.h
index e0a399f..7f6272b 100644
--- a/tensorflow/core/util/mkl_util.h
+++ b/tensorflow/core/util/mkl_util.h
@@ -36,6 +36,7 @@
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/util/env_var.h"
+#include "tensorflow/core/util/mkl_threadpool.h"
#include "tensorflow/core/util/mkl_types.h"
#include "tensorflow/core/util/padding.h"
#include "tensorflow/core/util/tensor_format.h"
@@ -48,7 +49,6 @@
using mkldnn::primitive;
using mkldnn::reorder;
using mkldnn::stream;
-
using CPUDevice = Eigen::ThreadPoolDevice;
using MemoryArgsMap = std::unordered_map<int, memory>;
using ReorderPd = mkldnn::reorder::primitive_desc;
@@ -232,6 +232,27 @@
return true;
}
+inline mkldnn::stream* CreateStream(OpKernelContext* ctx,
+ const engine& engine) {
+#ifdef ENABLE_MKLDNN_THREADPOOL
+ stream_attr tp_stream_attr(ENGINE_CPU);
+ if (ctx != nullptr) {
+ auto eigen_tp =
+ MklDnnThreadPoolWrapper::GetInstance().CreateThreadPoolPtr(ctx);
+ tp_stream_attr.set_threadpool(eigen_tp);
+ stream* tp_stream =
+ new stream(engine, stream::flags::default_flags, tp_stream_attr);
+ return tp_stream;
+ } else {
+ stream* tp_stream = new CPU_STREAM(engine);
+ return tp_stream;
+ }
+#else
+ stream* tp_stream = new CPU_STREAM(engine);
+ return tp_stream;
+#endif // ENABLE_MKLDNN_THREADPOOL
+}
+
class MklDnnShape {
private:
typedef struct {
@@ -679,20 +700,21 @@
// TODO merge with the execute_primitives.
inline void ExecutePrimitive(const std::vector<primitive>& net,
const std::vector<MemoryArgsMap>* net_args,
- const engine& cpu_engine) {
+ const engine& cpu_engine,
+ OpKernelContext* context = nullptr) {
#ifdef ENABLE_MKLDNN_V1
DCHECK(net_args);
DCHECK_EQ(net.size(), net_args->size());
- stream cpu_stream(cpu_engine);
+ stream* cpu_stream = CreateStream(context, cpu_engine);
for (size_t i = 0; i < net.size(); ++i) {
- net.at(i).execute(cpu_stream, net_args->at(i));
+ net.at(i).execute(*cpu_stream, net_args->at(i));
}
- cpu_stream.wait();
+ cpu_stream->wait();
+ delete cpu_stream;
#else
stream(stream::kind::eager_nostore).submit(net).wait();
#endif // ENABLE_MKLDNN_V1
}
-
template <typename T>
inline Status ConvertMklToTF(OpKernelContext* context,
const Tensor& input_mkl_tensor,
@@ -731,7 +753,7 @@
return Status(error::Code::INTERNAL,
"ConvertMklToTF(): Failed to create reorder for input");
}
- ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine);
+ ExecutePrimitive(net, NET_ARGS_PTR, cpu_engine, context);
} else {
// If not, just forward input tensor to output tensor.
bool status =
@@ -1301,8 +1323,8 @@
inline void CreateAndExecuteReorder(const ReorderPd& reorder_desc,
const memory& src_mem,
- const memory& dst_mem,
- const engine& engine) {
+ const memory& dst_mem, const engine& engine,
+ OpKernelContext* ctx = nullptr) {
std::vector<primitive> net;
#ifdef ENABLE_MKLDNN_V1
net.push_back(mkldnn::reorder(reorder_desc));
@@ -1311,7 +1333,7 @@
#else
net.push_back(mkldnn::reorder(reorder_desc, src_mem, dst_mem));
#endif // ENABLE_MKLDNN_V1
- ExecutePrimitive(net, NET_ARGS_PTR, engine);
+ ExecutePrimitive(net, NET_ARGS_PTR, engine, ctx);
}
class MklReorderPrimitive;
@@ -1629,22 +1651,26 @@
#ifdef ENABLE_MKLDNN_V1
inline bool CheckReorderToOpMem(const memory::desc& op_md,
- const engine& engine) {
+ const engine& engine,
+ OpKernelContext* context = nullptr) {
DCHECK(user_memory_);
if (IsReorderNeeded(op_md)) {
// TODO(nhasabni): can we remove dynamic memory allocation?
// primitive reuse don't allow two same reorder prim in
// one stream, so submit it immediately
reorder_memory_ = new memory(op_md, engine);
- std::vector<primitive> net;
auto* prim = FindOrCreateReorder<T>(user_memory_, reorder_memory_);
+ std::shared_ptr<stream> cpu_stream;
+ cpu_stream.reset(CreateStream(context, prim->GetEngine()));
+ std::vector<primitive> net;
net.push_back(*(prim->GetPrimitive()));
std::vector<MemoryArgsMap> net_args;
net_args.push_back({{MKLDNN_ARG_FROM, *user_memory_},
{MKLDNN_ARG_TO, *reorder_memory_}});
- execute_primitives(net, prim->GetStream(), net_args);
+ execute_primitives(net, cpu_stream, net_args);
#else
- inline bool CheckReorderToOpMem(const memory::primitive_desc& op_pd) {
+ inline bool CheckReorderToOpMem(const memory::primitive_desc& op_pd,
+ OpKernelContext* ctx = nullptr) {
CHECK_NOTNULL(user_memory_);
if (IsReorderNeeded(op_pd)) {
reorder_memory_ = new memory(op_pd);
@@ -1708,7 +1734,8 @@
/// TODO(bhavanis): Need to use reorder cache here for better performance.
inline bool CheckReorderToOpMem(const memory::desc& op_md,
void* reorder_data_handle,
- const engine& engine) {
+ const engine& engine,
+ OpKernelContext* context = nullptr) {
DCHECK(reorder_data_handle);
DCHECK(user_memory_);
if (IsReorderNeeded(op_md)) {
@@ -1716,16 +1743,19 @@
// primitive reuse don't allow two same reorder prim in
// one stream, so submit it immediately
reorder_memory_ = new memory(op_md, engine, reorder_data_handle);
- std::vector<primitive> net;
auto* prim = FindOrCreateReorder<T>(user_memory_, reorder_memory_);
+ std::shared_ptr<stream> cpu_stream;
+ cpu_stream.reset(CreateStream(context, prim->GetEngine()));
+ std::vector<primitive> net;
net.push_back(*(prim->GetPrimitive()));
std::vector<MemoryArgsMap> net_args;
net_args.push_back({{MKLDNN_ARG_FROM, *user_memory_},
{MKLDNN_ARG_TO, *reorder_memory_}});
- execute_primitives(net, prim->GetStream(), net_args);
+ execute_primitives(net, cpu_stream, net_args);
#else
inline bool CheckReorderToOpMem(const memory::primitive_desc& op_pd,
- void* reorder_data_handle) {
+ void* reorder_data_handle,
+ OpKernelContext* context = nullptr) {
CHECK_NOTNULL(reorder_data_handle);
CHECK_NOTNULL(user_memory_);
if (IsReorderNeeded(op_pd)) {
@@ -1778,13 +1808,14 @@
/// remove
/// slow path in the future
inline bool CheckReorderToOpMem(const MEMORY_PRIMITIVE_DESC& op_pd,
- Tensor* reorder_tensor) {
+ Tensor* reorder_tensor,
+ OpKernelContext* ctx = nullptr) {
DCHECK(reorder_tensor);
#ifdef ENABLE_MKLDNN_V1
return CheckReorderToOpMem(op_pd, GetTensorBuffer(reorder_tensor),
- *cpu_engine_);
+ *cpu_engine_, ctx);
#else
- return CheckReorderToOpMem(op_pd, GetTensorBuffer(reorder_tensor));
+ return CheckReorderToOpMem(op_pd, GetTensorBuffer(reorder_tensor), ctx);
#endif // ENABLE_MKLDNN_V1
}
@@ -1843,7 +1874,7 @@
/// TODO: this is a faster path with reorder primitive cache compared with
/// InsertReorderToUserMem(net, net_args), will remove
/// slow path in the future
- inline void InsertReorderToUserMem() {
+ inline void InsertReorderToUserMem(OpKernelContext* ctx = nullptr) {
DCHECK(user_memory_);
DCHECK(reorder_memory_);
DCHECK(cpu_engine_);
@@ -1857,8 +1888,8 @@
net_args.push_back(
{{MKLDNN_ARG_FROM, *reorder_memory_}, {MKLDNN_ARG_TO, *user_memory_}});
std::shared_ptr<stream> cpu_stream;
- cpu_stream.reset(new stream(*cpu_engine_));
- execute_primitives(net, prim->GetStream(), net_args);
+ cpu_stream.reset(CreateStream(ctx, prim->GetEngine()));
+ execute_primitives(net, cpu_stream, net_args);
#else
net.push_back(FindOrCreateReorder<T>(reorder_memory_, user_memory_));
ExecutePrimitive(net, NET_ARGS_PTR, *cpu_engine_);
@@ -1870,9 +1901,12 @@
class MklPrimitive {
public:
virtual ~MklPrimitive() {}
-
+ MklPrimitive() {}
+ MklPrimitive(const engine& cpu_engine) { cpu_engine_ = cpu_engine; }
// Dummy data which MKL DNN never operates on
unsigned char* DummyData = nullptr;
+ engine cpu_engine_ = engine(ENGINE_CPU, 0);
+ const engine& GetEngine() { return cpu_engine_; }
};
const mkldnn::memory::dims NONE_DIMS = {};
@@ -2058,7 +2092,8 @@
class MklReorderPrimitive : public MklPrimitive {
public:
- explicit MklReorderPrimitive(const memory* from, const memory* to) {
+ explicit MklReorderPrimitive(const memory* from, const memory* to)
+ : MklPrimitive(engine(ENGINE_CPU, 0)) {
Setup(from, to);
}
~MklReorderPrimitive() {}
@@ -2081,7 +2116,6 @@
: src_mem(nullptr), dst_mem(nullptr), reorder_prim(nullptr) {}
} context_;
- engine cpu_engine_ = engine(ENGINE_CPU, 0);
std::shared_ptr<mkldnn::stream> stream_;
void Setup(const memory* from, const memory* to) {
diff --git a/tensorflow/core/util/padding.h b/tensorflow/core/util/padding.h
index 6c109d1..90e353d 100644
--- a/tensorflow/core/util/padding.h
+++ b/tensorflow/core/util/padding.h
@@ -38,7 +38,8 @@
// dimensions as the input.
// EXPLICIT: The user specifies the pad values in the explicit_paddings
// attribute.
-// The padded area is zero-filled.
+// The padded area is typically zero-filled. For pooling ops, the padded area is
+// instead ignored. For max pool, this is equivalent to padding with -infinity.
enum Padding {
VALID = 1, // No padding.
SAME = 2, // Input and output layers have the same size.
diff --git a/tensorflow/core/util/tensor_slice_reader.h b/tensorflow/core/util/tensor_slice_reader.h
index 4aa9a47..0fb2e11 100644
--- a/tensorflow/core/util/tensor_slice_reader.h
+++ b/tensorflow/core/util/tensor_slice_reader.h
@@ -61,7 +61,7 @@
};
typedef std::function<Status(const string&, Table**)> OpenTableFunction;
- static const int kLoadAllShards = -1;
+ static constexpr int kLoadAllShards = -1;
TensorSliceReader(const string& filepattern);
TensorSliceReader(const string& filepattern, OpenTableFunction open_function);
TensorSliceReader(const string& filepattern, OpenTableFunction open_function,
diff --git a/tensorflow/core/util/tensor_slice_writer.h b/tensorflow/core/util/tensor_slice_writer.h
index b610565..86077a5 100644
--- a/tensorflow/core/util/tensor_slice_writer.h
+++ b/tensorflow/core/util/tensor_slice_writer.h
@@ -68,7 +68,7 @@
static size_t MaxBytesPerElement(DataType dt);
private:
- static const size_t kMaxMessageBytes = 1LL << 31;
+ static constexpr size_t kMaxMessageBytes = 1LL << 31;
// Filling in the TensorProto in a SavedSlice will add the following
// header bytes, in addition to the data:
// - 1 byte: TensorProto tag and wire format
@@ -77,7 +77,7 @@
// - <= 5 bytes: *_val length
// However, we add 1KB of slack, to be conservative and guard
// against other additions to the TensorProto.
- static const size_t kTensorProtoHeaderBytes = 1 << 10;
+ static constexpr size_t kTensorProtoHeaderBytes = 1 << 10;
const string filename_;
const CreateBuilderFunction create_builder_;
diff --git a/tensorflow/examples/saved_model/integration_tests/saved_model_test.py b/tensorflow/examples/saved_model/integration_tests/saved_model_test.py
index 6333e55..7088567 100644
--- a/tensorflow/examples/saved_model/integration_tests/saved_model_test.py
+++ b/tensorflow/examples/saved_model/integration_tests/saved_model_test.py
@@ -120,6 +120,7 @@
if regularization_loss_multiplier is not None:
use_kwargs[
"regularization_loss_multiplier"] = regularization_loss_multiplier
+ self.skipTest("b/154863403")
self.assertCommandSucceeded("use_mnist_cnn", **use_kwargs)
self.assertCommandSucceeded(
diff --git a/tensorflow/examples/speech_commands/freeze.py b/tensorflow/examples/speech_commands/freeze.py
index 4a48a44..44940b0 100644
--- a/tensorflow/examples/speech_commands/freeze.py
+++ b/tensorflow/examples/speech_commands/freeze.py
@@ -80,6 +80,9 @@
preprocess: How the spectrogram is processed to produce features, for
example 'mfcc', 'average', or 'micro'.
+ Returns:
+ Input and output tensor objects.
+
Raises:
Exception: If the preprocessing mode isn't recognized.
"""
@@ -150,7 +153,59 @@
runtime_settings=runtime_settings)
# Create an output to use for inference.
- tf.nn.softmax(logits, name='labels_softmax')
+ softmax = tf.nn.softmax(logits, name='labels_softmax')
+
+ return reshaped_input, softmax
+
+
+def save_graph_def(file_name, frozen_graph_def):
+ """Writes a graph def file out to disk.
+
+ Args:
+ file_name: Where to save the file.
+ frozen_graph_def: GraphDef proto object to save.
+ """
+ tf.io.write_graph(
+ frozen_graph_def,
+ os.path.dirname(file_name),
+ os.path.basename(file_name),
+ as_text=False)
+ tf.compat.v1.logging.info('Saved frozen graph to %s', file_name)
+
+
+def save_saved_model(file_name, sess, input_tensor, output_tensor):
+ """Writes a SavedModel out to disk.
+
+ Args:
+ file_name: Where to save the file.
+ sess: TensorFlow session containing the graph.
+ input_tensor: Tensor object defining the input's properties.
+ output_tensor: Tensor object defining the output's properties.
+ """
+ # Store the frozen graph as a SavedModel for v2 compatibility.
+ builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(file_name)
+ tensor_info_inputs = {
+ 'input': tf.compat.v1.saved_model.utils.build_tensor_info(input_tensor)
+ }
+ tensor_info_outputs = {
+ 'output': tf.compat.v1.saved_model.utils.build_tensor_info(output_tensor)
+ }
+ signature = (
+ tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
+ inputs=tensor_info_inputs,
+ outputs=tensor_info_outputs,
+ method_name=tf.compat.v1.saved_model.signature_constants
+ .PREDICT_METHOD_NAME))
+ builder.add_meta_graph_and_variables(
+ sess,
+ [tf.compat.v1.saved_model.tag_constants.SERVING],
+ signature_def_map={
+ tf.compat.v1.saved_model.signature_constants
+ .DEFAULT_SERVING_SIGNATURE_DEF_KEY:
+ signature,
+ },
+ )
+ builder.save()
def main(_):
@@ -167,7 +222,7 @@
# Create the model and load its weights.
sess = tf.compat.v1.InteractiveSession()
- create_inference_graph(
+ input_tensor, output_tensor = create_inference_graph(
FLAGS.wanted_words, FLAGS.sample_rate, FLAGS.clip_duration_ms,
FLAGS.clip_stride_ms, FLAGS.window_size_ms, FLAGS.window_stride_ms,
FLAGS.feature_bin_count, FLAGS.model_architecture, FLAGS.preprocess)
@@ -178,12 +233,14 @@
# Turn all the variables into inline constants inside the graph and save it.
frozen_graph_def = graph_util.convert_variables_to_constants(
sess, sess.graph_def, ['labels_softmax'])
- tf.io.write_graph(
- frozen_graph_def,
- os.path.dirname(FLAGS.output_file),
- os.path.basename(FLAGS.output_file),
- as_text=False)
- tf.compat.v1.logging.info('Saved frozen graph to %s', FLAGS.output_file)
+
+ if FLAGS.save_format == 'graph_def':
+ save_graph_def(FLAGS.output_file, frozen_graph_def)
+ elif FLAGS.save_format == 'saved_model':
+ save_saved_model(FLAGS.output_file, sess, input_tensor, output_tensor)
+ else:
+ raise Exception('Unknown save format "%s" (should be "graph_def" or'
+ ' "saved_model")' % (FLAGS.save_format))
if __name__ == '__main__':
@@ -246,5 +303,10 @@
type=str,
default='mfcc',
help='Spectrogram processing mode. Can be "mfcc" or "average"')
+ parser.add_argument(
+ '--save_format',
+ type=str,
+ default='graph_def',
+ help='How to save the result. Can be "graph_def" or "saved_model"')
FLAGS, unparsed = parser.parse_known_args()
tf.compat.v1.app.run(main=main, argv=[sys.argv[0]] + unparsed)
diff --git a/tensorflow/examples/speech_commands/freeze_test.py b/tensorflow/examples/speech_commands/freeze_test.py
index a242453..93a79b0 100644
--- a/tensorflow/examples/speech_commands/freeze_test.py
+++ b/tensorflow/examples/speech_commands/freeze_test.py
@@ -18,8 +18,12 @@
from __future__ import division
from __future__ import print_function
+import os.path
+
from tensorflow.examples.speech_commands import freeze
+from tensorflow.python.framework import graph_util
from tensorflow.python.framework import test_util
+from tensorflow.python.ops.variables import global_variables_initializer
from tensorflow.python.platform import test
@@ -103,6 +107,27 @@
ops = [node.op for node in sess.graph_def.node]
self.assertEqual(0, ops.count('Mfcc'))
+ @test_util.run_deprecated_v1
+ def testCreateSavedModel(self):
+ tmp_dir = self.get_temp_dir()
+ saved_model_path = os.path.join(tmp_dir, 'saved_model')
+ with self.cached_session() as sess:
+ input_tensor, output_tensor = freeze.create_inference_graph(
+ wanted_words='a,b,c,d',
+ sample_rate=16000,
+ clip_duration_ms=1000.0,
+ clip_stride_ms=30.0,
+ window_size_ms=30.0,
+ window_stride_ms=10.0,
+ feature_bin_count=40,
+ model_architecture='conv',
+ preprocess='micro')
+ global_variables_initializer().run()
+ graph_util.convert_variables_to_constants(
+ sess, sess.graph_def, ['labels_softmax'])
+ freeze.save_saved_model(saved_model_path, sess, input_tensor,
+ output_tensor)
+
if __name__ == '__main__':
test.main()
diff --git a/tensorflow/go/op/wrappers.go b/tensorflow/go/op/wrappers.go
index 0d1de42..d841e02 100644
--- a/tensorflow/go/op/wrappers.go
+++ b/tensorflow/go/op/wrappers.go
@@ -9878,12 +9878,26 @@
return op.Output(0)
}
+// DataServiceDatasetAttr is an optional argument to DataServiceDataset.
+type DataServiceDatasetAttr func(optionalAttr)
+
+// DataServiceDatasetTaskRefreshIntervalHintMs sets the optional task_refresh_interval_hint_ms attribute to value.
+// If not specified, defaults to -1
+func DataServiceDatasetTaskRefreshIntervalHintMs(value int64) DataServiceDatasetAttr {
+ return func(m optionalAttr) {
+ m["task_refresh_interval_hint_ms"] = value
+ }
+}
+
// Creates a dataset that reads data from the tf.data service.
-func DataServiceDataset(scope *Scope, address tf.Output, protocol tf.Output, max_outstanding_requests tf.Output, output_types []tf.DataType, output_shapes []tf.Shape) (handle tf.Output) {
+func DataServiceDataset(scope *Scope, address tf.Output, protocol tf.Output, max_outstanding_requests tf.Output, output_types []tf.DataType, output_shapes []tf.Shape, optional ...DataServiceDatasetAttr) (handle tf.Output) {
if scope.Err() != nil {
return
}
attrs := map[string]interface{}{"output_types": output_types, "output_shapes": output_shapes}
+ for _, a := range optional {
+ a(attrs)
+ }
opspec := tf.OpSpec{
Type: "DataServiceDataset",
Input: []tf.Input{
@@ -11879,75 +11893,6 @@
return op.Output(0)
}
-// Converts one or more images from RGB to HSV.
-//
-// Outputs a tensor of the same shape as the `images` tensor, containing the HSV
-// value of the pixels. The output is only well defined if the value in `images`
-// are in `[0,1]`.
-//
-// `output[..., 0]` contains hue, `output[..., 1]` contains saturation, and
-// `output[..., 2]` contains value. All HSV values are in `[0,1]`. A hue of 0
-// corresponds to pure red, hue 1/3 is pure green, and 2/3 is pure blue.
-//
-// Usage Example:
-//
-// >>> blue_image = tf.stack([
-// ... tf.zeros([5,5]),
-// ... tf.zeros([5,5]),
-// ... tf.ones([5,5])],
-// ... axis=-1)
-// >>> blue_hsv_image = tf.image.rgb_to_hsv(blue_image)
-// >>> blue_hsv_image[0,0].numpy()
-// array([0.6666667, 1. , 1. ], dtype=float32)
-//
-//
-// Arguments:
-// images: 1-D or higher rank. RGB data to convert. Last dimension must be size 3.
-//
-// Returns `images` converted to HSV.
-func RGBToHSV(scope *Scope, images tf.Output) (output tf.Output) {
- if scope.Err() != nil {
- return
- }
- opspec := tf.OpSpec{
- Type: "RGBToHSV",
- Input: []tf.Input{
- images,
- },
- }
- op := scope.AddOperation(opspec)
- return op.Output(0)
-}
-
-// Decode the frame(s) of a GIF-encoded image to a uint8 tensor.
-//
-// GIF images with frame or transparency compression are not supported.
-// On Linux and MacOS systems, convert animated GIFs from compressed to
-// uncompressed by running:
-//
-// convert $src.gif -coalesce $dst.gif
-//
-// This op also supports decoding JPEGs and PNGs, though it is cleaner to use
-// `tf.io.decode_image`.
-//
-// Arguments:
-// contents: 0-D. The GIF-encoded image.
-//
-// Returns 4-D with shape `[num_frames, height, width, 3]`. RGB channel order.
-func DecodeGif(scope *Scope, contents tf.Output) (image tf.Output) {
- if scope.Err() != nil {
- return
- }
- opspec := tf.OpSpec{
- Type: "DecodeGif",
- Input: []tf.Input{
- contents,
- },
- }
- op := scope.AddOperation(opspec)
- return op.Output(0)
-}
-
// SampleDistortedBoundingBoxAttr is an optional argument to SampleDistortedBoundingBox.
type SampleDistortedBoundingBoxAttr func(optionalAttr)
@@ -12104,6 +12049,75 @@
return op.Output(0), op.Output(1), op.Output(2)
}
+// Converts one or more images from RGB to HSV.
+//
+// Outputs a tensor of the same shape as the `images` tensor, containing the HSV
+// value of the pixels. The output is only well defined if the value in `images`
+// are in `[0,1]`.
+//
+// `output[..., 0]` contains hue, `output[..., 1]` contains saturation, and
+// `output[..., 2]` contains value. All HSV values are in `[0,1]`. A hue of 0
+// corresponds to pure red, hue 1/3 is pure green, and 2/3 is pure blue.
+//
+// Usage Example:
+//
+// >>> blue_image = tf.stack([
+// ... tf.zeros([5,5]),
+// ... tf.zeros([5,5]),
+// ... tf.ones([5,5])],
+// ... axis=-1)
+// >>> blue_hsv_image = tf.image.rgb_to_hsv(blue_image)
+// >>> blue_hsv_image[0,0].numpy()
+// array([0.6666667, 1. , 1. ], dtype=float32)
+//
+//
+// Arguments:
+// images: 1-D or higher rank. RGB data to convert. Last dimension must be size 3.
+//
+// Returns `images` converted to HSV.
+func RGBToHSV(scope *Scope, images tf.Output) (output tf.Output) {
+ if scope.Err() != nil {
+ return
+ }
+ opspec := tf.OpSpec{
+ Type: "RGBToHSV",
+ Input: []tf.Input{
+ images,
+ },
+ }
+ op := scope.AddOperation(opspec)
+ return op.Output(0)
+}
+
+// Decode the frame(s) of a GIF-encoded image to a uint8 tensor.
+//
+// GIF images with frame or transparency compression are not supported.
+// On Linux and MacOS systems, convert animated GIFs from compressed to
+// uncompressed by running:
+//
+// convert $src.gif -coalesce $dst.gif
+//
+// This op also supports decoding JPEGs and PNGs, though it is cleaner to use
+// `tf.io.decode_image`.
+//
+// Arguments:
+// contents: 0-D. The GIF-encoded image.
+//
+// Returns 4-D with shape `[num_frames, height, width, 3]`. RGB channel order.
+func DecodeGif(scope *Scope, contents tf.Output) (image tf.Output) {
+ if scope.Err() != nil {
+ return
+ }
+ opspec := tf.OpSpec{
+ Type: "DecodeGif",
+ Input: []tf.Input{
+ contents,
+ },
+ }
+ op := scope.AddOperation(opspec)
+ return op.Output(0)
+}
+
// DecodeBmpAttr is an optional argument to DecodeBmp.
type DecodeBmpAttr func(optionalAttr)
diff --git a/tensorflow/lite/c/BUILD b/tensorflow/lite/c/BUILD
index f9549fc..e1702d4 100644
--- a/tensorflow/lite/c/BUILD
+++ b/tensorflow/lite/c/BUILD
@@ -87,6 +87,7 @@
name = "c_api_test",
size = "small",
srcs = ["c_api_test.cc"],
+ copts = tflite_copts(),
data = [
"//tensorflow/lite:testdata/add.bin",
"//tensorflow/lite:testdata/add_quantized.bin",
@@ -103,6 +104,7 @@
name = "c_api_experimental_test",
size = "small",
srcs = ["c_api_experimental_test.cc"],
+ copts = tflite_copts(),
data = ["//tensorflow/lite:testdata/add.bin"],
deps = [
":c_api",
diff --git a/tensorflow/lite/c/c_api_experimental_test.cc b/tensorflow/lite/c/c_api_experimental_test.cc
index 71a08b5..18bc7bb 100644
--- a/tensorflow/lite/c/c_api_experimental_test.cc
+++ b/tensorflow/lite/c/c_api_experimental_test.cc
@@ -25,11 +25,10 @@
TfLiteRegistration* GetDummyRegistration() {
static TfLiteRegistration registration = {
- .init = nullptr,
- .free = nullptr,
- .prepare = nullptr,
- .invoke = [](TfLiteContext*, TfLiteNode*) { return kTfLiteOk; },
- };
+ /*init=*/nullptr,
+ /*free=*/nullptr,
+ /*prepare=*/nullptr,
+ /*invoke=*/[](TfLiteContext*, TfLiteNode*) { return kTfLiteOk; }};
return ®istration;
}
diff --git a/tensorflow/lite/c/common.h b/tensorflow/lite/c/common.h
index 39ec547..81ba071 100644
--- a/tensorflow/lite/c/common.h
+++ b/tensorflow/lite/c/common.h
@@ -41,7 +41,11 @@
extern "C" {
#endif // __cplusplus
-typedef enum TfLiteStatus { kTfLiteOk = 0, kTfLiteError = 1 } TfLiteStatus;
+typedef enum TfLiteStatus {
+ kTfLiteOk = 0,
+ kTfLiteError = 1,
+ kTfLiteDelegateError = 2
+} TfLiteStatus;
// The list of external context types known to TF Lite. This list exists solely
// to avoid conflicts and to ensure ops can share the external contexts they
@@ -178,8 +182,9 @@
#define TF_LITE_ENSURE_STATUS(a) \
do { \
- if ((a) != kTfLiteOk) { \
- return kTfLiteError; \
+ const TfLiteStatus s = (a); \
+ if (s != kTfLiteOk) { \
+ return s; \
} \
} while (0)
@@ -208,8 +213,9 @@
#define TF_LITE_ENSURE_OK(context, status) \
do { \
- if ((status) != kTfLiteOk) { \
- return kTfLiteError; \
+ const TfLiteStatus s = (status); \
+ if ((s) != kTfLiteOk) { \
+ return s; \
} \
} while (0)
diff --git a/tensorflow/lite/core/api/flatbuffer_conversions.h b/tensorflow/lite/core/api/flatbuffer_conversions.h
index df1ada8..d774afe 100644
--- a/tensorflow/lite/core/api/flatbuffer_conversions.h
+++ b/tensorflow/lite/core/api/flatbuffer_conversions.h
@@ -29,7 +29,7 @@
// Interface class for builtin data allocations.
class BuiltinDataAllocator {
public:
- virtual void* Allocate(size_t size) = 0;
+ virtual void* Allocate(size_t size, size_t alignment_hint) = 0;
virtual void Deallocate(void* data) = 0;
// Allocate a structure, but make sure it is a POD structure that doesn't
@@ -38,8 +38,10 @@
// deallocation.
template <typename T>
T* AllocatePOD() {
+ // TODO(b/154346074): Change this to is_trivially_destructible when all
+ // platform targets support that properly.
static_assert(std::is_pod<T>::value, "Builtin data structure must be POD.");
- void* allocated_memory = this->Allocate(sizeof(T));
+ void* allocated_memory = this->Allocate(sizeof(T), alignof(T));
return new (allocated_memory) T;
}
diff --git a/tensorflow/lite/core/api/flatbuffer_conversions_test.cc b/tensorflow/lite/core/api/flatbuffer_conversions_test.cc
index 24d7ec9..89ca3f5 100644
--- a/tensorflow/lite/core/api/flatbuffer_conversions_test.cc
+++ b/tensorflow/lite/core/api/flatbuffer_conversions_test.cc
@@ -47,7 +47,7 @@
class MockDataAllocator : public BuiltinDataAllocator {
public:
MockDataAllocator() : is_allocated_(false) {}
- void* Allocate(size_t size) override {
+ void* Allocate(size_t size, size_t alignment_hint) override {
EXPECT_FALSE(is_allocated_);
const int max_size = kBufferSize;
EXPECT_LE(size, max_size);
diff --git a/tensorflow/lite/core/subgraph.cc b/tensorflow/lite/core/subgraph.cc
index b3e6069..4cebd05 100644
--- a/tensorflow/lite/core/subgraph.cc
+++ b/tensorflow/lite/core/subgraph.cc
@@ -1366,15 +1366,11 @@
auto reset_delegation_if_not_ok = [this](TfLiteStatus status) {
if (status != kTfLiteOk) {
- // This will undo all delegate nodes currently in the graph.
- TF_LITE_ENSURE_STATUS(this->UndoAllDelegates());
- // This will call AllocateTensors, thus-reapplying any (successfully
- // applied) previous delegates.
- TF_LITE_ENSURE_STATUS(this->EnsureMemoryAllocations());
+ TF_LITE_ENSURE_STATUS(RemoveAllDelegates());
ReportError(
- "Restored previous execution plan after delegate application "
+ "Restored original execution plan after delegate application "
"failure.");
- return kTfLiteError;
+ return kTfLiteDelegateError;
}
return kTfLiteOk;
};
diff --git a/tensorflow/lite/core/subgraph.h b/tensorflow/lite/core/subgraph.h
index 845e960..0b0c1e3 100644
--- a/tensorflow/lite/core/subgraph.h
+++ b/tensorflow/lite/core/subgraph.h
@@ -531,6 +531,12 @@
// be reallocated if the graph was modified (i.e., the caller does *not* need
// to explicitly call |AllocateTensors()| again). If tensors were unallocated,
// they will remain unallocated after delegate application.
+ // Returns one of the following three status codes:
+ // 1. kTfLiteOk: Delegation succeeded
+ // 2. kTfLiteDelegateError: Delegation failed due to an error in the
+ // delegate. The Subgraph has been restored to its pre-delegation state.
+ // NOTE: This reverts all delegates previously applied to the Subgraph.
+ // 3. kTfLiteError: Unexpected/runtime failure.
TfLiteStatus ModifyGraphWithDelegate(TfLiteDelegate* delegate);
// This un-applies all delegates that have been applied till now, but retains
diff --git a/tensorflow/lite/delegates/gpu/api.h b/tensorflow/lite/delegates/gpu/api.h
index 921f2d5..2a531f1 100644
--- a/tensorflow/lite/delegates/gpu/api.h
+++ b/tensorflow/lite/delegates/gpu/api.h
@@ -54,7 +54,7 @@
// H - height
// W - width
// C - channels
-// D - depth := IntegralDivideRoundUp(C, 4)
+// D - depth := DivideRoundUp(C, 4)
// C4 - is the constant = 4.
enum class DataLayout {
UNKNOWN,
@@ -164,7 +164,7 @@
Dimensions(int32_t batch, int32_t height, int32_t width, int32_t channels)
: b(batch), h(height), w(width), c(channels) {}
- int32_t d() const { return IntegralDivideRoundUp(c, 4); }
+ int32_t d() const { return DivideRoundUp(c, 4); }
int32_t product() const { return b * h * w * c; }
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/add.cc b/tensorflow/lite/delegates/gpu/cl/kernels/add.cc
index 4eaff12..b5d8920 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/add.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/add.cc
@@ -70,10 +70,10 @@
Add::Add(const OperationDef& definition, const std::vector<int>& channels,
int dst_channels)
: ElementwiseOperation(definition),
- dst_depth_(IntegralDivideRoundUp(dst_channels, 4)) {
+ dst_depth_(DivideRoundUp(dst_channels, 4)) {
src_depthes_.resize(channels.size());
for (int i = 0; i < channels.size(); ++i) {
- src_depthes_[i] = IntegralDivideRoundUp(channels[i], 4);
+ src_depthes_[i] = DivideRoundUp(channels[i], 4);
}
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc b/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
index 3a7ec1c..f1970ce 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/concat_z.cc
@@ -79,7 +79,7 @@
// generation.
c += " int Z = 0;\n";
for (int i = 0; i < channels.size(); ++i) {
- const int depth = IntegralDivideRoundUp(channels[i], 4);
+ const int depth = DivideRoundUp(channels[i], 4);
if (depth % 2 == 0) {
// We can read more at once inside of loop in case depth % 2 == 0
// it should be better for reading latency hiding
@@ -112,7 +112,7 @@
int read_index = 0;
int z = 0;
for (int i = 0; i < channels.size(); ++i) {
- const int depth = IntegralDivideRoundUp(channels[i], 4);
+ const int depth = DivideRoundUp(channels[i], 4);
for (int d = 0; d < depth; ++d) {
const int channels_in_group = std::min(4, channels[i] - d * 4);
const std::string temp_name = "t" + std::to_string(read_index);
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.cc
index b79599d..564f0d1 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.cc
@@ -128,24 +128,24 @@
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Batch()));
}
RETURN_IF_ERROR(kernel_.SetBytesAuto(
- IntegralDivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.w)));
+ DivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.w)));
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHDS()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWBatchedHDS()));
return absl::OkStatus();
}
int3 Conv3D::GetGridSize() const {
- const int grid_x = IntegralDivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(),
- conv_params_.block_size.x);
+ const int grid_x = DivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(),
+ conv_params_.block_size.x);
const int grid_y =
- IntegralDivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
+ DivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
const int grid_z =
- IntegralDivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.w) *
- IntegralDivideRoundUp(dst_[0]->Depth(), conv_params_.block_size.z);
+ DivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.w) *
+ DivideRoundUp(dst_[0]->Depth(), conv_params_.block_size.z);
int3 wg;
- wg.x = IntegralDivideRoundUp(grid_x, conv_params_.work_group_size.x);
- wg.y = IntegralDivideRoundUp(grid_y, conv_params_.work_group_size.y);
- wg.z = IntegralDivideRoundUp(grid_z, conv_params_.work_group_size.z);
+ wg.x = DivideRoundUp(grid_x, conv_params_.work_group_size.x);
+ wg.y = DivideRoundUp(grid_y, conv_params_.work_group_size.y);
+ wg.z = DivideRoundUp(grid_z, conv_params_.work_group_size.z);
return int3(wg[conv_params_.work_group_launch_order[0]] *
conv_params_.work_group_size.x,
wg[conv_params_.work_group_launch_order[1]] *
@@ -885,8 +885,8 @@
Conv3D::ConvParams Conv3D::GuessBestParams(
const CLDevice& device, const OperationDef& definition,
const Convolution3DAttributes& attr) const {
- const int dst_slices = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_slices = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_slices = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_slices = DivideRoundUp(attr.weights.shape.i, 4);
const bool x_kernel_is_1 = attr.weights.shape.w == 1 && attr.strides.w == 1 &&
attr.dilations.w == 1 &&
attr.padding.prepended.w == 0 &&
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.h
index 8df7994..8dfeac1 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_3d.h
@@ -147,8 +147,8 @@
CLContext* context) {
const int block_size = conv_params_.block_size.w;
const int dst_slices =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size);
- const int src_slices = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size);
+ const int src_slices = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
const int kernel_z = kernel_size_.z;
@@ -219,8 +219,8 @@
absl::Span<T> dst) {
const int block_size = conv_params_.block_size.w;
const int dst_slices =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size);
- const int src_slices = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size);
+ const int src_slices = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
const int kernel_z = kernel_size_.z;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.cc
index 70bd1b5..4acd6b3 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.cc
@@ -307,7 +307,7 @@
RETURN_IF_ERROR(kernel_.SetMemoryAuto(biases_.GetMemoryPtr()));
RETURN_IF_ERROR(BindArgs(&kernel_, linked_operations_));
RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
- const int src_width_elements = IntegralDivideRoundUp(
+ const int src_width_elements = DivideRoundUp(
src_[0]->Width() * src_[0]->Batch(), (conv_params_.element_size / 4));
int4 src_size = int4(src_width_elements, src_[0]->Height(), src_[0]->Slices(),
src_width_elements * src_[0]->Height());
@@ -317,14 +317,14 @@
}
int3 ConvBuffer1x1::GetGridSize() const {
- const int dst_width_elements = IntegralDivideRoundUp(
+ const int dst_width_elements = DivideRoundUp(
dst_[0]->Width() * dst_[0]->Batch(), (conv_params_.element_size / 4));
const int grid_x =
- IntegralDivideRoundUp(dst_width_elements, conv_params_.block_size.x);
+ DivideRoundUp(dst_width_elements, conv_params_.block_size.x);
const int grid_y =
- IntegralDivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
+ DivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
const int grid_z =
- IntegralDivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.z);
+ DivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.z);
return int3(grid_x, grid_y, grid_z);
}
@@ -358,8 +358,8 @@
if (!IsConvBuffer1x1Supported(definition, attr)) {
return absl::InvalidArgumentError("ConvBuffer1x1 doesn't supported");
}
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
ConvBuffer1x1::ConvParams conv_params;
if (shape) {
conv_params = GetBestParams(*creation_context.device, definition, *shape,
@@ -376,8 +376,8 @@
const OperationDef& definition,
const FullyConnectedAttributes& attr,
ConvBuffer1x1* result, const BHWC* shape) {
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
ConvBuffer1x1::ConvParams conv_params;
if (shape) {
conv_params = GetBestParams(*creation_context.device, definition, *shape,
@@ -396,8 +396,8 @@
const CreationContext& creation_context, const OperationDef& definition,
const Convolution2DAttributes& attr, ConvBuffer1x1* result,
const BHWC* shape) {
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
ConvBuffer1x1::ConvParams conv_params;
if (shape) {
conv_params = GetBestParams(*creation_context.device, definition, *shape,
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.h
index 7059572..dbda924 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_buffer_1x1.h
@@ -135,8 +135,8 @@
template <DataType T>
absl::Status ConvBuffer1x1::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const bool f32_weights = definition_.precision == CalculationsPrecision::F32;
const int float4_size = f32_weights ? sizeof(float4) : sizeof(half4);
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc
index 07d2da9..d4dc206 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.cc
@@ -40,9 +40,9 @@
std::string c = GetCommonDefines(op_def.precision);
- const int out_z = IntegralDivideRoundUp(dst_channels, 4);
+ const int out_z = DivideRoundUp(dst_channels, 4);
const std::string kOutZ = std::to_string(out_z);
- const int src_depth = IntegralDivideRoundUp(src_channels, 4);
+ const int src_depth = DivideRoundUp(src_channels, 4);
const auto src_tensor_type = op_def.src_tensors[0].storage_type;
const bool manual_clamp = src_tensor_type == TensorStorageType::BUFFER ||
@@ -290,7 +290,7 @@
: sizeof(half);
const int filters_buffer_size = filters_count * float_size;
const int kConstantMaxSize = GetOptimalMaxConstantSize(device.GetInfo());
- const int flt4_registers = IntegralDivideRoundUp(w_shape.o, 4);
+ const int flt4_registers = DivideRoundUp(w_shape.o, 4);
return filters_buffer_size <= kConstantMaxSize && flt4_registers <= 8;
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h
index 15049cf..8d80d48 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_constants.h
@@ -88,7 +88,7 @@
template <DataType T>
absl::Status ConvConstants::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
@@ -112,8 +112,8 @@
template <DataType S, typename T>
void ConvConstants::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
index bd4f533..7ba12df 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.cc
@@ -205,8 +205,8 @@
kernel_dilation_.z * src_[0]->Batch(), kernel_dilation_.w)));
}
if (conv_params_.linear_hw) {
- const int grid_x = IntegralDivideRoundUp(
- dst_[0]->Width() * dst_[0]->Batch(), conv_params_.block_size.x);
+ const int grid_x = DivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(),
+ conv_params_.block_size.x);
RETURN_IF_ERROR(kernel_.SetBytesAuto(grid_x));
}
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWBatchedHSB()));
@@ -215,27 +215,26 @@
}
int3 ConvPowerVR::GetGridSize() const {
- const int grid_x = IntegralDivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(),
- conv_params_.block_size.x);
+ const int grid_x = DivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(),
+ conv_params_.block_size.x);
const int grid_y =
- IntegralDivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
+ DivideRoundUp(dst_[0]->Height(), conv_params_.block_size.y);
const int grid_z =
- IntegralDivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.z);
+ DivideRoundUp(dst_[0]->Slices(), conv_params_.block_size.z);
int3 wg;
if (conv_params_.linear_hw) {
- wg.x =
- IntegralDivideRoundUp(grid_x * grid_y, conv_params_.work_group_size.x);
- wg.y = IntegralDivideRoundUp(grid_z, conv_params_.work_group_size.y);
+ wg.x = DivideRoundUp(grid_x * grid_y, conv_params_.work_group_size.x);
+ wg.y = DivideRoundUp(grid_z, conv_params_.work_group_size.y);
return int3(wg[conv_params_.work_group_launch_order[0]] *
conv_params_.work_group_size.x,
wg[conv_params_.work_group_launch_order[1]] *
conv_params_.work_group_size.y,
1);
} else {
- wg.x = IntegralDivideRoundUp(grid_x, conv_params_.work_group_size.x);
- wg.y = IntegralDivideRoundUp(grid_y, conv_params_.work_group_size.y);
- wg.z = IntegralDivideRoundUp(grid_z, conv_params_.work_group_size.z);
+ wg.x = DivideRoundUp(grid_x, conv_params_.work_group_size.x);
+ wg.y = DivideRoundUp(grid_y, conv_params_.work_group_size.y);
+ wg.z = DivideRoundUp(grid_z, conv_params_.work_group_size.z);
return int3(wg[conv_params_.work_group_launch_order[0]] *
conv_params_.work_group_size.x,
wg[conv_params_.work_group_launch_order[1]] *
@@ -808,8 +807,8 @@
ConvPowerVR::ConvParams ConvPowerVR::GuessBestParams(
const CLDevice& device, const OperationDef& definition,
const Convolution2DAttributes& attr, const BHWC* dst_shape) const {
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
const bool x_kernel_is_1 = attr.weights.shape.w == 1 && attr.strides.w == 1 &&
attr.dilations.w == 1 &&
attr.padding.prepended.w == 0 &&
@@ -825,8 +824,8 @@
ConvPowerVR::ConvParams ConvPowerVR::GuessBestParams(
const CLDevice& device, const OperationDef& definition,
const FullyConnectedAttributes& attr, const BHWC* dst_shape) const {
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
ConvPowerVR::ConvParams params = GuessBestParams(
device, definition, src_depth, dst_depth, true, true, false, dst_shape);
params.work_group_size.x *= params.work_group_size.y;
@@ -839,8 +838,8 @@
ConvPowerVR::ConvParams ConvPowerVR::GuessBestParamsWinograd(
const CLDevice& device, const OperationDef& definition,
const Convolution2DAttributes& attr, const BHWC* dst_shape) const {
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
ConvPowerVR::ConvParams params = GuessBestParams(
device, definition, src_depth, dst_depth, true, true, true, dst_shape);
params.block_size.x *= params.block_size.y;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.h
index 3a1332a..01f77ee 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_powervr.h
@@ -188,8 +188,8 @@
template <DataType T>
absl::Status ConvPowerVR::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const bool f32_weights = conv_params_.weights_data_type == DataType::FLOAT32;
const int float4_size = f32_weights ? sizeof(float4) : sizeof(half4);
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
index 953f564..e92cc13 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.cc
@@ -433,9 +433,9 @@
int3 ConvTexture::GetGridSize() const {
const int grid_x =
- IntegralDivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(), block_size_.x);
- const int grid_y = IntegralDivideRoundUp(dst_[0]->Height(), block_size_.y);
- const int grid_z = IntegralDivideRoundUp(dst_[0]->Slices(), block_size_.z);
+ DivideRoundUp(dst_[0]->Width() * dst_[0]->Batch(), block_size_.x);
+ const int grid_y = DivideRoundUp(dst_[0]->Height(), block_size_.y);
+ const int grid_z = DivideRoundUp(dst_[0]->Slices(), block_size_.z);
return int3(grid_x, grid_y, grid_z);
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
index dd91572..42f7ecd 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/conv_texture.h
@@ -148,9 +148,9 @@
template <DataType T>
absl::Status ConvTexture::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ int dst_depth = DivideRoundUp(weights.shape.o, 4);
dst_depth = AlignByN(dst_depth, block_size_.z);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
@@ -206,9 +206,9 @@
void ConvTexture::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst_0,
absl::Span<T> dst_1, absl::Span<T> dst_2, absl::Span<T> dst_3) {
- int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ int dst_depth = DivideRoundUp(weights.shape.o, 4);
dst_depth = AlignByN(dst_depth, block_size_.z);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.cc
index 417fb63..7d7ebeb 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.cc
@@ -322,7 +322,7 @@
block_size_ = is_f16 ? int3(2, 2, 2) : int3(2, 2, 1);
}
}
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
if (dst_depth == 1 || dst_depth == 3) {
if (!device.IsMali()) {
block_size_.y *= block_size_.z;
@@ -406,10 +406,9 @@
int3 ConvolutionTransposed::GetGridSize() const {
const int aligned_w = AlignByN(dst_[0]->Width(), stride_.x * block_size_.x);
const int aligned_h = AlignByN(dst_[0]->Height(), stride_.y * block_size_.y);
- const int grid_x =
- IntegralDivideRoundUp(aligned_w, block_size_.x) * dst_[0]->Batch();
- const int grid_y = IntegralDivideRoundUp(aligned_h, block_size_.y);
- const int grid_z = IntegralDivideRoundUp(dst_[0]->Slices(), block_size_.z);
+ const int grid_x = DivideRoundUp(aligned_w, block_size_.x) * dst_[0]->Batch();
+ const int grid_y = DivideRoundUp(aligned_h, block_size_.y);
+ const int grid_z = DivideRoundUp(dst_[0]->Slices(), block_size_.z);
return int3(grid_x, grid_y, grid_z);
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.h
index 57fdad1..867966f 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed.h
@@ -91,8 +91,8 @@
absl::Status ConvolutionTransposed::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
const int dst_depth =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size_.z);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size_.z);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
int texture_width = dst_depth;
@@ -160,8 +160,8 @@
void ConvolutionTransposed::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
const int dst_depth =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size_.z);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size_.z);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
int texture_width = dst_depth;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.cc
index 9d3f0b2..4f024ee 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.cc
@@ -440,8 +440,8 @@
if (definition_.IsBatchSupported()) {
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->Batch()));
}
- RETURN_IF_ERROR(kernel_.SetBytesAuto(
- IntegralDivideRoundUp(dst_[0]->Slices(), block_size_.w)));
+ RETURN_IF_ERROR(
+ kernel_.SetBytesAuto(DivideRoundUp(dst_[0]->Slices(), block_size_.w)));
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHDS()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWHDS()));
return absl::OkStatus();
@@ -451,11 +451,10 @@
const int aligned_w = AlignByN(dst_[0]->Width(), stride_.x * block_size_.x);
const int aligned_h = AlignByN(dst_[0]->Height(), stride_.y * block_size_.y);
const int aligned_d = AlignByN(dst_[0]->Depth(), stride_.z * block_size_.z);
- const int grid_x =
- IntegralDivideRoundUp(aligned_w, block_size_.x) * dst_[0]->Batch();
- const int grid_y = IntegralDivideRoundUp(aligned_h, block_size_.y);
- const int grid_z = IntegralDivideRoundUp(dst_[0]->Slices(), block_size_.w) *
- IntegralDivideRoundUp(aligned_d, block_size_.z);
+ const int grid_x = DivideRoundUp(aligned_w, block_size_.x) * dst_[0]->Batch();
+ const int grid_y = DivideRoundUp(aligned_h, block_size_.y);
+ const int grid_z = DivideRoundUp(dst_[0]->Slices(), block_size_.w) *
+ DivideRoundUp(aligned_d, block_size_.z);
return int3(grid_x, grid_y, grid_z);
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.h
index c610d11..14757ef 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3d.h
@@ -91,8 +91,8 @@
absl::Status ConvolutionTransposed3D::UploadWeights(
const tflite::gpu::Tensor<OHWDI, T>& weights, CLContext* context) {
const int dst_depth =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size_.z);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size_.z);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
const int kernel_z = kernel_size_.z;
@@ -162,8 +162,8 @@
void ConvolutionTransposed3D::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWDI, S>& weights, absl::Span<T> dst) {
const int dst_depth =
- AlignByN(IntegralDivideRoundUp(weights.shape.o, 4), block_size_.w);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ AlignByN(DivideRoundUp(weights.shape.o, 4), block_size_.w);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
const int kernel_z = kernel_size_.z;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc
index 4be593b..4a68eda 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.cc
@@ -339,14 +339,13 @@
}
int3 ConvolutionTransposed3x3::GetGridSize() const {
- const int grid_x =
- IntegralDivideRoundUp(dst_[0]->Width(), 2) * dst_[0]->Batch();
- const int grid_y = IntegralDivideRoundUp(dst_[0]->Height(), 2);
+ const int grid_x = DivideRoundUp(dst_[0]->Width(), 2) * dst_[0]->Batch();
+ const int grid_y = DivideRoundUp(dst_[0]->Height(), 2);
const int grid_z = dst_[0]->Slices();
int3 wg;
- wg.x = IntegralDivideRoundUp(grid_x, work_group_size_.x);
- wg.y = IntegralDivideRoundUp(grid_y, work_group_size_.y);
- wg.z = IntegralDivideRoundUp(grid_z, work_group_size_.z);
+ wg.x = DivideRoundUp(grid_x, work_group_size_.x);
+ wg.y = DivideRoundUp(grid_y, work_group_size_.y);
+ wg.z = DivideRoundUp(grid_z, work_group_size_.z);
return int3(wg[work_group_launch_order_[0]] * work_group_size_.x,
wg[work_group_launch_order_[1]] * work_group_size_.y,
wg[work_group_launch_order_[2]] * work_group_size_.z);
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.h
index 7ff3109..fa44d6a 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3.h
@@ -84,8 +84,8 @@
template <DataType T>
absl::Status ConvolutionTransposed3x3::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int kernel_x = 3; // This operation support only 3x3 kernel
const int kernel_y = 3;
const int flt4_count = kernel_x * kernel_y * src_depth * dst_depth * 4;
@@ -109,8 +109,8 @@
template <DataType S, typename T>
void ConvolutionTransposed3x3::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int kernel_x = 3;
const int kernel_y = 3;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.cc
index b8e4b25..d65ff07 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.cc
@@ -224,8 +224,8 @@
absl::Status ConvolutionTransposed3x3Thin::Compile(
const CreationContext& creation_context) {
const auto code = GenerateConvolutionTransposedCode(
- definition_, biases_, IntegralDivideRoundUp(src_channels_, 4),
- IntegralDivideRoundUp(dst_channels_, 4), *creation_context.device,
+ definition_, biases_, DivideRoundUp(src_channels_, 4),
+ DivideRoundUp(dst_channels_, 4), *creation_context.device,
linked_operations_);
return creation_context.cache->GetOrCreateCLKernel(
code, "main_function", *creation_context.context,
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.h
index 2f60ac5..447afb6 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_3x3_thin.h
@@ -82,8 +82,8 @@
template <DataType T>
absl::Status ConvolutionTransposed3x3Thin::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(src_channels_, 4);
- const int dst_depth = IntegralDivideRoundUp(dst_channels_, 4);
+ const int src_depth = DivideRoundUp(src_channels_, 4);
+ const int dst_depth = DivideRoundUp(dst_channels_, 4);
const int kernel_x = 3; // This operation support only 3x3 kernel
const int kernel_y = 3;
const int flt4_count = kernel_x * kernel_y * src_depth * dst_depth * 4;
@@ -108,8 +108,8 @@
template <DataType S, typename T>
void ConvolutionTransposed3x3Thin::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
- const int src_depth = IntegralDivideRoundUp(src_channels_, 4);
- const int dst_depth = IntegralDivideRoundUp(dst_channels_, 4);
+ const int src_depth = DivideRoundUp(src_channels_, 4);
+ const int dst_depth = DivideRoundUp(dst_channels_, 4);
const int kernel_x = 3;
const int kernel_y = 3;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc
index a558fe6..0f7f909 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.cc
@@ -332,9 +332,8 @@
}
int3 ConvolutionTransposed4x4::GetGridSize() const {
- const int grid_x =
- IntegralDivideRoundUp(dst_[0]->Width() + 2, 2) * dst_[0]->Batch();
- const int grid_y = IntegralDivideRoundUp(dst_[0]->Height() + 2, 2);
+ const int grid_x = DivideRoundUp(dst_[0]->Width() + 2, 2) * dst_[0]->Batch();
+ const int grid_y = DivideRoundUp(dst_[0]->Height() + 2, 2);
const int grid_z = dst_[0]->Slices();
return int3(grid_x, grid_y, grid_z);
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.h
index 9f514d1..870c72f 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_4x4.h
@@ -82,8 +82,8 @@
template <DataType T>
absl::Status ConvolutionTransposed4x4::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int kernel_x = 4; // This operation support only 4x4 kernel
const int kernel_y = 4;
const int flt4_count = kernel_x * kernel_y * src_depth * dst_depth * 4;
@@ -107,8 +107,8 @@
template <DataType S, typename T>
void ConvolutionTransposed4x4::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int kernel_x = 4;
const int kernel_y = 4;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.cc b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.cc
index 8ea40be..8eca689 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.cc
@@ -187,8 +187,8 @@
absl::Status ConvolutionTransposedThin::Compile(
const CreationContext& creation_context) {
const auto code = GenerateConvolutionTransposedCode(
- definition_, IntegralDivideRoundUp(src_channels_, 4), dst_channels_,
- kernel_size_, *creation_context.device, linked_operations_);
+ definition_, DivideRoundUp(src_channels_, 4), dst_channels_, kernel_size_,
+ *creation_context.device, linked_operations_);
std::vector<CompilerOptions> options;
if (definition_.precision == CalculationsPrecision::F16 &&
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.h b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.h
index 8cf9a7a..db2ad8c 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/convolution_transposed_thin.h
@@ -82,7 +82,7 @@
template <DataType T>
absl::Status ConvolutionTransposedThin::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(src_channels_, 4);
+ const int src_depth = DivideRoundUp(src_channels_, 4);
const int elements_count =
kernel_size_.x * kernel_size_.y * src_depth * 4 * dst_channels_;
@@ -104,7 +104,7 @@
template <DataType S, typename T>
void ConvolutionTransposedThin::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
- const int src_depth = IntegralDivideRoundUp(src_channels_, 4);
+ const int src_depth = DivideRoundUp(src_channels_, 4);
const int kernel_x = kernel_size_.x;
const int kernel_y = kernel_size_.y;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h
index 9d3e336..7655f2a 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv.h
@@ -89,7 +89,7 @@
absl::Status DepthwiseConvolution::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
const int dst_channels = weights.shape.i * weights.shape.o;
- const int dst_depth = IntegralDivideRoundUp(dst_channels, 4);
+ const int dst_depth = DivideRoundUp(dst_channels, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
@@ -137,7 +137,7 @@
void DepthwiseConvolution::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
const int dst_channels = weights.shape.i * weights.shape.o;
- const int dst_depth = IntegralDivideRoundUp(dst_channels, 4);
+ const int dst_depth = DivideRoundUp(dst_channels, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h
index 53e38a3..3c87ba5 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3d.h
@@ -88,7 +88,7 @@
absl::Status DepthwiseConvolution3D::UploadWeights(
const tflite::gpu::Tensor<OHWDI, T>& weights, CLContext* context) {
const int dst_channels = weights.shape.i * weights.shape.o;
- const int dst_slices = IntegralDivideRoundUp(dst_channels, 4);
+ const int dst_slices = DivideRoundUp(dst_channels, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
const int kernel_z = weights.shape.d;
@@ -130,7 +130,7 @@
void DepthwiseConvolution3D::RearrangeWeightsData(
const tflite::gpu::Tensor<OHWDI, S>& weights, absl::Span<T> dst) {
const int dst_channels = weights.shape.i * weights.shape.o;
- const int dst_slices = IntegralDivideRoundUp(dst_channels, 4);
+ const int dst_slices = DivideRoundUp(dst_channels, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
const int kernel_z = weights.shape.d;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.cc b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.cc
index 348229e..c8ac825 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.cc
@@ -323,8 +323,8 @@
}
int3 DepthwiseConv3x3::GetGridSize() const {
- const int grid_x = IntegralDivideRoundUp(dst_[0]->Width(), 2);
- const int grid_y = IntegralDivideRoundUp(dst_[0]->Height(), 2);
+ const int grid_x = DivideRoundUp(dst_[0]->Width(), 2);
+ const int grid_y = DivideRoundUp(dst_[0]->Height(), 2);
const int grid_z = dst_[0]->Slices();
return int3(grid_x, grid_y, grid_z);
}
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.h b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.h
index ac7c316..1ab17e3 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/depthwise_conv_3x3.h
@@ -83,7 +83,7 @@
absl::Status DepthwiseConv3x3::UploadWeightsAndBiases(
const tflite::gpu::Tensor<OHWI, T>& weights,
const tflite::gpu::Tensor<Linear, T>& biases, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
int texture_width = 10; // 3x3 kernel + 1 bias
int texture_height = src_depth;
const int elements_count = texture_width * texture_height;
@@ -129,7 +129,7 @@
void DepthwiseConv3x3::RearrangeWeightsAndBiasesData(
const tflite::gpu::Tensor<OHWI, S>& weights,
const tflite::gpu::Tensor<Linear, S>& biases, absl::Span<T> dst) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
int counter = 0;
for (int s = 0; s < src_depth; ++s) {
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
index 180f676..0be5288 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/fully_connected.h
@@ -70,8 +70,8 @@
template <DataType T>
absl::Status FullyConnected::UploadWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, CLContext* context) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
const int elements_count = src_depth * dst_depth * 4;
const bool f32_weights = definition_.precision == CalculationsPrecision::F32;
@@ -94,8 +94,8 @@
template <DataType T, typename S>
void FullyConnected::RearrangeWeights(
const tflite::gpu::Tensor<OHWI, T>& weights, absl::Span<S> dst) {
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
int counter = 0;
for (int s = 0; s < src_depth; ++s) {
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc
index fb98546..e292f2d 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/pooling.cc
@@ -185,9 +185,11 @@
TensorCodeGenerator dst_tensor(
"dst_data", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"},
op_def.dst_tensors[0]);
+ const auto dst_ind_def =
+ output_indices ? op_def.dst_tensors[1] : op_def.dst_tensors[0];
TensorCodeGenerator indices_tensor(
"dst_indices", WHSPoint{"dst_size.x", "dst_size.y", "dst_size.z"},
- op_def.dst_tensors[1]);
+ dst_ind_def);
std::string c = GetCommonDefines(op_def.precision);
@@ -281,10 +283,12 @@
"dst_data",
WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"},
op_def.dst_tensors[0]);
+ const auto dst_ind_def =
+ output_indices ? op_def.dst_tensors[1] : op_def.dst_tensors[0];
TensorCodeGenerator indices_tensor(
"dst_indices",
WHDSPoint{"dst_size.x", "dst_size.y", "dst_size.z", "dst_size.w"},
- op_def.dst_tensors[1]);
+ dst_ind_def);
std::string c = GetCommonDefines(op_def.precision);
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/softmax1x1.cc b/tensorflow/lite/delegates/gpu/cl/kernels/softmax1x1.cc
index 09e6c97..192bee7 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/softmax1x1.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/softmax1x1.cc
@@ -129,8 +129,7 @@
RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
const int depth = src_[0]->Slices();
- RETURN_IF_ERROR(
- kernel_.SetBytesAuto(int2(depth, IntegralDivideRoundUp(depth, 32))));
+ RETURN_IF_ERROR(kernel_.SetBytesAuto(int2(depth, DivideRoundUp(depth, 32))));
RETURN_IF_ERROR(
kernel_.SetBytesAuto(GetMaskForLastPlane(src_[0]->Channels())));
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/util.h b/tensorflow/lite/delegates/gpu/cl/kernels/util.h
index aeb8ee1..836a95f 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/util.h
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/util.h
@@ -234,12 +234,12 @@
void RearrangeWeightsToOHWIOGroupI4O4(
const tflite::gpu::Tensor<OHWI, S>& weights, int out_group_size,
absl::Span<T> dst) {
- const int dst_slices = IntegralDivideRoundUp(weights.shape.o, 4);
- const int src_slices = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int dst_slices = DivideRoundUp(weights.shape.o, 4);
+ const int src_slices = DivideRoundUp(weights.shape.i, 4);
const int kernel_x = weights.shape.w;
const int kernel_y = weights.shape.h;
- const int dst_groups = IntegralDivideRoundUp(dst_slices, out_group_size);
+ const int dst_groups = DivideRoundUp(dst_slices, out_group_size);
int counter = 0;
for (int d = 0; d < dst_groups; ++d) {
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc b/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
index b6d04d4..6219952 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/winograd.cc
@@ -436,9 +436,9 @@
RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWHSB()));
- const int tiles_x = IntegralDivideRoundUp(
+ const int tiles_x = DivideRoundUp(
src_[0]->Width() + padding_.prepended.w + padding_.appended.w - 2, 4);
- const int tiles_y = IntegralDivideRoundUp(
+ const int tiles_y = DivideRoundUp(
src_[0]->Height() + padding_.prepended.h + padding_.appended.h - 2, 4);
const int tiles_total = tiles_x * tiles_y;
RETURN_IF_ERROR(
@@ -550,14 +550,14 @@
RETURN_IF_ERROR(kernel_.SetMemoryAuto(dst_[0]->GetMemoryPtrForWriting()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(src_[0]->GetWHSB()));
RETURN_IF_ERROR(kernel_.SetBytesAuto(dst_[0]->GetWHSB()));
- const int tiles_x = IntegralDivideRoundUp(dst_[0]->Width(), 4);
+ const int tiles_x = DivideRoundUp(dst_[0]->Width(), 4);
RETURN_IF_ERROR(kernel_.SetBytesAuto(tiles_x));
return absl::OkStatus();
}
int3 Winograd36To4x4::GetGridSize() const {
- const int tiles_x = IntegralDivideRoundUp(dst_[0]->Width(), 4);
- const int tiles_y = IntegralDivideRoundUp(dst_[0]->Height(), 4);
+ const int tiles_x = DivideRoundUp(dst_[0]->Width(), 4);
+ const int tiles_y = DivideRoundUp(dst_[0]->Height(), 4);
const int grid_x = tiles_x * tiles_y * dst_[0]->Batch();
const int grid_y = 4;
const int grid_z = dst_[0]->Slices();
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.cc b/tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.cc
index 20c119f..5e280d5 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/work_group_picking.cc
@@ -187,7 +187,7 @@
int3 GetWorkGroup(const int3& grid, int max_size) {
int wg_z = GetBiggestDividerWithPriority(grid.z, 8);
int wg_xy_size = max_size / wg_z;
- int wg_x = std::min(IntegralDivideRoundUp(grid.x, 2), wg_xy_size);
+ int wg_x = std::min(DivideRoundUp(grid.x, 2), wg_xy_size);
int wg_y = std::min(wg_xy_size / wg_x, grid.y);
return int3(wg_x, wg_y, wg_z);
}
@@ -231,12 +231,12 @@
}
bool XY128RequiresMoreWorkGroupsThenXY128Linear(int width, int height) {
- int planar_work_groups = IntegralDivideRoundUp(width * height, 128);
+ int planar_work_groups = DivideRoundUp(width * height, 128);
auto base_work_groups = Get2DWorkgroupsEqualTo128();
bool have_equal_work_groups = false;
for (auto& work_group : base_work_groups) {
- int x_groups = IntegralDivideRoundUp(width, work_group.x);
- int y_groups = IntegralDivideRoundUp(height, work_group.y);
+ int x_groups = DivideRoundUp(width, work_group.x);
+ int y_groups = DivideRoundUp(height, work_group.y);
int xy_groups = x_groups * y_groups;
if (xy_groups == planar_work_groups) {
have_equal_work_groups = true;
diff --git a/tensorflow/lite/delegates/gpu/cl/linear_storage.h b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
index 0406ca1..f461b08 100644
--- a/tensorflow/lite/delegates/gpu/cl/linear_storage.h
+++ b/tensorflow/lite/delegates/gpu/cl/linear_storage.h
@@ -101,7 +101,7 @@
CLContext* context, LinearStorage* result) {
int size = creation_info.aligned_size != 0 ? creation_info.aligned_size
: tensor.shape.v;
- const int depth = IntegralDivideRoundUp(size, 4);
+ const int depth = DivideRoundUp(size, 4);
if (creation_info.data_type == DataType::FLOAT32) {
std::vector<float4> gpu_data(depth);
CopyLinearFLT4(tensor, absl::MakeSpan(gpu_data));
diff --git a/tensorflow/lite/delegates/gpu/cl/model_hints.h b/tensorflow/lite/delegates/gpu/cl/model_hints.h
index 274064d..7661cc0 100644
--- a/tensorflow/lite/delegates/gpu/cl/model_hints.h
+++ b/tensorflow/lite/delegates/gpu/cl/model_hints.h
@@ -26,11 +26,11 @@
using ModelHint = uint64_t;
// By default we want the fastest inference
- static const ModelHint kFastestInference = 0x00000000;
+ static constexpr ModelHint kFastestInference = 0x00000000;
// Can improve compilation time, but inference can be slower
- static const ModelHint kReduceKernelsCount = 0x00000001;
+ static constexpr ModelHint kReduceKernelsCount = 0x00000001;
// Can improve tuning time, but inference can be slower
- static const ModelHint kFastTuning = 0x00000002;
+ static constexpr ModelHint kFastTuning = 0x00000002;
void Add(ModelHint hint) {
if (hint == kFastestInference) {
diff --git a/tensorflow/lite/delegates/gpu/cl/selectors/operation_selector.cc b/tensorflow/lite/delegates/gpu/cl/selectors/operation_selector.cc
index 7d8b2bf..b6fb11b 100644
--- a/tensorflow/lite/delegates/gpu/cl/selectors/operation_selector.cc
+++ b/tensorflow/lite/delegates/gpu/cl/selectors/operation_selector.cc
@@ -56,10 +56,10 @@
bool IsSuitableForWinograd4x4To6x6(const Convolution2DAttributes& attr,
const CLDevice& device,
const BHWC& dst_shape) {
- const int tiles_x = IntegralDivideRoundUp(dst_shape.w, 4);
- const int tiles_y = IntegralDivideRoundUp(dst_shape.h, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int tiles_x = DivideRoundUp(dst_shape.w, 4);
+ const int tiles_y = DivideRoundUp(dst_shape.h, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
const bool suitable_attributes =
attr.weights.shape.w == 3 && attr.weights.shape.h == 3 &&
attr.dilations == HW(1, 1) && attr.strides == HW(1, 1);
@@ -82,8 +82,8 @@
return absl::UnimplementedError("No implementation for this case.");
}
- const int tiles_x = IntegralDivideRoundUp(output_shape.w, 4);
- const int tiles_y = IntegralDivideRoundUp(output_shape.h, 4);
+ const int tiles_x = DivideRoundUp(output_shape.w, 4);
+ const int tiles_y = DivideRoundUp(output_shape.h, 4);
const BHWC shape_0{input_shape.b, 36, tiles_x * tiles_y, input_shape.c};
const BHWC shape_1{input_shape.b, 36, tiles_x * tiles_y, output_shape.c};
TensorDescriptor td_0;
diff --git a/tensorflow/lite/delegates/gpu/cl/storage_type_util.cc b/tensorflow/lite/delegates/gpu/cl/storage_type_util.cc
index f6201fa..755da0c 100644
--- a/tensorflow/lite/delegates/gpu/cl/storage_type_util.cc
+++ b/tensorflow/lite/delegates/gpu/cl/storage_type_util.cc
@@ -28,7 +28,7 @@
bool CanCreateTensorWithShape(const CLContext& context, const CLDevice& device,
const BHWDC& shape,
const TensorDescriptor& descriptor) {
- const int slices = IntegralDivideRoundUp(shape.c, 4);
+ const int slices = DivideRoundUp(shape.c, 4);
switch (descriptor.storage_type) {
case TensorStorageType::BUFFER: {
const int flt4_size =
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor.cc b/tensorflow/lite/delegates/gpu/cl/tensor.cc
index 308e1b6..f01975e 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor.cc
+++ b/tensorflow/lite/delegates/gpu/cl/tensor.cc
@@ -65,11 +65,10 @@
}
if (descriptor.storage_type == TensorStorageType::IMAGE_BUFFER) {
cl_mem image_memory;
- RETURN_IF_ERROR(
- CreateImageBufferFromBuffer(context, memory, descriptor.data_type,
- shape.b * shape.w * shape.h * shape.d *
- IntegralDivideRoundUp(shape.c, 4),
- &image_memory));
+ RETURN_IF_ERROR(CreateImageBufferFromBuffer(
+ context, memory, descriptor.data_type,
+ shape.b * shape.w * shape.h * shape.d * DivideRoundUp(shape.c, 4),
+ &image_memory));
*result = Tensor(memory, memory_owner, image_memory, shape, descriptor);
} else {
*result = Tensor(memory, memory_owner, shape, descriptor);
@@ -386,7 +385,7 @@
const CLDevice& device, const BHWDC& shape,
const TensorDescriptor& descriptor,
CLMemory* result) {
- const int slices = IntegralDivideRoundUp(shape.c, 4);
+ const int slices = DivideRoundUp(shape.c, 4);
switch (descriptor.storage_type) {
case TensorStorageType::BUFFER:
case TensorStorageType::IMAGE_BUFFER: {
diff --git a/tensorflow/lite/delegates/gpu/cl/tensor.h b/tensorflow/lite/delegates/gpu/cl/tensor.h
index a27c54a..d59ef83 100644
--- a/tensorflow/lite/delegates/gpu/cl/tensor.h
+++ b/tensorflow/lite/delegates/gpu/cl/tensor.h
@@ -61,7 +61,7 @@
int Height() const { return shape_.h; }
int Depth() const { return shape_.d; }
int Channels() const { return shape_.c; }
- int Slices() const { return IntegralDivideRoundUp(shape_.c, 4); }
+ int Slices() const { return DivideRoundUp(shape_.c, 4); }
int Batch() const { return shape_.b; }
// returns int4(width * batch, height, slices, batch)
diff --git a/tensorflow/lite/delegates/gpu/common/BUILD b/tensorflow/lite/delegates/gpu/common/BUILD
index a24cc07..94d7918 100644
--- a/tensorflow/lite/delegates/gpu/common/BUILD
+++ b/tensorflow/lite/delegates/gpu/common/BUILD
@@ -115,9 +115,7 @@
":shape",
":status",
":tensor",
- "@com_google_absl//absl/memory",
"@com_google_absl//absl/strings",
- "@com_google_absl//absl/strings:str_format",
"//tensorflow/lite:context",
"//tensorflow/lite:kernel_api",
"//tensorflow/lite:util",
diff --git a/tensorflow/lite/delegates/gpu/common/convert.cc b/tensorflow/lite/delegates/gpu/common/convert.cc
index cee2e8f..fb0caf9 100644
--- a/tensorflow/lite/delegates/gpu/common/convert.cc
+++ b/tensorflow/lite/delegates/gpu/common/convert.cc
@@ -44,12 +44,11 @@
}
float* output = out.data();
- for (int p = 0; p < IntegralDivideRoundUp(shape.o, kPhwo4i4ChannelsInPlane);
- ++p) {
+ for (int p = 0; p < DivideRoundUp(shape.o, kPhwo4i4ChannelsInPlane); ++p) {
for (int h = 0; h < shape.h; ++h) {
for (int w = 0; w < shape.w; ++w) {
- for (int c = 0;
- c < IntegralDivideRoundUp(shape.i, kPhwo4i4ChannelsInPlane); ++c) {
+ for (int c = 0; c < DivideRoundUp(shape.i, kPhwo4i4ChannelsInPlane);
+ ++c) {
for (int co = 0; co < kPhwo4i4ChannelsInPlane; ++co) {
for (int ci = 0; ci < kPhwo4i4ChannelsInPlane; ++ci) {
float value = 0;
@@ -106,7 +105,7 @@
uint3 Get3DSizeForPHWO4I4(const OHWI& shape) {
return uint3(AlignByN(shape.i, 4), shape.h * shape.w,
- IntegralDivideRoundUp(shape.o, 4));
+ DivideRoundUp(shape.o, 4));
}
// Layout is Po,H,W,OI4x4.
@@ -123,8 +122,8 @@
out.size(), " != ", GetElementsSizeForPHWO4I4(shape)));
}
- const int dst_depth = IntegralDivideRoundUp(shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(shape.i, 4);
+ const int dst_depth = DivideRoundUp(shape.o, 4);
+ const int src_depth = DivideRoundUp(shape.i, 4);
float* output = out.data();
for (int f = 0; f < dst_depth; ++f) {
@@ -178,8 +177,7 @@
}
int32_t output_channels = shape.o * shape.i;
- int32_t num_planes =
- IntegralDivideRoundUp(output_channels, kPiohw4ChannelsInPlane);
+ int32_t num_planes = DivideRoundUp(output_channels, kPiohw4ChannelsInPlane);
float* output = out.data();
for (int p = 0; p < num_planes; ++p) {
for (int h = 0; h < shape.h; ++h) {
@@ -232,7 +230,7 @@
return absl::OkStatus();
}
// Layout is Pc,H,W,C4 where P - is a plane based on channels.
- int num_planes = IntegralDivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
+ int num_planes = DivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
const int num_pixels = shape.h * shape.w;
// A layer is a set of kPhwc4ChannelsInPlane channels images.
const int num_full_planes = shape.c / kPhwc4ChannelsInPlane;
@@ -281,7 +279,7 @@
RETURN_IF_ERROR(ValidateConvertToPHWC4(in, shape, out));
// Layout is Pc,H,W,C4 where P - is a plane based on channels.
- int num_planes = IntegralDivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
+ int num_planes = DivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
const int num_pixels = shape.h * shape.w;
// A layer is a set of kPhwc4ChannelsInPlane channels images.
const int num_full_planes = shape.c / kPhwc4ChannelsInPlane;
@@ -407,7 +405,7 @@
return absl::OkStatus();
}
- int num_planes = IntegralDivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
+ int num_planes = DivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
const int num_pixels = shape.h * shape.w;
const int padded_size = num_pixels * num_planes * kPhwc4ChannelsInPlane;
// A layer is a set of kPhwc4ChannelsInPlane channels images.
@@ -449,7 +447,7 @@
absl::Status ConvertFromPHWC4Half(absl::Span<const HalfBits> in,
const BHWC& shape, absl::Span<float> out) {
RETURN_IF_ERROR(ValidateConvertFromPHWC4(in, shape, out));
- int num_planes = IntegralDivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
+ int num_planes = DivideRoundUp(shape.c, kPhwc4ChannelsInPlane);
const int num_pixels = shape.h * shape.w;
const int padded_size = num_pixels * num_planes * kPhwc4ChannelsInPlane;
// A layer is a set of kPhwc4ChannelsInPlane channels images.
diff --git a/tensorflow/lite/delegates/gpu/common/model_builder.cc b/tensorflow/lite/delegates/gpu/common/model_builder.cc
index fc4026e..7c701a2 100644
--- a/tensorflow/lite/delegates/gpu/common/model_builder.cc
+++ b/tensorflow/lite/delegates/gpu/common/model_builder.cc
@@ -26,9 +26,7 @@
#include <utility>
#include <vector>
-#include "absl/memory/memory.h"
#include "absl/strings/str_cat.h"
-#include "absl/strings/str_format.h"
#include "absl/strings/str_join.h"
#include "absl/strings/string_view.h"
#include "tensorflow/lite/builtin_op_data.h"
@@ -107,8 +105,8 @@
// If tensor id is in range, it's guaranteed that it'll be available.
if (idx >= tflite_node->inputs->size) {
return absl::OutOfRangeError(
- absl::StrFormat("Requested index goes beyond array size (%d vs %d).",
- idx, tflite_node->inputs->data[idx]));
+ absl::StrCat("Requested index goes beyond array size: ", idx, " vs ",
+ idx, tflite_node->inputs->size));
}
return absl::OkStatus();
}
@@ -229,7 +227,6 @@
attr->weights.shape.o = weights.shape.h;
attr->weights.shape.i = weights.shape.w;
reader->ReadTensor(bias_tensor_id, &attr->bias).IgnoreError(); // optional
-
return absl::OkStatus();
}
@@ -262,8 +259,8 @@
const int op_version = registration->version;
if (op_version > max_version) {
return absl::UnimplementedError(
- absl::StrFormat("Max version supported: %d. Requested version %d.",
- max_version, op_version));
+ absl::StrCat("Max version supported: ", max_version,
+ ". Requested version ", op_version));
}
return absl::OkStatus();
}
@@ -273,36 +270,35 @@
int op_version = registration->version;
if (op_version != expected_version) {
return absl::UnimplementedError(
- absl::StrFormat("Only version %d is supported. Requested version %d.",
- expected_version, op_version));
+ absl::StrCat("Only version ", expected_version,
+ " is supported. Requested version ", op_version));
}
return absl::OkStatus();
}
absl::Status CheckKernels(int kernel_h, int kernel_w) {
if (kernel_h <= 0 || kernel_w <= 0) {
- return absl::InvalidArgumentError(absl::StrFormat(
- "Incorrect kernel values: kernel_height = %d, kernel_width = %d.",
- kernel_h, kernel_w));
+ return absl::InvalidArgumentError(
+ absl::StrCat("Incorrect kernel values: kernel_height = ", kernel_h,
+ ", kernel_width = ", kernel_w));
}
return absl::OkStatus();
}
absl::Status CheckStrides(int strides_h, int strides_w) {
if (strides_h <= 0 || strides_w <= 0) {
- return absl::InvalidArgumentError(absl::StrFormat(
- "Incorrect stride values: stride_height = %d, stride_width = %d.",
- strides_h, strides_w));
+ return absl::InvalidArgumentError(
+ absl::StrCat("Incorrect stride values: stride_height = ", strides_h,
+ ", stride_width = ", strides_w));
}
return absl::OkStatus();
}
absl::Status CheckDilation(int dilation_h, int dilation_w) {
if (dilation_h <= 0 || dilation_w <= 0) {
- return absl::InvalidArgumentError(
- absl::StrFormat("Incorrect dilation values: dilation_factor = %d, "
- "dilation_factor = %d.",
- dilation_h, dilation_w));
+ return absl::InvalidArgumentError(absl::StrCat(
+ "Incorrect dilation values: dilation_factor = ", dilation_h,
+ ", dilation_factor = ", dilation_w));
}
return absl::OkStatus();
}
@@ -382,7 +378,7 @@
constant_dims = input0->dims;
}
RETURN_IF_ERROR(reader->AddInput(node, runtime_tensor));
- if (constant_dims->size <= 0) {
+ if (constant_dims->size <= 0 || NumElements(constant_dims) == 1) {
Tensor<Scalar, DataType::FLOAT32> tensor;
RETURN_IF_ERROR(reader->ReadTensor(constant_tensor, &tensor));
*tensor_or_scalar = tensor.data[0];
@@ -2449,134 +2445,133 @@
const auto builtin_code = registration->builtin_code;
switch (builtin_code) {
case kTfLiteBuiltinAbs:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::ABS);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::ABS);
case kTfLiteBuiltinAdd:
- return absl::make_unique<AddOperationParser>();
+ return std::make_unique<AddOperationParser>();
case kTfLiteBuiltinAveragePool2d:
- return absl::make_unique<Pooling2DOperationParser>(PoolingType::AVERAGE);
+ return std::make_unique<Pooling2DOperationParser>(PoolingType::AVERAGE);
case kTfLiteBuiltinConcatenation:
- return absl::make_unique<ConcatenationOperationParser>();
+ return std::make_unique<ConcatenationOperationParser>();
case kTfLiteBuiltinConv2d:
- return absl::make_unique<Conv2DOperationParser>();
+ return std::make_unique<Conv2DOperationParser>();
case kTfLiteBuiltinCos:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::COS);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::COS);
case kTfLiteBuiltinDepthwiseConv2d:
- return absl::make_unique<DepthwiseConvolutionOperationParser>();
+ return std::make_unique<DepthwiseConvolutionOperationParser>();
case kTfLiteBuiltinDequantize:
if (allow_quant_ops) {
- return absl::make_unique<DequantizeOperationParser>();
+ return std::make_unique<DequantizeOperationParser>();
}
break;
case kTfLiteBuiltinDiv:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::DIV);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::DIV);
case kTfLiteBuiltinFullyConnected:
- return absl::make_unique<FullyConnectedOperationParser>();
+ return std::make_unique<FullyConnectedOperationParser>();
case kTfLiteBuiltinHardSwish:
- return absl::make_unique<HardSwishOperationParser>();
+ return std::make_unique<HardSwishOperationParser>();
case kTfLiteBuiltinLogistic:
- return absl::make_unique<ElementwiseOperationParser>(
+ return std::make_unique<ElementwiseOperationParser>(
OperationType::SIGMOID);
case kTfLiteBuiltinLog:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::LOG);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::LOG);
case kTfLiteBuiltinLstm:
- return absl::make_unique<LSTMOperationParser>();
+ return std::make_unique<LSTMOperationParser>();
case kTfLiteBuiltinMaximum:
- return absl::make_unique<ElementwiseOperationParser>(
+ return std::make_unique<ElementwiseOperationParser>(
OperationType::MAXIMUM);
case kTfLiteBuiltinMaxPool2d:
- return absl::make_unique<Pooling2DOperationParser>(PoolingType::MAX);
+ return std::make_unique<Pooling2DOperationParser>(PoolingType::MAX);
case kTfLiteBuiltinMean:
- return absl::make_unique<MeanOperationParser>();
+ return std::make_unique<MeanOperationParser>();
case kTfLiteBuiltinMinimum:
- return absl::make_unique<ElementwiseOperationParser>(
+ return std::make_unique<ElementwiseOperationParser>(
OperationType::MINIMUM);
case kTfLiteBuiltinMirrorPad:
- return absl::make_unique<PadOperationParser>(/*mirror_pad=*/true);
+ return std::make_unique<PadOperationParser>(/*mirror_pad=*/true);
case kTfLiteBuiltinMul:
- return absl::make_unique<MulOperationParser>();
+ return std::make_unique<MulOperationParser>();
case kTfLiteBuiltinPad:
- return absl::make_unique<PadOperationParser>(/*mirror_pad=*/false);
+ return std::make_unique<PadOperationParser>(/*mirror_pad=*/false);
case kTfLiteBuiltinPow:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::POW);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::POW);
case kTfLiteBuiltinQuantize:
if (allow_quant_ops) {
- return absl::make_unique<QuantizeOperationParser>();
+ return std::make_unique<QuantizeOperationParser>();
}
break;
case kTfLiteBuiltinRelu:
- return absl::make_unique<ReLUOperationParser>(0);
+ return std::make_unique<ReLUOperationParser>(0);
case kTfLiteBuiltinRelu6:
- return absl::make_unique<ReLUOperationParser>(6);
+ return std::make_unique<ReLUOperationParser>(6);
case kTfLiteBuiltinLeakyRelu:
- return absl::make_unique<ReLUOperationParser>(0);
+ return std::make_unique<ReLUOperationParser>(0);
case kTfLiteBuiltinPrelu:
- return absl::make_unique<PReLUOperationParser>();
+ return std::make_unique<PReLUOperationParser>();
case kTfLiteBuiltinReshape:
- return absl::make_unique<ReshapeOperationParser>();
+ return std::make_unique<ReshapeOperationParser>();
case kTfLiteBuiltinResizeBilinear:
- return absl::make_unique<Resize2DOperationParser>(SamplingType::BILINEAR);
+ return std::make_unique<Resize2DOperationParser>(SamplingType::BILINEAR);
case kTfLiteBuiltinResizeNearestNeighbor:
- return absl::make_unique<Resize2DOperationParser>(SamplingType::NEAREST);
+ return std::make_unique<Resize2DOperationParser>(SamplingType::NEAREST);
case kTfLiteBuiltinRsqrt:
- return absl::make_unique<ElementwiseOperationParser>(
- OperationType::RSQRT);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::RSQRT);
case kTfLiteBuiltinSin:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::SIN);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::SIN);
case kTfLiteBuiltinSlice:
- return absl::make_unique<SliceOperationParser>();
+ return std::make_unique<SliceOperationParser>();
case kTfLiteBuiltinSoftmax:
- return absl::make_unique<SoftmaxOperationParser>();
+ return std::make_unique<SoftmaxOperationParser>();
case kTfLiteBuiltinSpaceToDepth:
- return absl::make_unique<SpaceToDepthOperationParser>();
+ return std::make_unique<SpaceToDepthOperationParser>();
case kTfLiteBuiltinSqrt:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::SQRT);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::SQRT);
case kTfLiteBuiltinSquare:
- return absl::make_unique<ElementwiseOperationParser>(
+ return std::make_unique<ElementwiseOperationParser>(
OperationType::SQUARE);
case kTfLiteBuiltinSquaredDifference:
- return absl::make_unique<ElementwiseOperationParser>(
+ return std::make_unique<ElementwiseOperationParser>(
OperationType::SQUARED_DIFF);
case kTfLiteBuiltinStridedSlice:
- return absl::make_unique<StridedSliceOperationParser>();
+ return std::make_unique<StridedSliceOperationParser>();
case kTfLiteBuiltinSub:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::SUB);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::SUB);
case kTfLiteBuiltinTanh:
- return absl::make_unique<ElementwiseOperationParser>(OperationType::TANH);
+ return std::make_unique<ElementwiseOperationParser>(OperationType::TANH);
case kTfLiteBuiltinTranspose:
- return absl::make_unique<TransposeOperationParser>();
+ return std::make_unique<TransposeOperationParser>();
case kTfLiteBuiltinTransposeConv:
- return absl::make_unique<TransposeConvOperationParser>();
+ return std::make_unique<TransposeConvOperationParser>();
case kTfLiteBuiltinCustom:
const absl::string_view custom_name = registration->custom_name;
if (custom_name == "Convolution2DTransposeBias") {
- return absl::make_unique<Convolution2DTransposeBiasParser>();
+ return std::make_unique<Convolution2DTransposeBiasParser>();
}
if (custom_name == "MaxPoolingWithArgmax2D") {
- return absl::make_unique<Pooling2DOperationParser>(PoolingType::MAX);
+ return std::make_unique<Pooling2DOperationParser>(PoolingType::MAX);
}
if (custom_name == "MaxUnpooling2D") {
- return absl::make_unique<Unpooling2DOperationParser>();
+ return std::make_unique<Unpooling2DOperationParser>();
}
if (custom_name == "RoIToTransformMatrix") {
- return absl::make_unique<RoIToTransformMatrixOperationParser>();
+ return std::make_unique<RoIToTransformMatrixOperationParser>();
}
if (custom_name == "TransformTensor") {
- return absl::make_unique<TransformTensorOperationParser>();
+ return std::make_unique<TransformTensorOperationParser>();
}
if (custom_name == "TransformLandmarks") {
- return absl::make_unique<TransformLandmarksOperationParser>();
+ return std::make_unique<TransformLandmarksOperationParser>();
}
if (custom_name == "Landmarks2TransformMatrix") {
- return absl::make_unique<Landmarks2TransformMatrixOperationParser>();
+ return std::make_unique<Landmarks2TransformMatrixOperationParser>();
}
if (custom_name == "AlignmentPointsToTransformMatrix") {
- return absl::make_unique<
+ return std::make_unique<
AlignmentPointsToTransformMatrixOperationParser>();
}
break;
}
- return absl::make_unique<UnsupportedOperationParser>();
+ return std::make_unique<UnsupportedOperationParser>();
}
absl::Status IsSupported(const TfLiteContext* context, TfLiteNode* node,
@@ -2607,7 +2602,8 @@
// TODO(impjdi): Check number of input/output tensors and their dimensions.
// TODO(impjdi): Check ops' parameters.
-TfLiteIntArray* GetOpsToReplace(TfLiteContext* context, bool allow_quant_ops) {
+TfLiteIntArray* GetOpsToReplace(TfLiteContext* context, bool allow_quant_ops,
+ int max_delegated_partitions) {
delegates::IsNodeSupportedFn node_supported_fn =
[=](TfLiteContext* context, TfLiteNode* node,
TfLiteRegistration* registration,
@@ -2638,11 +2634,11 @@
return TfLiteIntArrayCreate(0);
}
- // We simply get 1st largest partition, but we could later explore whether
- // getting more partitions could lead to better performance, i.e. by
- // parameterizing '1' here.
+ // By default, we simply get 1st largest partition as 'max_delegate_partions'
+ // is set to 1 by default.
std::vector<int> ops_to_replace =
- partition_helper.GetNodesOfFirstNLargestPartitions(1);
+ partition_helper.GetNodesOfFirstNLargestPartitions(
+ max_delegated_partitions);
if (!unsupported_nodes_info.empty()) {
std::string unsupported = absl::StrJoin(unsupported_nodes_info, "\n");
@@ -2650,11 +2646,9 @@
"Following operations are not supported by GPU delegate:\n",
unsupported, "\n");
if (!ops_to_replace.empty()) {
- absl::StrAppendFormat(
- &error_message,
- "%d operations will run on the GPU (first node: "
- "%d, last node: %d), and the remaining %d",
- ops_to_replace.size(), ops_to_replace.front(), ops_to_replace.back(),
+ absl::StrAppend(
+ &error_message, ops_to_replace.size(),
+ " operations will run on the GPU, and the remaining ",
partition_helper.num_total_nodes() - ops_to_replace.size());
} else {
absl::StrAppend(&error_message,
diff --git a/tensorflow/lite/delegates/gpu/common/model_builder.h b/tensorflow/lite/delegates/gpu/common/model_builder.h
index 4b2a2f5..1e5016d 100644
--- a/tensorflow/lite/delegates/gpu/common/model_builder.h
+++ b/tensorflow/lite/delegates/gpu/common/model_builder.h
@@ -29,8 +29,12 @@
// Validates which operations are supported and returns array of operations to
// replace with GPU kernels. The caller must free the pointer on TfLiteIntArray.
+// 'max_delegated_partitions' limits the maximum number of partitions to
+// delegate as a graph could possibly have multiple partitions (each partition
+// consists of a subset of ops) to be replaced.
TfLiteIntArray* GetOpsToReplace(TfLiteContext* context,
- bool allow_quant_ops = false);
+ bool allow_quant_ops = false,
+ int max_delegated_partitions = 1);
// Extracts TFLite delegate execution plan from the input TFLite context and
// converts it into generic graph format.
diff --git a/tensorflow/lite/delegates/gpu/common/model_builder_test.cc b/tensorflow/lite/delegates/gpu/common/model_builder_test.cc
index 7b12f46..f0525e5 100644
--- a/tensorflow/lite/delegates/gpu/common/model_builder_test.cc
+++ b/tensorflow/lite/delegates/gpu/common/model_builder_test.cc
@@ -502,6 +502,187 @@
TfLiteIntArrayFree(ops_to_replace);
}
+class Interpreter2Fp32 : public DelegatedInterpreter {
+ public:
+ Interpreter2Fp32() : DelegatedInterpreter(4) {
+ void* builtin_data = malloc(sizeof(int));
+ EXPECT_EQ(interpreter_.AddTensors(8), kTfLiteOk);
+ EXPECT_EQ(interpreter_.SetInputs({0, 2, 4, 6}), kTfLiteOk);
+ EXPECT_EQ(interpreter_.SetOutputs({7}), kTfLiteOk);
+
+ // Add a Dequantize Node with uint8 input.
+ const TfLiteRegistration reg_dequant = {/*init=*/nullptr,
+ /*free=*/nullptr,
+ /*prepare=*/nullptr,
+ /*invoke=*/nullptr,
+ /*profiling_string=*/nullptr,
+ kTfLiteBuiltinDequantize};
+ EXPECT_EQ(interpreter_.AddNodeWithParameters(
+ /*inputs=*/{0}, /*outputs=*/{1}, /*init_data=*/nullptr,
+ /*init_data_size=*/0, /*builtin_data=*/nullptr,
+ /*registration=*/®_dequant),
+ kTfLiteOk);
+
+ // Add an ADD node that GPU delegate can parse.
+ const TfLiteRegistration reg_add0 = {
+ [](TfLiteContext* context, const char* buffer, size_t length) {
+ return reinterpret_cast<void*>(new int(1));
+ },
+ [](TfLiteContext* context, void* buffer) {
+ delete reinterpret_cast<int*>(buffer);
+ },
+ nullptr,
+ nullptr,
+ nullptr,
+ kTfLiteBuiltinAdd};
+ EXPECT_EQ(interpreter_.AddNodeWithParameters(
+ /*inputs=*/{1, 2}, /*outputs=*/{3}, /*init_data=*/nullptr,
+ /*init_data_size=*/0,
+ /*builtin_data=*/builtin_data,
+ /*registration=*/®_add0),
+ kTfLiteOk);
+
+ // Add a Pack Node that GPU delegate doesn't support
+ const TfLiteRegistration reg_pack = {/*init=*/nullptr,
+ /*free=*/nullptr,
+ /*prepare=*/nullptr,
+ /*invoke=*/nullptr,
+ /*profiling_string=*/nullptr,
+ kTfLiteBuiltinPack};
+ EXPECT_EQ(interpreter_.AddNodeWithParameters(
+ /*inputs=*/{3, 4}, /*outputs=*/{5}, /*init_data=*/nullptr,
+ /*init_data_size=*/0, /*builtin_data=*/nullptr,
+ /*registration=*/®_pack),
+ kTfLiteOk);
+
+ const TfLiteRegistration reg_add1 = {
+ [](TfLiteContext* context, const char* buffer, size_t length) {
+ return reinterpret_cast<void*>(new int[2]);
+ },
+ [](TfLiteContext* context, void* buffer) {
+ delete reinterpret_cast<int*>(buffer);
+ },
+ nullptr,
+ nullptr,
+ nullptr,
+ kTfLiteBuiltinAdd};
+ EXPECT_EQ(interpreter_.AddNodeWithParameters(
+ /*inputs=*/{5, 6}, /*outputs=*/{7}, /*init_data=*/nullptr,
+ /*init_data_size=*/0,
+ /*builtin_data=*/builtin_data,
+ /*registration=*/®_add1),
+ kTfLiteOk);
+
+ std::vector<int> dims = {1};
+ TfLiteQuantization quantization;
+ quantization.type = kTfLiteNoQuantization;
+ EXPECT_EQ(interpreter_.SetTensorParametersReadWrite(
+ 0, TfLiteType::kTfLiteUInt8, "t0", dims, quantization, false),
+ kTfLiteOk);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 1, TfLiteType::kTfLiteFloat32, "t1", dims, quantization, false),
+ kTfLiteOk);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 2, TfLiteType::kTfLiteFloat32, "t2", dims, quantization, false),
+ kTfLiteOk);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 3, TfLiteType::kTfLiteFloat32, "t3", dims, quantization, false),
+ kTfLiteOk);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 4, TfLiteType::kTfLiteFloat32, "t4", dims, quantization, false),
+ kTfLiteOk);
+
+ dims.push_back(2);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 5, TfLiteType::kTfLiteFloat32, "t5", dims, quantization, false),
+ kTfLiteOk);
+ EXPECT_EQ(
+ interpreter_.SetTensorParametersReadWrite(
+ 6, TfLiteType::kTfLiteFloat32, "t6", dims, quantization, false),
+ kTfLiteOk);
+
+ exec_plan()->data[0] = 0;
+ exec_plan()->data[1] = 1;
+ exec_plan()->data[2] = 2;
+ exec_plan()->data[3] = 3;
+ }
+};
+
+Interpreter2Fp32* interpreter2_fp32 = new Interpreter2Fp32();
+
+TEST(ModelBuilderTest, GetOpsToReplaceMultiplePartitions) {
+ // A graph with a Dequant node with uint8 input, a Pack node are not pruned.
+ // As these ops are currently not supported on the GPU, they will be scheduled
+ // to run on the CPU while the remaining supported op Add on the GPU.
+ //
+ // t0 (uint8) -> Dequant(0) -> t1 (FP32) -> Add(1) -> t3 (FP32) -> PACK (2)
+ // t2 (FP32) -/ t4 (FP32) -/
+ // PACK (2) -> t5 (FP32) -> Add(3) -> t7
+ // -> t6 (FP32) -/
+ //
+ TfLiteContext* context = interpreter2_fp32->context();
+
+ // These functions are meant to be called inside delegates. Swap out
+ // for similar functions to permit direct calling of GetOpsToReplace.
+ context->GetExecutionPlan = [](struct TfLiteContext* context,
+ TfLiteIntArray** execution_plan) {
+ *execution_plan = interpreter2_fp32->exec_plan();
+ return kTfLiteOk;
+ };
+ context->GetNodeAndRegistration = [](struct TfLiteContext*, int node_index,
+ TfLiteNode** node,
+ TfLiteRegistration** registration) {
+ auto& node_and_reg =
+ interpreter2_fp32->nodes_and_registration()[node_index];
+ *node = &node_and_reg.first;
+ *registration = &node_and_reg.second;
+ return kTfLiteOk;
+ };
+ context->PreviewDelegatePartitioning =
+ [](struct TfLiteContext* context, const TfLiteIntArray* nodes_to_replace,
+ TfLiteDelegateParams** partition_params_array, int* num_partitions) {
+ auto params = interpreter2_fp32->add_delegate_params();
+ params->nodes_to_replace = TfLiteIntArrayCreate(1);
+ params->nodes_to_replace->data[0] = 1;
+ params->input_tensors = TfLiteIntArrayCreate(2);
+ params->input_tensors->data[0] = 1;
+ params->input_tensors->data[1] = 2;
+ params->output_tensors = TfLiteIntArrayCreate(1);
+ params->output_tensors->data[0] = 3;
+
+ params = interpreter2_fp32->add_delegate_params();
+ params->nodes_to_replace = TfLiteIntArrayCreate(1);
+ params->nodes_to_replace->data[0] = 3;
+ params->input_tensors = TfLiteIntArrayCreate(2);
+ params->input_tensors->data[0] = 5;
+ params->input_tensors->data[1] = 6;
+ params->output_tensors = TfLiteIntArrayCreate(1);
+ params->output_tensors->data[0] = 7;
+
+ *partition_params_array = interpreter2_fp32->delegate_params();
+ *num_partitions = interpreter2_fp32->num_delegate_params();
+ return kTfLiteOk;
+ };
+
+ TfLiteIntArray* ops_to_replace = GetOpsToReplace(
+ context, /*allow_quant_ops=*/false, /*max_delegated_partitions*/ 2);
+
+ // As the Dequant op is not pruned and the ADD op could run on GPU, we have
+ // 2 partitions.
+ EXPECT_EQ(ops_to_replace->size, 2);
+ // ADD at index 1.
+ EXPECT_EQ(1, ops_to_replace->data[0]);
+ // ADD at index 3.
+ EXPECT_EQ(3, ops_to_replace->data[1]);
+
+ TfLiteIntArrayFree(ops_to_replace);
+}
+
class InterpreterMultiNode : public DelegatedInterpreter {
public:
explicit InterpreterMultiNode(bool add_op_first = true)
diff --git a/tensorflow/lite/delegates/gpu/common/operations.cc b/tensorflow/lite/delegates/gpu/common/operations.cc
index 771ed73..28ce67b 100644
--- a/tensorflow/lite/delegates/gpu/common/operations.cc
+++ b/tensorflow/lite/delegates/gpu/common/operations.cc
@@ -209,7 +209,7 @@
namespace {
template <typename T>
-T IntegralDivideRoundUp(T n, T divisor) {
+T DivideRoundUp(T n, T divisor) {
return (n - 1) / divisor + 1;
}
@@ -272,7 +272,7 @@
}
inline int32_t StridedSize(int32_t size, int32_t stride) {
- return stride == 0 ? -1 : IntegralDivideRoundUp(size, stride);
+ return stride == 0 ? -1 : DivideRoundUp(size, stride);
}
template <Axis AxisT, typename AttrT>
diff --git a/tensorflow/lite/delegates/gpu/common/util.h b/tensorflow/lite/delegates/gpu/common/util.h
index 929d00b..6a1e793 100644
--- a/tensorflow/lite/delegates/gpu/common/util.h
+++ b/tensorflow/lite/delegates/gpu/common/util.h
@@ -24,24 +24,23 @@
// @param n must be non negative
// @param divisor must be greater than zero
template <typename T, typename N>
-T IntegralDivideRoundUp(T n, N divisor) {
+T DivideRoundUp(T n, N divisor) {
const T div = static_cast<T>(divisor);
const T q = n / div;
return n % div == 0 ? q : q + 1;
}
template <>
-inline uint3 IntegralDivideRoundUp(uint3 n, uint3 divisor) {
- return uint3(IntegralDivideRoundUp(n.x, divisor.x),
- IntegralDivideRoundUp(n.y, divisor.y),
- IntegralDivideRoundUp(n.z, divisor.z));
+inline uint3 DivideRoundUp(uint3 n, uint3 divisor) {
+ return uint3(DivideRoundUp(n.x, divisor.x), DivideRoundUp(n.y, divisor.y),
+ DivideRoundUp(n.z, divisor.z));
}
// @param number or its components must be greater than zero
// @param n must be greater than zero
template <typename T, typename N>
T AlignByN(T number, N n) {
- return IntegralDivideRoundUp(number, n) * n;
+ return DivideRoundUp(number, n) * n;
}
} // namespace gpu
diff --git a/tensorflow/lite/delegates/gpu/common/util_test.cc b/tensorflow/lite/delegates/gpu/common/util_test.cc
index 7c8cb81..ce170f4 100644
--- a/tensorflow/lite/delegates/gpu/common/util_test.cc
+++ b/tensorflow/lite/delegates/gpu/common/util_test.cc
@@ -24,16 +24,16 @@
using testing::Eq;
-TEST(UtilTest, IntegralDivideRoundUp) {
- EXPECT_THAT(IntegralDivideRoundUp(0, 256), Eq(0));
- EXPECT_THAT(IntegralDivideRoundUp(2u, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(2, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(255u, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(255, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(256u, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(256, 256), Eq(1));
- EXPECT_THAT(IntegralDivideRoundUp(257u, 256), Eq(2));
- EXPECT_THAT(IntegralDivideRoundUp(257, 256), Eq(2));
+TEST(UtilTest, DivideRoundUp) {
+ EXPECT_THAT(DivideRoundUp(0, 256), Eq(0));
+ EXPECT_THAT(DivideRoundUp(2u, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(2, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(255u, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(255, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(256u, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(256, 256), Eq(1));
+ EXPECT_THAT(DivideRoundUp(257u, 256), Eq(2));
+ EXPECT_THAT(DivideRoundUp(257, 256), Eq(2));
}
TEST(UtilTest, AlignByN) {
diff --git a/tensorflow/lite/delegates/gpu/common/workgroup_selection.cc b/tensorflow/lite/delegates/gpu/common/workgroup_selection.cc
index d18e372..3abab71 100644
--- a/tensorflow/lite/delegates/gpu/common/workgroup_selection.cc
+++ b/tensorflow/lite/delegates/gpu/common/workgroup_selection.cc
@@ -34,9 +34,9 @@
for (int x = 1; x <= 4; ++x) {
for (int y = 1; y <= 4; ++y) {
for (int z = 1; z <= 4; ++z) {
- int wg_x = IntegralDivideRoundUp(grid.x, x);
- int wg_y = IntegralDivideRoundUp(grid.y, y);
- int wg_z = IntegralDivideRoundUp(grid.z, z);
+ int wg_x = DivideRoundUp(grid.x, x);
+ int wg_y = DivideRoundUp(grid.y, y);
+ int wg_z = DivideRoundUp(grid.z, z);
if (wg_x > max_work_group_sizes.x || wg_y > max_work_group_sizes.y ||
wg_z > max_work_group_sizes.z ||
wg_x * wg_y * wg_z > max_work_group_total_size) {
diff --git a/tensorflow/lite/delegates/gpu/delegate.cc b/tensorflow/lite/delegates/gpu/delegate.cc
index 540c8ba..44d899b 100644
--- a/tensorflow/lite/delegates/gpu/delegate.cc
+++ b/tensorflow/lite/delegates/gpu/delegate.cc
@@ -70,17 +70,25 @@
class Delegate {
public:
- explicit Delegate(const TfLiteGpuDelegateOptionsV2* options) {
+ explicit Delegate(const TfLiteGpuDelegateOptionsV2* options)
+ : num_delegate_kernels_(0) {
options_ = options ? *options : TfLiteGpuDelegateOptionsV2Default();
+ if (options_.max_delegated_partitions <= 0) {
+ options_.max_delegated_partitions = 1;
+ }
}
TfLiteDelegate* tflite_delegate() { return &delegate_; }
const TfLiteGpuDelegateOptionsV2& options() const { return options_; }
- bool IsQuantOpsAllowed() {
+ bool IsQuantOpsAllowed() const {
return options_.experimental_flags &
TFLITE_GPU_EXPERIMENTAL_FLAGS_ENABLE_QUANT;
}
+ int MaxDelegatedPartitions() const {
+ return options_.max_delegated_partitions;
+ }
+ int num_delegate_kernels() const { return num_delegate_kernels_; }
private:
TfLiteDelegate delegate_ = {
@@ -93,13 +101,18 @@
};
TfLiteGpuDelegateOptionsV2 options_;
+ int num_delegate_kernels_ = 0;
+
+ friend class DelegateKernel;
};
// Represent the execution of a subset of nodes on GPU.
class DelegateKernel {
public:
- explicit DelegateKernel(const TfLiteGpuDelegateOptionsV2& options)
- : options_(options) {}
+ explicit DelegateKernel(Delegate* delegate) : delegate_(delegate) {
+ ++delegate_->num_delegate_kernels_;
+ }
+ ~DelegateKernel() { --delegate_->num_delegate_kernels_; }
absl::Status Prepare(TfLiteContext* context,
const TfLiteDelegateParams* delegate_params) {
@@ -108,44 +121,47 @@
// Extract TFLite delegate execution plan from the context and convert it
// into GraphFloat32.
GraphFloat32 graph;
- std::vector<uint32_t> input_refs;
- std::vector<uint32_t> output_refs;
- RETURN_IF_ERROR(InitializeGraph(context, delegate_params, &graph,
- &input_refs, &output_refs));
+ RETURN_IF_ERROR(InitializeGraph(context, delegate_params, &graph));
std::unique_ptr<InferenceBuilder> builder;
bool graph_is_destroyed;
- absl::Status status =
- InitializeOpenClApi(&graph, &builder, &graph_is_destroyed);
- if (!status.ok()) {
- TF_LITE_KERNEL_LOG(context, std::string(status.message()).c_str());
- context->ReportError(context, "Falling back to OpenGL");
-
- // Graph need to be re-created because it is moved above.
- GraphFloat32 graph2;
- if (graph_is_destroyed) {
- RETURN_IF_ERROR(InitializeGraph(context, delegate_params, &graph2,
- &input_refs, &output_refs));
- }
+ const int experimental_flags = delegate_->options().experimental_flags;
+ if (experimental_flags & TFLITE_GPU_EXPERIMENTAL_FLAGS_CL_ONLY) {
RETURN_IF_ERROR(
- InitializeOpenGlApi(graph_is_destroyed ? &graph2 : &graph, &builder));
+ InitializeOpenClApi(&graph, &builder, &graph_is_destroyed));
+ } else if (experimental_flags & TFLITE_GPU_EXPERIMENTAL_FLAGS_GL_ONLY) {
+ RETURN_IF_ERROR(InitializeOpenGlApi(&graph, &builder));
+ } else {
+ // By default, we try CL first & fall back to GL if that fails.
+ absl::Status status =
+ InitializeOpenClApi(&graph, &builder, &graph_is_destroyed);
+ if (!status.ok()) {
+ TF_LITE_KERNEL_LOG(context, std::string(status.message()).c_str());
+ TF_LITE_KERNEL_LOG(context, "Falling back to OpenGL");
+
+ // Graph needs to be re-created because it is moved above.
+ GraphFloat32 graph2;
+ if (graph_is_destroyed) {
+ RETURN_IF_ERROR(InitializeGraph(context, delegate_params, &graph2));
+ }
+ RETURN_IF_ERROR(InitializeOpenGlApi(
+ graph_is_destroyed ? &graph2 : &graph, &builder));
+ }
}
// At this point tflite didn't allocate tensors yet, therefore, collect
// indices and set all input and output tensors from tflite later.
- input_indices_.reserve(input_refs.size());
- for (uint32_t tensor_index : input_refs) {
- const int64_t object_index = input_indices_.size();
- input_indices_.push_back(tensor_index);
- RETURN_IF_ERROR(
- builder->SetInputObjectDef(object_index, GetObjectDef(tensor_index)));
+ input_indices_.resize(graph.inputs().size());
+ for (int i = 0; i < input_indices_.size(); ++i) {
+ const int64_t tflite_tensor_id = graph.inputs()[i]->tensor.ref;
+ input_indices_.push_back(tflite_tensor_id);
+ RETURN_IF_ERROR(builder->SetInputObjectDef(i, GetObjectDef()));
}
- output_indices_.reserve(output_refs.size());
- for (uint32_t tensor_index : output_refs) {
- const int64_t object_index = output_indices_.size();
- output_indices_.push_back(tensor_index);
- RETURN_IF_ERROR(builder->SetOutputObjectDef(object_index,
- GetObjectDef(tensor_index)));
+ output_indices_.resize(graph.outputs().size());
+ for (int i = 0; i < output_indices_.size(); ++i) {
+ const int64_t tflite_tensor_id = graph.outputs()[i]->tensor.ref;
+ output_indices_.push_back(tflite_tensor_id);
+ RETURN_IF_ERROR(builder->SetOutputObjectDef(i, GetObjectDef()));
}
return builder->Build(&runner_);
@@ -211,7 +227,7 @@
return absl::OkStatus();
}
- ObjectDef GetObjectDef(int index) const {
+ ObjectDef GetObjectDef() const {
ObjectDef default_object_def;
default_object_def.data_type = DataType::FLOAT32;
default_object_def.data_layout = DataLayout::BHWC;
@@ -228,31 +244,14 @@
private:
absl::Status InitializeGraph(TfLiteContext* context,
const TfLiteDelegateParams* delegate_params,
- GraphFloat32* graph,
- std::vector<uint32_t>* input_refs,
- std::vector<uint32_t>* output_refs) {
+ GraphFloat32* graph) {
quant_conversion_map_.clear();
- if (options_.experimental_flags &
- TFLITE_GPU_EXPERIMENTAL_FLAGS_ENABLE_QUANT) {
+ if (delegate_->IsQuantOpsAllowed()) {
RETURN_IF_ERROR(BuildFinalModel(context, delegate_params, graph,
&quant_conversion_map_));
} else {
RETURN_IF_ERROR(BuildFinalModel(context, delegate_params, graph));
}
-
- input_refs->clear();
- output_refs->clear();
- const auto& inputs = graph->inputs();
- input_refs->reserve(inputs.size());
- for (const auto& input : inputs) {
- input_refs->push_back(input->tensor.ref);
- }
- const auto& outputs = graph->outputs();
- output_refs->reserve(outputs.size());
- for (const auto& output : outputs) {
- output_refs->push_back(output->tensor.ref);
- }
-
return absl::OkStatus();
}
@@ -328,22 +327,23 @@
cl::InferenceEnvironmentProperties properties;
RETURN_IF_ERROR(cl::NewInferenceEnvironment(env_options, &cl_environment_,
&properties));
+ auto delegate_options = delegate_->options();
cl::InferenceOptions options;
// If is_precision_loss_allowed == -1, then just use priorities instead
// of paying attention to is_precision_loss_allowed value.
- if (options_.is_precision_loss_allowed == -1) {
- options.priority1 = ToPriority(options_.inference_priority1);
- options.priority2 = ToPriority(options_.inference_priority2);
- options.priority3 = ToPriority(options_.inference_priority3);
+ if (delegate_options.is_precision_loss_allowed == -1) {
+ options.priority1 = ToPriority(delegate_options.inference_priority1);
+ options.priority2 = ToPriority(delegate_options.inference_priority2);
+ options.priority3 = ToPriority(delegate_options.inference_priority3);
} else {
// Users set is_precision_loss_allowed explicitly, thus use it explicitly.
- if (options_.is_precision_loss_allowed == 0) {
+ if (delegate_options.is_precision_loss_allowed == 0) {
options.priority1 = InferencePriority::MAX_PRECISION;
} else {
options.priority1 = InferencePriority::MIN_LATENCY;
}
}
- options.usage = ToUsage(options_.inference_preference);
+ options.usage = ToUsage(delegate_options.inference_preference);
*graph_is_destroyed = true;
RETURN_IF_ERROR(cl_environment_->NewInferenceBuilder(
options, std::move(*graph), builder));
@@ -358,11 +358,12 @@
gl::InferenceEnvironmentProperties properties;
RETURN_IF_ERROR(
NewInferenceEnvironment(env_options, &gl_environment_, &properties));
+ auto delegate_options = delegate_->options();
gl::InferenceOptions options;
- options.usage = ToUsage(options_.inference_preference);
- options.priority1 = ToPriority(options_.inference_priority1);
- options.priority2 = ToPriority(options_.inference_priority2);
- options.priority3 = ToPriority(options_.inference_priority3);
+ options.usage = ToUsage(delegate_options.inference_preference);
+ options.priority1 = ToPriority(delegate_options.inference_priority1);
+ options.priority2 = ToPriority(delegate_options.inference_priority2);
+ options.priority3 = ToPriority(delegate_options.inference_priority3);
RETURN_IF_ERROR(gl_environment_->NewInferenceBuilder(std::move(*graph),
options, builder));
enforce_same_thread_ = true;
@@ -371,9 +372,8 @@
return absl::OkStatus();
}
- // Shared across all DelegateKernel instances, passed by the Delegate
- // instance.
- const TfLiteGpuDelegateOptionsV2& options_;
+ // The Delegate instance that's shared across all DelegateKernel instances.
+ Delegate* const delegate_; // doesn't own the memory.
std::unique_ptr<cl::InferenceEnvironment> cl_environment_;
std::unique_ptr<gl::InferenceEnvironment> gl_environment_;
std::unique_ptr<InferenceRunner> runner_;
@@ -405,7 +405,7 @@
// Everything below should happen in prepare function call, but TFLite
// for whatever reason forbids that.
auto gpu_delegate_kernel =
- absl::make_unique<DelegateKernel>(gpu_delegate->options());
+ absl::make_unique<DelegateKernel>(gpu_delegate);
const auto status = gpu_delegate_kernel->Prepare(context, params);
if (!status.ok()) {
context->ReportError(context, "TfLiteGpuDelegate Init: %s",
@@ -454,10 +454,15 @@
"TfLiteGpuDelegateV2", // .custom_name
1, // .version
};
- TfLiteIntArray* ops_to_replace = GetOpsToReplace(
- context, /*allow_quant_ops=*/GetDelegate(delegate)->IsQuantOpsAllowed());
+
+ auto* gpu_delegate = GetDelegate(delegate);
+ TfLiteIntArray* ops_to_replace =
+ GetOpsToReplace(context, gpu_delegate->IsQuantOpsAllowed(),
+ gpu_delegate->MaxDelegatedPartitions());
const auto status = context->ReplaceNodeSubsetsWithDelegateKernels(
context, kRegistration, ops_to_replace, delegate);
+ TFLITE_LOG_PROD(TFLITE_LOG_INFO, "Created %d GPU delegate kernels.",
+ gpu_delegate->num_delegate_kernels());
TfLiteIntArrayFree(ops_to_replace);
return status;
}
@@ -467,15 +472,17 @@
} // namespace tflite
TfLiteGpuDelegateOptionsV2 TfLiteGpuDelegateOptionsV2Default() {
- TfLiteGpuDelegateOptionsV2 options;
- // set it to -1 to detect whether it was later adjusted.
- options.is_precision_loss_allowed = -1;
- options.inference_preference =
- TFLITE_GPU_INFERENCE_PREFERENCE_FAST_SINGLE_ANSWER;
- options.inference_priority1 = TFLITE_GPU_INFERENCE_PRIORITY_MAX_PRECISION;
- options.inference_priority2 = TFLITE_GPU_INFERENCE_PRIORITY_AUTO;
- options.inference_priority3 = TFLITE_GPU_INFERENCE_PRIORITY_AUTO;
- options.experimental_flags = TFLITE_GPU_EXPERIMENTAL_FLAGS_NONE;
+ TfLiteGpuDelegateOptionsV2 options = {
+ // set it to -1 to detect whether it was later adjusted.
+ .is_precision_loss_allowed = -1,
+ .inference_preference =
+ TFLITE_GPU_INFERENCE_PREFERENCE_FAST_SINGLE_ANSWER,
+ .inference_priority1 = TFLITE_GPU_INFERENCE_PRIORITY_MAX_PRECISION,
+ .inference_priority2 = TFLITE_GPU_INFERENCE_PRIORITY_AUTO,
+ .inference_priority3 = TFLITE_GPU_INFERENCE_PRIORITY_AUTO,
+ .experimental_flags = TFLITE_GPU_EXPERIMENTAL_FLAGS_NONE,
+ .max_delegated_partitions = 1,
+ };
return options;
}
diff --git a/tensorflow/lite/delegates/gpu/delegate.h b/tensorflow/lite/delegates/gpu/delegate.h
index a60ebec..f03392d 100644
--- a/tensorflow/lite/delegates/gpu/delegate.h
+++ b/tensorflow/lite/delegates/gpu/delegate.h
@@ -65,7 +65,10 @@
enum TfLiteGpuExperimentalFlags {
TFLITE_GPU_EXPERIMENTAL_FLAGS_NONE = 0,
// Enables inference on quantized models with the delegate.
- TFLITE_GPU_EXPERIMENTAL_FLAGS_ENABLE_QUANT = 1 << 0
+ TFLITE_GPU_EXPERIMENTAL_FLAGS_ENABLE_QUANT = 1 << 0,
+ // Enforces execution with the provided backend.
+ TFLITE_GPU_EXPERIMENTAL_FLAGS_CL_ONLY = 1 << 1,
+ TFLITE_GPU_EXPERIMENTAL_FLAGS_GL_ONLY = 1 << 2
};
// IMPORTANT: Always use TfLiteGpuDelegateOptionsV2Default() method to create
@@ -106,6 +109,11 @@
// Bitmask flags. See the comments in TfLiteGpuExperimentalFlags.
int64_t experimental_flags;
+
+ // A graph could have multiple partitions that can be delegated to the GPU.
+ // This limits the maximum number of partitions to be delegated. By default,
+ // it's set to 1 in TfLiteGpuDelegateOptionsV2Default().
+ int32_t max_delegated_partitions;
} TfLiteGpuDelegateOptionsV2;
// Populates TfLiteGpuDelegateOptionsV2 as follows:
diff --git a/tensorflow/lite/delegates/gpu/gl/api.cc b/tensorflow/lite/delegates/gpu/gl/api.cc
index ae050bf..0240a5c 100644
--- a/tensorflow/lite/delegates/gpu/gl/api.cc
+++ b/tensorflow/lite/delegates/gpu/gl/api.cc
@@ -201,7 +201,7 @@
ShaderCode code) {
// Calculate workgroup size.
uint3 workgroup_size = workgroup_calculator.Calculate(code);
- uint3 num_workgroups = IntegralDivideRoundUp(code.workload, workgroup_size);
+ uint3 num_workgroups = DivideRoundUp(code.workload, workgroup_size);
for (const auto& object : code.objects) {
if (IsRef(object)) {
diff --git a/tensorflow/lite/delegates/gpu/gl/api2.cc b/tensorflow/lite/delegates/gpu/gl/api2.cc
index 4c9894f..c8bf6dd 100644
--- a/tensorflow/lite/delegates/gpu/gl/api2.cc
+++ b/tensorflow/lite/delegates/gpu/gl/api2.cc
@@ -569,7 +569,7 @@
} else {
shader_index = it->second;
}
- auto num_workgroups = IntegralDivideRoundUp(code.workload, workgroup);
+ auto num_workgroups = DivideRoundUp(code.workload, workgroup);
return runtime_ptr->AddProgram(shaders[shader_index], code.parameters,
code.objects, num_workgroups);
}));
diff --git a/tensorflow/lite/delegates/gpu/gl/compiler.cc b/tensorflow/lite/delegates/gpu/gl/compiler.cc
index 4ae9b10..d316505 100644
--- a/tensorflow/lite/delegates/gpu/gl/compiler.cc
+++ b/tensorflow/lite/delegates/gpu/gl/compiler.cc
@@ -190,8 +190,7 @@
"Workload uint3() requires all output sizes to match");
}
}
- attr.code.workload =
- uint3(shape.w, shape.h, IntegralDivideRoundUp(shape.c, 4));
+ attr.code.workload = uint3(shape.w, shape.h, DivideRoundUp(shape.c, 4));
}
int num_textures = 0;
diff --git a/tensorflow/lite/delegates/gpu/gl/converters/bhwc_to_phwc4.cc b/tensorflow/lite/delegates/gpu/gl/converters/bhwc_to_phwc4.cc
index fc86b0f..67094d2 100644
--- a/tensorflow/lite/delegates/gpu/gl/converters/bhwc_to_phwc4.cc
+++ b/tensorflow/lite/delegates/gpu/gl/converters/bhwc_to_phwc4.cc
@@ -88,8 +88,8 @@
return absl::UnimplementedError(
"BhwcToPhwc4: Batch size is not equal to 1.");
}
- uint3 workload = uint3(shape.w, shape.h, IntegralDivideRoundUp(shape.c, 4));
- uint3 num_workgroups = IntegralDivideRoundUp(workload, workgroup_size_);
+ uint3 workload = uint3(shape.w, shape.h, DivideRoundUp(shape.c, 4));
+ uint3 num_workgroups = DivideRoundUp(workload, workgroup_size_);
RETURN_IF_ERROR(program_.SetParameter(
{"sizes_",
diff --git a/tensorflow/lite/delegates/gpu/gl/converters/phwc4_to_bhwc.cc b/tensorflow/lite/delegates/gpu/gl/converters/phwc4_to_bhwc.cc
index 5a9f51c..15f859c 100644
--- a/tensorflow/lite/delegates/gpu/gl/converters/phwc4_to_bhwc.cc
+++ b/tensorflow/lite/delegates/gpu/gl/converters/phwc4_to_bhwc.cc
@@ -83,7 +83,7 @@
}
uint3 workload = uint3(shape.w, shape.h, shape.c);
- uint3 num_workgroups = IntegralDivideRoundUp(workload, workgroup_size_);
+ uint3 num_workgroups = DivideRoundUp(workload, workgroup_size_);
// TODO(akulik): simply pass workload as soon as UniformParameter
// supports uint3
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/add.cc b/tensorflow/lite/delegates/gpu/gl/kernels/add.cc
index 01128c2..a07e97e 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/add.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/add.cc
@@ -102,7 +102,7 @@
// Declare workload explicitly because shader depends on gid.z.
/*workload=*/
uint3(ctx.input_shapes[0][2], ctx.input_shapes[0][1],
- IntegralDivideRoundUp(ctx.input_shapes[0][3], 4)),
+ DivideRoundUp(ctx.input_shapes[0][3], 4)),
/*workgroup=*/uint3(),
/*source_code=*/"value_0 += $add_buffer[gid.z]$;",
/*input=*/IOStructure::AUTO,
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/concat.cc b/tensorflow/lite/delegates/gpu/gl/kernels/concat.cc
index c3fd62d..c368f75 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/concat.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/concat.cc
@@ -210,7 +210,7 @@
// * - you are going to write into these cells
// @ - you will fill these cells next cycles
// ^ - first elem you start writing from
- int blocks_amount = IntegralDivideRoundUp<int>(in_ch, 4);
+ int blocks_amount = DivideRoundUp<int>(in_ch, 4);
code += "// Aligned case\n";
code += "// I'm going to make " + std::to_string(blocks_amount) +
" write(s)\n\n";
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/conv.cc b/tensorflow/lite/delegates/gpu/gl/kernels/conv.cc
index 3dceb9f..7bb1266 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/conv.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/conv.cc
@@ -55,7 +55,7 @@
{"dilation_h", attr.dilations.h},
{"kernel_w", weights.w},
{"kernel_h", weights.h},
- {"src_depth", IntegralDivideRoundUp(weights.i, 4)},
+ {"src_depth", DivideRoundUp(weights.i, 4)},
{"stride", int2(attr.strides.w, attr.strides.h)},
};
} else {
@@ -71,7 +71,7 @@
{"input_data_0_w", static_cast<int>(ctx.input_shapes[0][2])},
{"offsets_count", offsets_count},
{"offsets", offsets},
- {"src_depth", IntegralDivideRoundUp(weights.i, 4)},
+ {"src_depth", DivideRoundUp(weights.i, 4)},
{"stride", int2(attr.strides.w, attr.strides.h)},
};
}
@@ -181,14 +181,14 @@
std::vector<Variable> parameters = {
{"src_depth",
- IntegralDivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)},
+ DivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)},
};
std::vector<std::pair<std::string, Object>> objects = {
- {"weights", MakeReadonlyObject(
- uint3(4, IntegralDivideRoundUp(attr.weights.shape.i, 4),
- IntegralDivideRoundUp(attr.weights.shape.o, 4)),
- ConvertToPHWO4I4(attr.weights))}};
+ {"weights",
+ MakeReadonlyObject(uint3(4, DivideRoundUp(attr.weights.shape.i, 4),
+ DivideRoundUp(attr.weights.shape.o, 4)),
+ ConvertToPHWO4I4(attr.weights))}};
std::string source;
for (int i = 0; i < multiplier; i++) {
absl::StrAppend(&source, "highp vec4 result", i, " = vec4(0);\n");
@@ -224,7 +224,7 @@
absl::StrAppend(&source, "value_0 = result0;\n");
}
- auto dst_depth = IntegralDivideRoundUp(ctx.output_shapes[0][3], 4);
+ auto dst_depth = DivideRoundUp(ctx.output_shapes[0][3], 4);
uint3 workgroup = uint3(16, 16, 1);
if (ctx.gpu_info->type == GpuType::ADRENO) {
if (dst_depth >= 2) {
@@ -265,7 +265,7 @@
/*shared_variables=*/{},
/*workload=*/
uint3(ctx.output_shapes[0][2] / multiplier, ctx.output_shapes[0][1],
- IntegralDivideRoundUp(ctx.output_shapes[0][3], 4)),
+ DivideRoundUp(ctx.output_shapes[0][3], 4)),
/*workgroup=*/
GetIdealWorkgroupIfPossible(
ctx.gpu_info->gpu_model, OperationType::CONVOLUTION_2D,
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/converter.cc b/tensorflow/lite/delegates/gpu/gl/kernels/converter.cc
index bc4c610..016be3c 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/converter.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/converter.cc
@@ -64,7 +64,7 @@
}
absl::Status Dispatch(const uint3& workload) {
- uint3 num_workgroups = IntegralDivideRoundUp(workload, workgroup_size_);
+ uint3 num_workgroups = DivideRoundUp(workload, workgroup_size_);
if (command_queue_) {
return command_queue_->Dispatch(program_, num_workgroups);
}
@@ -256,7 +256,7 @@
return absl::InvalidArgumentError(
"ToTensorConverter: output data size does not match expected size.");
}
- auto d = IntegralDivideRoundUp(shape_.c, 4);
+ auto d = DivideRoundUp(shape_.c, 4);
RETURN_IF_ERROR(program_.SetParameter(
{"sizes",
int4(static_cast<int32_t>(shape_.w), static_cast<int32_t>(shape_.h),
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/depthwise_conv.cc b/tensorflow/lite/delegates/gpu/gl/kernels/depthwise_conv.cc
index 9852221..71217a8 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/depthwise_conv.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/depthwise_conv.cc
@@ -54,7 +54,7 @@
{"dilation_h", attr.dilations.h},
{"kernel_w", weights.w},
{"kernel_h", weights.h},
- {"src_depth", IntegralDivideRoundUp(weights.i, 4)},
+ {"src_depth", DivideRoundUp(weights.i, 4)},
{"channel_multiplier", weights.o},
{"stride", int2(attr.strides.w, attr.strides.h)},
};
@@ -71,7 +71,7 @@
{"input_data_0_w", static_cast<int>(ctx.input_shapes[0][2])},
{"offsets_count", offsets_count},
{"offsets", offsets},
- {"src_depth", IntegralDivideRoundUp(weights.i, 4)},
+ {"src_depth", DivideRoundUp(weights.i, 4)},
{"channel_multiplier", weights.o},
{"stride", int2(attr.strides.w, attr.strides.h)},
};
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/fully_connected.cc b/tensorflow/lite/delegates/gpu/gl/kernels/fully_connected.cc
index 55ff886..0c3d2cd 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/fully_connected.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/fully_connected.cc
@@ -39,8 +39,8 @@
const auto& attr =
absl::any_cast<const FullyConnectedAttributes&>(ctx.op_attr);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
// This shader can work with any workgroup size, the values below work well
// for OpenGL.
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/mul.cc b/tensorflow/lite/delegates/gpu/gl/kernels/mul.cc
index 0dc0bba..c7d98e4 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/mul.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/mul.cc
@@ -105,10 +105,9 @@
/*shared_variables=*/{},
// Declare workload explicitly because shader depends on gid.z.
/*workload=*/
- uint3(
- static_cast<int>(ctx.input_shapes[0][2]),
- static_cast<int>(ctx.input_shapes[0][1]),
- IntegralDivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)),
+ uint3(static_cast<int>(ctx.input_shapes[0][2]),
+ static_cast<int>(ctx.input_shapes[0][1]),
+ DivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)),
/*workgroup=*/uint3(),
/*source_code=*/"value_0 *= $mul_buffer[gid.z]$;",
/*input=*/IOStructure::AUTO,
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/pad.cc b/tensorflow/lite/delegates/gpu/gl/kernels/pad.cc
index 2fd6c2c..d5fdaec 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/pad.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/pad.cc
@@ -98,8 +98,8 @@
source += " value_0 = $input_data_0[src_x, src_y, gid.z]$;\n";
} else if (attr.prepended.c % 4 == 0) {
parameters.push_back(
- {"src_slices", IntegralDivideRoundUp(
- static_cast<int>(ctx.input_shapes[0][3]), 4)});
+ {"src_slices",
+ DivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)});
source += R"(
int src_z = gid.z - $prepended.z$ / 4;
if (src_z >= 0 && src_z < $src_slices$) {
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/prelu.cc b/tensorflow/lite/delegates/gpu/gl/kernels/prelu.cc
index 1baac788..c3e9714 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/prelu.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/prelu.cc
@@ -69,8 +69,8 @@
/*workload=*/
uint3(static_cast<int>(ctx.output_shapes[0][2]),
static_cast<int>(ctx.output_shapes[0][1]),
- IntegralDivideRoundUp(
- static_cast<int>(ctx.output_shapes[0][3]), 4)),
+ DivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]),
+ 4)),
/*workgroup=*/uint3(),
/*source_code=*/
"value_0 = max(value_0, 0.0) + $alpha[gid.z]$ * min(value_0, "
@@ -98,10 +98,10 @@
"Alpha shape does not match input shape.");
}
- ObjectSize obj_size = uint3(
- static_cast<int>(ctx.output_shapes[0][2]),
- static_cast<int>(ctx.output_shapes[0][1]),
- IntegralDivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]), 4));
+ ObjectSize obj_size =
+ uint3(static_cast<int>(ctx.output_shapes[0][2]),
+ static_cast<int>(ctx.output_shapes[0][1]),
+ DivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]), 4));
*generated_code =
attr.clip
@@ -116,8 +116,8 @@
/*workload=*/
uint3(static_cast<int>(ctx.output_shapes[0][2]),
static_cast<int>(ctx.output_shapes[0][1]),
- IntegralDivideRoundUp(
- static_cast<int>(ctx.output_shapes[0][3]), 4)),
+ DivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]),
+ 4)),
/*workgroup=*/uint3(),
/*source_code=*/
"value_0 = clamp(value_0, 0.0, $clip$) + "
@@ -136,8 +136,8 @@
/*workload=*/
uint3(static_cast<int>(ctx.output_shapes[0][2]),
static_cast<int>(ctx.output_shapes[0][1]),
- IntegralDivideRoundUp(
- static_cast<int>(ctx.output_shapes[0][3]), 4)),
+ DivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]),
+ 4)),
/*workgroup=*/uint3(),
/*source_code=*/
"value_0 = max(value_0, 0.0) + $alpha[gid.x, gid.y, gid.z]$ "
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/softmax.cc b/tensorflow/lite/delegates/gpu/gl/kernels/softmax.cc
index ff5c028..f546fdd 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/softmax.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/softmax.cc
@@ -60,13 +60,13 @@
private:
absl::Status GenerateCodeFor1x1(const GenerationContext& ctx,
GeneratedCode* generated_code) const {
- const int depth = IntegralDivideRoundUp(ctx.output_shapes[0][3], 4);
+ const int depth = DivideRoundUp(ctx.output_shapes[0][3], 4);
std::vector<Variable> shared_variables = {
{"partial_sum", std::vector<float4>(8)},
};
std::vector<Variable> uniform_parameters = {
{"depth", depth},
- {"depth_div_32", IntegralDivideRoundUp(depth, 32)},
+ {"depth_div_32", DivideRoundUp(depth, 32)},
{"mask", GetMask(ctx.output_shapes[0][3])},
};
std::string source_code = R"(
@@ -138,7 +138,7 @@
GeneratedCode* generated_code) const {
std::vector<Variable> parameters = {
{"src_depth",
- IntegralDivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]), 4)},
+ DivideRoundUp(static_cast<int>(ctx.output_shapes[0][3]), 4)},
{"mask", GetMask(ctx.output_shapes[0][3])},
};
diff --git a/tensorflow/lite/delegates/gpu/gl/kernels/transpose_conv.cc b/tensorflow/lite/delegates/gpu/gl/kernels/transpose_conv.cc
index a183373..4b84843 100644
--- a/tensorflow/lite/delegates/gpu/gl/kernels/transpose_conv.cc
+++ b/tensorflow/lite/delegates/gpu/gl/kernels/transpose_conv.cc
@@ -44,7 +44,7 @@
std::vector<Variable> parameters = {
{"input_data_0_h", static_cast<int>(ctx.input_shapes[0][1])},
{"input_data_0_w", static_cast<int>(ctx.input_shapes[0][2])},
- {"src_depth", IntegralDivideRoundUp(weights.i, 4)},
+ {"src_depth", DivideRoundUp(weights.i, 4)},
{"kernel_size", int2(weights.w, weights.h)},
{"stride", int2(attr.stride.w, attr.stride.h)},
{"padding", int2(weights.w - 1 - attr.padding.prepended.w,
diff --git a/tensorflow/lite/delegates/gpu/gl/object.h b/tensorflow/lite/delegates/gpu/gl/object.h
index 7ea1614..3463d06 100644
--- a/tensorflow/lite/delegates/gpu/gl/object.h
+++ b/tensorflow/lite/delegates/gpu/gl/object.h
@@ -153,17 +153,17 @@
inline Object MakeReadonlyObject(const std::vector<float>& data) {
return MakeReadonlyObject(
- IntegralDivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
+ DivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
}
inline Object MakeReadonlyTexture(const std::vector<float>& data) {
return MakeReadonlyTexture(
- IntegralDivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
+ DivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
}
inline Object MakeReadonlyBuffer(const std::vector<float>& data) {
return MakeReadonlyBuffer(
- IntegralDivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
+ DivideRoundUp(static_cast<uint32_t>(data.size()), 4U), data);
}
// TODO(akulik): find better place for functions below.
@@ -172,7 +172,7 @@
uint3 size;
size.x = shape.w;
size.y = shape.h;
- size.z = shape.b * IntegralDivideRoundUp(shape.c, 4);
+ size.z = shape.b * DivideRoundUp(shape.c, 4);
return size;
}
diff --git a/tensorflow/lite/delegates/gpu/metal/api.cc b/tensorflow/lite/delegates/gpu/metal/api.cc
index 71e779f..2fbc3b1 100644
--- a/tensorflow/lite/delegates/gpu/metal/api.cc
+++ b/tensorflow/lite/delegates/gpu/metal/api.cc
@@ -167,10 +167,10 @@
bool IsSuitableForWinograd4x4To6x6(const Convolution2DAttributes& attr,
const BHWC& dst_shape) {
- const int tiles_x = IntegralDivideRoundUp(dst_shape.w, 4);
- const int tiles_y = IntegralDivideRoundUp(dst_shape.h, 4);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int tiles_x = DivideRoundUp(dst_shape.w, 4);
+ const int tiles_y = DivideRoundUp(dst_shape.h, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
const bool suitable_attributes =
attr.weights.shape.w == 3 && attr.weights.shape.h == 3 &&
attr.dilations == HW(1, 1) && attr.strides == HW(1, 1);
@@ -229,8 +229,8 @@
auto attr =
absl::any_cast<Convolution2DAttributes>(node->operation.attributes);
if (IsSuitableForWinograd4x4To6x6(attr, dst_shape)) {
- int tiles_x = IntegralDivideRoundUp(dst_shape.w, 4);
- int tiles_y = IntegralDivideRoundUp(dst_shape.h, 4);
+ int tiles_x = DivideRoundUp(dst_shape.w, 4);
+ int tiles_y = DivideRoundUp(dst_shape.h, 4);
Winograd4x4To36Attributes wino_up_attr;
wino_up_attr.padding = attr.padding;
diff --git a/tensorflow/lite/delegates/gpu/metal/buffer_convert.mm b/tensorflow/lite/delegates/gpu/metal/buffer_convert.mm
index 8ddf78e..1fdf97b 100644
--- a/tensorflow/lite/delegates/gpu/metal/buffer_convert.mm
+++ b/tensorflow/lite/delegates/gpu/metal/buffer_convert.mm
@@ -21,8 +21,8 @@
#include "tensorflow/lite/delegates/gpu/common/util.h"
#include "tensorflow/lite/delegates/gpu/metal/common.h"
-using ::tflite::gpu::IntegralDivideRoundUp;
using ::tflite::gpu::BHWC;
+using ::tflite::gpu::DivideRoundUp;
using ::tflite::gpu::metal::CreateComputeProgram;
@implementation TFLBufferConvert {
@@ -102,10 +102,10 @@
[encoder setBytes:uniforms.data() length:uniforms.size() * sizeof(int) atIndex:2];
MTLSize group_size = MTLSizeMake(16, 16, 1);
- int layers = IntegralDivideRoundUp(shape.c, 4);
- int groups_x = IntegralDivideRoundUp(shape.w, group_size.width);
- int groups_y = IntegralDivideRoundUp(shape.h, group_size.height);
- int groups_z = IntegralDivideRoundUp(layers, group_size.depth);
+ int layers = DivideRoundUp(shape.c, 4);
+ int groups_x = DivideRoundUp(shape.w, group_size.width);
+ int groups_y = DivideRoundUp(shape.h, group_size.height);
+ int groups_z = DivideRoundUp(layers, group_size.depth);
MTLSize groups_count = MTLSizeMake(groups_x, groups_y, groups_z);
[encoder dispatchThreadgroups:groups_count threadsPerThreadgroup:group_size];
}
diff --git a/tensorflow/lite/delegates/gpu/metal/compiled_model.cc b/tensorflow/lite/delegates/gpu/metal/compiled_model.cc
index ce8e5e7..74202ed 100644
--- a/tensorflow/lite/delegates/gpu/metal/compiled_model.cc
+++ b/tensorflow/lite/delegates/gpu/metal/compiled_model.cc
@@ -440,9 +440,9 @@
desc->resize_function = [input_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(input_id)->second;
uint3 groups_size{16, 16, 1};
- uint3 groups_count{IntegralDivideRoundUp(dimension.w, groups_size.x),
- IntegralDivideRoundUp(dimension.h, groups_size.y),
- IntegralDivideRoundUp(dimension.c, 4)};
+ uint3 groups_count{DivideRoundUp(dimension.w, groups_size.x),
+ DivideRoundUp(dimension.h, groups_size.y),
+ DivideRoundUp(dimension.c, 4)};
return std::make_pair(groups_size, groups_count);
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/concat.cc b/tensorflow/lite/delegates/gpu/metal/kernels/concat.cc
index c252ee0..56d270b 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/concat.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/concat.cc
@@ -71,7 +71,7 @@
// Also it is easy to write a loop in this case, to prevent long kernel
// generation.
for (int i = 0; i < channels.size(); ++i) {
- const int depth = IntegralDivideRoundUp(channels[i], 4);
+ const int depth = DivideRoundUp(channels[i], 4);
const std::string src_buffer = "src_buffer" + std::to_string(i);
c += " for (int i = 0; i < " + std::to_string(depth) + "; ++i) {\n";
c += " int src_index = i * U.src_size.w + xy_offset;\n";
@@ -88,7 +88,7 @@
int read_index = 0;
int z = 0;
for (int i = 0; i < channels.size(); ++i) {
- const int depth = IntegralDivideRoundUp(channels[i], 4);
+ const int depth = DivideRoundUp(channels[i], 4);
const std::string src_buffer = "src_buffer" + std::to_string(i);
for (int d = 0; d < depth; ++d) {
const int channels_in_group = std::min(4, channels[i] - d * 4);
@@ -168,11 +168,11 @@
std::vector<int> uniform_params{
src_shape.w,
src_shape.h,
- IntegralDivideRoundUp(src_shape.c, 4),
+ DivideRoundUp(src_shape.c, 4),
src_shape.w * src_shape.h,
dst_shape.w,
dst_shape.h,
- IntegralDivideRoundUp(dst_shape.c, 4),
+ DivideRoundUp(dst_shape.c, 4),
dst_shape.w * dst_shape.h,
};
return GetByteBuffer(uniform_params);
@@ -184,9 +184,9 @@
uint3 grid(dst_shape.w, dst_shape.h, 1);
uint3 group_size{8u, 4u, 1u};
uint3 groups;
- groups.x = IntegralDivideRoundUp(grid.x, group_size.x);
- groups.y = IntegralDivideRoundUp(grid.y, group_size.y);
- groups.z = IntegralDivideRoundUp(grid.z, group_size.z);
+ groups.x = DivideRoundUp(grid.x, group_size.x);
+ groups.y = DivideRoundUp(grid.y, group_size.y);
+ groups.z = DivideRoundUp(grid.z, group_size.z);
return std::make_pair(group_size, groups);
};
@@ -265,7 +265,7 @@
[output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(output_id)->second;
std::vector<int> uniform_params{dimension.w, dimension.h,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
/*padding=*/0};
return GetByteBuffer(uniform_params);
}},
@@ -274,9 +274,9 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& output_dims = buffers.find(output_id)->second;
const uint3 groups_size{8, 4, 1};
- int groups_x = IntegralDivideRoundUp(output_dims.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(output_dims.h, groups_size.y);
- int groups_z = IntegralDivideRoundUp(output_dims.c, 4);
+ int groups_x = DivideRoundUp(output_dims.w, groups_size.x);
+ int groups_y = DivideRoundUp(output_dims.h, groups_size.y);
+ int groups_z = DivideRoundUp(output_dims.c, 4);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
@@ -356,7 +356,7 @@
[output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(output_id)->second;
std::vector<int> uniform_params{dimension.w, dimension.h,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
/*padding=*/0};
return GetByteBuffer(uniform_params);
}},
@@ -365,9 +365,9 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& output_dims = buffers.find(output_id)->second;
const uint3 groups_size{8, 4, 1};
- int groups_x = IntegralDivideRoundUp(output_dims.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(output_dims.h, groups_size.y);
- int groups_z = IntegralDivideRoundUp(output_dims.c, 4);
+ int groups_x = DivideRoundUp(output_dims.w, groups_size.x);
+ int groups_y = DivideRoundUp(output_dims.h, groups_size.y);
+ int groups_z = DivideRoundUp(output_dims.c, 4);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc b/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
index f9ff87e..04cd95d 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/conv.cc
@@ -72,7 +72,7 @@
namespace {
int GetNumOutputSlices(int dst_channels) {
- const int dst_depth = IntegralDivideRoundUp(dst_channels, 4);
+ const int dst_depth = DivideRoundUp(dst_channels, 4);
if (dst_depth % 4 == 0 || dst_depth >= 16) {
return 4;
} else if (dst_depth % 2 == 0 || dst_depth >= 4) {
@@ -571,8 +571,8 @@
std::vector<float> ReorderWeightsForConv(
const tflite::gpu::Tensor<OHWI, DataType::FLOAT32>& weights,
const ConvParams& params) {
- const int dst_depth = IntegralDivideRoundUp(weights.shape.o, 4);
- const int src_depth = IntegralDivideRoundUp(weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(weights.shape.i, 4);
std::vector<float> weights_reordered(
weights.shape.w * weights.shape.h *
AlignByN(dst_depth, params.block_size.z) * 4 * src_depth * 4);
@@ -580,8 +580,7 @@
bool isO4I4 = params.weight_layout == WeightsInnerBlockLayout::O4I4;
int counter = 0;
- for (int d = 0; d < IntegralDivideRoundUp(dst_depth, params.block_size.z);
- ++d) {
+ for (int d = 0; d < DivideRoundUp(dst_depth, params.block_size.z); ++d) {
for (int y = 0; y < weights.shape.h; ++y) {
for (int x = 0; x < weights.shape.w; ++x) {
for (int s = 0; s < src_depth; ++s) {
@@ -618,17 +617,17 @@
const BHWC& dst_size,
const Convolution2DAttributes& attr,
const ConvParams& params) {
- const int grid_x = IntegralDivideRoundUp(dst_size.w, params.block_size.x);
- const int grid_y = IntegralDivideRoundUp(dst_size.h, params.block_size.y);
+ const int grid_x = DivideRoundUp(dst_size.w, params.block_size.x);
+ const int grid_y = DivideRoundUp(dst_size.h, params.block_size.y);
std::vector<int> uniform_params = {
src_size.w,
src_size.h,
src_size.w * src_size.h,
- IntegralDivideRoundUp(src_size.c, 4),
+ DivideRoundUp(src_size.c, 4),
dst_size.w,
dst_size.h,
dst_size.w * dst_size.h,
- IntegralDivideRoundUp(dst_size.c, 4),
+ DivideRoundUp(dst_size.c, 4),
attr.strides.w,
attr.strides.h,
-attr.padding.prepended.w,
@@ -652,17 +651,17 @@
std::vector<uint8_t> GetUniformBufferForWinograd(const BHWC& src_size,
const BHWC& dst_size,
const ConvParams& params) {
- const int grid_x = IntegralDivideRoundUp(dst_size.w, params.block_size.x);
- const int grid_y = IntegralDivideRoundUp(dst_size.h, params.block_size.y);
+ const int grid_x = DivideRoundUp(dst_size.w, params.block_size.x);
+ const int grid_y = DivideRoundUp(dst_size.h, params.block_size.y);
std::vector<int> uniform_params = {
src_size.w,
src_size.h,
src_size.w * src_size.h,
- IntegralDivideRoundUp(src_size.c, 4),
+ DivideRoundUp(src_size.c, 4),
dst_size.w,
dst_size.h,
dst_size.w * dst_size.h,
- IntegralDivideRoundUp(dst_size.c, 4),
+ DivideRoundUp(dst_size.c, 4),
1,
1,
0,
@@ -685,38 +684,37 @@
int GetGroupsCount(const BHWC& dst_shape, const int3& wg_size,
const int3& block_size) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
- int grid_x = IntegralDivideRoundUp(dst_shape.w, block_size.x);
- int grid_y = IntegralDivideRoundUp(dst_shape.h, block_size.y);
- int grid_z = IntegralDivideRoundUp(dst_slices, block_size.z);
+ int grid_x = DivideRoundUp(dst_shape.w, block_size.x);
+ int grid_y = DivideRoundUp(dst_shape.h, block_size.y);
+ int grid_z = DivideRoundUp(dst_slices, block_size.z);
- return IntegralDivideRoundUp(grid_x, wg_size.x) *
- IntegralDivideRoundUp(grid_y, wg_size.y) *
- IntegralDivideRoundUp(grid_z, wg_size.z);
+ return DivideRoundUp(grid_x, wg_size.x) * DivideRoundUp(grid_y, wg_size.y) *
+ DivideRoundUp(grid_z, wg_size.z);
}
int GetGroupsCountForLinearWH(const BHWC& dst_shape, const int3& wg_size,
const int3& block_size) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
- int grid_x = IntegralDivideRoundUp(dst_shape.w, block_size.x);
- int grid_y = IntegralDivideRoundUp(dst_shape.h, block_size.y);
- int grid_z = IntegralDivideRoundUp(dst_slices, block_size.z);
+ int grid_x = DivideRoundUp(dst_shape.w, block_size.x);
+ int grid_y = DivideRoundUp(dst_shape.h, block_size.y);
+ int grid_z = DivideRoundUp(dst_slices, block_size.z);
- return IntegralDivideRoundUp(grid_x * grid_y, wg_size.x) *
- IntegralDivideRoundUp(grid_z, wg_size.y);
+ return DivideRoundUp(grid_x * grid_y, wg_size.x) *
+ DivideRoundUp(grid_z, wg_size.y);
}
int GetGroupsCountForLinearWHS(const BHWC& dst_shape, const int3& wg_size,
const int3& block_size) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
- int grid_x = IntegralDivideRoundUp(dst_shape.w, block_size.x);
- int grid_y = IntegralDivideRoundUp(dst_shape.h, block_size.y);
- int grid_z = IntegralDivideRoundUp(dst_slices, block_size.z);
+ int grid_x = DivideRoundUp(dst_shape.w, block_size.x);
+ int grid_y = DivideRoundUp(dst_shape.h, block_size.y);
+ int grid_z = DivideRoundUp(dst_slices, block_size.z);
- return IntegralDivideRoundUp(grid_x * grid_y * grid_z, wg_size.x);
+ return DivideRoundUp(grid_x * grid_y * grid_z, wg_size.x);
}
bool IsKernelXIs1(const Convolution2DAttributes& attr) {
@@ -758,8 +756,8 @@
ConvParams GetConvParamsForA7A8(const AppleGPUInfo& apple_info,
const Convolution2DAttributes& attr,
const BHWC& dst_shape) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
- const int src_slices = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
+ const int src_slices = DivideRoundUp(attr.weights.shape.i, 4);
ConvParams params;
params.weights_upload_type = WeightsUploadType::LOCAL_MEM_BY_THREADS;
@@ -835,8 +833,8 @@
ConvParams GetConvParamsForA9AndHigher(const AppleGPUInfo& apple_info,
const Convolution2DAttributes& attr,
const BHWC& dst_shape) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
- const int src_slices = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
+ const int src_slices = DivideRoundUp(attr.weights.shape.i, 4);
int blk_total_size = GetRecommendedBlockSize(apple_info, dst_shape);
int3 block_size = int3(1, 1, 1);
if (blk_total_size >= 2 && apple_info.IsBionic()) {
@@ -917,8 +915,8 @@
ConvParams GetConvParamsForIntel(const Convolution2DAttributes& attr,
const RuntimeOptions& options,
const BHWC& dst_shape) {
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
- const int src_slices = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
+ const int src_slices = DivideRoundUp(attr.weights.shape.i, 4);
ConvParams params;
params.weights_upload_type = WeightsUploadType::PRIVATE_MEM_SIMD8_BROADCAST;
params.x_kernel_is_1 = IsKernelXIs1(attr);
@@ -1017,29 +1015,28 @@
std::pair<uint3, uint3> GetDispatchSizes(const ConvParams& params,
const BHWC& shape) {
- const int dst_slices = IntegralDivideRoundUp(shape.c, 4);
+ const int dst_slices = DivideRoundUp(shape.c, 4);
- int grid_x = IntegralDivideRoundUp(shape.w, params.block_size.x);
- int grid_y = IntegralDivideRoundUp(shape.h, params.block_size.y);
- int grid_z = IntegralDivideRoundUp(dst_slices, params.block_size.z);
+ int grid_x = DivideRoundUp(shape.w, params.block_size.x);
+ int grid_y = DivideRoundUp(shape.h, params.block_size.y);
+ int grid_z = DivideRoundUp(dst_slices, params.block_size.z);
const uint3 group_size(params.work_group_size.x, params.work_group_size.y,
params.work_group_size.z);
int3 wg;
uint3 groups_count;
if (params.linear_whs) {
- wg.x = IntegralDivideRoundUp(grid_x * grid_y * grid_z,
- params.work_group_size.x);
+ wg.x = DivideRoundUp(grid_x * grid_y * grid_z, params.work_group_size.x);
groups_count = uint3(wg.x, 1, 1);
} else if (params.linear_wh) {
- wg.x = IntegralDivideRoundUp(grid_x * grid_y, params.work_group_size.x);
- wg.y = IntegralDivideRoundUp(grid_z, params.work_group_size.y);
+ wg.x = DivideRoundUp(grid_x * grid_y, params.work_group_size.x);
+ wg.y = DivideRoundUp(grid_z, params.work_group_size.y);
groups_count = uint3(wg[params.work_group_launch_order.x],
wg[params.work_group_launch_order.y], 1);
} else {
- wg.x = IntegralDivideRoundUp(grid_x, params.work_group_size.x);
- wg.y = IntegralDivideRoundUp(grid_y, params.work_group_size.y);
- wg.z = IntegralDivideRoundUp(grid_z, params.work_group_size.z);
+ wg.x = DivideRoundUp(grid_x, params.work_group_size.x);
+ wg.y = DivideRoundUp(grid_y, params.work_group_size.y);
+ wg.z = DivideRoundUp(grid_z, params.work_group_size.z);
groups_count = uint3(wg[params.work_group_launch_order.x],
wg[params.work_group_launch_order.y],
wg[params.work_group_launch_order.z]);
@@ -1076,7 +1073,7 @@
std::string addr_space =
params.weights_upload_type == WeightsUploadType::CONSTANT_MEM ? "constant"
: "device";
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
desc->immutable_buffers = {
{addr_space + " FLT4* const filters",
GetByteBufferConverted(weights_reordered, options.storage_precision)},
@@ -1108,7 +1105,7 @@
int id, ValueId input_id, ValueId output_id, const BHWC& dst_shape,
const Convolution2DAttributes& attr, const DeviceInfo& device_info,
const RuntimeOptions& options) {
- const int dst_slices = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int dst_slices = DivideRoundUp(attr.weights.shape.o, 4);
ConvParams params;
params.work_group_launch_order = int3(2, 0, 1);
params.src_depth_loop_size = 1;
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/conv_test.mm b/tensorflow/lite/delegates/gpu/metal/kernels/conv_test.mm
index 0291cd7..fc9e015 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/conv_test.mm
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/conv_test.mm
@@ -34,6 +34,7 @@
using ::tflite::gpu::BHWC;
using ::tflite::gpu::Convolution2DAttributes;
using ::tflite::gpu::DataType;
+using ::tflite::gpu::DivideRoundUp;
using ::tflite::gpu::HW;
using ::tflite::gpu::Linear;
using ::tflite::gpu::OHWI;
@@ -44,7 +45,6 @@
using ::tflite::gpu::ValueId;
using ::tflite::gpu::metal::ConvolutionGeneric;
using ::tflite::gpu::metal::ConvolutionWino4x4To6x6;
-using ::tflite::gpu::IntegralDivideRoundUp;
using ::tflite::gpu::metal::CompareVectors;
using ::tflite::gpu::metal::SingleOpModel;
@@ -275,7 +275,7 @@
BHWC conv_shape;
conv_shape.b = dst_shape.b;
conv_shape.h = 36;
- conv_shape.w = IntegralDivideRoundUp(new_width, 4) * IntegralDivideRoundUp(new_height, 4);
+ conv_shape.w = DivideRoundUp(new_width, 4) * DivideRoundUp(new_height, 4);
conv_shape.c = dst_shape.c;
TensorFloat32 src_tensor;
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/depthwise_conv.cc b/tensorflow/lite/delegates/gpu/metal/kernels/depthwise_conv.cc
index 6c26a87..8c22394 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/depthwise_conv.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/depthwise_conv.cc
@@ -208,7 +208,7 @@
// DepthWiseConv3x3Stride1x1
std::vector<float> ReorderWeightsDepthWiseConv3x3Stride1x1(
const DepthwiseConvolution2DAttributes& attr) {
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
const int kernel_x = 3;
const int kernel_y = 3;
std::vector<float> weights_reordered((kernel_x * kernel_y + 1) * src_depth *
@@ -250,11 +250,11 @@
src_size.w,
src_size.h,
src_size.w * src_size.h,
- IntegralDivideRoundUp(src_size.c, 4),
+ DivideRoundUp(src_size.c, 4),
dst_size.w,
dst_size.h,
dst_size.w * dst_size.h,
- IntegralDivideRoundUp(dst_size.c, 4),
+ DivideRoundUp(dst_size.c, 4),
-params.padding.prepended.w,
-params.padding.prepended.h,
0, // dummy, for alignment
@@ -403,7 +403,7 @@
// DepthWiseConv3x3Stride2
std::vector<float> ReorderWeightsDepthWiseConv3x3Stride2(
const DepthwiseConvolution2DAttributes& attr) {
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
const int kernel_x = 3;
const int kernel_y = 3;
std::vector<float> weights_reordered((kernel_x * kernel_y + 1) * src_depth *
@@ -445,11 +445,11 @@
src_size.w,
src_size.h,
src_size.w * src_size.h,
- IntegralDivideRoundUp(src_size.c, 4),
+ DivideRoundUp(src_size.c, 4),
dst_size.w,
dst_size.h,
dst_size.w * dst_size.h,
- IntegralDivideRoundUp(dst_size.c, 4),
+ DivideRoundUp(dst_size.c, 4),
-attr.padding.prepended.w,
-attr.padding.prepended.h,
attr.strides.w,
@@ -586,11 +586,11 @@
std::vector<int> uniform_params{
dimension.w,
dimension.h,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
0,
output_dimension.w,
output_dimension.h,
- IntegralDivideRoundUp(output_dimension.c, 4),
+ DivideRoundUp(output_dimension.c, 4),
0,
attr.strides.w,
attr.strides.h,
@@ -612,9 +612,9 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(output_id)->second;
uint3 groups_size{8, 4, 1};
- uint3 groups_count{IntegralDivideRoundUp(dimension.w, groups_size.x),
- IntegralDivideRoundUp(dimension.h, groups_size.y),
- IntegralDivideRoundUp(dimension.c, 4)};
+ uint3 groups_count{DivideRoundUp(dimension.w, groups_size.x),
+ DivideRoundUp(dimension.h, groups_size.y),
+ DivideRoundUp(dimension.c, 4)};
return std::make_pair(groups_size, groups_count);
};
@@ -661,17 +661,17 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(output_id)->second;
- const int grid_x = IntegralDivideRoundUp(dimension.w, 2);
- const int grid_y = IntegralDivideRoundUp(dimension.h, 2);
- const int grid_z = IntegralDivideRoundUp(dimension.c, 4);
+ const int grid_x = DivideRoundUp(dimension.w, 2);
+ const int grid_y = DivideRoundUp(dimension.h, 2);
+ const int grid_z = DivideRoundUp(dimension.c, 4);
uint3 group_size{8, 4, 1};
if (grid_x <= 4) {
group_size.x = 4;
group_size.z = grid_z % 2 == 0 ? 2 : 1;
}
- const int groups_x = IntegralDivideRoundUp(grid_x, group_size.x);
- const int groups_y = IntegralDivideRoundUp(grid_y, group_size.y);
- const int groups_z = IntegralDivideRoundUp(grid_z, group_size.z);
+ const int groups_x = DivideRoundUp(grid_x, group_size.x);
+ const int groups_y = DivideRoundUp(grid_y, group_size.y);
+ const int groups_z = DivideRoundUp(grid_z, group_size.z);
return std::make_pair(group_size, uint3(groups_x, groups_y, groups_z));
};
@@ -726,12 +726,12 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& dimension = buffers.find(output_id)->second;
const int grid_x = dimension.w;
- const int grid_y = IntegralDivideRoundUp(dimension.h, 2);
- const int grid_z = IntegralDivideRoundUp(dimension.c, 4);
+ const int grid_y = DivideRoundUp(dimension.h, 2);
+ const int grid_z = DivideRoundUp(dimension.c, 4);
const uint3 group_size{8, 4, 1};
- const int groups_x = IntegralDivideRoundUp(grid_x, group_size.x);
- const int groups_y = IntegralDivideRoundUp(grid_y, group_size.y);
- const int groups_z = IntegralDivideRoundUp(grid_z, group_size.z);
+ const int groups_x = DivideRoundUp(grid_x, group_size.x);
+ const int groups_y = DivideRoundUp(grid_y, group_size.y);
+ const int groups_z = DivideRoundUp(grid_z, group_size.z);
return std::make_pair(group_size, uint3(groups_x, groups_y, groups_z));
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/fully_connected.cc b/tensorflow/lite/delegates/gpu/metal/kernels/fully_connected.cc
index 331f3cc..9872328 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/fully_connected.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/fully_connected.cc
@@ -45,7 +45,7 @@
const std::string barrier = device_info.IsWaveSizeEqualTo32()
? "SIMDGROUP_BARRIER"
: "threadgroup_barrier";
- const int src_depth = IntegralDivideRoundUp(src_channels, 4);
+ const int src_depth = DivideRoundUp(src_channels, 4);
std::stringstream code;
code << R"(
#include <metal_stdlib>
@@ -116,9 +116,8 @@
}
}
)";
- const int src_depth_sub_groups = shared_memory
- ? IntegralDivideRoundUp(src_depth, 32)
- : IntegralDivideRoundUp(src_depth, 4);
+ const int src_depth_sub_groups = shared_memory ? DivideRoundUp(src_depth, 32)
+ : DivideRoundUp(src_depth, 4);
return absl::Substitute(code.str(), src_depth_sub_groups, barrier);
}
} // namespace
@@ -146,7 +145,7 @@
bool shared_memory =
device_info.IsAppleGPU() &&
device_info.apple_info.IsLocalMemoryPreferredOverGlobal();
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
const int src_depth_aligned = AlignByN(src_depth, shared_memory ? 32 : 4);
const int dst_channels_aligned = AlignByN(attr.weights.shape.o, 8);
@@ -179,8 +178,7 @@
{"constant uniforms& params",
[attr](const std::map<ValueId, BHWC>& buffers) {
std::vector<uint32_t> uniform_params{
- static_cast<uint32_t>(
- IntegralDivideRoundUp(attr.weights.shape.i, 4)),
+ static_cast<uint32_t>(DivideRoundUp(attr.weights.shape.i, 4)),
static_cast<uint32_t>(AlignByN(attr.weights.shape.o, 8)),
static_cast<uint32_t>(attr.weights.shape.o),
static_cast<uint32_t>(0),
@@ -192,7 +190,7 @@
desc->resize_function = [attr](const std::map<ValueId, BHWC>& buffers) {
const uint3 groups_size{8, 4, 1};
const int dst_channels_aligned = AlignByN(attr.weights.shape.o, 8);
- int groups_x = IntegralDivideRoundUp(dst_channels_aligned, groups_size.x);
+ int groups_x = DivideRoundUp(dst_channels_aligned, groups_size.x);
return std::make_pair(groups_size, uint3{groups_x, 1, 1});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/max_unpooling.cc b/tensorflow/lite/delegates/gpu/metal/kernels/max_unpooling.cc
index a94bf1c..d0e326b 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/max_unpooling.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/max_unpooling.cc
@@ -131,9 +131,9 @@
const auto& src_shape = buffers.find(input_id)->second;
BHWC dst_shape = CalculateOutputShape(src_shape, params);
const uint3 groups_size{16, 16, 1};
- int groups_x = IntegralDivideRoundUp(dst_shape.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(dst_shape.h, groups_size.y);
- int groups_z = IntegralDivideRoundUp(dst_shape.c, 4);
+ int groups_x = DivideRoundUp(dst_shape.w, groups_size.x);
+ int groups_y = DivideRoundUp(dst_shape.h, groups_size.y);
+ int groups_z = DivideRoundUp(dst_shape.c, 4);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/mean.cc b/tensorflow/lite/delegates/gpu/metal/kernels/mean.cc
index b4e06fb..431b1e5 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/mean.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/mean.cc
@@ -133,7 +133,7 @@
[input_id, output_id,
work_group_size](const std::map<ValueId, BHWC>& buffers) {
const auto& src_shape = buffers.find(input_id)->second;
- const int src_slices = IntegralDivideRoundUp(src_shape.c, 4);
+ const int src_slices = DivideRoundUp(src_shape.c, 4);
struct uniforms {
int4 src_size;
float4 inv_multipliers;
@@ -153,8 +153,8 @@
desc->resize_function = [output_id, work_group_size](
const std::map<ValueId, BHWC>& buffers) {
BHWC dst_shape = buffers.find(output_id)->second;
- const int dst_slices = IntegralDivideRoundUp(dst_shape.c, 4);
- const int groups_z = IntegralDivideRoundUp(dst_slices, work_group_size.z);
+ const int dst_slices = DivideRoundUp(dst_shape.c, 4);
+ const int groups_z = DivideRoundUp(dst_slices, work_group_size.z);
return std::make_pair(work_group_size, uint3{1, 1, groups_z});
};
return {desc};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/padding.cc b/tensorflow/lite/delegates/gpu/metal/kernels/padding.cc
index bc63a23..b117df9 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/padding.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/padding.cc
@@ -177,12 +177,12 @@
dimension.w,
dimension.h,
dimension.c,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
// int4 dst_size
output_dimension.w,
output_dimension.h,
output_dimension.c,
- IntegralDivideRoundUp(output_dimension.c, 4),
+ DivideRoundUp(output_dimension.c, 4),
// int4 prepended padding
attr.prepended.w,
attr.prepended.h,
@@ -198,10 +198,10 @@
const uint3 groups_size{16, 16, 1};
const auto& src_shape = buffers.find(input_id)->second;
BHWC dst_shape = CalculateOutputShape(src_shape, attr);
- const int dst_layers = IntegralDivideRoundUp(dst_shape.c, 4);
- int groups_x = IntegralDivideRoundUp(dst_shape.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(dst_shape.h, groups_size.y);
- int groups_z = IntegralDivideRoundUp(dst_layers, groups_size.z);
+ const int dst_layers = DivideRoundUp(dst_shape.c, 4);
+ int groups_x = DivideRoundUp(dst_shape.w, groups_size.x);
+ int groups_y = DivideRoundUp(dst_shape.h, groups_size.y);
+ int groups_z = DivideRoundUp(dst_layers, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/pooling.cc b/tensorflow/lite/delegates/gpu/metal/kernels/pooling.cc
index 3ba8c90..eaf4c9d 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/pooling.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/pooling.cc
@@ -224,11 +224,11 @@
std::vector<int> uniform_params = {
dimension.w,
dimension.h,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
dimension.w * dimension.h,
output_dimension.w,
output_dimension.h,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
output_dimension.w * output_dimension.h,
params.strides.w,
params.strides.h,
@@ -242,11 +242,11 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
BHWC dst_shape = buffers.find(output_id)->second;
const uint3 grid =
- uint3(dst_shape.w, dst_shape.h, IntegralDivideRoundUp(dst_shape.c, 4));
+ uint3(dst_shape.w, dst_shape.h, DivideRoundUp(dst_shape.c, 4));
const uint3 groups_size = GetWorkGroupSizeForGrid(grid);
- int groups_x = IntegralDivideRoundUp(grid.x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid.y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid.z, groups_size.z);
+ int groups_x = DivideRoundUp(grid.x, groups_size.x);
+ int groups_y = DivideRoundUp(grid.y, groups_size.y);
+ int groups_z = DivideRoundUp(grid.z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/reshape.cc b/tensorflow/lite/delegates/gpu/metal/kernels/reshape.cc
index 42b8a73..3bf392d 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/reshape.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/reshape.cc
@@ -159,11 +159,11 @@
desc->resize_function = [attr](const std::map<ValueId, BHWC>& buffers) {
const uint3 grid = uint3(attr.new_shape.w, attr.new_shape.h,
- IntegralDivideRoundUp(attr.new_shape.c, 4));
+ DivideRoundUp(attr.new_shape.c, 4));
const uint3 groups_size = GetWorkGroupSizeForGrid(grid);
- int groups_x = IntegralDivideRoundUp(grid.x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid.y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid.z, groups_size.z);
+ int groups_x = DivideRoundUp(grid.x, groups_size.x);
+ int groups_y = DivideRoundUp(grid.y, groups_size.y);
+ int groups_z = DivideRoundUp(grid.z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
@@ -197,14 +197,14 @@
const auto& dst_dim = buffers.find(output_id)->second;
std::vector<int32_t> uniform_params{
// int4 src_size
- src_dim.w, src_dim.h, IntegralDivideRoundUp(src_dim.c, 4),
+ src_dim.w, src_dim.h, DivideRoundUp(src_dim.c, 4),
src_dim.w * src_dim.h,
// int4 dst_size
- dst_dim.w, dst_dim.h, IntegralDivideRoundUp(dst_dim.c, 4),
+ dst_dim.w, dst_dim.h, DivideRoundUp(dst_dim.c, 4),
dst_dim.w * dst_dim.h,
// int2 plane_xz
- src_dim.w * IntegralDivideRoundUp(src_dim.c, 4),
- dst_dim.w * IntegralDivideRoundUp(dst_dim.c, 4),
+ src_dim.w * DivideRoundUp(src_dim.c, 4),
+ dst_dim.w * DivideRoundUp(dst_dim.c, 4),
0, // dummy, for alignment
0, // dummy, for alignment
0, // dummy, for alignment
@@ -218,11 +218,11 @@
desc->resize_function = [attr](const std::map<ValueId, BHWC>& buffers) {
const uint3 grid = uint3(attr.new_shape.w, attr.new_shape.h,
- IntegralDivideRoundUp(attr.new_shape.c, 4));
+ DivideRoundUp(attr.new_shape.c, 4));
const uint3 groups_size = GetWorkGroupSizeForGrid(grid);
- int groups_x = IntegralDivideRoundUp(grid.x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid.y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid.z, groups_size.z);
+ int groups_x = DivideRoundUp(grid.x, groups_size.x);
+ int groups_y = DivideRoundUp(grid.y, groups_size.y);
+ int groups_z = DivideRoundUp(grid.z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/resize.cc b/tensorflow/lite/delegates/gpu/metal/kernels/resize.cc
index 24d7bcf..49a65c1 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/resize.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/resize.cc
@@ -153,10 +153,10 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const uint3 groups_size{16, 16, 1};
const auto& dst_dim = buffers.find(output_id)->second;
- int groups_x = IntegralDivideRoundUp(dst_dim.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(dst_dim.h, groups_size.y);
- const int dst_layers = IntegralDivideRoundUp(dst_dim.c, 4);
- int groups_z = IntegralDivideRoundUp(dst_layers, groups_size.z);
+ int groups_x = DivideRoundUp(dst_dim.w, groups_size.x);
+ int groups_y = DivideRoundUp(dst_dim.h, groups_size.y);
+ const int dst_layers = DivideRoundUp(dst_dim.c, 4);
+ int groups_z = DivideRoundUp(dst_layers, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
return {desc};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/slice.cc b/tensorflow/lite/delegates/gpu/metal/kernels/slice.cc
index 8db9dd0..b1d78dc 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/slice.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/slice.cc
@@ -157,12 +157,12 @@
dimension.w,
dimension.h,
dimension.c,
- IntegralDivideRoundUp(dimension.c, 4),
+ DivideRoundUp(dimension.c, 4),
// int4 dst_size
output_dimension.w,
output_dimension.h,
output_dimension.c,
- IntegralDivideRoundUp(output_dimension.c, 4),
+ DivideRoundUp(output_dimension.c, 4),
};
return GetByteBuffer(uniform_params);
}},
@@ -173,10 +173,10 @@
const uint3 groups_size{16, 16, 1};
const auto& src_shape = buffers.find(input_id)->second;
BHWC dst_shape = CalculateOutputShape(src_shape, attr);
- int groups_x = IntegralDivideRoundUp(dst_shape.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(dst_shape.h, groups_size.y);
- const int dst_layers = IntegralDivideRoundUp(dst_shape.c, 4);
- int groups_z = IntegralDivideRoundUp(dst_layers, groups_size.z);
+ int groups_x = DivideRoundUp(dst_shape.w, groups_size.x);
+ int groups_y = DivideRoundUp(dst_shape.h, groups_size.y);
+ const int dst_layers = DivideRoundUp(dst_shape.c, 4);
+ int groups_z = DivideRoundUp(dst_layers, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/softmax.cc b/tensorflow/lite/delegates/gpu/metal/kernels/softmax.cc
index 0ed2e06..0dfbbc8 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/softmax.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/softmax.cc
@@ -169,8 +169,8 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
uint3 groups_size{8, 4, 1};
const auto& dimension = buffers.find(output_id)->second;
- uint3 groups_count{IntegralDivideRoundUp(dimension.w, groups_size.x),
- IntegralDivideRoundUp(dimension.h, groups_size.y), 1};
+ uint3 groups_count{DivideRoundUp(dimension.w, groups_size.x),
+ DivideRoundUp(dimension.h, groups_size.y), 1};
return std::make_pair(groups_size, groups_count);
};
@@ -198,13 +198,13 @@
desc->uniform_buffers = {
{"constant uniforms& params",
[channels_count](const std::map<ValueId, BHWC>& buffers) {
- const int src_depth = IntegralDivideRoundUp(channels_count, 4);
+ const int src_depth = DivideRoundUp(channels_count, 4);
struct uniforms {
int4 size;
float4 mask;
};
uniforms params;
- params.size = {src_depth, IntegralDivideRoundUp(src_depth, 32), 1, 1};
+ params.size = {src_depth, DivideRoundUp(src_depth, 32), 1, 1};
params.mask = {0.0f, 0.0f, 0.0f, 0.0f};
const int reminder = channels_count % 4 == 0 ? 4 : channels_count % 4;
for (int i = 0; i < reminder; ++i) {
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/space_to_depth.cc b/tensorflow/lite/delegates/gpu/metal/kernels/space_to_depth.cc
index 3614174..4c11c43 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/space_to_depth.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/space_to_depth.cc
@@ -113,12 +113,12 @@
input_shape.h / attr.block_size,
input_shape.w / attr.block_size,
input_shape.c * attr.block_size * attr.block_size);
- const uint3 grid = uint3(output_shape.w, output_shape.h,
- IntegralDivideRoundUp(output_shape.c, 4));
+ const uint3 grid =
+ uint3(output_shape.w, output_shape.h, DivideRoundUp(output_shape.c, 4));
const uint3 groups_size = GetWorkGroupSizeForGrid(grid);
- const int groups_x = IntegralDivideRoundUp(grid.x, groups_size.x);
- const int groups_y = IntegralDivideRoundUp(grid.y, groups_size.y);
- const int groups_z = IntegralDivideRoundUp(grid.z, groups_size.z);
+ const int groups_x = DivideRoundUp(grid.x, groups_size.x);
+ const int groups_y = DivideRoundUp(grid.y, groups_size.y);
+ const int groups_z = DivideRoundUp(grid.z, groups_size.z);
return std::make_pair(groups_size, uint3(groups_x, groups_y, groups_z));
};
return {desc};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/transpose_conv.cc b/tensorflow/lite/delegates/gpu/metal/kernels/transpose_conv.cc
index 2282349..4a7f356 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/transpose_conv.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/transpose_conv.cc
@@ -130,8 +130,8 @@
constant_args, attr.padding.prepended.w, attr.padding.prepended.h,
attr.stride.w, attr.stride.h, kernel_x, kernel_y, inner_size_x,
inner_size_y, kernel_x - 1, kernel_y - 1);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
const int dst_channels_aligned = AlignByN(attr.weights.shape.o, 4);
return absl::Substitute(shader_source, src_depth * dst_channels_aligned,
src_depth, dst_depth, attr.weights.shape.o,
@@ -264,8 +264,8 @@
constant_args, attr.padding.prepended.w, attr.padding.prepended.h,
attr.stride.w, attr.stride.h, kernel_x, kernel_y, inner_size_x,
inner_size_y, kernel_x - 1, kernel_y - 1);
- const int src_depth = IntegralDivideRoundUp(attr.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(attr.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(attr.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(attr.weights.shape.o, 4);
const int dst_channels_aligned = AlignByN(attr.weights.shape.o, 4);
const int src_local_size_x = (workgroup_x + kernel_x) / attr.stride.w;
const int src_local_size_y = (workgroup_y + kernel_y) / attr.stride.h;
@@ -464,7 +464,7 @@
(kThreadGroupWidth + params.weights.shape.w) / params.stride.w;
const int src_local_size_y =
(kThreadGroupHeight + params.weights.shape.h) / params.stride.h;
- const int src_depth = IntegralDivideRoundUp(params.weights.shape.i, 4);
+ const int src_depth = DivideRoundUp(params.weights.shape.i, 4);
const int shared_size =
sizeof(float) * 4 * src_depth * src_local_size_x * src_local_size_y;
if (shared_size < 1000 * 16 &&
@@ -543,8 +543,8 @@
const uint3 groups_size{kThreadGroupWidth, kThreadGroupHeight, 1};
BHWC dst_shape =
CalculateOutputShape(buffers.find(input_id)->second, params);
- int groups_x = IntegralDivideRoundUp(dst_shape.w, groups_size.x);
- int groups_y = IntegralDivideRoundUp(dst_shape.h, groups_size.y);
+ int groups_x = DivideRoundUp(dst_shape.w, groups_size.x);
+ int groups_y = DivideRoundUp(dst_shape.h, groups_size.y);
int groups_z = 1;
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
@@ -556,8 +556,8 @@
int id, ValueId input_id, ValueId output_id,
const ConvolutionTransposedAttributes& params,
const DeviceInfo& device_info, const RuntimeOptions& options) {
- const int src_depth = IntegralDivideRoundUp(params.weights.shape.i, 4);
- const int dst_depth = IntegralDivideRoundUp(params.weights.shape.o, 4);
+ const int src_depth = DivideRoundUp(params.weights.shape.i, 4);
+ const int dst_depth = DivideRoundUp(params.weights.shape.o, 4);
const int kernel_x = 4;
const int kernel_y = 4;
@@ -645,7 +645,7 @@
[input_id, output_id, params](const std::map<ValueId, BHWC>& buffers) {
const auto& src_shape = buffers.find(input_id)->second;
const auto& dst_shape = buffers.find(output_id)->second;
- const int src_depth = IntegralDivideRoundUp(src_shape.c, 4);
+ const int src_depth = DivideRoundUp(src_shape.c, 4);
std::vector<int> uniform_params{
src_shape.w,
src_shape.h,
@@ -653,7 +653,7 @@
src_shape.w * src_shape.h,
dst_shape.w,
dst_shape.h,
- IntegralDivideRoundUp(dst_shape.c, 4),
+ DivideRoundUp(dst_shape.c, 4),
0,
4 * 16 * src_depth,
0,
@@ -667,13 +667,13 @@
desc->resize_function = [output_id, block_size,
params](const std::map<ValueId, BHWC>& buffers) {
const auto& dst_shape = buffers.find(output_id)->second;
- const int grid_x = IntegralDivideRoundUp(dst_shape.w + 2, 2 * block_size.x);
- const int grid_y = IntegralDivideRoundUp(dst_shape.h + 2, 2 * block_size.y);
- const int grid_z = IntegralDivideRoundUp(dst_shape.c, 4);
+ const int grid_x = DivideRoundUp(dst_shape.w + 2, 2 * block_size.x);
+ const int grid_y = DivideRoundUp(dst_shape.h + 2, 2 * block_size.y);
+ const int grid_z = DivideRoundUp(dst_shape.c, 4);
const uint3 group_size{8, 4, 1};
- int groups_x = IntegralDivideRoundUp(grid_x, group_size.x);
- int groups_y = IntegralDivideRoundUp(grid_y, group_size.y);
- int groups_z = IntegralDivideRoundUp(grid_z, group_size.z);
+ int groups_x = DivideRoundUp(grid_x, group_size.x);
+ int groups_y = DivideRoundUp(grid_y, group_size.y);
+ int groups_z = DivideRoundUp(grid_z, group_size.z);
return std::make_pair(group_size, uint3{groups_z, groups_x, groups_y});
};
diff --git a/tensorflow/lite/delegates/gpu/metal/kernels/winograd.cc b/tensorflow/lite/delegates/gpu/metal/kernels/winograd.cc
index 6d68e9e..2098155 100644
--- a/tensorflow/lite/delegates/gpu/metal/kernels/winograd.cc
+++ b/tensorflow/lite/delegates/gpu/metal/kernels/winograd.cc
@@ -486,8 +486,8 @@
BHWC dst_shape;
dst_shape.b = src_shape.b;
dst_shape.h = 36;
- dst_shape.w = IntegralDivideRoundUp(new_width, 4) *
- IntegralDivideRoundUp(new_height, 4);
+ dst_shape.w =
+ DivideRoundUp(new_width, 4) * DivideRoundUp(new_height, 4);
dst_shape.c = src_shape.c;
return dst_shape;
}};
@@ -501,16 +501,16 @@
attr.padding.appended.w - 2;
int new_height = src_shape.h + attr.padding.prepended.h +
attr.padding.appended.h - 2;
- int tiles_x = IntegralDivideRoundUp(new_width, 4);
- int tiles_y = IntegralDivideRoundUp(new_height, 4);
+ int tiles_x = DivideRoundUp(new_width, 4);
+ int tiles_y = DivideRoundUp(new_height, 4);
std::vector<int> sizes = {
src_shape.w,
src_shape.h,
- IntegralDivideRoundUp(src_shape.c, 4),
+ DivideRoundUp(src_shape.c, 4),
0,
dst_shape.w,
dst_shape.h,
- IntegralDivideRoundUp(dst_shape.c, 4),
+ DivideRoundUp(dst_shape.c, 4),
0,
-attr.padding.prepended.w,
-attr.padding.prepended.h,
@@ -529,12 +529,12 @@
src_shape.w + attr.padding.prepended.w + attr.padding.appended.w - 2;
int new_height =
src_shape.h + attr.padding.prepended.h + attr.padding.appended.h - 2;
- int grid_x = IntegralDivideRoundUp(new_width, 4);
- int grid_y = IntegralDivideRoundUp(new_height, 4);
- int grid_z = IntegralDivideRoundUp(src_shape.c, 4);
- int groups_x = IntegralDivideRoundUp(grid_x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid_y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid_z, groups_size.z);
+ int grid_x = DivideRoundUp(new_width, 4);
+ int grid_y = DivideRoundUp(new_height, 4);
+ int grid_z = DivideRoundUp(src_shape.c, 4);
+ int groups_x = DivideRoundUp(grid_x, groups_size.x);
+ int groups_y = DivideRoundUp(grid_y, groups_size.y);
+ int groups_z = DivideRoundUp(grid_z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
return {desc};
@@ -563,8 +563,8 @@
BHWC dst_shape;
dst_shape.b = src_shape.b;
dst_shape.h = 36;
- dst_shape.w = IntegralDivideRoundUp(new_width, 4) *
- IntegralDivideRoundUp(new_height, 4);
+ dst_shape.w =
+ DivideRoundUp(new_width, 4) * DivideRoundUp(new_height, 4);
dst_shape.c = src_shape.c;
return dst_shape;
}};
@@ -593,16 +593,16 @@
attr.padding.appended.w - 2;
int new_height = src_shape.h + attr.padding.prepended.h +
attr.padding.appended.h - 2;
- int tiles_x = IntegralDivideRoundUp(new_width, 4);
- int tiles_y = IntegralDivideRoundUp(new_height, 4);
+ int tiles_x = DivideRoundUp(new_width, 4);
+ int tiles_y = DivideRoundUp(new_height, 4);
std::vector<int> sizes = {
src_shape.w,
src_shape.h,
- IntegralDivideRoundUp(src_shape.c, 4),
+ DivideRoundUp(src_shape.c, 4),
0,
dst_shape.w,
dst_shape.h,
- IntegralDivideRoundUp(dst_shape.c, 4),
+ DivideRoundUp(dst_shape.c, 4),
0,
-attr.padding.prepended.w,
-attr.padding.prepended.h,
@@ -619,10 +619,10 @@
const auto& dst_shape = buffers.find(output_id)->second;
int grid_x = dst_shape.w;
int grid_y = 6;
- int grid_z = IntegralDivideRoundUp(dst_shape.c, 4);
- int groups_x = IntegralDivideRoundUp(grid_x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid_y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid_z, groups_size.z);
+ int grid_z = DivideRoundUp(dst_shape.c, 4);
+ int groups_x = DivideRoundUp(grid_x, groups_size.x);
+ int groups_y = DivideRoundUp(grid_y, groups_size.y);
+ int groups_z = DivideRoundUp(grid_z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
return {desc};
@@ -665,8 +665,8 @@
const auto& src_shape = buffers.find(input_id)->second;
const auto& dst_shape = buffers.find(output_id)->second;
std::vector<int> sizes = {
- src_shape.w, src_shape.h, IntegralDivideRoundUp(src_shape.c, 4), 0,
- dst_shape.w, dst_shape.h, IntegralDivideRoundUp(dst_shape.c, 4), 0,
+ src_shape.w, src_shape.h, DivideRoundUp(src_shape.c, 4), 0,
+ dst_shape.w, dst_shape.h, DivideRoundUp(dst_shape.c, 4), 0,
};
return GetByteBuffer(sizes);
}},
@@ -677,10 +677,10 @@
const auto& src_shape = buffers.find(input_id)->second;
int grid_x = src_shape.w;
int grid_y = 1;
- int grid_z = IntegralDivideRoundUp(src_shape.c, 4);
- int groups_x = IntegralDivideRoundUp(grid_x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid_y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid_z, groups_size.z);
+ int grid_z = DivideRoundUp(src_shape.c, 4);
+ int groups_x = DivideRoundUp(grid_x, groups_size.x);
+ int groups_y = DivideRoundUp(grid_y, groups_size.y);
+ int groups_z = DivideRoundUp(grid_z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
return {desc};
@@ -734,16 +734,16 @@
[input_id, output_id](const std::map<ValueId, BHWC>& buffers) {
const auto& src_shape = buffers.find(input_id)->second;
const auto& dst_shape = buffers.find(output_id)->second;
- const int tiles_x = IntegralDivideRoundUp(dst_shape.w, 4);
- const int tiles_y = IntegralDivideRoundUp(dst_shape.h, 4);
+ const int tiles_x = DivideRoundUp(dst_shape.w, 4);
+ const int tiles_y = DivideRoundUp(dst_shape.h, 4);
std::vector<int> sizes = {
src_shape.w,
src_shape.h,
- IntegralDivideRoundUp(src_shape.c, 4),
+ DivideRoundUp(src_shape.c, 4),
0,
dst_shape.w,
dst_shape.h,
- IntegralDivideRoundUp(dst_shape.c, 4),
+ DivideRoundUp(dst_shape.c, 4),
0,
tiles_x,
tiles_y,
@@ -757,14 +757,14 @@
desc->resize_function = [output_id](const std::map<ValueId, BHWC>& buffers) {
const uint3 groups_size{8, 4, 1};
const auto& dst_shape = buffers.find(output_id)->second;
- const int tiles_x = IntegralDivideRoundUp(dst_shape.w, 4);
- const int tiles_y = IntegralDivideRoundUp(dst_shape.h, 4);
+ const int tiles_x = DivideRoundUp(dst_shape.w, 4);
+ const int tiles_y = DivideRoundUp(dst_shape.h, 4);
int grid_x = tiles_x * tiles_y;
int grid_y = 4;
- int grid_z = IntegralDivideRoundUp(dst_shape.c, 4);
- int groups_x = IntegralDivideRoundUp(grid_x, groups_size.x);
- int groups_y = IntegralDivideRoundUp(grid_y, groups_size.y);
- int groups_z = IntegralDivideRoundUp(grid_z, groups_size.z);
+ int grid_z = DivideRoundUp(dst_shape.c, 4);
+ int groups_x = DivideRoundUp(grid_x, groups_size.x);
+ int groups_y = DivideRoundUp(grid_y, groups_size.y);
+ int groups_z = DivideRoundUp(grid_z, groups_size.z);
return std::make_pair(groups_size, uint3{groups_x, groups_y, groups_z});
};
return {desc};
diff --git a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
index 36d9da1..012da4a 100644
--- a/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
+++ b/tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
@@ -87,6 +87,16 @@
return "ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE";
case ANEURALNETWORKS_UNAVAILABLE_DEVICE:
return "ANEURALNETWORKS_UNAVAILABLE_DEVICE";
+ case ANEURALNETWORKS_MISSED_DEADLINE_TRANSIENT:
+ return "ANEURALNETWORKS_MISSED_DEADLINE_TRANSIENT";
+ case ANEURALNETWORKS_MISSED_DEADLINE_PERSISTENT:
+ return "ANEURALNETWORKS_MISSED_DEADLINE_PERSISTENT";
+ case ANEURALNETWORKS_RESOURCE_EXHAUSTED_TRANSIENT:
+ return "ANEURALNETWORKS_RESOURCE_EXHAUSTED_TRANSIENT";
+ case ANEURALNETWORKS_RESOURCE_EXHAUSTED_PERSISTENT:
+ return "ANEURALNETWORKS_RESOURCE_EXHAUSTED_PERSISTENT";
+ case ANEURALNETWORKS_DEAD_OBJECT:
+ return "ANEURALNETWORKS_DEAD_OBJECT";
default:
return "Unknown NNAPI error code: " + std::to_string(error_code);
}
@@ -1789,7 +1799,7 @@
" NNAPI only support float tanh.", &val_ctx);
} break;
case kTfLiteBuiltinSub: {
- ExpectMaxOpVersion(version, 2, &val_ctx);
+ ExpectMaxOpVersion(version, 3, &val_ctx);
const TfLiteType input_type =
context->tensors[node->inputs->data[0]].type;
Expect((android_sdk_version >= kMinSdkVersionForNNAPI11 &&
@@ -1798,6 +1808,13 @@
IsQuantized(input_type)),
NNAPIValidationFailureType::kUnsupportedInputType,
"NNAPI only support float sub.", &val_ctx);
+ const int input0_rank =
+ context->tensors[node->inputs->data[0]].dims->size;
+ const int input1_rank =
+ context->tensors[node->inputs->data[1]].dims->size;
+ Expect(input0_rank <= 4 && input1_rank <= 4,
+ NNAPIValidationFailureType::kUnsupportedOperandRank,
+ "Input rank must be <= 4", &val_ctx);
} break;
case kTfLiteBuiltinDiv: {
ExpectOpVersion(version, 1, &val_ctx);
@@ -2317,7 +2334,7 @@
"Unsupported operation type.", &val_ctx);
}
return val_ctx.is_valid;
-}
+} // NOLINT(readability/fn_size)
TfLiteStatus NNAPIDelegateKernel::Map(
TfLiteContext* context, int builtin_code, int version,
diff --git a/tensorflow/lite/delegates/xnnpack/BUILD b/tensorflow/lite/delegates/xnnpack/BUILD
index 2e0aa52..0e6e6c9 100644
--- a/tensorflow/lite/delegates/xnnpack/BUILD
+++ b/tensorflow/lite/delegates/xnnpack/BUILD
@@ -43,17 +43,20 @@
],
)
-############################## Integration tests ###############################
+################################ Tester classes ################################
cc_library(
- name = "test_main",
+ name = "binary_elementwise_tester",
testonly = 1,
- linkopts = select({
- "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
- "//conditions:default": [],
- }),
+ srcs = ["binary_elementwise_tester.cc"],
+ hdrs = ["binary_elementwise_tester.h"],
deps = [
- "@com_google_googletest//:gtest_main",
+ "//tensorflow/lite:framework",
+ "//tensorflow/lite:schema_fbs_version",
+ "//tensorflow/lite/kernels:builtin_ops",
+ "//tensorflow/lite/schema:schema_fbs",
+ "@com_google_googletest//:gtest",
+ "@flatbuffers",
],
)
@@ -87,15 +90,59 @@
],
)
-cc_test(
- name = "average_pool_2d_test",
- srcs = ["average_pool_2d_test.cc"],
+cc_library(
+ name = "softmax_tester",
+ testonly = 1,
+ srcs = ["softmax_tester.cc"],
+ hdrs = ["softmax_tester.h"],
+ deps = [
+ "//tensorflow/lite:framework",
+ "//tensorflow/lite:schema_fbs_version",
+ "//tensorflow/lite/kernels:builtin_ops",
+ "//tensorflow/lite/schema:schema_fbs",
+ "@com_google_googletest//:gtest",
+ "@flatbuffers",
+ ],
+)
+
+cc_library(
+ name = "unary_elementwise_tester",
+ testonly = 1,
+ srcs = ["unary_elementwise_tester.cc"],
+ hdrs = ["unary_elementwise_tester.h"],
+ deps = [
+ "//tensorflow/lite:framework",
+ "//tensorflow/lite:schema_fbs_version",
+ "//tensorflow/lite/kernels:builtin_ops",
+ "//tensorflow/lite/schema:schema_fbs",
+ "@com_google_googletest//:gtest",
+ "@flatbuffers",
+ ],
+)
+
+############################## Integration tests ###############################
+
+cc_library(
+ name = "test_main",
+ testonly = 1,
linkopts = select({
"//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
"//conditions:default": [],
}),
deps = [
- ":pool_2d_tester",
+ "@com_google_googletest//:gtest_main",
+ ],
+)
+
+cc_test(
+ name = "add_test",
+ srcs = ["add_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":binary_elementwise_tester",
":test_main",
":xnnpack_delegate_test_mode",
"@com_google_googletest//:gtest",
@@ -103,8 +150,8 @@
)
cc_test(
- name = "max_pool_2d_test",
- srcs = ["max_pool_2d_test.cc"],
+ name = "average_pool_2d_test",
+ srcs = ["average_pool_2d_test.cc"],
linkopts = select({
"//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
"//conditions:default": [],
@@ -152,4 +199,124 @@
],
)
+cc_test(
+ name = "hard_swish_test",
+ srcs = ["hard_swish_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":test_main",
+ ":unary_elementwise_tester",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "max_pool_2d_test",
+ srcs = ["max_pool_2d_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":pool_2d_tester",
+ ":test_main",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "logistic_test",
+ srcs = ["logistic_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":test_main",
+ ":unary_elementwise_tester",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "mul_test",
+ srcs = ["mul_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":binary_elementwise_tester",
+ ":test_main",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "relu_test",
+ srcs = ["relu_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":test_main",
+ ":unary_elementwise_tester",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "relu6_test",
+ srcs = ["relu6_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":test_main",
+ ":unary_elementwise_tester",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "relu_n1_to_1_test",
+ srcs = ["relu_n1_to_1_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":test_main",
+ ":unary_elementwise_tester",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
+cc_test(
+ name = "softmax_test",
+ srcs = ["softmax_test.cc"],
+ linkopts = select({
+ "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":softmax_tester",
+ ":test_main",
+ ":xnnpack_delegate_test_mode",
+ "@com_google_googletest//:gtest",
+ ],
+)
+
tflite_portable_test_suite_combined(combine_conditions = {"deps": [":test_main"]})
diff --git a/tensorflow/lite/delegates/xnnpack/add_test.cc b/tensorflow/lite/delegates/xnnpack/add_test.cc
new file mode 100644
index 0000000..dd2857e
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/add_test.cc
@@ -0,0 +1,811 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Add, 4DBy4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy4DBroadcastChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, 1, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, 1, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy4DBroadcastWidth) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, width, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, width, 1})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy4DBroadcastHeight) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, 1, 1})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy4DBroadcastBatch) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, 1, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, 1, 1, 1})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy4DBroadcastHeightWidthChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DBy0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DBy2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DBy1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DBy0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4DBroadcastChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, 1, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, 1, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4DBroadcastWidth) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, width, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, width, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4DBroadcastHeight) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, 1, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4DBroadcastBatch) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, 1, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, 1, 1, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic4DBroadcastHeightWidthChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 4DByStatic0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DByStatic2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DByStatic1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, 2DByStatic0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({})
+ .Input2Static(true)
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, ReluActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .ReluActivation()
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, Relu6Activation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Relu6Activation()
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, ReluMinus1To1Activation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .ReluMinus1To1Activation()
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, DISABLED_TanhActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .TanhActivation()
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, DISABLED_SignBitActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .SignBitActivation()
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+TEST(Add, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_ADD, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.cc b/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.cc
new file mode 100644
index 0000000..e846cbe
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.cc
@@ -0,0 +1,298 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h"
+
+#include <array>
+#include <cstdint>
+#include <functional>
+#include <numeric>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+std::vector<int32_t> BinaryElementwiseTester::OutputShape() const {
+ std::vector<int32_t> output_shape;
+ if (!input1_shape_.empty()) {
+ output_shape.insert(
+ output_shape.end(), input1_shape_.cbegin(),
+ input1_shape_.cbegin() +
+ std::max(input1_shape_.size(), input2_shape_.size()) -
+ input2_shape_.size());
+ }
+ if (!input2_shape_.empty()) {
+ output_shape.insert(
+ output_shape.end(), input2_shape_.cbegin(),
+ input2_shape_.cbegin() +
+ std::max(input2_shape_.size(), input1_shape_.size()) -
+ input1_shape_.size());
+ }
+ for (size_t i = std::min(input1_shape_.size(), input2_shape_.size()); i >= 1;
+ i--) {
+ output_shape.push_back(
+ std::max(*(input1_shape_.cend() - i), *(input2_shape_.cend() - i)));
+ }
+ return output_shape;
+}
+
+void BinaryElementwiseTester::Test(tflite::BuiltinOperator binary_op,
+ TfLiteDelegate* delegate) const {
+ if (Input1Static()) {
+ ASSERT_FALSE(Input2Static());
+ }
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ std::uniform_real_distribution<float> input1_distribution(-25.0f, 25.0f);
+ std::uniform_real_distribution<float> input2_distribution(-25.0f, 25.0f);
+ switch (binary_op) {
+ case BuiltinOperator_DIV:
+ input1_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ input2_distribution = std::uniform_real_distribution<float>(0.1f, 1.0f);
+ break;
+ case BuiltinOperator_MUL:
+ input1_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ input2_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ break;
+ default:
+ break;
+ }
+ auto input1_rng = std::bind(input1_distribution, std::ref(rng));
+ auto input2_rng = std::bind(input2_distribution, std::ref(rng));
+
+ std::vector<char> buffer = CreateTfLiteModel(binary_op);
+ const Model* model = GetModel(buffer.data());
+
+ std::unique_ptr<Interpreter> delegate_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &delegate_interpreter),
+ kTfLiteOk);
+ std::unique_ptr<Interpreter> default_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &default_interpreter),
+ kTfLiteOk);
+
+ ASSERT_TRUE(delegate_interpreter);
+ ASSERT_TRUE(default_interpreter);
+
+ if (Input1Static() || Input2Static()) {
+ ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
+ ASSERT_EQ(default_interpreter->inputs().size(), 1);
+ } else {
+ ASSERT_EQ(delegate_interpreter->inputs().size(), 2);
+ ASSERT_EQ(default_interpreter->inputs().size(), 2);
+ }
+
+ ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
+ ASSERT_EQ(default_interpreter->outputs().size(), 1);
+
+ ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
+ ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
+
+ ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
+
+ if (!Input1Static()) {
+ float* default_input1_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->inputs()[0]);
+ std::generate(default_input1_data,
+ default_input1_data + ComputeSize(Input1Shape()),
+ std::ref(input1_rng));
+
+ float* xnnpack_input1_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->inputs()[0]);
+ std::copy(default_input1_data,
+ default_input1_data + ComputeSize(Input1Shape()),
+ xnnpack_input1_data);
+ }
+
+ if (!Input2Static()) {
+ float* default_input2_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->inputs()[Input1Static() ? 0 : 1]);
+ std::generate(default_input2_data,
+ default_input2_data + ComputeSize(Input2Shape()),
+ std::ref(input2_rng));
+
+ float* xnnpack_input2_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->inputs()[Input1Static() ? 0 : 1]);
+ std::copy(default_input2_data,
+ default_input2_data + ComputeSize(Input2Shape()),
+ xnnpack_input2_data);
+ }
+
+ ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
+ ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
+
+ float* default_output_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->outputs()[0]);
+ float* xnnpack_output_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->outputs()[0]);
+
+ for (size_t i = 0; i < ComputeSize(OutputShape()); i++) {
+ ASSERT_NEAR(default_output_data[i], xnnpack_output_data[i],
+ std::numeric_limits<float>::epsilon() *
+ std::max(std::abs(default_output_data[i]) * 2.0f, 1.0f));
+ }
+}
+
+std::vector<char> BinaryElementwiseTester::CreateTfLiteModel(
+ tflite::BuiltinOperator binary_op) const {
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ std::uniform_real_distribution<float> input1_distribution(-25.0f, 25.0f);
+ std::uniform_real_distribution<float> input2_distribution(-25.0f, 25.0f);
+ switch (binary_op) {
+ case BuiltinOperator_DIV:
+ input1_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ input2_distribution = std::uniform_real_distribution<float>(0.1f, 1.0f);
+ break;
+ case BuiltinOperator_MUL:
+ input1_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ input2_distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f);
+ break;
+ default:
+ break;
+ }
+ auto input1_rng = std::bind(input1_distribution, std::ref(rng));
+ auto input2_rng = std::bind(input2_distribution, std::ref(rng));
+
+ flatbuffers::FlatBufferBuilder builder;
+ flatbuffers::Offset<OperatorCode> operator_code =
+ CreateOperatorCode(builder, binary_op);
+
+ std::vector<flatbuffers::Offset<Buffer>> buffers{{
+ CreateBuffer(builder, builder.CreateVector({})),
+ }};
+
+ int32_t input1_buffer = 0;
+ if (Input1Static()) {
+ std::vector<float> input1_data(ComputeSize(Input1Shape()));
+ std::generate(input1_data.begin(), input1_data.end(), input1_rng);
+
+ input1_buffer = buffers.size();
+ buffers.push_back(CreateBuffer(
+ builder, builder.CreateVector(
+ reinterpret_cast<const uint8_t*>(input1_data.data()),
+ sizeof(float) * input1_data.size())));
+ }
+
+ int32_t input2_buffer = 0;
+ if (Input2Static()) {
+ std::vector<float> input2_data(ComputeSize(Input2Shape()));
+ std::generate(input2_data.begin(), input2_data.end(), input2_rng);
+
+ input2_buffer = buffers.size();
+ buffers.push_back(CreateBuffer(
+ builder, builder.CreateVector(
+ reinterpret_cast<const uint8_t*>(input2_data.data()),
+ sizeof(float) * input2_data.size())));
+ }
+
+ const std::vector<int32_t> output_shape = OutputShape();
+ const std::array<flatbuffers::Offset<Tensor>, 3> tensors{{
+ CreateTensor(builder,
+ builder.CreateVector<int32_t>(Input1Shape().data(),
+ Input1Shape().size()),
+ TensorType_FLOAT32, input1_buffer),
+ CreateTensor(builder,
+ builder.CreateVector<int32_t>(Input2Shape().data(),
+ Input2Shape().size()),
+ TensorType_FLOAT32, input2_buffer),
+ CreateTensor(builder,
+ builder.CreateVector<int32_t>(output_shape.data(),
+ output_shape.size()),
+ TensorType_FLOAT32),
+ }};
+
+ tflite::BuiltinOptions builtin_options_type = tflite::BuiltinOptions_NONE;
+ flatbuffers::Offset<void> builtin_options = 0;
+ switch (binary_op) {
+ case BuiltinOperator_ADD:
+ builtin_options_type = BuiltinOptions_AddOptions;
+ builtin_options = CreateAddOptions(builder, Activation()).Union();
+ break;
+ case BuiltinOperator_DIV:
+ builtin_options_type = BuiltinOptions_DivOptions;
+ builtin_options = CreateDivOptions(builder, Activation()).Union();
+ break;
+ case BuiltinOperator_MUL:
+ builtin_options_type = BuiltinOptions_MulOptions;
+ builtin_options = CreateMulOptions(builder, Activation()).Union();
+ break;
+ case BuiltinOperator_SUB:
+ builtin_options_type = BuiltinOptions_SubOptions;
+ builtin_options = CreateSubOptions(builder, Activation()).Union();
+ break;
+ default:
+ EXPECT_EQ(Activation(), ActivationFunctionType_NONE);
+ }
+
+ const std::array<int32_t, 2> op_inputs{{0, 1}};
+ const std::array<int32_t, 1> op_outputs{{2}};
+ flatbuffers::Offset<Operator> op = CreateOperator(
+ builder, /*opcode_index=*/0,
+ builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
+ builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()),
+ builtin_options_type, builtin_options);
+
+ std::vector<int32_t> subgraph_inputs;
+ if (!Input1Static()) {
+ subgraph_inputs.push_back(0);
+ }
+ if (!Input2Static()) {
+ subgraph_inputs.push_back(1);
+ }
+ const std::array<int32_t, 1> subgraph_outputs{{2}};
+ flatbuffers::Offset<SubGraph> subgraph = CreateSubGraph(
+ builder, builder.CreateVector(tensors.data(), tensors.size()),
+ builder.CreateVector<int32_t>(subgraph_inputs.data(),
+ subgraph_inputs.size()),
+ builder.CreateVector<int32_t>(subgraph_outputs.data(),
+ subgraph_outputs.size()),
+ builder.CreateVector(&op, 1));
+
+ flatbuffers::Offset<flatbuffers::String> description =
+ builder.CreateString("Binary operator model");
+
+ flatbuffers::Offset<Model> model_buffer = CreateModel(
+ builder, TFLITE_SCHEMA_VERSION, builder.CreateVector(&operator_code, 1),
+ builder.CreateVector(&subgraph, 1), description,
+ builder.CreateVector(buffers.data(), buffers.size()));
+
+ builder.Finish(model_buffer);
+
+ return std::vector<char>(builder.GetBufferPointer(),
+ builder.GetBufferPointer() + builder.GetSize());
+}
+
+int32_t BinaryElementwiseTester::ComputeSize(
+ const std::vector<int32_t>& shape) {
+ return std::accumulate(shape.cbegin(), shape.cend(), 1,
+ std::multiplies<int32_t>());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h b/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h
new file mode 100644
index 0000000..6d9a8b6
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h
@@ -0,0 +1,130 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_DELEGATES_XNNPACK_BINARY_ELEMENTWISE_TESTER_H_
+#define TENSORFLOW_LITE_DELEGATES_XNNPACK_BINARY_ELEMENTWISE_TESTER_H_
+
+#include <cstdint>
+#include <functional>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+class BinaryElementwiseTester {
+ public:
+ BinaryElementwiseTester() = default;
+ BinaryElementwiseTester(const BinaryElementwiseTester&) = delete;
+ BinaryElementwiseTester& operator=(const BinaryElementwiseTester&) = delete;
+
+ inline BinaryElementwiseTester& Input1Shape(
+ std::initializer_list<int32_t> shape) {
+ for (auto it = shape.begin(); it != shape.end(); ++it) {
+ EXPECT_GT(*it, 0);
+ }
+ input1_shape_ = std::vector<int32_t>(shape.begin(), shape.end());
+ return *this;
+ }
+
+ inline const std::vector<int32_t>& Input1Shape() const {
+ return input1_shape_;
+ }
+
+ inline BinaryElementwiseTester& Input2Shape(
+ std::initializer_list<int32_t> shape) {
+ for (auto it = shape.begin(); it != shape.end(); ++it) {
+ EXPECT_GT(*it, 0);
+ }
+ input2_shape_ = std::vector<int32_t>(shape.begin(), shape.end());
+ return *this;
+ }
+
+ inline const std::vector<int32_t>& Input2Shape() const {
+ return input2_shape_;
+ }
+
+ std::vector<int32_t> OutputShape() const;
+
+ inline BinaryElementwiseTester& Input1Static(bool is_static) {
+ input1_static_ = is_static;
+ return *this;
+ }
+
+ inline bool Input1Static() const { return input1_static_; }
+
+ inline BinaryElementwiseTester& Input2Static(bool is_static) {
+ input2_static_ = is_static;
+ return *this;
+ }
+
+ inline bool Input2Static() const { return input2_static_; }
+
+ inline BinaryElementwiseTester& ReluActivation() {
+ activation_ = ::tflite::ActivationFunctionType_RELU;
+ return *this;
+ }
+
+ inline BinaryElementwiseTester& Relu6Activation() {
+ activation_ = ::tflite::ActivationFunctionType_RELU6;
+ return *this;
+ }
+
+ inline BinaryElementwiseTester& ReluMinus1To1Activation() {
+ activation_ = ::tflite::ActivationFunctionType_RELU_N1_TO_1;
+ return *this;
+ }
+
+ inline BinaryElementwiseTester& TanhActivation() {
+ activation_ = ::tflite::ActivationFunctionType_TANH;
+ return *this;
+ }
+
+ inline BinaryElementwiseTester& SignBitActivation() {
+ activation_ = ::tflite::ActivationFunctionType_SIGN_BIT;
+ return *this;
+ }
+
+ void Test(tflite::BuiltinOperator binary_op, TfLiteDelegate* delegate) const;
+
+ private:
+ std::vector<char> CreateTfLiteModel(tflite::BuiltinOperator binary_op) const;
+
+ inline ::tflite::ActivationFunctionType Activation() const {
+ return activation_;
+ }
+
+ static int32_t ComputeSize(const std::vector<int32_t>& shape);
+
+ std::vector<int32_t> input1_shape_;
+ std::vector<int32_t> input2_shape_;
+ bool input1_static_ = false;
+ bool input2_static_ = false;
+ ::tflite::ActivationFunctionType activation_ =
+ ::tflite::ActivationFunctionType_NONE;
+};
+
+} // namespace xnnpack
+} // namespace tflite
+
+#endif // TENSORFLOW_LITE_DELEGATES_XNNPACK_BINARY_ELEMENTWISE_TESTER_H_
diff --git a/tensorflow/lite/delegates/xnnpack/hard_swish_test.cc b/tensorflow/lite/delegates/xnnpack/hard_swish_test.cc
new file mode 100644
index 0000000..efc0d60
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/hard_swish_test.cc
@@ -0,0 +1,120 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(HardSwish, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_HARD_SWISH, xnnpack_delegate.get());
+}
+
+TEST(HardSwish, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, width, channels})
+ .Test(BuiltinOperator_HARD_SWISH, xnnpack_delegate.get());
+}
+
+TEST(HardSwish, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, channels})
+ .Test(BuiltinOperator_HARD_SWISH, xnnpack_delegate.get());
+}
+
+TEST(HardSwish, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ UnaryElementwiseTester().Shape({batch}).Test(BuiltinOperator_HARD_SWISH,
+ xnnpack_delegate.get());
+}
+
+TEST(HardSwish, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_HARD_SWISH, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/logistic_test.cc b/tensorflow/lite/delegates/xnnpack/logistic_test.cc
new file mode 100644
index 0000000..db87464
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/logistic_test.cc
@@ -0,0 +1,120 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Logistic, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_LOGISTIC, xnnpack_delegate.get());
+}
+
+TEST(Logistic, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, width, channels})
+ .Test(BuiltinOperator_LOGISTIC, xnnpack_delegate.get());
+}
+
+TEST(Logistic, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, channels})
+ .Test(BuiltinOperator_LOGISTIC, xnnpack_delegate.get());
+}
+
+TEST(Logistic, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ UnaryElementwiseTester().Shape({batch}).Test(BuiltinOperator_LOGISTIC,
+ xnnpack_delegate.get());
+}
+
+TEST(Logistic, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_LOGISTIC, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/mul_test.cc b/tensorflow/lite/delegates/xnnpack/mul_test.cc
new file mode 100644
index 0000000..6c0475e
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/mul_test.cc
@@ -0,0 +1,811 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/binary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Mul, 4DBy4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy4DBroadcastChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, 1, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, 1, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy4DBroadcastWidth) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, width, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, width, 1})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy4DBroadcastHeight) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, 1, 1})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy4DBroadcastBatch) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, 1, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, 1, 1, 1})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy4DBroadcastHeightWidthChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DBy0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DBy2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DBy1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DBy0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4DBroadcastChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, 1, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, 1, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4DBroadcastWidth) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, 1, width, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, 1, width, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4DBroadcastHeight) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, 1, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4DBroadcastBatch) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, 1, 1, 1})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, 1, 1, 1})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic4DBroadcastHeightWidthChannels) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({1, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({1, height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({height, width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({width, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 4DByStatic0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, height, width, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DByStatic2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({batch, channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DByStatic1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({channels})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({channels})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, 2DByStatic0D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({})
+ .Input2Shape({batch, channels})
+ .Input1Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, channels})
+ .Input2Shape({})
+ .Input2Static(true)
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, ReluActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .ReluActivation()
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, Relu6Activation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Relu6Activation()
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, ReluMinus1To1Activation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .ReluMinus1To1Activation()
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, DISABLED_TanhActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .TanhActivation()
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, DISABLED_SignBitActivation) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .SignBitActivation()
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+TEST(Mul, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ BinaryElementwiseTester()
+ .Input1Shape({batch, height, width, channels})
+ .Input2Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_MUL, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/relu6_test.cc b/tensorflow/lite/delegates/xnnpack/relu6_test.cc
new file mode 100644
index 0000000..75f32dc
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/relu6_test.cc
@@ -0,0 +1,120 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Relu6, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU6, xnnpack_delegate.get());
+}
+
+TEST(Relu6, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, width, channels})
+ .Test(BuiltinOperator_RELU6, xnnpack_delegate.get());
+}
+
+TEST(Relu6, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, channels})
+ .Test(BuiltinOperator_RELU6, xnnpack_delegate.get());
+}
+
+TEST(Relu6, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ UnaryElementwiseTester().Shape({batch}).Test(BuiltinOperator_RELU6,
+ xnnpack_delegate.get());
+}
+
+TEST(Relu6, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU6, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/relu_n1_to_1_test.cc b/tensorflow/lite/delegates/xnnpack/relu_n1_to_1_test.cc
new file mode 100644
index 0000000..9e79957
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/relu_n1_to_1_test.cc
@@ -0,0 +1,120 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(ReluMinus1To1, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU_N1_TO_1, xnnpack_delegate.get());
+}
+
+TEST(ReluMinus1To1, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, width, channels})
+ .Test(BuiltinOperator_RELU_N1_TO_1, xnnpack_delegate.get());
+}
+
+TEST(ReluMinus1To1, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, channels})
+ .Test(BuiltinOperator_RELU_N1_TO_1, xnnpack_delegate.get());
+}
+
+TEST(ReluMinus1To1, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ UnaryElementwiseTester().Shape({batch}).Test(BuiltinOperator_RELU_N1_TO_1,
+ xnnpack_delegate.get());
+}
+
+TEST(ReluMinus1To1, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU_N1_TO_1, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/relu_test.cc b/tensorflow/lite/delegates/xnnpack/relu_test.cc
new file mode 100644
index 0000000..8996ff5
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/relu_test.cc
@@ -0,0 +1,120 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Relu, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU, xnnpack_delegate.get());
+}
+
+TEST(Relu, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, width, channels})
+ .Test(BuiltinOperator_RELU, xnnpack_delegate.get());
+}
+
+TEST(Relu, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, channels})
+ .Test(BuiltinOperator_RELU, xnnpack_delegate.get());
+}
+
+TEST(Relu, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ UnaryElementwiseTester().Shape({batch}).Test(BuiltinOperator_RELU,
+ xnnpack_delegate.get());
+}
+
+TEST(Relu, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ UnaryElementwiseTester()
+ .Shape({batch, height, width, channels})
+ .Test(BuiltinOperator_RELU, xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/softmax_test.cc b/tensorflow/lite/delegates/xnnpack/softmax_test.cc
new file mode 100644
index 0000000..ae33a1a
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/softmax_test.cc
@@ -0,0 +1,140 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/softmax_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(Softmax, 4D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ SoftmaxTester()
+ .Shape({batch, height, width, channels})
+ .Test(xnnpack_delegate.get());
+}
+
+TEST(Softmax, 3D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ SoftmaxTester().Shape({batch, width, channels}).Test(xnnpack_delegate.get());
+}
+
+TEST(Softmax, 2D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto channels = shape_rng();
+
+ SoftmaxTester().Shape({batch, channels}).Test(xnnpack_delegate.get());
+}
+
+TEST(Softmax, 1D) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+
+ SoftmaxTester().Shape({batch}).Test(xnnpack_delegate.get());
+}
+
+TEST(Softmax, DISABLED_Beta) {
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ SoftmaxTester()
+ .Shape({batch, height, width, channels})
+ .Beta(0.1f)
+ .Test(xnnpack_delegate.get());
+
+ SoftmaxTester()
+ .Shape({batch, height, width, channels})
+ .Beta(10.0f)
+ .Test(xnnpack_delegate.get());
+}
+
+TEST(Softmax, MultiThreading) {
+ TfLiteXNNPackDelegateOptions delegate_options =
+ TfLiteXNNPackDelegateOptionsDefault();
+ delegate_options.num_threads = 2;
+ std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+ xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+ TfLiteXNNPackDelegateDelete);
+
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto shape_rng =
+ std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+ const auto batch = shape_rng();
+ const auto height = shape_rng();
+ const auto width = shape_rng();
+ const auto channels = shape_rng();
+
+ SoftmaxTester()
+ .Shape({batch, height, width, channels})
+ .Test(xnnpack_delegate.get());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/softmax_tester.cc b/tensorflow/lite/delegates/xnnpack/softmax_tester.cc
new file mode 100644
index 0000000..c93aa0d
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/softmax_tester.cc
@@ -0,0 +1,156 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/delegates/xnnpack/softmax_tester.h"
+
+#include <array>
+#include <cstdint>
+#include <functional>
+#include <numeric>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+void SoftmaxTester::Test(TfLiteDelegate* delegate) const {
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto input_rng = std::bind(
+ std::uniform_real_distribution<float>(-15.0f, 15.0f), std::ref(rng));
+
+ std::vector<char> buffer = CreateTfLiteModel();
+ const Model* model = GetModel(buffer.data());
+
+ std::unique_ptr<Interpreter> delegate_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &delegate_interpreter),
+ kTfLiteOk);
+ std::unique_ptr<Interpreter> default_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &default_interpreter),
+ kTfLiteOk);
+
+ ASSERT_TRUE(delegate_interpreter);
+ ASSERT_TRUE(default_interpreter);
+
+ ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
+ ASSERT_EQ(default_interpreter->inputs().size(), 1);
+
+ ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
+ ASSERT_EQ(default_interpreter->outputs().size(), 1);
+
+ ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
+ ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
+
+ ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
+
+ float* default_input_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->inputs()[0]);
+ std::generate(default_input_data, default_input_data + Size(),
+ std::ref(input_rng));
+
+ float* delegate_input_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->inputs()[0]);
+ std::copy(default_input_data, default_input_data + Size(),
+ delegate_input_data);
+
+ ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
+ ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
+
+ float* default_output_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->outputs()[0]);
+ float* delegate_output_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->outputs()[0]);
+
+ for (size_t i = 0; i < Size(); i++) {
+ ASSERT_NEAR(default_output_data[i], delegate_output_data[i],
+ std::numeric_limits<float>::epsilon() *
+ std::max(std::abs(default_output_data[i]) * 10.0f, 1.0f));
+ }
+}
+
+std::vector<char> SoftmaxTester::CreateTfLiteModel() const {
+ flatbuffers::FlatBufferBuilder builder;
+ flatbuffers::Offset<OperatorCode> operator_code =
+ CreateOperatorCode(builder, BuiltinOperator_SOFTMAX);
+
+ const std::array<flatbuffers::Offset<Buffer>, 1> buffers{{
+ CreateBuffer(builder, builder.CreateVector({})),
+ }};
+
+ const std::array<flatbuffers::Offset<Tensor>, 2> tensors{{
+ CreateTensor(
+ builder,
+ builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+ TensorType_FLOAT32),
+ CreateTensor(
+ builder,
+ builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+ TensorType_FLOAT32),
+ }};
+
+ flatbuffers::Offset<SoftmaxOptions> softmax_options =
+ CreateSoftmaxOptions(builder, Beta());
+
+ const std::array<int32_t, 1> op_inputs{{0}};
+ const std::array<int32_t, 1> op_outputs{{1}};
+ flatbuffers::Offset<Operator> op = CreateOperator(
+ builder, /*opcode_index=*/0,
+ builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
+ builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()),
+ BuiltinOptions_SoftmaxOptions, softmax_options.Union());
+
+ const std::array<int32_t, 1> subgraph_inputs{{0}};
+ const std::array<int32_t, 1> subgraph_outputs{{1}};
+ flatbuffers::Offset<SubGraph> subgraph = CreateSubGraph(
+ builder, builder.CreateVector(tensors.data(), tensors.size()),
+ builder.CreateVector<int32_t>(subgraph_inputs.data(),
+ subgraph_inputs.size()),
+ builder.CreateVector<int32_t>(subgraph_outputs.data(),
+ subgraph_outputs.size()),
+ builder.CreateVector(&op, 1));
+
+ flatbuffers::Offset<flatbuffers::String> description =
+ builder.CreateString("Softmax model");
+
+ flatbuffers::Offset<Model> model_buffer = CreateModel(
+ builder, TFLITE_SCHEMA_VERSION, builder.CreateVector(&operator_code, 1),
+ builder.CreateVector(&subgraph, 1), description,
+ builder.CreateVector(buffers.data(), buffers.size()));
+
+ builder.Finish(model_buffer);
+
+ return std::vector<char>(builder.GetBufferPointer(),
+ builder.GetBufferPointer() + builder.GetSize());
+}
+
+int32_t SoftmaxTester::ComputeSize(const std::vector<int32_t>& shape) {
+ return std::accumulate(shape.cbegin(), shape.cend(), 1,
+ std::multiplies<int32_t>());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/softmax_tester.h b/tensorflow/lite/delegates/xnnpack/softmax_tester.h
new file mode 100644
index 0000000..9f930a6
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/softmax_tester.h
@@ -0,0 +1,77 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_DELEGATES_XNNPACK_SOFTMAX_TESTER_H_
+#define TENSORFLOW_LITE_DELEGATES_XNNPACK_SOFTMAX_TESTER_H_
+
+#include <cstdint>
+#include <functional>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+class SoftmaxTester {
+ public:
+ SoftmaxTester() = default;
+ SoftmaxTester(const SoftmaxTester&) = delete;
+ SoftmaxTester& operator=(const SoftmaxTester&) = delete;
+
+ inline SoftmaxTester& Shape(std::initializer_list<int32_t> shape) {
+ EXPECT_GT(shape.size(), 0);
+ for (auto it = shape.begin(); it != shape.end(); ++it) {
+ EXPECT_GT(*it, 0);
+ }
+ shape_ = std::vector<int32_t>(shape.begin(), shape.end());
+ size_ = SoftmaxTester::ComputeSize(shape_);
+ return *this;
+ }
+
+ const std::vector<int32_t>& Shape() const { return shape_; }
+
+ int32_t Size() const { return size_; }
+
+ inline SoftmaxTester& Beta(float beta) {
+ beta_ = beta;
+ return *this;
+ }
+
+ float Beta() const { return beta_; }
+
+ void Test(TfLiteDelegate* delegate) const;
+
+ private:
+ std::vector<char> CreateTfLiteModel() const;
+
+ static int32_t ComputeSize(const std::vector<int32_t>& shape);
+
+ std::vector<int32_t> shape_;
+ int32_t size_;
+ float beta_ = 1.0f;
+};
+
+} // namespace xnnpack
+} // namespace tflite
+
+#endif // TENSORFLOW_LITE_DELEGATES_XNNPACK_SOFTMAX_TESTER_H_
diff --git a/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.cc b/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.cc
new file mode 100644
index 0000000..30897c7
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.cc
@@ -0,0 +1,154 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h"
+
+#include <array>
+#include <cstdint>
+#include <functional>
+#include <numeric>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+void UnaryElementwiseTester::Test(tflite::BuiltinOperator unary_op,
+ TfLiteDelegate* delegate) const {
+ std::random_device random_device;
+ auto rng = std::mt19937(random_device());
+ auto input_rng = std::bind(
+ std::uniform_real_distribution<float>(-15.0f, 15.0f), std::ref(rng));
+
+ std::vector<char> buffer = CreateTfLiteModel(unary_op);
+ const Model* model = GetModel(buffer.data());
+
+ std::unique_ptr<Interpreter> delegate_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &delegate_interpreter),
+ kTfLiteOk);
+ std::unique_ptr<Interpreter> default_interpreter;
+ ASSERT_EQ(
+ InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+ &default_interpreter),
+ kTfLiteOk);
+
+ ASSERT_TRUE(delegate_interpreter);
+ ASSERT_TRUE(default_interpreter);
+
+ ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
+ ASSERT_EQ(default_interpreter->inputs().size(), 1);
+
+ ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
+ ASSERT_EQ(default_interpreter->outputs().size(), 1);
+
+ ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
+ ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
+
+ ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
+
+ float* default_input_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->inputs()[0]);
+ std::generate(default_input_data, default_input_data + Size(),
+ std::ref(input_rng));
+
+ float* delegate_input_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->inputs()[0]);
+ std::copy(default_input_data, default_input_data + Size(),
+ delegate_input_data);
+
+ ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
+ ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
+
+ float* default_output_data = default_interpreter->typed_tensor<float>(
+ default_interpreter->outputs()[0]);
+ float* delegate_output_data = delegate_interpreter->typed_tensor<float>(
+ delegate_interpreter->outputs()[0]);
+
+ for (size_t i = 0; i < Size(); i++) {
+ ASSERT_NEAR(default_output_data[i], delegate_output_data[i],
+ std::numeric_limits<float>::epsilon() *
+ std::max(std::abs(default_output_data[i]) * 10.0f, 1.0f));
+ }
+}
+
+std::vector<char> UnaryElementwiseTester::CreateTfLiteModel(
+ tflite::BuiltinOperator unary_op) const {
+ flatbuffers::FlatBufferBuilder builder;
+ flatbuffers::Offset<OperatorCode> operator_code =
+ CreateOperatorCode(builder, unary_op);
+
+ const std::array<flatbuffers::Offset<Buffer>, 1> buffers{{
+ CreateBuffer(builder, builder.CreateVector({})),
+ }};
+
+ const std::array<flatbuffers::Offset<Tensor>, 2> tensors{{
+ CreateTensor(
+ builder,
+ builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+ TensorType_FLOAT32),
+ CreateTensor(
+ builder,
+ builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+ TensorType_FLOAT32),
+ }};
+
+ const std::array<int32_t, 1> op_inputs{{0}};
+ const std::array<int32_t, 1> op_outputs{{1}};
+ flatbuffers::Offset<Operator> op = CreateOperator(
+ builder, /*opcode_index=*/0,
+ builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
+ builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()));
+
+ const std::array<int32_t, 1> subgraph_inputs{{0}};
+ const std::array<int32_t, 1> subgraph_outputs{{1}};
+ flatbuffers::Offset<SubGraph> subgraph = CreateSubGraph(
+ builder, builder.CreateVector(tensors.data(), tensors.size()),
+ builder.CreateVector<int32_t>(subgraph_inputs.data(),
+ subgraph_inputs.size()),
+ builder.CreateVector<int32_t>(subgraph_outputs.data(),
+ subgraph_outputs.size()),
+ builder.CreateVector(&op, 1));
+
+ flatbuffers::Offset<flatbuffers::String> description =
+ builder.CreateString("Unary operator model");
+
+ flatbuffers::Offset<Model> model_buffer = CreateModel(
+ builder, TFLITE_SCHEMA_VERSION, builder.CreateVector(&operator_code, 1),
+ builder.CreateVector(&subgraph, 1), description,
+ builder.CreateVector(buffers.data(), buffers.size()));
+
+ builder.Finish(model_buffer);
+
+ return std::vector<char>(builder.GetBufferPointer(),
+ builder.GetBufferPointer() + builder.GetSize());
+}
+
+int32_t UnaryElementwiseTester::ComputeSize(const std::vector<int32_t>& shape) {
+ return std::accumulate(shape.cbegin(), shape.cend(), 1,
+ std::multiplies<int32_t>());
+}
+
+} // namespace xnnpack
+} // namespace tflite
diff --git a/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h b/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h
new file mode 100644
index 0000000..c2c9e44
--- /dev/null
+++ b/tensorflow/lite/delegates/xnnpack/unary_elementwise_tester.h
@@ -0,0 +1,68 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_DELEGATES_XNNPACK_UNARY_ELEMENTWISE_TESTER_H_
+#define TENSORFLOW_LITE_DELEGATES_XNNPACK_UNARY_ELEMENTWISE_TESTER_H_
+
+#include <cstdint>
+#include <functional>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h" // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+class UnaryElementwiseTester {
+ public:
+ UnaryElementwiseTester() = default;
+ UnaryElementwiseTester(const UnaryElementwiseTester&) = delete;
+ UnaryElementwiseTester& operator=(const UnaryElementwiseTester&) = delete;
+
+ inline UnaryElementwiseTester& Shape(std::initializer_list<int32_t> shape) {
+ for (auto it = shape.begin(); it != shape.end(); ++it) {
+ EXPECT_GT(*it, 0);
+ }
+ shape_ = std::vector<int32_t>(shape.begin(), shape.end());
+ size_ = UnaryElementwiseTester::ComputeSize(shape_);
+ return *this;
+ }
+
+ const std::vector<int32_t>& Shape() const { return shape_; }
+
+ int32_t Size() const { return size_; }
+
+ void Test(tflite::BuiltinOperator unary_op, TfLiteDelegate* delegate) const;
+
+ private:
+ std::vector<char> CreateTfLiteModel(tflite::BuiltinOperator unary_op) const;
+
+ static int32_t ComputeSize(const std::vector<int32_t>& shape);
+
+ std::vector<int32_t> shape_;
+ int32_t size_;
+};
+
+} // namespace xnnpack
+} // namespace tflite
+
+#endif // TENSORFLOW_LITE_DELEGATES_XNNPACK_UNARY_ELEMENTWISE_TESTER_H_
diff --git a/tensorflow/lite/experimental/delegates/coreml/builders/test_util.mm b/tensorflow/lite/experimental/delegates/coreml/builders/test_util.mm
index 465ecce..a57e766 100644
--- a/tensorflow/lite/experimental/delegates/coreml/builders/test_util.mm
+++ b/tensorflow/lite/experimental/delegates/coreml/builders/test_util.mm
@@ -27,7 +27,7 @@
delegate_ptr, [](TfLiteDelegate* delegate) { TfLiteCoreMlDelegateDelete(delegate); });
// Add delegate.
// TODO(karimnosseir): This doesn't actually make the test fail, switch to something else.
- ASSERT_TRUE(interpreter_->ModifyGraphWithDelegate(delegate_.get()) != kTfLiteError);
+ ASSERT_TRUE(interpreter_->ModifyGraphWithDelegate(delegate_.get()) == kTfLiteOk);
Invoke();
}
diff --git a/tensorflow/lite/experimental/delegates/hexagon/README.md b/tensorflow/lite/experimental/delegates/hexagon/README.md
index 6ad7d30..198326d 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/README.md
+++ b/tensorflow/lite/experimental/delegates/hexagon/README.md
@@ -82,6 +82,7 @@
* Mul (without any activation) (b/129276536)
* Neg
* Pad: Only supports 0 padding (b/139277813)
+* Quantize (8-bit inputs & outputs only)
* Relu
* Relu6
* Reshape
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/BUILD b/tensorflow/lite/experimental/delegates/hexagon/builders/BUILD
index 3c666e6..550748e 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/builders/BUILD
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/BUILD
@@ -21,6 +21,7 @@
"op_builder.cc",
"pad_builder.cc",
"pool_2d_builder.cc",
+ "quantize_builder.cc",
"reduce_builder.cc",
"reshape_builder.cc",
"resize_bilinear_builder.cc",
@@ -44,6 +45,7 @@
"op_builder.h",
"pad_builder.h",
"pool_2d_builder.h",
+ "quantize_builder.h",
"reduce_builder.h",
"reshape_builder.h",
"resize_bilinear_builder.h",
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/l2_normalization_builder.cc b/tensorflow/lite/experimental/delegates/hexagon/builders/l2_normalization_builder.cc
index ab91f65..9244089 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/builders/l2_normalization_builder.cc
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/l2_normalization_builder.cc
@@ -36,9 +36,7 @@
const auto& input_tensor = context->tensors[tensor_id];
AddInput(graph_builder_->GetHexagonTensorId(tensor_id));
TF_LITE_ENSURE_STATUS(
- ComputeMinAndMaxQuantValues(input_tensor, &input_min_, &input_max_,
- std::numeric_limits<uint8_t>::min(),
- std::numeric_limits<uint8_t>::max()));
+ ComputeMinAndMaxQuantValues(input_tensor, &input_min_, &input_max_));
auto* input_min_const = graph_builder_->AddConstNodeWithData(
quant_bound_shape, reinterpret_cast<char*>(&input_min_),
sizeof(input_min_));
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/op_builder.cc b/tensorflow/lite/experimental/delegates/hexagon/builders/op_builder.cc
index d737d4a..0cfe999 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/builders/op_builder.cc
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/op_builder.cc
@@ -87,6 +87,8 @@
return CreateSpaceToDepthBuilder(this, OP_SpaceToDepth_8);
case kTfLiteBuiltinDepthToSpace:
return CreateSpaceToDepthBuilder(this, OP_DepthToSpace_8);
+ case kTfLiteBuiltinQuantize:
+ return CreateQuantizeBuilder(this, OP_Requantize_8to8);
default:
context_->ReportError(context_, "Op not supported: %d", op_type);
return nullptr;
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/op_factory.h b/tensorflow/lite/experimental/delegates/hexagon/builders/op_factory.h
index 47e63f5..e2a4ef9 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/builders/op_factory.h
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/op_factory.h
@@ -50,6 +50,7 @@
int max_size_for_batch,
TfLiteIntArray* input_batch_dimensions,
TfLiteIntArray* output_batch_dimensions);
+OpBuilder* CreateQuantizeBuilder(GraphBuilder* graph_builder, int op_type);
} // namespace hexagon
} // namespace delegates
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.cc b/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.cc
new file mode 100644
index 0000000..e425864
--- /dev/null
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.cc
@@ -0,0 +1,91 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include "tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.h"
+
+#include <stdint.h>
+
+#include <limits>
+
+#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/experimental/delegates/hexagon/hexagon_nn/hexagon_nn.h"
+#include "tensorflow/lite/kernels/kernel_util.h"
+
+namespace tflite {
+namespace delegates {
+namespace hexagon {
+TfLiteStatus QuantizeOpBuilder::PopulateSubGraph(const TfLiteIntArray* inputs,
+ const TfLiteIntArray* outputs,
+ TfLiteContext* context) {
+ static int scalar_shape[] = {1, 1, 1, 1};
+
+ // Input.
+ float input_min = 0;
+ float input_max = 0;
+ const auto& input_tensor = context->tensors[inputs->data[0]];
+ ComputeMinAndMaxQuantValues(input_tensor, &input_min, &input_max);
+ auto* input_min_const = graph_builder_->AddConstNodeWithData(
+ scalar_shape, reinterpret_cast<char*>(&input_min), sizeof(input_min));
+ auto* input_max_const = graph_builder_->AddConstNodeWithData(
+ scalar_shape, reinterpret_cast<char*>(&input_max), sizeof(input_max));
+
+ // Output.
+ float output_min = 0;
+ float output_max = 0;
+ const auto& output_tensor = context->tensors[outputs->data[0]];
+ TF_LITE_ENSURE_STATUS(
+ ComputeMinAndMaxQuantValues(output_tensor, &output_min, &output_max));
+ int output_batch_size, output_height_size, output_width_size,
+ output_depth_size;
+ GetDims(&output_batch_size, &output_height_size, &output_width_size,
+ &output_depth_size, output_tensor.dims);
+ auto* requantized_min_const = graph_builder_->AddConstNodeWithData(
+ scalar_shape, reinterpret_cast<char*>(&output_min), sizeof(output_min));
+ auto* requantized_max_const = graph_builder_->AddConstNodeWithData(
+ scalar_shape, reinterpret_cast<char*>(&output_max), sizeof(output_max));
+
+ AddInput(graph_builder_->GetHexagonTensorId(inputs->data[0]));
+ AddInput(TensorID(input_min_const->GetID(), 0));
+ AddInput(TensorID(input_max_const->GetID(), 0));
+ AddInput(TensorID(requantized_min_const->GetID(), 0));
+ AddInput(TensorID(requantized_max_const->GetID(), 0));
+
+ // Hexagon outputs for this node.
+ node_output_ = AddOutput(sizeof(uint8_t), 4,
+ {output_batch_size, output_height_size,
+ output_width_size, output_depth_size});
+ AddOutput(sizeof(float), 4, {1, 1, 1, 1});
+ AddOutput(sizeof(float), 4, {1, 1, 1, 1});
+
+ return kTfLiteOk;
+}
+
+TfLiteStatus QuantizeOpBuilder::RegisterOutputs(const TfLiteIntArray* outputs,
+ TfLiteContext* context) {
+ // Should be only 1 output.
+ graph_builder_->AddTensorWithID(outputs->data[0], node_output_.first,
+ node_output_.second);
+
+ return kTfLiteOk;
+}
+
+QuantizeOpBuilder::~QuantizeOpBuilder() {}
+
+OpBuilder* CreateQuantizeBuilder(GraphBuilder* graph_builder, int op_type) {
+ return new QuantizeOpBuilder(graph_builder, op_type);
+}
+
+} // namespace hexagon
+} // namespace delegates
+} // namespace tflite
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.h b/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.h
new file mode 100644
index 0000000..9851ce4
--- /dev/null
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/quantize_builder.h
@@ -0,0 +1,48 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_EXPERIMENTAL_DELEGATES_HEXAGON_BUILDERS_QUANTIZE_BUILDER_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_DELEGATES_HEXAGON_BUILDERS_QUANTIZE_BUILDER_H_
+
+#include "tensorflow/lite/experimental/delegates/hexagon/builders/op_builder.h"
+
+namespace tflite {
+namespace delegates {
+namespace hexagon {
+
+class QuantizeOpBuilder : public OpBuilder {
+ public:
+ explicit QuantizeOpBuilder(GraphBuilder* graph_builder, int op_type)
+ : OpBuilder(graph_builder, op_type) {}
+ explicit QuantizeOpBuilder(GraphBuilder* graph_builder, int op_type,
+ int relu_value)
+ : OpBuilder(graph_builder, op_type) {}
+ TfLiteStatus PopulateSubGraph(const TfLiteIntArray* inputs,
+ const TfLiteIntArray* outputs,
+ TfLiteContext* context) override;
+
+ TfLiteStatus RegisterOutputs(const TfLiteIntArray* outputs,
+ TfLiteContext* context) override;
+
+ ~QuantizeOpBuilder() override;
+
+ private:
+ TensorID node_output_;
+};
+
+} // namespace hexagon
+} // namespace delegates
+} // namespace tflite
+
+#endif // TENSORFLOW_LITE_EXPERIMENTAL_DELEGATES_HEXAGON_BUILDERS_QUANTIZE_BUILDER_H_
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/tests/BUILD b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/BUILD
index 3bf9120..b1df59c 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/builders/tests/BUILD
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/BUILD
@@ -28,11 +28,13 @@
"arg_min_max_test.cc",
"concat_test.cc",
"conv_test.cc",
+ "l2_norm_test.cc",
"matmul_test.cc",
"mul_test.cc",
"neg_test.cc",
"pad_test.cc",
"pool_test.cc",
+ "quantize_test.cc",
"reduce_test.cc",
"reshape_test.cc",
"resize_test.cc",
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/tests/l2_norm_test.cc b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/l2_norm_test.cc
new file mode 100644
index 0000000..34d53d6
--- /dev/null
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/l2_norm_test.cc
@@ -0,0 +1,122 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include <gtest/gtest.h>
+#include "tensorflow/lite/experimental/delegates/hexagon/builders/tests/hexagon_delegate_op_model.h"
+
+namespace tflite {
+using testing::ElementsAreArray;
+
+class L2NormOpModel : public SingleOpModelWithHexagon {
+ public:
+ L2NormOpModel(const std::initializer_list<int> input_shape,
+ const TensorType tensor_type) {
+ TensorData data = TensorData{tensor_type};
+ data.min = -2.0;
+ data.max = 2.0;
+ data.scale = 2.0;
+ data.zero_point = 128;
+ input_ = AddInput(data);
+
+ data.min = -1.0;
+ data.max = 127.0 / 128.0;
+ output_ = AddOutput(data);
+
+ SetBuiltinOp(
+ BuiltinOperator_L2_NORMALIZATION, BuiltinOptions_L2NormOptions,
+ CreateL2NormOptions(builder_, ActivationFunctionType_NONE).Union());
+ BuildInterpreter({input_shape});
+ }
+
+ void SetInput(std::initializer_list<float> data) {
+ PopulateTensor(input_, data);
+ }
+
+ template <typename T>
+ std::vector<T> GetOutput() {
+ return ExtractVector<T>(output_);
+ }
+
+ template <typename T>
+ std::vector<float> GetDequantizedOutput() {
+ return Dequantize<T>(ExtractVector<T>(output_), GetScale(output_),
+ GetZeroPoint(output_));
+ }
+
+ int input() const { return input_; }
+
+ private:
+ int input_;
+ int output_;
+};
+
+TEST(L2NormOpTest, ZerosVectorUint8Test) {
+ L2NormOpModel m({1, 1, 1, 6}, TensorType_UINT8);
+
+ m.QuantizeAndPopulate<uint8_t>(m.input(), {0, 0, 0, 0, 0, 0});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetDequantizedOutput<uint8_t>(),
+ ElementsAreArray(ArrayFloatNear({0, 0, 0, 0, 0, 0}, 0.1)));
+}
+
+TEST(L2NormOpTest, ZerosVectorInt8Test) {
+ L2NormOpModel m({1, 1, 1, 6}, TensorType_INT8);
+
+ m.QuantizeAndPopulate<int8_t>(m.input(), {0, 0, 0, 0, 0, 0});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetDequantizedOutput<int8_t>(),
+ ElementsAreArray(ArrayFloatNear({0, 0, 0, 0, 0, 0}, 0.1)));
+}
+
+TEST(L2NormOpTest, MultipleBatchUint8Test) {
+ L2NormOpModel m({3, 1, 1, 6}, TensorType_UINT8);
+
+ m.QuantizeAndPopulate<uint8_t>(m.input(),
+ {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 1
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 2
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 3
+ });
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetDequantizedOutput<uint8_t>(),
+ ElementsAreArray(ArrayFloatNear(
+ {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 1
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 2
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 3
+ },
+ 0.1)));
+}
+
+TEST(L2NormOpTest, MultipleBatchInt8Test) {
+ L2NormOpModel m({3, 1, 1, 6}, TensorType_INT8);
+
+ m.QuantizeAndPopulate<int8_t>(m.input(),
+ {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 1
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 2
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 3
+ });
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetDequantizedOutput<int8_t>(),
+ ElementsAreArray(ArrayFloatNear(
+ {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 1
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 2
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 3
+ },
+ 0.1)));
+}
+
+} // namespace tflite
diff --git a/tensorflow/lite/experimental/delegates/hexagon/builders/tests/quantize_test.cc b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/quantize_test.cc
new file mode 100644
index 0000000..93cd138
--- /dev/null
+++ b/tensorflow/lite/experimental/delegates/hexagon/builders/tests/quantize_test.cc
@@ -0,0 +1,170 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include <gtest/gtest.h>
+#include "tensorflow/lite/experimental/delegates/hexagon/builders/tests/hexagon_delegate_op_model.h"
+
+namespace tflite {
+using testing::ElementsAreArray;
+
+class QuantizeOpModel : public SingleOpModelWithHexagon {
+ public:
+ explicit QuantizeOpModel(const TensorData& input, const TensorData& output) {
+ input_ = AddInput(input);
+ output_ = AddOutput(output);
+ SetBuiltinOp(BuiltinOperator_QUANTIZE, BuiltinOptions_QuantizeOptions,
+ CreateQuantizeOptions(builder_).Union());
+ BuildInterpreter({GetShape(input_)});
+ }
+
+ template <typename T>
+ void SetInput(const std::vector<float>& data) {
+ QuantizeAndPopulate<T>(input_, data);
+ }
+
+ template <typename T>
+ std::vector<T> GetOutput() {
+ return ExtractVector<T>(output_);
+ }
+
+ protected:
+ BuiltinOperator op_code_;
+
+ int input_;
+ int output_;
+};
+
+// Input scale 0.500000, output scale 0.500000, input zeropoint 127, output
+// zeropoint 127
+TEST(QuantizeOpTest, UInt8UInt8SameScale) {
+ QuantizeOpModel m({TensorType_UINT8, {1, 1, 2, 5}, -63.5, 64},
+ {TensorType_UINT8, {1, 1, 2, 5}, -63.5, 64});
+
+ // Input will quantized to {129,131,133,135,137,139,141,143,145,147}.
+ m.SetInput<uint8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(
+ m.GetOutput<uint8_t>(),
+ ElementsAreArray({129, 131, 133, 135, 137, 139, 141, 143, 145, 147}));
+}
+
+// Input scale 0.500000, output scale 1.000000, input zeropoint 127, output
+// zeropoint 127
+TEST(QuantizeOpTest, Uint8Uint8LargerScale) {
+ QuantizeOpModel m({TensorType_UINT8, {1, 1, 2, 5}, -63.5, 64},
+ {TensorType_UINT8, {1, 1, 2, 5}, -127, 128});
+
+ // Input will quantized to {129,131,133,135,137,139,141,143,145,147}.
+ m.SetInput<uint8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(
+ m.GetOutput<uint8_t>(),
+ ElementsAreArray({128, 129, 130, 131, 132, 133, 134, 135, 136, 137}));
+}
+
+// Input scale 1.000000, output scale 0.500000, input zeropoint 127, output
+// zeropoint 127
+TEST(QuantizeOpTest, Uint8Uint8SmallerScale) {
+ QuantizeOpModel m({TensorType_UINT8, {1, 1, 2, 5}, -127, 128},
+ {TensorType_UINT8, {1, 1, 2, 5}, -63.5, 64});
+
+ // Input will quantized to {128, 129, 130, 131, 132, 133, 134, 135, 136, 137}.
+ m.SetInput<uint8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(
+ m.GetOutput<uint8_t>(),
+ ElementsAreArray({129, 131, 133, 135, 137, 139, 141, 143, 145, 147}));
+}
+
+// Input scale 1.000000, output scale 0.500000, input zeropoint -1, output
+// zeropoint 127
+TEST(QuantizeOpTest, Int8Uint8SmallerScale) {
+ QuantizeOpModel m({TensorType_INT8, {1, 1, 2, 5}, -127, 128},
+ {TensorType_UINT8, {1, 1, 2, 5}, -63.5, 64});
+
+ // Input will quantized to {0,1,2,3,4,5,6,7,8,9}.
+ m.SetInput<int8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(
+ m.GetOutput<uint8_t>(),
+ ElementsAreArray({129, 131, 133, 135, 137, 139, 141, 143, 145, 147}));
+}
+
+// Input scale 1.000000, output scale 2.000000, input zeropoint -1, output
+// zeropoint 127
+TEST(QuantizeOpTest, Int8Uint8LargerScale) {
+ QuantizeOpModel m({TensorType_INT8, {1, 1, 2, 5}, -127, 128},
+ {TensorType_UINT8, {1, 1, 2, 5}, -254, 256});
+
+ // Input will quantized to {0,1,2,3,4,5,6,7,8,9}.
+ m.SetInput<int8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(
+ m.GetOutput<uint8_t>(),
+ ElementsAreArray({128, 128, 129, 129, 130, 130, 131, 131, 132, 132}));
+}
+
+// input scale 0.500000, output scale 0.500000, input zeropoint 127, output
+// zeropoint -1
+TEST(QuantizeOpTest, UInt8Int8SameScale128Diff) {
+ QuantizeOpModel m({TensorType_UINT8, {1, 1, 2, 5}, -127, 128},
+ {TensorType_INT8, {1, 1, 2, 5}, -127, 128});
+
+ // Input will quantized to {128, 129, 130, 131, 132, 133, 134, 135, 136, 137}.
+ m.SetInput<uint8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetOutput<int8_t>(),
+ ElementsAreArray({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
+}
+
+// Input scale 0.500000, output scale 0.500000, input zeropoint -1, output
+// zeropoint -1
+TEST(QuantizeOpTest, Int8Int8SameScale) {
+ QuantizeOpModel m({TensorType_INT8, {1, 1, 2, 5}, -63.5, 64},
+ {TensorType_INT8, {1, 1, 2, 5}, -63.5, 64});
+
+ // Input will quantized to {1,3,5,7,9,11,13,15,17,19}.
+ m.SetInput<int8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetOutput<int8_t>(),
+ ElementsAreArray({1, 3, 5, 7, 9, 11, 13, 15, 17, 19}));
+}
+
+// Input scale 0.500000, output scale 1.000000, input zeropoint -1, output
+// zeropoint -1
+TEST(QuantizeOpTest, Int8Int8LargerScale) {
+ QuantizeOpModel m({TensorType_INT8, {1, 1, 2, 5}, -63.5, 64},
+ {TensorType_INT8, {1, 1, 2, 5}, -127, 128});
+
+ // Input will quantized to {1,3,5,7,9,11,13,15,17,19}.
+ m.SetInput<int8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetOutput<int8_t>(),
+ ElementsAreArray({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
+}
+
+// Input scale 1.000000, output scale 0.500000, input zeropoint -1, output
+// zeropoint -1
+TEST(QuantizeOpTest, Int8Int8SmallerScale) {
+ QuantizeOpModel m({TensorType_INT8, {1, 1, 2, 5}, -127, 128},
+ {TensorType_INT8, {1, 1, 2, 5}, -63.5, 64});
+
+ // Input will quantized to {0,1,2,3,4,5,6,7,8,9}.
+ m.SetInput<int8_t>({1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
+ m.ApplyDelegateAndInvoke();
+ EXPECT_THAT(m.GetOutput<int8_t>(),
+ ElementsAreArray({1, 3, 5, 7, 9, 11, 13, 15, 17, 19}));
+}
+
+} // namespace tflite
diff --git a/tensorflow/lite/experimental/delegates/hexagon/hexagon_delegate.cc b/tensorflow/lite/experimental/delegates/hexagon/hexagon_delegate.cc
index dd98f58..e8ba5fa 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/hexagon_delegate.cc
+++ b/tensorflow/lite/experimental/delegates/hexagon/hexagon_delegate.cc
@@ -152,8 +152,6 @@
std::string* unsupported_details) -> bool {
return IsNodeSupportedByHexagon(registration, node, context);
};
- TfLiteIntArray* plan;
- TF_LITE_ENSURE_STATUS(context->GetExecutionPlan(context, &plan));
delegates::GraphPartitionHelper helper(context, node_supported_fn);
TF_LITE_ENSURE_STATUS(helper.Partition(nullptr));
@@ -175,7 +173,8 @@
supported_nodes[0] = supported_nodes.size() - 1;
auto* hexagon_delegate = static_cast<HexagonDelegate*>(delegate);
// Make sure dynamic batch is requested on fully delegated graph only.
- if (supported_nodes[0] != plan->size && hexagon_delegate != nullptr &&
+ if (supported_nodes[0] != helper.num_total_nodes() &&
+ hexagon_delegate != nullptr &&
hexagon_delegate->params()->enable_dynamic_batch_size) {
TF_LITE_KERNEL_LOG(
context, "Dynamic batch requested on non-fully delegated graph !!.");
diff --git a/tensorflow/lite/experimental/delegates/hexagon/utils.cc b/tensorflow/lite/experimental/delegates/hexagon/utils.cc
index 35fce1b..6ba1279 100644
--- a/tensorflow/lite/experimental/delegates/hexagon/utils.cc
+++ b/tensorflow/lite/experimental/delegates/hexagon/utils.cc
@@ -77,17 +77,19 @@
case kTfLiteBuiltinArgMin:
case kTfLiteBuiltinAveragePool2d:
case kTfLiteBuiltinConcatenation:
+ case kTfLiteBuiltinL2Normalization:
case kTfLiteBuiltinLogistic:
case kTfLiteBuiltinMaxPool2d:
case kTfLiteBuiltinMul:
case kTfLiteBuiltinPad:
- case kTfLiteBuiltinSub:
+ case kTfLiteBuiltinQuantize:
case kTfLiteBuiltinRelu6:
case kTfLiteBuiltinResizeBilinear:
case kTfLiteBuiltinResizeNearestNeighbor:
case kTfLiteBuiltinSoftmax:
case kTfLiteBuiltinSpaceToDepth:
case kTfLiteBuiltinSplit:
+ case kTfLiteBuiltinSub:
case kTfLiteBuiltinTanh:
case kTfLiteBuiltinTranspose:
case kTfLiteBuiltinTransposeConv:
@@ -301,8 +303,7 @@
IsConstantTensor(GetInput(context, node, 1));
}
case kTfLiteBuiltinL2Normalization: {
- // TODO(b/142009955): Support int8.
- if (!InputsWithCorrectTypes(node, context, {{kTfLiteUInt8}}))
+ if (!InputsWithCorrectTypes(node, context, {{kTfLiteUInt8, kTfLiteInt8}}))
return false;
const TfLiteL2NormParams* norm_params =
reinterpret_cast<const TfLiteL2NormParams*>(node->builtin_data);
@@ -347,6 +348,10 @@
return InputsWithCorrectTypes(node, context,
{{kTfLiteUInt8, kTfLiteInt8}});
}
+ case kTfLiteBuiltinQuantize: {
+ return InputsWithCorrectTypes(node, context,
+ {{kTfLiteUInt8, kTfLiteInt8}});
+ }
default:
return false;
}
diff --git a/tensorflow/lite/experimental/kernels/BUILD b/tensorflow/lite/experimental/kernels/BUILD
index e5d7896..70ae658 100644
--- a/tensorflow/lite/experimental/kernels/BUILD
+++ b/tensorflow/lite/experimental/kernels/BUILD
@@ -125,50 +125,3 @@
"@com_google_googletest//:gtest",
],
)
-
-cc_library(
- name = "hashtable_op_kernels",
- srcs = [
- "hashtable.cc",
- "hashtable_find.cc",
- "hashtable_import.cc",
- "hashtable_ops.cc",
- "hashtable_size.cc",
- ],
- hdrs = [
- "hashtable_ops.h",
- ],
- deps = [
- "//tensorflow/lite:framework",
- "//tensorflow/lite/c:common",
- "//tensorflow/lite/core/api",
- "//tensorflow/lite/experimental/resource",
- "//tensorflow/lite/kernels:kernel_util",
- "//tensorflow/lite/kernels:op_macros",
- "//tensorflow/lite/kernels/internal:tensor",
- "//tensorflow/lite/schema:schema_fbs",
- "@flatbuffers",
- ],
-)
-
-cc_test(
- name = "hashtable_op_test",
- size = "small",
- srcs = [
- "hashtable_ops_test.cc",
- ],
- deps = [
- ":hashtable_op_kernels", # buildcleaner: keep
- "//tensorflow/lite:framework",
- "//tensorflow/lite/core/api",
- "//tensorflow/lite/experimental/resource",
- "//tensorflow/lite/kernels:test_main",
- "//tensorflow/lite/kernels:test_util",
- "//tensorflow/lite/kernels/internal:tensor",
- "//tensorflow/lite/testing:util",
- "@com_google_absl//absl/memory",
- "@com_google_absl//absl/strings",
- "@com_google_googletest//:gtest",
- "@flatbuffers",
- ],
-)
diff --git a/tensorflow/lite/experimental/support/BUILD b/tensorflow/lite/experimental/support/BUILD
index 2799a4c..05673ab 100644
--- a/tensorflow/lite/experimental/support/BUILD
+++ b/tensorflow/lite/experimental/support/BUILD
@@ -3,8 +3,5 @@
package_group(
name = "users",
- packages = [
- "//learning/brain/mobile/lite/support/...",
- "//tensorflow/lite/experimental/support/...",
- ],
+ packages = ["//tensorflow/lite/experimental/support/..."],
)
diff --git a/tensorflow/lite/g3doc/guide/roadmap.md b/tensorflow/lite/g3doc/guide/roadmap.md
index ff98c85..35ef44a 100644
--- a/tensorflow/lite/g3doc/guide/roadmap.md
+++ b/tensorflow/lite/g3doc/guide/roadmap.md
@@ -1,81 +1,87 @@
# TensorFlow Lite 2019 Roadmap
-**Updated: August 29, 2019**
+**Updated: April 18, 2020**
-The following represents a high level overview of our 2019 plan. You should be
-conscious that this roadmap may change at anytime relative to a range of factors
-and the order below does not reflect any type of priority. As a matter of
-principle, we typically prioritize issues that the majority of our users are
-asking for and so this list fundamentally reflects that.
+The following represents a high level overview of our 2020 plan. You should be
+aware that this roadmap may change at any time and the order below does not
+reflect any type of priority. As a matter of principle, we typically prioritize
+issues based on the number of users affected.
We break our roadmap into four key segments: usability, performance,
optimization and portability. We strongly encourage you to comment on our
-roadmap and provide us feedback in the TF Lite discussion groups and forums.
+roadmap and provide us feedback in the
+[TF Lite discussion group](https://groups.google.com/a/tensorflow.org/g/tflite).
## Usability
-* **New model converter**
- * New MLIR-based TensorFlow Lite convertor that better handles graph
- conversion (e.g., control flow, conditionals, etc...)
- * Improved diagnostics and debugging of model conversion failures
* **Expanded ops coverage**
* Prioritized op additions based on user feedback
* **Improvements to using TensorFlow ops in TensorFlow Lite**
* Pre-built libraries available via Bintray (Android) and Cocoapods (iOS)
* Smaller binary size when using select TF ops via op stripping
* **LSTM / RNN support**
- * Full support of conversion for LSTMs and RNNs
-* **Pre-and-post processing support**
- * New support library for model-specific pre-and-post processing
- * Utilities for common platform-specific functionality, e.g., loading a
- model efficiently from assets, or converting a Bitmap to a tensor
+ * Full LSTM and RNN conversion support, including support in Keras
+* **Pre-and-post processing support libraries and codegen tool**
+ * Ready-to-use API building blocks for common ML tasks
+ * Support more models (e.g. NLP) and more platforms (e.g. iOS)
+* **Android Studio Integration**
+ * Drag & drop TFLite models into Android Studio to generate model binding
+ classes
* **Control Flow & Training on-device**
- * Support for control flow related ops
* Support for training on-device, focused on personalization and transfer
- learning
-* **Graph visualization tooling**
- * Provide enhanced graph visualization tooling
+ learning
+* **Visualization tooling with TensorBoard**
+ * Provide enhanced tooling with TensorBoard
+* **Model Maker**
+ * Support more tasks, including object detection and BERT-based NLP tasks
* **More models and examples**
- * More models on the support section of the site
- * Additional examples to demonstrate model usage as well as new features
- and APIs, covering different platforms.
- * Model customization libraries and tutorials to let beginners to
- customize those models easily.
+ * More examples to demonstrate model usage as well as new features and
+ APIs, covering different platforms.
## Performance
* **Better tooling**
- * Simpler benchmarking and profiling tools for understanding available
- accelerators and performance tradeoffs
* Public dashboard for tracking performance gains with each release
* **Improved CPU performance**
- * Continued optimization of float and quantized kernels
+ * New highly optimized floating-point kernel library for convolutional
+ models
* First-class x86 support
* **Updated NN API support**
- * Full support for new Android Q NN API features, ops and types
+ * Full support for new Android R NN API features, ops and types
* **GPU backend optimizations**
- * OpenCL and Vulkan support on Android
- * Metal and Objective-C CocoaPods for Metal acceleration
+ * Vulkan support on Android
+ * Support integer quantized models
* **Hexagon DSP backend**
- * Initial release of DSP acceleration for pre-Android P devices
+ * Per-channel quantization support for all models created through
+ post-training quantization
+ * Dynamic input batch size support
+ * Better op coverage, including LSTM
+* **Core ML backend**
+ * Optimizing start-up time
+ * Dynamic quantized models support
+ * Float16 quantized models support
+ * Better op coverage
## Optimization
* **Quantization**
- * Post training quantization for hybrid kernels -- [Launched](https://medium.com/tensorflow/introducing-the-model-optimization-toolkit-for-tensorflow-254aca1ba0a3){:.external}
- * Post training quantization for (8b) fixed-point kernels -- [Launched](https://medium.com/tensorflow/tensorflow-model-optimization-toolkit-post-training-integer-quantization-b4964a1ea9ba){:.external}
- * Training with quantization for (8b) fixed-point kernels
- * Extend post and during training APIs to (8b) fixed-point RNNs
- * Training with quantization for low bit-width (< 8b) fixed-point kernels
+
+ * Post-training quantization for (8b) fixed-point RNNs
+ * During-training quantization for (8b) fixed-point RNNs
+ * Quality and performance improvements for post-training dynamic-range
+ quantization
+
* **Pruning / sparsity**
- * Magnitude based weight pruning during training -- [Launched](https://medium.com/tensorflow/tensorflow-model-optimization-toolkit-pruning-api-42cac9157a6a){:.external}
- * Support for sparse model execution
+
+ * Sparse model execution support in TensorFlow Lite -
+ [WIP](https://github.com/tensorflow/model-optimization/issues/173)
+ * Weight clustering API
## Portability
* **Microcontroller Support**
- * Add support for a range of 32-bit MCU architecture use cases for Speech
- and Image Classification
+ * Add support for a range of 32-bit MCU architecture use cases for speech
+ and image classification
* Sample code and models for vision and audio data
* Full TF Lite op support on microcontrollers
* Support for more platforms, including CircuitPython support
diff --git a/tensorflow/lite/interpreter.cc b/tensorflow/lite/interpreter.cc
index d5aad23..c8ccf67 100644
--- a/tensorflow/lite/interpreter.cc
+++ b/tensorflow/lite/interpreter.cc
@@ -311,10 +311,19 @@
}
TfLiteStatus Interpreter::ModifyGraphWithDelegate(TfLiteDelegate* delegate) {
+ TfLiteStatus status = kTfLiteOk;
for (auto& subgraph : subgraphs_) {
- TF_LITE_ENSURE_OK(context_, subgraph->ModifyGraphWithDelegate(delegate));
+ status = subgraph->ModifyGraphWithDelegate(delegate);
+ if (status != kTfLiteOk) {
+ break;
+ }
}
- return kTfLiteOk;
+ // Delegate-specific errors can be recovered from by restoring Interpreter to
+ // its original state.
+ if (status == kTfLiteDelegateError) {
+ TF_LITE_ENSURE_STATUS(RemoveAllDelegates());
+ }
+ return status;
}
TfLiteStatus Interpreter::ModifyGraphWithDelegate(TfLiteDelegatePtr delegate) {
diff --git a/tensorflow/lite/interpreter.h b/tensorflow/lite/interpreter.h
index 6948c13..b93fd76 100644
--- a/tensorflow/lite/interpreter.h
+++ b/tensorflow/lite/interpreter.h
@@ -392,6 +392,12 @@
/// parts of the graph themselves. After this is called, the graph may
/// contain new nodes that replace 1 more nodes.
/// 'delegate' must outlive the interpreter.
+ /// Returns one of the following three status codes:
+ /// 1. kTfLiteOk: Success.
+ /// 2. kTfLiteDelegateError: Delegation failed due to an error in the
+ /// delegate. The Interpreter has been restored to its pre-delegation state.
+ /// NOTE: This undoes all delegates previously applied to the Interpreter.
+ /// 3. kTfLiteError: Unexpected/runtime failure.
/// WARNING: This is an experimental API and subject to change.
TfLiteStatus ModifyGraphWithDelegate(TfLiteDelegate* delegate);
diff --git a/tensorflow/lite/interpreter_builder.cc b/tensorflow/lite/interpreter_builder.cc
index 65f4b8e..e32e076 100644
--- a/tensorflow/lite/interpreter_builder.cc
+++ b/tensorflow/lite/interpreter_builder.cc
@@ -34,6 +34,16 @@
#include "tensorflow/lite/profiling/platform_profiler.h"
#endif
+// aligned_alloc is available (via cstdlib/stdlib.h) with C++17/C11.
+#if __cplusplus >= 201703L || __STDC_VERSION__ >= 201112L
+#if !defined(__ANDROID__) || __ANDROID_API__ >= 28
+// Neither Apple nor Windows provide aligned_alloc.
+#if !defined(__APPLE__) && !defined(_WIN32)
+#define TFLITE_USE_STD_ALIGNED_ALLOC
+#endif
+#endif
+#endif
+
namespace tflite {
namespace {
@@ -197,7 +207,13 @@
// Used to determine how the op data parsing function creates its working space.
class MallocDataAllocator : public BuiltinDataAllocator {
public:
- void* Allocate(size_t size) override { return malloc(size); }
+ void* Allocate(size_t size, size_t alignment_hint) override {
+#ifdef TFLITE_USE_STD_ALIGNED_ALLOC
+ return aligned_alloc(alignment_hint, size);
+#else
+ return malloc(size);
+#endif
+ }
void Deallocate(void* data) override { free(data); }
};
diff --git a/tensorflow/lite/interpreter_test.cc b/tensorflow/lite/interpreter_test.cc
index 7c8c7ef..4eccdf3 100644
--- a/tensorflow/lite/interpreter_test.cc
+++ b/tensorflow/lite/interpreter_test.cc
@@ -960,8 +960,7 @@
TEST(BasicInterpreter, TestUnsupportedDelegateFunctions) {
Interpreter interpreter;
ASSERT_EQ(interpreter.AddTensors(2), kTfLiteOk);
- TfLiteRegistration registration = {
- .init = nullptr, .free = nullptr, .prepare = nullptr, .invoke = nullptr};
+ TfLiteRegistration registration = {nullptr, nullptr, nullptr, nullptr};
// These functions are only supported inside Delegate's Prepare function.
// The test verifies that these functions returns `kTfLiteError`, but not
// `kTfLiteOk` or just crashes.
@@ -1094,7 +1093,7 @@
}
struct TestExternalContext : public TfLiteExternalContext {
- static const TfLiteExternalContextType kType = kTfLiteGemmLowpContext;
+ static constexpr TfLiteExternalContextType kType = kTfLiteGemmLowpContext;
static TestExternalContext* Get(TfLiteContext* context) {
return reinterpret_cast<TestExternalContext*>(
@@ -1612,7 +1611,7 @@
// TfLiteRegistration returns an error status.
ASSERT_EQ(
interpreter_->ModifyGraphWithDelegate(delegate_->get_tf_lite_delegate()),
- kTfLiteError);
+ kTfLiteDelegateError);
// Execution plan should remain unchanged.
ASSERT_EQ(interpreter_->execution_plan().size(), 3);
@@ -1680,21 +1679,20 @@
interpreter_->ModifyGraphWithDelegate(delegate_->get_tf_lite_delegate()),
kTfLiteOk);
ASSERT_EQ(interpreter_->execution_plan().size(), 2);
- // Second delegate won't get applied. However, we should be back to the
- // previous 2-node plan.
+ // Second delegate won't get applied.
+ // As a result, previous delegate should also get undone, restoring the
+ // execution plan to its original state.
ASSERT_EQ(
interpreter_->ModifyGraphWithDelegate(delegate2_->get_tf_lite_delegate()),
- kTfLiteError);
- ASSERT_EQ(interpreter_->execution_plan().size(), 2);
+ kTfLiteDelegateError);
+ ASSERT_EQ(interpreter_->execution_plan().size(), 3);
std::vector<float> input = {1.0f, 2.0f, 3.0f};
- // Node 0: tensor_2 = tensor0 + tensor0
- // Delegated node: tensor_2 + tensor_1
- std::vector<float> expected_output = {3.0f, 6.0f, 9.0f};
+ std::vector<float> expected_output = {2.0f, 4.0f, 6.0f};
constexpr int kOutputTensorIndex = 3;
TfLiteTensor* tensor = interpreter_->tensor(kOutputTensorIndex);
- // Verify Invoke() behavior to ensure Interpreter isn't broken.
+ // Verify Invoke() behavior.
memcpy(interpreter_->typed_tensor<float>(0), input.data(), 3 * sizeof(float));
memcpy(interpreter_->typed_tensor<float>(1), input.data(), 3 * sizeof(float));
interpreter_->Invoke();
diff --git a/tensorflow/lite/java/ovic/demo/app/OvicBenchmarkerActivity.java b/tensorflow/lite/java/ovic/demo/app/OvicBenchmarkerActivity.java
index 1445303..76e0ced 100644
--- a/tensorflow/lite/java/ovic/demo/app/OvicBenchmarkerActivity.java
+++ b/tensorflow/lite/java/ovic/demo/app/OvicBenchmarkerActivity.java
@@ -45,6 +45,7 @@
/** Name of the task-dependent data files stored in Assets. */
private static String labelPath = null;
+
private static String testImagePath = null;
private static String modelPath = null;
/**
@@ -91,7 +92,7 @@
labelPath = "labels.txt";
testImagePath = "test_image_224.jpg";
modelPath = "quantized_model.lite";
- } else { // Benchmarking detection.
+ } else { // Benchmarking detection.
benchmarker = new OvicDetectorBenchmarker(WALL_TIME);
labelPath = "coco_labels.txt";
testImagePath = "test_image_224.jpg";
@@ -145,6 +146,7 @@
public void detectPressed(View view) throws IOException {
benchmarkSession(false);
}
+
public void classifyPressed(View view) throws IOException {
benchmarkSession(true);
}
@@ -194,7 +196,7 @@
displayText
+ modelPath
+ ": Average latency="
- + df2.format(benchmarker.getTotalRunTime() / testIter)
+ + df2.format(benchmarker.getTotalRuntimeNano() * 1.0e-6 / testIter)
+ "ms after "
+ testIter
+ " runs.");
@@ -204,12 +206,15 @@
}
}
+ // TODO(b/153429929) Remove with resolution of issue (see below).
+ @SuppressWarnings("RuntimeExec")
private static void setProcessorAffinity(int mask) throws IOException {
int myPid = Process.myPid();
Log.i(TAG, String.format("Setting processor affinity to 0x%02x", mask));
String command = String.format("taskset -a -p %x %d", mask, myPid);
try {
+ // TODO(b/153429929) This is deprecated, but updating is not safe while verification is hard.
Runtime.getRuntime().exec(command).waitFor();
} catch (InterruptedException e) {
throw new IOException("Interrupted: " + e);
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicBenchmarker.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicBenchmarker.java
index 32bdd5a..49cf21d 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicBenchmarker.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicBenchmarker.java
@@ -43,6 +43,7 @@
/** Dimensions of inputs. */
protected static final int DIM_BATCH_SIZE = 1;
+
protected static final int DIM_PIXEL_SIZE = 3;
protected int imgHeight = 224;
protected int imgWidth = 224;
@@ -53,38 +54,38 @@
/** A ByteBuffer to hold image data, to be feed into classifier as inputs. */
protected ByteBuffer imgData = null;
- /** Total runtime in ms. */
- protected double totalRuntime = 0.0;
+ /** Total runtime in ns. */
+ protected double totalRuntimeNano = 0.0;
/** Total allowed runtime in ms. */
- protected double wallTime = 20000 * 30.0;
+ protected double wallTimeNano = 20000 * 30 * 1.0e6;
/** Record whether benchmark has started (used to skip the first image). */
protected boolean benchmarkStarted = false;
/**
* Initializes an {@link OvicBenchmarker}
*
- * @param wallTime: a double number specifying the total amount of time to benchmark.
+ * @param wallTimeNano: a double number specifying the total amount of time to benchmark.
*/
- public OvicBenchmarker(double wallTime) {
+ public OvicBenchmarker(double wallTimeNano) {
benchmarkStarted = false;
- totalRuntime = 0.0;
- this.wallTime = wallTime;
+ totalRuntimeNano = 0.0;
+ this.wallTimeNano = wallTimeNano;
}
/** Return the cumulative latency of all runs so far. */
- public double getTotalRunTime() {
- return totalRuntime;
+ public double getTotalRuntimeNano() {
+ return totalRuntimeNano;
}
/** Check whether the benchmarker should stop. */
public Boolean shouldStop() {
- if (totalRuntime >= wallTime) {
+ if (totalRuntimeNano >= wallTimeNano) {
Log.e(
TAG,
- "Total runtime "
- + Double.toString(totalRuntime)
- + " exceeded walltime "
- + Double.toString(wallTime));
+ "Total runtime (ms) "
+ + (totalRuntimeNano * 1.0e-6)
+ + " exceeded wall-time "
+ + (wallTimeNano * 1.0e-6));
return true;
}
return false;
@@ -120,9 +121,9 @@
public abstract String getLastResultString();
/**
- * Loads input buffer from intValues into ByteBuffer for the interpreter.
- * Input buffer must be loaded in intValues and output will be placed in imgData.
- */
+ * Loads input buffer from intValues into ByteBuffer for the interpreter. Input buffer must be
+ * loaded in intValues and output will be placed in imgData.
+ */
protected void loadsInputToByteBuffer() {
if (imgData == null || intValues == null) {
throw new RuntimeException("Benchmarker is not yet ready to test.");
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassificationResult.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassificationResult.java
index 5ab804e..9aad371 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassificationResult.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassificationResult.java
@@ -21,34 +21,39 @@
/** Top K classes and probabilities. */
public final ArrayList<String> topKClasses;
+
public final ArrayList<Float> topKProbs;
public final ArrayList<Integer> topKIndices;
/** Latency (ms). */
- public Long latency;
+ public Long latencyMilli;
+
+ /** Latency (ns). */
+ public Long latencyNano;
OvicClassificationResult() {
topKClasses = new ArrayList<>();
topKProbs = new ArrayList<>();
topKIndices = new ArrayList<>();
- latency = -1L;
+ latencyMilli = -1L;
+ latencyNano = -1L;
}
@Override
public String toString() {
- String textToShow = latency + "ms";
+ String textToShow = latencyMilli + "ms";
+ textToShow += "\n" + latencyNano + "ns";
for (int k = 0; k < topKProbs.size(); ++k) {
textToShow +=
"\nPrediction ["
+ k
+ "] = Class "
- + Integer.toString(topKIndices.get(k))
+ + topKIndices.get(k)
+ " ("
+ topKClasses.get(k)
+ ") : "
- + Float.toString(topKProbs.get(k));
+ + topKProbs.get(k);
}
return textToShow;
}
-
}
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifier.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifier.java
index d8a54c1..e27272a 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifier.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifier.java
@@ -14,13 +14,14 @@
==============================================================================*/
package org.tensorflow.ovic;
+import static java.nio.charset.StandardCharsets.UTF_8;
+
import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStream;
import java.io.InputStreamReader;
import java.nio.ByteBuffer;
import java.nio.MappedByteBuffer;
-import java.nio.charset.StandardCharsets;
import java.util.AbstractMap;
import java.util.ArrayList;
import java.util.Collections;
@@ -44,7 +45,7 @@
private Interpreter tflite;
/** Labels corresponding to the output of the vision model. */
- private List<String> labelList;
+ private final List<String> labelList;
/** An array to hold inference results, to be feed into Tensorflow Lite as outputs. */
private byte[][] inferenceOutputArray = null;
@@ -56,19 +57,18 @@
/** Whether the model runs as float or quantized. */
private Boolean outputIsFloat = null;
- private PriorityQueue<Map.Entry<Integer, Float>> sortedLabels =
+ private final PriorityQueue<Map.Entry<Integer, Float>> sortedLabels =
new PriorityQueue<>(
RESULTS_TO_SHOW,
new Comparator<Map.Entry<Integer, Float>>() {
@Override
public int compare(Map.Entry<Integer, Float> o1, Map.Entry<Integer, Float> o2) {
- return (o1.getValue()).compareTo(o2.getValue());
+ return o1.getValue().compareTo(o2.getValue());
}
});
/** Initializes an {@code OvicClassifier}. */
- public OvicClassifier(InputStream labelInputStream, MappedByteBuffer model)
- throws IOException, RuntimeException {
+ public OvicClassifier(InputStream labelInputStream, MappedByteBuffer model) throws IOException {
if (model == null) {
throw new RuntimeException("Input model is empty.");
}
@@ -80,12 +80,12 @@
throw new RuntimeException("The model's input dimensions must be 4 (BWHC).");
}
if (inputDims[0] != 1) {
- throw new RuntimeException("The model must have a batch size of 1, got "
- + inputDims[0] + " instead.");
+ throw new IllegalStateException(
+ "The model must have a batch size of 1, got " + inputDims[0] + " instead.");
}
if (inputDims[3] != 3) {
- throw new RuntimeException("The model must have three color channels, got "
- + inputDims[3] + " instead.");
+ throw new IllegalStateException(
+ "The model must have three color channels, got " + inputDims[3] + " instead.");
}
int minSide = Math.min(inputDims[1], inputDims[2]);
int maxSide = Math.max(inputDims[1], inputDims[2]);
@@ -93,12 +93,15 @@
throw new RuntimeException("The model's resolution must be between (0, 1000].");
}
String outputDataType = TestHelper.getOutputDataType(tflite, 0);
- if (outputDataType.equals("float")) {
- outputIsFloat = true;
- } else if (outputDataType.equals("byte")) {
- outputIsFloat = false;
- } else {
- throw new RuntimeException("Cannot process output type: " + outputDataType);
+ switch (outputDataType) {
+ case "float":
+ outputIsFloat = true;
+ break;
+ case "byte":
+ outputIsFloat = false;
+ break;
+ default:
+ throw new IllegalStateException("Cannot process output type: " + outputDataType);
}
inferenceOutputArray = new byte[1][labelList.size()];
labelProbArray = new float[1][labelList.size()];
@@ -123,7 +126,8 @@
}
}
OvicClassificationResult iterResult = computeTopKLabels();
- iterResult.latency = getLastNativeInferenceLatencyMilliseconds();
+ iterResult.latencyMilli = getLastNativeInferenceLatencyMilliseconds();
+ iterResult.latencyNano = getLastNativeInferenceLatencyNanoseconds();
return iterResult;
}
@@ -154,6 +158,18 @@
return (latency == null) ? null : (Long) (latency / 1000000);
}
+ /*
+ * Get native inference latency of last image classification run.
+ * @throws RuntimeException if model is uninitialized.
+ */
+ public Long getLastNativeInferenceLatencyNanoseconds() {
+ if (tflite == null) {
+ throw new IllegalStateException(
+ TAG + ": ImageNet classifier has not been initialized; Failed.");
+ }
+ return tflite.getLastNativeInferenceDurationNanoseconds();
+ }
+
/** Closes tflite to release resources. */
public void close() {
tflite.close();
@@ -162,9 +178,9 @@
/** Reads label list from Assets. */
private static List<String> loadLabelList(InputStream labelInputStream) throws IOException {
- List<String> labelList = new ArrayList<String>();
+ List<String> labelList = new ArrayList<>();
try (BufferedReader reader =
- new BufferedReader(new InputStreamReader(labelInputStream, StandardCharsets.UTF_8))) {
+ new BufferedReader(new InputStreamReader(labelInputStream, UTF_8))) {
String line;
while ((line = reader.readLine()) != null) {
labelList.add(line);
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifierBenchmarker.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifierBenchmarker.java
index b35b8ff..8eafd7a 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifierBenchmarker.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicClassifierBenchmarker.java
@@ -88,16 +88,17 @@
Log.e(TAG, e.getMessage());
Log.e(TAG, "Failed to classify image.");
}
- if (iterResult == null || iterResult.latency == null) {
+ if (iterResult == null || iterResult.latencyMilli == null || iterResult.latencyNano == null) {
throw new RuntimeException("Classification result or timing is invalid.");
}
- Log.d(TAG, "Native inference latency: " + iterResult.latency);
+ Log.d(TAG, "Native inference latency (ms): " + iterResult.latencyMilli);
+ Log.d(TAG, "Native inference latency (ns): " + iterResult.latencyNano);
Log.i(TAG, iterResult.toString());
if (!benchmarkStarted) { // Skip the first image to discount warming-up time.
benchmarkStarted = true;
} else {
- totalRuntime += ((double) iterResult.latency);
+ totalRuntimeNano += ((double) iterResult.latencyNano);
}
return true;
}
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectionResult.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectionResult.java
index cf2902a..15e62c5 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectionResult.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectionResult.java
@@ -22,7 +22,9 @@
// Top K classes and probabilities.
public final ArrayList<BoundingBox> detections;
// Latency (ms).
- public Long latency = -1L;
+ public Long latencyMilli = -1L;
+ // Latency (ns).
+ public Long latencyNano = -1L;
// id of the image.
public int id = -1;
// Number of valid detections (separately maintained, maybe different from detections.size()).
@@ -37,9 +39,10 @@
}
}
- public void resetTo(Long latency, int id) {
+ public void resetTo(Long latencyMilli, Long latencyNano, int id) {
count = 0;
- this.latency = latency;
+ this.latencyMilli = latencyMilli;
+ this.latencyNano = latencyNano;
this.id = id;
}
@@ -64,7 +67,8 @@
@Override
public String toString() {
- String textToShow = latency + "ms";
+ String textToShow = latencyMilli + "ms";
+ textToShow += "\n" + latencyNano + "ns";
int k = 0;
for (BoundingBox box : detections) {
textToShow +=
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetector.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetector.java
index 84c9816..c43eb13 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetector.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetector.java
@@ -14,13 +14,14 @@
==============================================================================*/
package org.tensorflow.ovic;
+import static java.nio.charset.StandardCharsets.UTF_8;
+
import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStream;
import java.io.InputStreamReader;
import java.nio.ByteBuffer;
import java.nio.MappedByteBuffer;
-import java.nio.charset.StandardCharsets;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.List;
@@ -105,7 +106,7 @@
private static List<String> loadLabelList(InputStream labelInputStream) throws IOException {
List<String> labelList = new ArrayList<>();
try (BufferedReader reader =
- new BufferedReader(new InputStreamReader(labelInputStream, StandardCharsets.UTF_8))) {
+ new BufferedReader(new InputStreamReader(labelInputStream, UTF_8))) {
String line;
while ((line = reader.readLine()) != null) {
labelList.add(line);
@@ -131,10 +132,11 @@
Object[] inputArray = {imgData};
tflite.runForMultipleInputsOutputs(inputArray, outputMap);
- Long latency = getLastNativeInferenceLatencyMilliseconds();
+ Long latencyMilli = getLastNativeInferenceLatencyMilliseconds();
+ Long latencyNano = getLastNativeInferenceLatencyNanoseconds();
// Update the results.
- result.resetTo(latency, imageId);
+ result.resetTo(latencyMilli, latencyNano, imageId);
for (int i = 0; i < NUM_RESULTS; i++) {
// The model returns normalized coordinates [start_y, start_x, end_y, end_x].
// The boxes expect pixel coordinates [x1, y1, x2, y2].
@@ -154,7 +156,7 @@
/*
* Get native inference latency of last image detection run.
* @throws RuntimeException if model is uninitialized.
- * @return The inference latency in millisecond.
+ * @return The inference latency in milliseconds.
*/
public Long getLastNativeInferenceLatencyMilliseconds() {
if (tflite == null) {
@@ -164,6 +166,19 @@
return (latency == null) ? null : (Long) (latency / 1000000);
}
+ /*
+ * Get native inference latency of last image detection run.
+ * @throws RuntimeException if model is uninitialized.
+ * @return The inference latency in nanoseconds.
+ */
+ public Long getLastNativeInferenceLatencyNanoseconds() {
+ if (tflite == null) {
+ throw new IllegalStateException(
+ TAG + ": ImageNet classifier has not been initialized; Failed.");
+ }
+ return tflite.getLastNativeInferenceDurationNanoseconds();
+ }
+
public int[] getInputDims() {
return inputDims;
}
diff --git a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectorBenchmarker.java b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectorBenchmarker.java
index 15a4c98..0c03269 100644
--- a/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectorBenchmarker.java
+++ b/tensorflow/lite/java/ovic/src/main/java/org/tensorflow/ovic/OvicDetectorBenchmarker.java
@@ -98,7 +98,7 @@
if (!benchmarkStarted) { // Skip the first image to discount warming-up time.
benchmarkStarted = true;
} else {
- totalRuntime += ((double) detector.result.latency);
+ totalRuntimeNano += ((double) detector.result.latencyNano);
}
return true; // Indicating that result is ready.
}
diff --git a/tensorflow/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java b/tensorflow/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java
index 37d7165..7ded4df 100644
--- a/tensorflow/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java
+++ b/tensorflow/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java
@@ -116,7 +116,7 @@
public void ovicClassifier_latencyNotNull() throws Exception {
classifier = new OvicClassifier(labelsInputStream, floatModel);
testResult = classifier.classifyByteBuffer(testImage);
- assertThat(testResult.latency).isNotNull();
+ assertThat(testResult.latencyNano).isNotNull();
}
@Test
diff --git a/tensorflow/lite/kernels/BUILD b/tensorflow/lite/kernels/BUILD
index 8073146..ae49d79 100644
--- a/tensorflow/lite/kernels/BUILD
+++ b/tensorflow/lite/kernels/BUILD
@@ -600,7 +600,7 @@
":op_macros",
"//tensorflow/lite:context",
"//tensorflow/lite/c:common",
- "//tensorflow/lite/experimental/kernels:hashtable_op_kernels",
+ "//tensorflow/lite/kernels/hashtable:hashtable_op_kernels",
"//tensorflow/lite/kernels/internal:kernel_utils",
"//tensorflow/lite/kernels/internal:tensor",
"//third_party/fft2d:fft2d_headers",
diff --git a/tensorflow/lite/kernels/comparisons.cc b/tensorflow/lite/kernels/comparisons.cc
index a8e3148..4e20efc 100644
--- a/tensorflow/lite/kernels/comparisons.cc
+++ b/tensorflow/lite/kernels/comparisons.cc
@@ -27,7 +27,8 @@
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;
-TfLiteStatus ComparisonPrepare(TfLiteContext* context, TfLiteNode* node) {
+TfLiteStatus ComparisonPrepareCommon(TfLiteContext* context, TfLiteNode* node,
+ bool is_string_allowed) {
TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
@@ -36,7 +37,9 @@
TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
// Don't support string.
- TF_LITE_ENSURE(context, input1->type != kTfLiteString);
+ if (!is_string_allowed) {
+ TF_LITE_ENSURE(context, input1->type != kTfLiteString);
+ }
// Currently only support tensors have the same type.
TF_LITE_ENSURE_TYPES_EQ(context, input1->type, input2->type);
output->type = kTfLiteBool;
@@ -54,6 +57,15 @@
return context->ResizeTensor(context, output, output_size);
}
+TfLiteStatus ComparisonPrepare(TfLiteContext* context, TfLiteNode* node) {
+ return ComparisonPrepareCommon(context, node, false);
+}
+
+TfLiteStatus ComparisonPrepareStringAllowed(TfLiteContext* context,
+ TfLiteNode* node) {
+ return ComparisonPrepareCommon(context, node, true);
+}
+
template <typename input_dtype, reference_ops::ComparisonFn<int32> opname>
void ComparisonQuantized(const TfLiteTensor* input1, const TfLiteTensor* input2,
TfLiteTensor* output, bool requires_broadcast) {
@@ -108,6 +120,21 @@
GetTensorShape(output), GetTensorData<bool>(output));
}
+template <bool (*opname)(const StringRef&, const StringRef&)>
+void ComparisonString(const TfLiteTensor* input1, const TfLiteTensor* input2,
+ TfLiteTensor* output, bool requires_broadcast) {
+ bool* output_data = GetTensorData<bool>(output);
+ if (requires_broadcast) {
+ reference_ops::BroadcastComparison4DSlowStringImpl<opname>(
+ GetTensorShape(input1), input1, GetTensorShape(input2), input2,
+ GetTensorShape(output), output_data);
+ } else {
+ reference_ops::ComparisonStringImpl<opname>(
+ GetTensorShape(input1), input1, GetTensorShape(input2), input2,
+ GetTensorShape(output), output_data);
+ }
+}
+
TfLiteStatus EqualEval(TfLiteContext* context, TfLiteNode* node) {
const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
@@ -138,9 +165,14 @@
ComparisonQuantized<int8_t, reference_ops::EqualFn>(
input1, input2, output, requires_broadcast);
break;
+ case kTfLiteString:
+ ComparisonString<reference_ops::StringRefEqualFn>(input1, input2, output,
+ requires_broadcast);
+ break;
default:
context->ReportError(
- context, "Does not support type %d, requires bool|float|int|uint8",
+ context,
+ "Does not support type %d, requires bool|float|int|uint8|string",
input1->type);
return kTfLiteError;
}
@@ -177,9 +209,14 @@
ComparisonQuantized<int8_t, reference_ops::NotEqualFn>(
input1, input2, output, requires_broadcast);
break;
+ case kTfLiteString:
+ ComparisonString<reference_ops::StringRefNotEqualFn>(
+ input1, input2, output, requires_broadcast);
+ break;
default:
context->ReportError(
- context, "Does not support type %d, requires bool|float|int|uint8",
+ context,
+ "Does not support type %d, requires bool|float|int|uint8|string",
input1->type);
return kTfLiteError;
}
@@ -330,14 +367,15 @@
} // namespace comparisons
TfLiteRegistration* Register_EQUAL() {
- static TfLiteRegistration r = {
- nullptr, nullptr, comparisons::ComparisonPrepare, comparisons::EqualEval};
+ static TfLiteRegistration r = {nullptr, nullptr,
+ comparisons::ComparisonPrepareStringAllowed,
+ comparisons::EqualEval};
return &r;
}
TfLiteRegistration* Register_NOT_EQUAL() {
static TfLiteRegistration r = {nullptr, nullptr,
- comparisons::ComparisonPrepare,
+ comparisons::ComparisonPrepareStringAllowed,
comparisons::NotEqualEval};
return &r;
}
diff --git a/tensorflow/lite/kernels/comparisons_test.cc b/tensorflow/lite/kernels/comparisons_test.cc
index 0fc49ea..986600c 100644
--- a/tensorflow/lite/kernels/comparisons_test.cc
+++ b/tensorflow/lite/kernels/comparisons_test.cc
@@ -125,6 +125,20 @@
EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 4));
}
+TEST(ComparisonsTest, EqualString) {
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ ComparisonOpModel model({1, 1, 1, 4, 1}, {1, 1, 1, 4, 1}, TensorType_STRING,
+ BuiltinOperator_EQUAL);
+ model.PopulateTensor<std::string>(model.input1(), {"A", "B", "C", "D"});
+ model.PopulateTensor<std::string>(model.input2(), {"A", "C", "B", "D"});
+ model.Invoke();
+
+ EXPECT_THAT(model.GetOutput(), ElementsAre(true, false, false, true));
+ EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 4, 1));
+}
+
TEST(ComparisonsTest, EqualBroadcast) {
ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 1}, TensorType_INT32,
BuiltinOperator_EQUAL);
@@ -148,6 +162,20 @@
EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 2, 4));
}
+TEST(ComparisonsTest, EqualBroadcastString) {
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 1}, TensorType_STRING,
+ BuiltinOperator_EQUAL);
+ model.PopulateTensor<std::string>(model.input1(), {"A", "B", "A", "B"});
+ model.PopulateTensor<std::string>(model.input2(), {"A"});
+ model.Invoke();
+
+ EXPECT_THAT(model.GetOutput(), ElementsAre(true, false, true, false));
+ EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 4));
+}
+
TEST(ComparisonsTest, NotEqualBool) {
ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 4}, TensorType_BOOL,
BuiltinOperator_NOT_EQUAL);
@@ -181,6 +209,20 @@
EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 4));
}
+TEST(ComparisonsTest, NotEqualString) {
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ ComparisonOpModel model({1, 1, 1, 1, 4}, {1, 1, 1, 1, 4}, TensorType_STRING,
+ BuiltinOperator_NOT_EQUAL);
+ model.PopulateTensor<std::string>(model.input1(), {"A", "B", "C", "D"});
+ model.PopulateTensor<std::string>(model.input2(), {"A", "C", "B", "D"});
+ model.Invoke();
+
+ EXPECT_THAT(model.GetOutput(), ElementsAre(false, true, true, false));
+ EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 1, 4));
+}
+
TEST(ComparisonsTest, NotEqualBroadcast) {
ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 1}, TensorType_INT32,
BuiltinOperator_NOT_EQUAL);
@@ -204,6 +246,20 @@
EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 2, 4));
}
+TEST(ComparisonsTest, NotEqualBroadcastString) {
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 1}, TensorType_STRING,
+ BuiltinOperator_NOT_EQUAL);
+ model.PopulateTensor<std::string>(model.input1(), {"A", "B", "A", "B"});
+ model.PopulateTensor<std::string>(model.input2(), {"A"});
+ model.Invoke();
+
+ EXPECT_THAT(model.GetOutput(), ElementsAre(false, true, false, true));
+ EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 1, 1, 4));
+}
+
TEST(ComparisonsTest, GreaterFloat) {
ComparisonOpModel model({1, 1, 1, 4}, {1, 1, 1, 4}, TensorType_FLOAT32,
BuiltinOperator_GREATER);
diff --git a/tensorflow/lite/kernels/conv.cc b/tensorflow/lite/kernels/conv.cc
index 0bb7ca7..403adc7 100644
--- a/tensorflow/lite/kernels/conv.cc
+++ b/tensorflow/lite/kernels/conv.cc
@@ -320,9 +320,9 @@
// Check types. (We assume that UINT8 refers to quantized tensors)
TfLiteType input_type = input->type;
- TF_LITE_ENSURE(context, input_type == kTfLiteFloat32 ||
- input_type == kTfLiteUInt8 ||
- input_type == kTfLiteInt8);
+ TF_LITE_ENSURE(context,
+ input_type == kTfLiteFloat32 || input_type == kTfLiteUInt8 ||
+ input_type == kTfLiteInt8 || input_type == kTfLiteInt16);
TF_LITE_ENSURE_EQ(context, output->type, input_type);
const TfLiteTensor* bias = nullptr;
@@ -336,6 +336,11 @@
if (input_type == kTfLiteUInt8 || input_type == kTfLiteInt8) {
TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteInt32);
TF_LITE_ENSURE_EQ(context, bias->params.zero_point, 0);
+ } else if (input_type == kTfLiteInt16) {
+ TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteInt64);
+ TF_LITE_ENSURE_EQ(context, bias->params.zero_point, 0);
+ TF_LITE_ENSURE_EQ(context, input->params.zero_point, 0);
+ TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
} else {
TF_LITE_ENSURE_EQ(context, bias->type, input_type);
}
@@ -678,6 +683,42 @@
}
template <KernelType kernel_type>
+void EvalQuantizedPerChannel16x8(TfLiteContext* context, TfLiteNode* node,
+ TfLiteConvParams* params, OpData* data,
+ const TfLiteTensor* input,
+ const TfLiteTensor* filter,
+ const TfLiteTensor* bias, TfLiteTensor* output,
+ TfLiteTensor* im2col) {
+ ConvParams op_params;
+ op_params.input_offset = -input->params.zero_point;
+ op_params.output_offset = output->params.zero_point;
+ op_params.stride_height = params->stride_height;
+ op_params.stride_width = params->stride_width;
+ op_params.dilation_height_factor = params->dilation_height_factor;
+ op_params.dilation_width_factor = params->dilation_width_factor;
+ op_params.padding_values.height = data->padding.height;
+ op_params.padding_values.width = data->padding.width;
+ op_params.quantized_activation_min = data->output_activation_min;
+ op_params.quantized_activation_max = data->output_activation_max;
+
+ switch (kernel_type) {
+ case kGenericOptimized:
+ case kMultithreadOptimized:
+ case kCblasOptimized:
+ case kReference: {
+ reference_integer_ops::ConvPerChannel(
+ op_params, data->per_channel_output_multiplier.data(),
+ data->per_channel_output_shift.data(), GetTensorShape(input),
+ GetTensorData<int16>(input), GetTensorShape(filter),
+ GetTensorData<int8>(filter), GetTensorShape(bias),
+ GetTensorData<std::int64_t>(bias), GetTensorShape(output),
+ GetTensorData<int16>(output));
+ break;
+ }
+ }
+}
+
+template <KernelType kernel_type>
void EvalFloat(TfLiteContext* context, TfLiteNode* node,
TfLiteConvParams* params, OpData* data,
const TfLiteTensor* input, const TfLiteTensor* filter,
@@ -938,6 +979,10 @@
EvalQuantizedPerChannel<kernel_type>(context, node, params, data, input,
filter, bias, output, im2col);
break;
+ case kTfLiteInt16:
+ EvalQuantizedPerChannel16x8<kernel_type>(
+ context, node, params, data, input, filter, bias, output, im2col);
+ break;
default:
context->ReportError(context, "Type %s currently not supported.",
TfLiteTypeGetName(input->type));
@@ -957,6 +1002,8 @@
return EvalImpl<kernel_type, kTfLiteUInt8>(context, node);
case kTfLiteInt8:
return EvalImpl<kernel_type, kTfLiteInt8>(context, node);
+ case kTfLiteInt16:
+ return EvalImpl<kernel_type, kTfLiteInt16>(context, node);
default:
context->ReportError(context, "Type %d not currently supported.",
input->type);
diff --git a/tensorflow/lite/kernels/conv_test.cc b/tensorflow/lite/kernels/conv_test.cc
index 10e014d..8569809 100644
--- a/tensorflow/lite/kernels/conv_test.cc
+++ b/tensorflow/lite/kernels/conv_test.cc
@@ -70,7 +70,12 @@
input.scale * filter.per_channel_quantization_scales[i];
bias_zero_points[i] = 0;
}
- TensorData bias{TensorType_INT32,
+ tflite::TensorType bias_type = TensorType_INT32;
+ if (input.type == TensorType_INT16) {
+ // In case of 16-bit, the bias type is set to be int 64.
+ bias_type = TensorType_INT64;
+ }
+ TensorData bias{bias_type,
{bias_size},
/*min=*/0,
/*max=*/0,
diff --git a/tensorflow/lite/kernels/cpu_backend_context.cc b/tensorflow/lite/kernels/cpu_backend_context.cc
index 5128421..0fa4175 100644
--- a/tensorflow/lite/kernels/cpu_backend_context.cc
+++ b/tensorflow/lite/kernels/cpu_backend_context.cc
@@ -56,7 +56,7 @@
gemmlowp_context_(new gemmlowp::GemmContext) {
SetMaxNumThreads(kDefaultNumThreadpoolThreads);
#ifdef TFLITE_WITH_RUY_GEMV
- ruy_context_->cache_policy = ruy::kCacheLHSOnNarrowMul;
+ ruy_context_->set_cache_policy(ruy::CachePolicy::kCacheLHSOnNarrowMul);
#endif
}
@@ -66,7 +66,7 @@
const int target_num_threads =
max_num_threads > -1 ? max_num_threads : kDefaultNumThreadpoolThreads;
max_num_threads_ = target_num_threads;
- ruy_context_->max_num_threads = target_num_threads;
+ ruy_context_->set_max_num_threads(target_num_threads);
gemmlowp_context_->set_max_num_threads(target_num_threads);
}
diff --git a/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h b/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
index 3769479..a5bcccd 100644
--- a/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
+++ b/tensorflow/lite/kernels/cpu_backend_gemm_ruy.h
@@ -78,19 +78,8 @@
ruy::MulParams<AccumScalar, DstScalar> ruy_mul_params;
MakeRuyMulParams(params, &ruy_mul_params);
-// If Ruy is not selected intentionally (TFLITE_WITH_RUY not defined)
-// and GEMMLOWP_NEON is absent, we fall back to Ruy for some quantized
-// kernels. Some Ruy paths are still experimental, so we restrict to reference
-// code in that case.
-#if !defined(TFLITE_WITH_RUY) && !defined(GEMMLOWP_NEON)
- constexpr ruy::Path kRuyPath =
- ruy::Path::kReference | ruy::Path::kStandardCpp;
-#else
- constexpr ruy::Path kRuyPath = ruy::kAllPaths;
-#endif
-
- ruy::Mul<kRuyPath>(ruy_lhs, ruy_rhs, ruy_mul_params, context->ruy_context(),
- &ruy_dst);
+ ruy::Mul(ruy_lhs, ruy_rhs, ruy_mul_params, context->ruy_context(),
+ &ruy_dst);
}
};
diff --git a/tensorflow/lite/kernels/cpu_backend_threadpool.h b/tensorflow/lite/kernels/cpu_backend_threadpool.h
index ff03d37..39eafd5 100644
--- a/tensorflow/lite/kernels/cpu_backend_threadpool.h
+++ b/tensorflow/lite/kernels/cpu_backend_threadpool.h
@@ -37,7 +37,8 @@
void Execute(int tasks_count, TaskType* tasks,
CpuBackendContext* cpu_backend_context) {
TFLITE_DCHECK_LE(tasks_count, cpu_backend_context->max_num_threads());
- cpu_backend_context->ruy_context()->workers_pool.Execute(tasks_count, tasks);
+ cpu_backend_context->ruy_context()->mutable_thread_pool()->Execute(
+ tasks_count, tasks);
}
#else // not TFLITE_WITH_RUY
diff --git a/tensorflow/lite/kernels/expand_dims_test.cc b/tensorflow/lite/kernels/expand_dims_test.cc
index eba5b88..5bb1d76 100644
--- a/tensorflow/lite/kernels/expand_dims_test.cc
+++ b/tensorflow/lite/kernels/expand_dims_test.cc
@@ -26,8 +26,8 @@
using ::testing::ElementsAreArray;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
template <typename InputType>
@@ -36,7 +36,7 @@
ExpandDimsOpModel(int axis, std::initializer_list<int> input_shape,
std::initializer_list<InputType> input_data,
TestType input_tensor_types) {
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
input_ = AddInput(GetTensorType<InputType>());
axis_ = AddInput(TensorType_INT32);
} else {
@@ -50,7 +50,7 @@
BuildInterpreter({input_shape, {1}});
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
PopulateTensor<InputType>(input_, input_data);
PopulateTensor<int32_t>(axis_, {axis});
}
@@ -69,18 +69,18 @@
template <typename T>
class ExpandDimsOpTest : public ::testing::Test {
public:
- static std::vector<TestType> _range_;
+ static std::vector<TestType> range_;
};
template <>
-std::vector<TestType> ExpandDimsOpTest<TestType>::_range_{TestType::CONST,
- TestType::DYNAMIC};
+std::vector<TestType> ExpandDimsOpTest<TestType>::range_{TestType::kConst,
+ TestType::kDynamic};
using DataTypes = ::testing::Types<float, int8_t, int16_t, int32_t>;
TYPED_TEST_SUITE(ExpandDimsOpTest, DataTypes);
TYPED_TEST(ExpandDimsOpTest, PositiveAxis) {
- for (TestType test_type : ExpandDimsOpTest<TestType>::_range_) {
+ for (TestType test_type : ExpandDimsOpTest<TestType>::range_) {
std::initializer_list<TypeParam> values = {-1, 1, -2, 2};
ExpandDimsOpModel<TypeParam> axis_0(0, {2, 2}, values, test_type);
@@ -101,7 +101,7 @@
}
TYPED_TEST(ExpandDimsOpTest, NegativeAxis) {
- for (TestType test_type : ExpandDimsOpTest<TestType>::_range_) {
+ for (TestType test_type : ExpandDimsOpTest<TestType>::range_) {
std::initializer_list<TypeParam> values = {-1, 1, -2, 2};
ExpandDimsOpModel<TypeParam> m(-1, {2, 2}, values, test_type);
@@ -115,7 +115,7 @@
std::initializer_list<std::string> values = {"abc", "de", "fghi"};
// this test will fail on TestType::CONST
- ExpandDimsOpModel<std::string> m(0, {3}, values, TestType::DYNAMIC);
+ ExpandDimsOpModel<std::string> m(0, {3}, values, TestType::kDynamic);
m.Invoke();
EXPECT_THAT(m.GetValues(), ElementsAreArray(values));
EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({1, 3}));
diff --git a/tensorflow/lite/kernels/fully_connected_test.cc b/tensorflow/lite/kernels/fully_connected_test.cc
index fbc02dd..6eda657 100644
--- a/tensorflow/lite/kernels/fully_connected_test.cc
+++ b/tensorflow/lite/kernels/fully_connected_test.cc
@@ -713,11 +713,13 @@
/*activation_func=*/ActivationFunctionType_NONE, weights_format);
std::mt19937 random_engine;
- std::uniform_int_distribution<uint8_t> weights_dist;
+ // Some compilers don't support uint8_t for uniform_distribution.
+ std::uniform_int_distribution<uint32_t> weights_dist(
+ 0, std::numeric_limits<uint8_t>::max());
std::vector<float> weights_data(input_depth * output_depth);
for (auto& w : weights_data) {
- uint8_t q = weights_dist(random_engine);
+ uint8_t q = static_cast<uint8_t>(weights_dist(random_engine));
w = (q - kWeightsZeroPoint) * kWeightsScale;
}
@@ -739,10 +741,12 @@
LOG(FATAL) << "Unhandled weights format";
}
- std::uniform_int_distribution<uint8_t> input_dist;
+ // Some compilers don't support uint8_t for uniform_distribution.
+ std::uniform_int_distribution<uint32_t> input_dist(
+ 0, std::numeric_limits<uint8_t>::max());
std::vector<float> input_data(input_depth * batches);
for (auto& i : input_data) {
- uint8_t q = input_dist(random_engine);
+ uint8_t q = static_cast<uint8_t>(input_dist(random_engine));
i = (q - kInputZeroPoint) * kInputScale;
}
diff --git a/tensorflow/lite/kernels/hashtable/BUILD b/tensorflow/lite/kernels/hashtable/BUILD
new file mode 100644
index 0000000..4ec3abe
--- /dev/null
+++ b/tensorflow/lite/kernels/hashtable/BUILD
@@ -0,0 +1,54 @@
+load("//tensorflow/lite/micro:build_def.bzl", "cc_library")
+
+package(
+ default_visibility = [
+ "//visibility:public",
+ ],
+ licenses = ["notice"], # Apache 2.0
+)
+
+cc_library(
+ name = "hashtable_op_kernels",
+ srcs = [
+ "hashtable.cc",
+ "hashtable_find.cc",
+ "hashtable_import.cc",
+ "hashtable_ops.cc",
+ "hashtable_size.cc",
+ ],
+ hdrs = [
+ "hashtable_ops.h",
+ ],
+ deps = [
+ "//tensorflow/lite:framework",
+ "//tensorflow/lite/c:common",
+ "//tensorflow/lite/core/api",
+ "//tensorflow/lite/experimental/resource",
+ "//tensorflow/lite/kernels:kernel_util",
+ "//tensorflow/lite/kernels:op_macros",
+ "//tensorflow/lite/kernels/internal:tensor",
+ "//tensorflow/lite/schema:schema_fbs",
+ "@flatbuffers",
+ ],
+)
+
+cc_test(
+ name = "hashtable_op_test",
+ size = "small",
+ srcs = [
+ "hashtable_ops_test.cc",
+ ],
+ deps = [
+ ":hashtable_op_kernels", # buildcleaner: keep
+ "//tensorflow/lite:framework",
+ "//tensorflow/lite/core/api",
+ "//tensorflow/lite/experimental/resource",
+ "//tensorflow/lite/kernels:test_main",
+ "//tensorflow/lite/kernels:test_util",
+ "//tensorflow/lite/kernels/internal:tensor",
+ "//tensorflow/lite/testing:util",
+ "@com_google_absl//absl/memory",
+ "@com_google_absl//absl/strings",
+ "@flatbuffers",
+ ],
+)
diff --git a/tensorflow/lite/kernels/hashtable/README.md b/tensorflow/lite/kernels/hashtable/README.md
new file mode 100644
index 0000000..77076a9
--- /dev/null
+++ b/tensorflow/lite/kernels/hashtable/README.md
@@ -0,0 +1,190 @@
+# How to use TF Lookup ops in TFLite
+
+The objective of this file is to provide examples to demonstrate how to use TF
+Lookup ops in TFLite.
+
+## Supported Tensorflow Lookup ops in TFLite
+
+Here is the supported status of TensorFlow Lookup ops.
+
+<table>
+ <tr>
+ <td><strong><em>TF Python lookup ops</em></strong>
+ </td>
+ <td colspan="5" ><strong><em>Supported status</em></strong>
+ </td>
+ </tr>
+ <tr>
+ <td rowspan="2" >tf.lookup.StaticHashTable
+ </td>
+ <td rowspan="2" colspan="5" >Supported only with tensor initializers.
+<p>
+Supported mapping type: string → int64, int64 → string
+ </td>
+ </tr>
+ <tr>
+ </tr>
+ <tr>
+ <td rowspan="2" >tf.lookup.Hashtable
+ </td>
+ <td rowspan="2" colspan="5" >Supported only with tensor initializers.
+<p>
+Supported mapping type: string → int64, int64 → string
+ </td>
+ </tr>
+ <tr>
+ </tr>
+ <tr>
+ <td rowspan="2" >tf.lookup.index_to_string_table_from_tensor
+ </td>
+ <td rowspan="2" colspan="5" >Supported.
+ </td>
+ </tr>
+ <tr>
+ </tr>
+ <tr>
+ <td rowspan="2" >tf.lookup.index_table_from_tensor
+ </td>
+ <td rowspan="2" colspan="5" >Supported natively when num_oov_bukcets=0 and dtype=dtypes.string.
+<p>
+For the oov concept, you will need a <a href="https://www.tensorflow.org/lite/guide/ops_select" title="Select TensorFlow operators to use in TensorFlow Lite">Flex delegate</a>.
+ </td>
+ </tr>
+ <tr>
+ </tr>
+ <tr>
+ <td>tf.lookup.StaticVocabularyTable
+ </td>
+ <td colspan="5" >Supported but you will need a <a href="https://www.tensorflow.org/lite/guide/ops_select" title="Select TensorFlow operators to use in TensorFlow Lite">Flex delegate</a>.
+<p>
+Use tf.index_table_from_tensor or tf.index_to_string_table_from_tensor instead if possible if you don’t want to use <a href="https://www.tensorflow.org/lite/guide/ops_select" title="Select TensorFlow operators to use in TensorFlow Lite">Flex delegate</a>.
+ </td>
+ </tr>
+ <tr>
+ <td>tf.lookup.experimental.DenseHashTable
+<p>
+tf.contrib.lookup.MutableHashTable
+<p>
+tf.contrib.lookup.MutableDenseHashTable
+ </td>
+ <td colspan="5" >Not supported yet.
+ </td>
+ </tr>
+ <tr>
+ <td>tf.lookup.IdTableWithHashBuckets
+ </td>
+ <td colspan="5" >Supported but you need a <a href="https://www.tensorflow.org/lite/guide/ops_select" title="Select TensorFlow operators to use in TensorFlow Lite">Flex delegate</a>.
+ </td>
+ </tr>
+</table>
+
+
+
+## Python Sample code
+
+Here, you can find the Python sample code:
+
+
+
+* Static hash table (string → int64)
+
+```
+int64_values = tf.constant([1, 2, 3], dtype=tf.int64)
+string_values = tf.constant(['bar', 'foo', 'baz'], dtype=tf.string)
+
+initializer = tf.lookup.KeyValueTensorInitializer(string_values, int64_values)
+table = tf.lookup.StaticHashTable(initializer, 4)
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ input_string_tensor = tf.compat.v1.placeholder(tf.string, shape=[1])
+ out_int64_tensor = table.lookup(input_string_tensor)
+```
+
+* Static hash table, initialized from a file (string → int64)
+
+```
+with open('/tmp/vocab.file', 'r') as f:
+ words = f.read().splitlines()
+
+string_values = tf.constant(words, dtype=tf.string)
+
+initializer = tf.lookup.KeyValueTensorInitializer(string_values, int64_values)
+table = tf.lookup.StaticHashTable(initializer, 4)
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ input_string_tensor = tf.placeholder(tf.string, shape=[1])
+ out_int64_tensor = table.lookup(input_string_tensor)
+```
+
+* Index table (string → int64)
+
+```
+UNK_ID = -1
+vocab = tf.constant(["emerson", "lake", "palmer"])
+vocab_table = tf.lookup.index_table_from_tensor(vocab, default_value=UNK_ID)
+
+input_tensor = tf.compat.v1.placeholder(tf.string, shape=[5])
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ out_tensor = vocab_table.lookup(input_tensor)
+```
+
+* Index table, initialized from a file (string → int64)
+
+```
+with open('/tmp/vocab.file', 'r') as f:
+ words = f.read().splitlines()
+
+UNK_ID = -1
+vocab = tf.constant(words)
+vocab_table = tf.lookup.index_table_from_tensor(vocab, default_value=UNK_ID)
+
+input_tensor = tf.compat.v1.placeholder(tf.string, shape=[5])
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ out_tensor = vocab_table.lookup(input_tensor)
+```
+
+* Index to string table (int64 → string)
+
+```
+UNK_WORD = "unknown"
+vocab = tf.constant(["emerson", "lake", "palmer"])
+vocab_table = tf.lookup.index_to_string_table_from_tensor(vocab, default_value=UNK_WORD)
+
+input_tensor = tf.compat.v1.placeholder(tf.int64, shape=[1])
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ out_tensor = vocab_table.lookup(input_tensor)
+```
+
+* Index to string table, initialized from a file (int64 → string)
+
+```
+with open('/tmp/vocab.file', 'r') as f:
+ words = f.read().splitlines()
+
+UNK_WORD = "unknown"
+vocab = tf.constant(words)
+vocab_table = tf.lookup.index_to_string_table_from_tensor(vocab, default_value=UNK_WORD)
+
+input_tensor = tf.compat.v1.placeholder(tf.int64, shape=[1])
+
+with tf.control_dependencies([tf.initializers.tables_initializer()]):
+ out_tensor = vocab_table.lookup(input_tensor)
+```
+
+## How to Include Hashtable ops in your TFLite.
+
+Currently, hashtable ops are not included in the builtin op set. You need to add
+hashtable ops manually by including the following dependency:
+
+`"//tensorflow/lite/kernels/hashtable:hashtable_op_kernels"`
+
+And then, your op resolver should add them like the following statements:
+
+
+```
+ // Add hashtable op handlers.
+ tflite::ops::custom::AddHashtableOps(&resolver);
+```
diff --git a/tensorflow/lite/experimental/kernels/hashtable.cc b/tensorflow/lite/kernels/hashtable/hashtable.cc
similarity index 100%
rename from tensorflow/lite/experimental/kernels/hashtable.cc
rename to tensorflow/lite/kernels/hashtable/hashtable.cc
diff --git a/tensorflow/lite/experimental/kernels/hashtable_find.cc b/tensorflow/lite/kernels/hashtable/hashtable_find.cc
similarity index 100%
rename from tensorflow/lite/experimental/kernels/hashtable_find.cc
rename to tensorflow/lite/kernels/hashtable/hashtable_find.cc
diff --git a/tensorflow/lite/experimental/kernels/hashtable_import.cc b/tensorflow/lite/kernels/hashtable/hashtable_import.cc
similarity index 100%
rename from tensorflow/lite/experimental/kernels/hashtable_import.cc
rename to tensorflow/lite/kernels/hashtable/hashtable_import.cc
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops.cc b/tensorflow/lite/kernels/hashtable/hashtable_ops.cc
similarity index 95%
rename from tensorflow/lite/experimental/kernels/hashtable_ops.cc
rename to tensorflow/lite/kernels/hashtable/hashtable_ops.cc
index 5b5973e..29c932c 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops.cc
+++ b/tensorflow/lite/kernels/hashtable/hashtable_ops.cc
@@ -13,7 +13,7 @@
limitations under the License.
==============================================================================*/
-#include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
+#include "tensorflow/lite/kernels/hashtable/hashtable_ops.h"
namespace tflite {
namespace ops {
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops.h b/tensorflow/lite/kernels/hashtable/hashtable_ops.h
similarity index 85%
rename from tensorflow/lite/experimental/kernels/hashtable_ops.h
rename to tensorflow/lite/kernels/hashtable/hashtable_ops.h
index 125db2a..7ed4ab3 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops.h
+++ b/tensorflow/lite/kernels/hashtable/hashtable_ops.h
@@ -13,8 +13,8 @@
limitations under the License.
==============================================================================*/
-#ifndef TENSORFLOW_LITE_EXPERIMENTAL_KERNELS_HASHTABLE_OPS_H_
-#define TENSORFLOW_LITE_EXPERIMENTAL_KERNELS_HASHTABLE_OPS_H_
+#ifndef TENSORFLOW_LITE_KERNELS_HASHTABLE_HASHTABLE_OPS_H_
+#define TENSORFLOW_LITE_KERNELS_HASHTABLE_HASHTABLE_OPS_H_
#include "tensorflow/lite/mutable_op_resolver.h"
@@ -33,4 +33,4 @@
} // namespace ops
} // namespace tflite
-#endif // TENSORFLOW_LITE_EXPERIMENTAL_KERNELS_HASHTABLE_OPS_H_
+#endif // TENSORFLOW_LITE_KERNELS_HASHTABLE_HASHTABLE_OPS_H_
diff --git a/tensorflow/lite/experimental/kernels/hashtable_ops_test.cc b/tensorflow/lite/kernels/hashtable/hashtable_ops_test.cc
similarity index 99%
rename from tensorflow/lite/experimental/kernels/hashtable_ops_test.cc
rename to tensorflow/lite/kernels/hashtable/hashtable_ops_test.cc
index 797b7b3..f4a0d3c 100644
--- a/tensorflow/lite/experimental/kernels/hashtable_ops_test.cc
+++ b/tensorflow/lite/kernels/hashtable/hashtable_ops_test.cc
@@ -15,7 +15,6 @@
#include <initializer_list>
#include <vector>
-#include <gtest/gtest.h>
#include "absl/memory/memory.h"
#include "absl/strings/match.h"
#include "flatbuffers/flexbuffers.h" // from @flatbuffers
diff --git a/tensorflow/lite/experimental/kernels/hashtable_size.cc b/tensorflow/lite/kernels/hashtable/hashtable_size.cc
similarity index 100%
rename from tensorflow/lite/experimental/kernels/hashtable_size.cc
rename to tensorflow/lite/kernels/hashtable/hashtable_size.cc
diff --git a/tensorflow/lite/kernels/internal/BUILD b/tensorflow/lite/kernels/internal/BUILD
index e838dc9..1095301 100644
--- a/tensorflow/lite/kernels/internal/BUILD
+++ b/tensorflow/lite/kernels/internal/BUILD
@@ -504,6 +504,7 @@
":tensor",
":tensor_utils",
":types",
+ "//tensorflow/lite:string_util",
"//tensorflow/lite/c:common",
"//tensorflow/lite/kernels:op_macros",
"//tensorflow/lite/tools/optimize/sparsity:format_converter",
@@ -566,6 +567,7 @@
"//tensorflow/lite/kernels:op_macros",
"@ruy//ruy/profiler:instrumentation",
"//tensorflow/lite/tools/optimize/sparsity:format_converter",
+ "//tensorflow/lite:string_util",
] + select({
":haswell": tflite_deps_intel,
":ios_x86_64": tflite_deps_intel,
@@ -813,6 +815,7 @@
}),
linkstatic = 1,
deps = [
+ ":common",
":quantization_util",
":tensor_utils",
"//tensorflow/lite/c:common",
@@ -892,6 +895,23 @@
)
cc_test(
+ name = "conv_per_channel_quantized_16x8_test",
+ srcs = [
+ "conv_per_channel_quantized_16x8_test.cc",
+ ],
+ shard_count = 2,
+ deps = [
+ ":common",
+ ":optimized_base",
+ ":quantization_util",
+ ":reference_base",
+ ":test_util",
+ ":types",
+ "@com_google_googletest//:gtest_main",
+ ],
+)
+
+cc_test(
name = "resize_bilinear_test",
srcs = ["resize_bilinear_test.cc"],
deps = [
diff --git a/tensorflow/lite/kernels/internal/common.h b/tensorflow/lite/kernels/internal/common.h
index ff8cc6d..483e4d0 100644
--- a/tensorflow/lite/kernels/internal/common.h
+++ b/tensorflow/lite/kernels/internal/common.h
@@ -160,6 +160,27 @@
right_shift);
}
+inline int32 MultiplyByQuantizedMultiplier(int64_t x,
+ int32 quantized_multiplier,
+ int shift) {
+ // Inputs:
+ // - quantized_multiplier has fixed point at bit 31
+ // - shift is -31 to +7 (negative for right shift)
+ //
+ // Assumptions: The following input ranges are assumed
+ // - quantize_scale>=0 (the usual range is (1<<30) to (1>>31)-1)
+ // - scaling is chosen so final scaled result fits in int32
+ // - input x is in the range -(1<<47) <= x < (1<<47)
+ assert(quantized_multiplier >= 0);
+ assert(shift >= -31 && shift < 8);
+
+ int32_t reduced_multiplier = (quantized_multiplier + (1 << 15)) >> 16;
+ int total_shift = 15 - shift;
+ x = (x * (int64_t)reduced_multiplier) + ((int64_t)1 << (total_shift - 1));
+ int32_t result = x >> total_shift;
+ return result;
+}
+
template <typename T>
int CountLeadingZeros(T integer_input) {
static_assert(std::is_unsigned<T>::value,
@@ -199,6 +220,20 @@
#endif
}
+// Use "count leading zeros" helper functions to do a fast Floor(log_2(x)).
+template <typename Integer>
+inline Integer FloorLog2(Integer n) {
+ static_assert(std::is_integral<Integer>::value, "");
+ static_assert(std::is_signed<Integer>::value, "");
+ static_assert(sizeof(Integer) == 4 || sizeof(Integer) == 8, "");
+ TFLITE_CHECK_GT(n, 0);
+ if (sizeof(Integer) == 4) {
+ return 30 - CountLeadingSignBits(n);
+ } else {
+ return 62 - CountLeadingSignBits(n);
+ }
+}
+
// generate INT16 LUT for function(), e.g., table exp(x) and 1/(1+x) used in
// softmax
inline void gen_lut(const std::function<double(double)>& func, double min,
diff --git a/tensorflow/lite/kernels/internal/conv_per_channel_quantized_16x8_test.cc b/tensorflow/lite/kernels/internal/conv_per_channel_quantized_16x8_test.cc
new file mode 100644
index 0000000..562797b
--- /dev/null
+++ b/tensorflow/lite/kernels/internal/conv_per_channel_quantized_16x8_test.cc
@@ -0,0 +1,332 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include <stdio.h>
+#include <sys/types.h>
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <iterator>
+#include <limits>
+#include <string>
+#include <type_traits>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/reference/conv.h"
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/conv.h"
+#include "tensorflow/lite/kernels/internal/test_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace {
+
+void PickOutputMultiplier(
+ const ConvParams& params, const RuntimeShape& input_shape,
+ const int16* input_data, const RuntimeShape& filter_shape,
+ const int8* filter_data, const RuntimeShape& bias_shape,
+ const std::int64_t* bias_data, const RuntimeShape& output_shape,
+ float* output_multiplier) {
+ const int stride_width = params.stride_width;
+ const int stride_height = params.stride_height;
+ const int dilation_width_factor = params.dilation_width_factor;
+ const int dilation_height_factor = params.dilation_height_factor;
+ const int pad_width = params.padding_values.width;
+ const int pad_height = params.padding_values.height;
+
+ const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+ const int input_height = input_shape.Dims(1);
+ const int input_width = input_shape.Dims(2);
+ const int input_depth = input_shape.Dims(3);
+ const int filter_height = filter_shape.Dims(1);
+ const int filter_width = filter_shape.Dims(2);
+ const int output_height = output_shape.Dims(1);
+ const int output_width = output_shape.Dims(2);
+ const int output_depth = output_shape.Dims(3);
+
+ std::int64_t output_accu_min = std::numeric_limits<std::int64_t>::max();
+ std::int64_t output_accu_max = std::numeric_limits<std::int64_t>::min();
+
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ for (int output_channel = 0; output_channel < output_depth;
+ ++output_channel) {
+ const int in_x_origin = (out_x * stride_width) - pad_width;
+ const int in_y_origin = (out_y * stride_height) - pad_height;
+ std::int64_t acc = 0;
+ for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+ for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+ for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+ const int in_x = in_x_origin + dilation_width_factor * filter_x;
+ const int in_y =
+ in_y_origin + dilation_height_factor * filter_y;
+ // Zero padding by omitting the areas outside the image.
+ const bool is_point_inside_image =
+ (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+ (in_y < input_height);
+ if (is_point_inside_image) {
+ int32 input_val = input_data[Offset(input_shape, batch, in_y,
+ in_x, in_channel)];
+ int32 filter_val =
+ filter_data[Offset(filter_shape, output_channel, filter_y,
+ filter_x, in_channel)];
+ acc += static_cast<std::int64_t>(filter_val) *
+ static_cast<std::int64_t>(input_val);
+ }
+ }
+ }
+ }
+ if (bias_data) {
+ acc += bias_data[output_channel];
+ }
+ output_accu_max = std::max(acc, output_accu_max);
+ output_accu_min = std::min(acc, output_accu_min);
+ }
+ }
+ }
+ }
+
+ // Since int16 ranges from -32768 to 32767, we need to squeeze the accumulator
+ // min/max fit in those ranges correspondingly as much as possible.
+ if (std::abs(output_accu_max) > std::abs(output_accu_min)) {
+ *output_multiplier = 32767.0f / std::abs(output_accu_max);
+ } else {
+ *output_multiplier = 32768.0f / std::abs(output_accu_min);
+ }
+}
+
+void PickReasonableMultiplier(
+ const ConvParams& params, int output_activation_min,
+ int output_activation_max, int output_depth,
+ const RuntimeShape& input_shape_inference, const std::int16_t* input_data,
+ const RuntimeShape& filter_shape_inference, const std::int8_t* filter_data,
+ const RuntimeShape& bias_shape_inference, const std::int64_t* bias_data,
+ const RuntimeShape& output_shape_inference,
+ std::int32_t* output_multiplier_ptr, std::int32_t* output_shift_ptr,
+ std::int16_t* output_data) {
+ float output_multiplier;
+ PickOutputMultiplier(params, input_shape_inference, input_data,
+ filter_shape_inference, filter_data,
+ bias_shape_inference, bias_data, output_shape_inference,
+ &output_multiplier);
+
+ int base_multiplier;
+ int base_shift;
+ QuantizeMultiplier(output_multiplier, &base_multiplier, &base_shift);
+ for (int i = 0; i < output_depth; ++i) {
+ // multipliers typically range in [2^30 ; 2^31 - 1].
+ // Values in [0, 2^30 - 1] are normally unused, but harmless.
+ // Thus a good way to randomize multipliers is to subtract from them
+ // a random value smaller than 2^30 but still significant compared to it.
+ output_multiplier_ptr[i] = base_multiplier - (std::rand() % (1 << 26));
+ output_shift_ptr[i] = base_shift - 1 + (std::rand() % 4);
+ }
+}
+
+bool GenerateValidShapeConfigurations(
+ int filter_width, int filter_height, int dilation_width_factor,
+ int dilation_height_factor, RuntimeShape* input_shape_inference,
+ RuntimeShape* filter_shape_inference, RuntimeShape* output_shape_inference,
+ int* pad_width, int* pad_height, int* stride) {
+ const int batch = UniformRandomInt(1, 3);
+ const int input_depth = 8 * ExponentialRandomPositiveInt(0.9f, 10, 50);
+ const int input_width = UniformRandomInt(5, 50);
+ const int input_height = UniformRandomInt(5, 50);
+ *stride = UniformRandomInt(1, 2);
+ const bool test_pad = UniformRandomInt(0, 1);
+ const auto padding_type = test_pad ? PaddingType::kValid : PaddingType::kSame;
+
+ const int output_depth = 8 * ExponentialRandomPositiveInt(0.9f, 10, 50);
+
+ input_shape_inference->BuildFrom(
+ {batch, input_height, input_width, input_depth});
+
+ filter_shape_inference->BuildFrom(
+ {output_depth, filter_height, filter_width, input_depth});
+
+ EXPECT_TRUE(ComputeConvSizes(
+ *input_shape_inference, output_depth, filter_width, filter_height,
+ *stride, dilation_width_factor, dilation_height_factor, padding_type,
+ output_shape_inference, pad_width, pad_height));
+
+ return true;
+}
+
+void IntToFloat(std::vector<float>* d, std::vector<std::int8_t>* s) {
+ for (unsigned int i = 0; i < s->size(); i++) {
+ d->data()[i] = (float)s->data()[i];
+ }
+}
+
+void IntToFloat(std::vector<float>* d, std::vector<std::int64_t>* s) {
+ for (unsigned int i = 0; i < s->size(); i++) {
+ d->data()[i] = (float)s->data()[i];
+ }
+}
+
+void TryTestOneConvFilter(int test_num) {
+ const int filter_width = UniformRandomInt(2, 5);
+ const int filter_height = UniformRandomInt(2, 5);
+ std::cout << "Test number " << test_num << " (" << filter_width << ","
+ << filter_height << ")\n";
+ // We don't support dilations in the 3x3 filter.
+ const int dilation_width_factor = 1;
+ const int dilation_height_factor = 1;
+
+ const int output_activation_min = -32768;
+ const int output_activation_max = 32767;
+
+ RuntimeShape input_shape_inference;
+ RuntimeShape filter_shape_inference;
+ RuntimeShape output_shape_inference;
+ int pad_width, pad_height;
+ int stride;
+
+ // Keeps trying until we get valid shape/configurations for 3x3 filter case.
+ bool generated_valid_configurations_for_3x3_kernel = false;
+ while (!generated_valid_configurations_for_3x3_kernel) {
+ generated_valid_configurations_for_3x3_kernel =
+ GenerateValidShapeConfigurations(
+ filter_width, filter_height, dilation_width_factor,
+ dilation_height_factor, &input_shape_inference,
+ &filter_shape_inference, &output_shape_inference, &pad_width,
+ &pad_height, &stride);
+ }
+
+ const int output_depth = output_shape_inference.Dims(3);
+
+ RuntimeShape bias_shape_inference({1, 1, 1, output_depth});
+ const int input_buffer_size = input_shape_inference.FlatSize();
+ const int filter_buffer_size = filter_shape_inference.FlatSize();
+ const int output_buffer_size = output_shape_inference.FlatSize();
+ std::vector<std::int16_t> input_data(input_buffer_size);
+ std::vector<std::int8_t> filter_data(filter_buffer_size);
+ std::vector<std::int64_t> bias_data(output_depth);
+
+ if (test_num & 1) {
+ // Use high values samples to give large accumulator
+ FillRandom(&input_data, (std::int16_t)32700, (std::int16_t)32767);
+ FillRandom(&filter_data, (std::int8_t)120, (std::int8_t)127);
+ } else {
+ FillRandom(&input_data);
+ FillRandom(&filter_data);
+ }
+ for (int i = 0; i < output_depth; i++) {
+ bias_data.data()[i] = 0;
+ }
+
+ ConvParams params;
+ params.stride_width = stride;
+ params.stride_height = stride;
+ params.dilation_height_factor = dilation_height_factor;
+ params.dilation_width_factor = dilation_width_factor;
+ params.padding_values.width = pad_width;
+ params.padding_values.height = pad_height;
+ params.weights_offset = 0;
+ params.quantized_activation_min = output_activation_min;
+ params.quantized_activation_max = output_activation_max;
+ params.float_activation_max = (float)(1LL << 40);
+ params.float_activation_min = -params.float_activation_max;
+
+ std::vector<std::int16_t> reference_output_data(output_buffer_size);
+ std::vector<std::int16_t> neon_output_data(output_buffer_size);
+
+ std::vector<std::int32_t> output_multiplier(output_depth);
+ std::vector<std::int32_t> output_shift(output_depth);
+
+ // It's hard to come up with a right multiplier, random guess basically makes
+ // all the results saturated and becomes meaningfulless, so we first use
+ // reference impl to poke the min/max value of the accumulation, then use that
+ // value as a guided suggestion for us to populate meaningful mulitplier &
+ // shift.
+ PickReasonableMultiplier(
+ params, output_activation_min, output_activation_max, output_depth,
+ input_shape_inference, input_data.data(), filter_shape_inference,
+ filter_data.data(), bias_shape_inference, bias_data.data(),
+ output_shape_inference, output_multiplier.data(), output_shift.data(),
+ reference_output_data.data());
+
+ // The following tests compare referene impl and Neon general impl agrees,
+ // and reference impl loosely agrees with fast kernel since they use different
+ // rounding strategy.
+ reference_integer_ops::ConvPerChannel(
+ params, output_multiplier.data(), output_shift.data(),
+ input_shape_inference, input_data.data(), filter_shape_inference,
+ filter_data.data(), bias_shape_inference, bias_data.data(),
+ output_shape_inference, reference_output_data.data());
+
+ std::vector<float> input_data_float(input_buffer_size);
+ std::vector<float> filter_data_float(filter_buffer_size);
+ std::vector<float> bias_data_float(output_depth);
+ std::vector<float> output_data_float(output_buffer_size);
+
+ for (int i = 0; i < input_buffer_size; i++) {
+ input_data_float.data()[i] = (float)(input_data.data()[i]);
+ }
+ IntToFloat(&filter_data_float, &filter_data);
+ IntToFloat(&bias_data_float, &bias_data);
+ RuntimeShape im2col_shape;
+ float im2col_data;
+
+ reference_ops::Conv(params, input_shape_inference, input_data_float.data(),
+ filter_shape_inference, filter_data_float.data(),
+ bias_shape_inference, bias_data_float.data(),
+ output_shape_inference, output_data_float.data(),
+ im2col_shape, &im2col_data);
+
+ for (int n = 0; n < output_shape_inference.Dims(0); n++) {
+ for (int h = 0; h < output_shape_inference.Dims(1); h++) {
+ for (int w = 0; w < output_shape_inference.Dims(2); w++) {
+ for (int c = 0; c < output_shape_inference.Dims(3); c++) {
+ int offset = Offset(output_shape_inference, n, h, w, c);
+ float float_res = output_data_float.data()[offset];
+ int16 int16_res = reference_output_data.data()[offset];
+ int32 output_mul = output_multiplier.data()[c];
+ int shift = output_shift.data()[c];
+ float scale = (float)output_mul / (float)(1ULL << 31);
+ if (shift > 0) scale = scale * (float)(1 << shift);
+ if (shift < 0) scale = scale / (float)(1 << -shift);
+ int ref_res = floor(float_res * scale + 0.5);
+ if (ref_res < output_activation_min) ref_res = output_activation_min;
+ if (ref_res > output_activation_max) ref_res = output_activation_max;
+ int e = (ref_res - int16_res);
+ if (e < 0) e = -e;
+ if (e > 2) {
+ ADD_FAILURE() << "(" << n << ", " << h << ", " << w << ", " << c
+ << ")"
+ << " scale=" << output_mul << " shift=" << shift
+ << " res=" << int16_res
+ << " float=" << float_res * scale << " (" << float_res
+ << ", " << scale << ")";
+ EXPECT_TRUE(false);
+ }
+ }
+ }
+ }
+ }
+}
+
+TEST(QuantizedConvPerChannelTest, FastKernelTest) {
+ for (int i = 0; i < 30; ++i) {
+ TryTestOneConvFilter(i);
+ }
+}
+
+} // namespace
+} // namespace tflite
diff --git a/tensorflow/lite/kernels/internal/optimized/integer_ops/transpose_conv.h b/tensorflow/lite/kernels/internal/optimized/integer_ops/transpose_conv.h
index 123e0a0..36519dd 100644
--- a/tensorflow/lite/kernels/internal/optimized/integer_ops/transpose_conv.h
+++ b/tensorflow/lite/kernels/internal/optimized/integer_ops/transpose_conv.h
@@ -25,16 +25,17 @@
const ConvParams& params, const int32* output_multiplier,
const int32* output_shift, const RuntimeShape& input_shape,
const int8_t* input_data, const RuntimeShape& hwoi_ordered_filter_shape,
- const int8_t* hwoi_ordered_filter_data, const RuntimeShape& output_shape,
+ const int8_t* hwoi_ordered_filter_data, const RuntimeShape& bias_shape,
+ const int32* bias_data, const RuntimeShape& output_shape,
int8_t* output_data, const RuntimeShape& col2im_shape, int32_t* col2im_data,
int32_t* scratch_data, CpuBackendContext* cpu_backend_context) {
ruy::profiler::ScopeLabel label("TransposeConvV2/int8");
TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_EQ(hwoi_ordered_filter_shape.DimensionsCount(), 4);
- const int batch_size = input_shape.Dims(0);
TFLITE_DCHECK(col2im_data);
TFLITE_DCHECK(hwoi_ordered_filter_data);
+ const int batch_size = MatchingDim(input_shape, 0, output_shape, 0);
const int input_image_size = input_shape.Dims(1) * input_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
@@ -93,6 +94,9 @@
scratch_data_p += output_offset;
}
+ scratch_data_p = scratch_data;
+ optimized_ops::BiasAdd(scratch_data_p, bias_data, batch_size, output_height,
+ output_width, output_depth);
const int32_t output_min = std::numeric_limits<int8_t>::min();
const int32_t output_max = std::numeric_limits<int8_t>::max();
diff --git a/tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h b/tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h
index bc8b9b2..f206dfa 100644
--- a/tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h
+++ b/tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h
@@ -2946,6 +2946,18 @@
output_data, DimsToShape(im2col_dims), im2col_data);
}
+inline void TransposeConvV2(
+ const ConvParams& params, const RuntimeShape& input_shape,
+ const float* input_data, const RuntimeShape& hwoi_ordered_filter_shape,
+ const float* hwoi_ordered_filter_data, const RuntimeShape& output_shape,
+ float* output_data, const RuntimeShape& col2im_shape, float* col2im_data,
+ CpuBackendContext* cpu_backend_context) {
+ TransposeConvV2(params, input_shape, input_data, hwoi_ordered_filter_shape,
+ hwoi_ordered_filter_data, /*bias_shape*/ RuntimeShape(),
+ /*bias_data*/ nullptr, output_shape, output_data,
+ col2im_shape, col2im_data, cpu_backend_context);
+}
+
template <typename T>
void TransposeIm2col(const T* input_data, const Dims<4>& input_dims,
const Dims<4>& filter_dims, int stride_width,
diff --git a/tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc b/tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc
index dc2204e..32584fc 100644
--- a/tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc
+++ b/tensorflow/lite/kernels/internal/optimized/neon_tensor_utils.cc
@@ -39,7 +39,8 @@
// aligned_alloc is available (via cstdlib/stdlib.h) with C++17/C11.
#if __cplusplus >= 201703L || __STDC_VERSION__ >= 201112L
#if !defined(__ANDROID__) || __ANDROID_API__ >= 28
-#if !defined(__APPLE__) // Apple does not provide aligned_alloc.
+// Neither Apple nor Windows provide aligned_alloc.
+#if !defined(__APPLE__) && !defined(_WIN32)
#define TFLITE_USE_STD_ALIGNED_ALLOC
#endif
#endif
@@ -92,6 +93,32 @@
#endif
}
+// Empirically determined breakpoints on when to use CpuBackendGemm vs.
+// standard MatrixBatchVectorMultiplyAccumulate. Briefly, if the batch size
+// is above 8 and the device does not have sdot, use CpuBackendGemm. Otherwise,
+// for large batch sizes, it makes sense to use CpuBackendGemm if the matrix
+// is not extremely rectangular.
+bool UseCpuBackendGemm(int rows, int cols, int batch) {
+ if (!HasSdotInstruction()) {
+ return batch >= 8;
+ }
+ if (batch < 16) {
+ return false;
+ }
+ constexpr int kCpuBackendGemmThreshold = 2;
+ // Calculate "rectangularness" as a measure of how far from square the
+ // the LHS matrix is.
+ int row_rect = rows / cols;
+ int col_rect = cols / rows;
+ int rectangularness_lg2 =
+ row_rect > 0 ? FloorLog2(row_rect) : FloorLog2(col_rect);
+ int batch_lg2 = FloorLog2(batch);
+ // Large batch sizes move us above the threshold, but can be offset
+ // by significant rectangularness.
+ int batch_lg2_minus_rect_lg2 = batch_lg2 - rectangularness_lg2;
+ return batch_lg2_minus_rect_lg2 > kCpuBackendGemmThreshold;
+}
+
inline int32_t AccumulateNeonLane(const int32x4_t lane) {
#ifdef __aarch64__
return vaddvq_s32(lane);
@@ -1404,15 +1431,18 @@
int n_batch, float* __restrict__ result, const float* per_channel_scale,
const int32_t* input_offset, int32_t* scratch, int32_t* row_sums,
bool* compute_row_sums, CpuBackendContext* context) {
- if (input_offset == nullptr) {
#ifdef TFLITE_WITH_RUY_GEMV
- if (context) {
+ const bool use_cpu_backend_gemm = true;
+#else
+ const bool use_cpu_backend_gemm = UseCpuBackendGemm(m_rows, m_cols, n_batch);
+#endif
+ if (input_offset == nullptr) {
+ if (use_cpu_backend_gemm && context) {
NeonMatrixBatchVectorMultiplyAccumulate(matrix, m_rows, m_cols, vectors,
scaling_factors, n_batch, scratch,
result, context);
return;
}
-#endif
NeonMatrixBatchVectorMultiplyAccumulate(matrix, m_rows, m_cols, vectors,
scaling_factors, n_batch, result);
return;
@@ -1426,59 +1456,60 @@
}
}
-#ifdef TFLITE_WITH_RUY_GEMV
- if (context != nullptr && m_rows % 4 == 0) {
- const int32_t* bias = static_cast<const int32_t*>(nullptr);
- NeonCpuBackendGemm(vectors, bias, matrix, n_batch, m_cols, m_rows, 0,
- scratch, context);
+ if (use_cpu_backend_gemm) {
+ if (context != nullptr && m_rows % 4 == 0) {
+ const int32_t* bias = static_cast<const int32_t*>(nullptr);
+ NeonCpuBackendGemm(vectors, bias, matrix, n_batch, m_cols, m_rows, 0,
+ scratch, context);
- // Multiply by float scaling factors and write to result
- const int total_size = n_batch * m_rows;
- int i = 0;
- int32_t* scratch_ptr = scratch;
- for (; i <= total_size - 8; i += 8, result += 8) {
- float batch_scaling_factor0 = scaling_factors[i / m_rows];
- float batch_scaling_factor1 = scaling_factors[(i + 4) / m_rows];
- if (per_channel_scale) {
- batch_scaling_factor0 *= per_channel_scale[i % m_rows];
- batch_scaling_factor1 *= per_channel_scale[(i + 4) % m_rows];
+ // Multiply by float scaling factors and write to result
+ const int total_size = n_batch * m_rows;
+ int i = 0;
+ int32_t* scratch_ptr = scratch;
+ for (; i <= total_size - 8; i += 8, result += 8) {
+ float batch_scaling_factor0 = scaling_factors[i / m_rows];
+ float batch_scaling_factor1 = scaling_factors[(i + 4) / m_rows];
+ if (per_channel_scale) {
+ batch_scaling_factor0 *= per_channel_scale[i % m_rows];
+ batch_scaling_factor1 *= per_channel_scale[(i + 4) % m_rows];
+ }
+ const int batch_input_offset0 = -input_offset[i / m_rows];
+ const int batch_input_offset1 = -input_offset[(i + 4) / m_rows];
+ const float32x4_t scaling_factor0 = vdupq_n_f32(batch_scaling_factor0);
+ const float32x4_t scaling_factor1 = vdupq_n_f32(batch_scaling_factor1);
+ const int32x4_t input_offset0 = vdupq_n_s32(batch_input_offset0);
+ const int32x4_t input_offset1 = vdupq_n_s32(batch_input_offset1);
+ const int32x4_t row_sum0 = vld1q_s32(row_sums + (i % m_rows));
+ const int32x4_t row_sum1 = vld1q_s32(row_sums + ((i + 4) % m_rows));
+ const int32x4_t scratch_val0 = vld1q_s32(scratch_ptr + i);
+ const int32x4_t scratch_val1 = vld1q_s32(scratch_ptr + i + 4);
+ const int32x4_t dotprod0 =
+ vmlaq_s32(scratch_val0, row_sum0, input_offset0);
+ const int32x4_t dotprod1 =
+ vmlaq_s32(scratch_val1, row_sum1, input_offset1);
+ const float32x4_t float_val0 = vcvtq_f32_s32(dotprod0);
+ const float32x4_t float_val1 = vcvtq_f32_s32(dotprod1);
+ const float32x4_t result0 =
+ vmlaq_f32(vld1q_f32(result), float_val0, scaling_factor0);
+ const float32x4_t result1 =
+ vmlaq_f32(vld1q_f32(result + 4), float_val1, scaling_factor1);
+ vst1q_f32(result, result0);
+ vst1q_f32(result + 4, result1);
}
- const int batch_input_offset0 = -input_offset[i / m_rows];
- const int batch_input_offset1 = -input_offset[(i + 4) / m_rows];
- const float32x4_t scaling_factor0 = vdupq_n_f32(batch_scaling_factor0);
- const float32x4_t scaling_factor1 = vdupq_n_f32(batch_scaling_factor1);
- const int32x4_t input_offset0 = vdupq_n_s32(batch_input_offset0);
- const int32x4_t input_offset1 = vdupq_n_s32(batch_input_offset1);
- const int32x4_t row_sum0 = vld1q_s32(row_sums + (i % m_rows));
- const int32x4_t row_sum1 = vld1q_s32(row_sums + ((i + 4) % m_rows));
- const int32x4_t scratch_val0 = vld1q_s32(scratch_ptr + i);
- const int32x4_t scratch_val1 = vld1q_s32(scratch_ptr + i + 4);
- const int32x4_t dotprod0 =
- vmlaq_s32(scratch_val0, row_sum0, input_offset0);
- const int32x4_t dotprod1 =
- vmlaq_s32(scratch_val1, row_sum1, input_offset1);
- const float32x4_t float_val0 = vcvtq_f32_s32(dotprod0);
- const float32x4_t float_val1 = vcvtq_f32_s32(dotprod1);
- const float32x4_t result0 =
- vmlaq_f32(vld1q_f32(result), float_val0, scaling_factor0);
- const float32x4_t result1 =
- vmlaq_f32(vld1q_f32(result + 4), float_val1, scaling_factor1);
- vst1q_f32(result, result0);
- vst1q_f32(result + 4, result1);
- }
- scratch_ptr += i;
- for (; i < total_size; i++) {
- const float batch_scaling_factor = scaling_factors[i / m_rows];
- const int32_t zero_point = input_offset[i / m_rows];
- int32_t dotprod = *(scratch_ptr++);
- dotprod -= row_sums[i % m_rows] * zero_point;
- *result += dotprod * batch_scaling_factor;
- ++result;
+ scratch_ptr += i;
+ for (; i < total_size; i++) {
+ const float batch_scaling_factor = scaling_factors[i / m_rows];
+ const int32_t zero_point = input_offset[i / m_rows];
+ int32_t dotprod = *(scratch_ptr++);
+ dotprod -= row_sums[i % m_rows] * zero_point;
+ *result += dotprod * batch_scaling_factor;
+ ++result;
+ }
+ return;
}
- return;
}
-#endif
+
NeonMatrixBatchVectorMultiplyAccumulateImpl(
matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result,
per_channel_scale, input_offset, row_sums);
diff --git a/tensorflow/lite/kernels/internal/optimized/optimized_ops.h b/tensorflow/lite/kernels/internal/optimized/optimized_ops.h
index ce90737..6e1f805 100644
--- a/tensorflow/lite/kernels/internal/optimized/optimized_ops.h
+++ b/tensorflow/lite/kernels/internal/optimized/optimized_ops.h
@@ -30,6 +30,7 @@
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
#include "tensorflow/lite/kernels/internal/reference/add.h"
+#include "tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h"
#if defined(TF_LITE_USE_CBLAS) && defined(__APPLE__)
#include <Accelerate/Accelerate.h>
@@ -5588,20 +5589,38 @@
}
}
+template <typename T>
+void BiasAdd(T* im_data, const T* bias_data, const int batch_size,
+ const int height, const int width, const int depth) {
+ if (bias_data) {
+ for (int n = 0; n < batch_size; ++n) {
+ for (int h = 0; h < height; ++h) {
+ for (int w = 0; w < width; ++w) {
+ for (int d = 0; d < depth; ++d) {
+ im_data[d] += bias_data[d];
+ }
+ im_data += depth;
+ }
+ }
+ }
+ }
+}
+
// TransposeConvV2 expect the weights in HWOI order.
inline void TransposeConvV2(
const ConvParams& params, const RuntimeShape& input_shape,
const float* input_data, const RuntimeShape& hwoi_ordered_filter_shape,
- const float* hwoi_ordered_filter_data, const RuntimeShape& output_shape,
- float* output_data, const RuntimeShape& col2im_shape, float* col2im_data,
- CpuBackendContext* cpu_backend_context) {
+ const float* hwoi_ordered_filter_data, const RuntimeShape& bias_shape,
+ const float* bias_data, const RuntimeShape& output_shape,
+ float* const output_data, const RuntimeShape& col2im_shape,
+ float* col2im_data, CpuBackendContext* cpu_backend_context) {
ruy::profiler::ScopeLabel label("TransposeConvV2/float");
TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_EQ(hwoi_ordered_filter_shape.DimensionsCount(), 4);
- const int batch_size = input_shape.Dims(0);
TFLITE_DCHECK(col2im_data);
TFLITE_DCHECK(hwoi_ordered_filter_data);
+ const int batch_size = MatchingDim(input_shape, 0, output_shape, 0);
const int input_image_size = input_shape.Dims(1) * input_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
@@ -5653,6 +5672,9 @@
output_data_p);
output_data_p += output_offset;
}
+ output_data_p = output_data;
+ BiasAdd(output_data_p, bias_data, batch_size, output_height, output_width,
+ output_depth);
}
inline void Quantize(int32_t multiplier, int32_t shift, int32_t total_size,
@@ -5813,17 +5835,18 @@
inline void TransposeConvV2(
const ConvParams& params, const RuntimeShape& input_shape,
const uint8_t* input_data, const RuntimeShape& hwoi_ordered_filter_shape,
- const uint8_t* hwoi_ordered_filter_data, const RuntimeShape& output_shape,
+ const uint8_t* hwoi_ordered_filter_data, const RuntimeShape& bias_shape,
+ const int32* bias_data, const RuntimeShape& output_shape,
uint8_t* output_data, const RuntimeShape& col2im_shape,
int32_t* col2im_data, int32_t* scratch_data,
CpuBackendContext* cpu_backend_context) {
ruy::profiler::ScopeLabel label("TransposeConvV2/uint8");
TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_EQ(hwoi_ordered_filter_shape.DimensionsCount(), 4);
- const int batch_size = input_shape.Dims(0);
TFLITE_DCHECK(col2im_data);
TFLITE_DCHECK(hwoi_ordered_filter_data);
+ const int batch_size = MatchingDim(input_shape, 0, output_shape, 0);
const int input_image_size = input_shape.Dims(1) * input_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
@@ -5881,6 +5904,9 @@
scratch_data_p += output_offset;
}
+ scratch_data_p = scratch_data;
+ BiasAdd(scratch_data_p, bias_data, batch_size, output_height, output_width,
+ output_depth);
Quantize(params.output_multiplier, params.output_shift,
output_shape.FlatSize(), params.output_offset, scratch_data,
@@ -5890,13 +5916,21 @@
// Integer-only version of ResizeNearestNeighbor. Since scales are represented
// in fixed-point and thus approximated, |in_x| or |in_y| may differ from the
// reference version. Debug checks are in place to test if this occurs.
+// NOTE: If align_corners or half_pixel_centers is true, we use the reference
+// version.
inline void ResizeNearestNeighbor(
const tflite::ResizeNearestNeighborParams& op_params,
const RuntimeShape& unextended_input_shape, const uint8* input_data,
const RuntimeShape& output_size_shape, const int32* output_size_data,
const RuntimeShape& unextended_output_shape, uint8* output_data) {
- // Align corners = true is not supported.
- TFLITE_DCHECK(!op_params.align_corners);
+ if (op_params.align_corners || op_params.half_pixel_centers) {
+ // TODO(b/149823713): Add support for align_corners & half_pixel_centers in
+ // this kernel.
+ reference_ops::ResizeNearestNeighbor(
+ op_params, unextended_input_shape, input_data, output_size_shape,
+ output_size_data, unextended_output_shape, output_data);
+ return;
+ }
TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
diff --git a/tensorflow/lite/kernels/internal/reference/comparisons.h b/tensorflow/lite/kernels/internal/reference/comparisons.h
index 19a968e..379a20f 100644
--- a/tensorflow/lite/kernels/internal/reference/comparisons.h
+++ b/tensorflow/lite/kernels/internal/reference/comparisons.h
@@ -15,8 +15,10 @@
#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
+#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/string_util.h"
namespace tflite {
@@ -49,6 +51,18 @@
return lhs <= rhs;
}
+inline bool StringRefEqualFn(const StringRef& lhs, const StringRef& rhs) {
+ if (lhs.len != rhs.len) return false;
+ for (int i = 0; i < lhs.len; ++i) {
+ if (lhs.str[i] != rhs.str[i]) return false;
+ }
+ return true;
+}
+
+inline bool StringRefNotEqualFn(const StringRef& lhs, const StringRef& rhs) {
+ return !StringRefEqualFn(lhs, rhs);
+}
+
template <typename T>
using ComparisonFn = bool (*)(T, T);
@@ -64,6 +78,22 @@
}
}
+template <bool (*F)(const StringRef&, const StringRef&)>
+inline void ComparisonStringImpl(const RuntimeShape& input1_shape,
+ const TfLiteTensor* input1,
+ const RuntimeShape& input2_shape,
+ const TfLiteTensor* input2,
+ const RuntimeShape& output_shape,
+ bool* output_data) {
+ const int64_t flatsize =
+ MatchingFlatSize(input1_shape, input2_shape, output_shape);
+ for (int64_t i = 0; i < flatsize; ++i) {
+ const auto lhs = GetString(input1, i);
+ const auto rhs = GetString(input2, i);
+ output_data[i] = F(lhs, rhs);
+ }
+}
+
template <ComparisonFn<float> F>
inline void Comparison(const ComparisonParams& op_params,
const RuntimeShape& input1_shape,
@@ -105,35 +135,76 @@
}
}
+struct BroadcastComparison4DSlowCommon {
+ const RuntimeShape output_shape;
+ NdArrayDesc<4> desc1;
+ NdArrayDesc<4> desc2;
+};
+
+inline BroadcastComparison4DSlowCommon BroadcastComparison4DSlowPreprocess(
+ const RuntimeShape& unextended_input1_shape,
+ const RuntimeShape& unextended_input2_shape,
+ const RuntimeShape& unextended_output_shape) {
+ TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+ TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+ TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+ NdArrayDesc<4> desc1;
+ NdArrayDesc<4> desc2;
+ NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+ unextended_input2_shape, &desc1, &desc2);
+ return {RuntimeShape::ExtendedShape(4, unextended_output_shape), desc1,
+ desc2};
+}
+
template <typename T, ComparisonFn<T> F>
inline void BroadcastComparison4DSlowImpl(
const ComparisonParams& op_params,
const RuntimeShape& unextended_input1_shape, const T* input1_data,
const RuntimeShape& unextended_input2_shape, const T* input2_data,
const RuntimeShape& unextended_output_shape, bool* output_data) {
- TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
- TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
- TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
- const RuntimeShape output_shape =
- RuntimeShape::ExtendedShape(4, unextended_output_shape);
+ const BroadcastComparison4DSlowCommon dims =
+ BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+ unextended_input2_shape,
+ unextended_output_shape);
- NdArrayDesc<4> desc1;
- NdArrayDesc<4> desc2;
- NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
- unextended_input2_shape, &desc1, &desc2);
-
- for (int b = 0; b < output_shape.Dims(0); ++b) {
- for (int y = 0; y < output_shape.Dims(1); ++y) {
- for (int x = 0; x < output_shape.Dims(2); ++x) {
- for (int c = 0; c < output_shape.Dims(3); ++c) {
- output_data[Offset(output_shape, b, y, x, c)] =
- F(input1_data[SubscriptToIndex(desc1, b, y, x, c)],
- input2_data[SubscriptToIndex(desc2, b, y, x, c)]);
+ for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+ for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+ for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+ for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+ output_data[Offset(dims.output_shape, b, y, x, c)] =
+ F(input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)],
+ input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)]);
}
}
}
}
}
+
+template <bool (*F)(const StringRef&, const StringRef&)>
+inline void BroadcastComparison4DSlowStringImpl(
+ const RuntimeShape& unextended_input1_shape, const TfLiteTensor* input1,
+ const RuntimeShape& unextended_input2_shape, const TfLiteTensor* input2,
+ const RuntimeShape& unextended_output_shape, bool* output_data) {
+ const BroadcastComparison4DSlowCommon dims =
+ BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+ unextended_input2_shape,
+ unextended_output_shape);
+
+ for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+ for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+ for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+ for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+ const auto lhs =
+ GetString(input1, SubscriptToIndex(dims.desc1, b, y, x, c));
+ const auto rhs =
+ GetString(input2, SubscriptToIndex(dims.desc2, b, y, x, c));
+ output_data[Offset(dims.output_shape, b, y, x, c)] = F(lhs, rhs);
+ }
+ }
+ }
+ }
+}
+
template <ComparisonFn<float> F>
inline void BroadcastComparison4DSlow(const ComparisonParams& op_params,
const RuntimeShape& input1_shape,
@@ -153,16 +224,10 @@
const RuntimeShape& unextended_input1_shape, const T* input1_data,
const RuntimeShape& unextended_input2_shape, const T* input2_data,
const RuntimeShape& unextended_output_shape, bool* output_data) {
- TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
- TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
- TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
- const RuntimeShape output_shape =
- RuntimeShape::ExtendedShape(4, unextended_output_shape);
-
- NdArrayDesc<4> desc1;
- NdArrayDesc<4> desc2;
- NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
- unextended_input2_shape, &desc1, &desc2);
+ const BroadcastComparison4DSlowCommon dims =
+ BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+ unextended_input2_shape,
+ unextended_output_shape);
int left_shift = op_params.left_shift;
int32 input1_offset = op_params.input1_offset;
@@ -172,14 +237,16 @@
int32 input2_multiplier = op_params.input2_multiplier;
int input2_shift = op_params.input2_shift;
- for (int b = 0; b < output_shape.Dims(0); ++b) {
- for (int y = 0; y < output_shape.Dims(1); ++y) {
- for (int x = 0; x < output_shape.Dims(2); ++x) {
- for (int c = 0; c < output_shape.Dims(3); ++c) {
+ for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+ for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+ for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+ for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
const int32 input1_val =
- input1_offset + input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+ input1_offset +
+ input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)];
const int32 input2_val =
- input2_offset + input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+ input2_offset +
+ input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)];
const int32 shifted_input1_val = input1_val * (1 << left_shift);
const int32 shifted_input2_val = input2_val * (1 << left_shift);
const int32 scaled_input1_val =
@@ -188,7 +255,7 @@
const int32 scaled_input2_val =
MultiplyByQuantizedMultiplierSmallerThanOneExp(
shifted_input2_val, input2_multiplier, input2_shift);
- output_data[Offset(output_shape, b, y, x, c)] =
+ output_data[Offset(dims.output_shape, b, y, x, c)] =
F(scaled_input1_val, scaled_input2_val);
}
}
diff --git a/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h b/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h
index 4b101f7..9131c7d 100644
--- a/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h
+++ b/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h
@@ -122,6 +122,95 @@
}
}
+// Fixed-point per-channel-quantization convolution reference kernel.
+// 16-bit data and 8-bit filter
+inline void ConvPerChannel(
+ const ConvParams& params, const int32* output_multiplier,
+ const int32* output_shift, const RuntimeShape& input_shape,
+ const int16* input_data, const RuntimeShape& filter_shape,
+ const int8* filter_data, const RuntimeShape& bias_shape,
+ const std::int64_t* bias_data, const RuntimeShape& output_shape,
+ int16* output_data) {
+ // Get parameters.
+ const int stride_width = params.stride_width;
+ const int stride_height = params.stride_height;
+ const int dilation_width_factor = params.dilation_width_factor;
+ const int dilation_height_factor = params.dilation_height_factor;
+ const int pad_width = params.padding_values.width;
+ const int pad_height = params.padding_values.height;
+
+ // Set min and max value of the output.
+ const int32 output_activation_min = params.quantized_activation_min;
+ const int32 output_activation_max = params.quantized_activation_max;
+
+ // Sanity check.
+ TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+ TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+ TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+ TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+ const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+ const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+ const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+ if (bias_data) {
+ TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+ }
+
+ // Check dimensions of the tensors.
+ const int input_height = input_shape.Dims(1);
+ const int input_width = input_shape.Dims(2);
+ const int filter_height = filter_shape.Dims(1);
+ const int filter_width = filter_shape.Dims(2);
+ const int output_height = output_shape.Dims(1);
+ const int output_width = output_shape.Dims(2);
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+ const int in_x_origin = (out_x * stride_width) - pad_width;
+ const int in_y_origin = (out_y * stride_height) - pad_height;
+ std::int64_t acc = 0;
+ for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+ for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+ for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+ const int in_x = in_x_origin + dilation_width_factor * filter_x;
+ const int in_y =
+ in_y_origin + dilation_height_factor * filter_y;
+ // Zero padding by omitting the areas outside the image.
+ const bool is_point_inside_image =
+ (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+ (in_y < input_height);
+ if (is_point_inside_image) {
+ int32 input_val = input_data[Offset(input_shape, batch, in_y,
+ in_x, in_channel)];
+ int32 filter_val =
+ filter_data[Offset(filter_shape, out_channel, filter_y,
+ filter_x, in_channel)];
+ // Accumulate with 64 bits accumulator.
+ // int64 += int8 * int16 so the highest value we can
+ // get from each accumulation is [-127, 127] * ([-32768,
+ // 32767] -
+ // [-32768, 32767]), which is [-8322945, 8322945].
+ // log2(8322945) = 22.99.
+ acc += filter_val * input_val;
+ }
+ }
+ }
+ }
+ if (bias_data) {
+ acc += bias_data[out_channel];
+ }
+ int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+ acc, output_multiplier[out_channel], output_shift[out_channel]);
+ scaled_acc = std::max(scaled_acc, output_activation_min);
+ scaled_acc = std::min(scaled_acc, output_activation_max);
+ output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+ static_cast<int16_t>(scaled_acc);
+ }
+ }
+ }
+ }
+}
+
} // namespace reference_integer_ops
} // namespace tflite
diff --git a/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h b/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h
index 1ad6e20..422adc2 100644
--- a/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h
+++ b/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h
@@ -25,8 +25,9 @@
const ConvParams& params, const int32* output_multiplier,
const int32* output_shift, const RuntimeShape& input_shape,
const int8* input_data, const RuntimeShape& filter_shape,
- const int8* filter_data, const RuntimeShape& output_shape,
- int8* output_data, const RuntimeShape& im2col_shape, int8* im2col_data,
+ const int8* filter_data, const RuntimeShape& bias_shape,
+ const int32* bias_data, const RuntimeShape& output_shape, int8* output_data,
+ const RuntimeShape& im2col_shape, int8* im2col_data,
int32* scratch_buffer) {
const int stride_width = params.stride_width;
const int stride_height = params.stride_height;
@@ -41,6 +42,9 @@
const int batches = MatchingDim(input_shape, 0, output_shape, 0);
const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+ if (bias_data) {
+ TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+ }
const int input_height = input_shape.Dims(1);
const int input_width = input_shape.Dims(2);
const int filter_height = filter_shape.Dims(1);
@@ -99,6 +103,9 @@
for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
int32 acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x,
out_channel)];
+ if (bias_data) {
+ acc += bias_data[out_channel];
+ }
acc = MultiplyByQuantizedMultiplier(
acc, output_multiplier[out_channel], output_shift[out_channel]);
acc += output_offset;
diff --git a/tensorflow/lite/kernels/internal/reference/l2normalization.h b/tensorflow/lite/kernels/internal/reference/l2normalization.h
index d93eb13..00697c2 100644
--- a/tensorflow/lite/kernels/internal/reference/l2normalization.h
+++ b/tensorflow/lite/kernels/internal/reference/l2normalization.h
@@ -15,6 +15,7 @@
#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
+#include <algorithm>
#include <cmath>
#include "tensorflow/lite/c/common.h"
@@ -76,7 +77,9 @@
int32 rescaled_diff = MultiplyByQuantizedMultiplierSmallerThanOneExp(
128 * diff, inv_l2norm_multiplier, inv_l2norm_shift);
int32 unclamped_output_val = 128 + rescaled_diff;
- int32 output_val = std::min(255, std::max(0, unclamped_output_val));
+ int32 output_val =
+ std::min(static_cast<int32>(255),
+ std::max(static_cast<int32>(0), unclamped_output_val));
output_data[depth * i + c] = static_cast<uint8>(output_val);
}
}
diff --git a/tensorflow/lite/kernels/internal/reference/legacy_reference_ops.h b/tensorflow/lite/kernels/internal/reference/legacy_reference_ops.h
index 2148be4..f62c9bd 100644
--- a/tensorflow/lite/kernels/internal/reference/legacy_reference_ops.h
+++ b/tensorflow/lite/kernels/internal/reference/legacy_reference_ops.h
@@ -387,8 +387,20 @@
op_params.stride_height = stride_height;
TransposeConv(op_params, DimsToShape(input_dims), input_data,
- DimsToShape(filter_dims), filter_data, DimsToShape(output_dims),
- output_data, DimsToShape(im2col_dims), im2col_data);
+ DimsToShape(filter_dims), filter_data,
+ /*bias_shape*/ RuntimeShape(), /*bias*/ nullptr,
+ DimsToShape(output_dims), output_data, DimsToShape(im2col_dims),
+ im2col_data);
+}
+
+inline void TransposeConv(
+ const ConvParams& params, const RuntimeShape& input_shape,
+ const float* input_data, const RuntimeShape& filter_shape,
+ const float* filter_data, const RuntimeShape& output_shape,
+ float* output_data, const RuntimeShape& im2col_shape, float* im2col_data) {
+ TransposeConv(params, input_shape, input_data, filter_shape, filter_data,
+ /*bias_shape*/ RuntimeShape(), /*bias*/ nullptr, output_shape,
+ output_data, im2col_shape, im2col_data);
}
inline void FullyConnected(const float* input_data, const Dims<4>& input_dims,
diff --git a/tensorflow/lite/kernels/internal/reference/reference_ops.h b/tensorflow/lite/kernels/internal/reference/reference_ops.h
index e991d4e..f40b268 100644
--- a/tensorflow/lite/kernels/internal/reference/reference_ops.h
+++ b/tensorflow/lite/kernels/internal/reference/reference_ops.h
@@ -2048,7 +2048,8 @@
inline void TransposeConv(
const ConvParams& params, const RuntimeShape& input_shape,
const float* input_data, const RuntimeShape& filter_shape,
- const float* filter_data, const RuntimeShape& output_shape,
+ const float* filter_data, const RuntimeShape& bias_shape,
+ const float* bias_data, const RuntimeShape& output_shape,
float* output_data, const RuntimeShape& im2col_shape, float* im2col_data) {
const int stride_width = params.stride_width;
const int stride_height = params.stride_height;
@@ -2069,6 +2070,9 @@
const int filter_width = filter_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
+ if (bias_data) {
+ TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+ }
// Although transpose convolution simplifies to convolution with transposed
// weights for strides of 1, non-unitary striding complicates matters. To
@@ -2116,16 +2120,27 @@
}
}
}
+ if (bias_data) {
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+ output_data[Offset(output_shape, batch, out_y, out_x,
+ out_channel)] += bias_data[out_channel];
+ }
+ }
+ }
+ }
+ }
}
-inline void TransposeConv(const ConvParams& params,
- const RuntimeShape& input_shape,
- const uint8* input_data,
- const RuntimeShape& filter_shape,
- const uint8* filter_data,
- const RuntimeShape& output_shape, uint8* output_data,
- const RuntimeShape& im2col_shape, uint8* im2col_data,
- int32* scratch_buffer) {
+inline void TransposeConv(
+ const ConvParams& params, const RuntimeShape& input_shape,
+ const uint8* input_data, const RuntimeShape& filter_shape,
+ const uint8* filter_data, const RuntimeShape& bias_shape,
+ const int32* bias_data, const RuntimeShape& output_shape,
+ uint8* output_data, const RuntimeShape& im2col_shape, uint8* im2col_data,
+ int32* scratch_buffer) {
const int stride_width = params.stride_width;
const int stride_height = params.stride_height;
const int pad_width = params.padding_values.width;
@@ -2153,6 +2168,9 @@
const int32 output_activation_min = params.quantized_activation_min;
const int32 output_activation_max = params.quantized_activation_max;
TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+ if (bias_data) {
+ TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+ }
const int num_elements = output_shape.FlatSize();
// We need to initialize scratch_buffer to all 0s, as we apply the same
@@ -2194,14 +2212,25 @@
}
}
}
- for (int i = 0; i < num_elements; ++i) {
- int32 acc = scratch_buffer[i];
- acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
- acc += output_offset;
- // Clamp the output before converting back to uint8.
- acc = std::max(acc, output_activation_min);
- acc = std::min(acc, output_activation_max);
- output_data[i] = static_cast<uint8>(acc);
+ for (int batch = 0; batch < batches; ++batch) {
+ for (int out_y = 0; out_y < output_height; ++out_y) {
+ for (int out_x = 0; out_x < output_width; ++out_x) {
+ for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+ int32 acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+ out_channel)];
+ if (bias_data) {
+ acc += bias_data[out_channel];
+ }
+ int32 scaled_acc = MultiplyByQuantizedMultiplier(
+ acc, output_multiplier, output_shift);
+ scaled_acc += output_offset;
+ scaled_acc = std::max(scaled_acc, output_activation_min);
+ scaled_acc = std::min(scaled_acc, output_activation_max);
+ output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+ static_cast<uint8>(scaled_acc);
+ }
+ }
+ }
}
}
diff --git a/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h b/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h
index 25623ca..ed87863 100644
--- a/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h
+++ b/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h
@@ -23,14 +23,32 @@
namespace reference_ops {
+inline int32 GetNearestNeighbor(const int input_value, const int32 input_size,
+ const int32 output_size,
+ const bool align_corners,
+ const bool half_pixel_centers) {
+ const float scale =
+ (align_corners && output_size > 1)
+ ? (input_size - 1) / static_cast<float>(output_size - 1)
+ : input_size / static_cast<float>(output_size);
+ const float offset = half_pixel_centers ? 0.5f : 0.0f;
+ int32 output_value = std::min(
+ align_corners
+ ? static_cast<int32>(std::round((input_value + offset) * scale))
+ : static_cast<int32>(std::floor((input_value + offset) * scale)),
+ input_size - 1);
+ if (half_pixel_centers) {
+ output_value = std::max(static_cast<int32>(0), output_value);
+ }
+ return output_value;
+}
+
template <typename T>
inline void ResizeNearestNeighbor(
const tflite::ResizeNearestNeighborParams& op_params,
const RuntimeShape& unextended_input_shape, const T* input_data,
const RuntimeShape& output_size_shape, const int32* output_size_data,
const RuntimeShape& unextended_output_shape, T* output_data) {
- // Align corners = true is not supported.
- TFLITE_DCHECK(!op_params.align_corners);
TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
@@ -50,10 +68,6 @@
int32 output_height = output_size_data[0];
int32 output_width = output_size_data[1];
- // We use float to ensure agreement with the Tensorflow implementation.
- const float height_scale = static_cast<float>(input_height) / output_height;
- const float width_scale = static_cast<float>(input_width) / output_width;
-
const int col_offset = input_shape.Dims(3);
const int row_offset = input_shape.Dims(2) * col_offset;
const int batch_offset = input_shape.Dims(1) * row_offset;
@@ -62,12 +76,14 @@
T* output_ptr = output_data;
for (int b = 0; b < batches; ++b) {
for (int y = 0; y < output_height; ++y) {
- int32 in_y = std::min(static_cast<int32>(std::floor(y * height_scale)),
- input_height - 1);
+ int32 in_y = GetNearestNeighbor(y, input_height, output_height,
+ op_params.align_corners,
+ op_params.half_pixel_centers);
const T* y_input_ptr = input_ptr + in_y * row_offset;
for (int x = 0; x < output_width; ++x) {
- int32 in_x = std::min(static_cast<int32>(std::floor(x * width_scale)),
- input_width - 1);
+ int32 in_x = GetNearestNeighbor(x, input_width, output_width,
+ op_params.align_corners,
+ op_params.half_pixel_centers);
const T* x_input_ptr = y_input_ptr + in_x * col_offset;
memcpy(output_ptr, x_input_ptr, depth * sizeof(T));
output_ptr += depth;
diff --git a/tensorflow/lite/kernels/internal/resize_nearest_neighbor_test.cc b/tensorflow/lite/kernels/internal/resize_nearest_neighbor_test.cc
index 102ee04..4659d3a 100644
--- a/tensorflow/lite/kernels/internal/resize_nearest_neighbor_test.cc
+++ b/tensorflow/lite/kernels/internal/resize_nearest_neighbor_test.cc
@@ -30,8 +30,9 @@
const RuntimeShape& input_shape, const std::vector<T>& input_data,
const std::vector<int32>& output_size_data,
const RuntimeShape& output_shape,
- const std::vector<T>& expected_output_data) {
- ResizeNearestNeighborParams op_params{/*align_corners=*/false};
+ const std::vector<T>& expected_output_data, bool align_corners = false,
+ bool half_pixel_centers = false) {
+ ResizeNearestNeighborParams op_params{align_corners, half_pixel_centers};
RuntimeShape output_size_shape({1, 1, 1, 2});
std::vector<T> output_data(expected_output_data.size());
@@ -55,6 +56,30 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test2x2To1x1_AlignCorners) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<float> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {1, 1};
+ RuntimeShape output_shape = {1, 1, 1, 1};
+ std::vector<float> output_data = {1};
+
+ TestReferenceResizeNearestNeighbor(input_shape, input_data, output_size_data,
+ output_shape, output_data,
+ /*align_corners=*/true);
+}
+
+TEST(ResizeNearestNeighborReference, Test2x2To1x1_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<float> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {1, 1};
+ RuntimeShape output_shape = {1, 1, 1, 1};
+ std::vector<float> output_data = {4};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test2x2To3x3) {
RuntimeShape input_shape = {1, 2, 2, 1};
std::vector<uint8> input_data = {1, 2, 3, 4};
@@ -66,6 +91,30 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test2x2To3x3_AlignCorners) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<uint8> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {1, 3, 3, 1};
+ std::vector<uint8> output_data = {1, 2, 2, 3, 4, 4, 3, 4, 4};
+
+ TestReferenceResizeNearestNeighbor(input_shape, input_data, output_size_data,
+ output_shape, output_data,
+ /*align_corners=*/true);
+}
+
+TEST(ResizeNearestNeighborReference, Test2x2To3x3_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<uint8> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {1, 3, 3, 1};
+ std::vector<uint8> output_data = {1, 2, 2, 3, 4, 4, 3, 4, 4};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test3x3To2x2) {
RuntimeShape input_shape = {1, 3, 3, 1};
std::vector<float> input_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
@@ -77,6 +126,30 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test3x3To2x2_AlignCorners) {
+ RuntimeShape input_shape = {1, 3, 3, 1};
+ std::vector<float> input_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+ std::vector<int32> output_size_data = {2, 2};
+ RuntimeShape output_shape = {1, 2, 2, 1};
+ std::vector<float> output_data = {1, 3, 7, 9};
+
+ TestReferenceResizeNearestNeighbor(input_shape, input_data, output_size_data,
+ output_shape, output_data,
+ /*align_corners=*/true);
+}
+
+TEST(ResizeNearestNeighborReference, Test3x3To2x2_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 3, 3, 1};
+ std::vector<float> input_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+ std::vector<int32> output_size_data = {2, 2};
+ RuntimeShape output_shape = {1, 2, 2, 1};
+ std::vector<float> output_data = {1, 3, 7, 9};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test2x2To2x5) {
RuntimeShape input_shape = {1, 2, 2, 1};
std::vector<uint8> input_data = {1, 2, 3, 4};
@@ -88,6 +161,18 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test2x2To2x5_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<uint8> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {2, 5};
+ RuntimeShape output_shape = {1, 2, 5, 1};
+ std::vector<uint8> output_data = {1, 1, 2, 2, 2, 3, 3, 4, 4, 4};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test4x4To3x3) {
RuntimeShape input_shape = {1, 4, 4, 1};
std::vector<uint8> input_data = {1, 2, 3, 4, 5, 6, 7, 8,
@@ -100,6 +185,32 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test4x4To3x3_AlignCorners) {
+ RuntimeShape input_shape = {1, 4, 4, 1};
+ std::vector<uint8> input_data = {1, 2, 3, 4, 5, 6, 7, 8,
+ 9, 10, 11, 12, 13, 14, 15, 16};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {1, 3, 3, 1};
+ std::vector<uint8> output_data = {1, 3, 4, 9, 11, 12, 13, 15, 16};
+
+ TestReferenceResizeNearestNeighbor(input_shape, input_data, output_size_data,
+ output_shape, output_data,
+ /*align_corners=*/true);
+}
+
+TEST(ResizeNearestNeighborReference, Test4x4To3x3_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 4, 4, 1};
+ std::vector<uint8> input_data = {1, 2, 3, 4, 5, 6, 7, 8,
+ 9, 10, 11, 12, 13, 14, 15, 16};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {1, 3, 3, 1};
+ std::vector<uint8> output_data = {1, 3, 4, 9, 11, 12, 13, 15, 16};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test2x2To5x2) {
RuntimeShape input_shape = {1, 2, 2, 1};
std::vector<float> input_data = {1, 2, 3, 4};
@@ -111,6 +222,31 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test2x2To5x2_HalfPixelCenters) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<float> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {5, 2};
+ RuntimeShape output_shape = {1, 5, 2, 1};
+ std::vector<float> output_data = {1, 2, 1, 2, 3, 4, 3, 4, 3, 4};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
+TEST(ResizeNearestNeighborReference,
+ Test2x2To5x2_HalfPixelCenters_AlignCorners) {
+ RuntimeShape input_shape = {1, 2, 2, 1};
+ std::vector<float> input_data = {1, 2, 3, 4};
+ std::vector<int32> output_size_data = {5, 2};
+ RuntimeShape output_shape = {1, 5, 2, 1};
+ std::vector<float> output_data = {2, 2, 2, 2, 4, 4, 4, 4, 4, 4};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/true, /*half_pixel_centers=*/true);
+}
+
TEST(ResizeNearestNeighborReference, Test2x2To4x4) {
RuntimeShape input_shape = {1, 2, 2, 1};
std::vector<uint8> input_data = {1, 2, 3, 4};
@@ -149,10 +285,56 @@
output_shape, output_data);
}
+TEST(ResizeNearestNeighborReference, Test2x2x2x2To2x3x3x2_AlignCorners) {
+ RuntimeShape input_shape = {2, 2, 2, 2};
+ std::vector<float> input_data = {1, 2, 3, 4, 5, 6, 7, 8,
+ 1, 2, 3, 4, 5, 6, 7, 8};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {2, 3, 3, 2};
+ std::vector<float> output_data = {
+ 1, 2, 3, 4, 3, 4, 5, 6, 7, 8, 7, 8, 5, 6, 7, 8, 7, 8,
+ 1, 2, 3, 4, 3, 4, 5, 6, 7, 8, 7, 8, 5, 6, 7, 8, 7, 8,
+ };
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/true, /*half_pixel_centers=*/false);
+}
+
+TEST(ResizeNearestNeighborReference, Test2x2x2x2To2x3x3x2_HalfPixelCenters) {
+ RuntimeShape input_shape = {2, 2, 2, 2};
+ std::vector<float> input_data = {1, 1, 2, 2, 3, 3, 4, 4,
+ 5, 5, 6, 6, 7, 7, 8, 8};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {2, 3, 3, 2};
+ std::vector<float> output_data = {1, 1, 2, 2, 2, 2, 3, 3, 4, 4, 4, 4,
+ 3, 3, 4, 4, 4, 4, 5, 5, 6, 6, 6, 6,
+ 7, 7, 8, 8, 8, 8, 7, 7, 8, 8, 8, 8};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/false, /*half_pixel_centers=*/true);
+}
+
+TEST(ResizeNearestNeighborReference,
+ Test2x2x2x2To2x3x3x2_HalfPixelCenters_AlignCorners) {
+ RuntimeShape input_shape = {2, 2, 2, 2};
+ std::vector<float> input_data = {1, 2, 3, 4, 5, 6, 7, 8,
+ 1, 2, 3, 4, 5, 6, 7, 8};
+ std::vector<int32> output_size_data = {3, 3};
+ RuntimeShape output_shape = {2, 3, 3, 2};
+ std::vector<float> output_data = {1, 2, 3, 4, 3, 4, 5, 6, 7, 8, 7, 8,
+ 5, 6, 7, 8, 7, 8, 1, 2, 3, 4, 3, 4,
+ 5, 6, 7, 8, 7, 8, 5, 6, 7, 8, 7, 8};
+
+ TestReferenceResizeNearestNeighbor(
+ input_shape, input_data, output_size_data, output_shape, output_data,
+ /*align_corners=*/true, /*half_pixel_centers=*/true);
+}
+
void TestOptimizedResizeNearestNeighbor(int batch, int depth, int input_width,
int input_height, int output_width,
int output_height) {
- ResizeNearestNeighborParams op_params{/*align_corners=*/false};
RuntimeShape output_size_shape({1, 1, 1, 2});
RuntimeShape input_shape({batch, input_height, input_width, depth});
@@ -167,6 +349,9 @@
std::vector<uint8> output_data(output_shape.FlatSize(), 3);
std::vector<int32> output_size_data = {output_height, output_width};
+ ResizeNearestNeighborParams op_params{/*align_corners=*/false,
+ /*half_pixel_centers=*/false};
+
// Test the optimized version against the reference version.
reference_ops::ResizeNearestNeighbor(
op_params, input_shape, input_data.data(), output_size_shape,
@@ -174,7 +359,35 @@
optimized_ops::ResizeNearestNeighbor(
op_params, input_shape, input_data.data(), output_size_shape,
output_size_data.data(), output_shape, output_data.data());
+ ASSERT_EQ(reference_output_data, output_data);
+ op_params.align_corners = true;
+ reference_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, reference_output_data.data());
+ optimized_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, output_data.data());
+ ASSERT_EQ(reference_output_data, output_data);
+
+ op_params.align_corners = false;
+ op_params.half_pixel_centers = true;
+ reference_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, reference_output_data.data());
+ optimized_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, output_data.data());
+ ASSERT_EQ(reference_output_data, output_data);
+
+ op_params.align_corners = true;
+ op_params.half_pixel_centers = true;
+ reference_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, reference_output_data.data());
+ optimized_ops::ResizeNearestNeighbor(
+ op_params, input_shape, input_data.data(), output_size_shape,
+ output_size_data.data(), output_shape, output_data.data());
ASSERT_EQ(reference_output_data, output_data);
}
@@ -214,7 +427,7 @@
TEST(ResizeNearestNeighborOptimized, TestReferenceParity) {
int invalid_count = 0;
- const int kTestsToRun = 100 * 1000;
+ const int kTestsToRun = 10000;
for (int i = 0; i < kTestsToRun; i++) {
const int batch = ExponentialRandomPositiveInt(0.9f, 3, 20);
const int depth = ExponentialRandomPositiveInt(0.9f, 6, 50);
diff --git a/tensorflow/lite/kernels/internal/tensor_utils_test.cc b/tensorflow/lite/kernels/internal/tensor_utils_test.cc
index e039fb8..9b047d3 100644
--- a/tensorflow/lite/kernels/internal/tensor_utils_test.cc
+++ b/tensorflow/lite/kernels/internal/tensor_utils_test.cc
@@ -14,8 +14,11 @@
==============================================================================*/
#include "tensorflow/lite/kernels/internal/tensor_utils.h"
+#include <math.h>
+
#include <gmock/gmock.h>
#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/test_util.h"
@@ -26,6 +29,13 @@
namespace tflite {
namespace tensor_utils {
+TEST(uKernels, FloorLog2Test) {
+ for (int i = 1; i < 257; ++i) {
+ EXPECT_EQ(::tflite::FloorLog2(i),
+ static_cast<int>(std::floor(std::log2(i))));
+ }
+}
+
TEST(uKernels, ClipTest) {
constexpr int kVectorSize = 10;
constexpr float kAbsLimit = 2.0;
@@ -457,6 +467,37 @@
&context);
EXPECT_THAT(output2, testing::ElementsAreArray(expected_output));
+
+ // Run with a large batch size to trigger the CpuBackendGemm path on any
+ // device.
+ constexpr int kBatchMultiplier = 8;
+ std::vector<int8_t> input_big_batch(input.size() * kBatchMultiplier);
+ std::vector<float> scaling_factors_big_batch(scaling_factors.size() *
+ kBatchMultiplier);
+ std::vector<int32_t> scratch_big_batch(scratch.size() * kBatchMultiplier);
+ std::vector<int32_t> input_offsets_big_batch(input_offsets.size() *
+ kBatchMultiplier);
+ for (int i = 0; i < kBatchMultiplier; i++) {
+ std::copy(input.begin(), input.end(),
+ input_big_batch.begin() + i * input.size());
+ std::copy(scaling_factors.begin(), scaling_factors.end(),
+ scaling_factors_big_batch.begin() + i * scaling_factors.size());
+ std::copy(input_offsets.begin(), input_offsets.end(),
+ input_offsets_big_batch.begin() + i * input_offsets.size());
+ }
+ std::vector<float> output_big_batch(output.size() * kBatchMultiplier, 0);
+ MatrixBatchVectorMultiplyAccumulate(
+ input_to_gate_weights.data(), /*m_rows=*/8, /*m_cols=*/32,
+ input_big_batch.data(), scaling_factors_big_batch.data(),
+ /*n_batch*/ 4 * kBatchMultiplier, output_big_batch.data(), nullptr,
+ input_offsets_big_batch.data(), scratch_big_batch.data(), row_sums,
+ &compute_row_sums, &context);
+ for (int i = 0; i < kBatchMultiplier; i++) {
+ std::vector<float> output_per_batch(
+ output_big_batch.begin() + i * output.size(),
+ output_big_batch.begin() + (i + 1) * output.size());
+ EXPECT_THAT(output_per_batch, testing::ElementsAreArray(expected_output));
+ }
}
// Qautnized matmul with 2 * 30 input and 9 * 30 matrix.
diff --git a/tensorflow/lite/kernels/internal/test_util.cc b/tensorflow/lite/kernels/internal/test_util.cc
index 4462775..4971ed2 100644
--- a/tensorflow/lite/kernels/internal/test_util.cc
+++ b/tensorflow/lite/kernels/internal/test_util.cc
@@ -105,6 +105,7 @@
void FillRandom(std::vector<float>* vec, float min, float max) {
std::uniform_real_distribution<float> dist(min, max);
+ // TODO(b/154540105): use std::ref to avoid copying the random engine.
auto gen = std::bind(dist, RandomEngine());
std::generate(std::begin(*vec), std::end(*vec), gen);
}
diff --git a/tensorflow/lite/kernels/internal/test_util.h b/tensorflow/lite/kernels/internal/test_util.h
index 766a627..6c9a341 100644
--- a/tensorflow/lite/kernels/internal/test_util.h
+++ b/tensorflow/lite/kernels/internal/test_util.h
@@ -59,12 +59,22 @@
// Fills a vector with random floats between |min| and |max|.
void FillRandom(std::vector<float>* vec, float min, float max);
+template <typename T>
+void FillRandom(typename std::vector<T>::iterator begin_it,
+ typename std::vector<T>::iterator end_it, T min, T max) {
+ // Workaround for compilers that don't support (u)int8_t uniform_distribution.
+ typedef typename std::conditional<sizeof(T) >= sizeof(int16_t), T,
+ std::int16_t>::type rand_type;
+ std::uniform_int_distribution<rand_type> dist(min, max);
+ // TODO(b/154540105): use std::ref to avoid copying the random engine.
+ auto gen = std::bind(dist, RandomEngine());
+ std::generate(begin_it, end_it, [&gen] { return static_cast<T>(gen()); });
+}
+
// Fills a vector with random numbers between |min| and |max|.
template <typename T>
void FillRandom(std::vector<T>* vec, T min, T max) {
- std::uniform_int_distribution<T> dist(min, max);
- auto gen = std::bind(dist, RandomEngine());
- std::generate(std::begin(*vec), std::end(*vec), gen);
+ return FillRandom(std::begin(*vec), std::end(*vec), min, max);
}
// Fills a vector with random numbers.
@@ -73,14 +83,6 @@
FillRandom(vec, std::numeric_limits<T>::min(), std::numeric_limits<T>::max());
}
-template <typename T>
-void FillRandom(typename std::vector<T>::iterator begin_it,
- typename std::vector<T>::iterator end_it, T min, T max) {
- std::uniform_int_distribution<T> dist(min, max);
- auto gen = std::bind(dist, RandomEngine());
- std::generate(begin_it, end_it, gen);
-}
-
// Fill with a "skyscraper" pattern, in which there is a central section (across
// the depth) with higher values than the surround.
template <typename T>
diff --git a/tensorflow/lite/kernels/internal/types.h b/tensorflow/lite/kernels/internal/types.h
index 0752f5c..cbdedd8 100644
--- a/tensorflow/lite/kernels/internal/types.h
+++ b/tensorflow/lite/kernels/internal/types.h
@@ -1007,6 +1007,7 @@
struct ResizeNearestNeighborParams {
bool align_corners;
+ bool half_pixel_centers;
};
struct SliceParams {
diff --git a/tensorflow/lite/kernels/kernel_util.cc b/tensorflow/lite/kernels/kernel_util.cc
index f9b0cbb..b30747e 100644
--- a/tensorflow/lite/kernels/kernel_util.cc
+++ b/tensorflow/lite/kernels/kernel_util.cc
@@ -62,8 +62,9 @@
TF_LITE_ENSURE(context, affine_quantization->scale);
const bool is_per_channel = affine_quantization->scale->size > 1;
if (is_per_channel) {
- // Currently only Int8 is supported for per channel quantization.
- TF_LITE_ENSURE_EQ(context, input->type, kTfLiteInt8);
+ // Currently only Int8/Int16 is supported for per channel quantization.
+ TF_LITE_ENSURE(context,
+ input->type == kTfLiteInt8 || input->type == kTfLiteInt16);
TF_LITE_ENSURE_EQ(context, filter->type, kTfLiteInt8);
TF_LITE_ENSURE_EQ(context, affine_quantization->scale->size, num_channels);
TF_LITE_ENSURE_EQ(
@@ -104,7 +105,8 @@
QuantizeMultiplier(real_multiplier, multiplier, &exponent);
*shift = -exponent;
}
- if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8) {
+ if (input->type == kTfLiteInt8 || input->type == kTfLiteUInt8 ||
+ input->type == kTfLiteInt16) {
TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
context, activation, output, output_activation_min,
output_activation_max));
diff --git a/tensorflow/lite/kernels/pack_test.cc b/tensorflow/lite/kernels/pack_test.cc
index 7d18bb6..bc6758c 100644
--- a/tensorflow/lite/kernels/pack_test.cc
+++ b/tensorflow/lite/kernels/pack_test.cc
@@ -195,10 +195,10 @@
struct PackOpTestInt : public ::testing::Test {
using TypeToTest = InputType;
TensorType TENSOR_TYPE =
- std::is_same<InputType, int16_t>::value
- ? TensorType_INT16
- : (std::is_same<InputType, uint8_t>::value ? TensorType_UINT8
- : TensorType_INT8);
+ (std::is_same<InputType, int16_t>::value
+ ? TensorType_INT16
+ : (std::is_same<InputType, uint8_t>::value ? TensorType_UINT8
+ : TensorType_INT8));
};
using TestTypes = testing::Types<int8_t, uint8_t, int16_t>;
diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc
index 392369e..28515ae 100644
--- a/tensorflow/lite/kernels/register.cc
+++ b/tensorflow/lite/kernels/register.cc
@@ -203,7 +203,7 @@
AddBuiltin(BuiltinOperator_COS, Register_COS());
AddBuiltin(BuiltinOperator_TRANSPOSE_CONV, Register_TRANSPOSE_CONV(),
/* min_version = */ 1,
- /* max_version = */ 2);
+ /* max_version = */ 3);
AddBuiltin(BuiltinOperator_TILE, Register_TILE(),
/* min_version = */ 1,
/* max_version = */ 2);
@@ -224,10 +224,10 @@
/* max_version = */ 3);
AddBuiltin(BuiltinOperator_EQUAL, Register_EQUAL(),
/* min_version = */ 1,
- /* max_version = */ 2);
+ /* max_version = */ 3);
AddBuiltin(BuiltinOperator_NOT_EQUAL, Register_NOT_EQUAL(),
/* min_version = */ 1,
- /* max_version = */ 2);
+ /* max_version = */ 3);
AddBuiltin(BuiltinOperator_SQRT, Register_SQRT());
AddBuiltin(BuiltinOperator_RSQRT, Register_RSQRT());
AddBuiltin(BuiltinOperator_SHAPE, Register_SHAPE());
diff --git a/tensorflow/lite/kernels/resize_bilinear_test.cc b/tensorflow/lite/kernels/resize_bilinear_test.cc
index 86dcaae..5cbba02 100644
--- a/tensorflow/lite/kernels/resize_bilinear_test.cc
+++ b/tensorflow/lite/kernels/resize_bilinear_test.cc
@@ -25,8 +25,8 @@
using uint8 = std::uint8_t;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
class ResizeBilinearOpModel : public SingleOpModel {
@@ -35,7 +35,7 @@
std::initializer_list<int> size_data,
TestType test_type,
bool half_pixel_centers = false) {
- bool const_size = (test_type == TestType::CONST);
+ bool const_size = (test_type == TestType::kConst);
input_ = AddInput(input);
if (const_size) {
@@ -332,7 +332,7 @@
}
INSTANTIATE_TEST_SUITE_P(ResizeBilinearOpTest, ResizeBilinearOpTest,
- testing::Values(TestType::CONST, TestType::DYNAMIC));
+ testing::Values(TestType::kConst, TestType::kDynamic));
} // namespace
} // namespace tflite
diff --git a/tensorflow/lite/kernels/resize_nearest_neighbor.cc b/tensorflow/lite/kernels/resize_nearest_neighbor.cc
index b783a0e..122f244 100644
--- a/tensorflow/lite/kernels/resize_nearest_neighbor.cc
+++ b/tensorflow/lite/kernels/resize_nearest_neighbor.cc
@@ -89,6 +89,7 @@
tflite::ResizeNearestNeighborParams op_params;
op_params.align_corners = params->align_corners;
+ op_params.half_pixel_centers = false;
if (output->type == kTfLiteFloat32) {
reference_ops::ResizeNearestNeighbor(
diff --git a/tensorflow/lite/kernels/resize_nearest_neighbor_test.cc b/tensorflow/lite/kernels/resize_nearest_neighbor_test.cc
index e8170c9..b894d3a 100644
--- a/tensorflow/lite/kernels/resize_nearest_neighbor_test.cc
+++ b/tensorflow/lite/kernels/resize_nearest_neighbor_test.cc
@@ -25,8 +25,8 @@
using uint8 = std::uint8_t;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
class ResizeNearestNeighborOpModel : public SingleOpModel {
@@ -34,7 +34,7 @@
explicit ResizeNearestNeighborOpModel(const TensorData& input,
std::initializer_list<int> size_data,
TestType test_type) {
- bool const_size = (test_type == TestType::CONST);
+ bool const_size = (test_type == TestType::kConst);
input_ = AddInput(input);
if (const_size) {
@@ -264,7 +264,7 @@
}
INSTANTIATE_TEST_SUITE_P(ResizeNearestNeighborOpTest,
ResizeNearestNeighborOpTest,
- testing::Values(TestType::CONST, TestType::DYNAMIC));
+ testing::Values(TestType::kConst, TestType::kDynamic));
} // namespace
} // namespace tflite
diff --git a/tensorflow/lite/kernels/slice_test.cc b/tensorflow/lite/kernels/slice_test.cc
index b372aec..1a31ae4 100644
--- a/tensorflow/lite/kernels/slice_test.cc
+++ b/tensorflow/lite/kernels/slice_test.cc
@@ -24,8 +24,8 @@
using ::testing::ElementsAreArray;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
template <typename input_type, typename index_type>
@@ -39,7 +39,7 @@
TensorType tensor_index_type, TensorType tensor_input_type,
TestType input_tensor_types) {
input_ = AddInput(tensor_input_type);
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
begin_ = AddInput(tensor_index_type);
size_ = AddInput(tensor_index_type);
} else {
@@ -52,7 +52,7 @@
CreateSliceOptions(builder_).Union());
BuildInterpreter({input_shape, begin_shape, size_shape});
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
PopulateTensor<index_type>(begin_, begin_data);
PopulateTensor<index_type>(size_, size_data);
}
@@ -239,7 +239,8 @@
}
INSTANTIATE_TEST_SUITE_P(SliceOpTest, SliceOpTest,
- ::testing::Values(TestType::CONST, TestType::DYNAMIC));
+ ::testing::Values(TestType::kConst,
+ TestType::kDynamic));
} // namespace
} // namespace tflite
diff --git a/tensorflow/lite/kernels/split_test.cc b/tensorflow/lite/kernels/split_test.cc
index 48c7a0a..7952396 100644
--- a/tensorflow/lite/kernels/split_test.cc
+++ b/tensorflow/lite/kernels/split_test.cc
@@ -26,8 +26,8 @@
constexpr int kAxisIsATensor = -1000;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
class SplitOpModel : public SingleOpModel {
@@ -83,7 +83,7 @@
<< " and num_splits=" << num_splits;
return ss.str();
};
- if (test_type == TestType::DYNAMIC) {
+ if (test_type == TestType::kDynamic) {
SplitOpModel m({type, input_shape}, num_splits);
m.SetInput(input_data);
m.SetAxis(axis);
@@ -110,18 +110,18 @@
template <typename T>
class SplitOpTest : public ::testing::Test {
public:
- static std::vector<TestType> _range_;
+ static std::vector<TestType> range_;
};
template <>
-std::vector<TestType> SplitOpTest<TestType>::_range_{TestType::CONST,
- TestType::DYNAMIC};
+std::vector<TestType> SplitOpTest<TestType>::range_{TestType::kConst,
+ TestType::kDynamic};
using DataTypes = ::testing::Types<float, int8_t, int16_t>;
TYPED_TEST_SUITE(SplitOpTest, DataTypes);
TYPED_TEST(SplitOpTest, FourDimensional) {
- for (TestType test_type : SplitOpTest<TestType>::_range_) {
+ for (TestType test_type : SplitOpTest<TestType>::range_) {
Check<TypeParam>(/*axis_as_tensor*/ test_type,
/*axis=*/0, /*num_splits=*/2, {2, 2, 2, 2}, {1, 2, 2, 2},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
@@ -158,7 +158,7 @@
}
TYPED_TEST(SplitOpTest, FourDimensionalInt8) {
- for (TestType test_type : SplitOpTest<TestType>::_range_) {
+ for (TestType test_type : SplitOpTest<TestType>::range_) {
Check<TypeParam>(/*axis_as_tensor*/ test_type,
/*axis=*/0, /*num_splits=*/2, {2, 2, 2, 2}, {1, 2, 2, 2},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
@@ -195,7 +195,7 @@
}
TYPED_TEST(SplitOpTest, FourDimensionalInt32) {
- for (TestType test_type : SplitOpTest<TestType>::_range_) {
+ for (TestType test_type : SplitOpTest<TestType>::range_) {
Check<TypeParam>(/*axis_as_tensor*/ test_type,
/*axis=*/0, /*num_splits=*/2, {2, 2, 2, 2}, {1, 2, 2, 2},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
@@ -232,7 +232,7 @@
}
TYPED_TEST(SplitOpTest, OneDimensional) {
- for (TestType test_type : SplitOpTest<TestType>::_range_) {
+ for (TestType test_type : SplitOpTest<TestType>::range_) {
Check<TypeParam>(
/*axis_as_tensor*/ test_type,
/*axis=*/0, /*num_splits=*/8, {8}, {1}, {1, 2, 3, 4, 5, 6, 7, 8},
@@ -241,7 +241,7 @@
}
TYPED_TEST(SplitOpTest, NegativeAxis) {
- for (TestType test_type : SplitOpTest<TestType>::_range_) {
+ for (TestType test_type : SplitOpTest<TestType>::range_) {
Check<TypeParam>(/*axis_as_tensor*/ test_type,
/*axis=*/-4, /*num_splits=*/2, {2, 2, 2, 2}, {1, 2, 2, 2},
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16},
diff --git a/tensorflow/lite/kernels/test_util.h b/tensorflow/lite/kernels/test_util.h
index 7b504e4..7cd01f0 100644
--- a/tensorflow/lite/kernels/test_util.h
+++ b/tensorflow/lite/kernels/test_util.h
@@ -251,32 +251,44 @@
quantized_output.data() + quantized_output.size());
}
+ template <typename T>
+ void PerChannelQuantizeBiasPopulateTensor(
+ const std::vector<float>& input_data, int index,
+ TfLiteAffineQuantization* params) {
+ const int32_t num_inputs = input_data.size();
+ std::vector<T> quantized_output(num_inputs);
+ for (int i = 0; i < num_inputs; ++i) {
+ const float scale = params->scale->size == 1 ? params->scale->data[0]
+ : params->scale->data[i];
+ quantized_output[i] = input_data[i] / scale;
+ }
+ }
+
+ template <typename T>
+ void PerChannelQuantizeBiasPopulateTensor(
+ int index, const std::vector<float>& input_data,
+ const TfLiteAffineQuantization* params) {
+ const int32_t num_inputs = input_data.size();
+ std::vector<T> quantized_output(num_inputs);
+ for (int i = 0; i < num_inputs; ++i) {
+ const float scale = params->scale->size == 1 ? params->scale->data[0]
+ : params->scale->data[i];
+ quantized_output[i] = input_data[i] / scale;
+ }
+ PopulateTensor(index, /*offset=*/0, quantized_output.data(),
+ quantized_output.data() + quantized_output.size());
+ }
+
// Quantize and populate data for bias with per channel quantization.
void PerChannelQuantizeBias(int index, const std::vector<float>& input_data) {
- const int32_t num_inputs = input_data.size();
- std::vector<int32_t> quantized_output(num_inputs);
TfLiteTensor* t = interpreter_->tensor(index);
auto* params =
reinterpret_cast<TfLiteAffineQuantization*>(t->quantization.params);
CHECK(t->type == kTfLiteInt32 || t->type == kTfLiteInt64);
if (t->type == kTfLiteInt32) {
- std::vector<int32_t> quantized_output(num_inputs);
- for (int i = 0; i < num_inputs; ++i) {
- const float scale = params->scale->size == 1 ? params->scale->data[0]
- : params->scale->data[i];
- quantized_output[i] = input_data[i] / scale;
- }
- PopulateTensor(index, /*offset=*/0, quantized_output.data(),
- quantized_output.data() + quantized_output.size());
+ PerChannelQuantizeBiasPopulateTensor<int32_t>(index, input_data, params);
} else {
- std::vector<int64_t> quantized_output(num_inputs);
- for (int i = 0; i < num_inputs; ++i) {
- const float scale = params->scale->size == 1 ? params->scale->data[0]
- : params->scale->data[i];
- quantized_output[i] = input_data[i] / scale;
- }
- PopulateTensor(index, /*offset=*/0, quantized_output.data(),
- quantized_output.data() + quantized_output.size());
+ PerChannelQuantizeBiasPopulateTensor<int64_t>(index, input_data, params);
}
}
@@ -368,6 +380,14 @@
template <typename T>
void PopulateTensor(int index, int offset, T* begin, T* end) {
T* v = interpreter_->typed_tensor<T>(index);
+ if (!v) {
+ auto* t = interpreter_->tensor(index);
+ CHECK(t) << "No tensor with index " << index << ".";
+ CHECK(t->data.raw) << "Empty data for tensor with index " << index << ".";
+ CHECK(v) << "Type mismatch for tensor with index " << index
+ << ". Requested " << typeToTfLiteType<T>() << ", got "
+ << t->type;
+ }
memcpy(v + offset, begin, (end - begin) * sizeof(T));
}
@@ -816,40 +836,40 @@
template <>
struct TypeUnion<float> {
public:
- static const TensorType tensor_type = TensorType::TensorType_FLOAT32;
- static const TfLiteType tflite_type = TfLiteType::kTfLiteFloat32;
+ static constexpr TensorType tensor_type = TensorType::TensorType_FLOAT32;
+ static constexpr TfLiteType tflite_type = TfLiteType::kTfLiteFloat32;
typedef float ScalarType;
};
template <>
struct TypeUnion<int32_t> {
public:
- static const TensorType tensor_type = TensorType::TensorType_INT32;
- static const TfLiteType tflite_type = TfLiteType::kTfLiteInt32;
+ static constexpr TensorType tensor_type = TensorType::TensorType_INT32;
+ static constexpr TfLiteType tflite_type = TfLiteType::kTfLiteInt32;
typedef int32_t ScalarType;
};
template <>
struct TypeUnion<int16_t> {
public:
- static const TensorType tensor_type = TensorType::TensorType_INT16;
- static const TfLiteType tflite_type = TfLiteType::kTfLiteInt16;
+ static constexpr TensorType tensor_type = TensorType::TensorType_INT16;
+ static constexpr TfLiteType tflite_type = TfLiteType::kTfLiteInt16;
typedef int16_t ScalarType;
};
template <>
struct TypeUnion<int8_t> {
public:
- static const TensorType tensor_type = TensorType::TensorType_INT8;
- static const TfLiteType tflite_type = TfLiteType::kTfLiteInt8;
+ static constexpr TensorType tensor_type = TensorType::TensorType_INT8;
+ static constexpr TfLiteType tflite_type = TfLiteType::kTfLiteInt8;
typedef int8_t ScalarType;
};
template <>
struct TypeUnion<uint8_t> {
public:
- static const TensorType tensor_type = TensorType::TensorType_UINT8;
- static const TfLiteType tflite_type = TfLiteType::kTfLiteUInt8;
+ static constexpr TensorType tensor_type = TensorType::TensorType_UINT8;
+ static constexpr TfLiteType tflite_type = TfLiteType::kTfLiteUInt8;
typedef uint8_t ScalarType;
};
diff --git a/tensorflow/lite/kernels/topk_v2_test.cc b/tensorflow/lite/kernels/topk_v2_test.cc
index c82e5a6..72ed82c 100644
--- a/tensorflow/lite/kernels/topk_v2_test.cc
+++ b/tensorflow/lite/kernels/topk_v2_test.cc
@@ -26,8 +26,8 @@
using ::testing::ElementsAreArray;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
template <typename InputType>
@@ -36,7 +36,7 @@
TopKV2OpModel(int top_k, std::initializer_list<int> input_shape,
std::initializer_list<InputType> input_data,
TestType input_tensor_types) {
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
input_ = AddInput(GetTensorType<InputType>());
top_k_ = AddInput(TensorType_INT32);
} else {
@@ -49,7 +49,7 @@
SetBuiltinOp(BuiltinOperator_TOPK_V2, BuiltinOptions_TopKV2Options, 0);
BuildInterpreter({input_shape, {1}});
- if (input_tensor_types == TestType::DYNAMIC) {
+ if (input_tensor_types == TestType::kDynamic) {
PopulateTensor<InputType>(input_, input_data);
PopulateTensor<int32_t>(top_k_, {top_k});
}
@@ -119,7 +119,8 @@
}
INSTANTIATE_TEST_SUITE_P(TopKV2OpTest, TopKV2OpTest,
- ::testing::Values(TestType::CONST, TestType::DYNAMIC));
+ ::testing::Values(TestType::kConst,
+ TestType::kDynamic));
// Check that uint8_t works.
TEST_P(TopKV2OpTest, TypeUint8) {
diff --git a/tensorflow/lite/kernels/transpose_conv.cc b/tensorflow/lite/kernels/transpose_conv.cc
index 114b9ae..9b2767f 100644
--- a/tensorflow/lite/kernels/transpose_conv.cc
+++ b/tensorflow/lite/kernels/transpose_conv.cc
@@ -50,6 +50,7 @@
constexpr int kOutputShapeTensor = 0;
constexpr int kWeightsTensor = 1;
constexpr int kDataInputTensor = 2;
+constexpr int kBiasTensor = 3;
constexpr int kOutputTensor = 0;
const int kTensorNotAllocated = -1;
@@ -232,7 +233,7 @@
GetTensorData<int8>(transposed_weights));
} else {
context->ReportError(
- context, "Transpose conv only support float & uint8 right now.");
+ context, "Transpose conv only support float & uint8 & int8 right now.");
return kTfLiteError;
}
@@ -243,8 +244,10 @@
TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
OpData* data = reinterpret_cast<OpData*>(node->user_data);
+ bool has_bias = NumInputs(node) == 4;
+
// Sanity checks on op
- TF_LITE_ENSURE_EQ(context, NumInputs(node), 3);
+ TF_LITE_ENSURE(context, has_bias || NumInputs(node) == 3);
TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
// Retrieve tensors
@@ -252,6 +255,8 @@
GetInput(context, node, kOutputShapeTensor);
const TfLiteTensor* weights = GetInput(context, node, kWeightsTensor);
const TfLiteTensor* input = GetInput(context, node, kDataInputTensor);
+ const TfLiteTensor* bias = nullptr;
+
TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
// Tensor sanity checks
@@ -261,7 +266,23 @@
TF_LITE_ENSURE(context, input->type == kTfLiteFloat32 ||
input->type == kTfLiteUInt8 ||
input->type == kTfLiteInt8);
- TF_LITE_ENSURE_EQ(context, weights->type, input->type);
+
+ if (has_bias) {
+ bias = GetOptionalInputTensor(context, node, kBiasTensor);
+ if (bias) {
+ if (input->type == kTfLiteUInt8 || input->type == kTfLiteInt8) {
+ TF_LITE_ENSURE_EQ(context, bias->type, kTfLiteInt32);
+ if (input->type == kTfLiteInt8) {
+ TF_LITE_ENSURE_EQ(context, bias->params.zero_point, 0);
+ }
+ } else {
+ TF_LITE_ENSURE_EQ(context, bias->type, input->type);
+ }
+ TF_LITE_ENSURE_EQ(context, NumElements(bias),
+ SizeOfDimension(weights, 0));
+ }
+ }
+
TF_LITE_ENSURE_EQ(context, output->type, input->type);
// Ensure that weights and inputs have the same channel dimension.
// Note: TOCO will reorder weights in the following format: OHWI.
@@ -330,7 +351,7 @@
data->per_channel_output_multiplier.resize(number_channel);
data->per_channel_output_shift.resize(number_channel);
TF_LITE_ENSURE_STATUS(tflite::PopulateConvolutionQuantizationParams(
- context, input, weights, nullptr, output, kTfLiteActNone,
+ context, input, weights, bias, output, kTfLiteActNone,
&data->output_multiplier, &data->output_shift,
&data->output_activation_min, &data->output_activation_max,
data->per_channel_output_multiplier.data(),
@@ -343,7 +364,7 @@
template <KernelType kernel_type>
void EvalFloat(TfLiteContext* context, const TfLiteTransposeConvParams* params,
const OpData* data, const TfLiteTensor* input,
- const TfLiteTensor* weights,
+ const TfLiteTensor* weights, const TfLiteTensor* bias,
const TfLiteTensor* transposed_weights, TfLiteTensor* col2im,
TfLiteTensor* output) {
tflite::ConvParams op_params;
@@ -354,11 +375,13 @@
op_params.padding_values.height_offset = data->padding.height_offset;
op_params.stride_width = params->stride_width;
op_params.stride_height = params->stride_height;
+
switch (kernel_type) {
case kReference: {
reference_ops::TransposeConv(
op_params, GetTensorShape(input), GetTensorData<float>(input),
GetTensorShape(weights), GetTensorData<float>(weights),
+ GetTensorShape(bias), GetTensorData<float>(bias),
GetTensorShape(output), GetTensorData<float>(output),
GetTensorShape(col2im), GetTensorData<float>(col2im));
break;
@@ -367,7 +390,8 @@
optimized_ops::TransposeConvV2(
op_params, GetTensorShape(input), GetTensorData<float>(input),
GetTensorShape(transposed_weights),
- GetTensorData<float>(transposed_weights), GetTensorShape(output),
+ GetTensorData<float>(transposed_weights), GetTensorShape(bias),
+ GetTensorData<float>(bias), GetTensorShape(output),
GetTensorData<float>(output), GetTensorShape(col2im),
GetTensorData<float>(col2im),
CpuBackendContext::GetFromContext(context));
@@ -380,7 +404,8 @@
void EvalQuantized(TfLiteContext* context,
const TfLiteTransposeConvParams* params, OpData* data,
const TfLiteTensor* input, const TfLiteTensor* weights,
- const TfLiteTensor* transposed_weights, TfLiteTensor* col2im,
+ const TfLiteTensor* transposed_weights,
+ const TfLiteTensor* bias, TfLiteTensor* col2im,
TfLiteTensor* output, TfLiteTensor* scratch_buffer) {
int32_t input_offset = -input->params.zero_point;
int32_t filter_offset = -weights->params.zero_point;
@@ -407,6 +432,7 @@
reference_ops::TransposeConv(
op_params, GetTensorShape(input), GetTensorData<uint8>(input),
GetTensorShape(weights), GetTensorData<uint8>(weights),
+ GetTensorShape(bias), GetTensorData<int32_t>(bias),
GetTensorShape(output), GetTensorData<uint8>(output),
GetTensorShape(col2im), GetTensorData<uint8>(col2im),
GetTensorData<int32_t>(scratch_buffer));
@@ -416,7 +442,8 @@
optimized_ops::TransposeConvV2(
op_params, GetTensorShape(input), GetTensorData<uint8>(input),
GetTensorShape(transposed_weights),
- GetTensorData<uint8>(transposed_weights), GetTensorShape(output),
+ GetTensorData<uint8>(transposed_weights), GetTensorShape(bias),
+ GetTensorData<int32>(bias), GetTensorShape(output),
GetTensorData<uint8>(output), GetTensorShape(col2im),
GetTensorData<int32>(col2im), GetTensorData<int32>(scratch_buffer),
CpuBackendContext::GetFromContext(context));
@@ -426,13 +453,11 @@
}
template <KernelType kernel_type>
-void EvalQuantizedPerChannel(TfLiteContext* context,
- const TfLiteTransposeConvParams* params,
- OpData* data, const TfLiteTensor* input,
- const TfLiteTensor* weights,
- const TfLiteTensor* transposed_weights,
- TfLiteTensor* col2im, TfLiteTensor* output,
- TfLiteTensor* scratch_buffer) {
+void EvalQuantizedPerChannel(
+ TfLiteContext* context, const TfLiteTransposeConvParams* params,
+ OpData* data, const TfLiteTensor* input, const TfLiteTensor* weights,
+ const TfLiteTensor* transposed_weights, const TfLiteTensor* bias,
+ TfLiteTensor* col2im, TfLiteTensor* output, TfLiteTensor* scratch_buffer) {
tflite::ConvParams op_params;
op_params.padding_type = PaddingType::kSame;
op_params.padding_values.width = data->padding.width;
@@ -454,7 +479,8 @@
op_params, data->per_channel_output_multiplier.data(),
data->per_channel_output_shift.data(), GetTensorShape(input),
GetTensorData<int8>(input), GetTensorShape(weights),
- GetTensorData<int8>(weights), GetTensorShape(output),
+ GetTensorData<int8>(weights), GetTensorShape(bias),
+ GetTensorData<int32>(bias), GetTensorShape(output),
GetTensorData<int8>(output), GetTensorShape(col2im),
GetTensorData<int8>(col2im), GetTensorData<int32_t>(scratch_buffer));
break;
@@ -464,7 +490,8 @@
op_params, data->per_channel_output_multiplier.data(),
data->per_channel_output_shift.data(), GetTensorShape(input),
GetTensorData<int8>(input), GetTensorShape(transposed_weights),
- GetTensorData<int8>(transposed_weights), GetTensorShape(output),
+ GetTensorData<int8>(transposed_weights), GetTensorShape(bias),
+ GetTensorData<int32>(bias), GetTensorShape(output),
GetTensorData<int8>(output), GetTensorShape(col2im),
GetTensorData<int32>(col2im), GetTensorData<int32>(scratch_buffer),
CpuBackendContext::GetFromContext(context));
@@ -480,6 +507,10 @@
GetInput(context, node, kOutputShapeTensor);
const TfLiteTensor* weights = GetInput(context, node, kWeightsTensor);
const TfLiteTensor* input = GetInput(context, node, kDataInputTensor);
+ const TfLiteTensor* bias =
+ (NumInputs(node) == 4)
+ ? GetOptionalInputTensor(context, node, kBiasTensor)
+ : nullptr;
TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
OpData* data = reinterpret_cast<OpData*>(node->user_data);
TfLiteTensor* col2im = data->has_col2im
@@ -522,7 +553,7 @@
ResizeAndTransposeWeights(context, weights, transposed_weights);
}
}
- EvalFloat<kernel_type>(context, params, data, input, weights,
+ EvalFloat<kernel_type>(context, params, data, input, weights, bias,
transposed_weights, col2im, output);
break;
}
@@ -539,7 +570,7 @@
}
}
EvalQuantized<kernel_type>(context, params, data, input, weights,
- transposed_weights, col2im, output,
+ transposed_weights, bias, col2im, output,
scratch_buffer);
break;
}
@@ -554,8 +585,8 @@
ResizeAndTransposeWeights(context, weights, transposed_weights);
}
EvalQuantizedPerChannel<kernel_type>(context, params, data, input,
- weights, transposed_weights, col2im,
- output, scratch_buffer);
+ weights, transposed_weights, bias,
+ col2im, output, scratch_buffer);
break;
}
default:
diff --git a/tensorflow/lite/kernels/transpose_conv_test.cc b/tensorflow/lite/kernels/transpose_conv_test.cc
index 1851c01..77dc22b 100644
--- a/tensorflow/lite/kernels/transpose_conv_test.cc
+++ b/tensorflow/lite/kernels/transpose_conv_test.cc
@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -37,8 +37,8 @@
using ::testing::ElementsAreArray;
enum class TestType {
- CONST = 0,
- DYNAMIC = 1,
+ kConst = 0,
+ kDynamic = 1,
};
template <typename InputType>
@@ -54,7 +54,7 @@
// Just to be confusing, transpose_conv has an _input_ named "output_shape"
// that sets the shape of the output tensor of the op :). It must always be
// an int32 1D four element tensor.
- if (test_type == TestType::DYNAMIC) {
+ if (test_type == TestType::kDynamic) {
output_shape_ = AddInput({TensorType_INT32, {4}});
filter_ = AddInput(filter);
} else {
@@ -74,7 +74,7 @@
BuildInterpreter(
{GetShape(output_shape_), GetShape(filter_), GetShape(input_)});
- if (test_type == TestType::DYNAMIC) {
+ if (test_type == TestType::kDynamic) {
PopulateTensor<int32_t>(output_shape_, output_shape_data);
PopulateTensor<InputType>(filter_, filter_data);
}
@@ -441,11 +441,241 @@
EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 6, 6, 1}));
}
+template <typename InputType>
+class BaseTransposeConvBiasOpModel : public SingleOpModel {
+ public:
+ BaseTransposeConvBiasOpModel(TfLiteRegistration* registration,
+ std::initializer_list<int> output_shape_data,
+ const TensorData& filter,
+ std::initializer_list<InputType> filter_data,
+ const TensorData& input,
+ const TensorData& output, Padding padding,
+ int stride_w, int stride_h, TestType test_type,
+ int version = 3) {
+ if (test_type == TestType::kDynamic) {
+ output_shape_ = AddInput({TensorType_INT32, {4}});
+ filter_ = AddInput(filter);
+ } else {
+ output_shape_ = AddConstInput(TensorType_INT32, output_shape_data, {4});
+ filter_ = AddConstInput(filter, filter_data);
+ }
+ input_ = AddInput(input);
+
+ int bias_size = GetShape(filter_)[0];
+ if (input.type == TensorType_FLOAT32) {
+ bias_ = AddInput({TensorType_FLOAT32, {bias_size}});
+ } else if (input.type == TensorType_INT8) {
+ // per channel quantization.
+ std::vector<float> bias_scale(
+ filter.per_channel_quantization_scales.size());
+ std::vector<int64_t> bias_zero_points(
+ filter.per_channel_quantization_scales.size());
+ for (size_t i = 0; i < filter.per_channel_quantization_scales.size();
+ ++i) {
+ bias_scale[i] = input.scale * filter.per_channel_quantization_scales[i];
+ bias_zero_points[i] = 0;
+ }
+ TensorData bias{TensorType_INT32,
+ {bias_size},
+ /*min=*/0,
+ /*max=*/0,
+ /*scale=*/0,
+ /*zero_point=*/0,
+ true,
+ /*per_channel_quantization_scales=*/bias_scale,
+ /*per_channel_quantization_offsets=*/bias_zero_points,
+ /*channel_index==*/0};
+ bias_ = AddInput(bias);
+ } else {
+ // per tensor quantization.
+ auto bias_scale = GetScale(input_) * GetScale(filter_);
+ TensorData bias{TensorType_INT32, {bias_size}, 0, 0, bias_scale};
+ bias_ = AddInput(bias);
+ }
+
+ output_ = AddOutput(output);
+
+ SetBuiltinOp(
+ BuiltinOperator_TRANSPOSE_CONV, BuiltinOptions_TransposeConvOptions,
+ CreateTransposeConvOptions(builder_, padding, stride_w, stride_h)
+ .Union());
+ resolver_ = absl::make_unique<SingleOpResolver>(
+ BuiltinOperator_TRANSPOSE_CONV, registration, version);
+ BuildInterpreter({GetShape(output_shape_), GetShape(filter_),
+ GetShape(input_), GetShape(bias_)});
+
+ if (test_type == TestType::kDynamic) {
+ PopulateTensor<int32_t>(output_shape_, output_shape_data);
+ PopulateTensor<InputType>(filter_, filter_data);
+ }
+ }
+
+ void SetInput(std::initializer_list<float> data) {
+ if (std::is_same<InputType, uint8_t>::value) {
+ QuantizeAndPopulate<uint8_t>(input_, data);
+ } else if (std::is_same<InputType, int8_t>::value) {
+ QuantizeAndPopulate<int8_t>(input_, data);
+ } else {
+ PopulateTensor(input_, data);
+ }
+ }
+
+ void SetBias(std::initializer_list<float> bias) {
+ if (std::is_same<InputType, uint8_t>::value) {
+ QuantizeAndPopulate<int32_t>(bias_, bias);
+ } else if (std::is_same<InputType, int8_t>::value) {
+ PerChannelQuantizeBias(bias_, bias);
+ } else {
+ PopulateTensor(bias_, bias);
+ }
+ }
+
+ std::vector<int> GetOutputShape() { return GetTensorShape(output_); }
+
+ protected:
+ int output_shape_;
+ int filter_;
+ int input_;
+ int bias_;
+ int output_;
+};
+
+class TransposeConvOpBiasModel : public BaseTransposeConvBiasOpModel<float> {
+ public:
+ using BaseTransposeConvBiasOpModel::BaseTransposeConvBiasOpModel;
+
+ std::vector<float> GetOutput() { return ExtractVector<float>(output_); }
+};
+
+// Test case:
+// input_data = np.arange(1, 5).reshape(1,2,2,1).astype(np.float32)
+// filter_data = np.arange(1, 19).reshape(3,3,2,1).astype(np.float32)
+// bias_data = np.array([3,4])
+// input = tf.keras.layers.Input(shape=(2, 2, 1))
+// output = tf.keras.layers.Convolution2DTranspose(filters=2,
+// kernel_size=[3, 3],
+// strides=[2, 2],
+// padding="valid")(input)
+// model = tf.keras.models.Model(input, output)
+// model.layers[1].set_weights([filter_data, bias_data])
+// output = model.predict(input_data)
+TEST_P(TransposeConvOpTest, MultiChannelBiasTest) {
+ // TODO(b/138722124): Enable these tests on NNAPI.
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ TransposeConvOpBiasModel model(
+ GetRegistration(), /*output_shape=*/{1, 5, 5, 2},
+ /*filter=*/{TensorType_FLOAT32, {2, 3, 3, 1}},
+ /*filter_data=*/
+ {1, 3, 5, 7, 9, 11, 13, 15, 17, 2, 4, 6, 8, 10, 12, 14, 16, 18},
+ /*input=*/{TensorType_FLOAT32, {1, 2, 2, 1}},
+ /*output=*/{TensorType_FLOAT32, {}}, Padding_VALID,
+ /*stride_w=*/2, /*stride_h=*/2, GetTestType(), /* version */ 3);
+ model.SetInput({1, 2, 3, 4});
+ model.SetBias({3, 4});
+ model.Invoke();
+
+ EXPECT_THAT(
+ model.GetOutput(),
+ ElementsAreArray({4, 6, 6, 8, 10, 14, 9, 12, 13, 16, 10, 12, 12,
+ 14, 28, 32, 21, 24, 25, 28, 19, 24, 27, 32, 65, 76,
+ 45, 52, 57, 64, 24, 28, 30, 34, 64, 72, 39, 44, 47,
+ 52, 42, 46, 48, 52, 106, 114, 63, 68, 71, 76}));
+ EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 5, 5, 2}));
+}
+
+class QuantizedTransposeConvBiasOpModel
+ : public BaseTransposeConvBiasOpModel<uint8_t> {
+ public:
+ using BaseTransposeConvBiasOpModel::BaseTransposeConvBiasOpModel;
+
+ std::vector<float> GetDequantizedOutput() {
+ return Dequantize<uint8_t>(ExtractVector<uint8_t>(output_),
+ GetScale(output_), GetZeroPoint(output_));
+ }
+};
+
+TEST_P(TransposeConvOpTest, SimpleBiasTestQuantized) {
+ // TODO(b/138722124): Enable these tests on NNAPI.
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+ // Float would be {1, 2, 3, 4, 5, 6, 7, 8, 9}
+ std::initializer_list<uint8_t> filter_data = {129, 131, 133, 135, 137,
+ 139, 141, 143, 145};
+ QuantizedTransposeConvBiasOpModel model(
+ GetRegistration(), {1, 4, 4, 1},
+ {TensorType_UINT8, {1, 3, 3, 1}, -63.5, 64}, filter_data,
+ {TensorType_UINT8, {1, 4, 4, 1}, -63.5, 64},
+ {TensorType_UINT8, {}, -508, 512}, Padding_SAME, 1, 1, GetTestType(),
+ /* version */ 3);
+ model.SetInput({1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
+ model.SetBias({1});
+ model.Invoke();
+
+ EXPECT_THAT(
+ model.GetDequantizedOutput(),
+ ElementsAreArray(ArrayFloatNear({32, 64, 84, 76, 100, 192, 240, 200, 208,
+ 372, 420, 332, 264, 448, 488, 368},
+ 1e-5)));
+
+ // GetOutputShape() should always be same as model.SetOutputShape(...);
+ EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 4, 4, 1}));
+}
+
+class PerChannelQuantizedTransposeConvBiasOpModel
+ : public BaseTransposeConvBiasOpModel<int8_t> {
+ public:
+ using BaseTransposeConvBiasOpModel::BaseTransposeConvBiasOpModel;
+
+ std::vector<float> GetDequantizedOutput() {
+ return Dequantize<int8_t>(ExtractVector<int8_t>(output_), GetScale(output_),
+ GetZeroPoint(output_));
+ }
+
+ void SetInput(const std::initializer_list<float>& data) {
+ QuantizeAndPopulate<int8_t>(input_, data);
+ }
+
+ void SetFilter(const std::initializer_list<float>& data) {
+ PerChannelSymmetricQuantizeAndPopulate(filter_, data);
+ }
+};
+
+TEST_P(TransposeConvOpTest, SimpleBiasTestQuantizedPerChannelSingleChannel) {
+ // TODO(b/138722124): Enable these tests on NNAPI.
+ if (SingleOpModel::GetForceUseNnapi()) {
+ return;
+ }
+
+ const std::initializer_list<float> filter_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
+ PerChannelQuantizedTransposeConvBiasOpModel model(
+ GetRegistration(), {1, 4, 4, 1},
+ {TensorType_INT8, {1, 3, 3, 1}, 0, 0, 0, 0, true, {9.0 / 127}, {0}, 0},
+ {}, {TensorType_INT8, {1, 4, 4, 1}, 0, 0, 16.0 / 255, -128},
+ {TensorType_INT8, {}, 0, 0, 2, -128}, Padding_SAME, 1, 1, GetTestType(),
+ /* version */ 3);
+ model.SetInput({1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16});
+ model.SetFilter(filter_data);
+ model.SetBias({1});
+ model.Invoke();
+
+ EXPECT_THAT(
+ model.GetDequantizedOutput(),
+ ElementsAreArray(ArrayFloatNear({30, 62, 84, 76, 100, 194, 238, 200, 208,
+ 372, 418, 330, 264, 446, 486, 366},
+ 1e-5)));
+
+ // GetOutputShape() should always be same as model.SetOutputShape(...);
+ EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 4, 4, 1}));
+}
+
INSTANTIATE_TEST_SUITE_P(
TransposeConvOpTest, TransposeConvOpTest,
::testing::Combine(
::testing::ValuesIn(SingleOpTest::GetKernelTags(*kKernelMap)),
- ::testing::Values(TestType::CONST, TestType::DYNAMIC)));
+ ::testing::Values(TestType::kConst, TestType::kDynamic)));
} // namespace
} // namespace tflite
diff --git a/tensorflow/lite/kernels/unpack_test.cc b/tensorflow/lite/kernels/unpack_test.cc
index 1c6d8ff..0c6b8fa 100644
--- a/tensorflow/lite/kernels/unpack_test.cc
+++ b/tensorflow/lite/kernels/unpack_test.cc
@@ -87,15 +87,15 @@
struct UnpackOpTest : public ::testing::Test {
using TypeToTest = InputType;
TensorType TENSOR_TYPE =
- std::is_same<InputType, int16_t>::value
- ? TensorType_INT16
- : std::is_same<InputType, uint8_t>::value
- ? TensorType_UINT8
- : std::is_same<InputType, int8_t>::value
- ? TensorType_INT8
- : std::is_same<InputType, int32_t>::value
- ? TensorType_INT32
- : TensorType_FLOAT32;
+ (std::is_same<InputType, int16_t>::value
+ ? TensorType_INT16
+ : (std::is_same<InputType, uint8_t>::value
+ ? TensorType_UINT8
+ : (std::is_same<InputType, int8_t>::value
+ ? TensorType_INT8
+ : (std::is_same<InputType, int32_t>::value
+ ? TensorType_INT32
+ : TensorType_FLOAT32))));
};
using TestTypes = testing::Types<float, int32_t, int8_t, uint8_t, int16_t>;
diff --git a/tensorflow/lite/micro/benchmarks/BUILD b/tensorflow/lite/micro/benchmarks/BUILD
index f15284b..4af3267 100644
--- a/tensorflow/lite/micro/benchmarks/BUILD
+++ b/tensorflow/lite/micro/benchmarks/BUILD
@@ -51,3 +51,20 @@
"//tensorflow/lite/micro/testing:micro_benchmark",
],
)
+
+cc_binary(
+ name = "person_detection_benchmark",
+ srcs = ["person_detection_benchmark.cc"],
+ deps = [
+ "//tensorflow/lite:schema_fbs_version",
+ "//tensorflow/lite/c:common",
+ "//tensorflow/lite/micro:micro_framework",
+ "//tensorflow/lite/micro:micro_utils",
+ "//tensorflow/lite/micro/examples/person_detection:model_settings",
+ "//tensorflow/lite/micro/examples/person_detection:person_detect_model_data",
+ "//tensorflow/lite/micro/examples/person_detection:simple_images_test_data",
+ "//tensorflow/lite/micro/kernels:micro_ops",
+ "//tensorflow/lite/micro/testing:micro_benchmark",
+ "//tensorflow/lite/schema:schema_fbs",
+ ],
+)
diff --git a/tensorflow/lite/micro/benchmarks/Makefile.inc b/tensorflow/lite/micro/benchmarks/Makefile.inc
index 140616e..2a7eefd 100644
--- a/tensorflow/lite/micro/benchmarks/Makefile.inc
+++ b/tensorflow/lite/micro/benchmarks/Makefile.inc
@@ -1,3 +1,5 @@
+$(eval $(call add_third_party_download,$(PERSON_MODEL_URL),$(PERSON_MODEL_MD5),person_model_grayscale,))
+
KEYWORD_BENCHMARK_SRCS := \
tensorflow/lite/micro/benchmarks/keyword_benchmark.cc \
tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cc
@@ -5,6 +7,18 @@
KEYWORD_BENCHMARK_HDRS := \
tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.h
+PERSON_DETECTION_BENCHMARK_SRCS := \
+tensorflow/lite/micro/benchmarks/person_detection_benchmark.cc \
+$(MAKEFILE_DIR)/downloads/person_model_grayscale/no_person_image_data.cc \
+$(MAKEFILE_DIR)/downloads/person_model_grayscale/person_detect_model_data.cc \
+$(MAKEFILE_DIR)/downloads/person_model_grayscale/person_image_data.cc
+
+PERSON_DETECTION_BENCHMARK_HDRS := \
+tensorflow/lite/micro/examples/person_detection/person_detect_model_data.h
+
# Builds a standalone binary.
$(eval $(call microlite_test,keyword_benchmark,\
$(KEYWORD_BENCHMARK_SRCS),$(KEYWORD_BENCHMARK_HDRS)))
+
+$(eval $(call microlite_test,person_detection_benchmark,\
+$(PERSON_DETECTION_BENCHMARK_SRCS),$(PERSON_DETECTION_BENCHMARK_HDRS)))
diff --git a/tensorflow/lite/micro/benchmarks/README.md b/tensorflow/lite/micro/benchmarks/README.md
index e6738bb..7219555 100644
--- a/tensorflow/lite/micro/benchmarks/README.md
+++ b/tensorflow/lite/micro/benchmarks/README.md
@@ -7,8 +7,10 @@
## Table of contents
- [Keyword Benchmark](#keyword-benchmark)
+- [Person Detection Benchmark](#person-detection-benchmark)
- [Run on x86](#run-on-x86)
- [Run on Xtensa XPG Simulator](#run-on-xtensa-xpg-simulator)
+- [Run on Sparkfun Edge](#run-on-sparkfun-edge)
## Keyword benchmark
@@ -17,17 +19,48 @@
Since the weights are scrambled, the output is meaningless. In order to validate
the accuracy of optimized kernels, please run the kernel tests.
+## Person detection benchmark
+
+The keyword benchmark provides a way to evaluate the performance of the 250KB
+visual wakewords model.
+
## Run on x86
To run the keyword benchmark on x86, run
+
```
make -f tensorflow/lite/micro/tools/make/Makefile TAGS=posix test_keyword_benchmark
```
+To run the person detection benchmark on x86, run
+
+```
+make -f tensorflow/lite/micro/tools/make/Makefile TAGS=posix test_person_detection_benchmark
+```
+
## Run on Xtensa XPG Simulator
To run the keyword benchmark on the Xtensa XPG simulator, you will need a valid
Xtensa toolchain and license. With these set up, run:
+
```
make -f tensorflow/lite/micro/tools/make/Makefile TARGET=xtensa-xpg XTENSA_CORE=<xtensa core> TAGS=xtensa_hifimini test_keyword_benchmark -j18
```
+
+## Run on Sparkfun Edge
+The following instructions will help you build and deploy this benchmark on the
+[SparkFun Edge development board](https://sparkfun.com/products/15170).
+
+
+If you're new to using this board, we recommend walking through the
+[AI on a microcontroller with TensorFlow Lite and SparkFun Edge](https://codelabs.developers.google.com/codelabs/sparkfun-tensorflow)
+codelab to get an understanding of the workflow.
+
+Build binary using
+
+```
+make -f tensorflow/lite/micro/tools/make/Makefile TARGET=sparkfun_edge person_detection_benchmark_bin
+```
+
+Refer to flashing instructions in the [Person Detection Example](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/person_detection/README.md#running-on-sparkfun-edge).
+
diff --git a/tensorflow/lite/micro/benchmarks/person_detection_benchmark.cc b/tensorflow/lite/micro/benchmarks/person_detection_benchmark.cc
new file mode 100644
index 0000000..5287a9c
--- /dev/null
+++ b/tensorflow/lite/micro/benchmarks/person_detection_benchmark.cc
@@ -0,0 +1,132 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/examples/person_detection/model_settings.h"
+#include "tensorflow/lite/micro/examples/person_detection/no_person_image_data.h"
+#include "tensorflow/lite/micro/examples/person_detection/person_detect_model_data.h"
+#include "tensorflow/lite/micro/examples/person_detection/person_image_data.h"
+#include "tensorflow/lite/micro/kernels/micro_ops.h"
+#include "tensorflow/lite/micro/micro_error_reporter.h"
+#include "tensorflow/lite/micro/micro_interpreter.h"
+#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+#include "tensorflow/lite/micro/micro_utils.h"
+#include "tensorflow/lite/micro/testing/micro_benchmark.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+// Create an area of memory to use for input, output, and intermediate arrays.
+constexpr int tensor_arena_size = 73 * 1024;
+uint8_t tensor_arena[tensor_arena_size];
+
+/*
+ * Person Detection benchmark. Evaluates runtime performance of the visual
+ * wakewords person detection model. This is the same model found in
+ * exmaples/person_detection.
+ */
+
+namespace {
+
+// Create an area of memory to use for input, output, and intermediate arrays.
+constexpr int tensor_arena_size = 73 * 1024;
+uint8_t tensor_arena[tensor_arena_size];
+
+class PersonDetectionRunner {
+ public:
+ PersonDetectionRunner()
+ : person_detection_model_(tflite::GetModel(g_person_detect_model_data)),
+ reporter_(µ_reporter_),
+ interpreter_(person_detection_model_, resolver_, tensor_arena,
+ tensor_arena_size, reporter_) {
+ resolver_.AddBuiltin(tflite::BuiltinOperator_DEPTHWISE_CONV_2D,
+ tflite::ops::micro::Register_DEPTHWISE_CONV_2D());
+ resolver_.AddBuiltin(tflite::BuiltinOperator_CONV_2D,
+ tflite::ops::micro::Register_CONV_2D());
+ resolver_.AddBuiltin(tflite::BuiltinOperator_AVERAGE_POOL_2D,
+ tflite::ops::micro::Register_AVERAGE_POOL_2D());
+ interpreter_.AllocateTensors();
+
+ TfLiteTensor* input = interpreter_.input(0);
+ TFLITE_CHECK_EQ(input->type, kTfLiteUInt8);
+ }
+
+ void RunSingleIterationWithPerson() {
+ // Populate input tensor with an image with a person
+ TfLiteTensor* input = interpreter_.input(0);
+ int8_t* input_buffer = tflite::GetTensorData<int8_t>(input);
+ int input_length = tflite::ElementCount(*input->dims);
+ for (int i = 0; i < input_length; i++) {
+ input_buffer[i] = g_person_data[i];
+ }
+
+ // Run the model on this input and make sure it succeeds.
+ TfLiteStatus invoke_status = interpreter_.Invoke();
+ if (invoke_status != kTfLiteOk) {
+ TF_LITE_REPORT_ERROR(reporter_, "Invoke failed.");
+ }
+ }
+
+ void RunSingleIterationWithoutPerson() {
+ // Populate input tensor with an image with no person.
+ TfLiteTensor* input = interpreter_.input(0);
+ int8_t* input_buffer = tflite::GetTensorData<int8_t>(input);
+ int input_length = tflite::ElementCount(*input->dims);
+ for (int i = 0; i < input_length; i++) {
+ input_buffer[i] = g_no_person_data[i];
+ }
+
+ // Run the model on this input and make sure it succeeds.
+ TfLiteStatus invoke_status = interpreter_.Invoke();
+ if (invoke_status != kTfLiteOk) {
+ TF_LITE_REPORT_ERROR(reporter_, "Invoke failed.");
+ }
+ }
+
+ private:
+ const tflite::Model* person_detection_model_;
+ tflite::MicroErrorReporter micro_reporter_;
+ tflite::ErrorReporter* reporter_;
+ tflite::MicroOpResolver<6> resolver_;
+ tflite::MicroInterpreter interpreter_;
+};
+
+// NOLINTNEXTLINE
+PersonDetectionRunner runner;
+
+void PersonDetectionFirstIteration() { runner.RunSingleIterationWithPerson(); }
+
+void PersonDetectionTenIerationsWithPerson() {
+ // TODO(b/152644476): Add a way to run more than a single deterministic input.
+ for (int i = 0; i < 10; i++) {
+ runner.RunSingleIterationWithPerson();
+ }
+}
+
+void PersonDetectionTenIerationsWithoutPerson() {
+ // TODO(b/152644476): Add a way to run more than a single deterministic input.
+ for (int i = 0; i < 10; i++) {
+ runner.RunSingleIterationWithoutPerson();
+ }
+}
+
+} // namespace
+
+TF_LITE_MICRO_BENCHMARKS_BEGIN
+
+TF_LITE_MICRO_BENCHMARK(PersonDetectionFirstIteration);
+TF_LITE_MICRO_BENCHMARK(PersonDetectionTenIerationsWithPerson);
+TF_LITE_MICRO_BENCHMARK(PersonDetectionTenIerationsWithoutPerson);
+
+TF_LITE_MICRO_BENCHMARKS_END
diff --git a/tensorflow/lite/micro/examples/hello_world/hello_world_test.cc b/tensorflow/lite/micro/examples/hello_world/hello_world_test.cc
index 75dd607..3d1155e 100644
--- a/tensorflow/lite/micro/examples/hello_world/hello_world_test.cc
+++ b/tensorflow/lite/micro/examples/hello_world/hello_world_test.cc
@@ -43,8 +43,8 @@
tflite::ops::micro::AllOpsResolver resolver;
// Create an area of memory to use for input, output, and intermediate arrays.
- // Finding the minimum value for your model may require some trial and error.
- const int tensor_arena_size = 2 * 1024;
+ // `arena_used_bytes` can be used to retrieve the optimal size.
+ const int tensor_arena_size = 2208 + 16 + 100 /* some reserved space */;
uint8_t tensor_arena[tensor_arena_size];
// Build an interpreter to run the model with
@@ -53,6 +53,10 @@
// Allocate memory from the tensor_arena for the model's tensors
TF_LITE_MICRO_EXPECT_EQ(interpreter.AllocateTensors(), kTfLiteOk);
+ // At the time of writing, the hello world model uses 2208 bytes, we leave
+ // 100 bytes head room here to make the test less fragile and in the same
+ // time, alert for substantial increase.
+ TF_LITE_MICRO_EXPECT_LE(interpreter.arena_used_bytes(), 2208 + 100);
// Obtain a pointer to the model's input tensor
TfLiteTensor* input = interpreter.input(0);
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/.gitignore b/tensorflow/lite/micro/examples/image_recognition_experimental/.gitignore
new file mode 100644
index 0000000..5762b70
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/.gitignore
@@ -0,0 +1 @@
+first_10_cifar_images.h
\ No newline at end of file
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/Makefile.inc b/tensorflow/lite/micro/examples/image_recognition_experimental/Makefile.inc
new file mode 100644
index 0000000..2fdfb0e
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/Makefile.inc
@@ -0,0 +1,38 @@
+$(eval $(call add_third_party_download,$(IMAGE_RECOGNITION_MODEL_URL),$(IMAGE_RECOGNITION_MODEL_MD5),image_recognition_model,))
+$(eval $(call add_third_party_download,$(CIFAR10_DATASET_URL),$(CIFAR10_DATASET_MD5),cifar10,patch_cifar10_dataset))
+
+IMAGE_RECOGNITION_HDRS := \
+tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h \
+tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h \
+tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h \
+tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h \
+tensorflow/lite/micro/examples/image_recognition_experimental/util.h
+
+IMAGE_RECOGNITION_SRCS := \
+$(MAKEFILE_DIR)/downloads/image_recognition_model/image_recognition_model.cc \
+tensorflow/lite/micro/examples/image_recognition_experimental/main.cc \
+tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_provider.cc \
+tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.cc \
+tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.cc
+
+IMAGE_RECOGNITION_TEST_SRCS := \
+tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc \
+$(MAKEFILE_DIR)/downloads/image_recognition_model/image_recognition_model.cc
+
+IMAGE_RECOGNITION_TEST_HDRS := \
+tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h \
+tensorflow/lite/micro/examples/image_recognition_experimental/util.h
+
+include $(wildcard tensorflow/lite/micro/examples/image_recognition_experimental/*/Makefile.inc)
+
+ifneq ($(filter disco_f746ng,$(ALL_TAGS)),)
+ MBED_PROJECT_FILES += \
+ BSP_DISCO_F746NG.lib \
+ LCD_DISCO_F746NG.lib
+endif
+
+$(eval $(call microlite_test,image_recognition,\
+$(IMAGE_RECOGNITION_SRCS),$(IMAGE_RECOGNITION_HDRS)))
+
+$(eval $(call microlite_test,image_recognition_test,\
+$(IMAGE_RECOGNITION_TEST_SRCS),$(IMAGE_RECOGNITION_TEST_HDRS)))
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/README.md b/tensorflow/lite/micro/examples/image_recognition_experimental/README.md
new file mode 100644
index 0000000..7a29d2f
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/README.md
@@ -0,0 +1,90 @@
+# Image Recognition Example
+
+## Table of Contents
+
+- [Introduction](#introduction)
+- [Hardware](#hardware)
+- [Building](#building)
+ - [Building the testcase](#building-the-testcase)
+ - [Building the image recognition application](#building-the-image-recognition-application)
+ - [Prerequisites](#prerequisites)
+ - [Compiling and flashing](#compiling-and-flashing)
+
+## Introduction
+
+This example shows how you can use Tensorflow Lite Micro to perform image
+recognition on a
+[STM32F746 discovery kit](https://www.st.com/en/evaluation-tools/32f746gdiscovery.html)
+with a STM32F4DIS-CAM camera module attached. It classifies the captured image
+into 1 of 10 different classes, and those classes are "Plane", "Car", "Bird",
+"Cat", "Deer", "Dog", "Frog", "Horse", "Ship", "Truck".
+
+## Hardware
+
+[STM32F746G-DISCO board (Cortex-M7)](https://www.st.com/en/evaluation-tools/32f746gdiscovery.html)
+\
+[STM32F4DIS-CAM Camera module](https://www.element14.com/community/docs/DOC-67585?ICID=knode-STM32F4-cameramore)
+
+## Building
+
+These instructions have been tested on Ubuntu 16.04.
+
+### Building the test case
+
+```
+$ make -f tensorflow/lite/micro/tools/make/Makefile image_recognition_test
+```
+
+This will build and run the test case. As input, the test case uses the first 10
+images of the test batch included in the
+[CIFAR10](https://www.cs.toronto.edu/~kriz/cifar.html) dataset. Details
+surrounding the dataset can be found in
+[this paper](https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf).
+
+### Building the image recognition application
+
+#### Prerequisites
+
+Install mbed-cli: `$ pip install mbed-cli`
+
+Install the arm-none-eabi-toolchain.
+
+For Ubuntu, this can be done by installing the package `gcc-arm-none-eabi`. In
+Ubuntu 16.04, the version included in the repository is 4.9.3 while the
+recommended version is 6 and up. Later versions can be downloaded from
+[here](https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-rm/downloads)
+for Windows, Mac OS X and Linux.
+
+#### Compiling and flashing
+
+In order to generate the mbed project, run the following command: `$ make -f
+tensorflow/lite/micro/tools/make/Makefile TAGS=disco_f746ng
+generate_image_recognition_mbed_project` This will copy all of the necessary
+files needed to build and flash the application.
+
+Navigate to the output folder: `$ cd
+tensorflow/lite/micro/tools/make/gen/linux_x86_64/prj/image_recognition/mbed/`
+
+The following instructions for compiling and flashing can also be found in the
+file README_MBED.md in the output folder.
+
+To load the dependencies required, run: `$ mbed config root . $ mbed deploy`
+
+In order to compile, run: `mbed compile -m auto -t GCC_ARM --profile release`
+
+`-m auto`: Automatically detects the correct target if the Discovery board is
+connected to the computer. If the board is not connected, replace `auto` with
+`DISCO_F746NG`. \
+`-t GCC_ARM`: Specifies the toolchain used to compile. `GCC_ARM` indicates that
+the arm-none-eabi-toolchain will be used. \
+`--profile release`: Build the `release` profile. The different profiles can be
+found under mbed-os/tools/profiles/.
+
+This will produce a file named `mbed.bin` in
+`BUILD/DISCO_F746NG/GCC_ARM-RELEASE/`. To flash it to the board, simply copy the
+file to the volume mounted as a USB drive. Alternatively, the `-f` option can be
+appended to flash automatically after compilation.
+
+On Ubuntu 16.04 (and possibly other Linux distributions) there may be an error
+message when running `mbed compile` saying that the Python module `pywin32`
+failed to install. This message can be ignored.
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h b/tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h
new file mode 100644
index 0000000..b466796
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h
@@ -0,0 +1,41 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_PROVIDER_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_PROVIDER_H_
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/micro_error_reporter.h"
+
+TfLiteStatus InitCamera(tflite::ErrorReporter* error_reporter);
+
+// This is an abstraction around an image source like a camera, and is
+// expected to return 8-bit sample data. The assumption is that this will be
+// called in a low duty-cycle fashion in a low-power application. In these
+// cases, the imaging sensor need not be run in a streaming mode, but rather can
+// be idled in a relatively low-power mode between calls to GetImage(). The
+// assumption is that the overhead and time of bringing the low-power sensor out
+// of this standby mode is commensurate with the expected duty cycle of the
+// application. The underlying sensor may actually be put into a streaming
+// configuration, but the image buffer provided to GetImage should not be
+// overwritten by the driver code until the next call to GetImage();
+//
+// The reference implementation can have no platform-specific dependencies, so
+// it just returns a static image. For real applications, you should
+// ensure there's a specialized implementation that accesses hardware APIs.
+TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int image_width,
+ int image_height, int channels, uint8_t* image_data);
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_PROVIDER_H_
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h
new file mode 100644
index 0000000..a32dcd0
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h
@@ -0,0 +1,27 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a standard TensorFlow Lite model file that has been converted into a
+// C data array, so it can be easily compiled into a binary for devices that
+// don't have a file system. It can be created using the command:
+// xxd -i image_recognition_model.tflite > image_recognition_model.cc
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_RECOGNITION_MODEL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_RECOGNITION_MODEL_H_
+
+extern const unsigned char image_recognition_model_data[];
+extern const unsigned int image_recognition_model_data_len;
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_IMAGE_RECOGNITION_MODEL_H_
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc
new file mode 100644
index 0000000..d4dfee4
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_test.cc
@@ -0,0 +1,105 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/first_10_cifar_images.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/util.h"
+#include "tensorflow/lite/micro/kernels/micro_ops.h"
+#include "tensorflow/lite/micro/micro_error_reporter.h"
+#include "tensorflow/lite/micro/micro_interpreter.h"
+#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+#include "tensorflow/lite/micro/testing/micro_test.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+#define IMAGE_BYTES 3072
+#define LABEL_BYTES 1
+#define ENTRY_BYTES (IMAGE_BYTES + LABEL_BYTES)
+
+TF_LITE_MICRO_TESTS_BEGIN
+
+TF_LITE_MICRO_TEST(TestImageRecognitionInvoke) {
+ tflite::MicroErrorReporter micro_error_reporter;
+ tflite::ErrorReporter* error_reporter = µ_error_reporter;
+
+ const tflite::Model* model = ::tflite::GetModel(image_recognition_model_data);
+ if (model->version() != TFLITE_SCHEMA_VERSION) {
+ TF_LITE_REPORT_ERROR(error_reporter,
+ "Model provided is schema version %d not equal "
+ "to supported version %d.\n",
+ model->version(), TFLITE_SCHEMA_VERSION);
+ }
+
+ tflite::MicroOpResolver<4> micro_op_resolver;
+
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D,
+ tflite::ops::micro::Register_CONV_2D());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D,
+ tflite::ops::micro::Register_MAX_POOL_2D());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED,
+ tflite::ops::micro::Register_FULLY_CONNECTED());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX,
+ tflite::ops::micro::Register_SOFTMAX());
+
+ const int tensor_arena_size = 45 * 1024;
+ uint8_t tensor_arena[tensor_arena_size];
+
+ tflite::MicroInterpreter interpreter(model, micro_op_resolver, tensor_arena,
+ tensor_arena_size, error_reporter);
+ interpreter.AllocateTensors();
+
+ TfLiteTensor* input = interpreter.input(0);
+ TF_LITE_MICRO_EXPECT_NE(nullptr, input);
+ TF_LITE_MICRO_EXPECT_EQ(4, input->dims->size);
+ TF_LITE_MICRO_EXPECT_EQ(1, input->dims->data[0]);
+ TF_LITE_MICRO_EXPECT_EQ(32, input->dims->data[1]);
+ TF_LITE_MICRO_EXPECT_EQ(32, input->dims->data[2]);
+ TF_LITE_MICRO_EXPECT_EQ(3, input->dims->data[3]);
+ TF_LITE_MICRO_EXPECT_EQ(kTfLiteUInt8, input->type);
+
+ int num_correct = 0;
+ int num_images = 10;
+ for (int image_num = 0; image_num < num_images; image_num++) {
+ memset(input->data.uint8, 0, input->bytes);
+
+ uint8_t correct_label = 0;
+
+ correct_label =
+ tensorflow_lite_micro_tools_make_downloads_cifar10_test_batch_bin
+ [image_num * ENTRY_BYTES];
+ memcpy(input->data.uint8,
+ &tensorflow_lite_micro_tools_make_downloads_cifar10_test_batch_bin
+ [image_num * ENTRY_BYTES + LABEL_BYTES],
+ IMAGE_BYTES);
+ reshape_cifar_image(input->data.uint8, IMAGE_BYTES);
+
+ TfLiteStatus invoke_status = interpreter.Invoke();
+ TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, invoke_status);
+ if (invoke_status != kTfLiteOk) {
+ TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed\n");
+ }
+
+ TfLiteTensor* output = interpreter.output(0);
+ int guess = get_top_prediction(output->data.uint8, 10);
+
+ if (correct_label == guess) {
+ num_correct++;
+ }
+ }
+
+ TF_LITE_MICRO_EXPECT_EQ(6, num_correct);
+}
+
+TF_LITE_MICRO_TESTS_END
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc
new file mode 100644
index 0000000..613c97f
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/main.cc
@@ -0,0 +1,108 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// NOLINTNEXTLINE
+#include "mbed.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/image_recognition_model.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h"
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/util.h"
+#include "tensorflow/lite/micro/kernels/micro_ops.h"
+#include "tensorflow/lite/micro/micro_error_reporter.h"
+#include "tensorflow/lite/micro/micro_interpreter.h"
+#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+#define NUM_OUT_CH 3
+#define CNN_IMG_SIZE 32
+
+uint8_t camera_buffer[NUM_IN_CH * IN_IMG_WIDTH * IN_IMG_HEIGHT]
+ __attribute__((aligned(4)));
+static const char* labels[] = {"Plane", "Car", "Bird", "Cat", "Deer",
+ "Dog", "Frog", "Horse", "Ship", "Truck"};
+
+int main(int argc, char** argv) {
+ init_lcd();
+ wait_ms(100);
+
+ tflite::MicroErrorReporter micro_error_reporter;
+ tflite::ErrorReporter* error_reporter = µ_error_reporter;
+
+ if (InitCamera(error_reporter) != kTfLiteOk) {
+ TF_LITE_REPORT_ERROR(error_reporter, "Failed to init camera.");
+ return 1;
+ }
+
+ const tflite::Model* model = ::tflite::GetModel(image_recognition_model_data);
+ if (model->version() != TFLITE_SCHEMA_VERSION) {
+ TF_LITE_REPORT_ERROR(error_reporter,
+ "Model provided is schema version %d not equal "
+ "to supported version %d.",
+ model->version(), TFLITE_SCHEMA_VERSION);
+ return 1;
+ }
+
+ tflite::MicroOpResolver<4> micro_op_resolver;
+
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_CONV_2D,
+ tflite::ops::micro::Register_CONV_2D());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_MAX_POOL_2D,
+ tflite::ops::micro::Register_MAX_POOL_2D());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_FULLY_CONNECTED,
+ tflite::ops::micro::Register_FULLY_CONNECTED());
+ micro_op_resolver.AddBuiltin(tflite::BuiltinOperator_SOFTMAX,
+ tflite::ops::micro::Register_SOFTMAX());
+
+ constexpr int tensor_arena_size = 45 * 1024;
+ uint8_t tensor_arena[tensor_arena_size];
+ tflite::MicroInterpreter interpreter(model, resolver, tensor_arena,
+ tensor_arena_size, error_reporter);
+ interpreter.AllocateTensors();
+
+ while (true) {
+ TfLiteTensor* input = interpreter.input(0);
+
+ GetImage(error_reporter, IN_IMG_WIDTH, IN_IMG_HEIGHT, NUM_OUT_CH,
+ camera_buffer);
+
+ ResizeConvertImage(error_reporter, IN_IMG_WIDTH, IN_IMG_HEIGHT, NUM_IN_CH,
+ CNN_IMG_SIZE, CNN_IMG_SIZE, NUM_OUT_CH, camera_buffer,
+ input->data.uint8);
+
+ if (input->type != kTfLiteUInt8) {
+ TF_LITE_REPORT_ERROR(error_reporter, "Wrong input type.");
+ }
+
+ TfLiteStatus invoke_status = interpreter.Invoke();
+ if (invoke_status != kTfLiteOk) {
+ TF_LITE_REPORT_ERROR(error_reporter, "Invoke failed.");
+ break;
+ }
+
+ display_image_rgb565(IN_IMG_WIDTH, IN_IMG_HEIGHT, camera_buffer, 40, 40);
+ display_image_rgb888(CNN_IMG_SIZE, CNN_IMG_SIZE, input->data.uint8, 300,
+ 100);
+
+ TfLiteTensor* output = interpreter.output(0);
+
+ int top_ind = get_top_prediction(output->data.uint8, 10);
+ print_prediction(labels[top_ind]);
+ print_confidence(output->data.uint8[top_ind]);
+ }
+
+ return 0;
+}
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.cc
new file mode 100644
index 0000000..22e03c6
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.cc
@@ -0,0 +1,79 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h"
+
+#include <stdint.h>
+
+#include "LCD_DISCO_F746NG/LCD_DISCO_F746NG.h"
+
+LCD_DISCO_F746NG lcd;
+
+extern "C" {
+// defined in stm32746g_discovery_camera.c
+extern DCMI_HandleTypeDef hDcmiHandler;
+void DCMI_IRQHandler(void) { HAL_DCMI_IRQHandler(&hDcmiHandler); }
+void DMA2_Stream1_IRQHandler(void) {
+ HAL_DMA_IRQHandler(hDcmiHandler.DMA_Handle);
+}
+}
+
+static char lcd_output_string[50];
+
+void init_lcd() { lcd.Clear(LCD_COLOR_WHITE); }
+
+void display_image_rgb888(int x_dim, int y_dim, const uint8_t* image_data,
+ int x_loc, int y_loc) {
+ for (int y = 0; y < y_dim; ++y) {
+ for (int x = 0; x < x_dim; ++x, image_data += 3) {
+ uint8_t a = 0xFF;
+ auto r = image_data[0];
+ auto g = image_data[1];
+ auto b = image_data[2];
+ int pixel = a << 24 | r << 16 | g << 8 | b;
+ lcd.DrawPixel(x_loc + x, y_loc + y, pixel);
+ }
+ }
+}
+
+void display_image_rgb565(int x_dim, int y_dim, const uint8_t* image_data,
+ int x_loc, int y_loc) {
+ for (int y = 0; y < y_dim; ++y) {
+ for (int x = 0; x < x_dim; ++x, image_data += 2) {
+ uint8_t a = 0xFF;
+ uint8_t pix_lo = image_data[0];
+ uint8_t pix_hi = image_data[1];
+ uint8_t r = (0xF8 & pix_hi);
+ uint8_t g = ((0x07 & pix_hi) << 5) | ((0xE0 & pix_lo) >> 3);
+ uint8_t b = (0x1F & pix_lo) << 3;
+ int pixel = a << 24 | r << 16 | g << 8 | b;
+ // inverted image, so draw from bottom-right to top-left
+ lcd.DrawPixel(x_loc + (x_dim - x), y_loc + (y_dim - y), pixel);
+ }
+ }
+}
+
+void print_prediction(const char* prediction) {
+ // NOLINTNEXTLINE
+ sprintf(lcd_output_string, " Prediction: %s ", prediction);
+ lcd.DisplayStringAt(0, LINE(8), (uint8_t*)lcd_output_string, LEFT_MODE);
+}
+
+void print_confidence(uint8_t max_score) {
+ // NOLINTNEXTLINE
+ sprintf(lcd_output_string, " Confidence: %.1f%% ",
+ (max_score / 255.0) * 100.0);
+ lcd.DisplayStringAt(0, LINE(9), (uint8_t*)lcd_output_string, LEFT_MODE);
+}
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h
new file mode 100644
index 0000000..b114812
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/display_util.h
@@ -0,0 +1,33 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_DISPLAY_UTIL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_DISPLAY_UTIL_H_
+
+#include <stdint.h>
+
+void init_lcd();
+
+void display_image_rgb888(int x_dim, int y_dim, const uint8_t* image_data,
+ int x_loc, int y_loc);
+
+void display_image_rgb565(int x_dim, int y_dim, const uint8_t* image_data,
+ int x_loc, int y_loc);
+
+void print_prediction(const char* prediction);
+
+void print_confidence(uint8_t max_score);
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_DISPLAY_UTIL_H_
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_provider.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_provider.cc
new file mode 100644
index 0000000..594af5b
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_provider.cc
@@ -0,0 +1,39 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/image_provider.h"
+
+#include "BSP_DISCO_F746NG/Drivers/BSP/STM32746G-Discovery/stm32746g_discovery_camera.h"
+
+TfLiteStatus InitCamera(tflite::ErrorReporter* error_reporter) {
+ if (BSP_CAMERA_Init(RESOLUTION_R160x120) != CAMERA_OK) {
+ TF_LITE_REPORT_ERROR(error_reporter, "Failed to init camera.\n");
+ return kTfLiteError;
+ }
+
+ return kTfLiteOk;
+}
+
+TfLiteStatus GetImage(tflite::ErrorReporter* error_reporter, int frame_width,
+ int frame_height, int channels, uint8_t* frame) {
+ // For consistency, the signature of this function is the
+ // same as the GetImage-function in micro_vision.
+ (void)error_reporter;
+ (void)frame_width;
+ (void)frame_height;
+ (void)channels;
+ BSP_CAMERA_SnapshotStart(frame);
+ return kTfLiteOk;
+}
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.cc b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.cc
new file mode 100644
index 0000000..49544fd
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.cc
@@ -0,0 +1,49 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h"
+
+void ResizeConvertImage(tflite::ErrorReporter* error_reporter,
+ int in_frame_width, int in_frame_height,
+ int num_in_channels, int out_frame_width,
+ int out_frame_height, int channels,
+ const uint8_t* in_image, uint8_t* out_image) {
+ // offset so that only the center part of rectangular image is selected for
+ // resizing
+ int width_offset = ((in_frame_width - in_frame_height) / 2) * num_in_channels;
+
+ int yresize_ratio = (in_frame_height / out_frame_height) * num_in_channels;
+ int xresize_ratio = (in_frame_width / out_frame_width) * num_in_channels;
+ int resize_ratio =
+ (xresize_ratio < yresize_ratio) ? xresize_ratio : yresize_ratio;
+
+ for (int y = 0; y < out_frame_height; y++) {
+ for (int x = 0; x < out_frame_width; x++) {
+ int orig_img_loc =
+ y * in_frame_width * resize_ratio + x * resize_ratio + width_offset;
+ // correcting the image inversion here
+ int out_img_loc = ((out_frame_height - 1 - y) * out_frame_width +
+ (out_frame_width - 1 - x)) *
+ channels;
+ uint8_t pix_lo = in_image[orig_img_loc];
+ uint8_t pix_hi = in_image[orig_img_loc + 1];
+ // convert RGB565 to RGB888
+ out_image[out_img_loc] = (0xF8 & pix_hi);
+ out_image[out_img_loc + 1] =
+ ((0x07 & pix_hi) << 5) | ((0xE0 & pix_lo) >> 3);
+ out_image[out_img_loc + 2] = (0x1F & pix_lo) << 3;
+ }
+ }
+}
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h
new file mode 100644
index 0000000..5e8a7e6
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/stm32f746_discovery/image_util.h
@@ -0,0 +1,32 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_IMAGE_UTIL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_IMAGE_UTIL_H_
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/micro_error_reporter.h"
+
+#define NUM_IN_CH 2
+#define IN_IMG_WIDTH 160
+#define IN_IMG_HEIGHT 120
+
+void ResizeConvertImage(tflite::ErrorReporter* error_reporter,
+ int in_frame_width, int in_frame_height,
+ int num_in_channels, int out_frame_width,
+ int out_frame_height, int channels,
+ const uint8_t* in_frame, uint8_t* out_frame);
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_STM32F746_DISCOVERY_IMAGE_UTIL_H_
diff --git a/tensorflow/lite/micro/examples/image_recognition_experimental/util.h b/tensorflow/lite/micro/examples/image_recognition_experimental/util.h
new file mode 100644
index 0000000..7927e1b
--- /dev/null
+++ b/tensorflow/lite/micro/examples/image_recognition_experimental/util.h
@@ -0,0 +1,64 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_UTIL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_UTIL_H_
+
+#include <stdint.h>
+#include <string.h>
+
+#define IMAGE_SIZE 3072
+#define CHANNEL_SIZE 1024
+#define R_CHANNEL_OFFSET 0
+#define G_CHANNEL_OFFSET CHANNEL_SIZE
+#define B_CHANNEL_OFFSET (CHANNEL_SIZE * 2)
+
+int get_top_prediction(const uint8_t* predictions, int num_categories) {
+ int max_score = predictions[0];
+ int guess = 0;
+
+ for (int category_index = 1; category_index < num_categories;
+ category_index++) {
+ const uint8_t category_score = predictions[category_index];
+ if (category_score > max_score) {
+ max_score = category_score;
+ guess = category_index;
+ }
+ }
+
+ return guess;
+}
+
+void reshape_cifar_image(uint8_t* image_data, int num_bytes) {
+ uint8_t temp_data[IMAGE_SIZE];
+
+ memcpy(temp_data, image_data, num_bytes);
+
+ int k = 0;
+ for (int i = 0; i < CHANNEL_SIZE; i++) {
+ int r_ind = R_CHANNEL_OFFSET + i;
+ int g_ind = G_CHANNEL_OFFSET + i;
+ int b_ind = B_CHANNEL_OFFSET + i;
+
+ image_data[k] = temp_data[r_ind];
+ k++;
+ image_data[k] = temp_data[g_ind];
+ k++;
+ image_data[k] = temp_data[b_ind];
+ k++;
+ }
+}
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_IMAGE_RECOGNITION_EXPERIMENTAL_UTIL_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/BUILD b/tensorflow/lite/micro/examples/micro_speech/BUILD
index cc80ce6..d724972 100644
--- a/tensorflow/lite/micro/examples/micro_speech/BUILD
+++ b/tensorflow/lite/micro/examples/micro_speech/BUILD
@@ -22,12 +22,12 @@
)
cc_library(
- name = "tiny_conv_simple_features_model_data",
+ name = "model",
srcs = [
- "simple_features/tiny_conv_simple_features_model_data.cc",
+ "simple_features/model.cc",
],
hdrs = [
- "simple_features/tiny_conv_simple_features_model_data.h",
+ "simple_features/model.h",
],
)
@@ -52,7 +52,7 @@
"//tensorflow/lite:schema_fbs_version",
"//tensorflow/lite/micro:micro_framework",
"//tensorflow/lite/micro/examples/micro_speech/micro_features:micro_features_test_data",
- "//tensorflow/lite/micro/examples/micro_speech/micro_features:tiny_conv_micro_features_model_data",
+ "//tensorflow/lite/micro/examples/micro_speech/micro_features:model",
"//tensorflow/lite/micro/kernels:all_ops_resolver",
"//tensorflow/lite/micro/kernels:micro_ops",
"//tensorflow/lite/micro/testing:micro_test",
@@ -355,7 +355,7 @@
"//tensorflow/lite:schema_fbs_version",
"//tensorflow/lite/micro:micro_framework",
"//tensorflow/lite/micro/examples/micro_speech/micro_features:micro_model_settings",
- "//tensorflow/lite/micro/examples/micro_speech/micro_features:tiny_conv_micro_features_model_data",
+ "//tensorflow/lite/micro/examples/micro_speech/micro_features:model",
"//tensorflow/lite/micro/kernels:micro_ops",
"//tensorflow/lite/schema:schema_fbs",
],
@@ -376,7 +376,7 @@
"//tensorflow/lite:schema_fbs_version",
"//tensorflow/lite/micro:micro_framework",
"//tensorflow/lite/micro/examples/micro_speech/micro_features:micro_model_settings",
- "//tensorflow/lite/micro/examples/micro_speech/micro_features:tiny_conv_micro_features_model_data",
+ "//tensorflow/lite/micro/examples/micro_speech/micro_features:model",
"//tensorflow/lite/micro/kernels:micro_ops",
"//tensorflow/lite/schema:schema_fbs",
],
diff --git a/tensorflow/lite/micro/examples/micro_speech/Makefile.inc b/tensorflow/lite/micro/examples/micro_speech/Makefile.inc
index 636cd04..18d5fa5 100644
--- a/tensorflow/lite/micro/examples/micro_speech/Makefile.inc
+++ b/tensorflow/lite/micro/examples/micro_speech/Makefile.inc
@@ -28,12 +28,12 @@
MICRO_SPEECH_TEST_SRCS := \
tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.cc
MICRO_SPEECH_TEST_HDRS := \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.h \
@@ -190,7 +190,7 @@
tensorflow/lite/micro/examples/micro_speech/feature_provider.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.cc \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc \
tensorflow/lite/micro/examples/micro_speech/recognize_commands.cc \
tensorflow/lite/micro/examples/micro_speech/command_responder.cc \
$(MICRO_FEATURES_GENERATOR_SRCS)
@@ -200,7 +200,7 @@
tensorflow/lite/micro/examples/micro_speech/feature_provider.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.h \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.h \
tensorflow/lite/micro/examples/micro_speech/recognize_commands.h \
tensorflow/lite/micro/examples/micro_speech/command_responder.h \
tensorflow/lite/micro/examples/micro_speech/main_functions.h \
@@ -215,7 +215,7 @@
tensorflow/lite/micro/examples/micro_speech/feature_provider.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.cc \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.cc \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc \
tensorflow/lite/micro/examples/micro_speech/recognize_commands.cc \
tensorflow/lite/micro/examples/micro_speech/command_responder.cc \
$(MICRO_FEATURES_GENERATOR_SRCS)
@@ -228,7 +228,7 @@
tensorflow/lite/micro/examples/micro_speech/micro_features/micro_features_generator.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.h \
tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.h \
-tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h \
+tensorflow/lite/micro/examples/micro_speech/micro_features/model.h \
tensorflow/lite/micro/examples/micro_speech/recognize_commands.h \
tensorflow/lite/micro/examples/micro_speech/command_responder.h \
tensorflow/lite/micro/examples/micro_speech/main_functions.h \
diff --git a/tensorflow/lite/micro/examples/micro_speech/README.md b/tensorflow/lite/micro/examples/micro_speech/README.md
index 4b75041..7ccaa80 100644
--- a/tensorflow/lite/micro/examples/micro_speech/README.md
+++ b/tensorflow/lite/micro/examples/micro_speech/README.md
@@ -1,21 +1,21 @@
-# Micro speech example
+# Micro Speech Example
-This example shows how you can use TensorFlow Lite to run a 20 kilobyte neural
-network model to recognize keywords in speech. It's designed to run on systems
-with very small amounts of memory such as microcontrollers and DSPs.
+This example shows how to run a 20 kB model that can recognize 2 keywords,
+"yes" and "no", from speech data.
-The example application listens to its surroundings with a microphone and
-indicates when it has detected a word by lighting an LED or displaying data on a
+The application listens to its surroundings with a microphone and indicates
+when it has detected a word by lighting an LED or displaying data on a
screen, depending on the capabilities of the device.
-The code has a small footprint (for example around 22 kilobytes on a Cortex
+
+
+The code has a small footprint (for example, around 22 kilobytes on a Cortex
M3) and only uses about 10 kilobytes of RAM for working memory, so it's able to
run on systems like an STM32F103 with only 20 kilobytes of total SRAM and 64
kilobytes of Flash.
## Table of contents
-- [Getting started](#getting-started)
- [Deploy to Arduino](#deploy-to-arduino)
- [Deploy to ESP32](#deploy-to-esp32)
- [Deploy to SparkFun Edge](#deploy-to-sparkfun-edge)
@@ -23,7 +23,6 @@
- [Deploy to NXP FRDM K66F](#deploy-to-nxp-frdm-k66f)
- [Run on macOS](#run-on-macos)
- [Run the tests on a development machine](#run-the-tests-on-a-development-machine)
-- [Calculating the input to the neural network](#calculating-the-input-to-the-neural-network)
- [Train your own model](#train-your-own-model)
## Deploy to Arduino
@@ -542,160 +541,8 @@
a model that's been compiled into the program, and then invokes the interpreter
with the model and sample inputs.
-## Calculating the input to the neural network
-
-The TensorFlow Lite model doesn't take in raw audio sample data. Instead it
-works with spectrograms, which are two dimensional arrays that are made up of
-slices of frequency information, each taken from a different time window. This
-test uses spectrograms that have been pre-calculated from one-second WAV files
-in the test data set. In a complete application these spectrograms would be
-calculated at runtime from microphone inputs, but the code for doing that is not
-yet included in this sample code.
-
-The recipe for creating the spectrogram data is that each frequency slice is
-created by running an FFT across a 30ms section of the audio sample data. The
-input samples are treated as being between -1 and +1 as real values (encoded as
--32,768 and 32,767 in 16-bit signed integer samples).
-
-This results in an FFT with 256 entries. Every sequence of six entries is
-averaged together, giving a total of 43 frequency buckets in the final slice.
-The results are stored as unsigned eight-bit values, where 0 represents a real
-number of zero, and 255 represents 127.5 as a real number.
-
-Each adjacent frequency entry is stored in ascending memory order (frequency
-bucket 0 at data[0], bucket 1 at data [1], etc). The window for the frequency
-analysis is then moved forward by 20ms, and the process repeated, storing the
-results in the next memory row (for example bucket 0 in this moved window would
-be in data[43 + 0], etc). This process happens 49 times in total, producing a
-single channel image that is 43 pixels wide, and 49 rows high.
-
-Here's an illustration of the process:
-
-
-
-The test data files have been generated by running the following commands. See
-the training instructions below to learn how to set up the environment to run
-them.
-
-```
-python tensorflow/tensorflow/examples/speech_commands/wav_to_features.py \
---input_wav=/tmp/speech_dataset/yes/f2e59fea_nohash_1.wav \
---output_c_file=/tmp/yes_features_data.cc \
---window_stride=20 --preprocess=average --quantize=1
-
-python tensorflow/tensorflow/examples/speech_commands/wav_to_features.py \
---input_wav=/tmp/speech_dataset/no/f9643d42_nohash_4.wav \
---output_c_file=/tmp/no_features_data.cc \
---window_stride=20 --preprocess=average --quantize=1
-```
-
## Train your own model
-The neural network model used in this example was built using the
-[TensorFlow speech commands tutorial](https://www.tensorflow.org/tutorials/sequences/audio_recognition).
-You can retrain it to recognize any combination of words from this list:
-
-```
-yes
-no
-up
-down
-left
-right
-on
-off
-stop
-go
-```
-
-### Use Google Colaboratory
-
-The easiest way to train your own speech model is by running [`train_speech_model.ipynb`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb)
-in Google Colaboratory. This avoids the need to install dependencies, and allows
-the use of GPUs for training. Total training time will be 1.5-2hrs.
-
-We strongly recommend trying this approach first.
-
-### Use your local machine
-
-You can use the following commands to train the model on your own machine. It
-may be easiest to run these commands in a
-[TensorFlow Docker container](https://www.tensorflow.org/install/docker).
-
-You must currently use the TensorFlow Nightly `pip` package. This version is
-confirmed to work:
-
-```
-tf-nightly-gpu==1.15.0.dev20190729
-```
-
-To begin training, run the following:
-
-```
-python tensorflow/tensorflow/examples/speech_commands/train.py \
---model_architecture=tiny_conv --window_stride=20 --preprocess=micro \
---wanted_words="yes,no" --silence_percentage=25 --unknown_percentage=25 \
---quantize=1 --verbosity=INFO --how_many_training_steps="15000,3000" \
---learning_rate="0.001,0.0001" --summaries_dir=/tmp/retrain_logs \
---data_dir=/tmp/speech_dataset --train_dir=/tmp/speech_commands_train
-```
-
-The training process is likely to take a couple of hours. Once it
-has completed, the next step is to freeze the variables:
-
-```
-python tensorflow/tensorflow/examples/speech_commands/freeze.py \
---model_architecture=tiny_conv --window_stride=20 --preprocess=micro \
---wanted_words="yes,no" --quantize=1 --output_file=/tmp/tiny_conv.pb \
---start_checkpoint=/tmp/speech_commands_train/tiny_conv.ckpt-18000
-```
-
-The next step is to create a TensorFlow Lite file from the frozen graph:
-
-```
-toco \
---graph_def_file=/tmp/tiny_conv.pb --output_file=/tmp/tiny_conv.tflite \
---input_shapes=1,49,40,1 --input_arrays=Reshape_2 --output_arrays='labels_softmax' \
---inference_type=QUANTIZED_UINT8 --mean_values=0 --std_dev_values=9.8077
-```
-
-Finally, convert the file into a C source file that can be compiled into an
-embedded system:
-
-```
-xxd -i /tmp/tiny_conv.tflite > /tmp/tiny_conv_micro_features_model_data.cc
-```
-
-### Use Google Cloud
-
-If want to train your model in Google Cloud you can do so by using
-pre-configured Deep Learning images.
-
-First create the VM:
-
-```
-export IMAGE_FAMILY="tf-latest-cpu"
-export ZONE="us-west1-b" # Or any other required region
-export INSTANCE_NAME="model-trainer"
-export INSTANCE_TYPE="n1-standard-8" # or any other instance type
-gcloud compute instances create $INSTANCE_NAME \
- --zone=$ZONE \
- --image-family=$IMAGE_FAMILY \
- --image-project=deeplearning-platform-release \
- --machine-type=$INSTANCE_TYPE \
- --boot-disk-size=120GB \
- --min-cpu-platform=Intel\ Skylake
-```
-
-As soon as instance has been created you can SSH to it(as a jupyter user!):
-
-```
-gcloud compute ssh "jupyter@${INSTANCE_NAME}"
-```
-
-Finally, follow the instructions in the previous section to train the model. Do
-not forget to remove the instance when training is done:
-
-```
-gcloud compute instances delete "${INSTANCE_NAME}" --zone="${ZONE}"
-```
+So far you have used an existing trained model to run inference on
+microcontrollers. If you wish to train your own model, follow the instructions
+in [train/README.md](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech/train/README.md).
diff --git a/tensorflow/lite/micro/examples/micro_speech/apollo3/pushbutton_test.cc b/tensorflow/lite/micro/examples/micro_speech/apollo3/pushbutton_test.cc
index 34b2b3b..1126d7d 100644
--- a/tensorflow/lite/micro/examples/micro_speech/apollo3/pushbutton_test.cc
+++ b/tensorflow/lite/micro/examples/micro_speech/apollo3/pushbutton_test.cc
@@ -17,8 +17,8 @@
* micro_speech_test.cc */
#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/examples/micro_speech/simple_features/model.h"
#include "tensorflow/lite/micro/examples/micro_speech/simple_features/simple_features_generator.h"
-#include "tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.h"
#include "tensorflow/lite/micro/kernels/all_ops_resolver.h"
#include "tensorflow/lite/micro/micro_error_reporter.h"
#include "tensorflow/lite/micro/micro_interpreter.h"
@@ -60,8 +60,7 @@
// Map the model into a usable data structure. This doesn't involve any
// copying or parsing, it's a very lightweight operation.
- const tflite::Model* model =
- ::tflite::GetModel(g_tiny_conv_simple_features_model_data);
+ const tflite::Model* model = ::tflite::GetModel(g_model);
if (model->version() != TFLITE_SCHEMA_VERSION) {
TF_LITE_REPORT_ERROR(error_reporter,
"Model provided is schema version %d not equal "
diff --git a/tensorflow/lite/micro/examples/micro_speech/images/animation_on_arduino.gif b/tensorflow/lite/micro/examples/micro_speech/images/animation_on_arduino.gif
new file mode 100644
index 0000000..66ab9c1
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/images/animation_on_arduino.gif
Binary files differ
diff --git a/tensorflow/lite/micro/examples/micro_speech/images/model_architecture.png b/tensorflow/lite/micro/examples/micro_speech/images/model_architecture.png
new file mode 100644
index 0000000..b1c8c02
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/images/model_architecture.png
Binary files differ
diff --git a/tensorflow/lite/micro/examples/micro_speech/main_functions.cc b/tensorflow/lite/micro/examples/micro_speech/main_functions.cc
index a684940..23c63a3 100644
--- a/tensorflow/lite/micro/examples/micro_speech/main_functions.cc
+++ b/tensorflow/lite/micro/examples/micro_speech/main_functions.cc
@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -19,7 +19,7 @@
#include "tensorflow/lite/micro/examples/micro_speech/command_responder.h"
#include "tensorflow/lite/micro/examples/micro_speech/feature_provider.h"
#include "tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h"
-#include "tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h"
+#include "tensorflow/lite/micro/examples/micro_speech/micro_features/model.h"
#include "tensorflow/lite/micro/examples/micro_speech/recognize_commands.h"
#include "tensorflow/lite/micro/kernels/micro_ops.h"
#include "tensorflow/lite/micro/micro_error_reporter.h"
@@ -57,7 +57,7 @@
// Map the model into a usable data structure. This doesn't involve any
// copying or parsing, it's a very lightweight operation.
- model = tflite::GetModel(g_tiny_conv_micro_features_model_data);
+ model = tflite::GetModel(g_model);
if (model->version() != TFLITE_SCHEMA_VERSION) {
TF_LITE_REPORT_ERROR(error_reporter,
"Model provided is schema version %d not equal "
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/BUILD b/tensorflow/lite/micro/examples/micro_speech/micro_features/BUILD
index da9f500..7101049 100644
--- a/tensorflow/lite/micro/examples/micro_speech/micro_features/BUILD
+++ b/tensorflow/lite/micro/examples/micro_speech/micro_features/BUILD
@@ -21,12 +21,12 @@
)
cc_library(
- name = "tiny_conv_micro_features_model_data",
+ name = "model",
srcs = [
- "tiny_conv_micro_features_model_data.cc",
+ "model.cc",
],
hdrs = [
- "tiny_conv_micro_features_model_data.h",
+ "model.h",
],
)
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h b/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h
index 270c5f3..e542213 100644
--- a/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h
+++ b/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h
@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -25,17 +25,19 @@
constexpr int kMaxAudioSampleSize = 512;
constexpr int kAudioSampleFrequency = 16000;
-// All of these values are derived from the values used during model training,
-// if you change your model you'll need to update these constants.
+// The following values are derived from values used during model training.
+// If you change the way you preprocess the input, update all these constants.
constexpr int kFeatureSliceSize = 40;
constexpr int kFeatureSliceCount = 49;
constexpr int kFeatureElementCount = (kFeatureSliceSize * kFeatureSliceCount);
constexpr int kFeatureSliceStrideMs = 20;
constexpr int kFeatureSliceDurationMs = 30;
-constexpr int kCategoryCount = 4;
+// Variables for the model's output categories.
constexpr int kSilenceIndex = 0;
constexpr int kUnknownIndex = 1;
+// If you modify the output categories, you need to update the following values.
+constexpr int kCategoryCount = 4;
extern const char* kCategoryLabels[kCategoryCount];
#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_MICRO_MODEL_SETTINGS_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc b/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc
new file mode 100644
index 0000000..45198c7
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc
@@ -0,0 +1,1560 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a standard TensorFlow Lite FlatBuffer model file that has been
+// converted into a C data array, so it can be easily compiled into a binary
+// for devices that don't have a file system. It was created using the command:
+// xxd -i model.tflite > model.cc
+
+#include "tensorflow/lite/micro/examples/micro_speech/micro_features/model.h"
+
+// We need to keep the data array aligned on some architectures.
+#ifdef __has_attribute
+#define HAVE_ATTRIBUTE(x) __has_attribute(x)
+#else
+#define HAVE_ATTRIBUTE(x) 0
+#endif
+#if HAVE_ATTRIBUTE(aligned) || (defined(__GNUC__) && !defined(__clang__))
+#define DATA_ALIGN_ATTRIBUTE __attribute__((aligned(4)))
+#else
+#define DATA_ALIGN_ATTRIBUTE
+#endif
+
+const unsigned char g_model[] DATA_ALIGN_ATTRIBUTE = {
+ 0x1c, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x12, 0x00,
+ 0x1c, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00,
+ 0x00, 0x00, 0x18, 0x00, 0x12, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+ 0x1c, 0x47, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
+ 0x2c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0xc0, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x94, 0x00, 0x00, 0x00,
+ 0x0f, 0x00, 0x00, 0x00, 0x54, 0x4f, 0x43, 0x4f, 0x20, 0x43, 0x6f, 0x6e,
+ 0x76, 0x65, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x00, 0x0a, 0x00, 0x00, 0x00,
+ 0x60, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00,
+ 0x3c, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00,
+ 0x20, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x0e, 0xba, 0xff, 0xff, 0x38, 0x00, 0x00, 0x00,
+ 0xbc, 0xb9, 0xff, 0xff, 0xc0, 0xb9, 0xff, 0xff, 0x1e, 0xba, 0xff, 0xff,
+ 0xe0, 0x01, 0x00, 0x00, 0xcc, 0xb9, 0xff, 0xff, 0xd0, 0xb9, 0xff, 0xff,
+ 0x2e, 0xba, 0xff, 0xff, 0x60, 0x03, 0x00, 0x00, 0x36, 0xba, 0xff, 0xff,
+ 0x7c, 0x06, 0x00, 0x00, 0x3e, 0xba, 0xff, 0xff, 0x68, 0x45, 0x00, 0x00,
+ 0xec, 0xb9, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x31, 0x2e, 0x35, 0x2e,
+ 0x30, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x04, 0x00, 0x08, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00,
+ 0x13, 0x00, 0x00, 0x00, 0x6d, 0x69, 0x6e, 0x5f, 0x72, 0x75, 0x6e, 0x74,
+ 0x69, 0x6d, 0x65, 0x5f, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x00,
+ 0x10, 0xfa, 0xff, 0xff, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x2c, 0x45, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x9c, 0x44, 0x00, 0x00,
+ 0x8c, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0xdc, 0x01, 0x00, 0x00,
+ 0x68, 0x01, 0x00, 0x00, 0x3c, 0x02, 0x00, 0x00, 0x50, 0x05, 0x00, 0x00,
+ 0x8e, 0xbb, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00,
+ 0x28, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00,
+ 0x52, 0x65, 0x73, 0x68, 0x61, 0x70, 0x65, 0x5f, 0x32, 0x00, 0x00, 0x00,
+ 0x94, 0xfa, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00,
+ 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0xc6, 0xd0, 0xd0, 0x3d, 0x01, 0x00, 0x00, 0x00, 0xf5, 0xff, 0xcf, 0x41,
+ 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xbc, 0xff, 0xff,
+ 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00,
+ 0x1c, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x52, 0x65, 0x6c, 0x75,
+ 0x00, 0x00, 0x00, 0x00, 0x04, 0xfb, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00,
+ 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x09, 0xf5, 0x83, 0x3d, 0x01, 0x00, 0x00, 0x00,
+ 0x14, 0x71, 0x83, 0x41, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x72, 0xbc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02, 0x10, 0x00, 0x00, 0x00,
+ 0x06, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00,
+ 0x43, 0x6f, 0x6e, 0x76, 0x32, 0x44, 0x5f, 0x62, 0x69, 0x61, 0x73, 0x00,
+ 0x64, 0xbc, 0xff, 0xff, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2d, 0x95, 0x98, 0x38,
+ 0x20, 0x00, 0x00, 0x00, 0x27, 0xff, 0xff, 0xff, 0x97, 0xff, 0xff, 0xff,
+ 0x58, 0x00, 0x00, 0x00, 0x66, 0xff, 0xff, 0xff, 0x13, 0xff, 0xff, 0xff,
+ 0x72, 0xfe, 0xff, 0xff, 0x5d, 0xfb, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00,
+ 0xea, 0xbc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,
+ 0x05, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x0e, 0x00, 0x00, 0x00, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x73, 0x5f, 0x73,
+ 0x6f, 0x66, 0x74, 0x6d, 0x61, 0x78, 0x00, 0x00, 0xec, 0xfb, 0xff, 0xff,
+ 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3b,
+ 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x7f, 0x3f, 0x01, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x5a, 0xbd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,
+ 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x1c, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x61, 0x64, 0x64, 0x5f,
+ 0x31, 0x00, 0x00, 0x00, 0x54, 0xfc, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00,
+ 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x8f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x9c, 0xd2, 0xb5, 0x3d, 0x01, 0x00, 0x00, 0x00,
+ 0x48, 0x18, 0x1f, 0x41, 0x01, 0x00, 0x00, 0x00, 0x4a, 0x21, 0x4b, 0xc1,
+ 0xc2, 0xbd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,
+ 0x03, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00,
+ 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,
+ 0x77, 0x65, 0x69, 0x67, 0x68, 0x74, 0x73, 0x5f, 0x71, 0x75, 0x61, 0x6e,
+ 0x74, 0x2f, 0x46, 0x61, 0x6b, 0x65, 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57,
+ 0x69, 0x74, 0x68, 0x4d, 0x69, 0x6e, 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72,
+ 0x73, 0x00, 0x00, 0x00, 0xe4, 0xfc, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00,
+ 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x8a, 0x0f, 0x3b, 0x3a,
+ 0x01, 0x00, 0x00, 0x00, 0xfc, 0x0b, 0xb4, 0x3d, 0x01, 0x00, 0x00, 0x00,
+ 0xd9, 0x26, 0xbf, 0xbd, 0x80, 0x02, 0x00, 0x00, 0x60, 0x38, 0xab, 0xcb,
+ 0xfa, 0x7e, 0xa2, 0x55, 0x6e, 0x87, 0xa5, 0x9b, 0xb4, 0x66, 0x5c, 0x6f,
+ 0xae, 0xdb, 0xcd, 0xb6, 0xc2, 0x60, 0xa9, 0x7d, 0xd4, 0xac, 0xa6, 0x90,
+ 0x87, 0x6b, 0x50, 0x95, 0xde, 0xcd, 0xaa, 0xa1, 0x9c, 0x65, 0xb5, 0x6d,
+ 0xb0, 0xa5, 0xa5, 0x7f, 0x73, 0x95, 0x63, 0x81, 0x7a, 0xc6, 0xaf, 0x82,
+ 0x69, 0x89, 0xc3, 0x3c, 0x47, 0x73, 0x89, 0x4f, 0x33, 0xbc, 0x85, 0x5d,
+ 0x69, 0x11, 0x5b, 0xb9, 0xf1, 0x95, 0x8f, 0x5c, 0x7c, 0x59, 0x6c, 0xa0,
+ 0xa5, 0x7c, 0x5a, 0x7c, 0xb5, 0xa9, 0x7e, 0xa1, 0xb8, 0x65, 0xb3, 0x86,
+ 0xc1, 0x9f, 0x5c, 0x86, 0x7f, 0x74, 0x52, 0xa8, 0xc9, 0xc5, 0x71, 0x96,
+ 0x7a, 0x65, 0xc7, 0x69, 0x94, 0xa7, 0x65, 0x68, 0x69, 0x8d, 0x6d, 0x9e,
+ 0x59, 0xd4, 0x75, 0x7a, 0x4f, 0x70, 0xca, 0x48, 0x25, 0x8a, 0x69, 0x4d,
+ 0x2a, 0xa6, 0x76, 0x69, 0x6a, 0x02, 0x3b, 0xa2, 0xea, 0xc2, 0x73, 0x6b,
+ 0x86, 0x4d, 0x3a, 0xa2, 0xa2, 0x88, 0x4e, 0x6c, 0xb3, 0x83, 0x39, 0x93,
+ 0xa6, 0x85, 0xb8, 0x7a, 0xa8, 0x7d, 0x2e, 0x7b, 0x7f, 0x69, 0x56, 0xb5,
+ 0xbb, 0xae, 0x23, 0x78, 0x67, 0x5c, 0xd2, 0x82, 0x7d, 0x96, 0x46, 0x74,
+ 0x70, 0x72, 0x6a, 0x90, 0x43, 0xce, 0x44, 0x75, 0x4a, 0x58, 0xc7, 0x5c,
+ 0x34, 0x84, 0x46, 0x4b, 0x41, 0x6c, 0x62, 0x83, 0x7e, 0x01, 0x9b, 0x9b,
+ 0xeb, 0xf7, 0x58, 0x6f, 0x8a, 0x43, 0xb3, 0x9f, 0x9c, 0x9e, 0x55, 0xa8,
+ 0xaa, 0x84, 0x8f, 0x8f, 0xb0, 0x9e, 0xc8, 0x81, 0xb6, 0x80, 0xa0, 0x81,
+ 0x86, 0x73, 0x5d, 0xdc, 0xb9, 0xae, 0xa2, 0x6c, 0x46, 0x67, 0xfa, 0x79,
+ 0x89, 0xaf, 0xa0, 0x74, 0x76, 0x85, 0x72, 0xb1, 0x2a, 0xbb, 0xa0, 0x6d,
+ 0x4f, 0x50, 0xc9, 0x5d, 0x2f, 0xaa, 0x9c, 0x63, 0x3f, 0x59, 0x63, 0x90,
+ 0x73, 0x1e, 0xb3, 0x94, 0xcd, 0xff, 0x3c, 0x63, 0x9b, 0x59, 0xc5, 0xa2,
+ 0x9f, 0x9a, 0x53, 0xab, 0xb0, 0x74, 0xb2, 0x6f, 0x8a, 0xa7, 0xd5, 0x8d,
+ 0xb8, 0x7e, 0x9e, 0x78, 0x84, 0x61, 0x66, 0xe7, 0xa7, 0x9f, 0xb7, 0x45,
+ 0x24, 0x61, 0xfd, 0x69, 0x87, 0xb8, 0xb2, 0x7a, 0x7c, 0x58, 0x64, 0xa3,
+ 0x07, 0xa9, 0xaf, 0x69, 0x49, 0x2f, 0xc2, 0x46, 0x3b, 0xaf, 0x9a, 0x70,
+ 0x6b, 0x25, 0x5f, 0x9d, 0x82, 0x33, 0xa1, 0x54, 0xae, 0xff, 0x31, 0x5d,
+ 0xaf, 0x51, 0xb2, 0x82, 0x9c, 0xa9, 0x5b, 0x8c, 0xab, 0x75, 0xb3, 0x32,
+ 0x42, 0xbd, 0xcd, 0x77, 0xb6, 0x67, 0x9a, 0x5f, 0x6c, 0x71, 0x6e, 0xc2,
+ 0xac, 0x97, 0x9f, 0x4b, 0x21, 0x6a, 0xfc, 0x77, 0x83, 0xa1, 0xa3, 0x6a,
+ 0x7a, 0x6d, 0x5e, 0x87, 0x02, 0xa6, 0x8f, 0x7f, 0x5c, 0x2e, 0xc1, 0x51,
+ 0x4a, 0xa7, 0x96, 0x79, 0x83, 0x2e, 0x5a, 0x84, 0x82, 0x5c, 0x61, 0x3a,
+ 0x4a, 0xff, 0x2a, 0x51, 0xa4, 0x6b, 0x82, 0x5e, 0x67, 0xb3, 0x71, 0x80,
+ 0xad, 0x62, 0x59, 0x40, 0x26, 0xd7, 0xcf, 0x68, 0xab, 0x7c, 0x6a, 0x69,
+ 0x5b, 0x7c, 0x84, 0xbc, 0x95, 0x68, 0x77, 0x63, 0x3f, 0x85, 0xed, 0x7b,
+ 0x71, 0xa0, 0x76, 0x90, 0x8c, 0x6c, 0x61, 0x81, 0x16, 0x74, 0x72, 0x94,
+ 0x74, 0x37, 0xb5, 0x3d, 0x55, 0x96, 0x86, 0xad, 0x87, 0x39, 0x59, 0x88,
+ 0x5b, 0x65, 0x60, 0x33, 0x33, 0xe6, 0x2b, 0x4a, 0xb6, 0x82, 0x50, 0x56,
+ 0x51, 0x97, 0x71, 0x83, 0xa6, 0x60, 0x57, 0x51, 0x58, 0xe4, 0xd0, 0x87,
+ 0xa1, 0x78, 0x4c, 0x67, 0x72, 0x74, 0x86, 0xc6, 0x60, 0x47, 0x50, 0x96,
+ 0x67, 0x96, 0xdd, 0x7d, 0x63, 0x85, 0x5e, 0x98, 0xa2, 0x64, 0x5f, 0x8a,
+ 0x3b, 0x40, 0x54, 0xcb, 0xa0, 0x61, 0xa7, 0x44, 0x5f, 0x6d, 0x57, 0xb3,
+ 0xb9, 0x2e, 0x61, 0x8e, 0x54, 0x78, 0x85, 0x58, 0x43, 0xb0, 0x27, 0x5d,
+ 0x8a, 0x7c, 0x8a, 0x58, 0x40, 0x83, 0x82, 0x9b, 0x6c, 0x60, 0x6b, 0x72,
+ 0x7f, 0xde, 0xc9, 0x7d, 0x6f, 0x5f, 0x90, 0x7e, 0x7e, 0x7e, 0x8b, 0xe5,
+ 0x51, 0x37, 0x7a, 0xa9, 0xa2, 0xc5, 0xd3, 0x81, 0x32, 0x4b, 0x80, 0xa9,
+ 0xc5, 0x76, 0x56, 0x99, 0x33, 0x19, 0x72, 0xe6, 0xdb, 0x90, 0xa8, 0x50,
+ 0x65, 0x44, 0x77, 0xdb, 0xc7, 0x48, 0x65, 0x8d, 0x3d, 0x7f, 0xa2, 0x7c,
+ 0x53, 0x55, 0x26, 0x49, 0x5d, 0x7d, 0xa2, 0x6d, 0x3b, 0x5b, 0x87, 0x64,
+ 0x3a, 0x5b, 0x8d, 0x93, 0x7a, 0xb4, 0xca, 0x6d, 0x16, 0x5a, 0x99, 0x82,
+ 0x8d, 0x6a, 0x92, 0xa0, 0x39, 0x2c, 0x95, 0xc8, 0xb8, 0xf5, 0xc8, 0x66,
+ 0x2a, 0x45, 0x84, 0x9c, 0xc7, 0x8e, 0x61, 0x7b, 0x43, 0x28, 0x86, 0xff,
+ 0xd2, 0xc8, 0x9c, 0x46, 0x65, 0x33, 0x82, 0xd8, 0xcb, 0x73, 0x63, 0x80,
+ 0xda, 0xc0, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xa0, 0x0f, 0x00, 0x00,
+ 0x31, 0x00, 0x00, 0x00, 0x77, 0x65, 0x69, 0x67, 0x68, 0x74, 0x73, 0x5f,
+ 0x71, 0x75, 0x61, 0x6e, 0x74, 0x5f, 0x31, 0x2f, 0x46, 0x61, 0x6b, 0x65,
+ 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57, 0x69, 0x74, 0x68, 0x4d, 0x69, 0x6e,
+ 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72, 0x73, 0x2f, 0x74, 0x72, 0x61, 0x6e,
+ 0x73, 0x70, 0x6f, 0x73, 0x65, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x14, 0x00,
+ 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x00, 0x00,
+ 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x7e, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x87, 0xff, 0xdb, 0x39,
+ 0x01, 0x00, 0x00, 0x00, 0xd8, 0xb2, 0x5d, 0x3d, 0x01, 0x00, 0x00, 0x00,
+ 0x37, 0xdc, 0x56, 0xbd, 0x80, 0x3e, 0x00, 0x00, 0x67, 0x6d, 0x74, 0x77,
+ 0x35, 0x66, 0x87, 0x95, 0x8e, 0x82, 0x5e, 0x70, 0x6e, 0xa7, 0x60, 0x64,
+ 0x86, 0x5e, 0x93, 0x7a, 0x76, 0x74, 0x71, 0x8c, 0x61, 0x71, 0x60, 0x8b,
+ 0x83, 0x48, 0x8b, 0x5f, 0x95, 0x99, 0x5b, 0x59, 0x49, 0x44, 0x79, 0x62,
+ 0x8e, 0x77, 0x71, 0x89, 0x64, 0x46, 0x8f, 0x8e, 0x80, 0x73, 0x71, 0x81,
+ 0x85, 0x4a, 0x73, 0x57, 0x66, 0x58, 0x75, 0x93, 0x99, 0x58, 0x8a, 0x7b,
+ 0x87, 0x81, 0xa1, 0x46, 0x79, 0x6c, 0x83, 0x7a, 0x92, 0x74, 0x6f, 0x6b,
+ 0x79, 0x77, 0x97, 0x8a, 0x95, 0x75, 0xa2, 0x49, 0x80, 0x4e, 0x7f, 0x6d,
+ 0xaa, 0xac, 0x6c, 0x5d, 0x57, 0x82, 0x97, 0x77, 0x6f, 0x75, 0x95, 0x73,
+ 0x7e, 0x51, 0x9f, 0x5b, 0x54, 0x92, 0x60, 0x72, 0x80, 0x6a, 0x92, 0x83,
+ 0x9b, 0x85, 0x7b, 0x4d, 0x55, 0x4d, 0xb2, 0x7d, 0x65, 0x95, 0x76, 0x42,
+ 0x61, 0x49, 0xa2, 0x73, 0x9f, 0x7d, 0x7c, 0x54, 0x51, 0x76, 0xa1, 0x7f,
+ 0x86, 0x69, 0x98, 0x59, 0x6d, 0x84, 0x9f, 0x7b, 0x86, 0x79, 0x88, 0x55,
+ 0x9c, 0x72, 0x95, 0x8a, 0x91, 0x7a, 0x77, 0x95, 0x7b, 0x87, 0x87, 0x85,
+ 0x95, 0x72, 0x77, 0x59, 0x7c, 0x80, 0x90, 0x8f, 0x8a, 0x62, 0x76, 0x9f,
+ 0x64, 0x84, 0x71, 0x7e, 0x7c, 0x66, 0x8e, 0x94, 0x6e, 0xaa, 0x77, 0x5c,
+ 0x6b, 0x63, 0x68, 0x82, 0x89, 0x46, 0x61, 0x74, 0x8e, 0x85, 0x6b, 0x57,
+ 0x74, 0x50, 0x87, 0x66, 0x87, 0x98, 0x59, 0x7d, 0xa2, 0x59, 0x75, 0x64,
+ 0x72, 0x8c, 0x6a, 0x92, 0x8c, 0x56, 0x88, 0x7a, 0x6e, 0x77, 0x9c, 0x82,
+ 0x7e, 0x5a, 0x91, 0x80, 0x9c, 0x9e, 0x60, 0x8b, 0x6d, 0x76, 0x8d, 0x68,
+ 0x6c, 0x70, 0x6f, 0x8b, 0x61, 0x6e, 0x86, 0x78, 0x81, 0x81, 0x77, 0x79,
+ 0x76, 0x69, 0x7d, 0x7b, 0x96, 0x8b, 0x95, 0x91, 0xa2, 0x7b, 0x86, 0x8d,
+ 0x8b, 0x89, 0x86, 0x5a, 0x5c, 0x4d, 0x96, 0x80, 0x81, 0x55, 0x80, 0x80,
+ 0x7a, 0x76, 0x99, 0x98, 0x61, 0x95, 0x5a, 0x78, 0x5a, 0x6c, 0x89, 0x81,
+ 0x98, 0x77, 0x62, 0x77, 0x93, 0x4d, 0x9f, 0x77, 0x72, 0x87, 0x95, 0x71,
+ 0x65, 0x72, 0xac, 0x8c, 0xa2, 0x89, 0x90, 0x7b, 0x67, 0x60, 0x8a, 0xb3,
+ 0x72, 0x8f, 0x5c, 0x82, 0x74, 0x76, 0x7c, 0x85, 0x78, 0x6b, 0x97, 0x6d,
+ 0x86, 0x82, 0x76, 0x84, 0x89, 0x89, 0x7f, 0x6a, 0x7a, 0x7f, 0x6c, 0x77,
+ 0x80, 0x35, 0x7d, 0x66, 0x96, 0x7e, 0x88, 0x55, 0x6b, 0x55, 0x7c, 0xa7,
+ 0x7f, 0x9f, 0x64, 0x8b, 0xa0, 0x81, 0x80, 0x97, 0xaf, 0x7a, 0x7d, 0x61,
+ 0x7a, 0x77, 0x6f, 0x8c, 0x5e, 0x69, 0x6b, 0x94, 0x70, 0x6a, 0x66, 0x5d,
+ 0x78, 0x6e, 0x76, 0x64, 0xa0, 0x73, 0x8f, 0xa2, 0x9d, 0x50, 0x8e, 0x52,
+ 0x51, 0x85, 0x78, 0x83, 0x8f, 0x94, 0x83, 0x7c, 0x9c, 0x64, 0x59, 0x7d,
+ 0x66, 0x6a, 0x73, 0x80, 0x6a, 0x9b, 0x92, 0x7e, 0x7a, 0x78, 0x7d, 0xa0,
+ 0x8a, 0x9b, 0x61, 0x9e, 0x6c, 0x64, 0x6c, 0x8e, 0x86, 0x75, 0x8a, 0x95,
+ 0x8e, 0x89, 0x87, 0x8a, 0x5d, 0x8b, 0x82, 0x7c, 0x60, 0x63, 0x85, 0x85,
+ 0x63, 0x96, 0xa3, 0x7f, 0x93, 0x78, 0x8c, 0x86, 0x7b, 0x78, 0x8e, 0x71,
+ 0x72, 0x8b, 0x8a, 0x5e, 0x8d, 0x75, 0x78, 0xa3, 0x84, 0x67, 0xa7, 0x54,
+ 0x6c, 0x80, 0x8e, 0xa8, 0x83, 0x51, 0x6e, 0x9f, 0x8b, 0x86, 0x75, 0x95,
+ 0x7f, 0x7a, 0x80, 0x81, 0x8d, 0x9c, 0x83, 0x8a, 0x7b, 0x8a, 0x74, 0x6f,
+ 0x8d, 0x96, 0x5b, 0x9c, 0x8d, 0x7b, 0x83, 0x79, 0x7f, 0x65, 0x7e, 0x87,
+ 0x7c, 0x5d, 0x71, 0x97, 0x77, 0x44, 0x9a, 0x7f, 0xaa, 0x56, 0x75, 0x5f,
+ 0x7c, 0x51, 0x8c, 0x90, 0x84, 0x9a, 0x49, 0x5d, 0x86, 0x52, 0x94, 0x95,
+ 0x5b, 0x86, 0x66, 0x7d, 0x51, 0x4f, 0x7a, 0x91, 0x6d, 0x6e, 0x72, 0x70,
+ 0x83, 0x4f, 0x9b, 0x9a, 0x8a, 0x77, 0x6a, 0xa1, 0x71, 0x60, 0x61, 0x98,
+ 0x67, 0x4e, 0x7a, 0x8a, 0x53, 0x6b, 0x99, 0xa0, 0x91, 0x46, 0x8a, 0x8b,
+ 0x47, 0x78, 0xa9, 0x7b, 0x71, 0x6c, 0x81, 0x68, 0x53, 0x73, 0xaf, 0x70,
+ 0x62, 0x6d, 0x69, 0x97, 0x70, 0x83, 0x5f, 0x7f, 0x81, 0x87, 0x65, 0x93,
+ 0x67, 0x87, 0x70, 0x82, 0x79, 0x9e, 0x80, 0x77, 0x6c, 0x80, 0x92, 0x81,
+ 0x8d, 0x8c, 0x89, 0x8b, 0x4e, 0x91, 0x77, 0x84, 0x99, 0x8c, 0x71, 0x88,
+ 0x57, 0x7a, 0x9a, 0x8c, 0x82, 0x9b, 0x97, 0x72, 0x69, 0xac, 0x7c, 0x62,
+ 0x85, 0x7d, 0x76, 0x7f, 0x59, 0x85, 0x68, 0x63, 0x94, 0x8b, 0x7b, 0x92,
+ 0x7b, 0x6f, 0x77, 0x98, 0x66, 0x78, 0x74, 0x99, 0x85, 0x8c, 0x94, 0x89,
+ 0x6c, 0x77, 0x89, 0x80, 0x79, 0x8a, 0xa6, 0x95, 0xa9, 0x86, 0x6f, 0x95,
+ 0x90, 0x69, 0x98, 0x85, 0xa0, 0x7f, 0x56, 0xab, 0x6f, 0x5a, 0x94, 0x8b,
+ 0x5a, 0x72, 0x61, 0x83, 0x54, 0x70, 0x8d, 0x8d, 0x9c, 0x5e, 0x36, 0x9b,
+ 0x84, 0x32, 0x6e, 0x84, 0x79, 0x72, 0x64, 0x95, 0x83, 0x58, 0x67, 0x6c,
+ 0x9e, 0x8d, 0x6e, 0x9e, 0x4f, 0x78, 0x71, 0x85, 0x75, 0x60, 0x4d, 0x7d,
+ 0x64, 0x89, 0x8e, 0x89, 0x6e, 0x92, 0x53, 0x7c, 0x86, 0x8f, 0xa9, 0xb0,
+ 0x8e, 0x5e, 0x76, 0x96, 0x65, 0x7c, 0x8a, 0x89, 0x75, 0x8f, 0x65, 0x94,
+ 0x6c, 0x6c, 0x8d, 0x6d, 0x66, 0x6a, 0x62, 0x98, 0x53, 0x8f, 0x67, 0x76,
+ 0x80, 0x89, 0x66, 0x60, 0x55, 0x81, 0x85, 0x61, 0x75, 0x78, 0x80, 0x92,
+ 0x6f, 0x79, 0x66, 0x64, 0x99, 0xa7, 0x88, 0xa1, 0x86, 0x6b, 0x94, 0x88,
+ 0x77, 0x83, 0x8f, 0x61, 0x72, 0x7c, 0x6f, 0x8f, 0x61, 0x56, 0x8a, 0x7b,
+ 0x66, 0x8b, 0x98, 0x9d, 0x82, 0x65, 0x77, 0x98, 0x55, 0x83, 0x7a, 0x8c,
+ 0x74, 0x79, 0x6e, 0x85, 0x82, 0x9a, 0x7d, 0x8d, 0x76, 0x72, 0x64, 0x81,
+ 0x9a, 0x8d, 0x9f, 0x7b, 0x7c, 0x7b, 0x7b, 0x84, 0x90, 0x6b, 0xa4, 0x84,
+ 0x98, 0x6f, 0x81, 0xb8, 0x6f, 0x6c, 0x87, 0x6d, 0x8c, 0x72, 0x53, 0x85,
+ 0x59, 0x4d, 0x9c, 0x94, 0x7d, 0x6f, 0x4f, 0x82, 0x5d, 0x71, 0x6e, 0x78,
+ 0x61, 0x61, 0x34, 0x71, 0x6a, 0x5a, 0x73, 0xa3, 0x89, 0x65, 0x4d, 0x80,
+ 0x5c, 0x51, 0x81, 0x8e, 0x6c, 0x53, 0x4a, 0x95, 0x3b, 0x72, 0xa7, 0x86,
+ 0x7f, 0x75, 0x61, 0xa3, 0x85, 0x6c, 0x99, 0x88, 0x7c, 0x64, 0x7a, 0x8d,
+ 0x81, 0x7b, 0x6a, 0x7b, 0x8f, 0x74, 0x6d, 0xae, 0x42, 0x67, 0x88, 0xa1,
+ 0x90, 0x4d, 0x7c, 0x7b, 0x62, 0x55, 0x9a, 0x80, 0x4d, 0x76, 0x5c, 0x88,
+ 0x60, 0x86, 0x6f, 0x65, 0x67, 0x77, 0x8a, 0x97, 0x99, 0x7c, 0x89, 0x78,
+ 0x92, 0xa7, 0x6a, 0x7f, 0x8e, 0x88, 0x9d, 0xa1, 0x7b, 0xb0, 0x69, 0x8c,
+ 0x7e, 0x51, 0x76, 0x84, 0x7d, 0x91, 0x7a, 0x88, 0x7b, 0x88, 0x92, 0x79,
+ 0x6d, 0x82, 0x6c, 0x8a, 0x99, 0x62, 0x82, 0x9d, 0x99, 0x97, 0x78, 0x6a,
+ 0x6e, 0x83, 0x64, 0x7d, 0x8c, 0x78, 0x7c, 0x7a, 0x7d, 0x7b, 0x77, 0x84,
+ 0x76, 0x57, 0x63, 0x85, 0x97, 0x94, 0x80, 0x92, 0x88, 0x73, 0x91, 0x91,
+ 0x8f, 0x6d, 0x99, 0x86, 0x91, 0x7f, 0x8b, 0x87, 0x98, 0x62, 0x84, 0x70,
+ 0x97, 0x7b, 0x2e, 0x9b, 0x6e, 0x2a, 0xa4, 0x9c, 0x79, 0x88, 0x54, 0x81,
+ 0x4f, 0x41, 0xa0, 0x85, 0xaf, 0x9a, 0x47, 0x5a, 0x7d, 0x62, 0x7a, 0x84,
+ 0x81, 0x6e, 0x41, 0xb4, 0x60, 0x47, 0x8f, 0x98, 0x6c, 0x3c, 0x3b, 0x73,
+ 0x59, 0x55, 0x7c, 0xb0, 0x6e, 0x5f, 0x61, 0x97, 0x73, 0x59, 0x9f, 0x92,
+ 0x89, 0x5c, 0x70, 0x96, 0x5c, 0x7c, 0x7c, 0x64, 0x7e, 0x54, 0x5c, 0x94,
+ 0x56, 0x73, 0x8d, 0x95, 0x59, 0x83, 0x6c, 0x99, 0x6e, 0x5e, 0x7a, 0x99,
+ 0x83, 0x93, 0x88, 0x76, 0x5a, 0x5a, 0xa5, 0x95, 0x5d, 0x63, 0x8f, 0x6e,
+ 0x74, 0x65, 0x85, 0x86, 0x98, 0x83, 0x7b, 0x8a, 0x5c, 0x5e, 0x7f, 0x88,
+ 0x78, 0x68, 0x8f, 0x9f, 0x94, 0x8d, 0x74, 0x7b, 0x6a, 0x91, 0x7a, 0x9a,
+ 0x70, 0x67, 0xb2, 0x92, 0x75, 0x4e, 0x74, 0xa3, 0x68, 0x74, 0x91, 0x80,
+ 0x55, 0x8e, 0x88, 0x73, 0x70, 0x81, 0xa1, 0xb8, 0x96, 0x48, 0x67, 0xb2,
+ 0x76, 0xa1, 0x98, 0xa9, 0x61, 0x6c, 0x5f, 0x98, 0x84, 0x92, 0xa9, 0x83,
+ 0x9e, 0x74, 0x7b, 0xa2, 0x6f, 0x72, 0x95, 0xa3, 0xb9, 0x80, 0x81, 0x7b,
+ 0x65, 0x6b, 0x96, 0x8b, 0xae, 0x79, 0x2b, 0x86, 0x5c, 0x2c, 0x8b, 0xa3,
+ 0x84, 0x74, 0x53, 0x7c, 0x54, 0x4a, 0x65, 0x89, 0xa6, 0x89, 0x47, 0x77,
+ 0x50, 0x6d, 0x8b, 0x94, 0x8a, 0x61, 0x32, 0x7c, 0x6f, 0x47, 0x78, 0xa2,
+ 0x9f, 0x42, 0x42, 0x71, 0x78, 0x76, 0x9e, 0x88, 0x70, 0x70, 0x56, 0x8a,
+ 0x83, 0x95, 0xa7, 0x9d, 0x9d, 0x88, 0x9a, 0x92, 0x48, 0x63, 0xaf, 0x91,
+ 0x6c, 0x75, 0x5d, 0x5e, 0x83, 0x86, 0xaa, 0x6f, 0x79, 0x84, 0x67, 0x79,
+ 0x63, 0x69, 0x8e, 0x81, 0x6a, 0x96, 0x8d, 0x86, 0x7b, 0x9f, 0xaa, 0x8e,
+ 0x63, 0x89, 0x9a, 0x7a, 0x5e, 0x7c, 0x87, 0x83, 0x81, 0x64, 0x7e, 0x59,
+ 0x6d, 0x5c, 0xa4, 0x72, 0x78, 0x85, 0x9b, 0x79, 0x85, 0x7d, 0x9c, 0x7d,
+ 0x9c, 0x5c, 0x66, 0x75, 0x66, 0x72, 0xb4, 0x7c, 0x83, 0x9e, 0x90, 0xae,
+ 0x69, 0x71, 0xb0, 0x84, 0x86, 0x50, 0x66, 0xab, 0x75, 0x96, 0xa8, 0x6c,
+ 0x87, 0x7b, 0x7e, 0x7c, 0x60, 0x55, 0x96, 0xb0, 0x6a, 0x79, 0x42, 0x9c,
+ 0x97, 0xa8, 0xb2, 0x9a, 0xa0, 0x84, 0x68, 0x90, 0x90, 0x98, 0x67, 0x9c,
+ 0xa3, 0x81, 0x71, 0xaa, 0x93, 0x6a, 0x84, 0x8c, 0x77, 0x79, 0x4d, 0x82,
+ 0x45, 0x1e, 0x7b, 0x94, 0x86, 0x86, 0x26, 0x82, 0x41, 0x6f, 0x8b, 0x86,
+ 0xa4, 0x80, 0x38, 0x71, 0x5e, 0x5b, 0x9a, 0x73, 0x86, 0x60, 0x5a, 0x9d,
+ 0x7b, 0x53, 0x89, 0xa0, 0x99, 0x76, 0x57, 0x81, 0x76, 0x5a, 0x9e, 0x85,
+ 0x5a, 0x7b, 0x56, 0x74, 0x71, 0x6a, 0x9c, 0x68, 0x7e, 0x76, 0x7d, 0x7f,
+ 0x52, 0x71, 0x85, 0xa2, 0x96, 0x63, 0x73, 0x7c, 0x7a, 0x97, 0x9f, 0x7c,
+ 0x77, 0x77, 0x59, 0x6b, 0x62, 0x77, 0xbc, 0x6b, 0x7c, 0x79, 0x75, 0x90,
+ 0x67, 0x82, 0x92, 0x9c, 0x81, 0x92, 0x84, 0x7a, 0x72, 0x5b, 0x86, 0x82,
+ 0x87, 0x73, 0x87, 0x7c, 0x57, 0x76, 0xa6, 0x7d, 0x7d, 0x94, 0x6a, 0x67,
+ 0x76, 0x89, 0x9a, 0x6d, 0x7d, 0xa4, 0x6d, 0x7e, 0x74, 0x7e, 0x8f, 0xad,
+ 0x99, 0x55, 0x5c, 0x82, 0x75, 0x9e, 0xae, 0x76, 0x6b, 0x93, 0x5d, 0x92,
+ 0x6e, 0x54, 0x88, 0x8f, 0x6a, 0x72, 0x64, 0x93, 0x6e, 0x63, 0x8c, 0xa7,
+ 0xa6, 0x7a, 0x57, 0x9f, 0x94, 0x91, 0xbd, 0xa4, 0x92, 0x7a, 0x68, 0x9d,
+ 0x7d, 0x6b, 0x6b, 0xbc, 0xad, 0x7a, 0x73, 0x92, 0x7b, 0x6d, 0x91, 0x6a,
+ 0x66, 0x8d, 0x34, 0x9b, 0x75, 0x3b, 0x93, 0x78, 0x88, 0x58, 0x1a, 0x7f,
+ 0x52, 0x61, 0xa3, 0xb1, 0x9c, 0x60, 0x1d, 0x90, 0x7b, 0x37, 0x9f, 0x84,
+ 0xa3, 0x6c, 0x2e, 0xac, 0x73, 0x62, 0x92, 0x9a, 0x94, 0x6b, 0x5c, 0x82,
+ 0x5f, 0x4c, 0x9a, 0x8c, 0x76, 0x69, 0x77, 0x5f, 0x5d, 0x91, 0x80, 0x9a,
+ 0x60, 0x4c, 0x7b, 0x57, 0x67, 0x6b, 0x92, 0x93, 0x64, 0x91, 0x55, 0x75,
+ 0x41, 0x82, 0x78, 0x68, 0xa2, 0x55, 0x6a, 0x69, 0x59, 0x70, 0x8a, 0x7b,
+ 0x70, 0x6e, 0x63, 0x83, 0x7f, 0xa4, 0x80, 0x85, 0x86, 0x93, 0x7e, 0x6f,
+ 0x7b, 0x94, 0xa4, 0xa7, 0x97, 0x7a, 0x87, 0x64, 0x4a, 0x97, 0x94, 0x6a,
+ 0x96, 0x73, 0x5e, 0x79, 0x6a, 0x99, 0x86, 0xa0, 0x93, 0xac, 0x79, 0x76,
+ 0x7f, 0x7b, 0xa7, 0x75, 0x8a, 0x71, 0x53, 0x87, 0x93, 0x7f, 0x9e, 0x7b,
+ 0x81, 0x70, 0x68, 0x8b, 0x8c, 0x9c, 0xaf, 0xa7, 0x6a, 0x9b, 0x49, 0x6d,
+ 0x67, 0x80, 0x8b, 0x86, 0x9f, 0x80, 0x74, 0x7a, 0x96, 0x74, 0xc8, 0x9d,
+ 0xa4, 0x74, 0x71, 0x6c, 0x75, 0x6a, 0x9a, 0x95, 0x97, 0x8c, 0x6e, 0x8a,
+ 0x85, 0x62, 0x5f, 0x7e, 0x9e, 0x6b, 0x48, 0x93, 0x44, 0x37, 0x83, 0xa2,
+ 0x97, 0x72, 0x25, 0x79, 0x32, 0x39, 0x68, 0x8f, 0x93, 0x61, 0x2b, 0x96,
+ 0x94, 0x43, 0x82, 0x6e, 0x8f, 0x6d, 0x53, 0x9b, 0x65, 0x50, 0x70, 0x9d,
+ 0x7d, 0x53, 0x3b, 0x86, 0x77, 0x6c, 0xa6, 0x90, 0x6b, 0x3e, 0x7b, 0x7a,
+ 0x50, 0x81, 0xb4, 0x76, 0xa5, 0x74, 0x8b, 0x73, 0x79, 0x69, 0xa8, 0x9a,
+ 0x82, 0x4a, 0x5e, 0x6c, 0x8d, 0x66, 0xa3, 0x80, 0x8d, 0x74, 0x5b, 0x7c,
+ 0x77, 0xaa, 0x82, 0x69, 0x5e, 0x7d, 0x7f, 0x63, 0xa3, 0x8c, 0xb3, 0x9a,
+ 0x81, 0x8f, 0x7b, 0x77, 0x60, 0x89, 0x6a, 0x82, 0x5a, 0x7a, 0x71, 0x61,
+ 0x93, 0x73, 0x8b, 0xb0, 0xa2, 0x92, 0x7c, 0x84, 0x8b, 0x72, 0x91, 0x8d,
+ 0x91, 0x80, 0x6c, 0x75, 0x7a, 0xb3, 0x95, 0x5e, 0xa5, 0x5d, 0x54, 0x8b,
+ 0x63, 0x91, 0xa7, 0x68, 0x96, 0x4c, 0x5a, 0x86, 0x76, 0x82, 0xb6, 0xa0,
+ 0x68, 0x6b, 0x53, 0x76, 0x60, 0x65, 0x90, 0xaf, 0x82, 0x66, 0x80, 0x7b,
+ 0x84, 0xa0, 0xb0, 0xb8, 0x81, 0x6e, 0x81, 0x8a, 0x74, 0x6e, 0x97, 0xa8,
+ 0x89, 0x7b, 0x7b, 0x6e, 0x63, 0x74, 0x5a, 0x7b, 0x7e, 0x84, 0x40, 0x95,
+ 0x73, 0x3c, 0x7c, 0x72, 0x9b, 0x92, 0x27, 0x87, 0x69, 0x5b, 0x99, 0x8a,
+ 0xa8, 0x65, 0x36, 0x8f, 0x86, 0x3e, 0xa1, 0x79, 0x9f, 0x4d, 0x41, 0xc5,
+ 0x8c, 0x6a, 0x7e, 0x7f, 0x68, 0x49, 0x5c, 0x91, 0x50, 0x6a, 0x8c, 0x81,
+ 0x75, 0x4c, 0x6a, 0x74, 0x8a, 0x87, 0xa0, 0x93, 0x7e, 0x6d, 0x52, 0x79,
+ 0x86, 0x6a, 0x68, 0x6c, 0x83, 0x67, 0x79, 0x73, 0x6f, 0x72, 0x97, 0x84,
+ 0x8b, 0x78, 0x64, 0x69, 0x8f, 0x92, 0x86, 0x61, 0x5d, 0x85, 0x70, 0x64,
+ 0x7d, 0xa3, 0x92, 0xa0, 0x72, 0x71, 0x5d, 0x63, 0x7c, 0x70, 0xaf, 0x6f,
+ 0x93, 0x6a, 0x7e, 0x7f, 0x64, 0xab, 0x85, 0x73, 0x8f, 0x8a, 0x7e, 0x5f,
+ 0x7a, 0x6f, 0xaa, 0x71, 0x97, 0x7d, 0x60, 0x7c, 0x48, 0x69, 0xa9, 0xaa,
+ 0x98, 0x7c, 0x61, 0x85, 0x66, 0x97, 0xa2, 0x73, 0x74, 0x65, 0x52, 0x67,
+ 0x79, 0x8a, 0x79, 0x71, 0x85, 0x6e, 0x6d, 0x67, 0x5e, 0x7f, 0xb9, 0x93,
+ 0x96, 0x53, 0x69, 0x6e, 0x7f, 0x8f, 0xab, 0x93, 0xa9, 0x70, 0x6e, 0x71,
+ 0x7e, 0x87, 0x98, 0x7a, 0xae, 0x90, 0x64, 0x88, 0x8a, 0x4f, 0x6d, 0x9e,
+ 0xac, 0x7e, 0x31, 0x92, 0x50, 0x26, 0x95, 0xb2, 0x90, 0x99, 0x0c, 0x84,
+ 0x40, 0x4f, 0x8f, 0x76, 0xa4, 0x46, 0x4c, 0x9d, 0x8b, 0x57, 0x81, 0x79,
+ 0x7b, 0x47, 0x4d, 0x9c, 0x5f, 0x3b, 0x6f, 0x90, 0x7a, 0x3f, 0x66, 0x9d,
+ 0x6c, 0x45, 0x8b, 0x71, 0x79, 0x62, 0x72, 0x78, 0x93, 0x95, 0x7e, 0x86,
+ 0x7a, 0x6b, 0x77, 0x74, 0x6b, 0x86, 0xa4, 0x7e, 0x84, 0x48, 0x78, 0x75,
+ 0x6e, 0x8b, 0x8e, 0x56, 0x69, 0x7b, 0x59, 0x68, 0x5d, 0x77, 0x69, 0x66,
+ 0x67, 0x9f, 0x75, 0x7b, 0x76, 0x64, 0xc1, 0x78, 0x7d, 0x74, 0x82, 0x73,
+ 0x73, 0x90, 0xb8, 0x82, 0x7e, 0x70, 0x7b, 0x7a, 0x64, 0xa1, 0x7e, 0x85,
+ 0x83, 0x81, 0x60, 0x7b, 0x91, 0x82, 0x6f, 0x95, 0xa0, 0x86, 0x6d, 0x88,
+ 0x75, 0x8d, 0x94, 0x90, 0x76, 0x6d, 0x6e, 0x79, 0x64, 0x74, 0xa8, 0xb1,
+ 0x92, 0x6e, 0x61, 0x79, 0x74, 0x91, 0x95, 0x74, 0x65, 0x74, 0x5e, 0x7f,
+ 0x8b, 0x60, 0x9b, 0x9f, 0x74, 0x77, 0x4c, 0x66, 0x7c, 0x80, 0x97, 0x98,
+ 0x9d, 0x86, 0x55, 0x8a, 0x8a, 0x79, 0x8c, 0x82, 0xb0, 0x7d, 0x63, 0x8c,
+ 0x5d, 0x5b, 0x82, 0x58, 0x84, 0x56, 0x51, 0x92, 0x75, 0x24, 0x97, 0x92,
+ 0x75, 0x6e, 0x19, 0x8e, 0x47, 0x3e, 0x7b, 0x7b, 0x87, 0x6b, 0x3f, 0xa9,
+ 0x59, 0x40, 0x86, 0x74, 0x69, 0x4a, 0x2d, 0xad, 0x91, 0x62, 0xb2, 0xa9,
+ 0x74, 0x6c, 0x47, 0x94, 0x51, 0x75, 0xb2, 0x6f, 0x75, 0x4b, 0x60, 0xa2,
+ 0x8e, 0x6a, 0xa4, 0x79, 0x6f, 0x57, 0x80, 0x8c, 0x6c, 0x8e, 0x9e, 0x74,
+ 0x70, 0x5f, 0x66, 0x80, 0x80, 0x89, 0xb5, 0x8a, 0x7a, 0x96, 0x87, 0x7a,
+ 0x7b, 0x85, 0x90, 0x79, 0x59, 0x6d, 0x77, 0x8c, 0x8f, 0x82, 0xb3, 0x9c,
+ 0x6a, 0x6a, 0x6b, 0x70, 0x77, 0x89, 0x96, 0x86, 0x94, 0x72, 0x7e, 0x72,
+ 0xa9, 0x93, 0x8d, 0x7a, 0x6d, 0x8f, 0x66, 0x72, 0x9a, 0x91, 0x9e, 0x98,
+ 0xa0, 0x8b, 0x50, 0x76, 0x5c, 0x74, 0xbc, 0x9a, 0x98, 0x73, 0x80, 0x7d,
+ 0x73, 0x7c, 0xc0, 0x8b, 0x86, 0x7a, 0x66, 0x86, 0x83, 0x72, 0x8f, 0x96,
+ 0x98, 0x56, 0x45, 0x7b, 0x77, 0x92, 0xac, 0x8a, 0xae, 0x43, 0x33, 0x73,
+ 0x78, 0x83, 0x98, 0x84, 0x86, 0x78, 0x54, 0x7e, 0x70, 0x5f, 0xa6, 0xa1,
+ 0x94, 0x81, 0x73, 0x8d, 0x83, 0x5b, 0x88, 0x71, 0xb2, 0x91, 0x50, 0x99,
+ 0x6b, 0x47, 0x72, 0x92, 0x87, 0x6d, 0x07, 0x99, 0x57, 0x3d, 0x8d, 0x83,
+ 0x9d, 0x49, 0x40, 0x9d, 0x5c, 0x57, 0x95, 0x73, 0x6e, 0x4b, 0x49, 0xab,
+ 0x97, 0x58, 0x8b, 0x7a, 0x7a, 0x48, 0x47, 0x8b, 0x7e, 0x5d, 0xa9, 0x6d,
+ 0x8a, 0x3f, 0x60, 0x82, 0x86, 0x98, 0xa9, 0x7c, 0x74, 0x59, 0x9b, 0x80,
+ 0x4e, 0x75, 0x9c, 0x5e, 0x75, 0x8c, 0x67, 0x7e, 0x78, 0x75, 0x87, 0x6c,
+ 0x79, 0x73, 0x63, 0x77, 0x6e, 0x7a, 0x8d, 0x73, 0x4e, 0x72, 0x4a, 0x7c,
+ 0x8f, 0x79, 0x70, 0x7a, 0x70, 0x73, 0x7b, 0x7a, 0x62, 0xa1, 0x7b, 0x63,
+ 0x9a, 0x89, 0x76, 0x64, 0x84, 0x7d, 0x9c, 0x94, 0xb0, 0x7f, 0x6c, 0x7b,
+ 0x8d, 0x89, 0x89, 0x7b, 0x9d, 0x99, 0x64, 0x8b, 0x5c, 0x88, 0xa6, 0x8e,
+ 0x81, 0x86, 0x7e, 0x85, 0x73, 0x72, 0xad, 0x5d, 0x5f, 0x7e, 0x63, 0x74,
+ 0x64, 0xa1, 0x9c, 0x83, 0x7c, 0x83, 0x7b, 0x7b, 0x71, 0xa0, 0x9e, 0xaf,
+ 0x89, 0x79, 0x4c, 0x7c, 0x8c, 0x78, 0x91, 0x87, 0x8a, 0x87, 0x5e, 0x85,
+ 0x7b, 0x61, 0x9c, 0x88, 0xa5, 0x8d, 0x7c, 0x9c, 0x6b, 0x47, 0x95, 0x85,
+ 0x81, 0x80, 0x59, 0xb2, 0x4f, 0x3d, 0xae, 0x8c, 0x8d, 0x71, 0x11, 0x95,
+ 0x31, 0x65, 0x9d, 0xa0, 0x8e, 0x64, 0x42, 0xb9, 0x6a, 0x5c, 0x91, 0x82,
+ 0x91, 0x50, 0x33, 0xb2, 0x7a, 0x54, 0xac, 0x88, 0x92, 0x61, 0x4e, 0xad,
+ 0x65, 0x5c, 0x91, 0xb0, 0x72, 0x65, 0x4a, 0x79, 0x68, 0x77, 0x75, 0x5f,
+ 0x79, 0x6d, 0x6f, 0x7c, 0x4d, 0x71, 0xb8, 0x78, 0x8a, 0x87, 0x6e, 0x72,
+ 0x7d, 0x79, 0x87, 0x80, 0x5a, 0x78, 0x77, 0x78, 0x80, 0x8f, 0x8c, 0x56,
+ 0x7a, 0x8b, 0x62, 0x82, 0x5a, 0x96, 0x82, 0x68, 0x71, 0x5d, 0x75, 0x65,
+ 0x93, 0xb5, 0x71, 0x82, 0x82, 0x8a, 0x4b, 0x7c, 0x62, 0x6f, 0xc1, 0x86,
+ 0x9d, 0x90, 0x63, 0x71, 0x86, 0x9e, 0x9f, 0x77, 0x90, 0x97, 0x68, 0x81,
+ 0x5a, 0x8c, 0xab, 0x5e, 0x81, 0x76, 0x83, 0x79, 0x8f, 0xa1, 0x89, 0x79,
+ 0x81, 0x8a, 0x7e, 0x6c, 0x65, 0x79, 0xc7, 0x89, 0x92, 0x68, 0x78, 0x70,
+ 0x65, 0x96, 0x9e, 0x82, 0x7d, 0x5f, 0x7b, 0x77, 0x72, 0x84, 0x7e, 0x92,
+ 0x97, 0x7b, 0x6e, 0x67, 0x81, 0xa1, 0x9a, 0xab, 0x8d, 0x78, 0x61, 0x78,
+ 0x52, 0x66, 0xaa, 0x77, 0x75, 0xa3, 0x5e, 0xa0, 0x51, 0x40, 0x68, 0xb0,
+ 0x9a, 0x93, 0x11, 0x82, 0x69, 0x48, 0x9c, 0x77, 0x8d, 0x62, 0x36, 0xac,
+ 0x6c, 0x4c, 0xa3, 0xab, 0x8f, 0x32, 0x4f, 0xa9, 0x80, 0x68, 0xab, 0x7a,
+ 0x90, 0x61, 0x5c, 0xa5, 0x84, 0x4c, 0x8c, 0x7a, 0x95, 0x54, 0x72, 0xa0,
+ 0x66, 0x85, 0xb3, 0x91, 0x69, 0x64, 0x68, 0x56, 0x66, 0x8d, 0xa0, 0x9f,
+ 0x7a, 0x88, 0x5d, 0x7d, 0x48, 0x80, 0x7f, 0x7c, 0x7c, 0x99, 0x65, 0x81,
+ 0x73, 0x8b, 0x8c, 0x61, 0x44, 0x60, 0x53, 0x8e, 0x64, 0x80, 0x9c, 0x74,
+ 0x5d, 0x70, 0x8f, 0x5a, 0x68, 0x7a, 0x82, 0xa1, 0x75, 0x7b, 0x83, 0x60,
+ 0x75, 0x5e, 0xa2, 0x94, 0x6a, 0x88, 0x78, 0x71, 0x95, 0x70, 0x8b, 0x86,
+ 0x7e, 0x94, 0x5f, 0x65, 0x5f, 0xb1, 0x97, 0x99, 0x94, 0x84, 0x88, 0x7d,
+ 0x50, 0x8c, 0xaa, 0x81, 0x7b, 0x7c, 0x77, 0x65, 0x5e, 0x91, 0x9c, 0x89,
+ 0x8c, 0x85, 0x75, 0x62, 0x7b, 0x78, 0xc3, 0x7a, 0x62, 0x8c, 0x66, 0x6f,
+ 0x79, 0x7a, 0x9c, 0x6d, 0x7c, 0x6b, 0x5c, 0x7d, 0x6d, 0x54, 0x93, 0x87,
+ 0x7a, 0x7a, 0x50, 0x85, 0x60, 0x56, 0x5e, 0x6b, 0x90, 0x7c, 0x52, 0xa5,
+ 0x54, 0x42, 0x7b, 0x75, 0x83, 0x8c, 0x2c, 0xa6, 0x6f, 0x62, 0x78, 0x78,
+ 0x86, 0x36, 0x4b, 0xaa, 0x86, 0x54, 0x92, 0x8d, 0x7f, 0x53, 0x37, 0xbe,
+ 0x86, 0x7a, 0x90, 0x7e, 0x8e, 0x50, 0x58, 0xa6, 0x82, 0x58, 0x73, 0x74,
+ 0x66, 0x5c, 0x6a, 0x7f, 0xa2, 0x69, 0xbd, 0xa9, 0x74, 0x76, 0x75, 0x6f,
+ 0x45, 0x6c, 0xa5, 0x79, 0x82, 0x67, 0x56, 0x7c, 0x7f, 0x81, 0x67, 0x6d,
+ 0x81, 0x87, 0x71, 0x69, 0x69, 0x81, 0x85, 0x84, 0x5a, 0x8c, 0x5f, 0x73,
+ 0x80, 0x9c, 0x9e, 0x90, 0x77, 0xa0, 0x9c, 0x6c, 0x73, 0x8a, 0x84, 0x72,
+ 0x87, 0xa1, 0x67, 0x64, 0x5d, 0x9b, 0x9d, 0x9b, 0x97, 0x83, 0x5f, 0x61,
+ 0x77, 0x91, 0xa0, 0x8f, 0x8a, 0x6c, 0x45, 0x5f, 0x6d, 0xa6, 0x9b, 0x76,
+ 0x86, 0x93, 0x91, 0x7d, 0x54, 0x61, 0xa4, 0x6a, 0x5b, 0x69, 0x5f, 0x6d,
+ 0x83, 0xaf, 0xa0, 0x78, 0x9d, 0x62, 0x65, 0x69, 0x5f, 0x78, 0xbf, 0x91,
+ 0x7b, 0x7b, 0x52, 0x5d, 0x70, 0x78, 0xa9, 0x87, 0x93, 0x74, 0x61, 0x74,
+ 0x8c, 0x61, 0x97, 0x86, 0x9b, 0x7c, 0x7d, 0x75, 0x4b, 0x64, 0xa7, 0x81,
+ 0x8a, 0x9c, 0x29, 0xa2, 0x5f, 0x38, 0x6a, 0xb0, 0x82, 0x53, 0x1a, 0xa7,
+ 0x38, 0x47, 0x97, 0x90, 0x8d, 0x41, 0x25, 0xa7, 0x65, 0x63, 0x8b, 0x79,
+ 0x8f, 0x3e, 0x21, 0xd0, 0x5e, 0x5d, 0x9d, 0x68, 0x75, 0x3e, 0x68, 0xb6,
+ 0x6a, 0x50, 0x9a, 0x71, 0x81, 0x45, 0x6d, 0x9a, 0x7f, 0x86, 0x9c, 0x63,
+ 0x7d, 0x74, 0x69, 0x7d, 0x5a, 0x6a, 0x8d, 0x72, 0x6b, 0x69, 0x4c, 0x6f,
+ 0x7c, 0x8e, 0xa6, 0x83, 0x70, 0x65, 0x5f, 0x78, 0x69, 0x67, 0x7f, 0x8d,
+ 0x58, 0x76, 0x4a, 0x85, 0x80, 0x89, 0x9f, 0x91, 0x52, 0x62, 0x72, 0x60,
+ 0x7b, 0x5c, 0x77, 0x6f, 0x9d, 0xa4, 0x98, 0x70, 0x6f, 0xad, 0x94, 0x9f,
+ 0x7b, 0x89, 0x74, 0x7e, 0x5d, 0x8d, 0xab, 0x98, 0x8f, 0x90, 0x82, 0x84,
+ 0x60, 0x7c, 0xb7, 0x8e, 0x79, 0x83, 0x56, 0x86, 0x87, 0x79, 0x95, 0x75,
+ 0x78, 0x71, 0x58, 0x73, 0x87, 0x5d, 0xc6, 0x9f, 0x75, 0x61, 0x4f, 0x71,
+ 0x91, 0x88, 0xb3, 0x8c, 0x7d, 0x7c, 0x6a, 0x75, 0x6d, 0x66, 0x8e, 0x94,
+ 0x96, 0x74, 0x59, 0x6f, 0x6d, 0x65, 0xb0, 0x8e, 0x7b, 0x89, 0x7a, 0x6a,
+ 0x7d, 0x57, 0x82, 0x7a, 0x61, 0x9f, 0x50, 0xab, 0x57, 0x46, 0x86, 0x8d,
+ 0xa3, 0x96, 0x18, 0xab, 0x51, 0x6e, 0xb3, 0x7e, 0x90, 0x6d, 0x6d, 0xc0,
+ 0x54, 0x35, 0x96, 0x84, 0x8e, 0x49, 0x28, 0xe4, 0x81, 0x5f, 0x9b, 0x87,
+ 0x8c, 0x33, 0x56, 0xb4, 0x61, 0x5e, 0x8b, 0x81, 0x99, 0x61, 0x6b, 0x96,
+ 0x75, 0x82, 0x9e, 0x7c, 0x90, 0x63, 0x64, 0x6b, 0x55, 0x6e, 0xb6, 0x7f,
+ 0x5f, 0x55, 0x65, 0x60, 0x35, 0x8a, 0x85, 0x91, 0x4d, 0x62, 0x90, 0x90,
+ 0x57, 0x5a, 0x9f, 0x7b, 0x4c, 0x86, 0x73, 0x83, 0x4a, 0x6d, 0xb0, 0x67,
+ 0x65, 0x89, 0x54, 0x68, 0x89, 0x7b, 0x72, 0x4f, 0x7a, 0x93, 0x61, 0x7e,
+ 0x79, 0x89, 0x8f, 0x9c, 0x7b, 0x70, 0x48, 0x67, 0x82, 0x75, 0xaa, 0x92,
+ 0x9a, 0x8f, 0x79, 0x8c, 0x64, 0x94, 0x98, 0x83, 0x7c, 0x8f, 0x5c, 0x77,
+ 0x70, 0x90, 0x91, 0x88, 0x7d, 0x51, 0x5d, 0x5d, 0x8b, 0x9f, 0xbc, 0x78,
+ 0x9e, 0x73, 0x67, 0x6d, 0x82, 0x8d, 0xc9, 0x86, 0x96, 0x6a, 0x5d, 0x79,
+ 0x7e, 0x6b, 0xb2, 0x79, 0x88, 0x85, 0x65, 0x73, 0x75, 0x6b, 0x9e, 0x7f,
+ 0x8e, 0x94, 0x8e, 0x7d, 0x74, 0x61, 0x97, 0x56, 0x97, 0x6b, 0x30, 0xb6,
+ 0x5f, 0x5a, 0xaa, 0xa5, 0x85, 0x5d, 0x01, 0xbc, 0x79, 0x63, 0x6e, 0x82,
+ 0x72, 0x26, 0x4f, 0xc8, 0x98, 0x56, 0x85, 0x9a, 0x81, 0x1f, 0x48, 0xcf,
+ 0x84, 0x74, 0x75, 0x87, 0xae, 0x43, 0x6f, 0xdf, 0x6a, 0x4e, 0x97, 0x5d,
+ 0x8f, 0x37, 0x55, 0x89, 0x7d, 0x82, 0xb1, 0x89, 0x6d, 0x52, 0x65, 0x8b,
+ 0x71, 0x87, 0x8d, 0x6a, 0x99, 0x5d, 0x65, 0x78, 0x67, 0x8d, 0x7b, 0x51,
+ 0x60, 0x8a, 0x59, 0x72, 0x78, 0x93, 0x88, 0x75, 0x46, 0x60, 0x6e, 0x79,
+ 0x7b, 0x9d, 0x9c, 0x8c, 0x5c, 0x7c, 0x69, 0x71, 0x60, 0x6f, 0xb0, 0x7d,
+ 0x4c, 0x5e, 0x88, 0x77, 0x74, 0x6a, 0x6f, 0x9a, 0xa2, 0x83, 0x48, 0x5a,
+ 0x6e, 0xa2, 0x8b, 0x7a, 0x65, 0x5b, 0x4b, 0x80, 0x5b, 0x8f, 0xaf, 0x8e,
+ 0x93, 0x4a, 0x59, 0x6e, 0x5e, 0x89, 0x91, 0x87, 0x73, 0x6a, 0x47, 0x6c,
+ 0x6c, 0x81, 0xad, 0x5a, 0x76, 0x51, 0x51, 0x6c, 0x80, 0x92, 0x9d, 0xae,
+ 0x90, 0x71, 0x6c, 0x7a, 0x7c, 0x84, 0xa7, 0x7d, 0x82, 0x7c, 0x80, 0x59,
+ 0x7d, 0x86, 0xa9, 0x94, 0x8e, 0x7b, 0x7c, 0x67, 0x67, 0x66, 0x8f, 0x49,
+ 0x5d, 0xa4, 0x4a, 0xbc, 0x5a, 0x34, 0xa7, 0xaa, 0x9e, 0x86, 0x17, 0xc0,
+ 0x53, 0x67, 0x76, 0xae, 0x8d, 0x37, 0x4a, 0xd6, 0x76, 0x69, 0x95, 0x7a,
+ 0x8a, 0x0e, 0x3f, 0xe8, 0x60, 0x4d, 0x9e, 0x90, 0xad, 0x44, 0x46, 0xc5,
+ 0x4c, 0x6e, 0x72, 0x8c, 0x89, 0x49, 0x51, 0xa0, 0x60, 0x84, 0x84, 0x9d,
+ 0xa4, 0x5a, 0x84, 0x8d, 0x69, 0x6a, 0x97, 0x78, 0x72, 0x66, 0x72, 0x9b,
+ 0x74, 0x7a, 0x95, 0x7c, 0x7a, 0x6e, 0x74, 0x7f, 0x65, 0x94, 0x77, 0x7e,
+ 0x85, 0x6d, 0x65, 0x7b, 0x63, 0x7b, 0x87, 0x49, 0x80, 0x74, 0x74, 0x85,
+ 0x6e, 0x78, 0xad, 0x66, 0x8a, 0x65, 0x54, 0x7c, 0x4e, 0x62, 0x97, 0x7f,
+ 0x82, 0x6c, 0x58, 0x79, 0x91, 0x94, 0xb3, 0x7a, 0x88, 0x82, 0x60, 0x7f,
+ 0x8c, 0xa7, 0x7b, 0x93, 0x77, 0x49, 0x6f, 0x6f, 0x5a, 0x8d, 0x93, 0x8b,
+ 0x87, 0x59, 0x7d, 0x5e, 0x83, 0x7e, 0x8c, 0x7a, 0x91, 0x4e, 0x6f, 0x89,
+ 0x8a, 0x87, 0x8b, 0x85, 0x8e, 0x43, 0x63, 0x8d, 0x90, 0x6c, 0xa5, 0x73,
+ 0x8a, 0x78, 0x5f, 0x73, 0x88, 0x57, 0x9e, 0x8f, 0x7f, 0x91, 0x70, 0x77,
+ 0x8a, 0x76, 0xa2, 0x77, 0x53, 0x86, 0x51, 0xd8, 0xa9, 0x5b, 0x9b, 0x96,
+ 0x7c, 0x71, 0x01, 0xd4, 0x56, 0x4a, 0x95, 0xab, 0x91, 0x54, 0x45, 0xe5,
+ 0x74, 0x4f, 0x87, 0x6a, 0xa2, 0x3e, 0x47, 0xff, 0x91, 0x4d, 0x94, 0x97,
+ 0x6d, 0x74, 0x77, 0xe0, 0x5d, 0x4e, 0x5f, 0x73, 0x70, 0x3a, 0x68, 0xb2,
+ 0x78, 0x61, 0x8c, 0x77, 0xa8, 0x57, 0x8c, 0x99, 0x23, 0x5a, 0x84, 0x78,
+ 0x9b, 0x7f, 0x5e, 0xa0, 0x49, 0x84, 0x83, 0x94, 0x99, 0x4d, 0x8d, 0x9a,
+ 0x86, 0x90, 0x9b, 0x51, 0x75, 0x73, 0x78, 0x89, 0x59, 0x64, 0x78, 0x91,
+ 0x72, 0x9c, 0x72, 0x7e, 0x65, 0x6a, 0x80, 0xaa, 0x94, 0x65, 0x6d, 0x87,
+ 0x73, 0x93, 0x97, 0x7d, 0x99, 0x63, 0x75, 0x89, 0x67, 0xa1, 0x90, 0x7f,
+ 0x88, 0x65, 0x6d, 0x8f, 0x7d, 0x62, 0x91, 0xa7, 0x8b, 0x73, 0x51, 0x88,
+ 0x66, 0x66, 0x99, 0xa7, 0x7c, 0x54, 0x82, 0x67, 0x64, 0x8a, 0x95, 0x7c,
+ 0x8a, 0x5d, 0x5e, 0x68, 0x4b, 0x75, 0x92, 0x7a, 0x9f, 0x66, 0x71, 0x8d,
+ 0x76, 0x72, 0x8e, 0x77, 0x76, 0x8c, 0x5b, 0x88, 0x9a, 0x92, 0x7c, 0x74,
+ 0x95, 0xaa, 0x71, 0x77, 0x97, 0x93, 0x9e, 0x62, 0x96, 0x6a, 0x49, 0xd8,
+ 0x81, 0x99, 0xae, 0x87, 0x6c, 0x76, 0x3e, 0xd9, 0x6e, 0x95, 0xa3, 0x86,
+ 0x60, 0x6c, 0x5c, 0xbe, 0x98, 0x8a, 0x99, 0x7c, 0x47, 0x45, 0x69, 0xeb,
+ 0x9d, 0x7d, 0xbb, 0x90, 0x66, 0x69, 0x70, 0xc6, 0x7b, 0x59, 0x9e, 0x87,
+ 0x58, 0x76, 0x7c, 0xae, 0x72, 0x7d, 0x9f, 0x92, 0x82, 0x58, 0x51, 0x7a,
+ 0x5d, 0x77, 0xa8, 0x7c, 0x56, 0x68, 0x88, 0x8a, 0x7e, 0x8a, 0x98, 0x68,
+ 0x64, 0x79, 0x6e, 0x7a, 0x60, 0x96, 0x98, 0x60, 0x60, 0x71, 0x60, 0x8e,
+ 0x7c, 0x8c, 0x92, 0x92, 0x77, 0x80, 0x90, 0x91, 0x81, 0x82, 0x9c, 0x80,
+ 0x61, 0x7f, 0x5a, 0x8e, 0x88, 0x7c, 0x8e, 0x79, 0x69, 0x8e, 0x4e, 0x7e,
+ 0x84, 0x9e, 0x67, 0x72, 0x5c, 0x78, 0x7b, 0x8c, 0x65, 0x7d, 0x8e, 0xa4,
+ 0x5e, 0x7a, 0x5c, 0x97, 0x6a, 0x81, 0xab, 0x85, 0x4d, 0x73, 0x83, 0x96,
+ 0x8b, 0x7d, 0xa6, 0x69, 0x74, 0x86, 0x73, 0x79, 0x52, 0x8c, 0xa0, 0x86,
+ 0x64, 0x7b, 0x84, 0x77, 0x87, 0x93, 0x7d, 0x6d, 0x98, 0x6d, 0x88, 0x5f,
+ 0x7c, 0x84, 0x92, 0x82, 0x81, 0x76, 0x85, 0x77, 0x98, 0x85, 0x88, 0x68,
+ 0x7d, 0x71, 0x3c, 0xf1, 0x83, 0x86, 0xa2, 0xb3, 0x6e, 0x77, 0x53, 0xe8,
+ 0xa8, 0xc7, 0xb3, 0x83, 0x93, 0x83, 0x63, 0xe8, 0x94, 0xb3, 0x86, 0x6e,
+ 0x75, 0x5d, 0x54, 0xf0, 0x89, 0xa7, 0x94, 0xb1, 0x7e, 0x91, 0x9a, 0xb8,
+ 0x91, 0x7e, 0x99, 0x50, 0x71, 0x82, 0x8a, 0x91, 0x7a, 0x8a, 0x8b, 0x80,
+ 0x64, 0x6a, 0x5f, 0xbe, 0x5d, 0x96, 0xb1, 0x82, 0x45, 0x71, 0x8b, 0x95,
+ 0x7c, 0x9b, 0x89, 0x6d, 0x5b, 0x73, 0x81, 0x90, 0x76, 0xab, 0xa6, 0x88,
+ 0x62, 0x7d, 0x75, 0x99, 0x7a, 0x8b, 0x6e, 0x9b, 0x83, 0x89, 0x99, 0x93,
+ 0x81, 0x9e, 0x8a, 0x76, 0x75, 0x7d, 0x6c, 0x93, 0x68, 0x7a, 0x8d, 0x78,
+ 0x88, 0x93, 0x66, 0xa5, 0x6c, 0xae, 0xb1, 0x83, 0x72, 0x8f, 0x6b, 0x7b,
+ 0x79, 0x9b, 0x98, 0x7c, 0x82, 0x84, 0x7d, 0x7d, 0x71, 0x7c, 0xb0, 0x81,
+ 0x74, 0x89, 0x72, 0x89, 0x98, 0xa0, 0x7d, 0x62, 0x2f, 0x50, 0x7d, 0x8b,
+ 0x4c, 0x83, 0x87, 0x89, 0x57, 0x9e, 0x92, 0x8c, 0x81, 0x7e, 0xb9, 0x95,
+ 0x7f, 0x76, 0x8e, 0x90, 0x9d, 0x68, 0x78, 0x95, 0x7d, 0xab, 0x84, 0x8a,
+ 0x64, 0x9f, 0x80, 0x94, 0x8d, 0x89, 0x76, 0x8e, 0x6f, 0x8b, 0x75, 0x7d,
+ 0x89, 0x74, 0x67, 0x8a, 0x7d, 0x63, 0x79, 0x6d, 0x79, 0x8a, 0x78, 0x7f,
+ 0x7a, 0x9b, 0x70, 0x70, 0x84, 0x86, 0x80, 0x95, 0x5a, 0x77, 0x80, 0x91,
+ 0x9c, 0x92, 0x76, 0x81, 0x69, 0x89, 0x78, 0xa5, 0x7a, 0x8d, 0x86, 0x64,
+ 0x8f, 0x8d, 0x7d, 0xa1, 0x8c, 0x7b, 0x77, 0x7e, 0x80, 0x93, 0x86, 0x68,
+ 0x90, 0x9c, 0x71, 0x8c, 0x68, 0x52, 0x85, 0x88, 0x89, 0x92, 0x64, 0x8f,
+ 0x74, 0x64, 0x7c, 0x88, 0x8d, 0x97, 0x77, 0x97, 0x91, 0xac, 0x74, 0x7f,
+ 0x60, 0x7e, 0x6e, 0x70, 0x86, 0x83, 0x7f, 0x81, 0x6f, 0x94, 0x62, 0xa4,
+ 0x86, 0x7d, 0x90, 0x7c, 0x89, 0x63, 0x7b, 0x89, 0x75, 0xa1, 0x67, 0x69,
+ 0xa6, 0x76, 0x69, 0x9c, 0x71, 0x79, 0x76, 0x7a, 0x8e, 0x78, 0x94, 0x75,
+ 0x5a, 0x76, 0x6b, 0x91, 0x84, 0x75, 0x72, 0x93, 0x79, 0x7e, 0x75, 0x9a,
+ 0x6f, 0x7a, 0x7b, 0x80, 0x5f, 0x90, 0x74, 0x7d, 0x9b, 0x76, 0x70, 0x89,
+ 0x8f, 0x5f, 0x7f, 0x9c, 0x93, 0x6d, 0x81, 0x7f, 0x8d, 0x7d, 0x74, 0x5d,
+ 0x75, 0x88, 0x7b, 0x91, 0x75, 0x6b, 0x7f, 0x8c, 0x71, 0x74, 0x87, 0x88,
+ 0x83, 0x75, 0x77, 0x96, 0x7f, 0x67, 0x7d, 0x95, 0x81, 0x5c, 0x71, 0x5c,
+ 0x6e, 0x75, 0x86, 0x92, 0x5d, 0x7a, 0x77, 0x9f, 0x6e, 0x79, 0x68, 0x60,
+ 0x94, 0x88, 0x88, 0x88, 0x79, 0x7e, 0x8a, 0x6d, 0x84, 0xa7, 0x5b, 0x8e,
+ 0x67, 0x9c, 0x7e, 0x75, 0x82, 0x96, 0x7c, 0x7b, 0x72, 0x85, 0x8c, 0xa3,
+ 0x96, 0x5b, 0x93, 0x67, 0x7e, 0x9f, 0x71, 0x82, 0x79, 0x8c, 0x93, 0x9d,
+ 0x6b, 0x90, 0x8a, 0x8a, 0x55, 0x82, 0x94, 0x74, 0x7d, 0xaa, 0x81, 0x78,
+ 0x8a, 0x8d, 0x83, 0x7b, 0x97, 0x92, 0x68, 0x64, 0x8c, 0x5d, 0x78, 0x9b,
+ 0x73, 0x95, 0x78, 0x77, 0x6f, 0x61, 0x7c, 0x9d, 0x85, 0x6e, 0x84, 0x4c,
+ 0x87, 0x57, 0x93, 0x68, 0x8e, 0x77, 0x78, 0x72, 0x87, 0x91, 0x5f, 0x7e,
+ 0xa6, 0x75, 0x66, 0x86, 0x7a, 0x7d, 0x70, 0x6f, 0x87, 0x8b, 0x74, 0x85,
+ 0x7d, 0x8b, 0x7f, 0x70, 0x7e, 0x82, 0x84, 0x75, 0x89, 0xa6, 0x7b, 0x7a,
+ 0xa5, 0x69, 0x73, 0x74, 0x82, 0x65, 0x8f, 0x98, 0x7b, 0x77, 0x84, 0x92,
+ 0x73, 0x8a, 0xa1, 0x93, 0x80, 0x81, 0x72, 0x8a, 0x6b, 0x75, 0x8f, 0x98,
+ 0x73, 0x74, 0x6f, 0x70, 0x51, 0x6a, 0x84, 0x9e, 0x78, 0x9b, 0x8c, 0x81,
+ 0x7e, 0x75, 0x80, 0x88, 0x73, 0x4e, 0x71, 0x74, 0x8c, 0x74, 0x6a, 0x84,
+ 0x7f, 0x6b, 0x78, 0xab, 0x77, 0xa2, 0x98, 0x93, 0x77, 0x75, 0x72, 0x5c,
+ 0x60, 0x74, 0x84, 0x67, 0x83, 0x7d, 0x7f, 0x7c, 0x5c, 0x72, 0x70, 0x7f,
+ 0x6c, 0x84, 0x90, 0xab, 0x97, 0x7f, 0x6b, 0x82, 0x7f, 0x78, 0x73, 0x7d,
+ 0x8f, 0x8e, 0x8a, 0x8f, 0x8d, 0xa3, 0x74, 0x6e, 0x5e, 0x8c, 0x94, 0x86,
+ 0x57, 0xb0, 0x79, 0xa8, 0x7b, 0x8d, 0x83, 0x77, 0x89, 0xb6, 0x60, 0x9d,
+ 0x77, 0x59, 0x72, 0x4d, 0x6f, 0x94, 0x71, 0x75, 0x61, 0x96, 0x86, 0x5d,
+ 0x84, 0x68, 0x86, 0x82, 0x8d, 0x70, 0x9a, 0x86, 0x73, 0x64, 0x74, 0x7d,
+ 0x80, 0x5a, 0x64, 0x81, 0xa1, 0x71, 0x77, 0x65, 0xa3, 0x76, 0xa3, 0x9d,
+ 0x73, 0x7b, 0x8f, 0x7b, 0x79, 0x7d, 0x6c, 0x85, 0x8e, 0x75, 0x65, 0x6a,
+ 0x87, 0x70, 0x68, 0x8e, 0x76, 0x5d, 0x66, 0x7c, 0x83, 0x83, 0x7e, 0x89,
+ 0x59, 0x8c, 0x75, 0x59, 0x87, 0x7e, 0x7f, 0x90, 0x6b, 0x7b, 0x7e, 0x6d,
+ 0x6e, 0x86, 0x69, 0x92, 0x83, 0x8f, 0x8a, 0x60, 0x78, 0x75, 0x61, 0x91,
+ 0x73, 0x66, 0x86, 0x86, 0x9f, 0x6f, 0x7b, 0x9a, 0x7c, 0x54, 0x75, 0x8e,
+ 0x7e, 0x72, 0x8e, 0x98, 0x94, 0x5f, 0x71, 0x7c, 0x95, 0x9f, 0x8e, 0x83,
+ 0x96, 0x4b, 0x8d, 0x84, 0x81, 0x7d, 0x70, 0x84, 0x70, 0x53, 0x8d, 0x84,
+ 0x5a, 0x91, 0x88, 0x9a, 0x8f, 0x69, 0x8b, 0x52, 0x85, 0x89, 0x6e, 0x99,
+ 0x79, 0x89, 0x9a, 0x82, 0x6e, 0x8b, 0x65, 0x62, 0x80, 0xa8, 0x8f, 0x8a,
+ 0x71, 0x61, 0x7e, 0x7d, 0x7e, 0xaa, 0x7f, 0xa0, 0x5e, 0x67, 0x90, 0x86,
+ 0x6d, 0xac, 0x74, 0x50, 0x61, 0x91, 0x7d, 0x69, 0x8b, 0x7f, 0x81, 0x7a,
+ 0x93, 0x8c, 0x72, 0x64, 0x98, 0x88, 0x91, 0x83, 0x69, 0x6d, 0x78, 0x7a,
+ 0x68, 0x7c, 0x76, 0x81, 0xa7, 0x88, 0x8f, 0x79, 0x7d, 0x6c, 0x8a, 0x60,
+ 0x88, 0x6d, 0x79, 0x9d, 0x80, 0x82, 0x66, 0x7d, 0x7e, 0x96, 0x78, 0x70,
+ 0x9b, 0x70, 0x7e, 0x90, 0x77, 0x94, 0x7b, 0x89, 0x78, 0x84, 0x74, 0x6d,
+ 0x7d, 0xa7, 0x75, 0x97, 0x85, 0x83, 0x86, 0x65, 0x75, 0x9a, 0x7c, 0x68,
+ 0x87, 0x82, 0x75, 0x68, 0x4c, 0x8a, 0x68, 0x93, 0x7d, 0x88, 0x84, 0x72,
+ 0x58, 0x81, 0x5d, 0x83, 0x89, 0x63, 0x83, 0x7d, 0x8e, 0x75, 0x8c, 0x88,
+ 0x7f, 0x57, 0x8c, 0x8f, 0xa6, 0x71, 0x8a, 0x95, 0x88, 0x51, 0x74, 0x8a,
+ 0x8a, 0x98, 0x72, 0x80, 0x8a, 0x52, 0x90, 0x66, 0x54, 0x8e, 0x7f, 0x94,
+ 0x81, 0x49, 0x84, 0x70, 0x5c, 0x93, 0x89, 0x6d, 0x82, 0x7f, 0x70, 0x5d,
+ 0x87, 0x8a, 0x71, 0x70, 0x6f, 0xa1, 0x90, 0x9f, 0x74, 0x7c, 0x8c, 0x8b,
+ 0x72, 0xbf, 0x89, 0x90, 0x5c, 0x8c, 0x75, 0x72, 0x6f, 0xb2, 0x84, 0x6d,
+ 0x61, 0x80, 0x7d, 0x7a, 0x66, 0xaa, 0x75, 0x71, 0x89, 0x6d, 0x69, 0x72,
+ 0x73, 0x98, 0x8c, 0x78, 0x5a, 0x8e, 0x8c, 0x81, 0x55, 0x81, 0x96, 0x67,
+ 0x6f, 0x71, 0x74, 0x7d, 0x8e, 0x66, 0x9a, 0x67, 0xaa, 0x81, 0x90, 0x79,
+ 0x89, 0x59, 0x86, 0x66, 0x8f, 0x7d, 0x7e, 0xa2, 0xa4, 0x99, 0x68, 0x7a,
+ 0x8c, 0x73, 0x85, 0x77, 0x8b, 0x74, 0x75, 0x66, 0xaa, 0x98, 0x59, 0x8b,
+ 0x91, 0x6c, 0x76, 0x73, 0x87, 0xa4, 0x82, 0x82, 0x63, 0x70, 0x7e, 0x73,
+ 0x96, 0x97, 0x6f, 0x86, 0x81, 0x6f, 0x83, 0x82, 0x7b, 0x82, 0xa3, 0xa7,
+ 0x95, 0x77, 0x84, 0x65, 0x9b, 0x94, 0x6e, 0xb0, 0x75, 0x66, 0x78, 0x82,
+ 0x9c, 0x7a, 0x5f, 0xab, 0x99, 0x2f, 0x7f, 0x68, 0xa4, 0x69, 0x8f, 0x9a,
+ 0x91, 0x56, 0x6e, 0x75, 0x63, 0x9b, 0x9e, 0x97, 0x95, 0x68, 0x80, 0x6a,
+ 0x40, 0x95, 0x53, 0x72, 0x6f, 0x6b, 0x91, 0x78, 0x7f, 0x93, 0x70, 0x8d,
+ 0x62, 0x83, 0x7e, 0x64, 0x5b, 0xaa, 0x70, 0x6c, 0x7e, 0x9c, 0x88, 0x76,
+ 0x60, 0x70, 0x66, 0x69, 0x84, 0x97, 0x9d, 0x63, 0x5e, 0x9a, 0x7e, 0x52,
+ 0x58, 0xb8, 0x95, 0x7c, 0x4d, 0x96, 0x8f, 0x70, 0x71, 0xbf, 0x83, 0x83,
+ 0x9e, 0x70, 0x6f, 0x57, 0x70, 0x9a, 0x8d, 0x6e, 0x98, 0x5a, 0x69, 0x6f,
+ 0x90, 0x71, 0x8a, 0x5d, 0x8e, 0x6e, 0x69, 0x7a, 0x90, 0x86, 0x89, 0x88,
+ 0xb6, 0x77, 0x84, 0x79, 0x76, 0x86, 0x86, 0x7c, 0xbf, 0x6d, 0x5c, 0x90,
+ 0xa1, 0x93, 0x72, 0x63, 0x9a, 0x82, 0x7b, 0x61, 0x91, 0x76, 0x82, 0x96,
+ 0xb9, 0x80, 0x77, 0x7f, 0xa0, 0x73, 0x61, 0x80, 0x83, 0xc1, 0x92, 0x67,
+ 0x7c, 0x81, 0x90, 0x67, 0x8b, 0xbe, 0x81, 0x91, 0x6c, 0x7e, 0x8d, 0x6c,
+ 0x62, 0x83, 0x7e, 0x72, 0x64, 0x8a, 0x83, 0x82, 0xaa, 0x8c, 0x74, 0xab,
+ 0x79, 0x85, 0x91, 0x79, 0x90, 0x68, 0x5c, 0x9a, 0x7c, 0x36, 0x80, 0x6e,
+ 0x93, 0x76, 0x5e, 0xa0, 0xa5, 0x63, 0x73, 0x7e, 0x8d, 0x94, 0x63, 0x99,
+ 0x8f, 0x6a, 0x7f, 0x57, 0x57, 0x6f, 0x6d, 0x86, 0x8e, 0x6b, 0x8d, 0x53,
+ 0x94, 0xba, 0x84, 0x6f, 0x5a, 0x7b, 0x8c, 0x5f, 0x73, 0x93, 0x8b, 0x87,
+ 0x6f, 0x9e, 0x8a, 0x87, 0x62, 0x97, 0x86, 0x7c, 0x69, 0xab, 0xa1, 0x95,
+ 0x42, 0x8c, 0x8b, 0x66, 0x68, 0x99, 0xa8, 0x74, 0x80, 0xa5, 0x7d, 0x82,
+ 0x55, 0xb3, 0x6f, 0x81, 0xa8, 0x9a, 0x80, 0x67, 0x62, 0x7f, 0x78, 0x93,
+ 0x90, 0x83, 0x83, 0x7b, 0x77, 0x73, 0x8c, 0x56, 0xa7, 0x85, 0x7b, 0x71,
+ 0x8f, 0x5d, 0x92, 0x69, 0xbe, 0x5e, 0x7f, 0x7f, 0x8e, 0x71, 0x84, 0x75,
+ 0x95, 0x69, 0x88, 0x6b, 0x96, 0x85, 0x78, 0x39, 0xc2, 0x86, 0x7c, 0x99,
+ 0xa1, 0x94, 0x6b, 0x86, 0xb5, 0x5e, 0x7e, 0x6e, 0x81, 0x95, 0x6a, 0x88,
+ 0x7b, 0x92, 0x8f, 0x68, 0x97, 0x77, 0x84, 0x73, 0x68, 0x96, 0x5a, 0x92,
+ 0x66, 0x74, 0x74, 0x6c, 0x7d, 0x81, 0x6c, 0x93, 0x7f, 0x72, 0x86, 0x74,
+ 0xbf, 0x8f, 0x53, 0xa4, 0x89, 0x76, 0xa0, 0x87, 0x97, 0x6a, 0x6b, 0xb1,
+ 0x91, 0x50, 0x74, 0x68, 0xa3, 0x60, 0x8d, 0xbc, 0xc1, 0x3e, 0x62, 0x59,
+ 0x71, 0x72, 0x6d, 0x80, 0x9f, 0x52, 0x82, 0x6b, 0x5d, 0x7f, 0x74, 0x7e,
+ 0x74, 0x84, 0x8a, 0x59, 0x5c, 0x85, 0x6d, 0x9c, 0x75, 0x9a, 0x88, 0x89,
+ 0x81, 0x9f, 0x81, 0x88, 0x6a, 0x94, 0x84, 0x5f, 0x6b, 0x9b, 0x83, 0x4f,
+ 0x7e, 0xca, 0x99, 0x6d, 0x45, 0x7f, 0x87, 0x71, 0x69, 0xad, 0x95, 0x53,
+ 0x6e, 0x9b, 0x90, 0x73, 0x5d, 0xb0, 0x8d, 0x67, 0x83, 0x82, 0xa3, 0x70,
+ 0x70, 0x92, 0x82, 0x9a, 0x8a, 0x69, 0x6a, 0x6e, 0x7f, 0x89, 0xa4, 0x76,
+ 0x97, 0x62, 0x94, 0x80, 0x87, 0x55, 0x80, 0x76, 0xb3, 0x7e, 0x7e, 0x71,
+ 0x94, 0x88, 0x8e, 0x74, 0xb6, 0x4d, 0x7b, 0x73, 0x90, 0x86, 0x7c, 0x66,
+ 0xb5, 0x80, 0x7f, 0x84, 0x87, 0x82, 0x67, 0x83, 0x97, 0x91, 0x8a, 0x78,
+ 0x8b, 0x83, 0x5d, 0x84, 0x82, 0x9f, 0x8c, 0x91, 0x84, 0x8b, 0x6a, 0x68,
+ 0x86, 0x82, 0x73, 0x77, 0x7b, 0x83, 0x6a, 0x84, 0x92, 0x93, 0x90, 0x8b,
+ 0x4c, 0x94, 0x98, 0x76, 0xb8, 0x7b, 0xa0, 0xa2, 0x7d, 0x3e, 0x95, 0x88,
+ 0xa3, 0x6f, 0x5e, 0xc8, 0x9a, 0x52, 0x81, 0x86, 0xa3, 0x79, 0x88, 0xc3,
+ 0xbd, 0x54, 0x6c, 0x5e, 0x83, 0x8a, 0x98, 0x88, 0x92, 0x66, 0x73, 0x5b,
+ 0x6c, 0x7f, 0x6e, 0x97, 0x8d, 0x58, 0x89, 0x6e, 0x65, 0x7a, 0x7d, 0x7c,
+ 0x7e, 0x89, 0x94, 0x89, 0x55, 0xb8, 0x8f, 0x82, 0x6c, 0x9c, 0x96, 0x5e,
+ 0x6f, 0xb2, 0x70, 0x76, 0x95, 0xc8, 0x86, 0x78, 0x49, 0xac, 0x7e, 0x6c,
+ 0x68, 0xb6, 0xaf, 0x89, 0x68, 0xa5, 0x72, 0x85, 0x69, 0x9c, 0x94, 0x84,
+ 0xa4, 0x97, 0x91, 0x61, 0x7a, 0xa3, 0x8f, 0x8e, 0x93, 0x80, 0x8d, 0x76,
+ 0x74, 0x84, 0x9b, 0x79, 0x97, 0x4e, 0x67, 0x87, 0x9b, 0x69, 0x85, 0x7d,
+ 0xb2, 0x68, 0x76, 0x63, 0xa2, 0x86, 0x97, 0x7f, 0xb5, 0x63, 0x79, 0x76,
+ 0x8a, 0x7c, 0x7c, 0x91, 0xb1, 0x42, 0x7d, 0x7a, 0x8c, 0x8e, 0x72, 0xab,
+ 0xb8, 0x76, 0xab, 0x81, 0x98, 0x85, 0x56, 0x98, 0x84, 0x9f, 0x70, 0x86,
+ 0x76, 0x88, 0x70, 0x8d, 0x71, 0x7b, 0x7a, 0x8d, 0x76, 0x75, 0x62, 0x80,
+ 0x81, 0x94, 0x82, 0x6e, 0x57, 0x8d, 0xaf, 0x84, 0xbf, 0x85, 0x82, 0xa7,
+ 0x80, 0x89, 0x95, 0x81, 0x91, 0x49, 0x72, 0xa1, 0xa7, 0x3f, 0x72, 0x8b,
+ 0x99, 0x72, 0x86, 0xb2, 0xc3, 0x61, 0x55, 0x77, 0x86, 0x77, 0x83, 0xa7,
+ 0x95, 0x5a, 0x68, 0x68, 0x6a, 0x63, 0x6a, 0x77, 0x93, 0x7c, 0x88, 0x62,
+ 0x79, 0x84, 0x8b, 0x82, 0x58, 0x8f, 0x9c, 0x56, 0x77, 0xb1, 0x65, 0x8c,
+ 0x76, 0x91, 0x83, 0x5b, 0x62, 0x91, 0x87, 0x68, 0x71, 0xb0, 0x87, 0x64,
+ 0x62, 0x91, 0x94, 0x58, 0x7f, 0xac, 0xa3, 0x84, 0x75, 0xaa, 0xa3, 0x4d,
+ 0x7a, 0xc2, 0x84, 0x8a, 0x6d, 0xa2, 0x76, 0x74, 0x8c, 0x9e, 0x7c, 0x71,
+ 0x86, 0x70, 0x6d, 0x79, 0x9a, 0x74, 0xb0, 0x8d, 0xa5, 0x7e, 0x6b, 0x63,
+ 0x96, 0x74, 0x99, 0x76, 0xd0, 0x62, 0x85, 0x9d, 0x8f, 0x6d, 0x83, 0x88,
+ 0xb0, 0x62, 0x9b, 0x87, 0x91, 0x82, 0x7a, 0x90, 0x9c, 0x61, 0x6d, 0x97,
+ 0x84, 0x7c, 0x74, 0x8e, 0x8b, 0x75, 0x9a, 0x7e, 0x7c, 0x7d, 0x96, 0x81,
+ 0x94, 0x69, 0x83, 0x6f, 0x8e, 0x7c, 0x7b, 0x7a, 0x73, 0x98, 0x74, 0x9e,
+ 0x72, 0x8c, 0x5f, 0x7d, 0x99, 0x79, 0x5b, 0x73, 0x65, 0x78, 0xa5, 0x7d,
+ 0xa2, 0x98, 0x91, 0x91, 0x87, 0x7b, 0x8c, 0x82, 0xb8, 0x6b, 0x82, 0xba,
+ 0xa5, 0x3f, 0x83, 0x7a, 0x9b, 0x73, 0x93, 0xa1, 0xbe, 0x55, 0x6b, 0x75,
+ 0x94, 0x7d, 0x9c, 0xa1, 0x82, 0x50, 0x75, 0x5a, 0x88, 0x6e, 0x72, 0x7f,
+ 0x99, 0x64, 0x72, 0x49, 0x69, 0x79, 0x6d, 0x94, 0x73, 0x79, 0x80, 0x6f,
+ 0x72, 0xbc, 0x9d, 0x71, 0x7a, 0x9d, 0x8a, 0x55, 0x74, 0xaa, 0xa1, 0x85,
+ 0x7e, 0xc4, 0xa0, 0x7e, 0x50, 0x99, 0x68, 0x8c, 0x8a, 0xb0, 0x99, 0x6c,
+ 0x6d, 0xaf, 0x7b, 0x7b, 0x79, 0xba, 0x8a, 0x7a, 0x9d, 0x8b, 0x67, 0x87,
+ 0x76, 0xa9, 0x7f, 0x7e, 0x8b, 0x7b, 0x87, 0x84, 0x82, 0x74, 0xa3, 0x91,
+ 0x9a, 0x6a, 0x93, 0x7e, 0x87, 0x5b, 0x95, 0x89, 0xbb, 0x5d, 0x74, 0x6c,
+ 0x88, 0x7e, 0x81, 0x7e, 0xb6, 0x6b, 0x91, 0x92, 0x83, 0x78, 0x79, 0x95,
+ 0x90, 0x5e, 0x68, 0x8f, 0xa8, 0x92, 0x66, 0x8e, 0x6b, 0x8c, 0x86, 0x80,
+ 0x7e, 0x7e, 0x70, 0x84, 0x7d, 0x71, 0x67, 0x94, 0x71, 0x69, 0x84, 0x8f,
+ 0x6c, 0x72, 0x85, 0x83, 0x69, 0x76, 0x57, 0x62, 0x83, 0x96, 0x83, 0x77,
+ 0x64, 0x5f, 0xae, 0x7c, 0xa7, 0x88, 0x91, 0x8c, 0x9e, 0x7f, 0xa8, 0x8a,
+ 0x93, 0x6f, 0x58, 0xae, 0xb4, 0x4b, 0x7f, 0x64, 0x9f, 0x5a, 0x9e, 0xb6,
+ 0xa6, 0x6b, 0x79, 0x84, 0x6b, 0x7c, 0x8b, 0x94, 0x85, 0x60, 0x6b, 0x55,
+ 0x79, 0x68, 0x77, 0x75, 0x85, 0x5c, 0x91, 0x5e, 0x5a, 0x71, 0x68, 0x7b,
+ 0x73, 0x91, 0x6c, 0x6e, 0x71, 0x8b, 0x76, 0x86, 0x99, 0xb8, 0x91, 0x68,
+ 0x51, 0xa7, 0x6f, 0x7a, 0x8a, 0xc3, 0x8e, 0x65, 0x64, 0x9e, 0x80, 0x78,
+ 0x6c, 0xc5, 0xa2, 0x75, 0x71, 0xa5, 0x96, 0x4f, 0x70, 0xa4, 0x7a, 0x7c,
+ 0x8c, 0x80, 0x89, 0x97, 0x9a, 0x9a, 0x85, 0x89, 0x92, 0x8f, 0x81, 0x6f,
+ 0x82, 0x6a, 0xb8, 0x74, 0x8f, 0x51, 0x7b, 0x8b, 0x8c, 0x55, 0x7e, 0x8c,
+ 0xb2, 0x41, 0x85, 0x77, 0x9c, 0x73, 0x75, 0x8d, 0x9f, 0x64, 0x92, 0x77,
+ 0xa0, 0x87, 0x5f, 0x71, 0x85, 0x68, 0x8a, 0x78, 0x91, 0x78, 0x75, 0x7a,
+ 0x81, 0x67, 0x96, 0x64, 0x96, 0x85, 0x7a, 0x7e, 0x83, 0x74, 0x82, 0x8f,
+ 0x98, 0x75, 0x77, 0x84, 0x7e, 0x88, 0x94, 0x7d, 0x79, 0x8c, 0x47, 0x79,
+ 0x96, 0x7f, 0x8e, 0x90, 0x50, 0x7f, 0xa3, 0x77, 0xa8, 0x7f, 0x65, 0x9f,
+ 0xb9, 0x4c, 0xa7, 0x7f, 0xaa, 0x6e, 0xa2, 0xb0, 0xb8, 0x51, 0x6b, 0x74,
+ 0xaa, 0x63, 0x6c, 0xa3, 0xb6, 0x5e, 0x74, 0x6a, 0x75, 0x69, 0x87, 0x7f,
+ 0x9d, 0x71, 0x73, 0x72, 0x70, 0x57, 0x5a, 0x7e, 0x8b, 0x64, 0x9a, 0x4d,
+ 0x97, 0x81, 0x7b, 0x75, 0x6e, 0x92, 0x5f, 0x67, 0x7e, 0xaa, 0x90, 0x7a,
+ 0x92, 0xae, 0x92, 0x68, 0x79, 0x9d, 0x4f, 0x6c, 0x79, 0xb4, 0x9c, 0x58,
+ 0x86, 0x8e, 0x62, 0x72, 0x71, 0xc1, 0xac, 0x7d, 0x7a, 0x94, 0x8f, 0x7b,
+ 0x88, 0xa8, 0x8d, 0x82, 0x75, 0x9b, 0x5f, 0x83, 0x82, 0xb3, 0x7a, 0x93,
+ 0x94, 0x76, 0x70, 0x7e, 0x72, 0x7e, 0x8f, 0x8c, 0xa7, 0x53, 0x72, 0x77,
+ 0x7a, 0x64, 0xa8, 0x83, 0xc5, 0x56, 0x71, 0x7b, 0x96, 0x73, 0x7c, 0x73,
+ 0x93, 0x49, 0x83, 0x99, 0xa2, 0x83, 0x74, 0x79, 0xa4, 0x61, 0x8e, 0x84,
+ 0x7a, 0x7d, 0x56, 0x98, 0x97, 0x6d, 0x87, 0x8c, 0x7a, 0x77, 0x6a, 0x67,
+ 0x8a, 0x6f, 0xa2, 0x82, 0x8d, 0x85, 0x6d, 0x8f, 0x7e, 0x74, 0x72, 0x74,
+ 0x91, 0x75, 0x58, 0x7f, 0x9e, 0x7c, 0x9c, 0x75, 0x61, 0x6f, 0x85, 0x7b,
+ 0xbe, 0x84, 0x85, 0x9b, 0x8c, 0x3b, 0x9a, 0x90, 0xab, 0x77, 0x8e, 0xa2,
+ 0xbd, 0x55, 0x96, 0x70, 0xa8, 0x78, 0x98, 0x9c, 0xc3, 0x67, 0x6e, 0x81,
+ 0x70, 0x75, 0x96, 0x9c, 0x8a, 0x5b, 0x73, 0x54, 0x69, 0x6c, 0x5d, 0x82,
+ 0x99, 0x5b, 0x8c, 0x6d, 0x87, 0x80, 0x67, 0x86, 0x88, 0x7c, 0x70, 0x6b,
+ 0x75, 0xab, 0x8e, 0x79, 0x90, 0x91, 0xaf, 0x67, 0x5c, 0xa1, 0x5c, 0x6f,
+ 0x75, 0xa1, 0x95, 0x5f, 0x82, 0x8f, 0x78, 0x5d, 0x7c, 0xb8, 0x8a, 0x8a,
+ 0x6a, 0x98, 0x6e, 0x51, 0x6b, 0xaa, 0x7d, 0x7c, 0x80, 0x94, 0x79, 0x6d,
+ 0xaa, 0x8a, 0x7e, 0x77, 0xa4, 0x78, 0xa5, 0x6d, 0x7c, 0x75, 0xa8, 0x6f,
+ 0xa6, 0x51, 0x8e, 0x80, 0x96, 0x5b, 0x9d, 0x7b, 0xb8, 0x4e, 0x6c, 0x87,
+ 0x95, 0x7c, 0x78, 0x71, 0xb0, 0x5a, 0x99, 0xa0, 0x90, 0x87, 0x65, 0x8b,
+ 0x98, 0x68, 0x92, 0x76, 0x82, 0x77, 0x6a, 0x8a, 0x91, 0x84, 0x87, 0x8b,
+ 0x87, 0x84, 0x7a, 0x81, 0x77, 0x55, 0x8e, 0x86, 0x7a, 0x74, 0x65, 0x88,
+ 0x62, 0x51, 0xa1, 0x91, 0x88, 0x76, 0x5f, 0x89, 0x9f, 0x86, 0x66, 0x67,
+ 0x64, 0x75, 0x9e, 0x74, 0xc1, 0x80, 0x58, 0xa9, 0x8f, 0x5e, 0x94, 0x88,
+ 0xaf, 0x6f, 0x6c, 0xa4, 0xa1, 0x4d, 0x68, 0x66, 0xc2, 0x6e, 0x89, 0x9b,
+ 0xa3, 0x5a, 0x63, 0x5b, 0x9c, 0x7a, 0x93, 0x76, 0x9d, 0x6d, 0x71, 0x5d,
+ 0x80, 0x66, 0x79, 0x80, 0x7c, 0x65, 0x74, 0x64, 0x88, 0x90, 0x79, 0x89,
+ 0x72, 0x88, 0x67, 0x75, 0x6a, 0x96, 0x56, 0x67, 0x88, 0xa1, 0x8c, 0x6c,
+ 0x55, 0xb2, 0x8a, 0x71, 0x88, 0xdc, 0x7a, 0x72, 0x94, 0x9d, 0x7c, 0x76,
+ 0x6a, 0xaa, 0xa8, 0x7f, 0x80, 0xa0, 0x6b, 0x6f, 0x84, 0xe0, 0x68, 0x93,
+ 0xa6, 0x99, 0x69, 0x68, 0x93, 0xa0, 0x93, 0x6b, 0x87, 0x8b, 0x80, 0x90,
+ 0x90, 0x89, 0x8f, 0x7f, 0xaf, 0x6f, 0x82, 0x6d, 0x94, 0x70, 0x97, 0x8f,
+ 0xb0, 0x40, 0x9b, 0x67, 0x78, 0x86, 0x90, 0x8b, 0xa7, 0x51, 0x7f, 0x79,
+ 0x90, 0x71, 0x6d, 0x80, 0x95, 0x63, 0x7d, 0x87, 0xa0, 0x7e, 0x7b, 0x85,
+ 0x8e, 0x6d, 0xa1, 0x76, 0x70, 0x7b, 0x66, 0x87, 0x90, 0x7a, 0x86, 0x88,
+ 0x89, 0x87, 0x6a, 0x91, 0x78, 0x74, 0x76, 0x8d, 0x7e, 0x86, 0x63, 0x90,
+ 0x98, 0x7d, 0x4a, 0x85, 0x4f, 0x9d, 0xa2, 0x7c, 0xb4, 0x88, 0x78, 0xb5,
+ 0x8f, 0x3f, 0xa7, 0x7d, 0xa4, 0x7c, 0x60, 0x9c, 0xa8, 0x41, 0x6b, 0x7f,
+ 0xa2, 0x7f, 0x68, 0xaa, 0xb4, 0x73, 0x56, 0x62, 0x87, 0x72, 0xa5, 0x7c,
+ 0x97, 0x69, 0x58, 0x6b, 0x89, 0x57, 0x51, 0x80, 0x92, 0x7a, 0x7c, 0x4c,
+ 0x7c, 0x7b, 0x69, 0x5f, 0x90, 0x77, 0x78, 0x67, 0x7a, 0xad, 0x79, 0x5c,
+ 0x9c, 0xbf, 0xa6, 0x64, 0x53, 0xb3, 0x5e, 0x59, 0x86, 0xb9, 0x94, 0x65,
+ 0x70, 0x9d, 0x7a, 0x80, 0x7c, 0xae, 0x9c, 0x7b, 0x66, 0xae, 0x83, 0x5f,
+ 0x81, 0xc5, 0x8b, 0x7e, 0x9b, 0x89, 0x84, 0x7f, 0x7c, 0xa5, 0x5c, 0x89,
+ 0x8a, 0x75, 0x99, 0x6d, 0x8e, 0x90, 0x9f, 0x81, 0x81, 0x6b, 0x87, 0x76,
+ 0x92, 0x6f, 0xab, 0x95, 0x95, 0x4c, 0x97, 0x72, 0x80, 0x87, 0x83, 0x87,
+ 0xa3, 0x59, 0xad, 0x74, 0x93, 0x7f, 0x77, 0x78, 0x8d, 0x66, 0x9b, 0x7a,
+ 0x7d, 0x95, 0x64, 0x7f, 0x6d, 0x5c, 0x8e, 0x94, 0x92, 0x82, 0x60, 0x8d,
+ 0x75, 0x55, 0x8c, 0x8b, 0x8f, 0x86, 0x7d, 0x7c, 0x74, 0x57, 0x78, 0x9d,
+ 0x71, 0x65, 0x66, 0x7f, 0xaa, 0x92, 0x66, 0x81, 0x5a, 0x71, 0xa6, 0x78,
+ 0x9d, 0x8a, 0x5a, 0x8a, 0x91, 0x59, 0xb7, 0x5c, 0xc3, 0x73, 0x89, 0x9d,
+ 0xa7, 0x62, 0x77, 0x72, 0x9f, 0x92, 0x6a, 0x9f, 0xaa, 0x71, 0x6b, 0x5e,
+ 0x7d, 0x73, 0x8d, 0x89, 0xba, 0x61, 0x73, 0x6e, 0x71, 0x8a, 0x79, 0x7c,
+ 0x94, 0x76, 0x76, 0x65, 0x81, 0x6f, 0x4e, 0x75, 0x6e, 0x8b, 0x7d, 0x50,
+ 0x56, 0xb8, 0x72, 0x67, 0x93, 0xc6, 0x88, 0x6f, 0x57, 0xb7, 0x80, 0x4c,
+ 0x97, 0xc4, 0xb6, 0x71, 0x72, 0x9e, 0x6f, 0x72, 0x8d, 0xa5, 0x8f, 0x89,
+ 0x74, 0xae, 0x78, 0x70, 0x6e, 0xbb, 0x8f, 0x73, 0x74, 0x8b, 0x5e, 0x86,
+ 0x8b, 0x8a, 0x72, 0x71, 0x84, 0x84, 0x77, 0xa3, 0xa6, 0x73, 0xa4, 0x7e,
+ 0xab, 0x5d, 0x75, 0x96, 0x94, 0x5f, 0x8b, 0x74, 0x9c, 0x63, 0x8d, 0x81,
+ 0x80, 0x6a, 0x91, 0x88, 0x93, 0x53, 0x80, 0x75, 0x79, 0x8d, 0x78, 0x74,
+ 0x7c, 0x73, 0xb2, 0x89, 0x8e, 0xab, 0x75, 0x6c, 0x7a, 0x79, 0x99, 0x77,
+ 0x7d, 0x89, 0x5a, 0x81, 0x7c, 0x75, 0x6a, 0x7e, 0x8c, 0x83, 0x78, 0x8e,
+ 0x62, 0x76, 0x77, 0x6b, 0x79, 0x66, 0x6e, 0x82, 0xa1, 0x8d, 0x52, 0x79,
+ 0x70, 0x7d, 0xa9, 0x6a, 0x95, 0x7f, 0x59, 0x94, 0x8f, 0x73, 0xb7, 0x85,
+ 0xb3, 0x80, 0x77, 0x9f, 0xb8, 0x4d, 0x82, 0x7c, 0xa0, 0xa4, 0x7b, 0x8c,
+ 0xa9, 0x78, 0x62, 0x6b, 0x8a, 0x93, 0x80, 0x68, 0x9b, 0x6d, 0x6b, 0x7b,
+ 0x84, 0x8f, 0x86, 0x70, 0x70, 0x73, 0x84, 0x4f, 0x7c, 0x75, 0x64, 0x8d,
+ 0x6e, 0x81, 0x7c, 0x72, 0x81, 0xb0, 0x74, 0x65, 0xa7, 0xae, 0x80, 0x70,
+ 0x5e, 0xa4, 0x58, 0x54, 0x8e, 0xa7, 0x96, 0x65, 0x66, 0x8b, 0x6c, 0x5d,
+ 0x6b, 0xbe, 0x94, 0x79, 0x80, 0xa1, 0x91, 0x78, 0x6d, 0xc2, 0x82, 0x85,
+ 0x81, 0x7d, 0x88, 0x79, 0x93, 0x96, 0x7f, 0x7e, 0x7d, 0x92, 0x75, 0xa2,
+ 0x9f, 0x7b, 0x92, 0x77, 0x8a, 0x7c, 0x80, 0x8b, 0x9b, 0x64, 0xa5, 0x74,
+ 0xa1, 0x74, 0x7f, 0x7e, 0x85, 0x78, 0x9c, 0x86, 0x9f, 0x62, 0x8f, 0x7f,
+ 0x8a, 0x90, 0x6d, 0x7d, 0x93, 0x61, 0x9d, 0x81, 0x9b, 0x99, 0x69, 0x87,
+ 0x74, 0x7d, 0x8e, 0x8e, 0x7b, 0x7c, 0x6a, 0x71, 0x7d, 0x7f, 0x74, 0x74,
+ 0x7b, 0x65, 0x6e, 0x91, 0x7c, 0x6e, 0x80, 0x8c, 0x8a, 0x6c, 0x6b, 0x76,
+ 0xad, 0x94, 0x64, 0x81, 0x69, 0x7b, 0xac, 0x76, 0x9f, 0x71, 0x85, 0x85,
+ 0x8b, 0x66, 0xb5, 0x87, 0xb3, 0x63, 0x8b, 0x95, 0x8e, 0x50, 0x91, 0x77,
+ 0xa1, 0x99, 0x64, 0x81, 0xb3, 0x63, 0x6e, 0x7a, 0x7f, 0x73, 0x7a, 0x7b,
+ 0x93, 0x6d, 0x75, 0x75, 0x7c, 0x7b, 0x59, 0x7c, 0x7c, 0x68, 0x67, 0x78,
+ 0x79, 0x75, 0x53, 0x86, 0x84, 0x84, 0x91, 0x71, 0x85, 0xb1, 0x84, 0x64,
+ 0x88, 0xc0, 0x94, 0x5f, 0x6f, 0x9b, 0x69, 0x67, 0x97, 0x94, 0x88, 0x6a,
+ 0x7e, 0x94, 0x9e, 0x7f, 0x81, 0x9c, 0xa7, 0x7f, 0x7a, 0xa2, 0x63, 0x69,
+ 0x82, 0xc2, 0x5e, 0x8d, 0x7c, 0x89, 0x63, 0x93, 0x84, 0xb8, 0x76, 0x89,
+ 0x96, 0x87, 0x79, 0x88, 0xa6, 0x8e, 0x9b, 0x93, 0x9c, 0x5d, 0x92, 0x92,
+ 0x82, 0x5e, 0x85, 0x88, 0xad, 0x73, 0xa4, 0x6f, 0x74, 0x8e, 0x77, 0x89,
+ 0x9b, 0x6e, 0x82, 0x76, 0x93, 0xae, 0x82, 0x87, 0x76, 0x6f, 0x80, 0x76,
+ 0x95, 0x8e, 0x5e, 0x85, 0x7b, 0x68, 0x7f, 0x7c, 0x82, 0x94, 0x80, 0x91,
+ 0x77, 0x71, 0x7c, 0x94, 0x80, 0x62, 0x65, 0x7c, 0x5e, 0x70, 0x76, 0x75,
+ 0x7b, 0x60, 0x5f, 0x69, 0xb3, 0x6e, 0x95, 0x9d, 0x5a, 0x5b, 0x9e, 0x6e,
+ 0xa6, 0x80, 0x5d, 0xa5, 0x83, 0x5b, 0xa4, 0x80, 0xb3, 0x79, 0x83, 0xb6,
+ 0xa3, 0x73, 0x84, 0x67, 0x8d, 0x8f, 0x9d, 0x78, 0xb8, 0x8a, 0x7b, 0x6c,
+ 0x85, 0x87, 0x6d, 0x75, 0xae, 0x75, 0x53, 0x71, 0x6b, 0x87, 0x67, 0x7b,
+ 0x7f, 0x86, 0x58, 0x73, 0x7d, 0x87, 0x5d, 0x7f, 0x7d, 0x63, 0x92, 0x65,
+ 0x7a, 0x9c, 0x6f, 0x87, 0x81, 0xa9, 0x91, 0x54, 0x66, 0x8e, 0x58, 0x6d,
+ 0x92, 0xc2, 0xa9, 0x7b, 0x6e, 0x96, 0x7c, 0x60, 0x7e, 0xa8, 0x85, 0x94,
+ 0x90, 0x8b, 0x77, 0x79, 0x77, 0xa7, 0x8f, 0x83, 0x80, 0x99, 0x8c, 0x80,
+ 0x93, 0x9c, 0x73, 0x9e, 0x75, 0x90, 0x67, 0x74, 0x99, 0x98, 0x7e, 0x76,
+ 0x9f, 0x82, 0x90, 0x95, 0x9d, 0x5f, 0x95, 0x98, 0x8c, 0x5f, 0x77, 0x83,
+ 0x7b, 0x72, 0x85, 0x7c, 0x97, 0x74, 0x81, 0x80, 0x8d, 0x89, 0x7d, 0x69,
+ 0x95, 0x85, 0x83, 0x5e, 0x95, 0x74, 0x54, 0x7f, 0x6c, 0x67, 0x9b, 0x83,
+ 0x88, 0x8e, 0x6f, 0x96, 0x81, 0x7f, 0x6e, 0x87, 0x8f, 0x6f, 0x61, 0x87,
+ 0x63, 0x66, 0x72, 0x77, 0x75, 0x6d, 0x59, 0x7d, 0xaa, 0x85, 0x62, 0x83,
+ 0x97, 0x94, 0x96, 0x89, 0x9d, 0x90, 0x7d, 0x91, 0x78, 0x57, 0xa0, 0x7f,
+ 0xa2, 0x62, 0x63, 0x99, 0x77, 0x71, 0x7f, 0x61, 0x99, 0x89, 0x6f, 0xa2,
+ 0xae, 0x92, 0x88, 0x51, 0x87, 0x7a, 0x6f, 0x89, 0xa8, 0x89, 0x64, 0x81,
+ 0x84, 0x79, 0x5b, 0x73, 0x82, 0x6e, 0x7e, 0x5d, 0x8f, 0x82, 0x51, 0x69,
+ 0x8e, 0x76, 0x8b, 0x58, 0x89, 0xb2, 0x52, 0x72, 0x7f, 0xae, 0x96, 0x5a,
+ 0x80, 0xa1, 0x74, 0x62, 0x8d, 0xbe, 0x87, 0x6c, 0x6d, 0xad, 0x83, 0x5a,
+ 0x6c, 0xa5, 0x7f, 0x7c, 0x7a, 0xa1, 0x75, 0x6d, 0x85, 0xbe, 0x91, 0x8e,
+ 0x96, 0x8c, 0x87, 0x74, 0x8b, 0x82, 0x96, 0x8f, 0x8f, 0x93, 0x8f, 0x8c,
+ 0x9a, 0x78, 0x73, 0x6e, 0x91, 0x8d, 0x7e, 0x81, 0x81, 0x52, 0x90, 0x85,
+ 0x77, 0x66, 0x7e, 0x75, 0x8a, 0x67, 0x72, 0x76, 0x82, 0x7b, 0x6e, 0x67,
+ 0x96, 0x7b, 0x75, 0x76, 0x8d, 0x76, 0x7f, 0x79, 0x84, 0x7b, 0x57, 0x81,
+ 0x76, 0x80, 0x67, 0x8c, 0x7c, 0x80, 0x67, 0x85, 0x79, 0x5b, 0x97, 0x74,
+ 0x91, 0x75, 0x82, 0x75, 0x6b, 0x94, 0x7e, 0x85, 0x8e, 0x77, 0x5d, 0x78,
+ 0xb5, 0x8b, 0x73, 0x7f, 0x62, 0x8f, 0xb1, 0x7d, 0xa2, 0x85, 0x6b, 0x92,
+ 0x75, 0x75, 0xb8, 0x7d, 0xb3, 0x67, 0x5f, 0xa6, 0x9b, 0x85, 0x9a, 0x67,
+ 0xbe, 0x8d, 0x92, 0x88, 0xa5, 0x7c, 0xaa, 0x5a, 0x71, 0x7b, 0x70, 0x77,
+ 0xa0, 0xa4, 0x5e, 0x55, 0x6b, 0x8e, 0x53, 0x89, 0x8a, 0x5a, 0x7c, 0x54,
+ 0x7c, 0x8b, 0x53, 0x77, 0x67, 0x77, 0x67, 0x5d, 0x91, 0xac, 0x78, 0x81,
+ 0x8e, 0xb5, 0x6d, 0x58, 0x78, 0xa6, 0x7c, 0x85, 0x87, 0xb3, 0x76, 0x5d,
+ 0x7c, 0x87, 0x57, 0x68, 0x82, 0x8f, 0x89, 0x76, 0x86, 0x9f, 0x6c, 0x68,
+ 0x7c, 0x87, 0x79, 0x9f, 0x86, 0x9e, 0x83, 0x70, 0x8d, 0xb2, 0x84, 0x71,
+ 0x71, 0x91, 0x9f, 0x8e, 0x83, 0x84, 0x87, 0x80, 0x94, 0x80, 0x7d, 0x8d,
+ 0x7c, 0x56, 0x5f, 0x80, 0x7d, 0x84, 0x61, 0x6e, 0x69, 0x80, 0x8b, 0x67,
+ 0xa4, 0x8b, 0x98, 0x7a, 0x8a, 0x6c, 0x77, 0x66, 0x7d, 0x6e, 0x84, 0x78,
+ 0x82, 0x7d, 0x61, 0x88, 0x6e, 0x53, 0x92, 0x75, 0x88, 0x77, 0x82, 0x9f,
+ 0x9e, 0x6f, 0x9c, 0x76, 0x91, 0x78, 0x69, 0x7f, 0x71, 0x6c, 0x6f, 0x7d,
+ 0x83, 0x6e, 0x3c, 0x84, 0x90, 0x8b, 0x71, 0x69, 0x75, 0x81, 0xc8, 0x84,
+ 0xa7, 0x8a, 0x8a, 0x90, 0x96, 0x86, 0x9e, 0x68, 0x99, 0x84, 0x8c, 0xa0,
+ 0x8a, 0x71, 0x7d, 0x41, 0xa1, 0x98, 0x77, 0x91, 0xaa, 0x86, 0x96, 0x5e,
+ 0x86, 0x76, 0xa7, 0x83, 0xac, 0x86, 0x66, 0x46, 0x6a, 0x81, 0x64, 0x77,
+ 0x67, 0x53, 0x80, 0x59, 0x73, 0x71, 0x63, 0x71, 0x76, 0x86, 0x62, 0x4f,
+ 0x83, 0xa4, 0x5d, 0x66, 0x93, 0x87, 0x87, 0x5b, 0x7f, 0x9d, 0x61, 0x9d,
+ 0x94, 0xa4, 0x84, 0x75, 0x67, 0xb3, 0x7b, 0x6d, 0x64, 0x98, 0x62, 0x77,
+ 0x7d, 0x98, 0x8e, 0x75, 0x7d, 0xa6, 0xa4, 0x8c, 0x83, 0x8b, 0x7a, 0x97,
+ 0x6c, 0x7f, 0x66, 0x7f, 0x8f, 0x98, 0x72, 0x6e, 0x75, 0x65, 0x80, 0x8d,
+ 0x88, 0x7d, 0x8c, 0x8d, 0x67, 0x68, 0xab, 0x8c, 0x8b, 0x76, 0x87, 0x69,
+ 0x88, 0x6c, 0x83, 0x6e, 0x88, 0x64, 0xa8, 0x67, 0xa5, 0x5b, 0x65, 0x60,
+ 0x6b, 0x62, 0x76, 0x78, 0x8c, 0x5b, 0x61, 0x6f, 0x66, 0x65, 0x92, 0x67,
+ 0x84, 0x7b, 0x80, 0x86, 0x7b, 0x6c, 0x86, 0x7a, 0x72, 0x7b, 0x4d, 0x94,
+ 0x80, 0x67, 0x8e, 0x8d, 0x7f, 0x79, 0x65, 0x78, 0xa3, 0x71, 0x80, 0x74,
+ 0xa7, 0xa8, 0x97, 0x78, 0x91, 0x77, 0x98, 0x86, 0x82, 0x64, 0xa5, 0x6e,
+ 0x7a, 0x5d, 0x6f, 0xad, 0x9b, 0x7a, 0x91, 0x4b, 0xa1, 0x75, 0x95, 0x76,
+ 0xac, 0x9d, 0xa3, 0x65, 0x65, 0x6a, 0x81, 0x8b, 0x9f, 0x67, 0x6b, 0x6a,
+ 0x60, 0x5b, 0x77, 0x96, 0x73, 0x78, 0x5a, 0x77, 0x5f, 0x68, 0x70, 0x72,
+ 0x78, 0x65, 0x81, 0x20, 0x86, 0x99, 0x80, 0x7a, 0xa5, 0xb1, 0x69, 0x45,
+ 0x7d, 0xa6, 0x7d, 0x85, 0xaa, 0xa9, 0x65, 0x60, 0x75, 0x9b, 0x61, 0x92,
+ 0x91, 0x8f, 0x8a, 0x81, 0x88, 0x9c, 0x81, 0x7d, 0x7b, 0x8f, 0x7e, 0x9e,
+ 0x82, 0x94, 0x95, 0x80, 0x73, 0xae, 0x7b, 0x7a, 0x79, 0x8c, 0x8b, 0x65,
+ 0x71, 0x75, 0x8d, 0x7a, 0x90, 0x83, 0x7b, 0x77, 0x71, 0x4f, 0x70, 0x95,
+ 0x87, 0x69, 0x97, 0x8e, 0x70, 0x92, 0x6e, 0x91, 0x9d, 0x72, 0x75, 0x82,
+ 0xad, 0x81, 0x78, 0x8d, 0x6f, 0x65, 0x88, 0x86, 0x8c, 0x8e, 0x59, 0x8b,
+ 0x67, 0x69, 0x8b, 0x78, 0x7f, 0x59, 0x73, 0x87, 0x6f, 0x86, 0x66, 0x7c,
+ 0x96, 0x68, 0x59, 0x78, 0x67, 0x92, 0x7b, 0x76, 0x80, 0x6e, 0x4a, 0x7b,
+ 0x99, 0x67, 0x72, 0x9c, 0x7a, 0x80, 0x76, 0x5f, 0x8e, 0x4f, 0x71, 0x77,
+ 0xab, 0x78, 0x99, 0x50, 0x83, 0x65, 0x78, 0x8c, 0xbb, 0x8d, 0x4e, 0x54,
+ 0x81, 0x6f, 0x7f, 0x91, 0xb9, 0x79, 0x9c, 0x65, 0x5a, 0x5a, 0x73, 0x8c,
+ 0x9a, 0xac, 0x99, 0x44, 0x7d, 0x4f, 0x78, 0x5a, 0x7d, 0x79, 0x57, 0x44,
+ 0x6f, 0x6a, 0x75, 0x7f, 0x5f, 0x6f, 0x72, 0x62, 0x7f, 0x89, 0x57, 0x91,
+ 0x8d, 0x83, 0x7e, 0x63, 0x8c, 0x95, 0x48, 0x78, 0xa9, 0x88, 0x84, 0x5b,
+ 0x8c, 0xa5, 0x65, 0x71, 0x88, 0x82, 0x7e, 0xa4, 0x8d, 0x7d, 0x7d, 0x8d,
+ 0x91, 0x7c, 0x73, 0x7d, 0x99, 0x89, 0x6d, 0xa1, 0x98, 0x84, 0x8b, 0x6b,
+ 0x89, 0x86, 0x84, 0x7e, 0x86, 0x87, 0x78, 0x8c, 0x96, 0x92, 0x5a, 0xa0,
+ 0x64, 0x73, 0x91, 0x88, 0x8f, 0x6b, 0x96, 0x5c, 0x99, 0x62, 0x78, 0x6c,
+ 0x87, 0x4d, 0x5d, 0x69, 0x7b, 0x81, 0x4a, 0x61, 0x71, 0x69, 0x7d, 0x91,
+ 0x67, 0x92, 0x68, 0x6f, 0x50, 0x5e, 0x61, 0x7e, 0x81, 0x70, 0x5f, 0x7b,
+ 0x6b, 0x55, 0x71, 0x6c, 0x70, 0x53, 0x3f, 0x80, 0x6e, 0x57, 0x96, 0x84,
+ 0x75, 0x51, 0x60, 0x9a, 0x7f, 0xa5, 0x80, 0x94, 0x95, 0x74, 0x7c, 0x83,
+ 0xa0, 0x93, 0x5d, 0x92, 0x83, 0x66, 0x67, 0x8a, 0x8b, 0x9b, 0x81, 0x69,
+ 0x73, 0x91, 0x6b, 0x79, 0x93, 0x88, 0x64, 0x68, 0x81, 0x8c, 0x6f, 0x81,
+ 0x6f, 0x80, 0x68, 0x5f, 0x9c, 0x95, 0x76, 0x93, 0x87, 0x68, 0x83, 0x94,
+ 0x8b, 0x85, 0x72, 0x7f, 0x64, 0x8c, 0x6a, 0x95, 0x8d, 0x80, 0x69, 0x6b,
+ 0x98, 0x86, 0x75, 0x92, 0x7a, 0x7f, 0x5b, 0x7f, 0x9b, 0x57, 0x99, 0x8d,
+ 0x8a, 0x7b, 0x58, 0x73, 0x88, 0x6d, 0x8a, 0x8c, 0x8e, 0x82, 0x85, 0xaa,
+ 0x72, 0xa6, 0x7f, 0x7a, 0x83, 0x59, 0x6d, 0x6e, 0x79, 0x83, 0x88, 0x84,
+ 0x74, 0x85, 0x74, 0x78, 0x80, 0x7c, 0x97, 0x86, 0x94, 0x65, 0x7e, 0x80,
+ 0x6f, 0x97, 0x70, 0x74, 0x92, 0x76, 0x71, 0x91, 0x85, 0x72, 0x6e, 0x84,
+ 0x78, 0x7e, 0x88, 0x79, 0x7f, 0x80, 0x83, 0x7a, 0x85, 0x75, 0x82, 0x81,
+ 0x82, 0x7b, 0x7a, 0xa0, 0x76, 0x7f, 0x75, 0xa7, 0x67, 0x8e, 0x81, 0x98,
+ 0xa5, 0x86, 0x77, 0x78, 0x7f, 0x97, 0x90, 0x86, 0x80, 0x6b, 0x89, 0x66,
+ 0x9b, 0x5c, 0x8b, 0x74, 0xac, 0x89, 0x89, 0x92, 0x92, 0xa8, 0x61, 0x85,
+ 0x8c, 0x86, 0x88, 0x91, 0x92, 0x66, 0x63, 0x6c, 0x7a, 0x80, 0x7d, 0x90,
+ 0x6f, 0x7f, 0x92, 0x94, 0x8e, 0x7a, 0x86, 0x98, 0xa1, 0x59, 0x71, 0x8c,
+ 0x63, 0xa3, 0x60, 0x7d, 0x88, 0x6a, 0x83, 0x6e, 0x7a, 0x94, 0x7b, 0x81,
+ 0x7d, 0x83, 0x77, 0x7e, 0x63, 0xab, 0x75, 0x7b, 0x71, 0x8f, 0x76, 0x6e,
+ 0x78, 0x7b, 0x79, 0x86, 0x69, 0x67, 0x67, 0x70, 0x6c, 0x7a, 0x6c, 0x84,
+ 0x74, 0xa2, 0x74, 0x77, 0x8a, 0x58, 0x7d, 0xa0, 0x65, 0x7b, 0x79, 0x71,
+ 0x7c, 0x3c, 0x85, 0x96, 0x59, 0x76, 0x6a, 0x94, 0xa5, 0x5b, 0x70, 0x99,
+ 0x7f, 0x9a, 0x69, 0x7c, 0x6f, 0x79, 0x72, 0x8b, 0x83, 0x6e, 0x73, 0x7f,
+ 0x6f, 0x6d, 0x7e, 0xa3, 0x72, 0x87, 0x83, 0x8c, 0x8c, 0x70, 0x77, 0x75,
+ 0xa4, 0x5a, 0x89, 0x7d, 0xa0, 0x97, 0x67, 0x80, 0x78, 0x7e, 0x86, 0x6a,
+ 0x7b, 0x9c, 0x77, 0x67, 0x7b, 0x74, 0x7f, 0xa5, 0x90, 0x94, 0x92, 0x4d,
+ 0x7a, 0x79, 0x9f, 0x87, 0x64, 0x6e, 0x6d, 0x59, 0x83, 0x54, 0x79, 0x82,
+ 0x6c, 0x74, 0x82, 0x98, 0x77, 0x90, 0x85, 0xa4, 0x88, 0x81, 0x71, 0x85,
+ 0x90, 0x8e, 0x88, 0x68, 0x51, 0x6d, 0x71, 0x7b, 0x80, 0xbc, 0xa5, 0x57,
+ 0x8f, 0x9f, 0x95, 0x89, 0xb1, 0x96, 0x69, 0x65, 0x61, 0x73, 0x6f, 0x6c,
+ 0x5b, 0x95, 0x99, 0x7f, 0x76, 0x9d, 0x7c, 0x7d, 0x8d, 0xb1, 0x8f, 0x6a,
+ 0x76, 0x95, 0x74, 0x7a, 0x7b, 0xae, 0x77, 0x76, 0x6d, 0x99, 0x7d, 0x80,
+ 0x6e, 0x89, 0x7f, 0x74, 0x6f, 0x72, 0x89, 0x8b, 0x86, 0x7b, 0x7c, 0x72,
+ 0x6b, 0x4f, 0x71, 0x94, 0x80, 0x96, 0x83, 0x7e, 0x75, 0x74, 0x68, 0x83,
+ 0x95, 0x8c, 0x85, 0x7a, 0x82, 0x74, 0x85, 0x83, 0x8c, 0x7e, 0x7a, 0xa0,
+ 0x8e, 0x67, 0x6b, 0x82, 0x9b, 0x66, 0x6c, 0x8a, 0x88, 0x7e, 0x74, 0x9e,
+ 0x88, 0x82, 0x73, 0x73, 0x79, 0x7c, 0x72, 0x6b, 0x74, 0x8b, 0xa4, 0xa4,
+ 0xa3, 0x73, 0x73, 0x88, 0x8d, 0x94, 0x84, 0x9a, 0x9e, 0x93, 0x6c, 0x86,
+ 0x7a, 0x7a, 0x7e, 0xaa, 0x66, 0x8f, 0x99, 0xa4, 0x70, 0x4c, 0x6f, 0x66,
+ 0x8a, 0xaa, 0x69, 0x80, 0x6a, 0x5e, 0x71, 0x8f, 0x8b, 0x84, 0x75, 0x9d,
+ 0x5c, 0x60, 0x61, 0x4a, 0x6f, 0x91, 0x78, 0x6e, 0x8c, 0x62, 0x88, 0x75,
+ 0x64, 0x7c, 0x7d, 0x92, 0x9b, 0x96, 0x62, 0x72, 0x6c, 0x6f, 0x87, 0x5d,
+ 0xa0, 0xa7, 0x7c, 0x58, 0x6e, 0x8c, 0x82, 0x84, 0x7f, 0x8b, 0x54, 0x77,
+ 0x5b, 0x9a, 0x6a, 0x78, 0x5d, 0xb9, 0x8e, 0x7d, 0x6e, 0xa1, 0x66, 0x7c,
+ 0x87, 0xd2, 0x7a, 0x6c, 0x82, 0xa1, 0x83, 0x59, 0x64, 0x9e, 0x65, 0x6d,
+ 0x77, 0x80, 0x7c, 0x9a, 0x50, 0x9f, 0x8b, 0x7a, 0x73, 0x80, 0x92, 0x6d,
+ 0x97, 0x7f, 0x74, 0x6a, 0x5f, 0x44, 0x7d, 0x99, 0x95, 0x91, 0x8f, 0x6a,
+ 0x63, 0x56, 0x89, 0x96, 0xba, 0xa6, 0x71, 0x98, 0x9d, 0x3a, 0x8f, 0x77,
+ 0x6d, 0x76, 0x68, 0xb4, 0x8d, 0x79, 0x7a, 0x83, 0x7f, 0x96, 0x75, 0x94,
+ 0x9e, 0x51, 0x83, 0x5b, 0x66, 0x73, 0xa1, 0xbc, 0x8c, 0x70, 0x88, 0x80,
+ 0x92, 0x60, 0x7d, 0xa9, 0x97, 0x74, 0x7d, 0x98, 0x7b, 0x78, 0x85, 0xa7,
+ 0x8f, 0x8c, 0x91, 0x9d, 0x6a, 0x80, 0x6c, 0x8e, 0x8e, 0x91, 0x76, 0x8b,
+ 0x79, 0x59, 0x7d, 0x9c, 0x69, 0x83, 0x8c, 0x95, 0x8e, 0x75, 0x9d, 0x83,
+ 0x92, 0x99, 0x8a, 0x59, 0x61, 0x54, 0x63, 0x86, 0x83, 0x86, 0x98, 0x83,
+ 0x73, 0x74, 0x91, 0x52, 0x60, 0x8a, 0x7c, 0x57, 0xbc, 0x9d, 0x86, 0x6b,
+ 0x63, 0xa2, 0x78, 0x80, 0x75, 0xb1, 0x74, 0x76, 0x69, 0x8b, 0x7e, 0x76,
+ 0x7b, 0xb3, 0x77, 0x5b, 0x6c, 0x8b, 0x83, 0x80, 0x7f, 0xd1, 0x7c, 0x58,
+ 0x6f, 0x98, 0x71, 0x57, 0x60, 0xd0, 0x84, 0x62, 0x74, 0xa6, 0x8f, 0x7b,
+ 0x70, 0xaa, 0x81, 0x6b, 0x7f, 0x89, 0x6a, 0x74, 0x5a, 0x8c, 0x9c, 0x77,
+ 0x5d, 0x84, 0x63, 0x94, 0x8e, 0x91, 0x83, 0x4a, 0x49, 0x74, 0x6b, 0x70,
+ 0xc0, 0xa0, 0x6a, 0x90, 0x8e, 0x5a, 0x70, 0x96, 0xab, 0x72, 0x7e, 0xba,
+ 0xa7, 0x46, 0x86, 0x5d, 0x90, 0x76, 0x95, 0x8d, 0xa5, 0x40, 0x82, 0x8a,
+ 0x7d, 0x5e, 0x73, 0x94, 0x9d, 0x58, 0x8c, 0x8b, 0x69, 0x6c, 0x9a, 0x90,
+ 0xaa, 0x6f, 0x85, 0x8d, 0x64, 0x58, 0x7b, 0x97, 0xa9, 0x79, 0xa5, 0xa2,
+ 0x5f, 0x57, 0x9a, 0xb4, 0x89, 0x70, 0x84, 0x73, 0x46, 0x6c, 0x6e, 0x87,
+ 0x70, 0x94, 0x8a, 0x8a, 0x69, 0x7b, 0x6c, 0x68, 0x8e, 0xa2, 0x90, 0x84,
+ 0x78, 0x45, 0x63, 0x78, 0x7f, 0x90, 0x9f, 0x90, 0x68, 0x43, 0x92, 0x77,
+ 0x78, 0x77, 0x82, 0x7d, 0x8f, 0x6a, 0x7a, 0x70, 0x76, 0x75, 0x87, 0x63,
+ 0xbc, 0x8e, 0x6a, 0x71, 0x51, 0x51, 0x75, 0x6b, 0x8a, 0xb4, 0x6a, 0x5b,
+ 0x99, 0x84, 0x76, 0x84, 0x74, 0xaf, 0x86, 0x6a, 0x53, 0x97, 0x6e, 0x8e,
+ 0x61, 0xc4, 0x7e, 0x5d, 0x4d, 0x96, 0x73, 0x73, 0x53, 0xc0, 0x8f, 0x68,
+ 0x58, 0xae, 0x81, 0x83, 0x62, 0x98, 0x7b, 0x89, 0x54, 0x86, 0x78, 0x67,
+ 0x70, 0x9b, 0x63, 0x5f, 0x2d, 0x77, 0x84, 0x79, 0x6b, 0xa4, 0x7b, 0x65,
+ 0x45, 0x65, 0x56, 0x86, 0xbb, 0x8a, 0x8e, 0x92, 0x86, 0x48, 0x7c, 0x6d,
+ 0xb4, 0x7d, 0x56, 0xa4, 0x86, 0x52, 0x8b, 0x6a, 0x8d, 0x5b, 0x9d, 0xa2,
+ 0xbf, 0x36, 0x7c, 0x99, 0x9d, 0x65, 0x75, 0xa4, 0x9f, 0x6a, 0x7c, 0x6b,
+ 0x6f, 0x55, 0x70, 0x7f, 0xc2, 0x38, 0x6e, 0xa4, 0x74, 0x4c, 0x75, 0xbb,
+ 0xa4, 0x75, 0x8e, 0x8f, 0x56, 0x65, 0x57, 0x92, 0x73, 0x7f, 0x7d, 0x86,
+ 0x65, 0x76, 0x92, 0x84, 0x70, 0xa8, 0x91, 0x5b, 0x69, 0x74, 0x8e, 0x82,
+ 0x78, 0x8a, 0xaa, 0x71, 0x70, 0x50, 0x85, 0x82, 0x7d, 0x94, 0xa0, 0x76,
+ 0x6d, 0x55, 0x86, 0x79, 0x71, 0x7f, 0x9b, 0x71, 0x8a, 0x42, 0x87, 0x64,
+ 0x57, 0x88, 0xa0, 0x77, 0xa8, 0x91, 0x72, 0x65, 0x7e, 0x6b, 0x7e, 0x81,
+ 0x8d, 0x97, 0x7e, 0x6a, 0x92, 0x88, 0x84, 0x7a, 0x61, 0xa9, 0x86, 0x59,
+ 0x6c, 0x87, 0x61, 0x72, 0x4f, 0xc8, 0x99, 0x6c, 0x66, 0xa3, 0x80, 0x8b,
+ 0x5c, 0xc0, 0x69, 0x7a, 0x6c, 0xb8, 0x8e, 0x91, 0x51, 0x9f, 0x8c, 0x85,
+ 0x75, 0x96, 0x8c, 0x84, 0x6b, 0xa6, 0x71, 0x62, 0x42, 0x60, 0x74, 0x72,
+ 0x92, 0x91, 0x70, 0x5b, 0x3d, 0x71, 0x5e, 0x91, 0xa3, 0xa5, 0x6a, 0x7c,
+ 0x60, 0x58, 0x82, 0x80, 0xa3, 0x73, 0x8f, 0xa0, 0xb2, 0x4b, 0x94, 0x5e,
+ 0x9f, 0x75, 0x4d, 0x83, 0xbc, 0x42, 0x5e, 0x80, 0x8f, 0x59, 0x53, 0xac,
+ 0xb2, 0x45, 0x68, 0x7d, 0x9a, 0x65, 0x8a, 0xaa, 0xa0, 0x4e, 0x77, 0x72,
+ 0x4d, 0x62, 0x6e, 0x98, 0x8c, 0x73, 0x92, 0x5a, 0x49, 0x55, 0x7b, 0x98,
+ 0x8d, 0x84, 0x80, 0x8e, 0x2e, 0x56, 0x78, 0x73, 0x7b, 0x8f, 0x9a, 0x69,
+ 0x73, 0x68, 0x7a, 0x88, 0x78, 0xa5, 0xb1, 0x5c, 0x8f, 0x55, 0x71, 0x99,
+ 0x7a, 0xa9, 0xb0, 0x75, 0x69, 0x44, 0x5f, 0x66, 0x81, 0x7d, 0x9e, 0x4f,
+ 0x66, 0x7f, 0x87, 0x7d, 0x5d, 0x7c, 0x95, 0x62, 0xa5, 0x86, 0x90, 0x6f,
+ 0x60, 0xa5, 0x6e, 0x70, 0x80, 0x96, 0x6f, 0x55, 0x77, 0x87, 0x99, 0x7b,
+ 0x21, 0xaa, 0x7f, 0x60, 0x63, 0xae, 0x47, 0x79, 0x44, 0xb5, 0x83, 0x6e,
+ 0x6d, 0x93, 0x76, 0x54, 0x4b, 0xad, 0x91, 0x6b, 0x6a, 0x9c, 0x8c, 0x83,
+ 0x62, 0x8a, 0x88, 0x71, 0x73, 0xa0, 0x75, 0x95, 0x54, 0x80, 0x92, 0x65,
+ 0x45, 0x80, 0x63, 0x9a, 0x93, 0x9b, 0x78, 0x4e, 0x4d, 0x5f, 0x69, 0x9e,
+ 0xbd, 0xa5, 0x75, 0x6b, 0x6e, 0x6a, 0x82, 0x97, 0xab, 0x60, 0x76, 0xb3,
+ 0xc1, 0x39, 0x82, 0x5b, 0x71, 0x31, 0x7b, 0x9c, 0xb5, 0x4f, 0x75, 0x79,
+ 0x6c, 0x5d, 0x80, 0xa6, 0x9c, 0x53, 0x6f, 0x85, 0x84, 0x5e, 0x7d, 0xb5,
+ 0x95, 0x5f, 0x7c, 0x98, 0x72, 0x7c, 0x67, 0x99, 0xbb, 0x6c, 0x73, 0x66,
+ 0x59, 0x5c, 0x6c, 0x9a, 0x9b, 0x72, 0x9b, 0x5f, 0x4b, 0x51, 0x63, 0x84,
+ 0x74, 0xa0, 0xb3, 0x6e, 0x63, 0xa0, 0x84, 0x90, 0x71, 0x91, 0xba, 0x64,
+ 0x6d, 0x72, 0x78, 0x83, 0x6f, 0x8e, 0xbd, 0x64, 0x69, 0x60, 0x95, 0x67,
+ 0x70, 0x93, 0x78, 0x4d, 0x91, 0x3f, 0x7b, 0x6d, 0x69, 0x87, 0x7d, 0x8a,
+ 0xa3, 0x95, 0x9d, 0x66, 0x6d, 0x8b, 0x7a, 0x75, 0x94, 0x7b, 0x89, 0x52,
+ 0x66, 0x65, 0x79, 0x84, 0x49, 0x9c, 0x60, 0x66, 0x3e, 0xab, 0x4a, 0x86,
+ 0x54, 0xcd, 0x7c, 0x83, 0x7c, 0xac, 0x8b, 0x53, 0x67, 0xbb, 0x7c, 0x6d,
+ 0x72, 0xb3, 0x83, 0x85, 0x4f, 0x97, 0x86, 0x60, 0x7d, 0x93, 0x70, 0x8b,
+ 0x64, 0x78, 0x82, 0x73, 0x54, 0x87, 0x6c, 0xaa, 0x6f, 0x97, 0x8d, 0x51,
+ 0x2d, 0x50, 0x75, 0xa9, 0xc2, 0x94, 0x8d, 0x6f, 0x6d, 0x71, 0x7b, 0x87,
+ 0x93, 0x67, 0x7d, 0xa5, 0xa2, 0x4f, 0x99, 0x83, 0x95, 0x49, 0x70, 0x9c,
+ 0xcf, 0x37, 0x84, 0x86, 0x94, 0x5c, 0x95, 0xa1, 0xb6, 0x73, 0x80, 0x8d,
+ 0x89, 0x62, 0x6f, 0xb4, 0xa1, 0x5b, 0x64, 0x91, 0x41, 0x4f, 0x53, 0xa6,
+ 0xae, 0x75, 0x84, 0x82, 0x58, 0x8e, 0x63, 0x95, 0xa3, 0x8d, 0x8b, 0x76,
+ 0x5d, 0x78, 0x80, 0x82, 0x6e, 0x9d, 0xb8, 0x7d, 0x64, 0x8a, 0x7e, 0x80,
+ 0x72, 0x99, 0xcf, 0x76, 0x66, 0x77, 0x7c, 0x81, 0x71, 0x6f, 0xa1, 0x6c,
+ 0x6b, 0x70, 0x80, 0x7c, 0x6d, 0x83, 0x8e, 0x74, 0x7a, 0x58, 0x69, 0x53,
+ 0x58, 0x7d, 0x7f, 0x84, 0x96, 0x9c, 0x75, 0x6e, 0x62, 0x7c, 0x88, 0x7e,
+ 0x7f, 0x98, 0x93, 0x61, 0x98, 0x98, 0x80, 0x83, 0x2e, 0x7d, 0x64, 0x69,
+ 0x50, 0xa5, 0x38, 0x96, 0x2e, 0xc5, 0x66, 0x56, 0x64, 0xaa, 0x63, 0x64,
+ 0x6d, 0xb3, 0x8a, 0x6c, 0x59, 0xb6, 0x69, 0x7a, 0x54, 0x91, 0x58, 0x96,
+ 0x6b, 0x9f, 0x6d, 0x88, 0x4a, 0x82, 0x94, 0x67, 0x38, 0x93, 0x60, 0x87,
+ 0x8c, 0x93, 0x8c, 0x52, 0x31, 0x43, 0x66, 0xa9, 0xb3, 0x7a, 0x88, 0x64,
+ 0x60, 0x5b, 0x80, 0x84, 0xb7, 0x5a, 0x7a, 0x9d, 0x92, 0x50, 0x89, 0x80,
+ 0x72, 0x51, 0x7f, 0x85, 0xae, 0x47, 0x76, 0x9a, 0x7a, 0x74, 0x6d, 0x93,
+ 0xbd, 0x42, 0x72, 0x6d, 0x58, 0x5e, 0x6e, 0xa4, 0xb5, 0x4e, 0x76, 0x8f,
+ 0x75, 0x9b, 0x5d, 0x92, 0xad, 0x77, 0x7f, 0x73, 0x62, 0x7d, 0x65, 0xaf,
+ 0x98, 0x87, 0x80, 0x7c, 0x61, 0x81, 0x45, 0xa0, 0x84, 0x99, 0xbb, 0x72,
+ 0x86, 0x8f, 0x70, 0x97, 0x6a, 0x8a, 0xd3, 0x70, 0x7c, 0x91, 0x77, 0x82,
+ 0x70, 0x8c, 0xd5, 0x6c, 0x7f, 0x51, 0x5f, 0x69, 0x72, 0x89, 0x9a, 0x68,
+ 0x79, 0x70, 0x8b, 0x80, 0x52, 0x98, 0x86, 0x7a, 0xa0, 0x7b, 0x61, 0x6e,
+ 0x66, 0x6f, 0x77, 0x78, 0x64, 0xac, 0x7e, 0x73, 0x5d, 0x71, 0x6f, 0x80,
+ 0x2e, 0xa9, 0x90, 0x5c, 0x56, 0xa1, 0x32, 0x88, 0x55, 0xb9, 0x67, 0x6f,
+ 0x5c, 0xa5, 0x87, 0x61, 0x6b, 0xbd, 0x77, 0x7c, 0x62, 0xae, 0x7c, 0x7a,
+ 0x66, 0xac, 0x7a, 0x62, 0x5c, 0x9a, 0x58, 0x89, 0x5a, 0x74, 0x72, 0x66,
+ 0x5c, 0x8e, 0x51, 0x8e, 0x99, 0x92, 0xa0, 0x49, 0x31, 0x55, 0x68, 0x99,
+ 0xba, 0x82, 0xa2, 0x7a, 0x5e, 0x6f, 0x84, 0x98, 0x96, 0x52, 0x73, 0x99,
+ 0xb4, 0x5e, 0x7c, 0x59, 0x7d, 0x4a, 0x7e, 0xa0, 0xbe, 0x63, 0x67, 0x8e,
+ 0x7f, 0x71, 0x80, 0xaf, 0x93, 0x4e, 0x78, 0x7e, 0x6d, 0x52, 0x66, 0xb3,
+ 0x94, 0x56, 0x84, 0x8f, 0x50, 0x6d, 0x65, 0xa8, 0xb3, 0x4b, 0x91, 0x7f,
+ 0x4c, 0x8d, 0x69, 0x79, 0x95, 0x8f, 0x8f, 0x7c, 0x66, 0x98, 0x75, 0x9b,
+ 0x73, 0x9b, 0xac, 0x79, 0x6e, 0x84, 0x69, 0x9e, 0x80, 0xa0, 0xb0, 0x6c,
+ 0x46, 0x8b, 0x3f, 0x7a, 0x79, 0x79, 0xb3, 0x62, 0x6b, 0x60, 0x67, 0x81,
+ 0x4a, 0x7e, 0xa7, 0x8c, 0x74, 0x7f, 0x67, 0x4c, 0x4b, 0x8c, 0x8e, 0x67,
+ 0x78, 0x9d, 0x94, 0x79, 0x75, 0x7c, 0x86, 0x7b, 0x67, 0x9f, 0xa4, 0x61,
+ 0x5b, 0x6e, 0x85, 0x70, 0x20, 0xa5, 0x66, 0x5e, 0x55, 0xad, 0x3e, 0x7c,
+ 0x2d, 0xb4, 0x78, 0x6f, 0x4c, 0xc6, 0x7e, 0x6d, 0x54, 0xb4, 0x71, 0x78,
+ 0x54, 0xc3, 0x66, 0x6e, 0x4a, 0xa0, 0x7b, 0x85, 0x66, 0x94, 0x75, 0x8d,
+ 0x34, 0x88, 0x71, 0x4e, 0x49, 0x8a, 0x3b, 0x9c, 0x88, 0x76, 0x7f, 0x6a,
+ 0x37, 0x64, 0x66, 0xb6, 0xa3, 0x82, 0x76, 0x82, 0x6d, 0x65, 0x6f, 0x8c,
+ 0x99, 0x5e, 0x77, 0xa1, 0x99, 0x51, 0xa1, 0x67, 0x6f, 0x4c, 0x7f, 0x9e,
+ 0xad, 0x40, 0x65, 0x82, 0x76, 0x66, 0x72, 0xb5, 0xb2, 0x5b, 0x71, 0x8a,
+ 0x76, 0x74, 0x52, 0xa0, 0x91, 0x37, 0x86, 0x72, 0x6c, 0x75, 0x62, 0xa5,
+ 0xb6, 0x57, 0x75, 0x90, 0x3e, 0x7f, 0x49, 0x9f, 0x8e, 0x92, 0x81, 0x87,
+ 0x69, 0x9e, 0x6b, 0x86, 0x8d, 0xb1, 0x9e, 0x65, 0x6f, 0x93, 0x70, 0x79,
+ 0x7b, 0x87, 0xbe, 0x59, 0x69, 0x7a, 0x56, 0x7a, 0x81, 0x7d, 0xb8, 0x67,
+ 0x67, 0x7f, 0x54, 0x8f, 0x71, 0x85, 0xa0, 0x74, 0x89, 0x5d, 0x67, 0x52,
+ 0x65, 0x96, 0x89, 0x84, 0x81, 0x83, 0x82, 0x9a, 0x85, 0x73, 0x78, 0x62,
+ 0x87, 0x98, 0x75, 0x6a, 0x73, 0x95, 0x86, 0x71, 0x11, 0x9a, 0x91, 0x66,
+ 0x6e, 0xa4, 0x35, 0x89, 0x47, 0xbb, 0x5e, 0x46, 0x3a, 0xa8, 0x70, 0x4a,
+ 0x65, 0xb9, 0x70, 0x96, 0x66, 0xcf, 0x80, 0x79, 0x60, 0xa4, 0x79, 0x70,
+ 0x68, 0x92, 0x7f, 0x89, 0x6b, 0x87, 0x77, 0x67, 0x5b, 0x74, 0x3f, 0x9e,
+ 0x94, 0x9b, 0xa1, 0x61, 0x4b, 0x66, 0x70, 0xad, 0xb7, 0x67, 0x70, 0x6c,
+ 0x3f, 0x5b, 0x94, 0x88, 0xb3, 0x4f, 0x97, 0x97, 0x8c, 0x55, 0xb8, 0x78,
+ 0x60, 0x25, 0x51, 0x91, 0xcd, 0x44, 0x6f, 0x85, 0x5c, 0x65, 0x67, 0xa5,
+ 0x9e, 0x5f, 0x6d, 0x85, 0x6d, 0x56, 0x80, 0xae, 0x79, 0x63, 0x4f, 0x7d,
+ 0x5f, 0x6b, 0x6e, 0xa7, 0x8e, 0x76, 0x8f, 0x90, 0x6e, 0x8c, 0x88, 0x92,
+ 0x81, 0x81, 0x96, 0x7d, 0x48, 0x6b, 0x3f, 0xa1, 0x8c, 0xa2, 0x9f, 0x7f,
+ 0x77, 0x97, 0x73, 0x9c, 0x67, 0x95, 0xae, 0x77, 0x7f, 0x7a, 0x52, 0x7e,
+ 0x91, 0x77, 0xa8, 0x54, 0x6a, 0x74, 0x52, 0x8a, 0x67, 0x8e, 0x90, 0x8d,
+ 0x8b, 0x52, 0x72, 0x5a, 0x73, 0x8f, 0x94, 0x87, 0x7c, 0x88, 0x89, 0x76,
+ 0x77, 0x88, 0x5c, 0x77, 0x8f, 0x94, 0xac, 0x58, 0x70, 0x79, 0x75, 0x8a,
+ 0x20, 0x9c, 0x91, 0x55, 0x55, 0xa4, 0x5b, 0x84, 0x30, 0xc6, 0x8a, 0x51,
+ 0x31, 0xc3, 0x72, 0x6b, 0x65, 0xb9, 0x79, 0x7d, 0x62, 0xad, 0x88, 0x75,
+ 0x37, 0xb0, 0x76, 0x8a, 0x7d, 0x85, 0x7f, 0xb4, 0x46, 0x9c, 0x83, 0x7b,
+ 0x79, 0x78, 0x56, 0xac, 0x8d, 0xa2, 0xa9, 0x54, 0x44, 0x5a, 0x63, 0xb2,
+ 0xa8, 0x72, 0xa4, 0x6b, 0x5d, 0x4d, 0x8e, 0x95, 0x9e, 0x4a, 0x98, 0x8c,
+ 0xb0, 0x5c, 0xa5, 0x75, 0x83, 0x3b, 0x46, 0x92, 0xa7, 0x3b, 0x6a, 0x75,
+ 0x59, 0x57, 0x52, 0xa1, 0xab, 0x54, 0x68, 0x7c, 0x94, 0x6e, 0x5b, 0x9a,
+ 0xa3, 0x5d, 0x73, 0x74, 0x5a, 0x63, 0x56, 0x9e, 0xc1, 0x71, 0x82, 0x79,
+ 0x49, 0x92, 0x63, 0xa6, 0x99, 0x7d, 0x71, 0x81, 0x5e, 0x90, 0x5c, 0x8b,
+ 0x7e, 0xb4, 0xa0, 0x8c, 0x67, 0x93, 0x4e, 0x72, 0x65, 0x83, 0xb5, 0x77,
+ 0x83, 0x92, 0x43, 0x67, 0x8c, 0x81, 0xb1, 0x75, 0x6a, 0x61, 0x66, 0x6f,
+ 0x5d, 0x7f, 0x8d, 0x7b, 0x6b, 0x68, 0x6f, 0x85, 0x6e, 0x87, 0x97, 0x89,
+ 0x9b, 0x81, 0x7e, 0x7e, 0x9d, 0x83, 0x6b, 0x6a, 0xa5, 0x92, 0x7e, 0x70,
+ 0x60, 0x8f, 0x6f, 0x8b, 0x15, 0xa6, 0x66, 0x4e, 0x61, 0xbc, 0x38, 0x67,
+ 0x46, 0xab, 0x84, 0x5e, 0x3a, 0xac, 0x74, 0x58, 0x76, 0xc4, 0x7a, 0x76,
+ 0x67, 0xc0, 0x76, 0x6f, 0x52, 0xa6, 0xa2, 0x97, 0x76, 0xa6, 0x7f, 0x99,
+ 0x5d, 0xa5, 0x5f, 0x60, 0x58, 0x88, 0x3f, 0x9e, 0x7d, 0x81, 0x71, 0x63,
+ 0x42, 0x55, 0x3e, 0xbd, 0xa9, 0x7a, 0xa5, 0x67, 0x62, 0x7a, 0x80, 0x9e,
+ 0xc3, 0x54, 0x7f, 0x9f, 0x93, 0x73, 0xbd, 0x79, 0x74, 0x2e, 0x54, 0x9e,
+ 0xaa, 0x76, 0x68, 0x80, 0x78, 0x64, 0x57, 0x93, 0xa4, 0x56, 0x75, 0x72,
+ 0x81, 0x7f, 0x48, 0xad, 0x89, 0x67, 0x60, 0x7e, 0x7a, 0x83, 0x6e, 0x95,
+ 0xb0, 0x57, 0x89, 0x91, 0x4d, 0x86, 0x78, 0x7b, 0x74, 0x8c, 0x8f, 0x8d,
+ 0x67, 0xa4, 0x64, 0x8d, 0x77, 0x9a, 0xa1, 0x88, 0x6e, 0x94, 0x33, 0x95,
+ 0x81, 0x76, 0xc6, 0x7d, 0x7d, 0x85, 0x5a, 0x6e, 0x8e, 0x69, 0x9e, 0x71,
+ 0x82, 0x81, 0x59, 0x5b, 0x71, 0x9a, 0x91, 0x8e, 0x80, 0x69, 0x71, 0x73,
+ 0x6e, 0x9a, 0x95, 0x94, 0x7b, 0x80, 0x82, 0x7e, 0x76, 0x84, 0x70, 0x72,
+ 0x9c, 0xa0, 0x77, 0x66, 0x55, 0xa1, 0x8c, 0x73, 0x35, 0xa0, 0x68, 0x4d,
+ 0x3b, 0xaa, 0x44, 0x6f, 0x3c, 0xc0, 0x96, 0x78, 0x33, 0xbd, 0x64, 0x5b,
+ 0x75, 0xd2, 0x83, 0x87, 0x59, 0xbd, 0x80, 0x80, 0x6e, 0x8e, 0x65, 0x7a,
+ 0x87, 0xb6, 0x8d, 0x94, 0x39, 0x95, 0x8b, 0x5d, 0x66, 0x71, 0x4e, 0x9f,
+ 0x96, 0x8a, 0x98, 0x47, 0x41, 0x6c, 0x4c, 0xac, 0x95, 0x81, 0x90, 0x75,
+ 0x59, 0x4c, 0xa2, 0x93, 0x99, 0x58, 0x7b, 0xaf, 0xa3, 0x52, 0xb0, 0x6c,
+ 0x5f, 0x47, 0x6e, 0x8e, 0xae, 0x3d, 0x81, 0x6d, 0x78, 0x52, 0x4f, 0x81,
+ 0x80, 0x68, 0x4b, 0x81, 0x74, 0x71, 0x67, 0xa7, 0x9a, 0x55, 0x84, 0x72,
+ 0x64, 0x6b, 0x6e, 0x9d, 0xab, 0x76, 0x79, 0x85, 0x40, 0x84, 0x80, 0x85,
+ 0x70, 0x91, 0x9a, 0x81, 0x5b, 0x89, 0x6b, 0x8a, 0x92, 0x8c, 0xa4, 0x7b,
+ 0x75, 0x89, 0x54, 0x76, 0x69, 0x69, 0xb3, 0x6c, 0x47, 0x7d, 0x4c, 0x7f,
+ 0x81, 0x86, 0x8f, 0x63, 0x71, 0x6a, 0x63, 0x67, 0x7c, 0x8f, 0xa0, 0x68,
+ 0x86, 0x58, 0x5b, 0x87, 0x6a, 0x82, 0x89, 0x78, 0x9d, 0x8d, 0xaa, 0x82,
+ 0x6e, 0xa4, 0x6f, 0x6d, 0x70, 0x9f, 0x7f, 0x77, 0x41, 0xa5, 0x86, 0x61,
+ 0x2d, 0x99, 0xa9, 0x5f, 0x5a, 0xb3, 0x51, 0x70, 0x5a, 0xce, 0x77, 0x68,
+ 0x2c, 0xb8, 0x90, 0x44, 0x58, 0xb9, 0x74, 0x8e, 0x70, 0xb3, 0x9a, 0x75,
+ 0x6d, 0xc0, 0x9e, 0x8e, 0x8d, 0xa8, 0x7b, 0xa8, 0x4a, 0x89, 0x6e, 0x7f,
+ 0x5d, 0x6e, 0x46, 0x91, 0x6d, 0x81, 0x89, 0x3e, 0x35, 0x69, 0x44, 0xaf,
+ 0x99, 0x8d, 0x94, 0x54, 0x60, 0x5b, 0xaf, 0x97, 0x92, 0x4e, 0x80, 0xae,
+ 0x9e, 0x62, 0xa3, 0x77, 0x6e, 0x5d, 0x71, 0xa0, 0xa6, 0x59, 0x84, 0x5d,
+ 0x65, 0x4a, 0x69, 0xa1, 0xa1, 0x40, 0x75, 0x65, 0x6b, 0x68, 0x60, 0xb3,
+ 0x92, 0x27, 0x70, 0x67, 0x9b, 0x5e, 0x50, 0xaf, 0xae, 0x64, 0x7a, 0x6e,
+ 0x61, 0x94, 0x3b, 0x8f, 0x86, 0x7f, 0x98, 0x88, 0x7a, 0x7f, 0x61, 0x7b,
+ 0x64, 0x96, 0x96, 0x79, 0x5c, 0x96, 0x52, 0x92, 0x76, 0x7e, 0xc4, 0x60,
+ 0x6d, 0x7b, 0x41, 0x8c, 0x7b, 0x8e, 0x9a, 0x66, 0x79, 0x95, 0x67, 0x6a,
+ 0x7a, 0x9b, 0xa9, 0x85, 0x6d, 0x66, 0x55, 0x65, 0x76, 0x8b, 0x90, 0x86,
+ 0x88, 0x8b, 0x8f, 0x7e, 0x83, 0x7c, 0x75, 0x5f, 0x78, 0x96, 0x76, 0x47,
+ 0x54, 0x9c, 0x8d, 0x7d, 0x24, 0x9f, 0x79, 0x5c, 0x55, 0xb2, 0x3b, 0x67,
+ 0x4e, 0xd2, 0x90, 0x79, 0x3c, 0xc3, 0x8b, 0x4a, 0x7c, 0xd7, 0x70, 0x75,
+ 0x5b, 0xaf, 0xa8, 0x6b, 0x59, 0xc1, 0x6d, 0x5f, 0x5d, 0x96, 0x87, 0x9a,
+ 0x5d, 0x7f, 0x8e, 0x6d, 0x5c, 0x75, 0x3f, 0xb6, 0x8e, 0x81, 0x7b, 0x31,
+ 0x47, 0x67, 0x56, 0xb6, 0x90, 0x71, 0x89, 0x63, 0x61, 0x75, 0x8d, 0x8b,
+ 0x97, 0x62, 0x62, 0x85, 0x9c, 0x64, 0xb7, 0x61, 0x71, 0x3f, 0x6c, 0x8b,
+ 0xaa, 0x43, 0x82, 0x70, 0x52, 0x52, 0x80, 0xaa, 0x9e, 0x5d, 0x90, 0x69,
+ 0x8a, 0x77, 0x6d, 0x9f, 0x9e, 0x5f, 0x84, 0x61, 0x87, 0x70, 0x43, 0xab,
+ 0x97, 0x6e, 0x84, 0x6c, 0x5d, 0x82, 0x64, 0x85, 0x83, 0x7e, 0x82, 0x7c,
+ 0x7b, 0x91, 0x55, 0x7e, 0x77, 0x88, 0xba, 0x71, 0x6d, 0x7b, 0x71, 0x8a,
+ 0x7f, 0x84, 0xb5, 0x63, 0x4a, 0x9a, 0x3c, 0x70, 0x7a, 0x99, 0xa3, 0x50,
+ 0x84, 0x82, 0x56, 0x4c, 0x74, 0x8e, 0xa3, 0x77, 0x8f, 0x4e, 0x5f, 0x6d,
+ 0x97, 0x89, 0xa0, 0x6b, 0x7c, 0x8c, 0x85, 0x82, 0x8e, 0xa1, 0x89, 0x5b,
+ 0x7f, 0x8b, 0x8f, 0x5e, 0x74, 0x96, 0x8a, 0x7d, 0x15, 0x7b, 0x8f, 0x88,
+ 0x5f, 0xa7, 0x63, 0x5b, 0x39, 0xbd, 0x96, 0x56, 0x4c, 0xb4, 0x7b, 0x53,
+ 0x5a, 0xaf, 0x79, 0x7b, 0x5c, 0xa6, 0xaa, 0x74, 0x5f, 0xa0, 0x76, 0x9e,
+ 0x71, 0x9a, 0x60, 0xa4, 0x33, 0x87, 0x66, 0x66, 0x64, 0x7d, 0x6d, 0xac,
+ 0x9e, 0x8c, 0x78, 0x4f, 0x3d, 0x7b, 0x53, 0xb1, 0x97, 0x8a, 0x96, 0x6e,
+ 0x60, 0x4b, 0xa9, 0x9e, 0x93, 0x6e, 0x93, 0xb7, 0xae, 0x46, 0xb9, 0x60,
+ 0x72, 0x46, 0x80, 0x95, 0xb5, 0x57, 0x82, 0x53, 0x6e, 0x4e, 0x5b, 0xa2,
+ 0x9a, 0x3d, 0x8b, 0x6c, 0x84, 0x65, 0x69, 0xa1, 0x8c, 0x60, 0x83, 0x74,
+ 0x73, 0x53, 0x5d, 0x7e, 0x7f, 0x79, 0x6e, 0x81, 0x89, 0x8f, 0x51, 0x81,
+ 0x99, 0x97, 0x81, 0x8a, 0x87, 0x83, 0x43, 0x90, 0x89, 0x94, 0x93, 0x7a,
+ 0x66, 0x80, 0x82, 0x82, 0x79, 0x85, 0xb0, 0x6b, 0x87, 0x7b, 0x53, 0x89,
+ 0x79, 0x9d, 0xab, 0x6e, 0x82, 0x84, 0x50, 0x8f, 0x7e, 0x74, 0x90, 0x74,
+ 0x6e, 0x65, 0x84, 0x70, 0x82, 0x7a, 0x9e, 0x6d, 0x8f, 0x62, 0xb2, 0x84,
+ 0x78, 0x7e, 0x72, 0x5a, 0x7a, 0x85, 0x8c, 0x4b, 0x70, 0x99, 0x87, 0x78,
+ 0x26, 0x95, 0xb9, 0x77, 0x4d, 0xb6, 0x51, 0x6a, 0x41, 0xbf, 0x76, 0x68,
+ 0x56, 0xb6, 0x80, 0x53, 0x83, 0xaf, 0x87, 0x79, 0x79, 0xb4, 0x89, 0x7d,
+ 0x47, 0x9d, 0xa0, 0x86, 0x89, 0xc3, 0x6d, 0x99, 0x41, 0x89, 0x9a, 0x59,
+ 0x54, 0x83, 0x79, 0x9d, 0x7b, 0x73, 0x88, 0x4a, 0x42, 0x64, 0x7a, 0x9f,
+ 0x7b, 0x6e, 0x71, 0x7b, 0x6a, 0x61, 0xae, 0xa3, 0xa0, 0x68, 0x95, 0x9d,
+ 0x94, 0x49, 0x8b, 0x70, 0x8a, 0x5f, 0x49, 0xbb, 0xa7, 0x4a, 0xa1, 0x59,
+ 0x59, 0x59, 0x6d, 0xa0, 0x9f, 0x50, 0xa0, 0x7b, 0x75, 0x49, 0x5a, 0x8c,
+ 0x84, 0x68, 0x78, 0x57, 0x7a, 0x6e, 0x6b, 0x87, 0x9c, 0x7b, 0x84, 0x83,
+ 0x79, 0x7d, 0x5a, 0x77, 0x77, 0x6f, 0x6f, 0x7c, 0x8f, 0x83, 0x40, 0x62,
+ 0x6a, 0x87, 0xab, 0x74, 0x86, 0x96, 0x7a, 0x7d, 0x7b, 0x81, 0x9a, 0x65,
+ 0x60, 0x82, 0x61, 0x73, 0x71, 0x77, 0xa7, 0x79, 0x87, 0x8c, 0x4e, 0x72,
+ 0x8d, 0x89, 0x94, 0x6d, 0x75, 0x6d, 0x6e, 0x82, 0x7a, 0x8d, 0xa9, 0x77,
+ 0x77, 0x7c, 0x74, 0xa7, 0xb7, 0x67, 0x75, 0x67, 0x7e, 0x9f, 0x73, 0x60,
+ 0x6c, 0x95, 0x7f, 0x62, 0x31, 0x70, 0x85, 0x7a, 0x5f, 0xc0, 0x69, 0x66,
+ 0x71, 0xb0, 0x81, 0x5d, 0x48, 0xc9, 0x86, 0x39, 0x93, 0xa4, 0x8e, 0x7c,
+ 0x5e, 0xbb, 0x98, 0x5c, 0x74, 0x9c, 0x89, 0x6d, 0x74, 0xbd, 0x8e, 0x6e,
+ 0x5f, 0x9a, 0x6d, 0x70, 0x57, 0x9c, 0x58, 0xb7, 0x8e, 0x94, 0xa0, 0x3f,
+ 0x39, 0x75, 0x6f, 0xb4, 0xa2, 0x94, 0xa9, 0x70, 0x61, 0x8a, 0x70, 0x92,
+ 0xa7, 0x7f, 0x7f, 0x8d, 0x7a, 0x73, 0xa1, 0x5f, 0x8a, 0x4a, 0x65, 0xaa,
+ 0x92, 0x6e, 0x98, 0x51, 0x81, 0x47, 0x57, 0xb8, 0x89, 0x50, 0x8a, 0x6d,
+ 0x8b, 0x50, 0x8a, 0x86, 0x9b, 0x7d, 0x5b, 0x4a, 0x68, 0x74, 0x53, 0x9b,
+ 0x94, 0x74, 0x7c, 0x6f, 0x62, 0x86, 0x5b, 0x8f, 0x82, 0x96, 0x6e, 0x7c,
+ 0x80, 0x8f, 0x47, 0x5b, 0x70, 0x95, 0x97, 0x77, 0x8d, 0x8e, 0x69, 0x62,
+ 0x78, 0x8f, 0xbf, 0x5e, 0x76, 0xae, 0x4d, 0x84, 0x73, 0x76, 0xab, 0x6f,
+ 0x7f, 0x8c, 0x4b, 0x7d, 0x96, 0x7d, 0xb3, 0x55, 0x78, 0x8d, 0x76, 0x73,
+ 0x8d, 0x8e, 0x98, 0x6a, 0x91, 0x86, 0x6d, 0x8c, 0x7d, 0x93, 0x97, 0x56,
+ 0x79, 0x8f, 0xa3, 0x7f, 0x7e, 0x82, 0xa0, 0x63, 0x3d, 0x6b, 0x88, 0x5e,
+ 0x61, 0xc0, 0x45, 0x5f, 0x66, 0xb0, 0x6c, 0x6d, 0x29, 0xd5, 0x95, 0x3b,
+ 0x77, 0xaa, 0x62, 0x70, 0x63, 0xce, 0x8c, 0x6e, 0x56, 0xaa, 0x77, 0x6e,
+ 0x90, 0xcc, 0x6d, 0x7e, 0x41, 0x9f, 0x88, 0x4f, 0x5d, 0xb4, 0x4c, 0x9b,
+ 0x80, 0x97, 0x98, 0x59, 0x4c, 0x71, 0x53, 0xb4, 0x90, 0x97, 0x93, 0x90,
+ 0x46, 0x63, 0xa6, 0x87, 0x9d, 0x56, 0x7f, 0xab, 0x8e, 0x68, 0xc6, 0x5d,
+ 0x6e, 0x58, 0x4b, 0x85, 0xa1, 0x70, 0x8a, 0x60, 0x84, 0x44, 0x68, 0x8e,
+ 0x9b, 0x3a, 0x8c, 0x57, 0x91, 0x4c, 0x6b, 0x9c, 0xa7, 0x64, 0x82, 0x5f,
+ 0x68, 0x6d, 0x4d, 0xa1, 0x6c, 0x91, 0x6c, 0x6b, 0x64, 0x97, 0x86, 0x81,
+ 0x8d, 0x8e, 0x80, 0x72, 0x88, 0x96, 0x5d, 0x6e, 0x7c, 0x67, 0x97, 0x69,
+ 0x95, 0x93, 0x61, 0x8b, 0x9b, 0x7d, 0xc8, 0x6f, 0x85, 0x80, 0x67, 0x68,
+ 0x90, 0x6b, 0xcc, 0x7c, 0xa3, 0xa0, 0x58, 0x81, 0x7a, 0x8d, 0x9f, 0x65,
+ 0x81, 0x82, 0x78, 0x6b, 0x85, 0x7b, 0x9b, 0x69, 0x86, 0x6c, 0x83, 0x6c,
+ 0x8e, 0x59, 0xab, 0x56, 0x7c, 0x7f, 0x7b, 0x84, 0x71, 0x63, 0x7d, 0x73,
+ 0x60, 0x8b, 0x7a, 0x7b, 0x5e, 0xbb, 0x4b, 0x40, 0x30, 0xcc, 0x80, 0x65,
+ 0x6c, 0xb7, 0x80, 0x35, 0x7d, 0xa3, 0x5c, 0x6c, 0x49, 0xa6, 0x9b, 0x7b,
+ 0x53, 0xba, 0x62, 0x76, 0x78, 0xa0, 0x72, 0x80, 0x78, 0x93, 0x87, 0x62,
+ 0x64, 0x84, 0x6f, 0xa1, 0x70, 0x90, 0x9a, 0x6b, 0x42, 0x55, 0x6d, 0xc5,
+ 0xa6, 0x8a, 0x79, 0x64, 0x4c, 0x72, 0x7b, 0xa9, 0xa3, 0x70, 0x84, 0x8f,
+ 0x63, 0x7a, 0x9c, 0x4e, 0x5a, 0x76, 0x91, 0x67, 0xaf, 0x76, 0xbf, 0x46,
+ 0x62, 0x3f, 0x7d, 0xa7, 0x8d, 0x62, 0x90, 0x5b, 0x9a, 0x44, 0x51, 0x80,
+ 0xa6, 0x7e, 0x8d, 0x6a, 0x73, 0x65, 0x72, 0x82, 0x99, 0xb4, 0x6a, 0x75,
+ 0x85, 0x90, 0x47, 0x62, 0x9e, 0x95, 0x94, 0x78, 0x89, 0x74, 0x5d, 0xa3,
+ 0x7f, 0x9d, 0x7d, 0x63, 0x96, 0x86, 0x8d, 0xa2, 0x95, 0xab, 0xae, 0x5d,
+ 0x93, 0x8d, 0x3d, 0x76, 0x9e, 0x9c, 0xc4, 0x71, 0x7d, 0xa3, 0x75, 0x7e,
+ 0x6d, 0x9d, 0xa3, 0x7f, 0x94, 0x89, 0x47, 0x71, 0x8b, 0x95, 0xb1, 0x72,
+ 0x90, 0x53, 0x7e, 0x8f, 0x8c, 0x90, 0xa1, 0x4d, 0x59, 0x62, 0x73, 0xa0,
+ 0x69, 0x88, 0x86, 0x71, 0x60, 0x3b, 0x81, 0x57, 0x7d, 0x86, 0x58, 0x63,
+ 0x7d, 0x98, 0x74, 0x67, 0x5d, 0xb0, 0x67, 0x45, 0x9b, 0xa9, 0x94, 0x68,
+ 0x43, 0x8b, 0x85, 0x56, 0x63, 0x96, 0x87, 0x78, 0x88, 0xbf, 0x92, 0x8d,
+ 0x60, 0xa8, 0x7e, 0x7e, 0x78, 0x80, 0x66, 0x92, 0x6e, 0x97, 0xab, 0x7f,
+ 0x4f, 0x65, 0x59, 0xb0, 0x9b, 0x6b, 0x9f, 0x70, 0x6f, 0x5c, 0xac, 0x95,
+ 0xa3, 0x54, 0x8e, 0xa9, 0x9e, 0x8c, 0xa5, 0x66, 0x5f, 0x5b, 0x6c, 0x83,
+ 0x90, 0x73, 0x85, 0x64, 0x61, 0x51, 0x4a, 0x63, 0xa1, 0x96, 0x7e, 0x4e,
+ 0x87, 0x60, 0x68, 0xb5, 0x9a, 0x8d, 0x75, 0x4e, 0x8a, 0x7a, 0x5f, 0x9f,
+ 0x74, 0x80, 0x69, 0x6d, 0x73, 0x92, 0x79, 0x7e, 0x85, 0x68, 0x83, 0x9d,
+ 0xb6, 0x9d, 0x6e, 0x8f, 0x78, 0x91, 0xaf, 0x8f, 0xa0, 0x9d, 0x73, 0x55,
+ 0x91, 0x8f, 0xb2, 0x76, 0x97, 0xab, 0x63, 0x63, 0x68, 0x7b, 0xab, 0x5c,
+ 0x77, 0xae, 0x4c, 0x72, 0x6e, 0x93, 0xb8, 0x51, 0x79, 0x84, 0x7d, 0x6b,
+ 0x7f, 0x8a, 0xba, 0x68, 0x7a, 0x43, 0x9a, 0x8d, 0x77, 0x8a, 0x6d, 0x56,
+ 0x79, 0x41, 0x7a, 0x4b, 0x81, 0x7a, 0x5c, 0x68, 0x58, 0x36, 0x6f, 0x6f,
+ 0x9f, 0xa6, 0x5f, 0x60, 0x4e, 0x67, 0x70, 0x4c, 0x69, 0x69, 0x94, 0x63,
+ 0x6d, 0x7b, 0x88, 0x9e, 0x6d, 0x98, 0x69, 0x68, 0x88, 0x80, 0x80, 0x7a,
+ 0x8e, 0x78, 0x5e, 0x8d, 0x7e, 0x91, 0x76, 0x64, 0x7e, 0x7f, 0x4e, 0xc9,
+ 0x79, 0x8f, 0x9c, 0x82, 0x3d, 0x62, 0x63, 0xc3, 0xb8, 0x7b, 0x72, 0x7b,
+ 0x50, 0x56, 0x95, 0x72, 0x8f, 0x6b, 0x90, 0x9d, 0x76, 0xa4, 0xa5, 0x79,
+ 0x54, 0x4f, 0x59, 0x85, 0xc5, 0x92, 0x97, 0x4d, 0x6f, 0x69, 0x77, 0x7f,
+ 0x71, 0x7c, 0x87, 0x59, 0x98, 0x61, 0x80, 0x81, 0x88, 0x6b, 0x6d, 0x7f,
+ 0x7f, 0x77, 0x60, 0xa2, 0x96, 0x73, 0x69, 0x86, 0x83, 0x8d, 0x60, 0x66,
+ 0x88, 0x8c, 0x93, 0x67, 0x98, 0x82, 0x7e, 0x91, 0x99, 0x59, 0x8e, 0x6e,
+ 0x90, 0xa1, 0x62, 0x8a, 0x98, 0x7b, 0xc8, 0x67, 0x85, 0x8d, 0x6c, 0xa1,
+ 0xa1, 0x92, 0xd0, 0x49, 0x85, 0x76, 0x89, 0x75, 0x88, 0x83, 0xa3, 0x77,
+ 0x85, 0x68, 0x82, 0x83, 0x7f, 0x79, 0xae, 0x85, 0x76, 0x84, 0x80, 0x9a,
+ 0x9d, 0x7b, 0x83, 0x90, 0x79, 0x88, 0x79, 0x9a, 0x93, 0x6c, 0x69, 0x79,
+ 0x5f, 0x90, 0x81, 0x7b, 0x87, 0x9d, 0x86, 0x82, 0x7a, 0x77, 0x71, 0x85,
+ 0x8b, 0x99, 0x8f, 0x7b, 0x58, 0x98, 0x84, 0x6e, 0x9a, 0xa1, 0x7a, 0x8c,
+ 0x77, 0xa8, 0x86, 0x93, 0x7b, 0x90, 0x79, 0x8a, 0x85, 0x8f, 0x84, 0x97,
+ 0x73, 0x83, 0x7b, 0x76, 0x8e, 0xa1, 0x89, 0x8a, 0x83, 0x9c, 0x65, 0x68,
+ 0x7b, 0x89, 0x92, 0x84, 0x6d, 0x90, 0x61, 0x78, 0x98, 0x8c, 0x8d, 0x87,
+ 0xa0, 0x99, 0x79, 0x7b, 0x69, 0xa4, 0x7a, 0x8d, 0x73, 0x71, 0x70, 0x80,
+ 0x82, 0x77, 0x81, 0x67, 0x75, 0x97, 0x71, 0x73, 0x85, 0x6d, 0x8e, 0x86,
+ 0x6e, 0x80, 0x86, 0x9e, 0x6f, 0x70, 0x67, 0x59, 0x65, 0x89, 0x67, 0x8b,
+ 0x7d, 0x68, 0x69, 0x7a, 0x5b, 0x7e, 0x87, 0xa1, 0x92, 0x7b, 0x64, 0x7e,
+ 0x76, 0x72, 0x71, 0xab, 0x7c, 0x83, 0x6f, 0xa1, 0x86, 0x76, 0x71, 0x6f,
+ 0x91, 0x77, 0x6c, 0x71, 0x92, 0x78, 0x70, 0x7f, 0x6e, 0x65, 0x77, 0x93,
+ 0x7e, 0x6c, 0x85, 0x9d, 0x78, 0x8b, 0x7c, 0x5f, 0x94, 0x86, 0x7c, 0x7f,
+ 0x83, 0x6e, 0x72, 0x9e, 0x6e, 0x6b, 0x8d, 0x91, 0x97, 0x8b, 0x7b, 0x72,
+ 0x86, 0x75, 0x7f, 0x96, 0x7d, 0x81, 0xa1, 0x55, 0xa6, 0x88, 0x96, 0x87,
+ 0x93, 0x68, 0x89, 0x72, 0x6f, 0x9c, 0x75, 0x7c, 0x79, 0x6c, 0x74, 0x84,
+ 0x7d, 0xa4, 0x86, 0x84, 0x84, 0x8d, 0x63, 0x7a, 0x63, 0xbc, 0x7e, 0x93,
+ 0x80, 0x8d, 0x71, 0x7a, 0x5f, 0x8c, 0x74, 0x96, 0x7e, 0x9b, 0x9d, 0x8d,
+ 0x5b, 0xa4, 0x71, 0x5e, 0x83, 0x78, 0x86, 0x7f, 0x70, 0x99, 0x87, 0x85,
+ 0x8e, 0x81, 0x93, 0x80, 0x89, 0xa0, 0x7a, 0x77, 0x8e, 0x73, 0x5f, 0x80,
+ 0x6d, 0x87, 0x5b, 0x7a, 0x85, 0x7c, 0x85, 0x63, 0x61, 0x9d, 0x6f, 0x68,
+ 0x77, 0x86, 0x61, 0x6d, 0x84, 0x98, 0x7c, 0x78, 0x69, 0x84, 0x91, 0x6d,
+ 0x81, 0xa1, 0x6c, 0x62, 0x95, 0x6d, 0x86, 0x8b, 0x95, 0x8f, 0x5e, 0x86,
+ 0x73, 0xa1, 0x83, 0x58, 0x5f, 0x8e, 0x76, 0x79, 0x9e, 0x92, 0x7c, 0x7b,
+ 0x81, 0x8b, 0x83, 0x7b, 0x78, 0x75, 0x70, 0x83, 0x70, 0x5a, 0x6a, 0x59,
+ 0xa3, 0x82, 0x7a, 0x91, 0x8b, 0x6e, 0x82, 0x8e, 0x70, 0x73, 0x91, 0x76,
+ 0xa5, 0x7f, 0x70, 0x81, 0x6f, 0x85, 0x94, 0xa6, 0x8c, 0x50, 0x76, 0x6e,
+ 0x64, 0x95, 0xa0, 0x64, 0x6c, 0x68, 0x8e, 0x8b, 0xa1, 0x7d, 0xa0, 0x7f,
+ 0x76, 0x8b, 0x7b, 0x93, 0x7b, 0x6e, 0x7e, 0x64, 0x8a, 0xa7, 0x78, 0x64,
+ 0x93, 0x67, 0x7d, 0x68, 0x5c, 0xa0, 0x76, 0x98, 0xaf, 0x80, 0x55, 0x96,
+ 0x97, 0x9c, 0x78, 0x75, 0x87, 0x85, 0x77, 0x77, 0x62, 0x93, 0x76, 0x68,
+ 0xa0, 0x80, 0x81, 0x7f, 0x9a, 0x68, 0x74, 0x69, 0x94, 0x77, 0x77, 0x72,
+ 0x90, 0x9a, 0x6f, 0x95, 0x89, 0x6b, 0x6b, 0x94, 0x7e, 0x9c, 0x6f, 0x67,
+ 0x8f, 0x82, 0x80, 0x92, 0x76, 0x80, 0x65, 0x9b, 0x6a, 0x7c, 0x75, 0x5a,
+ 0x87, 0xa1, 0x69, 0x7a, 0x79, 0x9e, 0x9a, 0x58, 0x81, 0x92, 0x72, 0x67,
+ 0x90, 0x80, 0x82, 0x61, 0x9f, 0x9e, 0x6a, 0x8d, 0x8d, 0x8a, 0x73, 0x81,
+ 0x68, 0x7f, 0x5b, 0x59, 0x98, 0x89, 0x71, 0x72, 0x58, 0x7b, 0x94, 0x5d,
+ 0xa9, 0x8b, 0x72, 0x7b, 0x65, 0x73, 0x5b, 0x8b, 0x7d, 0x86, 0x6e, 0x8c,
+ 0x66, 0x6f, 0x6b, 0x8b, 0x71, 0x80, 0x7f, 0x70, 0x70, 0x88, 0x70, 0x7e,
+ 0x84, 0x89, 0x7f, 0x81, 0x87, 0x77, 0x71, 0x88, 0x7f, 0x8f, 0x5e, 0x80,
+ 0x5d, 0xa1, 0x89, 0x77, 0x93, 0x8e, 0x55, 0x64, 0x88, 0x9a, 0x8b, 0x80,
+ 0x77, 0x6f, 0x91, 0x83, 0x6b, 0x9b, 0x85, 0x5c, 0x57, 0x7e, 0xa9, 0x63,
+ 0x83, 0xaa, 0x7c, 0xa1, 0x91, 0x5f, 0x68, 0x76, 0x7a, 0x97, 0x96, 0x84,
+ 0xca, 0x8d, 0x8c, 0x8b, 0x71, 0x81, 0x88, 0x92, 0xaa, 0x74, 0x49, 0x7a,
+ 0x90, 0x93, 0x7a, 0x61, 0x8c, 0x66, 0x71, 0xa0, 0xab, 0x7d, 0x86, 0x6c,
+ 0x9f, 0x77, 0x67, 0x6a, 0x89, 0x89, 0x88, 0x70, 0xad, 0x88, 0x69, 0x84,
+ 0x70, 0x8f, 0x79, 0x7c, 0x66, 0xa6, 0x71, 0x8d, 0x77, 0x99, 0x69, 0x76,
+ 0x79, 0x7d, 0x9c, 0x6f, 0x64, 0x8b, 0x70, 0x82, 0x69, 0xa4, 0x65, 0x6e,
+ 0x7f, 0x9e, 0x7e, 0x84, 0x8c, 0x9c, 0x6c, 0x5b, 0x6e, 0xa7, 0x6d, 0x7a,
+ 0x92, 0x78, 0x9a, 0x6f, 0x81, 0x91, 0x71, 0x7d, 0x6b, 0x99, 0x6b, 0x92,
+ 0x5e, 0x7e, 0x64, 0x95, 0x78, 0x90, 0x6f, 0x68, 0x8a, 0x85, 0x6f, 0x88,
+ 0x64, 0x66, 0x7f, 0x78, 0x7c, 0x95, 0x66, 0x6c, 0x76, 0x6a, 0x9b, 0x8f,
+ 0x9d, 0x78, 0x86, 0x95, 0x73, 0x66, 0x6d, 0x71, 0x8b, 0x7f, 0x6f, 0x70,
+ 0x64, 0x94, 0xa0, 0x83, 0x6b, 0x6d, 0x85, 0x89, 0x68, 0x92, 0x8e, 0x51,
+ 0x81, 0x85, 0x86, 0x6e, 0x83, 0x85, 0x8a, 0x5e, 0x68, 0xbf, 0xc4, 0xa5,
+ 0x8b, 0x67, 0x86, 0x59, 0x85, 0x9e, 0x96, 0x67, 0x82, 0x7c, 0x6c, 0x80,
+ 0x84, 0xae, 0x9d, 0x80, 0xc2, 0x58, 0x5d, 0x95, 0x85, 0x8b, 0x7f, 0x5d,
+ 0xc7, 0x75, 0x75, 0x87, 0xa2, 0x8c, 0x62, 0x71, 0x9c, 0x61, 0x7f, 0x9c,
+ 0xca, 0x8d, 0x89, 0x6e, 0x7c, 0x71, 0x81, 0x99, 0x95, 0xa4, 0x76, 0x6f,
+ 0x64, 0x7b, 0x6c, 0x72, 0x8b, 0x83, 0x70, 0x70, 0x8b, 0xa4, 0x69, 0x76,
+ 0x6e, 0x8d, 0x7a, 0x80, 0x8f, 0x9e, 0x73, 0x4b, 0x75, 0x78, 0x77, 0x7b,
+ 0x8e, 0x92, 0x88, 0x49, 0x54, 0x9f, 0x7a, 0x7f, 0x68, 0x9f, 0x7f, 0x57,
+ 0x6b, 0xad, 0x85, 0x6f, 0x81, 0xa1, 0x96, 0x6f, 0x73, 0x8d, 0x5e, 0x65,
+ 0x7a, 0x8c, 0x7c, 0x6a, 0x7e, 0x7a, 0x6a, 0x97, 0x59, 0x86, 0x62, 0x77,
+ 0x70, 0x7a, 0x68, 0x62, 0x68, 0x86, 0x7e, 0x76, 0x9a, 0x7f, 0x6c, 0x7e,
+ 0x8a, 0x76, 0x65, 0x8f, 0x7d, 0x65, 0x76, 0xa4, 0x95, 0x62, 0x78, 0x97,
+ 0x7a, 0x6e, 0x7a, 0x7a, 0x7e, 0x91, 0x8c, 0x8a, 0x91, 0x82, 0x89, 0x6d,
+ 0x87, 0x90, 0x69, 0x71, 0x96, 0xa6, 0x7c, 0x7c, 0xa8, 0xa8, 0x62, 0x77,
+ 0x76, 0x99, 0xdd, 0x76, 0x8a, 0x5c, 0x86, 0x6a, 0x69, 0x9c, 0xa5, 0x7d,
+ 0x78, 0x6a, 0x88, 0x77, 0x77, 0xae, 0x8a, 0x99, 0xcb, 0x85, 0x59, 0x84,
+ 0x7b, 0x97, 0x8a, 0x82, 0xc5, 0x65, 0x8c, 0x93, 0xc3, 0x8c, 0x87, 0x64,
+ 0x91, 0x41, 0x70, 0xa8, 0xd1, 0x8b, 0x82, 0x71, 0x9c, 0x71, 0x4e, 0x86,
+ 0x98, 0x86, 0x7f, 0x7e, 0x69, 0x99, 0x79, 0x78, 0x77, 0xb3, 0x6b, 0x80,
+ 0x84, 0x8b, 0x56, 0x73, 0x84, 0x95, 0x82, 0x94, 0x5b, 0x92, 0x83, 0x46,
+ 0x66, 0x89, 0x6d, 0x61, 0x99, 0xa6, 0x99, 0x3f, 0x6c, 0xab, 0x5d, 0x5f,
+ 0x6c, 0x8e, 0x6b, 0x4a, 0x72, 0xb6, 0x6c, 0x75, 0x78, 0xa6, 0x6f, 0x5b,
+ 0x56, 0x8b, 0x57, 0x74, 0x8f, 0xab, 0x53, 0x56, 0x5d, 0x63, 0x63, 0x8b,
+ 0x65, 0x78, 0x71, 0x67, 0x7a, 0x62, 0x8d, 0x78, 0x99, 0x76, 0x94, 0x7a,
+ 0xa3, 0x70, 0x55, 0x87, 0x7e, 0x7c, 0x57, 0x57, 0x6e, 0x79, 0x94, 0x8f,
+ 0x86, 0x80, 0x90, 0x7d, 0x7d, 0x7f, 0x7f, 0x68, 0x41, 0x86, 0x8c, 0x6f,
+ 0x8a, 0x7f, 0x87, 0x8a, 0x7e, 0x7f, 0x5d, 0x71, 0x91, 0x81, 0x93, 0x71,
+ 0x91, 0xc6, 0x70, 0x4a, 0x74, 0xa8, 0xf3, 0x72, 0xa7, 0x80, 0x7e, 0x41,
+ 0x84, 0xa3, 0xb6, 0x94, 0xba, 0x84, 0x70, 0x74, 0x71, 0xac, 0x9f, 0x9d,
+ 0xe4, 0x67, 0x6a, 0x87, 0x92, 0x8e, 0x92, 0x82, 0xdb, 0x5e, 0x9b, 0x90,
+ 0xd5, 0x87, 0x8d, 0x7c, 0x9c, 0x3c, 0x6c, 0xab, 0xc2, 0x86, 0x83, 0x79,
+ 0x6c, 0x61, 0x51, 0xa9, 0x99, 0x79, 0x72, 0x80, 0x6f, 0x85, 0x57, 0x6c,
+ 0x81, 0x86, 0x6e, 0x88, 0x87, 0x8d, 0x8e, 0x81, 0x67, 0x88, 0x62, 0x99,
+ 0x87, 0xab, 0x8f, 0x57, 0x60, 0x77, 0x64, 0x81, 0x96, 0xa3, 0x81, 0x3d,
+ 0x4e, 0xb9, 0x57, 0x6e, 0x99, 0xad, 0x6a, 0x3e, 0x74, 0x96, 0x7e, 0x79,
+ 0x65, 0xa4, 0x7c, 0x6a, 0x53, 0x87, 0x56, 0x6f, 0x5e, 0x97, 0x85, 0x42,
+ 0x56, 0x6b, 0x67, 0x78, 0x7d, 0xa6, 0x7c, 0x7c, 0x7d, 0x78, 0x7b, 0x84,
+ 0x99, 0x7b, 0x89, 0x71, 0x76, 0x8b, 0x76, 0x73, 0x7d, 0x83, 0x56, 0x4f,
+ 0x86, 0x72, 0x83, 0x88, 0x6a, 0x93, 0x69, 0x90, 0x6c, 0x73, 0x6f, 0x63,
+ 0x55, 0x88, 0x6b, 0x88, 0x7c, 0x86, 0x87, 0x7b, 0x6c, 0x7e, 0x60, 0x57,
+ 0xa8, 0x81, 0xa3, 0x72, 0xba, 0xbf, 0x66, 0x65, 0x70, 0xb9, 0xe4, 0x78,
+ 0x99, 0x67, 0x8c, 0x72, 0x88, 0x96, 0xb5, 0x72, 0x8a, 0x66, 0x81, 0x39,
+ 0x85, 0x93, 0xa0, 0x9c, 0xdf, 0x74, 0x8a, 0x6d, 0x93, 0xa1, 0x8c, 0x7a,
+ 0xb5, 0x4b, 0x89, 0xae, 0xba, 0x9c, 0x96, 0x9a, 0xb4, 0x33, 0x5a, 0xb1,
+ 0xcd, 0x88, 0x84, 0x63, 0x8c, 0x5e, 0x71, 0x6d, 0xa7, 0x8a, 0x62, 0x85,
+ 0x77, 0x75, 0x62, 0x79, 0x96, 0x73, 0x4f, 0x7d, 0x93, 0x8a, 0x88, 0x7e,
+ 0x59, 0x6c, 0x7f, 0x87, 0x6f, 0x91, 0x88, 0x59, 0x6d, 0x83, 0x70, 0x7c,
+ 0x7f, 0x8d, 0x7f, 0x26, 0x41, 0xcf, 0x6b, 0x6e, 0x75, 0xa3, 0x90, 0x5e,
+ 0x3a, 0x94, 0x61, 0x9a, 0x6f, 0x9f, 0x69, 0x7d, 0x55, 0x8c, 0x60, 0x7c,
+ 0x93, 0x85, 0x85, 0x4b, 0x54, 0x71, 0x60, 0x8a, 0x6d, 0x8c, 0x9c, 0x7e,
+ 0x5b, 0x79, 0x74, 0x7b, 0x7b, 0x9d, 0x5b, 0x65, 0x81, 0x82, 0x66, 0x89,
+ 0x82, 0x72, 0x77, 0x78, 0x75, 0x76, 0x6b, 0x74, 0x89, 0x73, 0x6c, 0x6b,
+ 0x77, 0x7e, 0x67, 0x84, 0x41, 0x90, 0x58, 0x87, 0x98, 0x60, 0x96, 0x81,
+ 0x6b, 0x74, 0x7d, 0x56, 0x72, 0x71, 0x9a, 0x7d, 0xc5, 0xd0, 0x88, 0x6e,
+ 0x4d, 0xbe, 0xef, 0x8a, 0xa7, 0x92, 0x82, 0x67, 0x7f, 0x91, 0xc5, 0x7d,
+ 0xad, 0x77, 0x6b, 0x4e, 0x8e, 0x99, 0x9b, 0x8e, 0xc7, 0x7f, 0x8a, 0x8e,
+ 0x8f, 0x87, 0x9c, 0x75, 0xb0, 0x53, 0x75, 0x97, 0xc7, 0x98, 0xa4, 0xa4,
+ 0x80, 0x41, 0x79, 0xc3, 0xdb, 0x86, 0x9d, 0x75, 0x7f, 0x67, 0x7a, 0x96,
+ 0xc3, 0x83, 0x54, 0x8e, 0x6f, 0xa8, 0x7c, 0x65, 0x78, 0x7e, 0x59, 0xa3,
+ 0x8a, 0x97, 0x8b, 0x82, 0x5e, 0x66, 0x82, 0x9b, 0x9e, 0x9f, 0x70, 0x49,
+ 0x55, 0x88, 0x8a, 0x7e, 0x90, 0xa7, 0x6b, 0x3b, 0x28, 0xc0, 0x63, 0x7e,
+ 0x60, 0x90, 0x7c, 0x3f, 0x54, 0x9c, 0x7d, 0x8a, 0x6a, 0xa9, 0x6f, 0x61,
+ 0x76, 0x86, 0x64, 0x88, 0x72, 0xa5, 0x6b, 0x4d, 0x56, 0x6c, 0x52, 0xa1,
+ 0x84, 0x69, 0x69, 0x5b, 0x71, 0x84, 0x76, 0x9b, 0x92, 0x70, 0x86, 0x8b,
+ 0x71, 0x68, 0x56, 0x92, 0x76, 0x8f, 0x8f, 0x72, 0x5a, 0x77, 0x6f, 0x92,
+ 0x72, 0x72, 0x5e, 0x7a, 0x70, 0x73, 0x60, 0x7d, 0x5a, 0x93, 0x7f, 0x6b,
+ 0x89, 0x6b, 0xa1, 0x85, 0x5c, 0x8d, 0x76, 0x7c, 0x6f, 0x73, 0x96, 0x6d,
+ 0xbb, 0xad, 0x53, 0x53, 0x5f, 0x9a, 0xe2, 0x8d, 0xa7, 0x6d, 0x8a, 0x5b,
+ 0x85, 0x9c, 0xb4, 0x7b, 0xb3, 0x52, 0x75, 0x7f, 0x7a, 0x8c, 0x91, 0x7e,
+ 0xca, 0x5f, 0x64, 0x71, 0x85, 0x9a, 0x91, 0x72, 0xbd, 0x6e, 0x9b, 0x81,
+ 0x8f, 0xa8, 0xac, 0x7d, 0xb4, 0x5f, 0x45, 0xc5, 0xc8, 0x7a, 0x93, 0x8e,
+ 0x7b, 0x41, 0x69, 0x94, 0x8b, 0x76, 0x59, 0x81, 0x73, 0x92, 0x8e, 0x63,
+ 0x8e, 0x74, 0x33, 0xa5, 0x9c, 0xa2, 0x88, 0x48, 0x5d, 0x8c, 0x7d, 0xa6,
+ 0x68, 0x9a, 0x6f, 0x58, 0x6c, 0x8f, 0x77, 0x65, 0x97, 0x9d, 0x7a, 0x37,
+ 0x59, 0xab, 0x6e, 0x8f, 0x7a, 0xae, 0x65, 0x3e, 0x46, 0xa9, 0x82, 0x82,
+ 0x9c, 0x9d, 0x62, 0x79, 0x66, 0x7f, 0x5e, 0x88, 0x9e, 0x8f, 0x84, 0x71,
+ 0x5d, 0x6d, 0x70, 0xa0, 0x69, 0x92, 0x7f, 0x70, 0x66, 0x6f, 0x75, 0x8c,
+ 0x96, 0x7a, 0x85, 0x6a, 0x5a, 0x7c, 0x72, 0x8a, 0x8d, 0x7b, 0x8b, 0x5c,
+ 0x76, 0x69, 0x70, 0x7f, 0x74, 0xa1, 0x71, 0x91, 0x5a, 0x8c, 0x6e, 0x83,
+ 0x52, 0x78, 0x71, 0x6d, 0xa9, 0x63, 0x9d, 0x81, 0x52, 0x9e, 0x5d, 0x60,
+ 0x76, 0x93, 0x97, 0x67, 0xce, 0xc1, 0x75, 0x5e, 0x5f, 0x8c, 0xea, 0x76,
+ 0xad, 0x7a, 0x7d, 0x62, 0x85, 0x92, 0xd0, 0x6a, 0xbc, 0x53, 0x55, 0x5c,
+ 0x6d, 0x89, 0x9e, 0x71, 0xd2, 0x8b, 0x64, 0x61, 0x85, 0x9a, 0x77, 0x75,
+ 0xb9, 0x67, 0x8a, 0xac, 0x90, 0x8a, 0xb4, 0x91, 0xbb, 0x58, 0x94, 0xaf,
+ 0xb2, 0x76, 0xa2, 0x71, 0x95, 0x5e, 0x73, 0xa5, 0x92, 0x8c, 0x52, 0x96,
+ 0x53, 0x95, 0x84, 0x91, 0x93, 0x7a, 0x40, 0x88, 0xab, 0xa5, 0x63, 0x70,
+ 0x66, 0x88, 0x7e, 0x92, 0x89, 0x84, 0x78, 0x57, 0x3d, 0x8d, 0x84, 0x77,
+ 0x9b, 0x87, 0x5e, 0x4e, 0x42, 0xa0, 0x76, 0x8a, 0x77, 0x90, 0x83, 0x4c,
+ 0x42, 0x9b, 0x75, 0x7a, 0x88, 0x94, 0x98, 0x69, 0x4c, 0xa2, 0x6b, 0x7b,
+ 0x6e, 0x9b, 0x5d, 0x5f, 0x53, 0x6a, 0x63, 0x95, 0x69, 0x8a, 0x61, 0x75,
+ 0x6c, 0x7a, 0x58, 0x89, 0x84, 0x8f, 0x6b, 0x5a, 0x71, 0x6f, 0x59, 0x89,
+ 0x7d, 0x87, 0x5f, 0x77, 0x4b, 0x61, 0x77, 0x92, 0x67, 0x8e, 0x5c, 0x6f,
+ 0x5b, 0x77, 0x76, 0x6b, 0x44, 0x9d, 0x9f, 0x7f, 0x8b, 0x94, 0x9e, 0x7c,
+ 0x62, 0x94, 0x60, 0x55, 0x77, 0x8f, 0xa6, 0x62, 0xb5, 0xb2, 0x3c, 0x61,
+ 0x5c, 0x99, 0xeb, 0x5b, 0x90, 0x6c, 0x7f, 0x5f, 0x75, 0xa6, 0xcf, 0x77,
+ 0x98, 0x5d, 0x75, 0x69, 0x7f, 0x8a, 0xa7, 0x73, 0xc8, 0x74, 0x70, 0x82,
+ 0x76, 0x8f, 0xa2, 0x7a, 0xa4, 0x7a, 0x66, 0x81, 0x9b, 0x8f, 0x9e, 0x8b,
+ 0xa1, 0x51, 0x7b, 0xba, 0xc8, 0x90, 0xab, 0x92, 0x72, 0x57, 0x5b, 0xa3,
+ 0xb0, 0x7f, 0x4c, 0x7d, 0x5f, 0x8e, 0x6c, 0x7d, 0x71, 0x7e, 0x4e, 0x87,
+ 0xb7, 0x97, 0x7a, 0x4c, 0x5f, 0x72, 0x78, 0x84, 0x82, 0x7e, 0x63, 0x65,
+ 0x68, 0x78, 0x73, 0x85, 0x90, 0x99, 0x80, 0x57, 0x42, 0x8b, 0x8a, 0x77,
+ 0x71, 0x97, 0x6d, 0x44, 0x41, 0x8f, 0x78, 0x7d, 0x95, 0x81, 0x95, 0x5f,
+ 0x64, 0x87, 0x66, 0x80, 0x89, 0x9a, 0x61, 0x4d, 0x68, 0x7b, 0x72, 0x73,
+ 0x85, 0x92, 0x77, 0x7d, 0x73, 0x77, 0x54, 0x7a, 0x77, 0x7d, 0x7d, 0x7a,
+ 0x6e, 0x8e, 0x4f, 0x7d, 0x80, 0x9a, 0x79, 0x8b, 0x7b, 0x68, 0x6e, 0x86,
+ 0x7f, 0x93, 0x7a, 0x76, 0x72, 0x85, 0x6a, 0x7b, 0x57, 0x84, 0x96, 0x9a,
+ 0x8f, 0x91, 0x9b, 0x72, 0x73, 0x91, 0x53, 0x66, 0x76, 0x80, 0xae, 0x63,
+ 0xbf, 0x99, 0x5e, 0x77, 0x73, 0x9c, 0xd8, 0x74, 0xa7, 0x79, 0x52, 0x64,
+ 0x82, 0x95, 0xc7, 0x4f, 0xa8, 0x4f, 0x6d, 0x42, 0x7c, 0x89, 0xab, 0x83,
+ 0xc0, 0x82, 0x6a, 0x5f, 0x83, 0x92, 0xa8, 0x76, 0xc1, 0x77, 0x6e, 0x7b,
+ 0xa3, 0x9b, 0xaf, 0x87, 0xab, 0x60, 0x8d, 0xc2, 0xd2, 0x83, 0xb2, 0x78,
+ 0x8d, 0x39, 0x57, 0x9c, 0x90, 0x8e, 0x6e, 0x6a, 0x74, 0x79, 0x81, 0x6d,
+ 0x6f, 0x8e, 0x77, 0x92, 0x93, 0x7d, 0x5f, 0x68, 0x6a, 0x6c, 0x80, 0x8f,
+ 0x99, 0x84, 0x4f, 0x64, 0x5c, 0x93, 0x7c, 0x91, 0x98, 0x82, 0x62, 0x3f,
+ 0x41, 0x9f, 0x5d, 0x89, 0x98, 0x89, 0x73, 0x50, 0x32, 0xa8, 0xa0, 0x7a,
+ 0xa0, 0x95, 0x78, 0x69, 0x74, 0x7c, 0x89, 0x7b, 0x80, 0x65, 0x56, 0x6b,
+ 0x69, 0x78, 0x62, 0x87, 0xaf, 0x94, 0x7a, 0x64, 0x53, 0x86, 0x45, 0x99,
+ 0x88, 0x79, 0x4d, 0x74, 0x59, 0x91, 0x5f, 0x7b, 0x88, 0x90, 0x80, 0x86,
+ 0x7d, 0x7b, 0x64, 0xa3, 0x7f, 0x74, 0x89, 0x80, 0x7d, 0x7c, 0x7a, 0x87,
+ 0x5f, 0x8a, 0x5a, 0x72, 0x79, 0x74, 0x8c, 0x7c, 0x86, 0x91, 0x6e, 0x5d,
+ 0x61, 0x8e, 0xa2, 0x68, 0xd4, 0x92, 0x67, 0x66, 0x62, 0xa1, 0xf3, 0x63,
+ 0x91, 0x81, 0x74, 0x5f, 0x88, 0x98, 0xbb, 0x5a, 0x9b, 0x54, 0x6a, 0x5c,
+ 0x75, 0x88, 0xad, 0x7c, 0xb4, 0x7c, 0x69, 0x74, 0x84, 0x76, 0x9d, 0x9a,
+ 0xb0, 0x91, 0x5d, 0xa3, 0xa4, 0x7f, 0xbb, 0x80, 0xa4, 0x5d, 0x83, 0xaf,
+ 0xb7, 0x66, 0xb0, 0x7f, 0x89, 0x4b, 0x72, 0x9e, 0x99, 0x7c, 0x66, 0x71,
+ 0x6a, 0x6f, 0x6d, 0x67, 0x8d, 0x6d, 0x46, 0xa5, 0x9b, 0x84, 0x7a, 0x61,
+ 0x64, 0x5c, 0x88, 0x89, 0x95, 0x8c, 0x70, 0x4b, 0x6c, 0x85, 0x83, 0x8b,
+ 0x98, 0x87, 0x6a, 0x44, 0x4d, 0x9d, 0x78, 0x71, 0x78, 0x7e, 0x91, 0x5b,
+ 0x3f, 0x9f, 0x80, 0x62, 0xa7, 0x95, 0x5d, 0x74, 0x65, 0x9c, 0x6d, 0x7a,
+ 0x98, 0x79, 0x80, 0x61, 0x49, 0x82, 0x65, 0x92, 0x80, 0x96, 0x7c, 0x72,
+ 0x4f, 0x76, 0x5e, 0x8d, 0x97, 0xa5, 0x72, 0x57, 0x79, 0x87, 0x67, 0x87,
+ 0x80, 0x84, 0x7c, 0x6f, 0x66, 0x6b, 0x70, 0x9b, 0x64, 0x90, 0x59, 0x96,
+ 0x7a, 0x6f, 0x75, 0x89, 0x4e, 0x8a, 0x62, 0x6e, 0x9c, 0x8c, 0x9a, 0x78,
+ 0x8e, 0x91, 0x3d, 0x50, 0x72, 0x92, 0x9f, 0x63, 0xda, 0x92, 0x72, 0x60,
+ 0x59, 0xa6, 0xd0, 0x56, 0xc1, 0x6b, 0x5e, 0x76, 0x6e, 0x81, 0xbb, 0x4b,
+ 0xbb, 0x59, 0x68, 0x4f, 0x77, 0x87, 0xa1, 0x73, 0xbf, 0x65, 0x56, 0x67,
+ 0x77, 0x84, 0x8a, 0x7e, 0xb8, 0x85, 0x66, 0xa6, 0x99, 0xa0, 0xa5, 0x73,
+ 0x8d, 0x4a, 0x7d, 0xab, 0xb0, 0x6a, 0x94, 0x84, 0x87, 0x4c, 0x74, 0xa3,
+ 0xb3, 0xa9, 0x62, 0x7a, 0x71, 0x7f, 0x53, 0x79, 0x7a, 0x7c, 0x5e, 0x8f,
+ 0xa0, 0x90, 0x5c, 0x76, 0x6c, 0x92, 0x70, 0x9c, 0xb3, 0x8b, 0x7e, 0x57,
+ 0x5b, 0x9d, 0x96, 0x85, 0x70, 0x93, 0x8b, 0x67, 0x4c, 0x9c, 0x6a, 0x83,
+ 0x84, 0x90, 0x8e, 0x60, 0x56, 0xb3, 0x87, 0x7d, 0x86, 0x88, 0x79, 0x5b,
+ 0x58, 0x94, 0x92, 0x8e, 0x90, 0x76, 0x58, 0x51, 0x52, 0x63, 0x57, 0x88,
+ 0x9b, 0x7a, 0x85, 0x6c, 0x8b, 0x87, 0x5f, 0x8b, 0x90, 0x92, 0x81, 0x64,
+ 0x52, 0x8b, 0x77, 0x94, 0x96, 0x98, 0x69, 0x5b, 0x79, 0x87, 0x61, 0x96,
+ 0x7b, 0x9a, 0x61, 0x74, 0x7e, 0x8b, 0x82, 0x92, 0x4f, 0x87, 0x7f, 0x80,
+ 0x74, 0x97, 0x98, 0x7a, 0x79, 0x97, 0x65, 0x67, 0x66, 0xb1, 0xb1, 0x49,
+ 0xd6, 0x97, 0x58, 0x47, 0x62, 0x94, 0xd5, 0x82, 0xa0, 0x60, 0x3f, 0x67,
+ 0x6c, 0x9d, 0xb6, 0x58, 0xb1, 0x6e, 0x58, 0x4e, 0x7c, 0x83, 0x8b, 0x83,
+ 0xd5, 0x62, 0x8d, 0x84, 0x84, 0x8c, 0xa9, 0x6e, 0xac, 0x7f, 0x6d, 0x88,
+ 0xab, 0x8b, 0xb1, 0x77, 0x9b, 0x46, 0x76, 0xa7, 0xb8, 0x7b, 0xc5, 0x6e,
+ 0x73, 0x62, 0x68, 0x95, 0xab, 0x7c, 0x6f, 0x74, 0x56, 0x71, 0x61, 0x83,
+ 0x8a, 0x73, 0x54, 0x94, 0x86, 0x91, 0x60, 0x69, 0x65, 0x6b, 0x76, 0x85,
+ 0xae, 0x87, 0x8f, 0x55, 0x41, 0x98, 0x68, 0x87, 0x5e, 0x7a, 0x80, 0x38,
+ 0x50, 0xaf, 0x93, 0x79, 0x57, 0x96, 0x7b, 0x53, 0x4e, 0xc0, 0xa0, 0x85,
+ 0x87, 0x95, 0x86, 0x70, 0x4c, 0x9f, 0x77, 0x7d, 0x8b, 0x7a, 0x7b, 0x6d,
+ 0x57, 0x74, 0x81, 0x7d, 0xa2, 0x79, 0x64, 0x6c, 0x55, 0x70, 0x3c, 0x88,
+ 0x8a, 0x7a, 0x58, 0x72, 0x71, 0x7d, 0x6a, 0x8d, 0x78, 0x7e, 0x95, 0x8b,
+ 0x84, 0x7e, 0x73, 0x7c, 0x7e, 0x67, 0x89, 0x8b, 0x6d, 0x68, 0x66, 0x73,
+ 0x5a, 0x93, 0x82, 0x85, 0x97, 0x6b, 0x9a, 0x72, 0x51, 0xa2, 0x4f, 0x67,
+ 0x67, 0x7e, 0xbb, 0x37, 0xe3, 0x9c, 0x57, 0x5b, 0x6f, 0xa0, 0xdc, 0x5c,
+ 0xa6, 0x7c, 0x71, 0x77, 0x72, 0x88, 0xd0, 0x4d, 0x93, 0x58, 0x74, 0x6d,
+ 0x8f, 0x77, 0xa3, 0x76, 0xb7, 0x76, 0x6d, 0x6d, 0x6f, 0x7b, 0xaa, 0x6d,
+ 0xaa, 0x6a, 0x72, 0x98, 0x8d, 0x98, 0xb0, 0x52, 0x76, 0x5d, 0x61, 0xb7,
+ 0xac, 0x90, 0xa5, 0x75, 0x7e, 0x3d, 0x5b, 0x9a, 0xbf, 0x81, 0x83, 0x7b,
+ 0x5c, 0x77, 0x74, 0x82, 0x8d, 0x7e, 0x4f, 0x9f, 0x8f, 0x97, 0x7c, 0x75,
+ 0x5b, 0x73, 0x97, 0x73, 0x85, 0x7f, 0x70, 0x5a, 0x53, 0x81, 0x81, 0x89,
+ 0x73, 0x8d, 0x8a, 0x5c, 0x5f, 0x84, 0x86, 0x6f, 0x76, 0x78, 0x82, 0x6d,
+ 0x4f, 0xbb, 0x91, 0x61, 0x7e, 0x97, 0x6c, 0x67, 0x62, 0x83, 0x61, 0x7d,
+ 0x89, 0x76, 0x7b, 0x67, 0x56, 0x74, 0x49, 0x7b, 0x6b, 0x8b, 0x89, 0x74,
+ 0x5b, 0x7f, 0x78, 0x7b, 0x80, 0x7e, 0x63, 0x71, 0x5e, 0x91, 0x81, 0x92,
+ 0x7b, 0x90, 0x9c, 0x7a, 0x73, 0x85, 0x79, 0x9b, 0x66, 0x93, 0x60, 0x87,
+ 0x79, 0x69, 0x73, 0x8b, 0x53, 0x8c, 0x8d, 0x68, 0x93, 0xa0, 0x91, 0x65,
+ 0x57, 0x8d, 0x71, 0x65, 0x6c, 0x7e, 0xb3, 0x4f, 0xc7, 0xaa, 0x5a, 0x77,
+ 0x6e, 0x85, 0xe4, 0x6c, 0xa3, 0x89, 0x69, 0x54, 0x6d, 0x99, 0xb9, 0x77,
+ 0xa0, 0x80, 0x85, 0x71, 0x70, 0x78, 0x99, 0x66, 0xaf, 0x8a, 0x59, 0x64,
+ 0x54, 0x62, 0xbf, 0x5c, 0xbd, 0x77, 0x7f, 0xab, 0x95, 0x85, 0xaa, 0x6e,
+ 0xaa, 0x5a, 0x7b, 0x9f, 0xc3, 0x65, 0x93, 0x64, 0x7c, 0x2d, 0x4e, 0x8f,
+ 0xb2, 0x5f, 0x4e, 0x61, 0x64, 0x73, 0x56, 0x75, 0x79, 0x90, 0x5c, 0x81,
+ 0x8a, 0x8c, 0x70, 0x64, 0x74, 0x86, 0x86, 0x82, 0xab, 0x7e, 0x62, 0x4f,
+ 0x51, 0x89, 0x7b, 0x88, 0x73, 0x97, 0x77, 0x75, 0x5c, 0x9e, 0x97, 0x70,
+ 0x5a, 0x98, 0x7a, 0x54, 0x47, 0x99, 0xab, 0x5d, 0x91, 0xa0, 0x64, 0x51,
+ 0x57, 0x88, 0x88, 0x85, 0x81, 0x83, 0xa1, 0x89, 0x6a, 0x88, 0x69, 0x81,
+ 0x92, 0x63, 0x6a, 0x71, 0x72, 0x6a, 0x75, 0x8e, 0x90, 0x9d, 0x69, 0x60,
+ 0x73, 0x95, 0x79, 0x7b, 0x79, 0x7f, 0x77, 0x6e, 0x69, 0x63, 0x60, 0xa0,
+ 0x84, 0x91, 0x80, 0x96, 0x92, 0x70, 0x69, 0x7c, 0x3f, 0x90, 0x5c, 0x79,
+ 0x82, 0x63, 0x8d, 0x63, 0x56, 0x8a, 0x8e, 0x7a, 0x5c, 0x8d, 0xb8, 0x4e,
+ 0xb6, 0x84, 0x57, 0x79, 0x59, 0x79, 0xe8, 0x7e, 0xa8, 0x71, 0x61, 0x62,
+ 0x89, 0x71, 0xb7, 0x83, 0x7b, 0x53, 0x86, 0x88, 0x74, 0x71, 0xb1, 0x61,
+ 0xae, 0x7e, 0x8f, 0x69, 0x6b, 0x69, 0xb2, 0x6d, 0xb1, 0x7f, 0x5c, 0x9f,
+ 0xaa, 0x8c, 0xbd, 0x74, 0xaa, 0x5b, 0x7f, 0xa5, 0xb0, 0x6e, 0xc1, 0x5c,
+ 0x94, 0x34, 0x5b, 0xa6, 0xbc, 0x49, 0x75, 0x5b, 0x6e, 0x74, 0x7a, 0x92,
+ 0x92, 0x79, 0x78, 0x8a, 0x9e, 0x97, 0x7c, 0x5f, 0x76, 0x86, 0x59, 0x81,
+ 0x83, 0x7a, 0x65, 0x5b, 0x42, 0x95, 0x84, 0x99, 0x81, 0x8d, 0x6a, 0x5e,
+ 0x59, 0xb7, 0x96, 0x8a, 0x77, 0x86, 0x7a, 0x67, 0x3b, 0xa8, 0xae, 0x7a,
+ 0xa0, 0x97, 0x6c, 0x73, 0x5b, 0x9b, 0x77, 0x84, 0x7a, 0x77, 0x75, 0x6f,
+ 0x7d, 0x7a, 0x71, 0x86, 0x6c, 0x6f, 0x7d, 0x71, 0x68, 0x60, 0x64, 0x86,
+ 0x90, 0x75, 0x6a, 0x61, 0x60, 0x87, 0x68, 0x99, 0x87, 0x7e, 0x92, 0x87,
+ 0x87, 0x5f, 0x60, 0x91, 0x68, 0x8c, 0x7b, 0x67, 0x79, 0x5d, 0x67, 0x77,
+ 0x47, 0x72, 0x76, 0x88, 0x82, 0xa2, 0x7a, 0x5d, 0x64, 0x87, 0x75, 0x78,
+ 0x5e, 0x6f, 0xa4, 0x52, 0xc2, 0x9d, 0x81, 0x89, 0x55, 0x86, 0xc9, 0x6f,
+ 0x95, 0x71, 0x9d, 0x87, 0x95, 0x74, 0xac, 0x7f, 0x95, 0x6c, 0x68, 0x66,
+ 0x8a, 0x5f, 0x96, 0x69, 0x95, 0x79, 0x7f, 0x71, 0x86, 0x7e, 0x98, 0x71,
+ 0xac, 0x8f, 0x75, 0xa5, 0xac, 0x7a, 0xca, 0x63, 0xa0, 0x63, 0x69, 0xbf,
+ 0xae, 0x62, 0xc9, 0x46, 0x74, 0x2c, 0x66, 0x96, 0xb7, 0x70, 0x7c, 0x6b,
+ 0x7b, 0x90, 0x72, 0x74, 0x8d, 0x5f, 0x63, 0x93, 0x97, 0x78, 0x79, 0x64,
+ 0x67, 0x84, 0x64, 0x82, 0x90, 0x83, 0x91, 0x5f, 0x72, 0x93, 0x91, 0xae,
+ 0x6d, 0x99, 0x5b, 0x69, 0x54, 0x9f, 0x97, 0x80, 0x80, 0xa4, 0x91, 0x66,
+ 0x65, 0xa4, 0xa7, 0x7b, 0x97, 0x87, 0x72, 0x68, 0x6a, 0x96, 0x7b, 0x79,
+ 0x69, 0x83, 0x6f, 0x85, 0x6b, 0x92, 0x7f, 0x71, 0x84, 0x87, 0x6a, 0x7b,
+ 0x63, 0x72, 0x5f, 0x87, 0x98, 0x7b, 0x96, 0x71, 0x62, 0x90, 0x71, 0xa3,
+ 0x8c, 0x77, 0x90, 0x6f, 0x83, 0x76, 0x65, 0x87, 0x72, 0x8a, 0x64, 0x87,
+ 0x75, 0x75, 0x6d, 0x84, 0x54, 0x89, 0x88, 0xa0, 0x87, 0x73, 0x7f, 0x6f,
+ 0x5f, 0x90, 0x5e, 0x94, 0x5d, 0x61, 0xa6, 0x56, 0xb3, 0x91, 0x95, 0x75,
+ 0x4d, 0x74, 0xd9, 0x87, 0x92, 0x74, 0x7f, 0x79, 0x97, 0x6e, 0x90, 0x54,
+ 0x84, 0x5d, 0x5f, 0x75, 0x8b, 0x84, 0xa6, 0x75, 0xb4, 0x77, 0x78, 0x85,
+ 0x90, 0x76, 0xbd, 0x78, 0xd1, 0xa0, 0x5d, 0x96, 0xa9, 0x7c, 0xc1, 0x61,
+ 0xc2, 0x71, 0x8b, 0xa5, 0xa5, 0x5b, 0xc8, 0x50, 0x7b, 0x4b, 0x93, 0x99,
+ 0xae, 0x72, 0x67, 0x54, 0x81, 0x89, 0x96, 0x81, 0x6e, 0x68, 0x55, 0x7f,
+ 0x93, 0x8c, 0x5e, 0x65, 0x6c, 0x84, 0x7f, 0x8f, 0x9e, 0x7b, 0x73, 0x7f,
+ 0x51, 0x63, 0x8a, 0x8b, 0x6b, 0x9b, 0x9d, 0x57, 0x68, 0x89, 0x98, 0x70,
+ 0x73, 0xa3, 0x7f, 0x69, 0x44, 0x89, 0xae, 0x68, 0x89, 0x80, 0x7e, 0x6d,
+ 0x70, 0x95, 0x85, 0x65, 0x91, 0x7f, 0x66, 0x74, 0x96, 0x72, 0x60, 0x7a,
+ 0x87, 0x85, 0x79, 0x54, 0x53, 0x6c, 0x88, 0x87, 0xa9, 0x90, 0x75, 0x8b,
+ 0x69, 0x98, 0x7d, 0x95, 0x85, 0x7a, 0x8b, 0x82, 0x87, 0x6f, 0x86, 0x7f,
+ 0x74, 0xab, 0x93, 0x6c, 0x8a, 0x78, 0x68, 0x81, 0x62, 0x88, 0x78, 0x91,
+ 0x8b, 0x55, 0xa7, 0x58, 0x64, 0x88, 0x71, 0x93, 0x7d, 0x69, 0xbc, 0x58,
+ 0xbe, 0x9a, 0x6f, 0x74, 0x6f, 0x7f, 0xeb, 0x9e, 0xb7, 0x60, 0x63, 0x98,
+ 0x82, 0x77, 0x94, 0x63, 0x80, 0x6f, 0x7d, 0x8f, 0x8b, 0x85, 0xa5, 0x62,
+ 0xad, 0x86, 0x5f, 0x76, 0x88, 0x74, 0xa5, 0x66, 0xa5, 0x94, 0x88, 0x9b,
+ 0x87, 0x9e, 0xa8, 0x5a, 0xc9, 0x81, 0x92, 0xcd, 0xb5, 0x67, 0xb9, 0x63,
+ 0x86, 0x65, 0x8d, 0xad, 0x98, 0x7c, 0x8a, 0x40, 0x67, 0x65, 0x60, 0x71,
+ 0x8e, 0x84, 0x73, 0x64, 0x98, 0x80, 0x73, 0x81, 0x48, 0x75, 0x71, 0x9e,
+ 0x73, 0x89, 0x89, 0x68, 0x73, 0xa6, 0x84, 0x8a, 0x7e, 0x9f, 0x78, 0x83,
+ 0x60, 0x77, 0xa1, 0x87, 0x76, 0xab, 0x74, 0x57, 0x6d, 0x99, 0xa5, 0x5e,
+ 0x9d, 0x91, 0x6d, 0x6a, 0x76, 0x9c, 0x7b, 0x66, 0x96, 0x84, 0x85, 0x6e,
+ 0x6c, 0x75, 0x86, 0x6a, 0x71, 0x67, 0x8a, 0x66, 0x66, 0x68, 0x73, 0x90,
+ 0x92, 0x68, 0x8f, 0x71, 0x82, 0x7e, 0x71, 0xad, 0x9f, 0x84, 0x9e, 0x7d,
+ 0x77, 0x6b, 0x67, 0x8f, 0x73, 0x9a, 0x91, 0x74, 0x8a, 0x74, 0x5a, 0x87,
+ 0x37, 0x80, 0x8c, 0x8f, 0x7f, 0x75, 0xa8, 0x49, 0x63, 0x9b, 0x67, 0x68,
+ 0x4f, 0x87, 0xbf, 0x59, 0x9c, 0xbe, 0x93, 0x7e, 0x6f, 0x8a, 0xea, 0x77,
+ 0x83, 0x7a, 0x75, 0x8e, 0x7d, 0x50, 0x95, 0x60, 0x74, 0x60, 0x6f, 0x97,
+ 0x72, 0x5c, 0xa3, 0x6d, 0xb9, 0x86, 0x7b, 0x89, 0x9a, 0x76, 0xc7, 0x56,
+ 0xba, 0x86, 0x8d, 0x93, 0xa9, 0x98, 0xbb, 0x6a, 0x97, 0x74, 0x68, 0x84,
+ 0xc3, 0x65, 0xb6, 0x68, 0x89, 0x58, 0x87, 0xa1, 0xac, 0x60, 0x65, 0x68,
+ 0x7d, 0x98, 0x67, 0x8f, 0x8e, 0x84, 0x50, 0x75, 0x83, 0x91, 0x8a, 0x90,
+ 0x66, 0x74, 0x96, 0x89, 0x81, 0x7a, 0x7a, 0x64, 0x7f, 0x73, 0x8f, 0x95,
+ 0x8c, 0x89, 0x96, 0x76, 0x7a, 0x6c, 0x89, 0x91, 0x6d, 0x84, 0x68, 0x8d,
+ 0x47, 0x94, 0x9a, 0x67, 0x8f, 0x89, 0x8e, 0x79, 0x73, 0xa8, 0x7f, 0x6c,
+ 0x80, 0x64, 0x75, 0x81, 0x96, 0x9c, 0x68, 0x65, 0x76, 0x68, 0x74, 0x72,
+ 0x68, 0x76, 0x62, 0x6d, 0x6e, 0x6a, 0x84, 0x65, 0x8a, 0x73, 0x76, 0x91,
+ 0x78, 0x7c, 0x7a, 0x88, 0x6a, 0x87, 0x60, 0x99, 0x88, 0x75, 0x7b, 0x71,
+ 0x81, 0x7b, 0x76, 0x7d, 0x58, 0x75, 0x65, 0xa3, 0x95, 0x7e, 0x96, 0x3e,
+ 0x4c, 0x97, 0x86, 0x7a, 0x62, 0x92, 0xd1, 0x72, 0x8e, 0xaa, 0x85, 0x8e,
+ 0x59, 0x5f, 0xec, 0x77, 0x96, 0x66, 0x91, 0x9a, 0x89, 0x6c, 0xa2, 0x69,
+ 0x7d, 0x6e, 0x76, 0x63, 0x82, 0x72, 0x9c, 0x72, 0xa3, 0x75, 0x85, 0x7b,
+ 0x6d, 0x96, 0xc2, 0x69, 0xa7, 0x6a, 0x6b, 0x83, 0xa2, 0x7d, 0xce, 0x5c,
+ 0x94, 0x61, 0x7d, 0xae, 0xc3, 0x6d, 0x9f, 0x3c, 0x52, 0x4d, 0x8e, 0x92,
+ 0xae, 0x6e, 0x70, 0x5a, 0x76, 0x84, 0x7f, 0x72, 0x92, 0x72, 0x76, 0x5e,
+ 0x73, 0x8e, 0x82, 0x6d, 0x72, 0x81, 0x79, 0x94, 0x81, 0x88, 0x8b, 0x81,
+ 0x72, 0x72, 0x69, 0x84, 0x59, 0x6e, 0x74, 0x7d, 0x66, 0x74, 0x8d, 0x7b,
+ 0x7d, 0x7e, 0x7a, 0x83, 0x4d, 0x7e, 0x6a, 0x5a, 0x87, 0x66, 0x84, 0xa5,
+ 0x50, 0x5d, 0x6a, 0x8e, 0x87, 0x74, 0x88, 0x7c, 0x7d, 0x6c, 0x93, 0x98,
+ 0x8c, 0x76, 0x7f, 0xa3, 0x6e, 0x5d, 0x7d, 0x9f, 0x7c, 0x7a, 0x98, 0x88,
+ 0x74, 0x73, 0x50, 0x8c, 0x78, 0x8b, 0x71, 0x77, 0x9d, 0x56, 0x71, 0x85,
+ 0x6b, 0x8a, 0x93, 0x82, 0x8c, 0x79, 0x68, 0x8b, 0x57, 0x7b, 0x7c, 0x8a,
+ 0x6c, 0x87, 0x98, 0x54, 0x63, 0x7e, 0x78, 0x6b, 0x63, 0x77, 0xc1, 0x52,
+ 0xcd, 0xab, 0x75, 0x8e, 0x64, 0x68, 0xce, 0x68, 0x88, 0x6d, 0x67, 0x6d,
+ 0x68, 0x76, 0xa7, 0x78, 0x83, 0x67, 0x65, 0x5b, 0x8f, 0x63, 0x90, 0x5b,
+ 0xa1, 0x6f, 0x6a, 0x88, 0x70, 0x5c, 0x78, 0x49, 0xbc, 0x85, 0x8d, 0x8e,
+ 0xa3, 0x90, 0x97, 0x84, 0xa2, 0x46, 0x7a, 0x8e, 0x9e, 0xb1, 0xaa, 0x53,
+ 0x7d, 0x6b, 0x72, 0x86, 0x8c, 0x67, 0x6b, 0x48, 0x6f, 0x9c, 0x51, 0x94,
+ 0x6d, 0x66, 0x8e, 0x90, 0x79, 0x81, 0x66, 0x9f, 0x82, 0x9f, 0x98, 0x97,
+ 0x7c, 0x86, 0x7f, 0x57, 0x57, 0x83, 0x97, 0x8f, 0x73, 0x6f, 0x75, 0x6c,
+ 0x56, 0x8f, 0x7f, 0x73, 0x71, 0x84, 0x7d, 0x5f, 0x69, 0x69, 0x8e, 0x67,
+ 0x8a, 0x7f, 0x8c, 0x5a, 0x7a, 0x67, 0x82, 0x5a, 0x7a, 0x68, 0x73, 0x58,
+ 0x84, 0x83, 0x8d, 0x6d, 0x83, 0x72, 0x80, 0x7a, 0x8e, 0x7a, 0x68, 0x88,
+ 0x65, 0x74, 0x78, 0x73, 0x83, 0x97, 0x7b, 0x84, 0x77, 0x6d, 0x95, 0x99,
+ 0x76, 0x69, 0x5f, 0x9b, 0x7c, 0x75, 0x91, 0x80, 0x7b, 0x73, 0x6f, 0x9f,
+ 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x08, 0x00, 0x07, 0x00, 0x0c, 0x00,
+ 0x10, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02,
+ 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
+ 0x28, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
+ 0x0b, 0x00, 0x00, 0x00, 0x4d, 0x61, 0x74, 0x4d, 0x75, 0x6c, 0x5f, 0x62,
+ 0x69, 0x61, 0x73, 0x00, 0x0c, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0xaa, 0xcc, 0xe2, 0x37, 0x10, 0x00, 0x00, 0x00, 0xd6, 0x01, 0x00, 0x00,
+ 0xfd, 0xfd, 0xff, 0xff, 0x53, 0xfe, 0xff, 0xff, 0x74, 0x01, 0x00, 0x00,
+ 0x03, 0x00, 0x00, 0x00, 0xb4, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0xc0, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09,
+ 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x24, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x04, 0x00,
+ 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x14, 0x00, 0x1c, 0x00,
+ 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x07, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x18, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08,
+ 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
+ 0x28, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, 0x14, 0x00, 0x18, 0x00,
+ 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x07, 0x00, 0x10, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x14, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02,
+ 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00,
+ 0x1c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
+ 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x10, 0x00,
+ 0x00, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x0c, 0x00, 0x00, 0x00,
+ 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
+ 0x03, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
+ 0x04, 0x00, 0x00, 0x00, 0xfa, 0xff, 0xff, 0xff, 0x00, 0x19, 0x06, 0x00,
+ 0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x09, 0x06, 0x00,
+ 0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04};
+const int g_model_len = 18288;
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/model.h b/tensorflow/lite/micro/examples/micro_speech/micro_features/model.h
new file mode 100644
index 0000000..deec2d6
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/micro_features/model.h
@@ -0,0 +1,27 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a standard TensorFlow Lite FlatBuffer model file that has been
+// converted into a C data array, so it can be easily compiled into a binary
+// for devices that don't have a file system. It was created using the command:
+// xxd -i model.tflite > model.cc
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_MODEL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_MODEL_H_
+
+extern const unsigned char g_model[];
+extern const int g_model_len;
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_MODEL_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc b/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc
deleted file mode 100644
index 16052ba..0000000
--- a/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.cc
+++ /dev/null
@@ -1,1554 +0,0 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// Automatically created from a TensorFlow Lite flatbuffer using the command:
-// xxd -i tiny_conv.tflite > tiny_conv_simple_features_model_data.cc
-// See the README for a full description of the creation process.
-
-#include "tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h"
-
-// We need to keep the data array aligned on some architectures.
-#ifdef __has_attribute
-#define HAVE_ATTRIBUTE(x) __has_attribute(x)
-#else
-#define HAVE_ATTRIBUTE(x) 0
-#endif
-#if HAVE_ATTRIBUTE(aligned) || (defined(__GNUC__) && !defined(__clang__))
-#define DATA_ALIGN_ATTRIBUTE __attribute__((aligned(4)))
-#else
-#define DATA_ALIGN_ATTRIBUTE
-#endif
-
-const unsigned char
- g_tiny_conv_micro_features_model_data[] DATA_ALIGN_ATTRIBUTE = {
- 0x18, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x0e, 0x00,
- 0x18, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00,
- 0x0e, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0xd0, 0x46, 0x00, 0x00,
- 0x0c, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0xb4, 0x41, 0x00, 0x00, 0x0f, 0x00, 0x00, 0x00,
- 0x54, 0x4f, 0x43, 0x4f, 0x20, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74,
- 0x65, 0x64, 0x2e, 0x00, 0x09, 0x00, 0x00, 0x00, 0x94, 0x41, 0x00, 0x00,
- 0x74, 0x41, 0x00, 0x00, 0x44, 0x41, 0x00, 0x00, 0xb4, 0x3e, 0x00, 0x00,
- 0xac, 0x3e, 0x00, 0x00, 0xa4, 0x3e, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x0c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xf0, 0xb9, 0xff, 0xff,
- 0xf4, 0xb9, 0xff, 0xff, 0x52, 0xba, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00,
- 0x80, 0x3e, 0x00, 0x00, 0x68, 0x95, 0x91, 0x7d, 0x9b, 0x85, 0x85, 0x81,
- 0x77, 0x85, 0x99, 0x89, 0x7e, 0x8a, 0x85, 0x92, 0xa5, 0x7e, 0x93, 0x97,
- 0x97, 0x91, 0xa3, 0x97, 0x88, 0x8b, 0xa6, 0x71, 0x77, 0x85, 0x95, 0x86,
- 0x6b, 0x93, 0xcb, 0x96, 0x7a, 0x9a, 0x7f, 0x85, 0x7a, 0x8e, 0xac, 0x98,
- 0x6d, 0x9d, 0x9b, 0x70, 0x9a, 0x90, 0xba, 0x99, 0x7b, 0x93, 0x6e, 0x68,
- 0x75, 0x86, 0xc4, 0x8b, 0x66, 0x5d, 0x96, 0x7f, 0x92, 0x91, 0xb6, 0x7b,
- 0x96, 0x95, 0x9a, 0x77, 0x9a, 0x96, 0xce, 0x80, 0x88, 0x65, 0x8e, 0x80,
- 0x88, 0x85, 0xb7, 0x9c, 0x7b, 0x93, 0x9d, 0x95, 0x83, 0x92, 0xd0, 0x7e,
- 0x68, 0x88, 0x6c, 0x78, 0x98, 0x81, 0xac, 0x95, 0x9e, 0x98, 0xa2, 0x99,
- 0x8d, 0x7d, 0xb8, 0x81, 0x6e, 0x68, 0xa1, 0x81, 0x9d, 0x99, 0xb4, 0x7d,
- 0x92, 0x86, 0x9d, 0x93, 0xa3, 0xb0, 0xd6, 0x79, 0x93, 0x76, 0x8d, 0x84,
- 0x91, 0x9d, 0xbe, 0x94, 0xb0, 0x70, 0x84, 0x80, 0x85, 0x99, 0x9e, 0xa2,
- 0x86, 0x8a, 0x7a, 0x76, 0x91, 0x8d, 0xa6, 0x76, 0x8d, 0x82, 0x98, 0x8c,
- 0x92, 0x8f, 0x8c, 0xb3, 0x78, 0x75, 0xa5, 0x88, 0x73, 0x8c, 0x91, 0x7c,
- 0x82, 0x7d, 0x93, 0x9e, 0x8b, 0x97, 0x7c, 0x90, 0x84, 0x95, 0x7e, 0x9e,
- 0xa4, 0x52, 0x8a, 0xb4, 0x97, 0x65, 0x7d, 0xb6, 0x83, 0x7d, 0x99, 0x80,
- 0x97, 0x85, 0x96, 0x5f, 0x8e, 0x87, 0x95, 0x6d, 0x76, 0x84, 0x97, 0x8c,
- 0x66, 0x97, 0xae, 0x6b, 0x93, 0xb3, 0xa8, 0x8b, 0xa1, 0x79, 0xa3, 0x94,
- 0x7e, 0xa8, 0x8d, 0xad, 0x78, 0x82, 0xa2, 0x7b, 0x90, 0xa4, 0x7d, 0xb3,
- 0xa0, 0x7b, 0x94, 0x85, 0x9a, 0x8d, 0x76, 0x82, 0x65, 0x73, 0xab, 0xa4,
- 0xaa, 0x74, 0x93, 0x9c, 0x83, 0x66, 0xbf, 0x7a, 0xaa, 0x81, 0x92, 0x89,
- 0x7e, 0x88, 0xa6, 0x66, 0xaf, 0x92, 0x9f, 0x97, 0x6c, 0x89, 0x9c, 0x74,
- 0x7e, 0x82, 0x8e, 0x88, 0xb2, 0x85, 0xba, 0x96, 0x90, 0x78, 0x8d, 0xa7,
- 0x9e, 0x87, 0xbc, 0x7f, 0xb2, 0x8b, 0x77, 0x9b, 0xab, 0x8f, 0xa4, 0x7d,
- 0x6f, 0x77, 0x8c, 0x98, 0x6f, 0x89, 0xb1, 0x9f, 0xa7, 0x94, 0x7d, 0xae,
- 0x88, 0x8a, 0xa9, 0x75, 0x7d, 0x7c, 0x88, 0x99, 0x90, 0x9d, 0x97, 0xa7,
- 0x8d, 0x7f, 0x73, 0xa1, 0xa3, 0x87, 0xa9, 0x92, 0x98, 0x7e, 0x9c, 0x88,
- 0x73, 0x6b, 0x78, 0x8e, 0x7d, 0x86, 0x6c, 0x7c, 0x92, 0x40, 0x86, 0xa7,
- 0x65, 0x93, 0x67, 0x91, 0x67, 0x71, 0x6c, 0xa8, 0x81, 0x70, 0x8e, 0xa8,
- 0x7b, 0x63, 0x89, 0x76, 0x69, 0x90, 0x73, 0x5e, 0x92, 0x78, 0x7e, 0x9d,
- 0x87, 0x86, 0x89, 0x64, 0x66, 0xa9, 0x92, 0x8d, 0x72, 0x7c, 0x63, 0x7f,
- 0x94, 0x5c, 0x92, 0x89, 0x87, 0x9d, 0x8b, 0x75, 0x93, 0x8c, 0x94, 0x68,
- 0x97, 0x87, 0x78, 0x7d, 0x7f, 0x84, 0x84, 0x77, 0x6b, 0x8e, 0x83, 0xab,
- 0x7e, 0x62, 0x90, 0x83, 0x8e, 0x71, 0x7e, 0x9b, 0x96, 0x6d, 0x83, 0x6a,
- 0x76, 0x68, 0x71, 0x90, 0x98, 0x90, 0x9b, 0x68, 0x89, 0x89, 0x95, 0x85,
- 0x6e, 0x75, 0x8e, 0x95, 0x83, 0x7a, 0x91, 0x7f, 0x8b, 0x71, 0x90, 0x7d,
- 0xad, 0x91, 0x6f, 0x74, 0x98, 0x8a, 0xb0, 0xa8, 0x80, 0xa3, 0x8e, 0x7c,
- 0xa5, 0x67, 0xa4, 0x66, 0xa9, 0x7b, 0x85, 0x9d, 0x88, 0xab, 0x7d, 0x81,
- 0x6e, 0x7f, 0x8f, 0x97, 0x97, 0x84, 0x89, 0x74, 0x9d, 0x5f, 0x9c, 0x88,
- 0x6f, 0x74, 0x96, 0x9e, 0x7e, 0x7e, 0xa4, 0x85, 0x94, 0x91, 0xaf, 0x99,
- 0x7a, 0xaa, 0x8c, 0x92, 0x85, 0x9d, 0x6c, 0x79, 0x57, 0x7a, 0x80, 0x84,
- 0x79, 0x79, 0x74, 0xa5, 0x55, 0xab, 0x73, 0x8c, 0x72, 0x9d, 0x72, 0xa9,
- 0x90, 0x73, 0x8f, 0xa0, 0x89, 0x6d, 0x68, 0x66, 0x61, 0x6f, 0x57, 0x7d,
- 0x66, 0x8c, 0x65, 0x87, 0x62, 0x76, 0x83, 0x77, 0x89, 0xa4, 0x73, 0x89,
- 0x7f, 0x70, 0x79, 0x6b, 0x86, 0x6f, 0x8d, 0x96, 0x65, 0x89, 0x66, 0x53,
- 0x73, 0xae, 0x6a, 0x72, 0x88, 0x97, 0x7a, 0x7f, 0x5d, 0xa1, 0x86, 0x88,
- 0x5f, 0x9f, 0x9b, 0x8a, 0x74, 0x9a, 0x7a, 0x7e, 0x8b, 0x71, 0x58, 0x74,
- 0x8f, 0x9b, 0x9b, 0x8d, 0x6b, 0x83, 0x60, 0x7f, 0x75, 0x91, 0x79, 0x93,
- 0x7a, 0x92, 0x8c, 0x7e, 0x7a, 0x95, 0x84, 0x69, 0x8f, 0x8c, 0x7c, 0x6e,
- 0x8b, 0x87, 0x82, 0x62, 0xa6, 0x97, 0x91, 0x65, 0xa2, 0xa4, 0x9b, 0x8b,
- 0x85, 0xa4, 0x84, 0x7b, 0x67, 0x93, 0x96, 0x84, 0x85, 0x75, 0x6d, 0x9e,
- 0x80, 0x80, 0x73, 0x8c, 0x81, 0x70, 0x8a, 0x68, 0x9c, 0x8e, 0x63, 0x91,
- 0x89, 0x79, 0x8d, 0x79, 0xa4, 0x9a, 0x96, 0xa0, 0x83, 0x63, 0x88, 0x8f,
- 0x76, 0xb4, 0xa8, 0x8e, 0x68, 0x8d, 0x8e, 0x95, 0x78, 0xae, 0x5d, 0x89,
- 0x66, 0x7e, 0x7b, 0x8a, 0x75, 0x86, 0x71, 0x97, 0x6d, 0xb3, 0x67, 0x76,
- 0x82, 0x7d, 0x70, 0x79, 0x8a, 0x9c, 0x82, 0xa7, 0x82, 0xab, 0x58, 0x86,
- 0x5c, 0x70, 0x8c, 0x71, 0x61, 0xa6, 0x74, 0xa8, 0x65, 0x78, 0x72, 0x9d,
- 0x6c, 0x92, 0x70, 0x88, 0x88, 0x79, 0x96, 0x6f, 0x68, 0xa4, 0x7a, 0x7b,
- 0x96, 0xac, 0x6d, 0x76, 0x6a, 0xab, 0x82, 0x7d, 0x71, 0x8d, 0x6b, 0x81,
- 0x6c, 0x9d, 0x71, 0x59, 0x5c, 0x71, 0x77, 0x6d, 0x6a, 0x96, 0x76, 0x69,
- 0x80, 0x83, 0x88, 0x70, 0x97, 0xb4, 0x8a, 0x6c, 0xa5, 0x6e, 0x64, 0x75,
- 0x73, 0xa2, 0x7f, 0x97, 0x9e, 0x75, 0x8f, 0x86, 0x68, 0xbb, 0x6b, 0x86,
- 0x8d, 0x80, 0x8e, 0x58, 0x6d, 0xb2, 0x76, 0x99, 0x8f, 0x70, 0x6c, 0x86,
- 0x78, 0x9e, 0x91, 0x90, 0xa2, 0x7c, 0x8c, 0x81, 0x80, 0xb4, 0x77, 0x7a,
- 0x8c, 0x5f, 0x85, 0x56, 0x7a, 0x93, 0x6b, 0x5c, 0x74, 0x59, 0x7e, 0x86,
- 0x8c, 0xae, 0x76, 0x7d, 0x76, 0x7e, 0x81, 0x5f, 0x81, 0x8e, 0x7b, 0x90,
- 0xaa, 0x99, 0x79, 0x89, 0x93, 0xbc, 0x86, 0x91, 0xa2, 0x88, 0x79, 0x82,
- 0x80, 0xb6, 0x4a, 0x93, 0x7b, 0x89, 0x75, 0x8d, 0x7a, 0x8d, 0x66, 0x7c,
- 0x81, 0x9f, 0x6e, 0x86, 0x4d, 0x82, 0x66, 0x88, 0x73, 0x89, 0x7d, 0xac,
- 0x89, 0x9f, 0x58, 0x7f, 0x6b, 0x8c, 0x6a, 0x82, 0x59, 0xb8, 0x83, 0x67,
- 0x8b, 0x8a, 0x84, 0x7b, 0x7f, 0xb5, 0x44, 0x57, 0x5a, 0x73, 0x8b, 0x6d,
- 0x7c, 0x9e, 0x71, 0x72, 0x8d, 0x93, 0x80, 0x60, 0x7f, 0xc5, 0x69, 0x5c,
- 0x67, 0x92, 0x6c, 0x75, 0x66, 0x8f, 0x91, 0x5a, 0x6c, 0x70, 0x90, 0x84,
- 0x88, 0xab, 0x90, 0x66, 0x9c, 0x64, 0x6e, 0x68, 0x92, 0x9e, 0x89, 0x8d,
- 0x82, 0x97, 0x77, 0x75, 0x7f, 0xa7, 0x91, 0x75, 0x8c, 0x89, 0xa4, 0x6b,
- 0x98, 0x99, 0x80, 0x7d, 0x6b, 0x7f, 0x7d, 0x88, 0x79, 0xa1, 0x87, 0x90,
- 0x81, 0x8e, 0x94, 0x96, 0x7d, 0xa8, 0x86, 0x84, 0x86, 0x79, 0x97, 0x6e,
- 0xaa, 0x95, 0x8a, 0x9f, 0x8c, 0x72, 0x99, 0x77, 0x81, 0x94, 0x91, 0x9f,
- 0x6e, 0x67, 0x87, 0x70, 0x7d, 0xad, 0x58, 0x7f, 0x6d, 0x96, 0x8e, 0x82,
- 0x7d, 0xa6, 0x77, 0x99, 0x87, 0x95, 0x89, 0x7e, 0xa6, 0x9e, 0x86, 0xac,
- 0x78, 0x9f, 0x9b, 0x85, 0x76, 0x99, 0x6a, 0x92, 0x66, 0x7b, 0x9a, 0x99,
- 0x83, 0x8b, 0x57, 0x65, 0x75, 0x9f, 0xa6, 0x8a, 0x8d, 0x96, 0x6f, 0x80,
- 0x65, 0x8f, 0x80, 0x9f, 0x82, 0x85, 0x55, 0x75, 0x5c, 0x84, 0x91, 0x86,
- 0x76, 0x96, 0x5a, 0x6c, 0x62, 0x7b, 0x92, 0x88, 0x61, 0xca, 0x75, 0x66,
- 0x70, 0x70, 0x8e, 0x7a, 0x75, 0xb2, 0x66, 0x81, 0x5b, 0x79, 0x92, 0x97,
- 0x94, 0xaf, 0x72, 0x8a, 0x9b, 0x5f, 0x65, 0x96, 0x81, 0xb6, 0x8a, 0x6f,
- 0x94, 0x7a, 0x96, 0x92, 0x79, 0x94, 0x8e, 0x53, 0x9a, 0x73, 0x6a, 0x9d,
- 0xa1, 0xa3, 0xa4, 0x8f, 0x6b, 0xa4, 0x8b, 0x82, 0x96, 0xb1, 0x8c, 0x92,
- 0x7f, 0x91, 0x5f, 0x98, 0x8a, 0xa4, 0x7e, 0x80, 0x97, 0x86, 0x86, 0x86,
- 0x8f, 0xa6, 0x77, 0x9a, 0x82, 0x80, 0x6e, 0x73, 0x83, 0xaf, 0x87, 0x6d,
- 0x77, 0x9a, 0x83, 0x9f, 0x7c, 0xa4, 0x71, 0x6f, 0x7d, 0x75, 0x9d, 0x82,
- 0x83, 0xaf, 0x85, 0x80, 0x8d, 0x7f, 0xa4, 0xa2, 0x88, 0xba, 0x76, 0x76,
- 0x94, 0x6b, 0x76, 0x83, 0x77, 0x96, 0x78, 0x8c, 0xb0, 0x8e, 0x83, 0x87,
- 0xa0, 0xcc, 0x7f, 0xa4, 0x8c, 0x77, 0x84, 0x8c, 0x80, 0xa0, 0x57, 0x76,
- 0x76, 0x71, 0x86, 0x9c, 0x7f, 0x88, 0x57, 0x95, 0x4d, 0x8c, 0x7f, 0x80,
- 0x66, 0x9e, 0x42, 0x8d, 0x6a, 0x8e, 0x8c, 0x80, 0x89, 0x9d, 0x4f, 0x83,
- 0x54, 0x8a, 0x5e, 0x64, 0x70, 0x94, 0x78, 0x90, 0x7d, 0x78, 0x8d, 0x71,
- 0x56, 0x9a, 0x8c, 0x65, 0x8b, 0x62, 0x88, 0x9a, 0x6c, 0x8e, 0x7b, 0x78,
- 0x68, 0x86, 0x64, 0x6b, 0x67, 0xaa, 0x8c, 0x7b, 0x67, 0x75, 0x58, 0x7e,
- 0x6b, 0x97, 0x92, 0x87, 0x9c, 0x79, 0x71, 0x76, 0x7d, 0xbb, 0x89, 0x75,
- 0x83, 0x57, 0x74, 0x98, 0xa1, 0x8f, 0xb0, 0x89, 0x76, 0x88, 0x69, 0x9c,
- 0x74, 0xb0, 0x86, 0x9c, 0x79, 0x6f, 0x84, 0x70, 0x94, 0xa1, 0x6e, 0x7a,
- 0xa3, 0x88, 0xa0, 0x7a, 0x94, 0xa1, 0x82, 0x93, 0x99, 0x95, 0x7f, 0xab,
- 0x97, 0x9d, 0x6e, 0x68, 0x79, 0x73, 0x76, 0x83, 0x76, 0xbd, 0x87, 0x87,
- 0x86, 0x74, 0x8f, 0x6e, 0x65, 0xba, 0x6a, 0x78, 0x91, 0x62, 0x72, 0x67,
- 0x75, 0xbd, 0x8c, 0x5e, 0x85, 0x6d, 0x72, 0x85, 0x7d, 0x96, 0x8f, 0xb9,
- 0x9f, 0x97, 0xa2, 0x8a, 0xa1, 0xc1, 0x8d, 0xbc, 0x85, 0x78, 0x93, 0x97,
- 0x99, 0x9f, 0x3a, 0x98, 0x65, 0x8d, 0x6a, 0x6c, 0x92, 0x85, 0x49, 0x7e,
- 0x6a, 0xaa, 0x8a, 0x94, 0x6b, 0x93, 0x40, 0x8a, 0x8c, 0x9c, 0x6f, 0xad,
- 0x72, 0xb0, 0x58, 0x88, 0x60, 0x8c, 0x86, 0x84, 0x74, 0x96, 0x8f, 0x97,
- 0x5e, 0x6c, 0x79, 0x92, 0x51, 0xa8, 0x92, 0x58, 0x62, 0x6f, 0x6c, 0x76,
- 0x5f, 0x9e, 0x86, 0x71, 0x9c, 0x69, 0x7e, 0x80, 0x8a, 0x97, 0x6f, 0x79,
- 0x8b, 0x6f, 0x6c, 0x88, 0x73, 0x9c, 0x6d, 0x91, 0x77, 0x73, 0x7f, 0x97,
- 0x86, 0xa9, 0xac, 0x71, 0x82, 0x90, 0x83, 0x8a, 0x80, 0x9d, 0xa8, 0x85,
- 0x78, 0x7f, 0x94, 0x99, 0x8e, 0xa3, 0x89, 0x70, 0x87, 0x62, 0x82, 0x87,
- 0x8c, 0x98, 0x7a, 0x88, 0x72, 0x7e, 0x78, 0xa0, 0x78, 0x95, 0x97, 0x8f,
- 0x7b, 0x7c, 0x83, 0x94, 0x93, 0xa7, 0x77, 0x97, 0x90, 0x5e, 0x76, 0x7c,
- 0x68, 0xaa, 0x69, 0x67, 0x76, 0x84, 0x7e, 0x64, 0xa3, 0xbe, 0x7e, 0x8b,
- 0x82, 0x50, 0x8a, 0x82, 0x89, 0xc0, 0x79, 0x78, 0x68, 0x7c, 0x6b, 0x77,
- 0x82, 0x99, 0x7b, 0x83, 0x80, 0x90, 0x96, 0x96, 0x87, 0xb7, 0xa5, 0x94,
- 0x82, 0x99, 0x95, 0x91, 0x7e, 0xa2, 0x49, 0x95, 0x6d, 0x8e, 0xa9, 0x89,
- 0x8e, 0x8f, 0x3d, 0x95, 0x6a, 0x8c, 0x8b, 0x8c, 0x7e, 0x88, 0x63, 0x94,
- 0x69, 0x94, 0x88, 0x92, 0x79, 0xa7, 0x68, 0x60, 0x76, 0x85, 0xa1, 0x6f,
- 0x54, 0x96, 0x63, 0x7a, 0x5c, 0x73, 0x74, 0x6e, 0x53, 0x99, 0x69, 0x76,
- 0x69, 0x57, 0x6a, 0x82, 0x55, 0x93, 0x82, 0x80, 0x65, 0x7f, 0x7b, 0x76,
- 0x72, 0x87, 0x8d, 0x97, 0x98, 0x78, 0x7e, 0x6d, 0x7a, 0x95, 0x78, 0x70,
- 0x90, 0x83, 0x89, 0x80, 0x7f, 0x9d, 0x73, 0x73, 0x84, 0x77, 0x8e, 0x77,
- 0x8e, 0x75, 0x9e, 0xa5, 0x86, 0x68, 0x89, 0x7d, 0x8d, 0x99, 0x79, 0x8f,
- 0x8e, 0x87, 0x87, 0x97, 0x8c, 0x91, 0xa1, 0x96, 0x83, 0x73, 0x87, 0xa9,
- 0x8c, 0xa6, 0x85, 0x8c, 0x96, 0x7d, 0x7f, 0x8e, 0x7e, 0xb0, 0x85, 0x8f,
- 0x7f, 0x7d, 0x95, 0x7d, 0x9c, 0xb3, 0x71, 0x86, 0x81, 0x69, 0x7b, 0x69,
- 0x76, 0xb6, 0x5d, 0x67, 0x8a, 0x68, 0x9c, 0xa6, 0x70, 0xbf, 0x79, 0x60,
- 0x8b, 0x7f, 0x7a, 0x7b, 0x8b, 0xaf, 0x8c, 0xa1, 0x86, 0x92, 0x76, 0x8d,
- 0x89, 0xa2, 0xa8, 0xa3, 0xa0, 0xa2, 0x96, 0x9d, 0x7c, 0x92, 0x3f, 0x9b,
- 0x6d, 0x8a, 0x80, 0x81, 0xa0, 0x92, 0x50, 0x7c, 0x82, 0x99, 0x80, 0xa6,
- 0x8e, 0x8d, 0x4f, 0x8d, 0x65, 0x71, 0x77, 0x81, 0x51, 0xa6, 0x3f, 0x5c,
- 0x63, 0x6f, 0x61, 0x93, 0x5c, 0xaa, 0x77, 0x8f, 0x5d, 0x53, 0x79, 0x74,
- 0x6b, 0x94, 0x86, 0x81, 0x85, 0x48, 0x81, 0x80, 0x6b, 0x85, 0x6c, 0x91,
- 0x92, 0x6a, 0x74, 0x78, 0x72, 0x87, 0x6c, 0x82, 0x88, 0x7b, 0x93, 0x71,
- 0x91, 0x8d, 0x67, 0x83, 0x86, 0x5b, 0x86, 0x79, 0x81, 0x9f, 0x95, 0x8a,
- 0x70, 0x66, 0x9e, 0x6b, 0x72, 0x98, 0x97, 0x95, 0x72, 0x93, 0x84, 0x92,
- 0x8c, 0x96, 0xa2, 0x65, 0x80, 0x75, 0xa2, 0xa7, 0x7d, 0x97, 0x71, 0x8f,
- 0x69, 0x65, 0x8f, 0xae, 0x9c, 0x97, 0x5d, 0xb3, 0x98, 0x83, 0x98, 0xa0,
- 0x5f, 0x7e, 0x7a, 0x7a, 0x87, 0x7c, 0x92, 0xa0, 0x81, 0xa6, 0x71, 0x8e,
- 0x88, 0x52, 0xa3, 0x88, 0x6a, 0x9d, 0x84, 0x82, 0x7c, 0x78, 0x9f, 0x92,
- 0x66, 0xa4, 0x53, 0x6a, 0x7e, 0x84, 0x60, 0x84, 0x92, 0xb0, 0x93, 0x9d,
- 0xa0, 0x5f, 0x95, 0x8c, 0x77, 0xa1, 0x8c, 0x90, 0xa0, 0x9c, 0x9a, 0x95,
- 0x85, 0xa1, 0x22, 0x8f, 0x57, 0x80, 0x96, 0x7d, 0x92, 0x8b, 0x41, 0xa6,
- 0x61, 0xa2, 0x6f, 0x80, 0x5d, 0x91, 0x66, 0xab, 0x6d, 0x7e, 0x88, 0x93,
- 0x5c, 0xa5, 0x75, 0x6e, 0x6c, 0x86, 0x69, 0x73, 0x4e, 0x8e, 0x77, 0x6b,
- 0x6c, 0x60, 0x67, 0x91, 0x75, 0x91, 0x6c, 0x7c, 0x53, 0x6e, 0x75, 0x8e,
- 0x79, 0x8c, 0x8b, 0x74, 0x6b, 0x57, 0x71, 0xa1, 0x7f, 0x83, 0x6c, 0x6b,
- 0x93, 0x99, 0x7a, 0x78, 0x71, 0x8c, 0x78, 0x88, 0x9f, 0x85, 0x77, 0x7b,
- 0x86, 0x85, 0xa1, 0x61, 0x78, 0x65, 0x61, 0x75, 0x82, 0x7d, 0xa9, 0xa2,
- 0x84, 0x82, 0x94, 0x95, 0x90, 0x9f, 0x83, 0x97, 0x76, 0x95, 0x8a, 0x83,
- 0x9b, 0x87, 0x8b, 0x7a, 0x6c, 0x6e, 0x75, 0x95, 0x85, 0x95, 0x84, 0x9e,
- 0x96, 0x74, 0x7d, 0xa5, 0x85, 0x8e, 0x7e, 0x73, 0x85, 0x8d, 0x87, 0x80,
- 0x8a, 0x96, 0x65, 0x87, 0x7c, 0x73, 0x80, 0x96, 0x73, 0x8d, 0x5e, 0x79,
- 0x7e, 0x8d, 0x79, 0x85, 0x63, 0xa0, 0x62, 0x89, 0x9d, 0x8c, 0x74, 0x7b,
- 0x9c, 0xa5, 0x71, 0x8c, 0x83, 0x91, 0x8e, 0x8d, 0x89, 0x8b, 0x8b, 0xa4,
- 0x78, 0x88, 0x9e, 0x85, 0x8b, 0x94, 0x38, 0x84, 0x7b, 0x86, 0x7d, 0xa2,
- 0x73, 0x8f, 0x47, 0x7b, 0x69, 0xb4, 0x85, 0x71, 0x61, 0x9d, 0x59, 0x95,
- 0x74, 0x93, 0x6a, 0x88, 0x62, 0xa2, 0x56, 0x93, 0x8d, 0x68, 0x7e, 0x80,
- 0x6b, 0xb7, 0x63, 0x90, 0x5d, 0x54, 0x6c, 0x90, 0x5a, 0x8e, 0x7e, 0x7d,
- 0x82, 0x73, 0x7f, 0x89, 0x94, 0x8e, 0x7a, 0x70, 0x6c, 0x79, 0x88, 0x88,
- 0x9b, 0x8b, 0x70, 0x81, 0x83, 0x83, 0x8b, 0x86, 0x64, 0x93, 0x82, 0x66,
- 0x66, 0x79, 0x74, 0x91, 0x92, 0x94, 0x7c, 0x87, 0x72, 0x79, 0x8d, 0xaa,
- 0xa2, 0x9e, 0xaf, 0x95, 0xb1, 0x8a, 0x95, 0x8b, 0x94, 0x7e, 0x79, 0x8e,
- 0x99, 0x98, 0x97, 0x9e, 0x94, 0x87, 0x74, 0x72, 0x63, 0x92, 0x92, 0x95,
- 0xb0, 0x94, 0x86, 0x91, 0x77, 0x8f, 0x91, 0x7e, 0x83, 0x88, 0x90, 0xa5,
- 0x79, 0x70, 0x85, 0x8f, 0x67, 0x90, 0x98, 0x8d, 0x8a, 0x5d, 0x8c, 0x9c,
- 0x94, 0x91, 0x80, 0x95, 0x6e, 0x95, 0x73, 0x8d, 0x63, 0x8e, 0x53, 0x8a,
- 0x77, 0x88, 0x8f, 0x6f, 0x87, 0x9e, 0x8b, 0xb7, 0x99, 0xb2, 0x85, 0x82,
- 0xa1, 0x89, 0x9b, 0xa7, 0x80, 0x81, 0xa0, 0x8e, 0x84, 0xa9, 0x27, 0x73,
- 0x5e, 0x85, 0x5f, 0x92, 0x8c, 0xa2, 0x34, 0x8e, 0x6e, 0xb2, 0x7b, 0x8c,
- 0x69, 0x93, 0x47, 0x9e, 0x58, 0x7e, 0x94, 0x86, 0x47, 0xa3, 0x53, 0x6b,
- 0x6e, 0x6a, 0x7f, 0x73, 0x5b, 0x8c, 0x7a, 0x99, 0x6c, 0x5d, 0x82, 0x82,
- 0x62, 0x8a, 0x7a, 0x8e, 0x88, 0x62, 0xa0, 0x8e, 0x5c, 0x9a, 0x72, 0x79,
- 0x66, 0x6b, 0x75, 0x78, 0x82, 0x8a, 0x59, 0x91, 0x93, 0x68, 0x78, 0xb4,
- 0x86, 0x7e, 0x8c, 0x6e, 0x88, 0x7f, 0x96, 0x8e, 0x6e, 0x8b, 0x8c, 0x73,
- 0xab, 0x79, 0x88, 0xa6, 0x86, 0x81, 0x9a, 0x80, 0x9a, 0x9e, 0x8b, 0x6d,
- 0x9a, 0x70, 0x8e, 0x8a, 0x84, 0x7a, 0xaf, 0xb8, 0x9e, 0x90, 0x89, 0xb3,
- 0x9b, 0x85, 0x94, 0xb6, 0x87, 0x8c, 0x6e, 0xa3, 0xac, 0x9e, 0x8c, 0x7c,
- 0x81, 0x83, 0x70, 0x8d, 0x7c, 0x81, 0x77, 0x82, 0x69, 0x8e, 0x5e, 0x80,
- 0x8a, 0x8e, 0x7c, 0x8a, 0x89, 0x90, 0x58, 0x59, 0x85, 0x88, 0x7a, 0x86,
- 0x73, 0x9c, 0x4a, 0x81, 0x8d, 0x89, 0x91, 0x95, 0x72, 0x83, 0x9d, 0x99,
- 0x8d, 0x6b, 0x95, 0x7e, 0x70, 0x94, 0x8c, 0x9f, 0x8a, 0x8f, 0xa7, 0x84,
- 0x87, 0xb6, 0x42, 0x81, 0x63, 0x8a, 0x79, 0x77, 0x74, 0x90, 0x23, 0x85,
- 0x74, 0x8f, 0x87, 0x80, 0x50, 0xa1, 0x4d, 0x9b, 0x55, 0x82, 0x74, 0x8e,
- 0x4a, 0xa7, 0x52, 0x4d, 0x77, 0x67, 0x77, 0x9e, 0x62, 0xa5, 0x7d, 0x96,
- 0x6f, 0x45, 0x80, 0x8c, 0x6c, 0x92, 0x99, 0x6f, 0x5d, 0x56, 0x93, 0xac,
- 0x94, 0x9c, 0x95, 0x92, 0x6e, 0x71, 0x87, 0x8c, 0x7b, 0xa9, 0x7f, 0x7a,
- 0x69, 0x6b, 0x7d, 0x90, 0x6f, 0x81, 0x9f, 0x80, 0x83, 0x67, 0x78, 0x85,
- 0x85, 0x91, 0x8a, 0x80, 0xaa, 0x86, 0x8c, 0x88, 0x8c, 0x8f, 0x9b, 0x85,
- 0x8b, 0x7e, 0x83, 0x82, 0x95, 0x75, 0x6b, 0x8f, 0x85, 0x8b, 0xb0, 0x9f,
- 0xa7, 0x8e, 0x61, 0x9d, 0x72, 0xac, 0x92, 0x87, 0x94, 0x96, 0x68, 0x8f,
- 0x63, 0x85, 0x9c, 0xa8, 0x82, 0x9b, 0x85, 0x9b, 0x6b, 0x72, 0x83, 0x85,
- 0x90, 0x87, 0x74, 0xa4, 0x88, 0x57, 0x63, 0x90, 0x8e, 0x7b, 0x80, 0x81,
- 0x94, 0x74, 0x68, 0x8a, 0x7f, 0x86, 0x78, 0x72, 0x75, 0x67, 0x7a, 0x8a,
- 0x7a, 0x74, 0x8c, 0xad, 0x75, 0xa2, 0x7d, 0x9a, 0x9e, 0x83, 0x92, 0xa2,
- 0xa3, 0x98, 0xa5, 0x91, 0x84, 0xb0, 0x21, 0x9a, 0x5f, 0x8c, 0x7e, 0x86,
- 0x80, 0xa0, 0x16, 0x9b, 0x5b, 0x9c, 0x76, 0x8d, 0x77, 0x9f, 0x62, 0x86,
- 0x6a, 0x6c, 0x6e, 0x8f, 0x4e, 0xc1, 0x61, 0x6f, 0x74, 0x79, 0x80, 0x5f,
- 0x59, 0x9e, 0x7c, 0x87, 0x7f, 0x4b, 0x6c, 0x8b, 0x5a, 0x8f, 0x65, 0x8a,
- 0x62, 0x58, 0x66, 0x8d, 0x83, 0x97, 0x8a, 0x7a, 0x77, 0x79, 0x6c, 0x83,
- 0x8c, 0x93, 0x82, 0x5e, 0x61, 0x8c, 0x82, 0x80, 0x88, 0x88, 0x85, 0x87,
- 0x77, 0x70, 0x8d, 0x7f, 0x7a, 0x89, 0x72, 0x7e, 0xa3, 0x99, 0x6b, 0xaa,
- 0x81, 0x87, 0x90, 0x6f, 0x7f, 0x77, 0x96, 0x83, 0x89, 0x89, 0x6a, 0x77,
- 0xa4, 0x6c, 0x97, 0x7e, 0x95, 0xa4, 0x63, 0x8d, 0x71, 0x96, 0x8a, 0xa4,
- 0x9f, 0x7c, 0x54, 0x94, 0x7a, 0x89, 0x8a, 0x90, 0x7e, 0x9d, 0x53, 0x7c,
- 0x9d, 0x83, 0x90, 0x84, 0xa1, 0x8e, 0x80, 0x74, 0x69, 0x7a, 0x69, 0x93,
- 0x8a, 0x90, 0x83, 0x76, 0x8b, 0x6f, 0x8e, 0x93, 0x82, 0x84, 0x7d, 0x94,
- 0xa1, 0x78, 0x7d, 0x68, 0x79, 0x83, 0x85, 0x9d, 0x89, 0xa0, 0x8a, 0x93,
- 0x90, 0x8c, 0x82, 0x86, 0x80, 0x71, 0xb3, 0xa1, 0x90, 0xb2, 0x27, 0xa3,
- 0x5e, 0xa3, 0xa6, 0x64, 0x75, 0xa0, 0x23, 0x8c, 0x7c, 0xc4, 0x7a, 0x8c,
- 0x4d, 0xa3, 0x4c, 0x93, 0x71, 0x7b, 0x71, 0x8b, 0x34, 0xa5, 0x47, 0x7f,
- 0x4e, 0x73, 0x51, 0x8a, 0x67, 0xa0, 0x9d, 0x7f, 0x65, 0x38, 0x61, 0x70,
- 0x71, 0x8d, 0x6a, 0x7e, 0x7e, 0x4c, 0x7d, 0x8d, 0x81, 0x80, 0xa5, 0x84,
- 0x6f, 0x57, 0x70, 0x91, 0x8b, 0x99, 0x9d, 0x84, 0x77, 0x7f, 0x6b, 0x7f,
- 0x76, 0x8f, 0x90, 0x72, 0x6c, 0x58, 0x6b, 0x85, 0xa6, 0x8a, 0xa2, 0x6d,
- 0x8a, 0x71, 0x71, 0x95, 0x92, 0x7c, 0x88, 0x67, 0x86, 0x6d, 0x8d, 0x95,
- 0x79, 0x8e, 0x65, 0x71, 0x71, 0x91, 0x85, 0x99, 0xa9, 0x87, 0x80, 0x88,
- 0x74, 0x86, 0x75, 0x83, 0x8b, 0x7f, 0x78, 0xb1, 0x90, 0xa8, 0x7b, 0x98,
- 0x8a, 0x7b, 0x5b, 0x99, 0x6f, 0x7f, 0xa0, 0x79, 0xa5, 0x93, 0x8b, 0x7b,
- 0x7e, 0x7a, 0x61, 0x9d, 0x98, 0x8b, 0x82, 0x7c, 0x76, 0x73, 0x81, 0x8a,
- 0x7e, 0x8d, 0x6e, 0x71, 0xa0, 0x65, 0x80, 0x62, 0x7d, 0x8d, 0x5e, 0x9b,
- 0x8f, 0x85, 0x89, 0xad, 0x71, 0x73, 0x7f, 0x89, 0x8d, 0x89, 0xb3, 0xa1,
- 0x7c, 0xaf, 0x43, 0x82, 0x49, 0x92, 0x62, 0x7f, 0x79, 0xa6, 0x23, 0x99,
- 0x6c, 0x9a, 0x8a, 0x90, 0x6c, 0xb9, 0x6f, 0x8a, 0x61, 0x7f, 0x8f, 0x8a,
- 0x57, 0xb9, 0x55, 0x65, 0x4b, 0x51, 0x66, 0x6e, 0x4a, 0xa1, 0x83, 0x8a,
- 0x73, 0x23, 0x8a, 0x6d, 0x46, 0xa7, 0x87, 0x64, 0x84, 0x5f, 0x6f, 0x6f,
- 0x9b, 0x9d, 0x76, 0x83, 0x60, 0x6e, 0x76, 0x8a, 0x9a, 0xa6, 0x75, 0x73,
- 0x86, 0x5b, 0x97, 0x88, 0x7b, 0x8e, 0x82, 0x5c, 0x97, 0x71, 0x74, 0x85,
- 0x83, 0x91, 0x89, 0x6f, 0x93, 0x94, 0x8b, 0xa9, 0x7d, 0x84, 0x80, 0x89,
- 0x97, 0x80, 0x65, 0x92, 0x9a, 0x85, 0x5a, 0x6a, 0x6b, 0x58, 0x6f, 0x8c,
- 0x9a, 0x8b, 0x6e, 0x81, 0x9d, 0xae, 0x8c, 0x86, 0x8d, 0x90, 0x6c, 0xb8,
- 0x91, 0x89, 0x98, 0xbd, 0x8b, 0x78, 0x7d, 0x87, 0x9c, 0x72, 0x73, 0x80,
- 0x9e, 0x92, 0x5d, 0x77, 0x78, 0x4f, 0x87, 0x7b, 0x7a, 0x9e, 0x74, 0x67,
- 0x6a, 0x58, 0x95, 0x80, 0x75, 0x97, 0x81, 0x75, 0x94, 0x75, 0x73, 0x92,
- 0x83, 0x7b, 0x6b, 0x8e, 0x82, 0x6e, 0x7d, 0x9b, 0x91, 0x7f, 0x9e, 0xaa,
- 0x8c, 0xa3, 0xa8, 0x8c, 0x9a, 0xc1, 0x28, 0xac, 0x49, 0x9b, 0x59, 0x8a,
- 0x60, 0xa7, 0x39, 0xa7, 0x75, 0x9b, 0x95, 0x94, 0x76, 0xb3, 0x4a, 0x6b,
- 0x60, 0x6c, 0xa5, 0x71, 0x40, 0xc4, 0x4c, 0x7c, 0x76, 0x7b, 0x67, 0x76,
- 0x76, 0xa4, 0x7b, 0x83, 0x67, 0x4d, 0x87, 0x87, 0x6e, 0x93, 0x84, 0x70,
- 0x78, 0x41, 0x87, 0x9f, 0x7a, 0x8c, 0x87, 0x69, 0x73, 0x6c, 0x93, 0x73,
- 0x77, 0xa2, 0x52, 0x72, 0x5c, 0x75, 0x6c, 0x8f, 0x65, 0x92, 0x87, 0x52,
- 0x67, 0x54, 0x54, 0x75, 0x90, 0x9c, 0x91, 0x6f, 0xa3, 0x86, 0x87, 0x9c,
- 0x99, 0x86, 0x9f, 0x71, 0x8a, 0x7a, 0x7a, 0x97, 0x7a, 0x86, 0x6c, 0x99,
- 0x89, 0x7e, 0x9c, 0x83, 0x98, 0x78, 0x73, 0x7f, 0x91, 0x96, 0x9a, 0x8d,
- 0xb0, 0x9e, 0x6a, 0x80, 0x92, 0x86, 0x95, 0x83, 0x94, 0x92, 0x6f, 0x86,
- 0x8a, 0x52, 0x6e, 0x82, 0x84, 0x8b, 0x77, 0x88, 0x70, 0x54, 0x8f, 0x7f,
- 0x7d, 0x7e, 0x57, 0x89, 0x6d, 0x6f, 0x9c, 0x93, 0x90, 0x93, 0x52, 0x70,
- 0x75, 0x92, 0x73, 0x88, 0x93, 0x77, 0x77, 0x91, 0x89, 0xa2, 0x9d, 0xa6,
- 0xae, 0x84, 0x7d, 0xab, 0x92, 0x7e, 0x9c, 0x98, 0x7b, 0xc3, 0x38, 0x98,
- 0x4f, 0x97, 0x8f, 0x93, 0x62, 0xb8, 0x23, 0xa4, 0x6d, 0x9c, 0x81, 0x8e,
- 0x6f, 0x9d, 0x56, 0x89, 0x50, 0x94, 0x70, 0x77, 0x5d, 0xb7, 0x60, 0x5b,
- 0x72, 0x45, 0x81, 0x8c, 0x66, 0xbc, 0x8f, 0x7f, 0x57, 0x43, 0x85, 0x96,
- 0x5a, 0xb2, 0x91, 0x7d, 0x6c, 0x3a, 0x73, 0x92, 0x63, 0x93, 0x89, 0x90,
- 0x7f, 0x52, 0x7f, 0x7b, 0xa1, 0xa6, 0x8f, 0x60, 0x78, 0x51, 0x5f, 0xac,
- 0x7b, 0x89, 0x88, 0x97, 0x7e, 0x64, 0x57, 0x72, 0x6c, 0x96, 0x74, 0x78,
- 0xab, 0x66, 0x62, 0x8d, 0x6f, 0x86, 0x91, 0x93, 0x7d, 0x74, 0x82, 0x80,
- 0x73, 0x84, 0x9c, 0x8e, 0x68, 0x69, 0x9e, 0xa1, 0x8a, 0x83, 0x7a, 0x87,
- 0x94, 0x8c, 0x83, 0x7e, 0x91, 0x92, 0x82, 0x7b, 0xa0, 0x8e, 0x73, 0x86,
- 0xa9, 0x95, 0x7c, 0xa5, 0x6c, 0x6f, 0x8c, 0x87, 0xa6, 0x8a, 0x77, 0x86,
- 0x7d, 0x79, 0x89, 0x75, 0x8f, 0x82, 0x54, 0x61, 0x82, 0x8e, 0x80, 0x84,
- 0x7b, 0x8e, 0x61, 0x82, 0x86, 0x77, 0x7d, 0x7c, 0x7e, 0x6c, 0x7b, 0xad,
- 0x7b, 0x90, 0x88, 0x80, 0x64, 0x83, 0x7e, 0xa7, 0x83, 0x7e, 0xb5, 0xbb,
- 0x88, 0xd9, 0x21, 0x9a, 0x4d, 0x9f, 0x91, 0x97, 0x64, 0xb5, 0x1c, 0x8a,
- 0x5f, 0xaf, 0x7e, 0x7b, 0x67, 0xad, 0x48, 0x7f, 0x4e, 0x87, 0x8f, 0x7c,
- 0x46, 0xab, 0x70, 0x7f, 0x4b, 0x4e, 0x48, 0x8c, 0x63, 0xc5, 0xa2, 0x7f,
- 0x68, 0x3b, 0x59, 0x7f, 0x53, 0xa1, 0x8e, 0x6e, 0x7a, 0x4a, 0x5f, 0x62,
- 0x5b, 0xa1, 0x62, 0x78, 0x74, 0x57, 0x78, 0x91, 0x7b, 0x9b, 0x75, 0x73,
- 0x73, 0x72, 0x94, 0x92, 0x79, 0xaa, 0x94, 0x75, 0x86, 0x58, 0x8c, 0x71,
- 0x77, 0x91, 0xa5, 0x74, 0x8f, 0x73, 0x89, 0x77, 0x68, 0x8e, 0x90, 0x96,
- 0x9f, 0x79, 0x77, 0x7d, 0x89, 0x9b, 0x8c, 0x94, 0x81, 0x88, 0x91, 0x8f,
- 0x9b, 0x91, 0x78, 0x87, 0x82, 0x72, 0xa7, 0xa2, 0x85, 0x98, 0xa3, 0x91,
- 0x83, 0x75, 0x72, 0x93, 0x80, 0x8f, 0x85, 0x70, 0x97, 0x58, 0x9f, 0x72,
- 0x91, 0x8e, 0x93, 0x74, 0x97, 0x73, 0x74, 0x91, 0x80, 0x84, 0x96, 0x94,
- 0x76, 0x69, 0x66, 0x9e, 0x81, 0x8a, 0x8b, 0x63, 0x65, 0x7c, 0xa1, 0x9a,
- 0x72, 0x84, 0x9e, 0x89, 0x9a, 0x86, 0x98, 0x7f, 0x77, 0x85, 0x82, 0xaa,
- 0xa3, 0x88, 0xac, 0x9e, 0x76, 0xca, 0x2b, 0xa0, 0x40, 0xad, 0x6f, 0x6c,
- 0x66, 0xc8, 0x07, 0x9e, 0x3e, 0x9f, 0x85, 0x9f, 0x5e, 0xb7, 0x53, 0x91,
- 0x56, 0x6d, 0x62, 0x95, 0x4c, 0xc7, 0x46, 0x56, 0x4b, 0x5d, 0x6f, 0x52,
- 0x4d, 0xa3, 0x8c, 0x90, 0x78, 0x4d, 0x58, 0x8d, 0x53, 0x93, 0x8e, 0x68,
- 0x6f, 0x3b, 0x49, 0x86, 0x6e, 0x9d, 0x76, 0x74, 0x5b, 0x44, 0x7b, 0x8c,
- 0x89, 0xb0, 0x64, 0x62, 0x6a, 0x6d, 0x7a, 0xae, 0x84, 0x95, 0x8c, 0x71,
- 0x8b, 0x60, 0x82, 0x9e, 0x8c, 0xa8, 0x90, 0x66, 0xa1, 0x7b, 0x65, 0x82,
- 0x8f, 0x7d, 0x8d, 0x78, 0x8e, 0x5f, 0x75, 0x88, 0x5d, 0x93, 0xa1, 0x93,
- 0x6b, 0x67, 0x7a, 0xa7, 0x92, 0x8c, 0x65, 0x88, 0x95, 0x93, 0x87, 0x81,
- 0x9c, 0x97, 0x62, 0x9d, 0x90, 0x62, 0xa1, 0x9f, 0x87, 0x94, 0x94, 0x99,
- 0x92, 0x8f, 0x71, 0x80, 0x77, 0x82, 0x92, 0x78, 0x67, 0x69, 0x7e, 0x81,
- 0x93, 0x89, 0x80, 0x9b, 0x71, 0x57, 0x63, 0x83, 0x7b, 0x9f, 0x5d, 0x92,
- 0x85, 0x96, 0x7e, 0x92, 0x84, 0x7f, 0x81, 0xa3, 0xa8, 0x96, 0x91, 0x8e,
- 0x8c, 0x8e, 0x7d, 0xb0, 0x86, 0x72, 0x9d, 0x8e, 0x8e, 0xd0, 0x05, 0x77,
- 0x45, 0xad, 0x91, 0x95, 0x71, 0xb8, 0x01, 0x9a, 0x41, 0xb8, 0x94, 0x6e,
- 0x63, 0xd3, 0x58, 0x8c, 0x5a, 0x89, 0x85, 0x83, 0x52, 0xc1, 0x7b, 0x6a,
- 0x65, 0x6e, 0x73, 0x63, 0x68, 0xba, 0x67, 0x78, 0x79, 0x4a, 0x73, 0x8f,
- 0x51, 0xc9, 0x85, 0x8a, 0x6b, 0x45, 0x6a, 0x8f, 0x6c, 0xad, 0x8a, 0x8d,
- 0x6a, 0x6e, 0x6b, 0x7f, 0x86, 0xb4, 0x88, 0x7d, 0xaa, 0x71, 0x5c, 0x69,
- 0x5d, 0xa8, 0x62, 0x7d, 0x6c, 0x6e, 0x6f, 0x6a, 0x7c, 0x9d, 0x7a, 0x83,
- 0x7d, 0x79, 0x7b, 0x9c, 0x73, 0x93, 0x7f, 0x9d, 0x8c, 0x75, 0x78, 0x83,
- 0x85, 0x88, 0x81, 0x81, 0x98, 0x79, 0xa3, 0xae, 0x5b, 0x90, 0x89, 0x9d,
- 0x6d, 0x90, 0xa3, 0x8e, 0x87, 0x96, 0x60, 0xa7, 0x76, 0x82, 0x81, 0x84,
- 0x84, 0x9c, 0x73, 0x8a, 0x6c, 0x58, 0x64, 0x96, 0x89, 0x8b, 0x76, 0x60,
- 0x91, 0x72, 0x7f, 0x86, 0x9a, 0x89, 0x67, 0x7d, 0x77, 0x84, 0x73, 0x5c,
- 0x67, 0x8a, 0x82, 0x8c, 0x8c, 0x94, 0x8a, 0xa2, 0xaa, 0x7e, 0x5f, 0x7f,
- 0x86, 0x90, 0x96, 0xab, 0x8d, 0x91, 0x7c, 0xb6, 0x82, 0x8d, 0xb8, 0xa9,
- 0x92, 0xea, 0x1b, 0x74, 0x25, 0xab, 0x8d, 0x61, 0x81, 0xd8, 0x2c, 0x86,
- 0x2f, 0xcf, 0xa2, 0x84, 0x7f, 0xa4, 0x36, 0x86, 0x47, 0x8d, 0x60, 0x8a,
- 0x62, 0xb1, 0x4a, 0x54, 0x48, 0x73, 0x64, 0x9d, 0x72, 0xb2, 0x76, 0x4c,
- 0x8e, 0x4e, 0x76, 0x94, 0x7c, 0xad, 0x74, 0x6c, 0x6c, 0x54, 0x7f, 0x63,
- 0x97, 0xb3, 0x74, 0x6c, 0x99, 0x5f, 0x86, 0x6a, 0xa3, 0x94, 0x7c, 0x83,
- 0x8d, 0x81, 0x79, 0xac, 0x61, 0x9b, 0x65, 0x7b, 0x66, 0x89, 0x60, 0x76,
- 0x8d, 0x93, 0x8d, 0x84, 0x71, 0x65, 0x82, 0x8c, 0x94, 0xa7, 0x59, 0xa1,
- 0x8b, 0x72, 0x84, 0x65, 0x75, 0x95, 0x62, 0x71, 0x71, 0x7e, 0x7b, 0x97,
- 0x9b, 0x9a, 0x80, 0xb1, 0x77, 0x7a, 0x73, 0x8e, 0x9c, 0x8c, 0x7d, 0x96,
- 0x89, 0x7d, 0x7e, 0x80, 0x8e, 0x93, 0x63, 0x72, 0x6b, 0x57, 0x78, 0x8f,
- 0x90, 0x86, 0x62, 0x75, 0x7e, 0x54, 0x7d, 0x95, 0x85, 0x84, 0x73, 0x7b,
- 0x8f, 0x9e, 0x72, 0x8c, 0x90, 0x96, 0x8e, 0x6c, 0x80, 0x8b, 0x9e, 0x8c,
- 0x87, 0x8e, 0x9b, 0x97, 0x8f, 0x94, 0xa3, 0x6b, 0xad, 0x93, 0x8a, 0x96,
- 0x8d, 0x91, 0xa6, 0x8a, 0x9e, 0xce, 0x6b, 0x98, 0x6d, 0xa9, 0x92, 0x92,
- 0x7c, 0xe2, 0x63, 0x97, 0x42, 0xc8, 0xa3, 0xa0, 0x88, 0xdc, 0x75, 0x9b,
- 0x51, 0x7d, 0x5c, 0x80, 0x89, 0xc0, 0x83, 0x5e, 0x5e, 0xa4, 0x3e, 0x74,
- 0x9b, 0xb6, 0x7f, 0x63, 0x78, 0x7d, 0x74, 0x57, 0x93, 0xa2, 0x83, 0x70,
- 0x5e, 0x7d, 0x60, 0x69, 0x93, 0x9e, 0x79, 0x86, 0x91, 0x67, 0x86, 0x95,
- 0xa2, 0xad, 0x62, 0x74, 0x68, 0x7e, 0x7e, 0x82, 0x8c, 0xb0, 0xa0, 0x63,
- 0x8b, 0x82, 0x8f, 0x8c, 0xa4, 0xa3, 0x76, 0x6c, 0x8e, 0x87, 0x72, 0x85,
- 0xaa, 0xa4, 0x7f, 0x7b, 0x8e, 0x9a, 0x69, 0x91, 0x9d, 0xa0, 0x81, 0x92,
- 0x90, 0x85, 0x66, 0x82, 0xa3, 0xa9, 0x7f, 0x8f, 0x83, 0x9d, 0x8b, 0x8d,
- 0x96, 0xa3, 0x8f, 0x7a, 0x6d, 0x89, 0x74, 0x8a, 0xa9, 0xa9, 0x7b, 0x77,
- 0x93, 0x8b, 0x63, 0x92, 0x99, 0x8b, 0x88, 0x4f, 0x87, 0x7c, 0x67, 0x78,
- 0x83, 0xa5, 0xa5, 0x58, 0x8d, 0x70, 0x86, 0x82, 0x9e, 0xa7, 0xa5, 0x96,
- 0x8d, 0x7b, 0x96, 0x8c, 0x95, 0xa3, 0x8d, 0x9c, 0x92, 0x95, 0x98, 0x94,
- 0x87, 0x90, 0x92, 0x92, 0x95, 0x96, 0xad, 0x6e, 0x97, 0x8c, 0x92, 0x7f,
- 0x95, 0x8b, 0x8a, 0x90, 0x9b, 0x87, 0x9e, 0x86, 0x91, 0xa0, 0x68, 0x82,
- 0x85, 0x8e, 0x82, 0xa8, 0x9f, 0x68, 0x87, 0x75, 0x9b, 0x70, 0x95, 0x91,
- 0x6c, 0x77, 0x8b, 0x7b, 0x95, 0x80, 0x99, 0x65, 0x95, 0x82, 0x92, 0x9a,
- 0x8a, 0x65, 0x70, 0x8c, 0x98, 0x9e, 0x80, 0x7b, 0xa5, 0x9b, 0x93, 0x94,
- 0x84, 0x6a, 0x69, 0x82, 0x80, 0x7a, 0x75, 0x72, 0x94, 0x79, 0xad, 0xb2,
- 0x81, 0x8b, 0x85, 0x6c, 0x86, 0x88, 0x9e, 0x79, 0x86, 0x9e, 0x7e, 0x91,
- 0x7b, 0x6d, 0x93, 0x91, 0x82, 0x97, 0x6b, 0xa6, 0xaa, 0x9f, 0xa8, 0x74,
- 0x94, 0x7f, 0x63, 0x98, 0x90, 0xa1, 0x8c, 0x7f, 0x71, 0x86, 0x89, 0x95,
- 0x88, 0x80, 0x77, 0x67, 0x85, 0x7d, 0x89, 0x6d, 0x9c, 0x76, 0x72, 0x8d,
- 0x96, 0x94, 0x88, 0x98, 0x9f, 0x94, 0x8e, 0x84, 0x7a, 0x88, 0x79, 0x9f,
- 0x81, 0xa1, 0x7c, 0x8b, 0x71, 0x79, 0x7d, 0x9d, 0x7b, 0x6a, 0x8c, 0x66,
- 0x9e, 0x7b, 0x77, 0x7a, 0xb0, 0x74, 0x7f, 0x8d, 0x8d, 0x71, 0x72, 0x84,
- 0x90, 0x98, 0x7b, 0x89, 0x9b, 0x8e, 0x85, 0x7a, 0x67, 0x8a, 0x72, 0x84,
- 0x82, 0x91, 0x91, 0x7a, 0x85, 0x8a, 0xae, 0x8a, 0x9a, 0x9a, 0x7f, 0x85,
- 0x8a, 0x90, 0x69, 0x7b, 0x76, 0x78, 0x98, 0x54, 0x94, 0x7e, 0x6c, 0x72,
- 0x89, 0x88, 0x82, 0x96, 0x59, 0x95, 0x76, 0x91, 0x94, 0x96, 0x83, 0x84,
- 0x72, 0x8d, 0x97, 0x71, 0x68, 0x8e, 0x88, 0x8b, 0x7c, 0xa9, 0x73, 0x8a,
- 0x95, 0x86, 0x87, 0x96, 0x91, 0x77, 0xb1, 0x88, 0x6e, 0x7d, 0x7c, 0x9f,
- 0x8f, 0x82, 0x79, 0x83, 0xa6, 0x81, 0x89, 0x83, 0x85, 0x9b, 0x7c, 0x68,
- 0x6f, 0x84, 0x7c, 0xa1, 0x8e, 0x80, 0x78, 0x8f, 0x96, 0x77, 0x7e, 0x7b,
- 0x8f, 0x81, 0xa5, 0x84, 0x86, 0x91, 0x7b, 0x73, 0x92, 0x85, 0xa3, 0x7e,
- 0x80, 0x95, 0x7d, 0x5f, 0x8c, 0x94, 0x95, 0x73, 0x95, 0x78, 0x87, 0xa1,
- 0x94, 0x6c, 0xac, 0x6c, 0x77, 0x89, 0x86, 0x9c, 0x82, 0x76, 0x99, 0x93,
- 0x92, 0x88, 0x80, 0x80, 0x85, 0x8a, 0xa8, 0x8f, 0x7a, 0x89, 0x9a, 0x7a,
- 0x8f, 0x91, 0x86, 0x82, 0x7f, 0x82, 0x91, 0x95, 0x85, 0x71, 0x7d, 0x8f,
- 0x83, 0x8c, 0x79, 0x97, 0x7a, 0x9b, 0x91, 0x88, 0xa2, 0x86, 0x8a, 0x80,
- 0xa0, 0x96, 0x8b, 0x7d, 0x76, 0x96, 0x9f, 0x8d, 0x95, 0x8a, 0x94, 0xa0,
- 0x80, 0x95, 0x9b, 0x96, 0x81, 0xa8, 0x59, 0x89, 0x92, 0xb2, 0x83, 0x89,
- 0x85, 0x81, 0x7e, 0x64, 0x77, 0x82, 0x90, 0x96, 0x7e, 0x9f, 0xab, 0x8a,
- 0x6e, 0x9b, 0x90, 0x89, 0x6e, 0x7d, 0x81, 0x65, 0x81, 0x86, 0xa1, 0x93,
- 0x8b, 0x83, 0x81, 0x89, 0x8b, 0x90, 0x7e, 0x97, 0x8e, 0x75, 0x7e, 0x7e,
- 0x7b, 0x81, 0x9a, 0x64, 0x90, 0xab, 0x90, 0x82, 0x8a, 0x82, 0x8d, 0xad,
- 0x90, 0x74, 0x7f, 0x9a, 0x88, 0x92, 0x83, 0x97, 0xa6, 0x6e, 0x9d, 0x81,
- 0xa2, 0x98, 0x74, 0x84, 0x93, 0x85, 0x84, 0x7d, 0xa2, 0x92, 0x92, 0x87,
- 0x73, 0x8b, 0x92, 0x74, 0x96, 0x70, 0x83, 0x86, 0x8a, 0x89, 0x86, 0x88,
- 0x87, 0x7c, 0x7d, 0x81, 0x8d, 0x71, 0x8c, 0x89, 0x70, 0x94, 0x8f, 0x9a,
- 0x83, 0x9d, 0x99, 0x78, 0x74, 0x88, 0x84, 0x9a, 0x95, 0x8b, 0x8e, 0x7f,
- 0xa2, 0xa0, 0x76, 0x93, 0x9b, 0x7c, 0x97, 0x81, 0x83, 0x8c, 0xa1, 0x99,
- 0x9d, 0x7f, 0x87, 0x75, 0xa7, 0x75, 0x89, 0x7e, 0x88, 0x80, 0x8f, 0x84,
- 0x9a, 0x77, 0x8d, 0x90, 0x9d, 0x6c, 0x88, 0x8d, 0x8e, 0x81, 0x97, 0x6d,
- 0x81, 0x88, 0x64, 0x8c, 0x77, 0x8e, 0x91, 0x8a, 0x7f, 0x8a, 0x94, 0x7a,
- 0x89, 0x93, 0x8c, 0x69, 0x85, 0x8c, 0x93, 0x61, 0x7e, 0x89, 0x7e, 0x8a,
- 0x65, 0x8a, 0xa9, 0x7f, 0x80, 0x86, 0x82, 0x90, 0x66, 0x7a, 0x99, 0x71,
- 0x7f, 0x73, 0x8d, 0x94, 0x7d, 0x73, 0x7a, 0x7d, 0x87, 0x7a, 0x97, 0x70,
- 0x81, 0x60, 0x61, 0x7a, 0x91, 0x88, 0x93, 0x7a, 0x9e, 0xa6, 0x92, 0x9d,
- 0x92, 0x67, 0x99, 0x9a, 0xae, 0x71, 0x89, 0xa5, 0x9f, 0xa6, 0x98, 0x89,
- 0x97, 0x90, 0x9b, 0x9a, 0xc0, 0x95, 0x8f, 0x9c, 0x95, 0x93, 0x88, 0x95,
- 0x95, 0xa0, 0x8e, 0x8c, 0xa8, 0x94, 0x6e, 0x9e, 0x6f, 0x7b, 0xa5, 0x96,
- 0x98, 0x90, 0x91, 0x89, 0x93, 0x8f, 0x84, 0xb2, 0x7f, 0x5e, 0xc2, 0x75,
- 0x8f, 0x90, 0x9c, 0xbf, 0x8a, 0x84, 0xa6, 0x85, 0x7d, 0x84, 0x8a, 0xad,
- 0x6f, 0x88, 0xac, 0x77, 0x91, 0x8d, 0x94, 0xac, 0x8f, 0x7f, 0xa1, 0xa5,
- 0x8e, 0x6d, 0x8a, 0x82, 0x85, 0x80, 0x9b, 0x7a, 0x9f, 0x60, 0x95, 0x97,
- 0x90, 0x67, 0x8f, 0x91, 0x86, 0x89, 0x88, 0x89, 0x96, 0x6c, 0x8b, 0x94,
- 0x8a, 0x75, 0x84, 0x96, 0x8a, 0x86, 0x7c, 0x91, 0x74, 0x8f, 0x97, 0x89,
- 0x8f, 0x8e, 0x6b, 0x97, 0x93, 0x89, 0x6b, 0x7e, 0x65, 0xa4, 0xa5, 0x63,
- 0x85, 0x88, 0x81, 0xa3, 0x70, 0x9b, 0x9e, 0x8c, 0x62, 0x73, 0x85, 0xb4,
- 0x88, 0x6e, 0x92, 0x6f, 0x91, 0x88, 0x79, 0x91, 0x7f, 0x7d, 0x9a, 0x6b,
- 0x78, 0x93, 0x7e, 0x79, 0x93, 0x7a, 0x74, 0x91, 0x8d, 0x92, 0xb3, 0x61,
- 0xa3, 0x76, 0x81, 0x99, 0x96, 0x8b, 0x93, 0x8f, 0xa7, 0x6f, 0x8f, 0xa6,
- 0xb2, 0x76, 0xa1, 0x83, 0xa8, 0x8b, 0xae, 0x99, 0x90, 0x6a, 0x97, 0x97,
- 0xaa, 0x95, 0x85, 0x7d, 0x97, 0x94, 0x86, 0x94, 0x89, 0xa4, 0xa9, 0x81,
- 0x89, 0x7c, 0x96, 0xb3, 0x92, 0x7d, 0xa4, 0x6f, 0x6d, 0x92, 0x83, 0xb4,
- 0x7b, 0x94, 0x8c, 0x79, 0x61, 0x6f, 0x8f, 0xb7, 0x88, 0x66, 0xaa, 0x7d,
- 0x89, 0x7f, 0x90, 0xbd, 0x99, 0xac, 0xb1, 0x96, 0x9c, 0x7c, 0x92, 0xb7,
- 0x73, 0x94, 0xad, 0x9d, 0x7c, 0x80, 0x87, 0x96, 0x73, 0x8d, 0xa8, 0x88,
- 0xa9, 0x83, 0x7b, 0x84, 0x9d, 0x99, 0x83, 0x89, 0x9d, 0x7f, 0x7e, 0x86,
- 0x75, 0x83, 0x77, 0x7d, 0x8b, 0x7d, 0x80, 0x9d, 0xa2, 0x94, 0x72, 0x92,
- 0x75, 0x95, 0x99, 0xa0, 0x7b, 0x83, 0x99, 0x89, 0x82, 0x92, 0x5b, 0x9e,
- 0x7c, 0x91, 0x95, 0x79, 0x61, 0x86, 0x60, 0xc7, 0x72, 0x91, 0xb5, 0x88,
- 0x71, 0x8d, 0x85, 0x91, 0x83, 0x74, 0xa8, 0x67, 0x79, 0x77, 0x7f, 0x79,
- 0x68, 0x84, 0x95, 0x69, 0x98, 0x88, 0x74, 0x72, 0x9c, 0x86, 0x87, 0x95,
- 0x90, 0x95, 0x9b, 0x8b, 0xc5, 0x7d, 0x81, 0x8f, 0x88, 0x8c, 0xb0, 0x95,
- 0xa8, 0x8c, 0x84, 0xa0, 0xb0, 0x89, 0x9a, 0x90, 0xaa, 0x88, 0x96, 0x9b,
- 0x88, 0xa9, 0x89, 0x99, 0xb7, 0x82, 0x99, 0xa0, 0x85, 0x70, 0x9c, 0x9a,
- 0x94, 0x74, 0x91, 0x81, 0x76, 0x70, 0x8f, 0xc2, 0x8c, 0x91, 0x8f, 0x69,
- 0x74, 0x7e, 0x6d, 0x9a, 0x80, 0x77, 0xa5, 0x94, 0x8b, 0x6d, 0x82, 0xcf,
- 0x8e, 0x74, 0xc4, 0x86, 0x7f, 0x78, 0x72, 0xb3, 0x78, 0x7a, 0xac, 0x9c,
- 0x7d, 0x77, 0x8d, 0xca, 0x67, 0x8c, 0xd5, 0x8f, 0x7f, 0x71, 0x70, 0x82,
- 0x7e, 0x9f, 0xb0, 0x7f, 0x75, 0x90, 0x79, 0x7b, 0x8d, 0x7b, 0xa6, 0x87,
- 0x98, 0x76, 0x84, 0x96, 0x81, 0x6a, 0x96, 0x86, 0x8e, 0x77, 0xa3, 0x83,
- 0x91, 0x83, 0x8a, 0x6c, 0x74, 0x83, 0x99, 0x7d, 0x7c, 0x8a, 0x88, 0x9a,
- 0x6b, 0x86, 0x59, 0xa3, 0x8a, 0x8e, 0xbb, 0x8a, 0x75, 0x78, 0x68, 0xb5,
- 0x9b, 0x7b, 0xa7, 0x93, 0x5b, 0x6c, 0x6b, 0xa0, 0x74, 0x99, 0xc0, 0x73,
- 0x8b, 0x7e, 0x8e, 0x83, 0x64, 0x7c, 0x7d, 0x7a, 0x98, 0x7d, 0x82, 0x7c,
- 0x8f, 0x7e, 0x74, 0x86, 0xa9, 0x84, 0xba, 0x8f, 0xc7, 0x6f, 0x87, 0xae,
- 0x97, 0x91, 0xad, 0x82, 0xb2, 0x70, 0x8a, 0xa0, 0xb0, 0x7d, 0x95, 0x8d,
- 0xc2, 0x85, 0x80, 0xad, 0x9f, 0x85, 0x8b, 0x76, 0xaa, 0xab, 0x8f, 0xa0,
- 0x89, 0x9b, 0x8a, 0xb3, 0xa0, 0x72, 0xbe, 0x8c, 0x93, 0x7a, 0xa0, 0xad,
- 0x99, 0x6f, 0xa2, 0x79, 0x78, 0x8b, 0x6d, 0xae, 0x75, 0x6f, 0xa1, 0x8d,
- 0x68, 0x81, 0x74, 0xb3, 0x8f, 0x81, 0xc6, 0x96, 0x77, 0x68, 0x85, 0xaf,
- 0x86, 0x9f, 0xbb, 0x8a, 0x7e, 0x8a, 0x86, 0xab, 0x8b, 0x87, 0x94, 0x96,
- 0x99, 0x82, 0x6a, 0xaa, 0x7b, 0x81, 0xa6, 0x9b, 0xb6, 0x73, 0x78, 0x9a,
- 0x8f, 0xaa, 0x93, 0x81, 0x97, 0x7a, 0x72, 0x82, 0x79, 0x81, 0x7c, 0x88,
- 0x8e, 0x79, 0x9d, 0x81, 0x9a, 0x75, 0x9b, 0x89, 0x73, 0x6a, 0xa6, 0x84,
- 0x5c, 0x6f, 0xa0, 0x9d, 0x81, 0x84, 0x3e, 0xaf, 0x94, 0xa1, 0xb8, 0x93,
- 0x81, 0x89, 0x68, 0xd4, 0x87, 0x99, 0x99, 0x95, 0x79, 0x72, 0x81, 0xa1,
- 0x78, 0x7d, 0x8f, 0x7e, 0x87, 0x78, 0x8e, 0x97, 0x7e, 0x96, 0x86, 0x86,
- 0x97, 0x74, 0x6f, 0x7d, 0xa5, 0x81, 0x6f, 0x8e, 0x9e, 0x8b, 0xad, 0xac,
- 0xbd, 0x75, 0x84, 0xa2, 0x93, 0x76, 0xc7, 0x9e, 0xb0, 0x75, 0x89, 0xa4,
- 0x95, 0x92, 0xb5, 0xaa, 0xb9, 0x7d, 0x79, 0xa5, 0x88, 0x70, 0x84, 0x70,
- 0xa3, 0x81, 0xa1, 0xa6, 0x8f, 0x96, 0x96, 0x8d, 0xa5, 0x83, 0xb2, 0x8f,
- 0x88, 0x74, 0x96, 0xbc, 0x8b, 0x81, 0xa4, 0x85, 0x7c, 0x87, 0x64, 0xb4,
- 0x80, 0x88, 0x92, 0x90, 0x78, 0x79, 0x77, 0xa5, 0x79, 0x8b, 0xbd, 0x7d,
- 0x84, 0x8c, 0x96, 0xd4, 0x78, 0x81, 0xa4, 0x8c, 0x97, 0x89, 0x78, 0xc4,
- 0x9f, 0x94, 0xb9, 0x83, 0x76, 0x78, 0x89, 0x86, 0x81, 0x8f, 0xbd, 0xa7,
- 0x88, 0x79, 0x8e, 0x92, 0x86, 0x88, 0xad, 0x8a, 0x7b, 0x7f, 0x80, 0xad,
- 0x7a, 0xaf, 0x8a, 0x93, 0xa6, 0x84, 0x92, 0x8e, 0x84, 0x99, 0x80, 0xae,
- 0x74, 0x7c, 0x95, 0x9c, 0x7b, 0x84, 0x84, 0x84, 0xa4, 0x82, 0x57, 0xb5,
- 0x95, 0xc1, 0xb7, 0xa0, 0x85, 0x7b, 0x69, 0xc3, 0xb1, 0x8e, 0xa0, 0x8e,
- 0x81, 0x88, 0x78, 0x9e, 0x81, 0x97, 0xb2, 0x74, 0x81, 0x84, 0x91, 0x87,
- 0x6f, 0x6f, 0x75, 0x78, 0x92, 0x7a, 0x6d, 0x80, 0x9a, 0x7e, 0x81, 0xa1,
- 0xa8, 0x6d, 0xb5, 0x98, 0xb4, 0x7f, 0x9a, 0xa4, 0x9d, 0x7b, 0xba, 0xaa,
- 0xce, 0x93, 0x79, 0xa5, 0x81, 0x95, 0xa6, 0x7f, 0x8c, 0x8b, 0x96, 0xa4,
- 0xa1, 0x8d, 0x91, 0x97, 0xce, 0x8e, 0x8e, 0x9d, 0x86, 0x7f, 0x97, 0xa3,
- 0x99, 0x75, 0xa3, 0xa0, 0x69, 0x6a, 0x87, 0xa0, 0x9a, 0x80, 0xa2, 0x72,
- 0x6d, 0x85, 0x6b, 0x94, 0x8d, 0x77, 0x9f, 0x84, 0x7f, 0x92, 0x64, 0xaa,
- 0x78, 0x82, 0xa7, 0x8f, 0x84, 0x79, 0x84, 0xb9, 0x92, 0x7c, 0xb6, 0x96,
- 0x9c, 0x99, 0x8f, 0xab, 0xab, 0x8a, 0xa2, 0xab, 0x6d, 0x97, 0x7b, 0xb1,
- 0x9e, 0x6c, 0x9a, 0x99, 0xaa, 0xa3, 0x70, 0x80, 0x81, 0x6f, 0xb6, 0x95,
- 0x93, 0x93, 0x8e, 0x80, 0x86, 0xb0, 0x87, 0x91, 0x8f, 0x8c, 0xa4, 0x86,
- 0x89, 0x8f, 0x93, 0x83, 0x75, 0x7d, 0x9b, 0x86, 0x7d, 0x5a, 0x9d, 0x67,
- 0x9f, 0x78, 0x5c, 0xa5, 0x8e, 0xa2, 0xc1, 0x95, 0x89, 0x84, 0x53, 0xd1,
- 0x7d, 0x9b, 0xc0, 0x8f, 0x73, 0x7f, 0x85, 0x9e, 0x8a, 0x7b, 0xa6, 0x84,
- 0x6c, 0x74, 0x95, 0x93, 0x7a, 0x7a, 0x81, 0x7d, 0x89, 0x86, 0x76, 0x8a,
- 0xad, 0x66, 0x90, 0x90, 0x9d, 0x77, 0xb4, 0xad, 0xac, 0x8e, 0xb3, 0xa5,
- 0x9d, 0x91, 0xd7, 0x94, 0xba, 0x8b, 0x72, 0xa4, 0x93, 0x7e, 0xa7, 0x86,
- 0xae, 0x83, 0x63, 0xa6, 0xa0, 0x78, 0x81, 0x8b, 0xc4, 0x82, 0x8f, 0x98,
- 0xa1, 0x8f, 0x79, 0x9a, 0x92, 0x85, 0x9d, 0x91, 0x92, 0x84, 0x8f, 0x84,
- 0x91, 0x6d, 0x7b, 0x69, 0x75, 0x87, 0x5d, 0x99, 0x92, 0x83, 0xab, 0x8f,
- 0x53, 0x90, 0x7b, 0xa0, 0x71, 0x89, 0xc2, 0x7f, 0x6a, 0x7c, 0x86, 0xb2,
- 0x8d, 0x89, 0xaf, 0x9c, 0x81, 0x8c, 0x84, 0xbe, 0x93, 0x9c, 0xa8, 0x97,
- 0x68, 0x9b, 0x84, 0xa3, 0x8a, 0x77, 0xa5, 0x79, 0x7b, 0x87, 0x86, 0xa5,
- 0x80, 0x83, 0x9e, 0x8d, 0xb1, 0x94, 0x7a, 0x8b, 0xa6, 0xa8, 0x80, 0x98,
- 0x8c, 0x73, 0xa9, 0x7b, 0x91, 0x8f, 0x71, 0x82, 0x68, 0x84, 0xa5, 0x96,
- 0x67, 0x63, 0xa6, 0x71, 0xa7, 0x85, 0x57, 0x9f, 0x91, 0xb2, 0xa6, 0x87,
- 0x80, 0x8f, 0x6a, 0xba, 0x9d, 0xb7, 0xb9, 0x8b, 0x75, 0x7c, 0x6f, 0x9f,
- 0x74, 0x8d, 0xaf, 0x6e, 0x7c, 0x65, 0x6c, 0x8a, 0x7c, 0x81, 0x89, 0x77,
- 0x8b, 0x74, 0x65, 0x9b, 0xa5, 0x6b, 0x92, 0x71, 0xbb, 0x70, 0x99, 0xbf,
- 0xb0, 0x7b, 0x92, 0xb4, 0xa4, 0x84, 0xc4, 0x92, 0xa8, 0x94, 0x7e, 0xcd,
- 0x83, 0x87, 0xaf, 0xa0, 0xa5, 0x94, 0x72, 0xb9, 0x90, 0xa6, 0x9e, 0x9e,
- 0x9b, 0x7a, 0x68, 0xc0, 0x8f, 0x89, 0x72, 0x94, 0x9b, 0x81, 0x81, 0x91,
- 0x88, 0x90, 0xa8, 0x8d, 0x90, 0x78, 0x7c, 0x67, 0x64, 0x8e, 0x55, 0xa1,
- 0x6d, 0x86, 0xa3, 0x6f, 0x5c, 0x7d, 0x79, 0xa3, 0x64, 0x71, 0xd4, 0x87,
- 0x73, 0x85, 0x76, 0xc7, 0x72, 0x86, 0xb2, 0x8c, 0x7b, 0x8d, 0x96, 0xc3,
- 0xad, 0x87, 0xac, 0xa8, 0x84, 0x94, 0x7b, 0xbf, 0x83, 0x74, 0x8e, 0x8c,
- 0x9c, 0x99, 0x88, 0x8e, 0x86, 0x88, 0xae, 0x7f, 0x70, 0x96, 0x6f, 0x74,
- 0x8f, 0x85, 0x7c, 0x86, 0x97, 0x83, 0xa0, 0x6a, 0x8b, 0x82, 0x88, 0x90,
- 0x72, 0x84, 0x9b, 0xa1, 0x6f, 0x72, 0xa4, 0x95, 0xa6, 0x7d, 0x65, 0xbd,
- 0x90, 0xb6, 0x9e, 0x98, 0xa1, 0x94, 0x66, 0xb3, 0x9c, 0xb3, 0xa7, 0x7f,
- 0x91, 0x69, 0x6e, 0xb1, 0x68, 0x7a, 0xaa, 0x91, 0x7c, 0x71, 0x9f, 0x95,
- 0x83, 0x86, 0x76, 0x69, 0x9b, 0x7f, 0x8c, 0x94, 0x9c, 0x89, 0x86, 0x93,
- 0xc1, 0x79, 0x98, 0x9e, 0xb1, 0x90, 0x9b, 0xb7, 0xab, 0x86, 0xc6, 0xa1,
- 0xa9, 0xaa, 0x86, 0xb0, 0x8b, 0x79, 0xb9, 0x85, 0xbe, 0x92, 0x60, 0xc0,
- 0x9f, 0x9a, 0x90, 0x8d, 0xb5, 0x77, 0x95, 0xad, 0x8b, 0x93, 0x8a, 0x93,
- 0x93, 0x7e, 0x86, 0xa6, 0x7d, 0x89, 0x6b, 0x81, 0x93, 0x75, 0x7f, 0x86,
- 0x66, 0x8f, 0x56, 0x8f, 0x84, 0x75, 0x9e, 0x77, 0x78, 0x89, 0x62, 0xb3,
- 0x78, 0x76, 0xb5, 0x92, 0x7f, 0x80, 0x7a, 0xb9, 0x7d, 0x80, 0xc2, 0xb9,
- 0x7d, 0x8f, 0x8f, 0x8c, 0xa0, 0x78, 0xa2, 0xaf, 0x68, 0x98, 0x77, 0xac,
- 0x96, 0x77, 0x96, 0x99, 0x84, 0xb1, 0x72, 0x8e, 0x96, 0xa4, 0xa9, 0x8e,
- 0x84, 0x7b, 0x85, 0x8d, 0x8f, 0x83, 0x83, 0x7f, 0x85, 0x6e, 0xa4, 0x98,
- 0xab, 0x83, 0x90, 0x8e, 0x77, 0x8e, 0xab, 0x9c, 0x73, 0x79, 0x8d, 0x6e,
- 0xa0, 0x97, 0x68, 0xa7, 0x8a, 0xbd, 0x95, 0x96, 0x96, 0x8b, 0x72, 0xc7,
- 0x8d, 0x8c, 0xa5, 0x83, 0x9b, 0x8b, 0x6c, 0xac, 0x62, 0x78, 0xae, 0x78,
- 0x71, 0x7a, 0x8d, 0xae, 0x91, 0x87, 0x90, 0x82, 0x9b, 0x83, 0x90, 0x97,
- 0xb0, 0x96, 0x82, 0xa5, 0xa9, 0x76, 0xa5, 0xa0, 0xac, 0xa1, 0x93, 0x94,
- 0xb7, 0x91, 0xbb, 0x9b, 0xa4, 0xa5, 0x8c, 0xb5, 0x95, 0x7b, 0x92, 0x91,
- 0xb0, 0x97, 0x73, 0xb9, 0x86, 0xa7, 0x92, 0x98, 0x9e, 0x70, 0x77, 0xba,
- 0x96, 0x7b, 0xa6, 0x86, 0x97, 0x85, 0x8e, 0xaa, 0x93, 0x97, 0x8f, 0x8b,
- 0x8d, 0x79, 0x84, 0x7e, 0x70, 0x95, 0x52, 0x8f, 0x62, 0x75, 0x8b, 0x8b,
- 0x7b, 0x8b, 0x79, 0xaf, 0x90, 0x6d, 0xc8, 0x8d, 0x84, 0x8c, 0x72, 0xaf,
- 0x70, 0x8d, 0xa5, 0x8a, 0x76, 0x97, 0x87, 0x8e, 0xa9, 0x83, 0xb2, 0x8d,
- 0x7e, 0x9b, 0x76, 0xc2, 0xa2, 0x72, 0xc5, 0x87, 0x75, 0xb7, 0x92, 0x95,
- 0x9e, 0xa0, 0xc3, 0x82, 0x8d, 0x8f, 0x7d, 0x85, 0x90, 0x99, 0x7b, 0x82,
- 0x87, 0x87, 0xa0, 0x87, 0x9a, 0x8b, 0xa2, 0xa4, 0x67, 0x93, 0xa5, 0xbb,
- 0x73, 0x5f, 0x8c, 0x60, 0xa5, 0x7d, 0x6c, 0xb3, 0xb2, 0xb3, 0xa9, 0xa9,
- 0x8d, 0x8d, 0x67, 0xd7, 0x63, 0x99, 0xaa, 0x83, 0x88, 0x6a, 0x6f, 0x9e,
- 0x5e, 0x9e, 0x9d, 0x81, 0x84, 0x6e, 0x98, 0x90, 0x89, 0x7c, 0x95, 0x7d,
- 0x81, 0x8a, 0xa2, 0x8c, 0x92, 0x85, 0x80, 0x92, 0xac, 0x80, 0x9b, 0x9b,
- 0xc3, 0x8c, 0x95, 0xbc, 0xaa, 0x7c, 0xb5, 0x8d, 0xa1, 0xb8, 0x70, 0xb6,
- 0x8c, 0x92, 0xa8, 0x8e, 0xa3, 0x76, 0x6c, 0xbe, 0xa0, 0x8c, 0x92, 0x8e,
- 0xa1, 0x83, 0x76, 0xb2, 0x91, 0x7b, 0x8e, 0x87, 0x7f, 0x89, 0x8a, 0xa1,
- 0x91, 0xa0, 0x7a, 0x95, 0x7b, 0x86, 0x99, 0x92, 0x78, 0x8a, 0x62, 0x9e,
- 0x7b, 0x7b, 0x89, 0x79, 0x78, 0x87, 0x82, 0x94, 0x7d, 0x91, 0x96, 0x79,
- 0x7b, 0x8d, 0x80, 0xa7, 0x88, 0x95, 0xa6, 0x8f, 0x7d, 0x95, 0x79, 0xa2,
- 0x91, 0x9b, 0x9d, 0x90, 0x79, 0xa4, 0x88, 0x98, 0x9b, 0x7a, 0xa5, 0x7f,
- 0x71, 0x9c, 0x87, 0x96, 0x8c, 0x8f, 0xbc, 0x74, 0x95, 0x99, 0x7f, 0x78,
- 0x8c, 0x63, 0x7c, 0x7a, 0x92, 0x8c, 0xa8, 0x78, 0xa8, 0x89, 0x9a, 0x86,
- 0x69, 0x7e, 0xa1, 0xc3, 0x57, 0x68, 0x84, 0x89, 0xa9, 0x8d, 0x6f, 0xa9,
- 0x8a, 0xab, 0xa5, 0xad, 0x94, 0x83, 0x6b, 0xa7, 0x7e, 0x95, 0x9b, 0x7f,
- 0x8b, 0x78, 0x73, 0x90, 0x65, 0x8d, 0xb1, 0x91, 0x84, 0x65, 0x90, 0xb4,
- 0x8c, 0x89, 0x94, 0x7c, 0x99, 0x8b, 0x98, 0xb7, 0xb0, 0x91, 0x9e, 0x88,
- 0xbd, 0xa0, 0xa4, 0xb9, 0xad, 0x96, 0x97, 0xa3, 0xb6, 0x81, 0xba, 0x9b,
- 0xbc, 0xa9, 0x94, 0xb9, 0xa0, 0x85, 0x8e, 0xa1, 0xac, 0x87, 0x65, 0xa6,
- 0x98, 0x8e, 0xaa, 0xa3, 0xa3, 0x7f, 0x79, 0xb4, 0x93, 0x76, 0x90, 0x99,
- 0x8b, 0x90, 0x84, 0xa6, 0x90, 0x8f, 0x88, 0xa6, 0x89, 0x83, 0x86, 0x7a,
- 0x5d, 0x96, 0x71, 0xa5, 0x64, 0x94, 0x9a, 0x85, 0x7c, 0xa1, 0x96, 0x9d,
- 0x76, 0x8f, 0x95, 0xa0, 0x7f, 0x8c, 0x80, 0xc7, 0x6c, 0x7d, 0xb7, 0xb2,
- 0x82, 0x8e, 0x82, 0xbd, 0xb3, 0x82, 0x99, 0x9b, 0x80, 0x94, 0x8c, 0x94,
- 0x94, 0x6b, 0xc6, 0xa9, 0x81, 0x9f, 0x8c, 0x7e, 0x87, 0x88, 0xb3, 0x7d,
- 0x88, 0x8c, 0x81, 0x81, 0x7e, 0x7e, 0x86, 0x87, 0x96, 0x85, 0xb4, 0x87,
- 0xab, 0x91, 0x8f, 0xa1, 0x72, 0x83, 0xa4, 0x89, 0x6b, 0x75, 0x85, 0x7c,
- 0x94, 0x85, 0x6f, 0xad, 0x91, 0xae, 0xa4, 0xa5, 0xa7, 0x8e, 0x6c, 0xb2,
- 0x73, 0x99, 0x96, 0x92, 0x89, 0x81, 0x7d, 0x88, 0x60, 0x8d, 0x94, 0x83,
- 0x99, 0x68, 0x86, 0xa2, 0x94, 0x8e, 0x82, 0x76, 0x89, 0x8d, 0x98, 0x86,
- 0x94, 0x90, 0x83, 0x7d, 0xad, 0x94, 0xa6, 0x90, 0xcb, 0x96, 0xa2, 0xb2,
- 0xb6, 0x89, 0xc4, 0x9d, 0xc7, 0xa5, 0x75, 0xc3, 0x92, 0x8c, 0x8e, 0xad,
- 0x96, 0x94, 0x8e, 0xab, 0x94, 0x90, 0xa8, 0x84, 0xb5, 0x84, 0x66, 0xce,
- 0x74, 0x8c, 0x93, 0x8d, 0x8f, 0x95, 0x8b, 0xa1, 0x7b, 0xa1, 0x79, 0x9e,
- 0x81, 0xa4, 0xa0, 0x98, 0x5f, 0x78, 0x8e, 0x97, 0x6f, 0x81, 0x96, 0x8d,
- 0x70, 0x93, 0x72, 0x9c, 0x7b, 0x98, 0x8b, 0x8a, 0x8f, 0x8b, 0x6c, 0xa9,
- 0x81, 0x99, 0xb3, 0xa3, 0x71, 0x9c, 0x8b, 0x94, 0xa6, 0x8a, 0xb8, 0xa0,
- 0x7b, 0x98, 0x74, 0x9f, 0x92, 0x92, 0xb2, 0x89, 0x81, 0xa8, 0x87, 0x97,
- 0x96, 0x86, 0xa4, 0x7b, 0x63, 0x8e, 0x86, 0x7d, 0x76, 0x81, 0x93, 0x94,
- 0x98, 0x8b, 0xaf, 0x6d, 0xab, 0x9b, 0x85, 0x9b, 0x91, 0x86, 0x95, 0x95,
- 0x65, 0x89, 0x9e, 0x6b, 0xa4, 0x82, 0x68, 0xb5, 0x8b, 0xd1, 0x9d, 0x93,
- 0x7d, 0x67, 0x5e, 0xba, 0x9b, 0x94, 0x93, 0x8d, 0x88, 0x73, 0x7c, 0x8e,
- 0x7d, 0x83, 0x9a, 0x82, 0xa4, 0x62, 0x9a, 0x8d, 0x86, 0xa0, 0x7b, 0x72,
- 0xa9, 0x84, 0xa7, 0x94, 0xb2, 0x98, 0x8f, 0x81, 0xbe, 0x84, 0x9d, 0x94,
- 0x9c, 0x9a, 0x94, 0x8f, 0xb1, 0x82, 0xb1, 0x82, 0xb1, 0xb2, 0x78, 0xa7,
- 0x95, 0x99, 0x8b, 0x8c, 0xb1, 0x81, 0x5b, 0xbb, 0x88, 0x7a, 0x90, 0xa3,
- 0x8d, 0x78, 0x6f, 0xbf, 0x8c, 0x93, 0xa1, 0x8e, 0x9f, 0x98, 0x88, 0xb3,
- 0x7e, 0x82, 0x8a, 0x8e, 0x7d, 0x8a, 0x96, 0x6a, 0x6c, 0x7b, 0x91, 0x94,
- 0x6f, 0x89, 0x9a, 0x84, 0x73, 0x8b, 0x8c, 0x91, 0x7d, 0x8e, 0x9e, 0x80,
- 0x88, 0x81, 0x78, 0xaf, 0x86, 0xa5, 0xa2, 0x8d, 0x6a, 0x8a, 0x75, 0xa1,
- 0x83, 0x87, 0xaf, 0x7d, 0x6c, 0xa3, 0x65, 0x77, 0x89, 0x91, 0x9a, 0xa1,
- 0xa1, 0xaf, 0x78, 0x94, 0x93, 0xb2, 0xaf, 0x92, 0x74, 0x7a, 0xa7, 0x7b,
- 0x8f, 0x9c, 0x86, 0x8d, 0x8f, 0x79, 0xb0, 0xb3, 0x97, 0x82, 0x8e, 0x92,
- 0x92, 0x81, 0xa7, 0xbc, 0x6e, 0x6e, 0x89, 0xa5, 0x9a, 0x8d, 0x84, 0xb6,
- 0x83, 0xae, 0xa5, 0xa7, 0xae, 0x86, 0x6b, 0xb9, 0x89, 0xb0, 0x8f, 0x82,
- 0x8f, 0x6f, 0x83, 0x98, 0x6a, 0x98, 0x9a, 0x85, 0x9f, 0x78, 0x93, 0x8d,
- 0x83, 0x88, 0x88, 0x7e, 0x97, 0x99, 0x8a, 0x9b, 0xb0, 0x90, 0x86, 0x88,
- 0xb5, 0x90, 0xb3, 0xaa, 0xad, 0x96, 0x93, 0xa3, 0x9d, 0x81, 0xa3, 0x9a,
- 0x9f, 0x99, 0x90, 0x9c, 0x9e, 0x8e, 0x88, 0x93, 0xa8, 0x94, 0x62, 0xa6,
- 0x94, 0x92, 0xa1, 0x86, 0xb7, 0x8a, 0x6a, 0xa6, 0x81, 0x7e, 0x7b, 0x80,
- 0x89, 0x8f, 0x74, 0xa6, 0x72, 0x91, 0xa6, 0x9b, 0x73, 0x97, 0x7e, 0x6f,
- 0x70, 0x8d, 0x73, 0x98, 0x80, 0x90, 0x8f, 0x7e, 0x83, 0x77, 0x84, 0x92,
- 0x7f, 0x8c, 0x91, 0xa6, 0x99, 0x90, 0x9d, 0xb1, 0x88, 0x85, 0x89, 0x85,
- 0x7c, 0x9f, 0x7e, 0xb0, 0xaa, 0x84, 0xa0, 0x8e, 0x74, 0x93, 0x78, 0x90,
- 0x9a, 0x8b, 0x8e, 0x97, 0x8f, 0x9f, 0x7c, 0x83, 0x8a, 0x88, 0xa5, 0x8f,
- 0x8b, 0x74, 0x84, 0x9a, 0x7f, 0x91, 0x88, 0x77, 0x9c, 0x91, 0xbc, 0x93,
- 0x9c, 0x82, 0x89, 0x9b, 0x8a, 0x7d, 0xb7, 0xb8, 0x6f, 0x68, 0xb5, 0x8e,
- 0xb4, 0x86, 0x8c, 0xb3, 0x94, 0xb6, 0xa4, 0x93, 0x98, 0x8b, 0x70, 0xb3,
- 0x96, 0xaa, 0x87, 0x89, 0x99, 0x68, 0x74, 0xa4, 0x69, 0x9e, 0x8e, 0x6b,
- 0x9f, 0x6b, 0x95, 0x9c, 0x88, 0x89, 0x8a, 0x86, 0x8d, 0x75, 0x94, 0x88,
- 0xa0, 0x94, 0x77, 0x8c, 0x9c, 0x8d, 0x8e, 0xa4, 0xac, 0xa7, 0x8a, 0x9b,
- 0xa9, 0x81, 0xab, 0xac, 0xaf, 0xaf, 0x87, 0xbb, 0x9b, 0x95, 0x8e, 0x9e,
- 0x9f, 0xa1, 0x6c, 0xb4, 0x98, 0x8f, 0x81, 0x8d, 0x98, 0x8f, 0x78, 0x96,
- 0x89, 0x86, 0x6c, 0x91, 0x8d, 0x9f, 0x95, 0x9f, 0x6b, 0x7f, 0x93, 0x7c,
- 0x96, 0x8e, 0x8a, 0x58, 0x80, 0x8e, 0x7a, 0x93, 0x8b, 0x78, 0x99, 0x92,
- 0x62, 0x8e, 0x83, 0x8e, 0x87, 0x83, 0x86, 0x99, 0x93, 0x92, 0x80, 0x95,
- 0xa2, 0x72, 0xa2, 0x97, 0x78, 0x87, 0x7b, 0xa3, 0x99, 0x78, 0x98, 0x9c,
- 0x80, 0x9b, 0x5e, 0x8a, 0x9c, 0x99, 0xa6, 0x7a, 0x8e, 0x99, 0x7a, 0x8e,
- 0x8b, 0x76, 0x9b, 0x89, 0x80, 0x8e, 0x83, 0x8a, 0x80, 0x7c, 0x80, 0x74,
- 0x95, 0x8c, 0xbf, 0x7e, 0xa8, 0x7a, 0x99, 0x7d, 0x7d, 0x73, 0xb4, 0xae,
- 0x88, 0x76, 0xae, 0x78, 0xaa, 0x65, 0x94, 0xbe, 0x97, 0xaf, 0xa4, 0x91,
- 0x9c, 0x95, 0x6c, 0xbe, 0x82, 0xb1, 0x9b, 0x91, 0x85, 0x7d, 0x66, 0x9c,
- 0x99, 0xbd, 0xa3, 0x88, 0xa8, 0x73, 0x81, 0x94, 0x92, 0x8e, 0x90, 0x8d,
- 0xaf, 0x75, 0x86, 0x9b, 0x8b, 0x8b, 0x8d, 0x74, 0xbd, 0x85, 0x97, 0x8b,
- 0x9d, 0xba, 0x90, 0xa8, 0x9d, 0x72, 0xa5, 0xa8, 0xbf, 0xbb, 0x7b, 0xb6,
- 0xad, 0x94, 0x6f, 0x9a, 0xa7, 0x97, 0x78, 0x9c, 0x98, 0x8d, 0x8c, 0x93,
- 0xb8, 0xa8, 0x7f, 0x9d, 0x98, 0x7f, 0x8f, 0x8a, 0x8d, 0xa8, 0x86, 0x7b,
- 0x5d, 0x89, 0x8a, 0x83, 0x8c, 0x8b, 0x81, 0x56, 0x7c, 0x87, 0x89, 0xa6,
- 0x75, 0x7c, 0x92, 0x74, 0x96, 0x92, 0x78, 0x8d, 0x8d, 0x98, 0xae, 0x7a,
- 0x95, 0x8f, 0x8b, 0x9c, 0x95, 0x9f, 0xae, 0x93, 0x7b, 0x93, 0x8c, 0x9a,
- 0x79, 0x74, 0x94, 0x6e, 0x7e, 0x8f, 0x64, 0x9f, 0x9c, 0x88, 0x8f, 0x8e,
- 0x84, 0x8d, 0x89, 0x95, 0x96, 0x8f, 0x9d, 0x60, 0x85, 0x86, 0x7c, 0x93,
- 0x8d, 0x68, 0x83, 0x7c, 0x94, 0x87, 0xb8, 0xa2, 0x9d, 0x82, 0x8e, 0x84,
- 0x6c, 0x73, 0xa8, 0xbc, 0x84, 0x85, 0xa2, 0x79, 0x92, 0x64, 0x69, 0xa9,
- 0x82, 0xa7, 0x9d, 0x95, 0x8e, 0x6f, 0x9f, 0xa7, 0x97, 0xb1, 0x9d, 0x8e,
- 0xa1, 0x70, 0x80, 0x9e, 0x8e, 0x91, 0xa0, 0xaa, 0x81, 0x5b, 0x98, 0x8f,
- 0xa0, 0xaa, 0x83, 0x7a, 0x91, 0x7a, 0x73, 0x80, 0xa6, 0x9a, 0x80, 0x7d,
- 0x9e, 0x75, 0x7b, 0xa3, 0xad, 0x92, 0x98, 0xc0, 0xa1, 0x80, 0x88, 0xa2,
- 0xa5, 0xa4, 0x7e, 0x9b, 0xa0, 0x80, 0x6e, 0xa0, 0x9f, 0xa3, 0x8a, 0x8f,
- 0xa2, 0x93, 0x86, 0x8d, 0x8f, 0x93, 0x7e, 0x90, 0x98, 0x83, 0x7d, 0x9b,
- 0x9f, 0x9a, 0x97, 0x83, 0x6e, 0x8d, 0x94, 0x6c, 0x7b, 0x7f, 0x73, 0x65,
- 0x6a, 0x93, 0x8a, 0x94, 0x83, 0x89, 0x7d, 0x7b, 0x77, 0x8a, 0x7a, 0x9b,
- 0x8e, 0x8d, 0x94, 0x89, 0x86, 0x83, 0x7c, 0x8e, 0x8b, 0x90, 0xab, 0x99,
- 0x81, 0x8e, 0x77, 0x9c, 0x8c, 0x82, 0x97, 0x8f, 0x78, 0x91, 0x5f, 0xa1,
- 0x8b, 0x83, 0xa9, 0x8d, 0x7b, 0x97, 0x77, 0x80, 0x84, 0x7e, 0x9e, 0x75,
- 0xa3, 0x86, 0x67, 0x7c, 0x80, 0x6d, 0x77, 0x75, 0x88, 0x75, 0xad, 0x7a,
- 0x93, 0x89, 0x8c, 0x87, 0x7a, 0x79, 0xb2, 0xa1, 0x69, 0x80, 0xb5, 0x7a,
- 0xa6, 0x7b, 0x95, 0xac, 0x95, 0xa9, 0x98, 0xa4, 0xad, 0x83, 0x8d, 0xbe,
- 0xa4, 0x98, 0xad, 0x7d, 0x8b, 0x65, 0x65, 0xad, 0x6a, 0xae, 0xa3, 0xa8,
- 0x9c, 0x63, 0x90, 0x91, 0x6d, 0x9a, 0x81, 0x98, 0x86, 0x6a, 0x83, 0x84,
- 0x94, 0x9c, 0x77, 0x86, 0xc2, 0x7f, 0x9b, 0xa9, 0xad, 0xae, 0xa7, 0xa6,
- 0xd4, 0x70, 0x9d, 0xb5, 0xaa, 0xdb, 0x8f, 0xa3, 0xa5, 0x87, 0x88, 0x9e,
- 0xa9, 0x9f, 0x62, 0xa7, 0xa2, 0x8e, 0x7d, 0x8a, 0x9d, 0xa2, 0x6b, 0xa7,
- 0x96, 0x6d, 0x76, 0x8c, 0x9b, 0x8c, 0x86, 0x86, 0x93, 0x7c, 0x9d, 0x7c,
- 0x7e, 0x93, 0x5c, 0x79, 0x76, 0x8c, 0x8a, 0x87, 0x79, 0x97, 0x9a, 0x7a,
- 0x85, 0x8c, 0x7f, 0x85, 0x7a, 0xa1, 0xa7, 0x72, 0x87, 0x7f, 0x96, 0x9e,
- 0x92, 0x92, 0x9e, 0xa0, 0x72, 0x99, 0x7a, 0xb0, 0x8c, 0x8d, 0xa3, 0x9b,
- 0x91, 0xa6, 0x63, 0x94, 0x8b, 0x81, 0xbb, 0x94, 0x79, 0x95, 0x99, 0x9a,
- 0xa0, 0x7a, 0x96, 0x72, 0x82, 0x9a, 0x83, 0x7f, 0x72, 0x7f, 0x6d, 0x75,
- 0x91, 0x7f, 0xbc, 0x84, 0x9a, 0x81, 0x95, 0x69, 0x7d, 0x6d, 0xa2, 0xa8,
- 0x7e, 0x64, 0xac, 0x86, 0x85, 0x6d, 0x99, 0xaa, 0x7e, 0x79, 0x9c, 0xa0,
- 0xa4, 0x77, 0x99, 0xac, 0xa8, 0x8d, 0xb7, 0xa2, 0xa3, 0x61, 0x82, 0x98,
- 0x84, 0x8e, 0xa1, 0x8c, 0x88, 0x82, 0x6f, 0x7d, 0x88, 0x80, 0x7a, 0x8a,
- 0x8c, 0x6d, 0x87, 0x6f, 0xab, 0x8f, 0x8b, 0x76, 0xa0, 0x7d, 0x9f, 0xab,
- 0xb0, 0xb8, 0x9c, 0x8d, 0xb8, 0x81, 0x89, 0x94, 0xa8, 0xc8, 0x92, 0x9b,
- 0x8d, 0x83, 0x7b, 0xaf, 0x97, 0x94, 0x6e, 0xa5, 0x9b, 0x97, 0x89, 0x8d,
- 0xaa, 0x8a, 0x66, 0x88, 0x93, 0x84, 0xa1, 0x88, 0xa0, 0x99, 0x85, 0x89,
- 0x7d, 0x84, 0x8b, 0x6a, 0x92, 0xa1, 0x74, 0x76, 0x73, 0x87, 0x7a, 0x9a,
- 0x77, 0x86, 0x89, 0x5f, 0x7f, 0x8b, 0x7f, 0x8d, 0x7e, 0x81, 0x95, 0x8a,
- 0x7d, 0x85, 0x74, 0x9a, 0x87, 0x8c, 0x9e, 0xae, 0x80, 0x88, 0x7d, 0x8b,
- 0xaa, 0x79, 0x7c, 0x97, 0x79, 0x90, 0x7b, 0x97, 0x97, 0x9f, 0xa1, 0xa2,
- 0xab, 0x97, 0x69, 0x7a, 0x8d, 0x9f, 0x9f, 0x89, 0x90, 0x8c, 0x66, 0x98,
- 0x6e, 0x86, 0x7b, 0x6e, 0x86, 0x8a, 0xb2, 0xa6, 0x93, 0x7d, 0x8c, 0x81,
- 0x7e, 0x84, 0xa6, 0xb6, 0x83, 0x92, 0xa0, 0x88, 0x90, 0x5f, 0x7c, 0x92,
- 0x98, 0x94, 0x92, 0x98, 0xa7, 0x65, 0x90, 0xa2, 0xa2, 0x9b, 0xa6, 0x7d,
- 0x8b, 0x5a, 0x94, 0x95, 0x9b, 0xa5, 0x99, 0xa5, 0x7e, 0x61, 0x9a, 0x7a,
- 0x8b, 0x77, 0x87, 0x76, 0x9d, 0x72, 0x9a, 0x84, 0x98, 0x94, 0x92, 0x73,
- 0xae, 0x78, 0x8e, 0xaa, 0xa0, 0xc3, 0x7a, 0xa4, 0xa0, 0x75, 0xa9, 0xae,
- 0x8c, 0xd6, 0x87, 0x8f, 0x9f, 0x8c, 0x9b, 0x90, 0x99, 0x97, 0x73, 0x8f,
- 0x9b, 0x9c, 0x8c, 0x89, 0xa5, 0x84, 0x8f, 0x7b, 0x8b, 0x7f, 0x97, 0x98,
- 0x8d, 0x7b, 0x94, 0x9d, 0x9c, 0x8e, 0x92, 0x89, 0x88, 0x8d, 0x6c, 0x63,
- 0x73, 0x81, 0x72, 0x8a, 0x88, 0x8a, 0x9f, 0x79, 0x81, 0x82, 0x9a, 0xa9,
- 0x7a, 0x92, 0x7d, 0x76, 0x7b, 0x7a, 0x6a, 0xbe, 0x91, 0x7d, 0x86, 0xad,
- 0x84, 0x86, 0x6c, 0x91, 0x91, 0x9f, 0x92, 0x6b, 0x95, 0x98, 0x84, 0xa0,
- 0x8f, 0x8b, 0x9e, 0x7f, 0x9f, 0x97, 0x7e, 0x87, 0x80, 0x9e, 0x79, 0x8d,
- 0x68, 0x87, 0x88, 0x7d, 0x89, 0x81, 0x6d, 0x85, 0x80, 0x82, 0xa0, 0x97,
- 0xa3, 0x72, 0x94, 0x74, 0x8e, 0x56, 0x96, 0x98, 0x91, 0x6f, 0xa0, 0xae,
- 0x7c, 0x6e, 0x8e, 0xa9, 0x7c, 0x80, 0x87, 0xa3, 0x9e, 0x57, 0x8e, 0xb5,
- 0x87, 0xa6, 0x87, 0x79, 0x8f, 0x55, 0x8a, 0x81, 0x97, 0x6c, 0x9b, 0x99,
- 0x78, 0x5c, 0x82, 0x80, 0x91, 0x76, 0x80, 0x91, 0x8b, 0x65, 0x89, 0x7d,
- 0xa9, 0x95, 0x89, 0x97, 0x96, 0x6a, 0x89, 0xad, 0x92, 0x9f, 0xb6, 0x82,
- 0x88, 0x79, 0x9d, 0xa5, 0x9c, 0xae, 0x9a, 0x93, 0x77, 0x8e, 0x8a, 0xb5,
- 0x84, 0xb0, 0x76, 0xa2, 0x89, 0xa0, 0x96, 0x7a, 0xa5, 0x8e, 0x7e, 0x74,
- 0x8d, 0x89, 0x89, 0x9e, 0x93, 0x95, 0x90, 0x78, 0x93, 0x8f, 0xa5, 0x7c,
- 0x9d, 0x7c, 0x77, 0x85, 0x81, 0x92, 0x7c, 0x87, 0x92, 0x82, 0x98, 0xa3,
- 0x63, 0x76, 0x9b, 0x91, 0x7b, 0x8e, 0x97, 0x7e, 0x66, 0x90, 0x63, 0xb4,
- 0x71, 0x88, 0x86, 0x8e, 0x6f, 0x89, 0x7a, 0x88, 0x93, 0x7f, 0x96, 0xa8,
- 0x7d, 0x88, 0x88, 0x86, 0x7b, 0x91, 0x88, 0x6b, 0xa6, 0x8b, 0x69, 0x78,
- 0x82, 0x80, 0x83, 0x6b, 0xaf, 0x81, 0x7b, 0x64, 0x8f, 0x78, 0x6e, 0x7f,
- 0x86, 0x91, 0x92, 0xa3, 0xa0, 0x97, 0x82, 0x88, 0x92, 0x90, 0x9e, 0x89,
- 0x9d, 0x7b, 0x96, 0x82, 0xa3, 0x8c, 0x7f, 0x84, 0x7a, 0x6c, 0x60, 0x85,
- 0xa9, 0x74, 0x83, 0xa2, 0x89, 0x87, 0x9b, 0x77, 0x9b, 0x9a, 0x99, 0x84,
- 0x7c, 0x9c, 0x8d, 0x90, 0x8d, 0x7b, 0x74, 0x77, 0x93, 0x8c, 0x6c, 0x8b,
- 0x85, 0x78, 0x7f, 0x7d, 0x75, 0x7f, 0x7e, 0x85, 0x8f, 0x7d, 0x62, 0x8c,
- 0x7c, 0xad, 0x7f, 0x83, 0xa1, 0xa1, 0x97, 0x7b, 0x72, 0x82, 0x9d, 0x81,
- 0x94, 0x81, 0x8d, 0x9f, 0x6f, 0x8f, 0x9d, 0x89, 0x6a, 0x7e, 0x7f, 0x7f,
- 0x8d, 0x7e, 0x91, 0x86, 0x7d, 0x8a, 0x7e, 0x70, 0x7b, 0x9b, 0x6e, 0x5f,
- 0xa8, 0x7a, 0x73, 0x8a, 0x7a, 0x71, 0x90, 0x95, 0x8d, 0x78, 0x7b, 0x72,
- 0x5e, 0x89, 0x62, 0xa1, 0x87, 0x7f, 0x83, 0x75, 0x98, 0x7f, 0x76, 0x72,
- 0x8f, 0x9b, 0x7a, 0x8b, 0xa1, 0x7f, 0x60, 0x99, 0x96, 0x6e, 0x67, 0x76,
- 0x88, 0x98, 0x6c, 0x7b, 0x9b, 0x8d, 0x5f, 0x89, 0x7c, 0x81, 0x79, 0x86,
- 0x69, 0x9e, 0x83, 0x65, 0x8e, 0x82, 0x83, 0x89, 0x85, 0x7f, 0x90, 0x80,
- 0xa2, 0x81, 0x85, 0x83, 0x8e, 0x94, 0x94, 0x75, 0x86, 0x87, 0x9a, 0xb2,
- 0x82, 0x99, 0x85, 0x7f, 0x8c, 0x7e, 0x81, 0x9a, 0x81, 0x7d, 0x87, 0x81,
- 0xa3, 0x8c, 0x8d, 0x85, 0x8d, 0x96, 0x86, 0x7c, 0xa7, 0x87, 0x7e, 0x9d,
- 0x63, 0xa8, 0x7c, 0x97, 0xa2, 0xa4, 0x7e, 0x87, 0x93, 0x9e, 0x89, 0x8d,
- 0x6b, 0x6d, 0x9d, 0x9b, 0x78, 0x8a, 0x8e, 0x7f, 0x7b, 0xa5, 0x6e, 0x8c,
- 0x89, 0x88, 0x73, 0x7e, 0x77, 0x9d, 0xa6, 0xa7, 0x77, 0x87, 0x7e, 0x7e,
- 0x97, 0x84, 0x6b, 0x59, 0x60, 0x90, 0x85, 0x76, 0x8f, 0x61, 0x7f, 0x94,
- 0x8f, 0x84, 0x8b, 0x7f, 0x73, 0x77, 0x73, 0x71, 0x8a, 0x9b, 0x7b, 0x89,
- 0x97, 0x8f, 0x76, 0x63, 0xa3, 0xa1, 0x6b, 0x7c, 0x62, 0x95, 0x8e, 0xa3,
- 0x9f, 0x89, 0x8f, 0x7f, 0x92, 0x7c, 0xa2, 0xa4, 0xa6, 0x92, 0x89, 0x93,
- 0x74, 0x73, 0x73, 0x96, 0xad, 0x9b, 0x87, 0xac, 0x91, 0x8a, 0xa0, 0x70,
- 0x70, 0x7e, 0x8f, 0x74, 0x75, 0xaf, 0x8d, 0x82, 0x8e, 0x82, 0x96, 0x7d,
- 0x69, 0x9c, 0x64, 0xa2, 0x82, 0x89, 0x83, 0x9d, 0x83, 0x88, 0x62, 0x92,
- 0x72, 0x89, 0x6d, 0x7f, 0x92, 0x70, 0x8e, 0x80, 0x7e, 0x8d, 0x91, 0x85,
- 0x8d, 0x89, 0x83, 0x96, 0x90, 0x96, 0x9c, 0xa6, 0x8a, 0x73, 0x89, 0x79,
- 0xa9, 0x70, 0x80, 0x78, 0x96, 0x80, 0x7b, 0x85, 0xa5, 0x80, 0x93, 0x95,
- 0xc5, 0x74, 0x81, 0x88, 0xa2, 0x93, 0x86, 0x9c, 0xa3, 0x6d, 0x92, 0x8a,
- 0x92, 0x99, 0x98, 0x65, 0xad, 0x63, 0x9d, 0x95, 0x99, 0x89, 0x7f, 0x7a,
- 0x99, 0x91, 0x7f, 0x78, 0x90, 0x8f, 0x80, 0x85, 0xa1, 0x68, 0x9d, 0x6c,
- 0x83, 0x8f, 0x7c, 0x5e, 0x99, 0x7b, 0x80, 0x91, 0x66, 0x8a, 0x92, 0xb3,
- 0x7a, 0x99, 0x91, 0x7e, 0x7d, 0x96, 0x69, 0x9e, 0x7c, 0x89, 0xad, 0x8f,
- 0x9d, 0x90, 0x85, 0x8e, 0x72, 0xa9, 0x89, 0x83, 0x7c, 0x82, 0x70, 0x82,
- 0x6b, 0x79, 0x75, 0x8d, 0x77, 0x9b, 0x7c, 0x8f, 0x8a, 0x95, 0x87, 0x9f,
- 0x7c, 0x90, 0x87, 0x70, 0x83, 0x83, 0x98, 0x9f, 0x85, 0x86, 0x8d, 0x81,
- 0x87, 0x87, 0x87, 0x9d, 0x8f, 0x9d, 0x7c, 0x98, 0xa2, 0xac, 0x88, 0x93,
- 0x88, 0x7d, 0x9b, 0x76, 0x82, 0x67, 0x69, 0x7f, 0x8c, 0x8d, 0x94, 0x7d,
- 0x7b, 0xae, 0x8c, 0x85, 0x8b, 0xa7, 0x8c, 0x87, 0x96, 0x7d, 0x8b, 0x90,
- 0x90, 0x7c, 0x92, 0xa8, 0x81, 0x87, 0xa4, 0xa4, 0x82, 0x8b, 0x8d, 0x89,
- 0x8f, 0x70, 0x9d, 0x7f, 0xa0, 0x84, 0x99, 0x65, 0x99, 0x78, 0x94, 0x8b,
- 0xc5, 0x8d, 0x8d, 0x55, 0xb3, 0x8d, 0x78, 0x93, 0xb4, 0x6d, 0x84, 0x90,
- 0xd5, 0x76, 0x7a, 0x9e, 0xc8, 0x8f, 0x86, 0x8a, 0xaa, 0x8b, 0x7f, 0x90,
- 0xaa, 0x95, 0x9c, 0x81, 0xb4, 0x6b, 0x64, 0x8a, 0x99, 0x84, 0x74, 0x6e,
- 0x95, 0x75, 0x98, 0x92, 0x9a, 0x91, 0x8c, 0x7d, 0x88, 0x6e, 0x89, 0x7d,
- 0x87, 0x80, 0x8e, 0x86, 0x78, 0x9f, 0x96, 0x75, 0x76, 0x82, 0x84, 0xaf,
- 0x8a, 0xb3, 0x93, 0x97, 0x86, 0x7c, 0x7e, 0x96, 0x7c, 0x6d, 0x90, 0x8e,
- 0x85, 0x88, 0x8a, 0x9f, 0x70, 0x89, 0x9f, 0x99, 0x95, 0x87, 0x91, 0x9d,
- 0x80, 0x74, 0x88, 0x7c, 0x7f, 0xa8, 0x93, 0x77, 0x66, 0xa6, 0x80, 0xa2,
- 0x88, 0xa0, 0xaf, 0x6f, 0x76, 0x70, 0x82, 0x9a, 0x73, 0x89, 0x9a, 0x75,
- 0x75, 0x8e, 0x5f, 0x85, 0x6a, 0x76, 0x98, 0x66, 0x87, 0xa3, 0x7a, 0x73,
- 0x9d, 0xa1, 0x98, 0x8e, 0x78, 0x91, 0x83, 0x8c, 0x82, 0x9e, 0x90, 0x87,
- 0x8f, 0x9b, 0x8b, 0x8f, 0x89, 0x62, 0x74, 0x82, 0x7b, 0x7f, 0x8a, 0x9d,
- 0x89, 0x93, 0x8c, 0x7a, 0x99, 0x77, 0xac, 0x75, 0x9b, 0x7f, 0x7f, 0x56,
- 0x8c, 0x96, 0x70, 0x79, 0xc2, 0x7d, 0x90, 0x64, 0xe9, 0x79, 0x68, 0xb2,
- 0xc2, 0xa6, 0xa7, 0x7e, 0xd9, 0x98, 0x79, 0x87, 0xc0, 0x97, 0x87, 0x66,
- 0xd0, 0x9f, 0x92, 0x82, 0xa4, 0xa8, 0x8d, 0x78, 0xa6, 0xa1, 0x76, 0x7d,
- 0xa4, 0x87, 0x89, 0x51, 0xae, 0x88, 0x5b, 0x76, 0x7d, 0x70, 0x74, 0x93,
- 0x89, 0x74, 0x9e, 0x7a, 0x79, 0x64, 0x9a, 0x94, 0x65, 0x93, 0xb0, 0x8d,
- 0x88, 0x7e, 0x8e, 0xa5, 0x63, 0x94, 0x94, 0x7d, 0x91, 0x87, 0x84, 0x95,
- 0x75, 0x9e, 0x81, 0x99, 0x65, 0x76, 0x82, 0x9c, 0x6a, 0xab, 0x84, 0x85,
- 0x88, 0x72, 0x92, 0x83, 0x82, 0xaf, 0x6d, 0x9d, 0x9e, 0x73, 0x98, 0x7f,
- 0x91, 0xb4, 0x62, 0x8d, 0x74, 0x6e, 0xb4, 0x94, 0x97, 0x9e, 0x6f, 0x9a,
- 0x83, 0x7b, 0xa9, 0x7d, 0x87, 0x97, 0x60, 0xa9, 0x7a, 0x75, 0xad, 0x6c,
- 0x77, 0xa4, 0x88, 0x82, 0x6f, 0x8a, 0x83, 0x74, 0x9a, 0xa7, 0x83, 0x91,
- 0x7c, 0x7c, 0x78, 0x77, 0x83, 0x92, 0x7a, 0x83, 0x90, 0x6f, 0x79, 0x6b,
- 0x9b, 0x8d, 0x99, 0x95, 0x7b, 0x89, 0x8e, 0x6c, 0x8e, 0x6c, 0x9b, 0x91,
- 0x97, 0x80, 0x83, 0x6f, 0xaa, 0x91, 0x66, 0x76, 0xc9, 0x77, 0x82, 0x4d,
- 0xd7, 0x5f, 0x58, 0x9a, 0xb1, 0x7a, 0xb1, 0x6b, 0xe5, 0x9d, 0x76, 0x89,
- 0xb6, 0x94, 0x90, 0x5b, 0xb8, 0x92, 0x7d, 0x90, 0xbd, 0x9a, 0x85, 0x4e,
- 0xb4, 0x84, 0x61, 0x82, 0x94, 0x8e, 0x70, 0x57, 0x90, 0x89, 0x6f, 0x60,
- 0x78, 0x90, 0x78, 0x85, 0x8e, 0x7c, 0x76, 0x74, 0x71, 0x5d, 0x94, 0x93,
- 0x71, 0x8f, 0xc2, 0x80, 0x75, 0x7d, 0x77, 0xa8, 0x70, 0x8f, 0xa6, 0x83,
- 0x74, 0x6b, 0x79, 0x97, 0x76, 0xa2, 0xad, 0x93, 0x5b, 0x8c, 0x7c, 0x7e,
- 0x82, 0x9b, 0xa0, 0x76, 0x71, 0x7a, 0xa3, 0x80, 0x87, 0x90, 0x92, 0xa6,
- 0x85, 0x71, 0x99, 0x91, 0x91, 0x8c, 0x99, 0x9b, 0x92, 0x74, 0xb2, 0x79,
- 0x9c, 0x7c, 0x7b, 0xa8, 0x8c, 0x6f, 0xb5, 0x69, 0x7a, 0x8a, 0x68, 0x9f,
- 0x82, 0x7d, 0xbd, 0x5f, 0xa1, 0x92, 0x83, 0x9f, 0x6f, 0xa1, 0x88, 0x61,
- 0x7b, 0x94, 0x89, 0x83, 0x6f, 0x6e, 0x92, 0x9d, 0x65, 0x7f, 0x97, 0x83,
- 0x87, 0x75, 0x92, 0x8a, 0x82, 0x82, 0x79, 0x92, 0x78, 0x89, 0x92, 0x7a,
- 0x91, 0x64, 0x8a, 0x93, 0x9d, 0x74, 0x78, 0x64, 0xab, 0x57, 0x7a, 0x84,
- 0xcf, 0x7d, 0x95, 0x4f, 0xde, 0x63, 0x78, 0x9a, 0xb7, 0x7a, 0x8b, 0x5b,
- 0xda, 0xa3, 0x94, 0x99, 0xbd, 0x88, 0xa4, 0x53, 0xad, 0x8b, 0x81, 0x96,
- 0xca, 0x8f, 0x76, 0x5e, 0xbd, 0x9d, 0x70, 0x81, 0x9b, 0x7d, 0x8a, 0x44,
- 0xa0, 0x77, 0x52, 0x6e, 0x82, 0x62, 0x6a, 0x6b, 0x9d, 0xaa, 0x81, 0x85,
- 0x7d, 0x5f, 0x7f, 0x9c, 0x65, 0x99, 0x97, 0x81, 0x7f, 0x65, 0x65, 0xa4,
- 0x84, 0x8c, 0xa1, 0x6d, 0x7a, 0x70, 0x79, 0x90, 0x98, 0xaa, 0x76, 0x95,
- 0x7f, 0x91, 0x95, 0x96, 0x6e, 0xa5, 0x95, 0xa2, 0x7d, 0x7e, 0x93, 0x87,
- 0x7d, 0x9b, 0x85, 0x9b, 0x85, 0x79, 0x96, 0x6b, 0x9d, 0x9d, 0x61, 0x99,
- 0x9c, 0x74, 0xcc, 0x7e, 0x9a, 0x83, 0x83, 0x98, 0x6f, 0x6d, 0xc5, 0x69,
- 0xb0, 0xa5, 0x5c, 0x91, 0x6c, 0x7b, 0xcc, 0x72, 0x9a, 0x9d, 0x7e, 0xa3,
- 0x8a, 0x96, 0x8e, 0x74, 0x7b, 0x80, 0x6b, 0x85, 0x84, 0x56, 0x92, 0x83,
- 0x64, 0x90, 0x86, 0x86, 0x88, 0x79, 0x8b, 0xa0, 0x86, 0x72, 0xab, 0x95,
- 0x80, 0x81, 0x96, 0x8f, 0x75, 0x7f, 0x71, 0x92, 0x9e, 0x75, 0x62, 0x5e,
- 0xc3, 0x7a, 0x6c, 0x84, 0xba, 0x81, 0x8f, 0x49, 0xc9, 0x76, 0x54, 0x89,
- 0xc2, 0x8c, 0xa2, 0x54, 0xd8, 0xa4, 0x72, 0x90, 0xb1, 0x91, 0xa0, 0x7a,
- 0xbf, 0x9a, 0x6f, 0x82, 0xbb, 0x81, 0x6a, 0x52, 0xc2, 0x82, 0x52, 0x65,
- 0x8d, 0x8a, 0x84, 0x46, 0xa2, 0x90, 0x45, 0x52, 0x82, 0x61, 0x8c, 0x77,
- 0x92, 0x6d, 0x87, 0x5b, 0x5e, 0x72, 0x76, 0x97, 0x73, 0x8d, 0x8d, 0x70,
- 0x7a, 0x66, 0x76, 0x89, 0x72, 0xbf, 0xb0, 0x84, 0x7d, 0x80, 0x71, 0x8f,
- 0x85, 0xa9, 0xa3, 0x7d, 0x7b, 0x84, 0x83, 0xa1, 0x97, 0xa7, 0xaf, 0x84,
- 0x86, 0x7d, 0x94, 0x78, 0x80, 0x98, 0x71, 0x84, 0x94, 0x73, 0xb0, 0x74,
- 0x99, 0xa2, 0x68, 0xa7, 0x8b, 0x86, 0xe0, 0x75, 0x9e, 0x93, 0x5c, 0xb2,
- 0xa2, 0x68, 0xb8, 0x61, 0x92, 0xa3, 0x68, 0xa4, 0x89, 0x59, 0xd0, 0x77,
- 0x97, 0xa9, 0x6a, 0x9b, 0x7d, 0x69, 0x9b, 0x79, 0x8c, 0x7c, 0x68, 0x8b,
- 0x7a, 0x53, 0x99, 0x9c, 0x7e, 0x8d, 0x89, 0x96, 0x9e, 0x83, 0x89, 0x74,
- 0x7f, 0x94, 0x92, 0x8f, 0x85, 0x8a, 0x8a, 0x80, 0x99, 0x87, 0x7a, 0x7d,
- 0xac, 0x93, 0x74, 0x68, 0xba, 0x87, 0x6a, 0x98, 0xc7, 0x79, 0x91, 0x54,
- 0xeb, 0x80, 0x45, 0x80, 0xc4, 0xb4, 0x94, 0x61, 0xd2, 0xa6, 0x7b, 0x95,
- 0xa4, 0xaa, 0x93, 0x7b, 0xb1, 0x74, 0x53, 0x7c, 0xaa, 0x91, 0x64, 0x51,
- 0xa9, 0x6e, 0x5e, 0x7c, 0x79, 0x82, 0x8b, 0x2e, 0x9d, 0x66, 0x61, 0x5e,
- 0x72, 0x7f, 0x6e, 0x6d, 0x8c, 0x79, 0x7d, 0x60, 0x76, 0x79, 0x68, 0x84,
- 0x4d, 0x8e, 0xa8, 0x8f, 0x78, 0x74, 0x69, 0xa4, 0x6e, 0xa9, 0xb9, 0x59,
- 0x83, 0x7f, 0x7a, 0x93, 0x90, 0x9b, 0x8d, 0x93, 0x78, 0x80, 0x77, 0x8b,
- 0x72, 0xa3, 0x97, 0x73, 0x91, 0x6c, 0x9a, 0x97, 0xa3, 0xad, 0x89, 0x96,
- 0x9e, 0x6d, 0xb5, 0x7c, 0xa4, 0x98, 0x61, 0x8a, 0x93, 0x5f, 0xdc, 0x63,
- 0xba, 0x92, 0x84, 0x94, 0xab, 0x6f, 0xbf, 0x66, 0x98, 0x93, 0x74, 0x85,
- 0x96, 0x63, 0xb8, 0x60, 0x94, 0xbb, 0x79, 0x94, 0x7b, 0x67, 0x8a, 0x64,
- 0x99, 0xac, 0x60, 0x98, 0xb0, 0x65, 0xa2, 0x73, 0x8f, 0x94, 0x8c, 0x92,
- 0x84, 0x84, 0x9b, 0x8f, 0x84, 0x8d, 0x9f, 0x90, 0x91, 0x85, 0x93, 0x74,
- 0x97, 0x66, 0x7f, 0x78, 0xa2, 0x95, 0x73, 0x6b, 0xc5, 0x6f, 0x62, 0x79,
- 0xbd, 0x81, 0x89, 0x4a, 0xbd, 0x93, 0x57, 0x81, 0xba, 0xb0, 0x9b, 0x4c,
- 0xe8, 0xa2, 0x85, 0xa2, 0x96, 0x92, 0x93, 0x62, 0xbe, 0x7a, 0x71, 0x8b,
- 0x8d, 0x97, 0x53, 0x56, 0xb1, 0x5f, 0x67, 0x60, 0x7a, 0x8e, 0x8a, 0x3a,
- 0x86, 0x67, 0x6d, 0x53, 0x6e, 0x91, 0x7b, 0x60, 0x99, 0x6d, 0x71, 0x5d,
- 0x67, 0x65, 0x63, 0x87, 0x71, 0x8a, 0x92, 0x6d, 0x8f, 0x6f, 0x6f, 0xae,
- 0x6c, 0xa2, 0x87, 0x6f, 0x99, 0x88, 0x78, 0x94, 0x8a, 0xb2, 0x93, 0x89,
- 0x90, 0x8d, 0x8c, 0x98, 0x81, 0x86, 0x90, 0x6d, 0xa2, 0x82, 0xa2, 0xa3,
- 0x9d, 0x8f, 0x7a, 0x9f, 0x87, 0x70, 0xbd, 0x8e, 0xa5, 0x99, 0x5d, 0x70,
- 0x8c, 0x60, 0xc7, 0x78, 0x97, 0xb0, 0x6f, 0x94, 0x92, 0x5a, 0xc3, 0x6e,
- 0x8b, 0x9f, 0x79, 0xa3, 0x8c, 0x5e, 0xbf, 0x79, 0x8e, 0x98, 0x76, 0x8e,
- 0x67, 0x31, 0x9b, 0x85, 0x8e, 0x85, 0x71, 0x99, 0x72, 0x77, 0x84, 0x81,
- 0x91, 0x95, 0x80, 0x98, 0x82, 0x6f, 0x90, 0xa0, 0x91, 0x91, 0x8e, 0x75,
- 0x8a, 0x89, 0x93, 0x69, 0x95, 0x7f, 0x9a, 0xa0, 0x9e, 0x9b, 0x88, 0x4e,
- 0xc3, 0x8d, 0x65, 0x74, 0xba, 0x8d, 0x97, 0x4d, 0xd6, 0x94, 0x73, 0xa0,
- 0xb1, 0xb3, 0x8c, 0x67, 0xdd, 0x9f, 0x7f, 0xaa, 0xaf, 0x9a, 0x88, 0x67,
- 0xc2, 0x8f, 0x71, 0x7b, 0x8f, 0x9f, 0x47, 0x52, 0x93, 0x72, 0x5a, 0x52,
- 0x97, 0x9d, 0x67, 0x3c, 0xa9, 0x59, 0x59, 0x5b, 0x88, 0x92, 0x82, 0x57,
- 0x83, 0x67, 0x94, 0x77, 0x52, 0x74, 0x60, 0x9e, 0x52, 0x84, 0xa2, 0x69,
- 0x71, 0x96, 0x73, 0xb0, 0x5e, 0xb0, 0x89, 0x71, 0x94, 0x8a, 0x66, 0xa0,
- 0x75, 0xc1, 0x99, 0x8e, 0x83, 0x8a, 0x91, 0x89, 0x6b, 0xa5, 0x79, 0x82,
- 0x8b, 0x73, 0x95, 0xb0, 0x77, 0x9b, 0x82, 0x7d, 0x8f, 0x60, 0xb9, 0x78,
- 0x8b, 0x8f, 0x7b, 0x74, 0x84, 0x6d, 0xbf, 0x76, 0x8f, 0xa3, 0x91, 0xa1,
- 0x81, 0x59, 0xcb, 0x69, 0xac, 0x90, 0x98, 0x92, 0xa7, 0x5d, 0xb4, 0x8b,
- 0xaa, 0xb1, 0x98, 0x8c, 0xa2, 0x4d, 0xa1, 0x69, 0x7f, 0xa0, 0x7d, 0x8a,
- 0x9b, 0x77, 0x8e, 0x71, 0x82, 0x8a, 0x78, 0x8d, 0x98, 0x78, 0x90, 0x91,
- 0x7e, 0x7f, 0x78, 0x85, 0x97, 0x8a, 0x97, 0x6d, 0xb3, 0x94, 0x89, 0xa3,
- 0xa5, 0x9a, 0x76, 0x6b, 0xbd, 0x79, 0x71, 0x95, 0xce, 0xab, 0x93, 0x1f,
- 0xe9, 0x97, 0x4c, 0x84, 0xd5, 0x9f, 0x98, 0x6e, 0xdd, 0x8d, 0x80, 0x9c,
- 0xa8, 0x9e, 0x8d, 0x75, 0xbc, 0x8c, 0x80, 0x89, 0xa1, 0x89, 0x74, 0x58,
- 0x92, 0x86, 0x55, 0x87, 0x91, 0x8d, 0x70, 0x33, 0xb8, 0x50, 0x63, 0x6b,
- 0x79, 0x99, 0x76, 0x71, 0x75, 0x59, 0x73, 0x6b, 0x62, 0x62, 0x74, 0x85,
- 0x73, 0xa3, 0xac, 0x78, 0x77, 0x88, 0x64, 0xa0, 0x73, 0xa1, 0xa8, 0x73,
- 0x91, 0x8e, 0x5f, 0x9a, 0x68, 0xc9, 0xa1, 0x92, 0x7a, 0x7c, 0x69, 0x77,
- 0x7d, 0x9e, 0x8f, 0x76, 0x88, 0x80, 0x92, 0x93, 0x91, 0x99, 0x8c, 0x85,
- 0x9f, 0x69, 0xa8, 0x9b, 0x9f, 0x9a, 0x64, 0x7a, 0x99, 0x70, 0xc4, 0x6d,
- 0x9a, 0x99, 0x82, 0xa0, 0x8b, 0x59, 0xc8, 0x61, 0x8f, 0x95, 0x72, 0x8c,
- 0x90, 0x63, 0xa9, 0x7e, 0x88, 0x8c, 0x85, 0x78, 0x76, 0x58, 0x8e, 0x72,
- 0xa3, 0x9a, 0x7c, 0xa0, 0x7f, 0x6d, 0xa6, 0x83, 0x7e, 0x8d, 0x83, 0x88,
- 0x86, 0x68, 0x8d, 0x96, 0xaa, 0x78, 0x90, 0xa5, 0x9c, 0x9d, 0x99, 0x88,
- 0xb0, 0x82, 0x6f, 0x7e, 0xad, 0xa9, 0x7b, 0x6a, 0xba, 0x6c, 0x6d, 0x89,
- 0xc1, 0x9e, 0x8e, 0x2f, 0xf2, 0x77, 0x50, 0x73, 0xdb, 0xc4, 0x9c, 0x6c,
- 0xd0, 0x90, 0x88, 0xbe, 0x97, 0xb9, 0x9e, 0x6e, 0xbe, 0x8e, 0x83, 0x8e,
- 0x96, 0x98, 0x4c, 0x4e, 0xa7, 0x8d, 0x43, 0x92, 0x8f, 0x92, 0x6d, 0x27,
- 0x94, 0x73, 0x5f, 0x42, 0x7c, 0xa7, 0x8a, 0x5a, 0x81, 0x60, 0x85, 0x66,
- 0x73, 0x72, 0x74, 0x9d, 0x5a, 0x9e, 0xa3, 0x71, 0x75, 0x91, 0x4f, 0xa2,
- 0x67, 0xa6, 0x91, 0x64, 0x92, 0x7e, 0x95, 0x8d, 0x6e, 0xbe, 0x9b, 0x57,
- 0x9b, 0x82, 0x89, 0x70, 0x6f, 0x9e, 0x7e, 0x86, 0x97, 0x81, 0x85, 0x8e,
- 0x70, 0x96, 0x6c, 0x72, 0xab, 0x6d, 0x9c, 0x91, 0xa0, 0x8a, 0x8d, 0x88,
- 0x9e, 0x75, 0xc6, 0x76, 0x7c, 0xa7, 0x6b, 0xa8, 0x94, 0x72, 0xb6, 0x78,
- 0x8d, 0x90, 0x7b, 0x8c, 0xa6, 0x65, 0xad, 0x9b, 0xaa, 0x94, 0x89, 0x7d,
- 0x90, 0x69, 0xaa, 0x7e, 0x9e, 0xad, 0x7f, 0x94, 0x81, 0x7d, 0xa1, 0x7b,
- 0x6c, 0x65, 0x83, 0x95, 0x89, 0x75, 0x93, 0x87, 0x94, 0x87, 0xa8, 0x92,
- 0x8d, 0xa6, 0x9f, 0x78, 0xaa, 0x72, 0x95, 0x94, 0xac, 0xa6, 0x91, 0x5a,
- 0xdb, 0x82, 0x55, 0xb6, 0xc1, 0xa3, 0x84, 0x4f, 0xc9, 0x88, 0x53, 0x8f,
- 0xbb, 0xae, 0x9b, 0x8a, 0xd8, 0xa9, 0x68, 0xc2, 0xa0, 0xa9, 0x87, 0x6b,
- 0xbd, 0x99, 0x7e, 0x86, 0x88, 0xa7, 0x5e, 0x53, 0xa4, 0x84, 0x6b, 0x6e,
- 0x89, 0x95, 0x84, 0x2d, 0xb5, 0x43, 0x3e, 0x50, 0x71, 0x96, 0x9a, 0x5b,
- 0xa1, 0x60, 0x80, 0x70, 0x6a, 0x73, 0x8f, 0x95, 0x52, 0x9b, 0xae, 0x71,
- 0x76, 0x7d, 0x61, 0x99, 0x5b, 0xc3, 0xa8, 0x76, 0x98, 0x72, 0x7f, 0x8a,
- 0x66, 0xc7, 0xa3, 0x7b, 0x8e, 0x8f, 0x70, 0x74, 0x6a, 0xae, 0x85, 0x83,
- 0x96, 0x7d, 0x98, 0xa7, 0x8f, 0x94, 0x7e, 0x84, 0x96, 0x7a, 0xab, 0x7d,
- 0x83, 0xb1, 0x6f, 0x7d, 0x9f, 0x80, 0xca, 0x8f, 0x9b, 0xa9, 0x69, 0x7a,
- 0x92, 0x73, 0xaa, 0x74, 0x88, 0x98, 0x87, 0x8f, 0xa7, 0x68, 0xa0, 0x74,
- 0x97, 0x95, 0x6e, 0x6f, 0x83, 0x53, 0x9b, 0x79, 0x71, 0x87, 0x7d, 0x8b,
- 0x79, 0x87, 0xa3, 0x75, 0x68, 0x73, 0x7e, 0x89, 0x8f, 0x81, 0x98, 0x7a,
- 0x9a, 0x83, 0x9d, 0x95, 0x90, 0x98, 0x97, 0x57, 0x93, 0x7e, 0xa2, 0x9a,
- 0xa8, 0x8a, 0x85, 0x53, 0xbd, 0x7a, 0x61, 0x8b, 0xca, 0xac, 0x9b, 0x2e,
- 0xe8, 0xa5, 0x66, 0x86, 0xca, 0xa7, 0xa0, 0x85, 0xcf, 0xa4, 0x6a, 0xc2,
- 0xb0, 0xaa, 0x76, 0x76, 0xb6, 0xa2, 0x72, 0xa9, 0xa1, 0xa1, 0x67, 0x67,
- 0xac, 0x90, 0x70, 0x6d, 0x8f, 0xb5, 0x6d, 0x3b, 0x85, 0x64, 0x4a, 0x6e,
- 0x72, 0x9f, 0x98, 0x5b, 0x97, 0x3e, 0x8a, 0x6a, 0x6c, 0x7d, 0x77, 0x98,
- 0x5a, 0x92, 0xa3, 0x81, 0x6f, 0x91, 0x7b, 0xa6, 0x6e, 0x9c, 0x9b, 0x5f,
- 0x9e, 0x7e, 0x77, 0x9d, 0x88, 0xc6, 0x81, 0x5a, 0x93, 0x8b, 0x6c, 0x71,
- 0x63, 0x9e, 0x78, 0x79, 0x70, 0x90, 0x95, 0x9f, 0x71, 0xa9, 0x90, 0x73,
- 0x98, 0x8a, 0xa5, 0x8e, 0x87, 0xb0, 0x79, 0x79, 0x92, 0x7d, 0xcc, 0xa8,
- 0x7a, 0x92, 0x82, 0x91, 0x90, 0x69, 0xa4, 0x9b, 0x97, 0x8f, 0x75, 0x7c,
- 0xa3, 0x69, 0xb5, 0x87, 0x8d, 0x88, 0x7b, 0x94, 0x8b, 0x55, 0xa2, 0x6d,
- 0x89, 0x8e, 0x81, 0x8a, 0x9e, 0x87, 0x86, 0x83, 0x8b, 0x84, 0x87, 0xa7,
- 0x8e, 0x79, 0xa4, 0x9c, 0x99, 0x82, 0xa3, 0x8f, 0x91, 0x9a, 0x95, 0x5b,
- 0x9f, 0x6e, 0x85, 0x93, 0xa6, 0x9a, 0x91, 0x4c, 0xd8, 0x6b, 0x6d, 0x85,
- 0xde, 0xaa, 0x97, 0x51, 0xcf, 0x8c, 0x5f, 0x9a, 0xc2, 0x9d, 0x9a, 0x7c,
- 0xc6, 0xb1, 0x84, 0xac, 0xba, 0xa5, 0x7c, 0x76, 0xbd, 0x93, 0x7f, 0xa0,
- 0x86, 0xae, 0x47, 0x41, 0x88, 0x82, 0x62, 0x62, 0x73, 0xad, 0x6b, 0x23,
- 0xa0, 0x48, 0x5a, 0x5a, 0x8f, 0x98, 0xbd, 0x5c, 0x9c, 0x72, 0x7c, 0x68,
- 0x50, 0x78, 0x91, 0xab, 0x5c, 0xc1, 0xc6, 0x66, 0x87, 0x86, 0x60, 0x99,
- 0x65, 0xac, 0x94, 0x91, 0x7e, 0x8c, 0x7d, 0x9b, 0x70, 0xb2, 0x9a, 0x7d,
- 0x82, 0x91, 0x6b, 0x86, 0x6f, 0xbb, 0x7f, 0x66, 0x7a, 0x79, 0x94, 0x96,
- 0x71, 0xa5, 0x75, 0x73, 0x95, 0x81, 0xa4, 0x8b, 0x87, 0xaa, 0x8e, 0x92,
- 0xa9, 0x82, 0xb0, 0x92, 0x89, 0xa7, 0x83, 0x81, 0x8c, 0x6d, 0xc4, 0x7a,
- 0x89, 0xa5, 0xa1, 0xa2, 0xa4, 0x6b, 0xa4, 0x82, 0x90, 0xb2, 0x8d, 0x72,
- 0x83, 0x60, 0xa7, 0x7a, 0x80, 0x97, 0x65, 0x90, 0x87, 0x85, 0xae, 0x71,
- 0x7d, 0x71, 0x98, 0xa8, 0x90, 0x75, 0xa9, 0x96, 0xa2, 0x91, 0x7b, 0x6b,
- 0xa0, 0x9d, 0x8d, 0x5d, 0xa4, 0x79, 0x8c, 0xa4, 0xad, 0x94, 0x7e, 0x77,
- 0xb6, 0x92, 0x74, 0xaf, 0xb5, 0x9b, 0x99, 0x67, 0xe7, 0x8e, 0x6a, 0x87,
- 0xc1, 0x98, 0x9b, 0x7e, 0xd7, 0x9b, 0x5b, 0xae, 0xc9, 0x94, 0x7a, 0x6d,
- 0x9e, 0xb4, 0x86, 0x8e, 0xa3, 0xa1, 0x5e, 0x5d, 0x8e, 0x8f, 0x6b, 0x59,
- 0xa5, 0xa9, 0x69, 0x20, 0xa4, 0x64, 0x35, 0x61, 0x83, 0x9d, 0x8a, 0x4e,
- 0x8b, 0x6c, 0x5e, 0x5b, 0x68, 0x76, 0x89, 0x94, 0x5f, 0x87, 0x98, 0x7a,
- 0x5d, 0x81, 0x89, 0xa6, 0x54, 0xa3, 0xb4, 0x7b, 0x83, 0x8a, 0x90, 0x8b,
- 0x86, 0xbc, 0x86, 0x59, 0x91, 0x79, 0x71, 0x6b, 0x7c, 0x94, 0x98, 0x7f,
- 0x81, 0x76, 0x85, 0xad, 0x69, 0xa8, 0x83, 0x8c, 0x8f, 0x70, 0x9a, 0x91,
- 0x78, 0xb3, 0x8f, 0x6d, 0x90, 0x86, 0xbd, 0x97, 0x7f, 0xaf, 0x7e, 0x90,
- 0x8f, 0x63, 0xa2, 0x93, 0x6e, 0xab, 0x75, 0x72, 0x8d, 0x74, 0xa1, 0x72,
- 0x82, 0xaa, 0x70, 0x82, 0x8d, 0x67, 0x94, 0x91, 0x92, 0xa5, 0x7f, 0xa5,
- 0x6f, 0x6d, 0xaf, 0x80, 0x89, 0x7d, 0x92, 0x99, 0x92, 0x72, 0x9d, 0x7d,
- 0x92, 0x78, 0xa9, 0x89, 0xa9, 0x9b, 0xa3, 0x73, 0x98, 0x71, 0x98, 0x86,
- 0x9e, 0x97, 0x9e, 0x6a, 0xb9, 0x6a, 0x6e, 0x90, 0xde, 0x94, 0x9a, 0x52,
- 0xdd, 0xa9, 0x6a, 0x79, 0xb9, 0xa3, 0xaa, 0x95, 0xba, 0xa2, 0x75, 0xc2,
- 0xbf, 0xb5, 0x6d, 0x8d, 0xae, 0x9b, 0x8d, 0x9a, 0x92, 0xb4, 0x5e, 0x4b,
- 0x8b, 0x99, 0x4f, 0x65, 0x94, 0xb6, 0x5d, 0x3a, 0xa3, 0x77, 0x51, 0x4e,
- 0x6d, 0xa3, 0x94, 0x59, 0x80, 0x56, 0x8c, 0x67, 0x67, 0x74, 0x99, 0x85,
- 0x57, 0x7b, 0x9e, 0x7e, 0x84, 0x85, 0x94, 0x96, 0x71, 0xbf, 0x97, 0x5f,
- 0x7d, 0x80, 0x93, 0x87, 0x6b, 0xb9, 0x7d, 0x8b, 0x84, 0x84, 0x6b, 0x8c,
- 0x6c, 0xc4, 0x85, 0x82, 0x87, 0x8d, 0x64, 0x90, 0x80, 0xb6, 0x9a, 0x70,
- 0x9c, 0x68, 0xa0, 0x88, 0x81, 0x9d, 0x83, 0x75, 0x9d, 0x84, 0xbf, 0x8f,
- 0x83, 0x9b, 0x75, 0x82, 0x9c, 0x76, 0xa4, 0x9d, 0x8a, 0xa7, 0x8e, 0x96,
- 0x9c, 0x64, 0xc0, 0x95, 0x88, 0xa5, 0x6f, 0x74, 0x7e, 0x5d, 0x9f, 0x7d,
- 0x89, 0x81, 0x71, 0xa8, 0x82, 0x6e, 0x9b, 0x9a, 0x6f, 0xa5, 0x88, 0x89,
- 0xa4, 0x7e, 0xa4, 0x90, 0xa1, 0x83, 0x8b, 0x9c, 0x9a, 0x89, 0xa2, 0x89,
- 0x9d, 0x5d, 0x86, 0xa5, 0xc4, 0x96, 0x9c, 0x85, 0xd6, 0x7c, 0x69, 0x88,
- 0xc9, 0xa5, 0x9b, 0x60, 0xea, 0xab, 0x62, 0x9f, 0xd1, 0xa5, 0x86, 0x7e,
- 0xb3, 0xbd, 0x7a, 0xa1, 0xbd, 0xa0, 0x7c, 0x92, 0xa6, 0xa3, 0x7d, 0xa9,
- 0x98, 0xa6, 0x71, 0x5c, 0x9b, 0x9b, 0x58, 0x6f, 0x8f, 0xaa, 0x5e, 0x3b,
- 0xa6, 0x5f, 0x3a, 0x79, 0x94, 0xa5, 0x84, 0x6f, 0x83, 0x5d, 0x75, 0x65,
- 0x6c, 0x77, 0x86, 0xad, 0x4a, 0x92, 0x8e, 0x8a, 0x8f, 0x7b, 0x72, 0x96,
- 0x79, 0xa6, 0xa8, 0x6d, 0x7b, 0x7b, 0x98, 0xa9, 0x79, 0xb9, 0x9e, 0x8f,
- 0x90, 0x6d, 0x76, 0x82, 0x81, 0xc1, 0x95, 0x7c, 0x97, 0x8d, 0x95, 0xa2,
- 0x7c, 0xa4, 0x7b, 0x9b, 0x7f, 0x6f, 0xac, 0x83, 0x7e, 0xa1, 0x7c, 0x7c,
- 0xa1, 0x7a, 0xa1, 0x6d, 0x95, 0x86, 0x77, 0x98, 0x8e, 0x58, 0xa2, 0x76,
- 0x8e, 0xa8, 0x94, 0x90, 0xa7, 0x62, 0xb8, 0x8a, 0x9f, 0xac, 0x87, 0x91,
- 0x88, 0x50, 0xa7, 0x83, 0x88, 0x65, 0x7a, 0x92, 0x9d, 0x70, 0xa9, 0x99,
- 0x7c, 0x87, 0x8c, 0x96, 0x8e, 0x73, 0xa4, 0xa7, 0x9b, 0x70, 0x99, 0x96,
- 0x8f, 0x88, 0xb4, 0x85, 0xa8, 0x6a, 0x9e, 0x78, 0xb0, 0x82, 0x9f, 0x89,
- 0xc9, 0x8d, 0x71, 0x7f, 0xc0, 0x98, 0xa0, 0x6d, 0xd2, 0x8e, 0x64, 0x9e,
- 0xb2, 0xa9, 0x93, 0x6e, 0xcc, 0xbb, 0x89, 0xb1, 0xc1, 0x9b, 0x86, 0x94,
- 0xb5, 0xb5, 0x95, 0xa0, 0x9c, 0x9b, 0x62, 0x5f, 0x7b, 0x91, 0x69, 0x74,
- 0x9e, 0xa3, 0x81, 0x30, 0x85, 0x59, 0x49, 0x5e, 0x83, 0x85, 0x7d, 0x6a,
- 0x90, 0x51, 0x80, 0x5e, 0x64, 0x6f, 0x99, 0x93, 0x75, 0x9a, 0xa7, 0x72,
- 0x6c, 0x5d, 0xa3, 0x93, 0x87, 0xa7, 0xbd, 0x6f, 0x92, 0x6d, 0x85, 0x98,
- 0x6f, 0xc7, 0xb6, 0x7c, 0x80, 0x71, 0x8a, 0x9f, 0x71, 0xb5, 0x8c, 0x6d,
- 0xac, 0x7b, 0x72, 0xb7, 0x69, 0xa6, 0x9d, 0x66, 0xab, 0x7a, 0x8b, 0x70,
- 0x8c, 0x9e, 0x86, 0x75, 0x96, 0x7b, 0xa3, 0x93, 0x8f, 0xb7, 0x84, 0x8c,
- 0x87, 0x56, 0xae, 0x82, 0x71, 0xa3, 0x8d, 0x93, 0xaf, 0x59, 0xb3, 0x8a,
- 0x97, 0x99, 0x75, 0x73, 0x8e, 0x51, 0xae, 0x84, 0x8b, 0x7a, 0x76, 0x77,
- 0x6e, 0x75, 0xa4, 0x8a, 0x75, 0x8e, 0x8f, 0xa2, 0x96, 0x76, 0x9a, 0x80,
- 0x96, 0x7d, 0x94, 0x71, 0x8a, 0x90, 0xac, 0x82, 0xa5, 0x61, 0xa3, 0x84,
- 0xac, 0x8f, 0x74, 0x5c, 0xb6, 0x77, 0x8b, 0x9b, 0xb5, 0x8b, 0xb6, 0x52,
- 0xd7, 0xaa, 0x4b, 0x8c, 0xbf, 0xb8, 0x9f, 0x6d, 0xcb, 0xa3, 0x6e, 0x97,
- 0xaa, 0x8d, 0x7c, 0x99, 0xc0, 0xd0, 0x9e, 0xb7, 0x93, 0xaa, 0x5a, 0x6a,
- 0x7d, 0x9a, 0x63, 0x71, 0x78, 0x8c, 0x67, 0x43, 0x87, 0x52, 0x64, 0x68,
- 0x68, 0x9c, 0x65, 0x60, 0x7a, 0x35, 0x68, 0x66, 0x63, 0x69, 0x8d, 0x8f,
- 0x72, 0x9b, 0x99, 0x5b, 0x80, 0x67, 0x93, 0xa2, 0x97, 0x9d, 0x8c, 0x68,
- 0x80, 0x86, 0x96, 0x91, 0x64, 0xbf, 0x98, 0x63, 0x83, 0x85, 0x61, 0x97,
- 0x6a, 0xac, 0xb4, 0x99, 0x8d, 0x7b, 0x7b, 0xad, 0x8b, 0xb2, 0x9e, 0x7f,
- 0x9a, 0x73, 0x91, 0x84, 0x89, 0x9f, 0x8a, 0x87, 0x8b, 0x72, 0x8e, 0x79,
- 0x86, 0xa7, 0x77, 0x84, 0x90, 0x58, 0xb2, 0x90, 0x93, 0xa0, 0x7f, 0x8a,
- 0x91, 0x5a, 0xb1, 0x80, 0x99, 0xc1, 0x80, 0x7d, 0x97, 0x5c, 0x9a, 0x8c,
- 0x71, 0x96, 0x7e, 0x7f, 0xad, 0x7b, 0xb9, 0x8a, 0x84, 0x84, 0x81, 0x97,
- 0x94, 0x64, 0x9f, 0x7e, 0x9b, 0x8d, 0x7d, 0x8d, 0x9a, 0x9e, 0xac, 0x72,
- 0xb2, 0x73, 0x81, 0x84, 0xc8, 0x81, 0x88, 0x72, 0xbe, 0x85, 0x86, 0x97,
- 0xd3, 0x8a, 0xc7, 0x75, 0xce, 0x9c, 0x69, 0xa6, 0xb0, 0xa1, 0x8e, 0x64,
- 0xb1, 0xa6, 0x67, 0xaa, 0xcd, 0x95, 0x97, 0xa2, 0xb2, 0xb2, 0x85, 0x9a,
- 0x9d, 0xa3, 0x5e, 0x73, 0x6e, 0xae, 0x50, 0x83, 0x8c, 0xab, 0x92, 0x43,
- 0x6b, 0x66, 0x43, 0x5c, 0x8f, 0x8a, 0x9a, 0x6c, 0x84, 0x48, 0x80, 0x6b,
- 0x8d, 0x82, 0xaf, 0x89, 0x71, 0x9f, 0xa4, 0x9a, 0x7b, 0x68, 0x91, 0xaa,
- 0x6b, 0xa3, 0x9c, 0x62, 0x8d, 0x6d, 0x87, 0x87, 0x81, 0x9a, 0x97, 0x6c,
- 0x9c, 0x76, 0x63, 0xbc, 0x62, 0xbc, 0xb0, 0x97, 0xa7, 0x81, 0x70, 0x8f,
- 0x7d, 0xb2, 0xa6, 0x98, 0xa1, 0x7b, 0x8e, 0x83, 0x8c, 0xa2, 0x7e, 0x73,
- 0x99, 0x65, 0xc1, 0x77, 0x8e, 0xbc, 0x72, 0xa6, 0x8c, 0x55, 0xab, 0x8e,
- 0x7d, 0xa3, 0x79, 0x80, 0x9e, 0x6b, 0xa9, 0x6c, 0x80, 0xb6, 0x81, 0xa6,
- 0x92, 0x5b, 0xb7, 0x99, 0x81, 0x7e, 0x8e, 0x89, 0x97, 0x86, 0x93, 0x86,
- 0x7b, 0x9a, 0x7f, 0x9a, 0x8e, 0x69, 0xa3, 0xa4, 0x9f, 0x8b, 0x96, 0x6f,
- 0x8b, 0x97, 0xb4, 0x74, 0x96, 0x53, 0x99, 0x91, 0xa7, 0xa8, 0x69, 0x72,
- 0xc9, 0x85, 0x99, 0x93, 0xc0, 0x90, 0xaa, 0x7f, 0xc7, 0x71, 0x74, 0x8d,
- 0xb7, 0xab, 0x91, 0x69, 0xb4, 0x9b, 0x7d, 0x95, 0xc3, 0xb0, 0x9b, 0xa9,
- 0xb3, 0x9f, 0x79, 0xa5, 0x9f, 0xad, 0x6b, 0x85, 0x90, 0xad, 0x69, 0x62,
- 0x7e, 0xa6, 0x69, 0x4e, 0x80, 0x7e, 0x52, 0x57, 0x5f, 0x95, 0x72, 0x4c,
- 0x87, 0x4e, 0x5a, 0x62, 0x7d, 0x70, 0x92, 0x98, 0x76, 0x8e, 0x99, 0x7d,
- 0x73, 0x6d, 0x86, 0x8e, 0x6b, 0x80, 0xa7, 0x9d, 0x91, 0x73, 0x95, 0x70,
- 0x80, 0xc3, 0x9f, 0x8b, 0x72, 0x86, 0x6b, 0xad, 0x76, 0xbe, 0xad, 0x8e,
- 0x9c, 0x78, 0x6a, 0xbf, 0x7d, 0xa8, 0x88, 0x8a, 0x8b, 0x8c, 0x9c, 0x8c,
- 0x8a, 0x85, 0x73, 0x92, 0xa2, 0x7b, 0xa5, 0x96, 0x9b, 0xa3, 0x6c, 0x80,
- 0xa6, 0x63, 0xac, 0x98, 0xa3, 0x9a, 0x83, 0x8a, 0x8c, 0x63, 0xb9, 0x8c,
- 0x99, 0xa1, 0x7a, 0x6c, 0x9e, 0x59, 0x90, 0x84, 0x8a, 0x93, 0x8f, 0x87,
- 0x98, 0x84, 0x99, 0xa4, 0x72, 0x6d, 0x95, 0xa2, 0x95, 0x72, 0xc3, 0x88,
- 0x8f, 0x6a, 0x77, 0x7d, 0x8b, 0xae, 0xa3, 0x7c, 0xa8, 0x5d, 0x7c, 0xa8,
- 0xa1, 0x85, 0x7e, 0x8c, 0xac, 0x8d, 0x73, 0x88, 0xc1, 0x89, 0xaa, 0x89,
- 0xb2, 0x92, 0x75, 0x9a, 0x9c, 0x8e, 0xb9, 0xaa, 0xaa, 0xac, 0x78, 0x85,
- 0xbc, 0x9f, 0x6d, 0xb7, 0x89, 0xa6, 0xb3, 0x8e, 0xa5, 0xbb, 0x6b, 0x9d,
- 0x8f, 0x8b, 0x69, 0x7a, 0x82, 0x99, 0x8c, 0x49, 0x87, 0x74, 0x37, 0x63,
- 0x5d, 0x92, 0x77, 0x66, 0x63, 0x56, 0x77, 0x5d, 0x7f, 0x68, 0x97, 0x74,
- 0x84, 0x94, 0x7d, 0x7d, 0x91, 0x78, 0x87, 0x96, 0x7f, 0x97, 0x94, 0x6f,
- 0x89, 0x6c, 0x96, 0x71, 0x83, 0x8f, 0x8a, 0x89, 0x7d, 0x84, 0x8a, 0xa6,
- 0x7b, 0x95, 0x89, 0x77, 0x94, 0x80, 0x7f, 0x93, 0x5e, 0xbb, 0x9c, 0xa8,
- 0xa2, 0x7e, 0xa6, 0x86, 0x7d, 0x8b, 0x92, 0x73, 0xac, 0x78, 0xaa, 0x98,
- 0xb1, 0x94, 0x79, 0x8b, 0x8f, 0x70, 0xa7, 0xae, 0x92, 0xad, 0xb1, 0x8b,
- 0xb0, 0x78, 0xbc, 0xa9, 0xa4, 0xa3, 0x9e, 0x76, 0x89, 0x67, 0xab, 0x98,
- 0x75, 0x8c, 0x86, 0x95, 0x9e, 0x77, 0x96, 0x85, 0x8c, 0x8e, 0x8b, 0x8a,
- 0x8a, 0x4b, 0x71, 0x8a, 0x9b, 0x6d, 0x6e, 0x89, 0x81, 0x82, 0xa7, 0x98,
- 0xa5, 0x66, 0x72, 0x8b, 0x99, 0x9a, 0x8b, 0x8b, 0x9f, 0x87, 0x79, 0x84,
- 0x99, 0x6d, 0x90, 0x7d, 0x9d, 0xa7, 0x81, 0xa3, 0x9d, 0x96, 0x82, 0x86,
- 0xa2, 0x8e, 0x8d, 0x7f, 0x84, 0x8c, 0x98, 0xbc, 0x83, 0xb4, 0xb5, 0x78,
- 0x7d, 0xab, 0x8d, 0x87, 0x71, 0x8d, 0x6e, 0x8f, 0x89, 0xaa, 0x7c, 0x6f,
- 0x71, 0x69, 0x65, 0x60, 0x81, 0x91, 0x94, 0x6d, 0x76, 0x66, 0x74, 0x5e,
- 0x77, 0x7c, 0xa2, 0xa6, 0x70, 0x90, 0xa3, 0x68, 0x83, 0x69, 0x71, 0x72,
- 0x6c, 0xa9, 0x85, 0x71, 0x88, 0x60, 0x90, 0x84, 0x8a, 0xba, 0x8b, 0x8c,
- 0x72, 0x8f, 0x98, 0x84, 0x8b, 0x8a, 0xb1, 0xa2, 0x93, 0x8d, 0x86, 0x99,
- 0xa2, 0x99, 0xb0, 0xa6, 0x92, 0x78, 0x86, 0x87, 0x9c, 0x9d, 0x6f, 0x92,
- 0x9a, 0x8a, 0xbf, 0xaa, 0xa3, 0xa2, 0x71, 0x8d, 0x93, 0x70, 0xb5, 0x9c,
- 0xa8, 0x97, 0xb4, 0x93, 0xa6, 0x75, 0xbb, 0xa3, 0x92, 0x95, 0x95, 0x94,
- 0x90, 0x5b, 0xbf, 0x92, 0x8a, 0x95, 0xa0, 0xa1, 0x68, 0x7e, 0x9a, 0x7f,
- 0x88, 0xa7, 0x93, 0xa1, 0x7a, 0x93, 0x95, 0x8b, 0x96, 0x94, 0x70, 0xa0,
- 0x70, 0x8f, 0x9d, 0x96, 0x8e, 0x9c, 0x90, 0x9f, 0x7e, 0x83, 0x84, 0x9e,
- 0x7f, 0x65, 0x72, 0x84, 0x64, 0x94, 0x75, 0xa7, 0x62, 0xa3, 0x8a, 0x9b,
- 0x82, 0x99, 0x87, 0x70, 0x81, 0x6d, 0xac, 0x7b, 0x74, 0x68, 0x5d, 0x95,
- 0xa0, 0x6e, 0x84, 0xab, 0x79, 0x8e, 0x8b, 0x79, 0x7b, 0x83, 0xa0, 0x7b,
- 0x96, 0x71, 0x5d, 0xad, 0xa4, 0x82, 0x79, 0x96, 0x73, 0x84, 0x7d, 0x98,
- 0x87, 0x93, 0x86, 0xa6, 0x7f, 0x7c, 0x71, 0x9d, 0xa4, 0x9b, 0x8a, 0x7c,
- 0x87, 0x6a, 0x7f, 0x8d, 0x97, 0x92, 0xa0, 0x88, 0x77, 0x7d, 0x70, 0x9c,
- 0x9f, 0xa0, 0x71, 0xa3, 0x73, 0x95, 0x76, 0x79, 0x94, 0x95, 0x83, 0x8b,
- 0x8d, 0x82, 0x7a, 0x77, 0xa6, 0x88, 0x72, 0x7a, 0x90, 0x76, 0x7f, 0x95,
- 0x83, 0x90, 0x9e, 0x7c, 0x8e, 0x9a, 0x6b, 0xa4, 0x98, 0x9f, 0x86, 0x8c,
- 0x76, 0x70, 0x74, 0x97, 0x7e, 0xa4, 0x5f, 0xa3, 0xa7, 0x7f, 0x67, 0x8d,
- 0x82, 0x95, 0x93, 0x99, 0x82, 0x70, 0x75, 0xa8, 0xa1, 0xaf, 0x8a, 0x8a,
- 0xb0, 0x89, 0x88, 0x6b, 0x98, 0xaf, 0x75, 0x7f, 0x86, 0x90, 0x8f, 0x8c,
- 0x84, 0x8d, 0x7f, 0x8b, 0x94, 0x9f, 0x80, 0x8b, 0x93, 0xa2, 0x98, 0xa5,
- 0x83, 0x81, 0x8a, 0xaa, 0x86, 0xa3, 0xb0, 0xac, 0x64, 0x9c, 0x7c, 0x93,
- 0xac, 0x85, 0x7f, 0x88, 0x7a, 0xa5, 0x75, 0x69, 0x94, 0xa8, 0x95, 0xa9,
- 0x6f, 0x9f, 0x85, 0x8a, 0xa5, 0x97, 0x98, 0xa9, 0x76, 0x80, 0x7e, 0x95,
- 0x89, 0xaf, 0x68, 0x7b, 0xb4, 0x8a, 0x6b, 0xa4, 0x7b, 0x90, 0x79, 0xba,
- 0x9f, 0x82, 0x7d, 0x89, 0x85, 0x82, 0x94, 0xa5, 0x78, 0x8f, 0x6f, 0x71,
- 0x62, 0x66, 0x73, 0x98, 0x8c, 0x7d, 0x81, 0xa2, 0x69, 0x7c, 0x76, 0xa4,
- 0x94, 0x8f, 0x6f, 0x8a, 0x94, 0x8e, 0x8a, 0x88, 0x8c, 0xa3, 0x6f, 0xa2,
- 0x7d, 0x90, 0x8f, 0x96, 0x6c, 0x76, 0x6e, 0x8e, 0x82, 0x85, 0x7f, 0x93,
- 0x81, 0x83, 0x7b, 0x9f, 0x91, 0x89, 0x75, 0x9c, 0x9f, 0x86, 0x7a, 0x8c,
- 0x7a, 0x7b, 0x82, 0xae, 0x6a, 0x7d, 0x82, 0x82, 0xa0, 0x85, 0x99, 0x9f,
- 0x88, 0x8b, 0x8c, 0x8f, 0x90, 0x96, 0x8e, 0x98, 0xa3, 0x87, 0x7f, 0x9b,
- 0x94, 0x73, 0x96, 0x86, 0x72, 0x7c, 0x75, 0x7c, 0x90, 0x79, 0x83, 0x80,
- 0x79, 0x9e, 0x9c, 0x8e, 0x99, 0x8c, 0x7a, 0x9c, 0x8d, 0x99, 0x9d, 0x84,
- 0xa5, 0x93, 0x85, 0x96, 0x88, 0x94, 0x80, 0x90, 0x73, 0xa3, 0x7c, 0xa1,
- 0x88, 0xa4, 0x98, 0x9f, 0x9e, 0x92, 0x6c, 0xa0, 0x84, 0x87, 0x8a, 0x83,
- 0x7b, 0x91, 0x8c, 0x9e, 0x73, 0xa6, 0x93, 0xa0, 0x8d, 0x98, 0x74, 0xa1,
- 0x83, 0x9a, 0x80, 0xbc, 0x62, 0x70, 0x9e, 0xad, 0x9e, 0x8f, 0x8f, 0x9e,
- 0x7e, 0xac, 0xb0, 0xa9, 0x79, 0x6f, 0x79, 0x8f, 0x7e, 0x71, 0x8d, 0xab,
- 0x97, 0x76, 0x86, 0xa2, 0x98, 0x95, 0x8b, 0x9b, 0x75, 0x7a, 0x71, 0x85,
- 0x7f, 0x61, 0x76, 0x8e, 0x99, 0x91, 0x88, 0x73, 0x71, 0x65, 0x82, 0xa0,
- 0x9b, 0x8f, 0x79, 0x70, 0x78, 0x66, 0x85, 0x94, 0x8b, 0x91, 0x75, 0x80,
- 0x9c, 0x94, 0x7f, 0xa5, 0x82, 0x91, 0x7d, 0x76, 0x80, 0x78, 0x83, 0x82,
- 0x79, 0x98, 0x83, 0x87, 0x94, 0x71, 0x73, 0x77, 0x71, 0x94, 0x6a, 0xa8,
- 0x9e, 0x8d, 0x90, 0x78, 0x7a, 0x81, 0x9c, 0x91, 0x96, 0x80, 0x79, 0x83,
- 0x92, 0x9f, 0x8a, 0x84, 0x8e, 0x97, 0x8c, 0x81, 0x87, 0x74, 0x8b, 0x8e,
- 0xa7, 0x86, 0x8b, 0x8a, 0x8e, 0x8f, 0x9b, 0x6b, 0x82, 0x8a, 0x9f, 0x7a,
- 0x96, 0x80, 0x91, 0x94, 0xa6, 0x8e, 0x7a, 0x97, 0x8a, 0x6c, 0xad, 0xa1,
- 0x78, 0x95, 0x9d, 0x9d, 0x88, 0x94, 0x99, 0x86, 0x80, 0x9b, 0x7c, 0x9c,
- 0x87, 0x7a, 0xa0, 0xa8, 0x83, 0x74, 0x8e, 0x9b, 0x65, 0x95, 0x83, 0xc2,
- 0x69, 0x88, 0x87, 0xa7, 0x86, 0x98, 0x9f, 0xc6, 0x5c, 0x7f, 0xb9, 0x9c,
- 0x8b, 0x6e, 0x95, 0xbd, 0x72, 0x83, 0xbf, 0xb1, 0x89, 0x6d, 0x89, 0x8e,
- 0x9d, 0x87, 0x95, 0x92, 0x76, 0x8d, 0x7f, 0x7f, 0x6d, 0x9d, 0x7b, 0x95,
- 0x86, 0x69, 0x90, 0xa0, 0x62, 0x7c, 0x56, 0xa0, 0x9c, 0x8b, 0x81, 0x79,
- 0xa6, 0x73, 0x69, 0xaa, 0x7b, 0x87, 0x8b, 0x7e, 0xa1, 0x9f, 0x6d, 0xa6,
- 0x7e, 0x7e, 0x87, 0x7c, 0xa5, 0x84, 0x7b, 0xa2, 0xae, 0x92, 0x8e, 0x67,
- 0x93, 0x88, 0x8b, 0xa2, 0x8d, 0x96, 0x92, 0x8e, 0x71, 0x7a, 0x82, 0x80,
- 0x9e, 0x8b, 0x7b, 0x87, 0x96, 0xa0, 0xa4, 0x92, 0x88, 0x7e, 0x77, 0x8e,
- 0x91, 0x7e, 0x81, 0x77, 0x79, 0x93, 0x8d, 0x9d, 0x8a, 0x71, 0x8d, 0x88,
- 0x9d, 0x89, 0x85, 0x94, 0x99, 0x80, 0x89, 0x8f, 0x87, 0x81, 0x83, 0x74,
- 0x8a, 0x89, 0x68, 0x7e, 0x99, 0x82, 0x8c, 0x76, 0xc6, 0x8f, 0x90, 0x7d,
- 0x6c, 0x68, 0xbd, 0x90, 0x78, 0x9d, 0x7b, 0xa3, 0x99, 0x76, 0xaf, 0x8d,
- 0x7d, 0x84, 0x7f, 0x9f, 0x8b, 0x7a, 0xaa, 0xa8, 0x79, 0x89, 0x8f, 0x8f,
- 0x71, 0x80, 0x7f, 0xaa, 0x85, 0x70, 0xa8, 0x96, 0x6c, 0x8c, 0xaf, 0xeb,
- 0x57, 0x7e, 0xcf, 0x8d, 0x93, 0x72, 0xa6, 0xd2, 0x52, 0xab, 0xbb, 0xa8,
- 0x8d, 0x82, 0x7a, 0xbc, 0x72, 0x95, 0xa3, 0xa7, 0x8b, 0x74, 0x84, 0x85,
- 0x6a, 0x85, 0x92, 0x9f, 0x91, 0x6b, 0x9b, 0x73, 0x77, 0xa2, 0x7f, 0x81,
- 0x8e, 0x8b, 0x71, 0x8c, 0x7f, 0x60, 0x86, 0x81, 0x9c, 0x86, 0x93, 0x65,
- 0x84, 0x84, 0x89, 0xa2, 0x98, 0x67, 0x88, 0x71, 0x92, 0x80, 0x65, 0xa2,
- 0xa5, 0x99, 0x85, 0x95, 0x8f, 0x85, 0x8f, 0x82, 0x7e, 0x9a, 0x8a, 0x74,
- 0x9d, 0x75, 0x88, 0x7e, 0xa2, 0x77, 0x82, 0x9e, 0x78, 0xa1, 0x74, 0x79,
- 0x7f, 0x87, 0x91, 0x8d, 0x7a, 0x73, 0x96, 0xa2, 0xa3, 0x81, 0x7d, 0x8a,
- 0x85, 0x75, 0x84, 0x81, 0x8b, 0x7f, 0x6c, 0x86, 0x8d, 0x7b, 0x79, 0x78,
- 0x89, 0x85, 0x8c, 0x9a, 0xa6, 0x96, 0x7a, 0x78, 0xa2, 0x85, 0x9b, 0x89,
- 0xc8, 0x97, 0xa3, 0x82, 0x8b, 0x7f, 0xe7, 0x8f, 0x8f, 0x74, 0x75, 0x83,
- 0x87, 0x79, 0xb3, 0xab, 0x70, 0x9a, 0x9a, 0xa6, 0x81, 0x7e, 0xb8, 0x91,
- 0x8b, 0x8d, 0x93, 0xa1, 0x79, 0x7d, 0x81, 0xb4, 0x79, 0x94, 0xa5, 0x89,
- 0x8e, 0x7c, 0x9b, 0xe2, 0x50, 0x94, 0xdf, 0xa0, 0x53, 0x5d, 0x90, 0xde,
- 0x67, 0x90, 0xaf, 0x8a, 0x8f, 0x73, 0x7b, 0xcb, 0x64, 0x9f, 0x91, 0x86,
- 0x95, 0x84, 0x83, 0x88, 0x76, 0x8b, 0x8a, 0x8f, 0x9c, 0x9a, 0x92, 0x96,
- 0x7f, 0x8e, 0x79, 0x80, 0x91, 0x6d, 0x86, 0x59, 0x74, 0x8a, 0x53, 0x88,
- 0xae, 0x7b, 0x80, 0x70, 0x87, 0x74, 0x75, 0x91, 0xa4, 0x74, 0x8d, 0x5a,
- 0x83, 0x95, 0x65, 0xa1, 0xb3, 0x74, 0x87, 0x7d, 0xaa, 0x82, 0x79, 0x78,
- 0x9b, 0x7c, 0x78, 0x74, 0x9e, 0x74, 0x92, 0x92, 0xa3, 0x6e, 0x75, 0x92,
- 0x6a, 0x6f, 0xa3, 0x7c, 0x9e, 0x7f, 0x92, 0x6b, 0x96, 0x79, 0x9a, 0x87,
- 0x83, 0x8c, 0x72, 0x79, 0x6a, 0xa3, 0x79, 0x7d, 0x6d, 0x6c, 0x81, 0x96,
- 0x98, 0x7f, 0x94, 0x81, 0x8a, 0x8a, 0xa7, 0x8c, 0x9a, 0x84, 0xa7, 0x89,
- 0x9d, 0x85, 0xa6, 0xa8, 0xd0, 0x92, 0x97, 0x9f, 0x76, 0x86, 0xe6, 0x6f,
- 0x7c, 0x84, 0x98, 0x8d, 0x80, 0x75, 0xc5, 0x86, 0x6b, 0x8d, 0x9e, 0x9e,
- 0x7f, 0x71, 0x97, 0xa1, 0x75, 0x92, 0xa9, 0x9e, 0x91, 0x5e, 0xa2, 0xa2,
- 0x68, 0xad, 0xa5, 0xa0, 0x7e, 0x68, 0xac, 0xdc, 0x50, 0xa2, 0xc1, 0x8a,
- 0x63, 0x74, 0x7e, 0xd9, 0x3f, 0xbb, 0xba, 0x9d, 0x7f, 0x76, 0x5f, 0xb0,
- 0x74, 0x8e, 0xb1, 0x95, 0x9a, 0x81, 0x63, 0x9f, 0x98, 0x74, 0x80, 0x89,
- 0x95, 0x8e, 0x9e, 0x78, 0x87, 0x82, 0x57, 0x87, 0x8d, 0x90, 0x79, 0x80,
- 0x76, 0x7c, 0x7d, 0x8a, 0xa6, 0x82, 0x98, 0x7a, 0x96, 0x97, 0x84, 0x87,
- 0xab, 0x7f, 0x87, 0x57, 0x83, 0x6a, 0x6a, 0x84, 0x9c, 0x8d, 0x74, 0x68,
- 0xa2, 0x92, 0x90, 0x98, 0x98, 0x8b, 0x6d, 0x72, 0x90, 0x8c, 0x7c, 0x7d,
- 0x9b, 0x6e, 0x71, 0x76, 0x6b, 0x7b, 0x63, 0x81, 0xad, 0x71, 0x78, 0x8e,
- 0x74, 0x87, 0x8e, 0x8a, 0xab, 0x8e, 0x83, 0x85, 0x7d, 0xa0, 0x67, 0x7f,
- 0x9c, 0x74, 0x6b, 0x88, 0x66, 0x92, 0x7f, 0x83, 0x94, 0x92, 0xa5, 0x82,
- 0xa1, 0x7b, 0x6f, 0x70, 0xab, 0x72, 0xb5, 0x91, 0xb7, 0x89, 0x91, 0x77,
- 0x77, 0x8a, 0xdb, 0x88, 0x8a, 0x8d, 0x89, 0x6c, 0x7b, 0x83, 0xc8, 0xb5,
- 0x4b, 0x96, 0x8b, 0x92, 0x91, 0x76, 0xa9, 0xae, 0x70, 0xa8, 0x74, 0x9d,
- 0x96, 0x6d, 0xa1, 0xba, 0x86, 0xbc, 0xbc, 0xa2, 0x8d, 0x6c, 0x96, 0xd8,
- 0x71, 0xb1, 0xae, 0xb0, 0x79, 0x7b, 0x71, 0xd8, 0x32, 0xaa, 0xae, 0xa7,
- 0x7c, 0x6b, 0x77, 0xc0, 0x7c, 0x9e, 0x9f, 0x89, 0x92, 0x8a, 0x76, 0xae,
- 0x97, 0x75, 0x87, 0x8c, 0x7f, 0x86, 0x8b, 0x73, 0x6b, 0x64, 0x87, 0x6d,
- 0x99, 0x8f, 0x8d, 0x66, 0x76, 0x87, 0x6d, 0x6e, 0x98, 0x7a, 0x91, 0x92,
- 0x8c, 0x7c, 0x89, 0x9b, 0x9e, 0x83, 0x86, 0x62, 0x90, 0x6e, 0x62, 0x82,
- 0xa3, 0x7e, 0x86, 0x6a, 0x93, 0x9b, 0x73, 0x6c, 0xa8, 0x99, 0x73, 0x99,
- 0x8c, 0x89, 0x85, 0x67, 0x98, 0x78, 0x63, 0x98, 0x77, 0xa6, 0x6e, 0x81,
- 0xa4, 0x64, 0x8f, 0x8a, 0x7f, 0x9b, 0x91, 0x91, 0x94, 0x82, 0x8b, 0x8b,
- 0x76, 0x66, 0x83, 0x81, 0x94, 0x71, 0x82, 0x9e, 0x93, 0x85, 0x80, 0x8c,
- 0xae, 0x94, 0x96, 0x74, 0x91, 0x9a, 0x6f, 0x9e, 0xa9, 0x76, 0xab, 0x8e,
- 0xd6, 0x9c, 0x7d, 0x98, 0x83, 0x6e, 0xfe, 0x83, 0x71, 0x82, 0x9f, 0x93,
- 0x7b, 0x67, 0xcb, 0xb9, 0x66, 0x89, 0x99, 0x8a, 0xac, 0x8c, 0xa0, 0x9c,
- 0x70, 0xaf, 0x81, 0x88, 0x9c, 0x7e, 0xa8, 0xa5, 0x65, 0x8c, 0xa1, 0x8c,
- 0x83, 0x85, 0x9d, 0xcb, 0x4b, 0xc1, 0xb5, 0xa2, 0x75, 0x63, 0x75, 0xbd,
- 0x34, 0xae, 0xca, 0xa2, 0x89, 0x7a, 0x69, 0xb0, 0x70, 0xae, 0x94, 0x76,
- 0x85, 0x93, 0x6a, 0x90, 0x6a, 0x8a, 0xac, 0x71, 0x7e, 0x81, 0xa2, 0x71,
- 0x98, 0x86, 0x99, 0x76, 0x8f, 0x6f, 0x90, 0x93, 0x7c, 0x72, 0x81, 0x8c,
- 0x78, 0x77, 0x97, 0x84, 0x98, 0x70, 0x96, 0x9a, 0x9b, 0x93, 0x92, 0x5f,
- 0xaa, 0x88, 0x5b, 0x74, 0xaa, 0x96, 0x6a, 0x73, 0x87, 0x83, 0x72, 0x89,
- 0xab, 0x8a, 0x5f, 0x71, 0xa4, 0x94, 0x92, 0x60, 0x96, 0x7b, 0x53, 0x88,
- 0x69, 0x8b, 0x5e, 0x7b, 0xa0, 0x83, 0x70, 0x95, 0x6d, 0x9b, 0x6d, 0x98,
- 0x99, 0x86, 0x6e, 0x7a, 0x87, 0x86, 0x68, 0x8a, 0x7e, 0x87, 0x90, 0x7d,
- 0x76, 0x93, 0x80, 0x8a, 0x8f, 0x97, 0xac, 0x71, 0xa2, 0x96, 0x7f, 0x8e,
- 0xc2, 0x71, 0xab, 0xa9, 0xd1, 0x85, 0x8c, 0x74, 0x70, 0x72, 0xff, 0x77,
- 0x6d, 0x77, 0x91, 0x5d, 0x71, 0x5d, 0xb2, 0xb1, 0x38, 0x76, 0xa6, 0x80,
- 0x91, 0x86, 0xa3, 0x9c, 0x85, 0x95, 0x99, 0xab, 0x8a, 0x6e, 0x9f, 0xa6,
- 0x75, 0xa9, 0xb3, 0x97, 0x69, 0x85, 0xa4, 0xc9, 0x59, 0xb4, 0xca, 0x8d,
- 0x5c, 0x67, 0x7d, 0xcd, 0x29, 0xca, 0xdb, 0x8c, 0x86, 0x8c, 0x70, 0xaa,
- 0x5c, 0x9e, 0x98, 0x86, 0x92, 0x7e, 0x6b, 0x8e, 0x8f, 0x6a, 0x84, 0x71,
- 0x9a, 0x76, 0x87, 0x84, 0x8b, 0x7f, 0x7f, 0x6e, 0xa3, 0x83, 0x85, 0x78,
- 0x6f, 0x7c, 0x6f, 0x96, 0x95, 0x8c, 0xa3, 0x72, 0x92, 0x66, 0x7b, 0x99,
- 0x9c, 0x9c, 0x9a, 0x63, 0xaa, 0x81, 0x7f, 0x90, 0x8c, 0xa0, 0x7e, 0x67,
- 0x94, 0x96, 0x7f, 0x8a, 0x95, 0x91, 0x5c, 0x73, 0x88, 0x9b, 0x85, 0x70,
- 0x87, 0x79, 0x56, 0x92, 0x69, 0x95, 0x62, 0x78, 0x93, 0x83, 0x63, 0x98,
- 0x7a, 0xa4, 0x95, 0x7c, 0x8e, 0x69, 0x86, 0x92, 0x7d, 0x6b, 0x69, 0x85,
- 0xa8, 0x90, 0x7c, 0x7b, 0x9e, 0x87, 0x7b, 0x90, 0x98, 0x7a, 0xa4, 0x92,
- 0xad, 0x97, 0xa0, 0x6d, 0xa6, 0x74, 0xb7, 0x7f, 0xb9, 0x94, 0x6c, 0x77,
- 0x65, 0x6f, 0xfc, 0x7d, 0x68, 0x74, 0xa1, 0x6c, 0x71, 0x61, 0xc3, 0xb5,
- 0x60, 0x86, 0x8b, 0x7d, 0x89, 0x8b, 0x93, 0xa4, 0x68, 0xa0, 0x8f, 0x73,
- 0x96, 0x6e, 0x81, 0x99, 0x81, 0x9d, 0xae, 0x93, 0x6a, 0x8b, 0x9a, 0xcb,
- 0x68, 0xaf, 0xca, 0x81, 0x73, 0x6e, 0x70, 0xd7, 0x49, 0xb9, 0xc5, 0x9d,
- 0x87, 0x8d, 0x61, 0xa8, 0x5e, 0xa4, 0xb7, 0xab, 0x96, 0x84, 0x76, 0x98,
- 0x84, 0x99, 0x8f, 0x70, 0x79, 0x94, 0xa5, 0x87, 0x6e, 0x73, 0x63, 0x7e,
- 0x83, 0x8c, 0x88, 0x71, 0x7a, 0x81, 0x7d, 0x94, 0x92, 0x89, 0xab, 0x7a,
- 0x96, 0x66, 0x7b, 0x8b, 0x8f, 0x8e, 0x94, 0x5b, 0xa0, 0x7f, 0x82, 0x84,
- 0x84, 0x80, 0x7d, 0x81, 0x89, 0x7b, 0x97, 0x78, 0x83, 0x93, 0x4c, 0x95,
- 0x7f, 0x93, 0x8e, 0x70, 0x89, 0x81, 0x69, 0x87, 0x76, 0x73, 0x9a, 0x74,
- 0xa2, 0x88, 0x5e, 0xac, 0x74, 0x8e, 0x74, 0x8e, 0x94, 0x85, 0x7b, 0x7a,
- 0x72, 0x82, 0x68, 0x77, 0x96, 0x8a, 0x7b, 0x6c, 0x88, 0x8b, 0x6b, 0x86,
- 0xa4, 0x88, 0xac, 0xa1, 0x90, 0x8e, 0x85, 0x6d, 0xb1, 0x69, 0xb1, 0xa2,
- 0xbe, 0x9a, 0x7c, 0xb4, 0x63, 0x56, 0xf2, 0x90, 0x5e, 0x71, 0xa3, 0x6a,
- 0x8b, 0x67, 0xbe, 0xa8, 0x6e, 0x8b, 0x90, 0x83, 0xa0, 0x78, 0x9f, 0xa5,
- 0x65, 0xa3, 0x8b, 0x94, 0x84, 0x6c, 0xa5, 0x97, 0x7d, 0xa7, 0x9f, 0x9c,
- 0x62, 0x7d, 0xb5, 0xb1, 0x58, 0x98, 0xba, 0x8d, 0x7f, 0x57, 0x86, 0xc5,
- 0x39, 0xb3, 0xc9, 0xa9, 0x89, 0x8e, 0x55, 0xaf, 0x54, 0xb4, 0xb0, 0x8f,
- 0x8b, 0x7c, 0x6e, 0x8e, 0x96, 0x90, 0x8a, 0x83, 0x84, 0x8c, 0x96, 0x7f,
- 0x89, 0x67, 0x99, 0x60, 0x74, 0x8d, 0x9b, 0x82, 0x6f, 0x61, 0x84, 0x9a,
- 0x7c, 0x85, 0x86, 0x7c, 0x9b, 0x5f, 0x81, 0x96, 0x90, 0x9b, 0xa0, 0x58,
- 0xaf, 0x78, 0x81, 0x8f, 0x96, 0x81, 0x77, 0x7d, 0xa2, 0x85, 0x74, 0x84,
- 0x99, 0x8d, 0x5f, 0x77, 0x8a, 0x8c, 0x85, 0x78, 0x8f, 0x80, 0x5c, 0x6f,
- 0x77, 0x73, 0x80, 0x99, 0x83, 0x89, 0x6f, 0x8e, 0x85, 0x7e, 0x6c, 0x81,
- 0x99, 0x89, 0x69, 0x70, 0x8c, 0x8f, 0x6b, 0x89, 0x80, 0x7a, 0x83, 0x7a,
- 0x96, 0x99, 0x73, 0x76, 0x9c, 0x67, 0xab, 0xab, 0xbd, 0x8b, 0x85, 0x90,
- 0xb0, 0x6b, 0xbd, 0x9c, 0xb9, 0xa0, 0x7c, 0x7d, 0x66, 0x78, 0xdb, 0x97,
- 0x55, 0x67, 0x96, 0x69, 0x80, 0x49, 0xc1, 0xbb, 0x6c, 0x91, 0x8a, 0x92,
- 0x9a, 0x98, 0xa5, 0x98, 0x51, 0xa6, 0x99, 0x8e, 0x73, 0x73, 0x9d, 0x9f,
- 0x77, 0xa6, 0xa4, 0x92, 0x64, 0x75, 0xac, 0xb2, 0x5d, 0xa1, 0xab, 0xa4,
- 0x5a, 0x5b, 0xb3, 0xb7, 0x2d, 0xca, 0xc8, 0x76, 0x94, 0x8e, 0x59, 0xb0,
- 0x52, 0x9d, 0xbd, 0x89, 0x97, 0x84, 0x5d, 0x9a, 0x87, 0x9b, 0x94, 0x6c,
- 0x7b, 0xaa, 0x8a, 0x8b, 0x79, 0x5d, 0x90, 0x5c, 0x8b, 0x7b, 0xbe, 0x68,
- 0x84, 0x6f, 0x75, 0x72, 0x98, 0x82, 0x92, 0x7a, 0xa2, 0x6e, 0x7b, 0x7d,
- 0x9c, 0x99, 0x97, 0x5d, 0x9b, 0x69, 0x80, 0xa3, 0x96, 0x8d, 0x7c, 0x82,
- 0xa3, 0x76, 0x95, 0x67, 0x93, 0x8e, 0x62, 0x7b, 0x78, 0x96, 0x69, 0x67,
- 0x84, 0x8f, 0x62, 0x80, 0x88, 0x7e, 0x6c, 0x94, 0xab, 0x8b, 0x82, 0x9e,
- 0x7e, 0x8c, 0x70, 0x83, 0x9c, 0x9c, 0x80, 0x87, 0x8f, 0xa1, 0x7f, 0x81,
- 0x95, 0x83, 0x6d, 0x7a, 0xa0, 0x77, 0x6d, 0x76, 0x91, 0x7e, 0xa3, 0x62,
- 0xa0, 0x93, 0x7e, 0x97, 0xb6, 0x6c, 0xad, 0x72, 0xb2, 0x95, 0x73, 0x83,
- 0x62, 0x56, 0xe2, 0x99, 0x6e, 0x66, 0xb0, 0x6c, 0x75, 0x4e, 0xb2, 0xc7,
- 0x51, 0x98, 0x90, 0x8c, 0x82, 0x63, 0xa8, 0x99, 0x54, 0xc1, 0x87, 0x80,
- 0x79, 0x62, 0xad, 0x81, 0x76, 0x99, 0xa9, 0x9b, 0x4e, 0x8c, 0xaf, 0xb6,
- 0x5d, 0x9b, 0xb4, 0x9f, 0x6d, 0x60, 0xa5, 0xb5, 0x3e, 0xb2, 0xc4, 0x96,
- 0x86, 0x6d, 0x48, 0x99, 0x50, 0xc1, 0xa8, 0x93, 0x8a, 0x92, 0x7d, 0x8f,
- 0x74, 0x87, 0x91, 0x71, 0x8c, 0x87, 0x90, 0x80, 0x80, 0x82, 0x7b, 0x85,
- 0x81, 0x7f, 0xa7, 0x6a, 0x78, 0x4e, 0x90, 0x85, 0x9f, 0x93, 0x91, 0x91,
- 0xa5, 0x6e, 0x9d, 0xa7, 0x9e, 0x7f, 0x9a, 0x66, 0xbe, 0x6f, 0x82, 0x81,
- 0x85, 0x86, 0x89, 0x6c, 0x88, 0x92, 0x6d, 0x6a, 0x8c, 0x95, 0x68, 0x70,
- 0x91, 0x9b, 0x76, 0x59, 0x87, 0x93, 0x6f, 0x79, 0x7a, 0x99, 0x7d, 0x76,
- 0xa3, 0x9c, 0x69, 0x75, 0x8f, 0x8e, 0x7e, 0x7a, 0x80, 0x8b, 0x76, 0x82,
- 0x70, 0x71, 0x77, 0x7a, 0x88, 0xa1, 0x79, 0x75, 0x9e, 0x7e, 0x6d, 0x6f,
- 0xa5, 0x84, 0xb1, 0x77, 0xad, 0x94, 0x98, 0x90, 0xa7, 0x5c, 0xb6, 0x84,
- 0x99, 0x91, 0x71, 0x7b, 0x6d, 0x54, 0xd2, 0x84, 0x5d, 0x75, 0xb4, 0x7e,
- 0x7d, 0x53, 0xc5, 0x98, 0x70, 0xaa, 0x9e, 0x81, 0x7d, 0x68, 0xa7, 0x8d,
- 0x63, 0xab, 0x9b, 0x96, 0x7e, 0x6b, 0xa3, 0x9e, 0x6d, 0x98, 0xaf, 0x9b,
- 0x78, 0x74, 0xae, 0xc7, 0x70, 0x98, 0xd4, 0x9a, 0x6e, 0x75, 0xa2, 0xcd,
- 0x42, 0xb0, 0xc9, 0x89, 0x88, 0x77, 0x6a, 0xa4, 0x66, 0xb5, 0xbc, 0x8a,
- 0x96, 0x87, 0x5e, 0xa5, 0x87, 0x95, 0x91, 0x5d, 0x85, 0x91, 0xaa, 0x8f,
- 0x99, 0x78, 0x79, 0x74, 0x7f, 0x81, 0xa1, 0x74, 0x77, 0x64, 0x6c, 0x94,
- 0xa0, 0x8b, 0x9b, 0x8e, 0xac, 0x6a, 0x98, 0x9c, 0x7a, 0x9f, 0xab, 0x7e,
- 0xa3, 0x8b, 0x68, 0x7f, 0x84, 0x9f, 0x93, 0x77, 0x90, 0x98, 0x8f, 0x87,
- 0x81, 0x8e, 0x76, 0x95, 0x66, 0x78, 0x85, 0x79, 0x95, 0x89, 0x64, 0x8e,
- 0x8a, 0x87, 0x6f, 0x65, 0xa4, 0x98, 0x7a, 0x83, 0x85, 0x7e, 0x6b, 0xaa,
- 0x81, 0x94, 0x7c, 0x6e, 0x78, 0x85, 0x87, 0x6d, 0x7a, 0x92, 0x67, 0x7a,
- 0x8d, 0x95, 0x77, 0x7f, 0x9f, 0x71, 0xb1, 0xa1, 0xb2, 0x91, 0x7f, 0xb0,
- 0xac, 0x5c, 0xaf, 0x6a, 0xae, 0x98, 0x63, 0x7e, 0x67, 0x6f, 0xc4, 0x8a,
- 0x75, 0x61, 0xac, 0x73, 0x86, 0x54, 0xc3, 0xa8, 0x5d, 0xa9, 0xb4, 0x9b,
- 0x80, 0x6d, 0xa1, 0x8d, 0x64, 0xaa, 0x86, 0x96, 0x86, 0x6c, 0x9b, 0x8b,
- 0x73, 0x9f, 0x9a, 0x87, 0x64, 0x6c, 0xad, 0xa6, 0x64, 0x8a, 0xbe, 0x88,
- 0x67, 0x67, 0xaf, 0xb0, 0x71, 0xae, 0xde, 0x95, 0x9f, 0x7c, 0x7d, 0xa1,
- 0x79, 0xb8, 0xaa, 0x9c, 0x84, 0x91, 0x6b, 0xac, 0x74, 0xa1, 0xad, 0x74,
- 0x88, 0x93, 0x94, 0x72, 0x97, 0x7a, 0x78, 0x86, 0x76, 0x93, 0xb1, 0x6f,
- 0x91, 0x44, 0x96, 0x8e, 0x8e, 0xa5, 0x9a, 0x70, 0x99, 0x79, 0x84, 0x82,
- 0x7f, 0x78, 0xac, 0x6f, 0x9c, 0x80, 0x7d, 0x87, 0x7f, 0x9d, 0x6a, 0x71,
- 0x7c, 0x92, 0x78, 0x7a, 0x93, 0x90, 0x55, 0x83, 0x7a, 0x8a, 0x9a, 0x65,
- 0x86, 0x9b, 0x7c, 0x6b, 0xa3, 0x85, 0x86, 0x71, 0xab, 0x9a, 0x86, 0x90,
- 0x86, 0x88, 0x88, 0x88, 0x99, 0x98, 0x77, 0x86, 0x88, 0x90, 0x79, 0x7c,
- 0x6e, 0x9f, 0x76, 0x70, 0x84, 0x67, 0x7e, 0x8b, 0xa5, 0x68, 0xa7, 0x9d,
- 0xb5, 0x9b, 0x8b, 0x8a, 0xc0, 0x60, 0x9e, 0x83, 0xb0, 0xb7, 0x65, 0x7f,
- 0x7a, 0x7e, 0xc3, 0x7b, 0x74, 0x8f, 0xa4, 0x68, 0x5f, 0x47, 0xbb, 0xa4,
- 0x74, 0x95, 0xab, 0x80, 0x70, 0x5c, 0x9a, 0x8a, 0x7d, 0xa5, 0x90, 0x7d,
- 0x86, 0x68, 0xb1, 0x73, 0x6d, 0xad, 0x93, 0x8d, 0x7b, 0x64, 0xbd, 0xae,
- 0x7a, 0x98, 0xcb, 0x97, 0x83, 0x67, 0xab, 0xb0, 0x61, 0xa7, 0xcd, 0x7e,
- 0x87, 0x78, 0x76, 0x95, 0x6a, 0xba, 0xa9, 0x84, 0x8f, 0x95, 0x7c, 0x8b,
- 0x90, 0x89, 0x8b, 0x81, 0x87, 0x8b, 0x76, 0x73, 0x6f, 0x61, 0x94, 0x73,
- 0x83, 0x97, 0xb3, 0x6b, 0x9c, 0x55, 0x7f, 0x96, 0x9a, 0x92, 0x85, 0x52,
- 0xc6, 0x73, 0x88, 0x9c, 0x7c, 0x86, 0x98, 0x6d, 0x99, 0x87, 0x80, 0x7c,
- 0x7d, 0x98, 0x74, 0x7c, 0x89, 0x8a, 0x7d, 0x7b, 0x83, 0x90, 0x7d, 0x81,
- 0x7a, 0xa0, 0x86, 0x5f, 0x74, 0x8e, 0x68, 0x7b, 0x6c, 0x86, 0x90, 0x84,
- 0x7e, 0xae, 0x73, 0x6f, 0x8d, 0x81, 0x7c, 0x93, 0xa0, 0xb3, 0x6b, 0x9a,
- 0x88, 0xab, 0x8a, 0x94, 0x9c, 0x87, 0x9c, 0x75, 0x7d, 0x8f, 0x7c, 0x7f,
- 0x9b, 0x69, 0xa8, 0x99, 0x9d, 0x89, 0x8f, 0x72, 0xba, 0x61, 0xac, 0x91,
- 0xb5, 0xa7, 0x84, 0x99, 0x71, 0x7e, 0xd0, 0x7c, 0x6d, 0x66, 0xb6, 0x72,
- 0x79, 0x61, 0xb6, 0xab, 0x69, 0xa0, 0xaa, 0x7d, 0x74, 0x61, 0x95, 0xa5,
- 0x71, 0xb0, 0x93, 0x95, 0x86, 0x7d, 0x9f, 0x7e, 0x6c, 0x97, 0x85, 0x87,
- 0x72, 0x7b, 0xb4, 0xad, 0x84, 0x7b, 0xcd, 0xa9, 0x7e, 0x6d, 0xc8, 0xc7,
- 0x7e, 0xb7, 0xcf, 0x98, 0x7b, 0x7c, 0x69, 0xaf, 0x64, 0xa6, 0xc1, 0x8e,
- 0x8f, 0x9c, 0x7d, 0x93, 0x7a, 0x96, 0x8a, 0x65, 0x92, 0x95, 0x8d, 0x6f,
- 0x9f, 0x7f, 0x65, 0x69, 0x7a, 0x92, 0x9f, 0x5c, 0x90, 0x4e, 0x69, 0x89,
- 0x8f, 0x9c, 0xa8, 0x7a, 0xb6, 0x7d, 0x84, 0x97, 0x7f, 0x91, 0x8d, 0x71,
- 0xae, 0x86, 0x80, 0x78, 0x81, 0x87, 0x6e, 0x88, 0x87, 0x7f, 0x8f, 0x9d,
- 0x78, 0x91, 0x74, 0x91, 0x7f, 0x7a, 0x80, 0x63, 0x93, 0xa0, 0x7f, 0x6f,
- 0xa3, 0x88, 0x76, 0x5c, 0x6e, 0xa1, 0x6e, 0x7f, 0x84, 0x8b, 0x87, 0x6d,
- 0x87, 0x9f, 0x79, 0x7c, 0x83, 0x89, 0x7e, 0x86, 0xa0, 0x82, 0x80, 0x8e,
- 0x8b, 0x6c, 0x6e, 0x69, 0x9f, 0x79, 0xaa, 0x6e, 0xa2, 0x8f, 0x9d, 0x87,
- 0xb4, 0x5d, 0xba, 0x6c, 0xaf, 0xa0, 0x84, 0x87, 0x8c, 0x89, 0xcb, 0x6f,
- 0x8e, 0x71, 0xae, 0x5d, 0x6c, 0x61, 0xb3, 0xaf, 0x7a, 0x94, 0xb1, 0x8a,
- 0x80, 0x65, 0x8a, 0x9d, 0x61, 0xb6, 0x8b, 0x97, 0x8a, 0x73, 0xa8, 0x82,
- 0x74, 0x8a, 0x9c, 0x73, 0x61, 0x69, 0xb8, 0x9f, 0x76, 0x90, 0xc5, 0xaa,
- 0x6b, 0x5f, 0xb7, 0xce, 0x6d, 0xb7, 0xcc, 0x97, 0x7a, 0x81, 0x95, 0xbe,
- 0x78, 0xb1, 0xb4, 0x97, 0x8e, 0x99, 0x70, 0xa2, 0x72, 0x8d, 0x8e, 0x7d,
- 0x90, 0x9f, 0x7b, 0x63, 0x87, 0x89, 0x7a, 0x5f, 0x81, 0x97, 0x8d, 0x78,
- 0x94, 0x64, 0x95, 0x9d, 0x90, 0x87, 0xb3, 0x6e, 0xc2, 0x80, 0x94, 0x86,
- 0x87, 0x93, 0xb3, 0x57, 0xb8, 0x73, 0x8a, 0x81, 0x6f, 0x95, 0x89, 0x82,
- 0x94, 0x7a, 0x8e, 0x97, 0x8a, 0x91, 0x7f, 0x77, 0x98, 0x72, 0x67, 0x5f,
- 0x7b, 0x8d, 0x78, 0x74, 0x91, 0x82, 0x86, 0x5c, 0x88, 0xa3, 0x73, 0x6f,
- 0x92, 0x78, 0x9c, 0x95, 0x99, 0x9d, 0x70, 0x89, 0x8f, 0xa7, 0x74, 0x89,
- 0x77, 0x90, 0x72, 0x8d, 0x9c, 0x6f, 0x7a, 0x6c, 0x9f, 0x72, 0xad, 0x6c,
- 0xa5, 0x7a, 0x9d, 0x78, 0xa4, 0x52, 0xbd, 0x94, 0xb5, 0x97, 0x75, 0x78,
- 0x86, 0x72, 0xdf, 0x6f, 0x98, 0x81, 0xab, 0x5d, 0x62, 0x65, 0x9d, 0xbc,
- 0x68, 0x8a, 0xc1, 0x7e, 0x67, 0x7f, 0x88, 0x95, 0x7f, 0xbd, 0x9c, 0x77,
- 0x7d, 0x7e, 0x96, 0x7c, 0x7f, 0xa1, 0xa4, 0x90, 0x7c, 0x74, 0xc0, 0xac,
- 0x7d, 0xa1, 0xdb, 0x85, 0x85, 0x51, 0xbc, 0xb1, 0x6c, 0xcb, 0xd1, 0xa7,
- 0x76, 0x70, 0x7d, 0xba, 0x88, 0xb6, 0xaf, 0xa2, 0x9d, 0x9b, 0x71, 0x96,
- 0x80, 0x89, 0xa3, 0x86, 0x89, 0x8f, 0x76, 0x77, 0xa9, 0x82, 0x8f, 0x69,
- 0x7f, 0x9d, 0xac, 0x80, 0x98, 0x6c, 0x70, 0x72, 0x81, 0x8b, 0xaf, 0x80,
- 0xb1, 0x6f, 0x7c, 0x90, 0x91, 0x82, 0xa5, 0x67, 0x9c, 0x76, 0x8c, 0x6b,
- 0x9c, 0x9b, 0x87, 0x8c, 0x8e, 0x8b, 0xb0, 0x9d, 0x89, 0x8f, 0x76, 0x87,
- 0x9b, 0x90, 0x8e, 0x74, 0x73, 0x91, 0x85, 0x80, 0x81, 0x72, 0x99, 0x84,
- 0x87, 0x95, 0x84, 0x8c, 0x8a, 0x6e, 0x8c, 0x82, 0xad, 0x9d, 0x80, 0x7f,
- 0x96, 0x9c, 0x7f, 0x67, 0xb0, 0x98, 0x69, 0x84, 0x94, 0xa9, 0x7e, 0x83,
- 0x9d, 0x62, 0x92, 0x6e, 0x95, 0x88, 0xa4, 0x90, 0x97, 0x4d, 0xae, 0x89,
- 0xb6, 0xa1, 0x88, 0x9f, 0x7a, 0x70, 0xc2, 0x71, 0x7f, 0x83, 0x90, 0x83,
- 0x5e, 0x50, 0xa9, 0x9f, 0x73, 0x8c, 0xb2, 0x80, 0x79, 0x65, 0x7c, 0x90,
- 0x6d, 0x9a, 0x91, 0x8d, 0x6f, 0x65, 0x97, 0x87, 0x82, 0xa0, 0xa4, 0x8c,
- 0x68, 0x76, 0xa8, 0xa2, 0x7f, 0xa4, 0xcd, 0x91, 0x70, 0x54, 0x95, 0xc6,
- 0x6e, 0x9c, 0xe2, 0xa1, 0x86, 0x82, 0x73, 0xbc, 0x89, 0xaa, 0xb2, 0x7d,
- 0x82, 0x84, 0x8b, 0x9e, 0x84, 0x94, 0xa0, 0x7a, 0x98, 0x9d, 0x99, 0x7b,
- 0x7b, 0x89, 0x8f, 0x66, 0x89, 0x9b, 0xa7, 0x8b, 0x9b, 0x62, 0x9b, 0x78,
- 0x8b, 0x95, 0xbd, 0x7a, 0x9e, 0x61, 0x80, 0x84, 0x89, 0x8e, 0xb4, 0x7b,
- 0xb8, 0x70, 0x75, 0x8e, 0x7b, 0x9c, 0x9e, 0x9f, 0x89, 0x86, 0x9b, 0x7a,
- 0x7b, 0x95, 0x83, 0x95, 0x80, 0x94, 0x85, 0x65, 0x8c, 0x81, 0x67, 0x77,
- 0x94, 0x8a, 0x92, 0x74, 0x72, 0x90, 0x6b, 0x74, 0x7e, 0x75, 0x71, 0x84,
- 0x9e, 0xa6, 0x64, 0x80, 0x8d, 0x7a, 0x8c, 0x82, 0x98, 0x96, 0x64, 0x7d,
- 0x8b, 0x82, 0x6a, 0x7f, 0x97, 0x4e, 0x91, 0x74, 0x94, 0x99, 0x6d, 0x6a,
- 0xb3, 0x5a, 0xb8, 0x64, 0xa3, 0x95, 0x5d, 0x95, 0x90, 0x87, 0xcc, 0x72,
- 0x85, 0x85, 0x8f, 0x55, 0x6f, 0x65, 0x84, 0xb6, 0x7b, 0x77, 0xce, 0x79,
- 0x82, 0x59, 0x8a, 0xa2, 0x68, 0x9b, 0xa3, 0x81, 0x9c, 0x7a, 0x97, 0x87,
- 0x6b, 0x8c, 0x9c, 0xaa, 0x5c, 0x69, 0xb8, 0xb7, 0x7c, 0xa0, 0xb5, 0x92,
- 0x8d, 0x67, 0x96, 0xd2, 0x77, 0xa6, 0xd9, 0xad, 0xaa, 0x79, 0x90, 0xc9,
- 0x81, 0xbf, 0xd0, 0x8d, 0x9d, 0x88, 0x9c, 0x91, 0x90, 0x94, 0x89, 0x8a,
- 0x91, 0x9b, 0x89, 0x79, 0x92, 0x80, 0x8f, 0x7b, 0x7e, 0x8b, 0xb1, 0x85,
- 0xa4, 0x5a, 0xb4, 0x7a, 0xa7, 0x8c, 0xa4, 0x75, 0xb9, 0x66, 0x93, 0x86,
- 0x8a, 0x87, 0xad, 0x64, 0xa2, 0x7e, 0x99, 0x9f, 0x81, 0xa2, 0x9b, 0x88,
- 0x9e, 0xa2, 0xb9, 0x8a, 0x78, 0x84, 0x91, 0x8e, 0x8b, 0x90, 0x83, 0x80,
- 0x64, 0x93, 0x77, 0x89, 0x81, 0x86, 0x96, 0x7a, 0x81, 0xab, 0x6d, 0x73,
- 0x7d, 0x7e, 0xaa, 0x85, 0x95, 0xac, 0x8b, 0x89, 0x8b, 0x77, 0xa3, 0x8b,
- 0xa3, 0xa0, 0x87, 0x86, 0x7a, 0x74, 0x6f, 0x7c, 0x90, 0x58, 0xa2, 0x64,
- 0x94, 0x8b, 0xa0, 0x88, 0xab, 0x53, 0xce, 0x67, 0xb7, 0x7f, 0x8d, 0x69,
- 0x84, 0x74, 0xaf, 0x72, 0xab, 0x70, 0x8f, 0x6e, 0x5d, 0x61, 0x96, 0xa1,
- 0x7b, 0x6f, 0xa2, 0x75, 0x8f, 0x5d, 0x93, 0x72, 0x82, 0x97, 0x76, 0x65,
- 0x7e, 0x96, 0xb3, 0x8b, 0x8d, 0x89, 0x8f, 0x7b, 0x6f, 0x71, 0xa1, 0x9e,
- 0x91, 0x7c, 0xc9, 0x9f, 0x7c, 0x71, 0xa1, 0xba, 0x77, 0xa5, 0xd4, 0xa6,
- 0xa0, 0x82, 0x7b, 0x95, 0x9d, 0xb7, 0xaa, 0x8d, 0x71, 0x87, 0x94, 0x7e,
- 0x88, 0x7f, 0x8b, 0x6e, 0x93, 0x9f, 0x82, 0x88, 0x94, 0x8a, 0x97, 0x7f,
- 0x7d, 0x8c, 0xa0, 0x84, 0xb4, 0x7c, 0x8c, 0x7f, 0x71, 0x8c, 0x8e, 0x7f,
- 0xc6, 0x64, 0x81, 0x8d, 0x89, 0x8d, 0xc4, 0x77, 0xaf, 0x75, 0x92, 0x7f,
- 0x84, 0xa1, 0x99, 0x94, 0x9e, 0x82, 0x7a, 0x98, 0x7e, 0x8e, 0x93, 0x8c,
- 0x6b, 0x93, 0x84, 0xaa, 0x7f, 0x8f, 0x6b, 0x94, 0xa3, 0x8a, 0x78, 0x82,
- 0x60, 0x92, 0x8b, 0x8d, 0x75, 0x8c, 0x8e, 0x6e, 0x7e, 0x9d, 0x6d, 0x8e,
- 0x79, 0x8d, 0x80, 0x89, 0xaa, 0x99, 0x7e, 0xa3, 0x83, 0x95, 0x83, 0x85,
- 0x9c, 0x60, 0x99, 0x78, 0x93, 0x8b, 0x80, 0x82, 0x9d, 0x6b, 0xc2, 0x54,
- 0xb9, 0x7a, 0x83, 0x98, 0x88, 0x65, 0xcb, 0x52, 0xa7, 0x8d, 0x7f, 0x81,
- 0x6b, 0x6d, 0x9e, 0x92, 0x85, 0x82, 0x9f, 0x67, 0x6f, 0x74, 0xaa, 0x75,
- 0x99, 0x9f, 0x8a, 0x8b, 0x88, 0x82, 0xb8, 0x6b, 0x85, 0x99, 0x93, 0x90,
- 0x8d, 0x7a, 0xaa, 0x9d, 0x86, 0x7f, 0xbd, 0x91, 0x67, 0x65, 0x8c, 0xb3,
- 0x87, 0x94, 0xa3, 0x9a, 0x7e, 0x73, 0x83, 0xaa, 0x7a, 0xba, 0xaa, 0x9e,
- 0x9e, 0x86, 0x9a, 0x63, 0x9c, 0x98, 0x5e, 0xa0, 0x9c, 0x9e, 0x8b, 0x85,
- 0xa2, 0x74, 0x80, 0x8d, 0x7e, 0x89, 0xc0, 0x75, 0xa5, 0x3f, 0x97, 0xa2,
- 0x8c, 0x8c, 0x9d, 0x88, 0xa4, 0x5e, 0x75, 0x5f, 0x87, 0x82, 0xbc, 0x72,
- 0xa3, 0x77, 0x83, 0x79, 0x82, 0x95, 0x8d, 0x77, 0x73, 0x81, 0x9d, 0x9b,
- 0x6c, 0x87, 0x93, 0x96, 0x83, 0x86, 0x8b, 0x89, 0x72, 0x7d, 0x96, 0x78,
- 0x67, 0xa2, 0x8d, 0x81, 0x6a, 0x98, 0x75, 0x80, 0x8a, 0x80, 0x9e, 0x82,
- 0x76, 0x9b, 0x6c, 0x94, 0x7a, 0x96, 0x74, 0x92, 0x78, 0x91, 0x7a, 0x7c,
- 0x9a, 0x98, 0x70, 0x5d, 0x9c, 0x4b, 0x70, 0x7d, 0xa9, 0x9b, 0x70, 0x96,
- 0xad, 0x59, 0xc4, 0x63, 0xbc, 0x8f, 0x5c, 0x86, 0x8e, 0x97, 0xa0, 0x7c,
- 0xa6, 0x77, 0xaa, 0x93, 0x68, 0x66, 0x93, 0x91, 0x7b, 0x7e, 0xa2, 0x7a,
- 0x98, 0x77, 0x97, 0x59, 0x84, 0x76, 0x9c, 0x7b, 0x8b, 0x76, 0x88, 0x7a,
- 0x8c, 0x7b, 0xa4, 0xae, 0x6e, 0x7d, 0xb3, 0x99, 0x8d, 0x68, 0x9e, 0x7e,
- 0x77, 0x59, 0x80, 0xbe, 0x80, 0x83, 0xd9, 0x9f, 0x7d, 0x60, 0x8b, 0x98,
- 0x7f, 0x9e, 0xa3, 0x8d, 0x7d, 0x81, 0x9e, 0x78, 0x99, 0x94, 0x70, 0x80,
- 0x9b, 0x89, 0x8c, 0x6d, 0x9c, 0x95, 0x76, 0x7c, 0x83, 0x87, 0x97, 0x93,
- 0x89, 0x6d, 0x77, 0x7e, 0x7e, 0x87, 0x8e, 0x7e, 0x94, 0x61, 0x94, 0xa2,
- 0x94, 0x91, 0xa1, 0x64, 0xc1, 0x78, 0x79, 0xaf, 0x67, 0x7a, 0x9b, 0xa1,
- 0x95, 0x8e, 0x97, 0x84, 0x7b, 0x85, 0x80, 0xa1, 0x6f, 0x87, 0x79, 0x83,
- 0x73, 0x9d, 0x81, 0x64, 0x7a, 0x7f, 0x8f, 0x91, 0x73, 0x97, 0x74, 0x8b,
- 0x7e, 0x88, 0x7f, 0x7e, 0x6e, 0xa1, 0x85, 0x8f, 0x77, 0x93, 0x7a, 0x6f,
- 0x7b, 0x91, 0x67, 0x73, 0x8b, 0x97, 0x6d, 0x87, 0x84, 0xf8, 0xff, 0xff,
- 0x88, 0xf8, 0xff, 0xff, 0xe6, 0xf8, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00,
- 0x80, 0x02, 0x00, 0x00, 0x73, 0x84, 0xbb, 0xa4, 0xa5, 0x44, 0x5c, 0xb1,
- 0x8e, 0x50, 0x82, 0x8b, 0x81, 0x86, 0x48, 0x80, 0xa9, 0x61, 0xa3, 0xa8,
- 0xca, 0x5a, 0x9d, 0x8a, 0x89, 0x7c, 0x65, 0x91, 0x5e, 0x70, 0x84, 0x71,
- 0xbc, 0x36, 0x8e, 0x8b, 0xa6, 0x63, 0xb7, 0x75, 0x92, 0x59, 0x60, 0x7e,
- 0x33, 0x8f, 0x90, 0x7a, 0xa9, 0x27, 0x72, 0x80, 0x62, 0x95, 0x93, 0x7b,
- 0x60, 0x46, 0x40, 0x55, 0x01, 0x9e, 0x8a, 0x6b, 0x58, 0x8a, 0xa6, 0xb7,
- 0x91, 0x39, 0x72, 0xb4, 0x6e, 0x67, 0x83, 0x91, 0x82, 0x7b, 0x64, 0x7a,
- 0x87, 0x6e, 0xb0, 0xa0, 0xd3, 0x53, 0xb7, 0x93, 0x76, 0xa6, 0x68, 0x8a,
- 0x74, 0x6a, 0x96, 0x6e, 0xb3, 0x53, 0xaa, 0x89, 0xf1, 0x76, 0xb8, 0x75,
- 0x8b, 0x66, 0x5f, 0x6e, 0x52, 0x92, 0x6f, 0x82, 0xbe, 0x45, 0x8d, 0x69,
- 0x98, 0x98, 0x80, 0x87, 0x73, 0x7d, 0x4d, 0x42, 0x1f, 0xa5, 0x6a, 0x73,
- 0x47, 0x87, 0x8a, 0xd1, 0x75, 0x30, 0x91, 0xae, 0x60, 0x82, 0x7a, 0x94,
- 0x75, 0x71, 0x6a, 0x7c, 0x74, 0x7a, 0xac, 0xa2, 0xb6, 0x51, 0xc6, 0x97,
- 0x63, 0xa0, 0x67, 0x7f, 0x80, 0x69, 0x88, 0x6b, 0xa5, 0x5e, 0xc2, 0x72,
- 0xf4, 0x6e, 0xaf, 0x76, 0x7f, 0x7c, 0x55, 0x68, 0x67, 0x97, 0x61, 0x7b,
- 0xbe, 0x5e, 0xab, 0x58, 0xca, 0xa2, 0x77, 0x7a, 0x8f, 0x6e, 0x54, 0x33,
- 0x4d, 0xa7, 0x5d, 0x66, 0x47, 0x92, 0x6f, 0xd6, 0x5c, 0x25, 0xa9, 0xbc,
- 0x5c, 0xb8, 0x64, 0x9b, 0x58, 0x6e, 0x77, 0x76, 0x6a, 0x94, 0xb2, 0xac,
- 0x9a, 0x51, 0xd0, 0x94, 0x62, 0xcc, 0x5a, 0x7f, 0x74, 0x6e, 0x7d, 0x71,
- 0x9b, 0x69, 0xd3, 0x64, 0xef, 0x76, 0xaa, 0x75, 0x89, 0x84, 0x50, 0x76,
- 0x72, 0x97, 0x5f, 0x77, 0xc5, 0x66, 0xce, 0x3a, 0xe5, 0xad, 0x5a, 0x81,
- 0x9e, 0x8e, 0x60, 0x3d, 0x6d, 0xa9, 0x46, 0x6b, 0x44, 0x89, 0x4d, 0xd8,
- 0x4c, 0x28, 0xb1, 0xb7, 0x60, 0xc7, 0x57, 0xb5, 0x50, 0x68, 0x88, 0x7c,
- 0x60, 0x98, 0xac, 0x9a, 0x7f, 0x51, 0xce, 0x8a, 0x5e, 0xd8, 0x51, 0x7d,
- 0x68, 0x6e, 0x7f, 0x6e, 0x90, 0x7b, 0xdf, 0x60, 0xda, 0x77, 0x91, 0x6f,
- 0x85, 0xa0, 0x58, 0x73, 0x70, 0x93, 0x51, 0x7d, 0xb9, 0x70, 0xf5, 0x31,
- 0xe9, 0xa3, 0x47, 0x76, 0xa7, 0x9b, 0x72, 0x3d, 0x90, 0xb2, 0x57, 0x64,
- 0x5b, 0x6f, 0x2b, 0xcf, 0x52, 0x28, 0xc1, 0xa7, 0x6a, 0x78, 0x51, 0xad,
- 0x49, 0x70, 0x90, 0x81, 0x5c, 0x7e, 0x9e, 0x99, 0x77, 0x50, 0xc0, 0x94,
- 0x63, 0xb7, 0x4d, 0x71, 0x58, 0x66, 0x76, 0x6d, 0x78, 0x6a, 0xe1, 0x40,
- 0xc7, 0x73, 0x7f, 0x65, 0x7c, 0x7f, 0x4d, 0x80, 0x64, 0x95, 0x57, 0x81,
- 0xb1, 0x5e, 0xff, 0x26, 0xd6, 0xa2, 0x3a, 0x73, 0xa7, 0x81, 0x76, 0x5d,
- 0x92, 0xb1, 0x58, 0x48, 0x4e, 0x5e, 0x1a, 0xc8, 0x58, 0x2c, 0xb6, 0xa7,
- 0x67, 0x89, 0x5e, 0xa0, 0x4f, 0x78, 0x93, 0x8b, 0x57, 0x7b, 0x95, 0x78,
- 0x6e, 0x46, 0xb2, 0x98, 0x55, 0xd3, 0x5e, 0x66, 0x56, 0x68, 0x74, 0x7e,
- 0x72, 0x74, 0xdd, 0x36, 0xa6, 0x64, 0x65, 0x6b, 0x81, 0x98, 0x56, 0x76,
- 0x65, 0x93, 0x58, 0x7d, 0x9b, 0x82, 0xef, 0x44, 0xbf, 0xa4, 0x3d, 0x57,
- 0xa0, 0xa7, 0x7a, 0x74, 0x9f, 0xa8, 0x70, 0x52, 0x55, 0x5f, 0x1a, 0x94,
- 0x64, 0x37, 0xa7, 0xa6, 0x80, 0x7d, 0x6e, 0x99, 0x5d, 0x81, 0x8a, 0x99,
- 0x5c, 0x76, 0x8f, 0x44, 0x68, 0x50, 0x94, 0x97, 0x63, 0xb6, 0x73, 0x56,
- 0x5b, 0x70, 0x66, 0x8b, 0x72, 0x78, 0xcc, 0x31, 0x8b, 0x68, 0x4a, 0x74,
- 0x7d, 0x99, 0x54, 0x91, 0x6a, 0x90, 0x5d, 0x80, 0x8c, 0x82, 0xcd, 0x4f,
- 0xb0, 0x96, 0x63, 0x56, 0x97, 0xb3, 0x7e, 0x97, 0xa4, 0x9d, 0x7a, 0x5d,
- 0x49, 0x36, 0x18, 0x64, 0x60, 0x43, 0x89, 0xa2, 0x6a, 0x49, 0x7f, 0x58,
- 0x6a, 0x83, 0x77, 0x9d, 0x70, 0x3b, 0x83, 0x21, 0x59, 0x52, 0x6d, 0x95,
- 0x48, 0xa8, 0x8a, 0x42, 0x50, 0x6d, 0x44, 0x95, 0x69, 0x50, 0xc1, 0x4b,
- 0x7c, 0x59, 0x42, 0x78, 0x77, 0x7f, 0x5b, 0x98, 0x67, 0x89, 0x55, 0x8b,
- 0x82, 0x47, 0xb7, 0x64, 0x9d, 0x83, 0x5c, 0x53, 0x89, 0x90, 0x79, 0xb2,
- 0x90, 0x98, 0x85, 0x5a, 0x4d, 0x2b, 0x19, 0x1e, 0x52, 0x50, 0x57, 0x8b,
- 0x73, 0x3a, 0x88, 0x1e, 0x65, 0x80, 0x4d, 0x9b, 0x6c, 0x3c, 0x86, 0x26,
- 0x5b, 0x56, 0x36, 0x98, 0x49, 0x87, 0x9f, 0x2a, 0x40, 0x61, 0x27, 0x9d,
- 0x63, 0x40, 0xa8, 0x46, 0x6b, 0x52, 0x52, 0x7f, 0x67, 0x6a, 0x58, 0xa1,
- 0x5d, 0x6d, 0x5f, 0x9a, 0x72, 0x3a, 0x99, 0x63, 0x8c, 0x80, 0x68, 0x58,
- 0x72, 0x6a, 0x7c, 0xbb, 0x7e, 0x78, 0x94, 0x60, 0x72, 0xfb, 0xff, 0xff,
- 0x04, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x96, 0xfe, 0xff, 0xff,
- 0x8f, 0x00, 0x00, 0x00, 0x8f, 0xfc, 0xff, 0xff, 0xb4, 0xfe, 0xff, 0xff,
- 0xc1, 0xfd, 0xff, 0xff, 0x59, 0xff, 0xff, 0xff, 0xbc, 0xfe, 0xff, 0xff,
- 0x09, 0xff, 0xff, 0xff, 0x9e, 0xfb, 0xff, 0xff, 0x04, 0x00, 0x00, 0x00,
- 0x10, 0x00, 0x00, 0x00, 0xe9, 0x03, 0x00, 0x00, 0x2b, 0xfd, 0xff, 0xff,
- 0x3b, 0xfd, 0xff, 0xff, 0x91, 0x01, 0x00, 0x00, 0x60, 0xfb, 0xff, 0xff,
- 0x04, 0xfd, 0xff, 0xff, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0xf0, 0x03, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x18, 0x03, 0x00, 0x00, 0x78, 0x03, 0x00, 0x00,
- 0x88, 0x01, 0x00, 0x00, 0xf8, 0x01, 0x00, 0x00, 0x80, 0x00, 0x00, 0x00,
- 0x0c, 0x00, 0x00, 0x00, 0x64, 0x02, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00,
- 0xb2, 0xfc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x24, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,
- 0x0e, 0x00, 0x00, 0x00, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x73, 0x5f, 0x73,
- 0x6f, 0x66, 0x74, 0x6d, 0x61, 0x78, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x84, 0xfd, 0xff, 0xff,
- 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3b,
- 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x7f, 0x3f, 0x01, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x22, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,
- 0x1c, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
- 0x1c, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x61, 0x64, 0x64, 0x5f,
- 0x31, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x04, 0x00, 0x00, 0x00, 0xec, 0xfd, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00,
- 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x7d, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x28, 0x17, 0xb1, 0x3d,
- 0x01, 0x00, 0x00, 0x00, 0x84, 0xdb, 0x33, 0x41, 0x01, 0x00, 0x00, 0x00,
- 0x9d, 0xf0, 0x2c, 0xc1, 0x8e, 0xfd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,
- 0x48, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
- 0x48, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00, 0x77, 0x65, 0x69, 0x67,
- 0x68, 0x74, 0x73, 0x5f, 0x71, 0x75, 0x61, 0x6e, 0x74, 0x5f, 0x31, 0x2f,
- 0x46, 0x61, 0x6b, 0x65, 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57, 0x69, 0x74,
- 0x68, 0x4d, 0x69, 0x6e, 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72, 0x73, 0x2f,
- 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x6f, 0x73, 0x65, 0x00, 0x00, 0x00,
- 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xa0, 0x0f, 0x00, 0x00,
- 0x84, 0xfe, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00,
- 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x89, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0xac, 0x5f, 0xf6, 0x39, 0x01, 0x00, 0x00, 0x00, 0x1d, 0xaf, 0x62, 0x3d,
- 0x01, 0x00, 0x00, 0x00, 0x5e, 0x1b, 0x83, 0xbd, 0x22, 0xfe, 0xff, 0xff,
- 0x00, 0x00, 0x00, 0x03, 0x1c, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x52, 0x65, 0x6c, 0x75, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0xf4, 0xfe, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00,
- 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x47, 0x0f, 0x72, 0x3d,
- 0x01, 0x00, 0x00, 0x00, 0x38, 0x1d, 0x71, 0x41, 0x01, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x96, 0xfe, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,
- 0x20, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
- 0x28, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00, 0x52, 0x65, 0x73, 0x68,
- 0x61, 0x70, 0x65, 0x5f, 0x31, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x6c, 0xff, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00,
- 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xc6, 0xd0, 0xd0, 0x3d,
- 0x01, 0x00, 0x00, 0x00, 0xf5, 0xff, 0xcf, 0x41, 0x01, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x0e, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,
- 0x3c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
- 0x50, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00, 0x77, 0x65, 0x69, 0x67,
- 0x68, 0x74, 0x73, 0x5f, 0x71, 0x75, 0x61, 0x6e, 0x74, 0x2f, 0x46, 0x61,
- 0x6b, 0x65, 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57, 0x69, 0x74, 0x68, 0x4d,
- 0x69, 0x6e, 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72, 0x73, 0x00, 0x00, 0x00,
- 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x14, 0x00,
- 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x00, 0x00,
- 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x7f, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xf7, 0x5e, 0x6c, 0x3a,
- 0x01, 0x00, 0x00, 0x00, 0x30, 0x42, 0xec, 0x3d, 0x01, 0x00, 0x00, 0x00,
- 0x42, 0xca, 0xe8, 0xbd, 0xaa, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02,
- 0x20, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
- 0x1c, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00, 0x43, 0x6f, 0x6e, 0x76,
- 0x32, 0x44, 0x5f, 0x62, 0x69, 0x61, 0x73, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x9c, 0xff, 0xff, 0xff, 0x18, 0x00, 0x00, 0x00,
- 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0xec, 0xcd, 0xc0, 0x38, 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x08, 0x00,
- 0x07, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x02, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00,
- 0x4d, 0x61, 0x74, 0x4d, 0x75, 0x6c, 0x5f, 0x62, 0x69, 0x61, 0x73, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x0c, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x00, 0x00,
- 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,
- 0x25, 0xf5, 0xe8, 0x37, 0x03, 0x00, 0x00, 0x00, 0xb4, 0x00, 0x00, 0x00,
- 0x5c, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xc0, 0xff, 0xff, 0xff,
- 0x00, 0x00, 0x00, 0x09, 0x02, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
- 0x10, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00,
- 0x08, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f,
- 0x14, 0x00, 0x1c, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x07, 0x00,
- 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x18, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x08, 0x01, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,
- 0x10, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x03, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00,
- 0x14, 0x00, 0x18, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x07, 0x00,
- 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x14, 0x00, 0x14, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x02, 0x1c, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,
- 0x30, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x01, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,
- 0x03, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
- 0x0c, 0x00, 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00,
- 0x0c, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00,
- 0x08, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,
- 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xfa, 0xff, 0xff, 0xff,
- 0x00, 0x19, 0x06, 0x00, 0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00,
- 0x00, 0x09, 0x06, 0x00, 0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00,
- 0x00, 0x00, 0x00, 0x04};
-const int g_tiny_conv_micro_features_model_data_len = 18208;
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h b/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h
deleted file mode 100644
index b14f464..0000000
--- a/tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h
+++ /dev/null
@@ -1,27 +0,0 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// This is a standard TensorFlow Lite model file that has been converted into a
-// C data array, so it can be easily compiled into a binary for devices that
-// don't have a file system. It was created using the command:
-// xxd -i tiny_conv.tflite > tiny_conv_simple_features_model_data.cc
-
-#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_TINY_CONV_MICRO_FEATURES_MODEL_DATA_H_
-#define TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_TINY_CONV_MICRO_FEATURES_MODEL_DATA_H_
-
-extern const unsigned char g_tiny_conv_micro_features_model_data[];
-extern const int g_tiny_conv_micro_features_model_data_len;
-
-#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_MICRO_FEATURES_TINY_CONV_MICRO_FEATURES_MODEL_DATA_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc b/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc
index b48cee1..ca090ec 100644
--- a/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc
+++ b/tensorflow/lite/micro/examples/micro_speech/micro_speech_test.cc
@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -13,8 +13,8 @@
limitations under the License.
==============================================================================*/
+#include "tensorflow/lite/micro/examples/micro_speech/micro_features/model.h"
#include "tensorflow/lite/micro/examples/micro_speech/micro_features/no_micro_features_data.h"
-#include "tensorflow/lite/micro/examples/micro_speech/micro_features/tiny_conv_micro_features_model_data.h"
#include "tensorflow/lite/micro/examples/micro_speech/micro_features/yes_micro_features_data.h"
#include "tensorflow/lite/micro/kernels/micro_ops.h"
#include "tensorflow/lite/micro/micro_error_reporter.h"
@@ -33,8 +33,7 @@
// Map the model into a usable data structure. This doesn't involve any
// copying or parsing, it's a very lightweight operation.
- const tflite::Model* model =
- ::tflite::GetModel(g_tiny_conv_micro_features_model_data);
+ const tflite::Model* model = ::tflite::GetModel(g_model);
if (model->version() != TFLITE_SCHEMA_VERSION) {
TF_LITE_REPORT_ERROR(error_reporter,
"Model provided is schema version %d not equal "
diff --git a/tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.cc b/tensorflow/lite/micro/examples/micro_speech/simple_features/model.cc
similarity index 99%
rename from tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.cc
rename to tensorflow/lite/micro/examples/micro_speech/simple_features/model.cc
index 2c66e39..e8fea5b 100644
--- a/tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.cc
+++ b/tensorflow/lite/micro/examples/micro_speech/simple_features/model.cc
@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
@@ -13,13 +13,14 @@
limitations under the License.
==============================================================================*/
-// Automatically created from a TensorFlow Lite flatbuffer using the command:
-// xxd -i tiny_conv.tflite > tiny_conv_simple_features_model_data.cc
-// See the README for a full description of the creation process.
+// This is a standard TensorFlow Lite FlatBuffer model file that has been
+// converted into a C data array, so it can be easily compiled into a binary
+// for devices that don't have a file system. It was created using the command:
+// xxd -i model.tflite > model.cc
-#include "tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.h"
+#include "tensorflow/lite/micro/examples/micro_speech/simple_features/model.h"
-const unsigned char g_tiny_conv_simple_features_model_data[] = {
+const unsigned char g_model[] = {
0x18, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x0e, 0x00,
0x18, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00,
0x0e, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x08, 0x4d, 0x00, 0x00,
@@ -1670,4 +1671,4 @@
0x04, 0x00, 0x00, 0x00, 0xfa, 0xff, 0xff, 0xff, 0x00, 0x19, 0x06, 0x00,
0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x09, 0x06, 0x00,
0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04};
-const int g_tiny_conv_simple_features_model_data_len = 19800;
+const int g_model_len = 19800;
diff --git a/tensorflow/lite/micro/examples/micro_speech/simple_features/model.h b/tensorflow/lite/micro/examples/micro_speech/simple_features/model.h
new file mode 100644
index 0000000..b3e705e
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/simple_features/model.h
@@ -0,0 +1,27 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This is a standard TensorFlow Lite FlatBuffer model file that has been
+// converted into a C data array, so it can be easily compiled into a binary
+// for devices that don't have a file system. It was created using the command:
+// xxd -i model.tflite > model.cc
+
+#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_MODEL_H_
+#define TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_MODEL_H_
+
+extern const unsigned char g_model[];
+extern const int g_model_len;
+
+#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_MODEL_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.h b/tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.h
deleted file mode 100644
index a97d790..0000000
--- a/tensorflow/lite/micro/examples/micro_speech/simple_features/tiny_conv_simple_features_model_data.h
+++ /dev/null
@@ -1,27 +0,0 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
- http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// This is a standard TensorFlow Lite model file that has been converted into a
-// C data array, so it can be easily compiled into a binary for devices that
-// don't have a file system. It was created using the command:
-// xxd -i tiny_conv.tflite > tiny_conv_simple_features_model_data.cc
-
-#ifndef TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_TINY_CONV_SIMPLE_FEATURES_MODEL_DATA_H_
-#define TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_TINY_CONV_SIMPLE_FEATURES_MODEL_DATA_H_
-
-extern const unsigned char g_tiny_conv_simple_features_model_data[];
-extern const int g_tiny_conv_simple_features_model_data_len;
-
-#endif // TENSORFLOW_LITE_MICRO_EXAMPLES_MICRO_SPEECH_SIMPLE_FEATURES_TINY_CONV_SIMPLE_FEATURES_MODEL_DATA_H_
diff --git a/tensorflow/lite/micro/examples/micro_speech/train/README.md b/tensorflow/lite/micro/examples/micro_speech/train/README.md
new file mode 100644
index 0000000..5793985
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/train/README.md
@@ -0,0 +1,210 @@
+
+# Micro Speech Training
+
+This example shows how to train a 20 kB model that can recognize 2 keywords,
+"yes" and "no", from speech data.
+
+If the input does not belong to either categories, it classifies it as "unknown"
+and if the input is silent, it classifies it as "silence".
+
+You can retrain it to recognize any combination of words (2 or more) from this
+list:
+
+```
+yes
+no
+up
+down
+left
+right
+on
+off
+stop
+go
+```
+
+## Table of contents
+
+- [Overview](#overview)
+- [Trained Models](#trained-models)
+- [Training](#training)
+- [Model Architecture](#model-architecture)
+- [Dataset](#dataset)
+- [Preprocessing Speech Input](#preprocessing-speech-input)
+
+## Overview
+
+1. Training Jupyter Notebook: [`train_micro_speech_model.ipynb`]
+(https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb)
+. The training scripts used in this notebook is defined the
+[Simple Audio Recognition](https://www.tensorflow.org/tutorials/sequences/audio_recognition)
+tutorial.
+2. Dataset Type: **Speech**
+3. Dataset: Speech Commands, Version 2. ([Download Link](https://storage.cloud.google.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz)
+, [Paper](https://arxiv.org/abs/1804.03209))
+4. Deep Learning Framework: **TensorFlow 1.5**
+5. Language: **Python 3.7**
+6. Model Size: **<20 kB**
+7. Model Category: **Multiclass Classification**
+
+## Trained Models
+
+| Download Link | [speech_commands.zip](https://storage.googleapis.com/download.tensorflow.org/models/tflite/micro/speech_commands_2020_04_13.zip) |
+| ------------- |-------------|
+
+
+The `models` directory in the above zip file can be generated by running the
+colab notebook in the [Training](#training) section below. It
+includes the following 3 model files:
+
+| Name | Format | Target Framework | Target Device |
+| :------------- |:-------------|:-------------|-----|
+| `model.pb` | Frozen GraphDef | TensorFlow | Large-Scale/Cloud/Servers |
+| `model.tflite` *(<20 kB)* | Fully Quantized* TFLite Model |
+TensorFlow Lite | Mobile Devices|
+| `model.cc` | C Source File | TensorFlow Lite for Microcontrollers |
+Microcontrollers |
+
+**Fully quantized implies that the model is **strictly int8** quantized
+including the input(s)and output(s).*
+<!-- **Fully quantized implies that the model is **strictly int8** quantized
+**except** the input(s) and output(s) which remain float.* -->
+
+
+## Training
+
+You can train your own models using any of the following methods. We recommend
+that you try these methods in the order mentioned below.
+
+### 1. Use [Google Colaboratory](https://colab.research.google.com)
+
+*We strongly recommend trying this approach first.*
+
+<table class="tfo-notebook-buttons">
+ <td>
+ <a target="_blank" href="https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb"><img src="https://www.tensorflow.org/images/colab_logo_32px.png" />Run in
+ Google Colab</a>
+ </td>
+</table>
+
+**Estimated Training Time:** ~2 hours.
+**Advantage:** It allows the use of a free Tesla K80 GPU for training and avoids
+the need to install dependencies.
+**Disadvantage:** Your training time is limited as the session can only run
+upto 12 hours in a row if you keep the browser open and 90 minutes if you close
+the browser.
+
+### 2. Use Google Cloud
+
+1. Create a Virtual Machine (VM) using a pre-configured Deep Learning VM Image.
+
+```
+export IMAGE_FAMILY="tf-latest-cpu"
+export ZONE="us-west1-b" # Or any other required region
+export INSTANCE_NAME="model-trainer"
+export INSTANCE_TYPE="n1-standard-8" # or any other instance type
+gcloud compute instances create $INSTANCE_NAME \
+ --zone=$ZONE \
+ --image-family=$IMAGE_FAMILY \
+ --image-project=deeplearning-platform-release \
+ --machine-type=$INSTANCE_TYPE \
+ --boot-disk-size=120GB \
+ --min-cpu-platform=Intel\ Skylake
+```
+
+2. As soon as instance has been created you can SSH to it:
+
+```
+gcloud compute ssh "jupyter@${INSTANCE_NAME}"
+```
+
+3. Train a model by following the instructions in the [`train_micro_speech_model.ipynb`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb)
+jupyter notebook.
+
+4. Finally, don't forget to remove the instance when training is done:
+
+```
+gcloud compute instances delete "${INSTANCE_NAME}" --zone="${ZONE}"
+```
+
+**Estimated Training Time:** ~2 hours (with GPU) and ~1 day (with CPU).
+**Advantage:** There are no time constraints on how long the training process
+can take and it avoids the need to install dependencies.
+**Disadvantage:** Google Cloud isn't free. You will have to pay a certain amount
+depending on how long you use run the VM and what resources you use.
+
+## Model Architecture
+
+This is a simple model comprising of a 2D Convolutional layer, a Fully Connected
+Layer (outputs: logits) and finally a Softmax layer (outputs: probabilities) as
+shown below. Refer to the [`tiny_conv`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/speech_commands/models.py#L673)
+model architecture.
+
+
+
+*This image was generated by visualizing the 'model.tflite' file in
+[Netron](https://github.com/lutzroeder/netron)*
+
+This doesn't produce a highly accurate model, but it's designed to be used as
+the first stage of a pipeline, running on a low-energy piece of hardware that
+can always be on, and then wake higher-power chips when a possible utterance has
+been found, so that more accurate analysis can be done. Additionally, the model
+takes in preprocessed speech input as a result of which we can leverage a
+simpler model for accurate results.
+
+## Dataset
+
+The Speech Commands Dataset. ([Download Link](https://storage.cloud.google.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz)
+,[Paper](https://arxiv.org/abs/1804.03209)) consists of over 105,000 WAVE audio
+files of people saying thirty different words. This data was collected by
+Google and released under a CC BY license. You can help improve it by
+contributing five minutes of your own voice. The archive is over 2GB, so this
+part may take a while, but you should see progress logs, and once it's been
+downloaded you won't need to do this again.
+
+## Preprocessing Speech Input
+
+In this section we discuss spectrograms, the preprocessed speech input to the
+model. Here's an illustration of the process:
+
+
+
+The model doesn't take in raw audio sample data, instead it works with
+spectrograms, which are two dimensional arrays that are made up of slices of
+frequency information, each taken from a different time window.
+
+The recipe for creating the spectrogram data is that each frequency slice is
+created by running an FFT across a 30ms section of the audio sample data. The
+input samples are treated as being between -1 and +1 as real values (encoded as
+-32,768 and 32,767 in 16-bit signed integer samples).
+
+This results in an FFT with 256 entries. Every sequence of six entries is
+averaged together, giving a total of 43 frequency buckets in the final slice.
+The results are stored as unsigned eight-bit values, where 0 represents a real
+number of zero, and 255 represents 127.5 as a real number.
+
+Each adjacent frequency entry is stored in ascending memory order (frequency
+bucket 0 at data[0], bucket 1 at data[1], etc). The window for the frequency
+analysis is then moved forward by 20ms, and the process repeated, storing the
+results in the next memory row (for example bucket 0 in this moved window would
+be in data[43 + 0], etc). This process happens 49 times in total, producing a
+single channel image that is 43 pixels wide, and 49 rows high.
+
+In a complete application these spectrograms would be calculated at runtime from
+microphone inputs, but the code for doing that is not yet included in this
+sample code. The test uses spectrograms that have been pre-calculated from
+one-second WAV files in the test dataset generated by running the following
+commands:
+
+```
+python tensorflow/tensorflow/examples/speech_commands/wav_to_features.py \
+--input_wav=/tmp/speech_dataset/yes/f2e59fea_nohash_1.wav \
+--output_c_file=/tmp/yes_features_data.cc \
+--window_stride=20 --preprocess=average --quantize=1
+
+python tensorflow/tensorflow/examples/speech_commands/wav_to_features.py \
+--input_wav=/tmp/speech_dataset/no/f9643d42_nohash_4.wav \
+--output_c_file=/tmp/no_features_data.cc \
+--window_stride=20 --preprocess=average --quantize=1
+```
+
diff --git a/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb b/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb
new file mode 100644
index 0000000..40f56f8
--- /dev/null
+++ b/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb
@@ -0,0 +1,2020 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+ "colab": {
+ "name": "train_micro_speech_model.ipynb",
+ "provenance": [],
+ "collapsed_sections": []
+ },
+ "kernelspec": {
+ "name": "python3",
+ "display_name": "Python 3"
+ },
+ "accelerator": "GPU"
+ },
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "pO4-CY_TCZZS",
+ "colab_type": "text"
+ },
+ "source": [
+ "# Train a Simple Audio Recognition Model"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "BaFfr7DHRmGF",
+ "colab_type": "text"
+ },
+ "source": [
+ "This notebook demonstrates how to train a 20 kB [Simple Audio Recognition](https://www.tensorflow.org/tutorials/sequences/audio_recognition) model to recognize keywords in speech.\n",
+ "\n",
+ "The model created in this notebook is used in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) example for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview).\n",
+ "\n",
+ "<table class=\"tfo-notebook-buttons\" align=\"left\">\n",
+ " <td>\n",
+ " <a target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />Run in Google Colab</a>\n",
+ " </td>\n",
+ " <td>\n",
+ " <a target=\"_blank\" href=\"https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb\"><img src=\"https://www.tensorflow.org/images/GitHub-Mark-32px.png\" />View source on GitHub</a>\n",
+ " </td>\n",
+ "</table>\n"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "XaVtYN4nlCft",
+ "colab_type": "text"
+ },
+ "source": [
+ "**Training is much faster using GPU acceleration.** Before you proceed, ensure you are using a GPU runtime by going to **Runtime -> Change runtime type** and set **Hardware accelerator: GPU**. Training 15,000 iterations will take 1.5 - 2 hours on a GPU runtime.\n",
+ "\n",
+ "## Configure Defaults\n",
+ "\n",
+ "**MODIFY** the following constants for your specific use case."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "ludfxbNIaegy",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "# A comma-delimited list of the words you want to train for.\n",
+ "# The options are: yes,no,up,down,left,right,on,off,stop,go\n",
+ "# All the other words will be used to train an \"unknown\" label and silent\n",
+ "# audio data with no spoken words will be used to train a \"silence\" label.\n",
+ "WANTED_WORDS = \"yes,no\"\n",
+ "\n",
+ "# The number of steps and learning rates can be specified as comma-separated\n",
+ "# lists to define the rate at each stage. For example,\n",
+ "# TRAINING_STEPS=12000,3000 and LEARNING_RATE=0.001,0.0001\n",
+ "# will run 12,000 training loops in total, with a rate of 0.001 for the first\n",
+ "# 8,000, and 0.0001 for the final 3,000.\n",
+ "TRAINING_STEPS = \"12000,3000\"\n",
+ "LEARNING_RATE = \"0.001,0.0001\"\n",
+ "\n",
+ "# Calculate the total number of steps, which is used to identify the checkpoint\n",
+ "# file name.\n",
+ "TOTAL_STEPS = str(sum(map(lambda string: int(string), TRAINING_STEPS.split(\",\"))))\n",
+ "\n",
+ "# Print the configuration to confirm it\n",
+ "!echo \"Training these words:\" $WANTED_WORDS\n",
+ "!echo \"Training steps in each stage:\" $TRAINING_STEPS\n",
+ "!echo \"Learning rate in each stage:\" $LEARNING_RATE\n",
+ "!echo \"Total number of training steps:\" $TOTAL_STEPS"
+ ],
+ "execution_count": 1,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "Training these words: yes,no\n",
+ "Training steps in each stage: 12000,3000\n",
+ "Learning rate in each stage: 0.001,0.0001\n",
+ "Total number of training steps: 15000\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "gCgeOpvY9pAi",
+ "colab_type": "text"
+ },
+ "source": [
+ "**DO NOT MODIFY** the following constants as they include filepaths used in this notebook and data that is shared during training and inference."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "Nd1iM1o2ymvA",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "# Calculate the percentage of 'silence' and 'unknown' training samples required\n",
+ "# to ensure that we have equal number of samples for each label.\n",
+ "number_of_labels = WANTED_WORDS.count(',') + 1\n",
+ "number_of_total_labels = number_of_labels + 2 # for 'silence' and 'unknown' label\n",
+ "equal_percentage_of_training_samples = int(100.0/(number_of_total_labels))\n",
+ "SILENT_PERCENTAGE = equal_percentage_of_training_samples\n",
+ "UNKNOWN_PERCENTAGE = equal_percentage_of_training_samples\n",
+ "\n",
+ "# Constants which are shared during training and inference\n",
+ "PREPROCESS = 'micro'\n",
+ "WINDOW_STRIDE ='20'\n",
+ "MODEL_ARCHITECTURE = 'tiny_conv' # Other options include: single_fc, conv,\n",
+ " # low_latency_conv, low_latency_svdf, tiny_embedding_conv\n",
+ "QUANTIZE = '1' # For booleans, we provide 1 or 0 (instead of True or False)\n",
+ "\n",
+ "# Constants used during training only\n",
+ "VERBOSITY = 'WARN'\n",
+ "EVAL_STEP_INTERVAL = '1000'\n",
+ "SAVE_STEP_INTERVAL = '5000'\n",
+ "\n",
+ "# Constants for training directories and filepaths\n",
+ "DATASET_DIR = 'dataset/'\n",
+ "LOGS_DIR = 'logs/'\n",
+ "TRAIN_DIR = 'train/' # for training checkpoints and other files.\n",
+ "\n",
+ "# Constants for inference directories and filepaths\n",
+ "import os\n",
+ "MODELS_DIR = 'models/'\n",
+ "os.mkdir(MODELS_DIR)\n",
+ "MODEL_TF = MODELS_DIR + 'model.pb'\n",
+ "MODEL_TFLITE = MODELS_DIR + 'model.tflite'\n",
+ "MODEL_TFLITE_MICRO = MODELS_DIR + 'model.cc'"
+ ],
+ "execution_count": 2,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "6rLYpvtg9P4o",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Setup Environment\n",
+ "\n",
+ "Install Dependencies"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "ed_XpUrU5DvY",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "%tensorflow_version 1.x\n",
+ "import tensorflow as tf"
+ ],
+ "execution_count": 3,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "TensorFlow 1.x selected.\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "T9Ty5mR58E4i",
+ "colab_type": "text"
+ },
+ "source": [
+ "**DELETE** any old data from previous runs\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "APGx0fEh7hFF",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "!rm -rf {DATASET_DIR} {LOGS_DIR} {TRAIN_DIR} {MODELS_DIR}"
+ ],
+ "execution_count": 5,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "GfEUlfFBizio",
+ "colab_type": "text"
+ },
+ "source": [
+ "Clone the TensorFlow Github Repository, which contains the relevant code required to run this tutorial."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "yZArmzT85SLq",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "!git clone -q https://github.com/tensorflow/tensorflow"
+ ],
+ "execution_count": 6,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "nS9swHLSi7Bi",
+ "colab_type": "text"
+ },
+ "source": [
+ "Load TensorBoard to visualize the accuracy and loss as training proceeds.\n"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "q4qF1VxP3UE4",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "%load_ext tensorboard\n",
+ "%tensorboard --logdir {LOGS_DIR}"
+ ],
+ "execution_count": 7,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "x1J96Ron-O4R",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Training\n",
+ "\n",
+ "The following script downloads the dataset and begin training."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "VJsEZx6lynbY",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "!python tensorflow/tensorflow/examples/speech_commands/train.py \\\n",
+ "--data_dir={DATASET_DIR} \\\n",
+ "--wanted_words={WANTED_WORDS} \\\n",
+ "--silence_percentage={SILENT_PERCENTAGE} \\\n",
+ "--unknown_percentage={UNKNOWN_PERCENTAGE} \\\n",
+ "--preprocess={PREPROCESS} \\\n",
+ "--window_stride={WINDOW_STRIDE} \\\n",
+ "--model_architecture={MODEL_ARCHITECTURE} \\\n",
+ "--quantize={QUANTIZE} \\\n",
+ "--how_many_training_steps={TRAINING_STEPS} \\\n",
+ "--learning_rate={LEARNING_RATE} \\\n",
+ "--train_dir={TRAIN_DIR} \\\n",
+ "--summaries_dir={LOGS_DIR} \\\n",
+ "--verbosity={VERBOSITY} \\\n",
+ "--eval_step_interval={EVAL_STEP_INTERVAL} \\\n",
+ "--save_step_interval={SAVE_STEP_INTERVAL} \\"
+ ],
+ "execution_count": 8,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "..\n",
+ "..\n",
+ "..\n",
+ "..\n",
+ "WARNING:tensorflow:Confusion Matrix:\n",
+ " [[205 0 0 1]\n",
+ " [ 3 162 13 28]\n",
+ " [ 3 9 401 6]\n",
+ " [ 2 22 6 375]]\n",
+ "W0402 00:25:28.115174 139938153863040 train.py:320] Confusion Matrix:\n",
+ " [[205 0 0 1]\n",
+ " [ 3 162 13 28]\n",
+ " [ 3 9 401 6]\n",
+ " [ 2 22 6 375]]\n",
+ "WARNING:tensorflow:Final test accuracy = 92.5% (N=1236)\n",
+ "W0402 00:25:28.115574 139938153863040 train.py:322] Final test accuracy = 92.5% (N=1236)\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "XQUJLrdS-ftl",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Generate a TensorFlow Model for Inference\n",
+ "\n",
+ "Combine relevant training results (graph, weights, etc) into a single file for inference. This process is known as freezing a model and the resulting model is known as a frozen model/graph, as it cannot be further re-trained after this process."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "xyc3_eLh9sAg",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "!python tensorflow/tensorflow/examples/speech_commands/freeze.py \\\n",
+ "--wanted_words=$WANTED_WORDS \\\n",
+ "--window_stride_ms=$WINDOW_STRIDE \\\n",
+ "--preprocess=$PREPROCESS \\\n",
+ "--model_architecture=$MODEL_ARCHITECTURE \\\n",
+ "--quantize=$QUANTIZE \\\n",
+ "--start_checkpoint=$TRAIN_DIR$MODEL_ARCHITECTURE'.ckpt-'$TOTAL_STEPS \\\n",
+ "--output_file=$MODEL_TF \\"
+ ],
+ "execution_count": 9,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "..\n",
+ "..\n",
+ "..\n",
+ "..\n",
+ "INFO:tensorflow:Restoring parameters from /content/train/tiny_conv.ckpt-15000\n",
+ "I0402 00:25:47.086113 140352379615104 saver.py:1284] Restoring parameters from /content/train/tiny_conv.ckpt-15000\n",
+ "INFO:tensorflow:Froze 12 variables.\n",
+ "I0402 00:25:47.663757 140352379615104 graph_util_impl.py:334] Froze 12 variables.\n",
+ "INFO:tensorflow:Converted 12 variables to const ops.\n",
+ "I0402 00:25:47.665771 140352379615104 graph_util_impl.py:394] Converted 12 variables to const ops.\n",
+ "INFO:tensorflow:Saved frozen graph to /content/models/model.pb\n",
+ "I0402 00:25:47.667117 140352379615104 freeze.py:186] Saved frozen graph to /content/models/model.pb\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "_DBGDxVI-nKG",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Generate a TensorFlow Lite Model\n",
+ "\n",
+ "Convert the frozen graph into a TensorFlow Lite model, which is fully quantized for use with embedded devices.\n",
+ "\n",
+ "The following cell will also print the model size, which will be under 20 kilobytes."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "lBj_AyCh1cC0",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "input_tensor = 'Reshape_2'\n",
+ "output_tensor = 'labels_softmax'\n",
+ "\n",
+ "converter = tf.lite.TFLiteConverter.from_frozen_graph(\n",
+ " MODEL_TF, [input_tensor], [output_tensor])\n",
+ "converter.inference_type = tf.uint8\n",
+ "converter.quantized_input_stats = {input_tensor: (0.0, 9.8077)} # (mean, standard deviation)\n",
+ "tflite_model = converter.convert()\n",
+ "\n",
+ "tflite_model_size = open(MODEL_TFLITE, \"wb\").write(tflite_model)\n",
+ "print(\"Model is %d bytes\" % tflite_model_size)\n"
+ ],
+ "execution_count": 10,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "Model is 18288 bytes\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "dt6Zqbxu-wIi",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Generate a TensorFlow Lite for MicroControllers Model\n",
+ "Convert the TensorFlow Lite model into a C source file that can be loaded by TensorFlow Lite for Microcontrollers."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "XohZOTjR8ZyE",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "# Install xxd if it is not available\n",
+ "!apt-get update && apt-get -qq install xxd\n",
+ "# Convert to a C source file\n",
+ "!xxd -i {MODEL_TFLITE} > {MODEL_TFLITE_MICRO}\n",
+ "# Update variable names\n",
+ "REPLACE_TEXT = MODEL_TFLITE.replace('/', '_').replace('.', '_')\n",
+ "!sed -i 's/'{REPLACE_TEXT}'/g_model/g' {MODEL_TFLITE_MICRO}"
+ ],
+ "execution_count": 11,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "id": "2pQnN0i_-0L2",
+ "colab_type": "text"
+ },
+ "source": [
+ "## Deploy to a Microcontroller\n",
+ "\n",
+ "Follow the instructions in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) README.md for [TensorFlow Lite for MicroControllers](https://www.tensorflow.org/lite/microcontrollers/overview) to deploy this model on a specific microcontroller.\n",
+ "\n",
+ "**Reference Model:** If you have not modified this notebook, you can follow the instructions as is, to deploy the model. Refer to the [`micro_speech/train/models`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train/models) directory to access the models generated in this notebook. \n",
+ "\n",
+ "**New Model:** If you have generated a new model to identify different words: (i) Update `kCategoryCount` and `kCategoryLabels` in [`micro_speech/micro_features/micro_model_settings.h`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/micro_model_settings.h) and (ii) Update the values assigned to the variables defined in [`micro_speech/micro_features/model.cc`](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/micro_features/model.cc) with values displayed after running the following cell."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "metadata": {
+ "id": "eoYyh0VU8pca",
+ "colab_type": "code",
+ "colab": {}
+ },
+ "source": [
+ "# Print the C source file\n",
+ "!cat {MODEL_TFLITE_MICRO}"
+ ],
+ "execution_count": 12,
+ "outputs": [
+ {
+ "output_type": "stream",
+ "text": [
+ "unsigned char g_model[] = {\n",
+ " 0x1c, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x00, 0x00, 0x12, 0x00,\n",
+ " 0x1c, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00,\n",
+ " 0x00, 0x00, 0x18, 0x00, 0x12, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,\n",
+ " 0x1c, 0x47, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,\n",
+ " 0x2c, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0xc0, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x94, 0x00, 0x00, 0x00,\n",
+ " 0x0f, 0x00, 0x00, 0x00, 0x54, 0x4f, 0x43, 0x4f, 0x20, 0x43, 0x6f, 0x6e,\n",
+ " 0x76, 0x65, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x00, 0x0a, 0x00, 0x00, 0x00,\n",
+ " 0x60, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00, 0x48, 0x00, 0x00, 0x00,\n",
+ " 0x3c, 0x00, 0x00, 0x00, 0x34, 0x00, 0x00, 0x00, 0x2c, 0x00, 0x00, 0x00,\n",
+ " 0x20, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x0e, 0xba, 0xff, 0xff, 0x38, 0x00, 0x00, 0x00,\n",
+ " 0xbc, 0xb9, 0xff, 0xff, 0xc0, 0xb9, 0xff, 0xff, 0x1e, 0xba, 0xff, 0xff,\n",
+ " 0xe0, 0x01, 0x00, 0x00, 0xcc, 0xb9, 0xff, 0xff, 0xd0, 0xb9, 0xff, 0xff,\n",
+ " 0x2e, 0xba, 0xff, 0xff, 0x60, 0x03, 0x00, 0x00, 0x36, 0xba, 0xff, 0xff,\n",
+ " 0x7c, 0x06, 0x00, 0x00, 0x3e, 0xba, 0xff, 0xff, 0x68, 0x45, 0x00, 0x00,\n",
+ " 0xec, 0xb9, 0xff, 0xff, 0x05, 0x00, 0x00, 0x00, 0x31, 0x2e, 0x35, 0x2e,\n",
+ " 0x30, 0x00, 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x04, 0x00, 0x08, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00,\n",
+ " 0x13, 0x00, 0x00, 0x00, 0x6d, 0x69, 0x6e, 0x5f, 0x72, 0x75, 0x6e, 0x74,\n",
+ " 0x69, 0x6d, 0x65, 0x5f, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x00,\n",
+ " 0x10, 0xfa, 0xff, 0xff, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x2c, 0x45, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x05, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x04, 0x01, 0x00, 0x00, 0x9c, 0x44, 0x00, 0x00,\n",
+ " 0x8c, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0xdc, 0x01, 0x00, 0x00,\n",
+ " 0x68, 0x01, 0x00, 0x00, 0x3c, 0x02, 0x00, 0x00, 0x50, 0x05, 0x00, 0x00,\n",
+ " 0x8e, 0xbb, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x31, 0x00, 0x00, 0x00,\n",
+ " 0x28, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x09, 0x00, 0x00, 0x00,\n",
+ " 0x52, 0x65, 0x73, 0x68, 0x61, 0x70, 0x65, 0x5f, 0x32, 0x00, 0x00, 0x00,\n",
+ " 0x94, 0xfa, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00,\n",
+ " 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0xc6, 0xd0, 0xd0, 0x3d, 0x01, 0x00, 0x00, 0x00, 0xf5, 0xff, 0xcf, 0x41,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xbc, 0xff, 0xff,\n",
+ " 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00,\n",
+ " 0x1c, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x52, 0x65, 0x6c, 0x75,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x04, 0xfb, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00,\n",
+ " 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x09, 0xf5, 0x83, 0x3d, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x14, 0x71, 0x83, 0x41, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x72, 0xbc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x02, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x06, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x0b, 0x00, 0x00, 0x00,\n",
+ " 0x43, 0x6f, 0x6e, 0x76, 0x32, 0x44, 0x5f, 0x62, 0x69, 0x61, 0x73, 0x00,\n",
+ " 0x64, 0xbc, 0xff, 0xff, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x2d, 0x95, 0x98, 0x38,\n",
+ " 0x20, 0x00, 0x00, 0x00, 0x27, 0xff, 0xff, 0xff, 0x97, 0xff, 0xff, 0xff,\n",
+ " 0x58, 0x00, 0x00, 0x00, 0x66, 0xff, 0xff, 0xff, 0x13, 0xff, 0xff, 0xff,\n",
+ " 0x72, 0xfe, 0xff, 0xff, 0x5d, 0xfb, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0xea, 0xbc, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x05, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x0e, 0x00, 0x00, 0x00, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x73, 0x5f, 0x73,\n",
+ " 0x6f, 0x66, 0x74, 0x6d, 0x61, 0x78, 0x00, 0x00, 0xec, 0xfb, 0xff, 0xff,\n",
+ " 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3b,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x7f, 0x3f, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x5a, 0xbd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03,\n",
+ " 0x10, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x1c, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00, 0x61, 0x64, 0x64, 0x5f,\n",
+ " 0x31, 0x00, 0x00, 0x00, 0x54, 0xfc, 0xff, 0xff, 0x2c, 0x00, 0x00, 0x00,\n",
+ " 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x8f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x9c, 0xd2, 0xb5, 0x3d, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x48, 0x18, 0x1f, 0x41, 0x01, 0x00, 0x00, 0x00, 0x4a, 0x21, 0x4b, 0xc1,\n",
+ " 0xc2, 0xbd, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x03, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x0a, 0x00, 0x00, 0x00,\n",
+ " 0x08, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00,\n",
+ " 0x77, 0x65, 0x69, 0x67, 0x68, 0x74, 0x73, 0x5f, 0x71, 0x75, 0x61, 0x6e,\n",
+ " 0x74, 0x2f, 0x46, 0x61, 0x6b, 0x65, 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57,\n",
+ " 0x69, 0x74, 0x68, 0x4d, 0x69, 0x6e, 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72,\n",
+ " 0x73, 0x00, 0x00, 0x00, 0xe4, 0xfc, 0xff, 0xff, 0x30, 0x00, 0x00, 0x00,\n",
+ " 0x24, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x8a, 0x0f, 0x3b, 0x3a,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0xfc, 0x0b, 0xb4, 0x3d, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0xd9, 0x26, 0xbf, 0xbd, 0x80, 0x02, 0x00, 0x00, 0x60, 0x38, 0xab, 0xcb,\n",
+ " 0xfa, 0x7e, 0xa2, 0x55, 0x6e, 0x87, 0xa5, 0x9b, 0xb4, 0x66, 0x5c, 0x6f,\n",
+ " 0xae, 0xdb, 0xcd, 0xb6, 0xc2, 0x60, 0xa9, 0x7d, 0xd4, 0xac, 0xa6, 0x90,\n",
+ " 0x87, 0x6b, 0x50, 0x95, 0xde, 0xcd, 0xaa, 0xa1, 0x9c, 0x65, 0xb5, 0x6d,\n",
+ " 0xb0, 0xa5, 0xa5, 0x7f, 0x73, 0x95, 0x63, 0x81, 0x7a, 0xc6, 0xaf, 0x82,\n",
+ " 0x69, 0x89, 0xc3, 0x3c, 0x47, 0x73, 0x89, 0x4f, 0x33, 0xbc, 0x85, 0x5d,\n",
+ " 0x69, 0x11, 0x5b, 0xb9, 0xf1, 0x95, 0x8f, 0x5c, 0x7c, 0x59, 0x6c, 0xa0,\n",
+ " 0xa5, 0x7c, 0x5a, 0x7c, 0xb5, 0xa9, 0x7e, 0xa1, 0xb8, 0x65, 0xb3, 0x86,\n",
+ " 0xc1, 0x9f, 0x5c, 0x86, 0x7f, 0x74, 0x52, 0xa8, 0xc9, 0xc5, 0x71, 0x96,\n",
+ " 0x7a, 0x65, 0xc7, 0x69, 0x94, 0xa7, 0x65, 0x68, 0x69, 0x8d, 0x6d, 0x9e,\n",
+ " 0x59, 0xd4, 0x75, 0x7a, 0x4f, 0x70, 0xca, 0x48, 0x25, 0x8a, 0x69, 0x4d,\n",
+ " 0x2a, 0xa6, 0x76, 0x69, 0x6a, 0x02, 0x3b, 0xa2, 0xea, 0xc2, 0x73, 0x6b,\n",
+ " 0x86, 0x4d, 0x3a, 0xa2, 0xa2, 0x88, 0x4e, 0x6c, 0xb3, 0x83, 0x39, 0x93,\n",
+ " 0xa6, 0x85, 0xb8, 0x7a, 0xa8, 0x7d, 0x2e, 0x7b, 0x7f, 0x69, 0x56, 0xb5,\n",
+ " 0xbb, 0xae, 0x23, 0x78, 0x67, 0x5c, 0xd2, 0x82, 0x7d, 0x96, 0x46, 0x74,\n",
+ " 0x70, 0x72, 0x6a, 0x90, 0x43, 0xce, 0x44, 0x75, 0x4a, 0x58, 0xc7, 0x5c,\n",
+ " 0x34, 0x84, 0x46, 0x4b, 0x41, 0x6c, 0x62, 0x83, 0x7e, 0x01, 0x9b, 0x9b,\n",
+ " 0xeb, 0xf7, 0x58, 0x6f, 0x8a, 0x43, 0xb3, 0x9f, 0x9c, 0x9e, 0x55, 0xa8,\n",
+ " 0xaa, 0x84, 0x8f, 0x8f, 0xb0, 0x9e, 0xc8, 0x81, 0xb6, 0x80, 0xa0, 0x81,\n",
+ " 0x86, 0x73, 0x5d, 0xdc, 0xb9, 0xae, 0xa2, 0x6c, 0x46, 0x67, 0xfa, 0x79,\n",
+ " 0x89, 0xaf, 0xa0, 0x74, 0x76, 0x85, 0x72, 0xb1, 0x2a, 0xbb, 0xa0, 0x6d,\n",
+ " 0x4f, 0x50, 0xc9, 0x5d, 0x2f, 0xaa, 0x9c, 0x63, 0x3f, 0x59, 0x63, 0x90,\n",
+ " 0x73, 0x1e, 0xb3, 0x94, 0xcd, 0xff, 0x3c, 0x63, 0x9b, 0x59, 0xc5, 0xa2,\n",
+ " 0x9f, 0x9a, 0x53, 0xab, 0xb0, 0x74, 0xb2, 0x6f, 0x8a, 0xa7, 0xd5, 0x8d,\n",
+ " 0xb8, 0x7e, 0x9e, 0x78, 0x84, 0x61, 0x66, 0xe7, 0xa7, 0x9f, 0xb7, 0x45,\n",
+ " 0x24, 0x61, 0xfd, 0x69, 0x87, 0xb8, 0xb2, 0x7a, 0x7c, 0x58, 0x64, 0xa3,\n",
+ " 0x07, 0xa9, 0xaf, 0x69, 0x49, 0x2f, 0xc2, 0x46, 0x3b, 0xaf, 0x9a, 0x70,\n",
+ " 0x6b, 0x25, 0x5f, 0x9d, 0x82, 0x33, 0xa1, 0x54, 0xae, 0xff, 0x31, 0x5d,\n",
+ " 0xaf, 0x51, 0xb2, 0x82, 0x9c, 0xa9, 0x5b, 0x8c, 0xab, 0x75, 0xb3, 0x32,\n",
+ " 0x42, 0xbd, 0xcd, 0x77, 0xb6, 0x67, 0x9a, 0x5f, 0x6c, 0x71, 0x6e, 0xc2,\n",
+ " 0xac, 0x97, 0x9f, 0x4b, 0x21, 0x6a, 0xfc, 0x77, 0x83, 0xa1, 0xa3, 0x6a,\n",
+ " 0x7a, 0x6d, 0x5e, 0x87, 0x02, 0xa6, 0x8f, 0x7f, 0x5c, 0x2e, 0xc1, 0x51,\n",
+ " 0x4a, 0xa7, 0x96, 0x79, 0x83, 0x2e, 0x5a, 0x84, 0x82, 0x5c, 0x61, 0x3a,\n",
+ " 0x4a, 0xff, 0x2a, 0x51, 0xa4, 0x6b, 0x82, 0x5e, 0x67, 0xb3, 0x71, 0x80,\n",
+ " 0xad, 0x62, 0x59, 0x40, 0x26, 0xd7, 0xcf, 0x68, 0xab, 0x7c, 0x6a, 0x69,\n",
+ " 0x5b, 0x7c, 0x84, 0xbc, 0x95, 0x68, 0x77, 0x63, 0x3f, 0x85, 0xed, 0x7b,\n",
+ " 0x71, 0xa0, 0x76, 0x90, 0x8c, 0x6c, 0x61, 0x81, 0x16, 0x74, 0x72, 0x94,\n",
+ " 0x74, 0x37, 0xb5, 0x3d, 0x55, 0x96, 0x86, 0xad, 0x87, 0x39, 0x59, 0x88,\n",
+ " 0x5b, 0x65, 0x60, 0x33, 0x33, 0xe6, 0x2b, 0x4a, 0xb6, 0x82, 0x50, 0x56,\n",
+ " 0x51, 0x97, 0x71, 0x83, 0xa6, 0x60, 0x57, 0x51, 0x58, 0xe4, 0xd0, 0x87,\n",
+ " 0xa1, 0x78, 0x4c, 0x67, 0x72, 0x74, 0x86, 0xc6, 0x60, 0x47, 0x50, 0x96,\n",
+ " 0x67, 0x96, 0xdd, 0x7d, 0x63, 0x85, 0x5e, 0x98, 0xa2, 0x64, 0x5f, 0x8a,\n",
+ " 0x3b, 0x40, 0x54, 0xcb, 0xa0, 0x61, 0xa7, 0x44, 0x5f, 0x6d, 0x57, 0xb3,\n",
+ " 0xb9, 0x2e, 0x61, 0x8e, 0x54, 0x78, 0x85, 0x58, 0x43, 0xb0, 0x27, 0x5d,\n",
+ " 0x8a, 0x7c, 0x8a, 0x58, 0x40, 0x83, 0x82, 0x9b, 0x6c, 0x60, 0x6b, 0x72,\n",
+ " 0x7f, 0xde, 0xc9, 0x7d, 0x6f, 0x5f, 0x90, 0x7e, 0x7e, 0x7e, 0x8b, 0xe5,\n",
+ " 0x51, 0x37, 0x7a, 0xa9, 0xa2, 0xc5, 0xd3, 0x81, 0x32, 0x4b, 0x80, 0xa9,\n",
+ " 0xc5, 0x76, 0x56, 0x99, 0x33, 0x19, 0x72, 0xe6, 0xdb, 0x90, 0xa8, 0x50,\n",
+ " 0x65, 0x44, 0x77, 0xdb, 0xc7, 0x48, 0x65, 0x8d, 0x3d, 0x7f, 0xa2, 0x7c,\n",
+ " 0x53, 0x55, 0x26, 0x49, 0x5d, 0x7d, 0xa2, 0x6d, 0x3b, 0x5b, 0x87, 0x64,\n",
+ " 0x3a, 0x5b, 0x8d, 0x93, 0x7a, 0xb4, 0xca, 0x6d, 0x16, 0x5a, 0x99, 0x82,\n",
+ " 0x8d, 0x6a, 0x92, 0xa0, 0x39, 0x2c, 0x95, 0xc8, 0xb8, 0xf5, 0xc8, 0x66,\n",
+ " 0x2a, 0x45, 0x84, 0x9c, 0xc7, 0x8e, 0x61, 0x7b, 0x43, 0x28, 0x86, 0xff,\n",
+ " 0xd2, 0xc8, 0x9c, 0x46, 0x65, 0x33, 0x82, 0xd8, 0xcb, 0x73, 0x63, 0x80,\n",
+ " 0xda, 0xc0, 0xff, 0xff, 0x00, 0x00, 0x00, 0x03, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x54, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0xa0, 0x0f, 0x00, 0x00,\n",
+ " 0x31, 0x00, 0x00, 0x00, 0x77, 0x65, 0x69, 0x67, 0x68, 0x74, 0x73, 0x5f,\n",
+ " 0x71, 0x75, 0x61, 0x6e, 0x74, 0x5f, 0x31, 0x2f, 0x46, 0x61, 0x6b, 0x65,\n",
+ " 0x51, 0x75, 0x61, 0x6e, 0x74, 0x57, 0x69, 0x74, 0x68, 0x4d, 0x69, 0x6e,\n",
+ " 0x4d, 0x61, 0x78, 0x56, 0x61, 0x72, 0x73, 0x2f, 0x74, 0x72, 0x61, 0x6e,\n",
+ " 0x73, 0x70, 0x6f, 0x73, 0x65, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x14, 0x00,\n",
+ " 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x0c, 0x00, 0x00, 0x00,\n",
+ " 0x2c, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x7e, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x87, 0xff, 0xdb, 0x39,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0xd8, 0xb2, 0x5d, 0x3d, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x37, 0xdc, 0x56, 0xbd, 0x80, 0x3e, 0x00, 0x00, 0x67, 0x6d, 0x74, 0x77,\n",
+ " 0x35, 0x66, 0x87, 0x95, 0x8e, 0x82, 0x5e, 0x70, 0x6e, 0xa7, 0x60, 0x64,\n",
+ " 0x86, 0x5e, 0x93, 0x7a, 0x76, 0x74, 0x71, 0x8c, 0x61, 0x71, 0x60, 0x8b,\n",
+ " 0x83, 0x48, 0x8b, 0x5f, 0x95, 0x99, 0x5b, 0x59, 0x49, 0x44, 0x79, 0x62,\n",
+ " 0x8e, 0x77, 0x71, 0x89, 0x64, 0x46, 0x8f, 0x8e, 0x80, 0x73, 0x71, 0x81,\n",
+ " 0x85, 0x4a, 0x73, 0x57, 0x66, 0x58, 0x75, 0x93, 0x99, 0x58, 0x8a, 0x7b,\n",
+ " 0x87, 0x81, 0xa1, 0x46, 0x79, 0x6c, 0x83, 0x7a, 0x92, 0x74, 0x6f, 0x6b,\n",
+ " 0x79, 0x77, 0x97, 0x8a, 0x95, 0x75, 0xa2, 0x49, 0x80, 0x4e, 0x7f, 0x6d,\n",
+ " 0xaa, 0xac, 0x6c, 0x5d, 0x57, 0x82, 0x97, 0x77, 0x6f, 0x75, 0x95, 0x73,\n",
+ " 0x7e, 0x51, 0x9f, 0x5b, 0x54, 0x92, 0x60, 0x72, 0x80, 0x6a, 0x92, 0x83,\n",
+ " 0x9b, 0x85, 0x7b, 0x4d, 0x55, 0x4d, 0xb2, 0x7d, 0x65, 0x95, 0x76, 0x42,\n",
+ " 0x61, 0x49, 0xa2, 0x73, 0x9f, 0x7d, 0x7c, 0x54, 0x51, 0x76, 0xa1, 0x7f,\n",
+ " 0x86, 0x69, 0x98, 0x59, 0x6d, 0x84, 0x9f, 0x7b, 0x86, 0x79, 0x88, 0x55,\n",
+ " 0x9c, 0x72, 0x95, 0x8a, 0x91, 0x7a, 0x77, 0x95, 0x7b, 0x87, 0x87, 0x85,\n",
+ " 0x95, 0x72, 0x77, 0x59, 0x7c, 0x80, 0x90, 0x8f, 0x8a, 0x62, 0x76, 0x9f,\n",
+ " 0x64, 0x84, 0x71, 0x7e, 0x7c, 0x66, 0x8e, 0x94, 0x6e, 0xaa, 0x77, 0x5c,\n",
+ " 0x6b, 0x63, 0x68, 0x82, 0x89, 0x46, 0x61, 0x74, 0x8e, 0x85, 0x6b, 0x57,\n",
+ " 0x74, 0x50, 0x87, 0x66, 0x87, 0x98, 0x59, 0x7d, 0xa2, 0x59, 0x75, 0x64,\n",
+ " 0x72, 0x8c, 0x6a, 0x92, 0x8c, 0x56, 0x88, 0x7a, 0x6e, 0x77, 0x9c, 0x82,\n",
+ " 0x7e, 0x5a, 0x91, 0x80, 0x9c, 0x9e, 0x60, 0x8b, 0x6d, 0x76, 0x8d, 0x68,\n",
+ " 0x6c, 0x70, 0x6f, 0x8b, 0x61, 0x6e, 0x86, 0x78, 0x81, 0x81, 0x77, 0x79,\n",
+ " 0x76, 0x69, 0x7d, 0x7b, 0x96, 0x8b, 0x95, 0x91, 0xa2, 0x7b, 0x86, 0x8d,\n",
+ " 0x8b, 0x89, 0x86, 0x5a, 0x5c, 0x4d, 0x96, 0x80, 0x81, 0x55, 0x80, 0x80,\n",
+ " 0x7a, 0x76, 0x99, 0x98, 0x61, 0x95, 0x5a, 0x78, 0x5a, 0x6c, 0x89, 0x81,\n",
+ " 0x98, 0x77, 0x62, 0x77, 0x93, 0x4d, 0x9f, 0x77, 0x72, 0x87, 0x95, 0x71,\n",
+ " 0x65, 0x72, 0xac, 0x8c, 0xa2, 0x89, 0x90, 0x7b, 0x67, 0x60, 0x8a, 0xb3,\n",
+ " 0x72, 0x8f, 0x5c, 0x82, 0x74, 0x76, 0x7c, 0x85, 0x78, 0x6b, 0x97, 0x6d,\n",
+ " 0x86, 0x82, 0x76, 0x84, 0x89, 0x89, 0x7f, 0x6a, 0x7a, 0x7f, 0x6c, 0x77,\n",
+ " 0x80, 0x35, 0x7d, 0x66, 0x96, 0x7e, 0x88, 0x55, 0x6b, 0x55, 0x7c, 0xa7,\n",
+ " 0x7f, 0x9f, 0x64, 0x8b, 0xa0, 0x81, 0x80, 0x97, 0xaf, 0x7a, 0x7d, 0x61,\n",
+ " 0x7a, 0x77, 0x6f, 0x8c, 0x5e, 0x69, 0x6b, 0x94, 0x70, 0x6a, 0x66, 0x5d,\n",
+ " 0x78, 0x6e, 0x76, 0x64, 0xa0, 0x73, 0x8f, 0xa2, 0x9d, 0x50, 0x8e, 0x52,\n",
+ " 0x51, 0x85, 0x78, 0x83, 0x8f, 0x94, 0x83, 0x7c, 0x9c, 0x64, 0x59, 0x7d,\n",
+ " 0x66, 0x6a, 0x73, 0x80, 0x6a, 0x9b, 0x92, 0x7e, 0x7a, 0x78, 0x7d, 0xa0,\n",
+ " 0x8a, 0x9b, 0x61, 0x9e, 0x6c, 0x64, 0x6c, 0x8e, 0x86, 0x75, 0x8a, 0x95,\n",
+ " 0x8e, 0x89, 0x87, 0x8a, 0x5d, 0x8b, 0x82, 0x7c, 0x60, 0x63, 0x85, 0x85,\n",
+ " 0x63, 0x96, 0xa3, 0x7f, 0x93, 0x78, 0x8c, 0x86, 0x7b, 0x78, 0x8e, 0x71,\n",
+ " 0x72, 0x8b, 0x8a, 0x5e, 0x8d, 0x75, 0x78, 0xa3, 0x84, 0x67, 0xa7, 0x54,\n",
+ " 0x6c, 0x80, 0x8e, 0xa8, 0x83, 0x51, 0x6e, 0x9f, 0x8b, 0x86, 0x75, 0x95,\n",
+ " 0x7f, 0x7a, 0x80, 0x81, 0x8d, 0x9c, 0x83, 0x8a, 0x7b, 0x8a, 0x74, 0x6f,\n",
+ " 0x8d, 0x96, 0x5b, 0x9c, 0x8d, 0x7b, 0x83, 0x79, 0x7f, 0x65, 0x7e, 0x87,\n",
+ " 0x7c, 0x5d, 0x71, 0x97, 0x77, 0x44, 0x9a, 0x7f, 0xaa, 0x56, 0x75, 0x5f,\n",
+ " 0x7c, 0x51, 0x8c, 0x90, 0x84, 0x9a, 0x49, 0x5d, 0x86, 0x52, 0x94, 0x95,\n",
+ " 0x5b, 0x86, 0x66, 0x7d, 0x51, 0x4f, 0x7a, 0x91, 0x6d, 0x6e, 0x72, 0x70,\n",
+ " 0x83, 0x4f, 0x9b, 0x9a, 0x8a, 0x77, 0x6a, 0xa1, 0x71, 0x60, 0x61, 0x98,\n",
+ " 0x67, 0x4e, 0x7a, 0x8a, 0x53, 0x6b, 0x99, 0xa0, 0x91, 0x46, 0x8a, 0x8b,\n",
+ " 0x47, 0x78, 0xa9, 0x7b, 0x71, 0x6c, 0x81, 0x68, 0x53, 0x73, 0xaf, 0x70,\n",
+ " 0x62, 0x6d, 0x69, 0x97, 0x70, 0x83, 0x5f, 0x7f, 0x81, 0x87, 0x65, 0x93,\n",
+ " 0x67, 0x87, 0x70, 0x82, 0x79, 0x9e, 0x80, 0x77, 0x6c, 0x80, 0x92, 0x81,\n",
+ " 0x8d, 0x8c, 0x89, 0x8b, 0x4e, 0x91, 0x77, 0x84, 0x99, 0x8c, 0x71, 0x88,\n",
+ " 0x57, 0x7a, 0x9a, 0x8c, 0x82, 0x9b, 0x97, 0x72, 0x69, 0xac, 0x7c, 0x62,\n",
+ " 0x85, 0x7d, 0x76, 0x7f, 0x59, 0x85, 0x68, 0x63, 0x94, 0x8b, 0x7b, 0x92,\n",
+ " 0x7b, 0x6f, 0x77, 0x98, 0x66, 0x78, 0x74, 0x99, 0x85, 0x8c, 0x94, 0x89,\n",
+ " 0x6c, 0x77, 0x89, 0x80, 0x79, 0x8a, 0xa6, 0x95, 0xa9, 0x86, 0x6f, 0x95,\n",
+ " 0x90, 0x69, 0x98, 0x85, 0xa0, 0x7f, 0x56, 0xab, 0x6f, 0x5a, 0x94, 0x8b,\n",
+ " 0x5a, 0x72, 0x61, 0x83, 0x54, 0x70, 0x8d, 0x8d, 0x9c, 0x5e, 0x36, 0x9b,\n",
+ " 0x84, 0x32, 0x6e, 0x84, 0x79, 0x72, 0x64, 0x95, 0x83, 0x58, 0x67, 0x6c,\n",
+ " 0x9e, 0x8d, 0x6e, 0x9e, 0x4f, 0x78, 0x71, 0x85, 0x75, 0x60, 0x4d, 0x7d,\n",
+ " 0x64, 0x89, 0x8e, 0x89, 0x6e, 0x92, 0x53, 0x7c, 0x86, 0x8f, 0xa9, 0xb0,\n",
+ " 0x8e, 0x5e, 0x76, 0x96, 0x65, 0x7c, 0x8a, 0x89, 0x75, 0x8f, 0x65, 0x94,\n",
+ " 0x6c, 0x6c, 0x8d, 0x6d, 0x66, 0x6a, 0x62, 0x98, 0x53, 0x8f, 0x67, 0x76,\n",
+ " 0x80, 0x89, 0x66, 0x60, 0x55, 0x81, 0x85, 0x61, 0x75, 0x78, 0x80, 0x92,\n",
+ " 0x6f, 0x79, 0x66, 0x64, 0x99, 0xa7, 0x88, 0xa1, 0x86, 0x6b, 0x94, 0x88,\n",
+ " 0x77, 0x83, 0x8f, 0x61, 0x72, 0x7c, 0x6f, 0x8f, 0x61, 0x56, 0x8a, 0x7b,\n",
+ " 0x66, 0x8b, 0x98, 0x9d, 0x82, 0x65, 0x77, 0x98, 0x55, 0x83, 0x7a, 0x8c,\n",
+ " 0x74, 0x79, 0x6e, 0x85, 0x82, 0x9a, 0x7d, 0x8d, 0x76, 0x72, 0x64, 0x81,\n",
+ " 0x9a, 0x8d, 0x9f, 0x7b, 0x7c, 0x7b, 0x7b, 0x84, 0x90, 0x6b, 0xa4, 0x84,\n",
+ " 0x98, 0x6f, 0x81, 0xb8, 0x6f, 0x6c, 0x87, 0x6d, 0x8c, 0x72, 0x53, 0x85,\n",
+ " 0x59, 0x4d, 0x9c, 0x94, 0x7d, 0x6f, 0x4f, 0x82, 0x5d, 0x71, 0x6e, 0x78,\n",
+ " 0x61, 0x61, 0x34, 0x71, 0x6a, 0x5a, 0x73, 0xa3, 0x89, 0x65, 0x4d, 0x80,\n",
+ " 0x5c, 0x51, 0x81, 0x8e, 0x6c, 0x53, 0x4a, 0x95, 0x3b, 0x72, 0xa7, 0x86,\n",
+ " 0x7f, 0x75, 0x61, 0xa3, 0x85, 0x6c, 0x99, 0x88, 0x7c, 0x64, 0x7a, 0x8d,\n",
+ " 0x81, 0x7b, 0x6a, 0x7b, 0x8f, 0x74, 0x6d, 0xae, 0x42, 0x67, 0x88, 0xa1,\n",
+ " 0x90, 0x4d, 0x7c, 0x7b, 0x62, 0x55, 0x9a, 0x80, 0x4d, 0x76, 0x5c, 0x88,\n",
+ " 0x60, 0x86, 0x6f, 0x65, 0x67, 0x77, 0x8a, 0x97, 0x99, 0x7c, 0x89, 0x78,\n",
+ " 0x92, 0xa7, 0x6a, 0x7f, 0x8e, 0x88, 0x9d, 0xa1, 0x7b, 0xb0, 0x69, 0x8c,\n",
+ " 0x7e, 0x51, 0x76, 0x84, 0x7d, 0x91, 0x7a, 0x88, 0x7b, 0x88, 0x92, 0x79,\n",
+ " 0x6d, 0x82, 0x6c, 0x8a, 0x99, 0x62, 0x82, 0x9d, 0x99, 0x97, 0x78, 0x6a,\n",
+ " 0x6e, 0x83, 0x64, 0x7d, 0x8c, 0x78, 0x7c, 0x7a, 0x7d, 0x7b, 0x77, 0x84,\n",
+ " 0x76, 0x57, 0x63, 0x85, 0x97, 0x94, 0x80, 0x92, 0x88, 0x73, 0x91, 0x91,\n",
+ " 0x8f, 0x6d, 0x99, 0x86, 0x91, 0x7f, 0x8b, 0x87, 0x98, 0x62, 0x84, 0x70,\n",
+ " 0x97, 0x7b, 0x2e, 0x9b, 0x6e, 0x2a, 0xa4, 0x9c, 0x79, 0x88, 0x54, 0x81,\n",
+ " 0x4f, 0x41, 0xa0, 0x85, 0xaf, 0x9a, 0x47, 0x5a, 0x7d, 0x62, 0x7a, 0x84,\n",
+ " 0x81, 0x6e, 0x41, 0xb4, 0x60, 0x47, 0x8f, 0x98, 0x6c, 0x3c, 0x3b, 0x73,\n",
+ " 0x59, 0x55, 0x7c, 0xb0, 0x6e, 0x5f, 0x61, 0x97, 0x73, 0x59, 0x9f, 0x92,\n",
+ " 0x89, 0x5c, 0x70, 0x96, 0x5c, 0x7c, 0x7c, 0x64, 0x7e, 0x54, 0x5c, 0x94,\n",
+ " 0x56, 0x73, 0x8d, 0x95, 0x59, 0x83, 0x6c, 0x99, 0x6e, 0x5e, 0x7a, 0x99,\n",
+ " 0x83, 0x93, 0x88, 0x76, 0x5a, 0x5a, 0xa5, 0x95, 0x5d, 0x63, 0x8f, 0x6e,\n",
+ " 0x74, 0x65, 0x85, 0x86, 0x98, 0x83, 0x7b, 0x8a, 0x5c, 0x5e, 0x7f, 0x88,\n",
+ " 0x78, 0x68, 0x8f, 0x9f, 0x94, 0x8d, 0x74, 0x7b, 0x6a, 0x91, 0x7a, 0x9a,\n",
+ " 0x70, 0x67, 0xb2, 0x92, 0x75, 0x4e, 0x74, 0xa3, 0x68, 0x74, 0x91, 0x80,\n",
+ " 0x55, 0x8e, 0x88, 0x73, 0x70, 0x81, 0xa1, 0xb8, 0x96, 0x48, 0x67, 0xb2,\n",
+ " 0x76, 0xa1, 0x98, 0xa9, 0x61, 0x6c, 0x5f, 0x98, 0x84, 0x92, 0xa9, 0x83,\n",
+ " 0x9e, 0x74, 0x7b, 0xa2, 0x6f, 0x72, 0x95, 0xa3, 0xb9, 0x80, 0x81, 0x7b,\n",
+ " 0x65, 0x6b, 0x96, 0x8b, 0xae, 0x79, 0x2b, 0x86, 0x5c, 0x2c, 0x8b, 0xa3,\n",
+ " 0x84, 0x74, 0x53, 0x7c, 0x54, 0x4a, 0x65, 0x89, 0xa6, 0x89, 0x47, 0x77,\n",
+ " 0x50, 0x6d, 0x8b, 0x94, 0x8a, 0x61, 0x32, 0x7c, 0x6f, 0x47, 0x78, 0xa2,\n",
+ " 0x9f, 0x42, 0x42, 0x71, 0x78, 0x76, 0x9e, 0x88, 0x70, 0x70, 0x56, 0x8a,\n",
+ " 0x83, 0x95, 0xa7, 0x9d, 0x9d, 0x88, 0x9a, 0x92, 0x48, 0x63, 0xaf, 0x91,\n",
+ " 0x6c, 0x75, 0x5d, 0x5e, 0x83, 0x86, 0xaa, 0x6f, 0x79, 0x84, 0x67, 0x79,\n",
+ " 0x63, 0x69, 0x8e, 0x81, 0x6a, 0x96, 0x8d, 0x86, 0x7b, 0x9f, 0xaa, 0x8e,\n",
+ " 0x63, 0x89, 0x9a, 0x7a, 0x5e, 0x7c, 0x87, 0x83, 0x81, 0x64, 0x7e, 0x59,\n",
+ " 0x6d, 0x5c, 0xa4, 0x72, 0x78, 0x85, 0x9b, 0x79, 0x85, 0x7d, 0x9c, 0x7d,\n",
+ " 0x9c, 0x5c, 0x66, 0x75, 0x66, 0x72, 0xb4, 0x7c, 0x83, 0x9e, 0x90, 0xae,\n",
+ " 0x69, 0x71, 0xb0, 0x84, 0x86, 0x50, 0x66, 0xab, 0x75, 0x96, 0xa8, 0x6c,\n",
+ " 0x87, 0x7b, 0x7e, 0x7c, 0x60, 0x55, 0x96, 0xb0, 0x6a, 0x79, 0x42, 0x9c,\n",
+ " 0x97, 0xa8, 0xb2, 0x9a, 0xa0, 0x84, 0x68, 0x90, 0x90, 0x98, 0x67, 0x9c,\n",
+ " 0xa3, 0x81, 0x71, 0xaa, 0x93, 0x6a, 0x84, 0x8c, 0x77, 0x79, 0x4d, 0x82,\n",
+ " 0x45, 0x1e, 0x7b, 0x94, 0x86, 0x86, 0x26, 0x82, 0x41, 0x6f, 0x8b, 0x86,\n",
+ " 0xa4, 0x80, 0x38, 0x71, 0x5e, 0x5b, 0x9a, 0x73, 0x86, 0x60, 0x5a, 0x9d,\n",
+ " 0x7b, 0x53, 0x89, 0xa0, 0x99, 0x76, 0x57, 0x81, 0x76, 0x5a, 0x9e, 0x85,\n",
+ " 0x5a, 0x7b, 0x56, 0x74, 0x71, 0x6a, 0x9c, 0x68, 0x7e, 0x76, 0x7d, 0x7f,\n",
+ " 0x52, 0x71, 0x85, 0xa2, 0x96, 0x63, 0x73, 0x7c, 0x7a, 0x97, 0x9f, 0x7c,\n",
+ " 0x77, 0x77, 0x59, 0x6b, 0x62, 0x77, 0xbc, 0x6b, 0x7c, 0x79, 0x75, 0x90,\n",
+ " 0x67, 0x82, 0x92, 0x9c, 0x81, 0x92, 0x84, 0x7a, 0x72, 0x5b, 0x86, 0x82,\n",
+ " 0x87, 0x73, 0x87, 0x7c, 0x57, 0x76, 0xa6, 0x7d, 0x7d, 0x94, 0x6a, 0x67,\n",
+ " 0x76, 0x89, 0x9a, 0x6d, 0x7d, 0xa4, 0x6d, 0x7e, 0x74, 0x7e, 0x8f, 0xad,\n",
+ " 0x99, 0x55, 0x5c, 0x82, 0x75, 0x9e, 0xae, 0x76, 0x6b, 0x93, 0x5d, 0x92,\n",
+ " 0x6e, 0x54, 0x88, 0x8f, 0x6a, 0x72, 0x64, 0x93, 0x6e, 0x63, 0x8c, 0xa7,\n",
+ " 0xa6, 0x7a, 0x57, 0x9f, 0x94, 0x91, 0xbd, 0xa4, 0x92, 0x7a, 0x68, 0x9d,\n",
+ " 0x7d, 0x6b, 0x6b, 0xbc, 0xad, 0x7a, 0x73, 0x92, 0x7b, 0x6d, 0x91, 0x6a,\n",
+ " 0x66, 0x8d, 0x34, 0x9b, 0x75, 0x3b, 0x93, 0x78, 0x88, 0x58, 0x1a, 0x7f,\n",
+ " 0x52, 0x61, 0xa3, 0xb1, 0x9c, 0x60, 0x1d, 0x90, 0x7b, 0x37, 0x9f, 0x84,\n",
+ " 0xa3, 0x6c, 0x2e, 0xac, 0x73, 0x62, 0x92, 0x9a, 0x94, 0x6b, 0x5c, 0x82,\n",
+ " 0x5f, 0x4c, 0x9a, 0x8c, 0x76, 0x69, 0x77, 0x5f, 0x5d, 0x91, 0x80, 0x9a,\n",
+ " 0x60, 0x4c, 0x7b, 0x57, 0x67, 0x6b, 0x92, 0x93, 0x64, 0x91, 0x55, 0x75,\n",
+ " 0x41, 0x82, 0x78, 0x68, 0xa2, 0x55, 0x6a, 0x69, 0x59, 0x70, 0x8a, 0x7b,\n",
+ " 0x70, 0x6e, 0x63, 0x83, 0x7f, 0xa4, 0x80, 0x85, 0x86, 0x93, 0x7e, 0x6f,\n",
+ " 0x7b, 0x94, 0xa4, 0xa7, 0x97, 0x7a, 0x87, 0x64, 0x4a, 0x97, 0x94, 0x6a,\n",
+ " 0x96, 0x73, 0x5e, 0x79, 0x6a, 0x99, 0x86, 0xa0, 0x93, 0xac, 0x79, 0x76,\n",
+ " 0x7f, 0x7b, 0xa7, 0x75, 0x8a, 0x71, 0x53, 0x87, 0x93, 0x7f, 0x9e, 0x7b,\n",
+ " 0x81, 0x70, 0x68, 0x8b, 0x8c, 0x9c, 0xaf, 0xa7, 0x6a, 0x9b, 0x49, 0x6d,\n",
+ " 0x67, 0x80, 0x8b, 0x86, 0x9f, 0x80, 0x74, 0x7a, 0x96, 0x74, 0xc8, 0x9d,\n",
+ " 0xa4, 0x74, 0x71, 0x6c, 0x75, 0x6a, 0x9a, 0x95, 0x97, 0x8c, 0x6e, 0x8a,\n",
+ " 0x85, 0x62, 0x5f, 0x7e, 0x9e, 0x6b, 0x48, 0x93, 0x44, 0x37, 0x83, 0xa2,\n",
+ " 0x97, 0x72, 0x25, 0x79, 0x32, 0x39, 0x68, 0x8f, 0x93, 0x61, 0x2b, 0x96,\n",
+ " 0x94, 0x43, 0x82, 0x6e, 0x8f, 0x6d, 0x53, 0x9b, 0x65, 0x50, 0x70, 0x9d,\n",
+ " 0x7d, 0x53, 0x3b, 0x86, 0x77, 0x6c, 0xa6, 0x90, 0x6b, 0x3e, 0x7b, 0x7a,\n",
+ " 0x50, 0x81, 0xb4, 0x76, 0xa5, 0x74, 0x8b, 0x73, 0x79, 0x69, 0xa8, 0x9a,\n",
+ " 0x82, 0x4a, 0x5e, 0x6c, 0x8d, 0x66, 0xa3, 0x80, 0x8d, 0x74, 0x5b, 0x7c,\n",
+ " 0x77, 0xaa, 0x82, 0x69, 0x5e, 0x7d, 0x7f, 0x63, 0xa3, 0x8c, 0xb3, 0x9a,\n",
+ " 0x81, 0x8f, 0x7b, 0x77, 0x60, 0x89, 0x6a, 0x82, 0x5a, 0x7a, 0x71, 0x61,\n",
+ " 0x93, 0x73, 0x8b, 0xb0, 0xa2, 0x92, 0x7c, 0x84, 0x8b, 0x72, 0x91, 0x8d,\n",
+ " 0x91, 0x80, 0x6c, 0x75, 0x7a, 0xb3, 0x95, 0x5e, 0xa5, 0x5d, 0x54, 0x8b,\n",
+ " 0x63, 0x91, 0xa7, 0x68, 0x96, 0x4c, 0x5a, 0x86, 0x76, 0x82, 0xb6, 0xa0,\n",
+ " 0x68, 0x6b, 0x53, 0x76, 0x60, 0x65, 0x90, 0xaf, 0x82, 0x66, 0x80, 0x7b,\n",
+ " 0x84, 0xa0, 0xb0, 0xb8, 0x81, 0x6e, 0x81, 0x8a, 0x74, 0x6e, 0x97, 0xa8,\n",
+ " 0x89, 0x7b, 0x7b, 0x6e, 0x63, 0x74, 0x5a, 0x7b, 0x7e, 0x84, 0x40, 0x95,\n",
+ " 0x73, 0x3c, 0x7c, 0x72, 0x9b, 0x92, 0x27, 0x87, 0x69, 0x5b, 0x99, 0x8a,\n",
+ " 0xa8, 0x65, 0x36, 0x8f, 0x86, 0x3e, 0xa1, 0x79, 0x9f, 0x4d, 0x41, 0xc5,\n",
+ " 0x8c, 0x6a, 0x7e, 0x7f, 0x68, 0x49, 0x5c, 0x91, 0x50, 0x6a, 0x8c, 0x81,\n",
+ " 0x75, 0x4c, 0x6a, 0x74, 0x8a, 0x87, 0xa0, 0x93, 0x7e, 0x6d, 0x52, 0x79,\n",
+ " 0x86, 0x6a, 0x68, 0x6c, 0x83, 0x67, 0x79, 0x73, 0x6f, 0x72, 0x97, 0x84,\n",
+ " 0x8b, 0x78, 0x64, 0x69, 0x8f, 0x92, 0x86, 0x61, 0x5d, 0x85, 0x70, 0x64,\n",
+ " 0x7d, 0xa3, 0x92, 0xa0, 0x72, 0x71, 0x5d, 0x63, 0x7c, 0x70, 0xaf, 0x6f,\n",
+ " 0x93, 0x6a, 0x7e, 0x7f, 0x64, 0xab, 0x85, 0x73, 0x8f, 0x8a, 0x7e, 0x5f,\n",
+ " 0x7a, 0x6f, 0xaa, 0x71, 0x97, 0x7d, 0x60, 0x7c, 0x48, 0x69, 0xa9, 0xaa,\n",
+ " 0x98, 0x7c, 0x61, 0x85, 0x66, 0x97, 0xa2, 0x73, 0x74, 0x65, 0x52, 0x67,\n",
+ " 0x79, 0x8a, 0x79, 0x71, 0x85, 0x6e, 0x6d, 0x67, 0x5e, 0x7f, 0xb9, 0x93,\n",
+ " 0x96, 0x53, 0x69, 0x6e, 0x7f, 0x8f, 0xab, 0x93, 0xa9, 0x70, 0x6e, 0x71,\n",
+ " 0x7e, 0x87, 0x98, 0x7a, 0xae, 0x90, 0x64, 0x88, 0x8a, 0x4f, 0x6d, 0x9e,\n",
+ " 0xac, 0x7e, 0x31, 0x92, 0x50, 0x26, 0x95, 0xb2, 0x90, 0x99, 0x0c, 0x84,\n",
+ " 0x40, 0x4f, 0x8f, 0x76, 0xa4, 0x46, 0x4c, 0x9d, 0x8b, 0x57, 0x81, 0x79,\n",
+ " 0x7b, 0x47, 0x4d, 0x9c, 0x5f, 0x3b, 0x6f, 0x90, 0x7a, 0x3f, 0x66, 0x9d,\n",
+ " 0x6c, 0x45, 0x8b, 0x71, 0x79, 0x62, 0x72, 0x78, 0x93, 0x95, 0x7e, 0x86,\n",
+ " 0x7a, 0x6b, 0x77, 0x74, 0x6b, 0x86, 0xa4, 0x7e, 0x84, 0x48, 0x78, 0x75,\n",
+ " 0x6e, 0x8b, 0x8e, 0x56, 0x69, 0x7b, 0x59, 0x68, 0x5d, 0x77, 0x69, 0x66,\n",
+ " 0x67, 0x9f, 0x75, 0x7b, 0x76, 0x64, 0xc1, 0x78, 0x7d, 0x74, 0x82, 0x73,\n",
+ " 0x73, 0x90, 0xb8, 0x82, 0x7e, 0x70, 0x7b, 0x7a, 0x64, 0xa1, 0x7e, 0x85,\n",
+ " 0x83, 0x81, 0x60, 0x7b, 0x91, 0x82, 0x6f, 0x95, 0xa0, 0x86, 0x6d, 0x88,\n",
+ " 0x75, 0x8d, 0x94, 0x90, 0x76, 0x6d, 0x6e, 0x79, 0x64, 0x74, 0xa8, 0xb1,\n",
+ " 0x92, 0x6e, 0x61, 0x79, 0x74, 0x91, 0x95, 0x74, 0x65, 0x74, 0x5e, 0x7f,\n",
+ " 0x8b, 0x60, 0x9b, 0x9f, 0x74, 0x77, 0x4c, 0x66, 0x7c, 0x80, 0x97, 0x98,\n",
+ " 0x9d, 0x86, 0x55, 0x8a, 0x8a, 0x79, 0x8c, 0x82, 0xb0, 0x7d, 0x63, 0x8c,\n",
+ " 0x5d, 0x5b, 0x82, 0x58, 0x84, 0x56, 0x51, 0x92, 0x75, 0x24, 0x97, 0x92,\n",
+ " 0x75, 0x6e, 0x19, 0x8e, 0x47, 0x3e, 0x7b, 0x7b, 0x87, 0x6b, 0x3f, 0xa9,\n",
+ " 0x59, 0x40, 0x86, 0x74, 0x69, 0x4a, 0x2d, 0xad, 0x91, 0x62, 0xb2, 0xa9,\n",
+ " 0x74, 0x6c, 0x47, 0x94, 0x51, 0x75, 0xb2, 0x6f, 0x75, 0x4b, 0x60, 0xa2,\n",
+ " 0x8e, 0x6a, 0xa4, 0x79, 0x6f, 0x57, 0x80, 0x8c, 0x6c, 0x8e, 0x9e, 0x74,\n",
+ " 0x70, 0x5f, 0x66, 0x80, 0x80, 0x89, 0xb5, 0x8a, 0x7a, 0x96, 0x87, 0x7a,\n",
+ " 0x7b, 0x85, 0x90, 0x79, 0x59, 0x6d, 0x77, 0x8c, 0x8f, 0x82, 0xb3, 0x9c,\n",
+ " 0x6a, 0x6a, 0x6b, 0x70, 0x77, 0x89, 0x96, 0x86, 0x94, 0x72, 0x7e, 0x72,\n",
+ " 0xa9, 0x93, 0x8d, 0x7a, 0x6d, 0x8f, 0x66, 0x72, 0x9a, 0x91, 0x9e, 0x98,\n",
+ " 0xa0, 0x8b, 0x50, 0x76, 0x5c, 0x74, 0xbc, 0x9a, 0x98, 0x73, 0x80, 0x7d,\n",
+ " 0x73, 0x7c, 0xc0, 0x8b, 0x86, 0x7a, 0x66, 0x86, 0x83, 0x72, 0x8f, 0x96,\n",
+ " 0x98, 0x56, 0x45, 0x7b, 0x77, 0x92, 0xac, 0x8a, 0xae, 0x43, 0x33, 0x73,\n",
+ " 0x78, 0x83, 0x98, 0x84, 0x86, 0x78, 0x54, 0x7e, 0x70, 0x5f, 0xa6, 0xa1,\n",
+ " 0x94, 0x81, 0x73, 0x8d, 0x83, 0x5b, 0x88, 0x71, 0xb2, 0x91, 0x50, 0x99,\n",
+ " 0x6b, 0x47, 0x72, 0x92, 0x87, 0x6d, 0x07, 0x99, 0x57, 0x3d, 0x8d, 0x83,\n",
+ " 0x9d, 0x49, 0x40, 0x9d, 0x5c, 0x57, 0x95, 0x73, 0x6e, 0x4b, 0x49, 0xab,\n",
+ " 0x97, 0x58, 0x8b, 0x7a, 0x7a, 0x48, 0x47, 0x8b, 0x7e, 0x5d, 0xa9, 0x6d,\n",
+ " 0x8a, 0x3f, 0x60, 0x82, 0x86, 0x98, 0xa9, 0x7c, 0x74, 0x59, 0x9b, 0x80,\n",
+ " 0x4e, 0x75, 0x9c, 0x5e, 0x75, 0x8c, 0x67, 0x7e, 0x78, 0x75, 0x87, 0x6c,\n",
+ " 0x79, 0x73, 0x63, 0x77, 0x6e, 0x7a, 0x8d, 0x73, 0x4e, 0x72, 0x4a, 0x7c,\n",
+ " 0x8f, 0x79, 0x70, 0x7a, 0x70, 0x73, 0x7b, 0x7a, 0x62, 0xa1, 0x7b, 0x63,\n",
+ " 0x9a, 0x89, 0x76, 0x64, 0x84, 0x7d, 0x9c, 0x94, 0xb0, 0x7f, 0x6c, 0x7b,\n",
+ " 0x8d, 0x89, 0x89, 0x7b, 0x9d, 0x99, 0x64, 0x8b, 0x5c, 0x88, 0xa6, 0x8e,\n",
+ " 0x81, 0x86, 0x7e, 0x85, 0x73, 0x72, 0xad, 0x5d, 0x5f, 0x7e, 0x63, 0x74,\n",
+ " 0x64, 0xa1, 0x9c, 0x83, 0x7c, 0x83, 0x7b, 0x7b, 0x71, 0xa0, 0x9e, 0xaf,\n",
+ " 0x89, 0x79, 0x4c, 0x7c, 0x8c, 0x78, 0x91, 0x87, 0x8a, 0x87, 0x5e, 0x85,\n",
+ " 0x7b, 0x61, 0x9c, 0x88, 0xa5, 0x8d, 0x7c, 0x9c, 0x6b, 0x47, 0x95, 0x85,\n",
+ " 0x81, 0x80, 0x59, 0xb2, 0x4f, 0x3d, 0xae, 0x8c, 0x8d, 0x71, 0x11, 0x95,\n",
+ " 0x31, 0x65, 0x9d, 0xa0, 0x8e, 0x64, 0x42, 0xb9, 0x6a, 0x5c, 0x91, 0x82,\n",
+ " 0x91, 0x50, 0x33, 0xb2, 0x7a, 0x54, 0xac, 0x88, 0x92, 0x61, 0x4e, 0xad,\n",
+ " 0x65, 0x5c, 0x91, 0xb0, 0x72, 0x65, 0x4a, 0x79, 0x68, 0x77, 0x75, 0x5f,\n",
+ " 0x79, 0x6d, 0x6f, 0x7c, 0x4d, 0x71, 0xb8, 0x78, 0x8a, 0x87, 0x6e, 0x72,\n",
+ " 0x7d, 0x79, 0x87, 0x80, 0x5a, 0x78, 0x77, 0x78, 0x80, 0x8f, 0x8c, 0x56,\n",
+ " 0x7a, 0x8b, 0x62, 0x82, 0x5a, 0x96, 0x82, 0x68, 0x71, 0x5d, 0x75, 0x65,\n",
+ " 0x93, 0xb5, 0x71, 0x82, 0x82, 0x8a, 0x4b, 0x7c, 0x62, 0x6f, 0xc1, 0x86,\n",
+ " 0x9d, 0x90, 0x63, 0x71, 0x86, 0x9e, 0x9f, 0x77, 0x90, 0x97, 0x68, 0x81,\n",
+ " 0x5a, 0x8c, 0xab, 0x5e, 0x81, 0x76, 0x83, 0x79, 0x8f, 0xa1, 0x89, 0x79,\n",
+ " 0x81, 0x8a, 0x7e, 0x6c, 0x65, 0x79, 0xc7, 0x89, 0x92, 0x68, 0x78, 0x70,\n",
+ " 0x65, 0x96, 0x9e, 0x82, 0x7d, 0x5f, 0x7b, 0x77, 0x72, 0x84, 0x7e, 0x92,\n",
+ " 0x97, 0x7b, 0x6e, 0x67, 0x81, 0xa1, 0x9a, 0xab, 0x8d, 0x78, 0x61, 0x78,\n",
+ " 0x52, 0x66, 0xaa, 0x77, 0x75, 0xa3, 0x5e, 0xa0, 0x51, 0x40, 0x68, 0xb0,\n",
+ " 0x9a, 0x93, 0x11, 0x82, 0x69, 0x48, 0x9c, 0x77, 0x8d, 0x62, 0x36, 0xac,\n",
+ " 0x6c, 0x4c, 0xa3, 0xab, 0x8f, 0x32, 0x4f, 0xa9, 0x80, 0x68, 0xab, 0x7a,\n",
+ " 0x90, 0x61, 0x5c, 0xa5, 0x84, 0x4c, 0x8c, 0x7a, 0x95, 0x54, 0x72, 0xa0,\n",
+ " 0x66, 0x85, 0xb3, 0x91, 0x69, 0x64, 0x68, 0x56, 0x66, 0x8d, 0xa0, 0x9f,\n",
+ " 0x7a, 0x88, 0x5d, 0x7d, 0x48, 0x80, 0x7f, 0x7c, 0x7c, 0x99, 0x65, 0x81,\n",
+ " 0x73, 0x8b, 0x8c, 0x61, 0x44, 0x60, 0x53, 0x8e, 0x64, 0x80, 0x9c, 0x74,\n",
+ " 0x5d, 0x70, 0x8f, 0x5a, 0x68, 0x7a, 0x82, 0xa1, 0x75, 0x7b, 0x83, 0x60,\n",
+ " 0x75, 0x5e, 0xa2, 0x94, 0x6a, 0x88, 0x78, 0x71, 0x95, 0x70, 0x8b, 0x86,\n",
+ " 0x7e, 0x94, 0x5f, 0x65, 0x5f, 0xb1, 0x97, 0x99, 0x94, 0x84, 0x88, 0x7d,\n",
+ " 0x50, 0x8c, 0xaa, 0x81, 0x7b, 0x7c, 0x77, 0x65, 0x5e, 0x91, 0x9c, 0x89,\n",
+ " 0x8c, 0x85, 0x75, 0x62, 0x7b, 0x78, 0xc3, 0x7a, 0x62, 0x8c, 0x66, 0x6f,\n",
+ " 0x79, 0x7a, 0x9c, 0x6d, 0x7c, 0x6b, 0x5c, 0x7d, 0x6d, 0x54, 0x93, 0x87,\n",
+ " 0x7a, 0x7a, 0x50, 0x85, 0x60, 0x56, 0x5e, 0x6b, 0x90, 0x7c, 0x52, 0xa5,\n",
+ " 0x54, 0x42, 0x7b, 0x75, 0x83, 0x8c, 0x2c, 0xa6, 0x6f, 0x62, 0x78, 0x78,\n",
+ " 0x86, 0x36, 0x4b, 0xaa, 0x86, 0x54, 0x92, 0x8d, 0x7f, 0x53, 0x37, 0xbe,\n",
+ " 0x86, 0x7a, 0x90, 0x7e, 0x8e, 0x50, 0x58, 0xa6, 0x82, 0x58, 0x73, 0x74,\n",
+ " 0x66, 0x5c, 0x6a, 0x7f, 0xa2, 0x69, 0xbd, 0xa9, 0x74, 0x76, 0x75, 0x6f,\n",
+ " 0x45, 0x6c, 0xa5, 0x79, 0x82, 0x67, 0x56, 0x7c, 0x7f, 0x81, 0x67, 0x6d,\n",
+ " 0x81, 0x87, 0x71, 0x69, 0x69, 0x81, 0x85, 0x84, 0x5a, 0x8c, 0x5f, 0x73,\n",
+ " 0x80, 0x9c, 0x9e, 0x90, 0x77, 0xa0, 0x9c, 0x6c, 0x73, 0x8a, 0x84, 0x72,\n",
+ " 0x87, 0xa1, 0x67, 0x64, 0x5d, 0x9b, 0x9d, 0x9b, 0x97, 0x83, 0x5f, 0x61,\n",
+ " 0x77, 0x91, 0xa0, 0x8f, 0x8a, 0x6c, 0x45, 0x5f, 0x6d, 0xa6, 0x9b, 0x76,\n",
+ " 0x86, 0x93, 0x91, 0x7d, 0x54, 0x61, 0xa4, 0x6a, 0x5b, 0x69, 0x5f, 0x6d,\n",
+ " 0x83, 0xaf, 0xa0, 0x78, 0x9d, 0x62, 0x65, 0x69, 0x5f, 0x78, 0xbf, 0x91,\n",
+ " 0x7b, 0x7b, 0x52, 0x5d, 0x70, 0x78, 0xa9, 0x87, 0x93, 0x74, 0x61, 0x74,\n",
+ " 0x8c, 0x61, 0x97, 0x86, 0x9b, 0x7c, 0x7d, 0x75, 0x4b, 0x64, 0xa7, 0x81,\n",
+ " 0x8a, 0x9c, 0x29, 0xa2, 0x5f, 0x38, 0x6a, 0xb0, 0x82, 0x53, 0x1a, 0xa7,\n",
+ " 0x38, 0x47, 0x97, 0x90, 0x8d, 0x41, 0x25, 0xa7, 0x65, 0x63, 0x8b, 0x79,\n",
+ " 0x8f, 0x3e, 0x21, 0xd0, 0x5e, 0x5d, 0x9d, 0x68, 0x75, 0x3e, 0x68, 0xb6,\n",
+ " 0x6a, 0x50, 0x9a, 0x71, 0x81, 0x45, 0x6d, 0x9a, 0x7f, 0x86, 0x9c, 0x63,\n",
+ " 0x7d, 0x74, 0x69, 0x7d, 0x5a, 0x6a, 0x8d, 0x72, 0x6b, 0x69, 0x4c, 0x6f,\n",
+ " 0x7c, 0x8e, 0xa6, 0x83, 0x70, 0x65, 0x5f, 0x78, 0x69, 0x67, 0x7f, 0x8d,\n",
+ " 0x58, 0x76, 0x4a, 0x85, 0x80, 0x89, 0x9f, 0x91, 0x52, 0x62, 0x72, 0x60,\n",
+ " 0x7b, 0x5c, 0x77, 0x6f, 0x9d, 0xa4, 0x98, 0x70, 0x6f, 0xad, 0x94, 0x9f,\n",
+ " 0x7b, 0x89, 0x74, 0x7e, 0x5d, 0x8d, 0xab, 0x98, 0x8f, 0x90, 0x82, 0x84,\n",
+ " 0x60, 0x7c, 0xb7, 0x8e, 0x79, 0x83, 0x56, 0x86, 0x87, 0x79, 0x95, 0x75,\n",
+ " 0x78, 0x71, 0x58, 0x73, 0x87, 0x5d, 0xc6, 0x9f, 0x75, 0x61, 0x4f, 0x71,\n",
+ " 0x91, 0x88, 0xb3, 0x8c, 0x7d, 0x7c, 0x6a, 0x75, 0x6d, 0x66, 0x8e, 0x94,\n",
+ " 0x96, 0x74, 0x59, 0x6f, 0x6d, 0x65, 0xb0, 0x8e, 0x7b, 0x89, 0x7a, 0x6a,\n",
+ " 0x7d, 0x57, 0x82, 0x7a, 0x61, 0x9f, 0x50, 0xab, 0x57, 0x46, 0x86, 0x8d,\n",
+ " 0xa3, 0x96, 0x18, 0xab, 0x51, 0x6e, 0xb3, 0x7e, 0x90, 0x6d, 0x6d, 0xc0,\n",
+ " 0x54, 0x35, 0x96, 0x84, 0x8e, 0x49, 0x28, 0xe4, 0x81, 0x5f, 0x9b, 0x87,\n",
+ " 0x8c, 0x33, 0x56, 0xb4, 0x61, 0x5e, 0x8b, 0x81, 0x99, 0x61, 0x6b, 0x96,\n",
+ " 0x75, 0x82, 0x9e, 0x7c, 0x90, 0x63, 0x64, 0x6b, 0x55, 0x6e, 0xb6, 0x7f,\n",
+ " 0x5f, 0x55, 0x65, 0x60, 0x35, 0x8a, 0x85, 0x91, 0x4d, 0x62, 0x90, 0x90,\n",
+ " 0x57, 0x5a, 0x9f, 0x7b, 0x4c, 0x86, 0x73, 0x83, 0x4a, 0x6d, 0xb0, 0x67,\n",
+ " 0x65, 0x89, 0x54, 0x68, 0x89, 0x7b, 0x72, 0x4f, 0x7a, 0x93, 0x61, 0x7e,\n",
+ " 0x79, 0x89, 0x8f, 0x9c, 0x7b, 0x70, 0x48, 0x67, 0x82, 0x75, 0xaa, 0x92,\n",
+ " 0x9a, 0x8f, 0x79, 0x8c, 0x64, 0x94, 0x98, 0x83, 0x7c, 0x8f, 0x5c, 0x77,\n",
+ " 0x70, 0x90, 0x91, 0x88, 0x7d, 0x51, 0x5d, 0x5d, 0x8b, 0x9f, 0xbc, 0x78,\n",
+ " 0x9e, 0x73, 0x67, 0x6d, 0x82, 0x8d, 0xc9, 0x86, 0x96, 0x6a, 0x5d, 0x79,\n",
+ " 0x7e, 0x6b, 0xb2, 0x79, 0x88, 0x85, 0x65, 0x73, 0x75, 0x6b, 0x9e, 0x7f,\n",
+ " 0x8e, 0x94, 0x8e, 0x7d, 0x74, 0x61, 0x97, 0x56, 0x97, 0x6b, 0x30, 0xb6,\n",
+ " 0x5f, 0x5a, 0xaa, 0xa5, 0x85, 0x5d, 0x01, 0xbc, 0x79, 0x63, 0x6e, 0x82,\n",
+ " 0x72, 0x26, 0x4f, 0xc8, 0x98, 0x56, 0x85, 0x9a, 0x81, 0x1f, 0x48, 0xcf,\n",
+ " 0x84, 0x74, 0x75, 0x87, 0xae, 0x43, 0x6f, 0xdf, 0x6a, 0x4e, 0x97, 0x5d,\n",
+ " 0x8f, 0x37, 0x55, 0x89, 0x7d, 0x82, 0xb1, 0x89, 0x6d, 0x52, 0x65, 0x8b,\n",
+ " 0x71, 0x87, 0x8d, 0x6a, 0x99, 0x5d, 0x65, 0x78, 0x67, 0x8d, 0x7b, 0x51,\n",
+ " 0x60, 0x8a, 0x59, 0x72, 0x78, 0x93, 0x88, 0x75, 0x46, 0x60, 0x6e, 0x79,\n",
+ " 0x7b, 0x9d, 0x9c, 0x8c, 0x5c, 0x7c, 0x69, 0x71, 0x60, 0x6f, 0xb0, 0x7d,\n",
+ " 0x4c, 0x5e, 0x88, 0x77, 0x74, 0x6a, 0x6f, 0x9a, 0xa2, 0x83, 0x48, 0x5a,\n",
+ " 0x6e, 0xa2, 0x8b, 0x7a, 0x65, 0x5b, 0x4b, 0x80, 0x5b, 0x8f, 0xaf, 0x8e,\n",
+ " 0x93, 0x4a, 0x59, 0x6e, 0x5e, 0x89, 0x91, 0x87, 0x73, 0x6a, 0x47, 0x6c,\n",
+ " 0x6c, 0x81, 0xad, 0x5a, 0x76, 0x51, 0x51, 0x6c, 0x80, 0x92, 0x9d, 0xae,\n",
+ " 0x90, 0x71, 0x6c, 0x7a, 0x7c, 0x84, 0xa7, 0x7d, 0x82, 0x7c, 0x80, 0x59,\n",
+ " 0x7d, 0x86, 0xa9, 0x94, 0x8e, 0x7b, 0x7c, 0x67, 0x67, 0x66, 0x8f, 0x49,\n",
+ " 0x5d, 0xa4, 0x4a, 0xbc, 0x5a, 0x34, 0xa7, 0xaa, 0x9e, 0x86, 0x17, 0xc0,\n",
+ " 0x53, 0x67, 0x76, 0xae, 0x8d, 0x37, 0x4a, 0xd6, 0x76, 0x69, 0x95, 0x7a,\n",
+ " 0x8a, 0x0e, 0x3f, 0xe8, 0x60, 0x4d, 0x9e, 0x90, 0xad, 0x44, 0x46, 0xc5,\n",
+ " 0x4c, 0x6e, 0x72, 0x8c, 0x89, 0x49, 0x51, 0xa0, 0x60, 0x84, 0x84, 0x9d,\n",
+ " 0xa4, 0x5a, 0x84, 0x8d, 0x69, 0x6a, 0x97, 0x78, 0x72, 0x66, 0x72, 0x9b,\n",
+ " 0x74, 0x7a, 0x95, 0x7c, 0x7a, 0x6e, 0x74, 0x7f, 0x65, 0x94, 0x77, 0x7e,\n",
+ " 0x85, 0x6d, 0x65, 0x7b, 0x63, 0x7b, 0x87, 0x49, 0x80, 0x74, 0x74, 0x85,\n",
+ " 0x6e, 0x78, 0xad, 0x66, 0x8a, 0x65, 0x54, 0x7c, 0x4e, 0x62, 0x97, 0x7f,\n",
+ " 0x82, 0x6c, 0x58, 0x79, 0x91, 0x94, 0xb3, 0x7a, 0x88, 0x82, 0x60, 0x7f,\n",
+ " 0x8c, 0xa7, 0x7b, 0x93, 0x77, 0x49, 0x6f, 0x6f, 0x5a, 0x8d, 0x93, 0x8b,\n",
+ " 0x87, 0x59, 0x7d, 0x5e, 0x83, 0x7e, 0x8c, 0x7a, 0x91, 0x4e, 0x6f, 0x89,\n",
+ " 0x8a, 0x87, 0x8b, 0x85, 0x8e, 0x43, 0x63, 0x8d, 0x90, 0x6c, 0xa5, 0x73,\n",
+ " 0x8a, 0x78, 0x5f, 0x73, 0x88, 0x57, 0x9e, 0x8f, 0x7f, 0x91, 0x70, 0x77,\n",
+ " 0x8a, 0x76, 0xa2, 0x77, 0x53, 0x86, 0x51, 0xd8, 0xa9, 0x5b, 0x9b, 0x96,\n",
+ " 0x7c, 0x71, 0x01, 0xd4, 0x56, 0x4a, 0x95, 0xab, 0x91, 0x54, 0x45, 0xe5,\n",
+ " 0x74, 0x4f, 0x87, 0x6a, 0xa2, 0x3e, 0x47, 0xff, 0x91, 0x4d, 0x94, 0x97,\n",
+ " 0x6d, 0x74, 0x77, 0xe0, 0x5d, 0x4e, 0x5f, 0x73, 0x70, 0x3a, 0x68, 0xb2,\n",
+ " 0x78, 0x61, 0x8c, 0x77, 0xa8, 0x57, 0x8c, 0x99, 0x23, 0x5a, 0x84, 0x78,\n",
+ " 0x9b, 0x7f, 0x5e, 0xa0, 0x49, 0x84, 0x83, 0x94, 0x99, 0x4d, 0x8d, 0x9a,\n",
+ " 0x86, 0x90, 0x9b, 0x51, 0x75, 0x73, 0x78, 0x89, 0x59, 0x64, 0x78, 0x91,\n",
+ " 0x72, 0x9c, 0x72, 0x7e, 0x65, 0x6a, 0x80, 0xaa, 0x94, 0x65, 0x6d, 0x87,\n",
+ " 0x73, 0x93, 0x97, 0x7d, 0x99, 0x63, 0x75, 0x89, 0x67, 0xa1, 0x90, 0x7f,\n",
+ " 0x88, 0x65, 0x6d, 0x8f, 0x7d, 0x62, 0x91, 0xa7, 0x8b, 0x73, 0x51, 0x88,\n",
+ " 0x66, 0x66, 0x99, 0xa7, 0x7c, 0x54, 0x82, 0x67, 0x64, 0x8a, 0x95, 0x7c,\n",
+ " 0x8a, 0x5d, 0x5e, 0x68, 0x4b, 0x75, 0x92, 0x7a, 0x9f, 0x66, 0x71, 0x8d,\n",
+ " 0x76, 0x72, 0x8e, 0x77, 0x76, 0x8c, 0x5b, 0x88, 0x9a, 0x92, 0x7c, 0x74,\n",
+ " 0x95, 0xaa, 0x71, 0x77, 0x97, 0x93, 0x9e, 0x62, 0x96, 0x6a, 0x49, 0xd8,\n",
+ " 0x81, 0x99, 0xae, 0x87, 0x6c, 0x76, 0x3e, 0xd9, 0x6e, 0x95, 0xa3, 0x86,\n",
+ " 0x60, 0x6c, 0x5c, 0xbe, 0x98, 0x8a, 0x99, 0x7c, 0x47, 0x45, 0x69, 0xeb,\n",
+ " 0x9d, 0x7d, 0xbb, 0x90, 0x66, 0x69, 0x70, 0xc6, 0x7b, 0x59, 0x9e, 0x87,\n",
+ " 0x58, 0x76, 0x7c, 0xae, 0x72, 0x7d, 0x9f, 0x92, 0x82, 0x58, 0x51, 0x7a,\n",
+ " 0x5d, 0x77, 0xa8, 0x7c, 0x56, 0x68, 0x88, 0x8a, 0x7e, 0x8a, 0x98, 0x68,\n",
+ " 0x64, 0x79, 0x6e, 0x7a, 0x60, 0x96, 0x98, 0x60, 0x60, 0x71, 0x60, 0x8e,\n",
+ " 0x7c, 0x8c, 0x92, 0x92, 0x77, 0x80, 0x90, 0x91, 0x81, 0x82, 0x9c, 0x80,\n",
+ " 0x61, 0x7f, 0x5a, 0x8e, 0x88, 0x7c, 0x8e, 0x79, 0x69, 0x8e, 0x4e, 0x7e,\n",
+ " 0x84, 0x9e, 0x67, 0x72, 0x5c, 0x78, 0x7b, 0x8c, 0x65, 0x7d, 0x8e, 0xa4,\n",
+ " 0x5e, 0x7a, 0x5c, 0x97, 0x6a, 0x81, 0xab, 0x85, 0x4d, 0x73, 0x83, 0x96,\n",
+ " 0x8b, 0x7d, 0xa6, 0x69, 0x74, 0x86, 0x73, 0x79, 0x52, 0x8c, 0xa0, 0x86,\n",
+ " 0x64, 0x7b, 0x84, 0x77, 0x87, 0x93, 0x7d, 0x6d, 0x98, 0x6d, 0x88, 0x5f,\n",
+ " 0x7c, 0x84, 0x92, 0x82, 0x81, 0x76, 0x85, 0x77, 0x98, 0x85, 0x88, 0x68,\n",
+ " 0x7d, 0x71, 0x3c, 0xf1, 0x83, 0x86, 0xa2, 0xb3, 0x6e, 0x77, 0x53, 0xe8,\n",
+ " 0xa8, 0xc7, 0xb3, 0x83, 0x93, 0x83, 0x63, 0xe8, 0x94, 0xb3, 0x86, 0x6e,\n",
+ " 0x75, 0x5d, 0x54, 0xf0, 0x89, 0xa7, 0x94, 0xb1, 0x7e, 0x91, 0x9a, 0xb8,\n",
+ " 0x91, 0x7e, 0x99, 0x50, 0x71, 0x82, 0x8a, 0x91, 0x7a, 0x8a, 0x8b, 0x80,\n",
+ " 0x64, 0x6a, 0x5f, 0xbe, 0x5d, 0x96, 0xb1, 0x82, 0x45, 0x71, 0x8b, 0x95,\n",
+ " 0x7c, 0x9b, 0x89, 0x6d, 0x5b, 0x73, 0x81, 0x90, 0x76, 0xab, 0xa6, 0x88,\n",
+ " 0x62, 0x7d, 0x75, 0x99, 0x7a, 0x8b, 0x6e, 0x9b, 0x83, 0x89, 0x99, 0x93,\n",
+ " 0x81, 0x9e, 0x8a, 0x76, 0x75, 0x7d, 0x6c, 0x93, 0x68, 0x7a, 0x8d, 0x78,\n",
+ " 0x88, 0x93, 0x66, 0xa5, 0x6c, 0xae, 0xb1, 0x83, 0x72, 0x8f, 0x6b, 0x7b,\n",
+ " 0x79, 0x9b, 0x98, 0x7c, 0x82, 0x84, 0x7d, 0x7d, 0x71, 0x7c, 0xb0, 0x81,\n",
+ " 0x74, 0x89, 0x72, 0x89, 0x98, 0xa0, 0x7d, 0x62, 0x2f, 0x50, 0x7d, 0x8b,\n",
+ " 0x4c, 0x83, 0x87, 0x89, 0x57, 0x9e, 0x92, 0x8c, 0x81, 0x7e, 0xb9, 0x95,\n",
+ " 0x7f, 0x76, 0x8e, 0x90, 0x9d, 0x68, 0x78, 0x95, 0x7d, 0xab, 0x84, 0x8a,\n",
+ " 0x64, 0x9f, 0x80, 0x94, 0x8d, 0x89, 0x76, 0x8e, 0x6f, 0x8b, 0x75, 0x7d,\n",
+ " 0x89, 0x74, 0x67, 0x8a, 0x7d, 0x63, 0x79, 0x6d, 0x79, 0x8a, 0x78, 0x7f,\n",
+ " 0x7a, 0x9b, 0x70, 0x70, 0x84, 0x86, 0x80, 0x95, 0x5a, 0x77, 0x80, 0x91,\n",
+ " 0x9c, 0x92, 0x76, 0x81, 0x69, 0x89, 0x78, 0xa5, 0x7a, 0x8d, 0x86, 0x64,\n",
+ " 0x8f, 0x8d, 0x7d, 0xa1, 0x8c, 0x7b, 0x77, 0x7e, 0x80, 0x93, 0x86, 0x68,\n",
+ " 0x90, 0x9c, 0x71, 0x8c, 0x68, 0x52, 0x85, 0x88, 0x89, 0x92, 0x64, 0x8f,\n",
+ " 0x74, 0x64, 0x7c, 0x88, 0x8d, 0x97, 0x77, 0x97, 0x91, 0xac, 0x74, 0x7f,\n",
+ " 0x60, 0x7e, 0x6e, 0x70, 0x86, 0x83, 0x7f, 0x81, 0x6f, 0x94, 0x62, 0xa4,\n",
+ " 0x86, 0x7d, 0x90, 0x7c, 0x89, 0x63, 0x7b, 0x89, 0x75, 0xa1, 0x67, 0x69,\n",
+ " 0xa6, 0x76, 0x69, 0x9c, 0x71, 0x79, 0x76, 0x7a, 0x8e, 0x78, 0x94, 0x75,\n",
+ " 0x5a, 0x76, 0x6b, 0x91, 0x84, 0x75, 0x72, 0x93, 0x79, 0x7e, 0x75, 0x9a,\n",
+ " 0x6f, 0x7a, 0x7b, 0x80, 0x5f, 0x90, 0x74, 0x7d, 0x9b, 0x76, 0x70, 0x89,\n",
+ " 0x8f, 0x5f, 0x7f, 0x9c, 0x93, 0x6d, 0x81, 0x7f, 0x8d, 0x7d, 0x74, 0x5d,\n",
+ " 0x75, 0x88, 0x7b, 0x91, 0x75, 0x6b, 0x7f, 0x8c, 0x71, 0x74, 0x87, 0x88,\n",
+ " 0x83, 0x75, 0x77, 0x96, 0x7f, 0x67, 0x7d, 0x95, 0x81, 0x5c, 0x71, 0x5c,\n",
+ " 0x6e, 0x75, 0x86, 0x92, 0x5d, 0x7a, 0x77, 0x9f, 0x6e, 0x79, 0x68, 0x60,\n",
+ " 0x94, 0x88, 0x88, 0x88, 0x79, 0x7e, 0x8a, 0x6d, 0x84, 0xa7, 0x5b, 0x8e,\n",
+ " 0x67, 0x9c, 0x7e, 0x75, 0x82, 0x96, 0x7c, 0x7b, 0x72, 0x85, 0x8c, 0xa3,\n",
+ " 0x96, 0x5b, 0x93, 0x67, 0x7e, 0x9f, 0x71, 0x82, 0x79, 0x8c, 0x93, 0x9d,\n",
+ " 0x6b, 0x90, 0x8a, 0x8a, 0x55, 0x82, 0x94, 0x74, 0x7d, 0xaa, 0x81, 0x78,\n",
+ " 0x8a, 0x8d, 0x83, 0x7b, 0x97, 0x92, 0x68, 0x64, 0x8c, 0x5d, 0x78, 0x9b,\n",
+ " 0x73, 0x95, 0x78, 0x77, 0x6f, 0x61, 0x7c, 0x9d, 0x85, 0x6e, 0x84, 0x4c,\n",
+ " 0x87, 0x57, 0x93, 0x68, 0x8e, 0x77, 0x78, 0x72, 0x87, 0x91, 0x5f, 0x7e,\n",
+ " 0xa6, 0x75, 0x66, 0x86, 0x7a, 0x7d, 0x70, 0x6f, 0x87, 0x8b, 0x74, 0x85,\n",
+ " 0x7d, 0x8b, 0x7f, 0x70, 0x7e, 0x82, 0x84, 0x75, 0x89, 0xa6, 0x7b, 0x7a,\n",
+ " 0xa5, 0x69, 0x73, 0x74, 0x82, 0x65, 0x8f, 0x98, 0x7b, 0x77, 0x84, 0x92,\n",
+ " 0x73, 0x8a, 0xa1, 0x93, 0x80, 0x81, 0x72, 0x8a, 0x6b, 0x75, 0x8f, 0x98,\n",
+ " 0x73, 0x74, 0x6f, 0x70, 0x51, 0x6a, 0x84, 0x9e, 0x78, 0x9b, 0x8c, 0x81,\n",
+ " 0x7e, 0x75, 0x80, 0x88, 0x73, 0x4e, 0x71, 0x74, 0x8c, 0x74, 0x6a, 0x84,\n",
+ " 0x7f, 0x6b, 0x78, 0xab, 0x77, 0xa2, 0x98, 0x93, 0x77, 0x75, 0x72, 0x5c,\n",
+ " 0x60, 0x74, 0x84, 0x67, 0x83, 0x7d, 0x7f, 0x7c, 0x5c, 0x72, 0x70, 0x7f,\n",
+ " 0x6c, 0x84, 0x90, 0xab, 0x97, 0x7f, 0x6b, 0x82, 0x7f, 0x78, 0x73, 0x7d,\n",
+ " 0x8f, 0x8e, 0x8a, 0x8f, 0x8d, 0xa3, 0x74, 0x6e, 0x5e, 0x8c, 0x94, 0x86,\n",
+ " 0x57, 0xb0, 0x79, 0xa8, 0x7b, 0x8d, 0x83, 0x77, 0x89, 0xb6, 0x60, 0x9d,\n",
+ " 0x77, 0x59, 0x72, 0x4d, 0x6f, 0x94, 0x71, 0x75, 0x61, 0x96, 0x86, 0x5d,\n",
+ " 0x84, 0x68, 0x86, 0x82, 0x8d, 0x70, 0x9a, 0x86, 0x73, 0x64, 0x74, 0x7d,\n",
+ " 0x80, 0x5a, 0x64, 0x81, 0xa1, 0x71, 0x77, 0x65, 0xa3, 0x76, 0xa3, 0x9d,\n",
+ " 0x73, 0x7b, 0x8f, 0x7b, 0x79, 0x7d, 0x6c, 0x85, 0x8e, 0x75, 0x65, 0x6a,\n",
+ " 0x87, 0x70, 0x68, 0x8e, 0x76, 0x5d, 0x66, 0x7c, 0x83, 0x83, 0x7e, 0x89,\n",
+ " 0x59, 0x8c, 0x75, 0x59, 0x87, 0x7e, 0x7f, 0x90, 0x6b, 0x7b, 0x7e, 0x6d,\n",
+ " 0x6e, 0x86, 0x69, 0x92, 0x83, 0x8f, 0x8a, 0x60, 0x78, 0x75, 0x61, 0x91,\n",
+ " 0x73, 0x66, 0x86, 0x86, 0x9f, 0x6f, 0x7b, 0x9a, 0x7c, 0x54, 0x75, 0x8e,\n",
+ " 0x7e, 0x72, 0x8e, 0x98, 0x94, 0x5f, 0x71, 0x7c, 0x95, 0x9f, 0x8e, 0x83,\n",
+ " 0x96, 0x4b, 0x8d, 0x84, 0x81, 0x7d, 0x70, 0x84, 0x70, 0x53, 0x8d, 0x84,\n",
+ " 0x5a, 0x91, 0x88, 0x9a, 0x8f, 0x69, 0x8b, 0x52, 0x85, 0x89, 0x6e, 0x99,\n",
+ " 0x79, 0x89, 0x9a, 0x82, 0x6e, 0x8b, 0x65, 0x62, 0x80, 0xa8, 0x8f, 0x8a,\n",
+ " 0x71, 0x61, 0x7e, 0x7d, 0x7e, 0xaa, 0x7f, 0xa0, 0x5e, 0x67, 0x90, 0x86,\n",
+ " 0x6d, 0xac, 0x74, 0x50, 0x61, 0x91, 0x7d, 0x69, 0x8b, 0x7f, 0x81, 0x7a,\n",
+ " 0x93, 0x8c, 0x72, 0x64, 0x98, 0x88, 0x91, 0x83, 0x69, 0x6d, 0x78, 0x7a,\n",
+ " 0x68, 0x7c, 0x76, 0x81, 0xa7, 0x88, 0x8f, 0x79, 0x7d, 0x6c, 0x8a, 0x60,\n",
+ " 0x88, 0x6d, 0x79, 0x9d, 0x80, 0x82, 0x66, 0x7d, 0x7e, 0x96, 0x78, 0x70,\n",
+ " 0x9b, 0x70, 0x7e, 0x90, 0x77, 0x94, 0x7b, 0x89, 0x78, 0x84, 0x74, 0x6d,\n",
+ " 0x7d, 0xa7, 0x75, 0x97, 0x85, 0x83, 0x86, 0x65, 0x75, 0x9a, 0x7c, 0x68,\n",
+ " 0x87, 0x82, 0x75, 0x68, 0x4c, 0x8a, 0x68, 0x93, 0x7d, 0x88, 0x84, 0x72,\n",
+ " 0x58, 0x81, 0x5d, 0x83, 0x89, 0x63, 0x83, 0x7d, 0x8e, 0x75, 0x8c, 0x88,\n",
+ " 0x7f, 0x57, 0x8c, 0x8f, 0xa6, 0x71, 0x8a, 0x95, 0x88, 0x51, 0x74, 0x8a,\n",
+ " 0x8a, 0x98, 0x72, 0x80, 0x8a, 0x52, 0x90, 0x66, 0x54, 0x8e, 0x7f, 0x94,\n",
+ " 0x81, 0x49, 0x84, 0x70, 0x5c, 0x93, 0x89, 0x6d, 0x82, 0x7f, 0x70, 0x5d,\n",
+ " 0x87, 0x8a, 0x71, 0x70, 0x6f, 0xa1, 0x90, 0x9f, 0x74, 0x7c, 0x8c, 0x8b,\n",
+ " 0x72, 0xbf, 0x89, 0x90, 0x5c, 0x8c, 0x75, 0x72, 0x6f, 0xb2, 0x84, 0x6d,\n",
+ " 0x61, 0x80, 0x7d, 0x7a, 0x66, 0xaa, 0x75, 0x71, 0x89, 0x6d, 0x69, 0x72,\n",
+ " 0x73, 0x98, 0x8c, 0x78, 0x5a, 0x8e, 0x8c, 0x81, 0x55, 0x81, 0x96, 0x67,\n",
+ " 0x6f, 0x71, 0x74, 0x7d, 0x8e, 0x66, 0x9a, 0x67, 0xaa, 0x81, 0x90, 0x79,\n",
+ " 0x89, 0x59, 0x86, 0x66, 0x8f, 0x7d, 0x7e, 0xa2, 0xa4, 0x99, 0x68, 0x7a,\n",
+ " 0x8c, 0x73, 0x85, 0x77, 0x8b, 0x74, 0x75, 0x66, 0xaa, 0x98, 0x59, 0x8b,\n",
+ " 0x91, 0x6c, 0x76, 0x73, 0x87, 0xa4, 0x82, 0x82, 0x63, 0x70, 0x7e, 0x73,\n",
+ " 0x96, 0x97, 0x6f, 0x86, 0x81, 0x6f, 0x83, 0x82, 0x7b, 0x82, 0xa3, 0xa7,\n",
+ " 0x95, 0x77, 0x84, 0x65, 0x9b, 0x94, 0x6e, 0xb0, 0x75, 0x66, 0x78, 0x82,\n",
+ " 0x9c, 0x7a, 0x5f, 0xab, 0x99, 0x2f, 0x7f, 0x68, 0xa4, 0x69, 0x8f, 0x9a,\n",
+ " 0x91, 0x56, 0x6e, 0x75, 0x63, 0x9b, 0x9e, 0x97, 0x95, 0x68, 0x80, 0x6a,\n",
+ " 0x40, 0x95, 0x53, 0x72, 0x6f, 0x6b, 0x91, 0x78, 0x7f, 0x93, 0x70, 0x8d,\n",
+ " 0x62, 0x83, 0x7e, 0x64, 0x5b, 0xaa, 0x70, 0x6c, 0x7e, 0x9c, 0x88, 0x76,\n",
+ " 0x60, 0x70, 0x66, 0x69, 0x84, 0x97, 0x9d, 0x63, 0x5e, 0x9a, 0x7e, 0x52,\n",
+ " 0x58, 0xb8, 0x95, 0x7c, 0x4d, 0x96, 0x8f, 0x70, 0x71, 0xbf, 0x83, 0x83,\n",
+ " 0x9e, 0x70, 0x6f, 0x57, 0x70, 0x9a, 0x8d, 0x6e, 0x98, 0x5a, 0x69, 0x6f,\n",
+ " 0x90, 0x71, 0x8a, 0x5d, 0x8e, 0x6e, 0x69, 0x7a, 0x90, 0x86, 0x89, 0x88,\n",
+ " 0xb6, 0x77, 0x84, 0x79, 0x76, 0x86, 0x86, 0x7c, 0xbf, 0x6d, 0x5c, 0x90,\n",
+ " 0xa1, 0x93, 0x72, 0x63, 0x9a, 0x82, 0x7b, 0x61, 0x91, 0x76, 0x82, 0x96,\n",
+ " 0xb9, 0x80, 0x77, 0x7f, 0xa0, 0x73, 0x61, 0x80, 0x83, 0xc1, 0x92, 0x67,\n",
+ " 0x7c, 0x81, 0x90, 0x67, 0x8b, 0xbe, 0x81, 0x91, 0x6c, 0x7e, 0x8d, 0x6c,\n",
+ " 0x62, 0x83, 0x7e, 0x72, 0x64, 0x8a, 0x83, 0x82, 0xaa, 0x8c, 0x74, 0xab,\n",
+ " 0x79, 0x85, 0x91, 0x79, 0x90, 0x68, 0x5c, 0x9a, 0x7c, 0x36, 0x80, 0x6e,\n",
+ " 0x93, 0x76, 0x5e, 0xa0, 0xa5, 0x63, 0x73, 0x7e, 0x8d, 0x94, 0x63, 0x99,\n",
+ " 0x8f, 0x6a, 0x7f, 0x57, 0x57, 0x6f, 0x6d, 0x86, 0x8e, 0x6b, 0x8d, 0x53,\n",
+ " 0x94, 0xba, 0x84, 0x6f, 0x5a, 0x7b, 0x8c, 0x5f, 0x73, 0x93, 0x8b, 0x87,\n",
+ " 0x6f, 0x9e, 0x8a, 0x87, 0x62, 0x97, 0x86, 0x7c, 0x69, 0xab, 0xa1, 0x95,\n",
+ " 0x42, 0x8c, 0x8b, 0x66, 0x68, 0x99, 0xa8, 0x74, 0x80, 0xa5, 0x7d, 0x82,\n",
+ " 0x55, 0xb3, 0x6f, 0x81, 0xa8, 0x9a, 0x80, 0x67, 0x62, 0x7f, 0x78, 0x93,\n",
+ " 0x90, 0x83, 0x83, 0x7b, 0x77, 0x73, 0x8c, 0x56, 0xa7, 0x85, 0x7b, 0x71,\n",
+ " 0x8f, 0x5d, 0x92, 0x69, 0xbe, 0x5e, 0x7f, 0x7f, 0x8e, 0x71, 0x84, 0x75,\n",
+ " 0x95, 0x69, 0x88, 0x6b, 0x96, 0x85, 0x78, 0x39, 0xc2, 0x86, 0x7c, 0x99,\n",
+ " 0xa1, 0x94, 0x6b, 0x86, 0xb5, 0x5e, 0x7e, 0x6e, 0x81, 0x95, 0x6a, 0x88,\n",
+ " 0x7b, 0x92, 0x8f, 0x68, 0x97, 0x77, 0x84, 0x73, 0x68, 0x96, 0x5a, 0x92,\n",
+ " 0x66, 0x74, 0x74, 0x6c, 0x7d, 0x81, 0x6c, 0x93, 0x7f, 0x72, 0x86, 0x74,\n",
+ " 0xbf, 0x8f, 0x53, 0xa4, 0x89, 0x76, 0xa0, 0x87, 0x97, 0x6a, 0x6b, 0xb1,\n",
+ " 0x91, 0x50, 0x74, 0x68, 0xa3, 0x60, 0x8d, 0xbc, 0xc1, 0x3e, 0x62, 0x59,\n",
+ " 0x71, 0x72, 0x6d, 0x80, 0x9f, 0x52, 0x82, 0x6b, 0x5d, 0x7f, 0x74, 0x7e,\n",
+ " 0x74, 0x84, 0x8a, 0x59, 0x5c, 0x85, 0x6d, 0x9c, 0x75, 0x9a, 0x88, 0x89,\n",
+ " 0x81, 0x9f, 0x81, 0x88, 0x6a, 0x94, 0x84, 0x5f, 0x6b, 0x9b, 0x83, 0x4f,\n",
+ " 0x7e, 0xca, 0x99, 0x6d, 0x45, 0x7f, 0x87, 0x71, 0x69, 0xad, 0x95, 0x53,\n",
+ " 0x6e, 0x9b, 0x90, 0x73, 0x5d, 0xb0, 0x8d, 0x67, 0x83, 0x82, 0xa3, 0x70,\n",
+ " 0x70, 0x92, 0x82, 0x9a, 0x8a, 0x69, 0x6a, 0x6e, 0x7f, 0x89, 0xa4, 0x76,\n",
+ " 0x97, 0x62, 0x94, 0x80, 0x87, 0x55, 0x80, 0x76, 0xb3, 0x7e, 0x7e, 0x71,\n",
+ " 0x94, 0x88, 0x8e, 0x74, 0xb6, 0x4d, 0x7b, 0x73, 0x90, 0x86, 0x7c, 0x66,\n",
+ " 0xb5, 0x80, 0x7f, 0x84, 0x87, 0x82, 0x67, 0x83, 0x97, 0x91, 0x8a, 0x78,\n",
+ " 0x8b, 0x83, 0x5d, 0x84, 0x82, 0x9f, 0x8c, 0x91, 0x84, 0x8b, 0x6a, 0x68,\n",
+ " 0x86, 0x82, 0x73, 0x77, 0x7b, 0x83, 0x6a, 0x84, 0x92, 0x93, 0x90, 0x8b,\n",
+ " 0x4c, 0x94, 0x98, 0x76, 0xb8, 0x7b, 0xa0, 0xa2, 0x7d, 0x3e, 0x95, 0x88,\n",
+ " 0xa3, 0x6f, 0x5e, 0xc8, 0x9a, 0x52, 0x81, 0x86, 0xa3, 0x79, 0x88, 0xc3,\n",
+ " 0xbd, 0x54, 0x6c, 0x5e, 0x83, 0x8a, 0x98, 0x88, 0x92, 0x66, 0x73, 0x5b,\n",
+ " 0x6c, 0x7f, 0x6e, 0x97, 0x8d, 0x58, 0x89, 0x6e, 0x65, 0x7a, 0x7d, 0x7c,\n",
+ " 0x7e, 0x89, 0x94, 0x89, 0x55, 0xb8, 0x8f, 0x82, 0x6c, 0x9c, 0x96, 0x5e,\n",
+ " 0x6f, 0xb2, 0x70, 0x76, 0x95, 0xc8, 0x86, 0x78, 0x49, 0xac, 0x7e, 0x6c,\n",
+ " 0x68, 0xb6, 0xaf, 0x89, 0x68, 0xa5, 0x72, 0x85, 0x69, 0x9c, 0x94, 0x84,\n",
+ " 0xa4, 0x97, 0x91, 0x61, 0x7a, 0xa3, 0x8f, 0x8e, 0x93, 0x80, 0x8d, 0x76,\n",
+ " 0x74, 0x84, 0x9b, 0x79, 0x97, 0x4e, 0x67, 0x87, 0x9b, 0x69, 0x85, 0x7d,\n",
+ " 0xb2, 0x68, 0x76, 0x63, 0xa2, 0x86, 0x97, 0x7f, 0xb5, 0x63, 0x79, 0x76,\n",
+ " 0x8a, 0x7c, 0x7c, 0x91, 0xb1, 0x42, 0x7d, 0x7a, 0x8c, 0x8e, 0x72, 0xab,\n",
+ " 0xb8, 0x76, 0xab, 0x81, 0x98, 0x85, 0x56, 0x98, 0x84, 0x9f, 0x70, 0x86,\n",
+ " 0x76, 0x88, 0x70, 0x8d, 0x71, 0x7b, 0x7a, 0x8d, 0x76, 0x75, 0x62, 0x80,\n",
+ " 0x81, 0x94, 0x82, 0x6e, 0x57, 0x8d, 0xaf, 0x84, 0xbf, 0x85, 0x82, 0xa7,\n",
+ " 0x80, 0x89, 0x95, 0x81, 0x91, 0x49, 0x72, 0xa1, 0xa7, 0x3f, 0x72, 0x8b,\n",
+ " 0x99, 0x72, 0x86, 0xb2, 0xc3, 0x61, 0x55, 0x77, 0x86, 0x77, 0x83, 0xa7,\n",
+ " 0x95, 0x5a, 0x68, 0x68, 0x6a, 0x63, 0x6a, 0x77, 0x93, 0x7c, 0x88, 0x62,\n",
+ " 0x79, 0x84, 0x8b, 0x82, 0x58, 0x8f, 0x9c, 0x56, 0x77, 0xb1, 0x65, 0x8c,\n",
+ " 0x76, 0x91, 0x83, 0x5b, 0x62, 0x91, 0x87, 0x68, 0x71, 0xb0, 0x87, 0x64,\n",
+ " 0x62, 0x91, 0x94, 0x58, 0x7f, 0xac, 0xa3, 0x84, 0x75, 0xaa, 0xa3, 0x4d,\n",
+ " 0x7a, 0xc2, 0x84, 0x8a, 0x6d, 0xa2, 0x76, 0x74, 0x8c, 0x9e, 0x7c, 0x71,\n",
+ " 0x86, 0x70, 0x6d, 0x79, 0x9a, 0x74, 0xb0, 0x8d, 0xa5, 0x7e, 0x6b, 0x63,\n",
+ " 0x96, 0x74, 0x99, 0x76, 0xd0, 0x62, 0x85, 0x9d, 0x8f, 0x6d, 0x83, 0x88,\n",
+ " 0xb0, 0x62, 0x9b, 0x87, 0x91, 0x82, 0x7a, 0x90, 0x9c, 0x61, 0x6d, 0x97,\n",
+ " 0x84, 0x7c, 0x74, 0x8e, 0x8b, 0x75, 0x9a, 0x7e, 0x7c, 0x7d, 0x96, 0x81,\n",
+ " 0x94, 0x69, 0x83, 0x6f, 0x8e, 0x7c, 0x7b, 0x7a, 0x73, 0x98, 0x74, 0x9e,\n",
+ " 0x72, 0x8c, 0x5f, 0x7d, 0x99, 0x79, 0x5b, 0x73, 0x65, 0x78, 0xa5, 0x7d,\n",
+ " 0xa2, 0x98, 0x91, 0x91, 0x87, 0x7b, 0x8c, 0x82, 0xb8, 0x6b, 0x82, 0xba,\n",
+ " 0xa5, 0x3f, 0x83, 0x7a, 0x9b, 0x73, 0x93, 0xa1, 0xbe, 0x55, 0x6b, 0x75,\n",
+ " 0x94, 0x7d, 0x9c, 0xa1, 0x82, 0x50, 0x75, 0x5a, 0x88, 0x6e, 0x72, 0x7f,\n",
+ " 0x99, 0x64, 0x72, 0x49, 0x69, 0x79, 0x6d, 0x94, 0x73, 0x79, 0x80, 0x6f,\n",
+ " 0x72, 0xbc, 0x9d, 0x71, 0x7a, 0x9d, 0x8a, 0x55, 0x74, 0xaa, 0xa1, 0x85,\n",
+ " 0x7e, 0xc4, 0xa0, 0x7e, 0x50, 0x99, 0x68, 0x8c, 0x8a, 0xb0, 0x99, 0x6c,\n",
+ " 0x6d, 0xaf, 0x7b, 0x7b, 0x79, 0xba, 0x8a, 0x7a, 0x9d, 0x8b, 0x67, 0x87,\n",
+ " 0x76, 0xa9, 0x7f, 0x7e, 0x8b, 0x7b, 0x87, 0x84, 0x82, 0x74, 0xa3, 0x91,\n",
+ " 0x9a, 0x6a, 0x93, 0x7e, 0x87, 0x5b, 0x95, 0x89, 0xbb, 0x5d, 0x74, 0x6c,\n",
+ " 0x88, 0x7e, 0x81, 0x7e, 0xb6, 0x6b, 0x91, 0x92, 0x83, 0x78, 0x79, 0x95,\n",
+ " 0x90, 0x5e, 0x68, 0x8f, 0xa8, 0x92, 0x66, 0x8e, 0x6b, 0x8c, 0x86, 0x80,\n",
+ " 0x7e, 0x7e, 0x70, 0x84, 0x7d, 0x71, 0x67, 0x94, 0x71, 0x69, 0x84, 0x8f,\n",
+ " 0x6c, 0x72, 0x85, 0x83, 0x69, 0x76, 0x57, 0x62, 0x83, 0x96, 0x83, 0x77,\n",
+ " 0x64, 0x5f, 0xae, 0x7c, 0xa7, 0x88, 0x91, 0x8c, 0x9e, 0x7f, 0xa8, 0x8a,\n",
+ " 0x93, 0x6f, 0x58, 0xae, 0xb4, 0x4b, 0x7f, 0x64, 0x9f, 0x5a, 0x9e, 0xb6,\n",
+ " 0xa6, 0x6b, 0x79, 0x84, 0x6b, 0x7c, 0x8b, 0x94, 0x85, 0x60, 0x6b, 0x55,\n",
+ " 0x79, 0x68, 0x77, 0x75, 0x85, 0x5c, 0x91, 0x5e, 0x5a, 0x71, 0x68, 0x7b,\n",
+ " 0x73, 0x91, 0x6c, 0x6e, 0x71, 0x8b, 0x76, 0x86, 0x99, 0xb8, 0x91, 0x68,\n",
+ " 0x51, 0xa7, 0x6f, 0x7a, 0x8a, 0xc3, 0x8e, 0x65, 0x64, 0x9e, 0x80, 0x78,\n",
+ " 0x6c, 0xc5, 0xa2, 0x75, 0x71, 0xa5, 0x96, 0x4f, 0x70, 0xa4, 0x7a, 0x7c,\n",
+ " 0x8c, 0x80, 0x89, 0x97, 0x9a, 0x9a, 0x85, 0x89, 0x92, 0x8f, 0x81, 0x6f,\n",
+ " 0x82, 0x6a, 0xb8, 0x74, 0x8f, 0x51, 0x7b, 0x8b, 0x8c, 0x55, 0x7e, 0x8c,\n",
+ " 0xb2, 0x41, 0x85, 0x77, 0x9c, 0x73, 0x75, 0x8d, 0x9f, 0x64, 0x92, 0x77,\n",
+ " 0xa0, 0x87, 0x5f, 0x71, 0x85, 0x68, 0x8a, 0x78, 0x91, 0x78, 0x75, 0x7a,\n",
+ " 0x81, 0x67, 0x96, 0x64, 0x96, 0x85, 0x7a, 0x7e, 0x83, 0x74, 0x82, 0x8f,\n",
+ " 0x98, 0x75, 0x77, 0x84, 0x7e, 0x88, 0x94, 0x7d, 0x79, 0x8c, 0x47, 0x79,\n",
+ " 0x96, 0x7f, 0x8e, 0x90, 0x50, 0x7f, 0xa3, 0x77, 0xa8, 0x7f, 0x65, 0x9f,\n",
+ " 0xb9, 0x4c, 0xa7, 0x7f, 0xaa, 0x6e, 0xa2, 0xb0, 0xb8, 0x51, 0x6b, 0x74,\n",
+ " 0xaa, 0x63, 0x6c, 0xa3, 0xb6, 0x5e, 0x74, 0x6a, 0x75, 0x69, 0x87, 0x7f,\n",
+ " 0x9d, 0x71, 0x73, 0x72, 0x70, 0x57, 0x5a, 0x7e, 0x8b, 0x64, 0x9a, 0x4d,\n",
+ " 0x97, 0x81, 0x7b, 0x75, 0x6e, 0x92, 0x5f, 0x67, 0x7e, 0xaa, 0x90, 0x7a,\n",
+ " 0x92, 0xae, 0x92, 0x68, 0x79, 0x9d, 0x4f, 0x6c, 0x79, 0xb4, 0x9c, 0x58,\n",
+ " 0x86, 0x8e, 0x62, 0x72, 0x71, 0xc1, 0xac, 0x7d, 0x7a, 0x94, 0x8f, 0x7b,\n",
+ " 0x88, 0xa8, 0x8d, 0x82, 0x75, 0x9b, 0x5f, 0x83, 0x82, 0xb3, 0x7a, 0x93,\n",
+ " 0x94, 0x76, 0x70, 0x7e, 0x72, 0x7e, 0x8f, 0x8c, 0xa7, 0x53, 0x72, 0x77,\n",
+ " 0x7a, 0x64, 0xa8, 0x83, 0xc5, 0x56, 0x71, 0x7b, 0x96, 0x73, 0x7c, 0x73,\n",
+ " 0x93, 0x49, 0x83, 0x99, 0xa2, 0x83, 0x74, 0x79, 0xa4, 0x61, 0x8e, 0x84,\n",
+ " 0x7a, 0x7d, 0x56, 0x98, 0x97, 0x6d, 0x87, 0x8c, 0x7a, 0x77, 0x6a, 0x67,\n",
+ " 0x8a, 0x6f, 0xa2, 0x82, 0x8d, 0x85, 0x6d, 0x8f, 0x7e, 0x74, 0x72, 0x74,\n",
+ " 0x91, 0x75, 0x58, 0x7f, 0x9e, 0x7c, 0x9c, 0x75, 0x61, 0x6f, 0x85, 0x7b,\n",
+ " 0xbe, 0x84, 0x85, 0x9b, 0x8c, 0x3b, 0x9a, 0x90, 0xab, 0x77, 0x8e, 0xa2,\n",
+ " 0xbd, 0x55, 0x96, 0x70, 0xa8, 0x78, 0x98, 0x9c, 0xc3, 0x67, 0x6e, 0x81,\n",
+ " 0x70, 0x75, 0x96, 0x9c, 0x8a, 0x5b, 0x73, 0x54, 0x69, 0x6c, 0x5d, 0x82,\n",
+ " 0x99, 0x5b, 0x8c, 0x6d, 0x87, 0x80, 0x67, 0x86, 0x88, 0x7c, 0x70, 0x6b,\n",
+ " 0x75, 0xab, 0x8e, 0x79, 0x90, 0x91, 0xaf, 0x67, 0x5c, 0xa1, 0x5c, 0x6f,\n",
+ " 0x75, 0xa1, 0x95, 0x5f, 0x82, 0x8f, 0x78, 0x5d, 0x7c, 0xb8, 0x8a, 0x8a,\n",
+ " 0x6a, 0x98, 0x6e, 0x51, 0x6b, 0xaa, 0x7d, 0x7c, 0x80, 0x94, 0x79, 0x6d,\n",
+ " 0xaa, 0x8a, 0x7e, 0x77, 0xa4, 0x78, 0xa5, 0x6d, 0x7c, 0x75, 0xa8, 0x6f,\n",
+ " 0xa6, 0x51, 0x8e, 0x80, 0x96, 0x5b, 0x9d, 0x7b, 0xb8, 0x4e, 0x6c, 0x87,\n",
+ " 0x95, 0x7c, 0x78, 0x71, 0xb0, 0x5a, 0x99, 0xa0, 0x90, 0x87, 0x65, 0x8b,\n",
+ " 0x98, 0x68, 0x92, 0x76, 0x82, 0x77, 0x6a, 0x8a, 0x91, 0x84, 0x87, 0x8b,\n",
+ " 0x87, 0x84, 0x7a, 0x81, 0x77, 0x55, 0x8e, 0x86, 0x7a, 0x74, 0x65, 0x88,\n",
+ " 0x62, 0x51, 0xa1, 0x91, 0x88, 0x76, 0x5f, 0x89, 0x9f, 0x86, 0x66, 0x67,\n",
+ " 0x64, 0x75, 0x9e, 0x74, 0xc1, 0x80, 0x58, 0xa9, 0x8f, 0x5e, 0x94, 0x88,\n",
+ " 0xaf, 0x6f, 0x6c, 0xa4, 0xa1, 0x4d, 0x68, 0x66, 0xc2, 0x6e, 0x89, 0x9b,\n",
+ " 0xa3, 0x5a, 0x63, 0x5b, 0x9c, 0x7a, 0x93, 0x76, 0x9d, 0x6d, 0x71, 0x5d,\n",
+ " 0x80, 0x66, 0x79, 0x80, 0x7c, 0x65, 0x74, 0x64, 0x88, 0x90, 0x79, 0x89,\n",
+ " 0x72, 0x88, 0x67, 0x75, 0x6a, 0x96, 0x56, 0x67, 0x88, 0xa1, 0x8c, 0x6c,\n",
+ " 0x55, 0xb2, 0x8a, 0x71, 0x88, 0xdc, 0x7a, 0x72, 0x94, 0x9d, 0x7c, 0x76,\n",
+ " 0x6a, 0xaa, 0xa8, 0x7f, 0x80, 0xa0, 0x6b, 0x6f, 0x84, 0xe0, 0x68, 0x93,\n",
+ " 0xa6, 0x99, 0x69, 0x68, 0x93, 0xa0, 0x93, 0x6b, 0x87, 0x8b, 0x80, 0x90,\n",
+ " 0x90, 0x89, 0x8f, 0x7f, 0xaf, 0x6f, 0x82, 0x6d, 0x94, 0x70, 0x97, 0x8f,\n",
+ " 0xb0, 0x40, 0x9b, 0x67, 0x78, 0x86, 0x90, 0x8b, 0xa7, 0x51, 0x7f, 0x79,\n",
+ " 0x90, 0x71, 0x6d, 0x80, 0x95, 0x63, 0x7d, 0x87, 0xa0, 0x7e, 0x7b, 0x85,\n",
+ " 0x8e, 0x6d, 0xa1, 0x76, 0x70, 0x7b, 0x66, 0x87, 0x90, 0x7a, 0x86, 0x88,\n",
+ " 0x89, 0x87, 0x6a, 0x91, 0x78, 0x74, 0x76, 0x8d, 0x7e, 0x86, 0x63, 0x90,\n",
+ " 0x98, 0x7d, 0x4a, 0x85, 0x4f, 0x9d, 0xa2, 0x7c, 0xb4, 0x88, 0x78, 0xb5,\n",
+ " 0x8f, 0x3f, 0xa7, 0x7d, 0xa4, 0x7c, 0x60, 0x9c, 0xa8, 0x41, 0x6b, 0x7f,\n",
+ " 0xa2, 0x7f, 0x68, 0xaa, 0xb4, 0x73, 0x56, 0x62, 0x87, 0x72, 0xa5, 0x7c,\n",
+ " 0x97, 0x69, 0x58, 0x6b, 0x89, 0x57, 0x51, 0x80, 0x92, 0x7a, 0x7c, 0x4c,\n",
+ " 0x7c, 0x7b, 0x69, 0x5f, 0x90, 0x77, 0x78, 0x67, 0x7a, 0xad, 0x79, 0x5c,\n",
+ " 0x9c, 0xbf, 0xa6, 0x64, 0x53, 0xb3, 0x5e, 0x59, 0x86, 0xb9, 0x94, 0x65,\n",
+ " 0x70, 0x9d, 0x7a, 0x80, 0x7c, 0xae, 0x9c, 0x7b, 0x66, 0xae, 0x83, 0x5f,\n",
+ " 0x81, 0xc5, 0x8b, 0x7e, 0x9b, 0x89, 0x84, 0x7f, 0x7c, 0xa5, 0x5c, 0x89,\n",
+ " 0x8a, 0x75, 0x99, 0x6d, 0x8e, 0x90, 0x9f, 0x81, 0x81, 0x6b, 0x87, 0x76,\n",
+ " 0x92, 0x6f, 0xab, 0x95, 0x95, 0x4c, 0x97, 0x72, 0x80, 0x87, 0x83, 0x87,\n",
+ " 0xa3, 0x59, 0xad, 0x74, 0x93, 0x7f, 0x77, 0x78, 0x8d, 0x66, 0x9b, 0x7a,\n",
+ " 0x7d, 0x95, 0x64, 0x7f, 0x6d, 0x5c, 0x8e, 0x94, 0x92, 0x82, 0x60, 0x8d,\n",
+ " 0x75, 0x55, 0x8c, 0x8b, 0x8f, 0x86, 0x7d, 0x7c, 0x74, 0x57, 0x78, 0x9d,\n",
+ " 0x71, 0x65, 0x66, 0x7f, 0xaa, 0x92, 0x66, 0x81, 0x5a, 0x71, 0xa6, 0x78,\n",
+ " 0x9d, 0x8a, 0x5a, 0x8a, 0x91, 0x59, 0xb7, 0x5c, 0xc3, 0x73, 0x89, 0x9d,\n",
+ " 0xa7, 0x62, 0x77, 0x72, 0x9f, 0x92, 0x6a, 0x9f, 0xaa, 0x71, 0x6b, 0x5e,\n",
+ " 0x7d, 0x73, 0x8d, 0x89, 0xba, 0x61, 0x73, 0x6e, 0x71, 0x8a, 0x79, 0x7c,\n",
+ " 0x94, 0x76, 0x76, 0x65, 0x81, 0x6f, 0x4e, 0x75, 0x6e, 0x8b, 0x7d, 0x50,\n",
+ " 0x56, 0xb8, 0x72, 0x67, 0x93, 0xc6, 0x88, 0x6f, 0x57, 0xb7, 0x80, 0x4c,\n",
+ " 0x97, 0xc4, 0xb6, 0x71, 0x72, 0x9e, 0x6f, 0x72, 0x8d, 0xa5, 0x8f, 0x89,\n",
+ " 0x74, 0xae, 0x78, 0x70, 0x6e, 0xbb, 0x8f, 0x73, 0x74, 0x8b, 0x5e, 0x86,\n",
+ " 0x8b, 0x8a, 0x72, 0x71, 0x84, 0x84, 0x77, 0xa3, 0xa6, 0x73, 0xa4, 0x7e,\n",
+ " 0xab, 0x5d, 0x75, 0x96, 0x94, 0x5f, 0x8b, 0x74, 0x9c, 0x63, 0x8d, 0x81,\n",
+ " 0x80, 0x6a, 0x91, 0x88, 0x93, 0x53, 0x80, 0x75, 0x79, 0x8d, 0x78, 0x74,\n",
+ " 0x7c, 0x73, 0xb2, 0x89, 0x8e, 0xab, 0x75, 0x6c, 0x7a, 0x79, 0x99, 0x77,\n",
+ " 0x7d, 0x89, 0x5a, 0x81, 0x7c, 0x75, 0x6a, 0x7e, 0x8c, 0x83, 0x78, 0x8e,\n",
+ " 0x62, 0x76, 0x77, 0x6b, 0x79, 0x66, 0x6e, 0x82, 0xa1, 0x8d, 0x52, 0x79,\n",
+ " 0x70, 0x7d, 0xa9, 0x6a, 0x95, 0x7f, 0x59, 0x94, 0x8f, 0x73, 0xb7, 0x85,\n",
+ " 0xb3, 0x80, 0x77, 0x9f, 0xb8, 0x4d, 0x82, 0x7c, 0xa0, 0xa4, 0x7b, 0x8c,\n",
+ " 0xa9, 0x78, 0x62, 0x6b, 0x8a, 0x93, 0x80, 0x68, 0x9b, 0x6d, 0x6b, 0x7b,\n",
+ " 0x84, 0x8f, 0x86, 0x70, 0x70, 0x73, 0x84, 0x4f, 0x7c, 0x75, 0x64, 0x8d,\n",
+ " 0x6e, 0x81, 0x7c, 0x72, 0x81, 0xb0, 0x74, 0x65, 0xa7, 0xae, 0x80, 0x70,\n",
+ " 0x5e, 0xa4, 0x58, 0x54, 0x8e, 0xa7, 0x96, 0x65, 0x66, 0x8b, 0x6c, 0x5d,\n",
+ " 0x6b, 0xbe, 0x94, 0x79, 0x80, 0xa1, 0x91, 0x78, 0x6d, 0xc2, 0x82, 0x85,\n",
+ " 0x81, 0x7d, 0x88, 0x79, 0x93, 0x96, 0x7f, 0x7e, 0x7d, 0x92, 0x75, 0xa2,\n",
+ " 0x9f, 0x7b, 0x92, 0x77, 0x8a, 0x7c, 0x80, 0x8b, 0x9b, 0x64, 0xa5, 0x74,\n",
+ " 0xa1, 0x74, 0x7f, 0x7e, 0x85, 0x78, 0x9c, 0x86, 0x9f, 0x62, 0x8f, 0x7f,\n",
+ " 0x8a, 0x90, 0x6d, 0x7d, 0x93, 0x61, 0x9d, 0x81, 0x9b, 0x99, 0x69, 0x87,\n",
+ " 0x74, 0x7d, 0x8e, 0x8e, 0x7b, 0x7c, 0x6a, 0x71, 0x7d, 0x7f, 0x74, 0x74,\n",
+ " 0x7b, 0x65, 0x6e, 0x91, 0x7c, 0x6e, 0x80, 0x8c, 0x8a, 0x6c, 0x6b, 0x76,\n",
+ " 0xad, 0x94, 0x64, 0x81, 0x69, 0x7b, 0xac, 0x76, 0x9f, 0x71, 0x85, 0x85,\n",
+ " 0x8b, 0x66, 0xb5, 0x87, 0xb3, 0x63, 0x8b, 0x95, 0x8e, 0x50, 0x91, 0x77,\n",
+ " 0xa1, 0x99, 0x64, 0x81, 0xb3, 0x63, 0x6e, 0x7a, 0x7f, 0x73, 0x7a, 0x7b,\n",
+ " 0x93, 0x6d, 0x75, 0x75, 0x7c, 0x7b, 0x59, 0x7c, 0x7c, 0x68, 0x67, 0x78,\n",
+ " 0x79, 0x75, 0x53, 0x86, 0x84, 0x84, 0x91, 0x71, 0x85, 0xb1, 0x84, 0x64,\n",
+ " 0x88, 0xc0, 0x94, 0x5f, 0x6f, 0x9b, 0x69, 0x67, 0x97, 0x94, 0x88, 0x6a,\n",
+ " 0x7e, 0x94, 0x9e, 0x7f, 0x81, 0x9c, 0xa7, 0x7f, 0x7a, 0xa2, 0x63, 0x69,\n",
+ " 0x82, 0xc2, 0x5e, 0x8d, 0x7c, 0x89, 0x63, 0x93, 0x84, 0xb8, 0x76, 0x89,\n",
+ " 0x96, 0x87, 0x79, 0x88, 0xa6, 0x8e, 0x9b, 0x93, 0x9c, 0x5d, 0x92, 0x92,\n",
+ " 0x82, 0x5e, 0x85, 0x88, 0xad, 0x73, 0xa4, 0x6f, 0x74, 0x8e, 0x77, 0x89,\n",
+ " 0x9b, 0x6e, 0x82, 0x76, 0x93, 0xae, 0x82, 0x87, 0x76, 0x6f, 0x80, 0x76,\n",
+ " 0x95, 0x8e, 0x5e, 0x85, 0x7b, 0x68, 0x7f, 0x7c, 0x82, 0x94, 0x80, 0x91,\n",
+ " 0x77, 0x71, 0x7c, 0x94, 0x80, 0x62, 0x65, 0x7c, 0x5e, 0x70, 0x76, 0x75,\n",
+ " 0x7b, 0x60, 0x5f, 0x69, 0xb3, 0x6e, 0x95, 0x9d, 0x5a, 0x5b, 0x9e, 0x6e,\n",
+ " 0xa6, 0x80, 0x5d, 0xa5, 0x83, 0x5b, 0xa4, 0x80, 0xb3, 0x79, 0x83, 0xb6,\n",
+ " 0xa3, 0x73, 0x84, 0x67, 0x8d, 0x8f, 0x9d, 0x78, 0xb8, 0x8a, 0x7b, 0x6c,\n",
+ " 0x85, 0x87, 0x6d, 0x75, 0xae, 0x75, 0x53, 0x71, 0x6b, 0x87, 0x67, 0x7b,\n",
+ " 0x7f, 0x86, 0x58, 0x73, 0x7d, 0x87, 0x5d, 0x7f, 0x7d, 0x63, 0x92, 0x65,\n",
+ " 0x7a, 0x9c, 0x6f, 0x87, 0x81, 0xa9, 0x91, 0x54, 0x66, 0x8e, 0x58, 0x6d,\n",
+ " 0x92, 0xc2, 0xa9, 0x7b, 0x6e, 0x96, 0x7c, 0x60, 0x7e, 0xa8, 0x85, 0x94,\n",
+ " 0x90, 0x8b, 0x77, 0x79, 0x77, 0xa7, 0x8f, 0x83, 0x80, 0x99, 0x8c, 0x80,\n",
+ " 0x93, 0x9c, 0x73, 0x9e, 0x75, 0x90, 0x67, 0x74, 0x99, 0x98, 0x7e, 0x76,\n",
+ " 0x9f, 0x82, 0x90, 0x95, 0x9d, 0x5f, 0x95, 0x98, 0x8c, 0x5f, 0x77, 0x83,\n",
+ " 0x7b, 0x72, 0x85, 0x7c, 0x97, 0x74, 0x81, 0x80, 0x8d, 0x89, 0x7d, 0x69,\n",
+ " 0x95, 0x85, 0x83, 0x5e, 0x95, 0x74, 0x54, 0x7f, 0x6c, 0x67, 0x9b, 0x83,\n",
+ " 0x88, 0x8e, 0x6f, 0x96, 0x81, 0x7f, 0x6e, 0x87, 0x8f, 0x6f, 0x61, 0x87,\n",
+ " 0x63, 0x66, 0x72, 0x77, 0x75, 0x6d, 0x59, 0x7d, 0xaa, 0x85, 0x62, 0x83,\n",
+ " 0x97, 0x94, 0x96, 0x89, 0x9d, 0x90, 0x7d, 0x91, 0x78, 0x57, 0xa0, 0x7f,\n",
+ " 0xa2, 0x62, 0x63, 0x99, 0x77, 0x71, 0x7f, 0x61, 0x99, 0x89, 0x6f, 0xa2,\n",
+ " 0xae, 0x92, 0x88, 0x51, 0x87, 0x7a, 0x6f, 0x89, 0xa8, 0x89, 0x64, 0x81,\n",
+ " 0x84, 0x79, 0x5b, 0x73, 0x82, 0x6e, 0x7e, 0x5d, 0x8f, 0x82, 0x51, 0x69,\n",
+ " 0x8e, 0x76, 0x8b, 0x58, 0x89, 0xb2, 0x52, 0x72, 0x7f, 0xae, 0x96, 0x5a,\n",
+ " 0x80, 0xa1, 0x74, 0x62, 0x8d, 0xbe, 0x87, 0x6c, 0x6d, 0xad, 0x83, 0x5a,\n",
+ " 0x6c, 0xa5, 0x7f, 0x7c, 0x7a, 0xa1, 0x75, 0x6d, 0x85, 0xbe, 0x91, 0x8e,\n",
+ " 0x96, 0x8c, 0x87, 0x74, 0x8b, 0x82, 0x96, 0x8f, 0x8f, 0x93, 0x8f, 0x8c,\n",
+ " 0x9a, 0x78, 0x73, 0x6e, 0x91, 0x8d, 0x7e, 0x81, 0x81, 0x52, 0x90, 0x85,\n",
+ " 0x77, 0x66, 0x7e, 0x75, 0x8a, 0x67, 0x72, 0x76, 0x82, 0x7b, 0x6e, 0x67,\n",
+ " 0x96, 0x7b, 0x75, 0x76, 0x8d, 0x76, 0x7f, 0x79, 0x84, 0x7b, 0x57, 0x81,\n",
+ " 0x76, 0x80, 0x67, 0x8c, 0x7c, 0x80, 0x67, 0x85, 0x79, 0x5b, 0x97, 0x74,\n",
+ " 0x91, 0x75, 0x82, 0x75, 0x6b, 0x94, 0x7e, 0x85, 0x8e, 0x77, 0x5d, 0x78,\n",
+ " 0xb5, 0x8b, 0x73, 0x7f, 0x62, 0x8f, 0xb1, 0x7d, 0xa2, 0x85, 0x6b, 0x92,\n",
+ " 0x75, 0x75, 0xb8, 0x7d, 0xb3, 0x67, 0x5f, 0xa6, 0x9b, 0x85, 0x9a, 0x67,\n",
+ " 0xbe, 0x8d, 0x92, 0x88, 0xa5, 0x7c, 0xaa, 0x5a, 0x71, 0x7b, 0x70, 0x77,\n",
+ " 0xa0, 0xa4, 0x5e, 0x55, 0x6b, 0x8e, 0x53, 0x89, 0x8a, 0x5a, 0x7c, 0x54,\n",
+ " 0x7c, 0x8b, 0x53, 0x77, 0x67, 0x77, 0x67, 0x5d, 0x91, 0xac, 0x78, 0x81,\n",
+ " 0x8e, 0xb5, 0x6d, 0x58, 0x78, 0xa6, 0x7c, 0x85, 0x87, 0xb3, 0x76, 0x5d,\n",
+ " 0x7c, 0x87, 0x57, 0x68, 0x82, 0x8f, 0x89, 0x76, 0x86, 0x9f, 0x6c, 0x68,\n",
+ " 0x7c, 0x87, 0x79, 0x9f, 0x86, 0x9e, 0x83, 0x70, 0x8d, 0xb2, 0x84, 0x71,\n",
+ " 0x71, 0x91, 0x9f, 0x8e, 0x83, 0x84, 0x87, 0x80, 0x94, 0x80, 0x7d, 0x8d,\n",
+ " 0x7c, 0x56, 0x5f, 0x80, 0x7d, 0x84, 0x61, 0x6e, 0x69, 0x80, 0x8b, 0x67,\n",
+ " 0xa4, 0x8b, 0x98, 0x7a, 0x8a, 0x6c, 0x77, 0x66, 0x7d, 0x6e, 0x84, 0x78,\n",
+ " 0x82, 0x7d, 0x61, 0x88, 0x6e, 0x53, 0x92, 0x75, 0x88, 0x77, 0x82, 0x9f,\n",
+ " 0x9e, 0x6f, 0x9c, 0x76, 0x91, 0x78, 0x69, 0x7f, 0x71, 0x6c, 0x6f, 0x7d,\n",
+ " 0x83, 0x6e, 0x3c, 0x84, 0x90, 0x8b, 0x71, 0x69, 0x75, 0x81, 0xc8, 0x84,\n",
+ " 0xa7, 0x8a, 0x8a, 0x90, 0x96, 0x86, 0x9e, 0x68, 0x99, 0x84, 0x8c, 0xa0,\n",
+ " 0x8a, 0x71, 0x7d, 0x41, 0xa1, 0x98, 0x77, 0x91, 0xaa, 0x86, 0x96, 0x5e,\n",
+ " 0x86, 0x76, 0xa7, 0x83, 0xac, 0x86, 0x66, 0x46, 0x6a, 0x81, 0x64, 0x77,\n",
+ " 0x67, 0x53, 0x80, 0x59, 0x73, 0x71, 0x63, 0x71, 0x76, 0x86, 0x62, 0x4f,\n",
+ " 0x83, 0xa4, 0x5d, 0x66, 0x93, 0x87, 0x87, 0x5b, 0x7f, 0x9d, 0x61, 0x9d,\n",
+ " 0x94, 0xa4, 0x84, 0x75, 0x67, 0xb3, 0x7b, 0x6d, 0x64, 0x98, 0x62, 0x77,\n",
+ " 0x7d, 0x98, 0x8e, 0x75, 0x7d, 0xa6, 0xa4, 0x8c, 0x83, 0x8b, 0x7a, 0x97,\n",
+ " 0x6c, 0x7f, 0x66, 0x7f, 0x8f, 0x98, 0x72, 0x6e, 0x75, 0x65, 0x80, 0x8d,\n",
+ " 0x88, 0x7d, 0x8c, 0x8d, 0x67, 0x68, 0xab, 0x8c, 0x8b, 0x76, 0x87, 0x69,\n",
+ " 0x88, 0x6c, 0x83, 0x6e, 0x88, 0x64, 0xa8, 0x67, 0xa5, 0x5b, 0x65, 0x60,\n",
+ " 0x6b, 0x62, 0x76, 0x78, 0x8c, 0x5b, 0x61, 0x6f, 0x66, 0x65, 0x92, 0x67,\n",
+ " 0x84, 0x7b, 0x80, 0x86, 0x7b, 0x6c, 0x86, 0x7a, 0x72, 0x7b, 0x4d, 0x94,\n",
+ " 0x80, 0x67, 0x8e, 0x8d, 0x7f, 0x79, 0x65, 0x78, 0xa3, 0x71, 0x80, 0x74,\n",
+ " 0xa7, 0xa8, 0x97, 0x78, 0x91, 0x77, 0x98, 0x86, 0x82, 0x64, 0xa5, 0x6e,\n",
+ " 0x7a, 0x5d, 0x6f, 0xad, 0x9b, 0x7a, 0x91, 0x4b, 0xa1, 0x75, 0x95, 0x76,\n",
+ " 0xac, 0x9d, 0xa3, 0x65, 0x65, 0x6a, 0x81, 0x8b, 0x9f, 0x67, 0x6b, 0x6a,\n",
+ " 0x60, 0x5b, 0x77, 0x96, 0x73, 0x78, 0x5a, 0x77, 0x5f, 0x68, 0x70, 0x72,\n",
+ " 0x78, 0x65, 0x81, 0x20, 0x86, 0x99, 0x80, 0x7a, 0xa5, 0xb1, 0x69, 0x45,\n",
+ " 0x7d, 0xa6, 0x7d, 0x85, 0xaa, 0xa9, 0x65, 0x60, 0x75, 0x9b, 0x61, 0x92,\n",
+ " 0x91, 0x8f, 0x8a, 0x81, 0x88, 0x9c, 0x81, 0x7d, 0x7b, 0x8f, 0x7e, 0x9e,\n",
+ " 0x82, 0x94, 0x95, 0x80, 0x73, 0xae, 0x7b, 0x7a, 0x79, 0x8c, 0x8b, 0x65,\n",
+ " 0x71, 0x75, 0x8d, 0x7a, 0x90, 0x83, 0x7b, 0x77, 0x71, 0x4f, 0x70, 0x95,\n",
+ " 0x87, 0x69, 0x97, 0x8e, 0x70, 0x92, 0x6e, 0x91, 0x9d, 0x72, 0x75, 0x82,\n",
+ " 0xad, 0x81, 0x78, 0x8d, 0x6f, 0x65, 0x88, 0x86, 0x8c, 0x8e, 0x59, 0x8b,\n",
+ " 0x67, 0x69, 0x8b, 0x78, 0x7f, 0x59, 0x73, 0x87, 0x6f, 0x86, 0x66, 0x7c,\n",
+ " 0x96, 0x68, 0x59, 0x78, 0x67, 0x92, 0x7b, 0x76, 0x80, 0x6e, 0x4a, 0x7b,\n",
+ " 0x99, 0x67, 0x72, 0x9c, 0x7a, 0x80, 0x76, 0x5f, 0x8e, 0x4f, 0x71, 0x77,\n",
+ " 0xab, 0x78, 0x99, 0x50, 0x83, 0x65, 0x78, 0x8c, 0xbb, 0x8d, 0x4e, 0x54,\n",
+ " 0x81, 0x6f, 0x7f, 0x91, 0xb9, 0x79, 0x9c, 0x65, 0x5a, 0x5a, 0x73, 0x8c,\n",
+ " 0x9a, 0xac, 0x99, 0x44, 0x7d, 0x4f, 0x78, 0x5a, 0x7d, 0x79, 0x57, 0x44,\n",
+ " 0x6f, 0x6a, 0x75, 0x7f, 0x5f, 0x6f, 0x72, 0x62, 0x7f, 0x89, 0x57, 0x91,\n",
+ " 0x8d, 0x83, 0x7e, 0x63, 0x8c, 0x95, 0x48, 0x78, 0xa9, 0x88, 0x84, 0x5b,\n",
+ " 0x8c, 0xa5, 0x65, 0x71, 0x88, 0x82, 0x7e, 0xa4, 0x8d, 0x7d, 0x7d, 0x8d,\n",
+ " 0x91, 0x7c, 0x73, 0x7d, 0x99, 0x89, 0x6d, 0xa1, 0x98, 0x84, 0x8b, 0x6b,\n",
+ " 0x89, 0x86, 0x84, 0x7e, 0x86, 0x87, 0x78, 0x8c, 0x96, 0x92, 0x5a, 0xa0,\n",
+ " 0x64, 0x73, 0x91, 0x88, 0x8f, 0x6b, 0x96, 0x5c, 0x99, 0x62, 0x78, 0x6c,\n",
+ " 0x87, 0x4d, 0x5d, 0x69, 0x7b, 0x81, 0x4a, 0x61, 0x71, 0x69, 0x7d, 0x91,\n",
+ " 0x67, 0x92, 0x68, 0x6f, 0x50, 0x5e, 0x61, 0x7e, 0x81, 0x70, 0x5f, 0x7b,\n",
+ " 0x6b, 0x55, 0x71, 0x6c, 0x70, 0x53, 0x3f, 0x80, 0x6e, 0x57, 0x96, 0x84,\n",
+ " 0x75, 0x51, 0x60, 0x9a, 0x7f, 0xa5, 0x80, 0x94, 0x95, 0x74, 0x7c, 0x83,\n",
+ " 0xa0, 0x93, 0x5d, 0x92, 0x83, 0x66, 0x67, 0x8a, 0x8b, 0x9b, 0x81, 0x69,\n",
+ " 0x73, 0x91, 0x6b, 0x79, 0x93, 0x88, 0x64, 0x68, 0x81, 0x8c, 0x6f, 0x81,\n",
+ " 0x6f, 0x80, 0x68, 0x5f, 0x9c, 0x95, 0x76, 0x93, 0x87, 0x68, 0x83, 0x94,\n",
+ " 0x8b, 0x85, 0x72, 0x7f, 0x64, 0x8c, 0x6a, 0x95, 0x8d, 0x80, 0x69, 0x6b,\n",
+ " 0x98, 0x86, 0x75, 0x92, 0x7a, 0x7f, 0x5b, 0x7f, 0x9b, 0x57, 0x99, 0x8d,\n",
+ " 0x8a, 0x7b, 0x58, 0x73, 0x88, 0x6d, 0x8a, 0x8c, 0x8e, 0x82, 0x85, 0xaa,\n",
+ " 0x72, 0xa6, 0x7f, 0x7a, 0x83, 0x59, 0x6d, 0x6e, 0x79, 0x83, 0x88, 0x84,\n",
+ " 0x74, 0x85, 0x74, 0x78, 0x80, 0x7c, 0x97, 0x86, 0x94, 0x65, 0x7e, 0x80,\n",
+ " 0x6f, 0x97, 0x70, 0x74, 0x92, 0x76, 0x71, 0x91, 0x85, 0x72, 0x6e, 0x84,\n",
+ " 0x78, 0x7e, 0x88, 0x79, 0x7f, 0x80, 0x83, 0x7a, 0x85, 0x75, 0x82, 0x81,\n",
+ " 0x82, 0x7b, 0x7a, 0xa0, 0x76, 0x7f, 0x75, 0xa7, 0x67, 0x8e, 0x81, 0x98,\n",
+ " 0xa5, 0x86, 0x77, 0x78, 0x7f, 0x97, 0x90, 0x86, 0x80, 0x6b, 0x89, 0x66,\n",
+ " 0x9b, 0x5c, 0x8b, 0x74, 0xac, 0x89, 0x89, 0x92, 0x92, 0xa8, 0x61, 0x85,\n",
+ " 0x8c, 0x86, 0x88, 0x91, 0x92, 0x66, 0x63, 0x6c, 0x7a, 0x80, 0x7d, 0x90,\n",
+ " 0x6f, 0x7f, 0x92, 0x94, 0x8e, 0x7a, 0x86, 0x98, 0xa1, 0x59, 0x71, 0x8c,\n",
+ " 0x63, 0xa3, 0x60, 0x7d, 0x88, 0x6a, 0x83, 0x6e, 0x7a, 0x94, 0x7b, 0x81,\n",
+ " 0x7d, 0x83, 0x77, 0x7e, 0x63, 0xab, 0x75, 0x7b, 0x71, 0x8f, 0x76, 0x6e,\n",
+ " 0x78, 0x7b, 0x79, 0x86, 0x69, 0x67, 0x67, 0x70, 0x6c, 0x7a, 0x6c, 0x84,\n",
+ " 0x74, 0xa2, 0x74, 0x77, 0x8a, 0x58, 0x7d, 0xa0, 0x65, 0x7b, 0x79, 0x71,\n",
+ " 0x7c, 0x3c, 0x85, 0x96, 0x59, 0x76, 0x6a, 0x94, 0xa5, 0x5b, 0x70, 0x99,\n",
+ " 0x7f, 0x9a, 0x69, 0x7c, 0x6f, 0x79, 0x72, 0x8b, 0x83, 0x6e, 0x73, 0x7f,\n",
+ " 0x6f, 0x6d, 0x7e, 0xa3, 0x72, 0x87, 0x83, 0x8c, 0x8c, 0x70, 0x77, 0x75,\n",
+ " 0xa4, 0x5a, 0x89, 0x7d, 0xa0, 0x97, 0x67, 0x80, 0x78, 0x7e, 0x86, 0x6a,\n",
+ " 0x7b, 0x9c, 0x77, 0x67, 0x7b, 0x74, 0x7f, 0xa5, 0x90, 0x94, 0x92, 0x4d,\n",
+ " 0x7a, 0x79, 0x9f, 0x87, 0x64, 0x6e, 0x6d, 0x59, 0x83, 0x54, 0x79, 0x82,\n",
+ " 0x6c, 0x74, 0x82, 0x98, 0x77, 0x90, 0x85, 0xa4, 0x88, 0x81, 0x71, 0x85,\n",
+ " 0x90, 0x8e, 0x88, 0x68, 0x51, 0x6d, 0x71, 0x7b, 0x80, 0xbc, 0xa5, 0x57,\n",
+ " 0x8f, 0x9f, 0x95, 0x89, 0xb1, 0x96, 0x69, 0x65, 0x61, 0x73, 0x6f, 0x6c,\n",
+ " 0x5b, 0x95, 0x99, 0x7f, 0x76, 0x9d, 0x7c, 0x7d, 0x8d, 0xb1, 0x8f, 0x6a,\n",
+ " 0x76, 0x95, 0x74, 0x7a, 0x7b, 0xae, 0x77, 0x76, 0x6d, 0x99, 0x7d, 0x80,\n",
+ " 0x6e, 0x89, 0x7f, 0x74, 0x6f, 0x72, 0x89, 0x8b, 0x86, 0x7b, 0x7c, 0x72,\n",
+ " 0x6b, 0x4f, 0x71, 0x94, 0x80, 0x96, 0x83, 0x7e, 0x75, 0x74, 0x68, 0x83,\n",
+ " 0x95, 0x8c, 0x85, 0x7a, 0x82, 0x74, 0x85, 0x83, 0x8c, 0x7e, 0x7a, 0xa0,\n",
+ " 0x8e, 0x67, 0x6b, 0x82, 0x9b, 0x66, 0x6c, 0x8a, 0x88, 0x7e, 0x74, 0x9e,\n",
+ " 0x88, 0x82, 0x73, 0x73, 0x79, 0x7c, 0x72, 0x6b, 0x74, 0x8b, 0xa4, 0xa4,\n",
+ " 0xa3, 0x73, 0x73, 0x88, 0x8d, 0x94, 0x84, 0x9a, 0x9e, 0x93, 0x6c, 0x86,\n",
+ " 0x7a, 0x7a, 0x7e, 0xaa, 0x66, 0x8f, 0x99, 0xa4, 0x70, 0x4c, 0x6f, 0x66,\n",
+ " 0x8a, 0xaa, 0x69, 0x80, 0x6a, 0x5e, 0x71, 0x8f, 0x8b, 0x84, 0x75, 0x9d,\n",
+ " 0x5c, 0x60, 0x61, 0x4a, 0x6f, 0x91, 0x78, 0x6e, 0x8c, 0x62, 0x88, 0x75,\n",
+ " 0x64, 0x7c, 0x7d, 0x92, 0x9b, 0x96, 0x62, 0x72, 0x6c, 0x6f, 0x87, 0x5d,\n",
+ " 0xa0, 0xa7, 0x7c, 0x58, 0x6e, 0x8c, 0x82, 0x84, 0x7f, 0x8b, 0x54, 0x77,\n",
+ " 0x5b, 0x9a, 0x6a, 0x78, 0x5d, 0xb9, 0x8e, 0x7d, 0x6e, 0xa1, 0x66, 0x7c,\n",
+ " 0x87, 0xd2, 0x7a, 0x6c, 0x82, 0xa1, 0x83, 0x59, 0x64, 0x9e, 0x65, 0x6d,\n",
+ " 0x77, 0x80, 0x7c, 0x9a, 0x50, 0x9f, 0x8b, 0x7a, 0x73, 0x80, 0x92, 0x6d,\n",
+ " 0x97, 0x7f, 0x74, 0x6a, 0x5f, 0x44, 0x7d, 0x99, 0x95, 0x91, 0x8f, 0x6a,\n",
+ " 0x63, 0x56, 0x89, 0x96, 0xba, 0xa6, 0x71, 0x98, 0x9d, 0x3a, 0x8f, 0x77,\n",
+ " 0x6d, 0x76, 0x68, 0xb4, 0x8d, 0x79, 0x7a, 0x83, 0x7f, 0x96, 0x75, 0x94,\n",
+ " 0x9e, 0x51, 0x83, 0x5b, 0x66, 0x73, 0xa1, 0xbc, 0x8c, 0x70, 0x88, 0x80,\n",
+ " 0x92, 0x60, 0x7d, 0xa9, 0x97, 0x74, 0x7d, 0x98, 0x7b, 0x78, 0x85, 0xa7,\n",
+ " 0x8f, 0x8c, 0x91, 0x9d, 0x6a, 0x80, 0x6c, 0x8e, 0x8e, 0x91, 0x76, 0x8b,\n",
+ " 0x79, 0x59, 0x7d, 0x9c, 0x69, 0x83, 0x8c, 0x95, 0x8e, 0x75, 0x9d, 0x83,\n",
+ " 0x92, 0x99, 0x8a, 0x59, 0x61, 0x54, 0x63, 0x86, 0x83, 0x86, 0x98, 0x83,\n",
+ " 0x73, 0x74, 0x91, 0x52, 0x60, 0x8a, 0x7c, 0x57, 0xbc, 0x9d, 0x86, 0x6b,\n",
+ " 0x63, 0xa2, 0x78, 0x80, 0x75, 0xb1, 0x74, 0x76, 0x69, 0x8b, 0x7e, 0x76,\n",
+ " 0x7b, 0xb3, 0x77, 0x5b, 0x6c, 0x8b, 0x83, 0x80, 0x7f, 0xd1, 0x7c, 0x58,\n",
+ " 0x6f, 0x98, 0x71, 0x57, 0x60, 0xd0, 0x84, 0x62, 0x74, 0xa6, 0x8f, 0x7b,\n",
+ " 0x70, 0xaa, 0x81, 0x6b, 0x7f, 0x89, 0x6a, 0x74, 0x5a, 0x8c, 0x9c, 0x77,\n",
+ " 0x5d, 0x84, 0x63, 0x94, 0x8e, 0x91, 0x83, 0x4a, 0x49, 0x74, 0x6b, 0x70,\n",
+ " 0xc0, 0xa0, 0x6a, 0x90, 0x8e, 0x5a, 0x70, 0x96, 0xab, 0x72, 0x7e, 0xba,\n",
+ " 0xa7, 0x46, 0x86, 0x5d, 0x90, 0x76, 0x95, 0x8d, 0xa5, 0x40, 0x82, 0x8a,\n",
+ " 0x7d, 0x5e, 0x73, 0x94, 0x9d, 0x58, 0x8c, 0x8b, 0x69, 0x6c, 0x9a, 0x90,\n",
+ " 0xaa, 0x6f, 0x85, 0x8d, 0x64, 0x58, 0x7b, 0x97, 0xa9, 0x79, 0xa5, 0xa2,\n",
+ " 0x5f, 0x57, 0x9a, 0xb4, 0x89, 0x70, 0x84, 0x73, 0x46, 0x6c, 0x6e, 0x87,\n",
+ " 0x70, 0x94, 0x8a, 0x8a, 0x69, 0x7b, 0x6c, 0x68, 0x8e, 0xa2, 0x90, 0x84,\n",
+ " 0x78, 0x45, 0x63, 0x78, 0x7f, 0x90, 0x9f, 0x90, 0x68, 0x43, 0x92, 0x77,\n",
+ " 0x78, 0x77, 0x82, 0x7d, 0x8f, 0x6a, 0x7a, 0x70, 0x76, 0x75, 0x87, 0x63,\n",
+ " 0xbc, 0x8e, 0x6a, 0x71, 0x51, 0x51, 0x75, 0x6b, 0x8a, 0xb4, 0x6a, 0x5b,\n",
+ " 0x99, 0x84, 0x76, 0x84, 0x74, 0xaf, 0x86, 0x6a, 0x53, 0x97, 0x6e, 0x8e,\n",
+ " 0x61, 0xc4, 0x7e, 0x5d, 0x4d, 0x96, 0x73, 0x73, 0x53, 0xc0, 0x8f, 0x68,\n",
+ " 0x58, 0xae, 0x81, 0x83, 0x62, 0x98, 0x7b, 0x89, 0x54, 0x86, 0x78, 0x67,\n",
+ " 0x70, 0x9b, 0x63, 0x5f, 0x2d, 0x77, 0x84, 0x79, 0x6b, 0xa4, 0x7b, 0x65,\n",
+ " 0x45, 0x65, 0x56, 0x86, 0xbb, 0x8a, 0x8e, 0x92, 0x86, 0x48, 0x7c, 0x6d,\n",
+ " 0xb4, 0x7d, 0x56, 0xa4, 0x86, 0x52, 0x8b, 0x6a, 0x8d, 0x5b, 0x9d, 0xa2,\n",
+ " 0xbf, 0x36, 0x7c, 0x99, 0x9d, 0x65, 0x75, 0xa4, 0x9f, 0x6a, 0x7c, 0x6b,\n",
+ " 0x6f, 0x55, 0x70, 0x7f, 0xc2, 0x38, 0x6e, 0xa4, 0x74, 0x4c, 0x75, 0xbb,\n",
+ " 0xa4, 0x75, 0x8e, 0x8f, 0x56, 0x65, 0x57, 0x92, 0x73, 0x7f, 0x7d, 0x86,\n",
+ " 0x65, 0x76, 0x92, 0x84, 0x70, 0xa8, 0x91, 0x5b, 0x69, 0x74, 0x8e, 0x82,\n",
+ " 0x78, 0x8a, 0xaa, 0x71, 0x70, 0x50, 0x85, 0x82, 0x7d, 0x94, 0xa0, 0x76,\n",
+ " 0x6d, 0x55, 0x86, 0x79, 0x71, 0x7f, 0x9b, 0x71, 0x8a, 0x42, 0x87, 0x64,\n",
+ " 0x57, 0x88, 0xa0, 0x77, 0xa8, 0x91, 0x72, 0x65, 0x7e, 0x6b, 0x7e, 0x81,\n",
+ " 0x8d, 0x97, 0x7e, 0x6a, 0x92, 0x88, 0x84, 0x7a, 0x61, 0xa9, 0x86, 0x59,\n",
+ " 0x6c, 0x87, 0x61, 0x72, 0x4f, 0xc8, 0x99, 0x6c, 0x66, 0xa3, 0x80, 0x8b,\n",
+ " 0x5c, 0xc0, 0x69, 0x7a, 0x6c, 0xb8, 0x8e, 0x91, 0x51, 0x9f, 0x8c, 0x85,\n",
+ " 0x75, 0x96, 0x8c, 0x84, 0x6b, 0xa6, 0x71, 0x62, 0x42, 0x60, 0x74, 0x72,\n",
+ " 0x92, 0x91, 0x70, 0x5b, 0x3d, 0x71, 0x5e, 0x91, 0xa3, 0xa5, 0x6a, 0x7c,\n",
+ " 0x60, 0x58, 0x82, 0x80, 0xa3, 0x73, 0x8f, 0xa0, 0xb2, 0x4b, 0x94, 0x5e,\n",
+ " 0x9f, 0x75, 0x4d, 0x83, 0xbc, 0x42, 0x5e, 0x80, 0x8f, 0x59, 0x53, 0xac,\n",
+ " 0xb2, 0x45, 0x68, 0x7d, 0x9a, 0x65, 0x8a, 0xaa, 0xa0, 0x4e, 0x77, 0x72,\n",
+ " 0x4d, 0x62, 0x6e, 0x98, 0x8c, 0x73, 0x92, 0x5a, 0x49, 0x55, 0x7b, 0x98,\n",
+ " 0x8d, 0x84, 0x80, 0x8e, 0x2e, 0x56, 0x78, 0x73, 0x7b, 0x8f, 0x9a, 0x69,\n",
+ " 0x73, 0x68, 0x7a, 0x88, 0x78, 0xa5, 0xb1, 0x5c, 0x8f, 0x55, 0x71, 0x99,\n",
+ " 0x7a, 0xa9, 0xb0, 0x75, 0x69, 0x44, 0x5f, 0x66, 0x81, 0x7d, 0x9e, 0x4f,\n",
+ " 0x66, 0x7f, 0x87, 0x7d, 0x5d, 0x7c, 0x95, 0x62, 0xa5, 0x86, 0x90, 0x6f,\n",
+ " 0x60, 0xa5, 0x6e, 0x70, 0x80, 0x96, 0x6f, 0x55, 0x77, 0x87, 0x99, 0x7b,\n",
+ " 0x21, 0xaa, 0x7f, 0x60, 0x63, 0xae, 0x47, 0x79, 0x44, 0xb5, 0x83, 0x6e,\n",
+ " 0x6d, 0x93, 0x76, 0x54, 0x4b, 0xad, 0x91, 0x6b, 0x6a, 0x9c, 0x8c, 0x83,\n",
+ " 0x62, 0x8a, 0x88, 0x71, 0x73, 0xa0, 0x75, 0x95, 0x54, 0x80, 0x92, 0x65,\n",
+ " 0x45, 0x80, 0x63, 0x9a, 0x93, 0x9b, 0x78, 0x4e, 0x4d, 0x5f, 0x69, 0x9e,\n",
+ " 0xbd, 0xa5, 0x75, 0x6b, 0x6e, 0x6a, 0x82, 0x97, 0xab, 0x60, 0x76, 0xb3,\n",
+ " 0xc1, 0x39, 0x82, 0x5b, 0x71, 0x31, 0x7b, 0x9c, 0xb5, 0x4f, 0x75, 0x79,\n",
+ " 0x6c, 0x5d, 0x80, 0xa6, 0x9c, 0x53, 0x6f, 0x85, 0x84, 0x5e, 0x7d, 0xb5,\n",
+ " 0x95, 0x5f, 0x7c, 0x98, 0x72, 0x7c, 0x67, 0x99, 0xbb, 0x6c, 0x73, 0x66,\n",
+ " 0x59, 0x5c, 0x6c, 0x9a, 0x9b, 0x72, 0x9b, 0x5f, 0x4b, 0x51, 0x63, 0x84,\n",
+ " 0x74, 0xa0, 0xb3, 0x6e, 0x63, 0xa0, 0x84, 0x90, 0x71, 0x91, 0xba, 0x64,\n",
+ " 0x6d, 0x72, 0x78, 0x83, 0x6f, 0x8e, 0xbd, 0x64, 0x69, 0x60, 0x95, 0x67,\n",
+ " 0x70, 0x93, 0x78, 0x4d, 0x91, 0x3f, 0x7b, 0x6d, 0x69, 0x87, 0x7d, 0x8a,\n",
+ " 0xa3, 0x95, 0x9d, 0x66, 0x6d, 0x8b, 0x7a, 0x75, 0x94, 0x7b, 0x89, 0x52,\n",
+ " 0x66, 0x65, 0x79, 0x84, 0x49, 0x9c, 0x60, 0x66, 0x3e, 0xab, 0x4a, 0x86,\n",
+ " 0x54, 0xcd, 0x7c, 0x83, 0x7c, 0xac, 0x8b, 0x53, 0x67, 0xbb, 0x7c, 0x6d,\n",
+ " 0x72, 0xb3, 0x83, 0x85, 0x4f, 0x97, 0x86, 0x60, 0x7d, 0x93, 0x70, 0x8b,\n",
+ " 0x64, 0x78, 0x82, 0x73, 0x54, 0x87, 0x6c, 0xaa, 0x6f, 0x97, 0x8d, 0x51,\n",
+ " 0x2d, 0x50, 0x75, 0xa9, 0xc2, 0x94, 0x8d, 0x6f, 0x6d, 0x71, 0x7b, 0x87,\n",
+ " 0x93, 0x67, 0x7d, 0xa5, 0xa2, 0x4f, 0x99, 0x83, 0x95, 0x49, 0x70, 0x9c,\n",
+ " 0xcf, 0x37, 0x84, 0x86, 0x94, 0x5c, 0x95, 0xa1, 0xb6, 0x73, 0x80, 0x8d,\n",
+ " 0x89, 0x62, 0x6f, 0xb4, 0xa1, 0x5b, 0x64, 0x91, 0x41, 0x4f, 0x53, 0xa6,\n",
+ " 0xae, 0x75, 0x84, 0x82, 0x58, 0x8e, 0x63, 0x95, 0xa3, 0x8d, 0x8b, 0x76,\n",
+ " 0x5d, 0x78, 0x80, 0x82, 0x6e, 0x9d, 0xb8, 0x7d, 0x64, 0x8a, 0x7e, 0x80,\n",
+ " 0x72, 0x99, 0xcf, 0x76, 0x66, 0x77, 0x7c, 0x81, 0x71, 0x6f, 0xa1, 0x6c,\n",
+ " 0x6b, 0x70, 0x80, 0x7c, 0x6d, 0x83, 0x8e, 0x74, 0x7a, 0x58, 0x69, 0x53,\n",
+ " 0x58, 0x7d, 0x7f, 0x84, 0x96, 0x9c, 0x75, 0x6e, 0x62, 0x7c, 0x88, 0x7e,\n",
+ " 0x7f, 0x98, 0x93, 0x61, 0x98, 0x98, 0x80, 0x83, 0x2e, 0x7d, 0x64, 0x69,\n",
+ " 0x50, 0xa5, 0x38, 0x96, 0x2e, 0xc5, 0x66, 0x56, 0x64, 0xaa, 0x63, 0x64,\n",
+ " 0x6d, 0xb3, 0x8a, 0x6c, 0x59, 0xb6, 0x69, 0x7a, 0x54, 0x91, 0x58, 0x96,\n",
+ " 0x6b, 0x9f, 0x6d, 0x88, 0x4a, 0x82, 0x94, 0x67, 0x38, 0x93, 0x60, 0x87,\n",
+ " 0x8c, 0x93, 0x8c, 0x52, 0x31, 0x43, 0x66, 0xa9, 0xb3, 0x7a, 0x88, 0x64,\n",
+ " 0x60, 0x5b, 0x80, 0x84, 0xb7, 0x5a, 0x7a, 0x9d, 0x92, 0x50, 0x89, 0x80,\n",
+ " 0x72, 0x51, 0x7f, 0x85, 0xae, 0x47, 0x76, 0x9a, 0x7a, 0x74, 0x6d, 0x93,\n",
+ " 0xbd, 0x42, 0x72, 0x6d, 0x58, 0x5e, 0x6e, 0xa4, 0xb5, 0x4e, 0x76, 0x8f,\n",
+ " 0x75, 0x9b, 0x5d, 0x92, 0xad, 0x77, 0x7f, 0x73, 0x62, 0x7d, 0x65, 0xaf,\n",
+ " 0x98, 0x87, 0x80, 0x7c, 0x61, 0x81, 0x45, 0xa0, 0x84, 0x99, 0xbb, 0x72,\n",
+ " 0x86, 0x8f, 0x70, 0x97, 0x6a, 0x8a, 0xd3, 0x70, 0x7c, 0x91, 0x77, 0x82,\n",
+ " 0x70, 0x8c, 0xd5, 0x6c, 0x7f, 0x51, 0x5f, 0x69, 0x72, 0x89, 0x9a, 0x68,\n",
+ " 0x79, 0x70, 0x8b, 0x80, 0x52, 0x98, 0x86, 0x7a, 0xa0, 0x7b, 0x61, 0x6e,\n",
+ " 0x66, 0x6f, 0x77, 0x78, 0x64, 0xac, 0x7e, 0x73, 0x5d, 0x71, 0x6f, 0x80,\n",
+ " 0x2e, 0xa9, 0x90, 0x5c, 0x56, 0xa1, 0x32, 0x88, 0x55, 0xb9, 0x67, 0x6f,\n",
+ " 0x5c, 0xa5, 0x87, 0x61, 0x6b, 0xbd, 0x77, 0x7c, 0x62, 0xae, 0x7c, 0x7a,\n",
+ " 0x66, 0xac, 0x7a, 0x62, 0x5c, 0x9a, 0x58, 0x89, 0x5a, 0x74, 0x72, 0x66,\n",
+ " 0x5c, 0x8e, 0x51, 0x8e, 0x99, 0x92, 0xa0, 0x49, 0x31, 0x55, 0x68, 0x99,\n",
+ " 0xba, 0x82, 0xa2, 0x7a, 0x5e, 0x6f, 0x84, 0x98, 0x96, 0x52, 0x73, 0x99,\n",
+ " 0xb4, 0x5e, 0x7c, 0x59, 0x7d, 0x4a, 0x7e, 0xa0, 0xbe, 0x63, 0x67, 0x8e,\n",
+ " 0x7f, 0x71, 0x80, 0xaf, 0x93, 0x4e, 0x78, 0x7e, 0x6d, 0x52, 0x66, 0xb3,\n",
+ " 0x94, 0x56, 0x84, 0x8f, 0x50, 0x6d, 0x65, 0xa8, 0xb3, 0x4b, 0x91, 0x7f,\n",
+ " 0x4c, 0x8d, 0x69, 0x79, 0x95, 0x8f, 0x8f, 0x7c, 0x66, 0x98, 0x75, 0x9b,\n",
+ " 0x73, 0x9b, 0xac, 0x79, 0x6e, 0x84, 0x69, 0x9e, 0x80, 0xa0, 0xb0, 0x6c,\n",
+ " 0x46, 0x8b, 0x3f, 0x7a, 0x79, 0x79, 0xb3, 0x62, 0x6b, 0x60, 0x67, 0x81,\n",
+ " 0x4a, 0x7e, 0xa7, 0x8c, 0x74, 0x7f, 0x67, 0x4c, 0x4b, 0x8c, 0x8e, 0x67,\n",
+ " 0x78, 0x9d, 0x94, 0x79, 0x75, 0x7c, 0x86, 0x7b, 0x67, 0x9f, 0xa4, 0x61,\n",
+ " 0x5b, 0x6e, 0x85, 0x70, 0x20, 0xa5, 0x66, 0x5e, 0x55, 0xad, 0x3e, 0x7c,\n",
+ " 0x2d, 0xb4, 0x78, 0x6f, 0x4c, 0xc6, 0x7e, 0x6d, 0x54, 0xb4, 0x71, 0x78,\n",
+ " 0x54, 0xc3, 0x66, 0x6e, 0x4a, 0xa0, 0x7b, 0x85, 0x66, 0x94, 0x75, 0x8d,\n",
+ " 0x34, 0x88, 0x71, 0x4e, 0x49, 0x8a, 0x3b, 0x9c, 0x88, 0x76, 0x7f, 0x6a,\n",
+ " 0x37, 0x64, 0x66, 0xb6, 0xa3, 0x82, 0x76, 0x82, 0x6d, 0x65, 0x6f, 0x8c,\n",
+ " 0x99, 0x5e, 0x77, 0xa1, 0x99, 0x51, 0xa1, 0x67, 0x6f, 0x4c, 0x7f, 0x9e,\n",
+ " 0xad, 0x40, 0x65, 0x82, 0x76, 0x66, 0x72, 0xb5, 0xb2, 0x5b, 0x71, 0x8a,\n",
+ " 0x76, 0x74, 0x52, 0xa0, 0x91, 0x37, 0x86, 0x72, 0x6c, 0x75, 0x62, 0xa5,\n",
+ " 0xb6, 0x57, 0x75, 0x90, 0x3e, 0x7f, 0x49, 0x9f, 0x8e, 0x92, 0x81, 0x87,\n",
+ " 0x69, 0x9e, 0x6b, 0x86, 0x8d, 0xb1, 0x9e, 0x65, 0x6f, 0x93, 0x70, 0x79,\n",
+ " 0x7b, 0x87, 0xbe, 0x59, 0x69, 0x7a, 0x56, 0x7a, 0x81, 0x7d, 0xb8, 0x67,\n",
+ " 0x67, 0x7f, 0x54, 0x8f, 0x71, 0x85, 0xa0, 0x74, 0x89, 0x5d, 0x67, 0x52,\n",
+ " 0x65, 0x96, 0x89, 0x84, 0x81, 0x83, 0x82, 0x9a, 0x85, 0x73, 0x78, 0x62,\n",
+ " 0x87, 0x98, 0x75, 0x6a, 0x73, 0x95, 0x86, 0x71, 0x11, 0x9a, 0x91, 0x66,\n",
+ " 0x6e, 0xa4, 0x35, 0x89, 0x47, 0xbb, 0x5e, 0x46, 0x3a, 0xa8, 0x70, 0x4a,\n",
+ " 0x65, 0xb9, 0x70, 0x96, 0x66, 0xcf, 0x80, 0x79, 0x60, 0xa4, 0x79, 0x70,\n",
+ " 0x68, 0x92, 0x7f, 0x89, 0x6b, 0x87, 0x77, 0x67, 0x5b, 0x74, 0x3f, 0x9e,\n",
+ " 0x94, 0x9b, 0xa1, 0x61, 0x4b, 0x66, 0x70, 0xad, 0xb7, 0x67, 0x70, 0x6c,\n",
+ " 0x3f, 0x5b, 0x94, 0x88, 0xb3, 0x4f, 0x97, 0x97, 0x8c, 0x55, 0xb8, 0x78,\n",
+ " 0x60, 0x25, 0x51, 0x91, 0xcd, 0x44, 0x6f, 0x85, 0x5c, 0x65, 0x67, 0xa5,\n",
+ " 0x9e, 0x5f, 0x6d, 0x85, 0x6d, 0x56, 0x80, 0xae, 0x79, 0x63, 0x4f, 0x7d,\n",
+ " 0x5f, 0x6b, 0x6e, 0xa7, 0x8e, 0x76, 0x8f, 0x90, 0x6e, 0x8c, 0x88, 0x92,\n",
+ " 0x81, 0x81, 0x96, 0x7d, 0x48, 0x6b, 0x3f, 0xa1, 0x8c, 0xa2, 0x9f, 0x7f,\n",
+ " 0x77, 0x97, 0x73, 0x9c, 0x67, 0x95, 0xae, 0x77, 0x7f, 0x7a, 0x52, 0x7e,\n",
+ " 0x91, 0x77, 0xa8, 0x54, 0x6a, 0x74, 0x52, 0x8a, 0x67, 0x8e, 0x90, 0x8d,\n",
+ " 0x8b, 0x52, 0x72, 0x5a, 0x73, 0x8f, 0x94, 0x87, 0x7c, 0x88, 0x89, 0x76,\n",
+ " 0x77, 0x88, 0x5c, 0x77, 0x8f, 0x94, 0xac, 0x58, 0x70, 0x79, 0x75, 0x8a,\n",
+ " 0x20, 0x9c, 0x91, 0x55, 0x55, 0xa4, 0x5b, 0x84, 0x30, 0xc6, 0x8a, 0x51,\n",
+ " 0x31, 0xc3, 0x72, 0x6b, 0x65, 0xb9, 0x79, 0x7d, 0x62, 0xad, 0x88, 0x75,\n",
+ " 0x37, 0xb0, 0x76, 0x8a, 0x7d, 0x85, 0x7f, 0xb4, 0x46, 0x9c, 0x83, 0x7b,\n",
+ " 0x79, 0x78, 0x56, 0xac, 0x8d, 0xa2, 0xa9, 0x54, 0x44, 0x5a, 0x63, 0xb2,\n",
+ " 0xa8, 0x72, 0xa4, 0x6b, 0x5d, 0x4d, 0x8e, 0x95, 0x9e, 0x4a, 0x98, 0x8c,\n",
+ " 0xb0, 0x5c, 0xa5, 0x75, 0x83, 0x3b, 0x46, 0x92, 0xa7, 0x3b, 0x6a, 0x75,\n",
+ " 0x59, 0x57, 0x52, 0xa1, 0xab, 0x54, 0x68, 0x7c, 0x94, 0x6e, 0x5b, 0x9a,\n",
+ " 0xa3, 0x5d, 0x73, 0x74, 0x5a, 0x63, 0x56, 0x9e, 0xc1, 0x71, 0x82, 0x79,\n",
+ " 0x49, 0x92, 0x63, 0xa6, 0x99, 0x7d, 0x71, 0x81, 0x5e, 0x90, 0x5c, 0x8b,\n",
+ " 0x7e, 0xb4, 0xa0, 0x8c, 0x67, 0x93, 0x4e, 0x72, 0x65, 0x83, 0xb5, 0x77,\n",
+ " 0x83, 0x92, 0x43, 0x67, 0x8c, 0x81, 0xb1, 0x75, 0x6a, 0x61, 0x66, 0x6f,\n",
+ " 0x5d, 0x7f, 0x8d, 0x7b, 0x6b, 0x68, 0x6f, 0x85, 0x6e, 0x87, 0x97, 0x89,\n",
+ " 0x9b, 0x81, 0x7e, 0x7e, 0x9d, 0x83, 0x6b, 0x6a, 0xa5, 0x92, 0x7e, 0x70,\n",
+ " 0x60, 0x8f, 0x6f, 0x8b, 0x15, 0xa6, 0x66, 0x4e, 0x61, 0xbc, 0x38, 0x67,\n",
+ " 0x46, 0xab, 0x84, 0x5e, 0x3a, 0xac, 0x74, 0x58, 0x76, 0xc4, 0x7a, 0x76,\n",
+ " 0x67, 0xc0, 0x76, 0x6f, 0x52, 0xa6, 0xa2, 0x97, 0x76, 0xa6, 0x7f, 0x99,\n",
+ " 0x5d, 0xa5, 0x5f, 0x60, 0x58, 0x88, 0x3f, 0x9e, 0x7d, 0x81, 0x71, 0x63,\n",
+ " 0x42, 0x55, 0x3e, 0xbd, 0xa9, 0x7a, 0xa5, 0x67, 0x62, 0x7a, 0x80, 0x9e,\n",
+ " 0xc3, 0x54, 0x7f, 0x9f, 0x93, 0x73, 0xbd, 0x79, 0x74, 0x2e, 0x54, 0x9e,\n",
+ " 0xaa, 0x76, 0x68, 0x80, 0x78, 0x64, 0x57, 0x93, 0xa4, 0x56, 0x75, 0x72,\n",
+ " 0x81, 0x7f, 0x48, 0xad, 0x89, 0x67, 0x60, 0x7e, 0x7a, 0x83, 0x6e, 0x95,\n",
+ " 0xb0, 0x57, 0x89, 0x91, 0x4d, 0x86, 0x78, 0x7b, 0x74, 0x8c, 0x8f, 0x8d,\n",
+ " 0x67, 0xa4, 0x64, 0x8d, 0x77, 0x9a, 0xa1, 0x88, 0x6e, 0x94, 0x33, 0x95,\n",
+ " 0x81, 0x76, 0xc6, 0x7d, 0x7d, 0x85, 0x5a, 0x6e, 0x8e, 0x69, 0x9e, 0x71,\n",
+ " 0x82, 0x81, 0x59, 0x5b, 0x71, 0x9a, 0x91, 0x8e, 0x80, 0x69, 0x71, 0x73,\n",
+ " 0x6e, 0x9a, 0x95, 0x94, 0x7b, 0x80, 0x82, 0x7e, 0x76, 0x84, 0x70, 0x72,\n",
+ " 0x9c, 0xa0, 0x77, 0x66, 0x55, 0xa1, 0x8c, 0x73, 0x35, 0xa0, 0x68, 0x4d,\n",
+ " 0x3b, 0xaa, 0x44, 0x6f, 0x3c, 0xc0, 0x96, 0x78, 0x33, 0xbd, 0x64, 0x5b,\n",
+ " 0x75, 0xd2, 0x83, 0x87, 0x59, 0xbd, 0x80, 0x80, 0x6e, 0x8e, 0x65, 0x7a,\n",
+ " 0x87, 0xb6, 0x8d, 0x94, 0x39, 0x95, 0x8b, 0x5d, 0x66, 0x71, 0x4e, 0x9f,\n",
+ " 0x96, 0x8a, 0x98, 0x47, 0x41, 0x6c, 0x4c, 0xac, 0x95, 0x81, 0x90, 0x75,\n",
+ " 0x59, 0x4c, 0xa2, 0x93, 0x99, 0x58, 0x7b, 0xaf, 0xa3, 0x52, 0xb0, 0x6c,\n",
+ " 0x5f, 0x47, 0x6e, 0x8e, 0xae, 0x3d, 0x81, 0x6d, 0x78, 0x52, 0x4f, 0x81,\n",
+ " 0x80, 0x68, 0x4b, 0x81, 0x74, 0x71, 0x67, 0xa7, 0x9a, 0x55, 0x84, 0x72,\n",
+ " 0x64, 0x6b, 0x6e, 0x9d, 0xab, 0x76, 0x79, 0x85, 0x40, 0x84, 0x80, 0x85,\n",
+ " 0x70, 0x91, 0x9a, 0x81, 0x5b, 0x89, 0x6b, 0x8a, 0x92, 0x8c, 0xa4, 0x7b,\n",
+ " 0x75, 0x89, 0x54, 0x76, 0x69, 0x69, 0xb3, 0x6c, 0x47, 0x7d, 0x4c, 0x7f,\n",
+ " 0x81, 0x86, 0x8f, 0x63, 0x71, 0x6a, 0x63, 0x67, 0x7c, 0x8f, 0xa0, 0x68,\n",
+ " 0x86, 0x58, 0x5b, 0x87, 0x6a, 0x82, 0x89, 0x78, 0x9d, 0x8d, 0xaa, 0x82,\n",
+ " 0x6e, 0xa4, 0x6f, 0x6d, 0x70, 0x9f, 0x7f, 0x77, 0x41, 0xa5, 0x86, 0x61,\n",
+ " 0x2d, 0x99, 0xa9, 0x5f, 0x5a, 0xb3, 0x51, 0x70, 0x5a, 0xce, 0x77, 0x68,\n",
+ " 0x2c, 0xb8, 0x90, 0x44, 0x58, 0xb9, 0x74, 0x8e, 0x70, 0xb3, 0x9a, 0x75,\n",
+ " 0x6d, 0xc0, 0x9e, 0x8e, 0x8d, 0xa8, 0x7b, 0xa8, 0x4a, 0x89, 0x6e, 0x7f,\n",
+ " 0x5d, 0x6e, 0x46, 0x91, 0x6d, 0x81, 0x89, 0x3e, 0x35, 0x69, 0x44, 0xaf,\n",
+ " 0x99, 0x8d, 0x94, 0x54, 0x60, 0x5b, 0xaf, 0x97, 0x92, 0x4e, 0x80, 0xae,\n",
+ " 0x9e, 0x62, 0xa3, 0x77, 0x6e, 0x5d, 0x71, 0xa0, 0xa6, 0x59, 0x84, 0x5d,\n",
+ " 0x65, 0x4a, 0x69, 0xa1, 0xa1, 0x40, 0x75, 0x65, 0x6b, 0x68, 0x60, 0xb3,\n",
+ " 0x92, 0x27, 0x70, 0x67, 0x9b, 0x5e, 0x50, 0xaf, 0xae, 0x64, 0x7a, 0x6e,\n",
+ " 0x61, 0x94, 0x3b, 0x8f, 0x86, 0x7f, 0x98, 0x88, 0x7a, 0x7f, 0x61, 0x7b,\n",
+ " 0x64, 0x96, 0x96, 0x79, 0x5c, 0x96, 0x52, 0x92, 0x76, 0x7e, 0xc4, 0x60,\n",
+ " 0x6d, 0x7b, 0x41, 0x8c, 0x7b, 0x8e, 0x9a, 0x66, 0x79, 0x95, 0x67, 0x6a,\n",
+ " 0x7a, 0x9b, 0xa9, 0x85, 0x6d, 0x66, 0x55, 0x65, 0x76, 0x8b, 0x90, 0x86,\n",
+ " 0x88, 0x8b, 0x8f, 0x7e, 0x83, 0x7c, 0x75, 0x5f, 0x78, 0x96, 0x76, 0x47,\n",
+ " 0x54, 0x9c, 0x8d, 0x7d, 0x24, 0x9f, 0x79, 0x5c, 0x55, 0xb2, 0x3b, 0x67,\n",
+ " 0x4e, 0xd2, 0x90, 0x79, 0x3c, 0xc3, 0x8b, 0x4a, 0x7c, 0xd7, 0x70, 0x75,\n",
+ " 0x5b, 0xaf, 0xa8, 0x6b, 0x59, 0xc1, 0x6d, 0x5f, 0x5d, 0x96, 0x87, 0x9a,\n",
+ " 0x5d, 0x7f, 0x8e, 0x6d, 0x5c, 0x75, 0x3f, 0xb6, 0x8e, 0x81, 0x7b, 0x31,\n",
+ " 0x47, 0x67, 0x56, 0xb6, 0x90, 0x71, 0x89, 0x63, 0x61, 0x75, 0x8d, 0x8b,\n",
+ " 0x97, 0x62, 0x62, 0x85, 0x9c, 0x64, 0xb7, 0x61, 0x71, 0x3f, 0x6c, 0x8b,\n",
+ " 0xaa, 0x43, 0x82, 0x70, 0x52, 0x52, 0x80, 0xaa, 0x9e, 0x5d, 0x90, 0x69,\n",
+ " 0x8a, 0x77, 0x6d, 0x9f, 0x9e, 0x5f, 0x84, 0x61, 0x87, 0x70, 0x43, 0xab,\n",
+ " 0x97, 0x6e, 0x84, 0x6c, 0x5d, 0x82, 0x64, 0x85, 0x83, 0x7e, 0x82, 0x7c,\n",
+ " 0x7b, 0x91, 0x55, 0x7e, 0x77, 0x88, 0xba, 0x71, 0x6d, 0x7b, 0x71, 0x8a,\n",
+ " 0x7f, 0x84, 0xb5, 0x63, 0x4a, 0x9a, 0x3c, 0x70, 0x7a, 0x99, 0xa3, 0x50,\n",
+ " 0x84, 0x82, 0x56, 0x4c, 0x74, 0x8e, 0xa3, 0x77, 0x8f, 0x4e, 0x5f, 0x6d,\n",
+ " 0x97, 0x89, 0xa0, 0x6b, 0x7c, 0x8c, 0x85, 0x82, 0x8e, 0xa1, 0x89, 0x5b,\n",
+ " 0x7f, 0x8b, 0x8f, 0x5e, 0x74, 0x96, 0x8a, 0x7d, 0x15, 0x7b, 0x8f, 0x88,\n",
+ " 0x5f, 0xa7, 0x63, 0x5b, 0x39, 0xbd, 0x96, 0x56, 0x4c, 0xb4, 0x7b, 0x53,\n",
+ " 0x5a, 0xaf, 0x79, 0x7b, 0x5c, 0xa6, 0xaa, 0x74, 0x5f, 0xa0, 0x76, 0x9e,\n",
+ " 0x71, 0x9a, 0x60, 0xa4, 0x33, 0x87, 0x66, 0x66, 0x64, 0x7d, 0x6d, 0xac,\n",
+ " 0x9e, 0x8c, 0x78, 0x4f, 0x3d, 0x7b, 0x53, 0xb1, 0x97, 0x8a, 0x96, 0x6e,\n",
+ " 0x60, 0x4b, 0xa9, 0x9e, 0x93, 0x6e, 0x93, 0xb7, 0xae, 0x46, 0xb9, 0x60,\n",
+ " 0x72, 0x46, 0x80, 0x95, 0xb5, 0x57, 0x82, 0x53, 0x6e, 0x4e, 0x5b, 0xa2,\n",
+ " 0x9a, 0x3d, 0x8b, 0x6c, 0x84, 0x65, 0x69, 0xa1, 0x8c, 0x60, 0x83, 0x74,\n",
+ " 0x73, 0x53, 0x5d, 0x7e, 0x7f, 0x79, 0x6e, 0x81, 0x89, 0x8f, 0x51, 0x81,\n",
+ " 0x99, 0x97, 0x81, 0x8a, 0x87, 0x83, 0x43, 0x90, 0x89, 0x94, 0x93, 0x7a,\n",
+ " 0x66, 0x80, 0x82, 0x82, 0x79, 0x85, 0xb0, 0x6b, 0x87, 0x7b, 0x53, 0x89,\n",
+ " 0x79, 0x9d, 0xab, 0x6e, 0x82, 0x84, 0x50, 0x8f, 0x7e, 0x74, 0x90, 0x74,\n",
+ " 0x6e, 0x65, 0x84, 0x70, 0x82, 0x7a, 0x9e, 0x6d, 0x8f, 0x62, 0xb2, 0x84,\n",
+ " 0x78, 0x7e, 0x72, 0x5a, 0x7a, 0x85, 0x8c, 0x4b, 0x70, 0x99, 0x87, 0x78,\n",
+ " 0x26, 0x95, 0xb9, 0x77, 0x4d, 0xb6, 0x51, 0x6a, 0x41, 0xbf, 0x76, 0x68,\n",
+ " 0x56, 0xb6, 0x80, 0x53, 0x83, 0xaf, 0x87, 0x79, 0x79, 0xb4, 0x89, 0x7d,\n",
+ " 0x47, 0x9d, 0xa0, 0x86, 0x89, 0xc3, 0x6d, 0x99, 0x41, 0x89, 0x9a, 0x59,\n",
+ " 0x54, 0x83, 0x79, 0x9d, 0x7b, 0x73, 0x88, 0x4a, 0x42, 0x64, 0x7a, 0x9f,\n",
+ " 0x7b, 0x6e, 0x71, 0x7b, 0x6a, 0x61, 0xae, 0xa3, 0xa0, 0x68, 0x95, 0x9d,\n",
+ " 0x94, 0x49, 0x8b, 0x70, 0x8a, 0x5f, 0x49, 0xbb, 0xa7, 0x4a, 0xa1, 0x59,\n",
+ " 0x59, 0x59, 0x6d, 0xa0, 0x9f, 0x50, 0xa0, 0x7b, 0x75, 0x49, 0x5a, 0x8c,\n",
+ " 0x84, 0x68, 0x78, 0x57, 0x7a, 0x6e, 0x6b, 0x87, 0x9c, 0x7b, 0x84, 0x83,\n",
+ " 0x79, 0x7d, 0x5a, 0x77, 0x77, 0x6f, 0x6f, 0x7c, 0x8f, 0x83, 0x40, 0x62,\n",
+ " 0x6a, 0x87, 0xab, 0x74, 0x86, 0x96, 0x7a, 0x7d, 0x7b, 0x81, 0x9a, 0x65,\n",
+ " 0x60, 0x82, 0x61, 0x73, 0x71, 0x77, 0xa7, 0x79, 0x87, 0x8c, 0x4e, 0x72,\n",
+ " 0x8d, 0x89, 0x94, 0x6d, 0x75, 0x6d, 0x6e, 0x82, 0x7a, 0x8d, 0xa9, 0x77,\n",
+ " 0x77, 0x7c, 0x74, 0xa7, 0xb7, 0x67, 0x75, 0x67, 0x7e, 0x9f, 0x73, 0x60,\n",
+ " 0x6c, 0x95, 0x7f, 0x62, 0x31, 0x70, 0x85, 0x7a, 0x5f, 0xc0, 0x69, 0x66,\n",
+ " 0x71, 0xb0, 0x81, 0x5d, 0x48, 0xc9, 0x86, 0x39, 0x93, 0xa4, 0x8e, 0x7c,\n",
+ " 0x5e, 0xbb, 0x98, 0x5c, 0x74, 0x9c, 0x89, 0x6d, 0x74, 0xbd, 0x8e, 0x6e,\n",
+ " 0x5f, 0x9a, 0x6d, 0x70, 0x57, 0x9c, 0x58, 0xb7, 0x8e, 0x94, 0xa0, 0x3f,\n",
+ " 0x39, 0x75, 0x6f, 0xb4, 0xa2, 0x94, 0xa9, 0x70, 0x61, 0x8a, 0x70, 0x92,\n",
+ " 0xa7, 0x7f, 0x7f, 0x8d, 0x7a, 0x73, 0xa1, 0x5f, 0x8a, 0x4a, 0x65, 0xaa,\n",
+ " 0x92, 0x6e, 0x98, 0x51, 0x81, 0x47, 0x57, 0xb8, 0x89, 0x50, 0x8a, 0x6d,\n",
+ " 0x8b, 0x50, 0x8a, 0x86, 0x9b, 0x7d, 0x5b, 0x4a, 0x68, 0x74, 0x53, 0x9b,\n",
+ " 0x94, 0x74, 0x7c, 0x6f, 0x62, 0x86, 0x5b, 0x8f, 0x82, 0x96, 0x6e, 0x7c,\n",
+ " 0x80, 0x8f, 0x47, 0x5b, 0x70, 0x95, 0x97, 0x77, 0x8d, 0x8e, 0x69, 0x62,\n",
+ " 0x78, 0x8f, 0xbf, 0x5e, 0x76, 0xae, 0x4d, 0x84, 0x73, 0x76, 0xab, 0x6f,\n",
+ " 0x7f, 0x8c, 0x4b, 0x7d, 0x96, 0x7d, 0xb3, 0x55, 0x78, 0x8d, 0x76, 0x73,\n",
+ " 0x8d, 0x8e, 0x98, 0x6a, 0x91, 0x86, 0x6d, 0x8c, 0x7d, 0x93, 0x97, 0x56,\n",
+ " 0x79, 0x8f, 0xa3, 0x7f, 0x7e, 0x82, 0xa0, 0x63, 0x3d, 0x6b, 0x88, 0x5e,\n",
+ " 0x61, 0xc0, 0x45, 0x5f, 0x66, 0xb0, 0x6c, 0x6d, 0x29, 0xd5, 0x95, 0x3b,\n",
+ " 0x77, 0xaa, 0x62, 0x70, 0x63, 0xce, 0x8c, 0x6e, 0x56, 0xaa, 0x77, 0x6e,\n",
+ " 0x90, 0xcc, 0x6d, 0x7e, 0x41, 0x9f, 0x88, 0x4f, 0x5d, 0xb4, 0x4c, 0x9b,\n",
+ " 0x80, 0x97, 0x98, 0x59, 0x4c, 0x71, 0x53, 0xb4, 0x90, 0x97, 0x93, 0x90,\n",
+ " 0x46, 0x63, 0xa6, 0x87, 0x9d, 0x56, 0x7f, 0xab, 0x8e, 0x68, 0xc6, 0x5d,\n",
+ " 0x6e, 0x58, 0x4b, 0x85, 0xa1, 0x70, 0x8a, 0x60, 0x84, 0x44, 0x68, 0x8e,\n",
+ " 0x9b, 0x3a, 0x8c, 0x57, 0x91, 0x4c, 0x6b, 0x9c, 0xa7, 0x64, 0x82, 0x5f,\n",
+ " 0x68, 0x6d, 0x4d, 0xa1, 0x6c, 0x91, 0x6c, 0x6b, 0x64, 0x97, 0x86, 0x81,\n",
+ " 0x8d, 0x8e, 0x80, 0x72, 0x88, 0x96, 0x5d, 0x6e, 0x7c, 0x67, 0x97, 0x69,\n",
+ " 0x95, 0x93, 0x61, 0x8b, 0x9b, 0x7d, 0xc8, 0x6f, 0x85, 0x80, 0x67, 0x68,\n",
+ " 0x90, 0x6b, 0xcc, 0x7c, 0xa3, 0xa0, 0x58, 0x81, 0x7a, 0x8d, 0x9f, 0x65,\n",
+ " 0x81, 0x82, 0x78, 0x6b, 0x85, 0x7b, 0x9b, 0x69, 0x86, 0x6c, 0x83, 0x6c,\n",
+ " 0x8e, 0x59, 0xab, 0x56, 0x7c, 0x7f, 0x7b, 0x84, 0x71, 0x63, 0x7d, 0x73,\n",
+ " 0x60, 0x8b, 0x7a, 0x7b, 0x5e, 0xbb, 0x4b, 0x40, 0x30, 0xcc, 0x80, 0x65,\n",
+ " 0x6c, 0xb7, 0x80, 0x35, 0x7d, 0xa3, 0x5c, 0x6c, 0x49, 0xa6, 0x9b, 0x7b,\n",
+ " 0x53, 0xba, 0x62, 0x76, 0x78, 0xa0, 0x72, 0x80, 0x78, 0x93, 0x87, 0x62,\n",
+ " 0x64, 0x84, 0x6f, 0xa1, 0x70, 0x90, 0x9a, 0x6b, 0x42, 0x55, 0x6d, 0xc5,\n",
+ " 0xa6, 0x8a, 0x79, 0x64, 0x4c, 0x72, 0x7b, 0xa9, 0xa3, 0x70, 0x84, 0x8f,\n",
+ " 0x63, 0x7a, 0x9c, 0x4e, 0x5a, 0x76, 0x91, 0x67, 0xaf, 0x76, 0xbf, 0x46,\n",
+ " 0x62, 0x3f, 0x7d, 0xa7, 0x8d, 0x62, 0x90, 0x5b, 0x9a, 0x44, 0x51, 0x80,\n",
+ " 0xa6, 0x7e, 0x8d, 0x6a, 0x73, 0x65, 0x72, 0x82, 0x99, 0xb4, 0x6a, 0x75,\n",
+ " 0x85, 0x90, 0x47, 0x62, 0x9e, 0x95, 0x94, 0x78, 0x89, 0x74, 0x5d, 0xa3,\n",
+ " 0x7f, 0x9d, 0x7d, 0x63, 0x96, 0x86, 0x8d, 0xa2, 0x95, 0xab, 0xae, 0x5d,\n",
+ " 0x93, 0x8d, 0x3d, 0x76, 0x9e, 0x9c, 0xc4, 0x71, 0x7d, 0xa3, 0x75, 0x7e,\n",
+ " 0x6d, 0x9d, 0xa3, 0x7f, 0x94, 0x89, 0x47, 0x71, 0x8b, 0x95, 0xb1, 0x72,\n",
+ " 0x90, 0x53, 0x7e, 0x8f, 0x8c, 0x90, 0xa1, 0x4d, 0x59, 0x62, 0x73, 0xa0,\n",
+ " 0x69, 0x88, 0x86, 0x71, 0x60, 0x3b, 0x81, 0x57, 0x7d, 0x86, 0x58, 0x63,\n",
+ " 0x7d, 0x98, 0x74, 0x67, 0x5d, 0xb0, 0x67, 0x45, 0x9b, 0xa9, 0x94, 0x68,\n",
+ " 0x43, 0x8b, 0x85, 0x56, 0x63, 0x96, 0x87, 0x78, 0x88, 0xbf, 0x92, 0x8d,\n",
+ " 0x60, 0xa8, 0x7e, 0x7e, 0x78, 0x80, 0x66, 0x92, 0x6e, 0x97, 0xab, 0x7f,\n",
+ " 0x4f, 0x65, 0x59, 0xb0, 0x9b, 0x6b, 0x9f, 0x70, 0x6f, 0x5c, 0xac, 0x95,\n",
+ " 0xa3, 0x54, 0x8e, 0xa9, 0x9e, 0x8c, 0xa5, 0x66, 0x5f, 0x5b, 0x6c, 0x83,\n",
+ " 0x90, 0x73, 0x85, 0x64, 0x61, 0x51, 0x4a, 0x63, 0xa1, 0x96, 0x7e, 0x4e,\n",
+ " 0x87, 0x60, 0x68, 0xb5, 0x9a, 0x8d, 0x75, 0x4e, 0x8a, 0x7a, 0x5f, 0x9f,\n",
+ " 0x74, 0x80, 0x69, 0x6d, 0x73, 0x92, 0x79, 0x7e, 0x85, 0x68, 0x83, 0x9d,\n",
+ " 0xb6, 0x9d, 0x6e, 0x8f, 0x78, 0x91, 0xaf, 0x8f, 0xa0, 0x9d, 0x73, 0x55,\n",
+ " 0x91, 0x8f, 0xb2, 0x76, 0x97, 0xab, 0x63, 0x63, 0x68, 0x7b, 0xab, 0x5c,\n",
+ " 0x77, 0xae, 0x4c, 0x72, 0x6e, 0x93, 0xb8, 0x51, 0x79, 0x84, 0x7d, 0x6b,\n",
+ " 0x7f, 0x8a, 0xba, 0x68, 0x7a, 0x43, 0x9a, 0x8d, 0x77, 0x8a, 0x6d, 0x56,\n",
+ " 0x79, 0x41, 0x7a, 0x4b, 0x81, 0x7a, 0x5c, 0x68, 0x58, 0x36, 0x6f, 0x6f,\n",
+ " 0x9f, 0xa6, 0x5f, 0x60, 0x4e, 0x67, 0x70, 0x4c, 0x69, 0x69, 0x94, 0x63,\n",
+ " 0x6d, 0x7b, 0x88, 0x9e, 0x6d, 0x98, 0x69, 0x68, 0x88, 0x80, 0x80, 0x7a,\n",
+ " 0x8e, 0x78, 0x5e, 0x8d, 0x7e, 0x91, 0x76, 0x64, 0x7e, 0x7f, 0x4e, 0xc9,\n",
+ " 0x79, 0x8f, 0x9c, 0x82, 0x3d, 0x62, 0x63, 0xc3, 0xb8, 0x7b, 0x72, 0x7b,\n",
+ " 0x50, 0x56, 0x95, 0x72, 0x8f, 0x6b, 0x90, 0x9d, 0x76, 0xa4, 0xa5, 0x79,\n",
+ " 0x54, 0x4f, 0x59, 0x85, 0xc5, 0x92, 0x97, 0x4d, 0x6f, 0x69, 0x77, 0x7f,\n",
+ " 0x71, 0x7c, 0x87, 0x59, 0x98, 0x61, 0x80, 0x81, 0x88, 0x6b, 0x6d, 0x7f,\n",
+ " 0x7f, 0x77, 0x60, 0xa2, 0x96, 0x73, 0x69, 0x86, 0x83, 0x8d, 0x60, 0x66,\n",
+ " 0x88, 0x8c, 0x93, 0x67, 0x98, 0x82, 0x7e, 0x91, 0x99, 0x59, 0x8e, 0x6e,\n",
+ " 0x90, 0xa1, 0x62, 0x8a, 0x98, 0x7b, 0xc8, 0x67, 0x85, 0x8d, 0x6c, 0xa1,\n",
+ " 0xa1, 0x92, 0xd0, 0x49, 0x85, 0x76, 0x89, 0x75, 0x88, 0x83, 0xa3, 0x77,\n",
+ " 0x85, 0x68, 0x82, 0x83, 0x7f, 0x79, 0xae, 0x85, 0x76, 0x84, 0x80, 0x9a,\n",
+ " 0x9d, 0x7b, 0x83, 0x90, 0x79, 0x88, 0x79, 0x9a, 0x93, 0x6c, 0x69, 0x79,\n",
+ " 0x5f, 0x90, 0x81, 0x7b, 0x87, 0x9d, 0x86, 0x82, 0x7a, 0x77, 0x71, 0x85,\n",
+ " 0x8b, 0x99, 0x8f, 0x7b, 0x58, 0x98, 0x84, 0x6e, 0x9a, 0xa1, 0x7a, 0x8c,\n",
+ " 0x77, 0xa8, 0x86, 0x93, 0x7b, 0x90, 0x79, 0x8a, 0x85, 0x8f, 0x84, 0x97,\n",
+ " 0x73, 0x83, 0x7b, 0x76, 0x8e, 0xa1, 0x89, 0x8a, 0x83, 0x9c, 0x65, 0x68,\n",
+ " 0x7b, 0x89, 0x92, 0x84, 0x6d, 0x90, 0x61, 0x78, 0x98, 0x8c, 0x8d, 0x87,\n",
+ " 0xa0, 0x99, 0x79, 0x7b, 0x69, 0xa4, 0x7a, 0x8d, 0x73, 0x71, 0x70, 0x80,\n",
+ " 0x82, 0x77, 0x81, 0x67, 0x75, 0x97, 0x71, 0x73, 0x85, 0x6d, 0x8e, 0x86,\n",
+ " 0x6e, 0x80, 0x86, 0x9e, 0x6f, 0x70, 0x67, 0x59, 0x65, 0x89, 0x67, 0x8b,\n",
+ " 0x7d, 0x68, 0x69, 0x7a, 0x5b, 0x7e, 0x87, 0xa1, 0x92, 0x7b, 0x64, 0x7e,\n",
+ " 0x76, 0x72, 0x71, 0xab, 0x7c, 0x83, 0x6f, 0xa1, 0x86, 0x76, 0x71, 0x6f,\n",
+ " 0x91, 0x77, 0x6c, 0x71, 0x92, 0x78, 0x70, 0x7f, 0x6e, 0x65, 0x77, 0x93,\n",
+ " 0x7e, 0x6c, 0x85, 0x9d, 0x78, 0x8b, 0x7c, 0x5f, 0x94, 0x86, 0x7c, 0x7f,\n",
+ " 0x83, 0x6e, 0x72, 0x9e, 0x6e, 0x6b, 0x8d, 0x91, 0x97, 0x8b, 0x7b, 0x72,\n",
+ " 0x86, 0x75, 0x7f, 0x96, 0x7d, 0x81, 0xa1, 0x55, 0xa6, 0x88, 0x96, 0x87,\n",
+ " 0x93, 0x68, 0x89, 0x72, 0x6f, 0x9c, 0x75, 0x7c, 0x79, 0x6c, 0x74, 0x84,\n",
+ " 0x7d, 0xa4, 0x86, 0x84, 0x84, 0x8d, 0x63, 0x7a, 0x63, 0xbc, 0x7e, 0x93,\n",
+ " 0x80, 0x8d, 0x71, 0x7a, 0x5f, 0x8c, 0x74, 0x96, 0x7e, 0x9b, 0x9d, 0x8d,\n",
+ " 0x5b, 0xa4, 0x71, 0x5e, 0x83, 0x78, 0x86, 0x7f, 0x70, 0x99, 0x87, 0x85,\n",
+ " 0x8e, 0x81, 0x93, 0x80, 0x89, 0xa0, 0x7a, 0x77, 0x8e, 0x73, 0x5f, 0x80,\n",
+ " 0x6d, 0x87, 0x5b, 0x7a, 0x85, 0x7c, 0x85, 0x63, 0x61, 0x9d, 0x6f, 0x68,\n",
+ " 0x77, 0x86, 0x61, 0x6d, 0x84, 0x98, 0x7c, 0x78, 0x69, 0x84, 0x91, 0x6d,\n",
+ " 0x81, 0xa1, 0x6c, 0x62, 0x95, 0x6d, 0x86, 0x8b, 0x95, 0x8f, 0x5e, 0x86,\n",
+ " 0x73, 0xa1, 0x83, 0x58, 0x5f, 0x8e, 0x76, 0x79, 0x9e, 0x92, 0x7c, 0x7b,\n",
+ " 0x81, 0x8b, 0x83, 0x7b, 0x78, 0x75, 0x70, 0x83, 0x70, 0x5a, 0x6a, 0x59,\n",
+ " 0xa3, 0x82, 0x7a, 0x91, 0x8b, 0x6e, 0x82, 0x8e, 0x70, 0x73, 0x91, 0x76,\n",
+ " 0xa5, 0x7f, 0x70, 0x81, 0x6f, 0x85, 0x94, 0xa6, 0x8c, 0x50, 0x76, 0x6e,\n",
+ " 0x64, 0x95, 0xa0, 0x64, 0x6c, 0x68, 0x8e, 0x8b, 0xa1, 0x7d, 0xa0, 0x7f,\n",
+ " 0x76, 0x8b, 0x7b, 0x93, 0x7b, 0x6e, 0x7e, 0x64, 0x8a, 0xa7, 0x78, 0x64,\n",
+ " 0x93, 0x67, 0x7d, 0x68, 0x5c, 0xa0, 0x76, 0x98, 0xaf, 0x80, 0x55, 0x96,\n",
+ " 0x97, 0x9c, 0x78, 0x75, 0x87, 0x85, 0x77, 0x77, 0x62, 0x93, 0x76, 0x68,\n",
+ " 0xa0, 0x80, 0x81, 0x7f, 0x9a, 0x68, 0x74, 0x69, 0x94, 0x77, 0x77, 0x72,\n",
+ " 0x90, 0x9a, 0x6f, 0x95, 0x89, 0x6b, 0x6b, 0x94, 0x7e, 0x9c, 0x6f, 0x67,\n",
+ " 0x8f, 0x82, 0x80, 0x92, 0x76, 0x80, 0x65, 0x9b, 0x6a, 0x7c, 0x75, 0x5a,\n",
+ " 0x87, 0xa1, 0x69, 0x7a, 0x79, 0x9e, 0x9a, 0x58, 0x81, 0x92, 0x72, 0x67,\n",
+ " 0x90, 0x80, 0x82, 0x61, 0x9f, 0x9e, 0x6a, 0x8d, 0x8d, 0x8a, 0x73, 0x81,\n",
+ " 0x68, 0x7f, 0x5b, 0x59, 0x98, 0x89, 0x71, 0x72, 0x58, 0x7b, 0x94, 0x5d,\n",
+ " 0xa9, 0x8b, 0x72, 0x7b, 0x65, 0x73, 0x5b, 0x8b, 0x7d, 0x86, 0x6e, 0x8c,\n",
+ " 0x66, 0x6f, 0x6b, 0x8b, 0x71, 0x80, 0x7f, 0x70, 0x70, 0x88, 0x70, 0x7e,\n",
+ " 0x84, 0x89, 0x7f, 0x81, 0x87, 0x77, 0x71, 0x88, 0x7f, 0x8f, 0x5e, 0x80,\n",
+ " 0x5d, 0xa1, 0x89, 0x77, 0x93, 0x8e, 0x55, 0x64, 0x88, 0x9a, 0x8b, 0x80,\n",
+ " 0x77, 0x6f, 0x91, 0x83, 0x6b, 0x9b, 0x85, 0x5c, 0x57, 0x7e, 0xa9, 0x63,\n",
+ " 0x83, 0xaa, 0x7c, 0xa1, 0x91, 0x5f, 0x68, 0x76, 0x7a, 0x97, 0x96, 0x84,\n",
+ " 0xca, 0x8d, 0x8c, 0x8b, 0x71, 0x81, 0x88, 0x92, 0xaa, 0x74, 0x49, 0x7a,\n",
+ " 0x90, 0x93, 0x7a, 0x61, 0x8c, 0x66, 0x71, 0xa0, 0xab, 0x7d, 0x86, 0x6c,\n",
+ " 0x9f, 0x77, 0x67, 0x6a, 0x89, 0x89, 0x88, 0x70, 0xad, 0x88, 0x69, 0x84,\n",
+ " 0x70, 0x8f, 0x79, 0x7c, 0x66, 0xa6, 0x71, 0x8d, 0x77, 0x99, 0x69, 0x76,\n",
+ " 0x79, 0x7d, 0x9c, 0x6f, 0x64, 0x8b, 0x70, 0x82, 0x69, 0xa4, 0x65, 0x6e,\n",
+ " 0x7f, 0x9e, 0x7e, 0x84, 0x8c, 0x9c, 0x6c, 0x5b, 0x6e, 0xa7, 0x6d, 0x7a,\n",
+ " 0x92, 0x78, 0x9a, 0x6f, 0x81, 0x91, 0x71, 0x7d, 0x6b, 0x99, 0x6b, 0x92,\n",
+ " 0x5e, 0x7e, 0x64, 0x95, 0x78, 0x90, 0x6f, 0x68, 0x8a, 0x85, 0x6f, 0x88,\n",
+ " 0x64, 0x66, 0x7f, 0x78, 0x7c, 0x95, 0x66, 0x6c, 0x76, 0x6a, 0x9b, 0x8f,\n",
+ " 0x9d, 0x78, 0x86, 0x95, 0x73, 0x66, 0x6d, 0x71, 0x8b, 0x7f, 0x6f, 0x70,\n",
+ " 0x64, 0x94, 0xa0, 0x83, 0x6b, 0x6d, 0x85, 0x89, 0x68, 0x92, 0x8e, 0x51,\n",
+ " 0x81, 0x85, 0x86, 0x6e, 0x83, 0x85, 0x8a, 0x5e, 0x68, 0xbf, 0xc4, 0xa5,\n",
+ " 0x8b, 0x67, 0x86, 0x59, 0x85, 0x9e, 0x96, 0x67, 0x82, 0x7c, 0x6c, 0x80,\n",
+ " 0x84, 0xae, 0x9d, 0x80, 0xc2, 0x58, 0x5d, 0x95, 0x85, 0x8b, 0x7f, 0x5d,\n",
+ " 0xc7, 0x75, 0x75, 0x87, 0xa2, 0x8c, 0x62, 0x71, 0x9c, 0x61, 0x7f, 0x9c,\n",
+ " 0xca, 0x8d, 0x89, 0x6e, 0x7c, 0x71, 0x81, 0x99, 0x95, 0xa4, 0x76, 0x6f,\n",
+ " 0x64, 0x7b, 0x6c, 0x72, 0x8b, 0x83, 0x70, 0x70, 0x8b, 0xa4, 0x69, 0x76,\n",
+ " 0x6e, 0x8d, 0x7a, 0x80, 0x8f, 0x9e, 0x73, 0x4b, 0x75, 0x78, 0x77, 0x7b,\n",
+ " 0x8e, 0x92, 0x88, 0x49, 0x54, 0x9f, 0x7a, 0x7f, 0x68, 0x9f, 0x7f, 0x57,\n",
+ " 0x6b, 0xad, 0x85, 0x6f, 0x81, 0xa1, 0x96, 0x6f, 0x73, 0x8d, 0x5e, 0x65,\n",
+ " 0x7a, 0x8c, 0x7c, 0x6a, 0x7e, 0x7a, 0x6a, 0x97, 0x59, 0x86, 0x62, 0x77,\n",
+ " 0x70, 0x7a, 0x68, 0x62, 0x68, 0x86, 0x7e, 0x76, 0x9a, 0x7f, 0x6c, 0x7e,\n",
+ " 0x8a, 0x76, 0x65, 0x8f, 0x7d, 0x65, 0x76, 0xa4, 0x95, 0x62, 0x78, 0x97,\n",
+ " 0x7a, 0x6e, 0x7a, 0x7a, 0x7e, 0x91, 0x8c, 0x8a, 0x91, 0x82, 0x89, 0x6d,\n",
+ " 0x87, 0x90, 0x69, 0x71, 0x96, 0xa6, 0x7c, 0x7c, 0xa8, 0xa8, 0x62, 0x77,\n",
+ " 0x76, 0x99, 0xdd, 0x76, 0x8a, 0x5c, 0x86, 0x6a, 0x69, 0x9c, 0xa5, 0x7d,\n",
+ " 0x78, 0x6a, 0x88, 0x77, 0x77, 0xae, 0x8a, 0x99, 0xcb, 0x85, 0x59, 0x84,\n",
+ " 0x7b, 0x97, 0x8a, 0x82, 0xc5, 0x65, 0x8c, 0x93, 0xc3, 0x8c, 0x87, 0x64,\n",
+ " 0x91, 0x41, 0x70, 0xa8, 0xd1, 0x8b, 0x82, 0x71, 0x9c, 0x71, 0x4e, 0x86,\n",
+ " 0x98, 0x86, 0x7f, 0x7e, 0x69, 0x99, 0x79, 0x78, 0x77, 0xb3, 0x6b, 0x80,\n",
+ " 0x84, 0x8b, 0x56, 0x73, 0x84, 0x95, 0x82, 0x94, 0x5b, 0x92, 0x83, 0x46,\n",
+ " 0x66, 0x89, 0x6d, 0x61, 0x99, 0xa6, 0x99, 0x3f, 0x6c, 0xab, 0x5d, 0x5f,\n",
+ " 0x6c, 0x8e, 0x6b, 0x4a, 0x72, 0xb6, 0x6c, 0x75, 0x78, 0xa6, 0x6f, 0x5b,\n",
+ " 0x56, 0x8b, 0x57, 0x74, 0x8f, 0xab, 0x53, 0x56, 0x5d, 0x63, 0x63, 0x8b,\n",
+ " 0x65, 0x78, 0x71, 0x67, 0x7a, 0x62, 0x8d, 0x78, 0x99, 0x76, 0x94, 0x7a,\n",
+ " 0xa3, 0x70, 0x55, 0x87, 0x7e, 0x7c, 0x57, 0x57, 0x6e, 0x79, 0x94, 0x8f,\n",
+ " 0x86, 0x80, 0x90, 0x7d, 0x7d, 0x7f, 0x7f, 0x68, 0x41, 0x86, 0x8c, 0x6f,\n",
+ " 0x8a, 0x7f, 0x87, 0x8a, 0x7e, 0x7f, 0x5d, 0x71, 0x91, 0x81, 0x93, 0x71,\n",
+ " 0x91, 0xc6, 0x70, 0x4a, 0x74, 0xa8, 0xf3, 0x72, 0xa7, 0x80, 0x7e, 0x41,\n",
+ " 0x84, 0xa3, 0xb6, 0x94, 0xba, 0x84, 0x70, 0x74, 0x71, 0xac, 0x9f, 0x9d,\n",
+ " 0xe4, 0x67, 0x6a, 0x87, 0x92, 0x8e, 0x92, 0x82, 0xdb, 0x5e, 0x9b, 0x90,\n",
+ " 0xd5, 0x87, 0x8d, 0x7c, 0x9c, 0x3c, 0x6c, 0xab, 0xc2, 0x86, 0x83, 0x79,\n",
+ " 0x6c, 0x61, 0x51, 0xa9, 0x99, 0x79, 0x72, 0x80, 0x6f, 0x85, 0x57, 0x6c,\n",
+ " 0x81, 0x86, 0x6e, 0x88, 0x87, 0x8d, 0x8e, 0x81, 0x67, 0x88, 0x62, 0x99,\n",
+ " 0x87, 0xab, 0x8f, 0x57, 0x60, 0x77, 0x64, 0x81, 0x96, 0xa3, 0x81, 0x3d,\n",
+ " 0x4e, 0xb9, 0x57, 0x6e, 0x99, 0xad, 0x6a, 0x3e, 0x74, 0x96, 0x7e, 0x79,\n",
+ " 0x65, 0xa4, 0x7c, 0x6a, 0x53, 0x87, 0x56, 0x6f, 0x5e, 0x97, 0x85, 0x42,\n",
+ " 0x56, 0x6b, 0x67, 0x78, 0x7d, 0xa6, 0x7c, 0x7c, 0x7d, 0x78, 0x7b, 0x84,\n",
+ " 0x99, 0x7b, 0x89, 0x71, 0x76, 0x8b, 0x76, 0x73, 0x7d, 0x83, 0x56, 0x4f,\n",
+ " 0x86, 0x72, 0x83, 0x88, 0x6a, 0x93, 0x69, 0x90, 0x6c, 0x73, 0x6f, 0x63,\n",
+ " 0x55, 0x88, 0x6b, 0x88, 0x7c, 0x86, 0x87, 0x7b, 0x6c, 0x7e, 0x60, 0x57,\n",
+ " 0xa8, 0x81, 0xa3, 0x72, 0xba, 0xbf, 0x66, 0x65, 0x70, 0xb9, 0xe4, 0x78,\n",
+ " 0x99, 0x67, 0x8c, 0x72, 0x88, 0x96, 0xb5, 0x72, 0x8a, 0x66, 0x81, 0x39,\n",
+ " 0x85, 0x93, 0xa0, 0x9c, 0xdf, 0x74, 0x8a, 0x6d, 0x93, 0xa1, 0x8c, 0x7a,\n",
+ " 0xb5, 0x4b, 0x89, 0xae, 0xba, 0x9c, 0x96, 0x9a, 0xb4, 0x33, 0x5a, 0xb1,\n",
+ " 0xcd, 0x88, 0x84, 0x63, 0x8c, 0x5e, 0x71, 0x6d, 0xa7, 0x8a, 0x62, 0x85,\n",
+ " 0x77, 0x75, 0x62, 0x79, 0x96, 0x73, 0x4f, 0x7d, 0x93, 0x8a, 0x88, 0x7e,\n",
+ " 0x59, 0x6c, 0x7f, 0x87, 0x6f, 0x91, 0x88, 0x59, 0x6d, 0x83, 0x70, 0x7c,\n",
+ " 0x7f, 0x8d, 0x7f, 0x26, 0x41, 0xcf, 0x6b, 0x6e, 0x75, 0xa3, 0x90, 0x5e,\n",
+ " 0x3a, 0x94, 0x61, 0x9a, 0x6f, 0x9f, 0x69, 0x7d, 0x55, 0x8c, 0x60, 0x7c,\n",
+ " 0x93, 0x85, 0x85, 0x4b, 0x54, 0x71, 0x60, 0x8a, 0x6d, 0x8c, 0x9c, 0x7e,\n",
+ " 0x5b, 0x79, 0x74, 0x7b, 0x7b, 0x9d, 0x5b, 0x65, 0x81, 0x82, 0x66, 0x89,\n",
+ " 0x82, 0x72, 0x77, 0x78, 0x75, 0x76, 0x6b, 0x74, 0x89, 0x73, 0x6c, 0x6b,\n",
+ " 0x77, 0x7e, 0x67, 0x84, 0x41, 0x90, 0x58, 0x87, 0x98, 0x60, 0x96, 0x81,\n",
+ " 0x6b, 0x74, 0x7d, 0x56, 0x72, 0x71, 0x9a, 0x7d, 0xc5, 0xd0, 0x88, 0x6e,\n",
+ " 0x4d, 0xbe, 0xef, 0x8a, 0xa7, 0x92, 0x82, 0x67, 0x7f, 0x91, 0xc5, 0x7d,\n",
+ " 0xad, 0x77, 0x6b, 0x4e, 0x8e, 0x99, 0x9b, 0x8e, 0xc7, 0x7f, 0x8a, 0x8e,\n",
+ " 0x8f, 0x87, 0x9c, 0x75, 0xb0, 0x53, 0x75, 0x97, 0xc7, 0x98, 0xa4, 0xa4,\n",
+ " 0x80, 0x41, 0x79, 0xc3, 0xdb, 0x86, 0x9d, 0x75, 0x7f, 0x67, 0x7a, 0x96,\n",
+ " 0xc3, 0x83, 0x54, 0x8e, 0x6f, 0xa8, 0x7c, 0x65, 0x78, 0x7e, 0x59, 0xa3,\n",
+ " 0x8a, 0x97, 0x8b, 0x82, 0x5e, 0x66, 0x82, 0x9b, 0x9e, 0x9f, 0x70, 0x49,\n",
+ " 0x55, 0x88, 0x8a, 0x7e, 0x90, 0xa7, 0x6b, 0x3b, 0x28, 0xc0, 0x63, 0x7e,\n",
+ " 0x60, 0x90, 0x7c, 0x3f, 0x54, 0x9c, 0x7d, 0x8a, 0x6a, 0xa9, 0x6f, 0x61,\n",
+ " 0x76, 0x86, 0x64, 0x88, 0x72, 0xa5, 0x6b, 0x4d, 0x56, 0x6c, 0x52, 0xa1,\n",
+ " 0x84, 0x69, 0x69, 0x5b, 0x71, 0x84, 0x76, 0x9b, 0x92, 0x70, 0x86, 0x8b,\n",
+ " 0x71, 0x68, 0x56, 0x92, 0x76, 0x8f, 0x8f, 0x72, 0x5a, 0x77, 0x6f, 0x92,\n",
+ " 0x72, 0x72, 0x5e, 0x7a, 0x70, 0x73, 0x60, 0x7d, 0x5a, 0x93, 0x7f, 0x6b,\n",
+ " 0x89, 0x6b, 0xa1, 0x85, 0x5c, 0x8d, 0x76, 0x7c, 0x6f, 0x73, 0x96, 0x6d,\n",
+ " 0xbb, 0xad, 0x53, 0x53, 0x5f, 0x9a, 0xe2, 0x8d, 0xa7, 0x6d, 0x8a, 0x5b,\n",
+ " 0x85, 0x9c, 0xb4, 0x7b, 0xb3, 0x52, 0x75, 0x7f, 0x7a, 0x8c, 0x91, 0x7e,\n",
+ " 0xca, 0x5f, 0x64, 0x71, 0x85, 0x9a, 0x91, 0x72, 0xbd, 0x6e, 0x9b, 0x81,\n",
+ " 0x8f, 0xa8, 0xac, 0x7d, 0xb4, 0x5f, 0x45, 0xc5, 0xc8, 0x7a, 0x93, 0x8e,\n",
+ " 0x7b, 0x41, 0x69, 0x94, 0x8b, 0x76, 0x59, 0x81, 0x73, 0x92, 0x8e, 0x63,\n",
+ " 0x8e, 0x74, 0x33, 0xa5, 0x9c, 0xa2, 0x88, 0x48, 0x5d, 0x8c, 0x7d, 0xa6,\n",
+ " 0x68, 0x9a, 0x6f, 0x58, 0x6c, 0x8f, 0x77, 0x65, 0x97, 0x9d, 0x7a, 0x37,\n",
+ " 0x59, 0xab, 0x6e, 0x8f, 0x7a, 0xae, 0x65, 0x3e, 0x46, 0xa9, 0x82, 0x82,\n",
+ " 0x9c, 0x9d, 0x62, 0x79, 0x66, 0x7f, 0x5e, 0x88, 0x9e, 0x8f, 0x84, 0x71,\n",
+ " 0x5d, 0x6d, 0x70, 0xa0, 0x69, 0x92, 0x7f, 0x70, 0x66, 0x6f, 0x75, 0x8c,\n",
+ " 0x96, 0x7a, 0x85, 0x6a, 0x5a, 0x7c, 0x72, 0x8a, 0x8d, 0x7b, 0x8b, 0x5c,\n",
+ " 0x76, 0x69, 0x70, 0x7f, 0x74, 0xa1, 0x71, 0x91, 0x5a, 0x8c, 0x6e, 0x83,\n",
+ " 0x52, 0x78, 0x71, 0x6d, 0xa9, 0x63, 0x9d, 0x81, 0x52, 0x9e, 0x5d, 0x60,\n",
+ " 0x76, 0x93, 0x97, 0x67, 0xce, 0xc1, 0x75, 0x5e, 0x5f, 0x8c, 0xea, 0x76,\n",
+ " 0xad, 0x7a, 0x7d, 0x62, 0x85, 0x92, 0xd0, 0x6a, 0xbc, 0x53, 0x55, 0x5c,\n",
+ " 0x6d, 0x89, 0x9e, 0x71, 0xd2, 0x8b, 0x64, 0x61, 0x85, 0x9a, 0x77, 0x75,\n",
+ " 0xb9, 0x67, 0x8a, 0xac, 0x90, 0x8a, 0xb4, 0x91, 0xbb, 0x58, 0x94, 0xaf,\n",
+ " 0xb2, 0x76, 0xa2, 0x71, 0x95, 0x5e, 0x73, 0xa5, 0x92, 0x8c, 0x52, 0x96,\n",
+ " 0x53, 0x95, 0x84, 0x91, 0x93, 0x7a, 0x40, 0x88, 0xab, 0xa5, 0x63, 0x70,\n",
+ " 0x66, 0x88, 0x7e, 0x92, 0x89, 0x84, 0x78, 0x57, 0x3d, 0x8d, 0x84, 0x77,\n",
+ " 0x9b, 0x87, 0x5e, 0x4e, 0x42, 0xa0, 0x76, 0x8a, 0x77, 0x90, 0x83, 0x4c,\n",
+ " 0x42, 0x9b, 0x75, 0x7a, 0x88, 0x94, 0x98, 0x69, 0x4c, 0xa2, 0x6b, 0x7b,\n",
+ " 0x6e, 0x9b, 0x5d, 0x5f, 0x53, 0x6a, 0x63, 0x95, 0x69, 0x8a, 0x61, 0x75,\n",
+ " 0x6c, 0x7a, 0x58, 0x89, 0x84, 0x8f, 0x6b, 0x5a, 0x71, 0x6f, 0x59, 0x89,\n",
+ " 0x7d, 0x87, 0x5f, 0x77, 0x4b, 0x61, 0x77, 0x92, 0x67, 0x8e, 0x5c, 0x6f,\n",
+ " 0x5b, 0x77, 0x76, 0x6b, 0x44, 0x9d, 0x9f, 0x7f, 0x8b, 0x94, 0x9e, 0x7c,\n",
+ " 0x62, 0x94, 0x60, 0x55, 0x77, 0x8f, 0xa6, 0x62, 0xb5, 0xb2, 0x3c, 0x61,\n",
+ " 0x5c, 0x99, 0xeb, 0x5b, 0x90, 0x6c, 0x7f, 0x5f, 0x75, 0xa6, 0xcf, 0x77,\n",
+ " 0x98, 0x5d, 0x75, 0x69, 0x7f, 0x8a, 0xa7, 0x73, 0xc8, 0x74, 0x70, 0x82,\n",
+ " 0x76, 0x8f, 0xa2, 0x7a, 0xa4, 0x7a, 0x66, 0x81, 0x9b, 0x8f, 0x9e, 0x8b,\n",
+ " 0xa1, 0x51, 0x7b, 0xba, 0xc8, 0x90, 0xab, 0x92, 0x72, 0x57, 0x5b, 0xa3,\n",
+ " 0xb0, 0x7f, 0x4c, 0x7d, 0x5f, 0x8e, 0x6c, 0x7d, 0x71, 0x7e, 0x4e, 0x87,\n",
+ " 0xb7, 0x97, 0x7a, 0x4c, 0x5f, 0x72, 0x78, 0x84, 0x82, 0x7e, 0x63, 0x65,\n",
+ " 0x68, 0x78, 0x73, 0x85, 0x90, 0x99, 0x80, 0x57, 0x42, 0x8b, 0x8a, 0x77,\n",
+ " 0x71, 0x97, 0x6d, 0x44, 0x41, 0x8f, 0x78, 0x7d, 0x95, 0x81, 0x95, 0x5f,\n",
+ " 0x64, 0x87, 0x66, 0x80, 0x89, 0x9a, 0x61, 0x4d, 0x68, 0x7b, 0x72, 0x73,\n",
+ " 0x85, 0x92, 0x77, 0x7d, 0x73, 0x77, 0x54, 0x7a, 0x77, 0x7d, 0x7d, 0x7a,\n",
+ " 0x6e, 0x8e, 0x4f, 0x7d, 0x80, 0x9a, 0x79, 0x8b, 0x7b, 0x68, 0x6e, 0x86,\n",
+ " 0x7f, 0x93, 0x7a, 0x76, 0x72, 0x85, 0x6a, 0x7b, 0x57, 0x84, 0x96, 0x9a,\n",
+ " 0x8f, 0x91, 0x9b, 0x72, 0x73, 0x91, 0x53, 0x66, 0x76, 0x80, 0xae, 0x63,\n",
+ " 0xbf, 0x99, 0x5e, 0x77, 0x73, 0x9c, 0xd8, 0x74, 0xa7, 0x79, 0x52, 0x64,\n",
+ " 0x82, 0x95, 0xc7, 0x4f, 0xa8, 0x4f, 0x6d, 0x42, 0x7c, 0x89, 0xab, 0x83,\n",
+ " 0xc0, 0x82, 0x6a, 0x5f, 0x83, 0x92, 0xa8, 0x76, 0xc1, 0x77, 0x6e, 0x7b,\n",
+ " 0xa3, 0x9b, 0xaf, 0x87, 0xab, 0x60, 0x8d, 0xc2, 0xd2, 0x83, 0xb2, 0x78,\n",
+ " 0x8d, 0x39, 0x57, 0x9c, 0x90, 0x8e, 0x6e, 0x6a, 0x74, 0x79, 0x81, 0x6d,\n",
+ " 0x6f, 0x8e, 0x77, 0x92, 0x93, 0x7d, 0x5f, 0x68, 0x6a, 0x6c, 0x80, 0x8f,\n",
+ " 0x99, 0x84, 0x4f, 0x64, 0x5c, 0x93, 0x7c, 0x91, 0x98, 0x82, 0x62, 0x3f,\n",
+ " 0x41, 0x9f, 0x5d, 0x89, 0x98, 0x89, 0x73, 0x50, 0x32, 0xa8, 0xa0, 0x7a,\n",
+ " 0xa0, 0x95, 0x78, 0x69, 0x74, 0x7c, 0x89, 0x7b, 0x80, 0x65, 0x56, 0x6b,\n",
+ " 0x69, 0x78, 0x62, 0x87, 0xaf, 0x94, 0x7a, 0x64, 0x53, 0x86, 0x45, 0x99,\n",
+ " 0x88, 0x79, 0x4d, 0x74, 0x59, 0x91, 0x5f, 0x7b, 0x88, 0x90, 0x80, 0x86,\n",
+ " 0x7d, 0x7b, 0x64, 0xa3, 0x7f, 0x74, 0x89, 0x80, 0x7d, 0x7c, 0x7a, 0x87,\n",
+ " 0x5f, 0x8a, 0x5a, 0x72, 0x79, 0x74, 0x8c, 0x7c, 0x86, 0x91, 0x6e, 0x5d,\n",
+ " 0x61, 0x8e, 0xa2, 0x68, 0xd4, 0x92, 0x67, 0x66, 0x62, 0xa1, 0xf3, 0x63,\n",
+ " 0x91, 0x81, 0x74, 0x5f, 0x88, 0x98, 0xbb, 0x5a, 0x9b, 0x54, 0x6a, 0x5c,\n",
+ " 0x75, 0x88, 0xad, 0x7c, 0xb4, 0x7c, 0x69, 0x74, 0x84, 0x76, 0x9d, 0x9a,\n",
+ " 0xb0, 0x91, 0x5d, 0xa3, 0xa4, 0x7f, 0xbb, 0x80, 0xa4, 0x5d, 0x83, 0xaf,\n",
+ " 0xb7, 0x66, 0xb0, 0x7f, 0x89, 0x4b, 0x72, 0x9e, 0x99, 0x7c, 0x66, 0x71,\n",
+ " 0x6a, 0x6f, 0x6d, 0x67, 0x8d, 0x6d, 0x46, 0xa5, 0x9b, 0x84, 0x7a, 0x61,\n",
+ " 0x64, 0x5c, 0x88, 0x89, 0x95, 0x8c, 0x70, 0x4b, 0x6c, 0x85, 0x83, 0x8b,\n",
+ " 0x98, 0x87, 0x6a, 0x44, 0x4d, 0x9d, 0x78, 0x71, 0x78, 0x7e, 0x91, 0x5b,\n",
+ " 0x3f, 0x9f, 0x80, 0x62, 0xa7, 0x95, 0x5d, 0x74, 0x65, 0x9c, 0x6d, 0x7a,\n",
+ " 0x98, 0x79, 0x80, 0x61, 0x49, 0x82, 0x65, 0x92, 0x80, 0x96, 0x7c, 0x72,\n",
+ " 0x4f, 0x76, 0x5e, 0x8d, 0x97, 0xa5, 0x72, 0x57, 0x79, 0x87, 0x67, 0x87,\n",
+ " 0x80, 0x84, 0x7c, 0x6f, 0x66, 0x6b, 0x70, 0x9b, 0x64, 0x90, 0x59, 0x96,\n",
+ " 0x7a, 0x6f, 0x75, 0x89, 0x4e, 0x8a, 0x62, 0x6e, 0x9c, 0x8c, 0x9a, 0x78,\n",
+ " 0x8e, 0x91, 0x3d, 0x50, 0x72, 0x92, 0x9f, 0x63, 0xda, 0x92, 0x72, 0x60,\n",
+ " 0x59, 0xa6, 0xd0, 0x56, 0xc1, 0x6b, 0x5e, 0x76, 0x6e, 0x81, 0xbb, 0x4b,\n",
+ " 0xbb, 0x59, 0x68, 0x4f, 0x77, 0x87, 0xa1, 0x73, 0xbf, 0x65, 0x56, 0x67,\n",
+ " 0x77, 0x84, 0x8a, 0x7e, 0xb8, 0x85, 0x66, 0xa6, 0x99, 0xa0, 0xa5, 0x73,\n",
+ " 0x8d, 0x4a, 0x7d, 0xab, 0xb0, 0x6a, 0x94, 0x84, 0x87, 0x4c, 0x74, 0xa3,\n",
+ " 0xb3, 0xa9, 0x62, 0x7a, 0x71, 0x7f, 0x53, 0x79, 0x7a, 0x7c, 0x5e, 0x8f,\n",
+ " 0xa0, 0x90, 0x5c, 0x76, 0x6c, 0x92, 0x70, 0x9c, 0xb3, 0x8b, 0x7e, 0x57,\n",
+ " 0x5b, 0x9d, 0x96, 0x85, 0x70, 0x93, 0x8b, 0x67, 0x4c, 0x9c, 0x6a, 0x83,\n",
+ " 0x84, 0x90, 0x8e, 0x60, 0x56, 0xb3, 0x87, 0x7d, 0x86, 0x88, 0x79, 0x5b,\n",
+ " 0x58, 0x94, 0x92, 0x8e, 0x90, 0x76, 0x58, 0x51, 0x52, 0x63, 0x57, 0x88,\n",
+ " 0x9b, 0x7a, 0x85, 0x6c, 0x8b, 0x87, 0x5f, 0x8b, 0x90, 0x92, 0x81, 0x64,\n",
+ " 0x52, 0x8b, 0x77, 0x94, 0x96, 0x98, 0x69, 0x5b, 0x79, 0x87, 0x61, 0x96,\n",
+ " 0x7b, 0x9a, 0x61, 0x74, 0x7e, 0x8b, 0x82, 0x92, 0x4f, 0x87, 0x7f, 0x80,\n",
+ " 0x74, 0x97, 0x98, 0x7a, 0x79, 0x97, 0x65, 0x67, 0x66, 0xb1, 0xb1, 0x49,\n",
+ " 0xd6, 0x97, 0x58, 0x47, 0x62, 0x94, 0xd5, 0x82, 0xa0, 0x60, 0x3f, 0x67,\n",
+ " 0x6c, 0x9d, 0xb6, 0x58, 0xb1, 0x6e, 0x58, 0x4e, 0x7c, 0x83, 0x8b, 0x83,\n",
+ " 0xd5, 0x62, 0x8d, 0x84, 0x84, 0x8c, 0xa9, 0x6e, 0xac, 0x7f, 0x6d, 0x88,\n",
+ " 0xab, 0x8b, 0xb1, 0x77, 0x9b, 0x46, 0x76, 0xa7, 0xb8, 0x7b, 0xc5, 0x6e,\n",
+ " 0x73, 0x62, 0x68, 0x95, 0xab, 0x7c, 0x6f, 0x74, 0x56, 0x71, 0x61, 0x83,\n",
+ " 0x8a, 0x73, 0x54, 0x94, 0x86, 0x91, 0x60, 0x69, 0x65, 0x6b, 0x76, 0x85,\n",
+ " 0xae, 0x87, 0x8f, 0x55, 0x41, 0x98, 0x68, 0x87, 0x5e, 0x7a, 0x80, 0x38,\n",
+ " 0x50, 0xaf, 0x93, 0x79, 0x57, 0x96, 0x7b, 0x53, 0x4e, 0xc0, 0xa0, 0x85,\n",
+ " 0x87, 0x95, 0x86, 0x70, 0x4c, 0x9f, 0x77, 0x7d, 0x8b, 0x7a, 0x7b, 0x6d,\n",
+ " 0x57, 0x74, 0x81, 0x7d, 0xa2, 0x79, 0x64, 0x6c, 0x55, 0x70, 0x3c, 0x88,\n",
+ " 0x8a, 0x7a, 0x58, 0x72, 0x71, 0x7d, 0x6a, 0x8d, 0x78, 0x7e, 0x95, 0x8b,\n",
+ " 0x84, 0x7e, 0x73, 0x7c, 0x7e, 0x67, 0x89, 0x8b, 0x6d, 0x68, 0x66, 0x73,\n",
+ " 0x5a, 0x93, 0x82, 0x85, 0x97, 0x6b, 0x9a, 0x72, 0x51, 0xa2, 0x4f, 0x67,\n",
+ " 0x67, 0x7e, 0xbb, 0x37, 0xe3, 0x9c, 0x57, 0x5b, 0x6f, 0xa0, 0xdc, 0x5c,\n",
+ " 0xa6, 0x7c, 0x71, 0x77, 0x72, 0x88, 0xd0, 0x4d, 0x93, 0x58, 0x74, 0x6d,\n",
+ " 0x8f, 0x77, 0xa3, 0x76, 0xb7, 0x76, 0x6d, 0x6d, 0x6f, 0x7b, 0xaa, 0x6d,\n",
+ " 0xaa, 0x6a, 0x72, 0x98, 0x8d, 0x98, 0xb0, 0x52, 0x76, 0x5d, 0x61, 0xb7,\n",
+ " 0xac, 0x90, 0xa5, 0x75, 0x7e, 0x3d, 0x5b, 0x9a, 0xbf, 0x81, 0x83, 0x7b,\n",
+ " 0x5c, 0x77, 0x74, 0x82, 0x8d, 0x7e, 0x4f, 0x9f, 0x8f, 0x97, 0x7c, 0x75,\n",
+ " 0x5b, 0x73, 0x97, 0x73, 0x85, 0x7f, 0x70, 0x5a, 0x53, 0x81, 0x81, 0x89,\n",
+ " 0x73, 0x8d, 0x8a, 0x5c, 0x5f, 0x84, 0x86, 0x6f, 0x76, 0x78, 0x82, 0x6d,\n",
+ " 0x4f, 0xbb, 0x91, 0x61, 0x7e, 0x97, 0x6c, 0x67, 0x62, 0x83, 0x61, 0x7d,\n",
+ " 0x89, 0x76, 0x7b, 0x67, 0x56, 0x74, 0x49, 0x7b, 0x6b, 0x8b, 0x89, 0x74,\n",
+ " 0x5b, 0x7f, 0x78, 0x7b, 0x80, 0x7e, 0x63, 0x71, 0x5e, 0x91, 0x81, 0x92,\n",
+ " 0x7b, 0x90, 0x9c, 0x7a, 0x73, 0x85, 0x79, 0x9b, 0x66, 0x93, 0x60, 0x87,\n",
+ " 0x79, 0x69, 0x73, 0x8b, 0x53, 0x8c, 0x8d, 0x68, 0x93, 0xa0, 0x91, 0x65,\n",
+ " 0x57, 0x8d, 0x71, 0x65, 0x6c, 0x7e, 0xb3, 0x4f, 0xc7, 0xaa, 0x5a, 0x77,\n",
+ " 0x6e, 0x85, 0xe4, 0x6c, 0xa3, 0x89, 0x69, 0x54, 0x6d, 0x99, 0xb9, 0x77,\n",
+ " 0xa0, 0x80, 0x85, 0x71, 0x70, 0x78, 0x99, 0x66, 0xaf, 0x8a, 0x59, 0x64,\n",
+ " 0x54, 0x62, 0xbf, 0x5c, 0xbd, 0x77, 0x7f, 0xab, 0x95, 0x85, 0xaa, 0x6e,\n",
+ " 0xaa, 0x5a, 0x7b, 0x9f, 0xc3, 0x65, 0x93, 0x64, 0x7c, 0x2d, 0x4e, 0x8f,\n",
+ " 0xb2, 0x5f, 0x4e, 0x61, 0x64, 0x73, 0x56, 0x75, 0x79, 0x90, 0x5c, 0x81,\n",
+ " 0x8a, 0x8c, 0x70, 0x64, 0x74, 0x86, 0x86, 0x82, 0xab, 0x7e, 0x62, 0x4f,\n",
+ " 0x51, 0x89, 0x7b, 0x88, 0x73, 0x97, 0x77, 0x75, 0x5c, 0x9e, 0x97, 0x70,\n",
+ " 0x5a, 0x98, 0x7a, 0x54, 0x47, 0x99, 0xab, 0x5d, 0x91, 0xa0, 0x64, 0x51,\n",
+ " 0x57, 0x88, 0x88, 0x85, 0x81, 0x83, 0xa1, 0x89, 0x6a, 0x88, 0x69, 0x81,\n",
+ " 0x92, 0x63, 0x6a, 0x71, 0x72, 0x6a, 0x75, 0x8e, 0x90, 0x9d, 0x69, 0x60,\n",
+ " 0x73, 0x95, 0x79, 0x7b, 0x79, 0x7f, 0x77, 0x6e, 0x69, 0x63, 0x60, 0xa0,\n",
+ " 0x84, 0x91, 0x80, 0x96, 0x92, 0x70, 0x69, 0x7c, 0x3f, 0x90, 0x5c, 0x79,\n",
+ " 0x82, 0x63, 0x8d, 0x63, 0x56, 0x8a, 0x8e, 0x7a, 0x5c, 0x8d, 0xb8, 0x4e,\n",
+ " 0xb6, 0x84, 0x57, 0x79, 0x59, 0x79, 0xe8, 0x7e, 0xa8, 0x71, 0x61, 0x62,\n",
+ " 0x89, 0x71, 0xb7, 0x83, 0x7b, 0x53, 0x86, 0x88, 0x74, 0x71, 0xb1, 0x61,\n",
+ " 0xae, 0x7e, 0x8f, 0x69, 0x6b, 0x69, 0xb2, 0x6d, 0xb1, 0x7f, 0x5c, 0x9f,\n",
+ " 0xaa, 0x8c, 0xbd, 0x74, 0xaa, 0x5b, 0x7f, 0xa5, 0xb0, 0x6e, 0xc1, 0x5c,\n",
+ " 0x94, 0x34, 0x5b, 0xa6, 0xbc, 0x49, 0x75, 0x5b, 0x6e, 0x74, 0x7a, 0x92,\n",
+ " 0x92, 0x79, 0x78, 0x8a, 0x9e, 0x97, 0x7c, 0x5f, 0x76, 0x86, 0x59, 0x81,\n",
+ " 0x83, 0x7a, 0x65, 0x5b, 0x42, 0x95, 0x84, 0x99, 0x81, 0x8d, 0x6a, 0x5e,\n",
+ " 0x59, 0xb7, 0x96, 0x8a, 0x77, 0x86, 0x7a, 0x67, 0x3b, 0xa8, 0xae, 0x7a,\n",
+ " 0xa0, 0x97, 0x6c, 0x73, 0x5b, 0x9b, 0x77, 0x84, 0x7a, 0x77, 0x75, 0x6f,\n",
+ " 0x7d, 0x7a, 0x71, 0x86, 0x6c, 0x6f, 0x7d, 0x71, 0x68, 0x60, 0x64, 0x86,\n",
+ " 0x90, 0x75, 0x6a, 0x61, 0x60, 0x87, 0x68, 0x99, 0x87, 0x7e, 0x92, 0x87,\n",
+ " 0x87, 0x5f, 0x60, 0x91, 0x68, 0x8c, 0x7b, 0x67, 0x79, 0x5d, 0x67, 0x77,\n",
+ " 0x47, 0x72, 0x76, 0x88, 0x82, 0xa2, 0x7a, 0x5d, 0x64, 0x87, 0x75, 0x78,\n",
+ " 0x5e, 0x6f, 0xa4, 0x52, 0xc2, 0x9d, 0x81, 0x89, 0x55, 0x86, 0xc9, 0x6f,\n",
+ " 0x95, 0x71, 0x9d, 0x87, 0x95, 0x74, 0xac, 0x7f, 0x95, 0x6c, 0x68, 0x66,\n",
+ " 0x8a, 0x5f, 0x96, 0x69, 0x95, 0x79, 0x7f, 0x71, 0x86, 0x7e, 0x98, 0x71,\n",
+ " 0xac, 0x8f, 0x75, 0xa5, 0xac, 0x7a, 0xca, 0x63, 0xa0, 0x63, 0x69, 0xbf,\n",
+ " 0xae, 0x62, 0xc9, 0x46, 0x74, 0x2c, 0x66, 0x96, 0xb7, 0x70, 0x7c, 0x6b,\n",
+ " 0x7b, 0x90, 0x72, 0x74, 0x8d, 0x5f, 0x63, 0x93, 0x97, 0x78, 0x79, 0x64,\n",
+ " 0x67, 0x84, 0x64, 0x82, 0x90, 0x83, 0x91, 0x5f, 0x72, 0x93, 0x91, 0xae,\n",
+ " 0x6d, 0x99, 0x5b, 0x69, 0x54, 0x9f, 0x97, 0x80, 0x80, 0xa4, 0x91, 0x66,\n",
+ " 0x65, 0xa4, 0xa7, 0x7b, 0x97, 0x87, 0x72, 0x68, 0x6a, 0x96, 0x7b, 0x79,\n",
+ " 0x69, 0x83, 0x6f, 0x85, 0x6b, 0x92, 0x7f, 0x71, 0x84, 0x87, 0x6a, 0x7b,\n",
+ " 0x63, 0x72, 0x5f, 0x87, 0x98, 0x7b, 0x96, 0x71, 0x62, 0x90, 0x71, 0xa3,\n",
+ " 0x8c, 0x77, 0x90, 0x6f, 0x83, 0x76, 0x65, 0x87, 0x72, 0x8a, 0x64, 0x87,\n",
+ " 0x75, 0x75, 0x6d, 0x84, 0x54, 0x89, 0x88, 0xa0, 0x87, 0x73, 0x7f, 0x6f,\n",
+ " 0x5f, 0x90, 0x5e, 0x94, 0x5d, 0x61, 0xa6, 0x56, 0xb3, 0x91, 0x95, 0x75,\n",
+ " 0x4d, 0x74, 0xd9, 0x87, 0x92, 0x74, 0x7f, 0x79, 0x97, 0x6e, 0x90, 0x54,\n",
+ " 0x84, 0x5d, 0x5f, 0x75, 0x8b, 0x84, 0xa6, 0x75, 0xb4, 0x77, 0x78, 0x85,\n",
+ " 0x90, 0x76, 0xbd, 0x78, 0xd1, 0xa0, 0x5d, 0x96, 0xa9, 0x7c, 0xc1, 0x61,\n",
+ " 0xc2, 0x71, 0x8b, 0xa5, 0xa5, 0x5b, 0xc8, 0x50, 0x7b, 0x4b, 0x93, 0x99,\n",
+ " 0xae, 0x72, 0x67, 0x54, 0x81, 0x89, 0x96, 0x81, 0x6e, 0x68, 0x55, 0x7f,\n",
+ " 0x93, 0x8c, 0x5e, 0x65, 0x6c, 0x84, 0x7f, 0x8f, 0x9e, 0x7b, 0x73, 0x7f,\n",
+ " 0x51, 0x63, 0x8a, 0x8b, 0x6b, 0x9b, 0x9d, 0x57, 0x68, 0x89, 0x98, 0x70,\n",
+ " 0x73, 0xa3, 0x7f, 0x69, 0x44, 0x89, 0xae, 0x68, 0x89, 0x80, 0x7e, 0x6d,\n",
+ " 0x70, 0x95, 0x85, 0x65, 0x91, 0x7f, 0x66, 0x74, 0x96, 0x72, 0x60, 0x7a,\n",
+ " 0x87, 0x85, 0x79, 0x54, 0x53, 0x6c, 0x88, 0x87, 0xa9, 0x90, 0x75, 0x8b,\n",
+ " 0x69, 0x98, 0x7d, 0x95, 0x85, 0x7a, 0x8b, 0x82, 0x87, 0x6f, 0x86, 0x7f,\n",
+ " 0x74, 0xab, 0x93, 0x6c, 0x8a, 0x78, 0x68, 0x81, 0x62, 0x88, 0x78, 0x91,\n",
+ " 0x8b, 0x55, 0xa7, 0x58, 0x64, 0x88, 0x71, 0x93, 0x7d, 0x69, 0xbc, 0x58,\n",
+ " 0xbe, 0x9a, 0x6f, 0x74, 0x6f, 0x7f, 0xeb, 0x9e, 0xb7, 0x60, 0x63, 0x98,\n",
+ " 0x82, 0x77, 0x94, 0x63, 0x80, 0x6f, 0x7d, 0x8f, 0x8b, 0x85, 0xa5, 0x62,\n",
+ " 0xad, 0x86, 0x5f, 0x76, 0x88, 0x74, 0xa5, 0x66, 0xa5, 0x94, 0x88, 0x9b,\n",
+ " 0x87, 0x9e, 0xa8, 0x5a, 0xc9, 0x81, 0x92, 0xcd, 0xb5, 0x67, 0xb9, 0x63,\n",
+ " 0x86, 0x65, 0x8d, 0xad, 0x98, 0x7c, 0x8a, 0x40, 0x67, 0x65, 0x60, 0x71,\n",
+ " 0x8e, 0x84, 0x73, 0x64, 0x98, 0x80, 0x73, 0x81, 0x48, 0x75, 0x71, 0x9e,\n",
+ " 0x73, 0x89, 0x89, 0x68, 0x73, 0xa6, 0x84, 0x8a, 0x7e, 0x9f, 0x78, 0x83,\n",
+ " 0x60, 0x77, 0xa1, 0x87, 0x76, 0xab, 0x74, 0x57, 0x6d, 0x99, 0xa5, 0x5e,\n",
+ " 0x9d, 0x91, 0x6d, 0x6a, 0x76, 0x9c, 0x7b, 0x66, 0x96, 0x84, 0x85, 0x6e,\n",
+ " 0x6c, 0x75, 0x86, 0x6a, 0x71, 0x67, 0x8a, 0x66, 0x66, 0x68, 0x73, 0x90,\n",
+ " 0x92, 0x68, 0x8f, 0x71, 0x82, 0x7e, 0x71, 0xad, 0x9f, 0x84, 0x9e, 0x7d,\n",
+ " 0x77, 0x6b, 0x67, 0x8f, 0x73, 0x9a, 0x91, 0x74, 0x8a, 0x74, 0x5a, 0x87,\n",
+ " 0x37, 0x80, 0x8c, 0x8f, 0x7f, 0x75, 0xa8, 0x49, 0x63, 0x9b, 0x67, 0x68,\n",
+ " 0x4f, 0x87, 0xbf, 0x59, 0x9c, 0xbe, 0x93, 0x7e, 0x6f, 0x8a, 0xea, 0x77,\n",
+ " 0x83, 0x7a, 0x75, 0x8e, 0x7d, 0x50, 0x95, 0x60, 0x74, 0x60, 0x6f, 0x97,\n",
+ " 0x72, 0x5c, 0xa3, 0x6d, 0xb9, 0x86, 0x7b, 0x89, 0x9a, 0x76, 0xc7, 0x56,\n",
+ " 0xba, 0x86, 0x8d, 0x93, 0xa9, 0x98, 0xbb, 0x6a, 0x97, 0x74, 0x68, 0x84,\n",
+ " 0xc3, 0x65, 0xb6, 0x68, 0x89, 0x58, 0x87, 0xa1, 0xac, 0x60, 0x65, 0x68,\n",
+ " 0x7d, 0x98, 0x67, 0x8f, 0x8e, 0x84, 0x50, 0x75, 0x83, 0x91, 0x8a, 0x90,\n",
+ " 0x66, 0x74, 0x96, 0x89, 0x81, 0x7a, 0x7a, 0x64, 0x7f, 0x73, 0x8f, 0x95,\n",
+ " 0x8c, 0x89, 0x96, 0x76, 0x7a, 0x6c, 0x89, 0x91, 0x6d, 0x84, 0x68, 0x8d,\n",
+ " 0x47, 0x94, 0x9a, 0x67, 0x8f, 0x89, 0x8e, 0x79, 0x73, 0xa8, 0x7f, 0x6c,\n",
+ " 0x80, 0x64, 0x75, 0x81, 0x96, 0x9c, 0x68, 0x65, 0x76, 0x68, 0x74, 0x72,\n",
+ " 0x68, 0x76, 0x62, 0x6d, 0x6e, 0x6a, 0x84, 0x65, 0x8a, 0x73, 0x76, 0x91,\n",
+ " 0x78, 0x7c, 0x7a, 0x88, 0x6a, 0x87, 0x60, 0x99, 0x88, 0x75, 0x7b, 0x71,\n",
+ " 0x81, 0x7b, 0x76, 0x7d, 0x58, 0x75, 0x65, 0xa3, 0x95, 0x7e, 0x96, 0x3e,\n",
+ " 0x4c, 0x97, 0x86, 0x7a, 0x62, 0x92, 0xd1, 0x72, 0x8e, 0xaa, 0x85, 0x8e,\n",
+ " 0x59, 0x5f, 0xec, 0x77, 0x96, 0x66, 0x91, 0x9a, 0x89, 0x6c, 0xa2, 0x69,\n",
+ " 0x7d, 0x6e, 0x76, 0x63, 0x82, 0x72, 0x9c, 0x72, 0xa3, 0x75, 0x85, 0x7b,\n",
+ " 0x6d, 0x96, 0xc2, 0x69, 0xa7, 0x6a, 0x6b, 0x83, 0xa2, 0x7d, 0xce, 0x5c,\n",
+ " 0x94, 0x61, 0x7d, 0xae, 0xc3, 0x6d, 0x9f, 0x3c, 0x52, 0x4d, 0x8e, 0x92,\n",
+ " 0xae, 0x6e, 0x70, 0x5a, 0x76, 0x84, 0x7f, 0x72, 0x92, 0x72, 0x76, 0x5e,\n",
+ " 0x73, 0x8e, 0x82, 0x6d, 0x72, 0x81, 0x79, 0x94, 0x81, 0x88, 0x8b, 0x81,\n",
+ " 0x72, 0x72, 0x69, 0x84, 0x59, 0x6e, 0x74, 0x7d, 0x66, 0x74, 0x8d, 0x7b,\n",
+ " 0x7d, 0x7e, 0x7a, 0x83, 0x4d, 0x7e, 0x6a, 0x5a, 0x87, 0x66, 0x84, 0xa5,\n",
+ " 0x50, 0x5d, 0x6a, 0x8e, 0x87, 0x74, 0x88, 0x7c, 0x7d, 0x6c, 0x93, 0x98,\n",
+ " 0x8c, 0x76, 0x7f, 0xa3, 0x6e, 0x5d, 0x7d, 0x9f, 0x7c, 0x7a, 0x98, 0x88,\n",
+ " 0x74, 0x73, 0x50, 0x8c, 0x78, 0x8b, 0x71, 0x77, 0x9d, 0x56, 0x71, 0x85,\n",
+ " 0x6b, 0x8a, 0x93, 0x82, 0x8c, 0x79, 0x68, 0x8b, 0x57, 0x7b, 0x7c, 0x8a,\n",
+ " 0x6c, 0x87, 0x98, 0x54, 0x63, 0x7e, 0x78, 0x6b, 0x63, 0x77, 0xc1, 0x52,\n",
+ " 0xcd, 0xab, 0x75, 0x8e, 0x64, 0x68, 0xce, 0x68, 0x88, 0x6d, 0x67, 0x6d,\n",
+ " 0x68, 0x76, 0xa7, 0x78, 0x83, 0x67, 0x65, 0x5b, 0x8f, 0x63, 0x90, 0x5b,\n",
+ " 0xa1, 0x6f, 0x6a, 0x88, 0x70, 0x5c, 0x78, 0x49, 0xbc, 0x85, 0x8d, 0x8e,\n",
+ " 0xa3, 0x90, 0x97, 0x84, 0xa2, 0x46, 0x7a, 0x8e, 0x9e, 0xb1, 0xaa, 0x53,\n",
+ " 0x7d, 0x6b, 0x72, 0x86, 0x8c, 0x67, 0x6b, 0x48, 0x6f, 0x9c, 0x51, 0x94,\n",
+ " 0x6d, 0x66, 0x8e, 0x90, 0x79, 0x81, 0x66, 0x9f, 0x82, 0x9f, 0x98, 0x97,\n",
+ " 0x7c, 0x86, 0x7f, 0x57, 0x57, 0x83, 0x97, 0x8f, 0x73, 0x6f, 0x75, 0x6c,\n",
+ " 0x56, 0x8f, 0x7f, 0x73, 0x71, 0x84, 0x7d, 0x5f, 0x69, 0x69, 0x8e, 0x67,\n",
+ " 0x8a, 0x7f, 0x8c, 0x5a, 0x7a, 0x67, 0x82, 0x5a, 0x7a, 0x68, 0x73, 0x58,\n",
+ " 0x84, 0x83, 0x8d, 0x6d, 0x83, 0x72, 0x80, 0x7a, 0x8e, 0x7a, 0x68, 0x88,\n",
+ " 0x65, 0x74, 0x78, 0x73, 0x83, 0x97, 0x7b, 0x84, 0x77, 0x6d, 0x95, 0x99,\n",
+ " 0x76, 0x69, 0x5f, 0x9b, 0x7c, 0x75, 0x91, 0x80, 0x7b, 0x73, 0x6f, 0x9f,\n",
+ " 0x00, 0x00, 0x0e, 0x00, 0x18, 0x00, 0x08, 0x00, 0x07, 0x00, 0x0c, 0x00,\n",
+ " 0x10, 0x00, 0x14, 0x00, 0x0e, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02,\n",
+ " 0x10, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x28, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00,\n",
+ " 0x0b, 0x00, 0x00, 0x00, 0x4d, 0x61, 0x74, 0x4d, 0x75, 0x6c, 0x5f, 0x62,\n",
+ " 0x69, 0x61, 0x73, 0x00, 0x0c, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0xaa, 0xcc, 0xe2, 0x37, 0x10, 0x00, 0x00, 0x00, 0xd6, 0x01, 0x00, 0x00,\n",
+ " 0xfd, 0xfd, 0xff, 0xff, 0x53, 0xfe, 0xff, 0xff, 0x74, 0x01, 0x00, 0x00,\n",
+ " 0x03, 0x00, 0x00, 0x00, 0xb4, 0x00, 0x00, 0x00, 0x5c, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0xc0, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x09,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x24, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x05, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x08, 0x00, 0x04, 0x00,\n",
+ " 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x14, 0x00, 0x1c, 0x00,\n",
+ " 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x07, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x18, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00,\n",
+ " 0x28, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x07, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x01, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, 0x14, 0x00, 0x18, 0x00,\n",
+ " 0x00, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x07, 0x00, 0x10, 0x00, 0x00, 0x00,\n",
+ " 0x00, 0x00, 0x14, 0x00, 0x14, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02,\n",
+ " 0x10, 0x00, 0x00, 0x00, 0x1c, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00,\n",
+ " 0x1c, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00,\n",
+ " 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0c, 0x00, 0x10, 0x00,\n",
+ " 0x00, 0x00, 0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x0c, 0x00, 0x00, 0x00,\n",
+ " 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,\n",
+ " 0x03, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x14, 0x00, 0x00, 0x00,\n",
+ " 0x04, 0x00, 0x00, 0x00, 0xfa, 0xff, 0xff, 0xff, 0x00, 0x19, 0x06, 0x00,\n",
+ " 0x06, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x09, 0x06, 0x00,\n",
+ " 0x08, 0x00, 0x07, 0x00, 0x06, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04\n",
+ "};\n",
+ "unsigned int g_model_len = 18288;\n"
+ ],
+ "name": "stdout"
+ }
+ ]
+ }
+ ]
+}
\ No newline at end of file
diff --git a/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb b/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb
deleted file mode 100644
index 0baaeac..0000000
--- a/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb
+++ /dev/null
@@ -1,324 +0,0 @@
-{
- "nbformat": 4,
- "nbformat_minor": 0,
- "metadata": {
- "colab": {
- "name": "Train simple audio recognition model",
- "version": "0.3.2",
- "provenance": [],
- "collapsed_sections": []
- },
- "kernelspec": {
- "name": "python3",
- "display_name": "Python 3"
- },
- "accelerator": "GPU"
- },
- "cells": [
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "pO4-CY_TCZZS",
- "colab_type": "text"
- },
- "source": [
- "# Train a Simple Audio Recognition model for microcontroller use"
- ]
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "BaFfr7DHRmGF",
- "colab_type": "text"
- },
- "source": [
- "This notebook demonstrates how to train a 20kb [Simple Audio Recognition](https://www.tensorflow.org/tutorials/sequences/audio_recognition) model for [TensorFlow Lite for Microcontrollers](https://tensorflow.org/lite/microcontrollers/overview). It will produce the same model used in the [micro_speech](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech) example application.\n",
- "\n",
- "The model is designed to be used with [Google Colaboratory](https://colab.research.google.com).\n",
- "\n",
- "<table class=\"tfo-notebook-buttons\" align=\"left\">\n",
- " <td>\n",
- " <a target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />Run in Google Colab</a>\n",
- " </td>\n",
- " <td>\n",
- " <a target=\"_blank\" href=\"https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/examples/micro_speech/train_speech_model.ipynb\"><img src=\"https://www.tensorflow.org/images/GitHub-Mark-32px.png\" />View source on GitHub</a>\n",
- " </td>\n",
- "</table>\n"
- ]
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "XaVtYN4nlCft",
- "colab_type": "text"
- },
- "source": [
- "The notebook runs Python scripts to train and freeze the model, and uses the TensorFlow Lite converter to convert it for use with TensorFlow Lite for Microcontrollers.\n",
- "\n",
- "**Training is much faster using GPU acceleration.** Before you proceed, ensure you are using a GPU runtime by going to **Runtime -> Change runtime type** and selecting **GPU**. Training 18,000 iterations will take 1.5-2 hours on a GPU runtime.\n",
- "\n",
- "## Configure training\n",
- "\n",
- "The following `os.environ` lines can be customized to set the words that will be trained for, and the steps and learning rate of the training. The default values will result in the same model that is used in the micro_speech example. Run the cell to set the configuration:"
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "ludfxbNIaegy",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "import os\n",
- "\n",
- "# A comma-delimited list of the words you want to train for.\n",
- "# The options are: yes,no,up,down,left,right,on,off,stop,go\n",
- "# All other words will be used to train an \"unknown\" category.\n",
- "os.environ[\"WANTED_WORDS\"] = \"yes,no\"\n",
- "\n",
- "# The number of steps and learning rates can be specified as comma-separated\n",
- "# lists to define the rate at each stage. For example,\n",
- "# TRAINING_STEPS=15000,3000 and LEARNING_RATE=0.001,0.0001\n",
- "# will run 18,000 training loops in total, with a rate of 0.001 for the first\n",
- "# 15,000, and 0.0001 for the final 3,000.\n",
- "os.environ[\"TRAINING_STEPS\"]=\"15000,3000\"\n",
- "os.environ[\"LEARNING_RATE\"]=\"0.001,0.0001\"\n",
- "\n",
- "# Calculate the total number of steps, which is used to identify the checkpoint\n",
- "# file name.\n",
- "total_steps = sum(map(lambda string: int(string),\n",
- " os.environ[\"TRAINING_STEPS\"].split(\",\")))\n",
- "os.environ[\"TOTAL_STEPS\"] = str(total_steps)\n",
- "\n",
- "# Print the configuration to confirm it\n",
- "!echo \"Training these words: ${WANTED_WORDS}\"\n",
- "!echo \"Training steps in each stage: ${TRAINING_STEPS}\"\n",
- "!echo \"Learning rate in each stage: ${LEARNING_RATE}\"\n",
- "!echo \"Total number of training steps: ${TOTAL_STEPS}\"\n"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "gCgeOpvY9pAi",
- "colab_type": "text"
- },
- "source": [
- "## Install dependencies\n",
- "\n",
- "Next, we'll install a GPU build of TensorFlow, so we can use GPU acceleration for training."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "Nd1iM1o2ymvA",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "# Replace Colab's default TensorFlow install with an older\n",
- "# build that contains the operations that are needed for training\n",
- "!pip uninstall -y tensorflow tensorflow_estimator tensorboard\n",
- "!pip install -q tensorflow==1.15"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "T9Ty5mR58E4i",
- "colab_type": "text"
- },
- "source": [
- "We'll also clone the TensorFlow repository, which contains the scripts that train and freeze the model."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "APGx0fEh7hFF",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "# Clone the repository from GitHub\n",
- "!git clone -q https://github.com/tensorflow/tensorflow\n"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "aV_0qkYh98LD",
- "colab_type": "text"
- },
- "source": [
- "## Load TensorBoard\n",
- "\n",
- "Now, set up TensorBoard so that we can graph our accuracy and loss as training proceeds."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "yZArmzT85SLq",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "# Delete any old logs from previous runs\n",
- "!rm -rf /content/retrain_logs\n",
- "# Load TensorBoard\n",
- "%load_ext tensorboard\n",
- "%tensorboard --logdir /content/retrain_logs"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "x1J96Ron-O4R",
- "colab_type": "text"
- },
- "source": [
- "## Begin training\n",
- "\n",
- "Next, run the following script to begin training. The script will first download the training data:"
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "VJsEZx6lynbY",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "!python tensorflow/tensorflow/examples/speech_commands/train.py \\\n",
- "--model_architecture=tiny_conv --window_stride=20 --preprocess=micro \\\n",
- "--wanted_words=${WANTED_WORDS} --silence_percentage=25 --unknown_percentage=25 \\\n",
- "--quantize=1 --verbosity=WARN --how_many_training_steps=${TRAINING_STEPS} \\\n",
- "--learning_rate=${LEARNING_RATE} --summaries_dir=/content/retrain_logs \\\n",
- "--data_dir=/content/speech_dataset --train_dir=/content/speech_commands_train\n"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "XQUJLrdS-ftl",
- "colab_type": "text"
- },
- "source": [
- "## Freeze the graph\n",
- "\n",
- "Once training is complete, run the following cell to freeze the graph."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "xyc3_eLh9sAg",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "!python tensorflow/tensorflow/examples/speech_commands/freeze.py \\\n",
- "--model_architecture=tiny_conv --window_stride=20 --preprocess=micro \\\n",
- "--wanted_words=${WANTED_WORDS} --quantize=1 --output_file=/content/tiny_conv.pb \\\n",
- "--start_checkpoint=/content/speech_commands_train/tiny_conv.ckpt-${TOTAL_STEPS}"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "_DBGDxVI-nKG",
- "colab_type": "text"
- },
- "source": [
- "## Convert the model\n",
- "\n",
- "Run this cell to use the TensorFlow Lite converter to convert the frozen graph into the TensorFlow Lite format, fully quantized for use with embedded devices."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "lBj_AyCh1cC0",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "!toco \\\n",
- "--graph_def_file=/content/tiny_conv.pb --output_file=/content/tiny_conv.tflite \\\n",
- "--input_shapes=1,49,40,1 --input_arrays=Reshape_2 --output_arrays='labels_softmax' \\\n",
- "--inference_type=QUANTIZED_UINT8 --mean_values=0 --std_dev_values=9.8077"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "dt6Zqbxu-wIi",
- "colab_type": "text"
- },
- "source": [
- "The following cell will print the model size, which will be under 20 kilobytes."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "XohZOTjR8ZyE",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "import os\n",
- "model_size = os.path.getsize(\"/content/tiny_conv.tflite\")\n",
- "print(\"Model is %d bytes\" % model_size)"
- ],
- "execution_count": 0,
- "outputs": []
- },
- {
- "cell_type": "markdown",
- "metadata": {
- "id": "2pQnN0i_-0L2",
- "colab_type": "text"
- },
- "source": [
- "Finally, we use xxd to transform the model into a source file that can be included in a C++ project and loaded by TensorFlow Lite for Microcontrollers."
- ]
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "eoYyh0VU8pca",
- "colab_type": "code",
- "colab": {}
- },
- "source": [
- "# Install xxd if it is not available\n",
- "!apt-get -qq install xxd\n",
- "# Save the file as a C source file\n",
- "!xxd -i /content/tiny_conv.tflite > /content/tiny_conv.cc\n",
- "# Print the source file\n",
- "!cat /content/tiny_conv.cc"
- ],
- "execution_count": 0,
- "outputs": []
- }
- ]
-}
\ No newline at end of file
diff --git a/tensorflow/lite/micro/kernels/BUILD b/tensorflow/lite/micro/kernels/BUILD
index b83154b..1ba500f 100644
--- a/tensorflow/lite/micro/kernels/BUILD
+++ b/tensorflow/lite/micro/kernels/BUILD
@@ -34,6 +34,7 @@
"dequantize.cc",
"elementwise.cc",
"floor.cc",
+ "l2norm.cc",
"logical.cc",
"logistic.cc",
"maximum_minimum.cc",
@@ -132,6 +133,7 @@
"elementwise.cc",
"floor.cc",
"fully_connected.cc",
+ "l2norm.cc",
"logical.cc",
"logistic.cc",
"maximum_minimum.cc",
@@ -669,3 +671,16 @@
"//tensorflow/lite/micro/testing:micro_test",
],
)
+
+tflite_micro_cc_test(
+ name = "l2norm_test",
+ srcs = [
+ "l2norm_test.cc",
+ ],
+ deps = [
+ ":all_ops_resolver",
+ ":micro_ops",
+ "//tensorflow/lite/c:common",
+ "//tensorflow/lite/micro/testing:micro_test",
+ ],
+)
diff --git a/tensorflow/lite/micro/kernels/all_ops_resolver.cc b/tensorflow/lite/micro/kernels/all_ops_resolver.cc
index 4863382..6ba2e1a 100644
--- a/tensorflow/lite/micro/kernels/all_ops_resolver.cc
+++ b/tensorflow/lite/micro/kernels/all_ops_resolver.cc
@@ -24,7 +24,7 @@
AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED(), 1, 4);
AddBuiltin(BuiltinOperator_MAX_POOL_2D, Register_MAX_POOL_2D(), 1, 2);
AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX(), 1, 2);
- AddBuiltin(BuiltinOperator_LOGISTIC, Register_LOGISTIC());
+ AddBuiltin(BuiltinOperator_LOGISTIC, Register_LOGISTIC(), 1, 2);
AddBuiltin(BuiltinOperator_SVDF, Register_SVDF(), 1, 3);
AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D(), 1, 3);
AddBuiltin(BuiltinOperator_CONCATENATION, Register_CONCATENATION(), 1, 3);
@@ -75,6 +75,7 @@
Register_RESIZE_NEAREST_NEIGHBOR(),
/* min_version = */ 1,
/* max_version = */ 2);
+ AddBuiltin(BuiltinOperator_L2_NORMALIZATION, Register_L2_NORMALIZATION());
}
} // namespace micro
diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/add.cc b/tensorflow/lite/micro/kernels/cmsis-nn/add.cc
index 8b6d119..6dbe4a6 100644
--- a/tensorflow/lite/micro/kernels/cmsis-nn/add.cc
+++ b/tensorflow/lite/micro/kernels/cmsis-nn/add.cc
@@ -67,14 +67,15 @@
data->output_offset = output->params.zero_point;
data->left_shift = 20;
const double twice_max_input_scale =
- 2 * std::max(input1->params.scale, input2->params.scale);
+ 2 * static_cast<double>(
+ std::max(input1->params.scale, input2->params.scale));
const double real_input1_multiplier =
- input1->params.scale / twice_max_input_scale;
+ static_cast<double>(input1->params.scale) / twice_max_input_scale;
const double real_input2_multiplier =
- input2->params.scale / twice_max_input_scale;
+ static_cast<double>(input2->params.scale) / twice_max_input_scale;
const double real_output_multiplier =
twice_max_input_scale /
- ((1 << data->left_shift) * output->params.scale);
+ ((1 << data->left_shift) * static_cast<double>(output->params.scale));
QuantizeMultiplierSmallerThanOneExp(
real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);
diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc b/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc
index fdf14fe..34d4e83 100644
--- a/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc
+++ b/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc
@@ -40,8 +40,6 @@
// https://www.tensorflow.org/lite/performance/quantization_spec
constexpr int kConvQuantizedDimension = 0;
-const int kTensorNotAllocated = -1;
-
struct OpData {
TfLitePaddingValues padding;
// The scaling factor from input to output (aka the 'real multiplier') can
diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc b/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc
index 995834d..78787ea 100644
--- a/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc
+++ b/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc
@@ -78,9 +78,13 @@
}
TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
-#if defined(__ARM_FEATURE_DSP)
+ const TfLiteTensor* input = GetInput(context, node, kInputTensor);
const TfLiteTensor* filter = GetInput(context, node, kWeightsTensor);
-
+ TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
+ TF_LITE_ENSURE_EQ(context, input->type, output->type);
+ TF_LITE_ENSURE_MSG(context, input->type == filter->type,
+ "Hybrid models are not supported on TFLite Micro.");
+#if defined(__ARM_FEATURE_DSP)
RuntimeShape filter_shape = GetTensorShape(filter);
const int filter_dim_count = filter_shape.DimensionsCount();
const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
@@ -228,7 +232,8 @@
TF_LITE_ENSURE_STATUS(CalculateOpData(context, params, data_type, input,
filter, bias, output, data));
- switch (filter->type) { // Already know in/out types are same.
+ // Checks in Prepare ensure input, output and filter types are all the same.
+ switch (input->type) {
case kTfLiteFloat32:
return EvalFloat(context, node, params, data, input, filter, bias,
output);
diff --git a/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc b/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc
index c581de5..8f42a78 100644
--- a/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc
+++ b/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc
@@ -46,7 +46,8 @@
int input_left_shift;
tflite::PreprocessSoftmaxScaling(
- params->beta, input->params.scale, kScaledDiffIntegerBits,
+ static_cast<double>(params->beta),
+ static_cast<double>(input->params.scale), kScaledDiffIntegerBits,
&op_data->input_multiplier, &input_left_shift);
op_data->input_left_shift = input_left_shift;
op_data->diff_min =
diff --git a/tensorflow/lite/micro/kernels/l2norm.cc b/tensorflow/lite/micro/kernels/l2norm.cc
new file mode 100644
index 0000000..4dd71fe
--- /dev/null
+++ b/tensorflow/lite/micro/kernels/l2norm.cc
@@ -0,0 +1,150 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h"
+#include "tensorflow/lite/kernels/internal/reference/l2normalization.h"
+#include "tensorflow/lite/kernels/internal/tensor.h"
+#include "tensorflow/lite/kernels/kernel_util.h"
+
+namespace tflite {
+namespace ops {
+namespace micro {
+namespace l2norm {
+
+// This file has two implementation of L2Norm.
+enum KernelType {
+ kReference,
+ kGenericOptimized,
+};
+
+constexpr int kInputTensor = 0;
+constexpr int kOutputTensor = 0;
+
+TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
+#if defined(DEBUG)
+ auto* params = reinterpret_cast<TfLiteL2NormParams*>(node->builtin_data);
+
+ TF_LITE_ENSURE_EQ(context, NumInputs(node), 1);
+ TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
+
+ const TfLiteTensor* input = GetInput(context, node, kInputTensor);
+ TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
+
+ TF_LITE_ENSURE(context, NumDimensions(input) <= 4);
+
+ TF_LITE_ENSURE(context, output->type == kTfLiteFloat32 ||
+ output->type == kTfLiteUInt8 ||
+ output->type == kTfLiteInt8);
+ TF_LITE_ENSURE_EQ(context, input->type, output->type);
+
+ if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
+ TF_LITE_ENSURE_EQ(context, output->params.scale, (1. / 128.));
+ if (output->type == kTfLiteUInt8) {
+ TF_LITE_ENSURE_EQ(context, output->params.zero_point, 128);
+ }
+ if (output->type == kTfLiteInt8) {
+ TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
+ }
+ }
+
+ // TODO(ahentz): For some reason our implementations don't support
+ // activations.
+ TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActNone);
+#endif
+
+ return kTfLiteOk;
+}
+
+TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
+ const TfLiteTensor* input = GetInput(context, node, kInputTensor);
+ TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
+
+ // TODO(b/143912164): instead of hardcode the epsilon here, we should read it
+ // from tensorflow, i.e., adding a params.
+ // We don't compute epsilon for quantized kernel:
+ //
+ // epsilon_float = (epsilon_quant - zp) * scale
+ // so
+ // espsilon_quant = epsilon_float / scale + zp
+ // We know epsilon_float is just a very small number to avoid division by
+ // zero error, and scale is > 1, so the integer value of epsilon for quant
+ // is just dominated by the zero point.
+ // Also, GetInvSqrtQuantizedMultiplierExp handles the scenario where the sum
+ // of input value squared is zero case well.
+ // So we don't even need to do handle the epsilon for quantized kernel case.
+ const float epsilon = 1e-6f;
+ if (output->type == kTfLiteFloat32) {
+#define TF_LITE_L2NORM(type) \
+ tflite::L2NormalizationParams op_params; \
+ op_params.input_zero_point = 0; \
+ type::L2Normalization(op_params, GetTensorShape(input), \
+ GetTensorData<float>(input), GetTensorShape(output), \
+ GetTensorData<float>(output), epsilon)
+
+ TF_LITE_L2NORM(reference_ops);
+#undef TF_LITE_L2NORM
+ } else if (output->type == kTfLiteUInt8) {
+#define TF_LITE_L2NORM(type) \
+ tflite::L2NormalizationParams op_params; \
+ op_params.input_zero_point = input->params.zero_point; \
+ type::L2Normalization(op_params, GetTensorShape(input), \
+ GetTensorData<uint8>(input), GetTensorShape(output), \
+ GetTensorData<uint8>(output))
+
+ TF_LITE_L2NORM(reference_ops);
+#undef TF_LITE_L2NORM
+ } else if (output->type == kTfLiteInt8) {
+ const auto input_shape = GetTensorShape(input);
+ const auto output_shape = GetTensorShape(output);
+ const int trailing_dim = input_shape.DimensionsCount() - 1;
+ const int depth =
+ MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+ const int outer_size =
+ MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+ reference_integer_ops::L2Normalization(input->params.zero_point, outer_size,
+ depth, GetTensorData<int8>(input),
+ GetTensorData<int8>(output));
+ } else {
+ TF_LITE_KERNEL_LOG(context, "Output type is %d, requires float.",
+ output->type);
+ return kTfLiteError;
+ }
+
+ return kTfLiteOk;
+}
+
+} // namespace l2norm
+
+TfLiteRegistration* Register_L2NORM_REF() {
+ static TfLiteRegistration r = {/*init=*/nullptr,
+ /*free=*/nullptr,
+ /*prepare=*/l2norm::Prepare,
+ /*invoke=*/l2norm::Eval,
+ /*profiling_string=*/nullptr,
+ /*builtin_code=*/0,
+ /*custom_name=*/nullptr,
+ /*version=*/0};
+
+ return &r;
+}
+
+TfLiteRegistration* Register_L2_NORMALIZATION() {
+ return Register_L2NORM_REF();
+}
+
+} // namespace micro
+} // namespace ops
+} // namespace tflite
diff --git a/tensorflow/lite/micro/kernels/l2norm_test.cc b/tensorflow/lite/micro/kernels/l2norm_test.cc
new file mode 100644
index 0000000..a4f2fff
--- /dev/null
+++ b/tensorflow/lite/micro/kernels/l2norm_test.cc
@@ -0,0 +1,332 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/micro/kernels/all_ops_resolver.h"
+#include "tensorflow/lite/micro/testing/micro_test.h"
+#include "tensorflow/lite/micro/testing/test_utils.h"
+
+
+namespace tflite {
+namespace testing {
+namespace {
+
+// used to set the quantization parameters for the int8 and uint8 tests
+constexpr float kInputMin = -2.0;
+constexpr float kInputMax = 2.0;
+constexpr float kOutputMin = -1.0;
+constexpr float kOutputMax = 127.0 / 128.0;
+
+
+void QuantizeInputData(const float input_data[], int length,
+ uint8_t* quantized_data) {
+ for (int i=0; i < 6; i++) {
+ quantized_data[i] = tflite::testing::F2Q(input_data[i],
+ tflite::testing::kInputMin,
+ tflite::testing::kInputMax);
+ }
+}
+
+void QuantizeInputData(const float input_data[], int length,
+ int8_t* quantized_data) {
+ for (int i=0; i < 6; i++) {
+ quantized_data[i] = tflite::testing::F2QS(input_data[i],
+ tflite::testing::kInputMin,
+ tflite::testing::kInputMax);
+ }
+}
+
+TfLiteTensor CreateL2NormTensor(const float* data, TfLiteIntArray* dims,
+ const char* name, bool is_input) {
+ return CreateFloatTensor(data, dims, name);
+}
+
+TfLiteTensor CreateL2NormTensor(const uint8* data, TfLiteIntArray* dims,
+ const char* name, bool is_input) {
+ TfLiteTensor tensor;
+
+ if (is_input) {
+ tensor = CreateQuantizedTensor(data, dims, name, kInputMin, kInputMax);
+ } else {
+ tensor = CreateQuantizedTensor(data, dims, name, kOutputMin, kOutputMax);
+ }
+
+ tensor.quantization.type = kTfLiteAffineQuantization;
+ return tensor;
+}
+
+TfLiteTensor CreateL2NormTensor(const int8* data, TfLiteIntArray* dims,
+ const char* name, bool is_input) {
+ TfLiteTensor tensor;
+
+ if (is_input) {
+ tensor = CreateQuantizedTensor(data, dims, name, kInputMin, kInputMax);
+ } else {
+ tensor = CreateQuantizedTensor(data, dims, name, kOutputMin, kOutputMax);
+ }
+
+ tensor.quantization.type = kTfLiteAffineQuantization;
+ return tensor;
+}
+
+template <typename T>
+inline float Dequantize(const T data, float scale, int32_t zero_point) {
+ return scale * (data - zero_point);
+}
+
+template<typename T>
+void TestL2Normalization(const int* input_dims_data,
+ const T* input_data,
+ const float* expected_output_data,
+ T* output_data, float variance) {
+ TfLiteIntArray* dims = IntArrayFromInts(input_dims_data);
+
+ const int output_dims_count = ElementCount(*dims);
+
+ constexpr int tensors_size = 2;
+ TfLiteTensor tensors[tensors_size] = {
+ CreateL2NormTensor(input_data, dims, "input_tensor", true),
+ CreateL2NormTensor(output_data, dims, "output_tensor", false),
+ };
+
+ TfLiteContext context;
+ PopulateContext(tensors, tensors_size, micro_test::reporter, &context);
+ ::tflite::ops::micro::AllOpsResolver resolver;
+ const TfLiteRegistration* registration =
+ resolver.FindOp(tflite::BuiltinOperator_L2_NORMALIZATION, 1);
+ TF_LITE_MICRO_EXPECT_NE(nullptr, registration);
+
+ TfLiteL2NormParams builtin_data = {
+ .activation = kTfLiteActNone,
+ };
+
+ int inputs_array_data[] = {1, 0};
+ TfLiteIntArray* inputs_array = IntArrayFromInts(inputs_array_data);
+ int outputs_array_data[] = {1, 1};
+ TfLiteIntArray* outputs_array = IntArrayFromInts(outputs_array_data);
+ int temporaries_array_data[] = {0};
+ TfLiteIntArray* temporaries_array = IntArrayFromInts(temporaries_array_data);
+
+ TfLiteNode node;
+ node.inputs = inputs_array;
+ node.outputs = outputs_array;
+ node.temporaries = temporaries_array;
+ node.user_data = nullptr;
+ node.builtin_data = reinterpret_cast<void*>(&builtin_data);
+ node.custom_initial_data = nullptr;
+ node.custom_initial_data_size = 0;
+ node.delegate = nullptr;
+ TF_LITE_MICRO_EXPECT_NE(nullptr, registration->invoke);
+ TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, registration->invoke(&context, &node));
+
+ // Compare the results from dequantization and expected outputs, and make
+ // sure the difference is within a threshold.
+ if (tensors[1].quantization.type != kTfLiteNoQuantization) {
+ TfLiteTensor* output_tensor = &tensors[1];
+ int32_t zero_point = output_tensor->params.zero_point;
+ float scale = output_tensor->params.scale;
+
+ for (int i = 0; i < output_dims_count; ++i) {
+ float output_val = Dequantize(output_data[i], scale, zero_point);
+
+ TF_LITE_MICRO_EXPECT_LE(expected_output_data[i] - variance, output_val);
+ TF_LITE_MICRO_EXPECT_GE(expected_output_data[i] + variance, output_val);
+ }
+ } else {
+ for (int i = 0; i < output_dims_count; ++i) {
+ float output_val = static_cast<float>(output_data[i]);
+ TF_LITE_MICRO_EXPECT_LE(expected_output_data[i] - variance, output_val);
+ TF_LITE_MICRO_EXPECT_GE(expected_output_data[i] + variance, output_val);
+ }
+ }
+}
+
+} // namespace
+} // namespace testing
+} // namespace tflite
+
+
+TF_LITE_MICRO_TESTS_BEGIN
+
+TF_LITE_MICRO_TEST(SimpleFloatTest) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ const float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1
+ };
+ const float expected_output_data[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05
+ };
+ float output_data[data_length];
+
+ tflite::testing::TestL2Normalization<float>(input_dims, input_data,
+ expected_output_data, output_data, 0);
+}
+
+TF_LITE_MICRO_TEST(ZerosVectorFloatTest) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ const float input_data[data_length] = {0, 0, 0, 0, 0, 0};
+ const float expected_output_data[data_length] = {0, 0, 0, 0, 0, 0};
+ float output_data[data_length];
+
+ tflite::testing::TestL2Normalization<float>(input_dims, input_data,
+ expected_output_data, output_data, 0);
+}
+
+TF_LITE_MICRO_TEST(SimpleFloatWithRankLessThanFourTest) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ const float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1
+ };
+ const float expected_output_data[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05
+ };
+ float output_data[data_length];
+
+ tflite::testing::TestL2Normalization<float>(input_dims, input_data,
+ expected_output_data, output_data, 0);
+}
+
+TF_LITE_MICRO_TEST(MultipleBatchFloatTest) {
+ const int input_dims[] = {4, 3, 1, 1, 6};
+ constexpr int data_length = 18;
+ const float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 1
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 2
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 3
+ };
+ const float expected_output_data[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 1
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 2
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 3
+ };
+ float output_data[data_length];
+
+ tflite::testing::TestL2Normalization<float>(input_dims, input_data,
+ expected_output_data, output_data, 0);
+}
+
+TF_LITE_MICRO_TEST(ZerosVectorUint8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ const float input_data[data_length] = {0};
+ const float expected_output_data[data_length] = {0};
+ uint8_t quantized_input[data_length];
+ uint8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<uint8_t>(input_dims, quantized_input,
+ expected_output_data, output_data, .1);
+}
+
+TF_LITE_MICRO_TEST(SimpleUint8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1
+ };
+ float expected_output[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05
+ };
+ uint8_t quantized_input[data_length];
+ uint8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<uint8_t>(input_dims, quantized_input,
+ expected_output, output_data, .1);
+}
+
+TF_LITE_MICRO_TEST(SimpleInt8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1
+ };
+ float expected_output[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05
+ };
+ int8_t quantized_input[data_length];
+ int8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<int8_t>(input_dims, quantized_input,
+ expected_output, output_data, .1);
+}
+
+TF_LITE_MICRO_TEST(ZerosVectorInt8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 6;
+ const float input_data[data_length] = {0};
+ const float expected_output_data[data_length] = {0};
+ int8_t quantized_input[data_length];
+ int8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<int8_t>(input_dims, quantized_input,
+ expected_output_data, output_data, .1);
+}
+
+TF_LITE_MICRO_TEST(MultipleBatchUint8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 18;
+ float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 1
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 2
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 3
+ };
+ float expected_output[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 1
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 2
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 3
+ };
+ uint8_t quantized_input[data_length];
+ uint8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<uint8_t>(input_dims, quantized_input,
+ expected_output, output_data, .1);
+}
+
+TF_LITE_MICRO_TEST(MultipleBatchInt8Test) {
+ const int input_dims[] = {4, 1, 1, 1, 6};
+ constexpr int data_length = 18;
+ float input_data[data_length] = {
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 1
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 2
+ -1.1, 0.6, 0.7, 1.2, -0.7, 0.1, // batch 3
+ };
+ float expected_output[data_length] = {
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 1
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 2
+ -0.55, 0.3, 0.35, 0.6, -0.35, 0.05, // batch 3
+ };
+ int8_t quantized_input[data_length];
+ int8_t output_data[data_length];
+
+ tflite::testing::QuantizeInputData(input_data, data_length, quantized_input);
+
+ tflite::testing::TestL2Normalization<int8_t>(input_dims, quantized_input,
+ expected_output, output_data, .1);
+}
+
+TF_LITE_MICRO_TESTS_END
diff --git a/tensorflow/lite/micro/kernels/micro_ops.h b/tensorflow/lite/micro/kernels/micro_ops.h
index 650f4e8..83cccca 100644
--- a/tensorflow/lite/micro/kernels/micro_ops.h
+++ b/tensorflow/lite/micro/kernels/micro_ops.h
@@ -80,6 +80,7 @@
TfLiteRegistration* Register_SUB();
TfLiteRegistration* Register_SVDF();
TfLiteRegistration* Register_UNPACK();
+TfLiteRegistration* Register_L2_NORMALIZATION();
} // namespace micro
} // namespace ops
diff --git a/tensorflow/lite/micro/kernels/resize_nearest_neighbor_test.cc b/tensorflow/lite/micro/kernels/resize_nearest_neighbor_test.cc
index b38c7b6..518eada 100644
--- a/tensorflow/lite/micro/kernels/resize_nearest_neighbor_test.cc
+++ b/tensorflow/lite/micro/kernels/resize_nearest_neighbor_test.cc
@@ -45,13 +45,11 @@
// Input data expects a 4-D tensor of [batch, height, width, channels]
// Output data should match input datas batch and channels
// Expected sizes should be a 1-D tensor with 2 elements: new_height & new_width
-template<typename T>
-void TestResizeNearestNeighbor(const int32* input_dims_data,
- const T* input_data,
+template <typename T>
+void TestResizeNearestNeighbor(const int* input_dims_data, const T* input_data,
const int32* expected_size_data,
const T* expected_output_data,
- const int32 *output_dims_data,
- T* output_data) {
+ const int* output_dims_data, T* output_data) {
TfLiteIntArray* input_dims = IntArrayFromInts(input_dims_data);
int expected_size_dims_data[] = {2, 1, 2};
@@ -113,73 +111,73 @@
TF_LITE_MICRO_TESTS_BEGIN
TF_LITE_MICRO_TEST(HorizontalResize) {
- const int32 input_dims[] = {4, 1, 1, 2, 1};
+ const int input_dims[] = {4, 1, 1, 2, 1};
const float input_data[] = {3, 6};
const int32 expected_size_data[] = {1, 3};
const float expected_output_data[] = {3, 3, 6};
- const int32 output_dims[] = {4, 1, 1, 3, 1};
+ const int output_dims[] = {4, 1, 1, 3, 1};
float output_data[3];
tflite::testing::TestResizeNearestNeighbor<float>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(HorizontalResizeUInt8) {
- const int32 input_dims[] = {4, 1, 1, 2, 1};
+ const int input_dims[] = {4, 1, 1, 2, 1};
const uint8 input_data[] = {3, 6};
const int32 expected_size_data[] = {1, 3};
const uint8 expected_output_data[] = {3, 3, 6};
- const int32 output_dims[] = {4, 1, 1, 3, 1};
+ const int output_dims[] = {4, 1, 1, 3, 1};
uint8 output_data[3];
tflite::testing::TestResizeNearestNeighbor<uint8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(HorizontalResizeInt8) {
- const int32 input_dims[] = {4, 1, 1, 2, 1};
+ const int input_dims[] = {4, 1, 1, 2, 1};
const int8 input_data[] = {-3, 6};
const int32 expected_size_data[] = {1, 3};
const int8 expected_output_data[] = {-3, -3, 6};
- const int32 output_dims[] = {4, 1, 1, 3, 1};
+ const int output_dims[] = {4, 1, 1, 3, 1};
int8 output_data[3];
tflite::testing::TestResizeNearestNeighbor<int8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(VerticalResize) {
- const int32 input_dims[] = {4, 1, 2, 1, 1};
+ const int input_dims[] = {4, 1, 2, 1, 1};
const float input_data[] = {3, 9};
const int32 expected_size_data[] = {3, 1};
const float expected_output_data[] = {3, 3, 9};
- const int32 output_dims[] = {4, 1, 3, 1, 1};
+ const int output_dims[] = {4, 1, 3, 1, 1};
float output_data[3];
tflite::testing::TestResizeNearestNeighbor<float>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(VerticalResizeUInt8) {
- const int32 input_dims[] = {4, 1, 2, 1, 1};
+ const int input_dims[] = {4, 1, 2, 1, 1};
const uint8 input_data[] = {3, 9};
const int32 expected_size_data[] = {3, 1};
const uint8 expected_output_data[] = {3, 3, 9};
- const int32 output_dims[] = {4, 1, 3, 1, 1};
+ const int output_dims[] = {4, 1, 3, 1, 1};
uint8 output_data[3];
tflite::testing::TestResizeNearestNeighbor<uint8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(VerticalResizeInt8) {
- const int32 input_dims[] = {4, 1, 2, 1, 1};
+ const int input_dims[] = {4, 1, 2, 1, 1};
const int8 input_data[] = {3, -9};
const int32 expected_size_data[] = {3, 1};
const int8 expected_output_data[] = {3, 3, -9};
- const int32 output_dims[] = {4, 1, 3, 1, 1};
+ const int output_dims[] = {4, 1, 3, 1, 1};
int8 output_data[3];
tflite::testing::TestResizeNearestNeighbor<int8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(TwoDimensionalResize) {
- const int32 input_dims[] = {4, 1, 2, 2, 1};
+ const int input_dims[] = {4, 1, 2, 2, 1};
const float input_data[] = {3, 6, //
9, 12, //
};
@@ -189,7 +187,7 @@
9, 9, 12 //
};
- const int32 output_dims[] = {4, 1, 3, 3, 1};
+ const int output_dims[] = {4, 1, 3, 3, 1};
float output_data[9];
tflite::testing::TestResizeNearestNeighbor<float>(input_dims, input_data,
@@ -205,14 +203,14 @@
3, 3, 6, //
9, 9, 12 //
};
- const int32 output_dims[] = {4, 1, 3, 3, 1};
+ const int output_dims[] = {4, 1, 3, 3, 1};
uint8 output_data[9];
tflite::testing::TestResizeNearestNeighbor<uint8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(TwoDimensionalResizeInt8) {
- const int32 input_dims[] = {4, 1, 2, 2, 1};
+ const int input_dims[] = {4, 1, 2, 2, 1};
const int8 input_data[] = {3, -6, //
9, 12, //
};
@@ -221,14 +219,14 @@
3, 3, -6, //
9, 9, 12, //
};
- const int32 output_dims[] = {4, 1, 3, 3, 1};
+ const int output_dims[] = {4, 1, 3, 3, 1};
int8 output_data[9];
tflite::testing::TestResizeNearestNeighbor<int8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(TwoDimensionalResizeWithTwoBatches) {
- const int32 input_dims[] = {4, 2, 2, 2, 1};
+ const int input_dims[] = {4, 2, 2, 2, 1};
const float input_data[] = {3, 6, //
9, 12, //
4, 10, //
@@ -242,14 +240,14 @@
4, 4, 10, //
10, 10, 16, //
};
- const int32 output_dims[] = {4, 2, 3, 3, 1};
+ const int output_dims[] = {4, 2, 3, 3, 1};
float output_data[18];
tflite::testing::TestResizeNearestNeighbor<float>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(TwoDimensionalResizeWithTwoBatchesUInt8) {
- const int32 input_dims[] = {4, 2, 2, 2, 1};
+ const int input_dims[] = {4, 2, 2, 2, 1};
const uint8 input_data[] = {3, 6, //
9, 12, //
4, 10, //
@@ -263,14 +261,14 @@
4, 4, 10, //
10, 10, 16, //
};
- const int32 output_dims[] = {4, 2, 3, 3, 1};
+ const int output_dims[] = {4, 2, 3, 3, 1};
uint8 output_data[18];
tflite::testing::TestResizeNearestNeighbor<uint8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(TwoDimensionalResizeWithTwoBatchesInt8) {
- const int32 input_dims[] = {4, 2, 2, 2, 1};
+ const int input_dims[] = {4, 2, 2, 2, 1};
const int8 input_data[] = {3, 6, //
9, -12, //
-4, 10, //
@@ -284,7 +282,7 @@
-4, -4, 10, //
10, 10, 16, //
};
- const int32 output_dims[] = {4, 2, 3, 3, 1};
+ const int output_dims[] = {4, 2, 3, 3, 1};
int8 output_data[18];
tflite::testing::TestResizeNearestNeighbor<int8>(input_dims, input_data,
@@ -300,14 +298,14 @@
3, 4, 3, 4, 6, 10, //
9, 10, 9, 10, 12, 16, //
};
- const int32 output_dims[] = {4, 1, 3, 3, 2};
+ const int output_dims[] = {4, 1, 3, 3, 2};
float output_data[18];
tflite::testing::TestResizeNearestNeighbor<float>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(ThreeDimensionalResizeUInt8) {
- const int32 input_dims[] = {4, 1, 2, 2, 2};
+ const int input_dims[] = {4, 1, 2, 2, 2};
const uint8 input_data[] = {3, 4, 6, 10, //
10, 12, 14, 16, //
};
@@ -316,14 +314,14 @@
3, 4, 3, 4, 6, 10, //
10, 12, 10, 12, 14, 16, //
};
- const int32 output_dims[] = {4, 1, 3, 3, 2};
+ const int output_dims[] = {4, 1, 3, 3, 2};
uint8 output_data[18];
tflite::testing::TestResizeNearestNeighbor<uint8>(input_dims, input_data,
expected_size_data, expected_output_data, output_dims, output_data);
}
TF_LITE_MICRO_TEST(ThreeDimensionalResizeInt8) {
- const int32 input_dims[] = {4, 1, 2, 2, 2};
+ const int input_dims[] = {4, 1, 2, 2, 2};
const int8 input_data[] = {3, 4, -6, 10, //
10, 12, -14, 16, //
};
@@ -332,7 +330,7 @@
3, 4, 3, 4, -6, 10, //
10, 12, 10, 12, -14, 16, //
};
- const int32 output_dims[] = {4, 1, 3, 3, 2};
+ const int output_dims[] = {4, 1, 3, 3, 2};
int8 output_data[18];
tflite::testing::TestResizeNearestNeighbor<int8>(input_dims, input_data,
diff --git a/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc b/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc
index faea73e..c5e2d57 100644
--- a/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc
+++ b/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc
@@ -41,13 +41,8 @@
GreedyMemoryPlanner::GreedyMemoryPlanner(unsigned char* scratch_buffer,
int scratch_buffer_size)
: buffer_count_(0), need_to_calculate_offsets_(true) {
- const int per_buffer_size = sizeof(BufferRequirements) + // requirements_
- sizeof(int) + // buffer_sizes_sorted_by_size_
- sizeof(int) + // buffer_ids_sorted_by_size_
- sizeof(ListEntry) + // buffers_sorted_by_offset_
- sizeof(int); // buffer_offsets_;
// Allocate the arrays we need within the scratch buffer arena.
- max_buffer_count_ = scratch_buffer_size / per_buffer_size;
+ max_buffer_count_ = scratch_buffer_size / per_buffer_size();
unsigned char* next_free = scratch_buffer;
requirements_ = reinterpret_cast<BufferRequirements*>(next_free);
diff --git a/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h b/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h
index f2c77ed..0cb8109 100644
--- a/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h
+++ b/tensorflow/lite/micro/memory_planner/greedy_memory_planner.h
@@ -86,6 +86,17 @@
int next_entry_index;
};
+ // Number of bytes required in order to plan a buffer.
+ static size_t per_buffer_size() {
+ const int per_buffer_size =
+ sizeof(BufferRequirements) + // requirements_
+ sizeof(int) + // buffer_sizes_sorted_by_size_
+ sizeof(int) + // buffer_ids_sorted_by_size_
+ sizeof(ListEntry) + // buffers_sorted_by_offset_
+ sizeof(int); // buffer_offsets_;
+ return per_buffer_size;
+ }
+
private:
// Whether a buffer is active in a given time range.
bool DoesEntryOverlapInTime(const ListEntry* entry, const int first_time_used,
diff --git a/tensorflow/lite/micro/micro_allocator.cc b/tensorflow/lite/micro/micro_allocator.cc
index 1bbcadf..28de77c 100644
--- a/tensorflow/lite/micro/micro_allocator.cc
+++ b/tensorflow/lite/micro/micro_allocator.cc
@@ -44,29 +44,13 @@
// requirement for SIMD extensions.
constexpr int kBufferAlignment = 16;
-// If building with GNU clib from GCC 4.8.x or lower, `max_align_t` is not a
-// member of `std`. If using a newer version of clib, we import `max_align_t`
-// into the local anonymous namespace to be able to use it like the global
-// `max_align_t` from the older clib.
-#if defined(__GNUC__) && defined(__GNUC_PREREQ)
-#if __GNUC_PREREQ(4, 9)
-using std::max_align_t;
-#endif
-#else
-// We assume other compiler/clib configurations don't have this issue.
-using std::max_align_t;
-#endif
-
class MicroBuiltinDataAllocator : public BuiltinDataAllocator {
public:
explicit MicroBuiltinDataAllocator(SimpleMemoryAllocator* memory_allocator)
: memory_allocator_(memory_allocator) {}
- void* Allocate(size_t size) override {
- // Align to an address that is proper for all primitive types, but no more
- // than the size.
- return memory_allocator_->AllocateFromTail(
- size, std::min(size, alignof(max_align_t)));
+ void* Allocate(size_t size, size_t alignment_hint) override {
+ return memory_allocator_->AllocateFromTail(size, alignment_hint);
}
void Deallocate(void* data) override {
// Do not deallocate, builtin data needs to be available for the life time
@@ -440,6 +424,13 @@
ErrorReporter* error_reporter)
: model_(model), error_reporter_(error_reporter), context_(context) {
uint8_t* aligned_arena = AlignPointerUp(tensor_arena, kBufferAlignment);
+ if (aligned_arena != tensor_arena) {
+ TF_LITE_REPORT_ERROR(
+ error_reporter_,
+ "%d bytes lost due to alignment. To avoid this loss, please make sure "
+ "the tensor_arena is 16 bytes aligned.",
+ aligned_arena - tensor_arena);
+ }
size_t aligned_arena_size = tensor_arena + arena_size - aligned_arena;
// Creates a root memory allocator managing the arena. The allocator itself
// also locates in the arena buffer. This allocator doesn't need to be
diff --git a/tensorflow/lite/micro/micro_allocator.h b/tensorflow/lite/micro/micro_allocator.h
index e7dd4f3..a846b0c 100644
--- a/tensorflow/lite/micro/micro_allocator.h
+++ b/tensorflow/lite/micro/micro_allocator.h
@@ -64,9 +64,10 @@
// This information could change in the future version.
// ************** .memory_allocator->GetBuffer()
// Tensors/Scratch buffers (head)
-// **************
+// ************** .head_watermark
// unused memory
-// ************** .memory_allocator->GetBuffer() + ->GetDataSize()
+// ************** .memory_allocator->GetBuffer() + ->GetMaxBufferSize()
+// - ->GetDataSize()
// persistent area (tail)
// ************** .memory_allocator->GetBuffer() + ->GetMaxBufferSize()
class MicroAllocator {
@@ -88,6 +89,15 @@
// called in this class.
TfLiteStatus FinishTensorAllocation();
+ // Returns the arena usage in bytes, only available after
+ // `FinishTensorAllocation`. Otherwise, it will return 0.
+ size_t used_bytes() const {
+ if (active_) {
+ return 0;
+ }
+ return memory_allocator_->GetUsedBytes();
+ }
+
// Run through the model to allocate nodes and registrations. We need to keep
// them for the entire life time of the model to allow persistent tensors.
// This method needs to be called before FinishTensorAllocation method.
@@ -115,6 +125,7 @@
TfLiteStatus Init();
const Model* model_;
+ // A simple memory allocator that always allocate from the arena tail.
SimpleMemoryAllocator* memory_allocator_;
ErrorReporter* error_reporter_;
TfLiteContext* context_;
diff --git a/tensorflow/lite/micro/micro_allocator_test.cc b/tensorflow/lite/micro/micro_allocator_test.cc
index 03aa7c0..78419ed 100644
--- a/tensorflow/lite/micro/micro_allocator_test.cc
+++ b/tensorflow/lite/micro/micro_allocator_test.cc
@@ -142,11 +142,15 @@
TF_LITE_MICRO_TEST(TestFinishTensorAllocation) {
const tflite::Model* model = tflite::testing::GetSimpleMockModel();
TfLiteContext context;
- constexpr size_t arena_size = 1024;
+ constexpr size_t arena_size =
+ 760 /* minimal arena size at the time of writting */ +
+ 16 /* alignment */ + 100 /* leave some headroom for future proof */;
uint8_t arena[arena_size];
tflite::MicroAllocator allocator(&context, model, arena, arena_size,
micro_test::reporter);
TF_LITE_MICRO_EXPECT_EQ(4, context.tensors_size);
+ // Memory planning hasn't been finalized, so the used bytes is unknown.
+ TF_LITE_MICRO_EXPECT_EQ(0, allocator.used_bytes());
TF_LITE_MICRO_EXPECT_EQ(kTfLiteOk, allocator.FinishTensorAllocation());
// No allocation to be done afterwards.
@@ -170,6 +174,7 @@
context.tensors[1].data.raw);
TF_LITE_MICRO_EXPECT_NE(context.tensors[3].data.raw,
context.tensors[2].data.raw);
+ TF_LITE_MICRO_EXPECT_LE(allocator.used_bytes(), 760 + 100);
}
TF_LITE_MICRO_TEST(TestAllocationForModelsWithBranches) {
diff --git a/tensorflow/lite/micro/micro_interpreter.h b/tensorflow/lite/micro/micro_interpreter.h
index e41f2e3..b204612 100644
--- a/tensorflow/lite/micro/micro_interpreter.h
+++ b/tensorflow/lite/micro/micro_interpreter.h
@@ -139,6 +139,14 @@
return node_and_registrations_[node_index];
}
+ // For debugging only.
+ // Returns the actual used arena in bytes. This method gives the optimal arena
+ // size. It's only available after `AllocateTensors` has been called.
+ // Note that normally `tensor_arena` requires 16 bytes alignment to fully
+ // utilize the space. If it's not the case, the optimial arena size would be
+ // arena_used_bytes() + 16.
+ size_t arena_used_bytes() const { return allocator_.used_bytes(); }
+
private:
void CorrectTensorEndianness(TfLiteTensor* tensorCorr);
diff --git a/tensorflow/lite/micro/micro_interpreter_test.cc b/tensorflow/lite/micro/micro_interpreter_test.cc
index 9517a80..36e8c00 100644
--- a/tensorflow/lite/micro/micro_interpreter_test.cc
+++ b/tensorflow/lite/micro/micro_interpreter_test.cc
@@ -174,7 +174,9 @@
const tflite::Model* model = tflite::testing::GetSimpleMockModel();
TF_LITE_MICRO_EXPECT_NE(nullptr, model);
tflite::MockOpResolver mock_resolver;
- constexpr size_t allocator_buffer_size = 1024;
+ constexpr size_t allocator_buffer_size =
+ 928 /* optimal arena size at the time of writting. */ +
+ 16 /* alignment */ + 100 /* some headroom */;
uint8_t allocator_buffer[allocator_buffer_size];
// Create a new scope so that we can test the destructor.
@@ -183,6 +185,7 @@
allocator_buffer_size,
micro_test::reporter);
TF_LITE_MICRO_EXPECT_EQ(interpreter.AllocateTensors(), kTfLiteOk);
+ TF_LITE_MICRO_EXPECT_LE(interpreter.arena_used_bytes(), 928 + 100);
TF_LITE_MICRO_EXPECT_EQ(1, interpreter.inputs_size());
TF_LITE_MICRO_EXPECT_EQ(2, interpreter.outputs_size());
@@ -266,12 +269,15 @@
TF_LITE_MICRO_EXPECT_NE(nullptr, model);
tflite::MockOpResolver mock_resolver;
- constexpr size_t allocator_buffer_size = 2048;
+ constexpr size_t allocator_buffer_size =
+ 2096 /* optimal arena size at the time of writting. */ +
+ 16 /* alignment */ + 100 /* some headroom */;
uint8_t allocator_buffer[allocator_buffer_size];
tflite::MicroInterpreter interpreter(model, mock_resolver, allocator_buffer,
allocator_buffer_size,
micro_test::reporter);
TF_LITE_MICRO_EXPECT_EQ(interpreter.AllocateTensors(), kTfLiteOk);
+ TF_LITE_MICRO_EXPECT_LE(interpreter.arena_used_bytes(), 2096 + 100);
TF_LITE_MICRO_EXPECT_EQ(1, interpreter.inputs_size());
TF_LITE_MICRO_EXPECT_EQ(1, interpreter.outputs_size());
diff --git a/tensorflow/lite/micro/simple_memory_allocator.h b/tensorflow/lite/micro/simple_memory_allocator.h
index 42c7d96..cf18186 100644
--- a/tensorflow/lite/micro/simple_memory_allocator.h
+++ b/tensorflow/lite/micro/simple_memory_allocator.h
@@ -31,6 +31,8 @@
SimpleMemoryAllocator(ErrorReporter* error_reporter, uint8_t* buffer_head,
uint8_t* buffer_tail)
: error_reporter_(error_reporter),
+ buffer_head_(buffer_head),
+ buffer_tail_(buffer_tail),
head_(buffer_head),
tail_(buffer_tail) {}
SimpleMemoryAllocator(ErrorReporter* error_reporter, uint8_t* buffer,
@@ -47,9 +49,14 @@
uint8_t* GetHead() const { return head_; }
uint8_t* GetTail() const { return tail_; }
size_t GetAvailableMemory() const { return tail_ - head_; }
+ size_t GetUsedBytes() const { return GetBufferSize() - GetAvailableMemory(); }
private:
+ size_t GetBufferSize() const { return buffer_tail_ - buffer_head_; }
+
ErrorReporter* error_reporter_;
+ uint8_t* buffer_head_;
+ uint8_t* buffer_tail_;
uint8_t* head_;
uint8_t* tail_;
};
diff --git a/tensorflow/lite/micro/testing/micro_benchmark.h b/tensorflow/lite/micro/testing/micro_benchmark.h
index 7d8e736..f059842 100644
--- a/tensorflow/lite/micro/testing/micro_benchmark.h
+++ b/tensorflow/lite/micro/testing/micro_benchmark.h
@@ -16,6 +16,8 @@
#ifndef TENSORFLOW_LITE_MICRO_TESTING_MICRO_BENCHMARK_H_
#define TENSORFLOW_LITE_MICRO_TESTING_MICRO_BENCHMARK_H_
+#include <climits>
+
#include "tensorflow/lite/micro/micro_error_reporter.h"
#include "tensorflow/lite/micro/micro_time.h"
@@ -35,7 +37,9 @@
int32_t duration_ticks; \
int32_t duration_ms;
-#define TF_LITE_MICRO_BENCHMARKS_END }
+#define TF_LITE_MICRO_BENCHMARKS_END \
+ return 0; \
+ }
#define TF_LITE_MICRO_BENCHMARK(func) \
if (tflite::ticks_per_second() == 0) { \
@@ -44,7 +48,11 @@
start_ticks = tflite::GetCurrentTimeTicks(); \
func(); \
duration_ticks = tflite::GetCurrentTimeTicks() - start_ticks; \
- duration_ms = (duration_ticks * 1000) / tflite::ticks_per_second(); \
+ if (duration_ticks > INT_MAX / 1000) { \
+ duration_ms = duration_ticks / (tflite::ticks_per_second() / 1000); \
+ } else { \
+ duration_ms = (duration_ticks * 1000) / tflite::ticks_per_second(); \
+ } \
TF_LITE_REPORT_ERROR(micro_benchmark::reporter, "%s took %d ticks (%d ms)", \
#func, duration_ticks, duration_ms);
diff --git a/tensorflow/lite/micro/tools/ci_build/ci_build_micro_projects.sh b/tensorflow/lite/micro/tools/ci_build/ci_build_micro_projects.sh
index de5b63d..96fcc0c 100755
--- a/tensorflow/lite/micro/tools/ci_build/ci_build_micro_projects.sh
+++ b/tensorflow/lite/micro/tools/ci_build/ci_build_micro_projects.sh
@@ -26,7 +26,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
pwd
make -f tensorflow/lite/micro/tools/make/Makefile \
diff --git a/tensorflow/lite/micro/tools/ci_build/test_all.sh b/tensorflow/lite/micro/tools/ci_build/test_all.sh
index 5172f95..817a4dc 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_all.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_all.sh
@@ -21,7 +21,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
pwd
make -f tensorflow/lite/micro/tools/make/Makefile \
@@ -49,8 +49,7 @@
echo "Running x86 tests at `date`"
tensorflow/lite/micro/tools/ci_build/test_x86.sh
-# TODO(b/149597202): Disabled until we can get Docker running inside Docker.
-#echo "Running stm32f4 tests at `date`"
-#tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh
+echo "Running stm32f4 tests at `date`"
+tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh
echo "Finished all micro tests at `date`"
diff --git a/tensorflow/lite/micro/tools/ci_build/test_arc.sh b/tensorflow/lite/micro/tools/ci_build/test_arc.sh
index 24bbb3e..de8f7c5 100644
--- a/tensorflow/lite/micro/tools/ci_build/test_arc.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_arc.sh
@@ -21,7 +21,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/ci_build/test_arduino.sh b/tensorflow/lite/micro/tools/ci_build/test_arduino.sh
index 996612a..ecb821b 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_arduino.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_arduino.sh
@@ -21,7 +21,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/ci_build/test_arduino_library.sh b/tensorflow/lite/micro/tools/ci_build/test_arduino_library.sh
index 087b08d..8770ea9 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_arduino_library.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_arduino_library.sh
@@ -30,10 +30,10 @@
rm -rf ${TEMP_BUILD_DIR}
-mkdir -p ${ARDUINO_HOME_DIR}/libraries
+mkdir -p "${ARDUINO_HOME_DIR}/libraries"
mkdir -p ${TEMP_BUILD_DIR}
-unzip -q ${LIBRARY_ZIP} -d ${ARDUINO_LIBRARIES_DIR}
+unzip -q ${LIBRARY_ZIP} -d "${ARDUINO_LIBRARIES_DIR}"
# Installs all dependencies for Arduino
InstallLibraryDependencies () {
@@ -52,7 +52,7 @@
# the defines in ArduCAM/memorysaver.h are correct.
wget -O /tmp/arducam-master.zip https://github.com/ArduCAM/Arduino/archive/e216049ba304048ec9bb29adfc2cc24c16f589b1/master.zip
unzip /tmp/arducam-master.zip -d /tmp
- cp -r /tmp/Arduino-e216049ba304048ec9bb29adfc2cc24c16f589b1/ArduCAM ${ARDUINO_LIBRARIES_DIR}
+ cp -r /tmp/Arduino-e216049ba304048ec9bb29adfc2cc24c16f589b1/ArduCAM "${ARDUINO_LIBRARIES_DIR}"
}
InstallLibraryDependencies
diff --git a/tensorflow/lite/micro/tools/ci_build/test_bluepill.sh b/tensorflow/lite/micro/tools/ci_build/test_bluepill.sh
index fc0fc18..1f957e9 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_bluepill.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_bluepill.sh
@@ -20,7 +20,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
pwd
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/ci_build/test_mbed.sh b/tensorflow/lite/micro/tools/ci_build/test_mbed.sh
index 894c85a..a4d4700 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_mbed.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_mbed.sh
@@ -29,7 +29,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
pwd
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/ci_build/test_mbed_library.sh b/tensorflow/lite/micro/tools/ci_build/test_mbed_library.sh
index 3bf2155..c1ec1e6 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_mbed_library.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_mbed_library.sh
@@ -21,7 +21,7 @@
set -e
PATH=`pwd`/tensorflow/lite/micro/tools/make/downloads/gcc_embedded/bin:${PATH}
-cd ${1}
+cd "${1}"
mbed config root .
mbed deploy
diff --git a/tensorflow/lite/micro/tools/ci_build/test_sparkfun.sh b/tensorflow/lite/micro/tools/ci_build/test_sparkfun.sh
index d013022..dc08c4e 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_sparkfun.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_sparkfun.sh
@@ -20,7 +20,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh b/tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh
index 14e229c..be706a3 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_stm32f4.sh
@@ -22,7 +22,7 @@
TAGS=cmsis-nn
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
pwd
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
@@ -35,6 +35,7 @@
# Build test binaries first
readable_run make -j8 -f tensorflow/lite/micro/tools/make/Makefile TAGS=${TAGS} TARGET=${TARGET} build
+# TODO(b/149597202): Disabled until we can get Docker running inside Docker.
# Parallell builds doesn't work very well with this
-readable_run make -f tensorflow/lite/micro/tools/make/Makefile TAGS=${TAGS} TARGET=${TARGET} test
+#readable_run make -f tensorflow/lite/micro/tools/make/Makefile TAGS=${TAGS} TARGET=${TARGET} test
diff --git a/tensorflow/lite/micro/tools/ci_build/test_x86.sh b/tensorflow/lite/micro/tools/ci_build/test_x86.sh
index 48ef94a..c150d82 100755
--- a/tensorflow/lite/micro/tools/ci_build/test_x86.sh
+++ b/tensorflow/lite/micro/tools/ci_build/test_x86.sh
@@ -20,7 +20,7 @@
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
ROOT_DIR=${SCRIPT_DIR}/../../../../..
-cd ${ROOT_DIR}
+cd "${ROOT_DIR}"
source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
diff --git a/tensorflow/lite/micro/tools/make/Makefile b/tensorflow/lite/micro/tools/make/Makefile
index 8d717ae..8599a27 100644
--- a/tensorflow/lite/micro/tools/make/Makefile
+++ b/tensorflow/lite/micro/tools/make/Makefile
@@ -94,7 +94,7 @@
# runtime that can be linked in to other programs.
MICROLITE_LIB_NAME := libtensorflow-microlite.a
-MICRO_LITE_EXAMPLE_TESTS := $(wildcard tensorflow/lite/micro/examples/*/Makefile.inc)
+MICRO_LITE_EXAMPLE_TESTS := $(shell find tensorflow/lite/micro/examples/ -name Makefile.inc)
MICRO_LITE_BENCHMARKS := $(wildcard tensorflow/lite/micro/benchmarks/Makefile.inc)
@@ -159,7 +159,9 @@
tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h \
tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h \
tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h \
+tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h \
tensorflow/lite/kernels/internal/reference/integer_ops/mul.h \
+tensorflow/lite/kernels/internal/reference/l2normalization.h \
tensorflow/lite/kernels/internal/reference/maximum_minimum.h \
tensorflow/lite/kernels/internal/reference/mul.h \
tensorflow/lite/kernels/internal/reference/neg.h \
diff --git a/tensorflow/lite/micro/tools/make/download_and_extract.sh b/tensorflow/lite/micro/tools/make/download_and_extract.sh
index 97f3698..2248031 100755
--- a/tensorflow/lite/micro/tools/make/download_and_extract.sh
+++ b/tensorflow/lite/micro/tools/make/download_and_extract.sh
@@ -82,6 +82,13 @@
echo "Finished patching kissfft"
}
+# Create a header file containing an array with the first 10 images from the
+# CIFAR10 test dataset.
+patch_cifar10_dataset() {
+ xxd -l 30730 -i ${1}/test_batch.bin ${1}/../../../../examples/image_recognition_experimental/first_10_cifar_images.h
+ sed -i "s/unsigned char/const unsigned char/g" ${1}/../../../../examples/image_recognition_experimental/first_10_cifar_images.h
+}
+
build_embarc_mli() {
gmake -j 4 -C ${1}/lib/make TCF_FILE=${2}
}
@@ -102,14 +109,15 @@
command -v curl >/dev/null 2>&1 || {
echo >&2 "The required 'curl' tool isn't installed. Try 'apt-get install curl'."; exit 1;
}
-
+
echo "downloading ${url}" >&2
mkdir -p "${dir}"
# We've been seeing occasional 56 errors from valid URLs, so set up a retry
# loop to attempt to recover from them.
for (( i=1; i<=$curl_retries; ++i ))
- do
- CURL_RESULT=$(curl -Ls --retry 5 "${url}" > ${tempfile} || true)
+ do
+ curl -Ls --fail --retry 5 "${url}" > ${tempfile}
+ CURL_RESULT=$?
if [[ $CURL_RESULT -eq 0 ]]
then
break
@@ -128,7 +136,7 @@
echo "Checksum error for '${url}'. Expected ${expected_md5} but found ${DOWNLOADED_MD5}"
exit 1
fi
-
+
if [[ "${url}" == *gz ]]; then
tar -C "${dir}" --strip-components=1 -xzf ${tempfile}
elif [[ "${url}" == *tar.xz ]]; then
@@ -140,7 +148,7 @@
unzip ${tempfile} -d ${tempdir2} 2>&1 1>/dev/null
# If the zip file contains nested directories, extract the files from the
# inner directory.
- if ls ${tempdir2}/*/* 1> /dev/null 2>&1; then
+ if [ $(find $tempdir2/* -maxdepth 0 | wc -l) = 1 ] && [ -d $tempdir2/* ]; then
# unzip has no strip components, so unzip to a temp dir, and move the
# files we want from the tempdir to destination.
cp -R ${tempdir2}/*/* ${dir}/
@@ -159,6 +167,8 @@
patch_am_sdk ${dir}
elif [[ ${action} == "patch_kissfft" ]]; then
patch_kissfft ${dir}
+ elif [[ ${action} == "patch_cifar10_dataset" ]]; then
+ patch_cifar10_dataset ${dir}
elif [[ ${action} == "build_embarc_mli" ]]; then
build_embarc_mli ${dir} ${action_param1}
elif [[ ${action} ]]; then
diff --git a/tensorflow/lite/micro/tools/make/fix_arduino_subfolders.py b/tensorflow/lite/micro/tools/make/fix_arduino_subfolders.py
index 8676794..29e388c 100755
--- a/tensorflow/lite/micro/tools/make/fix_arduino_subfolders.py
+++ b/tensorflow/lite/micro/tools/make/fix_arduino_subfolders.py
@@ -80,6 +80,27 @@
source_file.write(file_contents)
+def move_image_data_experimental(library_dir):
+ """Moves the downloaded image detection model into the examples folder."""
+ old_image_data_path = os.path.join(
+ library_dir, 'src/tensorflow/lite/micro/tools/make/downloads/' +
+ 'image_recognition_model/image_recognition_model.cpp')
+ new_image_data_path = os.path.join(
+ library_dir,
+ 'examples/image_recognition_experimental/image_recognition_model.cpp')
+ if os.path.exists(old_image_data_path):
+ os.rename(old_image_data_path, new_image_data_path)
+ # Update include.
+ with open(new_image_data_path, 'r') as source_file:
+ file_contents = source_file.read()
+ file_contents = file_contents.replace(
+ six.ensure_str('#include "tensorflow/lite/micro/examples/' +
+ 'image_recognition_example/image_recognition_model.h"'),
+ '#include "image_recognition_model.h"')
+ with open(new_image_data_path, 'w') as source_file:
+ source_file.write(file_contents)
+
+
def rename_example_main_inos(library_dir):
"""Makes sure the .ino sketch files match the example name."""
search_path = os.path.join(library_dir, 'examples/*', 'main.ino')
@@ -97,6 +118,7 @@
rename_example_main_inos(library_dir)
move_person_data(library_dir)
move_person_data_experimental(library_dir)
+ move_image_data_experimental(library_dir)
def parse_args():
diff --git a/tensorflow/lite/micro/tools/make/helper_functions.inc b/tensorflow/lite/micro/tools/make/helper_functions.inc
index aee04c6..256a845 100644
--- a/tensorflow/lite/micro/tools/make/helper_functions.inc
+++ b/tensorflow/lite/micro/tools/make/helper_functions.inc
@@ -384,7 +384,9 @@
$(call generate_arc_project,make,$(MAKE_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS) $(2),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS) $(3),$(LDFLAGS) $(GENERATED_PROJECT_LIBS),$(CXXFLAGS) $(GENERATED_PROJECT_INCLUDES), $(CCFLAGS) $(GENERATED_PROJECT_INCLUDES))
$(call generate_project,mbed,$(MBED_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS) $(2),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS) $(3),$(MICROLITE_LIBS),$(CXXFLAGS),$(CCFLAGS),$(TARGET_TOOLCHAIN_ROOT),$(TARGET_TOOLCHAIN_PREFIX))
$(call generate_project,keil,$(KEIL_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS) $(2),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS) $(3),$(MICROLITE_LIBS),$(CXXFLAGS),$(CCFLAGS),$(TARGET_TOOLCHAIN_ROOT),$(TARGET_TOOLCHAIN_PREFIX))
-$(call generate_arduino_project,$(ARDUINO_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS) $(2),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS) $(3),$(MICROLITE_LIBS),$(CXXFLAGS),$(CCFLAGS))
+ifeq (,$(findstring _benchmark,$(1)))
+ $(call generate_arduino_project,$(ARDUINO_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS) $(2),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS) $(3),$(MICROLITE_LIBS),$(CXXFLAGS),$(CCFLAGS))
+endif
$(call generate_esp_project,$(ESP_PROJECT_FILES),$(1),$(MICROLITE_CC_SRCS) $(THIRD_PARTY_CC_SRCS),$(MICROLITE_CC_HDRS) $(THIRD_PARTY_CC_HDRS) $(MICROLITE_TEST_HDRS),$(2),$(3),$(MICROLITE_LIBS),$(CXXFLAGS),$(CCFLAGS),$(PROJECT_INCLUDES))
endef
diff --git a/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc b/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc
index fa20ad9..8671df5 100644
--- a/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/apollo3evb_makefile.inc
@@ -60,6 +60,7 @@
-Wmissing-field-initializers \
-Wno-unused-parameter \
-Wno-write-strings \
+ -Wunused-function \
-fno-delete-null-pointer-checks \
-fno-threadsafe-statics \
-fomit-frame-pointer \
diff --git a/tensorflow/lite/micro/tools/make/targets/bluepill_makefile.inc b/tensorflow/lite/micro/tools/make/targets/bluepill_makefile.inc
index 0b7e63c..40e6570 100644
--- a/tensorflow/lite/micro/tools/make/targets/bluepill_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/bluepill_makefile.inc
@@ -37,6 +37,7 @@
-Wmissing-field-initializers \
-Wno-unused-parameter \
-Wno-write-strings \
+ -Wunused-function \
-fno-delete-null-pointer-checks \
-fomit-frame-pointer \
-fpermissive \
diff --git a/tensorflow/lite/micro/tools/make/targets/ecm3531_makefile.inc b/tensorflow/lite/micro/tools/make/targets/ecm3531_makefile.inc
index bb6a9f3..756915f 100644
--- a/tensorflow/lite/micro/tools/make/targets/ecm3531_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/ecm3531_makefile.inc
@@ -47,6 +47,7 @@
-Wmissing-field-initializers \
-Wno-unused-parameter \
-Wno-write-strings \
+ -Wunused-function \
-fno-delete-null-pointer-checks \
-fno-threadsafe-statics \
-fomit-frame-pointer \
diff --git a/tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc b/tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc
index 85e5aa7..155fff9 100644
--- a/tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc
@@ -35,6 +35,7 @@
-Wno-sign-compare \
-Wno-unused-parameter \
-Wno-write-strings \
+ -Wunused-function \
-Wvla \
-fdata-sections \
-ffunction-sections \
diff --git a/tensorflow/lite/micro/tools/make/targets/mcu_riscv_makefile.inc b/tensorflow/lite/micro/tools/make/targets/mcu_riscv_makefile.inc
index ddd0671..079c3c1 100644
--- a/tensorflow/lite/micro/tools/make/targets/mcu_riscv_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/mcu_riscv_makefile.inc
@@ -30,6 +30,7 @@
-Wmissing-field-initializers \
-Wno-unused-parameter \
-Wno-write-strings \
+ -Wunused-function \
-fno-delete-null-pointer-checks \
-fno-threadsafe-statics \
-fomit-frame-pointer \
diff --git a/tensorflow/lite/micro/tools/make/targets/stm32f4_makefile.inc b/tensorflow/lite/micro/tools/make/targets/stm32f4_makefile.inc
index 2bad89e..7abd3cc 100644
--- a/tensorflow/lite/micro/tools/make/targets/stm32f4_makefile.inc
+++ b/tensorflow/lite/micro/tools/make/targets/stm32f4_makefile.inc
@@ -31,6 +31,7 @@
-Wno-missing-field-initializers \
-Wno-write-strings \
-Wno-sign-compare \
+ -Wunused-function \
-fno-delete-null-pointer-checks \
-fomit-frame-pointer \
-fpermissive \
@@ -71,14 +72,11 @@
tensorflow/lite/micro/kernels/elementwise_test.cc \
tensorflow/lite/micro/kernels/strided_slice_test.cc \
tensorflow/lite/micro/kernels/prelu_test.cc \
- tensorflow/lite/micro/kernels/pooling_test.cc \
tensorflow/lite/micro/kernels/pack_test.cc \
tensorflow/lite/micro/kernels/activations_test.cc \
tensorflow/lite/micro/kernels/dequantize_test.cc \
tensorflow/lite/micro/kernels/unpack_test.cc \
tensorflow/lite/micro/kernels/split_test.cc \
- tensorflow/lite/micro/kernels/conv_test.cc \
- tensorflow/lite/micro/kernels/depthwise_conv_test.cc \
tensorflow/lite/micro/simple_tensor_allocator_test.cc
MICROLITE_TEST_SRCS := $(filter-out $(EXCLUDED_TESTS), $(MICROLITE_TEST_SRCS))
EXCLUDED_EXAMPLE_TESTS := \
diff --git a/tensorflow/lite/micro/tools/make/third_party_downloads.inc b/tensorflow/lite/micro/tools/make/third_party_downloads.inc
index a31c7cf..6ec1aab 100644
--- a/tensorflow/lite/micro/tools/make/third_party_downloads.inc
+++ b/tensorflow/lite/micro/tools/make/third_party_downloads.inc
@@ -3,12 +3,20 @@
GEMMLOWP_URL := "https://github.com/google/gemmlowp/archive/719139ce755a0f31cbf1c37f7f98adcc7fc9f425.zip"
GEMMLOWP_MD5 := "7e8191b24853d75de2af87622ad293ba"
-FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.tar.gz"
-FLATBUFFERS_MD5 := "c62ffefb3d4548b127cca14ce047f16c"
+ifeq ($(HOST_OS),windows)
+ FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.zip"
+ FLATBUFFERS_MD5 := "a1afdbf114dec01a861c1b8c917d0fc7"
+else
+ FLATBUFFERS_URL := "https://github.com/google/flatbuffers/archive/v1.12.0.tar.gz"
+ FLATBUFFERS_MD5 := "c62ffefb3d4548b127cca14ce047f16c"
+endif
ifeq ($(HOST_OS),osx)
GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-mac.tar.bz2"
GCC_EMBEDDED_MD5 := "a66be9828cf3c57d7d21178e07cd8904"
+else ifeq ($(HOST_OS),windows)
+ GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-win32.zip"
+ GCC_EMBEDDED_MD5 := "bc8ae26d7c429f30d583a605a4bcf9bc"
else
GCC_EMBEDDED_URL := "https://developer.arm.com/-/media/Files/downloads/gnu-rm/7-2018q2/gcc-arm-none-eabi-7-2018-q2-update-linux.tar.bz2"
GCC_EMBEDDED_MD5 := "299ebd3f1c2c90930d28ab82e5d8d6c0"
@@ -51,6 +59,12 @@
RUY_URL="https://github.com/google/ruy/archive/9f53ba413e6fc879236dcaa3e008915973d67a4f.zip"
RUY_MD5="ce2c2444cced9dcf6ca6bc908061faa8"
+CIFAR10_DATASET_URL="https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz"
+CIFAR10_DATASET_MD5="c32a1d4ab5d03f1284b67883e8d87530"
+
+IMAGE_RECOGNITION_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/models/tflite/cifar_image_recognition_model_2020_4_14.zip"
+IMAGE_RECOGNITION_MODEL_MD5 := "2b886156e7ef4d6e53d0f1a4bc800e56"
+
PERSON_MODEL_URL := "https://storage.googleapis.com/download.tensorflow.org/data/tf_lite_micro_person_data_grayscale_2019_11_21.zip"
PERSON_MODEL_MD5 := "fe2934bd0788f1dcc7af3f0a954542ab"
diff --git a/tensorflow/lite/nnapi/NeuralNetworksTypes.h b/tensorflow/lite/nnapi/NeuralNetworksTypes.h
index a5c1a7c..851c171 100644
--- a/tensorflow/lite/nnapi/NeuralNetworksTypes.h
+++ b/tensorflow/lite/nnapi/NeuralNetworksTypes.h
@@ -173,6 +173,11 @@
ANEURALNETWORKS_UNMAPPABLE = 7,
ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE = 8,
ANEURALNETWORKS_UNAVAILABLE_DEVICE = 9,
+ ANEURALNETWORKS_MISSED_DEADLINE_TRANSIENT = 10,
+ ANEURALNETWORKS_MISSED_DEADLINE_PERSISTENT = 11,
+ ANEURALNETWORKS_RESOURCE_EXHAUSTED_TRANSIENT = 12,
+ ANEURALNETWORKS_RESOURCE_EXHAUSTED_PERSISTENT = 13,
+ ANEURALNETWORKS_DEAD_OBJECT = 14,
};
// LINT.ThenChange(//tensorflow/lite/delegates/nnapi/nnapi_delegate.cc:NnApiErrorDescription)
diff --git a/tensorflow/lite/python/convert.py b/tensorflow/lite/python/convert.py
index bf9bee0..70bea53 100644
--- a/tensorflow/lite/python/convert.py
+++ b/tensorflow/lite/python/convert.py
@@ -122,6 +122,18 @@
return wrap_toco.wrapped_experimental_mlir_quantize(input_data_str)
+def mlir_sparsify(input_data_str):
+ """Sparsify `input_data_str` to encode sparse tensor with proper format.
+
+ Args:
+ input_data_str: Input data in serialized form (e.g. a TFLITE model).
+
+ Returns:
+ Sparsified model in serialized form (e.g. a TFLITE model).
+ """
+ return wrap_toco.wrapped_experimental_mlir_sparsify(input_data_str)
+
+
def toco_convert_protos(model_flags_str,
toco_flags_str,
input_data_str,
diff --git a/tensorflow/lite/python/lite.py b/tensorflow/lite/python/lite.py
index 96f3428ef..ea4caf5 100644
--- a/tensorflow/lite/python/lite.py
+++ b/tensorflow/lite/python/lite.py
@@ -38,6 +38,7 @@
from tensorflow.lite.python.convert import build_toco_convert_protos # pylint: disable=unused-import
from tensorflow.lite.python.convert import ConverterError # pylint: disable=unused-import
from tensorflow.lite.python.convert import mlir_quantize as _mlir_quantize
+from tensorflow.lite.python.convert import mlir_sparsify as _mlir_sparsify
from tensorflow.lite.python.convert import OpsSet
from tensorflow.lite.python.convert import toco_convert # pylint: disable=unused-import
from tensorflow.lite.python.convert import toco_convert_graph_def as _toco_convert_graph_def
@@ -292,6 +293,7 @@
self._saved_model_tags = None
self._saved_model_version = None
self._saved_model_exported_names = []
+ self._experimental_sparsify_model = False
def _grappler_config(self, optimizers=None):
"""Creates a tf.compat.v1.ConfigProto for configuring Grappler.
@@ -710,6 +712,9 @@
result = self._calibrate_quantize_model(result, constants.FLOAT,
constants.FLOAT, True)
+ if self._experimental_sparsify_model:
+ result = _mlir_sparsify(result)
+
return result
@@ -1338,6 +1343,9 @@
result = self._calibrate_quantize_model(result, inference_input_type,
inference_output_type, True)
+ if self._experimental_sparsify_model:
+ result = _mlir_sparsify(result)
+
return result
def get_input_arrays(self):
diff --git a/tensorflow/lite/python/lite_test.py b/tensorflow/lite/python/lite_test.py
index 445a8b4..1fa2c6e 100644
--- a/tensorflow/lite/python/lite_test.py
+++ b/tensorflow/lite/python/lite_test.py
@@ -2174,6 +2174,14 @@
converter.convert()
self.assertValidDebugInfo(converter._debug_info)
+ def testExperimentalSparsifyModel(self):
+ self._getSequentialModel()
+
+ converter = lite.TocoConverter.from_keras_model_file(self._keras_file)
+ converter._experimental_sparsify_model = True
+ tflite_model = converter.convert()
+ self.assertTrue(tflite_model)
+
class GrapplerTest(TestModels, parameterized.TestCase):
diff --git a/tensorflow/lite/python/wrap_toco.py b/tensorflow/lite/python/wrap_toco.py
index 2d33578..b8d3fc3 100644
--- a/tensorflow/lite/python/wrap_toco.py
+++ b/tensorflow/lite/python/wrap_toco.py
@@ -46,3 +46,8 @@
def wrapped_experimental_mlir_quantize(input_data_str):
"""Wraps experimental mlir quantize model."""
return _pywrap_toco_api.ExperimentalMlirQuantizeModel(input_data_str)
+
+
+def wrapped_experimental_mlir_sparsify(input_data_str):
+ """Wraps experimental mlir sparsify model."""
+ return _pywrap_toco_api.ExperimentalMlirSparsifyModel(input_data_str)
diff --git a/tensorflow/lite/testing/BUILD b/tensorflow/lite/testing/BUILD
index 5de8a68..9d50f1a 100644
--- a/tensorflow/lite/testing/BUILD
+++ b/tensorflow/lite/testing/BUILD
@@ -222,10 +222,10 @@
"@com_google_absl//absl/strings",
"//tensorflow/lite:builtin_op_data",
"//tensorflow/lite:framework",
- "//tensorflow/lite/experimental/kernels:hashtable_op_kernels",
"//tensorflow/lite:string_util",
"//tensorflow/lite/kernels:builtin_ops",
"//tensorflow/lite/kernels:custom_ops",
+ "//tensorflow/lite/kernels/hashtable:hashtable_op_kernels",
"//tensorflow/lite/kernels:reference_ops",
"//tensorflow/lite/tools/evaluation:utils",
] + select({
diff --git a/tensorflow/lite/testing/op_tests/batch_to_space_nd.py b/tensorflow/lite/testing/op_tests/batch_to_space_nd.py
index 794bc2d..e3f0569 100644
--- a/tensorflow/lite/testing/op_tests/batch_to_space_nd.py
+++ b/tensorflow/lite/testing/op_tests/batch_to_space_nd.py
@@ -46,15 +46,6 @@
"constant_block_shape": [True],
"constant_crops": [True],
},
- # Non-4D use case: 1 batch dimension, 3 spatial dimensions, 2 others.
- {
- "dtype": [tf.float32],
- "input_shape": [[8, 2, 2, 2, 1, 1]],
- "block_shape": [[2, 2, 2]],
- "crops": [[[0, 0], [0, 0], [0, 0]]],
- "constant_block_shape": [True, False],
- "constant_crops": [True, False],
- },
# 3D use case.
{
"dtype": [tf.float32],
@@ -66,6 +57,17 @@
},
]
+ if options.run_with_flex:
+ # Non-4D use case: 1 batch dimension, 3 spatial dimensions, 2 others.
+ test_parameters = test_parameters + [{
+ "dtype": [tf.float32],
+ "input_shape": [[8, 2, 2, 2, 1, 1]],
+ "block_shape": [[2, 2, 2]],
+ "crops": [[[0, 0], [0, 0], [0, 0]]],
+ "constant_block_shape": [True, False],
+ "constant_crops": [True, False],
+ }]
+
def build_graph(parameters):
"""Build a batch_to_space graph given `parameters`."""
input_tensor = tf.compat.v1.placeholder(
diff --git a/tensorflow/lite/testing/op_tests/equal.py b/tensorflow/lite/testing/op_tests/equal.py
index 76a3fed..ddbece1 100644
--- a/tensorflow/lite/testing/op_tests/equal.py
+++ b/tensorflow/lite/testing/op_tests/equal.py
@@ -28,7 +28,7 @@
"""Make a set of tests to do equal."""
test_parameters = [{
- "input_dtype": [tf.float32, tf.int32, tf.int64],
+ "input_dtype": [tf.float32, tf.int32, tf.int64, tf.string],
"input_shape_pair": [([], []), ([1, 1, 1, 3], [1, 1, 1, 3]),
([2, 3, 4, 5], [2, 3, 4, 5]), ([2, 3, 3], [2, 3]),
([5, 5], [1]), ([10], [2, 4, 10])],
@@ -60,4 +60,4 @@
test_parameters,
build_graph,
build_inputs,
- expected_tf_failures=3)
+ expected_tf_failures=4)
diff --git a/tensorflow/lite/testing/op_tests/not_equal.py b/tensorflow/lite/testing/op_tests/not_equal.py
index 7ecf6e2..e0f9d3c 100644
--- a/tensorflow/lite/testing/op_tests/not_equal.py
+++ b/tensorflow/lite/testing/op_tests/not_equal.py
@@ -28,7 +28,7 @@
"""Make a set of tests to do not equal."""
test_parameters = [{
- "input_dtype": [tf.float32, tf.int32, tf.int64],
+ "input_dtype": [tf.float32, tf.int32, tf.int64, tf.string],
"input_shape_pair": [([1, 1, 1, 3], [1, 1, 1, 3]),
([2, 3, 4, 5], [2, 3, 4, 5]), ([2, 3, 3], [2, 3]),
([5, 5], [1]), ([10], [2, 4, 10])],
@@ -60,4 +60,4 @@
test_parameters,
build_graph,
build_inputs,
- expected_tf_failures=3)
+ expected_tf_failures=4)
diff --git a/tensorflow/lite/testing/tflite_driver.cc b/tensorflow/lite/testing/tflite_driver.cc
index 004c715..93a1b81 100644
--- a/tensorflow/lite/testing/tflite_driver.cc
+++ b/tensorflow/lite/testing/tflite_driver.cc
@@ -24,8 +24,8 @@
#if !defined(__APPLE__)
#include "tensorflow/lite/delegates/flex/delegate.h"
#endif
-#include "tensorflow/lite/experimental/kernels/hashtable_ops.h"
#include "tensorflow/lite/kernels/custom_ops_register.h"
+#include "tensorflow/lite/kernels/hashtable/hashtable_ops.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/kernels/register_ref.h"
#include "tensorflow/lite/string_util.h"
diff --git a/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc b/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
index 62a4b52..fcad8bc 100644
--- a/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
+++ b/tensorflow/lite/toco/graph_transformations/ensure_bias_vectors.cc
@@ -17,10 +17,10 @@
#include <unordered_map>
#include <vector>
+#include "tensorflow/core/platform/logging.h"
#include "tensorflow/lite/toco/graph_transformations/graph_transformations.h"
#include "tensorflow/lite/toco/model.h"
#include "tensorflow/lite/toco/tooling_util.h"
-#include "tensorflow/core/platform/logging.h"
namespace toco {
@@ -30,7 +30,8 @@
const string& weights_name = op.inputs[1];
const auto& weights_shape = model.GetArray(weights_name).shape();
if (op.type == OperatorType::kConv ||
- op.type == OperatorType::kFullyConnected) {
+ op.type == OperatorType::kFullyConnected ||
+ op.type == OperatorType::kTransposeConv) {
return weights_shape.dims(0);
}
if (op.type == OperatorType::kDepthwiseConv) {
@@ -40,8 +41,19 @@
return 0;
}
+bool CheckOpInputSize(const Operator& op) {
+ if (op.type == OperatorType::kConv ||
+ op.type == OperatorType::kFullyConnected ||
+ op.type == OperatorType::kDepthwiseConv) {
+ return (op.inputs.size() >= 3);
+ } else if (op.type == OperatorType::kTransposeConv) {
+ return (op.inputs.size() >= 4);
+ }
+ return true;
+}
+
bool ProcessLinearOperator(Model* model, Operator* op) {
- if (op->inputs.size() >= 3) {
+ if (CheckOpInputSize(*op)) {
return false;
}
const string& output_name = op->outputs[0];
@@ -52,7 +64,6 @@
const int depth = GetOutputDepthFromWeights(*model, *op);
const string& bias_name = AvailableArrayName(*model, output_name + "_bias");
op->inputs.push_back(bias_name);
- DCHECK_EQ(op->inputs.size(), 3);
auto& bias_array = model->GetOrCreateArray(bias_name);
bias_array.data_type = ArrayDataType::kFloat;
bias_array.mutable_shape()->mutable_dims()->push_back(depth);
@@ -68,7 +79,8 @@
auto* op = model->operators[op_index].get();
if (op->type == OperatorType::kConv ||
op->type == OperatorType::kDepthwiseConv ||
- op->type == OperatorType::kFullyConnected) {
+ op->type == OperatorType::kFullyConnected ||
+ op->type == OperatorType::kTransposeConv) {
if (ProcessLinearOperator(model, op)) {
AddMessageF("Added bias vector to %s as %s", LogName(*op), op->inputs[2]);
*modified = true;
diff --git a/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc b/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
index 330ce1b..05a2fec 100644
--- a/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
+++ b/tensorflow/lite/toco/graph_transformations/fuse_binary_into_preceding_affine.cc
@@ -17,16 +17,28 @@
#include <unordered_map>
#include <vector>
+#include "tensorflow/core/platform/logging.h"
#include "tensorflow/lite/toco/graph_transformations/graph_transformations.h"
#include "tensorflow/lite/toco/model.h"
#include "tensorflow/lite/toco/runtime/types.h"
#include "tensorflow/lite/toco/tooling_util.h"
-#include "tensorflow/core/platform/logging.h"
namespace toco {
namespace {
+int GetBiasIndex(const Operator& op) {
+ if (op.type == OperatorType::kConv ||
+ op.type == OperatorType::kFullyConnected ||
+ op.type == OperatorType::kDepthwiseConv) {
+ return 2;
+ } else if (op.type == OperatorType::kTransposeConv) {
+ return 3;
+ }
+ LOG(FATAL) << "Unhandled operator type";
+ return 0;
+}
+
void FuseAddOrSubParamsIntoPrecedingAffine(Model* model, Operator* preceding_op,
const Operator* add_or_sub_op,
int index_of_constant_input) {
@@ -36,7 +48,8 @@
if (preceding_op->inputs.size() < 3) {
LOG(FATAL) << "Missing bias parameter";
}
- auto& bias = model->GetArray(preceding_op->inputs[2]);
+ const auto bias_ind = GetBiasIndex(*preceding_op);
+ auto& bias = model->GetArray(preceding_op->inputs[bias_ind]);
bias.minmax = nullptr;
const auto& operand =
model->GetArray(add_or_sub_op->inputs[index_of_constant_input]);
@@ -101,7 +114,8 @@
LOG(FATAL) << "Missing bias parameter";
}
const auto& weights_name = preceding_op->inputs[1];
- const auto& bias_name = preceding_op->inputs[2];
+ const auto bias_ind = GetBiasIndex(*preceding_op);
+ const auto& bias_name = preceding_op->inputs[bias_ind];
auto& weights = model->GetArray(weights_name);
DropMinMax(model, weights_name);
auto& bias = model->GetArray(bias_name);
@@ -136,7 +150,8 @@
int output_depth;
if (preceding_op->type == OperatorType::kConv ||
- preceding_op->type == OperatorType::kFullyConnected) {
+ preceding_op->type == OperatorType::kFullyConnected ||
+ preceding_op->type == OperatorType::kTransposeConv) {
output_depth = weights_shape.dims(0);
} else if (preceding_op->type == OperatorType::kDepthwiseConv) {
output_depth = weights_shape.dims(weights_shape.dimensions_count() - 1);
@@ -253,7 +268,8 @@
if (preceding_op->type != OperatorType::kConv &&
preceding_op->type != OperatorType::kFullyConnected &&
- preceding_op->type != OperatorType::kDepthwiseConv) {
+ preceding_op->type != OperatorType::kDepthwiseConv &&
+ preceding_op->type != OperatorType::kTransposeConv) {
AddMessageF(
"Not fusing %s because the preceding %s is not of one of the supported "
"types",
@@ -261,6 +277,13 @@
return ::tensorflow::Status::OK();
}
+ if (preceding_op->type == OperatorType::kTransposeConv &&
+ binary_op->type != OperatorType::kAdd) {
+ AddMessageF("Not fusing %s to preceding %s", LogName(*binary_op),
+ LogName(*preceding_op));
+ return ::tensorflow::Status::OK();
+ }
+
if (preceding_op->fused_activation_function !=
FusedActivationFunctionType::kNone) {
AddMessageF(
@@ -278,7 +301,8 @@
}
const auto& weights_name = preceding_op->inputs[1];
- const auto& bias_name = preceding_op->inputs[2];
+ const auto bias_ind = GetBiasIndex(*preceding_op);
+ const auto& bias_name = preceding_op->inputs[bias_ind];
const auto& weights = model->GetArray(weights_name);
const auto& bias = model->GetArray(bias_name);
diff --git a/tensorflow/lite/toco/graph_transformations/quantize.cc b/tensorflow/lite/toco/graph_transformations/quantize.cc
index 421bff6..e6fd88c 100644
--- a/tensorflow/lite/toco/graph_transformations/quantize.cc
+++ b/tensorflow/lite/toco/graph_transformations/quantize.cc
@@ -244,6 +244,13 @@
weights_input_index = 1;
}
}
+ if (op.type == OperatorType::kTransposeConv) {
+ if (input_index == 3) {
+ is_bias_vector = true;
+ activations_input_index = 2;
+ weights_input_index = 1;
+ }
+ }
if (op.type == OperatorType::kLstmCell) {
if (input_index == LstmCellOperator::BIASES_INPUT) {
is_bias_vector = true;
diff --git a/tensorflow/lite/toco/graph_transformations/tests/BUILD b/tensorflow/lite/toco/graph_transformations/tests/BUILD
index 0b7b9d6..d83e97e 100644
--- a/tensorflow/lite/toco/graph_transformations/tests/BUILD
+++ b/tensorflow/lite/toco/graph_transformations/tests/BUILD
@@ -99,3 +99,15 @@
"@com_google_googletest//:gtest_main",
],
)
+
+tf_cc_test(
+ name = "fuse_binary_into_preceding_affine_test",
+ srcs = ["fuse_binary_into_preceding_affine_test.cc"],
+ deps = [
+ "//tensorflow/lite/toco:graph_transformations",
+ "//tensorflow/lite/toco:model",
+ "//tensorflow/lite/toco:tooling_util",
+ "@com_google_absl//absl/memory",
+ "@com_google_googletest//:gtest_main",
+ ],
+)
diff --git a/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc
new file mode 100644
index 0000000..b5c321c
--- /dev/null
+++ b/tensorflow/lite/toco/graph_transformations/tests/fuse_binary_into_preceding_affine_test.cc
@@ -0,0 +1,115 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include <string>
+#include <vector>
+
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+#include "tensorflow/lite/toco/graph_transformations/graph_transformations.h"
+#include "tensorflow/lite/toco/model.h"
+#include "tensorflow/lite/toco/tooling_util.h"
+
+namespace toco {
+
+namespace {
+// A gmock matcher that check that elements of a float vector match to a given
+// tolerance.
+std::vector<testing::Matcher<float>> ArrayFloatNear(
+ const std::vector<float>& values, float max_abs_error = 1e-5) {
+ std::vector<testing::Matcher<float>> matchers;
+ matchers.reserve(values.size());
+ for (const float& v : values) {
+ matchers.emplace_back(testing::FloatNear(v, max_abs_error));
+ }
+ return matchers;
+}
+} // namespace
+
+class FuseBinaryIntoPrecedingAffineTest : public ::testing::Test {
+ protected:
+ FuseBinaryIntoPrecedingAffineTest() {}
+
+ void SetUp() override { model_.reset(new Model); }
+
+ void CreateArray(const string& name, const std::vector<int>& shape) {
+ Array& array = model_->GetOrCreateArray(name);
+ array.data_type = ArrayDataType::kFloat;
+ Shape* array_shape = array.mutable_shape();
+ *(array_shape->mutable_dims()) = shape;
+ }
+
+ void CreateConstantArray(const string& name, const std::vector<int>& shape,
+ const std::vector<float>& data) {
+ CreateArray(name, shape);
+ Array& array = model_->GetOrCreateArray(name);
+ auto& array_buffer = array.GetMutableBuffer<ArrayDataType::kFloat>();
+ int bufsize = 1;
+ for (int dim : shape) {
+ bufsize *= dim;
+ }
+ array_buffer.data.resize(bufsize);
+ float* buf_ptr = array_buffer.data.data();
+ for (int i = 0; i < bufsize; ++i) {
+ buf_ptr[i] = data[i];
+ }
+ }
+
+ std::unique_ptr<Model> model_;
+};
+
+TEST_F(FuseBinaryIntoPrecedingAffineTest, FuseAddIntoTransposeConv) {
+ // Creating a model.
+ {
+ CreateConstantArray(/*name=*/"OutputShape",
+ /*shape=*/{1, 2}, /*data=*/{2, 2});
+ CreateConstantArray("TransConvWeight", {2, 2}, {1.0, 2.0, 3.0, 4.0});
+ CreateConstantArray("TransConvBias", {1}, {1.0});
+ CreateArray(/*name=*/"TransConvInput",
+ /*shape=*/{2, 2});
+ CreateArray("TransConvOutput", {2, 2});
+ CreateConstantArray("AddInput2", {1}, {2.0});
+ CreateArray("AddOutput", {2, 2});
+
+ auto* tc_op = new TransposeConvOperator;
+ tc_op->inputs = {"OutputShape", "TransConvWeight", "TransConvInput",
+ "TransConvBias"};
+ tc_op->outputs = {"TransConvOutput"};
+ model_->operators.push_back(std::unique_ptr<Operator>(tc_op));
+
+ auto* add_op = new AddOperator;
+ add_op->inputs = {"TransConvOutput", "AddInput2"};
+ add_op->outputs = {"AddOutput"};
+ model_->operators.push_back(std::unique_ptr<Operator>(add_op));
+ }
+ toco::FuseBinaryIntoPrecedingAffine transformation;
+ bool modified;
+ ASSERT_TRUE(transformation.Run(model_.get(), /*op_index=*/1, &modified).ok());
+ EXPECT_TRUE(modified);
+
+ // `Add` should be fused into `TransposeConv`. Only 1 op is left.
+ ASSERT_EQ(model_->operators.size(), 1);
+ const auto& op = model_->operators[0];
+ ASSERT_EQ(op->type, OperatorType::kTransposeConv);
+ ASSERT_EQ(op->inputs.size(), 4);
+
+ auto& weights_array = model_->GetArray(op->inputs[1]);
+ EXPECT_THAT(weights_array.GetBuffer<toco::ArrayDataType::kFloat>().data,
+ ElementsAreArray(ArrayFloatNear({1.0, 2.0, 3.0, 4.0})));
+
+ auto& bias_array = model_->GetArray(op->inputs[3]);
+ EXPECT_THAT(bias_array.GetBuffer<toco::ArrayDataType::kFloat>().data,
+ ElementsAreArray(ArrayFloatNear({3.0})));
+}
+} // namespace toco
diff --git a/tensorflow/lite/toco/model.h b/tensorflow/lite/toco/model.h
index 11a4003..7207496 100644
--- a/tensorflow/lite/toco/model.h
+++ b/tensorflow/lite/toco/model.h
@@ -1200,6 +1200,8 @@
// inputs[0]: required: the output shape
// inputs[1]: required: the weights
// inputs[2]: required: the input activations array
+// inputs[3]: optional: the bias vector, specifying the biases for each output
+// channel.
// NOTE: The input activations is NOT the first input.
//
//
@@ -1212,6 +1214,7 @@
OUTPUT_SHAPE = 0,
WEIGHTS = 1,
DATA_INPUT = 2,
+ BIAS = 3,
};
TransposeConvOperator() : Operator(OperatorType::kTransposeConv) {}
diff --git a/tensorflow/lite/toco/python/BUILD b/tensorflow/lite/toco/python/BUILD
index 495c403..bea582d 100644
--- a/tensorflow/lite/toco/python/BUILD
+++ b/tensorflow/lite/toco/python/BUILD
@@ -53,6 +53,7 @@
"//tensorflow/compiler/mlir/lite/python:graphdef_to_tfl_flatbuffer",
"//tensorflow/compiler/mlir/lite/python:saved_model_to_tfl_flatbuffer",
"//tensorflow/compiler/mlir/lite/quantization/lite:quantize_model",
+ "//tensorflow/compiler/mlir/lite/sparsity:sparsify_model",
] + select({
# This is required when running `tflite_convert` from `bazel`.
# It requires to link with TensorFlow Ops to get the op definitions.
diff --git a/tensorflow/lite/toco/python/toco_python_api.cc b/tensorflow/lite/toco/python/toco_python_api.cc
index dd21e81..a19f5d2 100644
--- a/tensorflow/lite/toco/python/toco_python_api.cc
+++ b/tensorflow/lite/toco/python/toco_python_api.cc
@@ -23,6 +23,7 @@
#include "tensorflow/compiler/mlir/lite/python/graphdef_to_tfl_flatbuffer.h"
#include "tensorflow/compiler/mlir/lite/python/saved_model_to_tfl_flatbuffer.h"
#include "tensorflow/compiler/mlir/lite/quantization/lite/quantize_model.h"
+#include "tensorflow/compiler/mlir/lite/sparsity/sparsify_model.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/core/api/error_reporter.h"
@@ -262,4 +263,37 @@
builder.GetSize());
}
+PyObject* MlirSparsifyModel(PyObject* data) {
+ using tflite::interpreter_wrapper::PythonErrorReporter;
+ char* buf = nullptr;
+ Py_ssize_t length;
+ std::unique_ptr<PythonErrorReporter> error_reporter(new PythonErrorReporter);
+
+ if (tflite::python_utils::ConvertFromPyString(data, &buf, &length) == -1) {
+ PyErr_Format(PyExc_ValueError, "Failed to convert input PyObject");
+ return nullptr;
+ }
+ std::unique_ptr<tflite::FlatBufferModel> model =
+ tflite::FlatBufferModel::BuildFromBuffer(buf, length,
+ error_reporter.get());
+ if (!model) {
+ PyErr_Format(PyExc_ValueError, "Invalid model");
+ return nullptr;
+ }
+ auto tflite_model = absl::make_unique<tflite::ModelT>();
+ model->GetModel()->UnPackTo(tflite_model.get(), nullptr);
+
+ flatbuffers::FlatBufferBuilder builder;
+ auto status =
+ mlir::lite::SparsifyModel(*tflite_model, &builder, error_reporter.get());
+
+ if (status != kTfLiteOk) {
+ error_reporter->exception();
+ return nullptr;
+ }
+ return tflite::python_utils::ConvertToPyString(
+ reinterpret_cast<const char*>(builder.GetCurrentBufferPointer()),
+ builder.GetSize());
+}
+
} // namespace toco
diff --git a/tensorflow/lite/toco/python/toco_python_api.h b/tensorflow/lite/toco/python/toco_python_api.h
index ca67e3f..c7c7a35 100644
--- a/tensorflow/lite/toco/python/toco_python_api.h
+++ b/tensorflow/lite/toco/python/toco_python_api.h
@@ -44,6 +44,10 @@
// is specified by the calibration data are not sufficient to quantize the
// model.
PyObject* MlirQuantizeModel(PyObject* data, bool fully_quantize);
+
+// Sparsifies model to encode sparse tensors with proper format. Throws error if
+// sparsification fails.
+PyObject* MlirSparsifyModel(PyObject* data);
} // namespace toco
#endif // TENSORFLOW_LITE_TOCO_PYTHON_TOCO_PYTHON_API_H_
diff --git a/tensorflow/lite/toco/tflite/op_version.cc b/tensorflow/lite/toco/tflite/op_version.cc
index c5b5e3c..1b259b7 100644
--- a/tensorflow/lite/toco/tflite/op_version.cc
+++ b/tensorflow/lite/toco/tflite/op_version.cc
@@ -148,6 +148,8 @@
{{OperatorType::kArgMin, 1}, "1.9.0"},
{{OperatorType::kArgMin, 2}, "1.14.0"},
{{OperatorType::kTransposeConv, 1}, "1.9.0"},
+ {{OperatorType::kTransposeConv, 2}, kPendingReleaseOpVersion},
+ {{OperatorType::kTransposeConv, 3}, kPendingReleaseOpVersion},
{{OperatorType::kSparseToDense, 1}, "1.9.0"},
{{OperatorType::kSparseToDense, 2}, "1.14.0"},
{{OperatorType::kSparseToDense, 3}, "1.15.0"},
@@ -183,8 +185,10 @@
{{OperatorType::kReverseSequence, 1}, "1.14.0"},
{{OperatorType::kEqual, 1}, "1.14.0"},
{{OperatorType::kEqual, 2}, "1.14.0"},
+ {{OperatorType::kEqual, 3}, kPendingReleaseOpVersion},
{{OperatorType::kNotEqual, 1}, "1.14.0"},
{{OperatorType::kNotEqual, 2}, "1.14.0"},
+ {{OperatorType::kNotEqual, 3}, kPendingReleaseOpVersion},
{{OperatorType::kGreater, 1}, "1.14.0"},
{{OperatorType::kGreater, 2}, "1.14.0"},
{{OperatorType::kGreaterEqual, 1}, "1.14.0"},
diff --git a/tensorflow/lite/tools/BUILD b/tensorflow/lite/tools/BUILD
index 106b79b..a96c1c3 100644
--- a/tensorflow/lite/tools/BUILD
+++ b/tensorflow/lite/tools/BUILD
@@ -1,5 +1,4 @@
load("//tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite")
-load("//tensorflow:tensorflow.bzl", "tf_cc_binary")
load("//tensorflow/lite:build_def.bzl", "tflite_copts")
package(
@@ -126,13 +125,12 @@
],
)
-tf_cc_binary(
+cc_binary(
name = "generate_op_registrations",
srcs = ["gen_op_registration_main.cc"],
deps = [
+ ":command_line_flags",
":gen_op_registration",
- "//tensorflow/core:framework_internal",
- "//tensorflow/core:lib",
"@com_google_absl//absl/strings",
],
)
diff --git a/tensorflow/lite/tools/benchmark/BUILD b/tensorflow/lite/tools/benchmark/BUILD
index 0e0c3ab..3570722 100644
--- a/tensorflow/lite/tools/benchmark/BUILD
+++ b/tensorflow/lite/tools/benchmark/BUILD
@@ -98,13 +98,13 @@
deps = [
":benchmark_performance_options",
":benchmark_tflite_model_lib",
- ":delegate_provider_hdr",
"//tensorflow/lite:framework",
"//tensorflow/lite:string_util",
"//tensorflow/lite/c:common",
"//tensorflow/lite/testing:util",
"//tensorflow/lite/tools:command_line_flags",
"//tensorflow/lite/tools:logging",
+ "//tensorflow/lite/tools/delegates:delegate_provider_hdr",
"@com_google_absl//absl/algorithm",
"@com_google_absl//absl/memory",
"@com_google_absl//absl/strings:str_format",
@@ -139,10 +139,7 @@
deps = [
":benchmark_model_lib",
":benchmark_utils",
- ":coreml_delegate_provider",
- ":delegate_provider_hdr",
":profiling_listener",
- ":tflite_execution_providers",
"//tensorflow/lite:framework",
"//tensorflow/lite:string_util",
"//tensorflow/lite/kernels:builtin_ops",
@@ -150,6 +147,8 @@
"//tensorflow/lite/profiling:profile_summary_formatter",
"//tensorflow/lite/profiling:profiler",
"//tensorflow/lite/tools:logging",
+ "//tensorflow/lite/tools/delegates:delegate_provider_hdr",
+ "//tensorflow/lite/tools/delegates:tflite_execution_providers",
"//tensorflow/lite/tools/evaluation:utils",
"@com_google_absl//absl/base:core_headers",
"@com_google_absl//absl/strings",
@@ -211,152 +210,6 @@
)
cc_library(
- name = "delegate_provider_hdr",
- hdrs = [
- "delegate_provider.h",
- ],
- copts = common_copts,
- deps = [
- ":benchmark_params",
- "//tensorflow/lite/c:common",
- "//tensorflow/lite/tools:command_line_flags",
- "//tensorflow/lite/tools:logging",
- ],
-)
-
-# A convenient library for all inference execution providers.
-cc_library(
- name = "tflite_execution_providers",
- copts = tflite_copts(),
- deps = [
- ":default_execution_provider",
- ":external_delegate_provider",
- ":gpu_delegate_provider",
- ":hexagon_delegate_provider",
- ":nnapi_delegate_provider",
- ] + select({
- "//tensorflow:fuchsia": [],
- "//tensorflow:windows": [],
- "//conditions:default": [
- ":xnnpack_delegate_provider",
- ],
- }),
- alwayslink = 1,
-)
-
-cc_library(
- name = "default_execution_provider",
- srcs = ["default_execution_provider.cc"],
- copts = tflite_copts(),
- linkstatic = True,
- visibility = ["//visibility:public"],
- deps = [
- ":delegate_provider_hdr",
- ],
- alwayslink = 1,
-)
-
-cc_library(
- name = "gpu_delegate_provider",
- srcs = ["gpu_delegate_provider.cc"],
- copts = common_copts + select({
- "//tensorflow:ios": [
- "-xobjective-c++",
- ],
- "//conditions:default": [],
- }),
- deps = [
- ":delegate_provider_hdr",
- "//tensorflow/lite/tools/evaluation:utils",
- ] + select({
- "//tensorflow:android": [
- "//tensorflow/lite/delegates/gpu:delegate",
- ],
- "//tensorflow:ios": [
- "//tensorflow/lite/delegates/gpu:metal_delegate",
- ],
- "//conditions:default": [],
- }),
- alwayslink = 1,
-)
-
-cc_library(
- name = "nnapi_delegate_provider",
- srcs = ["nnapi_delegate_provider.cc"],
- copts = common_copts,
- deps = [
- ":delegate_provider_hdr",
- "//tensorflow/lite/tools/evaluation:utils",
- ],
- alwayslink = 1,
-)
-
-cc_library(
- name = "hexagon_delegate_provider",
- srcs = ["hexagon_delegate_provider.cc"],
- copts = common_copts,
- deps = [
- ":delegate_provider_hdr",
- "//tensorflow/lite/tools/evaluation:utils",
- ] + select({
- "//tensorflow:android_arm": [
- "//tensorflow/lite/experimental/delegates/hexagon:hexagon_delegate",
- ],
- "//tensorflow:android_arm64": [
- "//tensorflow/lite/experimental/delegates/hexagon:hexagon_delegate",
- ],
- "//conditions:default": [],
- }),
- alwayslink = 1,
-)
-
-cc_library(
- name = "coreml_delegate_provider",
- srcs = ["coreml_delegate_provider.cc"],
- copts = common_copts + select({
- "//tensorflow:ios": [
- "-xobjective-c++",
- ],
- "//conditions:default": [],
- }),
- deps = [
- ":delegate_provider_hdr",
- "//tensorflow/lite/tools/evaluation:utils",
- ] + select({
- "//tensorflow:ios": [
- "//tensorflow/lite/experimental/delegates/coreml:coreml_delegate",
- ],
- "//conditions:default": [],
- }),
- alwayslink = 1,
-)
-
-cc_library(
- name = "xnnpack_delegate_provider",
- srcs = ["xnnpack_delegate_provider.cc"],
- copts = tflite_copts(),
- linkstatic = True,
- visibility = ["//visibility:public"],
- deps = [
- ":delegate_provider_hdr",
- "//tensorflow/lite/tools/evaluation:utils",
- ],
- alwayslink = 1,
-)
-
-cc_library(
- name = "external_delegate_provider",
- srcs = ["external_delegate_provider.cc"],
- copts = tflite_copts(),
- linkstatic = True,
- visibility = ["//visibility:public"],
- deps = [
- ":delegate_provider_hdr",
- ],
- alwayslink = 1,
-)
-
-cc_library(
name = "benchmark_utils",
srcs = [
"benchmark_utils.cc",
diff --git a/tensorflow/lite/tools/benchmark/README.md b/tensorflow/lite/tools/benchmark/README.md
index 70728f4..a4f632c 100644
--- a/tensorflow/lite/tools/benchmark/README.md
+++ b/tensorflow/lite/tools/benchmark/README.md
@@ -34,32 +34,6 @@
* `run_delay`: `float` (default=-1.0) \
The delay in seconds between subsequent benchmark runs. Non-positive values
mean use no delay.
-* `use_xnnpack`: `bool` (default=false) \
- Whether to use the XNNPack delegate.
-* `use_hexagon`: `bool` (default=false) \
- Whether to use the Hexagon delegate. Not all devices may support the Hexagon
- delegate, refer to the TensorFlow Lite documentation for more information
- about which devices/chipsets are supported and about how to get the required
- libraries. To use the Hexagon delegate also build the
- hexagon_nn:libhexagon_interface.so target and copy the library to the
- device. All libraries should be copied to /data/local/tmp on the device.
-* `use_nnapi`: `bool` (default=false) \
- Whether to use
- [Android NNAPI](https://developer.android.com/ndk/guides/neuralnetworks/).
- This API is available on recent Android devices. Note that some Android P
- devices will fail to use NNAPI for models in `/data/local/tmp/` and this
- benchmark tool will not correctly use NNAPI. When on Android Q+, will also
- print the names of NNAPI accelerators accessible through the
- `nnapi_accelerator_name` flag.
-* `nnapi_accelerator_name`: `str` (default="") \
- The name of the NNAPI accelerator to use (requires Android Q+). If left
- blank, NNAPI will automatically select which of the available accelerators
- to use.
-* `nnapi_execution_preference`: `string` (default="") \
- Which
- [NNAPI execution preference](https://developer.android.com/ndk/reference/group/neural-networks.html#group___neural_networks_1gga034380829226e2d980b2a7e63c992f18af727c25f1e2d8dcc693c477aef4ea5f5)
- to use when executing using NNAPI. Should be one of the following:
- fast_single_answer, sustained_speed, low_power, undefined.
* `use_legacy_nnapi`: `bool` (default=false) \
Whether to use the legacy
[Android NNAPI](https://developer.android.com/ndk/guides/neuralnetworks/)
@@ -67,39 +41,6 @@
This is available on recent Android devices. Note that some Android P
devices will fail to use NNAPI for models in `/data/local/tmp/` and this
benchmark tool will not correctly use NNAPI.
-* `max_delegated_partitions`: `int` (default=0, i.e. no limit) \
- The maximum number of partitions that will be delegated. \
- Currently supported by the Hexagon delegate or the NNAPI delegate but won't
- work if `use_legacy_nnapi` has been selected.
-* `min_nodes_per_partition`: `int` (default=0, i.e. default choice implemented
- by each delegate) \
- The minimal number of TFLite graph nodes of a partition that needs to be
- reached to be delegated. A negative value or 0 means to use the default
- choice of each delegate. \
- This option is currently only supported by the Hexagon delegate.
-* `disable_nnapi_cpu`: `bool` (default=false) \
- Excludes the
- [NNAPI CPU reference implementation](https://developer.android.com/ndk/guides/neuralnetworks#device-assignment)
- from the possible devices to be used by NNAPI to execute the model. This
- option is ignored if `nnapi_accelerator_name` is specified.
-* `use_gpu`: `bool` (default=false) \
- Whether to use the
- [GPU accelerator delegate](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/delegates/gpu).
- This option is currently only available on Android and iOS devices.
-* `gpu_precision_loss_allowed`: `bool` (default=true) \
- Whethre to allow the GPU delegate to carry out computation with some
- precision loss (i.e. processing in FP16) or not. If allowed, the performance
- will increase.
-* `gpu_experimental_enable_quant`: `bool` (default=true) \
- Whether to allow the GPU delegate to run a quantized model or not. This
- option is currently only available on Android.
-* `gpu_wait_type`: `str` (default="") \
- Which GPU wait_type option to use, when using GPU delegate on iOS. Should be
- one of the following: passive, active, do_not_wait, aggressive. When left
- blank, passive mode is used by default.
-* `use_coreml`: `bool` (default=false) \
- Whether to use the [Core ML delegate](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/delegates/coreml).
- This option is only available in iOS.
* `enable_op_profiling`: `bool` (default=false) \
Whether to enable per-operator profiling measurement.
* `enable_platform_tracing`: `bool` (default=false) \
@@ -107,16 +48,49 @@
'enable_op_profiling'. Note, the platform-wide tracing might not work if the
tool runs as a commandline native binary. For example, on Android, the
ATrace-based tracing only works when the tool is launched as an APK.
-* `hexagon_profiling`: `bool` (default=false) \
- Whether to profile ops running on hexagon. Needs to be combined with
- `enable_op_profiling`. When this is set to true the profile of ops on
- hexagon DSP will be added to the profile table. Note that, the reported data
- on hexagon is in cycles, not in ms like on cpu.
-* `external_delegate_path`: `string` (default="") \
- Path to the external delegate library to use.
-* `external_delegate_options`: `string` (default="") \
- A list of options to be passed to the external delegate library. Options
- should be in the format of `option1:value1;option2:value2;optionN:valueN`
+
+### TFLite delegate parameters
+The tool supports all runtime/delegate parameters introduced by
+[the delegate registrar](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/delegates).
+The following simply lists the names of these parameters and additional notes
+where applicable. For details about each parameter, please refer to
+[this page](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tools/delegates/README.md#tflite-delegate-registrar).
+#### Common parameters
+* `max_delegated_partitions`: `int` (default=0) \
+Note when `use_legacy_nnapi` is selected, this parameter won't work.
+* `min_nodes_per_partition`:`int` (default=0)
+
+#### GPU delegate
+* `use_gpu`: `bool` (default=false)
+* `gpu_precision_loss_allowed`: `bool` (default=true)
+* `gpu_experimental_enable_quant`: `bool` (default=true)
+* `gpu_backend`: `string` (default="")
+* `gpu_wait_type`: `str` (default="")
+
+#### NNAPI delegate
+* `use_nnapi`: `bool` (default=false) \
+ Note some Android P devices will fail to use NNAPI for models in
+ `/data/local/tmp/` and this benchmark tool will not correctly use NNAPI.
+* `nnapi_accelerator_name`: `str` (default="")
+* `disable_nnapi_cpu`: `bool` (default=false)
+
+#### Hexagon delegate
+* `use_hexagon`: `bool` (default=false)
+* `hexagon_profiling`: `bool` (default=false) \
+Note enabling this option will not produce profiling results outputs unless
+`enable_op_profiling` is also turned on. When both parameters are set to true,
+the profile of ops on hexagon DSP will be added to the profile table. Note that,
+the reported data on hexagon is in cycles, not in ms like on cpu.
+
+#### XNNPACK delegate
+* `use_xnnpack`: `bool` (default=false)
+
+#### CoreML delegate
+* `use_coreml`: `bool` (default=false)
+
+#### External delegate
+* `external_delegate_path`: `string` (default="")
+* `external_delegate_options`: `string` (default="")
## To build/install/run
diff --git a/tensorflow/lite/tools/benchmark/benchmark_model.h b/tensorflow/lite/tools/benchmark/benchmark_model.h
index 0aca42d..912e54f 100644
--- a/tensorflow/lite/tools/benchmark/benchmark_model.h
+++ b/tensorflow/lite/tools/benchmark/benchmark_model.h
@@ -161,7 +161,7 @@
const std::string& usage) {
return Flag(
name, [params, name](const T& val) { params->Set<T>(name, val); },
- params->Get<T>(name), usage, Flag::OPTIONAL);
+ params->Get<T>(name), usage, Flag::kOptional);
}
// Benchmarks a model.
diff --git a/tensorflow/lite/tools/benchmark/benchmark_test.cc b/tensorflow/lite/tools/benchmark/benchmark_test.cc
index 8d7fa8d..33ccacc 100644
--- a/tensorflow/lite/tools/benchmark/benchmark_test.cc
+++ b/tensorflow/lite/tools/benchmark/benchmark_test.cc
@@ -29,8 +29,8 @@
#include "tensorflow/lite/testing/util.h"
#include "tensorflow/lite/tools/benchmark/benchmark_performance_options.h"
#include "tensorflow/lite/tools/benchmark/benchmark_tflite_model.h"
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
#include "tensorflow/lite/tools/command_line_flags.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/logging.h"
namespace {
@@ -88,7 +88,8 @@
params.AddParam("enable_platform_tracing",
BenchmarkParam::Create<bool>(false));
- for (const auto& delegate_provider : GetRegisteredDelegateProviders()) {
+ for (const auto& delegate_provider :
+ tools::GetRegisteredDelegateProviders()) {
params.Merge(delegate_provider->DefaultParams());
}
return params;
diff --git a/tensorflow/lite/tools/benchmark/benchmark_tflite_model.cc b/tensorflow/lite/tools/benchmark/benchmark_tflite_model.cc
index dc158c1..489780e 100644
--- a/tensorflow/lite/tools/benchmark/benchmark_tflite_model.cc
+++ b/tensorflow/lite/tools/benchmark/benchmark_tflite_model.cc
@@ -36,8 +36,8 @@
#include "tensorflow/lite/profiling/profile_summary_formatter.h"
#include "tensorflow/lite/string_util.h"
#include "tensorflow/lite/tools/benchmark/benchmark_utils.h"
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
#include "tensorflow/lite/tools/benchmark/profiling_listener.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/logging.h"
void RegisterSelectedOps(::tflite::MutableOpResolver* resolver);
@@ -272,8 +272,9 @@
default_params.AddParam("enable_platform_tracing",
BenchmarkParam::Create<bool>(false));
- for (const auto& delegate_util : GetRegisteredDelegateProviders()) {
- default_params.Merge(delegate_util->DefaultParams());
+ for (const auto& delegate_provider :
+ tools::GetRegisteredDelegateProviders()) {
+ default_params.Merge(delegate_provider->DefaultParams());
}
return default_params;
@@ -332,8 +333,9 @@
flags.insert(flags.end(), specific_flags.begin(), specific_flags.end());
- for (const auto& delegate_util : GetRegisteredDelegateProviders()) {
- auto delegate_flags = delegate_util->CreateFlags(¶ms_);
+ for (const auto& delegate_provider :
+ tools::GetRegisteredDelegateProviders()) {
+ auto delegate_flags = delegate_provider->CreateFlags(¶ms_);
flags.insert(flags.end(), delegate_flags.begin(), delegate_flags.end());
}
@@ -372,8 +374,9 @@
TFLITE_LOG(INFO) << "Enable platform-wide tracing: ["
<< params_.Get<bool>("enable_platform_tracing") << "]";
- for (const auto& delegate_util : GetRegisteredDelegateProviders()) {
- delegate_util->LogParams(params_);
+ for (const auto& delegate_provider :
+ tools::GetRegisteredDelegateProviders()) {
+ delegate_provider->LogParams(params_);
}
}
@@ -615,7 +618,8 @@
interpreter_->SetAllowFp16PrecisionForFp32(params_.Get<bool>("allow_fp16"));
owned_delegates_.clear();
- for (const auto& delegate_provider : GetRegisteredDelegateProviders()) {
+ for (const auto& delegate_provider :
+ tools::GetRegisteredDelegateProviders()) {
auto delegate = delegate_provider->CreateTfLiteDelegate(params_);
// It's possible that a delegate of certain type won't be created as
// user-specified benchmark params tells not to.
diff --git a/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h b/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
index 39ec547..81ba071 100644
--- a/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
+++ b/tensorflow/lite/tools/benchmark/experimental/c/c_api_types.h
@@ -41,7 +41,11 @@
extern "C" {
#endif // __cplusplus
-typedef enum TfLiteStatus { kTfLiteOk = 0, kTfLiteError = 1 } TfLiteStatus;
+typedef enum TfLiteStatus {
+ kTfLiteOk = 0,
+ kTfLiteError = 1,
+ kTfLiteDelegateError = 2
+} TfLiteStatus;
// The list of external context types known to TF Lite. This list exists solely
// to avoid conflicts and to ensure ops can share the external contexts they
@@ -178,8 +182,9 @@
#define TF_LITE_ENSURE_STATUS(a) \
do { \
- if ((a) != kTfLiteOk) { \
- return kTfLiteError; \
+ const TfLiteStatus s = (a); \
+ if (s != kTfLiteOk) { \
+ return s; \
} \
} while (0)
@@ -208,8 +213,9 @@
#define TF_LITE_ENSURE_OK(context, status) \
do { \
- if ((status) != kTfLiteOk) { \
- return kTfLiteError; \
+ const TfLiteStatus s = (status); \
+ if ((s) != kTfLiteOk) { \
+ return s; \
} \
} while (0)
diff --git a/tensorflow/lite/tools/command_line_flags.cc b/tensorflow/lite/tools/command_line_flags.cc
index 0db2d53..92ddb16 100644
--- a/tensorflow/lite/tools/command_line_flags.cc
+++ b/tensorflow/lite/tools/command_line_flags.cc
@@ -142,7 +142,7 @@
flag_type_(flag_type) {}
bool Flag::Parse(const std::string& arg, bool* value_parsing_ok) const {
- return ParseFlag(arg, name_, flag_type_ == POSITIONAL, value_hook_,
+ return ParseFlag(arg, name_, flag_type_ == kPositional, value_hook_,
value_parsing_ok);
}
@@ -195,7 +195,7 @@
result = false;
}
continue;
- } else if (flag.flag_type_ == Flag::REQUIRED) {
+ } else if (flag.flag_type_ == Flag::kRequired) {
TFLITE_LOG(ERROR) << "Required flag not provided: " << flag.name_;
// If the required flag isn't found, we immediately stop the whole flag
// parsing.
@@ -205,7 +205,7 @@
}
// Parses positional flags.
- if (flag.flag_type_ == Flag::POSITIONAL) {
+ if (flag.flag_type_ == Flag::kPositional) {
if (++positional_count >= *argc) {
TFLITE_LOG(ERROR) << "Too few command line arguments.";
return false;
@@ -245,7 +245,7 @@
// The flag isn't found, do some bookkeeping work.
processed_flags[flag.name_] = -1;
- if (flag.flag_type_ == Flag::REQUIRED) {
+ if (flag.flag_type_ == Flag::kRequired) {
TFLITE_LOG(ERROR) << "Required flag not provided: " << flag.name_;
result = false;
// If the required flag isn't found, we immediately stop the whole flag
@@ -280,7 +280,7 @@
// Prints usage for positional flag.
for (int i = 0; i < sorted_idx.size(); ++i) {
const Flag& flag = flag_list[sorted_idx[i]];
- if (flag.flag_type_ == Flag::POSITIONAL) {
+ if (flag.flag_type_ == Flag::kPositional) {
positional_count++;
usage_text << " <" << flag.name_ << ">";
} else {
@@ -295,7 +295,7 @@
std::vector<std::string> name_column(flag_list.size());
for (int i = 0; i < sorted_idx.size(); ++i) {
const Flag& flag = flag_list[sorted_idx[i]];
- if (flag.flag_type_ != Flag::POSITIONAL) {
+ if (flag.flag_type_ != Flag::kPositional) {
name_column[i] += "--";
name_column[i] += flag.name_;
name_column[i] += "=";
@@ -320,7 +320,8 @@
usage_text << "\t";
usage_text << std::left << std::setw(max_name_width) << name_column[i];
usage_text << "\t" << type_name << "\t";
- usage_text << (flag.flag_type_ != Flag::OPTIONAL ? "required" : "optional");
+ usage_text << (flag.flag_type_ != Flag::kOptional ? "required"
+ : "optional");
usage_text << "\t" << flag.usage_text_ << "\n";
}
return usage_text.str();
diff --git a/tensorflow/lite/tools/command_line_flags.h b/tensorflow/lite/tools/command_line_flags.h
index 941a1b8..95e64a1 100644
--- a/tensorflow/lite/tools/command_line_flags.h
+++ b/tensorflow/lite/tools/command_line_flags.h
@@ -65,16 +65,16 @@
class Flag {
public:
enum FlagType {
- POSITIONAL = 0,
- REQUIRED,
- OPTIONAL,
+ kPositional = 0,
+ kRequired,
+ kOptional,
};
// The order of the positional flags is the same as they are added.
// Positional flags are supposed to be required.
template <typename T>
static Flag CreateFlag(const char* name, T* val, const char* usage,
- FlagType flag_type = OPTIONAL) {
+ FlagType flag_type = kOptional) {
return Flag(
name, [val](const T& v) { *val = v; }, *val, usage, flag_type);
}
diff --git a/tensorflow/lite/tools/command_line_flags_test.cc b/tensorflow/lite/tools/command_line_flags_test.cc
index eb02379..0216d7a 100644
--- a/tensorflow/lite/tools/command_line_flags_test.cc
+++ b/tensorflow/lite/tools/command_line_flags_test.cc
@@ -55,8 +55,10 @@
Flag::CreateFlag("some_numeric_bool", &some_numeric_bool,
"some numeric bool"),
Flag::CreateFlag("some_int1", &some_int1, "some int"),
- Flag::CreateFlag("some_int2", &some_int2, "some int", Flag::REQUIRED),
- Flag::CreateFlag("float_1", &float_1, "some float", Flag::POSITIONAL),
+ Flag::CreateFlag("some_int2", &some_int2, "some int",
+ Flag::kRequired),
+ Flag::CreateFlag("float_1", &float_1, "some float",
+ Flag::kPositional),
});
EXPECT_TRUE(parsed_ok);
@@ -131,7 +133,7 @@
const char* argv_strings[] = {"program_name", "--flag=12"};
bool parsed_ok = Flags::Parse(
&argc, reinterpret_cast<const char**>(argv_strings),
- {Flag::CreateFlag("some_flag", &some_float, "", Flag::REQUIRED)});
+ {Flag::CreateFlag("some_flag", &some_float, "", Flag::kRequired)});
EXPECT_FALSE(parsed_ok);
EXPECT_NEAR(-23.23f, some_float, 1e-5f);
@@ -144,7 +146,7 @@
const char* argv_strings[] = {"program_name"};
bool parsed_ok = Flags::Parse(
&argc, reinterpret_cast<const char**>(argv_strings),
- {Flag::CreateFlag("some_flag", &some_float, "", Flag::REQUIRED)});
+ {Flag::CreateFlag("some_flag", &some_float, "", Flag::kRequired)});
EXPECT_FALSE(parsed_ok);
EXPECT_NEAR(-23.23f, some_float, 1e-5f);
@@ -157,7 +159,7 @@
const char* argv_strings[] = {"program_name"};
bool parsed_ok = Flags::Parse(
&argc, reinterpret_cast<const char**>(argv_strings),
- {Flag::CreateFlag("some_flag", &some_float, "", Flag::POSITIONAL)});
+ {Flag::CreateFlag("some_flag", &some_float, "", Flag::kPositional)});
EXPECT_FALSE(parsed_ok);
EXPECT_NEAR(-23.23f, some_float, 1e-5f);
@@ -170,7 +172,7 @@
const char* argv_strings[] = {"program_name", "string"};
bool parsed_ok = Flags::Parse(
&argc, reinterpret_cast<const char**>(argv_strings),
- {Flag::CreateFlag("some_flag", &some_float, "", Flag::POSITIONAL)});
+ {Flag::CreateFlag("some_flag", &some_float, "", Flag::kPositional)});
EXPECT_FALSE(parsed_ok);
EXPECT_NEAR(-23.23f, some_float, 1e-5f);
@@ -213,9 +215,9 @@
{Flag::CreateFlag("some_int", &some_int, "some int"),
Flag::CreateFlag("some_int64", &some_int64, "some int64"),
Flag::CreateFlag("some_switch", &some_switch, "some switch"),
- Flag::CreateFlag("some_name", &some_name, "some name", Flag::REQUIRED),
+ Flag::CreateFlag("some_name", &some_name, "some name", Flag::kRequired),
Flag::CreateFlag("some_int2", &some_int2, "some int",
- Flag::POSITIONAL)});
+ Flag::kPositional)});
// Match the usage message, being sloppy about whitespace.
const char* expected_usage =
" usage: some_tool_name <some_int2> <flags>\n"
@@ -307,8 +309,8 @@
const char* argv_strings[] = {"program_name", "--some_float=1.0"};
bool parsed_ok = Flags::Parse(
&argc, reinterpret_cast<const char**>(argv_strings),
- {Flag::CreateFlag("some_int", &some_int1, "some int1", Flag::OPTIONAL),
- Flag::CreateFlag("some_int", &some_int2, "some int2", Flag::REQUIRED)});
+ {Flag::CreateFlag("some_int", &some_int1, "some int1", Flag::kOptional),
+ Flag::CreateFlag("some_int", &some_int2, "some int2", Flag::kRequired)});
EXPECT_FALSE(parsed_ok);
EXPECT_EQ(-23, some_int1);
diff --git a/tensorflow/lite/tools/delegates/BUILD b/tensorflow/lite/tools/delegates/BUILD
new file mode 100644
index 0000000..d2eac9d
--- /dev/null
+++ b/tensorflow/lite/tools/delegates/BUILD
@@ -0,0 +1,152 @@
+load("//tensorflow/lite:build_def.bzl", "tflite_copts")
+
+package(
+ default_visibility = [
+ "//visibility:public",
+ ],
+ licenses = ["notice"], # Apache 2.0
+)
+
+common_copts = ["-Wall"] + tflite_copts()
+
+cc_library(
+ name = "delegate_provider_hdr",
+ hdrs = [
+ "delegate_provider.h",
+ ],
+ copts = common_copts,
+ deps = [
+ "//tensorflow/lite/c:common",
+ "//tensorflow/lite/tools:command_line_flags",
+ "//tensorflow/lite/tools:logging",
+ "//tensorflow/lite/tools:tool_params",
+ ],
+)
+
+# A convenient library for all inference execution providers.
+cc_library(
+ name = "tflite_execution_providers",
+ copts = tflite_copts(),
+ deps = [
+ ":coreml_delegate_provider",
+ ":default_execution_provider",
+ ":external_delegate_provider",
+ ":gpu_delegate_provider",
+ ":hexagon_delegate_provider",
+ ":nnapi_delegate_provider",
+ ":xnnpack_delegate_provider",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "default_execution_provider",
+ srcs = ["default_execution_provider.cc"],
+ copts = tflite_copts(),
+ linkstatic = True,
+ visibility = ["//visibility:public"],
+ deps = [
+ ":delegate_provider_hdr",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "gpu_delegate_provider",
+ srcs = ["gpu_delegate_provider.cc"],
+ copts = common_copts + select({
+ "//tensorflow:ios": [
+ "-xobjective-c++",
+ ],
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":delegate_provider_hdr",
+ "//tensorflow/lite/tools/evaluation:utils",
+ ] + select({
+ "//tensorflow:android": [
+ "//tensorflow/lite/delegates/gpu:delegate",
+ ],
+ "//tensorflow:ios": [
+ "//tensorflow/lite/delegates/gpu:metal_delegate",
+ ],
+ "//conditions:default": [],
+ }),
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "nnapi_delegate_provider",
+ srcs = ["nnapi_delegate_provider.cc"],
+ copts = common_copts,
+ deps = [
+ ":delegate_provider_hdr",
+ "//tensorflow/lite/tools/evaluation:utils",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "hexagon_delegate_provider",
+ srcs = ["hexagon_delegate_provider.cc"],
+ copts = common_copts,
+ deps = [
+ ":delegate_provider_hdr",
+ "//tensorflow/lite/tools/evaluation:utils",
+ ] + select({
+ "//tensorflow:android_arm": [
+ "//tensorflow/lite/experimental/delegates/hexagon:hexagon_delegate",
+ ],
+ "//tensorflow:android_arm64": [
+ "//tensorflow/lite/experimental/delegates/hexagon:hexagon_delegate",
+ ],
+ "//conditions:default": [],
+ }),
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "coreml_delegate_provider",
+ srcs = ["coreml_delegate_provider.cc"],
+ copts = common_copts + select({
+ "//tensorflow:ios": [
+ "-xobjective-c++",
+ ],
+ "//conditions:default": [],
+ }),
+ deps = [
+ ":delegate_provider_hdr",
+ "//tensorflow/lite/tools/evaluation:utils",
+ ] + select({
+ "//tensorflow:ios": [
+ "//tensorflow/lite/experimental/delegates/coreml:coreml_delegate",
+ ],
+ "//conditions:default": [],
+ }),
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "xnnpack_delegate_provider",
+ srcs = ["xnnpack_delegate_provider.cc"],
+ copts = tflite_copts(),
+ linkstatic = True,
+ visibility = ["//visibility:public"],
+ deps = [
+ ":delegate_provider_hdr",
+ "//tensorflow/lite/tools/evaluation:utils",
+ ],
+ alwayslink = 1,
+)
+
+cc_library(
+ name = "external_delegate_provider",
+ srcs = ["external_delegate_provider.cc"],
+ copts = tflite_copts(),
+ linkstatic = True,
+ visibility = ["//visibility:public"],
+ deps = [
+ ":delegate_provider_hdr",
+ ],
+ alwayslink = 1,
+)
diff --git a/tensorflow/lite/tools/delegates/README.md b/tensorflow/lite/tools/delegates/README.md
new file mode 100644
index 0000000..f0e15e9
--- /dev/null
+++ b/tensorflow/lite/tools/delegates/README.md
@@ -0,0 +1,102 @@
+# TFLite Delegate Utilities for Tooling
+
+## TFLite Delegate Registrar
+[A TFLite delegate registrar](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/tools/delegates/delegate_provider.h)
+is provided here. The registrar keeps a list of TFLite delegate providers, each
+of which defines a list parameters that could be initialized from commandline
+argumenents and provides a TFLite delegate instance creation based on those
+parameters. This delegate registrar has been used in TFLite evaluation tools and
+the benchmark model tool.
+
+A particular TFLite delegate provider can be used by
+linking the corresponding library, e.g. adding it to the `deps` of a BUILD rule.
+Note that each delegate provider library has been configured with
+`alwayslink=1` in the BUILD rule so that it will be linked to any binary that
+directly or indirectly depends on it.
+
+The following lists all implemented TFLite delegate providers and their
+corresponding list of parameters that each supports to create a particular
+TFLite delegate.
+
+### Common parameters
+* `num_threads`: `int` (default=1) \
+ The number of threads to use for running the inference on CPU.
+* `max_delegated_partitions`: `int` (default=0, i.e. no limit) \
+ The maximum number of partitions that will be delegated. \
+ Currently supported by the GPU, Hexagon, CoreML and NNAPI delegate.
+* `min_nodes_per_partition`: `int` (default=delegate's own choice) \
+ The minimal number of TFLite graph nodes of a partition that needs to be
+ reached to be delegated. A negative value or 0 means to use the default
+ choice of each delegate. \
+ This option is currently supported by the Hexagon and CoreML delegate.
+
+### GPU delegate provider
+* `use_gpu`: `bool` (default=false) \
+ Whether to use the
+ [GPU accelerator delegate](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/delegates/gpu).
+ This option is currently only available on Android and iOS devices.
+* `gpu_precision_loss_allowed`: `bool` (default=true) \
+ Whethre to allow the GPU delegate to carry out computation with some
+ precision loss (i.e. processing in FP16) or not. If allowed, the performance
+ will increase.
+* `gpu_experimental_enable_quant`: `bool` (default=true) \
+ Whether to allow the GPU delegate to run a quantized model or not. \
+ This option is currently only available on Android.
+* `gpu_backend`: `string` (default="") \
+ Force the GPU delegate to use a particular backend for execution, and fail
+ if unsuccessful. Should be one of: cl, gl. By default, the GPU delegate will
+ try OpenCL first and then OpenGL if the former fails.\
+ Note this option is only available on Android.
+* `gpu_wait_type`: `string` (default="") \
+ Which GPU wait_type option to use, when using GPU delegate on iOS. Should be
+ one of the following: passive, active, do_not_wait, aggressive. When left
+ blank, passive mode is used by default.
+
+### NNAPI delegate provider
+* `use_nnapi`: `bool` (default=false) \
+ Whether to use
+ [Android NNAPI](https://developer.android.com/ndk/guides/neuralnetworks/).
+ This API is available on recent Android devices. When on Android Q+, will
+ also print the names of NNAPI accelerators accessible through the
+ `nnapi_accelerator_name` flag.
+* `nnapi_accelerator_name`: `string` (default="") \
+ The name of the NNAPI accelerator to use (requires Android Q+). If left
+ blank, NNAPI will automatically select which of the available accelerators
+ to use.
+* `nnapi_execution_preference`: `string` (default="") \
+ Which
+ [NNAPI execution preference](https://developer.android.com/ndk/reference/group/neural-networks.html#group___neural_networks_1gga034380829226e2d980b2a7e63c992f18af727c25f1e2d8dcc693c477aef4ea5f5)
+ to use when executing using NNAPI. Should be one of the following:
+ fast_single_answer, sustained_speed, low_power, undefined.
+* `disable_nnapi_cpu`: `bool` (default=false) \
+ Excludes the
+ [NNAPI CPU reference implementation](https://developer.android.com/ndk/guides/neuralnetworks#device-assignment)
+ from the possible devices to be used by NNAPI to execute the model. This
+ option is ignored if `nnapi_accelerator_name` is specified.
+
+### Hexagon delegate provider
+* `use_hexagon`: `bool` (default=false) \
+ Whether to use the Hexagon delegate. Not all devices may support the Hexagon
+ delegate, refer to the [TensorFlow Lite documentation](https://www.tensorflow.org/lite/performance/hexagon_delegate) for more
+ information about which devices/chipsets are supported and about how to get
+ the required libraries. To use the Hexagon delegate also build the
+ hexagon_nn:libhexagon_interface.so target and copy the library to the
+ device. All libraries should be copied to /data/local/tmp on the device.
+* `hexagon_profiling`: `bool` (default=false) \
+ Whether to profile ops running on hexagon.
+
+### XNNPACK delegate provider
+* `use_xnnpack`: `bool` (default=false) \
+ Whether to use the XNNPack delegate.
+
+### CoreML delegate provider
+* `use_coreml`: `bool` (default=false) \
+ Whether to use the [Core ML delegate](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/delegates/coreml).
+ This option is only available in iOS.
+
+### External delegate provider
+* `external_delegate_path`: `string` (default="") \
+ Path to the external delegate library to use.
+* `external_delegate_options`: `string` (default="") \
+ A list of options to be passed to the external delegate library. Options
+ should be in the format of `option1:value1;option2:value2;optionN:valueN`
diff --git a/tensorflow/lite/tools/benchmark/coreml_delegate_provider.cc b/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc
similarity index 80%
rename from tensorflow/lite/tools/benchmark/coreml_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/coreml_delegate_provider.cc
index ee95f7f..0d1a8ad 100644
--- a/tensorflow/lite/tools/benchmark/coreml_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/coreml_delegate_provider.cc
@@ -14,7 +14,7 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#if defined(__APPLE__)
#if TARGET_OS_IPHONE && !TARGET_IPHONE_SIMULATOR
@@ -25,28 +25,27 @@
#endif
namespace tflite {
-namespace benchmark {
+namespace tools {
class CoreMlDelegateProvider : public DelegateProvider {
public:
CoreMlDelegateProvider() {
#if defined(REAL_IPHONE_DEVICE)
- default_params_.AddParam("use_coreml", BenchmarkParam::Create<bool>(true));
+ default_params_.AddParam("use_coreml", ToolParam::Create<bool>(true));
#endif
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "COREML"; }
};
REGISTER_DELEGATE_PROVIDER(CoreMlDelegateProvider);
std::vector<Flag> CoreMlDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+ ToolParams* params) const {
#if defined(REAL_IPHONE_DEVICE)
std::vector<Flag> flags = {
CreateFlag<bool>("use_coreml", params, "use Core ML"),
@@ -57,7 +56,7 @@
#endif
}
-void CoreMlDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void CoreMlDelegateProvider::LogParams(const ToolParams& params) const {
#if defined(REAL_IPHONE_DEVICE)
TFLITE_LOG(INFO) << "Use Core ML : [" << params.Get<bool>("use_coreml")
<< "]";
@@ -65,7 +64,7 @@
}
TfLiteDelegatePtr CoreMlDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
TfLiteDelegatePtr delegate(nullptr, [](TfLiteDelegate*) {});
#if defined(REAL_IPHONE_DEVICE)
@@ -88,5 +87,5 @@
return delegate;
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/default_execution_provider.cc b/tensorflow/lite/tools/delegates/default_execution_provider.cc
similarity index 77%
rename from tensorflow/lite/tools/benchmark/default_execution_provider.cc
rename to tensorflow/lite/tools/delegates/default_execution_provider.cc
index 1dd8aeb..f75fd79 100644
--- a/tensorflow/lite/tools/benchmark/default_execution_provider.cc
+++ b/tensorflow/lite/tools/delegates/default_execution_provider.cc
@@ -14,10 +14,10 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
namespace tflite {
-namespace benchmark {
+namespace tools {
// This class actually doesn't provide any TFLite delegate instances, it simply
// provides common params and flags that are common to all actual delegate
@@ -25,23 +25,22 @@
class DefaultExecutionProvider : public DelegateProvider {
public:
DefaultExecutionProvider() {
- default_params_.AddParam("num_threads", BenchmarkParam::Create<int32_t>(1));
+ default_params_.AddParam("num_threads", ToolParam::Create<int32_t>(1));
default_params_.AddParam("max_delegated_partitions",
- BenchmarkParam::Create<int32_t>(0));
+ ToolParam::Create<int32_t>(0));
default_params_.AddParam("min_nodes_per_partition",
- BenchmarkParam::Create<int32_t>(0));
+ ToolParam::Create<int32_t>(0));
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
+ void LogParams(const ToolParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "Default-NoDelegate"; }
};
REGISTER_DELEGATE_PROVIDER(DefaultExecutionProvider);
std::vector<Flag> DefaultExecutionProvider::CreateFlags(
- BenchmarkParams* params) const {
+ ToolParams* params) const {
std::vector<Flag> flags = {
CreateFlag<int32_t>("num_threads", params,
"number of threads used for inference on CPU."),
@@ -55,7 +54,7 @@
return flags;
}
-void DefaultExecutionProvider::LogParams(const BenchmarkParams& params) const {
+void DefaultExecutionProvider::LogParams(const ToolParams& params) const {
TFLITE_LOG(INFO) << "#threads used for CPU inference: ["
<< params.Get<int32_t>("num_threads") << "]";
TFLITE_LOG(INFO) << "Max number of delegated partitions : ["
@@ -65,9 +64,9 @@
}
TfLiteDelegatePtr DefaultExecutionProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
return TfLiteDelegatePtr(nullptr, [](TfLiteDelegate*) {});
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/delegate_provider.h b/tensorflow/lite/tools/delegates/delegate_provider.h
similarity index 73%
rename from tensorflow/lite/tools/benchmark/delegate_provider.h
rename to tensorflow/lite/tools/delegates/delegate_provider.h
index 6090b7f..91dd3b1 100644
--- a/tensorflow/lite/tools/benchmark/delegate_provider.h
+++ b/tensorflow/lite/tools/delegates/delegate_provider.h
@@ -13,19 +13,19 @@
limitations under the License.
==============================================================================*/
-#ifndef TENSORFLOW_LITE_TOOLS_BENCHMARK_DELEGATE_PROVIDER_H_
-#define TENSORFLOW_LITE_TOOLS_BENCHMARK_DELEGATE_PROVIDER_H_
+#ifndef TENSORFLOW_LITE_TOOLS_DELEGATES_DELEGATE_PROVIDER_H_
+#define TENSORFLOW_LITE_TOOLS_DELEGATES_DELEGATE_PROVIDER_H_
#include <string>
#include <vector>
#include "tensorflow/lite/c/common.h"
-#include "tensorflow/lite/tools/benchmark/benchmark_params.h"
#include "tensorflow/lite/tools/command_line_flags.h"
#include "tensorflow/lite/tools/logging.h"
+#include "tensorflow/lite/tools/tool_params.h"
namespace tflite {
-namespace benchmark {
+namespace tools {
// Same w/ Interpreter::TfLiteDelegatePtr to avoid pulling
// tensorflow/lite/interpreter.h dependency
@@ -36,31 +36,30 @@
public:
virtual ~DelegateProvider() {}
- // Create a list of command-line parsable flags based on benchmark params
- // inside 'params' whose value will be set to the corresponding runtime flag
- // value.
- virtual std::vector<Flag> CreateFlags(BenchmarkParams* params) const = 0;
+ // Create a list of command-line parsable flags based on tool params inside
+ // 'params' whose value will be set to the corresponding runtime flag value.
+ virtual std::vector<Flag> CreateFlags(ToolParams* params) const = 0;
- // Log benchmark params.
- virtual void LogParams(const BenchmarkParams& params) const = 0;
+ // Log tool params.
+ virtual void LogParams(const ToolParams& params) const = 0;
- // Create a TfLiteDelegate based on benchmark params.
+ // Create a TfLiteDelegate based on tool params.
virtual TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const = 0;
+ const ToolParams& params) const = 0;
virtual std::string GetName() const = 0;
- const BenchmarkParams& DefaultParams() const { return default_params_; }
+ const ToolParams& DefaultParams() const { return default_params_; }
protected:
template <typename T>
- Flag CreateFlag(const char* name, BenchmarkParams* params,
+ Flag CreateFlag(const char* name, ToolParams* params,
const std::string& usage) const {
return Flag(
name, [params, name](const T& val) { params->Set<T>(name, val); },
- default_params_.Get<T>(name), usage, Flag::OPTIONAL);
+ default_params_.Get<T>(name), usage, Flag::kOptional);
}
- BenchmarkParams default_params_;
+ ToolParams default_params_;
};
using DelegateProviderPtr = std::unique_ptr<DelegateProvider>;
@@ -102,7 +101,7 @@
inline const DelegateProviderList& GetRegisteredDelegateProviders() {
return DelegateProviderRegistrar::GetProviders();
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
-#endif // TENSORFLOW_LITE_TOOLS_BENCHMARK_DELEGATE_PROVIDER_H_
+#endif // TENSORFLOW_LITE_TOOLS_DELEGATES_DELEGATE_PROVIDER_H_
diff --git a/tensorflow/lite/tools/benchmark/external_delegate_provider.cc b/tensorflow/lite/tools/delegates/external_delegate_provider.cc
similarity index 89%
rename from tensorflow/lite/tools/benchmark/external_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/external_delegate_provider.cc
index a5d8a94..95b0e42 100644
--- a/tensorflow/lite/tools/benchmark/external_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/external_delegate_provider.cc
@@ -12,7 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#if defined(_WIN32)
#include <Windows.h>
@@ -25,7 +25,7 @@
#include <vector>
namespace tflite {
-namespace benchmark {
+namespace tools {
namespace {
// Library Support construct to handle dynamic library operations
#if defined(_WIN32)
@@ -97,24 +97,23 @@
public:
ExternalDelegateProvider() {
default_params_.AddParam("external_delegate_path",
- BenchmarkParam::Create<std::string>(""));
+ ToolParam::Create<std::string>(""));
default_params_.AddParam("external_delegate_options",
- BenchmarkParam::Create<std::string>(""));
+ ToolParam::Create<std::string>(""));
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "EXTERNAL"; }
};
REGISTER_DELEGATE_PROVIDER(ExternalDelegateProvider);
std::vector<Flag> ExternalDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+ ToolParams* params) const {
std::vector<Flag> flags = {
CreateFlag<std::string>("external_delegate_path", params,
"The library path for the underlying external."),
@@ -124,7 +123,7 @@
return flags;
}
-void ExternalDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void ExternalDelegateProvider::LogParams(const ToolParams& params) const {
TFLITE_LOG(INFO) << "External delegate path : ["
<< params.Get<std::string>("external_delegate_path") << "]";
TFLITE_LOG(INFO) << "External delegate options : ["
@@ -133,7 +132,7 @@
}
TfLiteDelegatePtr ExternalDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
TfLiteDelegatePtr delegate(nullptr, [](TfLiteDelegate*) {});
std::string lib_path = params.Get<std::string>("external_delegate_path");
if (!lib_path.empty()) {
@@ -167,5 +166,5 @@
}
return delegate;
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/gpu_delegate_provider.cc b/tensorflow/lite/tools/delegates/gpu_delegate_provider.cc
similarity index 75%
rename from tensorflow/lite/tools/benchmark/gpu_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/gpu_delegate_provider.cc
index 96e86d1..db1f32b 100644
--- a/tensorflow/lite/tools/benchmark/gpu_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/gpu_delegate_provider.cc
@@ -14,7 +14,7 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#if defined(__ANDROID__)
#include "tensorflow/lite/delegates/gpu/delegate.h"
@@ -28,39 +28,38 @@
#endif
namespace tflite {
-namespace benchmark {
+namespace tools {
class GpuDelegateProvider : public DelegateProvider {
public:
GpuDelegateProvider() {
- default_params_.AddParam("use_gpu", BenchmarkParam::Create<bool>(false));
+ default_params_.AddParam("use_gpu", ToolParam::Create<bool>(false));
#if defined(__ANDROID__) || defined(REAL_IPHONE_DEVICE)
default_params_.AddParam("gpu_precision_loss_allowed",
- BenchmarkParam::Create<bool>(true));
+ ToolParam::Create<bool>(true));
#endif
#if defined(__ANDROID__)
default_params_.AddParam("gpu_experimental_enable_quant",
- BenchmarkParam::Create<bool>(true));
+ ToolParam::Create<bool>(true));
+ default_params_.AddParam("gpu_backend", ToolParam::Create<std::string>(""));
#endif
#if defined(REAL_IPHONE_DEVICE)
default_params_.AddParam("gpu_wait_type",
- BenchmarkParam::Create<std::string>(""));
+ ToolParam::Create<std::string>(""));
#endif
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "GPU"; }
};
REGISTER_DELEGATE_PROVIDER(GpuDelegateProvider);
-std::vector<Flag> GpuDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+std::vector<Flag> GpuDelegateProvider::CreateFlags(ToolParams* params) const {
std::vector<Flag> flags = {
CreateFlag<bool>("use_gpu", params, "use gpu"),
#if defined(__ANDROID__) || defined(REAL_IPHONE_DEVICE)
@@ -72,6 +71,10 @@
CreateFlag<bool>("gpu_experimental_enable_quant", params,
"Whether to enable the GPU delegate to run quantized "
"models or not. By default, it's disabled."),
+ CreateFlag<std::string>(
+ "gpu_backend", params,
+ "Force the GPU delegate to use a particular backend for execution, and "
+ "fail if unsuccessful. Should be one of: cl, gl"),
#endif
#if defined(REAL_IPHONE_DEVICE)
CreateFlag<std::string>(
@@ -83,7 +86,7 @@
return flags;
}
-void GpuDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void GpuDelegateProvider::LogParams(const ToolParams& params) const {
TFLITE_LOG(INFO) << "Use gpu : [" << params.Get<bool>("use_gpu") << "]";
#if defined(__ANDROID__) || defined(REAL_IPHONE_DEVICE)
TFLITE_LOG(INFO) << "Allow lower precision in gpu : ["
@@ -92,6 +95,8 @@
#if defined(__ANDROID__)
TFLITE_LOG(INFO) << "Enable running quant models in gpu : ["
<< params.Get<bool>("gpu_experimental_enable_quant") << "]";
+ TFLITE_LOG(INFO) << "GPU backend : ["
+ << params.Get<std::string>("gpu_backend") << "]";
#endif
#if defined(REAL_IPHONE_DEVICE)
TFLITE_LOG(INFO) << "GPU delegate wait type : ["
@@ -100,7 +105,7 @@
}
TfLiteDelegatePtr GpuDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
TfLiteDelegatePtr delegate(nullptr, [](TfLiteDelegate*) {});
if (params.Get<bool>("use_gpu")) {
@@ -116,6 +121,16 @@
if (params.Get<bool>("gpu_experimental_enable_quant")) {
gpu_opts.experimental_flags |= TFLITE_GPU_EXPERIMENTAL_FLAGS_ENABLE_QUANT;
}
+ std::string gpu_backend = params.Get<std::string>("gpu_backend");
+ if (!gpu_backend.empty()) {
+ if (gpu_backend == "cl") {
+ gpu_opts.experimental_flags |= TFLITE_GPU_EXPERIMENTAL_FLAGS_CL_ONLY;
+ } else if (gpu_backend == "gl") {
+ gpu_opts.experimental_flags |= TFLITE_GPU_EXPERIMENTAL_FLAGS_GL_ONLY;
+ }
+ }
+ gpu_opts.max_delegated_partitions =
+ params.Get<int>("max_delegated_partitions");
delegate = evaluation::CreateGPUDelegate(&gpu_opts);
#elif defined(REAL_IPHONE_DEVICE)
TFLGpuDelegateOptions gpu_opts = {0};
@@ -139,8 +154,8 @@
delegate = TfLiteDelegatePtr(TFLGpuDelegateCreate(&gpu_opts),
&TFLGpuDelegateDelete);
#else
- TFLITE_LOG(WARN) << "The GPU delegate compile options are only supported "
- "to be benchmarked on Android or iOS platforms.";
+ TFLITE_LOG(WARN) << "The GPU delegate compile options are only supported on"
+ "Android or iOS platforms.";
delegate = evaluation::CreateGPUDelegate();
#endif
@@ -151,5 +166,5 @@
return delegate;
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/hexagon_delegate_provider.cc b/tensorflow/lite/tools/delegates/hexagon_delegate_provider.cc
similarity index 79%
rename from tensorflow/lite/tools/benchmark/hexagon_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/hexagon_delegate_provider.cc
index 1904dd5..0afb0b2 100644
--- a/tensorflow/lite/tools/benchmark/hexagon_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/hexagon_delegate_provider.cc
@@ -14,7 +14,7 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#if (defined(ANDROID) || defined(__ANDROID__)) && \
@@ -27,35 +27,32 @@
#endif
namespace tflite {
-namespace benchmark {
+namespace tools {
class HexagonDelegateProvider : public DelegateProvider {
public:
HexagonDelegateProvider() {
#if defined(TFLITE_ENABLE_HEXAGON)
- default_params_.AddParam("use_hexagon",
- BenchmarkParam::Create<bool>(false));
- default_params_.AddParam(
- "hexagon_lib_path",
- BenchmarkParam::Create<std::string>("/data/local/tmp"));
+ default_params_.AddParam("use_hexagon", ToolParam::Create<bool>(false));
+ default_params_.AddParam("hexagon_lib_path",
+ ToolParam::Create<std::string>("/data/local/tmp"));
default_params_.AddParam("hexagon_profiling",
- BenchmarkParam::Create<bool>(false));
+ ToolParam::Create<bool>(false));
#endif
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "Hexagon"; }
};
REGISTER_DELEGATE_PROVIDER(HexagonDelegateProvider);
std::vector<Flag> HexagonDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+ ToolParams* params) const {
#if defined(TFLITE_ENABLE_HEXAGON)
std::vector<Flag> flags = {
CreateFlag<bool>("use_hexagon", params, "Use Hexagon delegate"),
@@ -70,7 +67,7 @@
#endif
}
-void HexagonDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void HexagonDelegateProvider::LogParams(const ToolParams& params) const {
#if defined(TFLITE_ENABLE_HEXAGON)
TFLITE_LOG(INFO) << "Use Hexagon : [" << params.Get<bool>("use_hexagon")
<< "]";
@@ -82,7 +79,7 @@
}
TfLiteDelegatePtr HexagonDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
TfLiteDelegatePtr delegate(nullptr, [](TfLiteDelegate*) {});
#if defined(TFLITE_ENABLE_HEXAGON)
if (params.Get<bool>("use_hexagon")) {
@@ -105,5 +102,5 @@
return delegate;
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/nnapi_delegate_provider.cc b/tensorflow/lite/tools/delegates/nnapi_delegate_provider.cc
similarity index 87%
rename from tensorflow/lite/tools/benchmark/nnapi_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/nnapi_delegate_provider.cc
index 04aa318..f3ed874 100644
--- a/tensorflow/lite/tools/benchmark/nnapi_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/nnapi_delegate_provider.cc
@@ -14,40 +14,38 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#if defined(__ANDROID__)
#include "tensorflow/lite/nnapi/nnapi_util.h"
#endif
namespace tflite {
-namespace benchmark {
+namespace tools {
class NnapiDelegateProvider : public DelegateProvider {
public:
NnapiDelegateProvider() {
- default_params_.AddParam("use_nnapi", BenchmarkParam::Create<bool>(false));
+ default_params_.AddParam("use_nnapi", ToolParam::Create<bool>(false));
default_params_.AddParam("nnapi_execution_preference",
- BenchmarkParam::Create<std::string>(""));
+ ToolParam::Create<std::string>(""));
default_params_.AddParam("nnapi_accelerator_name",
- BenchmarkParam::Create<std::string>(""));
+ ToolParam::Create<std::string>(""));
default_params_.AddParam("disable_nnapi_cpu",
- BenchmarkParam::Create<bool>(false));
+ ToolParam::Create<bool>(false));
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "NNAPI"; }
};
REGISTER_DELEGATE_PROVIDER(NnapiDelegateProvider);
-std::vector<Flag> NnapiDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+std::vector<Flag> NnapiDelegateProvider::CreateFlags(ToolParams* params) const {
std::vector<Flag> flags = {
CreateFlag<bool>("use_nnapi", params, "use nnapi delegate api"),
CreateFlag<std::string>("nnapi_execution_preference", params,
@@ -63,7 +61,7 @@
return flags;
}
-void NnapiDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void NnapiDelegateProvider::LogParams(const ToolParams& params) const {
#if defined(__ANDROID__)
TFLITE_LOG(INFO) << "Use nnapi : [" << params.Get<bool>("use_nnapi") << "]";
if (params.Get<bool>("use_nnapi")) {
@@ -90,7 +88,7 @@
}
TfLiteDelegatePtr NnapiDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
TfLiteDelegatePtr delegate(nullptr, [](TfLiteDelegate*) {});
if (params.Get<bool>("use_nnapi")) {
StatefulNnApiDelegate::Options options;
@@ -150,5 +148,5 @@
return delegate;
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/benchmark/xnnpack_delegate_provider.cc b/tensorflow/lite/tools/delegates/xnnpack_delegate_provider.cc
similarity index 71%
rename from tensorflow/lite/tools/benchmark/xnnpack_delegate_provider.cc
rename to tensorflow/lite/tools/delegates/xnnpack_delegate_provider.cc
index 7222639..e9bdfb4 100644
--- a/tensorflow/lite/tools/benchmark/xnnpack_delegate_provider.cc
+++ b/tensorflow/lite/tools/delegates/xnnpack_delegate_provider.cc
@@ -14,44 +14,42 @@
==============================================================================*/
#include <string>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
namespace tflite {
-namespace benchmark {
+namespace tools {
class XnnpackDelegateProvider : public DelegateProvider {
public:
XnnpackDelegateProvider() {
- default_params_.AddParam("use_xnnpack",
- BenchmarkParam::Create<bool>(false));
+ default_params_.AddParam("use_xnnpack", ToolParam::Create<bool>(false));
}
- std::vector<Flag> CreateFlags(BenchmarkParams* params) const final;
+ std::vector<Flag> CreateFlags(ToolParams* params) const final;
- void LogParams(const BenchmarkParams& params) const final;
+ void LogParams(const ToolParams& params) const final;
- TfLiteDelegatePtr CreateTfLiteDelegate(
- const BenchmarkParams& params) const final;
+ TfLiteDelegatePtr CreateTfLiteDelegate(const ToolParams& params) const final;
std::string GetName() const final { return "XNNPACK"; }
};
REGISTER_DELEGATE_PROVIDER(XnnpackDelegateProvider);
std::vector<Flag> XnnpackDelegateProvider::CreateFlags(
- BenchmarkParams* params) const {
+ ToolParams* params) const {
std::vector<Flag> flags = {
CreateFlag<bool>("use_xnnpack", params, "use XNNPack")};
return flags;
}
-void XnnpackDelegateProvider::LogParams(const BenchmarkParams& params) const {
+void XnnpackDelegateProvider::LogParams(const ToolParams& params) const {
TFLITE_LOG(INFO) << "Use xnnpack : [" << params.Get<bool>("use_xnnpack")
<< "]";
}
TfLiteDelegatePtr XnnpackDelegateProvider::CreateTfLiteDelegate(
- const BenchmarkParams& params) const {
+ const ToolParams& params) const {
if (params.Get<bool>("use_xnnpack")) {
return evaluation::CreateXNNPACKDelegate(
params.Get<int32_t>("num_threads"));
@@ -59,5 +57,5 @@
return TfLiteDelegatePtr(nullptr, [](TfLiteDelegate*) {});
}
-} // namespace benchmark
+} // namespace tools
} // namespace tflite
diff --git a/tensorflow/lite/tools/evaluation/BUILD b/tensorflow/lite/tools/evaluation/BUILD
index 0955b6d..1bc3521 100644
--- a/tensorflow/lite/tools/evaluation/BUILD
+++ b/tensorflow/lite/tools/evaluation/BUILD
@@ -57,6 +57,7 @@
"//conditions:default": [],
}) + select({
"//tensorflow:fuchsia": [],
+ "//tensorflow:windows": [],
"//conditions:default": [
"//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
],
@@ -73,8 +74,8 @@
"//tensorflow/lite/tools:command_line_flags",
"//tensorflow/lite/tools:logging",
"//tensorflow/lite/tools:tool_params",
- "//tensorflow/lite/tools/benchmark:delegate_provider_hdr",
- "//tensorflow/lite/tools/benchmark:tflite_execution_providers",
+ "//tensorflow/lite/tools/delegates:delegate_provider_hdr",
+ "//tensorflow/lite/tools/delegates:tflite_execution_providers",
"//tensorflow/lite/tools/evaluation/proto:evaluation_stages_cc_proto",
],
)
diff --git a/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.cc b/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.cc
index 0a3c3a3..42f2666 100644
--- a/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.cc
+++ b/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.cc
@@ -66,7 +66,6 @@
return p;
}
case TfliteInferenceParams::NONE:
- if (error_msg) *error_msg = "No delegate type is specified.";
return TfLiteDelegatePtr(nullptr, [](TfLiteDelegate*) {});
default:
if (error_msg) {
@@ -79,7 +78,7 @@
}
DelegateProviders::DelegateProviders()
- : delegates_list_(benchmark::GetRegisteredDelegateProviders()),
+ : delegates_list_(tools::GetRegisteredDelegateProviders()),
delegates_map_([=]() -> std::unordered_map<std::string, int> {
std::unordered_map<std::string, int> delegates_map;
for (int i = 0; i < delegates_list_.size(); ++i) {
diff --git a/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.h b/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.h
index 9afba41..36f8046 100644
--- a/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.h
+++ b/tensorflow/lite/tools/evaluation/evaluation_delegate_provider.h
@@ -20,7 +20,7 @@
#include <unordered_map>
#include <vector>
-#include "tensorflow/lite/tools/benchmark/delegate_provider.h"
+#include "tensorflow/lite/tools/delegates/delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/proto/evaluation_stages.pb.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#include "tensorflow/lite/tools/tool_params.h"
@@ -73,7 +73,7 @@
// flags.
tools::ToolParams params_;
- const benchmark::DelegateProviderList& delegates_list_;
+ const tools::DelegateProviderList& delegates_list_;
// Key is the delegate name, and the value is the index to the
// 'delegates_list_'.
const std::unordered_map<std::string, int> delegates_map_;
diff --git a/tensorflow/lite/tools/evaluation/proto/BUILD b/tensorflow/lite/tools/evaluation/proto/BUILD
index a506e74..7ced075 100644
--- a/tensorflow/lite/tools/evaluation/proto/BUILD
+++ b/tensorflow/lite/tools/evaluation/proto/BUILD
@@ -47,6 +47,15 @@
srcs = [
"evaluation_stages.proto",
],
+ protodeps = [":preprocessing_steps"],
+ visibility = ["//visibility:public"],
+)
+
+tf_proto_library_py(
+ name = "preprocessing_steps",
+ srcs = [
+ "preprocessing_steps.proto",
+ ],
visibility = ["//visibility:public"],
)
diff --git a/tensorflow/lite/tools/evaluation/stages/object_detection_stage.cc b/tensorflow/lite/tools/evaluation/stages/object_detection_stage.cc
index 27aba70..8a7f920e 100644
--- a/tensorflow/lite/tools/evaluation/stages/object_detection_stage.cc
+++ b/tensorflow/lite/tools/evaluation/stages/object_detection_stage.cc
@@ -157,7 +157,7 @@
}
TfLiteStatus PopulateGroundTruth(
- const std::string& grouth_truth_pbtxt_file,
+ const std::string& grouth_truth_proto_file,
absl::flat_hash_map<std::string, ObjectDetectionResult>*
ground_truth_mapping) {
if (ground_truth_mapping == nullptr) {
@@ -166,7 +166,7 @@
ground_truth_mapping->clear();
// Read the ground truth dump.
- std::ifstream t(grouth_truth_pbtxt_file);
+ std::ifstream t(grouth_truth_proto_file);
std::string proto_str((std::istreambuf_iterator<char>(t)),
std::istreambuf_iterator<char>());
ObjectDetectionGroundTruth ground_truth_proto;
diff --git a/tensorflow/lite/tools/evaluation/stages/object_detection_stage.h b/tensorflow/lite/tools/evaluation/stages/object_detection_stage.h
index 1489d85..0623d21 100644
--- a/tensorflow/lite/tools/evaluation/stages/object_detection_stage.h
+++ b/tensorflow/lite/tools/evaluation/stages/object_detection_stage.h
@@ -97,7 +97,7 @@
// preprocess_coco_minival.py script in evaluation/tasks/coco_object_detection.
// Useful for wrappers/scripts that use ObjectDetectionStage.
TfLiteStatus PopulateGroundTruth(
- const std::string& grouth_truth_pbtxt_file,
+ const std::string& grouth_truth_proto_file,
absl::flat_hash_map<std::string, ObjectDetectionResult>*
ground_truth_mapping);
diff --git a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/BUILD b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/BUILD
index 519fbc8..6c3a91c 100644
--- a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/BUILD
+++ b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/BUILD
@@ -1,4 +1,5 @@
-load("//tensorflow/lite:build_def.bzl", "tflite_copts", "tflite_linkopts")
+load("//tensorflow/lite:build_def.bzl", "tflite_copts")
+load("//tensorflow/lite/tools/evaluation/tasks:build_def.bzl", "task_linkopts")
package(
default_visibility = [
@@ -16,18 +17,11 @@
deps = ["//tensorflow/lite/tools/evaluation/proto:evaluation_stages_py"],
)
-cc_binary(
- name = "run_eval",
+cc_library(
+ name = "run_eval_lib",
srcs = ["run_eval.cc"],
copts = tflite_copts(),
- linkopts = tflite_linkopts() + select({
- "//tensorflow:android": [
- "-pie", # Android 5.0 and later supports only PIE
- "-lm", # some builtin ops, e.g., tanh, need -lm
- "-Wl,--rpath=/data/local/tmp/", # Hexagon delegate libraries should be in /data/local/tmp
- ],
- "//conditions:default": [],
- }),
+ linkopts = task_linkopts(),
deps = [
"//tensorflow/lite/c:common",
"//tensorflow/lite/tools:command_line_flags",
@@ -38,6 +32,18 @@
"//tensorflow/lite/tools/evaluation/proto:evaluation_config_cc_proto",
"//tensorflow/lite/tools/evaluation/proto:evaluation_stages_cc_proto",
"//tensorflow/lite/tools/evaluation/stages:object_detection_stage",
+ "//tensorflow/lite/tools/evaluation/tasks:task_executor",
"@com_google_absl//absl/container:flat_hash_map",
+ "@com_google_absl//absl/types:optional",
+ ],
+)
+
+cc_binary(
+ name = "run_eval",
+ copts = tflite_copts(),
+ linkopts = task_linkopts(),
+ deps = [
+ ":run_eval_lib",
+ "//tensorflow/lite/tools/evaluation/tasks:task_executor_main",
],
)
diff --git a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/README.md b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/README.md
index 5b4617d..a5baff1 100644
--- a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/README.md
+++ b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/README.md
@@ -83,9 +83,10 @@
assumes that `libhexagon_interface.so` and Qualcomm libraries lie in
`/data/local/tmp`.
-This script also supports all applicable runtime/delegate arguments supported on
-the `benchmark_model` tool. If there is any conflict (for example, `num_threads`
-in `benchmark_model` vs `num_interpreter_threads` here), the parameters of this
+This script also supports runtime/delegate arguments introduced by the
+[delegate registrar](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/delegates).
+If there is any conflict (for example, `num_threads` vs
+`num_interpreter_threads` here), the parameters of this
script are given precedence.
### Debug Mode
diff --git a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/run_eval.cc b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/run_eval.cc
index d1384dc..765e8fc 100644
--- a/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/run_eval.cc
+++ b/tensorflow/lite/tools/evaluation/tasks/coco_object_detection/run_eval.cc
@@ -18,12 +18,14 @@
#include <vector>
#include "absl/container/flat_hash_map.h"
+#include "absl/types/optional.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/tools/command_line_flags.h"
#include "tensorflow/lite/tools/evaluation/evaluation_delegate_provider.h"
#include "tensorflow/lite/tools/evaluation/proto/evaluation_config.pb.h"
#include "tensorflow/lite/tools/evaluation/proto/evaluation_stages.pb.h"
#include "tensorflow/lite/tools/evaluation/stages/object_detection_stage.h"
+#include "tensorflow/lite/tools/evaluation/tasks/task_executor.h"
#include "tensorflow/lite/tools/evaluation/utils.h"
#include "tensorflow/lite/tools/logging.h"
@@ -45,37 +47,102 @@
return str.substr(pos + 1);
}
-bool EvaluateModel(const std::string& model_file_path,
- const std::vector<std::string>& model_labels,
- const std::vector<std::string>& image_paths,
- const std::string& ground_truth_proto_file,
- std::string delegate, std::string output_file_path,
- int num_interpreter_threads, bool debug_mode,
- const DelegateProviders& delegate_providers) {
+class CocoObjectDetection : public TaskExecutor {
+ public:
+ CocoObjectDetection(int* argc, char* argv[]);
+ ~CocoObjectDetection() override {}
+
+ // If the run is successful, the latest metrics will be returned.
+ absl::optional<EvaluationStageMetrics> Run() final;
+
+ private:
+ void OutputResult(const EvaluationStageMetrics& latest_metrics) const;
+ std::string model_file_path_;
+ std::string model_output_labels_path_;
+ std::string ground_truth_images_path_;
+ std::string ground_truth_proto_file_;
+ std::string output_file_path_;
+ bool debug_mode_;
+ std::string delegate_;
+ int num_interpreter_threads_;
+ DelegateProviders delegate_providers_;
+};
+
+CocoObjectDetection::CocoObjectDetection(int* argc, char* argv[])
+ : debug_mode_(false), num_interpreter_threads_(1) {
+ std::vector<tflite::Flag> flag_list = {
+ tflite::Flag::CreateFlag(kModelFileFlag, &model_file_path_,
+ "Path to test tflite model file."),
+ tflite::Flag::CreateFlag(
+ kModelOutputLabelsFlag, &model_output_labels_path_,
+ "Path to labels that correspond to output of model."
+ " E.g. in case of COCO-trained SSD model, this is the path to file "
+ "where each line contains a class detected by the model in correct "
+ "order, starting from background."),
+ tflite::Flag::CreateFlag(
+ kGroundTruthImagesPathFlag, &ground_truth_images_path_,
+ "Path to ground truth images. These will be evaluated in "
+ "alphabetical order of filenames"),
+ tflite::Flag::CreateFlag(kGroundTruthProtoFileFlag,
+ &ground_truth_proto_file_,
+ "Path to file containing "
+ "tflite::evaluation::ObjectDetectionGroundTruth "
+ "proto in binary serialized format. If left "
+ "empty, mAP numbers are not output."),
+ tflite::Flag::CreateFlag(
+ kOutputFilePathFlag, &output_file_path_,
+ "File to output to. Contains only metrics proto if debug_mode is "
+ "off, and per-image predictions also otherwise."),
+ tflite::Flag::CreateFlag(kDebugModeFlag, &debug_mode_,
+ "Whether to enable debug mode. Per-image "
+ "predictions are written to the output file "
+ "along with metrics."),
+ tflite::Flag::CreateFlag(
+ kInterpreterThreadsFlag, &num_interpreter_threads_,
+ "Number of interpreter threads to use for inference."),
+ tflite::Flag::CreateFlag(
+ kDelegateFlag, &delegate_,
+ "Delegate to use for inference, if available. "
+ "Must be one of {'nnapi', 'gpu', 'xnnpack', 'hexagon'}"),
+ };
+ tflite::Flags::Parse(argc, const_cast<const char**>(argv), flag_list);
+ DelegateProviders delegate_providers;
+ delegate_providers.InitFromCmdlineArgs(argc, const_cast<const char**>(argv));
+}
+
+absl::optional<EvaluationStageMetrics> CocoObjectDetection::Run() {
+ // Process images in filename-sorted order.
+ std::vector<std::string> image_paths;
+ if (GetSortedFileNames(StripTrailingSlashes(ground_truth_images_path_),
+ &image_paths) != kTfLiteOk) {
+ return absl::nullopt;
+ }
+
+ std::vector<std::string> model_labels;
+ if (!ReadFileLines(model_output_labels_path_, &model_labels)) {
+ TFLITE_LOG(ERROR) << "Could not read model output labels file";
+ return absl::nullopt;
+ }
+
EvaluationStageConfig eval_config;
eval_config.set_name("object_detection");
auto* detection_params =
eval_config.mutable_specification()->mutable_object_detection_params();
auto* inference_params = detection_params->mutable_inference_params();
- inference_params->set_model_file_path(model_file_path);
- inference_params->set_num_threads(num_interpreter_threads);
- inference_params->set_delegate(ParseStringToDelegateType(delegate));
- if (!delegate.empty() &&
- inference_params->delegate() == TfliteInferenceParams::NONE) {
- TFLITE_LOG(WARN) << "Unsupported TFLite delegate: " << delegate;
- return false;
- }
+ inference_params->set_model_file_path(model_file_path_);
+ inference_params->set_num_threads(num_interpreter_threads_);
+ inference_params->set_delegate(ParseStringToDelegateType(delegate_));
// Get ground truth data.
absl::flat_hash_map<std::string, ObjectDetectionResult> ground_truth_map;
- if (!ground_truth_proto_file.empty()) {
- PopulateGroundTruth(ground_truth_proto_file, &ground_truth_map);
+ if (!ground_truth_proto_file_.empty()) {
+ PopulateGroundTruth(ground_truth_proto_file_, &ground_truth_map);
}
ObjectDetectionStage eval(eval_config);
eval.SetAllLabels(model_labels);
- if (eval.Init(&delegate_providers) != kTfLiteOk) return false;
+ if (eval.Init(&delegate_providers_) != kTfLiteOk) return absl::nullopt;
const int step = image_paths.size() / 100;
for (int i = 0; i < image_paths.size(); ++i) {
@@ -85,9 +152,9 @@
const std::string image_name = GetNameFromPath(image_paths[i]);
eval.SetInputs(image_paths[i], ground_truth_map[image_name]);
- if (eval.Run() != kTfLiteOk) return false;
+ if (eval.Run() != kTfLiteOk) return absl::nullopt;
- if (debug_mode) {
+ if (debug_mode_) {
ObjectDetectionResult prediction = *eval.GetLatestPrediction();
TFLITE_LOG(INFO) << "Image: " << image_name << "\n";
for (int i = 0; i < prediction.objects_size(); ++i) {
@@ -113,15 +180,22 @@
// Write metrics to file.
EvaluationStageMetrics latest_metrics = eval.LatestMetrics();
- if (ground_truth_proto_file.empty()) {
- // mAP metrics are meaningless for no ground truth.
+ if (ground_truth_proto_file_.empty()) {
+ TFLITE_LOG(WARN) << "mAP metrics are meaningless w/o ground truth.";
latest_metrics.mutable_process_metrics()
->mutable_object_detection_metrics()
->clear_average_precision_metrics();
}
- if (!output_file_path.empty()) {
+
+ OutputResult(latest_metrics);
+ return absl::make_optional(latest_metrics);
+}
+
+void CocoObjectDetection::OutputResult(
+ const EvaluationStageMetrics& latest_metrics) const {
+ if (!output_file_path_.empty()) {
std::ofstream metrics_ofile;
- metrics_ofile.open(output_file_path, std::ios::out);
+ metrics_ofile.open(output_file_path_, std::ios::out);
metrics_ofile << latest_metrics.SerializeAsString();
metrics_ofile.close();
}
@@ -148,81 +222,11 @@
}
TFLITE_LOG(INFO) << "Overall mAP: "
<< precision_metrics.overall_mean_average_precision();
-
- return true;
}
-int Main(int argc, char* argv[]) {
- // Command Line Flags.
- std::string model_file_path;
- std::string ground_truth_images_path;
- std::string ground_truth_proto_file;
- std::string model_output_labels_path;
- std::string output_file_path;
- std::string delegate;
- int num_interpreter_threads = 1;
- bool debug_mode;
- std::vector<tflite::Flag> flag_list = {
- tflite::Flag::CreateFlag(kModelFileFlag, &model_file_path,
- "Path to test tflite model file."),
- tflite::Flag::CreateFlag(
- kModelOutputLabelsFlag, &model_output_labels_path,
- "Path to labels that correspond to output of model."
- " E.g. in case of COCO-trained SSD model, this is the path to file "
- "where each line contains a class detected by the model in correct "
- "order, starting from background."),
- tflite::Flag::CreateFlag(
- kGroundTruthImagesPathFlag, &ground_truth_images_path,
- "Path to ground truth images. These will be evaluated in "
- "alphabetical order of filenames"),
- tflite::Flag::CreateFlag(
- kGroundTruthProtoFileFlag, &ground_truth_proto_file,
- "Path to file containing "
- "tflite::evaluation::ObjectDetectionGroundTruth "
- "proto in text format. If left empty, mAP numbers are not output."),
- tflite::Flag::CreateFlag(
- kOutputFilePathFlag, &output_file_path,
- "File to output to. Contains only metrics proto if debug_mode is "
- "off, and per-image predictions also otherwise."),
- tflite::Flag::CreateFlag(kDebugModeFlag, &debug_mode,
- "Whether to enable debug mode. Per-image "
- "predictions are written to the output file "
- "along with metrics."),
- tflite::Flag::CreateFlag(
- kInterpreterThreadsFlag, &num_interpreter_threads,
- "Number of interpreter threads to use for inference."),
- tflite::Flag::CreateFlag(kDelegateFlag, &delegate,
- "Delegate to use for inference, if available. "
- "Must be one of {'nnapi', 'gpu'}"),
- };
- tflite::Flags::Parse(&argc, const_cast<const char**>(argv), flag_list);
- DelegateProviders delegate_providers;
- delegate_providers.InitFromCmdlineArgs(&argc, const_cast<const char**>(argv));
-
- // Process images in filename-sorted order.
- std::vector<std::string> image_paths;
- TF_LITE_ENSURE_STATUS(GetSortedFileNames(
- StripTrailingSlashes(ground_truth_images_path), &image_paths));
-
- std::vector<std::string> model_labels;
- if (!ReadFileLines(model_output_labels_path, &model_labels)) {
- TFLITE_LOG(ERROR) << "Could not read model output labels file";
- return EXIT_FAILURE;
- }
-
- if (!EvaluateModel(model_file_path, model_labels, image_paths,
- ground_truth_proto_file, delegate, output_file_path,
- num_interpreter_threads, debug_mode, delegate_providers)) {
- TFLITE_LOG(ERROR) << "Could not evaluate model";
- return EXIT_FAILURE;
- }
-
- return EXIT_SUCCESS;
+std::unique_ptr<TaskExecutor> CreateTaskExecutor(int* argc, char* argv[]) {
+ return std::unique_ptr<TaskExecutor>(new CocoObjectDetection(argc, argv));
}
} // namespace evaluation
} // namespace tflite
-
-int main(int argc, char* argv[]) {
- return tflite::evaluation::Main(argc, argv);
-}
diff --git a/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification/README.md b/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification/README.md
index ac8006b..8557de6 100644
--- a/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification/README.md
+++ b/tensorflow/lite/tools/evaluation/tasks/imagenet_image_classification/README.md
@@ -91,9 +91,10 @@
assumes that `libhexagon_interface.so` and Qualcomm libraries lie in
`/data/local/tmp`.
-This script also supports all applicable runtime/delegate arguments supported on
-the `benchmark_model` tool. If there is any conflict (for example, `num_threads`
-in `benchmark_model` vs `num_interpreter_threads` here), the parameters of this
+This script also supports runtime/delegate arguments introduced by the
+[delegate registrar](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/delegates).
+If there is any conflict (for example, `num_threads` vs
+`num_interpreter_threads` here), the parameters of this
script are given precedence.
## Downloading ILSVRC
diff --git a/tensorflow/lite/tools/evaluation/tasks/inference_diff/README.md b/tensorflow/lite/tools/evaluation/tasks/inference_diff/README.md
index c8873a6..64606ee 100644
--- a/tensorflow/lite/tools/evaluation/tasks/inference_diff/README.md
+++ b/tensorflow/lite/tools/evaluation/tasks/inference_diff/README.md
@@ -64,9 +64,10 @@
The final metrics are dumped into `output_file_path` as a serialized
instance of `tflite::evaluation::EvaluationStageMetrics`
-This script also supports all applicable runtime/delegate arguments supported on
-the `benchmark_model` tool. If there is any conflict (for example, `num_threads`
-in `benchmark_model` vs `num_interpreter_threads` here), the parameters of this
+This script also supports runtime/delegate arguments introduced by the
+[delegate registrar](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/delegates).
+If there is any conflict (for example, `num_threads` vs
+`num_interpreter_threads` here), the parameters of this
script are given precedence.
## Running the binary on Android
diff --git a/tensorflow/lite/tools/flatbuffer_utils.py b/tensorflow/lite/tools/flatbuffer_utils.py
index 5b513bb..f80daad 100644
--- a/tensorflow/lite/tools/flatbuffer_utils.py
+++ b/tensorflow/lite/tools/flatbuffer_utils.py
@@ -31,6 +31,8 @@
from flatbuffers.python import flatbuffers
from tensorflow.lite.python import schema_py_generated as schema_fb
+TFLITE_FILE_IDENTIFIER = b'TFL3'
+
def read_model(input_tflite_file):
"""Reads and parses a tflite model.
@@ -66,7 +68,7 @@
# Initial size of the buffer, which will grow automatically if needed
builder = flatbuffers.Builder(1024)
model_offset = model.Pack(builder)
- builder.Finish(model_offset)
+ builder.Finish(model_offset, file_identifier=TFLITE_FILE_IDENTIFIER)
model_data = builder.Output()
with open(output_tflite_file, 'wb') as out_file:
out_file.write(model_data)
diff --git a/tensorflow/lite/tools/flatbuffer_utils_test.py b/tensorflow/lite/tools/flatbuffer_utils_test.py
index d2e4fe6..60235b0 100644
--- a/tensorflow/lite/tools/flatbuffer_utils_test.py
+++ b/tensorflow/lite/tools/flatbuffer_utils_test.py
@@ -31,7 +31,7 @@
def testWriteReadModel(self):
# 1. SETUP
# Define the initial model
- initial_model = test_utils.build_mock_model_python_object()
+ initial_model = test_utils.build_mock_model()
# Define temporary files
tmp_dir = self.get_temp_dir()
model_filename = os.path.join(tmp_dir, 'model.tflite')
@@ -76,7 +76,7 @@
def testStripStrings(self):
# 1. SETUP
# Define the initial model
- initial_model = test_utils.build_mock_model_python_object()
+ initial_model = test_utils.build_mock_model()
final_model = copy.deepcopy(initial_model)
# 2. INVOKE
@@ -121,7 +121,7 @@
def testRandomizeWeights(self):
# 1. SETUP
# Define the initial model
- initial_model = test_utils.build_mock_model_python_object()
+ initial_model = test_utils.build_mock_model()
final_model = copy.deepcopy(initial_model)
# 2. INVOKE
diff --git a/tensorflow/lite/tools/gen_op_registration_main.cc b/tensorflow/lite/tools/gen_op_registration_main.cc
index 3e80831..464d0a5 100644
--- a/tensorflow/lite/tools/gen_op_registration_main.cc
+++ b/tensorflow/lite/tools/gen_op_registration_main.cc
@@ -13,7 +13,6 @@
limitations under the License.
==============================================================================*/
-#include <cassert>
#include <fstream>
#include <map>
#include <sstream>
@@ -21,8 +20,7 @@
#include <vector>
#include "absl/strings/strip.h"
-#include "tensorflow/core/platform/init_main.h"
-#include "tensorflow/core/util/command_line_flags.h"
+#include "tensorflow/lite/tools/command_line_flags.h"
#include "tensorflow/lite/tools/gen_op_registration.h"
const char kInputModelFlag[] = "input_model";
@@ -31,27 +29,26 @@
const char kTfLitePathFlag[] = "tflite_path";
const char kForMicro[] = "for_micro";
-using tensorflow::Flag;
-using tensorflow::Flags;
-using tensorflow::string;
-
void ParseFlagAndInit(int* argc, char** argv, string* input_model,
string* output_registration, string* tflite_path,
string* namespace_flag, bool* for_micro) {
- std::vector<tensorflow::Flag> flag_list = {
- Flag(kInputModelFlag, input_model, "path to the tflite model"),
- Flag(kOutputRegistrationFlag, output_registration,
- "filename for generated registration code"),
- Flag(kTfLitePathFlag, tflite_path, "Path to tensorflow lite dir"),
- Flag(kNamespace, namespace_flag,
- "Namespace in which to put RegisterSelectedOps."),
- Flag(kForMicro, for_micro,
- "By default this script generate TFL registration file, but can "
- "also generate TFLM files when this flag is set to true"),
+ std::vector<tflite::Flag> flag_list = {
+ tflite::Flag::CreateFlag(kInputModelFlag, input_model,
+ "path to the tflite model"),
+ tflite::Flag::CreateFlag(kOutputRegistrationFlag, output_registration,
+ "filename for generated registration code"),
+ tflite::Flag::CreateFlag(kTfLitePathFlag, tflite_path,
+ "Path to tensorflow lite dir"),
+ tflite::Flag::CreateFlag(
+ kNamespace, namespace_flag,
+ "Namespace in which to put RegisterSelectedOps."),
+ tflite::Flag::CreateFlag(
+ kForMicro, for_micro,
+ "By default this script generate TFL registration file, but can "
+ "also generate TFLM files when this flag is set to true"),
};
- Flags::Parse(argc, argv, flag_list);
- tensorflow::port::InitMain(argv[0], argc, &argv);
+ tflite::Flags::Parse(argc, const_cast<const char**>(argv), flag_list);
}
namespace {
@@ -66,11 +63,10 @@
if (for_micro) {
if (!builtin_ops.empty()) {
- fout << "#include \"" << tflite_path
- << "/experimental/micro/kernels/micro_ops.h\"\n";
+ fout << "#include \"" << tflite_path << "/micro/kernels/micro_ops.h\"\n";
}
fout << "#include \"" << tflite_path
- << "/experimental/micro/micro_mutable_op_resolver.h\"\n";
+ << "/micro/micro_mutable_op_resolver.h\"\n";
} else {
if (!builtin_ops.empty()) {
fout << "#include \"" << tflite_path
@@ -151,7 +147,7 @@
string output_registration;
string tflite_path;
string namespace_flag;
- bool for_micro;
+ bool for_micro = false;
ParseFlagAndInit(&argc, argv, &input_model, &output_registration,
&tflite_path, &namespace_flag, &for_micro);
diff --git a/tensorflow/lite/tools/make/Makefile b/tensorflow/lite/tools/make/Makefile
index 426ed63..ad3832f 100644
--- a/tensorflow/lite/tools/make/Makefile
+++ b/tensorflow/lite/tools/make/Makefile
@@ -184,6 +184,14 @@
CXXFLAGS += -DTFLITE_WITH_RUY
endif
+BUILD_WITH_RUY_PROFILER ?= false
+ifeq ($(BUILD_WITH_RUY_PROFILER),true)
+ CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/instrumentation.cc
+ CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/profiler.cc
+ CORE_CC_ALL_SRCS += tensorflow/lite/tools/make/downloads/ruy/ruy/profiler/treeview.cc
+ CXXFLAGS += -DRUY_PROFILER
+endif
+
# Not to include XNNPACK.
CXXFLAGS += -DTFLITE_WITHOUT_XNNPACK
@@ -212,7 +220,7 @@
# Benchmark sources
BENCHMARK_SRCS_DIR := tensorflow/lite/tools/benchmark
-DELEGATE_PROVIDER_SRCS_DIR := tensorflow/lite/tools/benchmark
+DELEGATE_PROVIDER_SRCS_DIR := tensorflow/lite/tools/delegates
EVALUATION_UTILS_SRCS := \
tensorflow/lite/tools/evaluation/utils.cc
BENCHMARK_ALL_SRCS := \
diff --git a/tensorflow/lite/tools/optimize/calibration/node_info_delegate_test.cc b/tensorflow/lite/tools/optimize/calibration/node_info_delegate_test.cc
index b110174..722bdbd 100644
--- a/tensorflow/lite/tools/optimize/calibration/node_info_delegate_test.cc
+++ b/tensorflow/lite/tools/optimize/calibration/node_info_delegate_test.cc
@@ -92,7 +92,7 @@
TfLiteDelegate delegate = CreateNodeInfoDelegate(¶ms);
auto status = interpreter->ModifyGraphWithDelegate(&delegate);
- EXPECT_EQ(kTfLiteError, status);
+ EXPECT_EQ(kTfLiteDelegateError, status);
}
TEST(NodeInfoDelegateTest, NodeInfoDelegateObserver) {
diff --git a/tensorflow/lite/tools/optimize/modify_model_interface_main.cc b/tensorflow/lite/tools/optimize/modify_model_interface_main.cc
index e76b46b..24674a1 100644
--- a/tensorflow/lite/tools/optimize/modify_model_interface_main.cc
+++ b/tensorflow/lite/tools/optimize/modify_model_interface_main.cc
@@ -25,12 +25,12 @@
return 1;
}
- if (!strcmp(argv[3], "uint8") && !strcmp(argv[3], "int8")) {
+ if (strcmp(argv[3], "uint8") && strcmp(argv[3], "int8")) {
printf("Only support uint8 and int8 for input interface");
return 1;
}
- if (!strcmp(argv[4], "uint8") && !strcmp(argv[4], "int8")) {
+ if (strcmp(argv[4], "uint8") && strcmp(argv[4], "int8")) {
printf("Only support uint8 and int8 for output interface");
return 1;
}
diff --git a/tensorflow/lite/tools/optimize/operator_property.cc b/tensorflow/lite/tools/optimize/operator_property.cc
index 6a32858..71fdad8 100644
--- a/tensorflow/lite/tools/optimize/operator_property.cc
+++ b/tensorflow/lite/tools/optimize/operator_property.cc
@@ -130,9 +130,10 @@
tensor_property.per_axis = true;
tensor_property.per_axis_index = 0;
tensor_property.symmetric = true;
- property.inputs = {{1, tensor_property}, {2, {}}};
+ property.inputs = {{2, {}}, {1, tensor_property}};
property.outputs = {{0, {}}};
- property.version = 2;
+ property.biases = {3};
+ property.version = 3;
break;
}
case BuiltinOperator_DEPTHWISE_CONV_2D: {
diff --git a/tensorflow/lite/tools/test_utils.py b/tensorflow/lite/tools/test_utils.py
index 75a649b..3950e3d 100644
--- a/tensorflow/lite/tools/test_utils.py
+++ b/tensorflow/lite/tools/test_utils.py
@@ -14,7 +14,7 @@
# ==============================================================================
"""Utility functions that support testing.
-All functions that can be commonly used by various tests are in this file.
+All functions that can be commonly used by various tests.
"""
from __future__ import absolute_import
@@ -24,8 +24,10 @@
from flatbuffers.python import flatbuffers
from tensorflow.lite.python import schema_py_generated as schema_fb
+TFLITE_SCHEMA_VERSION = 3
-def build_mock_model():
+
+def build_mock_flatbuffer_model():
"""Creates a flatbuffer containing an example model."""
builder = flatbuffers.Builder(1024)
@@ -194,6 +196,7 @@
string4_offset = builder.CreateString('model_description')
schema_fb.ModelStart(builder)
+ schema_fb.ModelAddVersion(builder, TFLITE_SCHEMA_VERSION)
schema_fb.ModelAddOperatorCodes(builder, codes_offset)
schema_fb.ModelAddSubgraphs(builder, subgraphs_offset)
schema_fb.ModelAddDescription(builder, string4_offset)
@@ -205,10 +208,14 @@
return model
-def build_mock_model_python_object():
- """Creates a python flatbuffer object containing an example model."""
- model_mock = build_mock_model()
- model_obj = schema_fb.Model.GetRootAsModel(model_mock, 0)
- model = schema_fb.ModelT.InitFromObj(model_obj)
-
+def load_model_from_flatbuffer(flatbuffer_model):
+ """Loads a model as a python object from a flatbuffer model."""
+ model = schema_fb.Model.GetRootAsModel(flatbuffer_model, 0)
+ model = schema_fb.ModelT.InitFromObj(model)
return model
+
+
+def build_mock_model():
+ """Creates an object containing an example model."""
+ model = build_mock_flatbuffer_model()
+ return load_model_from_flatbuffer(model)
diff --git a/tensorflow/lite/tools/versioning/op_version.cc b/tensorflow/lite/tools/versioning/op_version.cc
index 2b7d3f7..1107f04 100644
--- a/tensorflow/lite/tools/versioning/op_version.cc
+++ b/tensorflow/lite/tools/versioning/op_version.cc
@@ -198,6 +198,10 @@
return 1;
case BuiltinOperator_TRANSPOSE_CONV:
+ // If the op has 4 inputs, it is version 3.
+ if (op_sig.input_types.size() == 4) {
+ return 3;
+ }
// If the op takes int8 input, it is version 2.
if (op_sig.input_types.at(0) == TensorType_INT8) {
return 2;
@@ -407,6 +411,18 @@
}
return 1;
+ case BuiltinOperator_EQUAL:
+ case BuiltinOperator_NOT_EQUAL:
+ if (!op_sig.input_types.empty()) {
+ if (op_sig.input_types.at(0) == TensorType_STRING) {
+ return 3;
+ }
+ if (op_sig.input_types.at(0) == TensorType_INT8) {
+ return 2;
+ }
+ }
+ return 1;
+
case BuiltinOperator_ADD:
case BuiltinOperator_CONCATENATION:
case BuiltinOperator_PAD:
@@ -426,8 +442,6 @@
case BuiltinOperator_TOPK_V2:
case BuiltinOperator_ARG_MAX:
case BuiltinOperator_ARG_MIN:
- case BuiltinOperator_EQUAL:
- case BuiltinOperator_NOT_EQUAL:
case BuiltinOperator_GREATER:
case BuiltinOperator_GREATER_EQUAL:
case BuiltinOperator_LESS:
diff --git a/tensorflow/lite/tools/versioning/op_version_test.cc b/tensorflow/lite/tools/versioning/op_version_test.cc
index 5dde260..22417c7 100644
--- a/tensorflow/lite/tools/versioning/op_version_test.cc
+++ b/tensorflow/lite/tools/versioning/op_version_test.cc
@@ -86,10 +86,20 @@
TEST(OpVersionTest, VersioningEqualTest) {
SimpleVersioningTest(BuiltinOperator_EQUAL);
+ OpSignature fake_op_sig = {
+ .op = BuiltinOperator_EQUAL,
+ .input_types = std::vector<TensorType>{TensorType_STRING},
+ };
+ EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 3);
}
TEST(OpVersionTest, VersioningNotEqualTest) {
SimpleVersioningTest(BuiltinOperator_NOT_EQUAL);
+ OpSignature fake_op_sig = {
+ .op = BuiltinOperator_NOT_EQUAL,
+ .input_types = std::vector<TensorType>{TensorType_STRING},
+ };
+ EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 3);
}
TEST(OpVersionTest, VersioningLessTest) {
@@ -422,6 +432,13 @@
.input_types = std::vector<TensorType>{TensorType_INT8},
};
EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 2);
+
+ fake_op_sig = {
+ .op = BuiltinOperator_TRANSPOSE_CONV,
+ .input_types = std::vector<TensorType>{TensorType_INT32, TensorType_INT8,
+ TensorType_INT8, TensorType_INT32},
+ };
+ EXPECT_EQ(GetBuiltinOperatorVersion(fake_op_sig), 3);
}
TEST(OpVersionTest, VersioningSVDFOperatorTest) {
diff --git a/tensorflow/lite/tools/visualize_test.py b/tensorflow/lite/tools/visualize_test.py
index 8beb8f8..aa74891 100644
--- a/tensorflow/lite/tools/visualize_test.py
+++ b/tensorflow/lite/tools/visualize_test.py
@@ -35,9 +35,9 @@
self.assertEqual('HASHTABLE_LOOKUP', visualize.BuiltinCodeToName(10))
def testFlatbufferToDict(self):
- model_data = test_utils.build_mock_model()
- model_dict = visualize.CreateDictFromFlatbuffer(model_data)
- self.assertEqual(0, model_dict['version'])
+ model = test_utils.build_mock_flatbuffer_model()
+ model_dict = visualize.CreateDictFromFlatbuffer(model)
+ self.assertEqual(test_utils.TFLITE_SCHEMA_VERSION, model_dict['version'])
self.assertEqual(1, len(model_dict['subgraphs']))
self.assertEqual(1, len(model_dict['operator_codes']))
self.assertEqual(3, len(model_dict['buffers']))
@@ -45,12 +45,11 @@
self.assertEqual(0, model_dict['subgraphs'][0]['tensors'][0]['buffer'])
def testVisualize(self):
- model_data = test_utils.build_mock_model()
-
+ model = test_utils.build_mock_flatbuffer_model()
tmp_dir = self.get_temp_dir()
model_filename = os.path.join(tmp_dir, 'model.tflite')
with open(model_filename, 'wb') as model_file:
- model_file.write(model_data)
+ model_file.write(model)
html_filename = os.path.join(tmp_dir, 'visualization.html')
visualize.CreateHtmlFile(model_filename, html_filename)
diff --git a/tensorflow/opensource_only.files b/tensorflow/opensource_only.files
index 1e45bf6..d4df3df 100644
--- a/tensorflow/opensource_only.files
+++ b/tensorflow/opensource_only.files
@@ -8,6 +8,7 @@
tensorflow/python/autograph/core/config.py
tensorflow/python/distribute/cluster_resolver/tpu_cluster_resolver.py
tensorflow/python/distribute/cluster_resolver/tpu_cluster_resolver_test.py
+tensorflow/python/eager/benchmarks_test_base.py
tensorflow/python/tpu/profiler/pip_package/BUILD
tensorflow/python/tpu/profiler/pip_package/README
tensorflow/python/tpu/profiler/pip_package/build_pip_package.sh
@@ -262,6 +263,7 @@
tensorflow/third_party/toolchains/remote_config/rbe_config.bzl
tensorflow/third_party/wrapt.BUILD
tensorflow/third_party/zlib.BUILD
+tensorflow/tools/build_info/BUILD
tensorflow/tools/ci_build/release/common.sh
tensorflow/tools/ci_build/release/common_win.bat
tensorflow/tools/ci_build/release/macos/cpu_libtensorflow/build.sh
@@ -392,6 +394,7 @@
tensorflow/tools/def_file_filter/BUILD.tpl
tensorflow/tools/def_file_filter/def_file_filter.py.tpl
tensorflow/tools/def_file_filter/def_file_filter_configure.bzl
+tensorflow/tools/git/BUILD
tensorflow/tools/lib_package/BUILD
tensorflow/tools/lib_package/LibTensorFlowTest.java
tensorflow/tools/lib_package/README.md
diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD
index 8be4c14..8e680c1 100644
--- a/tensorflow/python/BUILD
+++ b/tensorflow/python/BUILD
@@ -26,6 +26,9 @@
# buildifier: disable=same-origin-load
load("//tensorflow:tensorflow.bzl", "tf_pybind_cc_library_wrapper")
+
+# buildifier: disable=same-origin-load
+load("//tensorflow:tensorflow.bzl", "tf_local_platform_constraint")
load("//tensorflow/core/platform:build_config.bzl", "pyx_library", "tf_additional_all_protos", "tf_additional_lib_deps", "tf_proto_library", "tf_proto_library_py", "tf_protos_grappler") # @unused
load("//tensorflow/core/platform:build_config_root.bzl", "if_static", "tf_additional_plugin_deps", "tf_additional_xla_deps_py")
load("//tensorflow/python:build_defs.bzl", "tf_gen_op_wrapper_private_py")
@@ -233,9 +236,12 @@
],
)
+# TODO(gunan): Investigate making this action hermetic so we do not need
+# to run it locally.
tf_py_build_info_genrule(
name = "py_build_info_gen",
out = "platform/build_info.py",
+ exec_compatible_with = tf_local_platform_constraint(),
)
py_library(
@@ -506,7 +512,7 @@
hdrs = ["lib/core/ndarray_tensor_bridge.h"],
visibility = tf_external_workspace_visible(
visibility + [
- "//learning/deepmind/courier:__subpackages__",
+ "//tensorflow:ndarray_tensor_allow_list",
],
),
deps = [
@@ -980,7 +986,7 @@
srcs = ["lib/core/ndarray_tensor.cc"],
hdrs = ["lib/core/ndarray_tensor.h"],
visibility = tf_external_workspace_visible(visibility + [
- "//learning/deepmind/courier:__subpackages__",
+ "//tensorflow:ndarray_tensor_allow_list",
]),
deps = [
":bfloat16_lib",
@@ -1550,7 +1556,6 @@
":platform",
":registry",
":tensor_conversion_registry",
- ":tensor_like",
":tensor_shape",
":tf2",
":traceable_stack",
@@ -1590,13 +1595,6 @@
)
py_library(
- name = "tensor_like",
- srcs = ["framework/tensor_like.py"],
- srcs_version = "PY2AND3",
- deps = [],
-)
-
-py_library(
name = "indexed_slices",
srcs = ["framework/indexed_slices.py"],
srcs_version = "PY2AND3",
@@ -1608,6 +1606,7 @@
":type_spec",
":util",
"//tensorflow/python/eager:context",
+ "//tensorflow/python/types",
],
)
@@ -1768,9 +1767,9 @@
":composite_tensor",
":dtypes",
":framework_ops",
- ":tensor_like",
":tensor_util",
":type_spec",
+ "//tensorflow/python/types",
],
)
@@ -1878,7 +1877,6 @@
srcs = ["framework/tensor_util.py"],
srcs_version = "PY2AND3",
deps = [
- ":tensor_like",
":tensor_shape",
":util",
"//tensorflow/core:protos_all_py",
@@ -2471,9 +2469,11 @@
main = "framework/sparse_tensor_test.py",
python_version = "PY3",
deps = [
+ ":array_ops",
":framework",
":framework_for_generated_wrappers",
":framework_test_lib",
+ ":math_ops",
":platform_test",
"//tensorflow/core:protos_all_py",
],
@@ -3222,6 +3222,27 @@
],
)
+tf_py_test(
+ name = "collective_ops_xla_test",
+ size = "small",
+ srcs = ["ops/collective_ops_xla_test.py"],
+ python_version = "PY3",
+ tags = [
+ "no_pip",
+ "no_rocm",
+ "no_windows",
+ "nomac",
+ ],
+ xla_enable_strict_auto_jit = True,
+ deps = [
+ ":client_testlib",
+ ":collective_ops",
+ ":framework_for_generated_wrappers",
+ ":kernels",
+ "//third_party/py/numpy",
+ ],
+)
+
cuda_py_test(
name = "collective_ops_gpu_test",
size = "small",
@@ -5932,7 +5953,11 @@
"//tensorflow/compiler/jit:flags", #tfe
"//tensorflow/compiler/mlir/python:mlir", # mlir
"//tensorflow/core/common_runtime:core_cpu_base_no_ops", # tf_session
- "//tensorflow/core:core_cpu_impl", # device_lib
+ "//tensorflow/core/common_runtime:core_cpu_rump_impl", # quantize_training
+ "//tensorflow/core/common_runtime:device", # device_lib, tfe, tf_session
+ "//tensorflow/core/common_runtime:device_factory", # device_lib, tfe, tf_session
+ "//tensorflow/core/common_runtime:session_options", # device_lib, tfe, tf_session
+ "//tensorflow/core/common_runtime:session_state", # tf_session
"//tensorflow/core/data/service:server_lib", # server_lib
"//tensorflow/core:framework_internal_impl", # op_def_registry
"//tensorflow/core:lib_internal_impl", # device_lib
@@ -6447,6 +6472,7 @@
python_version = "PY3",
tags = [
"gpu_cupti",
+ "no_gpu", # b/154742661
],
xla_enable_strict_auto_jit = False, # Graph structure is different with autojit
deps = [
@@ -8020,6 +8046,29 @@
],
)
+tf_python_pybind_extension(
+ name = "_pywrap_parallel_device",
+ srcs = [
+ "lib/core/safe_ptr.h",
+ "//tensorflow/c:headers",
+ "//tensorflow/c/eager:headers",
+ "//tensorflow/c/eager/parallel_device:headers",
+ "//tensorflow/c/eager/parallel_device:sources",
+ "//tensorflow/python/distribute/parallel_device:pywrap_parallel_device.cc",
+ ],
+ module_name = "_pywrap_parallel_device",
+ visibility = ["//tensorflow/python/distribute/parallel_device:__pkg__"],
+ deps = [
+ "//tensorflow/core:framework_headers_lib",
+ "//tensorflow/core:lib_headers_for_pybind",
+ "//tensorflow/core:protos_all_cc",
+ "//tensorflow/python:pybind11_lib",
+ "//tensorflow/python:pybind11_status",
+ "//third_party/python_runtime:headers",
+ "@pybind11",
+ ],
+)
+
pyx_library(
name = "framework_fast_tensor_util",
srcs = ["framework/fast_tensor_util.pyx"],
diff --git a/tensorflow/python/autograph/converters/break_statements.py b/tensorflow/python/autograph/converters/break_statements.py
index 60e65e9..8de4865 100644
--- a/tensorflow/python/autograph/converters/break_statements.py
+++ b/tensorflow/python/autograph/converters/break_statements.py
@@ -53,7 +53,7 @@
return block
template = """
- if ag__.not_(var_name):
+ if not var_name:
block
"""
node = templates.replace(
@@ -100,7 +100,7 @@
template = """
var_name = False
- while ag__.and_(lambda: test, lambda: ag__.not_(var_name)):
+ while not var_name and test:
body
orelse
"""
@@ -150,7 +150,7 @@
# break did not trigger).
guarded_orelse = self._guard_if_present(node.orelse, break_var)
extra_test = templates.replace_as_expression(
- 'ag__.not_(var_name)', var_name=break_var)
+ 'not var_name', var_name=break_var)
# The extra test is hidden in the AST, which will confuse the static
# analysis. To mitigate that, we insert a no-op statement that ensures
diff --git a/tensorflow/python/autograph/converters/continue_statements.py b/tensorflow/python/autograph/converters/continue_statements.py
index 551ee19..2a1b56a 100644
--- a/tensorflow/python/autograph/converters/continue_statements.py
+++ b/tensorflow/python/autograph/converters/continue_statements.py
@@ -83,7 +83,7 @@
block.create_guard_next = False
if should_wrap_current:
template = """
- if ag__.not_(var_name):
+ if not var_name:
original_node
"""
cond, = templates.replace(
diff --git a/tensorflow/python/autograph/converters/return_statements.py b/tensorflow/python/autograph/converters/return_statements.py
index 8341187..4d262d9 100644
--- a/tensorflow/python/autograph/converters/return_statements.py
+++ b/tensorflow/python/autograph/converters/return_statements.py
@@ -256,7 +256,7 @@
state = self.state[_Block]
if state.create_guard_now:
template = """
- if ag__.not_(do_return_var_name):
+ if not do_return_var_name:
original_node
"""
cond, = templates.replace(
@@ -285,7 +285,7 @@
node.body = self._visit_statement_block(node, node.body)
if self.state[_Block].return_used:
node.test = templates.replace_as_expression(
- 'ag__.and_(lambda: ag__.not_(control_var), lambda: test)',
+ 'not control_var and test',
test=node.test,
control_var=self.state[_Function].do_return_var_name)
@@ -302,12 +302,12 @@
extra_test = anno.getanno(node, anno.Basic.EXTRA_LOOP_TEST, default=None)
if extra_test is not None:
extra_test = templates.replace_as_expression(
- 'ag__.and_(lambda: ag__.not_(control_var), lambda: extra_test)',
+ 'not control_var and extra_test',
extra_test=extra_test,
control_var=self.state[_Function].do_return_var_name)
else:
extra_test = templates.replace_as_expression(
- 'ag__.not_(control_var)',
+ 'not control_var',
control_var=self.state[_Function].do_return_var_name)
anno.setanno(node, anno.Basic.EXTRA_LOOP_TEST, extra_test)
diff --git a/tensorflow/python/autograph/g3doc/pyct_tutorial.ipynb b/tensorflow/python/autograph/g3doc/pyct_tutorial.ipynb
new file mode 100644
index 0000000..51abdf6
--- /dev/null
+++ b/tensorflow/python/autograph/g3doc/pyct_tutorial.ipynb
@@ -0,0 +1,512 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "fWfkYsCgPvqR"
+ },
+ "source": [
+ "# Short intro to the SCT library of AutoGraph\n",
+ "\n",
+ "**Work in progress, use with care and expect changes.**\n",
+ "\n",
+ "The `pyct` module packages the source code transformation APIs used by AutoGraph.\n",
+ "\n",
+ "This tutorial is just a preview - there is no PIP package yet, and the API has not been finalized, although most of those shown here are quite stable.\n",
+ "\n",
+ "[Run in Colab](https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/python/autograph/g3doc/pyct_tutorial.ipynb)\n",
+ "\n",
+ "Requires `tf-nightly`:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "wq1DRamRlqoB"
+ },
+ "outputs": [],
+ "source": [
+ "!pip install tf-nightly"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "r7Q78WIKe2cu"
+ },
+ "source": [
+ "### Writing a custom code generator\n",
+ "\n",
+ "[transformer.CodeGenerator](https://github.com/tensorflow/tensorflow/blob/40802bcdb5c8a4379da2145441f51051402bd29b/tensorflow/python/autograph/pyct/transformer.py#L480) is an AST visitor that outputs a string. This makes it useful in the final stage of translating Python to another language."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "HHaCMFOpuoVx"
+ },
+ "source": [
+ "Here's a toy C++ code generator written using a `transformer.CodeGenerator`, which is just a fancy subclass of [ast.NodeVisitor](https://docs.python.org/3/library/ast.html#ast.NodeVisitor):"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "PJlTIbJlurpm"
+ },
+ "outputs": [],
+ "source": [
+ "import gast\n",
+ "from tensorflow.python.autograph.pyct import transformer\n",
+ "\n",
+ "class BasicCppCodegen(transformer.CodeGenerator):\n",
+ "\n",
+ " def visit_Name(self, node):\n",
+ " self.emit(node.id)\n",
+ "\n",
+ " def visit_arguments(self, node):\n",
+ " self.visit(node.args[0])\n",
+ " for arg in node.args[1:]:\n",
+ " self.emit(', ')\n",
+ " self.visit(arg)\n",
+ "\n",
+ " def visit_FunctionDef(self, node):\n",
+ " self.emit('void {}'.format(node.name))\n",
+ " self.emit('(')\n",
+ " self.visit(node.args)\n",
+ " self.emit(') {\\n')\n",
+ " self.visit_block(node.body)\n",
+ " self.emit('\\n}')\n",
+ "\n",
+ " def visit_Call(self, node):\n",
+ " self.emit(node.func.id)\n",
+ " self.emit('(')\n",
+ " self.visit(node.args[0])\n",
+ " for arg in node.args[1:]:\n",
+ " self.emit(', ')\n",
+ " self.visit(arg)\n",
+ " self.emit(');')\n"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "nUhlScyOjlYM"
+ },
+ "source": [
+ "Let's try it on a simple function:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "ty9q853QvUqo"
+ },
+ "outputs": [],
+ "source": [
+ "def f(x, y):\n",
+ " print(x, y)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "R8N15UpVvbmu"
+ },
+ "source": [
+ "First, parse the Python code and annotate the AST. This is easily done with standard libraries, but [parser.parse_entity](https://github.com/tensorflow/tensorflow/blob/40802bcdb5c8a4379da2145441f51051402bd29b/tensorflow/python/autograph/pyct/parser.py#L182) makes it a single call. It returns a [gast](https://github.com/serge-sans-paille/gast) AST, so you don't have to worry about Python version:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "Cs_Ls0MesvBp"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import parser\n",
+ "\n",
+ "node, source = parser.parse_entity(f, ())"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "kuT7J-xps_2Y"
+ },
+ "source": [
+ "There are a couple of context objects that most transformer objects like `CodeGenerator` use.\n",
+ "\n",
+ "Of note here is `EntityInfo.namespace`, which contains the runtime values for all the global and closure names that the function has access to. Inside a transformer object, this is available under `self.ctx.info.namespace`.\n",
+ "\n",
+ "For example, if a function uses NumPy, its namespace will typically include `'np'`."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "pnB63kpttIVU"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import inspect_utils\n",
+ "\n",
+ "f_info = transformer.EntityInfo(\n",
+ " name='f',\n",
+ " source_code=source,\n",
+ " source_file=None,\n",
+ " future_features=(),\n",
+ " namespace=inspect_utils.getnamespace(f))\n",
+ "ctx = transformer.Context(f_info, None, None)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "kCjcucmiwW98"
+ },
+ "source": [
+ "Finally, it's just a matter of running the generator:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "SdHjliuuwaaJ"
+ },
+ "outputs": [],
+ "source": [
+ "codegen = BasicCppCodegen(ctx)\n",
+ "codegen.visit(node)\n",
+ "\n",
+ "print(codegen.code_buffer)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "rmRI9dG_ydE_"
+ },
+ "source": [
+ "### Helpful static analysis passes\n",
+ "\n",
+ "The `static_analysis` module contains various helper passes for dataflow analyis.\n",
+ "\n",
+ "All these passes annotate the AST. These annotations can be extracted using [anno.getanno](https://github.com/tensorflow/tensorflow/blob/40802bcdb5c8a4379da2145441f51051402bd29b/tensorflow/python/autograph/pyct/anno.py#L111). Most of them rely on the `qual_names` annotations, which just simplify the way more complex identifiers like `a.b.c` are accessed.\n",
+ "\n",
+ "The most useful is the activity analysis which just inventories symbols read, modified, etc.:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "GEJ30Wea4Xfy"
+ },
+ "outputs": [],
+ "source": [
+ "def get_node_and_ctx(f):\n",
+ " node, source = parser.parse_entity(f, ())\n",
+ " f_info = transformer.EntityInfo(\n",
+ " name='f',\n",
+ " source_code=source,\n",
+ " source_file=None,\n",
+ " future_features=(),\n",
+ " namespace=None)\n",
+ " ctx = transformer.Context(f_info, None, None)\n",
+ " return node, ctx"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "BiwPJrDd0aAX"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import anno\n",
+ "from tensorflow.python.autograph.pyct import qual_names\n",
+ "from tensorflow.python.autograph.pyct.static_analysis import annos\n",
+ "from tensorflow.python.autograph.pyct.static_analysis import activity\n",
+ "\n",
+ "\n",
+ "def f(a):\n",
+ " b = a + 1\n",
+ " return b\n",
+ "\n",
+ "\n",
+ "node, ctx = get_node_and_ctx(f)\n",
+ "\n",
+ "node = qual_names.resolve(node)\n",
+ "node = activity.resolve(node, ctx)\n",
+ "\n",
+ "fn_scope = anno.getanno(node, annos.NodeAnno.BODY_SCOPE) # Note: tag will be changed soon.\n",
+ "\n",
+ "\n",
+ "print('read:', fn_scope.read)\n",
+ "print('modified:', fn_scope.modified)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "w8dBRlKkFNIP"
+ },
+ "source": [
+ "Another useful utility is the control flow graph builder.\n",
+ "\n",
+ "Of course, a CFG that fully accounts for all effects is impractical to build in a late-bound language like Python without creating an almost fully-connected graph. However, one can be reasonably built if we ignore the potential for functions to raise arbitrary exceptions."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "KvLe9lWnFg7N"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import cfg\n",
+ "\n",
+ "\n",
+ "def f(a):\n",
+ " if a \u003e 0:\n",
+ " return a\n",
+ " b = -a\n",
+ "\n",
+ "node, ctx = get_node_and_ctx(f)\n",
+ "\n",
+ "node = qual_names.resolve(node)\n",
+ "cfgs = cfg.build(node)\n",
+ "cfgs[node]"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "Cro-jfPA2oxR"
+ },
+ "source": [
+ "Other useful analyses include liveness analysis. Note that these make simplifying assumptions, because in general the CFG of a Python program is a graph that's almost complete. The only robust assumption is that execution can't jump backwards."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "73dARy4_2oAI"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import anno\n",
+ "from tensorflow.python.autograph.pyct import cfg\n",
+ "from tensorflow.python.autograph.pyct import qual_names\n",
+ "from tensorflow.python.autograph.pyct.static_analysis import annos\n",
+ "from tensorflow.python.autograph.pyct.static_analysis import liveness\n",
+ "\n",
+ "\n",
+ "def f(a):\n",
+ " b = a + 1\n",
+ " return b\n",
+ "\n",
+ "\n",
+ "node, ctx = get_node_and_ctx(f)\n",
+ "\n",
+ "node = qual_names.resolve(node)\n",
+ "cfgs = cfg.build(node)\n",
+ "node = activity.resolve(node, ctx)\n",
+ "node = liveness.resolve(node, ctx, cfgs)\n",
+ "\n",
+ "print('live into `b = a + 1`:', anno.getanno(node.body[0], anno.Static.LIVE_VARS_IN))\n",
+ "print('live into `return b`:', anno.getanno(node.body[1], anno.Static.LIVE_VARS_IN))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "GKSaqLbKQI_v"
+ },
+ "source": [
+ "### Writing a custom Python transpiler\n",
+ "\n",
+ "`transpiler.FunctionTranspiler` is a generic class for a Python [source-to-source compiler](https://en.wikipedia.org/wiki/Source-to-source_compiler). It operates on Python ASTs. Subclasses override its [transform_ast](https://github.com/tensorflow/tensorflow/blob/95ea3404528afcb1a74dd5f0946ea8d17beda28b/tensorflow/python/autograph/pyct/transpiler.py#L261) method.\n",
+ "\n",
+ "Unlike the `transformer` module, which have an AST as input/output, the `transpiler` APIs accept and return actual Python objects, handling the tasks associated with parsing, unparsing and loading of code."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "eicHoYlzRhnc"
+ },
+ "source": [
+ "Here's a transpiler that does nothing:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "edaG6dWEPvUI"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import transpiler\n",
+ "\n",
+ "\n",
+ "class NoopTranspiler(transpiler.FunctionTranspiler):\n",
+ "\n",
+ " def transform_ast(self, ast, transformer_context):\n",
+ " return ast\n",
+ "\n",
+ "tr = NoopTranspiler()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "hKxmlWeQSQyN"
+ },
+ "source": [
+ "The main method is [transform_function](https://github.com/tensorflow/tensorflow/blob/95ea3404528afcb1a74dd5f0946ea8d17beda28b/tensorflow/python/autograph/pyct/transpiler.py#L384), which as its name suggests, operates on functions."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "HXTIYsunSVr1"
+ },
+ "outputs": [],
+ "source": [
+ "def f(x, y):\n",
+ " return x + y\n",
+ "\n",
+ "\n",
+ "new_f, _, _ = tr.transform_function(f, None, None, {})\n",
+ "\n",
+ "print(new_f(1, 1))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "metadata": {
+ "colab_type": "text",
+ "id": "aKO42LBXw3SD"
+ },
+ "source": [
+ "### Adding new variables to the transformed code\n",
+ "\n",
+ "The transformed function has the same global and local variables as the original function. You can of course generate local imports to add any new references into the generated code, but an easier method is to use the `extra_locals` arg of `transform_function`:"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "_Wl0n5I_1NJZ"
+ },
+ "outputs": [],
+ "source": [
+ "from tensorflow.python.autograph.pyct import parser\n",
+ "\n",
+ "\n",
+ "class HelloTranspiler(transpiler.FunctionTranspiler):\n",
+ "\n",
+ " def transform_ast(self, ast, transformer_context):\n",
+ " print_code = parser.parse('print(\"Hello\", name)')\n",
+ " ast.body = [print_code] + ast.body\n",
+ " return ast\n",
+ "\n",
+ "\n",
+ "def f(x, y):\n",
+ " pass\n",
+ "\n",
+ "\n",
+ "extra_locals = {'name': 'you'}\n",
+ "new_f, _, _ = HelloTranspiler().transform_function(f, None, None, extra_locals)\n",
+ "\n",
+ "_ = new_f(1, 1)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 0,
+ "metadata": {
+ "colab": {},
+ "colab_type": "code",
+ "id": "JcMSHJXK6pO2"
+ },
+ "outputs": [],
+ "source": [
+ "import inspect\n",
+ "\n",
+ "print(inspect.getsource(new_f))"
+ ]
+ }
+ ],
+ "metadata": {
+ "colab": {
+ "collapsed_sections": [],
+ "name": "pyctr_tutorial.ipynb",
+ "provenance": [
+ {
+ "file_id": "1dT93XRkt7vUpVp7GZech8LB0u1OytKff",
+ "timestamp": 1586205976756
+ }
+ ]
+ },
+ "kernelspec": {
+ "display_name": "Python 3",
+ "name": "python3"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}
diff --git a/tensorflow/python/autograph/operators/py_builtins.py b/tensorflow/python/autograph/operators/py_builtins.py
index d444722..4dbe25a 100644
--- a/tensorflow/python/autograph/operators/py_builtins.py
+++ b/tensorflow/python/autograph/operators/py_builtins.py
@@ -29,6 +29,7 @@
from tensorflow.python.autograph.utils import py_func
from tensorflow.python.autograph.utils import tensors
+from tensorflow.python.data.experimental.ops import cardinality
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.data.ops import iterator_ops
from tensorflow.python.framework import constant_op
@@ -234,6 +235,8 @@
return _tf_tensor_list_len(s)
elif tensor_util.is_tensor(s):
return _tf_tensor_len(s)
+ if isinstance(s, dataset_ops.DatasetV2):
+ return _tf_dataset_len(s)
return _py_len(s)
@@ -278,6 +281,26 @@
raise_zero_rank_error)
+def _tf_dataset_len(s):
+ l = cardinality.cardinality(s)
+ msg = gen_string_ops.string_join([
+ 'len requires dataset with definitive cardinality, got ',
+ gen_string_ops.as_string(l)
+ ])
+ # TODO (yongtang): UNKNOWN is treated as an error.
+ # In case there are more UNKNOWN cases for dataset, we could
+ # use dataset.reduce() to find out the length (in an expensive way).
+ with ops.control_dependencies([
+ control_flow_ops.Assert(
+ math_ops.logical_and(
+ math_ops.not_equal(l, cardinality.INFINITE),
+ math_ops.not_equal(l, cardinality.UNKNOWN)), [msg])
+ ]):
+ l = array_ops.identity(l)
+
+ return l
+
+
def _py_len(s):
return len(s)
diff --git a/tensorflow/python/autograph/operators/py_builtins_test.py b/tensorflow/python/autograph/operators/py_builtins_test.py
index a98c9cc..43feb0d 100644
--- a/tensorflow/python/autograph/operators/py_builtins_test.py
+++ b/tensorflow/python/autograph/operators/py_builtins_test.py
@@ -123,6 +123,46 @@
tl = py_builtins.len_(data_structures.tf_tensor_list_new([3, 4, 5]))
self.assertEqual(self.evaluate(tl), 3)
+ def test_len_dataset(self):
+ dataset = dataset_ops.DatasetV2.from_tensor_slices([3, 2, 1])
+ self.assertEqual(self.evaluate(py_builtins.len_(dataset)), 3)
+
+ # graph mode
+ @def_function.function(autograph=False)
+ def test_fn():
+ dataset = dataset_ops.DatasetV2.from_tensor_slices([3, 2, 1])
+ return py_builtins.len_(dataset)
+
+ self.assertEqual(self.evaluate(test_fn()), 3)
+
+ def test_len_dataset_infinite(self):
+ dataset = dataset_ops.DatasetV2.range(5).repeat().batch(2)
+ with self.assertRaises(errors_impl.InvalidArgumentError):
+ _ = self.evaluate(py_builtins.len_(dataset))
+
+ # graph mode
+ @def_function.function
+ def test_fn():
+ dataset = dataset_ops.DatasetV2.range(5).repeat().batch(2)
+ return py_builtins.len_(dataset)
+
+ with self.assertRaises(errors_impl.InvalidArgumentError):
+ self.evaluate(test_fn())
+
+ def test_len_dataset_unknown(self):
+ dataset = dataset_ops.DatasetV2.range(5).filter(lambda _: True).batch(2)
+ with self.assertRaises(errors_impl.InvalidArgumentError):
+ _ = self.evaluate(py_builtins.len_(dataset))
+
+ # graph mode
+ @def_function.function(autograph=False)
+ def test_fn():
+ dataset = dataset_ops.DatasetV2.range(5).filter(lambda _: True).batch(2)
+ return py_builtins.len_(dataset)
+
+ with self.assertRaises(errors_impl.InvalidArgumentError):
+ self.evaluate(test_fn())
+
def test_len_scalar(self):
with self.assertRaises(ValueError):
py_builtins.len_(constant_op.constant(1))
diff --git a/tensorflow/python/autograph/pyct/cfg.py b/tensorflow/python/autograph/pyct/cfg.py
index cae838c..ea22392 100644
--- a/tensorflow/python/autograph/pyct/cfg.py
+++ b/tensorflow/python/autograph/pyct/cfg.py
@@ -43,10 +43,8 @@
import weakref
from enum import Enum
-# pylint:disable=g-bad-import-order
-
import gast
-# pylint:enable=g-bad-import-order
+import six
from tensorflow.python.autograph.pyct import anno
from tensorflow.python.autograph.pyct import parser
@@ -207,6 +205,18 @@
node: self.init_state(node) for node in self.graph.index.values()
}
+ def can_ignore(self, node):
+ """Returns True if the node can safely be assumed not to touch variables."""
+ ast_node = node.ast_node
+ if anno.hasanno(ast_node, anno.Basic.SKIP_PROCESSING):
+ return True
+ if six.PY2:
+ if (isinstance(ast_node, gast.Name) and
+ ast_node.id in ('None', 'True', 'False')):
+ return True
+ return isinstance(ast_node,
+ (gast.Break, gast.Continue, gast.Raise, gast.Pass))
+
def _visit_internal(self, mode):
"""Visits the CFG, depth-first."""
assert mode in (_WalkMode.FORWARD, _WalkMode.REVERSE)
@@ -679,6 +689,7 @@
def _process_exit_statement(
self, node, exits_nodes_of_type, may_exit_via_except=False):
+ self.generic_visit(node)
# Note: this is safe because we process functions separately.
try_node, guards = self._get_enclosing_finally_scopes(exits_nodes_of_type)
assert try_node is not None, '{} that is not enclosed by any of {}'.format(
@@ -727,11 +738,9 @@
# TODO(mdan): Track the CFG local to the class definition as well?
self.builder = self.builder_stack.pop()
- def visit_FunctionDef(self, node):
- # We also keep the FunctionDef node in the CFG. This allows us to determine
- # things like reaching definitions via closure. Note that the function body
- # will be stored in a separate graph, because function definitions are not
- # the same as function calls.
+ def _process_function_def(self, node, is_lambda):
+ # The function body is stored in a separate graph, because function
+ # definitions have effects very different from function calls.
if self.builder is not None:
self.builder.add_ordinary_node(node)
@@ -742,8 +751,11 @@
self.builder.enter_section(node)
self._process_basic_statement(node.args)
- for stmt in node.body:
- self.visit(stmt)
+ if is_lambda:
+ self._process_exit_statement(node.body, (gast.Lambda,))
+ else:
+ for stmt in node.body:
+ self.visit(stmt)
self.builder.exit_section(node)
self._exit_lexical_scope(node)
@@ -751,6 +763,12 @@
self.cfgs[node] = self.builder.build()
self.builder = self.builder_stack.pop()
+ def visit_FunctionDef(self, node):
+ self._process_function_def(node, is_lambda=False)
+
+ def visit_Lambda(self, node):
+ self._process_function_def(node, is_lambda=True)
+
def visit_Return(self, node):
self._process_exit_statement(node, (gast.FunctionDef,))
@@ -824,6 +842,7 @@
self.builder.enter_section(node)
+ self.generic_visit(node.test)
self.builder.enter_loop_section(node, node.test)
for stmt in node.body:
self.visit(stmt)
@@ -849,6 +868,7 @@
# Note: Strictly speaking, this should be node.target + node.iter.
# However, the activity analysis accounts for this inconsistency,
# so dataflow analysis produces the correct values.
+ self.generic_visit(node.iter)
self.builder.enter_loop_section(node, node.iter)
# Also include the "extra loop test" annotation, to capture things like the
# control variable for return and break in for loops.
diff --git a/tensorflow/python/autograph/pyct/cfg_test.py b/tensorflow/python/autograph/pyct/cfg_test.py
index d0b88c8..242a33b 100644
--- a/tensorflow/python/autograph/pyct/cfg_test.py
+++ b/tensorflow/python/autograph/pyct/cfg_test.py
@@ -18,6 +18,8 @@
from __future__ import division
from __future__ import print_function
+import gast
+
from tensorflow.python.autograph.pyct import cfg
from tensorflow.python.autograph.pyct import parser
from tensorflow.python.platform import test
@@ -1030,16 +1032,117 @@
a = lambda b: a + b
return a
- graph, = self._build_cfg(test_fn).values()
+ graphs = self._build_cfg(test_fn)
+ for k, v in graphs.items():
+ if isinstance(k, gast.Lambda):
+ lam_graph = v
+ else:
+ fn_graph = v
self.assertGraphMatches(
- graph,
+ fn_graph,
(
- ('a', 'a = (lambda b: (a + b))', 'return a'),
+ ('a', '(lambda b: (a + b))', 'a = (lambda b: (a + b))'),
+ ('(lambda b: (a + b))', 'a = (lambda b: (a + b))', 'return a'),
('a = (lambda b: (a + b))', 'return a', None),
),
)
- self.assertGraphEnds(graph, 'a', ('return a',))
+ self.assertGraphEnds(fn_graph, 'a', ('return a',))
+ self.assertGraphMatches(
+ lam_graph,
+ (
+ ('b', '(a + b)', None),
+ ),
+ )
+ self.assertGraphEnds(lam_graph, 'b', ('(a + b)',))
+
+ def test_lambda_in_return(self):
+
+ def test_fn(a):
+ return lambda b: a + b
+
+ graphs = self._build_cfg(test_fn)
+ for k, v in graphs.items():
+ if isinstance(k, gast.Lambda):
+ lam_graph = v
+ else:
+ fn_graph = v
+
+ self.assertGraphMatches(
+ fn_graph,
+ (
+ ('a', '(lambda b: (a + b))', 'return (lambda b: (a + b))'),
+ ('(lambda b: (a + b))', 'return (lambda b: (a + b))', None),
+ ),
+ )
+ self.assertGraphEnds(fn_graph, 'a', ('return (lambda b: (a + b))',))
+ self.assertGraphMatches(
+ lam_graph,
+ (
+ ('b', '(a + b)', None),
+ ),
+ )
+ self.assertGraphEnds(lam_graph, 'b', ('(a + b)',))
+
+ def test_lambda_in_while_loop_test(self):
+
+ def test_fn(a):
+ while (lambda b: a + b)(a):
+ pass
+
+ graphs = self._build_cfg(test_fn)
+ for k, v in graphs.items():
+ if isinstance(k, gast.Lambda):
+ lam_graph = v
+ else:
+ fn_graph = v
+
+ self.assertGraphMatches(
+ fn_graph,
+ (
+ ('a', '(lambda b: (a + b))', '(lambda b: (a + b))(a)'),
+ (('(lambda b: (a + b))', 'pass'), '(lambda b: (a + b))(a)', 'pass'),
+ ('(lambda b: (a + b))(a)', 'pass', '(lambda b: (a + b))(a)'),
+ ),
+ )
+ self.assertGraphEnds(fn_graph, 'a', ('(lambda b: (a + b))(a)',))
+ self.assertGraphMatches(
+ lam_graph,
+ (
+ ('b', '(a + b)', None),
+ ),
+ )
+ self.assertGraphEnds(lam_graph, 'b', ('(a + b)',))
+
+ def test_lambda_in_for_loop_test(self):
+
+ def test_fn(a):
+ for _ in (lambda b: a + b)(a):
+ pass
+
+ graphs = self._build_cfg(test_fn)
+ for k, v in graphs.items():
+ if isinstance(k, gast.Lambda):
+ lam_graph = v
+ else:
+ fn_graph = v
+
+ self.assertGraphMatches(
+ fn_graph,
+ (
+ ('a', '(lambda b: (a + b))', '(lambda b: (a + b))(a)'),
+ (('(lambda b: (a + b))', 'pass'), '(lambda b: (a + b))(a)', 'pass'),
+ ('(lambda b: (a + b))(a)', 'pass', '(lambda b: (a + b))(a)'),
+ ),
+ )
+ self.assertGraphEnds(fn_graph, 'a', ('(lambda b: (a + b))(a)',))
+ self.assertGraphMatches(
+ lam_graph,
+ (
+ ('b', '(a + b)', None),
+ ),
+ )
+ self.assertGraphEnds(lam_graph, 'b', ('(a + b)',))
def test_pass(self):
diff --git a/tensorflow/python/autograph/pyct/error_utils.py b/tensorflow/python/autograph/pyct/error_utils.py
index 3f7ace0..3e9b875 100644
--- a/tensorflow/python/autograph/pyct/error_utils.py
+++ b/tensorflow/python/autograph/pyct/error_utils.py
@@ -131,6 +131,7 @@
RuntimeError,
StopIteration,
TypeError,
+ UnboundLocalError,
ValueError,
)
diff --git a/tensorflow/python/autograph/pyct/static_analysis/activity.py b/tensorflow/python/autograph/pyct/static_analysis/activity.py
index 4e0c812..b9e398a 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/activity.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/activity.py
@@ -59,6 +59,8 @@
block.
isolated_names: Set[qual_names.QN], identifiers that are isolated to this
scope (even if the scope is not isolated).
+ annotations: Set[qual_names.QN], identifiers used as type annotations
+ in this scope.
read: Set[qual_names.QN], identifiers read in this scope.
modified: Set[qual_names.QN], identifiers modified in this scope.
deleted: Set[qual_names.QN], identifiers deleted in this scope.
@@ -109,6 +111,7 @@
self.bound = set()
self.globals = set()
+ self.annotations = set()
self.params = weakref.WeakValueDictionary()
@@ -145,6 +148,7 @@
self.read = copy.copy(other.read)
self.deleted = copy.copy(other.deleted)
self.bound = copy.copy(other.bound)
+ self.annotations = copy.copy(other.annotations)
self.params = copy.copy(other.params)
@classmethod
@@ -159,6 +163,7 @@
return new_copy
def merge_from(self, other):
+ """Adds all activity from another scope to this scope."""
assert not self.is_final
if self.parent is not None:
assert other.parent is not None
@@ -167,6 +172,7 @@
self.read.update(other.read)
self.modified.update(other.modified)
self.bound.update(other.deleted)
+ self.annotations.update(other.annotations)
self.params.update(other.params)
def finalize(self):
@@ -180,9 +186,11 @@
self.parent.modified.update(self.modified - self.isolated_names)
self.parent.bound.update(self.bound - self.isolated_names)
self.parent.globals.update(self.globals)
+ self.parent.annotations.update(self.annotations)
else:
# TODO(mdan): This is not accurate.
self.parent.read.update(self.read - self.bound)
+ self.parent.annotations.update(self.annotations - self.bound)
self.is_final = True
def __repr__(self):
@@ -223,11 +231,14 @@
def __init__(self, context, parent_scope=None):
super(ActivityAnalyzer, self).__init__(context)
+ self.allow_skips = False
self.scope = Scope(parent_scope, isolated=True)
# Note: all these flags crucially rely on the respective nodes are
# leaves in the AST, that is, they cannot contain other statements.
self._in_aug_assign = False
+ self._in_annotation = False
+ self._track_annotations_only = False
@property
def _in_constructor(self):
@@ -249,6 +260,9 @@
return False
def _track_symbol(self, node, composite_writes_alter_parent=False):
+ if self._track_annotations_only and not self._in_annotation:
+ return
+
# A QN may be missing when we have an attribute (or subscript) on a function
# call. Example: a().b
if not anno.hasanno(node, anno.Basic.QN):
@@ -282,6 +296,8 @@
elif isinstance(node.ctx, gast.Load):
self.scope.read.add(qn)
+ if self._in_annotation:
+ self.scope.annotations.add(qn)
elif isinstance(node.ctx, gast.Param):
self.scope.bound.add(qn)
@@ -320,6 +336,12 @@
self._exit_and_record_scope(node)
return node
+ def _process_annotation(self, node):
+ self._in_annotation = True
+ node = self.visit(node)
+ self._in_annotation = False
+ return node
+
def visit_Import(self, node):
return self._process_statement(node)
@@ -327,8 +349,21 @@
return self._process_statement(node)
def visit_Global(self, node):
+ self._enter_scope(False)
for name in node.names:
- self.scope.globals.add(qual_names.QN(name))
+ qn = qual_names.QN(name)
+ self.scope.read.add(qn)
+ self.scope.globals.add(qn)
+ self._exit_and_record_scope(node)
+ return node
+
+ def visit_Nonlocal(self, node):
+ self._enter_scope(False)
+ for name in node.names:
+ qn = qual_names.QN(name)
+ self.scope.read.add(qn)
+ self.scope.bound.add(qn)
+ self._exit_and_record_scope(node)
return node
def visit_Expr(self, node):
@@ -344,7 +379,13 @@
return self._process_statement(node)
def visit_AnnAssign(self, node):
- return self._process_statement(node)
+ self._enter_scope(False)
+ node.target = self.visit(node.target)
+ node.value = self.visit(node.value)
+ if node.annotation:
+ node.annotation = self._process_annotation(node.annotation)
+ self._exit_and_record_scope(node)
+ return node
def visit_AugAssign(self, node):
# Special rules for AugAssign. Here, the AST only shows the target as
@@ -364,7 +405,8 @@
return self._process_statement(node)
def visit_Name(self, node):
- node = self.generic_visit(node)
+ if node.annotation:
+ node.annotation = self._process_annotation(node.annotation)
self._track_symbol(node)
return node
@@ -474,9 +516,6 @@
def visit_GeneratorExp(self, node):
return self._process_comprehension(node)
- def visit_arguments(self, node):
- return self._process_statement(node)
-
def visit_ClassDef(self, node):
with self.state[_FunctionOrClass] as fn:
fn.node = node
@@ -496,6 +535,27 @@
self._exit_scope()
return node
+ def _visit_node_list(self, nodes):
+ return [(None if n is None else self.visit(n)) for n in nodes]
+
+ def _visit_arg_annotations(self, node):
+ node.args.kw_defaults = self._visit_node_list(node.args.kw_defaults)
+ node.args.defaults = self._visit_node_list(node.args.defaults)
+ self._track_annotations_only = True
+ node = self._visit_arg_declarations(node)
+ self._track_annotations_only = False
+ return node
+
+ def _visit_arg_declarations(self, node):
+ node.args.posonlyargs = self._visit_node_list(node.args.posonlyargs)
+ node.args.args = self._visit_node_list(node.args.args)
+ if node.args.vararg is not None:
+ node.args.vararg = self.visit(node.args.vararg)
+ node.args.kwonlyargs = self._visit_node_list(node.args.kwonlyargs)
+ if node.args.kwarg is not None:
+ node.args.kwarg = self.visit(node.args.kwarg)
+ return node
+
def visit_FunctionDef(self, node):
with self.state[_FunctionOrClass] as fn:
fn.node = node
@@ -503,6 +563,11 @@
# of its name, along with the usage of any decorator accompanying it.
self._enter_scope(False)
node.decorator_list = self.visit_block(node.decorator_list)
+ if node.returns:
+ node.returns = self._process_annotation(node.returns)
+ # Argument annotartions (includeing defaults) affect the defining context.
+ node = self._visit_arg_annotations(node)
+
function_name = qual_names.QN(node.name)
self.scope.modified.add(function_name)
self.scope.bound.add(function_name)
@@ -510,7 +575,15 @@
# A separate Scope tracks the actual function definition.
self._enter_scope(True)
- node.args = self.visit(node.args)
+
+ # Keep a separate scope for the arguments node, which is used in the CFG.
+ self._enter_scope(False)
+
+ # Arg declarations only affect the function itself, and have no effect
+ # in the defining context whatsoever.
+ node = self._visit_arg_declarations(node)
+
+ self._exit_and_record_scope(node.args)
# Track the body separately. This is for compatibility reasons, it may not
# be strictly needed.
@@ -518,16 +591,35 @@
node.body = self.visit_block(node.body)
self._exit_and_record_scope(node, NodeAnno.BODY_SCOPE)
- self._exit_scope()
+ self._exit_and_record_scope(node, NodeAnno.ARGS_AND_BODY_SCOPE)
return node
def visit_Lambda(self, node):
# Lambda nodes are treated in roughly the same way as FunctionDef nodes.
with self.state[_FunctionOrClass] as fn:
fn.node = node
- self._enter_scope(True)
- node = self.generic_visit(node)
+ # The Lambda node itself has a Scope object that tracks the creation
+ # of its name, along with the usage of any decorator accompanying it.
+ self._enter_scope(False)
+ node = self._visit_arg_annotations(node)
self._exit_and_record_scope(node)
+
+ # A separate Scope tracks the actual function definition.
+ self._enter_scope(True)
+
+ # Keep a separate scope for the arguments node, which is used in the CFG.
+ self._enter_scope(False)
+ node = self._visit_arg_declarations(node)
+ self._exit_and_record_scope(node.args)
+
+ # Track the body separately. This is for compatibility reasons, it may not
+ # be strictly needed.
+ # TODO(mdan): Do remove it, it's confusing.
+ self._enter_scope(False)
+ node.body = self.visit(node.body)
+ self._exit_and_record_scope(node, NodeAnno.BODY_SCOPE)
+
+ self._exit_and_record_scope(node, NodeAnno.ARGS_AND_BODY_SCOPE)
return node
def visit_With(self, node):
diff --git a/tensorflow/python/autograph/pyct/static_analysis/activity_py3_test.py b/tensorflow/python/autograph/pyct/static_analysis/activity_py3_test.py
index 30ce2f7..b29e08f 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/activity_py3_test.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/activity_py3_test.py
@@ -43,7 +43,10 @@
node, _ = self._parse_and_analyze(test_fn)
fn_node = node
body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('nonlocal_b', 'c'), ('nonlocal_a',))
+ self.assertScopeIs(
+ body_scope, ('nonlocal_a', 'nonlocal_b', 'c'), ('nonlocal_a',))
+ nonlocal_a_scope = anno.getanno(fn_node.body[0], anno.Static.SCOPE)
+ self.assertScopeIs(nonlocal_a_scope, ('nonlocal_a',), ())
def test_annotated_assign(self):
b = int
@@ -54,10 +57,33 @@
node, _ = self._parse_and_analyze(test_fn)
fn_node = node
+
body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
self.assertScopeIs(body_scope, ('b', 'c', 'a'), ('a',))
+ self.assertSymbolSetsAre(('b',), body_scope.annotations, 'annotations')
+
ann_assign_scope = anno.getanno(fn_node.body[0], anno.Static.SCOPE)
self.assertScopeIs(ann_assign_scope, ('b', 'c'), ('a',))
+ self.assertSymbolSetsAre(
+ ('b',), ann_assign_scope.annotations, 'annotations')
+
+ def test_function_def_annotations(self):
+ b = int
+ c = int
+
+ def test_fn(a: b) -> c:
+ return a
+
+ node, _ = self._parse_and_analyze(test_fn)
+ fn_node = node
+
+ fn_scope = anno.getanno(fn_node, anno.Static.SCOPE)
+ self.assertScopeIs(fn_scope, ('b', 'c'), ('test_fn',))
+ self.assertSymbolSetsAre(('b', 'c'), fn_scope.annotations, 'annotations')
+
+ body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(body_scope, ('a',), ())
+ self.assertSymbolSetsAre((), body_scope.annotations, 'annotations')
if __name__ == '__main__':
diff --git a/tensorflow/python/autograph/pyct/static_analysis/activity_test.py b/tensorflow/python/autograph/pyct/static_analysis/activity_test.py
index e4a93db..7a6bfd4 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/activity_test.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/activity_test.py
@@ -373,17 +373,48 @@
y = x * x
return y
- b = a
- for i in a:
- c = b
- b -= f(i)
- return b, c
+ return f(a)
+
+ node, _ = self._parse_and_analyze(test_fn)
+
+ fn_node = node
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'f'), ('f',))
+
+ fn_def_node = node.body[0]
+
+ scope = anno.getanno(fn_def_node, anno.Static.SCOPE)
+ self.assertScopeIs(scope, (), ('f'))
+
+ scope = anno.getanno(fn_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('x', 'y'), ('y',))
+
+ scope = anno.getanno(fn_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('x', 'y'), ('y',))
+ self.assertSymbolSetsAre(('x', 'y'), scope.bound, 'BOUND')
+
+ def test_nested_function_arg_defaults(self):
+
+ def test_fn(a):
+
+ def f(x=a):
+ y = x * x
+ return y
+
+ return f(a)
node, _ = self._parse_and_analyze(test_fn)
fn_def_node = node.body[0]
self.assertScopeIs(
- anno.getanno(fn_def_node, NodeAnno.BODY_SCOPE), ('x', 'y'), ('y',))
+ anno.getanno(fn_def_node, anno.Static.SCOPE), ('a',), ('f',))
+
+ scope = anno.getanno(fn_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('x', 'y'), ('y',))
+
+ scope = anno.getanno(fn_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('x', 'y'), ('y',))
+ self.assertSymbolSetsAre(('x', 'y'), scope.bound, 'BOUND')
def test_constructor_attributes(self):
@@ -482,64 +513,154 @@
self.assertScopeIs(
anno.getanno(fn_node, NodeAnno.BODY_SCOPE), ('foo', 'x'), ())
- def test_params(self):
-
- def test_fn(a, b): # pylint: disable=unused-argument
- return b
-
- node, _ = self._parse_and_analyze(test_fn)
- fn_node = node
- body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('b',), ())
- self.assertScopeIs(body_scope.parent, ('b',), ())
-
- args_scope = anno.getanno(fn_node.args, anno.Static.SCOPE)
- self.assertSymbolSetsAre(('a', 'b'), args_scope.params.keys(), 'params')
-
- def test_lambda_captures_reads(self):
+ def test_lambda(self):
def test_fn(a, b):
- return lambda: a + b
+ return lambda: (a + b)
node, _ = self._parse_and_analyze(test_fn)
- fn_node = node
- body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('a', 'b'), ())
- # Nothing local to the lambda is tracked.
- self.assertSymbolSetsAre((), body_scope.params.keys(), 'params')
- def test_lambda_params_are_isolated(self):
+ fn_node = node
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+
+ lam_def_node = node.body[0].value
+
+ scope = anno.getanno(lam_def_node, anno.Static.SCOPE)
+ self.assertScopeIs(scope, (), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+ self.assertSymbolSetsAre((), scope.bound, 'BOUND')
+
+ scope = anno.getanno(lam_def_node.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre((), scope.params.keys(), 'lambda params')
+
+ def test_lambda_params_args(self):
def test_fn(a, b): # pylint: disable=unused-argument
return lambda a: a + b
node, _ = self._parse_and_analyze(test_fn)
+
fn_node = node
- body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('b',), ())
- self.assertSymbolSetsAre((), body_scope.params.keys(), 'params')
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ # Note: `a` in `a + b` is not "read" here because it's hidden by the `a`
+ # argument.
+ self.assertScopeIs(scope, ('b',), ())
+
+ lam_def_node = node.body[0].value
+
+ scope = anno.getanno(lam_def_node, anno.Static.SCOPE)
+ self.assertScopeIs(scope, (), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+ self.assertSymbolSetsAre(('a',), scope.bound, 'BOUND')
+
+ scope = anno.getanno(lam_def_node.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre(('a',), scope.params.keys(), 'lambda params')
+
+ def test_lambda_params_arg_defaults(self):
+
+ def test_fn(a, b, c): # pylint: disable=unused-argument
+ return lambda b=c: a + b
+
+ node, _ = self._parse_and_analyze(test_fn)
+
+ fn_node = node
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ # Note: `b` is not "read" here because it's hidden by the argument.
+ self.assertScopeIs(scope, ('a', 'c'), ())
+
+ lam_def_node = node.body[0].value
+
+ scope = anno.getanno(lam_def_node, anno.Static.SCOPE)
+ self.assertScopeIs(scope, ('c',), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b'), ())
+ self.assertSymbolSetsAre(('b',), scope.bound, 'BOUND')
+
+ scope = anno.getanno(lam_def_node.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre(('b',), scope.params.keys(), 'lambda params')
def test_lambda_complex(self):
- def test_fn(a, b, c, d): # pylint: disable=unused-argument
- a = (lambda a, b, c: a + b + c)(d, 1, 2) + b
+ def test_fn(a, b, c, d, e): # pylint: disable=unused-argument
+ a = (lambda a, b, c=e: a + b + c)(d, 1, 2) + b
node, _ = self._parse_and_analyze(test_fn)
+
fn_node = node
- body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('b', 'd'), ('a',))
- self.assertSymbolSetsAre((), body_scope.params.keys(), 'params')
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('d', 'b', 'e'), ('a',))
+
+ lam_def_node = node.body[0].value.left.func
+
+ scope = anno.getanno(lam_def_node, anno.Static.SCOPE)
+ self.assertScopeIs(scope, ('e',), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b', 'c'), ())
+
+ scope = anno.getanno(lam_def_node, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b', 'c'), ())
+ self.assertSymbolSetsAre(('a', 'b', 'c'), scope.bound, 'BOUND')
+
+ scope = anno.getanno(lam_def_node.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre(
+ ('a', 'b', 'c'), scope.params.keys(), 'lambda params')
def test_lambda_nested(self):
- def test_fn(a, b, c, d, e): # pylint: disable=unused-argument
- a = lambda a, b: d(lambda b: a + b + c) # pylint: disable=undefined-variable
+ def test_fn(a, b, c, d, e, f): # pylint: disable=unused-argument
+ a = lambda a, b: d(lambda b=f: a + b + c) # pylint: disable=undefined-variable
node, _ = self._parse_and_analyze(test_fn)
+
fn_node = node
- body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('c', 'd'), ('a',))
- self.assertSymbolSetsAre((), body_scope.params.keys(), 'params')
+ scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('d', 'c', 'f'), ('a',))
+
+ outer_lam_def = node.body[0].value
+
+ scope = anno.getanno(outer_lam_def, anno.Static.SCOPE)
+ self.assertScopeIs(scope, (), ())
+
+ scope = anno.getanno(outer_lam_def, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('d', 'f', 'a', 'c'), ())
+
+ scope = anno.getanno(outer_lam_def, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('d', 'f', 'a', 'c'), ())
+ self.assertSymbolSetsAre(('a', 'b'), scope.bound, 'BOUND')
+
+ scope = anno.getanno(outer_lam_def.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre(('a', 'b'), scope.params.keys(), 'lambda params')
+
+ inner_lam_def = outer_lam_def.body.args[0]
+
+ scope = anno.getanno(inner_lam_def, anno.Static.SCOPE)
+ self.assertScopeIs(scope, ('f',), ())
+
+ scope = anno.getanno(inner_lam_def, NodeAnno.BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b', 'c'), ())
+
+ scope = anno.getanno(inner_lam_def, NodeAnno.ARGS_AND_BODY_SCOPE)
+ self.assertScopeIs(scope, ('a', 'b', 'c'), ())
+ self.assertSymbolSetsAre(('b',), scope.bound, 'BOUND')
+
+ scope = anno.getanno(inner_lam_def.args, anno.Static.SCOPE)
+ self.assertSymbolSetsAre(('b',), scope.params.keys(), 'lambda params')
def test_comprehension_targets_are_isolated(self):
@@ -607,9 +728,11 @@
node, _ = self._parse_and_analyze(test_fn)
fn_node = node
body_scope = anno.getanno(fn_node, NodeAnno.BODY_SCOPE)
- self.assertScopeIs(body_scope, ('global_b', 'c'), ('global_a',))
+ self.assertScopeIs(body_scope, ('global_a', 'global_b', 'c'), ('global_a',))
self.assertSetEqual(body_scope.globals, set(
(QN('global_a'), QN('global_b'))))
+ global_a_scope = anno.getanno(fn_node.body[0], anno.Static.SCOPE)
+ self.assertScopeIs(global_a_scope, ('global_a',), ())
def test_class_definition_basic(self):
diff --git a/tensorflow/python/autograph/pyct/static_analysis/annos.py b/tensorflow/python/autograph/pyct/static_analysis/annos.py
index cc7ad61..948684a 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/annos.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/annos.py
@@ -48,6 +48,9 @@
ARGS_SCOPE = 'The scope for the argument list of a function call.'
COND_SCOPE = 'The scope for the test node of a conditional statement.'
ITERATE_SCOPE = 'The scope for the iterate assignment of a for loop.'
+ ARGS_AND_BODY_SCOPE = (
+ 'The scope for the main body of a function or lambda, including its'
+ ' arguments.')
BODY_SCOPE = (
'The scope for the main body of a statement (True branch for if '
'statements, main body for loops).')
diff --git a/tensorflow/python/autograph/pyct/static_analysis/liveness.py b/tensorflow/python/autograph/pyct/static_analysis/liveness.py
index c237509..7d64a93 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/liveness.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/liveness.py
@@ -40,11 +40,12 @@
class Analyzer(cfg.GraphVisitor):
"""CFG visitor that performs liveness analysis at statement level."""
- def __init__(self, graph):
+ def __init__(self, graph, include_annotations):
super(Analyzer, self).__init__(graph)
# This allows communicating that nodes generate extra symbols,
# e.g. those that a function definition closes over.
self.extra_gen = {}
+ self.include_annotations = include_annotations
def init_state(self, _):
return set()
@@ -56,6 +57,8 @@
node_scope = anno.getanno(node.ast_node, anno.Static.SCOPE)
gen = node_scope.read | self.extra_gen.get(node.ast_node, frozenset())
+ if not self.include_annotations:
+ gen -= node_scope.annotations
# TODO(mdan): verify whether composites' parents need to be added.
# E.g. whether x needs to be added if x.y is live. Theoretically the
# activity analysis should have both so that wouldn't be needed.
@@ -67,12 +70,8 @@
live_in = gen | (live_out - kill)
else:
- # Nodes that don't have a scope annotation are assumed not to touch any
- # symbols.
- # This Name node below is a literal name, e.g. False
- assert isinstance(node.ast_node,
- (gast.Name, gast.Continue, gast.Break, gast.Pass,
- gast.Global, gast.Nonlocal)), type(node.ast_node)
+ assert self.can_ignore(node), (node.ast_node, node)
+
live_out = set()
for n in node.next:
live_out |= self.in_[n]
@@ -103,8 +102,10 @@
for the effect above.
"""
- def __init__(self, source_info, graphs):
+ def __init__(self, source_info, graphs, include_annotations):
super(WholeTreeAnalyzer, self).__init__(source_info)
+ self.include_annotations = include_annotations
+ self.allow_skips = False
self.graphs = graphs
self.current_analyzer = None
self.analyzers = {}
@@ -118,7 +119,7 @@
# 2. recursively walk the subtree; this will initialize the analyzer's
# in_ state properly (done in a block below)
# 3. run the final analysis
- analyzer = Analyzer(subgraph)
+ analyzer = Analyzer(subgraph, self.include_annotations)
self.current_analyzer = analyzer
node = self.generic_visit(node)
analyzer.visit_reverse()
@@ -232,17 +233,21 @@
return node
-def resolve(node, source_info, graphs):
+# TODO(mdan): Investigate the possibility of removing include_annotations.
+def resolve(node, source_info, graphs, include_annotations=True):
"""Resolves the live symbols at the exit of control flow statements.
Args:
node: ast.AST
source_info: transformer.SourceInfo
graphs: Dict[ast.FunctionDef, cfg.Graph]
+ include_annotations: Bool, whether type annotations should be included in
+ the analysis.
Returns:
ast.AST
"""
- cross_function_analyzer = WholeTreeAnalyzer(source_info, graphs)
+ cross_function_analyzer = WholeTreeAnalyzer(
+ source_info, graphs, include_annotations)
node = cross_function_analyzer.visit(node)
visitor = Annotator(source_info, cross_function_analyzer)
node = visitor.visit(node)
diff --git a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions.py b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions.py
index dda132c..008e4b8 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions.py
@@ -34,9 +34,7 @@
from tensorflow.python.autograph.pyct import anno
from tensorflow.python.autograph.pyct import cfg
-from tensorflow.python.autograph.pyct import qual_names
from tensorflow.python.autograph.pyct import transformer
-from tensorflow.python.autograph.pyct.static_analysis import annos
class Definition(object):
@@ -137,8 +135,12 @@
# their ids are used in equality checks.
if node not in self.gen_map:
node_symbols = {}
- # Every modification receives a definition.
- for s in node_scope.modified:
+ # Every binding operation (assign, nonlocal, global, etc.) counts as a
+ # definition, with the exception of del, which only deletes without
+ # creating a new variable.
+ newly_defined = ((node_scope.bound | node_scope.globals) -
+ node_scope.deleted)
+ for s in newly_defined:
def_ = self._definition_factory()
node_symbols[s] = def_
# Every param receives a definition. Params are not necessarily
@@ -153,41 +155,16 @@
kill = node_scope.modified | node_scope.deleted
defs_out = gen | (defs_in - kill)
- elif isinstance(node.ast_node, (gast.Global, gast.Nonlocal)):
- # Special case for global and nonlocal: they generate a definition,
- # but are not tracked by activity analysis.
- if node not in self.gen_map:
- node_symbols = {}
- kill = set()
- for s in node.ast_node.names:
- qn = qual_names.QN(s)
- # TODO(mdan): If definitions exist, should we preserve those instead?
- # Incoming definitions may be present when this is a local function.
- # In that case, the definitions of the nonlocal symbol from the
- # enclosing function are available here. See self.extra_in.
- kill.add(qn)
- def_ = self._definition_factory()
- node_symbols[qn] = def_
- self.gen_map[node] = _NodeState(node_symbols)
-
gen = self.gen_map[node]
defs_out = gen | (defs_in - kill)
else:
- # Nodes that don't have a scope annotation are assumed not to touch any
- # symbols.
- # This Name node below is a literal name, e.g. False
- # This can also happen if activity.py forgot to annotate the node with a
- # scope object.
- assert isinstance(node.ast_node,
- (gast.Name, gast.Break, gast.Continue, gast.Raise,
- gast.Pass)), (node.ast_node, node)
+ assert self.can_ignore(node), (node.ast_node, node)
defs_out = defs_in
self.in_[node] = defs_in
self.out[node] = defs_out
- # TODO(mdan): Move this to the superclass?
return prev_defs_out != defs_out
@@ -205,6 +182,7 @@
def __init__(self, source_info, graphs, definition_factory):
super(TreeAnnotator, self).__init__(source_info)
+ self.allow_skips = False
self.definition_factory = definition_factory
self.graphs = graphs
self.current_analyzer = None
@@ -214,28 +192,11 @@
parent_analyzer = self.current_analyzer
subgraph = self.graphs[node]
- # Preorder tree processing:
- # 1. if this is a child function, the parent was already analyzed and it
- # has the proper state value for the subgraph's entry
- # 2. analyze the current function body
- # 2. recursively walk the subtree; child functions will be processed
analyzer = Analyzer(subgraph, self.definition_factory)
- if parent_analyzer is not None:
- # Wire the state between the two subgraphs' analyzers.
- parent_out_state = parent_analyzer.out[parent_analyzer.graph.index[node]]
- # Exception: symbols modified in the child function are local to it
- body_scope = anno.getanno(node, annos.NodeAnno.BODY_SCOPE)
- parent_out_state -= body_scope.modified
- analyzer.extra_in[node.args] = parent_out_state
-
- # Complete the analysis for the local function and annotate its body.
analyzer.visit_forward()
# Recursively process any remaining subfunctions.
self.current_analyzer = analyzer
- # Note: not visiting name, decorator_list and returns because they don't
- # apply to this analysis.
- # TODO(mdan): Should we still process the function name?
node.args = self.visit(node.args)
node.body = self.visit_block(node.body)
self.current_analyzer = parent_analyzer
diff --git a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_py3_test.py b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_py3_test.py
index 8ac642b..7333ec0 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_py3_test.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_py3_test.py
@@ -78,7 +78,7 @@
self.assertSameDef(local_body[1].test, local_body[2].value.elts[0])
- self.assertHasDefinedIn(local_body[1], ('a', 'b', 'local_fn'))
+ self.assertHasDefinedIn(local_body[1], ('a', 'b'))
if __name__ == '__main__':
diff --git a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_test.py b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_test.py
index 8b00b5c..c4e7cbd 100644
--- a/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_test.py
+++ b/tensorflow/python/autograph/pyct/static_analysis/reaching_definitions_test.py
@@ -254,7 +254,8 @@
self.assertHasDefs(fn_body[2].value, 2)
inner_fn_body = fn_body[1].body[1].body
- self.assertSameDef(inner_fn_body[0].value, def_of_a_in_if)
+ def_of_a_in_foo = inner_fn_body[0].value
+ self.assertHasDefs(def_of_a_in_foo, 0)
def test_nested_functions_isolation(self):
diff --git a/tensorflow/python/autograph/pyct/transformer.py b/tensorflow/python/autograph/pyct/transformer.py
index 3370050..c8d5c9d 100644
--- a/tensorflow/python/autograph/pyct/transformer.py
+++ b/tensorflow/python/autograph/pyct/transformer.py
@@ -23,7 +23,7 @@
import gast
from tensorflow.python.autograph.pyct import anno
-from tensorflow.python.autograph.pyct import loader
+from tensorflow.python.autograph.pyct import parser
from tensorflow.python.autograph.pyct import pretty_printer
from tensorflow.python.autograph.pyct import templates
@@ -267,7 +267,7 @@
def debug_print_src(self, node):
"""Helper method useful for debugging. Prints the AST as code."""
if __debug__:
- print(loader.load_ast(node))
+ print(parser.unparse(node))
return node
def visit_block(self, nodes, before_visit=None, after_visit=None):
@@ -346,17 +346,6 @@
node_destination = new_destination
return results
- def _get_source(self, node):
- try:
- source, _ = loader.load_ast(node)
- return source
- # pylint: disable=broad-except
- # This function is used for error reporting. If an exception occurs here,
- # it should be suppressed, in favor of emitting as informative a message
- # about the original error as possible.
- except Exception:
- return '<could not convert AST to source>'
-
# TODO(mdan): Rename to PythonCodeTransformer.
class Base(NodeStateTracker, gast.NodeTransformer):
diff --git a/tensorflow/python/client/timeline.py b/tensorflow/python/client/timeline.py
index c3f3829..db9e1a0 100644
--- a/tensorflow/python/client/timeline.py
+++ b/tensorflow/python/client/timeline.py
@@ -359,7 +359,8 @@
graph: (Optional) The 'Graph' that was executed.
"""
- self._step_stats = step_stats
+ self._origin_step_stats = step_stats
+ self._step_stats = None
self._graph = graph
self._chrome_trace = _ChromeTraceFormatter()
self._next_pid = 0
@@ -396,6 +397,17 @@
inputs = inputs.split(', ')
return nn, op, inputs
+ def _parse_kernel_label(self, label, node_name):
+ """Parses the fields in a node timeline label."""
+ # Expects labels of the form: retval (arg) detail @@annotation
+ match = re.match(r'.*@@(.*)\#id.*', label)
+ if match is not None:
+ node_name = match.group(1)
+ # Node names should always have the form 'name:op'.
+ fields = node_name.split(':') + ['unknown']
+ name, op = fields[:2]
+ return name, op
+
def _assign_lanes(self):
"""Assigns non-overlapping lanes for the activities on each device."""
for device_stats in self._step_stats.dev_stats:
@@ -427,9 +439,8 @@
tid = nodestats.thread_id
inputs = []
if is_gputrace:
- # Node names should always have the form 'name:op'.
- fields = node_name.split(':') + ['unknown']
- node_name, op = fields[:2]
+ node_name, op = self._parse_kernel_label(nodestats.timeline_label,
+ node_name)
elif node_name == 'RecvTensor':
# RPC tracing does not use the standard timeline_label format.
op = 'RecvTensor'
@@ -607,7 +618,81 @@
total_bytes)
self._allocator_maximums = alloc_maxes
- def analyze_step_stats(self, show_dataflow=True, show_memory=True):
+ def _preprocess_op_time(self, op_time):
+ """Update the start and end time of ops in step stats.
+
+ Args:
+ op_time: How the execution time of op is shown in timeline. Possible values
+ are "schedule", "gpu" and "all". "schedule" will show op from the time it
+ is scheduled to the end of the scheduling. Notice by the end of its
+ scheduling its async kernels may not start yet. It is shown using the
+ default value from step_stats. "gpu" will show op with the execution time
+ of its kernels on GPU. "all" will show op from the start of its scheduling
+ to the end of its last kernel.
+ """
+ if op_time == 'schedule':
+ self._step_stats = self._origin_step_stats
+ return
+ self._step_stats = copy.deepcopy(self._origin_step_stats)
+ # Separate job task and gpu tracer stream
+ stream_all_stats = []
+ job_stats = []
+ for stats in self._step_stats.dev_stats:
+ if '/stream:all' in stats.device:
+ stream_all_stats.append(stats)
+ elif '/job' in stats.device:
+ job_stats.append(stats)
+
+ # Record the start time of the first kernel and the end time of
+ # the last gpu kernel for all ops.
+ op_gpu_start = {}
+ op_gpu_end = {}
+ for stats in stream_all_stats:
+ for kernel in stats.node_stats:
+ name, _ = self._parse_kernel_label(kernel.timeline_label,
+ kernel.node_name)
+ start = kernel.all_start_micros
+ end = kernel.all_start_micros + kernel.all_end_rel_micros
+ if name in op_gpu_start:
+ op_gpu_start[name] = min(op_gpu_start[name], start)
+ op_gpu_end[name] = max(op_gpu_end[name], end)
+ else:
+ op_gpu_start[name] = start
+ op_gpu_end[name] = end
+
+ # Update the start and end time of each op according to the op_time
+ for stats in job_stats:
+ for op in stats.node_stats:
+ if op.node_name in op_gpu_start:
+ end = max(op_gpu_end[op.node_name],
+ op.all_start_micros + op.all_end_rel_micros)
+ if op_time == 'gpu':
+ op.all_start_micros = op_gpu_start[op.node_name]
+ op.all_end_rel_micros = end - op.all_start_micros
+
+ def analyze_step_stats(self,
+ show_dataflow=True,
+ show_memory=True,
+ op_time='schedule'):
+ """Analyze the step stats and format it into Chrome Trace Format.
+
+ Args:
+ show_dataflow: (Optional.) If True, add flow events to the trace
+ connecting producers and consumers of tensors.
+ show_memory: (Optional.) If True, add object snapshot events to the trace
+ showing the sizes and lifetimes of tensors.
+ op_time: (Optional.) How the execution time of op is shown in timeline.
+ Possible values are "schedule", "gpu" and "all". "schedule" will show op
+ from the time it is scheduled to the end of the scheduling. Notice by
+ the end of its scheduling its async kernels may not start yet. It is
+ shown using the default value from step_stats. "gpu" will show op with
+ the execution time of its kernels on GPU. "all" will show op from the
+ start of its scheduling to the end of its last kernel.
+
+ Returns:
+ A 'StepStatsAnalysis' object.
+ """
+ self._preprocess_op_time(op_time)
self._allocate_pids()
self._assign_lanes()
self._analyze_tensors(show_memory)
@@ -618,7 +703,10 @@
chrome_trace=self._chrome_trace,
allocator_maximums=self._allocator_maximums)
- def generate_chrome_trace_format(self, show_dataflow=True, show_memory=False):
+ def generate_chrome_trace_format(self,
+ show_dataflow=True,
+ show_memory=False,
+ op_time='schedule'):
"""Produces a trace in Chrome Trace Format.
Args:
@@ -626,11 +714,20 @@
connecting producers and consumers of tensors.
show_memory: (Optional.) If True, add object snapshot events to the trace
showing the sizes and lifetimes of tensors.
+ op_time: (Optional.) How the execution time of op is shown in timeline.
+ Possible values are "schedule", "gpu" and "all".
+ "schedule" will show op from the time it is scheduled to the end of
+ the scheduling.
+ Notice by the end of its scheduling its async kernels may not start
+ yet. It is shown using the default value from step_stats.
+ "gpu" will show op with the execution time of its kernels on GPU.
+ "all" will show op from the start of its scheduling to the end of
+ its last kernel.
Returns:
A JSON formatted string in Chrome Trace format.
"""
step_stats_analysis = self.analyze_step_stats(
- show_dataflow=show_dataflow, show_memory=show_memory)
+ show_dataflow=show_dataflow, show_memory=show_memory, op_time=op_time)
return step_stats_analysis.chrome_trace.format_to_string(pretty=True)
diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
index 82fc781..1cbf092 100644
--- a/tensorflow/python/compat/compat.py
+++ b/tensorflow/python/compat/compat.py
@@ -31,7 +31,7 @@
# This value changes every day with an automatic CL. It can be modified in code
# via `forward_compatibility_horizon()` or with the environment variable
# TF_FORWARD_COMPATIBILITY_DELTA_DAYS, which is added to the compatibility date.
-_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 4, 20)
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2020, 4, 23)
_FORWARD_COMPATIBILITY_DELTA_DAYS_VAR_NAME = "TF_FORWARD_COMPATIBILITY_DELTA_DAYS"
_FORWARD_COMPATIBILITY_DATE_NUMBER = None
diff --git a/tensorflow/python/data/experimental/ops/data_service_ops.py b/tensorflow/python/data/experimental/ops/data_service_ops.py
index 744ee78..f8e9ac1 100644
--- a/tensorflow/python/data/experimental/ops/data_service_ops.py
+++ b/tensorflow/python/data/experimental/ops/data_service_ops.py
@@ -52,7 +52,8 @@
dataset_id,
address,
protocol,
- max_outstanding_requests=None):
+ max_outstanding_requests=None,
+ task_refresh_interval_hint_ms=None):
"""Constructs a _DataServiceDatasetV2.
Args:
@@ -66,21 +67,27 @@
requested at the same time. You can use this option to control the
amount of memory used, since `distribute` won't use more than
`element_size` * `max_outstanding_requests` of memory.
+ task_refresh_interval_hint_ms: (Optional.) A hint for how often to query
+ the master for task changes.
"""
if max_outstanding_requests is None:
max_outstanding_requests = dataset_ops.AUTOTUNE
+ if task_refresh_interval_hint_ms is None:
+ task_refresh_interval_hint_ms = dataset_ops.AUTOTUNE
self._element_spec = input_dataset.element_spec
self._dataset_id = dataset_id
self._address = address
self._protocol = protocol
self._max_outstanding_requests = max_outstanding_requests
+ self._task_refresh_interval_hint_ms = task_refresh_interval_hint_ms
variant_tensor = gen_experimental_dataset_ops.data_service_dataset(
address=address,
protocol=protocol,
max_outstanding_requests=max_outstanding_requests,
+ task_refresh_interval_hint_ms=task_refresh_interval_hint_ms,
**self._flat_structure)
super(_DataServiceDatasetV2, self).__init__(variant_tensor)
@@ -106,14 +113,15 @@
@functools.wraps(_DataServiceDatasetV2.__init__)
def __init__(self, input_dataset, dataset_id, address, protocol,
- max_outstanding_requests):
+ max_outstanding_requests, task_refresh_interval_hint_ms):
self._wrapped = _DataServiceDatasetV2(
input_dataset=input_dataset,
dataset_id=dataset_id,
address=address,
protocol=protocol,
- max_outstanding_requests=max_outstanding_requests)
+ max_outstanding_requests=max_outstanding_requests,
+ task_refresh_interval_hint_ms=task_refresh_interval_hint_ms)
super(_DataServiceDatasetV1, self).__init__(self._wrapped)
@property
@@ -135,29 +143,13 @@
_DataServiceDataset = _DataServiceDatasetV1
-def distribute(service, max_outstanding_requests=None):
+def _distribute(service,
+ max_outstanding_requests=None,
+ task_refresh_interval_hint_ms=None):
"""A transformation that moves dataset processing to the tf.data service.
- ```
- dataset = tf.data.Dataset.range(10)
- dataset = dataset.map(lambda x: x*x)
- dataset = dataset.apply(
- tf.data.experimental.service.distribute("grpc://dataservice:5000"))
- dataset = dataset.map(lambda x: x+10)
-
- job_token = tf.data.experimental.service.create_job(dataset)
- it = tf.data.experimental.service.create_iterator(dataset, job_token)
- for element in it:
- # process element
- ```
-
- In the above example, the first two lines (before the call to `distribute`)
- will be executed on tf.data workers, and the elements provided over
- RPC. The remaining transformations (after the call to `distribute`) will be
- executed locally.
-
- The token returned from `create_job` may be used to create multiple
- coordinated iterators which consume data from the same job.
+ This transformation is similar to `distribute`, but supports additional
+ parameters which we do not yet want to add to the public Python API.
Args:
service: A string indicating how to connect to the tf.data service. The
@@ -167,6 +159,8 @@
requested at the same time. You can use this option to control the amount
of memory used, since `distribute` won't use more than `element_size` *
`max_outstanding_requests` of memory.
+ task_refresh_interval_hint_ms: (Optional.) A hint for how often to query the
+ master for task changes.
Returns:
Dataset: A `Dataset` of the elements produced by the data service.
@@ -205,11 +199,51 @@
dataset_id=dataset_id,
address=address,
protocol=protocol,
- max_outstanding_requests=max_outstanding_requests)
+ max_outstanding_requests=max_outstanding_requests,
+ task_refresh_interval_hint_ms=task_refresh_interval_hint_ms)
return _apply_fn
+def distribute(service, max_outstanding_requests=None):
+ """A transformation that moves dataset processing to the tf.data service.
+
+ ```
+ dataset = tf.data.Dataset.range(10)
+ dataset = dataset.map(lambda x: x*x)
+ dataset = dataset.apply(
+ tf.data.experimental.service.distribute("grpc://dataservice:5000"))
+ dataset = dataset.map(lambda x: x+10)
+
+ job_token = tf.data.experimental.service.create_job(dataset)
+ it = tf.data.experimental.service.create_iterator(dataset, job_token)
+ for element in it:
+ # process element
+ ```
+
+ In the above example, the first two lines (before the call to `distribute`)
+ will be executed on tf.data workers, and the elements provided over
+ RPC. The remaining transformations (after the call to `distribute`) will be
+ executed locally.
+
+ The token returned from `create_job` may be used to create multiple
+ coordinated iterators which consume data from the same job.
+
+ Args:
+ service: A string indicating how to connect to the tf.data service. The
+ string should be in the format <protocol>://<address>, e.g.
+ grpc://localhost:5000.
+ max_outstanding_requests: (Optional.) A limit on how many elements may be
+ requested at the same time. You can use this option to control the amount
+ of memory used, since `distribute` won't use more than `element_size` *
+ `max_outstanding_requests` of memory.
+
+ Returns:
+ Dataset: A `Dataset` of the elements produced by the data service.
+ """
+ return _distribute(service, max_outstanding_requests)
+
+
def create_job(dataset, processing_mode):
"""Creates a job for reading a dataset through the tf.data service.
diff --git a/tensorflow/python/data/experimental/ops/distribute_options.py b/tensorflow/python/data/experimental/ops/distribute_options.py
index 9a4c42d..40c9b2e 100644
--- a/tensorflow/python/data/experimental/ops/distribute_options.py
+++ b/tensorflow/python/data/experimental/ops/distribute_options.py
@@ -77,14 +77,6 @@
"files to shard.",
default_factory=lambda: AutoShardPolicy.AUTO)
- _make_stateless = options.create_option(
- name="_make_stateless",
- ty=bool,
- docstring=
- "Determines whether the input pipeline should be rewritten to not "
- "contain stateful transformations (so that its graph can be moved "
- "between devices).")
-
num_devices = options.create_option(
name="num_devices",
ty=int,
diff --git a/tensorflow/python/data/kernel_tests/BUILD b/tensorflow/python/data/kernel_tests/BUILD
index 673b366..2e01021 100644
--- a/tensorflow/python/data/kernel_tests/BUILD
+++ b/tensorflow/python/data/kernel_tests/BUILD
@@ -81,7 +81,7 @@
tf_py_test(
name = "data_service_ops_test",
- size = "small",
+ size = "medium",
srcs = ["data_service_ops_test.py"],
deps = [
"//tensorflow:tensorflow_py",
@@ -91,6 +91,7 @@
"//tensorflow/python:framework_test_lib",
"//tensorflow/python:platform_test",
"//tensorflow/python/data",
+ "//tensorflow/python/data/experimental/ops:testing",
"//tensorflow/python/data/kernel_tests:test_base",
"//tensorflow/python/data/service:server_lib",
],
@@ -117,6 +118,21 @@
)
tf_py_test(
+ name = "dataset_spec_test",
+ size = "small",
+ srcs = ["dataset_spec_test.py"],
+ deps = [
+ ":test_base",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python:dtypes",
+ "//tensorflow/python:errors",
+ "//tensorflow/python/data/ops:dataset_ops",
+ "//third_party/py/numpy",
+ "@absl_py//absl/testing:parameterized",
+ ],
+)
+
+tf_py_test(
name = "enumerate_test",
size = "small",
srcs = ["enumerate_test.py"],
diff --git a/tensorflow/python/data/kernel_tests/cache_test.py b/tensorflow/python/data/kernel_tests/cache_test.py
index 00068a7..a95424b 100644
--- a/tensorflow/python/data/kernel_tests/cache_test.py
+++ b/tensorflow/python/data/kernel_tests/cache_test.py
@@ -244,8 +244,6 @@
dataset_ops.Dataset.from_tensor_slices(components).repeat(0))
cache_dataset = repeat_dataset.cache()
- # Create initialization ops for iterators without and with
- # caching, respectively.
self.assertDatasetProduces(cache_dataset, expected_output=[])
@combinations.generate(test_base.default_test_combinations())
diff --git a/tensorflow/python/data/kernel_tests/data_service_ops_test.py b/tensorflow/python/data/kernel_tests/data_service_ops_test.py
index 4faec6c..55fad6f 100644
--- a/tensorflow/python/data/kernel_tests/data_service_ops_test.py
+++ b/tensorflow/python/data/kernel_tests/data_service_ops_test.py
@@ -17,6 +17,8 @@
from __future__ import division
from __future__ import print_function
+import time
+
from absl.testing import parameterized
from tensorflow.python.data.experimental.ops import data_service_ops
@@ -35,6 +37,12 @@
PROTOCOL = "grpc"
+def _make_distributed_dataset(dataset, service):
+ """Creates a distributed dataset with a short task refresh interval."""
+ return dataset.apply(
+ data_service_ops._distribute(service, task_refresh_interval_hint_ms=20))
+
+
class DataServiceOpsTest(test_base.DatasetTestBase, parameterized.TestCase):
def create_cluster(self, num_workers):
@@ -61,7 +69,7 @@
def testMultipleEpochs(self):
service = self.create_cluster(1)
ds = dataset_ops.Dataset.range(3)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
for _ in range(10):
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
it = data_service_ops.create_iterator(ds, token)
@@ -72,7 +80,7 @@
num_elements = 10
service = self.create_cluster(1)
ds = dataset_ops.Dataset.range(num_elements)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
it = data_service_ops.create_iterator(ds, token)
results = [t.numpy() for t in it]
@@ -87,7 +95,7 @@
results = []
for _ in range(num_datasets):
ds = dataset_ops.Dataset.range(num_elements)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
it = data_service_ops.create_iterator(ds, token)
iterators.append(it)
@@ -106,7 +114,7 @@
num_iterators = 3
service = self.create_cluster(1)
ds = dataset_ops.Dataset.range(num_elements)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
result = []
iterators = []
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
@@ -131,13 +139,77 @@
num_elements = 10
service = self.create_cluster(num_workers)
ds = dataset_ops.Dataset.range(num_elements)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
iterator = data_service_ops.create_iterator(ds, token)
results = [elem.numpy() for elem in iterator]
self.assertCountEqual(num_workers * list(range(num_elements)), results)
@combinations.generate(test_base.eager_only_combinations())
+ def testAddWorkerMidJob(self):
+ self._master = server_lib.MasterServer(PROTOCOL)
+ master_address = self._master.target[len(PROTOCOL + "://"):]
+ self._worker = server_lib.WorkerServer(
+ PROTOCOL, master_address=master_address)
+ num_elements = 100
+ ds = dataset_ops.Dataset.range(num_elements)
+ ds = _make_distributed_dataset(ds, self._master.target)
+ token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
+ iterator = data_service_ops.create_iterator(ds, token)
+ results = []
+ # Read halfway through the dataset.
+ for _ in range(num_elements // 2):
+ results.append(next(iterator).numpy())
+
+ self._new_worker = server_lib.WorkerServer(
+ PROTOCOL, master_address=master_address)
+
+ # Wait for the new worker to register with the master.
+ while self._master.num_tasks() < 2:
+ time.sleep(10 / 1000) # 10ms
+
+ for elem in iterator:
+ results.append(elem.numpy())
+
+ self.assertCountEqual(2 * list(range(num_elements)), results)
+
+ @combinations.generate(
+ combinations.times(test_base.eager_only_combinations(),
+ combinations.combine(use_same_port=[True, False])))
+ def testRestartWorker(self, use_same_port):
+ self._master = server_lib.MasterServer(PROTOCOL)
+ master_address = self._master.target[len(PROTOCOL + "://"):]
+ self._worker = server_lib.WorkerServer(
+ PROTOCOL, master_address=master_address)
+ num_elements = 100
+ ds = dataset_ops.Dataset.range(num_elements)
+ ds = _make_distributed_dataset(ds, self._master.target)
+ token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
+ iterator = data_service_ops.create_iterator(ds, token)
+ # Read halfway through the dataset.
+ midpoint = num_elements // 2
+ for i in range(midpoint):
+ self.assertEqual(i, next(iterator).numpy())
+
+ # Stop the original worker and start a new one.
+ port = 0
+ if use_same_port:
+ worker_address = self._worker.target[len(PROTOCOL + "://"):]
+ port = int(worker_address.split(":")[1])
+ self._worker.stop()
+ self._new_worker = server_lib.WorkerServer(
+ PROTOCOL, master_address=master_address, port=port)
+
+ # The dataset starts over now that we read from the new worker.
+ for i in range(num_elements):
+ val = next(iterator).numpy()
+ if val == midpoint and i != midpoint:
+ # There may have been one last element prefetched from the first worker
+ # before it was stopped.
+ val = next(iterator).numpy()
+ self.assertEqual(i, val)
+
+ @combinations.generate(test_base.eager_only_combinations())
def testInsideFunction(self):
num_workers = 3
num_elements = 10
@@ -146,7 +218,7 @@
@def_function.function
def f():
ds = dataset_ops.Dataset.range(num_elements)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
it = data_service_ops.create_iterator(ds, token)
result = tensor_array_ops.TensorArray(
@@ -170,7 +242,7 @@
ds = ds.with_options(options)
service = self.create_cluster(3)
- ds = ds.apply(data_service_ops.distribute(service))
+ ds = _make_distributed_dataset(ds, service)
token = data_service_ops.create_job(ds, processing_mode="parallel_epochs")
iterator = data_service_ops.create_iterator(ds, token)
next(iterator)
@@ -202,9 +274,9 @@
def testMultipleDistributeCalls(self):
service = self.create_cluster(1)
ds1 = dataset_ops.Dataset.range(1)
- ds1 = ds1.apply(data_service_ops.distribute(service))
+ ds1 = _make_distributed_dataset(ds1, service)
ds2 = dataset_ops.Dataset.range(1)
- ds2 = ds2.apply(data_service_ops.distribute(service))
+ ds2 = _make_distributed_dataset(ds2, service)
ds = dataset_ops.Dataset.zip((ds1, ds2))
with self.assertRaisesWithLiteralMatch(
ValueError, "Datasets containing multiple calls to .distribute(...) "
@@ -218,7 +290,7 @@
def interleave_fn(_):
ds = dataset_ops.Dataset.range(2)
- ds = ds.apply(data_service_ops.distribute(service))
+ _make_distributed_dataset(ds, service)
return ds
with self.assertRaisesRegex(
diff --git a/tensorflow/python/data/kernel_tests/dataset_spec_test.py b/tensorflow/python/data/kernel_tests/dataset_spec_test.py
new file mode 100644
index 0000000..781a972
--- /dev/null
+++ b/tensorflow/python/data/kernel_tests/dataset_spec_test.py
@@ -0,0 +1,54 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for `tf.data.DatasetSpec`."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from absl.testing import parameterized
+import numpy as np
+
+from tensorflow.python.data.kernel_tests import test_base
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.eager import def_function
+from tensorflow.python.framework import combinations
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import tensor_shape
+from tensorflow.python.framework import tensor_spec
+from tensorflow.python.platform import test
+
+
+class DatasetSpecTest(test_base.DatasetTestBase, parameterized.TestCase):
+
+ @combinations.generate(test_base.default_test_combinations())
+ def testInputSignature(self):
+ dataset = dataset_ops.Dataset.from_tensor_slices(
+ np.arange(10).astype(np.int32)).batch(5)
+
+ @def_function.function(input_signature=[
+ dataset_ops.DatasetSpec(
+ tensor_spec.TensorSpec(
+ shape=(None,), dtype=dtypes.int32, name=None),
+ tensor_shape.TensorShape([]))
+ ])
+ def fn(_):
+ pass
+
+ fn(dataset)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/data/kernel_tests/memory_cleanup_test.py b/tensorflow/python/data/kernel_tests/memory_cleanup_test.py
index 8ba9d4c..583a2d4 100644
--- a/tensorflow/python/data/kernel_tests/memory_cleanup_test.py
+++ b/tensorflow/python/data/kernel_tests/memory_cleanup_test.py
@@ -32,9 +32,9 @@
from tensorflow.python.framework import combinations
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
-from tensorflow.python.framework import tensor_like
from tensorflow.python.platform import test
from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.types import internal
# memory_profiler might not be available in the OSS version of TensorFlow.
@@ -116,7 +116,7 @@
gc.collect()
tensors = [
- o for o in gc.get_objects() if isinstance(o, tensor_like.TensorLike)
+ o for o in gc.get_objects() if isinstance(o, internal.NativeObject)
]
self.assertEmpty(tensors, "%d Tensors are still alive." % len(tensors))
diff --git a/tensorflow/python/data/kernel_tests/shuffle_test.py b/tensorflow/python/data/kernel_tests/shuffle_test.py
index 81a9786..eaa4afb 100644
--- a/tensorflow/python/data/kernel_tests/shuffle_test.py
+++ b/tensorflow/python/data/kernel_tests/shuffle_test.py
@@ -23,6 +23,8 @@
from absl.testing import parameterized
import numpy as np
+from tensorflow.python import tf2
+from tensorflow.python.compat import compat
from tensorflow.python.data.kernel_tests import test_base
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.eager import function
@@ -334,9 +336,12 @@
@combinations.generate(
combinations.times(
test_base.graph_only_combinations() +
- combinations.combine(mode=["eager"], tf_api_version=1),
+ combinations.combine(mode=["eager"]),
combinations.combine(reshuffle=[True, False])))
def testRerandomizeOnReplicate(self, reshuffle):
+ if tf2.enabled() and not compat.forward_compatible(2020, 5, 22):
+ self.skipTest("Functionality currently not supported.")
+
random_seed.set_random_seed(None)
# When no seeds are fixed, each instantiation of the shuffle dataset should
# produce elements in a different order.
diff --git a/tensorflow/python/data/ops/dataset_ops.py b/tensorflow/python/data/ops/dataset_ops.py
index 2eddeb6..932ad03 100644
--- a/tensorflow/python/data/ops/dataset_ops.py
+++ b/tensorflow/python/data/ops/dataset_ops.py
@@ -2824,9 +2824,6 @@
result.append("latency_all_edges")
if self.experimental_slack:
result.append("slack")
- if (self.experimental_distribute and
- self.experimental_distribute._make_stateless): # pylint: disable=protected-access
- result.append("make_stateless")
return result
def _graph_rewrite_configs(self):
@@ -3050,7 +3047,7 @@
@property
def value_type(self):
- return _VariantDataset
+ return Dataset
def _serialize(self):
return (self._element_spec, self._dataset_shape)
@@ -3526,6 +3523,8 @@
return self._structure
+# This can be deleted after the forward compatibility window for switching
+# to using dummy resource expires on 5/20.
class _MemoryCacheDeleter(object):
"""An object which cleans up an anonymous memory cache resource.
@@ -3552,15 +3551,20 @@
handle=self._handle, deleter=self._deleter)
+# This can be deleted after the forward compatibility window for switching
+# to using dummy resource expires on 5/20.
class _MemoryCache(object):
"""Represents a memory cache resource."""
def __init__(self):
super(_MemoryCache, self).__init__()
- self._device = context.context().device_name
- self._handle, self._deleter = (gen_dataset_ops.anonymous_memory_cache())
- self._resource_deleter = _MemoryCacheDeleter(
- handle=self._handle, device=self._device, deleter=self._deleter)
+ if compat.forward_compatible(2020, 5, 20):
+ self._handle = gen_dataset_ops.dummy_memory_cache()
+ else:
+ self._device = context.context().device_name
+ self._handle, self._deleter = gen_dataset_ops.anonymous_memory_cache()
+ self._resource_deleter = _MemoryCacheDeleter(
+ handle=self._handle, device=self._device, deleter=self._deleter)
@property
def handle(self):
@@ -3590,6 +3594,8 @@
super(CacheDataset, self).__init__(input_dataset, variant_tensor)
+# This can be deleted after the forward compatibility window for switching
+# to using dummy resource expires on 5/22.
class _SeedGeneratorDeleter(object):
"""An object which cleans up an anonymous seed generator resource.
@@ -3616,63 +3622,22 @@
handle=self._handle, deleter=self._deleter)
+# This can be deleted after the forward compatibility window for switching
+# to using dummy resource expires on 5/22.
class _SeedGenerator(object):
"""Represents a fixed seed generator resource."""
def __init__(self, seed, seed2, reshuffle):
super(_SeedGenerator, self).__init__()
- self._device = context.context().device_name
- self._handle, self._deleter = (
- gen_dataset_ops.anonymous_seed_generator(
- seed=seed, seed2=seed2, reshuffle=reshuffle))
- self._resource_deleter = _SeedGeneratorDeleter(
- handle=self._handle, device=self._device, deleter=self._deleter)
-
- @property
- def handle(self):
- return self._handle
-
-
-# TODO(b/151115950): Remove this class after forward compatibility window
-# expires
-class _RandomSeedGeneratorDeleter(object):
- """An object which cleans up an anonymous random seed generator resource.
-
- An alternative to defining a __del__ method on an object. Even if the parent
- object is part of a reference cycle, the cycle will be collectable.
- """
-
- def __init__(self, handle, device, deleter):
- self._deleter = deleter
- self._handle = handle
- self._device = device
- self._eager_mode = context.executing_eagerly()
-
- def __del__(self):
- with ops.device(self._device):
- # Make sure the resource is deleted in the same mode as it was created in.
- if self._eager_mode:
- with context.eager_mode():
- gen_dataset_ops.delete_random_seed_generator(
- handle=self._handle, deleter=self._deleter)
- else:
- with context.graph_mode():
- gen_dataset_ops.delete_random_seed_generator(
- handle=self._handle, deleter=self._deleter)
-
-
-# TODO(b/151115950): Remove this class after forward compatibility window
-# expires
-class _RandomSeedGenerator(object):
- """Represents a random seed generator resource."""
-
- def __init__(self, seed, seed2):
- super(_RandomSeedGenerator, self).__init__()
- self._device = context.context().device_name
- self._handle, self._deleter = (
- gen_dataset_ops.anonymous_random_seed_generator(seed=seed, seed2=seed2))
- self._resource_deleter = _RandomSeedGeneratorDeleter(
- handle=self._handle, device=self._device, deleter=self._deleter)
+ if compat.forward_compatible(2020, 5, 22):
+ self._handle = gen_dataset_ops.dummy_seed_generator()
+ else:
+ self._device = context.context().device_name
+ self._handle, self._deleter = (
+ gen_dataset_ops.anonymous_seed_generator(
+ seed=seed, seed2=seed2, reshuffle=reshuffle))
+ self._resource_deleter = _SeedGeneratorDeleter(
+ handle=self._handle, device=self._device, deleter=self._deleter)
@property
def handle(self):
@@ -3710,25 +3675,29 @@
self._buffer_size = ops.convert_to_tensor(
buffer_size, dtype=dtypes.int64, name="buffer_size")
self._seed, self._seed2 = random_seed.get_seed(seed)
-
if reshuffle_each_iteration is None:
- self._reshuffle_each_iteration = True
- else:
- self._reshuffle_each_iteration = reshuffle_each_iteration
+ reshuffle_each_iteration = True
+ self._reshuffle_each_iteration = reshuffle_each_iteration
- if (tf2.enabled() and (self._reshuffle_each_iteration or
- compat.forward_compatible(2020, 4, 10)) and
+ if (tf2.enabled() and
(context.executing_eagerly() or ops.inside_function())):
- if compat.forward_compatible(2020, 4, 10):
- self._seed_generator = _SeedGenerator(self._seed, self._seed2,
- self._reshuffle_each_iteration)
+ self._seed_generator = _SeedGenerator(self._seed, self._seed2,
+ self._reshuffle_each_iteration)
+ if compat.forward_compatible(2020, 5, 22):
+ variant_tensor = gen_dataset_ops.shuffle_dataset_v3(
+ input_dataset._variant_tensor, # pylint: disable=protected-access
+ buffer_size=self._buffer_size,
+ seed=self._seed,
+ seed2=self._seed2,
+ seed_generator=self._seed_generator.handle,
+ reshuffle_each_iteration=self._reshuffle_each_iteration,
+ **self._flat_structure)
else:
- self._seed_generator = _RandomSeedGenerator(self._seed, self._seed2)
- variant_tensor = gen_dataset_ops.shuffle_dataset_v2(
- input_dataset._variant_tensor, # pylint: disable=protected-access
- buffer_size=self._buffer_size,
- seed_generator=self._seed_generator.handle,
- **self._flat_structure)
+ variant_tensor = gen_dataset_ops.shuffle_dataset_v2(
+ input_dataset._variant_tensor, # pylint: disable=protected-access
+ buffer_size=self._buffer_size,
+ seed_generator=self._seed_generator.handle,
+ **self._flat_structure)
else:
variant_tensor = gen_dataset_ops.shuffle_dataset(
input_dataset._variant_tensor, # pylint: disable=protected-access
@@ -4138,29 +4107,17 @@
else:
self._deterministic = "false"
self._preserve_cardinality = preserve_cardinality
- if deterministic is not None or compat.forward_compatible(2020, 3, 6):
- self._num_parallel_calls = ops.convert_to_tensor(
- num_parallel_calls, dtype=dtypes.int64, name="num_parallel_calls")
- variant_tensor = gen_dataset_ops.parallel_map_dataset_v2(
- input_dataset._variant_tensor, # pylint: disable=protected-access
- self._map_func.function.captured_inputs,
- f=self._map_func.function,
- num_parallel_calls=self._num_parallel_calls,
- deterministic=self._deterministic,
- use_inter_op_parallelism=self._use_inter_op_parallelism,
- preserve_cardinality=self._preserve_cardinality,
- **self._flat_structure)
- else:
- self._num_parallel_calls = ops.convert_to_tensor(
- num_parallel_calls, dtype=dtypes.int32, name="num_parallel_calls")
- variant_tensor = gen_dataset_ops.parallel_map_dataset(
- input_dataset._variant_tensor, # pylint: disable=protected-access
- self._map_func.function.captured_inputs,
- f=self._map_func.function,
- num_parallel_calls=self._num_parallel_calls,
- use_inter_op_parallelism=self._use_inter_op_parallelism,
- preserve_cardinality=self._preserve_cardinality,
- **self._flat_structure)
+ self._num_parallel_calls = ops.convert_to_tensor(
+ num_parallel_calls, dtype=dtypes.int64, name="num_parallel_calls")
+ variant_tensor = gen_dataset_ops.parallel_map_dataset_v2(
+ input_dataset._variant_tensor, # pylint: disable=protected-access
+ self._map_func.function.captured_inputs,
+ f=self._map_func.function,
+ num_parallel_calls=self._num_parallel_calls,
+ deterministic=self._deterministic,
+ use_inter_op_parallelism=self._use_inter_op_parallelism,
+ preserve_cardinality=self._preserve_cardinality,
+ **self._flat_structure)
super(ParallelMapDataset, self).__init__(input_dataset, variant_tensor)
def _functions(self):
@@ -4287,30 +4244,17 @@
else:
deterministic_string = "false"
- if (buffer_output_elements != AUTOTUNE or
- prefetch_input_elements != AUTOTUNE or
- compat.forward_compatible(2020, 3, 6)):
- variant_tensor = gen_dataset_ops.parallel_interleave_dataset_v4(
- input_dataset._variant_tensor, # pylint: disable=protected-access
- self._map_func.function.captured_inputs, # pylint: disable=protected-access
- self._cycle_length,
- self._block_length,
- self._buffer_output_elements,
- self._prefetch_input_elements,
- self._num_parallel_calls,
- f=self._map_func.function,
- deterministic=deterministic_string,
- **self._flat_structure)
- else:
- variant_tensor = gen_dataset_ops.parallel_interleave_dataset_v3(
- input_dataset._variant_tensor, # pylint: disable=protected-access
- self._map_func.function.captured_inputs, # pylint: disable=protected-access
- self._cycle_length,
- self._block_length,
- self._num_parallel_calls,
- f=self._map_func.function,
- deterministic=deterministic_string,
- **self._flat_structure)
+ variant_tensor = gen_dataset_ops.parallel_interleave_dataset_v4(
+ input_dataset._variant_tensor, # pylint: disable=protected-access
+ self._map_func.function.captured_inputs, # pylint: disable=protected-access
+ self._cycle_length,
+ self._block_length,
+ self._buffer_output_elements,
+ self._prefetch_input_elements,
+ self._num_parallel_calls,
+ f=self._map_func.function,
+ deterministic=deterministic_string,
+ **self._flat_structure)
super(ParallelInterleaveDataset, self).__init__(input_dataset,
variant_tensor)
diff --git a/tensorflow/python/data/service/server_lib.py b/tensorflow/python/data/service/server_lib.py
index 45b1924..ab6fb35 100644
--- a/tensorflow/python/data/service/server_lib.py
+++ b/tensorflow/python/data/service/server_lib.py
@@ -39,6 +39,7 @@
`data/service:local_credentials`.
"""
self._server = _pywrap_server_lib.TF_DATA_NewMasterServer(0, protocol)
+ self._running = True
@property
def target(self):
@@ -47,21 +48,30 @@
The returned string will be in the form protocol://address:port, e.g.
"grpc://localhost:1000".
"""
- return _pywrap_server_lib.TF_DATA_ServerTarget(self._server)
+ return _pywrap_server_lib.TF_DATA_MasterServerTarget(self._server)
- def __del__(self):
+ def num_tasks(self):
+ """Returns the number of tasks on the master."""
+ return _pywrap_server_lib.TF_DATA_MasterServerNumTasks(self._server)
+
+ def stop(self):
"""Shuts down and deletes the server.
This method will block until all outstanding rpcs have completed and the
server has been shut down.
"""
- _pywrap_server_lib.TF_DATA_DeleteServer(self._server)
+ if self._running:
+ self._running = False
+ _pywrap_server_lib.TF_DATA_DeleteMasterServer(self._server)
+
+ def __del__(self):
+ self.stop()
class WorkerServer(object):
"""An in-process tf.data service worker, for use in testing."""
- def __init__(self, protocol, master_address):
+ def __init__(self, protocol, master_address, port=0):
"""Creates and starts a new tf.data worker server.
The server will choose an available port. Use `target()` to get the string
@@ -73,9 +83,11 @@
"grpc+local", and make sure your binary links in
`data/service:local_credentials`.
master_address: The address of the tf.data master server to register with.
+ port: The port to bind to.
"""
self._server = _pywrap_server_lib.TF_DATA_NewWorkerServer(
- 0, protocol, master_address)
+ port, protocol, master_address)
+ self._running = True
@property
def target(self):
@@ -84,12 +96,17 @@
The returned string will be in the form protocol://address:port, e.g.
"grpc://localhost:1000".
"""
- return _pywrap_server_lib.TF_DATA_ServerTarget(self._server)
+ return _pywrap_server_lib.TF_DATA_WorkerServerTarget(self._server)
- def __del__(self):
+ def stop(self):
"""Shuts down and deletes the server.
This method will block until all outstanding rpcs have completed and the
server has been shut down.
"""
- _pywrap_server_lib.TF_DATA_DeleteServer(self._server)
+ if self._running:
+ self._running = False
+ _pywrap_server_lib.TF_DATA_DeleteWorkerServer(self._server)
+
+ def __del__(self):
+ self.stop()
diff --git a/tensorflow/python/data/service/server_lib_wrapper.cc b/tensorflow/python/data/service/server_lib_wrapper.cc
index 16a12ee..de3cd4e 100644
--- a/tensorflow/python/data/service/server_lib_wrapper.cc
+++ b/tensorflow/python/data/service/server_lib_wrapper.cc
@@ -16,6 +16,7 @@
#include "Python.h"
#include "pybind11/chrono.h"
#include "pybind11/complex.h"
+#include "pybind11/detail/common.h"
#include "pybind11/functional.h"
#include "pybind11/pybind11.h"
#include "pybind11/pytypes.h"
@@ -27,13 +28,14 @@
namespace py = pybind11;
PYBIND11_MODULE(_pywrap_server_lib, m) {
- py::class_<tensorflow::data::GrpcDataServer>(m, "GrpcDataServer");
+ py::class_<tensorflow::data::MasterGrpcDataServer>(m, "MasterGrpcDataServer");
+ py::class_<tensorflow::data::WorkerGrpcDataServer>(m, "WorkerGrpcDataServer");
m.def(
"TF_DATA_NewMasterServer",
[](int port, std::string protocol)
- -> std::unique_ptr<tensorflow::data::GrpcDataServer> {
- std::unique_ptr<tensorflow::data::GrpcDataServer> server;
+ -> std::unique_ptr<tensorflow::data::MasterGrpcDataServer> {
+ std::unique_ptr<tensorflow::data::MasterGrpcDataServer> server;
tensorflow::Status status =
tensorflow::data::NewMasterServer(port, protocol, &server);
tensorflow::MaybeRaiseFromStatus(status);
@@ -41,12 +43,29 @@
return server;
},
py::return_value_policy::reference);
+ m.def(
+ "TF_DATA_MasterServerTarget",
+ [](tensorflow::data::MasterGrpcDataServer* server) -> std::string {
+ return server->Target();
+ },
+ py::return_value_policy::copy);
+ m.def("TF_DATA_DeleteMasterServer",
+ [](tensorflow::data::MasterGrpcDataServer* server) { server->Stop(); });
+ m.def(
+ "TF_DATA_MasterServerNumTasks",
+ [](tensorflow::data::MasterGrpcDataServer* server) -> int {
+ int num_tasks;
+ tensorflow::Status status = server->NumTasks(&num_tasks);
+ tensorflow::MaybeRaiseFromStatus(status);
+ return num_tasks;
+ },
+ py::return_value_policy::copy);
m.def(
"TF_DATA_NewWorkerServer",
[](int port, std::string protocol, std::string master_address)
- -> std::unique_ptr<tensorflow::data::GrpcDataServer> {
- std::unique_ptr<tensorflow::data::GrpcDataServer> server;
+ -> std::unique_ptr<tensorflow::data::WorkerGrpcDataServer> {
+ std::unique_ptr<tensorflow::data::WorkerGrpcDataServer> server;
tensorflow::Status status = tensorflow::data::NewWorkerServer(
port, protocol, master_address, &server);
tensorflow::MaybeRaiseFromStatus(status);
@@ -55,11 +74,11 @@
},
py::return_value_policy::reference);
m.def(
- "TF_DATA_ServerTarget",
- [](tensorflow::data::GrpcDataServer* server) -> std::string {
+ "TF_DATA_WorkerServerTarget",
+ [](tensorflow::data::WorkerGrpcDataServer* server) -> std::string {
return server->Target();
},
py::return_value_policy::copy);
- m.def("TF_DATA_DeleteServer",
- [](tensorflow::data::GrpcDataServer* server) { server->Stop(); });
+ m.def("TF_DATA_DeleteWorkerServer",
+ [](tensorflow::data::WorkerGrpcDataServer* server) { server->Stop(); });
};
diff --git a/tensorflow/python/debug/lib/check_numerics_callback_test.py b/tensorflow/python/debug/lib/check_numerics_callback_test.py
index ea5d70f..5f578da 100644
--- a/tensorflow/python/debug/lib/check_numerics_callback_test.py
+++ b/tensorflow/python/debug/lib/check_numerics_callback_test.py
@@ -94,10 +94,16 @@
dataset = dataset_ops.Dataset.from_tensor_slices(tensor).batch(2).map(
map_fn)
- iterator = dataset_ops.make_one_shot_iterator(dataset)
- self.assertAllClose(self.evaluate(iterator.get_next()), np.log([1.25, 2]))
- self.assertAllClose(self.evaluate(iterator.get_next()), np.log([3.25, 5]))
+ @def_function.function
+ def get_batches():
+ iterator = iter(dataset)
+ return [next(iterator), next(iterator)]
+
+ batches = self.evaluate(get_batches())
+ self.assertLen(batches, 2)
+ self.assertAllClose(batches[0], np.log([1.25, 2]))
+ self.assertAllClose(batches[1], np.log([3.25, 5]))
class CheckNumericsCallbackUnhealthyTest(test_util.TensorFlowTestCase):
@@ -268,6 +274,23 @@
self.assertIn("accum.assign(accum * 2.0)", message)
@test_util.run_in_graph_and_eager_modes
+ def testNanInConstIsCaptured(self):
+ check_numerics_callback.enable_check_numerics()
+ v = variables.Variable(3.0, dtype=dtypes.float32)
+ @def_function.function
+ def add_a_bad_constant(x):
+ c = constant_op.constant(np.nan)
+ return x + c
+ if not context.executing_eagerly():
+ self.evaluate(v.initializer)
+ message = self._assertRaisesInvalidArgumentErrorAndGetMessage(
+ lambda: self.evaluate(add_a_bad_constant(v)))
+ self.assertTrue(re.search(r"graph op.*\"Const\"", message))
+ self.assertTrue(re.search(r"dtype:.*float32", message))
+ self.assertTrue(re.search(r"shape:.*\(\)", message))
+ self.assertTrue(re.search(r"Graph name:.*add_a_bad_constant", message))
+
+ @test_util.run_in_graph_and_eager_modes
def testCatchInfinityInDatasetMapFunction(self):
"""Test that callback catches NaN in a tf.dataset map function."""
check_numerics_callback.enable_check_numerics()
diff --git a/tensorflow/python/debug/lib/debug_events_monitors_test.py b/tensorflow/python/debug/lib/debug_events_monitors_test.py
index 05eaa51..e8dcd6e 100644
--- a/tensorflow/python/debug/lib/debug_events_monitors_test.py
+++ b/tensorflow/python/debug/lib/debug_events_monitors_test.py
@@ -173,7 +173,8 @@
self.assertLen(traces[1].debug_tensor_value, 11)
self.assertLen(traces[2].debug_tensor_value, 11)
elif tensor_debug_mode == "FULL_TENSOR":
- self.assertLen(traces, 4) # [Placeholder:0, Unique:0, Unique:1, Sum:0].
+ # [Placeholder:0, Unique:0, Unique:1, Const:0, Sum:0].
+ self.assertLen(traces, 5)
self.assertEqual(traces[0].op_type, "Placeholder")
self.assertEqual(traces[0].output_slot, 0)
self.assertIsNone(traces[0].debug_tensor_value)
@@ -192,11 +193,16 @@
self.assertAllEqual(
reader.graph_execution_trace_to_tensor_value(traces[2]),
[0, 1, 2, 3, 0])
- self.assertEqual(traces[3].op_type, "Sum")
+ self.assertEqual(traces[3].op_type, "Const")
self.assertEqual(traces[3].output_slot, 0)
self.assertIsNone(traces[3].debug_tensor_value)
self.assertAllClose(
- reader.graph_execution_trace_to_tensor_value(traces[3]), 17.)
+ reader.graph_execution_trace_to_tensor_value(traces[3]), [0])
+ self.assertEqual(traces[4].op_type, "Sum")
+ self.assertEqual(traces[4].output_slot, 0)
+ self.assertIsNone(traces[4].debug_tensor_value)
+ self.assertAllClose(
+ reader.graph_execution_trace_to_tensor_value(traces[4]), 17.)
class AlertDataObjectsTest(test_util.TensorFlowTestCase):
diff --git a/tensorflow/python/debug/lib/debug_events_reader.py b/tensorflow/python/debug/lib/debug_events_reader.py
index 10fd0d6..4ff7b26 100644
--- a/tensorflow/python/debug/lib/debug_events_reader.py
+++ b/tensorflow/python/debug/lib/debug_events_reader.py
@@ -1030,48 +1030,61 @@
else:
return self._graph_op_digests
- def graph_execution_traces(self, digest=False):
+ def graph_execution_traces(self, digest=False, begin=None, end=None):
"""Get all the intra-graph execution tensor traces read so far.
- TODO(cais): Support begin and end to enable partial loading.
-
Args:
digest: Whether the results will be returned in the more light-weight
digest form.
+ begin: Optional beginning index for the requested traces or their digests.
+ Python-style negative indices are supported.
+ end: Optional ending index for the requested traces or their digests.
+ Python-style negative indices are supported.
Returns:
If `digest`: a `list` of `GraphExecutionTraceDigest` objects.
Else: a `list` of `GraphExecutionTrace` objects.
"""
+ digests = self._graph_execution_trace_digests
+ if begin is not None or end is not None:
+ begin = begin or 0
+ end = end or len(digests)
+ digests = digests[begin:end]
if digest:
- return self._graph_execution_trace_digests
+ return digests
else:
- return [self.read_graph_execution_trace(digest)
- for digest in self._graph_execution_trace_digests]
+ return [self.read_graph_execution_trace(digest) for digest in digests]
def num_graph_execution_traces(self):
"""Get the number of graph execution traces read so far."""
return len(self._graph_execution_trace_digests)
- def executions(self, digest=False):
+ def executions(self, digest=False, begin=None, end=None):
"""Get `Execution`s or `ExecutionDigest`s this reader has read so far.
- # TODO(cais): Support begin index and end index to support partial loading.
-
Args:
digest: Whether the results are returned in a digest form, i.e.,
`ExecutionDigest` format, instead of the more detailed `Execution`
format.
+ begin: Optional beginning index for the requested execution data objects
+ or their digests. Python-style negative indices are supported.
+ end: Optional ending index for the requested execution data objects or
+ their digests. Python-style negative indices are supported.
Returns:
If `digest`: a `list` of `ExecutionDigest` objects.
Else: a `list` of `Execution` objects.
"""
+ digests = self._execution_digests
+ if begin is not None or end is not None:
+ begin = begin or 0
+ end = end or len(digests)
+ digests = digests[begin:end]
if digest:
- return self._execution_digests
+ return digests
else:
# TODO(cais): Optimizer performance removing repeated file open/close.
- return [self.read_execution(digest) for digest in self._execution_digests]
+ return [self.read_execution(digest) for digest in digests]
def num_executions(self):
"""Get the number of execution events read so far."""
diff --git a/tensorflow/python/debug/lib/debug_events_writer_test.py b/tensorflow/python/debug/lib/debug_events_writer_test.py
index 45b7f16..c953b5f 100644
--- a/tensorflow/python/debug/lib/debug_events_writer_test.py
+++ b/tensorflow/python/debug/lib/debug_events_writer_test.py
@@ -24,6 +24,8 @@
import threading
import time
+from absl.testing import parameterized
+
from tensorflow.core.protobuf import debug_event_pb2
from tensorflow.python.debug.lib import debug_events_reader
from tensorflow.python.debug.lib import debug_events_writer
@@ -34,7 +36,8 @@
from tensorflow.python.platform import googletest
-class DebugEventsWriterTest(dumping_callback_test_lib.DumpingCallbackTestBase):
+class DebugEventsWriterTest(dumping_callback_test_lib.DumpingCallbackTestBase,
+ parameterized.TestCase):
def testMultiThreadedConstructorCallWorks(self):
def init_writer():
@@ -51,15 +54,15 @@
# Verify that there is only one debug event file of each type.
metadata_paths = glob.glob(os.path.join(self.dump_root, "*.metadata"))
- self.assertEqual(len(metadata_paths), 1)
+ self.assertLen(metadata_paths, 1)
source_files_paths = glob.glob(
os.path.join(self.dump_root, "*.source_files"))
- self.assertEqual(len(source_files_paths), 1)
+ self.assertLen(source_files_paths, 1)
stack_frames_paths = glob.glob(
os.path.join(self.dump_root, "*.stack_frames"))
- self.assertEqual(len(stack_frames_paths), 1)
+ self.assertLen(stack_frames_paths, 1)
graphs_paths = glob.glob(os.path.join(self.dump_root, "*.graphs"))
- self.assertEqual(len(graphs_paths), 1)
+ self.assertLen(graphs_paths, 1)
self._readAndCheckMetadataFile()
def testWriteSourceFilesAndStackFrames(self):
@@ -256,7 +259,7 @@
actuals = list(reader.graph_execution_traces_iterator())
# Before FlushExecutionFiles() is called. No data should have been written
# to the file.
- self.assertEqual(len(actuals), 0)
+ self.assertEmpty(actuals)
writer.FlushExecutionFiles()
actuals = list(item.debug_event.graph_execution_trace
@@ -520,6 +523,65 @@
for i in range(100):
self.assertEqual(traces[i].op_name, "Op%d" % i)
+ @parameterized.named_parameters(
+ ("Begin1End3", 1, 3, 1, 3),
+ ("Begin0End3", 0, 3, 0, 3),
+ ("Begin0EndNeg1", 0, -1, 0, 4),
+ ("BeginNoneEnd3", None, 3, 0, 3),
+ ("Begin2EndNone", 2, None, 2, 5),
+ ("BeginNoneEndNone", None, None, 0, 5),
+ )
+ def testRangeReadingExecutions(self, begin, end, expected_begin,
+ expected_end):
+ writer = debug_events_writer.DebugEventsWriter(
+ self.dump_root, circular_buffer_size=-1)
+ for i in range(5):
+ execution = debug_event_pb2.Execution(op_type="OpType%d" % i)
+ writer.WriteExecution(execution)
+ writer.FlushExecutionFiles()
+ writer.Close()
+
+ with debug_events_reader.DebugDataReader(self.dump_root) as reader:
+ reader.update()
+ executions = reader.executions(begin=begin, end=end)
+ self.assertLen(executions, expected_end - expected_begin)
+ self.assertEqual(executions[0].op_type, "OpType%d" % expected_begin)
+ self.assertEqual(executions[-1].op_type, "OpType%d" % (expected_end - 1))
+
+ @parameterized.named_parameters(
+ ("Begin1End3", 1, 3, 1, 3),
+ ("Begin0End3", 0, 3, 0, 3),
+ ("Begin0EndNeg1", 0, -1, 0, 4),
+ ("BeginNoneEnd3", None, 3, 0, 3),
+ ("Begin2EndNone", 2, None, 2, 5),
+ ("BeginNoneEndNone", None, None, 0, 5),
+ )
+ def testRangeReadingGraphExecutionTraces(self, begin, end, expected_begin,
+ expected_end):
+ writer = debug_events_writer.DebugEventsWriter(
+ self.dump_root, circular_buffer_size=-1)
+ debugged_graph = debug_event_pb2.DebuggedGraph(
+ graph_id="graph1", graph_name="graph1")
+ writer.WriteDebuggedGraph(debugged_graph)
+ for i in range(5):
+ op_name = "Op_%d" % i
+ graph_op_creation = debug_event_pb2.GraphOpCreation(
+ op_name=op_name, graph_id="graph1")
+ writer.WriteGraphOpCreation(graph_op_creation)
+ trace = debug_event_pb2.GraphExecutionTrace(
+ op_name=op_name, tfdbg_context_id="graph1")
+ writer.WriteGraphExecutionTrace(trace)
+ writer.FlushNonExecutionFiles()
+ writer.FlushExecutionFiles()
+ writer.Close()
+
+ with debug_events_reader.DebugDataReader(self.dump_root) as reader:
+ reader.update()
+ traces = reader.graph_execution_traces(begin=begin, end=end)
+ self.assertLen(traces, expected_end - expected_begin)
+ self.assertEqual(traces[0].op_name, "Op_%d" % expected_begin)
+ self.assertEqual(traces[-1].op_name, "Op_%d" % (expected_end - 1))
+
class DataObjectsTest(test_util.TensorFlowTestCase):
diff --git a/tensorflow/python/debug/lib/dumping_callback.py b/tensorflow/python/debug/lib/dumping_callback.py
index 9218910..efc5caa 100644
--- a/tensorflow/python/debug/lib/dumping_callback.py
+++ b/tensorflow/python/debug/lib/dumping_callback.py
@@ -292,7 +292,12 @@
# TODO(cais): Evaluate performance optimization options. For the
# `NO_TENSOR` debug mode, an alternative is to add `debug_tensor` as a
# control dependency of `tensor.op` without an additional identity op.
- if tensor_debug_mode == debug_event_pb2.TensorDebugMode.FULL_TENSOR:
+ if (tensor_debug_mode == debug_event_pb2.TensorDebugMode.FULL_TENSOR and
+ op_type != "Const"):
+ # NOTE(b/153716279): Under v1 graph mode, overriding the output tensor
+ # of Const ops can lead to downstream errors related to shapes. We opt
+ # to use an identity op to avoid this issue at the cost of slightly
+ # larger graph size.
return debug_tensor
else:
identity = array_ops.identity(tensor)
@@ -530,8 +535,8 @@
is_v1_graph_mode = not ops.executing_eagerly_outside_functions()
context_id = self._get_context_id(graph) # Innermost context ID.
output_tensor_ids = self._get_symbolic_tensor_ids(len(outputs))
- if op_type in ("Placeholder", "PlaceholderWithDefault"):
- # In some cases, the op name of a Placeholder op in a graph
+ if op_type in ("Const", "Placeholder", "PlaceholderWithDefault"):
+ # In some cases, the op name of a Const or Placeholder op in a graph
# can be duplicate (e.g., with the name "resource").
# When this happens, we give the op an debugger-generated name
# in order to prevent problems and check failures down the pipe.
diff --git a/tensorflow/python/debug/lib/dumping_callback_test.py b/tensorflow/python/debug/lib/dumping_callback_test.py
index 5f932ef..3486430 100644
--- a/tensorflow/python/debug/lib/dumping_callback_test.py
+++ b/tensorflow/python/debug/lib/dumping_callback_test.py
@@ -289,7 +289,8 @@
with debug_events_reader.DebugDataReader(self.dump_root) as reader:
reader.update()
graph_exec_traces = reader.graph_execution_traces()
- executed_op_types = [trace.op_type for trace in graph_exec_traces]
+ executed_op_types = [trace.op_type for trace in graph_exec_traces
+ if trace.op_type != "Const"]
self.assertCountEqual(
executed_op_types,
["Placeholder", "Placeholder", "AddV2", "Sub", "RealDiv"])
@@ -345,6 +346,46 @@
[tensor_id, 19, 1, 8, 8, 0, 0, 0, 0, 0])
@parameterized.named_parameters(
+ ("CurtHealth", "CURT_HEALTH"),
+ ("FullTensor", "FULL_TENSOR"),
+ )
+ @test_util.run_in_graph_and_eager_modes
+ def testConstTensorsAreCaptured(self, tensor_debug_mode):
+ writer = dumping_callback.enable_dump_debug_info(
+ self.dump_root, tensor_debug_mode=tensor_debug_mode)
+ @def_function.function
+ def times_two_plus_three(x):
+ return x * constant_op.constant(2.0) + constant_op.constant(3.0)
+ self.assertAllEqual(
+ self.evaluate(times_two_plus_three(10.0)), 23.0)
+ writer.FlushNonExecutionFiles()
+ writer.FlushExecutionFiles()
+
+ with debug_events_reader.DebugDataReader(self.dump_root) as reader:
+ reader.update()
+ const_traces = [trace for trace in reader.graph_execution_traces()
+ if trace.op_type == "Const"]
+ self.assertGreaterEqual(len(const_traces), 3)
+ if tensor_debug_mode == "CURT_HEALTH":
+ # Under CURT_HEALTH, each debug tensor value has the form
+ # [tensor_id, has_inf_or_nan].
+ self.assertLen(const_traces[0].debug_tensor_value, 2)
+ self.assertEqual(const_traces[0].debug_tensor_value[1], 0)
+ self.assertLen(const_traces[1].debug_tensor_value, 2)
+ self.assertEqual(const_traces[1].debug_tensor_value[1], 0)
+ self.assertLen(const_traces[2].debug_tensor_value, 2)
+ self.assertEqual(const_traces[2].debug_tensor_value[1], 0)
+ else: # FULL_TENSOR.
+ const_tensor_values = [
+ reader.graph_execution_trace_to_tensor_value(const_trace)
+ for const_trace in const_traces]
+ # Avoid making assertion on the particular order of the debug tensors
+ # for the three Consts because it may be indeterminate.
+ self.assertIn(10.0, const_tensor_values)
+ self.assertIn(2.0, const_tensor_values)
+ self.assertIn(3.0, const_tensor_values)
+
+ @parameterized.named_parameters(
("Shape", "SHAPE"),
)
@test_util.run_in_graph_and_eager_modes
@@ -367,7 +408,8 @@
with debug_events_reader.DebugDataReader(self.dump_root) as reader:
reader.update()
graph_exec_traces = reader.graph_execution_traces()
- executed_op_types = [trace.op_type for trace in graph_exec_traces]
+ executed_op_types = [trace.op_type for trace in graph_exec_traces
+ if trace.op_type != "Const"]
self.assertEqual(
executed_op_types,
["Placeholder", "Placeholder", "LogicalAnd", "LogicalNot"])
@@ -489,7 +531,8 @@
_, stack_frames = reader.read_graph_op_creation_stack_trace(op_digest)
self._verifyStackFrames(stack_frames)
- graph_exec_traces = reader.graph_execution_traces()
+ graph_exec_traces = [trace for trace in reader.graph_execution_traces()
+ if trace.op_type != "Const"]
executed_op_types = [digest.op_type for digest in graph_exec_traces]
self.assertEqual(
executed_op_types,
@@ -902,10 +945,10 @@
reader.update()
graph_exec_digests = reader.graph_execution_traces(digest=True)
executed_op_types = [digest.op_type for digest in graph_exec_digests
- if digest.op_type != "Placeholder"]
+ if digest.op_type not in ("Const", "Placeholder")]
tensor_values = [reader.graph_execution_trace_to_tensor_value(digest)
for digest in graph_exec_digests
- if digest.op_type != "Placeholder"]
+ if digest.op_type not in ("Const", "Placeholder")]
if tensor_dtypes == [dtypes.float32] and not op_regex:
self.assertEqual(executed_op_types, ["Unique", "Sum"])
@@ -1003,7 +1046,8 @@
self.assertAllClose(tensor_values, [8.0])
graph_exec_traces = reader.graph_execution_traces()
- executed_op_types = [trace.op_type for trace in graph_exec_traces]
+ executed_op_types = [trace.op_type for trace in graph_exec_traces
+ if trace.op_type != "Const"]
if tensor_debug_mode != "CURT_HEALTH":
# Less outputs a boolean tensor, which is not tracked under CURT_HEALTH.
# The Less op should have been executed 5 times.
diff --git a/tensorflow/python/distribute/BUILD b/tensorflow/python/distribute/BUILD
index 2c6bbe1..6aad253 100644
--- a/tensorflow/python/distribute/BUILD
+++ b/tensorflow/python/distribute/BUILD
@@ -848,9 +848,7 @@
main = "input_lib_test.py",
shard_count = 10,
tags = [
- "manual",
"multi_and_single_gpu",
- "notap", # TODO(b/151467526)
],
deps = [
":collective_all_reduce_strategy",
@@ -1242,6 +1240,7 @@
cuda_py_test(
name = "remote_mirrored_strategy_eager_test",
srcs = ["remote_mirrored_strategy_eager_test.py"],
+ tags = ["no_oss"], # b/154743849
deps = [
":combinations",
":distribute_lib",
diff --git a/tensorflow/python/distribute/input_lib.py b/tensorflow/python/distribute/input_lib.py
index 6cf6bd0..9919884 100644
--- a/tensorflow/python/distribute/input_lib.py
+++ b/tensorflow/python/distribute/input_lib.py
@@ -511,11 +511,6 @@
else:
raise
- # TODO(b/138745411): Remove once stateful transformations are supported.
- options = dataset_ops.Options()
- options.experimental_distribute._make_stateless = True # pylint: disable=protected-access
- dataset = dataset.with_options(options)
-
self._cloned_datasets = []
if input_context:
# Between-graph where we rely on the input_context for sharding
@@ -1034,10 +1029,6 @@
worker = input_workers.worker_devices[i]
with ops.device(worker):
dataset = dataset_fn(ctx)
- # TODO(b/138745411): Remove once stateful transformations are supported.
- options = dataset_ops.Options()
- options.experimental_distribute._make_stateless = True # pylint: disable=protected-access
- dataset = dataset.with_options(options)
devices = input_workers.compute_devices_for_worker(i)
iterator = _SingleWorkerDatasetIterator(dataset, worker, devices)
iterators.append(iterator)
diff --git a/tensorflow/python/distribute/input_lib_test.py b/tensorflow/python/distribute/input_lib_test.py
index 8995704..be78bf1 100644
--- a/tensorflow/python/distribute/input_lib_test.py
+++ b/tensorflow/python/distribute/input_lib_test.py
@@ -26,6 +26,7 @@
import numpy as np
from tensorflow.python import tf2
+from tensorflow.python.compat import compat
from tensorflow.python.data.experimental.ops.distribute_options import AutoShardPolicy
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.distribute import collective_all_reduce_strategy
@@ -231,6 +232,8 @@
strategy_combinations.mirrored_strategy_with_gpu_and_cpu
]))
def testMultiDeviceIterInitialize(self, distribution):
+ if tf2.enabled():
+ self.skipTest("Only V1 is supported.")
worker_device_pairs = [("", ["/device:GPU:0", "/device:CPU:0"])]
dataset_fn = lambda _: dataset_ops.DatasetV1.range(10)
@@ -511,6 +514,63 @@
sess=None,
split_batch_by=split_batch_by)
+ @combinations.generate(
+ combinations.combine(
+ mode=["eager"],
+ distribution=[
+ strategy_combinations.one_device_strategy,
+ strategy_combinations.mirrored_strategy_with_one_cpu,
+ strategy_combinations.mirrored_strategy_with_gpu_and_cpu,
+ strategy_combinations.tpu_strategy,
+ strategy_combinations.central_storage_strategy_with_two_gpus,
+ ],
+ ))
+ def testCacheAcrossIteration(self, distribution):
+ if not tf2.enabled():
+ self.skipTest("Only V2 is supported.")
+
+ dataset = dataset_ops.Dataset.range(10).shuffle(10).cache().batch(2)
+ dist_dataset = distribution.experimental_distribute_dataset(dataset)
+
+ first_epoch = list(
+ distribution.experimental_local_results(x) for x in dist_dataset)
+ second_epoch = list(
+ distribution.experimental_local_results(x) for x in dist_dataset)
+
+ self.assertAllEqual(first_epoch, second_epoch)
+
+ @combinations.generate(
+ combinations.combine(
+ mode=["eager"],
+ distribution=[
+ strategy_combinations.one_device_strategy,
+ strategy_combinations.mirrored_strategy_with_one_cpu,
+ strategy_combinations.mirrored_strategy_with_gpu_and_cpu,
+ strategy_combinations.tpu_strategy,
+ strategy_combinations.central_storage_strategy_with_two_gpus,
+ ],
+ reshuffle=[True, False]))
+ def testShuffleAcrossIterations(self, distribution, reshuffle):
+ if not tf2.enabled():
+ self.skipTest("Only V2 is supported.")
+
+ if not reshuffle and not compat.forward_compatible(2020, 5, 22):
+ self.skipTest("Functionality currently not supported.")
+
+ dataset = dataset_ops.Dataset.range(10).shuffle(
+ 10, reshuffle_each_iteration=reshuffle).batch(2)
+ dist_dataset = distribution.experimental_distribute_dataset(dataset)
+
+ first_epoch = list(
+ distribution.experimental_local_results(x) for x in dist_dataset)
+ second_epoch = list(
+ distribution.experimental_local_results(x) for x in dist_dataset)
+
+ if reshuffle:
+ self.assertNotAllEqual(first_epoch, second_epoch)
+ else:
+ self.assertAllEqual(first_epoch, second_epoch)
+
class DistributedIteratorTensorTypeTest(DistributedIteratorTestBase,
parameterized.TestCase):
@@ -1078,6 +1138,5 @@
for x in dist_dataset:
process_inputs(x)
-
if __name__ == "__main__":
test.main()
diff --git a/tensorflow/python/distribute/mirrored_run.py b/tensorflow/python/distribute/mirrored_run.py
index 2cd139c..aed7b36 100644
--- a/tensorflow/python/distribute/mirrored_run.py
+++ b/tensorflow/python/distribute/mirrored_run.py
@@ -364,7 +364,7 @@
This pauses the current replica thread and passes `fn` and its arguments to
the main thread. The main thread will wait until all replicas pause, then
- invoke `fn` with grouped arugments. The current replica thread will continue
+ invoke `fn` with grouped arguments. The current replica thread will continue
after `fn` completes.
See `_call_for_each_replica` for the logic in the main thread.
diff --git a/tensorflow/python/distribute/parallel_device/BUILD b/tensorflow/python/distribute/parallel_device/BUILD
new file mode 100644
index 0000000..e7526a5
--- /dev/null
+++ b/tensorflow/python/distribute/parallel_device/BUILD
@@ -0,0 +1,45 @@
+package(
+ licenses = ["notice"], # Apache 2.0
+)
+
+# Pybind rules must live in tensorflow/python due to header rule visibility.
+exports_files(
+ ["pywrap_parallel_device.cc"],
+ visibility = ["//tensorflow/python:__pkg__"],
+)
+
+py_library(
+ name = "parallel_device",
+ srcs = ["parallel_device.py"],
+ srcs_version = "PY2AND3",
+ deps = [
+ ":saving",
+ "//tensorflow/python:_pywrap_parallel_device",
+ ],
+)
+
+py_library(
+ name = "saving",
+ srcs = ["saving.py"],
+ srcs_version = "PY2AND3",
+ deps = ["//tensorflow/python:framework_ops"],
+)
+
+py_test(
+ name = "parallel_device_test",
+ srcs = ["parallel_device_test.py"],
+ python_version = "PY3",
+ tags = [
+ # Dependencies aren't otherwise included in the pip package yet.
+ "no_pip",
+ ],
+ deps = [
+ ":parallel_device",
+ ":saving",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python:collective_ops",
+ "//tensorflow/python:framework_ops",
+ "//tensorflow/python/module",
+ "//tensorflow/python/tpu",
+ ],
+)
diff --git a/tensorflow/python/distribute/parallel_device/parallel_device.py b/tensorflow/python/distribute/parallel_device/parallel_device.py
new file mode 100644
index 0000000..982b061
--- /dev/null
+++ b/tensorflow/python/distribute/parallel_device/parallel_device.py
@@ -0,0 +1,95 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Utility for eagerly executing operations in parallel on multiple devices."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import contextlib
+import threading
+
+from tensorflow.python import _pywrap_parallel_device
+from tensorflow.python.distribute.parallel_device import saving
+from tensorflow.python.eager import context
+from tensorflow.python.framework import ops
+from tensorflow.python.tpu.ops import tpu_ops
+
+_next_device_number = 0
+_next_device_number_lock = threading.Lock()
+
+
+# TODO(allenl): Expand this docstring once things like getting components on and
+# off the device are stable.
+class ParallelDevice(object):
+ """A device which executes operations in parallel."""
+
+ def __init__(self, components):
+ """Creates a device which executes operations in parallel on `components`.
+
+ Args:
+ components: A list of device names. Each operation executed on the
+ returned device executes on these component devices.
+
+ Returns:
+ A string with the name of the newly created device.
+ """
+ global _next_device_number, _next_device_number_lock
+ self.components = tuple(components)
+ ctx = context.context()
+ with _next_device_number_lock:
+ # TODO(allenl): Better names for parallel devices (right now "CUSTOM" is
+ # special-cased).
+ self.name = "{}/device:CUSTOM:{}".format(
+ ctx.host_address_space(), _next_device_number)
+ _next_device_number += 1
+ device, device_info = _pywrap_parallel_device.GetParallelDeviceCapsules(
+ self.name, self.components)
+ context.register_custom_device(device, self.name, device_info)
+
+ def pack(self, tensors):
+ """Create a tensor on the parallel device from a sequence of tensors.
+
+ Args:
+ tensors: A flat list of tensors, one per device in `self.components`.
+
+ Returns:
+ A single tensor placed on `self.name`.
+ """
+ with ops.device(self.name):
+ return tpu_ops.tpu_replicated_input(inputs=tensors)
+
+ def unpack(self, parallel_tensor):
+ """Unpack a parallel tensor into its components.
+
+ Args:
+ parallel_tensor: A tensor placed on `self.name`.
+
+ Returns:
+ A flat list of tensors, one per `self.components`.
+ """
+ with ops.device(self.name):
+ return tpu_ops.tpu_replicated_output(
+ parallel_tensor, num_replicas=len(self.components))
+
+ # TODO(allenl): Fixing saving in Python is a bit odd. One alternative would be
+ # to provide a hook for the custom device to create save specs/etc., then call
+ # that hook from the default variable implementation if the variable is on a
+ # custom device. We'll likely want similar hooks for repr() and such.
+ @contextlib.contextmanager
+ def scope(self):
+ """Runs ops in parallel, makes variables which save independent buffers."""
+ with ops.device(self.name), saving.independent_buffers(self):
+ yield
diff --git a/tensorflow/python/distribute/parallel_device/parallel_device_test.py b/tensorflow/python/distribute/parallel_device/parallel_device_test.py
new file mode 100644
index 0000000..d3f3417
--- /dev/null
+++ b/tensorflow/python/distribute/parallel_device/parallel_device_test.py
@@ -0,0 +1,254 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import threading
+
+from tensorflow.python.distribute.parallel_device import parallel_device
+from tensorflow.python.eager import backprop
+from tensorflow.python.eager import context
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import ops
+from tensorflow.python.module import module
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import collective_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import variables
+from tensorflow.python.platform import test
+from tensorflow.python.training import checkpoint_management
+from tensorflow.python.training.tracking import util as tracking
+from tensorflow.python.util import nest
+
+# When running collectives asynchronously, we need to give each parallel device
+# execution a unique ID so the collectives don't interfere. Since the op is
+# replicated with group/instance key intact, the replicated nodes will
+# communicate.
+# TODO(allenl): Switch to using a collective manager.
+_COUNTER_LOCK = threading.Lock()
+_COUNTER = 0
+
+
+def _collective_reduce(inputs, operation, num_replicas):
+
+ def _reduce_tensor(tensor):
+ with _COUNTER_LOCK:
+ global _COUNTER
+ keys = _COUNTER
+ _COUNTER += 1
+ return collective_ops.all_reduce(
+ t=tensor,
+ group_size=num_replicas,
+ merge_op=operation,
+ group_key=keys,
+ instance_key=keys,
+ final_op="Id")
+
+ return nest.map_structure(_reduce_tensor, inputs)
+
+
+def _collective_sum(inputs, num_replicas):
+ return _collective_reduce(
+ inputs=inputs, operation="Add", num_replicas=num_replicas)
+
+
+class _Dense(module.Module):
+
+ def __init__(self, output_size):
+ self.output_size = output_size
+ self.kernel = None
+ self.bias = None
+
+ def __call__(self, x):
+ if self.kernel is None:
+ self.kernel = variables.Variable(
+ array_ops.ones(
+ array_ops.stack([self.output_size,
+ array_ops.shape(x)[-1]])))
+ self.bias = variables.Variable(array_ops.ones([self.output_size]))
+ return math_ops.matmul(x, self.kernel, transpose_b=True) + self.bias
+
+
+class _VirtualDeviceTestCase(test.TestCase):
+
+ def setUp(self):
+ super(_VirtualDeviceTestCase, self).setUp()
+ cpus = context.context().list_physical_devices("CPU")
+ # Set 4 virtual CPUs
+ context.context().set_logical_device_configuration(cpus[0], [
+ context.LogicalDeviceConfiguration(),
+ context.LogicalDeviceConfiguration(),
+ context.LogicalDeviceConfiguration(),
+ context.LogicalDeviceConfiguration()
+ ])
+
+ # TODO(allenl): Make CPU:0 and CPU:1 work (right now "CPU:1" soft-places
+ # onto CPU:0, which seems wrong).
+ components = [
+ "/job:localhost/replica:0/task:0/device:CPU:0",
+ "/job:localhost/replica:0/task:0/device:CPU:1"
+ ]
+ self.device = parallel_device.ParallelDevice(components)
+
+
+class ParallelDeviceTests(_VirtualDeviceTestCase):
+
+ def test_register_parallel_device(self):
+ with ops.device(self.device.name):
+ c = constant_op.constant(1.)
+ d = constant_op.constant(2.)
+ e = c + d
+ outputs = self.device.unpack(e)
+ self.assertAllClose([3., 3.], outputs)
+
+ self.assertIn(self.device.components[0], outputs[0].backing_device)
+ self.assertIn(self.device.components[1], outputs[1].backing_device)
+
+ def test_collective_reduce(self):
+ with ops.device(self.device.name):
+ x = self.device.pack(
+ [constant_op.constant(-1.5),
+ constant_op.constant(3.5)])
+ reduced = _collective_sum(x, num_replicas=2)
+ outputs = self.device.unpack(reduced)
+ self.assertAllClose([2., 2.], outputs)
+ self.assertIn(self.device.components[0], outputs[0].backing_device)
+ self.assertIn(self.device.components[1], outputs[1].backing_device)
+
+ def test_checkpointing(self):
+ prefix = os.path.join(self.get_temp_dir(), "ckpt")
+ with self.device.scope():
+ different_values = self.device.pack(
+ [constant_op.constant(-1.),
+ constant_op.constant(3.)])
+ v = variables.Variable(different_values)
+ checkpoint = tracking.Checkpoint(v=v)
+ save_path = checkpoint.save(prefix)
+ with ops.device(self.device.name):
+ v.assign(constant_op.constant(0.))
+ # Make sure the checkpoint is actually written before we try to read it
+ context.async_wait()
+ checkpoint.restore(save_path).assert_consumed()
+ with ops.device(self.device.name):
+ outputs = self.device.unpack(v)
+ self.assertAllClose([-1., 3.], outputs)
+
+
+class LayerTests(_VirtualDeviceTestCase):
+
+ def test_layer_forward(self):
+ with ops.device(self.device.name):
+ layer = _Dense(5)
+ x = constant_op.constant([[2.]])
+ y = layer(x)
+ outputs = self.device.unpack(y)
+ self.assertAllClose([[3.] * 5], outputs[0])
+ self.assertAllClose([[3.] * 5], outputs[1])
+ self.assertIn(self.device.components[0], outputs[0].backing_device)
+ self.assertIn(self.device.components[1], outputs[1].backing_device)
+
+ # With different Layer inputs we get different outputs
+ with ops.device(self.device.name):
+ x = self.device.pack(
+ [constant_op.constant([[-0.5]]),
+ constant_op.constant([[0.5]])])
+ y = layer(x)
+ outputs = self.device.unpack(y)
+ self.assertGreater(
+ math_ops.reduce_max(math_ops.abs(outputs[0] - outputs[1])), 1e-5)
+ self.assertIn(self.device.components[0], outputs[0].backing_device)
+ self.assertIn(self.device.components[1], outputs[1].backing_device)
+
+ def test_layer_sync_training(self):
+ with ops.device(self.device.name):
+ layer = _Dense(5)
+
+ with backprop.GradientTape() as tape:
+ x = self.device.pack(
+ [constant_op.constant([[-0.5]]),
+ constant_op.constant([[0.5]])])
+ y = layer(x)
+ loss = (y - math_ops.range(5.))**2.
+ parameters = layer.trainable_variables
+ unreduced_gradients = tape.gradient(loss, parameters)
+ reduced_gradients = _collective_sum(unreduced_gradients, num_replicas=2)
+ for grad, param in zip(reduced_gradients, parameters):
+ param.assign_sub(0.01 * grad)
+ final_kernels = self.device.unpack(layer.kernel)
+ self.assertAllClose(final_kernels[0], final_kernels[1])
+ final_bias = self.device.unpack(layer.bias)
+ expected_bias = (1. - 0.01 * 2. * (1. + .5 - math_ops.range(5.)) -
+ 0.01 * 2. * (1. - .5 - math_ops.range(5.)))
+ self.assertAllClose(expected_bias, final_bias[0])
+ self.assertAllClose(expected_bias, final_bias[1])
+ self.assertIn(self.device.components[0], final_kernels[0].backing_device)
+ self.assertIn(self.device.components[1], final_kernels[1].backing_device)
+
+ def test_layer_divergent_buffer_training(self):
+ with ops.device(self.device.name):
+ layer = _Dense(5)
+
+ with backprop.GradientTape() as tape:
+ x = self.device.pack(
+ [constant_op.constant([[-0.5]]),
+ constant_op.constant([[0.5]])])
+ y = layer(x)
+ loss = (y - math_ops.range(5.))**2.
+ parameters = layer.trainable_variables
+ unreduced_gradients = tape.gradient(loss, parameters)
+ for grad, param in zip(unreduced_gradients, parameters):
+ param.assign_sub(0.01 * grad)
+ final_kernels = self.device.unpack(layer.kernel)
+ self.assertNotAllClose(final_kernels[0], final_kernels[1])
+ final_bias = self.device.unpack(layer.bias)
+ self.assertAllClose(1. - 0.01 * 2. * (1. - .5 - math_ops.range(5.)),
+ final_bias[0])
+ self.assertAllClose(1. - 0.01 * 2. * (1. + .5 - math_ops.range(5.)),
+ final_bias[1])
+ self.assertIn(self.device.components[0], final_kernels[0].backing_device)
+ self.assertIn(self.device.components[1], final_kernels[1].backing_device)
+
+ def test_training_loop(self):
+ for _ in range(5):
+ layer = _Dense(5)
+ checkpoint = tracking.Checkpoint(layer=layer)
+ manager = checkpoint_management.CheckpointManager(
+ checkpoint, directory=self.get_temp_dir(), max_to_keep=5)
+ manager.restore_or_initialize()
+
+ for _ in range(10):
+ with self.device.scope():
+ with backprop.GradientTape() as tape:
+ x = self.device.pack(
+ [constant_op.constant([[-0.5]]),
+ constant_op.constant([[0.5]])])
+ y = layer(x)
+ loss = (y - math_ops.range(5.))**2.
+ parameters = layer.trainable_variables
+ unreduced_gradients = tape.gradient(loss, parameters)
+ reduced_gradients = _collective_sum(
+ unreduced_gradients, num_replicas=len(self.device.components))
+ for grad, param in zip(reduced_gradients, parameters):
+ param.assign_sub(0.01 * grad)
+
+ manager.save()
+
+
+if __name__ == "__main__":
+ ops.enable_eager_execution()
+ test.main()
diff --git a/tensorflow/python/distribute/parallel_device/pywrap_parallel_device.cc b/tensorflow/python/distribute/parallel_device/pywrap_parallel_device.cc
new file mode 100644
index 0000000..62488cb
--- /dev/null
+++ b/tensorflow/python/distribute/parallel_device/pywrap_parallel_device.cc
@@ -0,0 +1,70 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "Python.h"
+#include "pybind11/pybind11.h"
+#include "pybind11/stl.h"
+#include "tensorflow/c/c_api.h"
+#include "tensorflow/c/c_api_experimental.h"
+#include "tensorflow/c/eager/c_api.h"
+#include "tensorflow/c/eager/c_api_experimental.h"
+#include "tensorflow/c/eager/parallel_device/parallel_device.h"
+#include "tensorflow/python/lib/core/py_exception_registry.h"
+#include "tensorflow/python/lib/core/pybind11_lib.h"
+#include "tensorflow/python/lib/core/pybind11_status.h"
+#include "tensorflow/python/lib/core/safe_ptr.h"
+
+namespace py = pybind11;
+
+void CallDelete_Device(PyObject* capsule) {
+ delete reinterpret_cast<TFE_CustomDevice*>(
+ PyCapsule_GetPointer(capsule, "TFE_CustomDevice"));
+}
+
+void CallDelete_DeviceInfo(PyObject* capsule) {
+ void (*destructor)(void*) =
+ reinterpret_cast<void (*)(void*)>(PyCapsule_GetContext(capsule));
+ destructor(PyCapsule_GetPointer(capsule, "TFE_CustomDevice_DeviceInfo"));
+}
+
+PYBIND11_MODULE(_pywrap_parallel_device, m) {
+ m.def("GetParallelDeviceCapsules",
+ [](const char* name, std::vector<std::string> underlying_devices) {
+ std::vector<const char*> underlying_devices_c;
+ underlying_devices_c.reserve(underlying_devices.size());
+ for (const std::string& element : underlying_devices) {
+ underlying_devices_c.push_back(element.c_str());
+ }
+ // `device` is owned by `device_capsule`.
+ TFE_CustomDevice* device = new TFE_CustomDevice;
+ tensorflow::Safe_PyObjectPtr device_capsule(
+ PyCapsule_New(device, "TFE_CustomDevice", &CallDelete_Device));
+ void* device_info;
+ tensorflow::eager::AllocateParallelDevice(
+ name, underlying_devices_c.data(), underlying_devices_c.size(),
+ device, &device_info);
+ if (PyErr_Occurred()) throw py::error_already_set();
+ tensorflow::Safe_PyObjectPtr device_info_capsule(
+ PyCapsule_New(device_info, "TFE_CustomDevice_DeviceInfo",
+ &CallDelete_DeviceInfo));
+ if (PyErr_Occurred()) throw py::error_already_set();
+ // The PyCapsule destructor needs a pointer to the destructor for
+ // DeviceInfo.
+ PyCapsule_SetContext(device_info_capsule.get(),
+ reinterpret_cast<void*>(device->delete_device));
+ return tensorflow::PyoOrThrow(
+ PyTuple_Pack(2, device_capsule.get(), device_info_capsule.get()));
+ });
+}
diff --git a/tensorflow/python/distribute/parallel_device/saving.py b/tensorflow/python/distribute/parallel_device/saving.py
new file mode 100644
index 0000000..f2e7dad
--- /dev/null
+++ b/tensorflow/python/distribute/parallel_device/saving.py
@@ -0,0 +1,131 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Special-cased checkpointing for variables on a parallel device."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import contextlib
+import functools
+
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import gen_resource_variable_ops
+from tensorflow.python.ops import resource_variable_ops
+from tensorflow.python.ops import variable_scope
+from tensorflow.python.ops import variables
+from tensorflow.python.training.saving import saveable_object
+
+
+def _read_component(handle, dtype, replica_id, parallel_device):
+ """Read one component of a parallel variable and discard the rest."""
+ with ops.device(handle.device):
+ read = gen_resource_variable_ops.read_variable_op(
+ resource=handle, dtype=dtype)
+ all_components = parallel_device.unpack(read)
+ # We're pretending that parallel variables have a first axis with length
+ # num_components, so we need to add a dummy first axis to the shape that gets
+ # saved.
+ return all_components[replica_id][None, ...]
+
+
+class _ParallelDeviceSaveable(saveable_object.SaveableObject):
+ """Saves and restores a parallel variable."""
+
+ def __init__(self, name, handle, dtype, component_shape, parallel_device):
+ # Each component device gets one spec with a tensor to save.
+ specs = []
+ for replica_id, device_name in enumerate(parallel_device.components):
+ # TODO(b/151773535): SaveableObjects with SaveSpecs on different devices
+ # will cause extra copying at the moment. We should fix that before doing
+ # anything serious with this code.
+ specs.append(
+ saveable_object.SaveSpec(
+ tensor=functools.partial(
+ _read_component,
+ handle=handle,
+ dtype=dtype,
+ replica_id=replica_id,
+ parallel_device=parallel_device),
+ slice_spec=variables.Variable.SaveSliceInfo(
+ full_shape=([len(parallel_device.components)] +
+ component_shape),
+ var_offset=[replica_id] + [0] * len(component_shape),
+ var_shape=[1] + component_shape).spec,
+ device=device_name,
+ dtype=dtype,
+ name=name))
+ self._handle = handle
+ self._parallel_device = parallel_device
+ self._component_shape = component_shape
+ super(_ParallelDeviceSaveable, self).__init__(None, specs, name)
+
+ def restore(self, tensors, restored_shapes=None):
+ with ops.device(self._handle.device):
+ # Combine the restored tensors into one parallel tensor to assign.
+ bundled = self._parallel_device.pack(tensors)
+ gen_resource_variable_ops.assign_variable_op(
+ resource=self._handle,
+ # Squeeze out the dummy first axis we added when saving.
+ value=array_ops.squeeze(bundled, axis=0))
+
+
+class VariableWithFixedCheckpointing(resource_variable_ops.ResourceVariable):
+ """Overrides checkpointing behavior to save like a partitioned variable."""
+
+ def __init__(self, parallel_device, **kwargs):
+ self._parallel_device = parallel_device
+ kwargs = {k: v for k, v in kwargs.items()
+ if k not in ["use_resource", "expected_shape"]}
+ super(VariableWithFixedCheckpointing, self).__init__(**kwargs)
+
+ def _gather_saveables_for_checkpoint(self):
+ # Note VARIABLE_VALUE is the usual attribute name for variables. Using
+ # something different means (a) the checkpointing infrastructure won't try
+ # doing restore-on-create (which has shape issues), and (b) the saved
+ # variables won't be compatible with regular variables. Both of those are
+ # good in this case.
+ return dict(
+ PARALLEL_VARIABLE_VALUE=functools.partial(
+ _ParallelDeviceSaveable,
+ handle=self.handle,
+ dtype=self.dtype,
+ component_shape=self.shape,
+ parallel_device=self._parallel_device))
+
+
+def _variable_creator(next_creator, parallel_device, **kwargs):
+ del next_creator
+ return VariableWithFixedCheckpointing(
+ parallel_device=parallel_device, **kwargs)
+
+
+@contextlib.contextmanager
+def independent_buffers(parallel_device):
+ """Context manager which saves parallel buffers independently.
+
+ Creates a ParallelDevice-aware variable subclass which saves buffers for each
+ device separately.
+
+ Args:
+ parallel_device: A ParallelDevice object on which variables are placed.
+
+ Yields:
+ Nothing.
+ """
+ with variable_scope.variable_creator_scope(
+ functools.partial(_variable_creator, parallel_device=parallel_device)):
+ yield
diff --git a/tensorflow/python/distribute/tpu_values.py b/tensorflow/python/distribute/tpu_values.py
index 2ea2481..cab4815 100644
--- a/tensorflow/python/distribute/tpu_values.py
+++ b/tensorflow/python/distribute/tpu_values.py
@@ -24,7 +24,6 @@
import contextlib
-from tensorflow.python.distribute import distribution_strategy_context as ds_context
from tensorflow.python.distribute import values
from tensorflow.python.eager import context
from tensorflow.python.eager import tape
@@ -186,20 +185,10 @@
class TPUMirroredVariable(TPUVariableMixin, values.MirroredVariable):
"""Holds a map from replica to TPU variables whose values are kept in sync."""
- def _mirrored_update(self, update_fn, **kwargs):
- with ds_context.enter_or_assert_strategy(self._distribute_strategy):
- if (ds_context.in_cross_replica_context() and
- (enclosing_tpu_context() is not None)):
- return self._distribute_strategy.extended.update(
- self, update_fn, kwargs=kwargs)
- else:
- return values.MirroredVariable._mirrored_update(self, update_fn,
- **kwargs)
-
def assign_sub(self, value, use_locking=False, name=None, read_value=True):
assign_sub_fn = _make_raw_assign_fn(
gen_resource_variable_ops.assign_sub_variable_op)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_sub_fn,
value=value,
use_locking=use_locking,
@@ -209,7 +198,7 @@
def assign_add(self, value, use_locking=False, name=None, read_value=True):
assign_add_fn = _make_raw_assign_fn(
gen_resource_variable_ops.assign_add_variable_op)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_add_fn,
value=value,
use_locking=use_locking,
@@ -219,7 +208,7 @@
def assign(self, value, use_locking=False, name=None, read_value=True):
assign_fn = _make_raw_assign_fn(
gen_resource_variable_ops.assign_variable_op)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_fn,
value=value,
use_locking=use_locking,
diff --git a/tensorflow/python/distribute/values.py b/tensorflow/python/distribute/values.py
index 1a5b87f..fda2585 100644
--- a/tensorflow/python/distribute/values.py
+++ b/tensorflow/python/distribute/values.py
@@ -433,10 +433,6 @@
self._aggregation = aggregation
super(DistributedVariable, self).__init__(values)
self._common_name = self._primary.name.split(":")[0]
- # Use a weakref to make it easy to map from the contained values
- # to the container without introducing a reference cycle.
- for v in values:
- v._distributed_container = weakref.ref(self) # pylint: disable=protected-access
# tf.keras keeps track of variables initialized using this attribute. When
# tf.keras gets the default session, it initializes all uninitialized vars.
# We need to make _keras_initialized a member of DistributedVariable because
@@ -591,6 +587,65 @@
def value(self):
return self._get_closest().value()
+ def _update_cross_replica(self, update_fn, value, **kwargs):
+ """Applies updates across replicas.
+
+ Args:
+ update_fn: A callable to pass to `strategy.extended.update` to update the
+ variable. It should has the same signature as `Variable.assign()`.
+ value: value to be passed to `update_fn`.
+ **kwargs: remaining arguments to `update_fn`.
+
+ Returns:
+ Updated variable or `tf.Operation`.
+ """
+ return self.distribute_strategy.extended.update(
+ self, update_fn, args=(value,), kwargs=kwargs, group=True)
+
+ def _update_replica(self, update_fn, value, **kwargs):
+ """Applies updates in one replica.
+
+ Args:
+ update_fn: A callable to update the variable. It should has the same
+ signature as `Variable.assign()`.
+ value: value to be passed to `update_fn`.
+ **kwargs: remaining arguments to `update_fn`.
+
+ Returns:
+ Updated variable or `tf.Operation`.
+ """
+ raise NotImplementedError("should be implemented by subclass.")
+
+ def _update(self, update_fn, value, **kwargs):
+ """Applies updates depending on the context.
+
+ The method calls `_update_replica` in replica context,
+ `_update_cross_replica` in cross replica context, and `update_fn` in update
+ context.
+
+ If `read_value` is True, the method returns the updated Variable. If
+ `read_value` is False, the method returns the update `tf.Operation`.
+
+ Args:
+ update_fn: A callable to pass to `strategy.extended.update` to update the
+ variable. It should have the same signature as `Variable.assign()`.
+ value: value to be passed to `update_fn`.
+ **kwargs: keyword arguments to `update_fn`.
+
+ Returns:
+ Updated variable or `tf.Operation`.
+
+ """
+ with ds_context.enter_or_assert_strategy(self.distribute_strategy):
+ if ds_context.in_cross_replica_context():
+ update_replica_id = distribute_lib.get_update_replica_id()
+ if update_replica_id is not None:
+ return update_fn(self._values[update_replica_id], value, **kwargs)
+ return self._update_cross_replica(update_fn, value, **kwargs)
+ else:
+ _assert_replica_context(self.distribute_strategy)
+ return self._update_replica(update_fn, value, **kwargs)
+
def _should_act_as_resource_variable(self):
"""Pass resource_variable_ops.is_resource_variable check."""
pass
@@ -715,6 +770,13 @@
value_list = real_mirrored_creator(**kwargs)
var_cls = sync_on_read_cls if is_sync_on_read else mirrored_cls
result = var_cls(strategy, value_list, aggregation)
+ # Install the created DistributedVariable as _distributed_container property
+ # of the underlying variables, to make it easy to map back to the container.
+ for v in result.values:
+ # Hold a strong reference to avoid the container from being GC-ed. After
+ # v = v.assign(), the user code may no longer holds references to the
+ # original container, since v.assign() returns a new DistributedVariable.
+ v._distributed_container = result # pylint: disable=protected-access
# Add the wrapped variable to the requested collections.
# The handling of eager mode and the global step matches
@@ -745,66 +807,38 @@
class MirroredVariable(DistributedVariable, Mirrored):
"""Holds a map from replica to variables whose values are kept in sync."""
- def _mirrored_update(self, update_fn, value, **kwargs):
- """Apply identical updates using `update_fn` to variables on each replica."""
- with ds_context.enter_or_assert_strategy(self._distribute_strategy):
- if ds_context.in_cross_replica_context():
- update_replica_id = distribute_lib.get_update_replica_id()
- if update_replica_id is not None:
- # We are calling an update function on the mirrored variable in an
- # update context.
- #
- # The arguments to update() are automatically unwrapped so the
- # update() function would normally see regular variables, not
- # MirroredVariables. However, the update function can still operate on
- # wrapped MirroredVariables through object members, captured arguments
- # , etc. This is more likely in an update_non_slot() function
- # , which can update several non-slot variables in one call.
- return update_fn(self._values[update_replica_id], value, **kwargs)
+ def _update_replica(self, update_fn, value, **kwargs):
+ if self.aggregation == vs.VariableAggregation.NONE:
+ raise ValueError(
+ _aggregation_error_msg.format(variable_type="MirroredVariable"))
- # We are calling update on the mirrored variable in cross replica
- # context, use `strategy.extended.update()` to update the variable.
- return self._distribute_strategy.extended.update(
- self, update_fn, args=(value,), kwargs=kwargs)
- else:
- _assert_replica_context(self._distribute_strategy)
- # We are calling an update function on the mirrored variable in replica
- # context.
- # We reduce the value we want to update. More details about how
- # we handle the different use cases can be found in the _reduce method.
- # We call the function on each of the mirrored variables with the
- # reduced value.
- if self._aggregation == vs.VariableAggregation.NONE:
- raise ValueError(
- _aggregation_error_msg.format(variable_type="MirroredVariable"))
+ def merge_fn(strategy, value, **kwargs):
+ """Aggregate values and update all variables in cross replica context."""
+ # Don't allow MEAN with non float dtype, since it may cause unexpected
+ # precision loss. Python3 and NumPy automatically upcast integers to
+ # float in division, but we should always preserve the type.
+ #
+ # Note that to be backward compatible we allow the case when the value
+ # is *always* the same on each replica. I.E. value is not a
+ # PerReplica. Refer to regroup() to see how values are grouped.
+ if self._aggregation == vs.VariableAggregation.MEAN and (
+ not self.dtype.is_floating) and isinstance(value, PerReplica):
+ raise ValueError(
+ "Cannot update non-float variables with "
+ "tf.VariableAggregation.MEAN aggregation in replica context. "
+ "Either change the variable dtype to float or update it in "
+ "cross-replica context.")
- def merge_fn(strategy, value, **other_kwargs):
- """Aggregate across replicas and update MV with aggregated value."""
- # Don't allow MEAN with non float dtype, since it may cause unexpected
- # precision loss. Python3 and NumPy automatically upcast integers to
- # float in division, but we should always preserve the type.
- #
- # Note that to be backward compatible we allow the case when the value
- # is *always* the same on each replica. I.E. value is not a
- # PerReplica. Refer to regroup() to see how values are grouped.
- if self._aggregation == vs.VariableAggregation.MEAN and (
- not self.dtype.is_floating) and isinstance(value, PerReplica):
- raise ValueError(
- "Cannot update non-float variables with "
- "tf.VariableAggregation.MEAN aggregation in replica context. "
- "Either change the variable dtype to float or update it in "
- "cross-replica context.")
+ assert strategy == self.distribute_strategy
+ v = _apply_aggregation(strategy, value, self.aggregation, self)
+ return self._update_cross_replica(update_fn, v, **kwargs)
- v = _apply_aggregation(strategy, value, self._aggregation, self)
- return strategy.extended.update(
- self, update_fn, args=(v,), kwargs=other_kwargs)
-
- return ds_context.get_replica_context().merge_call(
- merge_fn, args=(value,), kwargs=kwargs)
+ return ds_context.get_replica_context().merge_call(
+ merge_fn, args=(value,), kwargs=kwargs)
def assign_sub(self, value, use_locking=False, name=None, read_value=True):
assign_sub_fn = lambda var, *a, **kw: var.assign_sub(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_sub_fn,
value=value,
use_locking=use_locking,
@@ -813,7 +847,7 @@
def assign_add(self, value, use_locking=False, name=None, read_value=True):
assign_add_fn = lambda var, *a, **kw: var.assign_add(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_add_fn,
value=value,
use_locking=use_locking,
@@ -822,7 +856,7 @@
def assign(self, value, use_locking=False, name=None, read_value=True):
assign_fn = lambda var, *a, **kw: var.assign(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=assign_fn,
value=value,
use_locking=use_locking,
@@ -831,7 +865,7 @@
def scatter_sub(self, sparse_delta, use_locking=False, name=None):
scatter_sub_fn = lambda var, *a, **kw: var.scatter_sub(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_sub_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -839,7 +873,7 @@
def scatter_add(self, sparse_delta, use_locking=False, name=None):
scatter_add_fn = lambda var, *a, **kw: var.scatter_add(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_add_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -847,7 +881,7 @@
def scatter_mul(self, sparse_delta, use_locking=False, name=None):
scatter_mul_fn = lambda var, *a, **kw: var.scatter_mul(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_mul_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -855,7 +889,7 @@
def scatter_div(self, sparse_delta, use_locking=False, name=None):
scatter_div_fn = lambda var, *a, **kw: var.scatter_div(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_div_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -870,7 +904,7 @@
"`ONLY_FIRST_REPLICA` aggregation, got: %s" %
self._aggregation)
scatter_min_fn = lambda var, *a, **kw: var.scatter_min(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_min_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -885,7 +919,7 @@
"`ONLY_FIRST_REPLICA` aggregation, got: %s" %
self._aggregation)
scatter_max_fn = lambda var, *a, **kw: var.scatter_max(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_max_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -900,7 +934,7 @@
"`ONLY_FIRST_REPLICA` aggregation, got: %s" %
self._aggregation)
scatter_update_fn = lambda var, *a, **kw: var.scatter_update(*a, **kw)
- return self._mirrored_update(
+ return self._update(
update_fn=scatter_update_fn,
value=sparse_delta,
use_locking=use_locking,
@@ -1209,10 +1243,10 @@
# pylint: disable=protected-access
assert not isinstance(v0, MirroredVariable), (
"ids = %s, values = %s" % ([id(v) for v in values], values))
- distributed_container = v0._distributed_container()
+ distributed_container = v0._distributed_container
assert distributed_container is not None
for v in values[1:]:
- assert distributed_container is v._distributed_container()
+ assert distributed_container is v._distributed_container
return distributed_container
# pylint: enable=protected-access
@@ -1300,7 +1334,7 @@
# DistributedVariable has _distributed_container defined
# but we don't want to return it.
not isinstance(val, DistributedVariable)):
- container = val._distributed_container() # pylint: disable=protected-access
+ container = val._distributed_container # pylint: disable=protected-access
if container is not None:
return container
return val
diff --git a/tensorflow/python/distribute/values_test.py b/tensorflow/python/distribute/values_test.py
index ad29b74..36fc465 100644
--- a/tensorflow/python/distribute/values_test.py
+++ b/tensorflow/python/distribute/values_test.py
@@ -365,7 +365,7 @@
return mirrored
-class RegroupAndSelectDeviceTest(test.TestCase):
+class RegroupAndSelectDeviceTest(test.TestCase, parameterized.TestCase):
def _is_per_replica(self, result, expected, klass=values.PerReplica):
self.assertIsInstance(result, klass)
@@ -375,12 +375,12 @@
def testNested(self):
result = values.regroup((_nested_value("1"), _nested_value("2")))
self.assertIsInstance(result, tuple)
- self.assertEqual(3, len(result))
+ self.assertLen(result, 3)
self._is_per_replica(result[0], ["a1", "a2"])
self._is_per_replica(result[2], ["h1", "h2"])
self.assertIsInstance(result[1], list)
- self.assertEqual(3, len(result[1]))
+ self.assertLen(result[1], 3)
self._is_per_replica(result[1][0], ["b1", "b2"])
self._is_per_replica(result[1][2], ["g1", "g2"])
@@ -416,12 +416,12 @@
result = values.regroup((_nested_value("1"), _nested_value("2")),
values.Mirrored)
self.assertIsInstance(result, tuple)
- self.assertEqual(3, len(result))
+ self.assertLen(result, 3)
self._is_per_replica(result[0], ["a1", "a2"], values.Mirrored)
self._is_per_replica(result[2], ["h1", "h2"], values.Mirrored)
self.assertIsInstance(result[1], list)
- self.assertEqual(3, len(result[1]))
+ self.assertLen(result[1], 3)
self._is_per_replica(result[1][0], ["b1", "b2"], values.Mirrored)
self._is_per_replica(result[1][2], ["g1", "g2"], values.Mirrored)
@@ -444,34 +444,42 @@
def testWrapAListOfTwoTuples(self):
result = values.regroup([("1", "2"), ("3", "4")])
self.assertIsInstance(result, tuple)
- self.assertEqual(2, len(result))
+ self.assertLen(result, 2)
self._is_per_replica(result[0], ("1", "3"), values.PerReplica)
self._is_per_replica(result[1], ("2", "4"), values.PerReplica)
- def testMirroredContainer(self):
- if context.num_gpus() < 1 and context.executing_eagerly():
- self.skipTest("A GPU is not available for this test in eager mode.")
- mirrored = _make_mirrored()
- result = values.regroup(mirrored.values)
- self.assertIs(mirrored, result)
+ @combinations.generate(
+ combinations.combine(
+ distribution=[
+ strategy_combinations.mirrored_strategy_with_gpu_and_cpu,
+ strategy_combinations.mirrored_strategy_with_one_cpu,
+ ],
+ mode=["graph", "eager"],
+ ))
+ def testMirroredContainer(self, distribution):
+ with distribution.scope():
+ v = variable_scope.variable(
+ 1., aggregation=variable_scope.VariableAggregation.SUM)
+ self.assertTrue(values.is_distributed_variable(v))
+ self.assertTrue(values.is_distributed_variable(values.regroup(v.values)))
def testSameId(self):
foo = object()
result = values.regroup((("a", foo), ("b", foo)))
self.assertIsInstance(result, tuple)
- self.assertEqual(2, len(result))
+ self.assertLen(result, 2)
self._is_per_replica(result[0], ["a", "b"])
self.assertIs(foo, result[1])
# Test select_replica(), should undo the merge done by regroup().
result_0 = values.select_replica(0, result)
self.assertIsInstance(result_0, tuple)
- self.assertEqual(2, len(result_0))
+ self.assertLen(result_0, 2)
self.assertEqual("a", result_0[0])
self.assertIs(foo, result_0[1])
result_1 = values.select_replica(1, result)
self.assertIsInstance(result_1, tuple)
- self.assertEqual(2, len(result_1))
+ self.assertLen(result_1, 2)
self.assertEqual("b", result_1[0])
self.assertIs(foo, result_1[1])
@@ -479,18 +487,7 @@
result = values.regroup((_nested_value("1"),))
# On one device regroup() and select_replica() are basically identity.
self.assertEqual(_nested_value("1"), result)
- self.assertEqual(_nested_value("1"),
- values.select_replica(0, result))
-
- # The one exception has to do with MirroredVariables.
- d = "/device:CPU:0"
- with ops.device(d):
- v = variable_scope.get_variable(
- name="v", initializer=1., use_resource=True)
- mirrored = values.MirroredVariable(None, (v,),
- variable_scope.VariableAggregation.SUM)
- result = values.regroup((v,))
- self.assertIs(mirrored, result)
+ self.assertEqual(_nested_value("1"), values.select_replica(0, result))
def testNamedTuple(self):
diff --git a/tensorflow/python/eager/BUILD b/tensorflow/python/eager/BUILD
index 30cc424..345bbae 100644
--- a/tensorflow/python/eager/BUILD
+++ b/tensorflow/python/eager/BUILD
@@ -81,6 +81,7 @@
visibility = ["//tensorflow:internal"],
deps = [
":backprop",
+ ":benchmarks_test_base",
":cancellation",
":context",
":core",
@@ -612,6 +613,14 @@
],
)
+py_library(
+ name = "benchmarks_test_base",
+ srcs = ["benchmarks_test_base.py"],
+ srcs_version = "PY2AND3",
+ visibility = ["//tensorflow:internal"],
+ deps = [":test"],
+)
+
# TODO(b/153582383): Move tf_ops_alwayslink dependency to c_api_tfrt instead.
cuda_py_test(
name = "benchmarks_test",
@@ -620,6 +629,7 @@
tfrt_enabled = True,
deps = [
":backprop",
+ ":benchmarks_test_base",
":context",
":forwardprop",
":function",
@@ -638,6 +648,7 @@
python_version = "PY3",
deps = [
":backprop",
+ ":benchmarks_test_base",
":context",
":forwardprop",
":function",
diff --git a/tensorflow/python/eager/benchmarks_test.py b/tensorflow/python/eager/benchmarks_test.py
index 794a7e2..9630ce0 100644
--- a/tensorflow/python/eager/benchmarks_test.py
+++ b/tensorflow/python/eager/benchmarks_test.py
@@ -40,6 +40,7 @@
from tensorflow.python import pywrap_tfe
from tensorflow.python.eager import backprop # pylint: disable=unused-import
+from tensorflow.python.eager import benchmarks_test_base
from tensorflow.python.eager import context
from tensorflow.python.eager import core
from tensorflow.python.eager import def_function
@@ -106,7 +107,7 @@
return end - start
-class MicroBenchmarks(test.Benchmark):
+class MicroBenchmarks(benchmarks_test_base.MicroBenchmarksBase):
def __init__(self):
# TODO(b/153054118): Add tf.RandomUniform
@@ -145,20 +146,7 @@
return name
def _run(self, func, num_iters, execution_mode=None):
- total_time = run_benchmark(func, num_iters, execution_mode)
- mean_us = total_time * 1e6 / num_iters
- benchmark_name = self._get_benchmark_name()
-
- self.report_benchmark(
- iters=num_iters,
- wall_time=mean_us,
- extras={
- "examples_per_sec":
- float("{0:.3f}".format(num_iters / total_time)),
- "us_per_example":
- float("{0:.3f}".format(total_time * 1e6 / num_iters))
- },
- name=benchmark_name)
+ self.run_report(run_benchmark, func, num_iters, execution_mode)
def benchmark_create_np_array(self):
func = lambda: np.array([3.0])
@@ -166,7 +154,6 @@
def _benchmark_create_tensor(self, value, dtype, device):
"""Benchmark overheads of creating a Tensor object."""
- ctx = context.context()
if device == GPU:
# Warmup the GPU
ops.EagerTensor(value, device=device)
diff --git a/tensorflow/python/eager/benchmarks_test_base.py b/tensorflow/python/eager/benchmarks_test_base.py
new file mode 100644
index 0000000..552d844
--- /dev/null
+++ b/tensorflow/python/eager/benchmarks_test_base.py
@@ -0,0 +1,35 @@
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+r"""Benchmark base to run and report benchmark results."""
+
+from __future__ import absolute_import as _absolute_import
+from __future__ import division as _division
+from __future__ import print_function as _print_function
+
+from tensorflow.python.eager import test
+
+
+class MicroBenchmarksBase(test.Benchmark):
+ """Run and report benchmark results."""
+
+ def run_report(self, run_benchmark, func, num_iters, execution_mode=None):
+ """Run and report benchmark results."""
+ total_time = run_benchmark(func, num_iters, execution_mode)
+ mean_us = total_time * 1e6 / num_iters
+ extras = {
+ "examples_per_sec": float("{0:.3f}".format(num_iters / total_time)),
+ "us_per_example": float("{0:.3f}".format(total_time * 1e6 / num_iters))
+ }
+ self.report_benchmark(iters=num_iters, wall_time=mean_us, extras=extras)
diff --git a/tensorflow/python/eager/def_function.py b/tensorflow/python/eager/def_function.py
index 39f77e7..cdc675d 100644
--- a/tensorflow/python/eager/def_function.py
+++ b/tensorflow/python/eager/def_function.py
@@ -830,6 +830,13 @@
def function_spec(self):
return self._function_spec
+ def pretty_printed_concrete_signatures(self, verbose=True):
+ joiner = "\n\n" if verbose else "\n"
+ return joiner.join([
+ c.pretty_printed_signature(verbose=verbose)
+ for c in self._list_all_concrete_functions()
+ ])
+
def _initialize_uninitialized_variables(self, initializers):
"""Make and call a `ConcreteFunction` which initializes variables."""
@@ -913,12 +920,8 @@
return initialize_variables.get_concrete_function()
- def _list_all_concrete_functions_for_serialization(self):
- """Returns all concrete functions for serialization.
-
- Returns:
- A list of instances of `ConcreteFunction`.
- """
+ def _list_all_concrete_functions(self):
+ """Returns all concrete functions."""
if self.input_signature is not None:
self.get_concrete_function()
concrete_functions = []
@@ -930,6 +933,15 @@
concrete_functions.extend(
self._stateless_fn._function_cache.all_values())
# pylint: enable=protected-access
+ return concrete_functions
+
+ def _list_all_concrete_functions_for_serialization(self):
+ """Returns all concrete functions for serialization.
+
+ Returns:
+ A list of instances of `ConcreteFunction`.
+ """
+ concrete_functions = self._list_all_concrete_functions()
seen_signatures = []
for concrete_function in concrete_functions:
signature = concrete_function.structured_input_signature
diff --git a/tensorflow/python/eager/def_function_xla_jit_test.py b/tensorflow/python/eager/def_function_xla_jit_test.py
index 7ed964e..13b4649 100644
--- a/tensorflow/python/eager/def_function_xla_jit_test.py
+++ b/tensorflow/python/eager/def_function_xla_jit_test.py
@@ -286,7 +286,7 @@
tape.watch(x)
with self.assertRaisesRegexp(
errors.UnimplementedError,
- 'TensorList or Stack crossing the XLA/TF boundary'):
+ 'TensorList crossing the XLA/TF boundary'):
y = f(x)
tape.gradient(y, x)
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index 6a3f99b..57a915a 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -54,6 +54,7 @@
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import tensor_util
+from tensorflow.python.framework import type_spec
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import custom_gradient
@@ -345,7 +346,7 @@
if t.name == compat.as_str(self._func.name):
g = self._func.graph
elif g:
- next_func = g._get_function(t.name)
+ next_func = g._get_function(t.name) # pylint: disable=protected-access
if next_func is not None and isinstance(next_func,
_EagerDefinedFunction):
g = next_func.graph
@@ -1504,6 +1505,12 @@
flat_outputs, self._inference_args, self._input_tangents)
+# Sentinel value used by with ConcreteFunction's structured signature to
+# indicate that a non-tensor parameter should use the value that was
+# specified when the concrete function was created.
+_BOUND_VALUE = object()
+
+
class ConcreteFunction(object):
"""Callable object encapsulating a function definition and its gradient.
@@ -1511,7 +1518,11 @@
is differentiable under `tf.GradientTape` objects.
"""
- def __init__(self, func_graph, attrs=None, shared_func_graph=True):
+ def __init__(self,
+ func_graph,
+ attrs=None,
+ shared_func_graph=True,
+ function_spec=None):
"""Initialize a `ConcreteFunction`.
Args:
@@ -1522,16 +1533,25 @@
shared_func_graph: If False, the ConcreteFunction takes ownership of
`func_graph` and will break reference cycles when it is deleted. This
makes the FuncGraph inoperable.
+ function_spec: FunctionSpec for the original function. If not specified,
+ then this ConcreteFunction may only be called using the flat signature.
Raises:
ValueError: If number of input_placeholders is not equal to the number
of function inputs.
"""
+ # _arg_keywords and _num_positional_args define the flat signature. They
+ # are assigned after construction.
self._arg_keywords = None
self._num_positional_args = None
+
self._func_graph = func_graph
self._captured_inputs = self._func_graph.external_captures
self._captured_closures = self._func_graph.deferred_external_captures
+
+ # function_spec defines the structured signature.
+ self._set_function_spec(function_spec)
+
if attrs and IMPLEMENTS_ATTRIBUTE_NAME in attrs:
# The alternative is to silently drop "implements" tag
# but it seems likely it would lead to hard to catch bugs.
@@ -1581,6 +1601,52 @@
# building gradients.
self._inference_function = self._delayed_rewrite_functions.forward()
+ def _set_function_spec(self, function_spec):
+ """Enables the structured signature by supplying a function_spec."""
+ self._function_spec = None
+ self._pre_initialized_function_spec = function_spec
+
+ # Note: when ConcreteFunctions are built by recreate_function() in
+ # function_deserialization.py, they don't have a structured_input_signature
+ # yet. In that case, _initialize_function_spec() gets called by
+ # _setup_functions_structures() in load.py.
+ if (function_spec is not None and
+ self.structured_input_signature is not None):
+ self._initialize_function_spec()
+
+ def _initialize_function_spec(self):
+ """Updates `self._function_spec` to include varargs and bound variables.
+
+ Adds new positional arguments for any varargs (i.e., for args that are
+ in `structured_input_signature`, but not in the original fullargspec.args).
+
+ Replaces `defaults` and `kwonlydefaults` with the `_BOUND_VALUE`, for
+ all args and kwargs in `structured_input_signature`.
+
+ Sets `varkw` and `varargs` to None.
+ """
+ if self._pre_initialized_function_spec is None:
+ return # e.g., SavedBareConcreteFunction doesn't have function_spec yet.
+ assert not self._function_spec, "already initialized"
+ function_spec = self._pre_initialized_function_spec
+ args = function_spec.fullargspec.args
+ arg_specs, kwarg_specs = self.structured_input_signature
+ fullargspec = tf_inspect.FullArgSpec(
+ args=list(args) +
+ ["<arg{}>".format(i + 1) for i in range(len(args), len(arg_specs))],
+ varargs=None,
+ varkw=None,
+ defaults=[_BOUND_VALUE] * len(arg_specs),
+ kwonlyargs=list(sorted(kwarg_specs)),
+ kwonlydefaults=dict((k, _BOUND_VALUE) for k in kwarg_specs),
+ annotations=function_spec.fullargspec.annotations)
+ self._function_spec = FunctionSpec(
+ fullargspec,
+ function_spec.is_method,
+ function_spec.input_signature,
+ function_spec.is_pure,
+ name=self._func_graph.name)
+
@property
def variables(self):
"""Sequence of variables for this function."""
@@ -1594,15 +1660,44 @@
def __call__(self, *args, **kwargs):
"""Executes the wrapped function.
+ ConcreteFunctions have two signatures:
+
+ * The signature of the original function wrapped by this ConcreteFunction.
+ * A flat signature, where each argument accepts a single Tensor.
+
+ The original function signature is generally preferred, but the flat input
+ signature is supported for backward compatibility.
+
+ ### Original Function Signature
+
+ When calling a ConcreteFunction with the signature of the original function,
+ each argument must match the type or value that was used when the
+ ConcreteFunction's graph was traced. In particular:
+
+ * Tensor arguments (including CompositeTensors, such as RaggedTensor) must
+ have matching `TypeSpec`s.
+ * Non-Tensor arguments (such as booleans or ints) must have equal values.
+ * Nested arguments (such as lists, tuples, or dictionaries) must have the
+ same nesting structure; and each nested value must have a matching type
+ or value.
+
+ The default value for any arguments that were traced with non-Tensor values
+ is the value that was used in the trace. Arguments that were traced with
+ tensor arguments do not have a default value (even if the original function
+ had a default value for that argument).
+
+ ### Flat Signature
+
+ When calling a ConcreteFunction with the flat signature, the arguments
+ correspond to the flattened component tensors of the arguments that were
+ used to construct the ConcreteFunction. Parameter names are assigned based
+ on `TensorSpec.name` (when specified) or the original argument names (with
+ suffixes automatically added for nested arguments or composite tensors with
+ multiple components).
+
Args:
- *args: Tensors or Variables. Positional arguments are only accepted when
- they correspond one-to-one with arguments of the traced Python function.
- **kwargs: Tensors or Variables specified by name. When
- `get_concrete_function` was called to create this `ConcreteFunction`,
- each Tensor input was given a name, defaulting to the name of the Python
- function's argument but possibly overridden by the `name=` argument to
- `tf.TensorSpec`. These names become the argument names for the concrete
- function.
+ *args: Positional arguments to the concrete function.
+ **kwargs: Keyword arguments to the concrete function.
Returns:
The result of applying the TF function on the given Tensors.
@@ -1610,9 +1705,7 @@
Raises:
AssertionError: If this `ConcreteFunction` was not created through
`get_concrete_function`.
- ValueError: If arguments contains anything other than Tensors or
- Variables.
- TypeError: For invalid positional/keyword argument combinations.
+ TypeError: If the arguments do not match the function's signature.
"""
return self._call_impl(args, kwargs)
@@ -1620,40 +1713,174 @@
"""See `__call__` for details."""
with traceme.TraceMe(self._func_graph.name,
tf_function_call="concrete"):
- if self._arg_keywords is None or self._num_positional_args is None:
- raise AssertionError(
- "Tried to call a concrete function obtained from an internal API "
- "through the public interface. Use get_concrete_function instead.")
- if len(args) > self._num_positional_args:
- raise TypeError(
- ("Expected at most {} positional arguments (and the rest keywords, "
- "of {}), got {}. When calling a concrete function, positional "
- "arguments may not be bound to Tensors within nested structures."
- ).format(self._num_positional_args, self._arg_keywords, args))
- args = list(args)
- for keyword in self._arg_keywords[len(args):]:
+ # Construct the list of input tensors: check if the structured signature
+ # applies first; and if not, then use the flat signature.
+ if self._function_spec is not None:
try:
- args.append(kwargs.pop(compat.as_str(keyword)))
- except KeyError:
- specified_keywords = (list(self._arg_keywords[:len(args)])
- + list(kwargs.keys()))
- raise TypeError(
- "Expected argument names {} but got values for {}. Missing: {}."
- .format(
- list(self._arg_keywords),
- specified_keywords,
- list(set(self._arg_keywords) - set(specified_keywords))))
- if kwargs:
- positional_arg_keywords = set(self._arg_keywords[:len(args)])
- for unused_key in kwargs:
- if unused_key in positional_arg_keywords:
- raise TypeError("Got two values for keyword '{}'.".format(
- unused_key))
- raise TypeError("Keyword arguments {} unknown. Expected {}.".format(
- list(kwargs.keys()), list(self._arg_keywords)))
- return self._call_flat(args, self.captured_inputs, cancellation_manager)
+ return self._call_with_structured_signature(args, kwargs,
+ cancellation_manager)
+ except TypeError as structured_err:
+ try:
+ return self._call_with_flat_signature(args, kwargs,
+ cancellation_manager)
+ except TypeError:
+ raise structured_err
- def _filtered_call(self, args, kwargs):
+ return self._call_with_flat_signature(args, kwargs, cancellation_manager)
+
+ def _call_with_flat_signature(self, args, kwargs, cancellation_manager):
+ """Executes the wrapped function with the flat signature.
+
+ Args:
+ args: Positional arguments to the concrete function.
+ kwargs: Keyword arguments to the concrete function.
+ cancellation_manager: A `CancellationManager` that can be used to cancel
+ function invocation.
+
+ Returns:
+ The result of applying the function on the Tensors/Variables contained in
+ `args` and `kwargs`.
+ Raises:
+ TypeError: if `args` and `kwargs` do not match the flat signature of this
+ `ConcreteFunction`.
+ """
+ if len(args) > self._num_positional_args:
+ raise TypeError(
+ "{} takes {} positional arguments but {} were given".format(
+ self._flat_signature_summary(), self._num_positional_args,
+ len(args)))
+ args = list(args)
+ kwargs = dict(kwargs)
+ for keyword in self._arg_keywords[len(args):]:
+ try:
+ args.append(kwargs.pop(compat.as_str(keyword)))
+ except KeyError:
+ specified_keywords = (
+ list(self._arg_keywords[:len(args)]) + list(kwargs.keys()))
+ raise TypeError("{} missing required arguments: {}".format(
+ self._flat_signature_summary(), ", ".join(
+ sorted(set(self._arg_keywords) - set(specified_keywords)))))
+ if kwargs:
+ positional_arg_keywords = set(self._arg_keywords[:len(args)])
+ for unused_key in kwargs:
+ if unused_key in positional_arg_keywords:
+ raise TypeError("{} got two values for argument '{}'".format(
+ self._flat_signature_summary(), unused_key))
+ raise TypeError("{} got unexpected keyword arguments: {}.".format(
+ self._flat_signature_summary(), ", ".join(sorted(kwargs))))
+
+ for i, arg in enumerate(args):
+ if not isinstance(
+ arg, (ops.Tensor, resource_variable_ops.BaseResourceVariable)):
+ raise TypeError("{}: expected argument #{}(zero-based) to be a Tensor; "
+ "got {} ({})".format(self._flat_signature_summary(), i,
+ type(arg).__name__, str(arg)))
+ return self._call_flat(args, self.captured_inputs, cancellation_manager)
+
+ def _call_with_structured_signature(self, args, kwargs, cancellation_manager):
+ """Executes the wrapped function with the structured signature.
+
+ Args:
+ args: Positional arguments to the concrete function.
+ kwargs: Keyword arguments to the concrete function.
+ cancellation_manager: A `CancellationManager` that can be used to cancel
+ function invocation.
+
+ Returns:
+ The result of applying the function on the Tensors/Variables contained in
+ `args` and `kwargs`.
+ Raises:
+ TypeError: if `args` and `kwargs` do not match the structured signature
+ of this `ConcreteFunction`.
+ """
+ args, kwargs = self._function_spec.canonicalize_function_inputs(
+ *args, **kwargs)
+ self._structured_signature_check_missing_args(args, kwargs)
+ self._structured_signature_check_unexpected_args(args, kwargs)
+ self._structured_signature_check_arg_types(args, kwargs)
+ return self._filtered_call(args, kwargs, cancellation_manager)
+
+ def _structured_signature_check_missing_args(self, args, kwargs):
+ """Raises a TypeError if any args are missing."""
+ arg_specs, kwarg_specs = self.structured_input_signature
+ missing_arguments = []
+ for i, (arg, spec) in enumerate(zip(args, arg_specs)):
+ if arg is _BOUND_VALUE and _contains_type_spec(spec):
+ missing_arguments.append(self._function_spec.arg_names[i])
+ for (name, arg) in kwargs.items():
+ if arg is _BOUND_VALUE and _contains_type_spec(kwarg_specs[name]):
+ missing_arguments.append(name)
+ if missing_arguments:
+ raise TypeError("{} missing required arguments: {}".format(
+ self._structured_signature_summary(),
+ ", ".join(sorted(missing_arguments))))
+
+ def _structured_signature_check_unexpected_args(self, args, kwargs):
+ """Raises a TypeError if there are any extra args."""
+ arg_specs, kwarg_specs = self.structured_input_signature
+ if len(args) > len(arg_specs):
+ raise TypeError(
+ "{} takes {} positional arguments but {} were given".format(
+ self._structured_signature_summary(),
+ len(self._function_spec.arg_names), len(args)))
+ if len(kwargs) > len(kwarg_specs):
+ extra_args = set(kwargs) - set(kwarg_specs)
+ raise TypeError("{} got unexpected keyword arguments: {}".format(
+ self._structured_signature_summary(), ", ".join(extra_args)))
+
+ def _structured_signature_check_arg_types(self, args, kwargs):
+ """Raises a TypeError if any args have the wrong type."""
+ # Check argument types
+ arg_specs, kwarg_specs = self.structured_input_signature
+ for i, (arg, spec) in enumerate(zip(args, arg_specs)):
+ name = self._function_spec.arg_names[i]
+ self._structured_signature_check_arg_type(arg, spec, name)
+ for (name, arg) in kwargs.items():
+ self._structured_signature_check_arg_type(arg, kwarg_specs[name], name)
+
+ def _structured_signature_check_arg_type(self, arg, spec, name):
+ """Raise TypeError if `arg`'s type doesn't match `spec`."""
+ if arg is _BOUND_VALUE:
+ return
+
+ # Check the overall nested structure of the argument.
+ try:
+ nest.assert_same_structure(arg, spec, expand_composites=True)
+ except (ValueError, TypeError):
+ try:
+ nest.assert_same_structure(arg, spec, expand_composites=False)
+ expected, got = spec, arg
+ except (ValueError, TypeError):
+ expected, got = _structure_summary(spec), _structure_summary(arg)
+ raise TypeError("{}: argument {} had incorrect type\n"
+ " expected: {}\n got: {}".format(
+ self._structured_signature_summary(), name, expected,
+ got))
+
+ # Check the type for each leaf in the nested structure.
+ arg_pieces = nest.flatten(arg, expand_composites=True)
+ spec_pieces = nest.flatten(spec, expand_composites=True)
+ for (arg_piece, spec_piece) in zip(arg_pieces, spec_pieces):
+ if isinstance(spec_piece, tensor_spec.DenseSpec):
+ # TODO(edloper): Consider calling convert_to_tensor on non-tensor
+ # values here. That would match the behavior of
+ # _call_concrete_function() in function_deserialization.py. If
+ # we do, then we need to change the nest assert_same_structure and
+ # flatten calls above to use shallow variants.
+ tensor_types = (ops.Tensor, resource_variable_ops.BaseResourceVariable)
+ if not isinstance(arg_piece, tensor_types):
+ raise TypeError(
+ "{} expected a Tensor in {}, but got {} value {}".format(
+ self._structured_signature_summary(), name,
+ type(arg_piece).__name__, arg_piece))
+ elif arg_piece is not _BOUND_VALUE and arg_piece != spec_piece:
+ raise TypeError("ConcreteFunction {} was constructed with {} value "
+ "{} in {}, but was called with {} value {}".format(
+ self._structured_signature_summary(),
+ type(spec_piece).__name__, spec_piece, name,
+ type(arg_piece).__name__, arg_piece))
+
+ def _filtered_call(self, args, kwargs, cancellation_manager=None):
"""Executes the function, filtering arguments from the Python function.
Objects aside from Tensors, CompositeTensors, and Variables are ignored.
@@ -1662,6 +1889,8 @@
Args:
args: Canonicalized positional arguments of the Python function.
kwargs: Canonicalized keyword arguments of the Python function.
+ cancellation_manager: (Optional.) A `CancellationManager` that can be
+ used to cancel function invocation.
Returns:
The result of applying the function on the Tensors/Variables contained in
@@ -1671,7 +1900,8 @@
(t for t in nest.flatten((args, kwargs), expand_composites=True)
if isinstance(t, (ops.Tensor,
resource_variable_ops.BaseResourceVariable))),
- self.captured_inputs)
+ captured_inputs=self.captured_inputs,
+ cancellation_manager=cancellation_manager)
def _call_flat(self, args, captured_inputs, cancellation_manager=None):
"""Executes the wrapped function.
@@ -1800,7 +2030,26 @@
@property
def structured_input_signature(self):
- """Returns structured signature of the original function."""
+ """Returns structured signature for this concrete function.
+
+ Returns:
+ A tuple `(args, kwargs)`, where:
+
+ * `args` is a tuple that specifies the expected type or value each for
+ positional argument.
+ * `kwargs` is a dictionary that specifies the expected type or value
+ for each keyword-only argument.
+
+ The type or value for each argument is specified using one of the
+ following:
+
+ * A `tf.TypeSpec`, indicating that a Tensor or other TensorFlow-native
+ value is expected.
+ * A Python value, such as an integer, indicating that an equal value
+ is expected.
+ * A nested structure of `tf.TypeSpec`s and Python values, indicating
+ that a corresponding nested structure is expected.
+ """
return self._func_graph.structured_input_signature
@property
@@ -1987,6 +2236,103 @@
ret.attr[name].CopyFrom(value)
return ret
+ def _structured_signature_summary(self, default_values=False):
+ """Returns a string summarizing this function's structured signature.
+
+ Args:
+ default_values: If true, then include default values in the signature.
+
+ Returns:
+ A `string`.
+ """
+ # Note: we can't just use self._funcion_spec.signature_summary(), because
+ # that would show "_BOUND_VALUE" as the default value for all arguments.
+ assert self._function_spec is not None
+ arg_specs, kwarg_specs = self.structured_input_signature
+ arg_names = list(self._function_spec.arg_names)
+ if default_values:
+ for i in range(len(arg_names)):
+ if not _contains_type_spec(arg_specs[i]):
+ arg_names[i] += "={}".format(arg_specs[i])
+ if kwarg_specs:
+ arg_names.append("*")
+ for name, spec in kwarg_specs.items():
+ arg_names.append(name)
+ if default_values and not _contains_type_spec(spec):
+ arg_names[-1] += "={}".format(spec)
+ signature = "{}({})".format(self._func_graph.name, ", ".join(arg_names))
+
+ return signature
+
+ def _flat_signature_summary(self):
+ """Returns a string summarizing this function's flat signature."""
+ assert self._arg_keywords is not None
+ assert self._num_positional_args is not None
+ arg_names = self._arg_keywords
+ if self._num_positional_args > len(arg_names):
+ arg_names.extend(
+ "<arg{}>".format(i + 1)
+ for i in range(len(arg_names), self._num_positional_args))
+ return "{}({})".format(self._func_graph.name, ", ".join(arg_names))
+
+ def pretty_printed_signature(self, verbose=True):
+ """Returns a string summarizing the signature of this concrete function."""
+ if not verbose:
+ return self._structured_signature_summary(default_values=True)
+
+ def pretty_print_spec(spec):
+ """Returns a string describing the spec for a single argument."""
+ if isinstance(spec, tensor_spec.TensorSpec):
+ return "{} Tensor, shape={}".format(spec.dtype.name, spec.shape)
+ elif nest.is_sequence(spec):
+ pieces = nest.flatten(spec, expand_composites=False)
+ markers = [_Marker("<{}>".format(i + 1)) for i in range(len(pieces))]
+ structure = nest.pack_sequence_as(spec, markers)
+ result = "{}".format(structure)
+ for (marker, piece) in zip(markers, pieces):
+ result += "\n {}: {}".format(marker, pretty_print_spec(piece))
+ return result
+ else:
+ return repr(spec)
+
+ lines = [self._structured_signature_summary(default_values=True)]
+ arg_specs, kwarg_specs = self.structured_input_signature
+ names = list(self._function_spec.arg_names)
+ names.extend(sorted(kwarg_specs))
+ specs = list(arg_specs) + list(kwarg_specs.values())
+ # note: we can skip bound args, since we already displayed thier bound
+ # value in the signature summary.
+ arg_details = []
+ for (name, spec) in zip(names, specs):
+ if _contains_type_spec(spec):
+ arg_details.append(" {}: {}".format(name, pretty_print_spec(spec)))
+ if arg_details:
+ lines.append(" Args:")
+ lines.extend(arg_details)
+ lines.append(" Returns:")
+ lines.append(" {}".format(
+ pretty_print_spec(
+ nest.map_structure(type_spec.type_spec_from_value,
+ self.structured_outputs))))
+
+ return "\n".join(lines)
+
+ def __repr__(self):
+ if self._function_spec is not None:
+ return "<ConcreteFunction {} at 0x{:X}>".format(
+ self.pretty_printed_signature(verbose=False), id(self))
+ elif not (self._num_positional_args is None or self._arg_keywords is None):
+ return "<ConcreteFunction {} at 0x{:X}>".format(
+ self._flat_signature_summary(), id(self))
+ else:
+ return object.__repr__(self)
+
+ def __str__(self):
+ if self._function_spec is not None:
+ return "ConcreteFunction {}".format(self.pretty_printed_signature())
+ else:
+ return self.__repr__()
+
_pywrap_utils.RegisterType("Tensor", ops.Tensor)
_pywrap_utils.RegisterType("EagerTensor", ops.EagerTensor)
@@ -2080,17 +2426,37 @@
kwonlydefaults={},
annotations=fullargspec.annotations)
is_method = tf_inspect.ismethod(python_function)
- return FunctionSpec(fullargspec, is_method, [], {}, input_signature,
- is_pure=is_pure)
- def __init__(self, fullargspec, is_method, args_to_prepend, kwargs_to_include,
- input_signature, is_pure=False):
+ # Get the function's name. Remove functools.partial wrappers if necessary.
+ while isinstance(python_function, functools.partial):
+ python_function = python_function.func
+ name = getattr(python_function, "__name__", "f")
+
+ return FunctionSpec(
+ fullargspec, is_method, input_signature, is_pure=is_pure, name=name)
+
+ def __init__(self,
+ fullargspec,
+ is_method,
+ input_signature,
+ is_pure=False,
+ name=None):
+ """Constructs a FunctionSpec describing a python function.
+
+ Args:
+ fullargspec: `tf_inspect.FullArgSpec` object describing the function.
+ is_method: True if the function is a method.
+ input_signature: a signature of the function (None, if variable)
+ is_pure: if True all input arguments (including variables and constants)
+ will be converted to tensors and no variable changes allowed.
+ name: Name of the function
+ """
self._fullargspec = fullargspec
self._is_method = is_method
self._is_pure = is_pure
- del args_to_prepend
- del kwargs_to_include
- self._default_values = fullargspec.defaults
+
+ # TODO(edloper): Include name when serializing for SavedModel?
+ self._name = name or "f"
if self._is_method:
# Remove `self`: default arguments shouldn't be matched to it.
@@ -2103,21 +2469,21 @@
# A cache mapping from argument name to index, for canonicalizing
# arguments that are called in a keyword-like fashion.
self._args_to_indices = {arg: i for i, arg in enumerate(args)}
- self.arg_names = args
- self.vararg_name = fullargspec.varargs
+ self._arg_names = args
# A cache mapping from arg index to default value, for canonicalization.
- offset = len(args) - len(self._default_values or [])
+ default_values = fullargspec.defaults
+ offset = len(args) - len(default_values or [])
self._arg_indices_to_default_values = {
offset + index: default
- for index, default in enumerate(self._default_values or [])
+ for index, default in enumerate(default_values or [])
}
if input_signature is None:
self._input_signature = None
else:
- if fullargspec.kwonlyargs:
+ if set(fullargspec.kwonlyargs) - set(fullargspec.kwonlydefaults or ()):
raise ValueError("Cannot define a TensorFlow function from a Python "
- "function with keyword arguments when "
+ "function with keyword-only arguments when "
"input_signature is provided.")
if not isinstance(input_signature, (tuple, list)):
@@ -2137,8 +2503,8 @@
return self._is_method
@property
- def args_to_prepend(self):
- return self._args_to_prepend
+ def args_to_indices(self):
+ return self._args_to_indices
@property
def kwargs_to_include(self):
@@ -2152,6 +2518,43 @@
def flat_input_signature(self):
return self._flat_input_signature
+ @property
+ def is_pure(self):
+ return self._is_pure
+
+ @property
+ def arg_names(self):
+ return self._arg_names
+
+ @property
+ def vararg_name(self):
+ return self._fullargspec.varargs
+
+ @property
+ def varkw_name(self):
+ return self._fullargspec.varkw
+
+ def signature_summary(self, default_values=False):
+ """Returns a string summarizing this function's signature.
+
+ Args:
+ default_values: If true, then include default values in the signature.
+
+ Returns:
+ A `string`.
+ """
+ args = list(self._arg_names)
+ if default_values:
+ for (i, default) in self._arg_indices_to_default_values.items():
+ args[i] += "={}".format(default)
+ if self._fullargspec.kwonlyargs:
+ args.append("*")
+ for arg_name in self._fullargspec.kwonlyargs:
+ args.append(arg_name)
+ if default_values and arg_name in self._fullargspec.kwonlydefaults:
+ args[-1] += "={}".format(self._fullargspec.kwonlydefaults[arg_name])
+ return "{}({})".format(self._name, ", ".join(args))
+
def _convert_variables_to_tensors(self, args, kwargs):
args = [ops.convert_to_tensor(x) for x in args]
kwargs = {kw: ops.convert_to_tensor(x) for kw, x in kwargs.items()}
@@ -2164,7 +2567,13 @@
instance. In particular, we parse the varags and kwargs that the
original function was called with into a tuple corresponding to the
Python function's positional (named) arguments and a dictionary
- corresponding to its kwargs.
+ corresponding to its kwargs. Missing default arguments are added.
+
+ If this `FunctionSpec` has an input signature, then it is used to convert
+ arguments to tensors; otherwise, any inputs containing numpy arrays are
+ converted to tensors.
+
+ Additionally, any inputs containing numpy arrays are converted to Tensors.
Args:
*args: The varargs this object was called with.
@@ -2185,29 +2594,38 @@
args, kwargs = self._convert_variables_to_tensors(args, kwargs)
if self._input_signature is not None:
if len(args) > len(self._input_signature):
- raise TypeError(
- "When input_signature is provided, only pass arguments "
- "covered by it. Received %d argument(s)." % len(args))
+ raise TypeError("{} takes {} positional arguments (as specified by the "
+ "input_signature) but {} were given".format(
+ self.signature_summary(),
+ len(self._input_signature), len(args)))
for arg in six.iterkeys(kwargs):
index = self._args_to_indices.get(arg, None)
if index is None:
- raise TypeError(
- "Function got an unexpected keyword argument %s" % arg)
+ raise TypeError("{} got unexpected keyword argument `{}`".format(
+ self.signature_summary(), arg))
if index >= len(self._input_signature):
raise TypeError(
- "When input_signature is provided, only pass arguments "
- "covered by it. Received argument %s." % arg)
+ "{} got keyword argument `{}` that was not included in "
+ "input_signature".format(self.signature_summary(), arg))
if not kwargs:
inputs = args
- default_keys = sorted(self._arg_indices_to_default_values.keys())
- if default_keys:
- assert min(default_keys) <= len(
- args), "Not enough arguments (%s, %s, %s)" % (args, default_keys,
- self.arg_names)
- for index in default_keys:
- if index >= len(args):
- inputs += (self._arg_indices_to_default_values[index],)
+ if self._arg_indices_to_default_values:
+ try:
+ inputs += tuple(
+ self._arg_indices_to_default_values[i]
+ for i in range(len(args), len(self._arg_names)))
+ except KeyError:
+ missing_args = [
+ self._arg_names[i]
+ for i in range(len(args), len(self._arg_names))
+ if i not in self._arg_indices_to_default_values
+ ]
+ raise TypeError("{} missing required arguments: {}".format(
+ self.signature_summary(), ", ".join(missing_args)))
+
+ if self._fullargspec.kwonlydefaults:
+ kwargs.update(self._fullargspec.kwonlydefaults)
else:
# Maps from index of arg to its corresponding value, according to `args`
# and `kwargs`; seeded with the default values for the named args that
@@ -2220,18 +2638,28 @@
for arg, value in six.iteritems(kwargs):
index = self._args_to_indices.get(arg, None)
if index is not None:
+ if index < len(args):
+ raise TypeError("{} got two values for argument '{}'".format(
+ self.signature_summary(), arg))
arg_indices_to_values[index] = value
consumed_args.append(arg)
- elif self._input_signature is not None:
- raise ValueError("Cannot define a TensorFlow function from a Python "
- "function with keyword arguments when "
- "input_signature is provided.")
for arg in consumed_args:
- # After this loop, `kwargs` will only contain true keyword arguments, as
- # opposed to named arguments called in a keyword-like fashion.
+ # After this loop, `kwargs` will only contain keyword_only arguments,
+ # and all positional_or_keyword arguments have been moved to `inputs`.
kwargs.pop(arg)
inputs = args + _deterministic_dict_values(arg_indices_to_values)
+ if kwargs and self._input_signature is not None:
+ raise TypeError(
+ "{} got unexpected keyword arguments: {}\n(Cannot define a "
+ "TensorFlow function from a Python function with keyword arguments "
+ "when input_signature is provided.)".format(
+ self.signature_summary(), ", ".join(kwargs)))
+
+ if self._fullargspec.kwonlydefaults:
+ for (kwarg, default) in self._fullargspec.kwonlydefaults.items():
+ kwargs.setdefault(kwarg, default)
+
if self._input_signature is None:
inputs = _convert_numpy_inputs(inputs)
kwargs = _convert_numpy_inputs(kwargs)
@@ -2258,7 +2686,9 @@
resource_variable_ops.is_resource_variable(value)
or tensor_util.is_tensor(value)
# For legacy reasons we do not automatically promote Numpy strings.
- or isinstance(value, np.str_))
+ or isinstance(value, np.str_)
+ # NumPy dtypes have __array__ as unbound methods.
+ or isinstance(value, type))
def _convert_numpy_inputs(inputs):
@@ -2472,6 +2902,7 @@
graph_function, _, _ = self._maybe_define_function(args, kwargs)
return graph_function
+ # XX TODO: make sure we fix up this path as well!?
def _get_concrete_function_internal(self, *args, **kwargs):
"""Bypasses error checking when getting a graph function."""
graph_function = self._get_concrete_function_internal_garbage_collected(
@@ -2689,6 +3120,7 @@
override_flat_arg_shapes=override_flat_arg_shapes,
capture_by_value=self._capture_by_value),
self._function_attributes,
+ function_spec=self.function_spec,
# Tell the ConcreteFunction to clean up its graph once it goes out of
# scope. This is not the default behavior since it gets used in some
# places (like Keras) where the FuncGraph lives longer than the
@@ -3375,3 +3807,30 @@
func_graph_module.dismantle_func_graph(self._func_graph)
except: # pylint: disable=bare-except
pass
+
+
+class _Marker(object):
+ """Markers used to pretty-print nested args in function signatures."""
+
+ def __init__(self, s):
+ self._s = s
+
+ def __repr__(self):
+ return str(self._s)
+
+
+def _structure_summary(structure):
+ """Displays a summary of the nesting structure of the given value."""
+
+ def type_name(x):
+ if isinstance(x, type_spec.TypeSpec):
+ return x.value_type.__name__
+ else:
+ return type(x).__name__
+
+ markers = [_Marker(type_name(v)) for v in nest.flatten(structure)]
+ return str(nest.pack_sequence_as(structure, markers))
+
+
+def _contains_type_spec(value):
+ return any(isinstance(x, type_spec.TypeSpec) for x in nest.flatten(value))
diff --git a/tensorflow/python/eager/function_test.py b/tensorflow/python/eager/function_test.py
index be6524c..fd66871 100644
--- a/tensorflow/python/eager/function_test.py
+++ b/tensorflow/python/eager/function_test.py
@@ -36,6 +36,7 @@
from tensorflow.python.eager import context
from tensorflow.python.eager import def_function
from tensorflow.python.eager import function
+from tensorflow.python.framework import composite_tensor
from tensorflow.python.framework import config
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
@@ -50,6 +51,7 @@
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import test_ops
from tensorflow.python.framework import test_util
+from tensorflow.python.framework import type_spec
from tensorflow.python.layers import convolutional
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import check_ops
@@ -65,6 +67,7 @@
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import random_ops
from tensorflow.python.ops import resource_variable_ops
+from tensorflow.python.ops import string_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.ops import variables
from tensorflow.python.ops.ragged import ragged_factory_ops
@@ -99,6 +102,16 @@
constant_op.constant([1, 2]), constant_op.constant([0, 1]))
+def _spec_for_value(value):
+ """Returns the (nested) TypeSpec for a value."""
+ if nest.is_sequence(value):
+ return nest.map_structure(_spec_for_value, value)
+ elif isinstance(value, (ops.Tensor, composite_tensor.CompositeTensor)):
+ return type_spec.type_spec_from_value(value)
+ else:
+ return value
+
+
class FunctionTest(test.TestCase, parameterized.TestCase):
def setUp(self):
@@ -791,6 +804,16 @@
# We should not have triggered any re-tracing of the python function.
self.assertLen(total_function_cache(defined), 2)
+ def testNumpyDtypeInputSupported(self):
+ @function.defun
+ def f(x, dtype):
+ return constant_op.constant(dtype(x))
+
+ self.assertEqual(f(1, numpy.float32).numpy(), numpy.float32(1))
+ self.assertEqual(f(2, numpy.float32).numpy(), numpy.float32(2))
+ self.assertEqual(f(1, numpy.int32).numpy(), numpy.int32(1))
+ self.assertEqual(f(2, numpy.int32).numpy(), numpy.int32(2))
+
def testDefunNumpyArraysConvertedToTensorsInKwargs(self):
def f(**kwargs):
@@ -1812,6 +1835,18 @@
with self.assertRaisesRegexp(ValueError, 'incompatible'):
func([['wrong dtype']])
+ def testNoKeywordOnlyArgumentsWithInputSignature(self):
+ if sys.version_info[0] < 3:
+ self.skipTest('keyword_only arguments only exist in Python 3.')
+
+ func = eval('lambda x, *, y: x') # pylint: disable=eval-used
+ signature = [tensor_spec.TensorSpec(None, dtypes.int32)]
+ with self.assertRaisesRegexp(
+ ValueError, 'Cannot define a TensorFlow function from a Python '
+ 'function with keyword-only arguments when input_signature is '
+ 'provided.'):
+ def_function.function(func, signature)
+
def testNestedInputSignatures(self):
def expected_foo(a, b):
@@ -1928,7 +1963,9 @@
defined(array_ops.ones([2, 1]))
# Wrong number of arguments.
- with self.assertRaisesRegexp(TypeError, r'Received 2 argument\(s\)'):
+ with self.assertRaisesRegexp(
+ TypeError, r'takes 1 positional arguments \(as specified by the '
+ r'input_signature\) but 2 were given'):
defined(array_ops.ones([2]), array_ops.ones([2]))
with self.assertRaisesRegexp(ValueError,
'Structure of Python function inputs.*'):
@@ -1969,10 +2006,14 @@
return -1.0 * a
x = constant_op.constant(1.0)
- with self.assertRaisesRegexp(TypeError, 'only pass arguments'):
+ with self.assertRaisesRegexp(
+ TypeError, 'got keyword argument `training` '
+ 'that was not included in input_signature'):
foo(x, training=True)
- with self.assertRaisesRegexp(TypeError, 'only pass arguments'):
+ with self.assertRaisesRegexp(
+ TypeError, 'got keyword argument `training` '
+ 'that was not included in input_signature'):
foo(x, training=False)
self.assertAllEqual(x.numpy(), foo(x).numpy())
@@ -2495,8 +2536,7 @@
return x
graph_function = foo.get_concrete_function(constant_op.constant(1.0))
- with self.assertRaisesRegexp(
- ValueError, 'All inputs to `ConcreteFunction`s must be Tensors;.*'):
+ with self.assertRaises((TypeError, ValueError)):
graph_function('Not a Tensor.')
def testSwapImplementationWithGrapplerPlugin(self):
@@ -3171,6 +3211,432 @@
function.clear_function_callbacks()
self.assertEmpty(function._function_callbacks) # pylint:disable=protected-access
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionWithNestedTensorInputs(self):
+
+ @def_function.function
+ def f(x, y):
+ return (x['a'] + x['b'], y[0] + y[1])
+
+ a = constant_op.constant(1000)
+ b = constant_op.constant(200)
+ c = constant_op.constant(30)
+ d = {'a': a, 'b': b}
+ e = (c, 4)
+
+ # Test different argument signatures when constructing the concrete func.
+ for cf in [
+ f.get_concrete_function(d, e),
+ f.get_concrete_function(d, y=e),
+ f.get_concrete_function(y=e, x=d),
+ f.get_concrete_function(_spec_for_value(d), _spec_for_value(e)),
+ f.get_concrete_function(_spec_for_value(d), y=_spec_for_value(e)),
+ f.get_concrete_function(y=_spec_for_value(e), x=_spec_for_value(d))
+ ]:
+ # Test different calling conventions when calling the concrete func.
+ for output in [
+ cf(d, e), # structured signature
+ cf(d, y=e), # structured signature w/ kwarg
+ cf(y=e, x=d), # structured signature w/ 2 kwargs
+ cf(a, b, c), # flat signature
+ cf(x=a, x_1=b, y=c) # flat signature w/ kwargs
+ ]:
+ self.assertIsInstance(output, tuple)
+ self.assertLen(output, 2)
+ self.assertAllEqual(output[0], 1200)
+ self.assertAllEqual(output[1], 34)
+
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionWithNestedNonTensorInputs(self):
+
+ @def_function.function
+ def f(x, y):
+ return (x['a'] + x['b'], y[0] + y[1])
+
+ a = {'a': constant_op.constant(1000), 'b': constant_op.constant(200)}
+ b = (50, 3)
+
+ for cf in [ # argument y is bound to non-Tensor value (50, 3).
+ f.get_concrete_function(a, b),
+ f.get_concrete_function(a, y=b),
+ f.get_concrete_function(x=a, y=b)
+ ]:
+ for output in [cf(a), cf(x=a), cf(a, b), cf(x=a, y=b)]:
+ self.assertAllEqual(output[0] + output[1], 1253)
+
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionWithBoundNestedNonTensorInputs(self):
+
+ @def_function.function
+ def f(x, y):
+ return (x['a'] + x['b'], y[0] + y[1])
+
+ a = {'a': 3000, 'b': 200, 'c': 9000}
+ b = (constant_op.constant(30), 4)
+
+ for cf in [ # argument x is bound to non-tensor value `a`
+ f.get_concrete_function(a, b),
+ f.get_concrete_function(a, y=b),
+ f.get_concrete_function(x=a, y=b)
+ ]:
+ for output in [cf(a, b), cf(a, y=b), cf(y=b), cf(x=a, y=b)]:
+ self.assertAllEqual(output[0] + output[1], 3234)
+
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionWithAllBoundNestedNonTensorInputs(self):
+
+ @def_function.function
+ def f(x, y):
+ return (x['a'] + x['b'], y[0] + y[1])
+
+ a = {'a': 5000, 'b': 500}
+ b = (50, 5)
+
+ cf = f.get_concrete_function(a, b)
+ for output in [cf(), cf(a), cf(y=b)]:
+ self.assertAllEqual(output[0] + output[1], 5555)
+
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionStructuredSignatureKeywordOrder(self):
+ # Check that keyword-only arguments are sorted appropriately, so that they
+ # feed the right tensor into each input.
+ @def_function.function
+ def g(**kwargs):
+ return string_ops.reduce_join(
+ string_ops.reduce_join(
+ ops.convert_to_tensor(sorted(kwargs.items())),
+ axis=1,
+ separator='='),
+ axis=0,
+ separator=', ')
+
+ s = constant_op.constant('s')
+ g.get_concrete_function(q=s, a=s, p=s, r=s, v=s, m=s, l=s)
+ self.assertAllEqual(
+ g(m='a', r='b', v='c', q='d', l='e', a='f', p='g'),
+ b'a=f, l=e, m=a, p=g, q=d, r=b, v=c')
+ self.assertAllEqual(
+ g(q='d', a='f', p='g', r='b', v='c', m='a', l='e'),
+ b'a=f, l=e, m=a, p=g, q=d, r=b, v=c')
+ self.assertAllEqual(
+ g(a='f', l='e', m='a', p='g', q='d', r='b', v='c'),
+ b'a=f, l=e, m=a, p=g, q=d, r=b, v=c')
+
+ # pylint: disable=g-long-lambda
+ @parameterized.named_parameters([
+ dict(
+ testcase_name='MissingArg',
+ conc_args=lambda: (1, constant_op.constant(2)),
+ call_args=lambda: (1,),
+ error=r'func\(x, y\) missing required arguments: y'),
+ dict(
+ testcase_name='MissingVararg',
+ conc_args=lambda: (1, 2, constant_op.constant(1.0)),
+ call_args=lambda: (1, 2),
+ error=r'func\(x, y, <arg3>\) missing required arguments: <arg3>'),
+ dict(
+ testcase_name='ExtraPositionalArg',
+ conc_args=lambda: (1, 2),
+ call_args=lambda: (1, 2, 3),
+ error=r'func\(x, y\) takes 2 positional arguments but 3 were given'),
+ dict(
+ testcase_name='MissingKeywordOnlyArg',
+ conc_args=lambda: (1, 2),
+ conc_kwargs=lambda: {'c': constant_op.constant(1.0)},
+ call_args=lambda: (1, 2),
+ error=r'func\(x, y, \*, c\) missing required arguments: c'),
+ dict(
+ testcase_name='ExtraKeywordArg',
+ conc_args=lambda: (1, 2),
+ call_args=lambda: (1, 2),
+ call_kwargs=lambda: {'c': constant_op.constant(1.0)},
+ error=r'func\(x, y\) got unexpected keyword arguments: c'),
+ dict(
+ testcase_name='ExpectedRaggedGotNest',
+ conc_args=lambda: (ragged_factory_ops.constant([[1, 2], [3]]),),
+ call_args=lambda: ({
+ 'a': constant_op.constant([1, 2, 3])
+ },),
+ error=r'func\(x, y\): argument x had incorrect type\n'
+ r' expected: RaggedTensor\n'
+ r" got: {'a': (Eager)?Tensor}"),
+ dict(
+ testcase_name='WrongRaggedRank',
+ conc_args=lambda: (ragged_factory_ops.constant([[1, 2], [3]]),),
+ call_args=lambda: (ragged_factory_ops.constant([[[1]]]),),
+ error=r'func\(x, y\): argument x had incorrect type\n'),
+ dict(
+ testcase_name='WrongRaggedDType',
+ conc_args=lambda: (ragged_factory_ops.constant([[1]]),),
+ call_args=lambda: (ragged_factory_ops.constant([[1.0]]),),
+ error=r'func\(x, y\): argument x had incorrect type\n'),
+ dict(
+ testcase_name='ExpectedDictGotTensor',
+ conc_args=lambda: ({
+ 'a': constant_op.constant(1),
+ 'b': constant_op.constant(1)
+ },),
+ call_args=lambda: (constant_op.constant(1),),
+ error=r'func\(x, y\): argument x had incorrect type\n'),
+ dict(
+ testcase_name='ExpectedTupleGotTensor',
+ conc_args=lambda:
+ ((constant_op.constant(1), constant_op.constant(2)),),
+ call_args=lambda: (constant_op.constant(1),),
+ error=r'func\(x, y\): argument x had incorrect type\n'),
+ dict(
+ testcase_name='WrongDType',
+ conc_args=lambda: (constant_op.constant(1),),
+ call_args=lambda: (constant_op.constant(1.0),),
+ exception=(ValueError, errors.InvalidArgumentError,
+ # on xla_gpu, we get InternalError instead.
+ errors.InternalError)),
+ dict(
+ testcase_name='ExpectedTensorGotInt',
+ conc_args=lambda: (constant_op.constant(1),),
+ call_args=lambda: (5,),
+ error=r'func\(x, y\) expected a Tensor in x, but got int value 5'),
+ dict(
+ testcase_name='ExpectedIntGotDifferentInt',
+ conc_args=lambda: (5,),
+ call_args=lambda: (8,),
+ error=r'ConcreteFunction func\(x, y\) was constructed with int '
+ r'value 5 in x, but was called with int value 8'),
+ dict(
+ testcase_name='ExpectedIntGotTensor',
+ conc_args=lambda: (5,),
+ call_args=lambda: (constant_op.constant(6),),
+ error=r'ConcreteFunction func\(x, y\) was constructed with int '
+ 'value 5 in x, but was called with (Eager)?Tensor value .*'),
+ dict(
+ testcase_name='TwoValuesForArgument',
+ conc_args=lambda: (1, 2),
+ call_args=lambda: (1, 2),
+ call_kwargs=lambda: {'x': 3},
+ error=r"func\(x, y\) got two values for argument 'x'"),
+ ])
+ # pylint: enable=g-long-lambda
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionStructuredSignatureError(self,
+ conc_args=(),
+ conc_kwargs=None,
+ call_args=(),
+ call_kwargs=None,
+ error='.*',
+ exception=TypeError):
+ """Tests for errors in the structrued signature.
+
+ Args:
+ conc_args: Positional arguments used for get_concrete_function.
+ conc_kwargs: Keyword arguments used for get_concrete_function.
+ call_args: Positional arguments used to call the function.
+ call_kwargs: Keyword arguments used to call the function.
+ error: Expected exception message.
+ exception: Expected exception type.
+ """
+ conc_args = conc_args() if callable(conc_args) else conc_args
+ conc_kwargs = conc_kwargs() if callable(conc_kwargs) else conc_kwargs or {}
+ call_args = call_args() if callable(call_args) else call_args
+ call_kwargs = call_kwargs() if callable(call_kwargs) else call_kwargs or {}
+ self.assertIsInstance(conc_args, tuple)
+ self.assertIsInstance(call_args, tuple)
+ self.assertIsInstance(conc_kwargs, dict)
+ self.assertIsInstance(call_kwargs, dict)
+
+ @def_function.function
+ def func(x, y=5, *varargs, **kwargs): # pylint: disable=keyword-arg-before-vararg
+ del y, varargs, kwargs
+ return x
+
+ conc = func.get_concrete_function(*conc_args, **conc_kwargs)
+ with self.assertRaisesRegexp(exception, error):
+ self.evaluate(conc(*call_args, **call_kwargs))
+
+ # pylint: disable=g-long-lambda
+ @parameterized.named_parameters([
+ dict(
+ testcase_name='MissingArg',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ call_args=lambda: (constant_op.constant(1),),
+ error=r'func\(x, y\) missing required arguments: y'),
+ dict(
+ testcase_name='TwoValuesForArg',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ call_args=lambda: (constant_op.constant(1),),
+ call_kwargs=lambda: {
+ 'x': constant_op.constant(1),
+ 'y': constant_op.constant(1)
+ },
+ error=r"func\(x, y\) got two values for argument 'x'"),
+ dict(
+ testcase_name='ExtraPositionalArg',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ call_args=lambda: (constant_op.constant(1), constant_op.constant(2),
+ constant_op.constant(3)),
+ error=r'func\(x, y\) takes 2 positional arguments but 3 were given'),
+ dict(
+ testcase_name='UnexpectedKeywordArg',
+ conc_args=lambda: (constant_op.constant(1),),
+ call_args=lambda: (constant_op.constant(1),),
+ call_kwargs=lambda: {'c': constant_op.constant(1)},
+ error=r'func\(x\) got unexpected keyword arguments: c'),
+ dict(
+ testcase_name='MissingVararg',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2),
+ constant_op.constant(3)),
+ call_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ error=r'func\(x, y, varargs_0\) missing required '
+ r'arguments: varargs_0'),
+ dict(
+ testcase_name='MissingKeywordArg',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ conc_kwargs=lambda: {'c': constant_op.constant(1)},
+ call_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ error=r'func\(x, y, c\) missing required arguments: c'),
+ dict(
+ testcase_name='ExpectedTensorGotInt',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ call_args=lambda: (5, constant_op.constant(2)),
+ error=r'func\(x, y\): expected argument #0\(zero-based\) to be '
+ r'a Tensor; got int \(5\)'),
+ dict(
+ testcase_name='WrongDType',
+ conc_args=lambda: (constant_op.constant(1),),
+ call_args=lambda: (constant_op.constant(1.0),),
+ exception=(ValueError, errors.InvalidArgumentError,
+ # on xla_gpu, we get InternalError instead.
+ errors.InternalError)),
+ dict(
+ testcase_name='MissingKeywordArgNestPiece',
+ conc_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ conc_kwargs=lambda: {'c': ragged_factory_ops.constant([[1]])},
+ call_args=lambda: (constant_op.constant(1), constant_op.constant(2)),
+ call_kwargs=lambda: {'c': constant_op.constant(1)},
+ error=r'func\(x, y, c, c_1\) missing required arguments: c_1'),
+ ])
+ # pylint: enable=g-long-lambda
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionFlatSignatureError(self,
+ conc_args=(),
+ conc_kwargs=None,
+ call_args=(),
+ call_kwargs=None,
+ error='.*',
+ exception=TypeError):
+ """Tests for errors in the flat signature.
+
+ Args:
+ conc_args: Positional arguments used for get_concrete_function.
+ conc_kwargs: Keyword arguments used for get_concrete_function.
+ call_args: Positional arguments used to call the function.
+ call_kwargs: Keyword arguments used to call the function.
+ error: Expected exception message.
+ exception: Expected exception type.
+ """
+ conc_args = conc_args() if callable(conc_args) else conc_args
+ conc_kwargs = conc_kwargs() if callable(conc_kwargs) else conc_kwargs or {}
+ call_args = call_args() if callable(call_args) else call_args
+ call_kwargs = call_kwargs() if callable(call_kwargs) else call_kwargs or {}
+ self.assertIsInstance(conc_args, tuple)
+ self.assertIsInstance(call_args, tuple)
+ self.assertIsInstance(conc_kwargs, dict)
+ self.assertIsInstance(call_kwargs, dict)
+
+ @def_function.function
+ def func(x, y=5, *varargs, **kwargs): # pylint: disable=keyword-arg-before-vararg
+ del y, varargs, kwargs
+ return x
+
+ conc = func.get_concrete_function(*conc_args, **conc_kwargs)
+
+ # Remove _function_spec, to disable the structured signature.
+ conc._set_function_spec(None) # pylint: disable=protected-access
+
+ with self.assertRaisesRegexp(exception, error):
+ self.evaluate(conc(*call_args, **call_kwargs))
+
+ @test_util.run_in_graph_and_eager_modes
+ def testConcreteFunctionAmbiguousSignature(self):
+ # When both the flat & structured signatures are applicable, but they
+ # give different results, we use the structured signature. Note: we expect
+ # this to be extremely rare.
+ @def_function.function
+ def f(x, y):
+ return x * 10 + y
+
+ conc = f.get_concrete_function(
+ x=tensor_spec.TensorSpec(None, dtypes.int32, name='y'),
+ y=tensor_spec.TensorSpec(None, dtypes.int32, name='x'))
+
+ result = conc(x=constant_op.constant(5), y=constant_op.constant(6))
+ self.assertAllEqual(result, 56)
+
+ def testPrettyPrintedSignature(self):
+
+ @def_function.function
+ def func(x, kangaroo=None, octopus=7):
+ del octopus, kangaroo
+ return x
+
+ scalar = constant_op.constant(5)
+ vector = constant_op.constant([10, 10, 20])
+ ragged = ragged_factory_ops.constant([[10, 20], [40]])
+
+ c1 = func.get_concrete_function(scalar, vector)
+ c1_summary = r'func\(x, kangaroo, octopus=7\)'
+ c1_details = (r' Args:\n'
+ r' x: int32 Tensor, shape=\(\)\n'
+ r' kangaroo: int32 Tensor, shape=\(3,\)\n'
+ r' Returns:\n'
+ r' int32 Tensor, shape=\(\)')
+ self.assertRegexpMatches(
+ c1.pretty_printed_signature(verbose=False), c1_summary)
+ self.assertRegexpMatches(
+ c1.pretty_printed_signature(verbose=True),
+ c1_summary + '\n' + c1_details)
+ self.assertRegexpMatches(
+ repr(c1), r'<ConcreteFunction func\(x, kangaroo, octopus=7\) at .*>')
+ self.assertRegexpMatches(
+ str(c1), 'ConcreteFunction {}\n{}'.format(c1_summary, c1_details))
+
+ c2 = func.get_concrete_function(scalar, ragged, 3)
+ c2_summary = r'func\(x, kangaroo, octopus=3\)'
+ c2_details = (r' Args:\n'
+ r' x: int32 Tensor, shape=\(\)\n'
+ r' kangaroo: RaggedTensorSpec\(.*\)\n'
+ r' Returns:\n'
+ r' int32 Tensor, shape=\(\)')
+ self.assertRegexpMatches(c2.pretty_printed_signature(),
+ c2_summary + '\n' + c2_details)
+
+ c3 = func.get_concrete_function({'a': scalar, 'b': [ragged, ragged]})
+ c3_summary = r'func\(x, kangaroo=None, octopus=7\)'
+ c3_details = (r' Args:\n'
+ r" x: {'a': <1>, 'b': \[<2>, <3>\]}\n"
+ r' <1>: int32 Tensor, shape=\(\)\n'
+ r' <2>: RaggedTensorSpec\(.*\)\n'
+ r' <3>: RaggedTensorSpec\(.*\)\n'
+ r' Returns:\n'
+ r" {'a': <1>, 'b': \[<2>, <3>\]}\n"
+ r' <1>: int32 Tensor, shape=\(\)\n'
+ r' <2>: RaggedTensorSpec\(.*\)\n'
+ r' <3>: RaggedTensorSpec\(.*\)')
+ self.assertRegexpMatches(c3.pretty_printed_signature(),
+ c3_summary + '\n' + c3_details)
+
+ # pylint: disable=keyword-arg-before-vararg
+ @def_function.function
+ def func2(x, y=3, *args, **kwargs):
+ return (x, y, args, kwargs)
+
+ c4 = func2.get_concrete_function(scalar, 4, 5, a=scalar)
+ c4_summary = 'func2(x, y=4, <arg3>=5, *, a)'
+ self.assertEqual(c4.pretty_printed_signature(verbose=False), c4_summary)
+
+ c5 = func2.get_concrete_function(8, vector)
+ c5_summary = 'func2(x=8, y)'
+ self.assertEqual(c5.pretty_printed_signature(verbose=False), c5_summary)
+
class MultiDeviceTest(test.TestCase, parameterized.TestCase):
diff --git a/tensorflow/python/feature_column/BUILD b/tensorflow/python/feature_column/BUILD
index 4e26c52..c9f923b 100644
--- a/tensorflow/python/feature_column/BUILD
+++ b/tensorflow/python/feature_column/BUILD
@@ -80,16 +80,20 @@
"//tensorflow/python:sparse_tensor",
"//tensorflow/python:string_ops",
"//tensorflow/python:tensor_shape",
+ "//tensorflow/python:tf_export",
"//tensorflow/python:training",
"//tensorflow/python:util",
"//tensorflow/python:variable_scope",
"//tensorflow/python:variables",
"//tensorflow/python/eager:context",
- "//tensorflow/python/keras:base_layer",
- "//tensorflow/python/keras:engine",
+ "//tensorflow/python/keras:backend",
"//tensorflow/python/keras:initializers",
+ "//tensorflow/python/keras/engine",
+ "//tensorflow/python/keras/engine:base_layer",
+ "//tensorflow/python/keras/layers",
"//tensorflow/python/keras/utils:generic_utils",
"//tensorflow/python/training/tracking",
+ "//tensorflow/python/training/tracking:data_structures",
"//third_party/py/numpy",
"@six_archive//:six",
],
@@ -295,8 +299,8 @@
srcs = ["save_test.py"],
python_version = "PY3",
deps = [
+ ":feature_column_v2",
"//tensorflow/python:client_testlib",
- "//tensorflow/python/feature_column:feature_column_v2",
"//tensorflow/python/keras",
"//tensorflow/python/keras:combinations",
"//third_party/py/numpy",
diff --git a/tensorflow/python/feature_column/feature_column.py b/tensorflow/python/feature_column/feature_column.py
index f538253..87420d0 100644
--- a/tensorflow/python/feature_column/feature_column.py
+++ b/tensorflow/python/feature_column/feature_column.py
@@ -1797,7 +1797,26 @@
`__gt__` is called when the "other" object being compared during the sort
does not have `__lt__` defined.
- Example: http://gpaste/4803354716798976
+ Example:
+ ```
+ # __lt__ only class
+ class A():
+ def __lt__(self, other): return str(self) < str(other)
+
+ a = A()
+ a < "b" # True
+ "0" < a # Error
+
+ # __lt__ and __gt__ class
+ class B():
+ def __lt__(self, other): return str(self) < str(other)
+ def __gt__(self, other): return str(self) > str(other)
+
+ b = B()
+ b < "c" # True
+ "0" < b # True
+ ```
+
Args:
other: The other object to compare to.
diff --git a/tensorflow/python/feature_column/feature_column_v2.py b/tensorflow/python/feature_column/feature_column_v2.py
index ab9866b..9e7df66 100644
--- a/tensorflow/python/feature_column/feature_column_v2.py
+++ b/tensorflow/python/feature_column/feature_column_v2.py
@@ -2254,7 +2254,25 @@
`__gt__` is called when the "other" object being compared during the sort
does not have `__lt__` defined.
- Example: http://gpaste/4803354716798976
+ Example:
+ ```
+ # __lt__ only class
+ class A():
+ def __lt__(self, other): return str(self) < str(other)
+
+ a = A()
+ a < "b" # True
+ "0" < a # Error
+
+ # __lt__ and __gt__ class
+ class B():
+ def __lt__(self, other): return str(self) < str(other)
+ def __gt__(self, other): return str(self) > str(other)
+
+ b = B()
+ b < "c" # True
+ "0" < b # True
+ ```
Args:
other: The other object to compare to.
diff --git a/tensorflow/python/framework/constant_op.py b/tensorflow/python/framework/constant_op.py
index 9736bb8..af9a0f7 100644
--- a/tensorflow/python/framework/constant_op.py
+++ b/tensorflow/python/framework/constant_op.py
@@ -28,6 +28,7 @@
from tensorflow.python.eager import context
from tensorflow.python.eager import execute
from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import op_callbacks
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_util
@@ -299,11 +300,17 @@
value, dtype=dtype, shape=shape, verify_shape=verify_shape,
allow_broadcast=allow_broadcast))
dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)
+ attrs = {"value": tensor_value, "dtype": dtype_value}
const_tensor = g._create_op_internal( # pylint: disable=protected-access
- "Const", [], [dtype_value.type],
- attrs={"value": tensor_value,
- "dtype": dtype_value},
- name=name).outputs[0]
+ "Const", [], [dtype_value.type], attrs=attrs, name=name).outputs[0]
+
+ if op_callbacks.should_invoke_op_callbacks():
+ # TODO(b/147670703): Once the special-op creation code paths
+ # are unified. Remove this `if` block.
+ callback_outputs = op_callbacks.invoke_op_callbacks(
+ "Const", tuple(), attrs, (const_tensor,), op_name=name, graph=g)
+ if callback_outputs is not None:
+ const_tensor, = callback_outputs
return const_tensor
diff --git a/tensorflow/python/framework/indexed_slices.py b/tensorflow/python/framework/indexed_slices.py
index 8e9c6f6..f85d0e7 100644
--- a/tensorflow/python/framework/indexed_slices.py
+++ b/tensorflow/python/framework/indexed_slices.py
@@ -29,9 +29,9 @@
from tensorflow.python.framework import composite_tensor
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import tensor_conversion_registry
-from tensorflow.python.framework import tensor_like
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import type_spec
+from tensorflow.python.types import internal
from tensorflow.python.util.lazy_loader import LazyLoader
from tensorflow.python.util.tf_export import tf_export
@@ -55,8 +55,9 @@
"tensorflow.python.framework.tensor_util")
+# TODO(mdan): Should IndexedSlices be a "tensor"?
@tf_export("IndexedSlices")
-class IndexedSlices(tensor_like.TensorLike, composite_tensor.CompositeTensor):
+class IndexedSlices(internal.NativeObject, composite_tensor.CompositeTensor):
"""A sparse representation of a set of tensor slices at given indices.
This class is a simple wrapper for a pair of `Tensor` objects:
@@ -305,7 +306,8 @@
"""
if isinstance(value, ops.EagerTensor) and not context.executing_eagerly():
return ops.convert_to_tensor(value, dtype=dtype, name=name, as_ref=as_ref)
- elif isinstance(value, tensor_like.TensorLike):
+ # TODO(mdan): Name says tensor_or_indexed_slices. So do explicitly just that?
+ elif isinstance(value, internal.NativeObject):
if dtype and not dtypes.as_dtype(dtype).is_compatible_with(value.dtype):
raise ValueError(
"Tensor conversion requested dtype %s for Tensor with dtype %s: %r" %
diff --git a/tensorflow/python/framework/op_callbacks_test.py b/tensorflow/python/framework/op_callbacks_test.py
index 31b6a58..8868ffd 100644
--- a/tensorflow/python/framework/op_callbacks_test.py
+++ b/tensorflow/python/framework/op_callbacks_test.py
@@ -109,7 +109,8 @@
if compat.as_bytes(op_type) in (_ENTER_OP, _EXIT_OP, _IF_OP, _MERGE_OP,
_NEXT_ITERATION_OP, _STATELESS_IF_OP,
_SWITCH_OP, _WHILE_OP, _IDENTITY_OP,
- _VAR_HANDLE_OP, _PLACEHOLDER_OP):
+ _VAR_HANDLE_OP, _PLACEHOLDER_OP,
+ _CONSTANT_OP):
# TODO(cais): Overriding the output of StatelessIf, If and While ops
# currently fails with error. Investigate (b/139668453).
# Avoid instrumenting Identity ops as well, as they are inserted
@@ -724,7 +725,7 @@
def testOverrideDTypeInFuncGraph(self):
def to_float64(op_type, inputs, attrs, outputs, op_name=None, graph=None):
del inputs, attrs, op_name, graph # Unused.
- if op_type == "Placeholder":
+ if op_type in ("Const", "Placeholder"):
return outputs
else:
return [math_ops.cast(output, dtypes.float64) for output in outputs]
@@ -751,6 +752,17 @@
self.assertIsNone(w)
self.assertEqual(instrument.eager_op_types, [_ADD_OP])
+ def testOpCallbackCapturesConstTensors(self):
+ instrument = _NumpyFunctionCallback()
+ op_callbacks.add_op_callback(instrument.callback)
+
+ @def_function.function
+ def times_two_plus_three(x):
+ return x * 2.0 + 3.0
+
+ self.assertAllClose(times_two_plus_three(constant_op.constant(10.0)), 23.0)
+ self.assertEqual(instrument.graph_op_types.count(b"Const"), 2)
+
@test_util.run_in_graph_and_eager_modes
def testOpCallbackWorksWithGradientTape(self):
instrument = _NumpyFunctionCallback()
diff --git a/tensorflow/python/framework/ops.py b/tensorflow/python/framework/ops.py
index 914dc21..7f5754b 100644
--- a/tensorflow/python/framework/ops.py
+++ b/tensorflow/python/framework/ops.py
@@ -56,13 +56,13 @@
from tensorflow.python.framework import indexed_slices
from tensorflow.python.framework import registry
from tensorflow.python.framework import tensor_conversion_registry
-from tensorflow.python.framework import tensor_like
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import traceable_stack
from tensorflow.python.framework import versions
from tensorflow.python.ops import control_flow_util
from tensorflow.python.platform import app
from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.types import internal
from tensorflow.python.util import compat
from tensorflow.python.util import decorator_utils
from tensorflow.python.util import deprecation
@@ -302,8 +302,9 @@
Tensor._USE_EQUALITY = False # pylint: disable=protected-access
+# TODO(mdan): This object should subclass Symbol, not just Tensor.
@tf_export("Tensor")
-class Tensor(tensor_like.TensorLike):
+class Tensor(internal.NativeObject):
"""A tensor is a multidimensional array of elements represented by a
`tf.Tensor` object. All elements are of a single known data type.
@@ -1007,6 +1008,7 @@
# TODO(agarwal): consider getting rid of this.
+# TODO(mdan): This object should not subclass ops.Tensor.
class _EagerTensorBase(Tensor):
"""Base class for EagerTensor."""
@@ -5295,6 +5297,11 @@
See `tf.Graph.control_dependencies`
for more details.
+ Note: *In TensorFlow 2 with eager and/or Autograph, you should not require
+ this method, as code executes in the expected order.* Only use
+ `tf.control_dependencies` when working with v1-style code or in a graph
+ context such as inside `Dataset.map`.
+
When eager execution is enabled, any callable object in the `control_inputs`
list will be called.
@@ -6057,7 +6064,7 @@
# TODO(josh11b): Note that we exclude subclasses of Tensor. Need to clean this
# up.
graph_element = None
- if (isinstance(op_input, (Operation, tensor_like.TensorLike)) and
+ if (isinstance(op_input, (Operation, internal.NativeObject)) and
((not isinstance(op_input, Tensor)) or type(op_input) == Tensor)): # pylint: disable=unidiomatic-typecheck
graph_element = op_input
else:
diff --git a/tensorflow/python/framework/sparse_tensor.py b/tensorflow/python/framework/sparse_tensor.py
index d085dfd..76cb24f 100644
--- a/tensorflow/python/framework/sparse_tensor.py
+++ b/tensorflow/python/framework/sparse_tensor.py
@@ -29,12 +29,12 @@
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
-from tensorflow.python.framework import tensor_like
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import tensor_util
from tensorflow.python.framework import type_spec
from tensorflow.python.ops import gen_sparse_ops
+from tensorflow.python.types import internal
from tensorflow.python.util.tf_export import tf_export
# pylint: disable=protected-access
@@ -44,7 +44,7 @@
@tf_export("sparse.SparseTensor", "SparseTensor")
-class SparseTensor(tensor_like.TensorLike, composite_tensor.CompositeTensor):
+class SparseTensor(internal.NativeObject, composite_tensor.CompositeTensor):
"""Represents a sparse tensor.
TensorFlow represents a sparse tensor as three separate dense tensors:
@@ -132,38 +132,9 @@
# is a VariableOp and updating users of SparseTensor.
values = ops.convert_to_tensor(values, name="values")
- # Can't check `if context.executing_eagerly()` here because sparse
- # placeholders can still be used in eager context, when building a
- # functional model.
- if isinstance(indices, ops.EagerTensor):
- try:
- dense_shape = ops.convert_to_tensor(
- dense_shape, name="dense_shape", dtype=dtypes.int64)
- dense_shape_default = tensor_shape.TensorShape(dense_shape)
- except ValueError:
- raise ValueError("Unable to create eager SparseTensor. Check that "
- "your shape is correctly defined. Eager "
- "SparseTensors don't support unknown dimesions.\n"
- "got shape:\n {}".format(dense_shape))
- else:
- if isinstance(dense_shape, ops.Tensor):
- dense_shape_default = tensor_util.constant_value_as_shape(dense_shape)
- else:
- dense_shape_default = []
- for dim in dense_shape:
- if isinstance(dim, ops.Tensor):
- # There is code passing lists of constant tensors.
- dim = tensor_util.constant_value(dim)
- if dim == -1:
- # -1 may be passed for unknown shapes.
- dim = None
-
- dense_shape_default.append(dim)
-
- dense_shape_default = tensor_shape.TensorShape(dense_shape_default)
-
- dense_shape = ops.convert_to_tensor(
- dense_shape, name="dense_shape", dtype=dtypes.int64)
+ dense_shape = ops.convert_to_tensor(
+ dense_shape, name="dense_shape", dtype=dtypes.int64)
+ dense_shape_default = tensor_util.constant_value_as_shape(dense_shape)
self._indices = indices
self._values = values
diff --git a/tensorflow/python/framework/sparse_tensor_test.py b/tensorflow/python/framework/sparse_tensor_test.py
index f7ecf00..0d18af1 100644
--- a/tensorflow/python/framework/sparse_tensor_test.py
+++ b/tensorflow/python/framework/sparse_tensor_test.py
@@ -29,6 +29,8 @@
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import test_util
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import math_ops
from tensorflow.python.ops import sparse_ops
from tensorflow.python.platform import googletest
@@ -124,6 +126,84 @@
sparse_tensor_value.dense_shape, convertee.dense_shape)
+class SparseTensorShapeTest(test_util.TensorFlowTestCase):
+
+ def test_simple(self):
+ indices = [[0, 2]]
+ values = [1]
+ dense_shape = [5, 5]
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+
+ self.assertIsInstance(sp.shape, tensor_shape.TensorShape)
+ self.assertIsInstance(sp.dense_shape, ops.Tensor)
+ self.assertEqual(sp.shape.as_list(), [5, 5])
+
+ def test_unknown_shape(self):
+
+ @def_function.function
+ def my_func(dense_shape):
+ indices = [[0, 2]]
+ values = [1]
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+ self.assertEqual(sp.shape.as_list(), [None, None])
+ return sp
+
+ my_func.get_concrete_function(
+ dense_shape=tensor_spec.TensorSpec(
+ dtype=dtypes.int64, shape=[2,]))
+
+ def test_partial_shape(self):
+
+ @def_function.function
+ def my_func(x):
+ indices = [[0, 2]]
+ values = [1]
+ y = ops.convert_to_tensor(3, dtype=dtypes.int64)
+ dense_shape = [x, y]
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+ self.assertEqual(sp.shape.as_list(), [None, 3])
+ return sp
+
+ my_func.get_concrete_function(
+ x=tensor_spec.TensorSpec(dtype=dtypes.int64, shape=[]))
+
+ def test_neg_shape(self):
+ indices = [[0, 2]]
+ values = [1]
+ dense_shape = [-1, 5]
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+ self.assertEqual(sp.shape.as_list(), [None, 5])
+
+ def test_unknown_tensor_shape(self):
+
+ @def_function.function
+ def my_func(x):
+ indices = [[0, 0]]
+ values = [1]
+ dense_shape = array_ops.shape(x)
+ dense_shape = math_ops.cast(dense_shape, dtypes.int64)
+
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+ self.assertEqual(sp.shape.as_list(), [None, None])
+ return sp
+
+ my_func.get_concrete_function(
+ x=tensor_spec.TensorSpec(dtype=dtypes.int64, shape=[None, None]))
+
+ def test_unknown_rank(self):
+
+ @def_function.function
+ def my_func(dense_shape):
+ indices = [[0, 0]]
+ values = [1]
+ sp = sparse_tensor.SparseTensor(indices, values, dense_shape)
+ self.assertEqual(sp.shape.rank, None)
+ return sp
+
+ my_func.get_concrete_function(
+ dense_shape=tensor_spec.TensorSpec(dtype=dtypes.int64, shape=[None]))
+
+
@test_util.run_all_in_graph_and_eager_modes
class SparseTensorSpecTest(test_util.TensorFlowTestCase,
parameterized.TestCase):
diff --git a/tensorflow/python/framework/tensor_util.py b/tensorflow/python/framework/tensor_util.py
index 8f22cad..4aa0334 100644
--- a/tensorflow/python/framework/tensor_util.py
+++ b/tensorflow/python/framework/tensor_util.py
@@ -25,9 +25,8 @@
from tensorflow.python.eager import context
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
-from tensorflow.python.framework import tensor_like
from tensorflow.python.framework import tensor_shape
-from tensorflow.python.types import core
+from tensorflow.python.types import internal
from tensorflow.python.util import compat
from tensorflow.python.util import nest
from tensorflow.python.util.tf_export import tf_export
@@ -791,6 +790,10 @@
return np.not_equal(value1, value2)
elif tensor.op.type == "StopGradient":
return constant_value(tensor.op.inputs[0], partial)
+ elif tensor.op.type == "Identity":
+ return constant_value(tensor.op.inputs[0], partial)
+ elif tensor.op.type in ("CheckNumericsV2", "DebugIdentityV2"):
+ return constant_value(tensor.op.inputs[0], partial)
else:
return None
@@ -976,6 +979,7 @@
return ret
+# TODO(mdan): Deprecate in favor of more static-friendly types.
@tf_export("is_tensor")
def is_tensor(x): # pylint: disable=invalid-name
"""Checks whether `x` is a TF-native type that can be passed to many TF ops.
@@ -1002,7 +1006,7 @@
Returns:
`True` if `x` is a tensor or "tensor-like", `False` if not.
"""
- return (isinstance(x, (tensor_like.TensorLike, core.Tensor)) or
+ return (isinstance(x, internal.NativeObject) or
ops.is_dense_tensor_like(x) or
getattr(x, "is_tensor_like", False))
diff --git a/tensorflow/python/framework/type_spec.py b/tensorflow/python/framework/type_spec.py
index 490574b..8da3265 100644
--- a/tensorflow/python/framework/type_spec.py
+++ b/tensorflow/python/framework/type_spec.py
@@ -83,7 +83,11 @@
@abc.abstractproperty
def value_type(self):
- """The Python type for values that are compatible with this TypeSpec."""
+ """The Python type for values that are compatible with this TypeSpec.
+
+ In particular, all values that are compatible with this TypeSpec must be an
+ instance of this type.
+ """
raise NotImplementedError("%s.value_type" % type(self).__name__)
def is_compatible_with(self, spec_or_value):
diff --git a/tensorflow/python/keras/activations.py b/tensorflow/python/keras/activations.py
index b69080d..1b68f42 100644
--- a/tensorflow/python/keras/activations.py
+++ b/tensorflow/python/keras/activations.py
@@ -155,7 +155,15 @@
@keras_export('keras.activations.softplus')
def softplus(x):
"""Softplus activation function, `softplus(x) = log(exp(x) + 1)`.
-
+
+ Example Usage:
+
+ >>> a = tf.constant([-20, -1.0, 0.0, 1.0, 20], dtype = tf.float32)
+ >>> b = tf.keras.activations.softplus(a)
+ >>> b.numpy()
+ array([2.0611537e-09, 3.1326166e-01, 6.9314718e-01, 1.3132616e+00,
+ 2.0000000e+01], dtype=float32)
+
Arguments:
x: Input tensor.
@@ -168,6 +176,13 @@
@keras_export('keras.activations.softsign')
def softsign(x):
"""Softsign activation function, `softsign(x) = x / (abs(x) + 1)`.
+
+ Example Usage:
+
+ >>> a = tf.constant([-1.0, 0.0, 1.0], dtype = tf.float32)
+ >>> b = tf.keras.activations.softsign(a)
+ >>> b.numpy()
+ array([-0.5, 0. , 0.5], dtype=float32)
Arguments:
x: Input tensor.
@@ -180,7 +195,21 @@
@keras_export('keras.activations.swish')
def swish(x):
- """Swish activation function.
+ """Swish activation function, `swish(x) = x * sigmoid(x)`.
+
+ Swish activation function which returns `x*sigmoid(x)`.
+ It is a smooth, non-monotonic function that consistently matches
+ or outperforms ReLU on deep networks, it is unbounded above and
+ bounded below.
+
+
+ Example Usage:
+
+ >>> a = tf.constant([-20, -1.0, 0.0, 1.0, 20], dtype = tf.float32)
+ >>> b = tf.keras.activations.swish(a)
+ >>> b.numpy()
+ array([-4.1223075e-08, -2.6894143e-01, 0.0000000e+00, 7.3105860e-01,
+ 2.0000000e+01], dtype=float32)
Arguments:
x: Input tensor.
@@ -264,14 +293,16 @@
the result of the function gets close to 1.
Sigmoid is equivalent to a 2-element Softmax, where the second element is
- assumed to be zero.
+ assumed to be zero. The sigmoid function always returns a value between
+ 0 and 1.
For example:
>>> a = tf.constant([-20, -1.0, 0.0, 1.0, 20], dtype = tf.float32)
>>> b = tf.keras.activations.sigmoid(a)
- >>> b.numpy() >= 0
- array([ True, True, True, True, True])
+ >>> b.numpy()
+ array([2.0611537e-09, 2.6894143e-01, 5.0000000e-01, 7.3105860e-01,
+ 1.0000000e+00], dtype=float32)
Arguments:
x: Input tensor.
diff --git a/tensorflow/python/keras/callbacks.py b/tensorflow/python/keras/callbacks.py
index 1dca7ad..6748a57 100644
--- a/tensorflow/python/keras/callbacks.py
+++ b/tensorflow/python/keras/callbacks.py
@@ -27,7 +27,6 @@
import json
import os
import re
-import tempfile
import time
import numpy as np
@@ -225,12 +224,7 @@
if params:
self.set_params(params)
- self._queue_length = 10
- self._reset_batch_timing()
-
- # Determines if batch-level hooks need to be called.
- # This is important for performance, because processing batch-level logs
- # will cause async eager to block on each batch.
+ # Performance optimization: determines if batch hooks need to be called.
# pylint: disable=protected-access
self._should_call_train_batch_hooks = any(
cb._implements_train_batch_hooks() for cb in self.callbacks)
@@ -240,6 +234,11 @@
cb._implements_predict_batch_hooks() for cb in self.callbacks)
# pylint: enable=protected-access
+ # Performance check: Check batch hooks for slowness compared to batch time.
+ self._timing = {}
+ self._check_timing = False
+ self._batch_start_time = None
+
def _add_default_callbacks(self, add_history, add_progbar):
"""Adds `Callback`s that are always present."""
self._progbar = None
@@ -259,11 +258,6 @@
self._history = History()
self.callbacks.append(self._history)
- def _reset_batch_timing(self):
- self._delta_t_batch = 0.
- self._delta_ts = collections.defaultdict(
- lambda: collections.deque([], maxlen=self._queue_length))
-
def append(self, callback):
self.callbacks.append(callback)
@@ -283,33 +277,65 @@
"""Helper function for all batch_{begin | end} methods."""
if not self.callbacks:
return
- hook_name = 'on_{mode}_batch_{hook}'.format(mode=mode, hook=hook)
- if hook == 'begin':
- self._t_enter_batch = time.time()
- if hook == 'end':
- # Batch is ending, calculate batch time.
- self._delta_t_batch = time.time() - self._t_enter_batch
+ if hook == 'begin':
+ self._call_batch_begin_hook(mode, batch, logs)
+ elif hook == 'end':
+ self._call_batch_end_hook(mode, batch, logs)
+ else:
+ raise ValueError('Unrecognized hook: {}'.format(hook))
+
+ def _call_batch_begin_hook(self, mode, batch, logs):
+ """Helper function for `on_*_batch_begin` methods."""
+ hook_name = 'on_{mode}_batch_begin'.format(mode=mode)
+ self._check_timing = batch == 1 and hook_name not in self._timing
+ self._call_batch_hook_helper(hook_name, batch, logs)
+
+ if self._check_timing:
+ self._batch_start_time = time.time()
+
+ def _call_batch_end_hook(self, mode, batch, logs):
+ """Helper function for `on_*_batch_end` methods."""
+ hook_name = 'on_{mode}_batch_end'.format(mode=mode)
+
+ if self._check_timing:
+ batch_time = time.time() - self._batch_start_time
+
+ self._call_batch_hook_helper(hook_name, batch, logs)
+
+ if self._check_timing:
+ end_hook_name = hook_name
+ begin_hook_name = 'on_{mode}_batch_begin'.format(mode=mode)
+
+ threshold_time = 0.5 * batch_time
+ warning_msg = ('Callbacks method `{hook}` is slow compared to '
+ 'the batch time. Check your callbacks.')
+ if self._timing[begin_hook_name] > threshold_time:
+ logging.warning(warning_msg.format(hook=begin_hook_name))
+ if self._timing[end_hook_name] > threshold_time:
+ logging.warning(warning_msg.format(hook=end_hook_name))
+
+ self._check_timing = False
+ self._batch_start_time = None
+
+ def _call_batch_hook_helper(self, hook_name, batch, logs):
+ """Helper function for `on_*_batch_*` methods."""
logs = logs or {}
- t_before_callbacks = time.time()
numpy_logs = None
+ if self._check_timing:
+ start_time = time.time()
+
for callback in self.callbacks:
- batch_hook = getattr(callback, hook_name)
+ hook = getattr(callback, hook_name)
if getattr(callback, '_supports_tf_logs', False):
- batch_hook(batch, logs)
+ hook(batch, logs)
else:
if numpy_logs is None: # Only convert once.
numpy_logs = tf_utils.to_numpy_or_python_type(logs)
- batch_hook(batch, numpy_logs)
- self._delta_ts[hook_name].append(time.time() - t_before_callbacks)
+ hook(batch, numpy_logs)
- delta_t_median = np.median(self._delta_ts[hook_name])
- if (self._delta_t_batch > 0. and
- delta_t_median > 0.95 * self._delta_t_batch and delta_t_median > 0.1):
- logging.warning(
- 'Method (%s) is slow compared '
- 'to the batch update (%f). Check your callbacks.', hook_name,
- delta_t_median)
+ if self._check_timing:
+ self._timing[hook_name] = time.time() - start_time
def _call_begin_hook(self, mode):
"""Helper function for on_{train|test|predict}_begin methods."""
@@ -356,7 +382,6 @@
if numpy_logs is None: # Only convert once.
numpy_logs = tf_utils.to_numpy_or_python_type(logs)
callback.on_epoch_begin(epoch, numpy_logs)
- self._reset_batch_timing()
def on_epoch_end(self, epoch, logs=None):
"""Calls the `on_epoch_end` methods of its callbacks.
@@ -1303,36 +1328,24 @@
def _get_file_path(self, epoch, logs):
"""Returns the file path for checkpoint."""
# pylint: disable=protected-access
- if not self.model._in_multi_worker_mode(
- ) or multi_worker_util.should_save_checkpoint():
- try:
- # `filepath` may contain placeholders such as `{epoch:02d}` and
- # `{mape:.2f}`. A mismatch between logged metrics and the path's
- # placeholders can cause formatting to fail.
- return self.filepath.format(epoch=epoch + 1, **logs)
- except KeyError as e:
- raise KeyError('Failed to format this callback filepath: "{}". '
- 'Reason: {}'.format(self.filepath, e))
- else:
- # If this is multi-worker training, and this worker should not
- # save checkpoint, we use a temp filepath to store a dummy checkpoint, so
- # it writes to a file that will be removed at the end of `_save_model()`
- # call. This is because the SyncOnReadVariable needs to be synced across
- # all the workers in order to be read, and all workers need to initiate
- # that.
- self._temp_file_dir = tempfile.mkdtemp()
- extension = os.path.splitext(self.filepath)[1]
- return os.path.join(self._temp_file_dir, 'temp' + extension)
+ try:
+ # `filepath` may contain placeholders such as `{epoch:02d}` and
+ # `{mape:.2f}`. A mismatch between logged metrics and the path's
+ # placeholders can cause formatting to fail.
+ file_path = self.filepath.format(epoch=epoch + 1, **logs)
+ except KeyError as e:
+ raise KeyError('Failed to format this callback filepath: "{}". '
+ 'Reason: {}'.format(self.filepath, e))
+ self._write_filepath = distributed_file_utils.write_filepath(
+ file_path, self.model.distribute_strategy)
+ return self._write_filepath
def _maybe_remove_file(self):
# Remove the checkpoint directory in multi-worker training where this worker
# should not checkpoint. It is a dummy directory previously saved for sync
# distributed training.
-
- if (self.model._in_multi_worker_mode() and # pylint: disable=protected-access
- not multi_worker_util.should_save_checkpoint()):
- file_io.delete_recursively(self._temp_file_dir)
- del self._temp_file_dir
+ distributed_file_utils.remove_temp_dir_with_filepath(
+ self._write_filepath, self.model.distribute_strategy)
def _get_most_recently_modified_file_matching_pattern(self, pattern):
"""Returns the most recently modified filepath matching pattern.
diff --git a/tensorflow/python/keras/callbacks_test.py b/tensorflow/python/keras/callbacks_test.py
index 7c9bef5..9d15f87 100644
--- a/tensorflow/python/keras/callbacks_test.py
+++ b/tensorflow/python/keras/callbacks_test.py
@@ -274,6 +274,37 @@
model.fit(dataset, epochs=2, steps_per_epoch=10)
self.assertRegexpMatches(printed.contents(), expected_log)
+ @keras_parameterized.run_all_keras_modes
+ def test_callback_warning(self):
+
+ class SleepCallback(keras.callbacks.Callback):
+
+ def on_train_batch_end(self, batch, logs=None):
+ time.sleep(1)
+
+ model = sequential.Sequential()
+ model.add(keras.layers.Dense(1, activation='sigmoid'))
+ model.compile(
+ 'sgd',
+ loss='binary_crossentropy',
+ run_eagerly=testing_utils.should_run_eagerly())
+
+ warning_messages = []
+
+ def warning(msg):
+ warning_messages.append(msg)
+
+ with test.mock.patch.object(logging, 'warning', warning):
+ model.fit(
+ np.ones((10, 10), 'float32'),
+ np.ones((10, 1), 'float32'),
+ batch_size=5,
+ epochs=10,
+ callbacks=[SleepCallback()])
+ warning_msg = ('Callbacks method `on_train_batch_end` is slow compared '
+ 'to the batch time. Check your callbacks.')
+ self.assertIn(warning_msg, warning_messages)
+
@keras_parameterized.run_with_all_model_types(exclude_models='functional')
@keras_parameterized.run_all_keras_modes
def test_progbar_logging_deferred_model_build(self):
diff --git a/tensorflow/python/keras/distribute/BUILD b/tensorflow/python/keras/distribute/BUILD
index 7281e17..14647df 100644
--- a/tensorflow/python/keras/distribute/BUILD
+++ b/tensorflow/python/keras/distribute/BUILD
@@ -440,3 +440,15 @@
"//tensorflow/python/keras/optimizer_v2",
],
)
+
+py_library(
+ name = "tpu_strategy_test_utils",
+ srcs = ["tpu_strategy_test_utils.py"],
+ deps = [
+ "//tensorflow/python:platform",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/distribute/cluster_resolver:tpu_cluster_resolver_py",
+ "//tensorflow/python/eager:remote",
+ "//tensorflow/python/tpu:tpu_strategy_util",
+ ],
+)
diff --git a/tensorflow/python/keras/distribute/multi_worker_callback_tf2_test.py b/tensorflow/python/keras/distribute/multi_worker_callback_tf2_test.py
index c99b6db..7ea385e 100644
--- a/tensorflow/python/keras/distribute/multi_worker_callback_tf2_test.py
+++ b/tensorflow/python/keras/distribute/multi_worker_callback_tf2_test.py
@@ -23,6 +23,7 @@
from tensorflow.python.distribute import collective_all_reduce_strategy as collective_strategy
from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import distributed_file_utils
from tensorflow.python.distribute import multi_process_runner
from tensorflow.python.distribute import multi_worker_test_base as test_base
from tensorflow.python.keras import callbacks
@@ -106,6 +107,16 @@
training_state.checkpoint_exists(saving_filepath),
test_base.is_chief())
+ # If it's chief, the model should be saved (`write_filepath` should
+ # simply return `saving_filepath`); if not, i.e. for non-chief workers,
+ # the temporary path generated by `write_filepath` should no longer
+ # contain the checkpoint that has been deleted.
+ test_obj.assertEqual(
+ training_state.checkpoint_exists(
+ distributed_file_utils.write_filepath(
+ saving_filepath, model._distribution_strategy)),
+ test_base.is_chief())
+
multi_process_runner.run(
proc_model_checkpoint_saves_on_chief_but_not_otherwise,
cluster_spec=test_base.create_cluster_spec(num_workers=2),
diff --git a/tensorflow/python/keras/distribute/tpu_strategy_test_utils.py b/tensorflow/python/keras/distribute/tpu_strategy_test_utils.py
new file mode 100644
index 0000000..14fdf92
--- /dev/null
+++ b/tensorflow/python/keras/distribute/tpu_strategy_test_utils.py
@@ -0,0 +1,47 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Utility functions for tests using TPUStrategy."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import google_type_annotations
+from __future__ import print_function
+
+from tensorflow.python.distribute import tpu_strategy
+from tensorflow.python.distribute.cluster_resolver import tpu_cluster_resolver
+from tensorflow.python.eager import remote
+from tensorflow.python.platform import flags
+from tensorflow.python.tpu import tpu_strategy_util
+
+FLAGS = flags.FLAGS
+flags.DEFINE_string("tpu", "", "Name of TPU to connect to.")
+flags.DEFINE_string("project", None, "Name of GCP project with TPU.")
+flags.DEFINE_string("zone", None, "Name of GCP zone with TPU.")
+
+
+def get_tpu_cluster_resolver():
+ resolver = tpu_cluster_resolver.TPUClusterResolver(
+ tpu=FLAGS.tpu,
+ zone=FLAGS.zone,
+ project=FLAGS.project,
+ )
+ return resolver
+
+
+def get_tpu_strategy():
+ resolver = get_tpu_cluster_resolver()
+ remote.connect_to_cluster(resolver)
+ tpu_strategy_util.initialize_tpu_system(resolver)
+ return tpu_strategy.TPUStrategy(resolver)
diff --git a/tensorflow/python/keras/engine/base_layer.py b/tensorflow/python/keras/engine/base_layer.py
index eb98022..7a6ba18 100644
--- a/tensorflow/python/keras/engine/base_layer.py
+++ b/tensorflow/python/keras/engine/base_layer.py
@@ -413,7 +413,7 @@
Note here that `call()` method in `tf.keras` is little bit different
from `keras` API. In `keras` API, you can pass support masking for
- layers as additional arguements. Whereas `tf.keras` has `compute_mask()`
+ layers as additional arguments. Whereas `tf.keras` has `compute_mask()`
method to support masking.
Arguments:
@@ -893,7 +893,7 @@
input_spec.assert_input_compatibility(self.input_spec, inputs,
self.name)
if (any(isinstance(x, ragged_tensor.RaggedTensor) for x in input_list)
- and self._supports_ragged_inputs is False): # pylint: disable=g-bool-id-comparison
+ and not self._supports_ragged_inputs):
raise ValueError('Layer %s does not support RaggedTensors as input. '
'Inputs received: %s. You can try converting your '
'input to an uniform tensor.' % (self.name, inputs))
@@ -1484,9 +1484,6 @@
dependent on `a` and some on `b`. This method automatically keeps track
of dependencies.
- The `get_updates_for` method allows to retrieve the updates relevant to a
- specific set of inputs.
-
This call is ignored when eager execution is enabled (in that case, variable
updates are run on the fly and thus do not need to be tracked for later
execution).
@@ -1518,12 +1515,6 @@
update()
return
- if call_context.in_call:
- relevant_inputs = call_context.inputs
- else:
- inbound_nodes = getattr(self, '_inbound_nodes', [])
- relevant_inputs = [node.input_tensors for node in inbound_nodes]
-
def process_update(x):
"""Standardize update ops.
@@ -1545,9 +1536,6 @@
update = x.op
else:
update = ops.convert_to_tensor_v2(x)
-
- reachable = tf_utils.get_reachable_from_inputs(relevant_inputs, [update])
- update._unconditional_update = update not in reachable
return update
updates = [process_update(x) for x in updates]
@@ -1691,15 +1679,7 @@
Returns:
List of update ops of the layer that depend on `inputs`.
"""
- if inputs is None:
- # Requesting unconditional updates.
- return [u for u in self.updates if u._unconditional_update]
-
- # Requesting input-conditional updates.
- updates = [u for u in self.updates if not u._unconditional_update]
- inputs = nest.flatten(inputs)
- reachable = tf_utils.get_reachable_from_inputs(inputs, updates)
- return [u for u in updates if u in reachable]
+ return self.updates
@doc_controls.do_not_doc_inheritable
def get_losses_for(self, inputs):
diff --git a/tensorflow/python/keras/engine/compile_utils.py b/tensorflow/python/keras/engine/compile_utils.py
index fd792e0..a6fb44f 100644
--- a/tensorflow/python/keras/engine/compile_utils.py
+++ b/tensorflow/python/keras/engine/compile_utils.py
@@ -192,6 +192,7 @@
loss_values = [] # Used for gradient calculation.
loss_metric_values = [] # Used for loss metric calculation.
+ batch_dim = None
zip_args = (y_true, y_pred, sample_weight, self._losses, self._loss_weights,
self._per_output_metrics)
for y_t, y_p, sw, loss_obj, loss_weight, metric_obj in zip(*zip_args):
@@ -207,8 +208,11 @@
# Correct for the `Mean` loss metrics counting each replica as a batch.
if loss_obj.reduction == losses_utils.ReductionV2.SUM:
loss_metric_value *= ds_context.get_strategy().num_replicas_in_sync
+
+ if batch_dim is None:
+ batch_dim = array_ops.shape(y_t)[0]
if metric_obj is not None:
- metric_obj.update_state(loss_metric_value)
+ metric_obj.update_state(loss_metric_value, sample_weight=batch_dim)
if loss_weight is not None:
loss_value *= loss_weight
@@ -232,7 +236,8 @@
loss_metric_values = losses_utils.cast_losses_to_common_dtype(
loss_metric_values)
total_loss_metric_value = math_ops.add_n(loss_metric_values)
- self._loss_metric.update_state(total_loss_metric_value)
+ self._loss_metric.update_state(
+ total_loss_metric_value, sample_weight=batch_dim)
loss_values = losses_utils.cast_losses_to_common_dtype(loss_values)
total_loss = math_ops.add_n(loss_values)
diff --git a/tensorflow/python/keras/engine/network.py b/tensorflow/python/keras/engine/network.py
index 313f0f2..770f046 100644
--- a/tensorflow/python/keras/engine/network.py
+++ b/tensorflow/python/keras/engine/network.py
@@ -192,7 +192,7 @@
# checkpoints, but may cause "all Python objects matched" assertions to fail
# (in which case less strict assertions may be substituted if necessary).
@trackable.no_automatic_dependency_tracking
- def _base_init(self, name=None, **kwargs):
+ def _base_init(self, **kwargs):
# The following are implemented as property functions:
# self.trainable_weights
# self.non_trainable_weights
@@ -201,12 +201,12 @@
# self.updates
generic_utils.validate_kwargs(kwargs, {'trainable', 'dtype', 'dynamic',
- 'autocast'})
+ 'name', 'autocast'})
- super(Network, self).__init__(name=name, **kwargs)
+ super(Network, self).__init__(**kwargs)
- self.output_names = None
self.input_names = None
+ self.output_names = None
self._saved_model_inputs_spec = None
# This is True for Sequential networks and Functional networks.
@@ -219,20 +219,13 @@
self._maybe_create_attribute('_is_compiled', False)
self._maybe_create_attribute('optimizer', None)
- self._scope = None # Never used.
- self._reuse = None # Never used.
- if context.executing_eagerly():
- self._graph = None
- else:
- self._graph = ops.get_default_graph() # Used in symbolic mode only.
-
self._trackable_saver = (
trackable_utils.saver_with_op_caching(self))
@trackable.no_automatic_dependency_tracking
- def _init_graph_network(self, inputs, outputs, name=None, **kwargs):
+ def _init_graph_network(self, inputs, outputs, **kwargs):
generic_utils.validate_kwargs(
- kwargs, {'trainable'},
+ kwargs, {'name', 'trainable'},
'Functional models may only specify `name` and `trainable` keyword '
'arguments during initialization. Got an unexpected argument:')
# Normalize and set self.inputs, self.outputs.
@@ -256,7 +249,7 @@
if any(not hasattr(tensor, '_keras_history') for tensor in self.outputs):
base_layer_utils.create_keras_history(self._nested_outputs)
- self._base_init(name=name, **kwargs)
+ self._base_init(**kwargs)
self._validate_graph_inputs_and_outputs()
# A Network does not create weights of its own, thus it is already
@@ -278,8 +271,6 @@
self._input_coordinates = []
self._output_coordinates = []
- self._supports_ragged_inputs = None
-
# This is for performance optimization when calling the Network on new
# inputs. Every time the Network is called on a set on input tensors,
# we compute the output tensors, output masks and output shapes in one pass,
@@ -366,17 +357,15 @@
self.output_names = uniquified
@trackable.no_automatic_dependency_tracking
- def _init_subclassed_network(self, name=None, **kwargs):
- self._base_init(name=name, **kwargs)
+ def _init_subclassed_network(self, **kwargs):
+ self._base_init(**kwargs)
self._is_graph_network = False
- self._init_call_fn_args()
- self._autocast = kwargs.get('autocast',
- base_layer_utils.v2_dtype_behavior_enabled())
- self._supports_ragged_inputs = None
- self.outputs = None
self.inputs = None
- self.built = False
- self._build_input_shape = None
+ self.outputs = None
+ # Since we don't know whether the subclass model support ragged inputs,
+ # we leave it as True, otherwise the layer will raise error when a ragged
+ # tensor is called as input.
+ self._supports_ragged_inputs = True
@property
@trackable_layer_utils.cache_recursive_attribute('dynamic')
diff --git a/tensorflow/python/keras/engine/network_test.py b/tensorflow/python/keras/engine/network_test.py
index ad62071..429a096 100644
--- a/tensorflow/python/keras/engine/network_test.py
+++ b/tensorflow/python/keras/engine/network_test.py
@@ -83,21 +83,14 @@
_ = layer(x1)
self.assertEqual(len(layer.updates), 2)
- self.assertEqual(len(layer.get_updates_for(x1)), 1)
- self.assertEqual(len(layer.get_updates_for(None)), 1)
x2 = input_layer_lib.Input(shape=(1,))
y2 = layer(x2)
self.assertEqual(len(layer.updates), 3)
- self.assertEqual(len(layer.get_updates_for(x1)), 1)
- self.assertEqual(len(layer.get_updates_for(x2)), 1)
- self.assertEqual(len(layer.get_updates_for(None)), 1)
network = network_lib.Network(x2, y2)
self.assertEqual(len(network.updates), 3)
- self.assertEqual(len(network.get_updates_for(x2)), 1)
- self.assertEqual(len(network.get_updates_for(None)), 1)
x3 = input_layer_lib.Input(shape=(1,))
_ = layer(x3)
@@ -106,17 +99,12 @@
x4 = input_layer_lib.Input(shape=(1,))
_ = network(x4)
self.assertEqual(len(network.updates), 5)
- self.assertEqual(len(network.get_updates_for(x2)), 1)
- self.assertEqual(len(network.get_updates_for(x4)), 1)
- self.assertEqual(len(network.get_updates_for(None)), 1)
network.add_update(state_ops.assign_add(layer.a, [[1]]))
self.assertEqual(len(network.updates), 6)
- self.assertEqual(len(network.get_updates_for(None)), 2)
network.add_update(state_ops.assign_add(layer.b, x4), inputs=True)
self.assertEqual(len(network.updates), 7)
- self.assertEqual(len(network.get_updates_for(x4)), 2)
@combinations.generate(combinations.combine(mode=['graph', 'eager']))
def test_get_updates_bn(self):
@@ -125,8 +113,6 @@
_ = layer(x1)
self.assertEqual(len(layer.updates), 2)
- self.assertEqual(len(layer.get_updates_for(x1)), 2)
- self.assertEqual(len(layer.get_updates_for(None)), 0)
def test_get_layer(self):
# create a simple network
@@ -1572,7 +1558,6 @@
output_shape = network.compute_output_shape([(None, 1), (None, 1)])
self.assertListEqual(output_shape.as_list(), [None, 1])
- @combinations.generate(combinations.combine(mode=['graph', 'eager']))
def test_updates_with_direct_call(self):
inputs = input_layer_lib.Input(shape=(10,))
x = layers.BatchNormalization()(inputs)
@@ -1582,8 +1567,7 @@
ph = backend.placeholder(shape=(10, 10))
model(ph)
- self.assertLen(model.get_updates_for(ph), 2)
- self.assertLen(model.get_updates_for(None), 0)
+ self.assertLen(model.updates, 4)
def test_dict_mapping_input(self):
diff --git a/tensorflow/python/keras/engine/sequential.py b/tensorflow/python/keras/engine/sequential.py
index 30ec7d0..9edfa4f 100644
--- a/tensorflow/python/keras/engine/sequential.py
+++ b/tensorflow/python/keras/engine/sequential.py
@@ -122,6 +122,10 @@
self._input_dtype = None
self._layer_call_argspecs = {}
self._created_nodes = set()
+ # Flag that indicate whether the sequential network topology has been
+ # created. It is false when there isn't any layer, or the layers doesn't
+ # have input shape.
+ self._graph_initialized = False
# Unfortunately some Sequential models using custom layers or FeatureColumn
# layers have multiple inputs. This is fundamentally incompatible with
@@ -228,8 +232,9 @@
self.outputs = [output_tensor]
self.built = True
- if set_inputs or self._is_graph_network:
+ if set_inputs or self._graph_initialized:
self._init_graph_network(self.inputs, self.outputs, name=self.name)
+ self._graph_initialized = True
else:
self._layers.append(layer)
self._handle_deferred_layer_dependencies([layer])
@@ -258,7 +263,8 @@
self.built = False
self._inferred_input_shape = None
self._has_explicit_input_shape = False
- elif self._is_graph_network:
+ self._graph_initialized = False
+ elif self._graph_initialized:
self.layers[-1]._outbound_nodes = []
self.outputs = [self.layers[-1].output]
self._init_graph_network(self.inputs, self.outputs, name=self.name)
@@ -285,9 +291,7 @@
if (new_shape is not None and new_shape != self._inferred_input_shape):
# A novel shape has been received: we need to rebuild the model.
# In case we are inside a graph function, we step out of it.
- # We also open a CPU device scope to avoid allocating memory on GPU.
- # The graph we create here is never used for execution.
- with ops.init_scope(), ops.device('/cpu:0'):
+ with ops.init_scope():
inputs = input_layer.Input(
batch_shape=new_shape,
dtype=input_dtype,
@@ -338,13 +342,14 @@
# TODO(fchollet): consider raising here, as we should not be
# supporting such layers.
self._init_graph_network(inputs, outputs, name=self.name)
+ self._graph_initialized = True
except: # pylint:disable=bare-except
self._use_legacy_deferred_behavior = True
self._inferred_input_shape = new_shape
@generic_utils.default
def build(self, input_shape=None):
- if self._is_graph_network:
+ if self._graph_initialized:
self._init_graph_network(self.inputs, self.outputs, name=self.name)
else:
if input_shape is None:
@@ -373,7 +378,7 @@
else:
self._build_graph_network_for_inferred_shape(inputs.shape, inputs.dtype)
- if self._is_graph_network:
+ if self._graph_initialized:
if not self.built:
self._init_graph_network(self.inputs, self.outputs, name=self.name)
return super(Sequential, self).call(inputs, training=training, mask=mask)
diff --git a/tensorflow/python/keras/engine/sequential_test.py b/tensorflow/python/keras/engine/sequential_test.py
index c65ac09..9589d24 100644
--- a/tensorflow/python/keras/engine/sequential_test.py
+++ b/tensorflow/python/keras/engine/sequential_test.py
@@ -491,6 +491,30 @@
y = np.random.random((2, 5))
model.fit(x, y, epochs=1)
+ @keras_parameterized.run_all_keras_modes
+ def test_build_empty_network(self):
+ x = np.random.random((2, 6))
+ y = np.random.random((2, 5))
+ model = keras.Sequential()
+
+ # Make sure an empty sequential model can still work with build().
+ model.build((None, 6))
+ self.assertTrue(model.built)
+
+ model.add(keras.layers.Dense(5, input_shape=(6,)))
+
+ model.compile(
+ loss='mse',
+ optimizer='rmsprop',
+ run_eagerly=testing_utils.should_run_eagerly())
+ model.fit(x, y)
+
+ model.pop()
+ self.assertFalse(model.built)
+
+ model.build((None, 6))
+ self.assertTrue(model.built)
+
if __name__ == '__main__':
test.main()
diff --git a/tensorflow/python/keras/engine/training.py b/tensorflow/python/keras/engine/training.py
index c41f4f6..5d8c3dc 100644
--- a/tensorflow/python/keras/engine/training.py
+++ b/tensorflow/python/keras/engine/training.py
@@ -350,7 +350,7 @@
_keras_api_gauge.get_cell('compile').set(True)
with self.distribute_strategy.scope():
self._validate_compile(optimizer, metrics, **kwargs)
- self._run_eagerly = kwargs.pop('run_eagerly', None)
+ self._run_eagerly = run_eagerly
self.optimizer = self._get_optimizer(optimizer)
self.compiled_loss = compile_utils.LossesContainer(
@@ -853,6 +853,7 @@
version_utils.disallow_legacy_graph('Model', 'fit')
self._assert_compile_was_called()
self._check_call_args('fit')
+ _disallow_inside_tf_function('fit')
if validation_split:
# Create the validation data using the training data. Only supported for
@@ -1130,6 +1131,7 @@
version_utils.disallow_legacy_graph('Model', 'evaluate')
self._assert_compile_was_called()
self._check_call_args('evaluate')
+ _disallow_inside_tf_function('evaluate')
with self.distribute_strategy.scope():
# Creates a `tf.data.Dataset` and handles batch and epoch iteration.
@@ -1329,6 +1331,7 @@
_keras_api_gauge.get_cell('predict').set(True)
version_utils.disallow_legacy_graph('Model', 'predict')
self._check_call_args('predict')
+ _disallow_inside_tf_function('predict')
outputs = None
with self.distribute_strategy.scope():
@@ -1449,6 +1452,7 @@
"""
self._assert_compile_was_called()
self._check_call_args('train_on_batch')
+ _disallow_inside_tf_function('train_on_batch')
with self.distribute_strategy.scope(), \
training_utils.RespectCompiledTrainableState(self):
iterator = data_adapter.single_batch_iterator(self.distribute_strategy, x,
@@ -1508,6 +1512,7 @@
"""
self._assert_compile_was_called()
self._check_call_args('test_on_batch')
+ _disallow_inside_tf_function('test_on_batch')
with self.distribute_strategy.scope():
iterator = data_adapter.single_batch_iterator(self.distribute_strategy, x,
y, sample_weight)
@@ -1541,6 +1546,7 @@
expectations of the model.
"""
self._check_call_args('predict_on_batch')
+ _disallow_inside_tf_function('predict_on_batch')
with self.distribute_strategy.scope():
iterator = data_adapter.single_batch_iterator(self.distribute_strategy, x)
predict_function = self.make_predict_function()
@@ -1944,3 +1950,15 @@
if not isinstance(out, variables.Variable):
return [out] # Return first Tensor or Op from outputs.
return [] # No viable Tensor or Op to use for control deps.
+
+
+def _disallow_inside_tf_function(method_name):
+ if ops.inside_function():
+ error_msg = (
+ 'Detected a call to `Model.{method_name}` inside a `tf.function`. '
+ '`Model.{method_name} is a high-level endpoint that manages its own '
+ '`tf.function`. Please move the call to `Model.{method_name}` outside '
+ 'of all enclosing `tf.function`s. Note that you can call a `Model` '
+ 'directly on `Tensor`s inside a `tf.function` like: `model(x)`.'
+ ).format(method_name=method_name)
+ raise RuntimeError(error_msg)
diff --git a/tensorflow/python/keras/engine/training_test.py b/tensorflow/python/keras/engine/training_test.py
index 475a370..e4c1ff6 100644
--- a/tensorflow/python/keras/engine/training_test.py
+++ b/tensorflow/python/keras/engine/training_test.py
@@ -28,6 +28,7 @@
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.eager import context
+from tensorflow.python.eager import def_function
from tensorflow.python.eager import function
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
@@ -89,6 +90,35 @@
self.assertAllClose(hist.history['loss'][0], 10000)
@keras_parameterized.run_all_keras_modes
+ def test_run_eagerly_setting(self):
+ model = sequential.Sequential([layers_module.Dense(1)])
+ run_eagerly = testing_utils.should_run_eagerly()
+ model.compile('sgd', 'mse', run_eagerly=run_eagerly)
+ self.assertEqual(model.run_eagerly, run_eagerly)
+
+ @keras_parameterized.run_all_keras_modes(always_skip_v1=True)
+ @parameterized.named_parameters(
+ ('train_on_batch', 'train_on_batch'),
+ ('test_on_batch', 'test_on_batch'),
+ ('predict_on_batch', 'predict_on_batch'),
+ ('fit', 'fit'),
+ ('evaluate', 'evaluate'),
+ ('predict', 'predict'),
+ )
+ def test_disallow_methods_inside_tf_function(self, method_name):
+ model = sequential.Sequential([layers_module.Dense(1)])
+ run_eagerly = testing_utils.should_run_eagerly()
+ model.compile('sgd', 'mse', run_eagerly=run_eagerly)
+
+ @def_function.function
+ def my_fn():
+ getattr(model, method_name)(1)
+
+ error_msg = 'inside a `tf.function`'
+ with self.assertRaisesRegexp(RuntimeError, error_msg):
+ my_fn()
+
+ @keras_parameterized.run_all_keras_modes
def test_fit_and_validate_learning_phase(self):
class ReturnTraining(layers_module.Layer):
diff --git a/tensorflow/python/keras/estimator/__init__.py b/tensorflow/python/keras/estimator/__init__.py
index 3560520..895dd04 100644
--- a/tensorflow/python/keras/estimator/__init__.py
+++ b/tensorflow/python/keras/estimator/__init__.py
@@ -169,6 +169,11 @@
estimator.train(input_fn, steps=1)
```
+ Note: We do not support creating weighted metrics in Keras and converting them
+ to weighted metrics in the Estimator API using `model_to_estimator`.
+ You will have to create these metrics directly on the estimator spec using the
+ `add_metrics` function.
+
To customize the estimator `eval_metric_ops` names, you can pass in the
`metric_names_map` dictionary mapping the keras model output metric names
to the custom names as follows:
diff --git a/tensorflow/python/keras/layers/normalization_test.py b/tensorflow/python/keras/layers/normalization_test.py
index 56f140f..6a615c2 100644
--- a/tensorflow/python/keras/layers/normalization_test.py
+++ b/tensorflow/python/keras/layers/normalization_test.py
@@ -403,8 +403,6 @@
model.train_on_batch(x, x)
self.assertLen(bn.updates, 4)
- self.assertLen(bn.get_updates_for(x1), 2)
- self.assertLen(model.get_updates_for(x2), 2)
# Test model-level reuse
x3 = keras.layers.Input(shape=(10,))
@@ -413,7 +411,6 @@
self.assertLen(new_model.updates, 6)
self.assertLen(model.updates, 6)
- self.assertLen(new_model.get_updates_for(x3), 2)
new_model.compile(gradient_descent.GradientDescentOptimizer(0.01), 'mse')
new_model.train_on_batch(x, x)
diff --git a/tensorflow/python/keras/layers/preprocessing/BUILD b/tensorflow/python/keras/layers/preprocessing/BUILD
index 8c346da..898e8d7 100644
--- a/tensorflow/python/keras/layers/preprocessing/BUILD
+++ b/tensorflow/python/keras/layers/preprocessing/BUILD
@@ -3,6 +3,7 @@
load("//tensorflow:tensorflow.bzl", "tf_py_test")
load("//tensorflow:tensorflow.bzl", "cuda_py_test")
+load("//tensorflow/python/tpu:tpu.bzl", "tpu_py_test")
package(
default_visibility = [
@@ -231,19 +232,6 @@
],
)
-tf_py_test(
- name = "discretization_test",
- size = "small",
- srcs = ["discretization_test.py"],
- python_version = "PY3",
- deps = [
- ":discretization",
- ":preprocessing_test_utils",
- "//tensorflow/python:client_testlib",
- "@absl_py//absl/testing:parameterized",
- ],
-)
-
cuda_py_test(
name = "categorical_crossing_test",
size = "medium",
@@ -261,6 +249,63 @@
],
)
+tf_py_test(
+ name = "categorical_encoding_test",
+ size = "medium",
+ srcs = ["categorical_encoding_test.py"],
+ python_version = "PY3",
+ deps = [
+ ":categorical_encoding",
+ ":preprocessing_test_utils",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/utils:generic_utils",
+ "//tensorflow/python/ops/ragged:ragged_string_ops",
+ "@absl_py//absl/testing:parameterized",
+ ],
+)
+
+tpu_py_test(
+ name = "categorical_encoding_tpu_test",
+ srcs = ["categorical_encoding_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":categorical_encoding",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
+tf_py_test(
+ name = "discretization_test",
+ size = "small",
+ srcs = ["discretization_test.py"],
+ python_version = "PY3",
+ deps = [
+ ":discretization",
+ ":preprocessing_test_utils",
+ "//tensorflow/python:client_testlib",
+ "@absl_py//absl/testing:parameterized",
+ ],
+)
+
+tpu_py_test(
+ name = "discretization_tpu_test",
+ srcs = ["discretization_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":discretization",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
cuda_py_test(
name = "hashing_test",
size = "medium",
@@ -275,6 +320,20 @@
],
)
+tpu_py_test(
+ name = "hashing_tpu_test",
+ srcs = ["hashing_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":hashing",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
tf_py_test(
name = "index_lookup_test",
size = "medium",
@@ -291,6 +350,20 @@
],
)
+tpu_py_test(
+ name = "index_lookup_tpu_test",
+ srcs = ["index_lookup_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":index_lookup",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
cuda_py_test(
name = "image_preprocessing_test",
size = "medium",
@@ -318,6 +391,20 @@
],
)
+tpu_py_test(
+ name = "normalization_tpu_test",
+ srcs = ["normalization_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":normalization",
+ "//tensorflow/python/distribute:tpu_strategy",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
tf_py_test(
name = "text_vectorization_test",
size = "medium",
@@ -334,6 +421,20 @@
],
)
+tpu_py_test(
+ name = "text_vectorization_tpu_test",
+ srcs = ["text_vectorization_tpu_test.py"],
+ disable_experimental = True,
+ python_version = "PY3",
+ tags = ["no_oss"],
+ deps = [
+ ":text_vectorization",
+ "//tensorflow/python/eager:test",
+ "//tensorflow/python/keras",
+ "//tensorflow/python/keras/distribute:tpu_strategy_test_utils",
+ ],
+)
+
tf_py_test(
name = "reduction_test",
srcs = ["reduction_test.py"],
@@ -346,22 +447,6 @@
)
tf_py_test(
- name = "categorical_encoding_test",
- size = "medium",
- srcs = ["categorical_encoding_test.py"],
- python_version = "PY3",
- deps = [
- ":categorical_encoding",
- ":preprocessing_test_utils",
- "//tensorflow/python:client_testlib",
- "//tensorflow/python/keras",
- "//tensorflow/python/keras/utils:generic_utils",
- "//tensorflow/python/ops/ragged:ragged_string_ops",
- "@absl_py//absl/testing:parameterized",
- ],
-)
-
-tf_py_test(
name = "preprocessing_stage_test",
srcs = ["preprocessing_stage_test.py"],
python_version = "PY3",
diff --git a/tensorflow/python/keras/layers/preprocessing/categorical_encoding_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/categorical_encoding_tpu_test.py
new file mode 100644
index 0000000..c3bba2f
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/categorical_encoding_tpu_test.py
@@ -0,0 +1,60 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.framework import dtypes
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import categorical_encoding
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class CategoricalEncodingDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_tpu_distribution(self):
+ input_array = np.array([[1, 2, 3, 1], [0, 3, 1, 0]])
+
+ # pyformat: disable
+ expected_output = [[0, 1, 1, 1, 0, 0],
+ [1, 1, 0, 1, 0, 0]]
+ # pyformat: enable
+ max_tokens = 6
+
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(4,), dtype=dtypes.int32)
+ layer = categorical_encoding.CategoricalEncoding(
+ max_tokens=max_tokens, output_mode=categorical_encoding.BINARY)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+ output_dataset = model.predict(input_array)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/preprocessing/discretization_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/discretization_tpu_test.py
new file mode 100644
index 0000000..005f8b0
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/discretization_tpu_test.py
@@ -0,0 +1,57 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import discretization
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class DiscretizationDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_tpu_distribution(self):
+ input_array = np.array([[-1.5, 1.0, 3.4, .5], [0.0, 3.0, 1.3, 0.0]])
+
+ expected_output = [[0, 2, 3, 1], [1, 3, 2, 1]]
+ expected_output_shape = [None, None]
+
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,))
+ layer = discretization.Discretization(
+ bins=[0., 1., 2.], output_mode=discretization.INTEGER)
+ bucket_data = layer(input_data)
+ self.assertAllEqual(expected_output_shape, bucket_data.shape.as_list())
+
+ model = keras.Model(inputs=input_data, outputs=bucket_data)
+ output_dataset = model.predict(input_array)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/preprocessing/hashing_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/hashing_tpu_test.py
new file mode 100644
index 0000000..e2e6d98
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/hashing_tpu_test.py
@@ -0,0 +1,58 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.framework import config
+from tensorflow.python.framework import dtypes
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import hashing
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class HashingDistributionTest(keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_tpu_distribution(self):
+ input_data = np.asarray([["omar"], ["stringer"], ["marlo"], ["wire"]])
+ input_dataset = dataset_ops.Dataset.from_tensor_slices(input_data).batch(
+ 2, drop_remainder=True)
+ expected_output = [[0], [0], [1], [0]]
+
+ config.set_soft_device_placement(True)
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = hashing.Hashing(num_bins=2)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+ output_dataset = model.predict(input_dataset)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py b/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
index feee0b3..4f909b6 100644
--- a/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
+++ b/tensorflow/python/keras/layers/preprocessing/image_preprocessing.py
@@ -205,7 +205,7 @@
self.input_spec = InputSpec(ndim=4)
super(RandomCrop, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -370,7 +370,7 @@
self._rng = make_generator(self.seed)
self.input_spec = InputSpec(ndim=4)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -493,7 +493,7 @@
self.input_spec = InputSpec(ndim=4)
super(RandomTranslation, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -771,7 +771,7 @@
self.input_spec = InputSpec(ndim=4)
super(RandomRotation, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -899,7 +899,7 @@
self.input_spec = InputSpec(ndim=4)
super(RandomZoom, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -1035,7 +1035,7 @@
self.input_spec = InputSpec(ndim=4)
super(RandomContrast, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -1113,7 +1113,7 @@
self._rng = make_generator(self.seed)
super(RandomHeight, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
@@ -1208,7 +1208,7 @@
self._rng = make_generator(self.seed)
super(RandomWidth, self).__init__(name=name, **kwargs)
- def call(self, inputs, training=None):
+ def call(self, inputs, training=True):
if training is None:
training = K.learning_phase()
diff --git a/tensorflow/python/keras/layers/preprocessing/index_lookup_test.py b/tensorflow/python/keras/layers/preprocessing/index_lookup_test.py
index d9990dd..ca34bf9 100644
--- a/tensorflow/python/keras/layers/preprocessing/index_lookup_test.py
+++ b/tensorflow/python/keras/layers/preprocessing/index_lookup_test.py
@@ -30,6 +30,7 @@
from tensorflow.python import tf2
from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.distribute import one_device_strategy
from tensorflow.python.eager import context
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import sparse_tensor
@@ -404,6 +405,29 @@
@keras_parameterized.run_all_keras_modes
+class IndexLookupDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_cpu_distribution(self):
+ vocab_data = ["earth", "wind", "and", "fire"]
+ input_array = np.array([["earth", "wind", "and", "fire"],
+ ["fire", "and", "earth", "michigan"]])
+ expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
+
+ strategy = one_device_strategy.OneDeviceStrategy("/cpu:0")
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = get_layer_class()()
+ layer.set_vocabulary(vocab_data)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+ output_dataset = model.predict(input_array)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+@keras_parameterized.run_all_keras_modes
class IndexLookupOutputTest(keras_parameterized.TestCase,
preprocessing_test_utils.PreprocessingLayerTest):
diff --git a/tensorflow/python/keras/layers/preprocessing/index_lookup_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/index_lookup_tpu_test.py
new file mode 100644
index 0000000..b371eec
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/index_lookup_tpu_test.py
@@ -0,0 +1,66 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.framework import config
+from tensorflow.python.framework import dtypes
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import index_lookup
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class IndexLookupDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_tpu_distribution(self):
+ vocab_data = [[
+ "earth", "earth", "earth", "earth", "wind", "wind", "wind", "and",
+ "and", "fire"
+ ]]
+ vocab_dataset = dataset_ops.Dataset.from_tensors(vocab_data)
+ input_array = np.array([["earth", "wind", "and", "fire"],
+ ["fire", "and", "earth", "michigan"]])
+ input_dataset = dataset_ops.Dataset.from_tensor_slices(input_array).batch(
+ 2, drop_remainder=True)
+ expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
+
+ config.set_soft_device_placement(True)
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = index_lookup.IndexLookup()
+ layer.adapt(vocab_dataset)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+ output_dataset = model.predict(input_dataset)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/preprocessing/normalization_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/normalization_tpu_test.py
new file mode 100644
index 0000000..b9c7b41
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/normalization_tpu_test.py
@@ -0,0 +1,128 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from absl.testing import parameterized
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import normalization
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.platform import test
+
+
+def _get_layer_computation_test_cases():
+ test_cases = ({
+ "adapt_data": np.array([[1.], [2.], [3.], [4.], [5.]], dtype=np.float32),
+ "axis": -1,
+ "test_data": np.array([[1.], [2.], [3.]], np.float32),
+ "expected": np.array([[-1.414214], [-.707107], [0]], np.float32),
+ "testcase_name": "2d_single_element"
+ }, {
+ "adapt_data": np.array([[1.], [2.], [3.], [4.], [5.]], dtype=np.float32),
+ "axis": None,
+ "test_data": np.array([[1.], [2.], [3.]], np.float32),
+ "expected": np.array([[-1.414214], [-.707107], [0]], np.float32),
+ "testcase_name": "2d_single_element_none_axis"
+ }, {
+ "adapt_data": np.array([[1., 2., 3., 4., 5.]], dtype=np.float32),
+ "axis": None,
+ "test_data": np.array([[1.], [2.], [3.]], np.float32),
+ "expected": np.array([[-1.414214], [-.707107], [0]], np.float32),
+ "testcase_name": "2d_single_element_none_axis_flat_data"
+ }, {
+ "adapt_data":
+ np.array([[[1., 2., 3.], [2., 3., 4.]], [[3., 4., 5.], [4., 5., 6.]]],
+ np.float32),
+ "axis":
+ 1,
+ "test_data":
+ np.array([[[1., 2., 3.], [2., 3., 4.]], [[3., 4., 5.], [4., 5., 6.]]],
+ np.float32),
+ "expected":
+ np.array([[[-1.549193, -0.774597, 0.], [-1.549193, -0.774597, 0.]],
+ [[0., 0.774597, 1.549193], [0., 0.774597, 1.549193]]],
+ np.float32),
+ "testcase_name":
+ "3d_internal_axis"
+ }, {
+ "adapt_data":
+ np.array(
+ [[[1., 0., 3.], [2., 3., 4.]], [[3., -1., 5.], [4., 5., 8.]]],
+ np.float32),
+ "axis": (1, 2),
+ "test_data":
+ np.array(
+ [[[3., 1., -1.], [2., 5., 4.]], [[3., 0., 5.], [2., 5., 8.]]],
+ np.float32),
+ "expected":
+ np.array(
+ [[[1., 3., -5.], [-1., 1., -1.]], [[1., 1., 1.], [-1., 1., 1.]]],
+ np.float32),
+ "testcase_name":
+ "3d_multiple_axis"
+ })
+
+ crossed_test_cases = []
+ # Cross above test cases with use_dataset in (True, False)
+ for use_dataset in (True, False):
+ for case in test_cases:
+ case = case.copy()
+ if use_dataset:
+ case["testcase_name"] = case["testcase_name"] + "_with_dataset"
+ case["use_dataset"] = use_dataset
+ crossed_test_cases.append(case)
+
+ return crossed_test_cases
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class NormalizationTest(keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ @parameterized.named_parameters(*_get_layer_computation_test_cases())
+ def test_layer_computation(self, adapt_data, axis, test_data, use_dataset,
+ expected):
+ input_shape = tuple([None for _ in range(test_data.ndim - 1)])
+ if use_dataset:
+ # Keras APIs expect batched datasets
+ adapt_data = dataset_ops.Dataset.from_tensor_slices(adapt_data).batch(
+ test_data.shape[0] // 2)
+ test_data = dataset_ops.Dataset.from_tensor_slices(test_data).batch(
+ test_data.shape[0] // 2)
+
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=input_shape)
+ layer = normalization.Normalization(axis=axis)
+ layer.adapt(adapt_data)
+ output = layer(input_data)
+ model = keras.Model(input_data, output)
+ output_data = model.predict(test_data)
+ self.assertAllClose(expected, output_data)
+
+
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/preprocessing/text_vectorization.py b/tensorflow/python/keras/layers/preprocessing/text_vectorization.py
index 4e8edf5..89dd48e 100644
--- a/tensorflow/python/keras/layers/preprocessing/text_vectorization.py
+++ b/tensorflow/python/keras/layers/preprocessing/text_vectorization.py
@@ -592,6 +592,9 @@
return inputs
def call(self, inputs):
+ if inputs.shape.rank == 1:
+ inputs = array_ops.expand_dims(inputs, axis=-1)
+
self._called = True
inputs = self._preprocess(inputs)
diff --git a/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py b/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py
index c50f31b..d8325f3 100644
--- a/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py
+++ b/tensorflow/python/keras/layers/preprocessing/text_vectorization_test.py
@@ -27,6 +27,7 @@
from tensorflow.python import tf2
from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.distribute import one_device_strategy
from tensorflow.python.eager import context
from tensorflow.python.framework import dtypes
from tensorflow.python.keras import backend
@@ -569,6 +570,33 @@
@keras_parameterized.run_all_keras_modes
+class TextVectorizationDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_distribution_strategy_output(self):
+ vocab_data = ["earth", "wind", "and", "fire"]
+ input_array = np.array([["earth", "wind", "and", "fire"],
+ ["fire", "and", "earth", "michigan"]])
+ expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
+
+ strategy = one_device_strategy.OneDeviceStrategy("/cpu:0")
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = get_layer_class()(
+ max_tokens=None,
+ standardize=None,
+ split=None,
+ output_mode=text_vectorization.INT)
+ layer.set_vocabulary(vocab_data)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+
+ output_dataset = model.predict(input_array)
+ self.assertAllEqual(expected_output, output_dataset)
+
+
+@keras_parameterized.run_all_keras_modes
class TextVectorizationOutputTest(
keras_parameterized.TestCase,
preprocessing_test_utils.PreprocessingLayerTest):
@@ -1110,6 +1138,16 @@
output_dataset = model.predict(input_array)
self.assertAllClose(expected_output, output_dataset)
+ def test_accept_1D_input(self):
+ input_array = np.array(["earth wind and fire",
+ "fire and earth michigan"])
+ layer = get_layer_class()(
+ standardize=None,
+ split=None,
+ output_mode="int")
+ layer.adapt(input_array)
+ _ = layer(input_array)
+
@keras_parameterized.run_all_keras_modes
class TextVectorizationModelBuildingTest(
diff --git a/tensorflow/python/keras/layers/preprocessing/text_vectorization_tpu_test.py b/tensorflow/python/keras/layers/preprocessing/text_vectorization_tpu_test.py
new file mode 100644
index 0000000..93e9796
--- /dev/null
+++ b/tensorflow/python/keras/layers/preprocessing/text_vectorization_tpu_test.py
@@ -0,0 +1,96 @@
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for keras.layers.preprocessing.normalization."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python import keras
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.framework import config
+from tensorflow.python.framework import dtypes
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.distribute import tpu_strategy_test_utils
+from tensorflow.python.keras.layers.preprocessing import preprocessing_test_utils
+from tensorflow.python.keras.layers.preprocessing import text_vectorization
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes(
+ always_skip_v1=True, always_skip_eager=True)
+class TextVectorizationTPUDistributionTest(
+ keras_parameterized.TestCase,
+ preprocessing_test_utils.PreprocessingLayerTest):
+
+ def test_distribution_strategy_output(self):
+ vocab_data = ["earth", "wind", "and", "fire"]
+ input_array = np.array([["earth", "wind", "and", "fire"],
+ ["fire", "and", "earth", "michigan"]])
+ input_dataset = dataset_ops.Dataset.from_tensor_slices(input_array).batch(
+ 2, drop_remainder=True)
+
+ expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
+
+ config.set_soft_device_placement(True)
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = text_vectorization.TextVectorization(
+ max_tokens=None,
+ standardize=None,
+ split=None,
+ output_mode=text_vectorization.INT)
+ layer.set_vocabulary(vocab_data)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+
+ output_dataset = model.predict(input_dataset)
+ self.assertAllEqual(expected_output, output_dataset)
+
+ def test_distribution_strategy_output_with_adapt(self):
+ vocab_data = [[
+ "earth", "earth", "earth", "earth", "wind", "wind", "wind", "and",
+ "and", "fire"
+ ]]
+ vocab_dataset = dataset_ops.Dataset.from_tensors(vocab_data)
+ input_array = np.array([["earth", "wind", "and", "fire"],
+ ["fire", "and", "earth", "michigan"]])
+ input_dataset = dataset_ops.Dataset.from_tensor_slices(input_array).batch(
+ 2, drop_remainder=True)
+
+ expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
+
+ config.set_soft_device_placement(True)
+ strategy = tpu_strategy_test_utils.get_tpu_strategy()
+
+ with strategy.scope():
+ input_data = keras.Input(shape=(None,), dtype=dtypes.string)
+ layer = text_vectorization.TextVectorization(
+ max_tokens=None,
+ standardize=None,
+ split=None,
+ output_mode=text_vectorization.INT)
+ layer.adapt(vocab_dataset)
+ int_data = layer(input_data)
+ model = keras.Model(inputs=input_data, outputs=int_data)
+
+ output_dataset = model.predict(input_dataset)
+ self.assertAllEqual(expected_output, output_dataset)
+if __name__ == "__main__":
+ test.main()
diff --git a/tensorflow/python/keras/layers/recurrent_test.py b/tensorflow/python/keras/layers/recurrent_test.py
index a6e77f8..b6afe2a 100644
--- a/tensorflow/python/keras/layers/recurrent_test.py
+++ b/tensorflow/python/keras/layers/recurrent_test.py
@@ -602,7 +602,7 @@
self.assertEqual(layer.get_losses_for(None), [loss_2])
self.assertEqual(layer.get_losses_for(x), [loss_1])
- # Test `get_updates_for` and `updates`
+ # Test `updates`
cells = [keras.layers.LSTMCell(1),
keras.layers.LSTMCell(1)]
layer = keras.layers.RNN(cells)
@@ -618,8 +618,6 @@
cells[0].add_update(update_1, inputs=x)
cells[0].add_update(update_2)
self.assertEqual(len(layer.updates), 2)
- self.assertEqual(len(layer.get_updates_for(None)), 1)
- self.assertEqual(len(layer.get_updates_for(x)), 1)
def test_rnn_dynamic_trainability(self):
layer_class = keras.layers.SimpleRNN
diff --git a/tensorflow/python/keras/layers/wrappers_test.py b/tensorflow/python/keras/layers/wrappers_test.py
index 3dc122e..a3173f4 100644
--- a/tensorflow/python/keras/layers/wrappers_test.py
+++ b/tensorflow/python/keras/layers/wrappers_test.py
@@ -787,8 +787,6 @@
layer = keras.layers.Bidirectional(keras.layers.SimpleRNN(3))
_ = layer(x)
assert not layer.updates
- assert not layer.get_updates_for(None)
- assert not layer.get_updates_for(x)
# TODO(b/128684069): Remove when Wrapper sublayers are __call__'d.
with base_layer_utils.call_context().enter(layer, x, True, None):
layer.forward_layer.add_update(x_reachable_update, inputs=x)
@@ -796,8 +794,6 @@
layer.backward_layer.add_update(x_reachable_update, inputs=x)
layer.backward_layer.add_update(1, inputs=None)
assert len(layer.updates) == 4
- assert len(layer.get_updates_for(None)) == 2
- assert len(layer.get_updates_for(x)) == 2
def test_Bidirectional_losses(self):
x = keras.layers.Input(shape=(3, 2))
diff --git a/tensorflow/python/keras/metrics.py b/tensorflow/python/keras/metrics.py
index 7530673..c67b047 100644
--- a/tensorflow/python/keras/metrics.py
+++ b/tensorflow/python/keras/metrics.py
@@ -1925,7 +1925,8 @@
# Add an endpoint "threshold" below zero and above one for either
# threshold method to account for floating point imprecisions.
- self.thresholds = [0.0 - K.epsilon()] + thresholds + [1.0 + K.epsilon()]
+ self._thresholds = np.array([0.0 - K.epsilon()] + thresholds +
+ [1.0 + K.epsilon()])
if isinstance(curve, metrics_utils.AUCCurve):
self.curve = curve
@@ -1959,6 +1960,11 @@
else:
self._build(None)
+ @property
+ def thresholds(self):
+ """The thresholds used for evaluating AUC."""
+ return list(self._thresholds)
+
def _build(self, shape):
"""Initialize TP, FP, TN, and FN tensors, given the shape of the data."""
if self.multi_label:
@@ -2056,7 +2062,7 @@
},
y_true,
y_pred,
- self.thresholds,
+ self._thresholds,
sample_weight=sample_weight,
multi_label=self.multi_label,
label_weights=label_weights)
@@ -3462,4 +3468,3 @@
def is_built_in(cls):
return cls.__module__ == Metric.__module__
-
diff --git a/tensorflow/python/keras/metrics_correctness_test.py b/tensorflow/python/keras/metrics_correctness_test.py
index e209a81..4b71cb5 100644
--- a/tensorflow/python/keras/metrics_correctness_test.py
+++ b/tensorflow/python/keras/metrics_correctness_test.py
@@ -47,15 +47,14 @@
def custom_generator_multi_io(sample_weights=None):
batch_size = 2
- num_samples = 4
- inputs = np.asarray([[1.], [2.], [3.], [4.]])
- targets_1 = np.asarray([[2.], [4.], [6.], [8.]])
- targets_2 = np.asarray([[1.], [2.], [3.], [4.]])
- i = 0
+ num_samples = 5
+ inputs = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ targets_1 = np.asarray([[2.], [4.], [6.], [8.], [10.]])
+ targets_2 = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ start = 0
while True:
- batch_index = i * batch_size % num_samples
- i += 1
- start = batch_index
+ if start > num_samples:
+ start = 0
end = start + batch_size
x = [inputs[start:end], inputs[start:end]]
y = [targets_1[start:end], targets_2[start:end]]
@@ -63,6 +62,7 @@
sw = nest.map_structure(lambda w: w[start:end], sample_weights)
else:
sw = None
+ start = end
yield x, y, sw
@@ -84,97 +84,103 @@
def setUp(self):
super(TestMetricsCorrectnessMultiIO, self).setUp()
- self.x = np.asarray([[1.], [2.], [3.], [4.]])
- self.y1 = np.asarray([[2.], [4.], [6.], [8.]])
- self.y2 = np.asarray([[1.], [2.], [3.], [4.]])
- self.sample_weight_1 = np.asarray([2., 3., 4., 5.])
- self.sample_weight_2 = np.asarray([3.5, 2.5, 1.5, 0.5])
+ self.x = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ self.y1 = np.asarray([[2.], [4.], [6.], [8.], [10.]])
+ self.y2 = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ self.sample_weight_1 = np.asarray([2., 3., 4., 5., 6.])
+ self.sample_weight_2 = np.asarray([3.5, 2.5, 1.5, 0.5, 3.])
- # y_true_1 = [[2.], [4.], [6.], [8.]], y_pred = [[3.], [6.], [9.], [12.]]
- # y_true_2 = [[1.], [2.], [3.], [4.]], y_pred = [[3.], [6.], [9.], [12.]]
+ # y_true_1 = [[2.], [4.], [6.], [8.], [10.]]
+ # y_pred_1 = [[3.], [6.], [9.], [12.], [15.]]
+ # y_true_2 = [[1.], [2.], [3.], [4.], [5.]]
+ # y_pred_2 = [[3.], [6.], [9.], [12.], [15.]]
# Weighted metric `output_1`:
- # Total = ((3 - 2)^2 * 2 + (6 - 4)^2 * 3) +
- # ((9 - 6)^2 * 4 + (12 - 8)^2 * 5)
- # = 130
- # Count = (2 + 3) + (4 + 5)
- # Result = 9.2857141
+ # Total = ((3 - 2)^2 * 2 + (6 - 4)^2 * 3) +
+ # ((9 - 6)^2 * 4 + (12 - 8)^2 * 5) +
+ # ((15 - 10)^2 * 6)
+ # = 280
+ # Count = (2 + 3) + (4 + 5) + 6 = 20
+ # Result = 14
# Weighted metric `output_2`:
# Total = ((3 - 1)^2 * 3.5 + (6 - 2)^2 * 2.5) +
- # ((9 - 3)^2 * 1.5 + (12 - 4)^2 * 0.5)
- # = 140
- # Count = (3.5 + 2.5) + (1.5 + 0.5)
- # Result = 17.5
+ # ((9 - 3)^2 * 1.5 + (12 - 4)^2 * 0.5) +
+ # (15 - 5)^2 * 3.0
+ # = 440
+ # Count = (3.5 + 2.5) + (1.5 + 0.5) + 3.0 = 11.0
+ # Result = 40
# Loss `output_1` with weights:
# Total = ((3 - 2)^2 * 2 + (6 - 4)^2 * 3) +
- # ((9 - 6)^2 * 4 + (12 - 8)^2 * 5)
- # = 130
- # Count = 2 + 2
- # Result = 32.5
+ # ((9 - 6)^2 * 4 + (12 - 8)^2 * 5) +
+ # ((15 - 10)^2 * 6)
+ # = 280
+ # Count = 2 + 2 + 1
+ # Result = 56
# Loss `output_1` without weights/Metric `output_1`:
- # Total = ((3 - 2)^2 + (6 - 4)^2) + ((9 - 6)^2 + (12 - 8)^2) = 30
- # Count = 2 + 2
- # Result = 7.5
+ # Total = ((3 - 2)^2 + (6 - 4)^2) + ((9 - 6)^2 + (12 - 8)^2) + (15 - 10)^2
+ # = 55
+ # Count = 2 + 2 + 1
+ # Result = 11
# Loss `output_2` with weights:
# Total = ((3 - 1)^2 * 3.5 + (6 - 2)^2 * 2.5) +
- # ((9 - 3)^2 * 1.5 + (12 - 4)^2 * 0.5)
- # = 140
- # Count = 2 + 2
- # Result = 35
+ # ((9 - 3)^2 * 1.5 + (12 - 4)^2 * 0.5) +
+ # (15 - 5)^2 * 3.0
+ # = 440
+ # Count = 2 + 2 + 1
+ # Result = 88
# Loss `output_2` without weights/Metric `output_2`:
- # Total = ((3 - 1)^2 + (6 - 2)^2) + ((9 - 3)^2 + (12 - 4)^2) = 120
- # Count = 2 + 2
- # Result = 30
+ # Total = ((3 - 1)^2 + (6 - 2)^2) + ((9 - 3)^2 + (12 - 4)^2) + (15 - 5)^2
+ # = 220
+ # Count = 2 + 2 + 1
+ # Result = 44
- # Total loss with weights = 32.5 + 35 = 67.5
- # Total loss without weights = 7.5 + 30 = 37.5
+ # Total loss with weights = 56 + 88 = 144
+ # Total loss without weights = 11 + 44 = 55
self.wmse = 'mean_squared_error_2'
self.expected_fit_result_with_weights = {
- 'output_1_mean_squared_error': [7.5, 7.5],
- 'output_2_mean_squared_error': [30, 30],
- 'output_1_' + self.wmse: [9.286, 9.286],
- 'output_2_' + self.wmse: [17.5, 17.5],
- 'loss': [67.5, 67.5],
- 'output_1_loss': [32.5, 32.5],
- 'output_2_loss': [35, 35],
+ 'output_1_mean_squared_error': [11, 11],
+ 'output_2_mean_squared_error': [44, 44],
+ 'output_1_' + self.wmse: [14, 14],
+ 'output_2_' + self.wmse: [40, 40],
+ 'loss': [144, 144],
+ 'output_1_loss': [56, 56],
+ 'output_2_loss': [88, 88],
}
self.expected_fit_result_with_weights_output_2 = {
- 'output_1_mean_squared_error': [7.5, 7.5],
- 'output_2_mean_squared_error': [30, 30],
- 'output_1_' + self.wmse: [7.5, 7.5],
- 'output_2_' + self.wmse: [17.5, 17.5],
- 'loss': [42.5, 42.5],
- 'output_1_loss': [7.5, 7.5],
- 'output_2_loss': [35, 35],
+ 'output_1_mean_squared_error': [11, 11],
+ 'output_2_mean_squared_error': [44, 44],
+ 'output_1_' + self.wmse: [11, 11],
+ 'output_2_' + self.wmse: [40, 40],
+ 'loss': [99, 99],
+ 'output_1_loss': [11, 11],
+ 'output_2_loss': [88, 88],
}
self.expected_fit_result = {
- 'output_1_mean_squared_error': [7.5, 7.5],
- 'output_2_mean_squared_error': [30, 30],
- 'output_1_' + self.wmse: [7.5, 7.5],
- 'output_2_' + self.wmse: [30, 30],
- 'loss': [37.5, 37.5],
- 'output_1_loss': [7.5, 7.5],
- 'output_2_loss': [30, 30],
+ 'output_1_mean_squared_error': [11, 11],
+ 'output_2_mean_squared_error': [44, 44],
+ 'output_1_' + self.wmse: [11, 11],
+ 'output_2_' + self.wmse: [44, 44],
+ 'loss': [55, 55],
+ 'output_1_loss': [11, 11],
+ 'output_2_loss': [44, 44],
}
# In the order: 'loss', 'output_1_loss', 'output_2_loss',
# 'output_1_mean_squared_error', 'output_1_mean_squared_error_2',
# 'output_2_mean_squared_error', 'output_2_mean_squared_error_2'
- self.expected_batch_result_with_weights = [
- 67.5, 32.5, 35, 7.5, 9.286, 30, 17.5
- ]
+ self.expected_batch_result_with_weights = [144, 56, 88, 11, 14, 44, 40]
self.expected_batch_result_with_weights_output_2 = [
- 42.5, 7.5, 35, 7.5, 7.5, 30, 17.5
+ 99, 11, 88, 11, 11, 44, 40
]
- self.expected_batch_result = [37.5, 7.5, 30, 7.5, 7.5, 30, 30]
+ self.expected_batch_result = [55, 11, 44, 11, 11, 44, 44]
def test_fit(self):
model = self._get_compiled_multi_io_model()
@@ -291,7 +297,7 @@
def test_fit_generator(self):
model = self._get_compiled_multi_io_model()
history = model.fit_generator(
- custom_generator_multi_io(), steps_per_epoch=2, epochs=2)
+ custom_generator_multi_io(), steps_per_epoch=3, epochs=2)
for key, value in self.expected_fit_result.items():
self.assertAllClose(history.history[key], value, 1e-3)
@@ -300,7 +306,7 @@
history = model.fit_generator(
custom_generator_multi_io(
sample_weights=[self.sample_weight_1, self.sample_weight_2]),
- steps_per_epoch=2,
+ steps_per_epoch=3,
epochs=2)
for key, value in self.expected_fit_result_with_weights.items():
self.assertAllClose(history.history[key], value, 1e-3)
@@ -309,14 +315,14 @@
history = model.fit_generator(
custom_generator_multi_io(
sample_weights={'output_2': self.sample_weight_2}),
- steps_per_epoch=2,
+ steps_per_epoch=3,
epochs=2)
for key, value in self.expected_fit_result_with_weights_output_2.items():
self.assertAllClose(history.history[key], value, 1e-3)
def test_eval_generator(self):
model = self._get_compiled_multi_io_model()
- eval_result = model.evaluate_generator(custom_generator_multi_io(), steps=2)
+ eval_result = model.evaluate_generator(custom_generator_multi_io(), steps=3)
self.assertAllClose(eval_result, self.expected_batch_result, 1e-3)
def test_eval_generator_with_sample_weight(self):
@@ -324,7 +330,7 @@
eval_result = model.evaluate_generator(
custom_generator_multi_io(
sample_weights=[self.sample_weight_1, self.sample_weight_2]),
- steps=2)
+ steps=3)
self.assertAllClose(eval_result, self.expected_batch_result_with_weights,
1e-3)
@@ -332,7 +338,7 @@
eval_result = model.evaluate_generator(
custom_generator_multi_io(
sample_weights={'output_2': self.sample_weight_2}),
- steps=2)
+ steps=3)
self.assertAllClose(eval_result,
self.expected_batch_result_with_weights_output_2, 1e-3)
@@ -549,7 +555,7 @@
@keras_parameterized.run_with_all_model_types(exclude_models=['sequential'])
-@keras_parameterized.run_all_keras_modes
+@keras_parameterized.run_all_keras_modes(always_skip_v1=True)
@parameterized.parameters([
loss_reduction.ReductionV2.SUM_OVER_BATCH_SIZE,
loss_reduction.ReductionV2.AUTO,
@@ -567,29 +573,34 @@
def setUp(self):
super(TestOutputLossMetrics, self).setUp()
- self.x = np.asarray([[1.], [2.], [3.], [4.]])
- self.y1 = np.asarray([[2.], [4.], [6.], [8.]])
- self.y2 = np.asarray([[1.], [2.], [3.], [4.]])
- self.sample_weight_1 = np.asarray([2., 3., 4., 5.])
- self.sample_weight_2 = np.asarray([3.5, 2.5, 1.5, 0.5])
+ self.x = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ self.y1 = np.asarray([[2.], [4.], [6.], [8.], [10.]])
+ self.y2 = np.asarray([[1.], [2.], [3.], [4.], [5.]])
+ self.sample_weight_1 = np.asarray([2., 3., 4., 5., 6.])
+ self.sample_weight_2 = np.asarray([3.5, 2.5, 1.5, 0.5, 3.])
- # y_true = [[2.], [4.], [6.], [8.]], y_pred = [[3.], [6.], [9.], [12.]]
+ # y_true_1 = [[2.], [4.], [6.], [8.], [10.]]
+ # y_pred_1 = [[3.], [6.], [9.], [12.], [15.]]
+ # y_true_2 = [[1.], [2.], [3.], [4.], [5.]]
+ # y_pred_2 = [[3.], [6.], [9.], [12.], [15.]]
# Loss `output_1`:
# Per-sample weighted losses
# Batch 1 = [(3 - 2)^2 * 2, (6 - 4)^2 * 3)] = [2, 12]
# Batch 2 = [((9 - 6)^2 * 4, (12 - 8)^2 * 5)] = [36, 80]
+ # Batch 3 = [(15 - 10)^2 * 6] = [150]
- # Result (reduction=SUM) = ((2 + 12) + (36 + 80))/2 = 65
- # Result (reduction=SUM_OVER_BATCH_SIZE/AUTO/NONE) = 130 / 4 = 32.5
+ # Result (reduction=SUM) = ((2 + 12)*2 + (36 + 80)*2 + 150) / 5 = 82
+ # Result (reduction=SUM_OVER_BATCH_SIZE/AUTO/NONE) = 280 / 5 = 56
# Loss `output_2`:
# Per-sample weighted losses
# Batch 1 = [(3 - 1)^2 * 3.5, (6 - 2)^2 * 2.5)] = [14, 40]
# Batch 2 = [(9 - 3)^2 * 1.5, (12 - 4)^2 * 0.5)] = [54, 32]
+ # Batch 3 = [(15 - 5)^2 * 3] = [300]
- # Result (reduction=SUM) = ((14 + 40) + (54 + 32))/2 = 70
- # Result (reduction=SUM_OVER_BATCH_SIZE/AUTO/NONE) = 140 / 4 = 35
+ # Result (reduction=SUM) = ((14 + 40)*2 + (54 + 32)*2 + 300) / 5 = 116
+ # Result (reduction=SUM_OVER_BATCH_SIZE/AUTO/NONE) = 440 / 5 = 88
# When reduction is 'NONE' loss value that is passed to the optimizer will
# be vector loss but what is reported is a scalar, which is an average of
@@ -598,18 +609,18 @@
# Total loss = Output_loss_1 + Output_loss_2
sum_over_batch_size_fit_result = {
- 'loss': [67.5, 67.5],
- 'output_1_loss': [32.5, 32.5],
- 'output_2_loss': [35, 35],
+ 'loss': [144, 144],
+ 'output_1_loss': [56, 56],
+ 'output_2_loss': [88, 88],
}
self.expected_fit_result = {
loss_reduction.ReductionV2.NONE:
sum_over_batch_size_fit_result,
loss_reduction.ReductionV2.SUM: {
- 'loss': [135, 135],
- 'output_1_loss': [65, 65],
- 'output_2_loss': [70, 70],
+ 'loss': [198, 198],
+ 'output_1_loss': [82, 82],
+ 'output_2_loss': [116, 116],
},
loss_reduction.ReductionV2.AUTO:
sum_over_batch_size_fit_result,
@@ -619,12 +630,16 @@
# In the order: 'loss', 'output_1_loss', 'output_2_loss',
self.expected_batch_result = {
- loss_reduction.ReductionV2.NONE: [67.5, 32.5, 35],
- loss_reduction.ReductionV2.SUM: [135, 65, 70],
- loss_reduction.ReductionV2.AUTO: [67.5, 32.5, 35],
- loss_reduction.ReductionV2.SUM_OVER_BATCH_SIZE: [67.5, 32.5, 35],
+ loss_reduction.ReductionV2.NONE: [144, 56, 88],
+ loss_reduction.ReductionV2.SUM: [198, 82, 116],
+ loss_reduction.ReductionV2.AUTO: [144, 56, 88],
+ loss_reduction.ReductionV2.SUM_OVER_BATCH_SIZE: [144, 56, 88],
}
+ # 2 + 12 + 36 + 80 + 150 = 280
+ # 14 + 40 + 54 + 32 + 300 = 440
+ self.expected_single_batch_result = [720, 280, 440]
+
def test_fit(self, reduction):
model = self._get_compiled_multi_io_model(
loss=losses.MeanSquaredError(reduction=reduction))
@@ -661,8 +676,7 @@
expected_values = self.expected_batch_result[reduction]
if reduction == loss_reduction.ReductionV2.SUM:
- # We are taking all the data as one batch, so undo the averaging here.
- expected_values = [x * 2 for x in self.expected_batch_result[reduction]]
+ expected_values = self.expected_single_batch_result
self.assertAllClose(result, expected_values)
def test_test_on_batch(self, reduction):
@@ -675,8 +689,7 @@
})
expected_values = self.expected_batch_result[reduction]
if reduction == loss_reduction.ReductionV2.SUM:
- # We are taking all the data as one batch, so undo the averaging here.
- expected_values = [x * 2 for x in self.expected_batch_result[reduction]]
+ expected_values = self.expected_single_batch_result
self.assertAllClose(result, expected_values)
def test_fit_generator(self, reduction):
@@ -685,7 +698,7 @@
history = model.fit_generator(
custom_generator_multi_io(
sample_weights=[self.sample_weight_1, self.sample_weight_2]),
- steps_per_epoch=2,
+ steps_per_epoch=3,
epochs=2)
for key, value in self.expected_fit_result[reduction].items():
self.assertAllClose(history.history[key], value)
@@ -696,7 +709,7 @@
eval_result = model.evaluate_generator(
custom_generator_multi_io(
sample_weights=[self.sample_weight_1, self.sample_weight_2]),
- steps=2)
+ steps=3)
self.assertAllClose(eval_result, self.expected_batch_result[reduction])
diff --git a/tensorflow/python/keras/models_test.py b/tensorflow/python/keras/models_test.py
index 17e0d9e..d7ef7d7 100644
--- a/tensorflow/python/keras/models_test.py
+++ b/tensorflow/python/keras/models_test.py
@@ -124,7 +124,7 @@
self.assertEqual(new_model._is_graph_network, model._is_graph_network)
if input_shape:
# update ops from batch norm needs to be included
- self.assertEqual(len(new_model.get_updates_for(new_model.inputs)), 2)
+ self.assertGreaterEqual(len(new_model.updates), 2)
# On top of new tensor -- clone model should always have an InputLayer.
input_a = keras.Input(shape=(4,))
@@ -173,7 +173,7 @@
# With placeholder creation
new_model = clone_fn(model)
- self.assertEqual(len(new_model.get_updates_for(new_model.inputs)), 2)
+ self.assertGreaterEqual(len(new_model.updates), 2)
new_model.compile(
testing_utils.get_v2_optimizer('rmsprop'),
'mse',
@@ -185,7 +185,7 @@
input_b = keras.Input(shape=(4,), name='b')
new_model = keras.models.clone_model(
model, input_tensors=[input_a, input_b])
- self.assertEqual(len(new_model.get_updates_for(new_model.inputs)), 2)
+ self.assertLen(new_model.updates, 2)
new_model.compile(
testing_utils.get_v2_optimizer('rmsprop'),
'mse',
@@ -199,7 +199,7 @@
input_a = keras.backend.variable(val_a)
input_b = keras.backend.variable(val_b)
new_model = clone_fn(model, input_tensors=[input_a, input_b])
- self.assertEqual(len(new_model.get_updates_for(new_model.inputs)), 2)
+ self.assertGreaterEqual(len(new_model.updates), 2)
new_model.compile(
testing_utils.get_v2_optimizer('rmsprop'),
'mse',
diff --git a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
index 1c1db4c..4cf0703 100644
--- a/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
+++ b/tensorflow/python/keras/optimizer_v2/optimizer_v2.py
@@ -1274,7 +1274,7 @@
# pylint: disable=protected-access
# Get the distributed variable if it exists.
if hasattr(var, "_distributed_container"):
- var = var._distributed_container()
+ var = var._distributed_container
if var._in_graph_mode:
return var._shared_name
return var._unique_id
diff --git a/tensorflow/python/keras/preprocessing/BUILD b/tensorflow/python/keras/preprocessing/BUILD
index 3cfdb1e..403bc6e 100644
--- a/tensorflow/python/keras/preprocessing/BUILD
+++ b/tensorflow/python/keras/preprocessing/BUILD
@@ -31,6 +31,7 @@
py_library(
name = "image",
srcs = [
+ "dataset_utils.py",
"image.py",
"image_dataset.py",
],
@@ -69,7 +70,9 @@
py_library(
name = "text",
srcs = [
+ "dataset_utils.py",
"text.py",
+ "text_dataset.py",
],
deps = ["//tensorflow/python:util"],
)
@@ -125,6 +128,19 @@
)
tf_py_test(
+ name = "text_dataset_test",
+ size = "small",
+ srcs = ["text_dataset_test.py"],
+ python_version = "PY3",
+ deps = [
+ ":text",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python/compat:v2_compat",
+ "//tensorflow/python/keras",
+ ],
+)
+
+tf_py_test(
name = "timeseries_test",
size = "small",
srcs = ["timeseries_test.py"],
diff --git a/tensorflow/python/keras/preprocessing/dataset_utils.py b/tensorflow/python/keras/preprocessing/dataset_utils.py
new file mode 100644
index 0000000..70d9566
--- /dev/null
+++ b/tensorflow/python/keras/preprocessing/dataset_utils.py
@@ -0,0 +1,201 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Keras image dataset loading utilities."""
+# pylint: disable=g-classes-have-attributes
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import multiprocessing
+import os
+
+import numpy as np
+
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import math_ops
+
+
+def index_directory(directory,
+ labels,
+ formats,
+ class_names=None,
+ shuffle=True,
+ seed=None,
+ follow_links=False):
+ """Make list of all files in the subdirs of `directory`, with their labels.
+
+ Args:
+ directory: The target directory (string).
+ labels: Either "inferred"
+ (labels are generated from the directory structure),
+ or a list/tuple of integer labels of the same size as the number of
+ valid files found in the directory. Labels should be sorted according
+ to the alphanumeric order of the image file paths
+ (obtained via `os.walk(directory)` in Python).
+ formats: Whitelist of file extensions to index (e.g. ".jpg", ".txt").
+ class_names: Only valid if "labels" is "inferred". This is the explict
+ list of class names (must match names of subdirectories). Used
+ to control the order of the classes
+ (otherwise alphanumerical order is used).
+ shuffle: Whether to shuffle the data. Default: True.
+ If set to False, sorts the data in alphanumeric order.
+ seed: Optional random seed for shuffling.
+ follow_links: Whether to visits subdirectories pointed to by symlinks.
+
+ Returns:
+ tuple (file_paths, labels, class_names).
+ file_paths: list of file paths (strings).
+ labels: list of matching integer labels (same length as file_paths)
+ class_names: names of the classes corresponding to these labels, in order.
+ """
+ inferred_class_names = []
+ for subdir in sorted(os.listdir(directory)):
+ if os.path.isdir(os.path.join(directory, subdir)):
+ inferred_class_names.append(subdir)
+ if not class_names:
+ class_names = inferred_class_names
+ else:
+ if set(class_names) != set(inferred_class_names):
+ raise ValueError(
+ 'The `class_names` passed did not match the '
+ 'names of the subdirectories of the target directory. '
+ 'Expected: %s, but received: %s' %
+ (inferred_class_names, class_names))
+ class_indices = dict(zip(class_names, range(len(class_names))))
+
+ # Build an index of the files
+ # in the different class subfolders.
+ pool = multiprocessing.pool.ThreadPool()
+ results = []
+ filenames = []
+ for dirpath in (os.path.join(directory, subdir) for subdir in class_names):
+ results.append(
+ pool.apply_async(index_subdirectory,
+ (dirpath, class_indices, follow_links, formats)))
+ labels_list = []
+ for res in results:
+ partial_filenames, partial_labels = res.get()
+ labels_list.append(partial_labels)
+ filenames += partial_filenames
+ if labels != 'inferred':
+ if len(labels) != len(filenames):
+ raise ValueError('Expected the lengths of `labels` to match the number '
+ 'of files in the target directory. len(labels) is %s '
+ 'while we found %s files in %s.' % (
+ len(labels), len(filenames), directory))
+ else:
+ i = 0
+ labels = np.zeros((len(filenames),), dtype='int32')
+ for partial_labels in labels_list:
+ labels[i:i + len(partial_labels)] = partial_labels
+ i += len(partial_labels)
+
+ print('Found %d files belonging to %d classes.' %
+ (len(filenames), len(class_names)))
+ pool.close()
+ pool.join()
+ file_paths = [os.path.join(directory, fname) for fname in filenames]
+
+ if shuffle:
+ # Shuffle globally to erase macro-structure
+ if seed is None:
+ seed = np.random.randint(1e6)
+ rng = np.random.RandomState(seed)
+ rng.shuffle(file_paths)
+ rng = np.random.RandomState(seed)
+ rng.shuffle(labels)
+ return file_paths, labels, class_names
+
+
+def iter_valid_files(directory, follow_links, formats):
+ walk = os.walk(directory, followlinks=follow_links)
+ for root, _, files in sorted(walk, key=lambda x: x[0]):
+ for fname in sorted(files):
+ if fname.lower().endswith(formats):
+ yield root, fname
+
+
+def index_subdirectory(directory, class_indices, follow_links, formats):
+ """Recursively walks directory and list image paths and their class index.
+
+ Arguments:
+ directory: string, target directory.
+ class_indices: dict mapping class names to their index.
+ follow_links: boolean, whether to recursively follow subdirectories
+ (if False, we only list top-level images in `directory`).
+ formats: Whitelist of file extensions to index (e.g. ".jpg", ".txt").
+
+ Returns:
+ tuple `(filenames, labels)`. `filenames` is a list of relative file
+ paths, and `labels` is a list of integer labels corresponding to these
+ files.
+ """
+ dirname = os.path.basename(directory)
+ valid_files = iter_valid_files(directory, follow_links, formats)
+ labels = []
+ filenames = []
+ for root, fname in valid_files:
+ labels.append(class_indices[dirname])
+ absolute_path = os.path.join(root, fname)
+ relative_path = os.path.join(
+ dirname, os.path.relpath(absolute_path, directory))
+ filenames.append(relative_path)
+ return filenames, labels
+
+
+def get_training_or_validation_split(samples, labels, validation_split, subset):
+ """Potentially restict samples & labels to a training or validation split.
+
+ Args:
+ samples: List of elements.
+ labels: List of corresponding labels.
+ validation_split: Float, fraction of data to reserve for validation.
+ subset: Subset of the data to return.
+ Either "training", "validation", or None. If None, we return all of the
+ data.
+
+ Returns:
+ tuple (samples, labels), potentially restricted to the specified subset.
+ """
+ if validation_split:
+ if not 0 < validation_split < 1:
+ raise ValueError(
+ '`validation_split` must be between 0 and 1, received: %s' %
+ (validation_split,))
+ if subset is None:
+ return samples, labels
+
+ num_val_samples = int(validation_split * len(samples))
+ if subset == 'training':
+ samples = samples[:-num_val_samples]
+ labels = labels[:-num_val_samples]
+ elif subset == 'validation':
+ samples = samples[-num_val_samples:]
+ labels = labels[-num_val_samples:]
+ else:
+ raise ValueError('`subset` must be either "training" '
+ 'or "validation", received: %s' % (subset,))
+ return samples, labels
+
+
+def labels_to_dataset(labels, label_mode, num_classes):
+ label_ds = dataset_ops.Dataset.from_tensor_slices(labels)
+ if label_mode == 'binary':
+ label_ds = label_ds.map(
+ lambda x: array_ops.expand_dims(math_ops.cast(x, 'float32'), axis=-1))
+ elif label_mode == 'categorical':
+ label_ds = label_ds.map(lambda x: array_ops.one_hot(x, num_classes))
+ return label_ds
diff --git a/tensorflow/python/keras/preprocessing/image_dataset.py b/tensorflow/python/keras/preprocessing/image_dataset.py
index 500a41f..2e24ef8 100644
--- a/tensorflow/python/keras/preprocessing/image_dataset.py
+++ b/tensorflow/python/keras/preprocessing/image_dataset.py
@@ -18,17 +18,13 @@
from __future__ import division
from __future__ import print_function
-import multiprocessing
-import os
-
import numpy as np
from tensorflow.python.data.ops import dataset_ops
from tensorflow.python.keras.layers.preprocessing import image_preprocessing
-from tensorflow.python.ops import array_ops
+from tensorflow.python.keras.preprocessing import dataset_utils
from tensorflow.python.ops import image_ops
from tensorflow.python.ops import io_ops
-from tensorflow.python.ops import math_ops
from tensorflow.python.util.tf_export import keras_export
@@ -49,7 +45,7 @@
subset=None,
interpolation='bilinear',
follow_links=False):
- """Generates a Dataset from image files in a directory.
+ """Generates a `tf.data.Dataset` from image files in a directory.
If your directory structure is:
@@ -63,10 +59,10 @@
......b_image_2.jpg
```
- Then calling `from_directory(main_directory, labels='inferred')`
- will return a Dataset that yields batches of images from
+ Then calling `image_dataset_from_directory(main_directory, labels='inferred')`
+ will return a `tf.data.Dataset` that yields batches of images from
the subdirectories `class_a` and `class_b`, together with labels
- 0 and 1 (0 corresponding to class_a and 1 corresponding to class_b).
+ 0 and 1 (0 corresponding to `class_a` and 1 corresponding to `class_b`).
Supported image formats: jpeg, png, bmp, gif.
Animated gifs are truncated to the first frame.
@@ -126,22 +122,22 @@
has shape `(batch_size, image_size[0], image_size[1], num_channels)`,
and `labels` follows the format described below.
- Rules regarding labels format:
- - if `label_mode` is `int`, the labels are an `int32` tensor of shape
- `(batch_size,)`.
- - if `label_mode` is `binary`, the labels are a `float32` tensor of
- 1s and 0s of shape `(batch_size, 1)`.
- - if `label_mode` is `categorial`, the labels are a `float32` tensor
- of shape `(batch_size, num_classes)`, representing a one-hot
- encoding of the class index.
+ Rules regarding labels format:
+ - if `label_mode` is `int`, the labels are an `int32` tensor of shape
+ `(batch_size,)`.
+ - if `label_mode` is `binary`, the labels are a `float32` tensor of
+ 1s and 0s of shape `(batch_size, 1)`.
+ - if `label_mode` is `categorial`, the labels are a `float32` tensor
+ of shape `(batch_size, num_classes)`, representing a one-hot
+ encoding of the class index.
- Rules regarding number of channels in the yielded images:
- - if `color_mode` is `grayscale`,
- there's 1 channel in the image tensors.
- - if `color_mode` is `rgb`,
- there are 3 channel in the image tensors.
- - if `color_mode` is `rgba`,
- there are 4 channel in the image tensors.
+ Rules regarding number of channels in the yielded images:
+ - if `color_mode` is `grayscale`,
+ there's 1 channel in the image tensors.
+ - if `color_mode` is `rgb`,
+ there are 3 channel in the image tensors.
+ - if `color_mode` is `rgba`,
+ there are 4 channel in the image tensors.
"""
if labels != 'inferred':
if not isinstance(labels, (list, tuple)):
@@ -172,85 +168,25 @@
'Received: %s' % (color_mode,))
interpolation = image_preprocessing.get_interpolation(interpolation)
- inferred_class_names = []
- for subdir in sorted(os.listdir(directory)):
- if os.path.isdir(os.path.join(directory, subdir)):
- inferred_class_names.append(subdir)
- if not class_names:
- class_names = inferred_class_names
- else:
- if set(class_names) != set(inferred_class_names):
- raise ValueError(
- 'The `class_names` passed did not match the '
- 'names of the subdirectories of the target directory. '
- 'Expected: %s, but received: %s' %
- (inferred_class_names, class_names))
- class_indices = dict(zip(class_names, range(len(class_names))))
+ if seed is None:
+ seed = np.random.randint(1e6)
+ image_paths, labels, class_names = dataset_utils.index_directory(
+ directory,
+ labels,
+ formats=WHITELIST_FORMATS,
+ class_names=class_names,
+ shuffle=shuffle,
+ seed=seed,
+ follow_links=follow_links)
if label_mode == 'binary' and len(class_names) != 2:
raise ValueError(
'When passing `label_mode="binary", there must exactly 2 classes. '
'Found the following classes: %s' % (class_names,))
- # Build an index of the images
- # in the different class subfolders.
- pool = multiprocessing.pool.ThreadPool()
- results = []
- filenames = []
- for dirpath in (os.path.join(directory, subdir) for subdir in class_names):
- results.append(
- pool.apply_async(list_labeled_images_in_directory,
- (dirpath, class_indices, follow_links)))
- labels_list = []
- for res in results:
- partial_labels, partial_filenames = res.get()
- labels_list.append(partial_labels)
- filenames += partial_filenames
- if labels != 'inferred':
- if len(labels) != len(filenames):
- raise ValueError('Expected the lengths of `labels` to match the number '
- 'of images in the target directory. len(labels) is %s '
- 'while we found %s images in %s.' % (
- len(labels), len(filenames), directory))
- else:
- i = 0
- labels = np.zeros((len(filenames),), dtype='int32')
- for partial_labels in labels_list:
- labels[i:i + len(partial_labels)] = partial_labels
- i += len(partial_labels)
+ image_paths, labels = dataset_utils.get_training_or_validation_split(
+ image_paths, labels, validation_split, subset)
- print('Found %d images belonging to %d classes.' %
- (len(filenames), len(class_names)))
- pool.close()
- pool.join()
- image_paths = [os.path.join(directory, fname) for fname in filenames]
-
- if shuffle:
- # Shuffle globally to erase macro-structure
- # (the dataset will be further shuffled within a local buffer
- # at each iteration)
- if seed is None:
- seed = np.random.randint(1e6)
- rng = np.random.RandomState(seed)
- rng.shuffle(image_paths)
- rng = np.random.RandomState(seed)
- rng.shuffle(labels)
-
- if validation_split:
- if not 0 < validation_split < 1:
- raise ValueError(
- '`validation_split` must be between 0 and 1, received: %s' %
- (validation_split,))
- num_val_samples = int(validation_split * len(image_paths))
- if subset == 'training':
- image_paths = image_paths[:-num_val_samples]
- labels = labels[:-num_val_samples]
- elif subset == 'validation':
- image_paths = image_paths[-num_val_samples:]
- labels = labels[-num_val_samples:]
- else:
- raise ValueError('`subset` must be either "training" '
- 'or "validation", received: %s' % (subset,))
dataset = paths_and_labels_to_dataset(
image_paths=image_paths,
image_size=image_size,
@@ -263,6 +199,8 @@
# Shuffle locally at each iteration
dataset = dataset.shuffle(buffer_size=batch_size * 8, seed=seed)
dataset = dataset.batch(batch_size)
+ # Users may need to reference `class_names`.
+ dataset.class_names = class_names
return dataset
@@ -279,51 +217,11 @@
img_ds = path_ds.map(
lambda x: path_to_image(x, image_size, num_channels, interpolation))
if label_mode:
- label_ds = dataset_ops.Dataset.from_tensor_slices(labels)
- if label_mode == 'binary':
- label_ds = label_ds.map(
- lambda x: array_ops.expand_dims(math_ops.cast(x, 'float32'), axis=-1))
- elif label_mode == 'categorical':
- label_ds = label_ds.map(lambda x: array_ops.one_hot(x, num_classes))
+ label_ds = dataset_utils.labels_to_dataset(labels, label_mode, num_classes)
img_ds = dataset_ops.Dataset.zip((img_ds, label_ds))
return img_ds
-def iter_valid_files(directory, follow_links):
- walk = os.walk(directory, followlinks=follow_links)
- for root, _, files in sorted(walk, key=lambda x: x[0]):
- for fname in sorted(files):
- if fname.lower().endswith(WHITELIST_FORMATS):
- yield root, fname
-
-
-def list_labeled_images_in_directory(directory, class_indices, follow_links):
- """Recursively walks directory and list image paths and their class index.
-
- Arguments:
- directory: string, target directory.
- class_indices: dict mapping class names to their index.
- follow_links: boolean, whether to recursively follow subdirectories
- (if False, we only list top-level images in `directory`).
-
- Returns:
- tuple `(labels, filenames)`. `labels` is a list of integer
- labels and `filenames` is a list of relative image paths corresponding
- to these labels.
- """
- dirname = os.path.basename(directory)
- valid_files = iter_valid_files(directory, follow_links)
- labels = []
- filenames = []
- for root, fname in valid_files:
- labels.append(class_indices[dirname])
- absolute_path = os.path.join(root, fname)
- relative_path = os.path.join(
- dirname, os.path.relpath(absolute_path, directory))
- filenames.append(relative_path)
- return labels, filenames
-
-
def path_to_image(path, image_size, num_channels, interpolation):
img = io_ops.read_file(path)
img = image_ops.decode_image(
diff --git a/tensorflow/python/keras/preprocessing/image_dataset_test.py b/tensorflow/python/keras/preprocessing/image_dataset_test.py
index 629f03d..aa10c1c 100644
--- a/tensorflow/python/keras/preprocessing/image_dataset_test.py
+++ b/tensorflow/python/keras/preprocessing/image_dataset_test.py
@@ -35,7 +35,7 @@
PIL = None
-class DatasetFromDirectoryTest(keras_parameterized.TestCase):
+class ImageDatasetFromDirectoryTest(keras_parameterized.TestCase):
def _get_images(self, count=16, color_mode='rgb'):
width = height = 24
@@ -262,7 +262,7 @@
with self.assertRaisesRegex(
ValueError,
- 'Expected the lengths of `labels` to match the number of images'):
+ 'Expected the lengths of `labels` to match the number of files'):
_ = image_dataset.image_dataset_from_directory(
directory, labels=[0, 0, 1, 1])
diff --git a/tensorflow/python/keras/preprocessing/text.py b/tensorflow/python/keras/preprocessing/text.py
index e501789a..5a343e0 100644
--- a/tensorflow/python/keras/preprocessing/text.py
+++ b/tensorflow/python/keras/preprocessing/text.py
@@ -21,6 +21,7 @@
from keras_preprocessing import text
+from tensorflow.python.keras.preprocessing.text_dataset import text_dataset_from_directory # pylint: disable=unused-import
from tensorflow.python.util.tf_export import keras_export
hashing_trick = text.hashing_trick
diff --git a/tensorflow/python/keras/preprocessing/text_dataset.py b/tensorflow/python/keras/preprocessing/text_dataset.py
new file mode 100644
index 0000000..6a57e99
--- /dev/null
+++ b/tensorflow/python/keras/preprocessing/text_dataset.py
@@ -0,0 +1,188 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Keras text dataset generation utilities."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+from tensorflow.python.data.ops import dataset_ops
+from tensorflow.python.keras.preprocessing import dataset_utils
+from tensorflow.python.ops import io_ops
+from tensorflow.python.ops import string_ops
+from tensorflow.python.util.tf_export import keras_export
+
+
+@keras_export('keras.preprocessing.text_dataset_from_directory', v1=[])
+def text_dataset_from_directory(directory,
+ labels='inferred',
+ label_mode='int',
+ class_names=None,
+ batch_size=32,
+ max_length=None,
+ shuffle=True,
+ seed=None,
+ validation_split=None,
+ subset=None,
+ follow_links=False):
+ """Generates a `tf.data.Dataset` from text files in a directory.
+
+ If your directory structure is:
+
+ ```
+ main_directory/
+ ...class_a/
+ ......a_text_1.txt
+ ......a_text_2.txt
+ ...class_b/
+ ......b_text_1.txt
+ ......b_text_2.txt
+ ```
+
+ Then calling `text_dataset_from_directory(main_directory, labels='inferred')`
+ will return a `tf.data.Dataset` that yields batches of texts from
+ the subdirectories `class_a` and `class_b`, together with labels
+ 0 and 1 (0 corresponding to `class_a` and 1 corresponding to `class_b`).
+
+ Only `.txt` files are supported at this time.
+
+ Arguments:
+ directory: Directory where the data is located.
+ If `labels` is "inferred", it should contain
+ subdirectories, each containing text files for a class.
+ Otherwise, the directory structure is ignored.
+ labels: Either "inferred"
+ (labels are generated from the directory structure),
+ or a list/tuple of integer labels of the same size as the number of
+ text files found in the directory. Labels should be sorted according
+ to the alphanumeric order of the text file paths
+ (obtained via `os.walk(directory)` in Python).
+ label_mode:
+ - 'int': means that the labels are encoded as integers
+ (e.g. for `sparse_categorical_crossentropy` loss).
+ - 'categorical' means that the labels are
+ encoded as a categorical vector
+ (e.g. for `categorical_crossentropy` loss).
+ - 'binary' means that the labels (there can be only 2)
+ are encoded as `float32` scalars with values 0 or 1
+ (e.g. for `binary_crossentropy`).
+ - None (no labels).
+ class_names: Only valid if "labels" is "inferred". This is the explict
+ list of class names (must match names of subdirectories). Used
+ to control the order of the classes
+ (otherwise alphanumerical order is used).
+ batch_size: Size of the batches of data. Default: 32.
+ max_length: Maximum size of a text string. Texts longer than this will
+ be truncated to `max_length`.
+ shuffle: Whether to shuffle the data. Default: True.
+ If set to False, sorts the data in alphanumeric order.
+ seed: Optional random seed for shuffling and transformations.
+ validation_split: Optional float between 0 and 1,
+ fraction of data to reserve for validation.
+ subset: One of "training" or "validation".
+ Only used if `validation_split` is set.
+ follow_links: Whether to visits subdirectories pointed to by symlinks.
+ Defaults to False.
+
+ Returns:
+ A `tf.data.Dataset` object.
+ - If `label_mode` is None, it yields `string` tensors of shape
+ `(batch_size,)`, containing the contents of a batch of text files.
+ - Otherwise, it yields a tuple `(texts, labels)`, where `texts`
+ has shape `(batch_size,)` and `labels` follows the format described
+ below.
+
+ Rules regarding labels format:
+ - if `label_mode` is `int`, the labels are an `int32` tensor of shape
+ `(batch_size,)`.
+ - if `label_mode` is `binary`, the labels are a `float32` tensor of
+ 1s and 0s of shape `(batch_size, 1)`.
+ - if `label_mode` is `categorial`, the labels are a `float32` tensor
+ of shape `(batch_size, num_classes)`, representing a one-hot
+ encoding of the class index.
+ """
+ if labels != 'inferred':
+ if not isinstance(labels, (list, tuple)):
+ raise ValueError(
+ '`labels` argument should be a list/tuple of integer labels, of '
+ 'the same size as the number of text files in the target '
+ 'directory. If you wish to infer the labels from the subdirectory '
+ 'names in the target directory, pass `labels="inferred"`. '
+ 'If you wish to get a dataset that only contains text samples '
+ '(no labels), pass `labels=None`.')
+ if class_names:
+ raise ValueError('You can only pass `class_names` if the labels are '
+ 'inferred from the subdirectory names in the target '
+ 'directory (`labels="inferred"`).')
+ if label_mode not in {'int', 'categorical', 'binary', None}:
+ raise ValueError(
+ '`label_mode` argument must be one of "int", "categorical", "binary", '
+ 'or None. Received: %s' % (label_mode,))
+
+ if seed is None:
+ seed = np.random.randint(1e6)
+ file_paths, labels, class_names = dataset_utils.index_directory(
+ directory,
+ labels,
+ formats=('.txt',),
+ class_names=class_names,
+ shuffle=shuffle,
+ seed=seed,
+ follow_links=follow_links)
+
+ if label_mode == 'binary' and len(class_names) != 2:
+ raise ValueError(
+ 'When passing `label_mode="binary", there must exactly 2 classes. '
+ 'Found the following classes: %s' % (class_names,))
+
+ file_paths, labels = dataset_utils.get_training_or_validation_split(
+ file_paths, labels, validation_split, subset)
+
+ dataset = paths_and_labels_to_dataset(
+ file_paths=file_paths,
+ labels=labels,
+ label_mode=label_mode,
+ num_classes=len(class_names),
+ max_length=max_length)
+ if shuffle:
+ # Shuffle locally at each iteration
+ dataset = dataset.shuffle(buffer_size=batch_size * 8, seed=seed)
+ dataset = dataset.batch(batch_size)
+ # Users may need to reference `class_names`.
+ dataset.class_names = class_names
+ return dataset
+
+
+def paths_and_labels_to_dataset(file_paths,
+ labels,
+ label_mode,
+ num_classes,
+ max_length):
+ """Constructs a dataset of text strings and labels."""
+ path_ds = dataset_ops.Dataset.from_tensor_slices(file_paths)
+ string_ds = path_ds.map(
+ lambda x: path_to_string_content(x, max_length))
+ if label_mode:
+ label_ds = dataset_utils.labels_to_dataset(labels, label_mode, num_classes)
+ string_ds = dataset_ops.Dataset.zip((string_ds, label_ds))
+ return string_ds
+
+
+def path_to_string_content(path, max_length):
+ txt = io_ops.read_file(path)
+ if max_length is not None:
+ txt = string_ops.substr(txt, 0, max_length)
+ return txt
diff --git a/tensorflow/python/keras/preprocessing/text_dataset_test.py b/tensorflow/python/keras/preprocessing/text_dataset_test.py
new file mode 100644
index 0000000..c0e231e
--- /dev/null
+++ b/tensorflow/python/keras/preprocessing/text_dataset_test.py
@@ -0,0 +1,218 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for text_dataset."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import random
+import shutil
+import string
+
+from tensorflow.python.compat import v2_compat
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.preprocessing import text_dataset
+from tensorflow.python.platform import test
+
+
+class TextDatasetFromDirectoryTest(keras_parameterized.TestCase):
+
+ def _prepare_directory(self,
+ num_classes=2,
+ nested_dirs=False,
+ count=16,
+ length=20):
+ # Get a unique temp directory
+ temp_dir = os.path.join(self.get_temp_dir(), str(random.randint(0, 1e6)))
+ os.mkdir(temp_dir)
+ self.addCleanup(shutil.rmtree, temp_dir)
+
+ # Generate paths to class subdirectories
+ paths = []
+ for class_index in range(num_classes):
+ class_directory = 'class_%s' % (class_index,)
+ if nested_dirs:
+ class_paths = [
+ class_directory, os.path.join(class_directory, 'subfolder_1'),
+ os.path.join(class_directory, 'subfolder_2'), os.path.join(
+ class_directory, 'subfolder_1', 'sub-subfolder')
+ ]
+ else:
+ class_paths = [class_directory]
+ for path in class_paths:
+ os.mkdir(os.path.join(temp_dir, path))
+ paths += class_paths
+
+ for i in range(count):
+ path = paths[count % len(paths)]
+ filename = os.path.join(path, 'text_%s.txt' % (i,))
+ f = open(os.path.join(temp_dir, filename), 'w')
+ text = ''.join([random.choice(string.printable) for _ in range(length)])
+ f.write(text)
+ f.close()
+ return temp_dir
+
+ def test_text_dataset_from_directory_binary(self):
+ directory = self._prepare_directory(num_classes=2)
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode='int', max_length=10)
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ self.assertEqual(batch[0].dtype.name, 'string')
+ self.assertEqual(len(batch[0].numpy()[0]), 10) # Test max_length
+ self.assertEqual(batch[1].shape, (8,))
+ self.assertEqual(batch[1].dtype.name, 'int32')
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode='binary')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ self.assertEqual(batch[0].dtype.name, 'string')
+ self.assertEqual(batch[1].shape, (8, 1))
+ self.assertEqual(batch[1].dtype.name, 'float32')
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode='categorical')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ self.assertEqual(batch[0].dtype.name, 'string')
+ self.assertEqual(batch[1].shape, (8, 2))
+ self.assertEqual(batch[1].dtype.name, 'float32')
+
+ def test_sample_count(self):
+ directory = self._prepare_directory(num_classes=4, count=15)
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode=None)
+ sample_count = 0
+ for batch in dataset:
+ sample_count += batch.shape[0]
+ self.assertEqual(sample_count, 15)
+
+ def test_text_dataset_from_directory_multiclass(self):
+ directory = self._prepare_directory(num_classes=4, count=15)
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode=None)
+ batch = next(iter(dataset))
+ self.assertEqual(batch.shape, (8,))
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode=None)
+ sample_count = 0
+ iterator = iter(dataset)
+ for batch in dataset:
+ sample_count += next(iterator).shape[0]
+ self.assertEqual(sample_count, 15)
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode='int')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ self.assertEqual(batch[0].dtype.name, 'string')
+ self.assertEqual(batch[1].shape, (8,))
+ self.assertEqual(batch[1].dtype.name, 'int32')
+
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode='categorical')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ self.assertEqual(batch[0].dtype.name, 'string')
+ self.assertEqual(batch[1].shape, (8, 4))
+ self.assertEqual(batch[1].dtype.name, 'float32')
+
+ def test_text_dataset_from_directory_validation_split(self):
+ directory = self._prepare_directory(num_classes=2, count=10)
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=10, validation_split=0.2, subset='training')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (8,))
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=10, validation_split=0.2, subset='validation')
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertEqual(batch[0].shape, (2,))
+
+ def test_text_dataset_from_directory_manual_labels(self):
+ directory = self._prepare_directory(num_classes=2, count=2)
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, labels=[0, 1], shuffle=False)
+ batch = next(iter(dataset))
+ self.assertLen(batch, 2)
+ self.assertAllClose(batch[1], [0, 1])
+
+ def test_text_dataset_from_directory_follow_links(self):
+ directory = self._prepare_directory(num_classes=2, count=25,
+ nested_dirs=True)
+ dataset = text_dataset.text_dataset_from_directory(
+ directory, batch_size=8, label_mode=None, follow_links=True)
+ sample_count = 0
+ for batch in dataset:
+ sample_count += batch.shape[0]
+ self.assertEqual(sample_count, 25)
+
+ def test_text_dataset_from_directory_errors(self):
+ directory = self._prepare_directory(num_classes=3, count=5)
+
+ with self.assertRaisesRegex(ValueError, '`labels` argument should be'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, labels=None)
+
+ with self.assertRaisesRegex(ValueError, '`label_mode` argument must be'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, label_mode='other')
+
+ with self.assertRaisesRegex(
+ ValueError, 'only pass `class_names` if the labels are inferred'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, labels=[0, 0, 1, 1, 1],
+ class_names=['class_0', 'class_1', 'class_2'])
+
+ with self.assertRaisesRegex(
+ ValueError,
+ 'Expected the lengths of `labels` to match the number of files'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, labels=[0, 0, 1, 1])
+
+ with self.assertRaisesRegex(
+ ValueError, '`class_names` passed did not match'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, class_names=['class_0', 'class_2'])
+
+ with self.assertRaisesRegex(ValueError, 'there must exactly 2 classes'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, label_mode='binary')
+
+ with self.assertRaisesRegex(ValueError,
+ '`validation_split` must be between 0 and 1'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, validation_split=2)
+
+ with self.assertRaisesRegex(ValueError,
+ '`subset` must be either "training" or'):
+ _ = text_dataset.text_dataset_from_directory(
+ directory, validation_split=0.2, subset='other')
+
+
+if __name__ == '__main__':
+ v2_compat.enable_v2_behavior()
+ test.main()
diff --git a/tensorflow/python/keras/preprocessing/timeseries.py b/tensorflow/python/keras/preprocessing/timeseries.py
index 373594b..64e2d065 100644
--- a/tensorflow/python/keras/preprocessing/timeseries.py
+++ b/tensorflow/python/keras/preprocessing/timeseries.py
@@ -106,8 +106,8 @@
```python
input_data = data[:-10]
targets = data[10:]
- dataset = tf.keras.preprocessing.timeseries.dataset_from_array(
- input_data, targets, sequence_length=10)
+ dataset = tf.keras.preprocessing.timeseries_dataset_from_array(
+ input_data, targets, sequence_length=10)
for batch in dataset:
inputs, targets = batch
assert np.array_equal(inputs[0], data[:10]) # First sequence: steps [0-9]
diff --git a/tensorflow/python/keras/tests/BUILD b/tensorflow/python/keras/tests/BUILD
index 16d3d84..da2b24d 100644
--- a/tensorflow/python/keras/tests/BUILD
+++ b/tensorflow/python/keras/tests/BUILD
@@ -14,6 +14,22 @@
exports_files(["LICENSE"])
tf_py_test(
+ name = "get_config_test",
+ srcs = ["get_config_test.py"],
+ python_version = "PY3",
+ shard_count = 4,
+ tags = [
+ "no_pip",
+ ],
+ deps = [
+ ":get_config_samples",
+ "//tensorflow/python:client_testlib",
+ "//tensorflow/python/keras",
+ "@absl_py//absl/testing:parameterized",
+ ],
+)
+
+tf_py_test(
name = "add_loss_correctness_test",
srcs = ["add_loss_correctness_test.py"],
python_version = "PY3",
@@ -272,3 +288,10 @@
"//third_party/py/numpy",
],
)
+
+py_library(
+ name = "get_config_samples",
+ srcs = ["get_config_samples.py"],
+ srcs_version = "PY2AND3",
+ deps = [],
+)
diff --git a/tensorflow/python/keras/tests/get_config_samples.py b/tensorflow/python/keras/tests/get_config_samples.py
new file mode 100644
index 0000000..ca622e8
--- /dev/null
+++ b/tensorflow/python/keras/tests/get_config_samples.py
@@ -0,0 +1,491 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+# pylint: disable=protected-access
+"""Sample `get_config` results for testing backwards compatibility."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+# inputs = tf.keras.Input(10)
+# x = tf.keras.layers.Dense(10, activation='relu')(inputs)
+# outputs = tf.keras.layers.Dense(1)(x)
+# model = tf.keras.Model(inputs, outputs)
+FUNCTIONAL_DNN = {
+ 'input_layers': [['input_1', 0, 0]],
+ 'layers': [{
+ 'class_name': 'InputLayer',
+ 'config': {
+ 'batch_input_shape': (None, 10),
+ 'dtype': 'float32',
+ 'name': 'input_1',
+ 'ragged': False,
+ 'sparse': False
+ },
+ 'inbound_nodes': [],
+ 'name': 'input_1'
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'relu',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense',
+ 'trainable': True,
+ 'units': 10,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['input_1', 0, 0, {}]]],
+ 'name': 'dense'
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_1',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['dense', 0, 0, {}]]],
+ 'name': 'dense_1'
+ }],
+ 'name': 'model',
+ 'output_layers': [['dense_1', 0, 0]]
+}
+
+# inputs = tf.keras.Input((256, 256, 3))
+# x = tf.keras.layers.Conv2D(filters=3, kernel_size=(3, 3))(inputs)
+# x = tf.keras.layers.Flatten()(x)
+# outputs = tf.keras.layers.Dense(1)(x)
+# model = tf.keras.Model(inputs, outputs)
+FUNCTIONAL_CNN = {
+ 'input_layers': [['input_2', 0, 0]],
+ 'layers': [{
+ 'class_name': 'InputLayer',
+ 'config': {
+ 'batch_input_shape': (None, 256, 256, 3),
+ 'dtype': 'float32',
+ 'name': 'input_2',
+ 'ragged': False,
+ 'sparse': False
+ },
+ 'inbound_nodes': [],
+ 'name': 'input_2'
+ }, {
+ 'class_name': 'Conv2D',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'data_format': 'channels_last',
+ 'dilation_rate': (1, 1),
+ 'dtype': 'float32',
+ 'filters': 3,
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'kernel_size': (3, 3),
+ 'name': 'conv2d',
+ 'padding': 'valid',
+ 'strides': (1, 1),
+ 'trainable': True,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['input_2', 0, 0, {}]]],
+ 'name': 'conv2d'
+ }, {
+ 'class_name': 'Flatten',
+ 'config': {
+ 'data_format': 'channels_last',
+ 'dtype': 'float32',
+ 'name': 'flatten',
+ 'trainable': True
+ },
+ 'inbound_nodes': [[['conv2d', 0, 0, {}]]],
+ 'name': 'flatten'
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_2',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['flatten', 0, 0, {}]]],
+ 'name': 'dense_2'
+ }],
+ 'name': 'model_1',
+ 'output_layers': [['dense_2', 0, 0]]
+}
+
+# inputs = tf.keras.Input((10, 3))
+# x = tf.keras.layers.LSTM(10)(inputs)
+# outputs = tf.keras.layers.Dense(1)(x)
+# model = tf.keras.Model(inputs, outputs)
+FUNCTIONAL_LSTM = {
+ 'input_layers': [['input_5', 0, 0]],
+ 'layers': [{
+ 'class_name': 'InputLayer',
+ 'config': {
+ 'batch_input_shape': (None, 10, 3),
+ 'dtype': 'float32',
+ 'name': 'input_5',
+ 'ragged': False,
+ 'sparse': False
+ },
+ 'inbound_nodes': [],
+ 'name': 'input_5'
+ }, {
+ 'class_name': 'LSTM',
+ 'config': {
+ 'activation': 'tanh',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dropout': 0.0,
+ 'dtype': 'float32',
+ 'go_backwards': False,
+ 'implementation': 2,
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'lstm_2',
+ 'recurrent_activation': 'sigmoid',
+ 'recurrent_constraint': None,
+ 'recurrent_dropout': 0.0,
+ 'recurrent_initializer': {
+ 'class_name': 'Orthogonal',
+ 'config': {
+ 'gain': 1.0,
+ 'seed': None
+ }
+ },
+ 'recurrent_regularizer': None,
+ 'return_sequences': False,
+ 'return_state': False,
+ 'stateful': False,
+ 'time_major': False,
+ 'trainable': True,
+ 'unit_forget_bias': True,
+ 'units': 10,
+ 'unroll': False,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['input_5', 0, 0, {}]]],
+ 'name': 'lstm_2'
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_4',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ },
+ 'inbound_nodes': [[['lstm_2', 0, 0, {}]]],
+ 'name': 'dense_4'
+ }],
+ 'name': 'model_3',
+ 'output_layers': [['dense_4', 0, 0]]
+}
+
+# model = tf.keras.Sequential()
+# model.add(tf.keras.layers.Dense(10))
+# model.add(tf.keras.layers.Dense(1))
+SEQUENTIAL_DNN = {
+ 'layers': [{
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_2',
+ 'trainable': True,
+ 'units': 10,
+ 'use_bias': True
+ }
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_3',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ }
+ }],
+ 'name': 'sequential_1'
+}
+
+# model = tf.keras.Sequential()
+# model.add(tf.keras.layers.Conv2D(32, (3, 3)))
+# model.add(tf.keras.layers.Flatten())
+# model.add(tf.keras.layers.Dense(1))
+SEQUENTIAL_CNN = {
+ 'layers': [{
+ 'class_name': 'Conv2D',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'data_format': 'channels_last',
+ 'dilation_rate': (1, 1),
+ 'dtype': 'float32',
+ 'filters': 32,
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'kernel_size': (3, 3),
+ 'name': 'conv2d_1',
+ 'padding': 'valid',
+ 'strides': (1, 1),
+ 'trainable': True,
+ 'use_bias': True
+ }
+ }, {
+ 'class_name': 'Flatten',
+ 'config': {
+ 'data_format': 'channels_last',
+ 'dtype': 'float32',
+ 'name': 'flatten_1',
+ 'trainable': True
+ }
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_6',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ }
+ }],
+ 'name': 'sequential_4'
+}
+
+# model = tf.keras.Sequential()
+# model.add(tf.keras.layers.LSTM(10))
+# model.add(tf.keras.layers.Dense(1))
+SEQUENTIAL_LSTM = {
+ 'layers': [{
+ 'class_name': 'LSTM',
+ 'config': {
+ 'activation': 'tanh',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dropout': 0.0,
+ 'dtype': 'float32',
+ 'go_backwards': False,
+ 'implementation': 2,
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'lstm',
+ 'recurrent_activation': 'sigmoid',
+ 'recurrent_constraint': None,
+ 'recurrent_dropout': 0.0,
+ 'recurrent_initializer': {
+ 'class_name': 'Orthogonal',
+ 'config': {
+ 'gain': 1.0,
+ 'seed': None
+ }
+ },
+ 'recurrent_regularizer': None,
+ 'return_sequences': False,
+ 'return_state': False,
+ 'stateful': False,
+ 'time_major': False,
+ 'trainable': True,
+ 'unit_forget_bias': True,
+ 'units': 10,
+ 'unroll': False,
+ 'use_bias': True
+ }
+ }, {
+ 'class_name': 'Dense',
+ 'config': {
+ 'activation': 'linear',
+ 'activity_regularizer': None,
+ 'bias_constraint': None,
+ 'bias_initializer': {
+ 'class_name': 'Zeros',
+ 'config': {}
+ },
+ 'bias_regularizer': None,
+ 'dtype': 'float32',
+ 'kernel_constraint': None,
+ 'kernel_initializer': {
+ 'class_name': 'GlorotUniform',
+ 'config': {
+ 'seed': None
+ }
+ },
+ 'kernel_regularizer': None,
+ 'name': 'dense_4',
+ 'trainable': True,
+ 'units': 1,
+ 'use_bias': True
+ }
+ }],
+ 'name': 'sequential_2'
+}
diff --git a/tensorflow/python/keras/tests/get_config_test.py b/tensorflow/python/keras/tests/get_config_test.py
new file mode 100644
index 0000000..3274447
--- /dev/null
+++ b/tensorflow/python/keras/tests/get_config_test.py
@@ -0,0 +1,58 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#,============================================================================
+"""Tests for `get_config` backwards compatibility."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.keras.engine import sequential
+from tensorflow.python.keras.engine import training
+from tensorflow.python.keras.tests import get_config_samples
+from tensorflow.python.platform import test
+
+
+@keras_parameterized.run_all_keras_modes
+class TestGetConfigBackwardsCompatible(keras_parameterized.TestCase):
+
+ def test_functional_dnn(self):
+ model = training.Model.from_config(get_config_samples.FUNCTIONAL_DNN)
+ self.assertLen(model.layers, 3)
+
+ def test_functional_cnn(self):
+ model = training.Model.from_config(get_config_samples.FUNCTIONAL_CNN)
+ self.assertLen(model.layers, 4)
+
+ def test_functional_lstm(self):
+ model = training.Model.from_config(get_config_samples.FUNCTIONAL_LSTM)
+ self.assertLen(model.layers, 3)
+
+ def test_sequential_dnn(self):
+ model = sequential.Sequential.from_config(get_config_samples.SEQUENTIAL_DNN)
+ self.assertLen(model.layers, 2)
+
+ def test_sequential_cnn(self):
+ model = sequential.Sequential.from_config(get_config_samples.SEQUENTIAL_CNN)
+ self.assertLen(model.layers, 3)
+
+ def test_sequential_lstm(self):
+ model = sequential.Sequential.from_config(
+ get_config_samples.SEQUENTIAL_LSTM)
+ self.assertLen(model.layers, 2)
+
+
+if __name__ == '__main__':
+ test.main()
diff --git a/tensorflow/python/keras/tests/integration_test.py b/tensorflow/python/keras/tests/integration_test.py
index dbb6f75..8e4d38c 100644
--- a/tensorflow/python/keras/tests/integration_test.py
+++ b/tensorflow/python/keras/tests/integration_test.py
@@ -112,9 +112,9 @@
optimizer=keras.optimizer_v2.adam.Adam(0.005),
metrics=['acc'],
run_eagerly=testing_utils.should_run_eagerly())
- if not testing_utils.should_run_eagerly():
- self.assertEqual(len(model.get_losses_for(None)), 2)
- self.assertEqual(len(model.get_updates_for(x)), 2)
+ self.assertLen(model.losses, 2)
+ if not context.executing_eagerly():
+ self.assertLen(model.get_updates_for(x), 2)
history = model.fit(x_train, y_train, epochs=10, batch_size=10,
validation_data=(x_train, y_train),
verbose=2)
diff --git a/tensorflow/python/keras/tests/model_subclassing_test.py b/tensorflow/python/keras/tests/model_subclassing_test.py
index 5af1148..d2f4ee8 100644
--- a/tensorflow/python/keras/tests/model_subclassing_test.py
+++ b/tensorflow/python/keras/tests/model_subclassing_test.py
@@ -477,8 +477,6 @@
self.assertEqual(0, len(model.updates))
else:
self.assertEqual(2, len(model.updates))
- self.assertEqual(1, len(model.get_updates_for(None)))
- self.assertEqual(1, len(model.get_updates_for(x)))
class GraphSpecificModelSubclassingTests(test.TestCase):
@@ -536,8 +534,8 @@
x = array_ops.ones(shape=[100, 784], dtype='float32')
model(x)
- self.assertEqual(len(model.get_updates_for(x)), 2)
- self.assertEqual(len(model.get_losses_for(x)), 1)
+ self.assertLen(model.updates, 2)
+ self.assertLen(model.losses, 1)
# Case 2: placeholder-sequential nested in subclass.
class TestModel2(keras.Model):
diff --git a/tensorflow/python/keras/utils/metrics_utils.py b/tensorflow/python/keras/utils/metrics_utils.py
index 1d769a0..58fff40 100644
--- a/tensorflow/python/keras/utils/metrics_utils.py
+++ b/tensorflow/python/keras/utils/metrics_utils.py
@@ -267,8 +267,8 @@
y_true: A `Tensor` whose shape matches `y_pred`. Will be cast to `bool`.
y_pred: A floating point `Tensor` of arbitrary shape and whose values are in
the range `[0, 1]`.
- thresholds: A float value or a python list or tuple of float thresholds in
- `[0, 1]`, or NEG_INF (used when top_k is set).
+ thresholds: A float value, float tensor, python list, or tuple of float
+ thresholds in `[0, 1]`, or NEG_INF (used when top_k is set).
top_k: Optional int, indicates that the positive labels should be limited to
the top k predictions.
class_id: Optional int, limits the prediction and labels to the class
@@ -301,9 +301,9 @@
return
y_true = math_ops.cast(y_true, dtype=dtypes.float32)
y_pred = math_ops.cast(y_pred, dtype=dtypes.float32)
+ thresholds = ops.convert_to_tensor_v2(thresholds, dtype=dtypes.float32)
+ num_thresholds = thresholds.shape[0]
if multi_label:
- thresh_shape = array_ops.shape(thresholds)
- num_thresholds = thresh_shape[0]
one_thresh = math_ops.equal(
math_ops.cast(1, dtype=dtypes.int32),
array_ops.rank(thresholds),
@@ -312,7 +312,6 @@
[y_pred,
y_true], _ = ragged_assert_compatible_and_get_flat_values([y_pred, y_true],
sample_weight)
- num_thresholds = len(to_list(thresholds))
one_thresh = math_ops.cast(True, dtype=dtypes.bool)
if not any(
@@ -388,9 +387,8 @@
data_tiles = [num_thresholds, 1]
thresh_tiled = array_ops.tile(
- array_ops.reshape(
- array_ops.constant(thresholds, dtype=dtypes.float32),
- thresh_pretile_shape), array_ops.stack(thresh_tiles))
+ array_ops.reshape(thresholds, thresh_pretile_shape),
+ array_ops.stack(thresh_tiles))
# Tile the predictions for every threshold.
preds_tiled = array_ops.tile(predictions_extra_dim, data_tiles)
diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD
index 7d32085..a8d9946 100644
--- a/tensorflow/python/kernel_tests/BUILD
+++ b/tensorflow/python/kernel_tests/BUILD
@@ -1014,6 +1014,7 @@
"//tensorflow/python:framework_for_generated_wrappers",
"//tensorflow/python:sparse_ops",
"//third_party/py/numpy",
+ "@absl_py//absl/testing:parameterized",
],
)
diff --git a/tensorflow/python/kernel_tests/confusion_matrix_test.py b/tensorflow/python/kernel_tests/confusion_matrix_test.py
index c117825..8ea9b9f 100644
--- a/tensorflow/python/kernel_tests/confusion_matrix_test.py
+++ b/tensorflow/python/kernel_tests/confusion_matrix_test.py
@@ -188,7 +188,7 @@
def testLabelsTooLarge(self):
labels = np.asarray([1, 1, 0, 3, 5], dtype=np.int32)
predictions = np.asarray([2, 1, 0, 2, 2], dtype=np.int32)
- with self.assertRaisesOpError("`labels`.*x < y"):
+ with self.assertRaisesOpError("`labels`.*out of bound"):
self._testConfMatrix(
labels=labels, predictions=predictions, num_classes=3, truth=None)
@@ -203,7 +203,7 @@
def testPredictionsTooLarge(self):
labels = np.asarray([1, 1, 0, 2, 2], dtype=np.int32)
predictions = np.asarray([2, 1, 0, 3, 5], dtype=np.int32)
- with self.assertRaisesOpError("`predictions`.*x < y"):
+ with self.assertRaisesOpError("`predictions`.*out of bound"):
self._testConfMatrix(
labels=labels, predictions=predictions, num_classes=3, truth=None)
diff --git a/tensorflow/python/kernel_tests/cumulative_logsumexp_test.py b/tensorflow/python/kernel_tests/cumulative_logsumexp_test.py
index aae624f..2b0309f 100644
--- a/tensorflow/python/kernel_tests/cumulative_logsumexp_test.py
+++ b/tensorflow/python/kernel_tests/cumulative_logsumexp_test.py
@@ -55,6 +55,11 @@
reverse=reverse, exclusive=exclusive,
axis=axis)
+ def testMinusInfinity(self):
+ x = np.log([0., 0., 1., 1., 1., 1., 0., 0.])
+ self._testLogSumExpAllArgs(x, use_gpu=False)
+ self._testLogSumExpAllArgs(x, use_gpu=True)
+
def test1D(self):
x = np.arange(10) / 10.0 - 0.5
self._testLogSumExpAllArgs(x, use_gpu=False)
diff --git a/tensorflow/python/kernel_tests/cwise_ops_binary_test.py b/tensorflow/python/kernel_tests/cwise_ops_binary_test.py
index 71006f2..4c6a41b 100644
--- a/tensorflow/python/kernel_tests/cwise_ops_binary_test.py
+++ b/tensorflow/python/kernel_tests/cwise_ops_binary_test.py
@@ -953,6 +953,44 @@
"Incompatible shapes|Dimensions must be equal"):
f(x.astype(t), y.astype(t))
+ def testEqualDType(self):
+ dtypes = [
+ np.float16,
+ np.float32,
+ np.float64,
+ np.int8,
+ np.int16,
+ np.int32,
+ np.int64,
+ np.uint8,
+ np.uint16,
+ np.uint32,
+ np.uint64,
+ np.bool,
+ ]
+ x = np.asarray([0, 1, 2, 3, 4])
+ y = np.asarray([0, 1, 2, 3, 4])
+ for dtype in dtypes:
+ xt = x.astype(dtype)
+ yt = y.astype(dtype)
+ cmp_eq = math_ops.equal(xt, yt)
+ cmp_ne = math_ops.not_equal(xt, yt)
+ values = self.evaluate([cmp_eq, cmp_ne])
+ self.assertAllEqual(
+ [[True, True, True, True, True], [False, False, False, False, False]],
+ values)
+ for dtype in [np.complex64, np.complex128]:
+ xt = x.astype(dtype)
+ xt -= 1j * xt
+ yt = y.astype(dtype)
+ yt -= 1j * yt
+ cmp_eq = math_ops.equal(xt, yt)
+ cmp_ne = math_ops.not_equal(xt, yt)
+ values = self.evaluate([cmp_eq, cmp_ne])
+ self.assertAllEqual(
+ [[True, True, True, True, True], [False, False, False, False, False]],
+ values)
+
if __name__ == "__main__":
test.main()
diff --git a/tensorflow/python/kernel_tests/cwise_ops_test.py b/tensorflow/python/kernel_tests/cwise_ops_test.py
index 4f9e35a..d01d647 100644
--- a/tensorflow/python/kernel_tests/cwise_ops_test.py
+++ b/tensorflow/python/kernel_tests/cwise_ops_test.py
@@ -733,7 +733,8 @@
def testBasic(self):
x = np.random.rand(1, 3, 2) * 100.
y = np.random.rand(1, 3, 2) * 100.
- for t in [np.float16, np.float32, np.float64, np.int16, np.int32, np.int64]:
+ for t in [np.float16, np.float32, np.float64, np.uint8, np.int16, np.int32,
+ np.int64]:
self._compare(x.astype(t), y.astype(t), use_gpu=False)
self._compare(x.astype(t), y.astype(t), use_gpu=True)
diff --git a/tensorflow/python/kernel_tests/segment_reduction_ops_test.py b/tensorflow/python/kernel_tests/segment_reduction_ops_test.py
index 98a4c0c..e993ae2 100644
--- a/tensorflow/python/kernel_tests/segment_reduction_ops_test.py
+++ b/tensorflow/python/kernel_tests/segment_reduction_ops_test.py
@@ -516,6 +516,9 @@
dtypes_lib.int32
]
+ index_dtypes = [dtypes_lib.int32, dtypes_lib.int64]
+ segment_ids_dtypes = [dtypes_lib.int32, dtypes_lib.int64]
+
mean_dtypes = [dtypes_lib.float32, dtypes_lib.float64]
# Each item is np_op1, np_op2, tf_op
@@ -531,22 +534,29 @@
segment_indices.append(i)
num_indices = len(segment_indices)
for dtype in dtypes:
- with self.cached_session(use_gpu=False):
- tf_indices, np_indices, tf_x, np_x = self._sparse_input(
- shape, num_indices, dtype=dtype)
- for np_op1, np_op2, tf_op in ops_list:
- if tf_op == math_ops.sparse_segment_mean and dtype not in mean_dtypes:
- continue
- np_ans = self._sparseSegmentReduce(np_x, np_indices, segment_indices,
- np_op1, np_op2)
- s = tf_op(data=tf_x, indices=tf_indices, segment_ids=segment_indices)
- tf_ans = self.evaluate(s)
- self.assertAllClose(np_ans, tf_ans)
- # NOTE(mrry): The static shape inference that computes
- # `tf_ans.shape` can only infer that sizes from dimension 1
- # onwards, because the size of dimension 0 is data-dependent
- # and may therefore vary dynamically.
- self.assertAllEqual(np_ans.shape[1:], tf_ans.shape[1:])
+ for index_dtype in index_dtypes:
+ for segment_ids_dtype in segment_ids_dtypes:
+ with self.cached_session(use_gpu=False):
+ tf_indices, np_indices, tf_x, np_x = self._sparse_input(
+ shape, num_indices, dtype=dtype)
+ for np_op1, np_op2, tf_op in ops_list:
+ if (tf_op == math_ops.sparse_segment_mean
+ and dtype not in mean_dtypes):
+ continue
+ np_ans = self._sparseSegmentReduce(np_x, np_indices,
+ segment_indices, np_op1,
+ np_op2)
+ s = tf_op(
+ data=tf_x,
+ indices=math_ops.cast(tf_indices, index_dtype),
+ segment_ids=math_ops.cast(segment_indices, segment_ids_dtype))
+ tf_ans = self.evaluate(s)
+ self.assertAllClose(np_ans, tf_ans)
+ # NOTE(mrry): The static shape inference that computes
+ # `tf_ans.shape` can only infer that sizes from dimension 1
+ # onwards, because the size of dimension 0 is data-dependent
+ # and may therefore vary dynamically.
+ self.assertAllEqual(np_ans.shape[1:], tf_ans.shape[1:])
def testSegmentIdsHole(self):
tf_x, np_x = self._input([10, 4], dtype=dtypes_lib.float32)
diff --git a/tensorflow/python/kernel_tests/sparse_reshape_op_test.py b/tensorflow/python/kernel_tests/sparse_reshape_op_test.py
index 56aaf4c..6ec51bb 100644
--- a/tensorflow/python/kernel_tests/sparse_reshape_op_test.py
+++ b/tensorflow/python/kernel_tests/sparse_reshape_op_test.py
@@ -18,10 +18,12 @@
from __future__ import division
from __future__ import print_function
+from absl.testing import parameterized
import numpy as np
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import test_util
from tensorflow.python.ops import array_ops
@@ -41,7 +43,6 @@
ind = np.array([[0, 0], [1, 0], [1, 3], [1, 4], [3, 2],
[3, 3]]).astype(np.int64)
val = np.array([0, 10, 13, 14, 32, 33]).astype(np.float64)
-
shape = np.array([5, 6]).astype(np.int64)
return sparse_tensor.SparseTensorValue(ind, val, shape)
@@ -329,5 +330,73 @@
self.assertAllEqual(output_val.dense_shape, new_shape)
+class EmptySparseTensorReshapeTest(test.TestCase, parameterized.TestCase):
+ """Tests for reshaping 0-sized SparseTensors, compared w/ dense tensors."""
+
+ def _MakeAndReshapeTensor(self, tensor_class, original_shape, target_shape):
+ if tensor_class == "sparse":
+ ind = np.zeros([0, len(original_shape)]).astype(np.int64)
+ val = np.array([]).astype(np.float64)
+ shape = np.array(original_shape).astype(np.int64)
+ sp_input = sparse_tensor.SparseTensorValue(ind, val, shape)
+ sp_output = self.evaluate(
+ sparse_ops.sparse_reshape(sp_input, target_shape))
+ return sp_output.dense_shape
+ else:
+ dense_input = array_ops.zeros(original_shape)
+ dense_output = self.evaluate(array_ops.reshape(dense_input, target_shape))
+ return dense_output.shape
+
+ @parameterized.named_parameters([
+ ("Dense", "dense"),
+ ("Sparse", "sparse"),
+ ])
+ def testImpliedReshapeEmpty1DTensor(self, tensor_class):
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [0], [-1, 1]), [0, 1])
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [0], [-1, 1, 2]), [0, 1, 2])
+
+ @parameterized.named_parameters([
+ ("Dense", "dense"),
+ ("Sparse", "sparse"),
+ ])
+ def testImpliedReshapeEmpty2DTensor(self, tensor_class):
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [1, 0], [-1, 1]), [0, 1])
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [1, 0], [-1, 2, 3]), [0, 2, 3])
+
+ @parameterized.named_parameters([
+ ("Dense", "dense"),
+ ("Sparse", "sparse"),
+ ])
+ def testImpliedReshapeEmpty3DTensor(self, tensor_class):
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [1, 0, 0], [-1, 2, 3]),
+ [0, 2, 3])
+
+ @parameterized.named_parameters([
+ ("Dense", "dense"),
+ ("Sparse", "sparse"),
+ ])
+ def testImpliedReshapeEmpty4DTensor(self, tensor_class):
+ self.assertAllEqual(
+ self._MakeAndReshapeTensor(tensor_class, [2, 4, 0, 6], [-1, 4, 6, 2]),
+ [0, 4, 6, 2])
+
+ def testImpliedDimTogetherWithZeroDimCausesError(self):
+ # NOTE: When implied dimensions and zero dimensions coexist in the target
+ # shape, the behavior currently differs between sparse and regular tensors.
+ with self.assertRaises(errors.InvalidArgumentError):
+ self._MakeAndReshapeTensor("sparse", [0], [-1, 0])
+ with self.assertRaises(errors.InvalidArgumentError):
+ self._MakeAndReshapeTensor("sparse", [1, 0], [-1, 0])
+ with self.assertRaises(errors.InvalidArgumentError):
+ self._MakeAndReshapeTensor("sparse", [1, 2, 0], [2, -1, 0])
+ with self.assertRaises(errors.InvalidArgumentError):
+ self._MakeAndReshapeTensor("sparse", [1, 2, 3, 0], [2, 0, -1, 3])
+
+
if __name__ == "__main__":
test.main()
diff --git a/tensorflow/python/kernel_tests/unique_op_test.py b/tensorflow/python/kernel_tests/unique_op_test.py
index 3d08bf0..7d9e875 100644
--- a/tensorflow/python/kernel_tests/unique_op_test.py
+++ b/tensorflow/python/kernel_tests/unique_op_test.py
@@ -205,6 +205,28 @@
for value, count in zip(tf_y, tf_count):
self.assertEqual(count, np.sum(x == value))
+ def testFloat(self):
+ # NOTE(mrry): The behavior when a key is NaN is inherited from
+ # `std::unordered_map<float, ...>`: each NaN becomes a unique key in the
+ # map.
+ x = [0.0, 1.0, np.nan, np.nan]
+ y, idx, count = gen_array_ops.unique_with_counts_v2(
+ x, axis=np.array([], np.int32))
+ tf_y, tf_idx, tf_count = self.evaluate([y, idx, count])
+
+ self.assertEqual(len(x), len(tf_idx))
+ self.assertEqual(len(tf_y), len(np.unique(x)))
+ for i in range(len(x)):
+ if np.isnan(x[i]):
+ self.assertTrue(np.isnan(tf_y[tf_idx[i]]))
+ else:
+ self.assertEqual(x[i], tf_y[tf_idx[i]])
+ for value, count in zip(tf_y, tf_count):
+ if np.isnan(value):
+ self.assertEqual(count, 1)
+ else:
+ self.assertEqual(count, np.sum(x == value))
+
if __name__ == '__main__':
test.main()
diff --git a/tensorflow/python/kernel_tests/unstack_op_test.py b/tensorflow/python/kernel_tests/unstack_op_test.py
index 7a15888..13611b2 100644
--- a/tensorflow/python/kernel_tests/unstack_op_test.py
+++ b/tensorflow/python/kernel_tests/unstack_op_test.py
@@ -39,22 +39,49 @@
class UnstackOpTest(test.TestCase):
+ def randn(self, shape, dtype):
+ data = np.random.randn(*shape)
+ if dtype == np.bool:
+ return data < 0 # Naive casting yields True with P(1)!
+ else:
+ return data.astype(dtype)
+
+ def unstackReference(self, data, axis):
+ """Use numpy primitives to implement unstack equivalent."""
+ result = []
+ rank = len(data.shape)
+ axis = axis + rank if axis < 0 else axis
+ for k in range(data.shape[axis]):
+ axis = rank + axis if axis < 0 else axis
+ # Slice in axis dimension of k'th slice.
+ # e.g. if rank=4 k=2, axis=2 then equivalent of data[:,:,2,:]
+ # Give error with loop context
+ slice_spec = tuple(
+ slice(None) if i != axis else k for i in range(rank))
+ result.append(data.__getitem__(slice_spec))
+ return result
+
def testSimple(self):
np.random.seed(7)
for shape in (2,), (3,), (2, 3), (3, 2), (4, 3, 2):
- for dtype in [
- np.bool, np.float16, np.float32, np.float64, np.uint8, np.int32,
- np.int64
- ]:
- data = np.random.randn(*shape).astype(dtype)
- # Convert data to a single tensorflow tensor
- x = constant_op.constant(data)
- # Unstack into a list of tensors
- cs = array_ops.unstack(x, num=shape[0])
- self.assertEqual(type(cs), list)
- self.assertEqual(len(cs), shape[0])
- cs = [self.evaluate(c) for c in cs]
- self.assertAllEqual(cs, data)
+ rank = len(shape)
+ for axis in range(-rank, rank):
+ for dtype in [
+ np.bool, np.float16, np.float32, np.float64, np.uint8, np.int32,
+ np.int64
+ ]:
+ data = self.randn(shape, dtype)
+ # Convert data to a single tensorflow tensor
+ x = constant_op.constant(data)
+
+ # Unstack into a list of tensors
+ ref = self.unstackReference(data, axis)
+ cs = array_ops.unstack(x, axis=axis)
+ self.assertEqual(type(cs), list)
+ self.assertEqual(len(cs), shape[axis])
+ for k, c in enumerate(cs):
+ with self.subTest(shape=shape, k=k, axis=axis, dtype=dtype):
+ self.assertAllEqual(ref[k], self.evaluate(c))
def testSimpleGpu(self):
if not test_util.is_gpu_available():
@@ -63,19 +90,24 @@
np.random.seed(7)
with test_util.force_gpu():
for shape in (2,), (3,), (2, 3), (3, 2), (4, 3, 2):
- for dtype in [
- np.bool, np.float16, np.float32, np.float64, np.uint8, np.int32,
- np.int64
- ]:
- data = np.random.randn(*shape).astype(dtype)
- # Convert data to a single tensorflow tensor
- x = constant_op.constant(data)
- # Unstack into a list of tensors
- cs = array_ops.unstack(x, num=shape[0])
- self.assertEqual(type(cs), list)
- self.assertEqual(len(cs), shape[0])
- cs = [self.evaluate(c) for c in cs]
- self.assertAllEqual(cs, data)
+ rank = len(shape)
+ for axis in range(-rank, rank):
+ for dtype in [
+ np.bool, np.float16, np.float32, np.float64, np.uint8, np.int32,
+ np.int64
+ ]:
+ data = self.randn(shape, dtype)
+ # Convert data to a single tensorflow tensor
+ x = constant_op.constant(data)
+ # Unstack into a list of tensors
+ ref = self.unstackReference(data, axis)
+ cs = array_ops.unstack(x, axis=axis)
+ self.assertEqual(type(cs), list)
+ self.assertEqual(len(cs), shape[axis])
+ for k, c in enumerate(cs):
+ # Give error with loop context
+ with self.subTest(shape=shape, k=k, axis=axis, dtype=dtype):
+ self.assertAllEqual(ref[k], self.evaluate(c))
@test_util.run_deprecated_v1
def testGradientsAxis0(self):
diff --git a/tensorflow/python/kernel_tests/while_v2_test.py b/tensorflow/python/kernel_tests/while_v2_test.py
index 3f53f49..b1e5957 100644
--- a/tensorflow/python/kernel_tests/while_v2_test.py
+++ b/tensorflow/python/kernel_tests/while_v2_test.py
@@ -1204,6 +1204,24 @@
F()
+ @test_util.run_deprecated_v1 # Need to pass RunMetadata.
+ def testDisableLowering(self):
+ old = control_flow_util_v2._DISABLE_LOWER_USING_SWITCH_MERGE
+ control_flow_util_v2._DISABLE_LOWER_USING_SWITCH_MERGE = True
+ with self.session() as sess:
+ x = constant_op.constant(2.)
+ ret = while_loop_v2(
+ lambda v: v < 8., lambda v: v * v, [x], return_same_structure=False)
+
+ opts = config_pb2.RunOptions(trace_level=config_pb2.RunOptions.FULL_TRACE)
+ run_metadata = config_pb2.RunMetadata()
+ self.assertEqual(sess.run(ret, options=opts, run_metadata=run_metadata),
+ 16)
+ for dev_stat in run_metadata.step_stats.dev_stats:
+ for ns in dev_stat.node_stats:
+ self.assertNotIn("switch", ns.node_name)
+ control_flow_util_v2._DISABLE_LOWER_USING_SWITCH_MERGE = old
+
def ScalarShape():
return ops.convert_to_tensor([], dtype=dtypes.int32)
diff --git a/tensorflow/python/lite/toco_python_api_wrapper.cc b/tensorflow/python/lite/toco_python_api_wrapper.cc
index b8d15b1..2c6cee5 100644
--- a/tensorflow/python/lite/toco_python_api_wrapper.cc
+++ b/tensorflow/python/lite/toco_python_api_wrapper.cc
@@ -64,4 +64,14 @@
R"pbdoc(
Returns a quantized model.
)pbdoc");
+ m.def(
+ "ExperimentalMlirSparsifyModel",
+ [](py::object input_contents_txt_raw) {
+ return tensorflow::PyoOrThrow(
+ toco::MlirSparsifyModel(input_contents_txt_raw.ptr()));
+ },
+ py::arg("input_contents_txt_raw"),
+ R"pbdoc(
+ Returns a sparsified model.
+ )pbdoc");
}
diff --git a/tensorflow/python/ops/collective_ops_xla_test.py b/tensorflow/python/ops/collective_ops_xla_test.py
new file mode 100644
index 0000000..613dd25
--- /dev/null
+++ b/tensorflow/python/ops/collective_ops_xla_test.py
@@ -0,0 +1,79 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for Collective Operations with XLA."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.core.protobuf import config_pb2
+from tensorflow.core.protobuf import rewriter_config_pb2
+from tensorflow.python.eager import def_function
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import test_util
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import collective_ops
+from tensorflow.python.platform import test
+
+
+class CollectiveOpXlaTest(test.TestCase):
+
+ @test_util.run_deprecated_v1
+ def testScopedAllocatorWithXla(self):
+ group_size = 2
+ group_key = 1
+ instance_key1 = 1
+ instance_key2 = 2
+ tensor_size = 10
+
+ graph_options = config_pb2.GraphOptions(
+ optimizer_options=config_pb2.OptimizerOptions(
+ do_constant_folding=False))
+ cfg = config_pb2.ConfigProto(device_count={'CPU': group_size},
+ graph_options=graph_options)
+ rewrite_options = cfg.graph_options.rewrite_options
+ rewrite_options.scoped_allocator_optimization = (
+ rewriter_config_pb2.RewriterConfig.ON)
+ del rewrite_options.scoped_allocator_opts.enable_op[:]
+ rewrite_options.scoped_allocator_opts.enable_op.append('CollectiveReduce')
+
+ with self.session(config=cfg) as sess:
+ run_ops = []
+ for i in range(group_size):
+ with ops.device('CPU:%d' % i):
+ tensor_val = [i + 1.] * tensor_size
+ constant = constant_op.constant(tensor_val)
+
+ @def_function.function(experimental_compile=True)
+ def f(x):
+ return 2 * x + 1
+
+ input_tensor1 = array_ops.identity(f(constant))
+ input_tensor2 = array_ops.identity(f(constant))
+ reduced_tensor1 = collective_ops.all_reduce(
+ input_tensor1, group_size, group_key, instance_key1, 'Add', 'Id')
+ reduced_tensor2 = collective_ops.all_reduce(
+ input_tensor2, group_size, group_key, instance_key2, 'Add', 'Id')
+ run_ops.append(array_ops.identity(reduced_tensor1))
+ run_ops.append(array_ops.identity(reduced_tensor2))
+ results = sess.run(run_ops)
+ for result in results:
+ for result_val in result:
+ self.assertEqual(result_val, 8.)
+
+
+if __name__ == '__main__':
+ test.main()
diff --git a/tensorflow/python/ops/control_flow_ops.py b/tensorflow/python/ops/control_flow_ops.py
index bbb4f91..58948f7 100644
--- a/tensorflow/python/ops/control_flow_ops.py
+++ b/tensorflow/python/ops/control_flow_ops.py
@@ -2870,6 +2870,23 @@
When this op finishes, all ops in `inputs` have finished. This op has no
output.
+ Note: *In TensorFlow 2 with eager and/or Autograph, you should not require
+ this method, as code executes in your expected order.* Only use tf.group when
+ working with v1-style code or in a graph context such as inside `Dataset.map`.
+
+ When operating in a v1-style graph context, ops are not executed in the same
+ order as specified in the code; TensorFlow will attempt to execute ops in
+ parallel or in an order convienient to the result it is computing. `tf.group`
+ allows you to request that one or more results finish before execution
+ continues.
+
+ `tf.group` creates a single op (of type `NoOp`), and then adds appropriate
+ control dependencies. Thus, `c = tf.group(a, b)` will compute the same graph
+ as this:
+
+ with tf.control_dependencies([a, b]):
+ c = tf.no_op()
+
See also `tf.tuple` and
`tf.control_dependencies`.
diff --git a/tensorflow/python/ops/control_flow_util_v2.py b/tensorflow/python/ops/control_flow_util_v2.py
index b64aec1..98a1db7 100644
--- a/tensorflow/python/ops/control_flow_util_v2.py
+++ b/tensorflow/python/ops/control_flow_util_v2.py
@@ -33,6 +33,7 @@
_EXPERIMENTAL_OUTPUT_ALL_INTERMEDIATES_OVERRIDE = None
_KERAS_LAYER_CONTEXT_FUNCTION = None
+_DISABLE_LOWER_USING_SWITCH_MERGE = False
CondBranchFuncGraph = control_flow_v2_func_graphs.CondBranchFuncGraph
@@ -111,7 +112,8 @@
Args:
op: An `If` or `While` Operation.
"""
- if (not control_flow_util.GraphOrParentsInXlaContext(op.graph) and
+ if (not _DISABLE_LOWER_USING_SWITCH_MERGE and
+ not control_flow_util.GraphOrParentsInXlaContext(op.graph) and
context.context().function_call_options.executor_type !=
"SINGLE_THREADED_EXECUTOR"):
# pylint: disable=protected-access
diff --git a/tensorflow/python/ops/embedding_ops.py b/tensorflow/python/ops/embedding_ops.py
index 731599d..cb41802 100644
--- a/tensorflow/python/ops/embedding_ops.py
+++ b/tensorflow/python/ops/embedding_ops.py
@@ -19,6 +19,7 @@
from six.moves import xrange # pylint: disable=redefined-builtin
+from tensorflow.python.compat import compat
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
@@ -481,8 +482,6 @@
with ops.name_scope(name, "embedding_lookup_sparse",
params + [sp_ids]) as name:
segment_ids = sp_ids.indices[:, 0]
- if segment_ids.dtype != dtypes.int32:
- segment_ids = math_ops.cast(segment_ids, dtypes.int32)
ids = sp_ids.values
ids, idx = array_ops.unique(ids)
@@ -492,6 +491,9 @@
if embeddings.dtype in (dtypes.float16, dtypes.bfloat16):
embeddings = math_ops.cast(embeddings, dtypes.float32)
if not ignore_weights:
+ if segment_ids.dtype != dtypes.int32:
+ segment_ids = math_ops.cast(segment_ids, dtypes.int32)
+
weights = sp_weights.values
if weights.dtype != embeddings.dtype:
weights = math_ops.cast(weights, embeddings.dtype)
@@ -531,6 +533,12 @@
else:
assert False, "Unrecognized combiner"
else:
+ if compat.forward_compatible(2020, 5, 14):
+ if segment_ids.dtype not in (dtypes.int32, dtypes.int64):
+ segment_ids = math_ops.cast(segment_ids, dtypes.int32)
+ else:
+ if segment_ids.dtype != dtypes.int32:
+ segment_ids = math_ops.cast(segment_ids, dtypes.int32)
assert idx is not None
if combiner == "sum":
embeddings = math_ops.sparse_segment_sum(
diff --git a/tensorflow/python/ops/ragged/ragged_tensor.py b/tensorflow/python/ops/ragged/ragged_tensor.py
index 32e388e..1e1ef27 100644
--- a/tensorflow/python/ops/ragged/ragged_tensor.py
+++ b/tensorflow/python/ops/ragged/ragged_tensor.py
@@ -30,7 +30,6 @@
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.framework import sparse_tensor
-from tensorflow.python.framework import tensor_like
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import tensor_spec
from tensorflow.python.framework import tensor_util
@@ -44,6 +43,7 @@
from tensorflow.python.ops.ragged import ragged_tensor_value
from tensorflow.python.ops.ragged import ragged_util
from tensorflow.python.ops.ragged.row_partition import RowPartition
+from tensorflow.python.types import internal as internal_types
from tensorflow.python.util.tf_export import tf_export
# pylint: disable=protected-access
@@ -56,7 +56,8 @@
@tf_export("RaggedTensor")
-class RaggedTensor(composite_tensor.CompositeTensor, tensor_like.TensorLike):
+class RaggedTensor(composite_tensor.CompositeTensor,
+ internal_types.NativeObject):
"""Represents a ragged tensor.
A `RaggedTensor` is a tensor with one or more *ragged dimensions*, which are
diff --git a/tensorflow/python/ops/sparse_ops.py b/tensorflow/python/ops/sparse_ops.py
index 5096b33..844aa3c 100644
--- a/tensorflow/python/ops/sparse_ops.py
+++ b/tensorflow/python/ops/sparse_ops.py
@@ -860,14 +860,19 @@
original_reshaped_shape = list(reshaped_shape_const) # A copy
in_shape_size = np.prod(sp_input.shape.as_list())
num_implied = sum(dim is None for dim in reshaped_shape_const)
- if num_implied == 1:
+
+ # If there is a 0 dim in the user-provided shape, we cannot infer the
+ # unknown dim reliably. This is why we skip the `if` branch below when
+ # a 0 is present in `reshaped_shape_const`. Same below.
+ if num_implied == 1 and 0 not in reshaped_shape_const:
implied_idx = original_reshaped_shape.index(None)
non_implied_idx = (
original_reshaped_shape[:implied_idx] +
original_reshaped_shape[implied_idx + 1:])
reshaped_shape_const[implied_idx] = int(
in_shape_size // np.prod(non_implied_idx))
- if num_implied <= 1:
+ if num_implied == 0 or (num_implied == 1 and
+ 0 not in reshaped_shape_const):
reshaped_size = np.prod(reshaped_shape_const)
if reshaped_size != in_shape_size:
raise ValueError(
diff --git a/tensorflow/python/ops/summary_ops_v2.py b/tensorflow/python/ops/summary_ops_v2.py
index 6f4472d..91b8e61 100644
--- a/tensorflow/python/ops/summary_ops_v2.py
+++ b/tensorflow/python/ops/summary_ops_v2.py
@@ -93,6 +93,11 @@
if _summary_state.writer is None:
return constant_op.constant(False)
+ if not callable(_summary_state.is_recording):
+ static_cond = tensor_util.constant_value(_summary_state.is_recording)
+ if static_cond is not None and not static_cond:
+ return constant_op.constant(False)
+
resolve = lambda x: x() if callable(x) else x
cond_distributed = resolve(_summary_state.is_recording_distribution_strategy)
cond = resolve(_summary_state.is_recording)
@@ -110,6 +115,7 @@
return _should_record_summaries_internal(default_state=True)
+@tf_export("summary.should_record_summaries", v1=[])
def should_record_summaries():
"""Returns boolean Tensor which is true if summaries should be recorded."""
return _should_record_summaries_internal(default_state=False)
diff --git a/tensorflow/python/saved_model/BUILD b/tensorflow/python/saved_model/BUILD
index f99340e..2e5db7e 100644
--- a/tensorflow/python/saved_model/BUILD
+++ b/tensorflow/python/saved_model/BUILD
@@ -310,6 +310,7 @@
"//tensorflow/python/eager:context",
"//tensorflow/python/eager:def_function",
"//tensorflow/python/eager:function",
+ "//tensorflow/python/training/saving:checkpoint_options",
"//tensorflow/python/training/saving:functional_saver",
"//tensorflow/python/training/tracking",
"//tensorflow/python/training/tracking:base",
diff --git a/tensorflow/python/saved_model/function_deserialization.py b/tensorflow/python/saved_model/function_deserialization.py
index 9fcffc8..b36a1f2 100644
--- a/tensorflow/python/saved_model/function_deserialization.py
+++ b/tensorflow/python/saved_model/function_deserialization.py
@@ -43,6 +43,8 @@
return isinstance(t, (ops.Tensor, resource_variable_ops.BaseResourceVariable))
+# TODO(edloper): Update this to just use ConcreteFunction.__call__ with the
+# structured signature.
def _call_concrete_function(function, inputs):
"""Calls a restored Function with structured inputs.
@@ -137,8 +139,6 @@
input_signature = coder.decode_proto(function_spec_proto.input_signature)
return function_lib.FunctionSpec(fullargspec=fullargspec,
is_method=False,
- args_to_prepend=[],
- kwargs_to_include={},
input_signature=input_signature)
@@ -191,6 +191,8 @@
Args:
saved_function: `SavedFunction` proto.
concrete_functions: map from function name to `ConcreteFunction`.
+ As a side effect of this function, the `FunctionSpec` from
+ `saved_function` is added to each `ConcreteFunction` in this map.
Returns:
A `Function`.
@@ -254,6 +256,9 @@
for concrete_function_name in saved_function.concrete_functions:
concrete_function_objects.append(concrete_functions[concrete_function_name])
+ for cf in concrete_function_objects:
+ cf._set_function_spec(function_spec) # pylint: disable=protected-access
+
restored_function = RestoredFunction(
restored_function_body,
restored_function_body.__name__,
@@ -317,6 +322,11 @@
for dep in _list_function_deps(fdef, library_function_names):
functions[dep].add_to_graph(func_graph)
+
+ # We do not initialize the new ConcreteFunction's function_spec or
+ # arg_keywords here (which are used to parse the structured and flat
+ # signatures, respectively). function_spec is set up later by
+ # recreate_function(); and arg_keywords by setup_bare_concrete_function().
func = function_lib.ConcreteFunction(func_graph)
func.add_to_graph(graph)
diff --git a/tensorflow/python/saved_model/function_serialization.py b/tensorflow/python/saved_model/function_serialization.py
index a6b84d1..13280d9 100644
--- a/tensorflow/python/saved_model/function_serialization.py
+++ b/tensorflow/python/saved_model/function_serialization.py
@@ -77,6 +77,10 @@
def serialize_bare_concrete_function(concrete_function, name_map):
"""Build a SavedBareConcreteFunction."""
+ # TODO(edloper): Currently, bare concrete functions don't have access to a
+ # function_spec, so they can't be called with the structured signature.
+ # Update the serialization to include a function_spec.
+
# pylint: disable=protected-access
name = name_map.get(compat.as_text(concrete_function.name),
concrete_function.name)
@@ -151,7 +155,8 @@
func_graph_module.func_graph_from_py_func(
None, wrap_function, args=tuple(args), kwargs={},
func_graph=outer_graph)
- fn = defun.ConcreteFunction(outer_graph)
+ fn = defun.ConcreteFunction(
+ outer_graph, function_spec=concrete_function._function_spec) # pylint: disable=protected-access
fn._arg_keywords = concrete_function._arg_keywords # pylint: disable=protected-access
fn._num_positional_args = concrete_function._num_positional_args # pylint: disable=protected-access
diff --git a/tensorflow/python/saved_model/load.py b/tensorflow/python/saved_model/load.py
index 05f9c21..16f1ebd 100644
--- a/tensorflow/python/saved_model/load.py
+++ b/tensorflow/python/saved_model/load.py
@@ -173,12 +173,12 @@
# The original_outputs here had Tensors converted to TensorSpecs, so
# the restored function's structured_outputs field will not be
# exactly the same. Fortunately the repacking logic cares only about
- # the structure.
- # TODO(vbardiovsky): Should we just replicate the structures, with
- # Nones instead of real objects?
+ # the structure; and the unpacking logic cares only about structure
+ # and types.
concrete_function._func_graph.structured_outputs = original_outputs # pylint: disable=protected-access
concrete_function._func_graph.structured_input_signature = ( # pylint: disable=protected-access
coder.decode_proto(proto.canonicalized_input_signature))
+ concrete_function._initialize_function_spec() # pylint: disable=protected-access
def _setup_functions_captures(self):
"""Setup captures and variables in restored functions."""
diff --git a/tensorflow/python/saved_model/save.py b/tensorflow/python/saved_model/save.py
index a5d6353..9553fb5 100644
--- a/tensorflow/python/saved_model/save.py
+++ b/tensorflow/python/saved_model/save.py
@@ -52,6 +52,7 @@
from tensorflow.python.saved_model import signature_serialization
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.saved_model import utils_impl
+from tensorflow.python.training.saving import checkpoint_options
from tensorflow.python.training.saving import functional_saver
from tensorflow.python.training.tracking import base
from tensorflow.python.training.tracking import graph_view
@@ -941,6 +942,7 @@
May not be called from within a function body.
@end_compatibility
"""
+ options = options or save_options.SaveOptions()
# TODO(allenl): Factor out some subset of SavedModelBuilder which is 2.x
# compatible (no sessions) and share it with this export API rather than
# making a SavedModel proto and writing it directly.
@@ -954,7 +956,10 @@
# Write the checkpoint, copy assets into the assets directory, and write out
# the SavedModel proto itself.
utils_impl.get_or_create_variables_dir(export_dir)
- object_saver.save(utils_impl.get_variables_path(export_dir))
+ ckpt_options = checkpoint_options.CheckpointOptions(
+ experimental_io_device=options.experimental_io_device)
+ object_saver.save(utils_impl.get_variables_path(export_dir),
+ options=ckpt_options)
builder_impl.copy_assets_to_destination_dir(asset_info.asset_filename_map,
export_dir)
# Note that this needs to be the last file operation when saving the
@@ -976,6 +981,7 @@
def export_meta_graph(obj, filename, signatures=None, options=None):
"""Exports the MetaGraph proto to a file."""
+ options = options or save_options.SaveOptions()
export_dir = os.path.dirname(filename)
meta_graph_def, exported_graph, _, _ = _build_meta_graph(
obj, export_dir, signatures, options)
@@ -1001,7 +1007,6 @@
if not isinstance(obj, base.Trackable):
raise ValueError(
"Expected a Trackable object for export, got {}.".format(obj))
- options = options or save_options.SaveOptions()
meta_graph_def = meta_graph_def or meta_graph_pb2.MetaGraphDef()
checkpoint_graph_view = _AugmentedGraphView(obj)
diff --git a/tensorflow/python/saved_model/save_options.py b/tensorflow/python/saved_model/save_options.py
index a8528c0..748ae76 100644
--- a/tensorflow/python/saved_model/save_options.py
+++ b/tensorflow/python/saved_model/save_options.py
@@ -33,12 +33,14 @@
"""
# Define object attributes in __slots__ for improved memory and performance.
- __slots__ = ("namespace_whitelist", "save_debug_info", "function_aliases")
+ __slots__ = ("namespace_whitelist", "save_debug_info", "function_aliases",
+ "experimental_io_device")
def __init__(self,
namespace_whitelist=None,
save_debug_info=False,
- function_aliases=None):
+ function_aliases=None,
+ experimental_io_device=None):
"""Creates an object that stores options for SavedModel saving.
Args:
@@ -46,16 +48,15 @@
when saving a model. Saving an object that uses namespaced ops must
explicitly add all namespaces to the whitelist. The namespaced ops must
be registered into the framework when loading the SavedModel.
- save_debug_info: Boolean indicating whether debug information is saved.
- If True, then a debug/saved_model_debug_info.pb file will be written
- with the contents of a GraphDebugInfo binary protocol buffer containing
- stack trace information for all ops and functions that are saved.
+ save_debug_info: Boolean indicating whether debug information is saved. If
+ True, then a debug/saved_model_debug_info.pb file will be written with
+ the contents of a GraphDebugInfo binary protocol buffer containing stack
+ trace information for all ops and functions that are saved.
function_aliases: Python dict. Mapping from string to object returned by
- @tf.function.
- A single tf.function can generate many ConcreteFunctions. If a
- downstream tool wants to refer to all concrete functions generated by a
- single tf.function you can use the `function_aliases` argument to store
- a map from the alias name to all concrete function names.
+ @tf.function. A single tf.function can generate many ConcreteFunctions.
+ If a downstream tool wants to refer to all concrete functions generated
+ by a single tf.function you can use the `function_aliases` argument to
+ store a map from the alias name to all concrete function names.
E.g.
```python
class MyModel:
@@ -77,11 +78,21 @@
})
tf.saved_model.save(model, export_dir, signatures, options)
```
+ experimental_io_device: string. Applies in a distributed setting.
+ Tensorflow device to use to access the filesystem. If `None` (default)
+ then for each variable the filesystem is accessed from the CPU:0 device
+ of the host where that variable is assigned. If specified, the
+ filesystem is instead accessed from that device for all variables.
+
+ This is for example useful if you want to save to a local directory,
+ such as "/tmp" when running in a distributed setting. In that case pass
+ a device for the host where the "/tmp" directory is accessible.
"""
self.namespace_whitelist = _validate_namespace_whitelist(
namespace_whitelist)
self.save_debug_info = save_debug_info
self.function_aliases = function_aliases if function_aliases else dict()
+ self.experimental_io_device = experimental_io_device
def _validate_namespace_whitelist(namespace_whitelist):
diff --git a/tensorflow/python/saved_model/save_test.py b/tensorflow/python/saved_model/save_test.py
index cae8c4c..09e7296 100644
--- a/tensorflow/python/saved_model/save_test.py
+++ b/tensorflow/python/saved_model/save_test.py
@@ -577,6 +577,12 @@
self.assertEqual(function_cache[0].name.decode("utf-8"),
list(function_aliases.keys())[0])
+ def test_accepts_io_device(self):
+ options = save_options.SaveOptions()
+ self.assertEqual(None, options.experimental_io_device)
+ options = save_options.SaveOptions(experimental_io_device="/job:localhost")
+ self.assertEqual("/job:localhost", options.experimental_io_device)
+
class AssetTests(test.TestCase):
diff --git a/tensorflow/python/tpu/client/client.py b/tensorflow/python/tpu/client/client.py
index fdfda90..bc693cb 100644
--- a/tensorflow/python/tpu/client/client.py
+++ b/tensorflow/python/tpu/client/client.py
@@ -29,7 +29,7 @@
_GOOGLE_API_CLIENT_INSTALLED = True
try:
- from apiclient import discovery # pylint: disable=g-import-not-at-top
+ from googleapiclient import discovery # pylint: disable=g-import-not-at-top
from oauth2client import client # pylint: disable=g-import-not-at-top
except ImportError:
_GOOGLE_API_CLIENT_INSTALLED = False
diff --git a/tensorflow/python/tpu/client/pip_package/setup.py b/tensorflow/python/tpu/client/pip_package/setup.py
index e27e006..74f81f4 100644
--- a/tensorflow/python/tpu/client/pip_package/setup.py
+++ b/tensorflow/python/tpu/client/pip_package/setup.py
@@ -50,5 +50,5 @@
],
license='Apache 2.0',
keywords='tensorflow tpu',
- install_requires=['google-api-python-client', 'oauth2client']
+ install_requires=['google-api-python-client==1.8.0', 'oauth2client']
)
diff --git a/tensorflow/python/tpu/client/version.py b/tensorflow/python/tpu/client/version.py
index 3cb8087..42d04fa 100644
--- a/tensorflow/python/tpu/client/version.py
+++ b/tensorflow/python/tpu/client/version.py
@@ -18,4 +18,4 @@
from __future__ import division
from __future__ import print_function
-__version__ = "0.7"
+__version__ = "0.8"
diff --git a/tensorflow/python/training/optimizer.py b/tensorflow/python/training/optimizer.py
index f1a31d0..9732ea0 100644
--- a/tensorflow/python/training/optimizer.py
+++ b/tensorflow/python/training/optimizer.py
@@ -759,7 +759,7 @@
if hasattr(var, "_distributed_container"):
# NOTE: If this isn't patched, then there is no `handle` in
# `_resource_apply_dense`.
- distributed_container = var._distributed_container()
+ distributed_container = var._distributed_container
assert distributed_container is not None
if ops.executing_eagerly_outside_functions():
key = distributed_container._unique_id
diff --git a/tensorflow/python/training/saving/BUILD b/tensorflow/python/training/saving/BUILD
index a8f595f..670a4c3 100644
--- a/tensorflow/python/training/saving/BUILD
+++ b/tensorflow/python/training/saving/BUILD
@@ -13,10 +13,19 @@
exports_files(["LICENSE"])
py_library(
+ name = "checkpoint_options",
+ srcs = ["checkpoint_options.py"],
+ deps = [
+ "//tensorflow/python:tf_export",
+ ],
+)
+
+py_library(
name = "functional_saver",
srcs = ["functional_saver.py"],
srcs_version = "PY2AND3",
deps = [
+ ":checkpoint_options",
":saveable_hook",
":saveable_object",
":saveable_object_util",
@@ -31,6 +40,7 @@
"functional_saver_test.py",
],
deps = [
+ ":checkpoint_options",
":functional_saver",
":saveable_hook",
"//tensorflow/python/eager:test",
diff --git a/tensorflow/python/training/saving/checkpoint_options.py b/tensorflow/python/training/saving/checkpoint_options.py
new file mode 100644
index 0000000..92fd679
--- /dev/null
+++ b/tensorflow/python/training/saving/checkpoint_options.py
@@ -0,0 +1,58 @@
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Options for saving Checkpoints."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.util.tf_export import tf_export
+
+
+@tf_export("train.CheckpointOptions")
+class CheckpointOptions(object):
+ """Options for constructing a Checkpoint.
+
+ Used as the `_options` argument to the `tf.Checkpoint` constructor to adjust
+ how variables are saved.
+
+ Example: Run IO ops on "localhost" while saving a checkpoint:
+
+ ```
+ step = tf.Variable(0, name="step")
+ checkpoint = tf.Checkpoint(step=step)
+ options = tf.CheckpointOptions(experimental_io_device="/job:localhost")
+ checkpoint.save("/tmp/ckpt", options=options)
+ ```
+ """
+
+ # Define object attributes in __slots__ for improved memory and performance.
+ __slots__ = ("experimental_io_device",)
+
+ def __init__(self, experimental_io_device=None):
+ """Creates an object that stores options for a Checkpoint.
+
+ Args:
+ experimental_io_device: string. Applies in a distributed setting.
+ Tensorflow device to use to access the filesystem. If `None` (default)
+ then for each variable the filesystem is accessed from the CPU:0 device
+ of the host where that variable is assigned. If specified, the
+ filesystem is instead accessed from that device for all variables.
+
+ This is for example useful if you want to save to a local directory,
+ such as "/tmp" when running in a distributed setting. In that case pass
+ a device for the host where the "/tmp" directory is accessible.
+ """
+ self.experimental_io_device = experimental_io_device
diff --git a/tensorflow/python/training/saving/functional_saver.py b/tensorflow/python/training/saving/functional_saver.py
index d85852d..c4334e0 100644
--- a/tensorflow/python/training/saving/functional_saver.py
+++ b/tensorflow/python/training/saving/functional_saver.py
@@ -30,6 +30,7 @@
from tensorflow.python.ops import gen_io_ops
from tensorflow.python.ops import io_ops
from tensorflow.python.ops import string_ops
+from tensorflow.python.training.saving import checkpoint_options
from tensorflow.python.training.saving import saveable_hook
from tensorflow.python.training.saving import saveable_object
from tensorflow.python.training.saving import saveable_object_util
@@ -52,15 +53,17 @@
"Expected a list of SaveableObjects, got %s." % (saveable,))
self._saveable_objects = saveable_objects
- def save(self, file_prefix):
+ def save(self, file_prefix, options=None):
"""Save the saveable objects to a checkpoint with `file_prefix`.
Args:
file_prefix: A string or scalar string Tensor containing the prefix to
save under.
+ options: Optional `CheckpointOptions` object.
Returns:
An `Operation`, or None when executing eagerly.
"""
+ options = options or checkpoint_options.CheckpointOptions()
tensor_names = []
tensors = []
tensor_slices = []
@@ -69,19 +72,22 @@
tensor_names.append(spec.name)
tensors.append(spec.tensor)
tensor_slices.append(spec.slice_spec)
- with ops.device("cpu:0"):
+ save_device = options.experimental_io_device or "cpu:0"
+ with ops.device(save_device):
return io_ops.save_v2(file_prefix, tensor_names, tensor_slices, tensors)
- def restore(self, file_prefix):
+ def restore(self, file_prefix, options=None):
"""Restore the saveable objects from a checkpoint with `file_prefix`.
Args:
file_prefix: A string or scalar string Tensor containing the prefix for
files to read from.
+ options: Optional `CheckpointOptions` object.
Returns:
A dictionary mapping from SaveableObject names to restore operations.
"""
+ options = options or checkpoint_options.CheckpointOptions()
restore_specs = []
tensor_structure = []
for saveable in self._saveable_objects:
@@ -91,7 +97,8 @@
saveable_tensor_structure.append(spec.name)
restore_specs.append((spec.name, spec.slice_spec, spec.dtype))
tensor_names, tensor_slices, tensor_dtypes = zip(*restore_specs)
- with ops.device("cpu:0"):
+ restore_device = options.experimental_io_device or "cpu:0"
+ with ops.device(restore_device):
restored_tensors = io_ops.restore_v2(
file_prefix, tensor_names, tensor_slices, tensor_dtypes)
structured_restored_tensors = nest.pack_sequence_as(
@@ -190,15 +197,17 @@
with ops.control_dependencies(restore_ops.values()):
return array_ops.identity(file_prefix)
- def save(self, file_prefix):
+ def save(self, file_prefix, options=None):
"""Save the saveable objects to a checkpoint with `file_prefix`.
Args:
file_prefix: A string or scalar string Tensor containing the prefix to
save under.
+ options: Optional `CheckpointOptions` object.
Returns:
An `Operation`, or None when executing eagerly.
"""
+ options = options or checkpoint_options.CheckpointOptions()
for callback in self._before_save_callbacks:
callback()
@@ -253,32 +262,37 @@
with ops.device(device):
# _SingleDeviceSaver will use the CPU device when necessary, but initial
# read operations should be placed on the SaveableObject's device.
- sharded_saves.append(saver.save(shard_prefix))
+ sharded_saves.append(saver.save(shard_prefix, options))
with ops.control_dependencies(sharded_saves):
- # Co-locates the merge step with the last device.
- with ops.device(saveable_object_util.set_cpu0(last_device)):
+ # Merge on the io_device if specified, otherwise co-locates the merge op
+ # with the last device used.
+ merge_device = (options.experimental_io_device or
+ saveable_object_util.set_cpu0(last_device))
+ with ops.device(merge_device):
# V2 format write path consists of a metadata merge step. Once merged,
# attempts to delete the temporary directory, "<user-fed prefix>_temp".
return gen_io_ops.merge_v2_checkpoints(
sharded_prefixes, file_prefix, delete_old_dirs=True)
- def restore(self, file_prefix):
+ def restore(self, file_prefix, options=None):
"""Restore the saveable objects from a checkpoint with `file_prefix`.
Args:
file_prefix: A string or scalar string Tensor containing the prefix for
files to read from.
+ options: Optional `CheckpointOptions` object.
Returns:
A dictionary mapping from SaveableObject names to restore operations.
"""
+ options = options or checkpoint_options.CheckpointOptions()
restore_ops = {}
# Sort by device name to avoid propagating non-deterministic dictionary
# ordering in some Python versions.
for device, saver in sorted(self._single_device_savers.items()):
with ops.device(device):
- restore_ops.update(saver.restore(file_prefix))
+ restore_ops.update(saver.restore(file_prefix, options))
for callback in self._after_restore_callbacks:
callback()
diff --git a/tensorflow/python/training/saving/functional_saver_test.py b/tensorflow/python/training/saving/functional_saver_test.py
index dfa2023..7db32ff 100644
--- a/tensorflow/python/training/saving/functional_saver_test.py
+++ b/tensorflow/python/training/saving/functional_saver_test.py
@@ -20,21 +20,37 @@
import os
-from tensorflow.core.protobuf import config_pb2
+from tensorflow.python.eager import context
from tensorflow.python.eager import test
from tensorflow.python.eager import wrap_function
+from tensorflow.python.framework import config
from tensorflow.python.framework import constant_op
from tensorflow.python.framework import ops
from tensorflow.python.framework import test_util
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.platform import gfile
+from tensorflow.python.training.saving import checkpoint_options
from tensorflow.python.training.saving import functional_saver
from tensorflow.python.training.saving import saveable_hook
from tensorflow.python.training.saving import saveable_object_util
+LOCALHOST = "/job:localhost/replica:0/task:0/device:CPU:0"
+
class SaverTest(test.TestCase):
+ def setUp(self):
+ super(SaverTest, self).setUp()
+ cpus = config.list_physical_devices("CPU")
+ # Set 3 virtual CPUs
+ config.set_logical_device_configuration(cpus[0], [
+ context.LogicalDeviceConfiguration(),
+ context.LogicalDeviceConfiguration(),
+ context.LogicalDeviceConfiguration()
+ ])
+ self.local_options = checkpoint_options.CheckpointOptions(
+ experimental_io_device=LOCALHOST)
+
@test_util.run_in_graph_and_eager_modes
def test_resource_variable(self):
v1 = resource_variable_ops.ResourceVariable(2.)
@@ -55,6 +71,33 @@
self.evaluate(second_saver.restore(prefix))
self.assertEqual(2., self.evaluate(v2))
+ @test_util.run_in_graph_and_eager_modes
+ def test_resource_variable_use_localhost(self):
+ v1 = resource_variable_ops.ResourceVariable(2.)
+ self.evaluate(v1.initializer)
+ saver = functional_saver._SingleDeviceSaver(
+ saveable_object_util.saveable_objects_for_op(v1, "x"))
+ prefix = os.path.join(self.get_temp_dir(), "ckpt")
+ self.evaluate(saver.save(constant_op.constant(prefix), self.local_options))
+ self.assertEqual(2, len(gfile.Glob(prefix + "*")))
+ self.evaluate(v1.assign(1.))
+ self.evaluate(saver.restore(prefix, self.local_options))
+ self.assertEqual(2., self.evaluate(v1))
+
+ v2 = resource_variable_ops.ResourceVariable(3.)
+ self.evaluate(v2.initializer)
+ second_saver = functional_saver._SingleDeviceSaver(
+ saveable_object_util.saveable_objects_for_op(v2, "x"))
+ self.evaluate(second_saver.restore(prefix, self.local_options))
+ self.assertEqual(2., self.evaluate(v2))
+
+ # In graph mode, verify that the save and restore ops were set to run on
+ # localhost.
+ if not context.executing_eagerly():
+ for op in ops.get_default_graph().get_operations():
+ if op.type in ("SaveV2", "RestoreV2"):
+ self.assertEqual(LOCALHOST, op.device)
+
def test_to_proto(self):
v1 = resource_variable_ops.ResourceVariable(2.)
saver = functional_saver.MultiDeviceSaver(
@@ -83,12 +126,7 @@
second_saver.restore(save_path)
self.assertEqual(2., self.evaluate(v2))
- @test_util.run_v1_only(
- "Needs an API to setup multiple devices, b/124805129")
- # Set up multiple devices when graph building. Before test.main() we configure
- # the devices for eager execution.
- @test_util.run_in_graph_and_eager_modes(
- config=config_pb2.ConfigProto(device_count={"CPU": 3}))
+ @test_util.run_in_graph_and_eager_modes
def test_checkpoint_is_sharded_by_device(self):
with ops.device("cpu:0"):
v0 = resource_variable_ops.ResourceVariable(0.)
@@ -99,9 +137,9 @@
self.evaluate([v0.initializer, v1.initializer, v2.initializer])
saver = functional_saver.MultiDeviceSaver(
- list(saveable_object_util.saveable_objects_for_op(v0, "v0"))
- + list(saveable_object_util.saveable_objects_for_op(v1, "v1"))
- + list(saveable_object_util.saveable_objects_for_op(v2, "v2")))
+ list(saveable_object_util.saveable_objects_for_op(v0, "v0")) +
+ list(saveable_object_util.saveable_objects_for_op(v1, "v1")) +
+ list(saveable_object_util.saveable_objects_for_op(v2, "v2")))
prefix = os.path.join(self.get_temp_dir(), "ckpt")
self.evaluate(saver.save(constant_op.constant(prefix)))
self.assertEqual(4, len(gfile.Glob(prefix + "*")))
@@ -113,8 +151,38 @@
self.assertEqual(1., self.evaluate(v1))
self.assertEqual(2., self.evaluate(v2))
+ @test_util.run_in_graph_and_eager_modes
+ def test_checkpoint_multi_device_using_localhost(self):
+ with ops.device("cpu:0"):
+ v0 = resource_variable_ops.ResourceVariable(0.)
+ with ops.device("cpu:1"):
+ v1 = resource_variable_ops.ResourceVariable(1.)
+ with ops.device("cpu:2"):
+ v2 = resource_variable_ops.ResourceVariable(2.)
-class SaveableHookTest(test.TestCase):
+ self.evaluate([v0.initializer, v1.initializer, v2.initializer])
+ saver = functional_saver.MultiDeviceSaver(
+ list(saveable_object_util.saveable_objects_for_op(v0, "v0")) +
+ list(saveable_object_util.saveable_objects_for_op(v1, "v1")) +
+ list(saveable_object_util.saveable_objects_for_op(v2, "v2")))
+ prefix = os.path.join(self.get_temp_dir(), "ckpt")
+ self.evaluate(saver.save(constant_op.constant(prefix), self.local_options))
+ self.assertEqual(4, len(gfile.Glob(prefix + "*")))
+ self.evaluate(v0.assign(-1.))
+ self.evaluate(v1.assign(-1.))
+ self.evaluate(v2.assign(-1.))
+ self.evaluate(
+ saver.restore(constant_op.constant(prefix), self.local_options))
+ self.assertEqual(0., self.evaluate(v0))
+ self.assertEqual(1., self.evaluate(v1))
+ self.assertEqual(2., self.evaluate(v2))
+
+ # In graph mode, verify that the save and restore ops were set to run on
+ # localhost.
+ if not context.executing_eagerly():
+ for op in ops.get_default_graph().get_operations():
+ if op.type in ("SaveV2", "RestoreV2", "MergeV2Checkpoints"):
+ self.assertEqual(LOCALHOST, op.device)
def test_callbacks_run(self):
# Use dict because an int would be shadowed inside callback.
@@ -144,6 +212,5 @@
if __name__ == "__main__":
- ops.enable_eager_execution(
- config=config_pb2.ConfigProto(device_count={"CPU": 3}))
+ ops.enable_eager_execution()
test.main()
diff --git a/tensorflow/python/training/tracking/BUILD b/tensorflow/python/training/tracking/BUILD
index 9434902..f893e29 100644
--- a/tensorflow/python/training/tracking/BUILD
+++ b/tensorflow/python/training/tracking/BUILD
@@ -150,6 +150,7 @@
"//tensorflow/python/eager:context",
"//tensorflow/python/eager:def_function",
"//tensorflow/python/keras:backend",
+ "//tensorflow/python/training/saving:checkpoint_options",
"//tensorflow/python/training/saving:functional_saver",
"//tensorflow/python/training/saving:saveable_object_util",
"@six_archive//:six",
@@ -191,6 +192,7 @@
"//tensorflow/python/keras:engine",
"//tensorflow/python/keras/layers",
"//tensorflow/python/keras/optimizer_v2",
+ "//tensorflow/python/training/saving:checkpoint_options",
"@absl_py//absl/testing:parameterized",
"@six_archive//:six",
],
diff --git a/tensorflow/python/training/tracking/util.py b/tensorflow/python/training/tracking/util.py
index 24a28e9..7b603ed 100644
--- a/tensorflow/python/training/tracking/util.py
+++ b/tensorflow/python/training/tracking/util.py
@@ -44,6 +44,7 @@
from tensorflow.python.training import checkpoint_management
from tensorflow.python.training import py_checkpoint_reader
from tensorflow.python.training import saver as v1_saver_lib
+from tensorflow.python.training.saving import checkpoint_options
from tensorflow.python.training.saving import functional_saver
from tensorflow.python.training.saving import saveable_object_util
from tensorflow.python.training.tracking import base
@@ -168,7 +169,7 @@
"""Holds the status of an object-based checkpoint load."""
def __init__(self, object_graph_proto, save_path, save_path_tensor,
- restore_op_cache, graph_view):
+ restore_op_cache, graph_view, options):
"""Specify the checkpoint being loaded.
Args:
@@ -184,7 +185,9 @@
`restore()` calls.
graph_view: A graph_view_lib.ObjectGraphView object for the restored
objects.
+ options: A CheckpointOptions object.
"""
+ self.options = options
self.object_graph_proto = object_graph_proto
self.restore_uid = ops.uid()
# Maps from proto ids to lists of attributes which were in the checkpoint
@@ -291,7 +294,7 @@
("Saveable keys changed when validating. Got back %s, was "
"expecting %s") % (tensor_saveables.keys(), validated_names))
new_restore_ops = functional_saver.MultiDeviceSaver(
- validated_saveables).restore(self.save_path_tensor)
+ validated_saveables).restore(self.save_path_tensor, self.options)
if not context.executing_eagerly():
for name, restore_op in sorted(new_restore_ops.items()):
restore_ops.append(restore_op)
@@ -1113,13 +1116,15 @@
def _save_cached_when_graph_building(self,
file_prefix,
- object_graph_tensor=None):
+ object_graph_tensor,
+ options):
"""Create or retrieve save ops.
Args:
file_prefix: The prefix for saved checkpoint files.
object_graph_tensor: A `Tensor` to which the current object graph will be
fed.
+ options: `CheckpointOptions` object.
Returns:
A two-element tuple with a filename tensor and a feed_dict of tensors to
@@ -1137,14 +1142,15 @@
# var_list.
or context.executing_eagerly() or ops.inside_function()):
saver = functional_saver.MultiDeviceSaver(named_saveable_objects)
- save_op = saver.save(file_prefix)
+ save_op = saver.save(file_prefix, options=options)
with ops.device("/cpu:0"):
with ops.control_dependencies([save_op]):
self._cached_save_operation = array_ops.identity(file_prefix)
self._last_save_object_graph = graph_proto
return self._cached_save_operation, feed_additions
- def save(self, file_prefix, checkpoint_number=None, session=None):
+ def save(self, file_prefix, checkpoint_number=None, session=None,
+ options=None):
"""Save a training checkpoint.
The saved checkpoint includes variables created by this object and any
@@ -1162,10 +1168,12 @@
session: The session to evaluate variables in. Ignored when executing
eagerly. If not provided when graph building, the default session is
used.
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
The full path to the checkpoint.
"""
+ options = options or checkpoint_options.CheckpointOptions()
feed_dict = {}
use_session = (not context.executing_eagerly() and
not ops.inside_function())
@@ -1189,7 +1197,7 @@
file_io.recursive_create_dir(os.path.dirname(file_prefix))
save_path, new_feed_additions = self._save_cached_when_graph_building(
- file_prefix=file_prefix_tensor, object_graph_tensor=object_graph_tensor)
+ file_prefix_tensor, object_graph_tensor, options)
if new_feed_additions:
feed_dict.update(new_feed_additions)
if not use_session:
@@ -1202,7 +1210,7 @@
else:
return save_path
- def restore(self, save_path):
+ def restore(self, save_path, options=None):
"""Restore a training checkpoint.
Restores `root_trackable` and any objects that it tracks
@@ -1250,6 +1258,7 @@
object which may run initializers for objects in the dependency graph.
If the checkpoint was written by the name-based
`tf.compat.v1.train.Saver`, names are used to match variables.
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
A load status object, which can be used to make assertions about the
@@ -1260,6 +1269,7 @@
If `save_path` points to a name-based checkpoint, a `NameBasedSaverStatus`
object is returned which runs restore ops from a name-based saver.
"""
+ options = options or checkpoint_options.CheckpointOptions()
if save_path is None:
return InitializationOnlyStatus(self._graph_view, ops.uid())
reader = py_checkpoint_reader.NewCheckpointReader(save_path)
@@ -1304,7 +1314,8 @@
save_path=save_path,
save_path_tensor=file_prefix_tensor,
restore_op_cache=self._restore_op_cache,
- graph_view=self._graph_view)
+ graph_view=self._graph_view,
+ options=options)
base.CheckpointPosition(
checkpoint=checkpoint, proto_id=0).restore(self._graph_view.root)
load_status = CheckpointLoadStatus(
@@ -1736,6 +1747,8 @@
checkpoint_directory = "/tmp/training_checkpoints"
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
+ # Create a Checkpoint that will manage two objects with trackable state,
+ # one we name "optimizer" and the other we name "model".
checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
status = checkpoint.restore(tf.train.latest_checkpoint(checkpoint_directory))
for _ in range(num_training_steps):
@@ -1744,7 +1757,7 @@
checkpoint.save(file_prefix=checkpoint_prefix)
```
- `Checkpoint.save` and `Checkpoint.restore` write and read object-based
+ `Checkpoint.save()` and `Checkpoint.restore()` write and read object-based
checkpoints, in contrast to TensorFlow 1.x's `tf.compat.v1.train.Saver` which
writes and
reads `variable.name` based checkpoints. Object-based checkpointing saves a
@@ -1757,7 +1770,7 @@
arguments to their constructors, and each dependency is given a name that is
identical to the name of the keyword argument for which it was created.
TensorFlow classes like `Layer`s and `Optimizer`s will automatically add
- dependencies on their variables (e.g. "kernel" and "bias" for
+ dependencies on their own variables (e.g. "kernel" and "bias" for
`tf.keras.layers.Dense`). Inheriting from `tf.keras.Model` makes managing
dependencies easy in user-defined classes, since `Model` hooks into attribute
assignment. For example:
@@ -1840,7 +1853,7 @@
dtype=dtypes.int64,
trainable=False))
- def write(self, file_prefix):
+ def write(self, file_prefix, options=None):
"""Writes a training checkpoint.
The checkpoint includes variables created by this object and any
@@ -1854,14 +1867,35 @@
Checkpoints written with `write` must be read with `read`.
+ Example usage:
+
+ ```
+ step = tf.Variable(0, name="step")
+ checkpoint = tf.Checkpoint(step=step)
+ checkpoint.write("/tmp/ckpt")
+
+ # Later, read the checkpoint with read()
+ checkpoint.read("/tmp/ckpt").assert_consumed()
+
+ # You can also pass options to write() and read(). For example this
+ # runs the IO ops on the localhost:
+ options = tf.CheckpointOptions(experimental_io_device="/job:localhost")
+ checkpoint.write("/tmp/ckpt", options=options)
+
+ # Later, read the checkpoint with read()
+ checkpoint.read("/tmp/ckpt", options=options).assert_consumed()
+ ```
+
Args:
file_prefix: A prefix to use for the checkpoint filenames
(/path/to/directory/and_a_prefix).
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
The full path to the checkpoint (i.e. `file_prefix`).
"""
- output = self._saver.save(file_prefix=file_prefix)
+ options = options or checkpoint_options.CheckpointOptions()
+ output = self._saver.save(file_prefix=file_prefix, options=options)
if tensor_util.is_tensor(output):
if context.executing_eagerly():
return compat.as_str(output.numpy())
@@ -1884,7 +1918,7 @@
self._maybe_create_save_counter()
return self._save_counter
- def save(self, file_prefix):
+ def save(self, file_prefix, options=None):
"""Saves a training checkpoint and provides basic checkpoint management.
The saved checkpoint includes variables created by this object and any
@@ -1898,14 +1932,33 @@
provided by other utilities which also wrap `write` and `read`.
(`tf.train.CheckpointManager` for example).
+ ```
+ step = tf.Variable(0, name="step")
+ checkpoint = tf.Checkpoint(step=step)
+ checkpoint.save("/tmp/ckpt")
+
+ # Later, read the checkpoint with restore()
+ checkpoint.restore("/tmp/ckpt").assert_consumed()
+
+ # You can also pass options to save() and restore(). For example this
+ # runs the IO ops on the localhost:
+ options = tf.CheckpointOptions(experimental_io_device="/job:localhost")
+ checkpoint.save("/tmp/ckpt", options=options)
+
+ # Later, read the checkpoint with restore()
+ checkpoint.restore("/tmp/ckpt", options=options).assert_consumed()
+ ```
+
Args:
file_prefix: A prefix to use for the checkpoint filenames
(/path/to/directory/and_a_prefix). Names are generated based on this
prefix and `Checkpoint.save_counter`.
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
The full path to the checkpoint.
"""
+ options = options or checkpoint_options.CheckpointOptions()
graph_building = not context.executing_eagerly()
if graph_building:
if ops.inside_function():
@@ -1931,7 +1984,8 @@
checkpoint_number = session.run(self._save_assign_op)
else:
checkpoint_number = assign_op.numpy()
- file_path = self.write("%s-%d" % (file_prefix, checkpoint_number))
+ file_path = self.write("%s-%d" % (file_prefix, checkpoint_number),
+ options=options)
checkpoint_management.update_checkpoint_state_internal(
save_dir=os.path.dirname(file_prefix),
model_checkpoint_path=file_path,
@@ -1939,7 +1993,7 @@
save_relative_paths=True)
return file_path
- def read(self, save_path):
+ def read(self, save_path, options=None):
"""Read a training checkpoint written with `write`.
Reads this `Checkpoint` and any objects it depends on.
@@ -1962,18 +2016,25 @@
# Later, load the checkpoint with read()
# With restore() assert_consumed() would have failed.
checkpoint.read(path).assert_consumed()
+
+ # You can also pass options to restore(). For example this
+ # runs the IO ops on the localhost:
+ options = tf.CheckpointOptions(experimental_io_device="/job:localhost")
+ checkpoint.read(path, options=options)
```
Args:
save_path: The path to the checkpoint as returned by `write`.
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
A load status object, which can be used to make assertions about the
status of a checkpoint restoration. See `restore` for details.
"""
- return self._saver.restore(save_path=save_path)
+ options = options or checkpoint_options.CheckpointOptions()
+ return self._saver.restore(save_path=save_path, options=options)
- def restore(self, save_path):
+ def restore(self, save_path, options=None):
"""Restore a training checkpoint.
Restores this `Checkpoint` and any objects it depends on.
@@ -1995,6 +2056,10 @@
```python
checkpoint = tf.train.Checkpoint( ... )
checkpoint.restore(path).assert_consumed()
+
+ # You can additionally pass options to restore():
+ options = tf.CheckpointOptions(experimental_io_device="/job:localhost")
+ checkpoint.restore(path, options=options).assert_consumed()
```
An exception will be raised if any Python objects in the dependency graph
@@ -2011,6 +2076,7 @@
`tf.train.latest_checkpoint`. If the checkpoint was written by the
name-based `tf.compat.v1.train.Saver`, names are used to match
variables.
+ options: Optional `tf.train.CheckpointOptions` object.
Returns:
A load status object, which can be used to make assertions about the
@@ -2049,7 +2115,7 @@
checkpoint file or object when the `Checkpoint` object is deleted
(often at program shutdown).
"""
- status = self.read(save_path)
+ status = self.read(save_path, options=options)
# Create the save counter now so it gets initialized with other variables
# when graph building. Creating it earlier would lead to errors when using,
# say, train.Saver() to save the model before initializing it.
diff --git a/tensorflow/python/training/tracking/util_test.py b/tensorflow/python/training/tracking/util_test.py
index a69a34c..7a96fed 100644
--- a/tensorflow/python/training/tracking/util_test.py
+++ b/tensorflow/python/training/tracking/util_test.py
@@ -47,6 +47,7 @@
from tensorflow.python.training import checkpoint_management
from tensorflow.python.training import saver as saver_lib
from tensorflow.python.training import training_util
+from tensorflow.python.training.saving import checkpoint_options
from tensorflow.python.training.tracking import base
from tensorflow.python.training.tracking import graph_view
from tensorflow.python.training.tracking import tracking
@@ -410,6 +411,28 @@
status.assert_consumed()
@test_util.run_in_graph_and_eager_modes
+ def testPassingCheckpointOptions(self):
+ localhost = "/job:localhost/device:CPU:0"
+ options = checkpoint_options.CheckpointOptions(
+ experimental_io_device=localhost)
+ prefix = os.path.join(self.get_temp_dir(), "ckpt")
+ v = variable_scope.get_variable(name="v", initializer=0.)
+ self.evaluate(v.initializer)
+ ckpt = trackable_utils.Checkpoint(v=v)
+ self.evaluate(trackable_utils.gather_initializers(ckpt))
+ save_path = ckpt.save(file_prefix=prefix, options=options)
+ status = ckpt.restore(save_path=save_path, options=options)
+ del ckpt
+ status.assert_consumed()
+
+ # In graph mode, verify that the save and restore ops were set to run on
+ # localhost.
+ if not context.executing_eagerly():
+ for op in ops.get_default_graph().get_operations():
+ if op.type in ("SaveV2", "RestoreV2"):
+ self.assertEqual(localhost, op.device)
+
+ @test_util.run_in_graph_and_eager_modes
def testSaveRestore(self):
model = MyModel()
optimizer = adam.Adam(0.001)
diff --git a/tensorflow/python/types/BUILD b/tensorflow/python/types/BUILD
index 040555b..3e6be59 100644
--- a/tensorflow/python/types/BUILD
+++ b/tensorflow/python/types/BUILD
@@ -23,9 +23,9 @@
srcs = [
"__init__.py",
"core.py",
+ "internal.py",
],
srcs_version = "PY2AND3",
visibility = ["//tensorflow:__subpackages__"],
- deps = [
- ],
+ deps = [],
)
diff --git a/tensorflow/python/framework/tensor_like.py b/tensorflow/python/types/internal.py
similarity index 65%
rename from tensorflow/python/framework/tensor_like.py
rename to tensorflow/python/types/internal.py
index e8fe2f2..892fd96 100644
--- a/tensorflow/python/framework/tensor_like.py
+++ b/tensorflow/python/types/internal.py
@@ -1,4 +1,4 @@
-# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
@@ -12,17 +12,19 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
-"""Base class for tensor-like objects."""
+"""Types internal to TensorFlow.
+
+These types should not be exported. External code should not rely on these.
+"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
-class TensorLike(object):
- """TF-specific types TF operations are expected to natively support.
+# TODO(mdan): Is this strictly needed? Only ops.py really uses it.
+class NativeObject(object):
+ """Types natively supported by various TF operations.
- Do not check this with isinstance directly; prefer instead using
- `tf.is_tensor` to check whether converting to a tensor is necessary.
+ The most notable example of NativeObject is Tensor.
"""
- pass
diff --git a/tensorflow/python/util/nest.py b/tensorflow/python/util/nest.py
index 5170301..695cc4c 100644
--- a/tensorflow/python/util/nest.py
+++ b/tensorflow/python/util/nest.py
@@ -231,7 +231,7 @@
yield field, getattr(iterable, field)
elif _is_composite_tensor(iterable):
type_spec = iterable._type_spec # pylint: disable=protected-access
- yield type(iterable).__name__, type_spec._to_components(iterable) # pylint: disable=protected-access
+ yield type_spec.value_type.__name__, type_spec._to_components(iterable) # pylint: disable=protected-access
elif _is_type_spec(iterable):
# Note: to allow CompositeTensors and their TypeSpecs to have matching
# structures, we need to use the same key string here.
diff --git a/tensorflow/stream_executor/device_options.h b/tensorflow/stream_executor/device_options.h
index b195bc8..00eb8c8 100644
--- a/tensorflow/stream_executor/device_options.h
+++ b/tensorflow/stream_executor/device_options.h
@@ -39,19 +39,19 @@
// this flag prevents it from ever being deallocated. Potentially saves
// thrashing the thread stack memory allocation, but at the potential cost of
// some memory space.
- static const unsigned kDoNotReclaimStackAllocation = 0x1;
+ static constexpr unsigned kDoNotReclaimStackAllocation = 0x1;
// The following options refer to synchronization options when
// using SynchronizeStream or SynchronizeContext.
// Synchronize with spinlocks.
- static const unsigned kScheduleSpin = 0x02;
+ static constexpr unsigned kScheduleSpin = 0x02;
// Synchronize with spinlocks that also call CPU yield instructions.
- static const unsigned kScheduleYield = 0x04;
+ static constexpr unsigned kScheduleYield = 0x04;
// Synchronize with a "synchronization primitive" (e.g. mutex).
- static const unsigned kScheduleBlockingSync = 0x08;
+ static constexpr unsigned kScheduleBlockingSync = 0x08;
- static const unsigned kMask = 0xf; // Mask of all available flags.
+ static constexpr unsigned kMask = 0xf; // Mask of all available flags.
// Constructs an or-d together set of device options.
explicit DeviceOptions(unsigned flags) : flags_(flags) {
diff --git a/tensorflow/stream_executor/gpu/redzone_allocator.h b/tensorflow/stream_executor/gpu/redzone_allocator.h
index 77755cc..e5e42df 100644
--- a/tensorflow/stream_executor/gpu/redzone_allocator.h
+++ b/tensorflow/stream_executor/gpu/redzone_allocator.h
@@ -39,10 +39,10 @@
// memory for cudnn convolutions.
class RedzoneAllocator : public ScratchAllocator {
public:
- static const int64 kDefaultMemoryLimit = 1LL << 32; // 4GB
- static const int64 kDefaultRedzoneSize =
+ static constexpr int64 kDefaultMemoryLimit = 1LL << 32; // 4GB
+ static constexpr int64 kDefaultRedzoneSize =
1LL << 23; // 8MiB per side, 16MiB total.
- static const uint8 kDefaultRedzonePattern = -1;
+ static constexpr uint8 kDefaultRedzonePattern = -1;
RedzoneAllocator(Stream* stream, DeviceMemoryAllocator* memory_allocator,
GpuAsmOpts gpu_compilation_opts_,
int64 memory_limit = kDefaultMemoryLimit,
diff --git a/tensorflow/stream_executor/rng.h b/tensorflow/stream_executor/rng.h
index acbf8fc..3dee347 100644
--- a/tensorflow/stream_executor/rng.h
+++ b/tensorflow/stream_executor/rng.h
@@ -40,8 +40,8 @@
// thread-hostility.
class RngSupport {
public:
- static const int kMinSeedBytes = 16;
- static const int kMaxSeedBytes = INT_MAX;
+ static constexpr int kMinSeedBytes = 16;
+ static constexpr int kMaxSeedBytes = INT_MAX;
// Releases any random-number-generation resources associated with this
// support object in the underlying platform implementation.
diff --git a/tensorflow/stream_executor/rocm/rocm_blas.cc b/tensorflow/stream_executor/rocm/rocm_blas.cc
index 328cce5..5ddad13 100644
--- a/tensorflow/stream_executor/rocm/rocm_blas.cc
+++ b/tensorflow/stream_executor/rocm/rocm_blas.cc
@@ -1519,7 +1519,7 @@
float beta, DeviceMemory<Eigen::half> *c, int ldc) {
blas_log("DoBlasGemm");
VLOG(1) << absl::StreamFormat(
- "doing rocBLAS SGEMM: at=%d bt=%d m=%u n=%u "
+ "doing rocBLAS SGEMM<half>: at=%d bt=%d m=%u n=%u "
"k=%llu alpha=%f a=%p lda=%d b=%p ldb=%d beta=%f "
"c=%p ldc=%d",
static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha,
@@ -1565,7 +1565,7 @@
DeviceMemory<float> *c, int ldc) {
blas_log("DoBlasGemm");
VLOG(1) << absl::StreamFormat(
- "doing rocBLAS SGEMM: at=%d bt=%d m=%u n=%u "
+ "doing rocBLAS SGEMM<float>: at=%d bt=%d m=%u n=%u "
"k=%llu alpha=%f a=%p lda=%d b=%p ldb=%d beta=%f "
"c=%p ldc=%d",
static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha,
@@ -2473,7 +2473,12 @@
int64 stride_a, const DeviceMemory<float> &b, int ldb, int64 stride_b,
float beta, DeviceMemory<float> *c, int ldc, int64 stride_c,
int batch_count) {
- blas_log("DoBlasGemmStridedBatched");
+ VLOG(1) << absl::StreamFormat(
+ "doing rocBLAS SGEMM Strided Batched<float>: at=%d bt=%d m=%u n=%u "
+ "k=%llu alpha=%f a=%p lda=%d b=%p ldb=%d beta=%f "
+ "c=%p ldc=%d",
+ static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha,
+ a.opaque(), lda, b.opaque(), ldb, beta, c->opaque(), ldc);
return DoBlasInternal(wrap::rocblas_sgemm_strided_batched, stream,
false, /* pointer_mode_host */
ROCMBlasTranspose(transa), ROCMBlasTranspose(transb), m,
@@ -2487,7 +2492,12 @@
int64 stride_a, const DeviceMemory<double> &b, int ldb, int64 stride_b,
double beta, DeviceMemory<double> *c, int ldc, int64 stride_c,
int batch_count) {
- blas_log("DoBlasGemmStridedBatched");
+ VLOG(1) << absl::StreamFormat(
+ "doing rocBLAS SGEMM Strided Batched<double>: at=%d bt=%d m=%u n=%u "
+ "k=%llu alpha=%f a=%p lda=%d b=%p ldb=%d beta=%f "
+ "c=%p ldc=%d",
+ static_cast<int>(transa), static_cast<int>(transb), m, n, k, alpha,
+ a.opaque(), lda, b.opaque(), ldb, beta, c->opaque(), ldc);
return DoBlasInternal(wrap::rocblas_dgemm_strided_batched, stream,
false, /* pointer_mode_host */
ROCMBlasTranspose(transa), ROCMBlasTranspose(transb), m,
@@ -2502,10 +2512,13 @@
const DeviceMemory<std::complex<float>> &b, int ldb, int64 stride_b,
std::complex<float> beta, DeviceMemory<std::complex<float>> *c, int ldc,
int64 stride_c, int batch_count) {
- LOG(ERROR) << "rocBLAS does not currently support the "
- "DoBlasGemmStridedBatched operation "
- << "for the \"complex<float>\" datatype";
- return false;
+ return DoBlasInternal(wrap::rocblas_cgemm_strided_batched, stream,
+ false, /* pointer_mode_host */
+ ROCMBlasTranspose(transa), ROCMBlasTranspose(transb), m,
+ n, k, complex_cast(alpha), complex_cast(a), lda,
+ stride_a, complex_cast(b), ldb, stride_b,
+ complex_cast(beta), complex_cast(c), ldc, stride_c,
+ batch_count);
}
bool ROCMBlas::DoBlasGemmStridedBatched(
Stream *stream, blas::Transpose transa, blas::Transpose transb, uint64 m,
@@ -2514,10 +2527,13 @@
const DeviceMemory<std::complex<double>> &b, int ldb, int64 stride_b,
std::complex<double> beta, DeviceMemory<std::complex<double>> *c, int ldc,
int64 stride_c, int batch_count) {
- LOG(ERROR) << "rocBLAS does not currently support the "
- "DoBlasGemmStridedBatched operation "
- << "for the \"complex<double>\" datatype";
- return false;
+ return DoBlasInternal(wrap::rocblas_zgemm_strided_batched, stream,
+ false, /* pointer_mode_host */
+ ROCMBlasTranspose(transa), ROCMBlasTranspose(transb), m,
+ n, k, complex_cast(alpha), complex_cast(a), lda,
+ stride_a, complex_cast(b), ldb, stride_b,
+ complex_cast(beta), complex_cast(c), ldc, stride_c,
+ batch_count);
}
port::Status ROCMBlas::GetVersion(string *version) {
diff --git a/tensorflow/stream_executor/stream_executor_pimpl.h b/tensorflow/stream_executor/stream_executor_pimpl.h
index eeb0710..f7f69f7 100644
--- a/tensorflow/stream_executor/stream_executor_pimpl.h
+++ b/tensorflow/stream_executor/stream_executor_pimpl.h
@@ -723,7 +723,7 @@
// Only one worker thread is needed; little work will be done by the
// executor.
- static const int kNumBackgroundThreads = 1;
+ static constexpr int kNumBackgroundThreads = 1;
// Indicates if StreamExecutor operation tracing should be performed.
bool tracing_enabled_;
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index d00ac4b..f5a32d4 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -2223,14 +2223,11 @@
deps = deps + tf_additional_xla_deps_py()
if grpc_enabled:
deps = deps + tf_additional_grpc_deps_py()
- if tfrt_enabled:
- deps = deps + ["//tensorflow/python:is_tfrt_test_true"]
# NOTE(ebrevdo): This is a workaround for depset() not being able to tell
# the difference between 'dep' and 'clean_dep(dep)'.
for to_add in [
"//tensorflow/python:extra_py_tests_deps",
- "//tensorflow/python:gradient_checker",
]:
if to_add not in deps and clean_dep(to_add) not in deps:
deps.append(clean_dep(to_add))
@@ -2253,6 +2250,23 @@
deps = depset(deps + xla_test_true_list),
**kwargs
)
+ if tfrt_enabled:
+ py_test(
+ name = name + "_tfrt",
+ size = size,
+ srcs = srcs,
+ args = args,
+ data = data,
+ flaky = flaky,
+ kernels = kernels,
+ main = main,
+ shard_count = shard_count,
+ tags = tags,
+ visibility = [clean_dep("//tensorflow:internal")] +
+ additional_visibility,
+ deps = depset(deps + xla_test_true_list + ["//tensorflow/python:is_tfrt_test_true"]),
+ **kwargs
+ )
register_extension_info(
extension_name = "tf_py_test",
@@ -2504,7 +2518,12 @@
return ("cat $(SRCS) > $(@) && " + "echo >> $(@) && " + "echo " + to_append +
" >> $(@)")
+def tf_local_platform_constraint():
+ return ["@local_execution_config_platform//:platform_constraint"]
+
def tf_version_info_genrule(name, out):
+ # TODO(gunan): Investigate making this action hermetic so we do not need
+ # to run it locally.
native.genrule(
name = name,
srcs = [
@@ -2517,9 +2536,10 @@
"$(location //tensorflow/tools/git:gen_git_source) --generate $(SRCS) \"$@\" --git_tag_override=$${GIT_TAG_OVERRIDE:-}",
local = 1,
exec_tools = [clean_dep("//tensorflow/tools/git:gen_git_source")],
+ exec_compatible_with = tf_local_platform_constraint(),
)
-def tf_py_build_info_genrule(name, out, **kwargs):
+def tf_py_build_info_genrule(name, out, exec_compatible_with, **kwargs):
native.genrule(
name = name,
outs = [out],
diff --git a/tensorflow/tf_exported_symbols.lds b/tensorflow/tf_exported_symbols.lds
index 734b090..911363b 100644
--- a/tensorflow/tf_exported_symbols.lds
+++ b/tensorflow/tf_exported_symbols.lds
@@ -1,4 +1,5 @@
*tensorflow*
+*absl*kSeed*;
*toco*
*perftools*gputools*
*tf_*
diff --git a/tensorflow/tf_version_script.lds b/tensorflow/tf_version_script.lds
index a32da32..5796385 100644
--- a/tensorflow/tf_version_script.lds
+++ b/tensorflow/tf_version_script.lds
@@ -1,6 +1,7 @@
tensorflow {
global:
*tensorflow*;
+ *absl*kSeed*;
*toco*;
*perftools*gputools*;
*tf_*;
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.-indexed-slices.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.-indexed-slices.pbtxt
index e9e8054..4da8280 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.-indexed-slices.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.-indexed-slices.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.IndexedSlices"
tf_class {
is_instance: "<class \'tensorflow.python.framework.indexed_slices.IndexedSlices\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.-ragged-tensor.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.-ragged-tensor.pbtxt
index 44a6687..7a41cfe 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.-ragged-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.-ragged-tensor.pbtxt
@@ -2,7 +2,7 @@
tf_class {
is_instance: "<class \'tensorflow.python.ops.ragged.ragged_tensor.RaggedTensor\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.-sparse-tensor.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.-sparse-tensor.pbtxt
index d71812c..aa89308 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.-sparse-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.-sparse-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.SparseTensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.sparse_tensor.SparseTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.-tensor.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.-tensor.pbtxt
index 33742e3..4a30fae 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.Tensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.ops.Tensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<type \'object\'>"
member {
name: "OVERLOADABLE_OPERATORS"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
index c8cc33f..44da86f 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
index 2c6b4bc..29a82a1 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
index 782a7d5..6dc385a 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
index 769fbd0..52cec45 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
index f539ee3..7096d94 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
index 57b20ce..7e4089f 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
index d4b19e2..0d982f2 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.keras.metrics.-a-u-c.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.keras.metrics.-a-u-c.pbtxt
index 64ccf7c..5e263f7 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.keras.metrics.-a-u-c.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.keras.metrics.-a-u-c.pbtxt
@@ -89,6 +89,10 @@
mtype: "<type \'property\'>"
}
member {
+ name: "thresholds"
+ mtype: "<type \'property\'>"
+ }
+ member {
name: "trainable"
mtype: "<type \'property\'>"
}
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
index 2fe82b3..1f050e9 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
@@ -938,7 +938,7 @@
}
member_method {
name: "DataServiceDataset"
- argspec: "args=[\'address\', \'protocol\', \'max_outstanding_requests\', \'output_types\', \'output_shapes\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'address\', \'protocol\', \'max_outstanding_requests\', \'output_types\', \'output_shapes\', \'task_refresh_interval_hint_ms\', \'name\'], varargs=None, keywords=None, defaults=[\'-1\', \'None\'], "
}
member_method {
name: "DatasetCardinality"
@@ -1177,6 +1177,14 @@
argspec: "args=[\'images\', \'boxes\', \'colors\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
+ name: "DummyMemoryCache"
+ argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+ member_method {
+ name: "DummySeedGenerator"
+ argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+ member_method {
name: "DynamicPartition"
argspec: "args=[\'data\', \'partitions\', \'num_partitions\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
@@ -3901,6 +3909,10 @@
argspec: "args=[\'input_dataset\', \'buffer_size\', \'seed_generator\', \'output_types\', \'output_shapes\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
+ name: "ShuffleDatasetV3"
+ argspec: "args=[\'input_dataset\', \'buffer_size\', \'seed\', \'seed2\', \'seed_generator\', \'output_types\', \'output_shapes\', \'reshuffle_each_iteration\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], "
+ }
+ member_method {
name: "ShutdownDistributedTPU"
argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.saved_model.-save-options.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.saved_model.-save-options.pbtxt
index 9846232..6a8163c 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.saved_model.-save-options.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.saved_model.-save-options.pbtxt
@@ -3,6 +3,10 @@
is_instance: "<class \'tensorflow.python.saved_model.save_options.SaveOptions\'>"
is_instance: "<type \'object\'>"
member {
+ name: "experimental_io_device"
+ mtype: "<type \'member_descriptor\'>"
+ }
+ member {
name: "function_aliases"
mtype: "<type \'member_descriptor\'>"
}
@@ -16,6 +20,6 @@
}
member_method {
name: "__init__"
- argspec: "args=[\'self\', \'namespace_whitelist\', \'save_debug_info\', \'function_aliases\'], varargs=None, keywords=None, defaults=[\'None\', \'False\', \'None\'], "
+ argspec: "args=[\'self\', \'namespace_whitelist\', \'save_debug_info\', \'function_aliases\', \'experimental_io_device\'], varargs=None, keywords=None, defaults=[\'None\', \'False\', \'None\', \'None\'], "
}
}
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.sparse.-sparse-tensor.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.sparse.-sparse-tensor.pbtxt
index a3ea216..a49cd1c 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.sparse.-sparse-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.sparse.-sparse-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.sparse.SparseTensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.sparse_tensor.SparseTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.train.-checkpoint-options.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.train.-checkpoint-options.pbtxt
new file mode 100644
index 0000000..b86e4cb
--- /dev/null
+++ b/tensorflow/tools/api/golden/v1/tensorflow.train.-checkpoint-options.pbtxt
@@ -0,0 +1,13 @@
+path: "tensorflow.train.CheckpointOptions"
+tf_class {
+ is_instance: "<class \'tensorflow.python.training.saving.checkpoint_options.CheckpointOptions\'>"
+ is_instance: "<type \'object\'>"
+ member {
+ name: "experimental_io_device"
+ mtype: "<type \'member_descriptor\'>"
+ }
+ member_method {
+ name: "__init__"
+ argspec: "args=[\'self\', \'experimental_io_device\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+}
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.train.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.train.pbtxt
index c71bc4a..f89c502 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.train.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.train.pbtxt
@@ -29,6 +29,10 @@
mtype: "<type \'type\'>"
}
member {
+ name: "CheckpointOptions"
+ mtype: "<type \'type\'>"
+ }
+ member {
name: "CheckpointSaverHook"
mtype: "<type \'type\'>"
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.-indexed-slices.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.-indexed-slices.pbtxt
index e9e8054..4da8280 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.-indexed-slices.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.-indexed-slices.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.IndexedSlices"
tf_class {
is_instance: "<class \'tensorflow.python.framework.indexed_slices.IndexedSlices\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.-ragged-tensor.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.-ragged-tensor.pbtxt
index 44a6687..7a41cfe 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.-ragged-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.-ragged-tensor.pbtxt
@@ -2,7 +2,7 @@
tf_class {
is_instance: "<class \'tensorflow.python.ops.ragged.ragged_tensor.RaggedTensor\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<type \'object\'>"
member {
name: "dtype"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.-sparse-tensor.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.-sparse-tensor.pbtxt
index d71812c..aa89308 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.-sparse-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.-sparse-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.SparseTensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.sparse_tensor.SparseTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.-tensor.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.-tensor.pbtxt
index 33742e3..4a30fae 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.Tensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.ops.Tensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<type \'object\'>"
member {
name: "OVERLOADABLE_OPERATORS"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
index c8cc33f..44da86f 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-contrast.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
index 2c6b4bc..29a82a1 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-crop.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
index 782a7d5..6dc385a 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-flip.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
index 769fbd0..52cec45 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-height.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
index f539ee3..7096d94 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-rotation.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
index 57b20ce..7e4089f 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-translation.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
index d4b19e2..0d982f2 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.layers.experimental.preprocessing.-random-width.pbtxt
@@ -145,7 +145,7 @@
}
member_method {
name: "call"
- argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'self\', \'inputs\', \'training\'], varargs=None, keywords=None, defaults=[\'True\'], "
}
member_method {
name: "compute_mask"
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.metrics.-a-u-c.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.metrics.-a-u-c.pbtxt
index 64ccf7c..5e263f7 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.metrics.-a-u-c.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.metrics.-a-u-c.pbtxt
@@ -89,6 +89,10 @@
mtype: "<type \'property\'>"
}
member {
+ name: "thresholds"
+ mtype: "<type \'property\'>"
+ }
+ member {
name: "trainable"
mtype: "<type \'property\'>"
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.keras.preprocessing.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.keras.preprocessing.pbtxt
index aa78190..3189c50 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.keras.preprocessing.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.keras.preprocessing.pbtxt
@@ -17,6 +17,10 @@
argspec: "args=[\'directory\', \'labels\', \'label_mode\', \'class_names\', \'color_mode\', \'batch_size\', \'image_size\', \'shuffle\', \'seed\', \'validation_split\', \'subset\', \'interpolation\', \'follow_links\'], varargs=None, keywords=None, defaults=[\'inferred\', \'int\', \'None\', \'rgb\', \'32\', \'(256, 256)\', \'True\', \'None\', \'None\', \'None\', \'bilinear\', \'False\'], "
}
member_method {
+ name: "text_dataset_from_directory"
+ argspec: "args=[\'directory\', \'labels\', \'label_mode\', \'class_names\', \'batch_size\', \'max_length\', \'shuffle\', \'seed\', \'validation_split\', \'subset\', \'follow_links\'], varargs=None, keywords=None, defaults=[\'inferred\', \'int\', \'None\', \'32\', \'None\', \'True\', \'None\', \'None\', \'None\', \'False\'], "
+ }
+ member_method {
name: "timeseries_dataset_from_array"
argspec: "args=[\'data\', \'targets\', \'sequence_length\', \'sequence_stride\', \'sampling_rate\', \'batch_size\', \'shuffle\', \'seed\', \'start_index\', \'end_index\'], varargs=None, keywords=None, defaults=[\'1\', \'1\', \'128\', \'False\', \'None\', \'None\', \'None\'], "
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.metrics.-a-u-c.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.metrics.-a-u-c.pbtxt
index 66ab6b5..45e32e8 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.metrics.-a-u-c.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.metrics.-a-u-c.pbtxt
@@ -89,6 +89,10 @@
mtype: "<type \'property\'>"
}
member {
+ name: "thresholds"
+ mtype: "<type \'property\'>"
+ }
+ member {
name: "trainable"
mtype: "<type \'property\'>"
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
index 2fe82b3..1f050e9 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
@@ -938,7 +938,7 @@
}
member_method {
name: "DataServiceDataset"
- argspec: "args=[\'address\', \'protocol\', \'max_outstanding_requests\', \'output_types\', \'output_shapes\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ argspec: "args=[\'address\', \'protocol\', \'max_outstanding_requests\', \'output_types\', \'output_shapes\', \'task_refresh_interval_hint_ms\', \'name\'], varargs=None, keywords=None, defaults=[\'-1\', \'None\'], "
}
member_method {
name: "DatasetCardinality"
@@ -1177,6 +1177,14 @@
argspec: "args=[\'images\', \'boxes\', \'colors\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
+ name: "DummyMemoryCache"
+ argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+ member_method {
+ name: "DummySeedGenerator"
+ argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+ member_method {
name: "DynamicPartition"
argspec: "args=[\'data\', \'partitions\', \'num_partitions\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
@@ -3901,6 +3909,10 @@
argspec: "args=[\'input_dataset\', \'buffer_size\', \'seed_generator\', \'output_types\', \'output_shapes\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
+ name: "ShuffleDatasetV3"
+ argspec: "args=[\'input_dataset\', \'buffer_size\', \'seed\', \'seed2\', \'seed_generator\', \'output_types\', \'output_shapes\', \'reshuffle_each_iteration\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], "
+ }
+ member_method {
name: "ShutdownDistributedTPU"
argspec: "args=[\'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.saved_model.-save-options.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.saved_model.-save-options.pbtxt
index 9846232..6a8163c 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.saved_model.-save-options.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.saved_model.-save-options.pbtxt
@@ -3,6 +3,10 @@
is_instance: "<class \'tensorflow.python.saved_model.save_options.SaveOptions\'>"
is_instance: "<type \'object\'>"
member {
+ name: "experimental_io_device"
+ mtype: "<type \'member_descriptor\'>"
+ }
+ member {
name: "function_aliases"
mtype: "<type \'member_descriptor\'>"
}
@@ -16,6 +20,6 @@
}
member_method {
name: "__init__"
- argspec: "args=[\'self\', \'namespace_whitelist\', \'save_debug_info\', \'function_aliases\'], varargs=None, keywords=None, defaults=[\'None\', \'False\', \'None\'], "
+ argspec: "args=[\'self\', \'namespace_whitelist\', \'save_debug_info\', \'function_aliases\', \'experimental_io_device\'], varargs=None, keywords=None, defaults=[\'None\', \'False\', \'None\', \'None\'], "
}
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.sparse.-sparse-tensor.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.sparse.-sparse-tensor.pbtxt
index a3ea216..a49cd1c 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.sparse.-sparse-tensor.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.sparse.-sparse-tensor.pbtxt
@@ -1,7 +1,7 @@
path: "tensorflow.sparse.SparseTensor"
tf_class {
is_instance: "<class \'tensorflow.python.framework.sparse_tensor.SparseTensor\'>"
- is_instance: "<class \'tensorflow.python.framework.tensor_like.TensorLike\'>"
+ is_instance: "<class \'tensorflow.python.types.internal.NativeObject\'>"
is_instance: "<class \'tensorflow.python.framework.composite_tensor.CompositeTensor\'>"
is_instance: "<type \'object\'>"
member {
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.summary.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.summary.pbtxt
index a81480f..6cadb56 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.summary.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.summary.pbtxt
@@ -41,6 +41,10 @@
argspec: "args=[\'name\', \'data\', \'step\', \'description\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
}
member_method {
+ name: "should_record_summaries"
+ argspec: "args=[], varargs=None, keywords=None, defaults=None"
+ }
+ member_method {
name: "text"
argspec: "args=[\'name\', \'data\', \'step\', \'description\'], varargs=None, keywords=None, defaults=[\'None\', \'None\'], "
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint-options.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint-options.pbtxt
new file mode 100644
index 0000000..b86e4cb
--- /dev/null
+++ b/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint-options.pbtxt
@@ -0,0 +1,13 @@
+path: "tensorflow.train.CheckpointOptions"
+tf_class {
+ is_instance: "<class \'tensorflow.python.training.saving.checkpoint_options.CheckpointOptions\'>"
+ is_instance: "<type \'object\'>"
+ member {
+ name: "experimental_io_device"
+ mtype: "<type \'member_descriptor\'>"
+ }
+ member_method {
+ name: "__init__"
+ argspec: "args=[\'self\', \'experimental_io_device\'], varargs=None, keywords=None, defaults=[\'None\'], "
+ }
+}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint.pbtxt
index d7e93a0..5665127 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.train.-checkpoint.pbtxt
@@ -14,18 +14,18 @@
}
member_method {
name: "read"
- argspec: "args=[\'self\', \'save_path\'], varargs=None, keywords=None, defaults=None"
+ argspec: "args=[\'self\', \'save_path\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
name: "restore"
- argspec: "args=[\'self\', \'save_path\'], varargs=None, keywords=None, defaults=None"
+ argspec: "args=[\'self\', \'save_path\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
name: "save"
- argspec: "args=[\'self\', \'file_prefix\'], varargs=None, keywords=None, defaults=None"
+ argspec: "args=[\'self\', \'file_prefix\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
member_method {
name: "write"
- argspec: "args=[\'self\', \'file_prefix\'], varargs=None, keywords=None, defaults=None"
+ argspec: "args=[\'self\', \'file_prefix\', \'options\'], varargs=None, keywords=None, defaults=[\'None\'], "
}
}
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.train.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.train.pbtxt
index 13dc982..f354e5d 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.train.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.train.pbtxt
@@ -13,6 +13,10 @@
mtype: "<type \'type\'>"
}
member {
+ name: "CheckpointOptions"
+ mtype: "<type \'type\'>"
+ }
+ member {
name: "ClusterDef"
mtype: "<class \'google.protobuf.pyext.cpp_message.GeneratedProtocolMessageType\'>"
}
diff --git a/tensorflow/tools/build_info/BUILD b/tensorflow/tools/build_info/BUILD
index 91627d4..eb5f909 100644
--- a/tensorflow/tools/build_info/BUILD
+++ b/tensorflow/tools/build_info/BUILD
@@ -9,6 +9,7 @@
py_binary(
name = "gen_build_info",
srcs = ["gen_build_info.py"],
+ exec_compatible_with = ["@local_execution_config_platform//:platform_constraint"],
python_version = "PY3",
srcs_version = "PY2AND3",
deps = [
diff --git a/tensorflow/tools/ci_build/Dockerfile.micro b/tensorflow/tools/ci_build/Dockerfile.micro
index ead38b2..5da2621 100644
--- a/tensorflow/tools/ci_build/Dockerfile.micro
+++ b/tensorflow/tools/ci_build/Dockerfile.micro
@@ -5,5 +5,5 @@
LABEL maintainer="Pete Warden <petewarden@google.com>"
-RUN apt-get update && apt-get install -y zip
+RUN apt-get update && apt-get install -y zip xxd
RUN pip install six
\ No newline at end of file
diff --git a/tensorflow/tools/def_file_filter/symbols_pybind.txt b/tensorflow/tools/def_file_filter/symbols_pybind.txt
index 40ada68..c71bdc6 100644
--- a/tensorflow/tools/def_file_filter/symbols_pybind.txt
+++ b/tensorflow/tools/def_file_filter/symbols_pybind.txt
@@ -76,18 +76,27 @@
tensorflow::Status::error_message
tensorflow::Status::ok()
-[core_cpu_impl] # device_lib, tfe, tf_session
+[device] # device_lib, tfe, tf_session
tensorflow::Device::attributes
+
+[device_factory] # device_lib, tfe, tf_session
tensorflow::DeviceFactory::AddDevices
-tensorflow::SessionOptions::SessionOptions
-tensorflow::DoQuantizeTrainingOnSerializedGraphDef
tensorflow::DeviceFactory::ListAllPhysicalDevices
+
+[session_options] # device_lib, tfe, tf_session
+tensorflow::SessionOptions::SessionOptions
+
+[core_cpu_rump_impl] # quantize_training
+tensorflow::DoQuantizeTrainingOnSerializedGraphDef
+
+[session_state] # tf_session
tensorflow::SessionState::kTensorHandleResourceTypeName
[server_lib] # server_lib
-tensorflow::data::GrpcDataServer::Start
-tensorflow::data::GrpcDataServer::Stop
-tensorflow::data::GrpcDataServer::Target
+tensorflow::data::GrpcDataServerBase::Start
+tensorflow::data::GrpcDataServerBase::Stop
+tensorflow::data::GrpcDataServerBase::Target
+tensorflow::data::MasterGrpcDataServer::NumTasks
tensorflow::data::NewMasterServer
tensorflow::data::NewWorkerServer
@@ -108,6 +117,7 @@
toco::TocoConvert
toco::TocoGetPotentiallySupportedOps
toco::MlirQuantizeModel
+toco::MlirSparsifyModel
[transform_graph_lib] # transform_graph
tensorflow::graph_transforms::TransformGraph
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/cpu-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/cpu-jupyter.Dockerfile
index 14261d3..107d1b4 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/cpu-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/cpu-jupyter.Dockerfile
@@ -53,9 +53,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/devel-cpu-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/devel-cpu-jupyter.Dockerfile
index b1cc8b3..1024f64 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/devel-cpu-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/devel-cpu-jupyter.Dockerfile
@@ -104,9 +104,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/devel-gpu-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/devel-gpu-jupyter.Dockerfile
index 676862b..091ac0d 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/devel-gpu-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/devel-gpu-jupyter.Dockerfile
@@ -146,9 +146,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/gpu-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/gpu-jupyter.Dockerfile
index bb8bf7e..d4d913c 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/gpu-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/gpu-jupyter.Dockerfile
@@ -102,9 +102,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod-jupyter.Dockerfile
index b74c776..db66938 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod-jupyter.Dockerfile
@@ -93,7 +93,7 @@
enum34
# Install bazel
-ARG BAZEL_VERSION=2.0.0
+ARG BAZEL_VERSION=3.0.0
RUN mkdir /bazel && \
wget -O /bazel/installer.sh "https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/bazel-${BAZEL_VERSION}-installer-linux-x86_64.sh" && \
wget -O /bazel/LICENSE.txt "https://raw.githubusercontent.com/bazelbuild/bazel/master/LICENSE" && \
@@ -156,9 +156,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod.Dockerfile
index bfd1fd1..5d90624 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/devel-horovod.Dockerfile
@@ -93,7 +93,7 @@
enum34
# Install bazel
-ARG BAZEL_VERSION=2.0.0
+ARG BAZEL_VERSION=3.0.0
RUN mkdir /bazel && \
wget -O /bazel/installer.sh "https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/bazel-${BAZEL_VERSION}-installer-linux-x86_64.sh" && \
wget -O /bazel/LICENSE.txt "https://raw.githubusercontent.com/bazelbuild/bazel/master/LICENSE" && \
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/horovod-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/horovod-jupyter.Dockerfile
index 4f07894..00c21e2 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/horovod-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/mkl_horovod/horovod-jupyter.Dockerfile
@@ -105,9 +105,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le-jupyter.Dockerfile
index 0e37354..0a284f4 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le-jupyter.Dockerfile
@@ -46,7 +46,7 @@
# tf-nightly-gpu
ARG TF_PACKAGE=tensorflow
RUN apt-get update && apt-get install -y curl libhdf5-dev wget
-RUN python3 -m pip install --global-option=build_ext \
+RUN python3 -m pip install --no-cache-dir --global-option=build_ext \
--global-option=-I/usr/include/hdf5/serial/ \
--global-option=-L/usr/lib/powerpc64le-linux-gnu/hdf5/serial \
h5py
@@ -66,14 +66,14 @@
MINOR=`python3 -c 'import sys; print(sys.version_info[1])'`; \
PACKAGE=$(wget -qO- ${BASE}"api/xml?xpath=//fileName&wrapper=artifacts" | grep -o "[^<>]*cp${MAJOR}${MINOR}[^<>]*.whl"); \
wget ${BASE}"artifact/tensorflow_pkg/"${PACKAGE}; \
- python3 -m pip install ${PACKAGE}
+ python3 -m pip install --no-cache-dir ${PACKAGE}
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le.Dockerfile
index 47bfeab..831e5ae 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/cpu-ppc64le.Dockerfile
@@ -46,7 +46,7 @@
# tf-nightly-gpu
ARG TF_PACKAGE=tensorflow
RUN apt-get update && apt-get install -y curl libhdf5-dev wget
-RUN python3 -m pip install --global-option=build_ext \
+RUN python3 -m pip install --no-cache-dir --global-option=build_ext \
--global-option=-I/usr/include/hdf5/serial/ \
--global-option=-L/usr/lib/powerpc64le-linux-gnu/hdf5/serial \
h5py
@@ -66,7 +66,7 @@
MINOR=`python3 -c 'import sys; print(sys.version_info[1])'`; \
PACKAGE=$(wget -qO- ${BASE}"api/xml?xpath=//fileName&wrapper=artifacts" | grep -o "[^<>]*cp${MAJOR}${MINOR}[^<>]*.whl"); \
wget ${BASE}"artifact/tensorflow_pkg/"${PACKAGE}; \
- python3 -m pip install ${PACKAGE}
+ python3 -m pip install --no-cache-dir ${PACKAGE}
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le-jupyter.Dockerfile
index b6f0c3c..53ccffd 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le-jupyter.Dockerfile
@@ -91,7 +91,7 @@
enum34
# Build and install bazel
-ENV BAZEL_VERSION 0.15.0
+ENV BAZEL_VERSION 3.0.0
WORKDIR /
RUN mkdir /bazel && \
cd /bazel && \
@@ -105,9 +105,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le.Dockerfile
index 4f83322..1bbe712 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-cpu-ppc64le.Dockerfile
@@ -91,7 +91,7 @@
enum34
# Build and install bazel
-ENV BAZEL_VERSION 0.15.0
+ENV BAZEL_VERSION 3.0.0
WORKDIR /
RUN mkdir /bazel && \
cd /bazel && \
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le-jupyter.Dockerfile
index 8a3fc26..0700a35 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le-jupyter.Dockerfile
@@ -133,7 +133,7 @@
enum34
# Build and install bazel
-ENV BAZEL_VERSION 0.15.0
+ENV BAZEL_VERSION 3.0.0
WORKDIR /
RUN mkdir /bazel && \
cd /bazel && \
@@ -147,9 +147,9 @@
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le.Dockerfile
index d748f70..b6d8ff8 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/devel-gpu-ppc64le.Dockerfile
@@ -133,7 +133,7 @@
enum34
# Build and install bazel
-ENV BAZEL_VERSION 0.15.0
+ENV BAZEL_VERSION 3.0.0
WORKDIR /
RUN mkdir /bazel && \
cd /bazel && \
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le-jupyter.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le-jupyter.Dockerfile
index fe568c0..6ef0810 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le-jupyter.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le-jupyter.Dockerfile
@@ -95,7 +95,7 @@
# tf-nightly-gpu
ARG TF_PACKAGE=tensorflow
RUN apt-get update && apt-get install -y curl libhdf5-dev wget
-RUN python3 -m pip install --global-option=build_ext \
+RUN python3 -m pip install --no-cache-dir --global-option=build_ext \
--global-option=-I/usr/include/hdf5/serial/ \
--global-option=-L/usr/lib/powerpc64le-linux-gnu/hdf5/serial \
h5py
@@ -115,14 +115,14 @@
MINOR=`python3 -c 'import sys; print(sys.version_info[1])'`; \
PACKAGE=$(wget -qO- ${BASE}"api/xml?xpath=//fileName&wrapper=artifacts" | grep -o "[^<>]*cp${MAJOR}${MINOR}[^<>]*.whl"); \
wget ${BASE}"artifact/tensorflow_pkg/"${PACKAGE}; \
- python3 -m pip install ${PACKAGE}
+ python3 -m pip install --no-cache-dir ${PACKAGE}
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le.Dockerfile b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le.Dockerfile
index 1fce2e5..f10e9f9 100644
--- a/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le.Dockerfile
+++ b/tensorflow/tools/dockerfiles/dockerfiles/ppc64le/gpu-ppc64le.Dockerfile
@@ -95,7 +95,7 @@
# tf-nightly-gpu
ARG TF_PACKAGE=tensorflow
RUN apt-get update && apt-get install -y curl libhdf5-dev wget
-RUN python3 -m pip install --global-option=build_ext \
+RUN python3 -m pip install --no-cache-dir --global-option=build_ext \
--global-option=-I/usr/include/hdf5/serial/ \
--global-option=-L/usr/lib/powerpc64le-linux-gnu/hdf5/serial \
h5py
@@ -115,7 +115,7 @@
MINOR=`python3 -c 'import sys; print(sys.version_info[1])'`; \
PACKAGE=$(wget -qO- ${BASE}"api/xml?xpath=//fileName&wrapper=artifacts" | grep -o "[^<>]*cp${MAJOR}${MINOR}[^<>]*.whl"); \
wget ${BASE}"artifact/tensorflow_pkg/"${PACKAGE}; \
- python3 -m pip install ${PACKAGE}
+ python3 -m pip install --no-cache-dir ${PACKAGE}
COPY bashrc /etc/bash.bashrc
RUN chmod a+rwx /etc/bash.bashrc
diff --git a/tensorflow/tools/dockerfiles/partials/jupyter.partial.Dockerfile b/tensorflow/tools/dockerfiles/partials/jupyter.partial.Dockerfile
index bbe58b7..cd84872 100644
--- a/tensorflow/tools/dockerfiles/partials/jupyter.partial.Dockerfile
+++ b/tensorflow/tools/dockerfiles/partials/jupyter.partial.Dockerfile
@@ -1,6 +1,6 @@
-RUN python3 -m pip install jupyter matplotlib
+RUN python3 -m pip install --no-cache-dir jupyter matplotlib
# Pin ipykernel and nbformat; see https://github.com/ipython/ipykernel/issues/422
-RUN python3 -m pip install jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
+RUN python3 -m pip install --no-cache-dir jupyter_http_over_ws ipykernel==5.1.1 nbformat==4.4.0
RUN jupyter serverextension enable --py jupyter_http_over_ws
RUN mkdir -p /tf/tensorflow-tutorials && chmod -R a+rwx /tf/
diff --git a/tensorflow/tools/dockerfiles/partials/tensorflow-ppc64le.partial.Dockerfile b/tensorflow/tools/dockerfiles/partials/tensorflow-ppc64le.partial.Dockerfile
index 1772a39..f28f4d5 100644
--- a/tensorflow/tools/dockerfiles/partials/tensorflow-ppc64le.partial.Dockerfile
+++ b/tensorflow/tools/dockerfiles/partials/tensorflow-ppc64le.partial.Dockerfile
@@ -5,7 +5,7 @@
# tf-nightly-gpu
ARG TF_PACKAGE=tensorflow
RUN apt-get update && apt-get install -y curl libhdf5-dev wget
-RUN python3 -m pip install --global-option=build_ext \
+RUN python3 -m pip install --no-cache-dir --global-option=build_ext \
--global-option=-I/usr/include/hdf5/serial/ \
--global-option=-L/usr/lib/powerpc64le-linux-gnu/hdf5/serial \
h5py
@@ -25,4 +25,4 @@
MINOR=`python3 -c 'import sys; print(sys.version_info[1])'`; \
PACKAGE=$(wget -qO- ${BASE}"api/xml?xpath=//fileName&wrapper=artifacts" | grep -o "[^<>]*cp${MAJOR}${MINOR}[^<>]*.whl"); \
wget ${BASE}"artifact/tensorflow_pkg/"${PACKAGE}; \
- python3 -m pip install ${PACKAGE}
+ python3 -m pip install --no-cache-dir ${PACKAGE}
diff --git a/tensorflow/tools/docs/generate2.py b/tensorflow/tools/docs/generate2.py
index 12c0b51..6c10cda 100644
--- a/tensorflow/tools/docs/generate2.py
+++ b/tensorflow/tools/docs/generate2.py
@@ -227,7 +227,7 @@
out_path = pathlib.Path(output_dir)
num_files = len(list(out_path.rglob("*")))
- if num_files < 2500:
+ if num_files < 2000:
raise ValueError("The TensorFlow api should be more than 2500 files"
"(found {}).".format(num_files))
expected_path_contents = {
diff --git a/tensorflow/tools/git/BUILD b/tensorflow/tools/git/BUILD
index c1f0577..90eb300 100644
--- a/tensorflow/tools/git/BUILD
+++ b/tensorflow/tools/git/BUILD
@@ -11,6 +11,7 @@
py_binary(
name = "gen_git_source",
srcs = ["gen_git_source.py"],
+ exec_compatible_with = ["@local_execution_config_platform//:platform_constraint"],
python_version = "PY3",
srcs_version = "PY2AND3",
)
diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 0977fb2..3814e9c 100755
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -77,10 +77,18 @@
tf_repositories(path_prefix, tf_repo_name)
tf_bind()
+# Toolchains & platforms required by Tensorflow to build.
+def tf_toolchains():
+ native.register_execution_platforms("@local_execution_config_platform//:platform")
+ native.register_toolchains("@local_execution_config_python//:py_toolchain")
+
# Define all external repositories required by TensorFlow
def tf_repositories(path_prefix = "", tf_repo_name = ""):
"""All external dependencies for TF builds."""
+ # Initialize toolchains and platforms.
+ tf_toolchains()
+
# Loads all external repos to configure RBE builds.
initialize_rbe_configs()
@@ -156,31 +164,31 @@
tf_http_archive(
name = "XNNPACK",
- sha256 = "583e408c9ab9a6ec241a54e0775bc170ed2ea3d1073668c9379dbfe282fa8acc",
- strip_prefix = "XNNPACK-24d9a03a9ee036949f8c56878ecec17ab400fc23",
+ sha256 = "c002c961fd73b87b68074f9fda49d0dcbd0627c783e487a445da16bcd8dfdee6",
+ strip_prefix = "XNNPACK-10a38087936d84ab2879a2e39fc7e204757ff3e8",
urls = [
- "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/24d9a03a9ee036949f8c56878ecec17ab400fc23.zip",
- "https://github.com/google/XNNPACK/archive/24d9a03a9ee036949f8c56878ecec17ab400fc23.zip",
+ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/10a38087936d84ab2879a2e39fc7e204757ff3e8.zip",
+ "https://github.com/google/XNNPACK/archive/10a38087936d84ab2879a2e39fc7e204757ff3e8.zip",
],
)
tf_http_archive(
name = "FXdiv",
- sha256 = "8224ff187cdfa178b8c54d36eea70520391781eda16d13a418ab5ae53289e1ab",
- strip_prefix = "FXdiv-561254d968e5679460e6a0a743206410284d9f46",
+ sha256 = "ab7dfb08829bee33dca38405d647868fb214ac685e379ec7ef2bebcd234cd44d",
+ strip_prefix = "FXdiv-b408327ac2a15ec3e43352421954f5b1967701d1",
urls = [
- "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/FXdiv/archive/561254d968e5679460e6a0a743206410284d9f46.zip",
- "https://github.com/Maratyszcza/FXdiv/archive/561254d968e5679460e6a0a743206410284d9f46.zip",
+ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/FXdiv/archive/b408327ac2a15ec3e43352421954f5b1967701d1.zip",
+ "https://github.com/Maratyszcza/FXdiv/archive/b408327ac2a15ec3e43352421954f5b1967701d1.zip",
],
)
tf_http_archive(
name = "pthreadpool",
- sha256 = "27c039e73846d0bdfe393833e91afafe45e61ba792cc60e1c62808090554ce4d",
- strip_prefix = "pthreadpool-a61ed1ab70389c62f6f699ca1a30a2421d3ea594",
+ sha256 = "c4b148fba41fc937fdf96bc195caadf0cf0be83f1c3e335ef5355934d4501f83",
+ strip_prefix = "pthreadpool-e918b206d26b1f3b2100b0edabf445c18708d2b7",
urls = [
- "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/pthreadpool/archive/a61ed1ab70389c62f6f699ca1a30a2421d3ea594.zip",
- "https://github.com/Maratyszcza/pthreadpool/archive/a61ed1ab70389c62f6f699ca1a30a2421d3ea594.zip",
+ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/pthreadpool/archive/e918b206d26b1f3b2100b0edabf445c18708d2b7.zip",
+ "https://github.com/Maratyszcza/pthreadpool/archive/e918b206d26b1f3b2100b0edabf445c18708d2b7.zip",
],
)
@@ -203,11 +211,11 @@
tf_http_archive(
name = "mkl_dnn_v1",
build_file = clean_dep("//third_party/mkl_dnn:mkldnn_v1.BUILD"),
- sha256 = "a71ec1f27c30b8a176605e8a78444f1f12301a3c313b70ff93290926c140509c",
- strip_prefix = "mkl-dnn-1.2.2",
+ sha256 = "54737bcb4dc1961d32ee75da3ecc529fa48198f8b2ca863a079e19a9c4adb70f",
+ strip_prefix = "oneDNN-1.4",
urls = [
- "https://storage.googleapis.com/mirror.tensorflow.org/github.com/intel/mkl-dnn/archive/v1.2.2.tar.gz",
- "https://github.com/intel/mkl-dnn/archive/v1.2.2.tar.gz",
+ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/oneapi-src/oneDNN/archive/v1.4.tar.gz",
+ "https://github.com/oneapi-src/oneDNN/archive/v1.4.tar.gz",
],
)
@@ -658,8 +666,8 @@
)
# Check out LLVM and MLIR from llvm-project.
- LLVM_COMMIT = "b6d77e792c3339425a733756b970dbac0da119fb"
- LLVM_SHA256 = "9d299e918ee5850afd834ed62f93101b4777f6e0d72c37e672f5a8b3558f8dd4"
+ LLVM_COMMIT = "35cf2f42dda4d708741e06570b2dbe91cec4dc41"
+ LLVM_SHA256 = "c522ae8860a3cc5807d195d521cc1fc5ef3d27c4598762ea7e38185cef71fe05"
LLVM_URLS = [
"https://storage.googleapis.com/mirror.tensorflow.org/github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),
"https://github.com/llvm/llvm-project/archive/{commit}.tar.gz".format(commit = LLVM_COMMIT),
diff --git a/third_party/clog/BUILD.bazel b/third_party/clog/BUILD.bazel
index 6431f98..ee601b8 100644
--- a/third_party/clog/BUILD.bazel
+++ b/third_party/clog/BUILD.bazel
@@ -15,15 +15,13 @@
hdrs = [
"deps/clog/include/clog.h",
],
- copts = [
- "-Wno-unused-result",
- ],
+ copts = select({
+ ":windows": [],
+ "//conditions:default": ["-Wno-unused-result"],
+ }),
linkopts = select({
- ":android": [
- "-llog",
- ],
- "//conditions:default": [
- ],
+ ":android": ["-llog"],
+ "//conditions:default": [],
}),
linkstatic = True,
strip_include_prefix = "deps/clog/include",
@@ -32,5 +30,9 @@
config_setting(
name = "android",
values = {"crosstool_top": "//external:android/crosstool"},
- visibility = ["//visibility:public"],
+)
+
+config_setting(
+ name = "windows",
+ values = {"cpu": "x64_windows"},
)
diff --git a/third_party/cpuinfo/BUILD.bazel b/third_party/cpuinfo/BUILD.bazel
index 67787e7..22ec4e2 100644
--- a/third_party/cpuinfo/BUILD.bazel
+++ b/third_party/cpuinfo/BUILD.bazel
@@ -98,9 +98,8 @@
srcs = select({
":linux_x86_64": COMMON_SRCS + X86_SRCS + LINUX_SRCS + LINUX_X86_SRCS,
":linux_aarch64": COMMON_SRCS + ARM_SRCS + LINUX_SRCS + LINUX_ARM64_SRCS,
- ":linux_armhf": COMMON_SRCS + ARM_SRCS + LINUX_SRCS + LINUX_ARM32_SRCS,
- ":linux_arm": COMMON_SRCS + ARM_SRCS + LINUX_SRCS + LINUX_ARM32_SRCS,
":macos_x86_64": COMMON_SRCS + X86_SRCS + MACH_SRCS + MACH_X86_SRCS,
+ ":windows_x86_64": COMMON_SRCS + X86_SRCS + WINDOWS_X86_SRCS,
":android_armv7": COMMON_SRCS + ARM_SRCS + LINUX_SRCS + LINUX_ARM32_SRCS + ANDROID_ARM_SRCS,
":android_arm64": COMMON_SRCS + ARM_SRCS + LINUX_SRCS + LINUX_ARM64_SRCS + ANDROID_ARM_SRCS,
":android_x86": COMMON_SRCS + X86_SRCS + LINUX_SRCS + LINUX_X86_SRCS,
@@ -116,9 +115,11 @@
":watchos_arm64_32": COMMON_SRCS + MACH_SRCS + MACH_ARM_SRCS,
":tvos_x86_64": COMMON_SRCS + X86_SRCS + MACH_SRCS + MACH_X86_SRCS,
":tvos_arm64": COMMON_SRCS + MACH_SRCS + MACH_ARM_SRCS,
- ":emscripten_wasm": COMMON_SRCS + EMSCRIPTEN_SRCS,
}),
- copts = C99OPTS + [
+ copts = select({
+ ":windows_x86_64": [],
+ "//conditions:default": C99OPTS,
+ }) + [
"-Iexternal/cpuinfo/include",
"-Iexternal/cpuinfo/src",
],
@@ -170,20 +171,6 @@
)
config_setting(
- name = "linux_arm",
- values = {"cpu": "arm"},
- visibility = ["//visibility:public"],
-)
-
-config_setting(
- name = "linux_armhf",
- values = {
- "cpu": "armhf",
- },
- visibility = ["//visibility:public"],
-)
-
-config_setting(
name = "macos_x86_64",
values = {
"apple_platform_type": "macos",
@@ -192,6 +179,11 @@
)
config_setting(
+ name = "windows_x86_64",
+ values = {"cpu": "x64_windows"},
+)
+
+config_setting(
name = "android_armv7",
values = {
"crosstool_top": "//external:android/crosstool",
diff --git a/third_party/gpus/cuda/cuda_config.h.tpl b/third_party/gpus/cuda/cuda_config.h.tpl
index 916315d..dbd8463 100644
--- a/third_party/gpus/cuda/cuda_config.h.tpl
+++ b/third_party/gpus/cuda/cuda_config.h.tpl
@@ -16,8 +16,6 @@
#ifndef CUDA_CUDA_CONFIG_H_
#define CUDA_CUDA_CONFIG_H_
-#define TF_CUDA_CAPABILITIES %{cuda_compute_capabilities}
-
#define TF_CUDA_VERSION "%{cuda_version}"
#define TF_CUDA_LIB_VERSION "%{cuda_lib_version}"
#define TF_CUDNN_VERSION "%{cudnn_version}"
diff --git a/third_party/gpus/cuda_configure.bzl b/third_party/gpus/cuda_configure.bzl
index 347d117..545aeeb 100644
--- a/third_party/gpus/cuda_configure.bzl
+++ b/third_party/gpus/cuda_configure.bzl
@@ -771,10 +771,6 @@
"%{cuda_version}": "",
"%{cuda_lib_version}": "",
"%{cudnn_version}": "",
- "%{cuda_compute_capabilities}": ",".join([
- "CudaVersion(\"%s\")" % c
- for c in _DEFAULT_CUDA_COMPUTE_CAPABILITIES
- ]),
"%{cuda_toolkit_path}": "",
},
"cuda/cuda/cuda_config.h",
@@ -1140,10 +1136,6 @@
"%{cuda_version}": cuda_config.cuda_version,
"%{cuda_lib_version}": cuda_config.cuda_lib_version,
"%{cudnn_version}": cuda_config.cudnn_version,
- "%{cuda_compute_capabilities}": ", ".join([
- "CudaVersion(\"%s\")" % c
- for c in cuda_config.compute_capabilities
- ]),
"%{cuda_toolkit_path}": cuda_config.cuda_toolkit_path,
},
)
diff --git a/third_party/llvm/llvm.autogenerated.BUILD b/third_party/llvm/llvm.autogenerated.BUILD
index 92c2469..d283e68 100644
--- a/third_party/llvm/llvm.autogenerated.BUILD
+++ b/third_party/llvm/llvm.autogenerated.BUILD
@@ -1934,6 +1934,7 @@
]),
copts = llvm_copts,
deps = [
+ ":binary_format",
":config",
":debug_info_code_view",
":debug_info_msf",
@@ -4330,7 +4331,6 @@
":x86_defs",
":x86_desc",
":x86_info",
- ":x86_utils",
],
)
@@ -4355,7 +4355,6 @@
":mc_disassembler",
":support",
":x86_info",
- ":x86_utils",
],
)
@@ -4405,27 +4404,6 @@
)
cc_library(
- name = "x86_utils",
- srcs = glob([
- "lib/Target/X86/Utils/*.c",
- "lib/Target/X86/Utils/*.cpp",
- "lib/Target/X86/Utils/*.inc",
- ]),
- hdrs = glob([
- "include/llvm/Target/X86/Utils/*.h",
- "include/llvm/Target/X86/Utils/*.def",
- "include/llvm/Target/X86/Utils/*.inc",
- "lib/Target/X86/Utils/*.h",
- ]),
- copts = llvm_copts + ["-Iexternal/llvm-project/llvm/lib/Target/X86"],
- deps = [
- ":code_gen",
- ":config",
- ":support",
- ],
-)
-
-cc_library(
name = "x_core_code_gen",
srcs = glob([
"lib/Target/XCore/*.c",
diff --git a/third_party/mkl_dnn/mkldnn_v1.BUILD b/third_party/mkl_dnn/mkldnn_v1.BUILD
index b916738..243ec00 100644
--- a/third_party/mkl_dnn/mkldnn_v1.BUILD
+++ b/third_party/mkl_dnn/mkldnn_v1.BUILD
@@ -43,7 +43,7 @@
out = "include/dnnl_version.h",
substitutions = {
"@DNNL_VERSION_MAJOR@": "1",
- "@DNNL_VERSION_MINOR@": "2",
+ "@DNNL_VERSION_MINOR@": "4",
"@DNNL_VERSION_PATCH@": "0",
"@DNNL_VERSION_HASH@": "N/A",
},
diff --git a/third_party/mlir/BUILD b/third_party/mlir/BUILD
index eedac45..74dfe01 100644
--- a/third_party/mlir/BUILD
+++ b/third_party/mlir/BUILD
@@ -2,6 +2,7 @@
# The MLIR "Multi-Level Intermediate Representation" Compiler Infrastructure
load("@org_tensorflow//third_party/mlir:tblgen.bzl", "gentbl")
+load("@org_tensorflow//third_party/mlir:linalggen.bzl", "genlinalg")
licenses(["notice"])
@@ -278,6 +279,7 @@
name = "LoopOpsTdFiles",
srcs = [
"include/mlir/Dialect/LoopOps/LoopOps.td",
+ "include/mlir/Interfaces/ControlFlowInterfaces.td",
"include/mlir/Interfaces/LoopLikeInterface.td",
"include/mlir/Interfaces/SideEffects.td",
":OpBaseTdFiles",
@@ -578,6 +580,7 @@
]),
includes = ["include"],
deps = [
+ ":ControlFlowInterfaces",
":EDSC",
":IR",
":LoopLikeInterface",
@@ -1898,6 +1901,7 @@
deps = [
":Affine",
":Analysis",
+ ":ControlFlowInterfaces",
":IR",
":LoopLikeInterface",
":LoopOps",
@@ -1907,7 +1911,6 @@
":Support",
":TransformUtils",
":TransformsPassIncGen",
- ":VectorOps",
"@llvm-project//llvm:support",
],
)
@@ -2689,6 +2692,10 @@
srcs = glob([
"tools/mlir-linalg-ods-gen/mlir-linalg-ods-gen.cpp",
]),
+ linkopts = [
+ "-lm",
+ "-lpthread",
+ ],
deps = [
":IR",
":Support",
@@ -2869,9 +2876,26 @@
],
)
+genlinalg(
+ name = "LinalgNamedStructuredOpsIncGen",
+ src = "include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOpsSpec.tc",
+ linalg_outs = [
+ (
+ "-gen-impl",
+ "include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.cpp.inc",
+ ),
+ (
+ "-gen-ods-decl",
+ "include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.td",
+ ),
+ ],
+ linalggen = ":mlir-linalg-ods-gen",
+)
+
filegroup(
name = "LinalgStructuredOpsTdFiles",
srcs = [
+ "include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.td",
"include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td",
"include/mlir/Dialect/Linalg/IR/LinalgStructuredOpsInterface.td",
":AffineOpsTdFiles",
@@ -2905,6 +2929,7 @@
td_file = "include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td",
td_srcs = [
":LinalgStructuredOpsTdFiles",
+ "include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.td",
],
)
@@ -3020,14 +3045,18 @@
"lib/Dialect/Linalg/IR/LinalgTypes.cpp",
],
hdrs = [
+ "include/mlir/Dialect/Linalg/EDSC/Intrinsics.h",
"include/mlir/Dialect/Linalg/IR/LinalgOps.h",
"include/mlir/Dialect/Linalg/IR/LinalgTraits.h",
"include/mlir/Dialect/Linalg/IR/LinalgTypes.h",
],
includes = ["include"],
deps = [
+ ":Affine",
":DialectUtils",
+ ":EDSC",
":IR",
+ ":LinalgNamedStructuredOpsIncGen",
":LinalgOpsIncGen",
":LinalgStructuredOpsIncGen",
":Parser",
@@ -3068,7 +3097,6 @@
"include/mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h",
"include/mlir/Dialect/Linalg/EDSC/Builders.h",
"include/mlir/Dialect/Linalg/EDSC/FoldedIntrinsics.h",
- "include/mlir/Dialect/Linalg/EDSC/Intrinsics.h",
"include/mlir/Dialect/Linalg/Passes.h",
"include/mlir/Dialect/Linalg/Transforms/LinalgTransforms.h",
"include/mlir/Dialect/Linalg/Utils/Utils.h",
diff --git a/third_party/mlir/linalggen.bzl b/third_party/mlir/linalggen.bzl
new file mode 100644
index 0000000..5162911
--- /dev/null
+++ b/third_party/mlir/linalggen.bzl
@@ -0,0 +1,39 @@
+"""BUILD extensions for MLIR linalg generation."""
+
+def genlinalg(name, linalggen, src, linalg_outs):
+ """genlinalg() generates code from a tc spec file.
+
+ Args:
+ name: The name of the build rule for use in dependencies.
+ linalggen: The binary used to produce the output.
+ src: The tc spec file.
+ linalg_outs: A list of tuples (opts, out), where each opts is a string of
+ options passed to linalggen, and the out is the corresponding output file
+ produced.
+ """
+
+ for (opts, out) in linalg_outs:
+ # All arguments to generate the output except output destination.
+ base_args = [
+ "$(location %s)" % linalggen,
+ "%s" % opts,
+ "$(location %s)" % src,
+ ]
+ rule_suffix = "_".join(opts.replace("-", "_").replace("=", "_").split(" "))
+
+ # Rule to generate code using generated shell script.
+ native.genrule(
+ name = "%s_%s_genrule" % (name, rule_suffix),
+ srcs = [src],
+ outs = [out],
+ tools = [linalggen],
+ cmd = ("echo " + " ".join(base_args) + " -o $@; " + " ".join(base_args) + " -o $@"),
+ )
+
+ # List of opts that do not generate cc files.
+ hdrs = [f for (opts, f) in linalg_outs]
+ native.cc_library(
+ name = name,
+ hdrs = hdrs,
+ textual_hdrs = hdrs,
+ )
diff --git a/third_party/mlir/tblgen.bzl b/third_party/mlir/tblgen.bzl
index 6434bba..bbe64c1 100644
--- a/third_party/mlir/tblgen.bzl
+++ b/third_party/mlir/tblgen.bzl
@@ -21,6 +21,7 @@
srcs += [td_file]
td_includes_cmd = ["-I external/llvm-project/mlir/include -I external/org_tensorflow"]
+ td_includes_cmd += ["-I $(GENDIR)/external/llvm-project/mlir/include"]
for td_include in td_includes:
td_includes_cmd += ["-I%s" % td_include]
local_inc = "-I $$(dirname $(location %s))" % td_file
diff --git a/third_party/mlir/test.BUILD b/third_party/mlir/test.BUILD
index 5e761ac..0902a48 100644
--- a/third_party/mlir/test.BUILD
+++ b/third_party/mlir/test.BUILD
@@ -121,6 +121,7 @@
":TestOpsIncGen",
"@llvm-project//llvm:support",
"@llvm-project//mlir:ControlFlowInterfaces",
+ "@llvm-project//mlir:DerivedAttributeOpInterface",
"@llvm-project//mlir:Dialect",
"@llvm-project//mlir:IR",
"@llvm-project//mlir:InferTypeOpInterface",
diff --git a/third_party/psimd/workspace.bzl b/third_party/psimd/workspace.bzl
index ca0bca7..03d010c 100644
--- a/third_party/psimd/workspace.bzl
+++ b/third_party/psimd/workspace.bzl
@@ -5,11 +5,11 @@
def repo():
third_party_http_archive(
name = "psimd",
- strip_prefix = "psimd-10b4ffc6ea9e2e11668f86969586f88bc82aaefa",
- sha256 = "1fefd66702cb2eb3462b962f33d4fb23d59a55d5889ee6372469d286c4512df4",
+ strip_prefix = "psimd-85427dd4c8521cc037a1ffa6fcd25c55fafc8a00",
+ sha256 = "db23c2bc4a58d6f40c181797e43103300edac7cf9d286ca81590543f66ab95d2",
urls = [
- "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/psimd/archive/10b4ffc6ea9e2e11668f86969586f88bc82aaefa.tar.gz",
- "https://github.com/Maratyszcza/psimd/archive/10b4ffc6ea9e2e11668f86969586f88bc82aaefa.tar.gz",
+ "https://storage.googleapis.com/mirror.tensorflow.org/github.com/Maratyszcza/psimd/archive/85427dd4c8521cc037a1ffa6fcd25c55fafc8a00.zip",
+ "https://github.com/Maratyszcza/psimd/archive/85427dd4c8521cc037a1ffa6fcd25c55fafc8a00.zip",
],
build_file = "//third_party/psimd:BUILD.bazel",
)
diff --git a/third_party/py/BUILD.tpl b/third_party/py/BUILD.tpl
index 3a0be1b..08ba167 100644
--- a/third_party/py/BUILD.tpl
+++ b/third_party/py/BUILD.tpl
@@ -28,6 +28,8 @@
name = "py_toolchain",
toolchain = ":py_runtime_pair",
toolchain_type = "@bazel_tools//tools/python:toolchain_type",
+ target_compatible_with = [%{PLATFORM_CONSTRAINT}],
+ exec_compatible_with = [%{PLATFORM_CONSTRAINT}],
)
# To build Python C/C++ extension on Windows, we need to link to python import library pythonXY.lib
diff --git a/third_party/py/python_configure.bzl b/third_party/py/python_configure.bzl
index 6e9a22f..2f75262 100644
--- a/third_party/py/python_configure.bzl
+++ b/third_party/py/python_configure.bzl
@@ -240,11 +240,15 @@
"numpy_include",
)
+ platform_constraint = ""
+ if repository_ctx.attr.platform_constraint:
+ platform_constraint = "\"%s\"" % repository_ctx.attr.platform_constraint
repository_ctx.template("BUILD", build_tpl, {
"%{PYTHON_BIN_PATH}": python_bin,
"%{PYTHON_INCLUDE_GENRULE}": python_include_rule,
"%{PYTHON_IMPORT_LIB_GENRULE}": python_import_lib_genrule,
"%{NUMPY_INCLUDE_GENRULE}": numpy_include_rule,
+ "%{PLATFORM_CONSTRAINT}": platform_constraint,
})
def _create_remote_python_repository(repository_ctx, remote_config_repo):
@@ -268,18 +272,31 @@
PYTHON_LIB_PATH,
]
+local_python_configure = repository_rule(
+ implementation = _create_local_python_repository,
+ environ = _ENVIRONS,
+ attrs = {
+ "environ": attr.string_dict(),
+ "platform_constraint": attr.string(),
+ },
+)
+
remote_python_configure = repository_rule(
implementation = _create_local_python_repository,
environ = _ENVIRONS,
remotable = True,
attrs = {
"environ": attr.string_dict(),
+ "platform_constraint": attr.string(),
},
)
python_configure = repository_rule(
implementation = _python_autoconf_impl,
environ = _ENVIRONS + [TF_PYTHON_CONFIG_REPO],
+ attrs = {
+ "platform_constraint": attr.string(),
+ },
)
"""Detects and configures the local Python.
diff --git a/third_party/remote_config/BUILD.tpl b/third_party/remote_config/BUILD.tpl
index 7bcee41..e7f9300 100644
--- a/third_party/remote_config/BUILD.tpl
+++ b/third_party/remote_config/BUILD.tpl
@@ -1,8 +1,25 @@
+# Each platform creates a constraint @<platform>//:platform_constraint that
+# is listed in its constraint_values; rule that want to select a specific
+# platform to run on can put @<platform>//:platform_constraing into their
+# exec_compatible_with attribute.
+# Toolchains can similarly be marked with target_compatible_with or
+# exec_compatible_with to bind them to this platform.
+constraint_setting(
+ name = "platform_setting"
+)
+
+constraint_value(
+ name = "platform_constraint",
+ constraint_setting = ":platform_setting",
+ visibility = ["//visibility:public"],
+)
+
platform(
name = "platform",
constraint_values = [
"@bazel_tools//platforms:x86_64",
"@bazel_tools//platforms:%{platform}",
+ ":platform_constraint",
],
exec_properties = %{exec_properties},
)
diff --git a/third_party/remote_config/remote_platform_configure.bzl b/third_party/remote_config/remote_platform_configure.bzl
index 5c2918b..1031fbf 100644
--- a/third_party/remote_config/remote_platform_configure.bzl
+++ b/third_party/remote_config/remote_platform_configure.bzl
@@ -2,6 +2,14 @@
def _remote_platform_configure_impl(repository_ctx):
platform = repository_ctx.attr.platform
+ if platform == "local":
+ os = repository_ctx.os.name.lower()
+ if os.startswith("windows"):
+ platform = "windows"
+ elif os.startswith("mac os"):
+ platform = "osx"
+ else:
+ platform = "linux"
exec_properties = repository_ctx.attr.platform_exec_properties
serialized_exec_properties = "{"
@@ -22,6 +30,6 @@
implementation = _remote_platform_configure_impl,
attrs = {
"platform_exec_properties": attr.string_dict(mandatory = True),
- "platform": attr.string(default = "linux", values = ["linux", "windows"]),
+ "platform": attr.string(default = "linux", values = ["linux", "windows", "local"]),
},
)
diff --git a/third_party/toolchains/remote_config/configs.bzl b/third_party/toolchains/remote_config/configs.bzl
index 4c94abf..c97d47f 100644
--- a/third_party/toolchains/remote_config/configs.bzl
+++ b/third_party/toolchains/remote_config/configs.bzl
@@ -1,8 +1,12 @@
"""Configurations of RBE builds used with remote config."""
-load("//third_party/toolchains/remote_config:rbe_config.bzl", "tensorflow_rbe_config", "tensorflow_rbe_win_config")
+load("//third_party/toolchains/remote_config:rbe_config.bzl", "tensorflow_local_config", "tensorflow_rbe_config", "tensorflow_rbe_win_config")
def initialize_rbe_configs():
+ tensorflow_local_config(
+ name = "local_execution",
+ )
+
tensorflow_rbe_config(
name = "ubuntu16.04-manylinux2010-py3",
os = "ubuntu16.04-manylinux2010",
diff --git a/third_party/toolchains/remote_config/rbe_config.bzl b/third_party/toolchains/remote_config/rbe_config.bzl
index 6709cad..597ab55 100644
--- a/third_party/toolchains/remote_config/rbe_config.bzl
+++ b/third_party/toolchains/remote_config/rbe_config.bzl
@@ -1,6 +1,6 @@
"""Macro that creates external repositories for remote config."""
-load("//third_party/py:python_configure.bzl", "remote_python_configure")
+load("//third_party/py:python_configure.bzl", "local_python_configure", "remote_python_configure")
load("//third_party/gpus:cuda_configure.bzl", "remote_cuda_configure")
load("//third_party/nccl:nccl_configure.bzl", "remote_nccl_configure")
load("//third_party/gpus:rocm_configure.bzl", "remote_rocm_configure")
@@ -113,6 +113,7 @@
name = "%s_config_python" % name,
environ = env,
exec_properties = exec_properties,
+ platform_constraint = "@%s_config_platform//:platform_constraint" % name,
)
remote_rocm_configure(
@@ -127,10 +128,17 @@
"Pool": "default",
}
+ remote_platform_configure(
+ name = "%s_config_platform" % name,
+ platform = "linux",
+ platform_exec_properties = exec_properties,
+ )
+
remote_python_configure(
name = "%s_config_python" % name,
environ = env,
exec_properties = exec_properties,
+ platform_constraint = "@%s_config_platform//:platform_constraint" % name,
)
else:
fail("Neither cuda_version, rocm_version nor python_version specified.")
@@ -156,7 +164,20 @@
name = "%s_config_python" % name,
environ = env,
exec_properties = exec_properties,
+ platform_constraint = "@%s_config_platform//:platform_constraint" % name,
+ )
+
+def _tensorflow_local_config(name):
+ remote_platform_configure(
+ name = "%s_config_platform" % name,
+ platform = "local",
+ platform_exec_properties = {},
+ )
+ local_python_configure(
+ name = "%s_config_python" % name,
+ platform_constraint = "@%s_config_platform//:platform_constraint" % name,
)
tensorflow_rbe_config = _tensorflow_rbe_config
tensorflow_rbe_win_config = _tensorflow_rbe_win_config
+tensorflow_local_config = _tensorflow_local_config
