arm_compute v19.05
diff --git a/src/graph/Graph.cpp b/src/graph/Graph.cpp
index 88e2682..9d437b1 100644
--- a/src/graph/Graph.cpp
+++ b/src/graph/Graph.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -152,7 +152,7 @@
return true;
}
-TensorID Graph::create_tensor(TensorDescriptor desc)
+TensorID Graph::create_tensor(const TensorDescriptor &desc)
{
TensorID tid = _tensors.size();
auto tensor = support::cpp14::make_unique<Tensor>(tid, desc);
diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index a944d2c..5db9540 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -30,15 +30,19 @@
#include "support/ToolchainSupport.h"
-#define CHECK_NODEIDX_PAIR(pair, g) \
- ARM_COMPUTE_ERROR_ON(((pair).node_id >= (g).nodes().size()) || ((g).node((pair).node_id) == nullptr) || ((pair).index >= (g).node((pair).node_id)->num_outputs()));
-
namespace arm_compute
{
namespace graph
{
namespace
{
+inline void check_nodeidx_pair(const NodeIdxPair &pair, const Graph &g)
+{
+ ARM_COMPUTE_UNUSED(pair);
+ ARM_COMPUTE_UNUSED(g);
+ ARM_COMPUTE_ERROR_ON((pair.node_id >= g.nodes().size()) || (g.node((pair).node_id) == nullptr) || (pair.index >= g.node(pair.node_id)->num_outputs()));
+}
+
Status set_node_params(Graph &g, NodeID nid, NodeParams ¶ms)
{
INode *node = g.node(nid);
@@ -62,10 +66,10 @@
return Status{};
}
-NodeID add_const_node_with_name(Graph &g, NodeParams params, const std::string &name, TensorDescriptor desc, ITensorAccessorUPtr accessor)
+NodeID add_const_node_with_name(Graph &g, NodeParams params, const std::string &name, const TensorDescriptor &desc, ITensorAccessorUPtr accessor)
{
params.name = params.name.empty() ? "" : params.name + name;
- auto nid = GraphBuilder::add_const_node(g, params, std::move(desc), std::move(accessor));
+ auto nid = GraphBuilder::add_const_node(g, params, desc, std::move(accessor));
set_node_params(g, nid, params);
return nid;
}
@@ -73,7 +77,7 @@
template <typename NT, typename... Args>
NodeID create_simple_single_input_output_node(Graph &g, NodeParams ¶ms, NodeIdxPair input, Args &&... args)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
NodeID nid = g.add_node<NT>(std::forward<Args>(args)...);
g.add_connection(input.node_id, input.index, nid, 0);
@@ -81,9 +85,27 @@
return nid;
}
+
+template <typename NT, typename... Args>
+NodeID create_simple_multiple_input_single_output_node(Graph &g, NodeParams ¶ms, const std::vector<NodeIdxPair> &inputs, Args &&... args)
+{
+ ARM_COMPUTE_ERROR_ON(inputs.size() == 0);
+
+ NodeID nid = g.add_node<NT>(std::forward<Args>(args)...);
+
+ unsigned int i = 0;
+ for(const auto &input : inputs)
+ {
+ check_nodeidx_pair(input, g);
+ g.add_connection(input.node_id, input.index, nid, i++);
+ }
+ set_node_params(g, nid, params);
+
+ return nid;
+}
} // namespace
-NodeID GraphBuilder::add_const_node(Graph &g, NodeParams params, TensorDescriptor desc, ITensorAccessorUPtr accessor)
+NodeID GraphBuilder::add_const_node(Graph &g, NodeParams params, const TensorDescriptor &desc, ITensorAccessorUPtr accessor)
{
auto nid = g.add_node<ConstNode>(desc);
set_node_params(g, nid, params);
@@ -91,7 +113,7 @@
return nid;
}
-NodeID GraphBuilder::add_input_node(Graph &g, NodeParams params, TensorDescriptor desc, ITensorAccessorUPtr accessor)
+NodeID GraphBuilder::add_input_node(Graph &g, NodeParams params, const TensorDescriptor &desc, ITensorAccessorUPtr accessor)
{
auto nid = g.add_node<InputNode>(desc);
set_node_params(g, nid, params);
@@ -101,7 +123,7 @@
NodeID GraphBuilder::add_output_node(Graph &g, NodeParams params, NodeIdxPair input, ITensorAccessorUPtr accessor)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
NodeID nid = g.add_node<OutputNode>();
g.add_connection(input.node_id, input.index, nid, 0);
@@ -111,16 +133,17 @@
return nid;
}
-NodeID GraphBuilder::add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info)
+NodeID GraphBuilder::add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info,
+ const QuantizationInfo out_quant_info)
{
- return create_simple_single_input_output_node<ActivationLayerNode>(g, params, input, act_info);
+ return create_simple_single_input_output_node<ActivationLayerNode>(g, params, input, act_info, out_quant_info);
}
NodeID GraphBuilder::add_batch_normalization_node(Graph &g, NodeParams params, NodeIdxPair input, float epsilon,
ITensorAccessorUPtr mean_accessor, ITensorAccessorUPtr var_accessor,
ITensorAccessorUPtr beta_accessor, ITensorAccessorUPtr gamma_accessor)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
bool has_beta = (beta_accessor != nullptr);
bool has_gamma = (gamma_accessor != nullptr);
@@ -170,8 +193,8 @@
NodeID GraphBuilder::add_bounding_box_transform_node(Graph &g, NodeParams params, NodeIdxPair input, NodeIdxPair deltas, BoundingBoxTransformInfo info)
{
- CHECK_NODEIDX_PAIR(input, g);
- CHECK_NODEIDX_PAIR(deltas, g);
+ check_nodeidx_pair(input, g);
+ check_nodeidx_pair(deltas, g);
NodeID nid = g.add_node<BoundingBoxTransformLayerNode>(info);
@@ -194,7 +217,7 @@
const QuantizationInfo weights_quant_info,
const QuantizationInfo out_quant_info)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
ARM_COMPUTE_ERROR_ON(depth == 0);
ARM_COMPUTE_ERROR_ON((kernel_spatial_extend.width == 0) || (kernel_spatial_extend.height == 0));
@@ -202,14 +225,15 @@
// Get input tensor descriptor
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
+ const DataLayout input_data_layout = input_tensor_desc.layout;
// Create weights node
TensorDescriptor w_desc = input_tensor_desc;
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::CHANNEL),
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::CHANNEL),
get_dimension_size(input_tensor_desc, DataLayoutDimension::CHANNEL) / num_groups);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::BATCHES), depth);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::BATCHES), depth);
if(!weights_quant_info.empty())
{
w_desc.quant_info = weights_quant_info;
@@ -248,7 +272,7 @@
Size2D inner_border, ITensorAccessorUPtr weights_accessor,
ITensorAccessorUPtr bias_accessor)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
ARM_COMPUTE_ERROR_ON(depth == 0);
ARM_COMPUTE_ERROR_ON((kernel_spatial_extend.width == 0) || (kernel_spatial_extend.height == 0));
@@ -256,14 +280,15 @@
// Get input tensor descriptor
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
+ const DataLayout input_data_layout = input_tensor_desc.layout;
// Create weights node
TensorDescriptor w_desc = input_tensor_desc;
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::CHANNEL),
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::CHANNEL),
get_dimension_size(input_tensor_desc, DataLayoutDimension::CHANNEL));
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::BATCHES), depth);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::BATCHES), depth);
NodeID w_nid = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
@@ -293,40 +318,29 @@
return deconv_nid;
}
-NodeID GraphBuilder::add_concatenate_node(Graph &g, NodeParams params, std::vector<NodeIdxPair> inputs, DataLayoutDimension axis)
+NodeID GraphBuilder::add_concatenate_node(Graph &g, NodeParams params, const std::vector<NodeIdxPair> &inputs, descriptors::ConcatLayerDescriptor concat_descriptor)
{
- ARM_COMPUTE_ERROR_ON(inputs.size() == 0);
-
- NodeID nid = g.add_node<ConcatenateLayerNode>(inputs.size(), axis);
-
- unsigned int i = 0;
- for(const auto &input : inputs)
- {
- CHECK_NODEIDX_PAIR(input, g);
- g.add_connection(input.node_id, input.index, nid, i++);
- }
- set_node_params(g, nid, params);
-
- return nid;
+ return create_simple_multiple_input_single_output_node<ConcatenateLayerNode>(g, params, inputs, inputs.size(), concat_descriptor);
}
NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params, NodeIdxPair input, Size2D kernel_spatial_extend,
PadStrideInfo conv_info, int depth_multiplier, DepthwiseConvolutionMethod method,
- ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const QuantizationInfo quant_info)
+ ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const QuantizationInfo quant_info, const QuantizationInfo out_quant_info)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
ARM_COMPUTE_ERROR_ON((kernel_spatial_extend.width == 0) || (kernel_spatial_extend.height == 0));
bool has_bias = (bias_accessor != nullptr);
// Get input tensor descriptor
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
+ const DataLayout input_data_layout = input_tensor_desc.layout;
// Create weights node
TensorDescriptor w_desc = input_tensor_desc;
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
- w_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::CHANNEL),
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::WIDTH), kernel_spatial_extend.width);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::HEIGHT), kernel_spatial_extend.height);
+ w_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::CHANNEL),
get_dimension_size(input_tensor_desc, DataLayoutDimension::CHANNEL) * depth_multiplier);
if(!quant_info.empty())
{
@@ -351,7 +365,7 @@
}
// Create convolution node and connect
- NodeID conv_nid = g.add_node<DepthwiseConvolutionLayerNode>(conv_info, depth_multiplier, method);
+ NodeID conv_nid = g.add_node<DepthwiseConvolutionLayerNode>(conv_info, depth_multiplier, method, out_quant_info);
g.add_connection(input.node_id, input.index, conv_nid, 0);
g.add_connection(w_nid, 0, conv_nid, 1);
if(has_bias)
@@ -362,11 +376,11 @@
return conv_nid;
}
-NodeID GraphBuilder::add_detection_output_node(Graph &g, NodeParams params, NodeIdxPair input_loc, NodeIdxPair input_conf, NodeIdxPair input_priorbox, DetectionOutputLayerInfo detect_info)
+NodeID GraphBuilder::add_detection_output_node(Graph &g, NodeParams params, NodeIdxPair input_loc, NodeIdxPair input_conf, NodeIdxPair input_priorbox, const DetectionOutputLayerInfo &detect_info)
{
- CHECK_NODEIDX_PAIR(input_loc, g);
- CHECK_NODEIDX_PAIR(input_conf, g);
- CHECK_NODEIDX_PAIR(input_priorbox, g);
+ check_nodeidx_pair(input_loc, g);
+ check_nodeidx_pair(input_conf, g);
+ check_nodeidx_pair(input_priorbox, g);
// Create detection_output node and connect
NodeID detect_nid = g.add_node<DetectionOutputLayerNode>(detect_info);
@@ -386,8 +400,8 @@
NodeID GraphBuilder::add_elementwise_node(Graph &g, NodeParams params, NodeIdxPair input0, NodeIdxPair input1, EltwiseOperation operation)
{
- CHECK_NODEIDX_PAIR(input0, g);
- CHECK_NODEIDX_PAIR(input1, g);
+ check_nodeidx_pair(input0, g);
+ check_nodeidx_pair(input1, g);
NodeID nid = g.add_node<EltwiseLayerNode>(operation);
@@ -405,11 +419,38 @@
}
NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs,
+ NodeID weights_nid, NodeID bias_nid,
+ const FullyConnectedLayerInfo fc_info, const QuantizationInfo out_quant_info)
+{
+ check_nodeidx_pair(input, g);
+ ARM_COMPUTE_ERROR_ON(num_outputs == 0);
+ ARM_COMPUTE_ERROR_ON(weights_nid == EmptyNodeID);
+
+ const bool has_bias = (bias_nid != EmptyNodeID);
+
+ // Get input tensor descriptor
+ const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
+
+ // Create fully connected node and connect
+ NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info);
+ g.add_connection(input.node_id, input.index, fc_nid, 0);
+ g.add_connection(weights_nid, 0, fc_nid, 1);
+ if(has_bias)
+ {
+ g.add_connection(bias_nid, 0, fc_nid, 2);
+ }
+
+ set_node_params(g, fc_nid, params);
+
+ return fc_nid;
+}
+
+NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs,
ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor,
const FullyConnectedLayerInfo fc_info,
const QuantizationInfo weights_quant_info, const QuantizationInfo out_quant_info)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
ARM_COMPUTE_ERROR_ON(num_outputs == 0);
bool has_bias = (bias_accessor != nullptr);
@@ -450,9 +491,9 @@
NodeID GraphBuilder::add_generate_proposals_node(Graph &g, NodeParams params, NodeIdxPair scores, NodeIdxPair deltas, NodeIdxPair anchors, GenerateProposalsInfo info)
{
- CHECK_NODEIDX_PAIR(scores, g);
- CHECK_NODEIDX_PAIR(deltas, g);
- CHECK_NODEIDX_PAIR(anchors, g);
+ check_nodeidx_pair(scores, g);
+ check_nodeidx_pair(deltas, g);
+ check_nodeidx_pair(anchors, g);
NodeID nid = g.add_node<GenerateProposalsLayerNode>(info);
@@ -472,7 +513,7 @@
NodeID GraphBuilder::add_normalize_planar_yuv_node(Graph &g, NodeParams params, NodeIdxPair input,
ITensorAccessorUPtr mean_accessor, ITensorAccessorUPtr std_accessor)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
// Get input tensor descriptor
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
@@ -510,10 +551,10 @@
return create_simple_single_input_output_node<PoolingLayerNode>(g, params, input, pool_info);
}
-NodeID GraphBuilder::add_priorbox_node(Graph &g, NodeParams params, NodeIdxPair input0, NodeIdxPair input1, PriorBoxLayerInfo prior_info)
+NodeID GraphBuilder::add_priorbox_node(Graph &g, NodeParams params, NodeIdxPair input0, NodeIdxPair input1, const PriorBoxLayerInfo &prior_info)
{
- CHECK_NODEIDX_PAIR(input0, g);
- CHECK_NODEIDX_PAIR(input1, g);
+ check_nodeidx_pair(input0, g);
+ check_nodeidx_pair(input1, g);
// Create priorbox node and connect
NodeID prior_nid = g.add_node<PriorBoxLayerNode>(prior_info);
@@ -543,8 +584,8 @@
NodeID GraphBuilder::add_roi_align_node(Graph &g, NodeParams params, NodeIdxPair input, NodeIdxPair rois, ROIPoolingLayerInfo pool_info)
{
- CHECK_NODEIDX_PAIR(input, g);
- CHECK_NODEIDX_PAIR(rois, g);
+ check_nodeidx_pair(input, g);
+ check_nodeidx_pair(rois, g);
NodeID nid = g.add_node<ROIAlignLayerNode>(pool_info);
@@ -557,17 +598,18 @@
NodeID GraphBuilder::add_scale_layer(Graph &g, const NodeParams ¶ms, NodeIdxPair input, ITensorAccessorUPtr mul_accessor, ITensorAccessorUPtr add_accessor)
{
- CHECK_NODEIDX_PAIR(input, g);
+ check_nodeidx_pair(input, g);
// Get input tensor descriptor
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
+ const DataLayout input_data_layout = input_tensor_desc.layout;
// Create mul node
TensorDescriptor mul_desc = input_tensor_desc;
- const size_t C = input_tensor_desc.shape[get_dimension_idx(mul_desc, DataLayoutDimension::CHANNEL)];
- mul_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::WIDTH), 1);
- mul_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::HEIGHT), 1);
- mul_desc.shape.set(get_dimension_idx(input_tensor_desc, DataLayoutDimension::CHANNEL), C);
+ const size_t C = input_tensor_desc.shape[get_dimension_idx(input_data_layout, DataLayoutDimension::CHANNEL)];
+ mul_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::WIDTH), 1);
+ mul_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::HEIGHT), 1);
+ mul_desc.shape.set(get_dimension_idx(input_data_layout, DataLayoutDimension::CHANNEL), C);
NodeID mul_const_nid = add_const_node_with_name(g, params, "Mul", mul_desc, std::move(mul_accessor));
NodeIdxPair mul_const_nidxp = { mul_const_nid, 0 };
@@ -599,6 +641,11 @@
return create_simple_single_input_output_node<SplitLayerNode>(g, params, input, num_splits, axis);
}
+NodeID GraphBuilder::add_stack_node(Graph &g, NodeParams params, const std::vector<NodeIdxPair> &inputs, int axis)
+{
+ return create_simple_multiple_input_single_output_node<StackLayerNode>(g, params, inputs, inputs.size(), axis);
+}
+
NodeID GraphBuilder::add_upsample_node(Graph &g, NodeParams params, NodeIdxPair input, Size2D info, InterpolationPolicy upsampling_policy)
{
return create_simple_single_input_output_node<UpsampleLayerNode>(g, params, input, info, upsampling_policy);
diff --git a/src/graph/GraphManager.cpp b/src/graph/GraphManager.cpp
index 57c5f9d..4f942b9 100644
--- a/src/graph/GraphManager.cpp
+++ b/src/graph/GraphManager.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -45,9 +45,6 @@
void GraphManager::finalize_graph(Graph &graph, GraphContext &ctx, PassManager &pm, Target target)
{
- // Setup graph context if not done manually
- setup_default_graph_context(ctx);
-
// Check if graph has been registered
if(_workloads.find(graph.id()) != std::end(_workloads))
{
@@ -55,7 +52,7 @@
}
// Force target to all graph construct
- // TODO (geopin01) : Support heterogeneous execution
+ // TODO (COMPMID-2014) : Support heterogeneous execution
Target forced_target = target;
if(!is_target_supported(target))
{
@@ -64,6 +61,10 @@
}
force_target_to_graph(graph, forced_target);
+ // Setup backend context
+ // TODO (COMPMID-2014) : Setup all backends needed by the graph
+ setup_requested_backend_context(ctx, forced_target);
+
// Configure all tensors
detail::configure_all_tensors(graph);
diff --git a/src/graph/Tensor.cpp b/src/graph/Tensor.cpp
index 9850128..205ef11 100644
--- a/src/graph/Tensor.cpp
+++ b/src/graph/Tensor.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -108,7 +108,7 @@
_bound_edges.erase(eid);
}
-const std::set<EdgeID> Tensor::bound_edges() const
+std::set<EdgeID> Tensor::bound_edges() const
{
return _bound_edges;
}
diff --git a/src/graph/TypeLoader.cpp b/src/graph/TypeLoader.cpp
index e0ba7e2..b63672b 100644
--- a/src/graph/TypeLoader.cpp
+++ b/src/graph/TypeLoader.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -17,7 +17,7 @@
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWNISE, ARISING FROM,
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
@@ -100,5 +100,55 @@
}
#endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */
}
+
+ConvolutionMethod Convolution_method_from_name(const std::string &name)
+{
+ static const std::map<std::string, ConvolutionMethod> methods =
+ {
+ { "default", ConvolutionMethod::Default },
+ { "direct", ConvolutionMethod::Direct },
+ { "gemm", ConvolutionMethod::GEMM },
+ { "winograd", ConvolutionMethod::Winograd },
+ };
+
+#ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED
+ try
+ {
+#endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */
+ return methods.at(arm_compute::utility::tolower(name));
+
+#ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED
+ }
+ catch(const std::out_of_range &)
+ {
+ throw std::invalid_argument(name);
+ }
+#endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */
+}
+
+DepthwiseConvolutionMethod depthwise_convolution_method_from_name(const std::string &name)
+{
+ static const std::map<std::string, DepthwiseConvolutionMethod> methods =
+ {
+ { "default", DepthwiseConvolutionMethod::Default },
+ { "gemv", DepthwiseConvolutionMethod::GEMV },
+ { "optimized3x3", DepthwiseConvolutionMethod::Optimized3x3 },
+ };
+
+#ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED
+ try
+ {
+#endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */
+ return methods.at(arm_compute::utility::tolower(name));
+
+#ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED
+ }
+ catch(const std::out_of_range &)
+ {
+ throw std::invalid_argument(name);
+ }
+#endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */
+}
+
} // namespace graph
} // namespace arm_compute
diff --git a/src/graph/Utils.cpp b/src/graph/Utils.cpp
index 71ec548..4c34dd8 100644
--- a/src/graph/Utils.cpp
+++ b/src/graph/Utils.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -104,13 +104,14 @@
}
}
-void setup_default_graph_context(GraphContext &ctx)
+void setup_requested_backend_context(GraphContext &ctx, Target target)
{
- for(const auto &backend : backends::BackendRegistry::get().backends())
+ if(backends::BackendRegistry::get().contains(target))
{
- if(backend.second->is_backend_supported())
+ const auto &backend = backends::BackendRegistry::get().find_backend(target);
+ if(backend->is_backend_supported())
{
- backend.second->setup_backend_context(ctx);
+ backend->setup_backend_context(ctx);
}
}
}
@@ -118,12 +119,12 @@
size_t get_dimension_size(const TensorDescriptor &descriptor, const DataLayoutDimension data_layout_dimension)
{
ARM_COMPUTE_ERROR_ON_MSG(descriptor.layout == DataLayout::UNKNOWN, "Cannot retrieve the dimension index for an unknown layout!");
- return descriptor.shape[get_dimension_idx(descriptor, data_layout_dimension)];
+ return descriptor.shape[get_dimension_idx(descriptor.layout, data_layout_dimension)];
}
-size_t get_dimension_idx(const TensorDescriptor &descriptor, const DataLayoutDimension data_layout_dimension)
+size_t get_dimension_idx(DataLayout data_layout, const DataLayoutDimension data_layout_dimension)
{
- ARM_COMPUTE_ERROR_ON_MSG(descriptor.layout == DataLayout::UNKNOWN, "Cannot retrieve the dimension index for an unknown layout!");
+ ARM_COMPUTE_ERROR_ON_MSG(data_layout == DataLayout::UNKNOWN, "Cannot retrieve the dimension index for an unknown layout!");
/* Return the index based on the data layout
* [N C H W]
@@ -133,13 +134,13 @@
switch(data_layout_dimension)
{
case DataLayoutDimension::CHANNEL:
- return (descriptor.layout == DataLayout::NCHW) ? 2 : 0;
+ return (data_layout == DataLayout::NCHW) ? 2 : 0;
break;
case DataLayoutDimension::HEIGHT:
- return (descriptor.layout == DataLayout::NCHW) ? 1 : 2;
+ return (data_layout == DataLayout::NCHW) ? 1 : 2;
break;
case DataLayoutDimension::WIDTH:
- return (descriptor.layout == DataLayout::NCHW) ? 0 : 1;
+ return (data_layout == DataLayout::NCHW) ? 0 : 1;
break;
case DataLayoutDimension::BATCHES:
return 3;
diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp
index ae7f0a5..0666ec0 100644
--- a/src/graph/backends/CL/CLDeviceBackend.cpp
+++ b/src/graph/backends/CL/CLDeviceBackend.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -81,6 +81,11 @@
_tuner.set_tune_new_kernels(enable_tuning);
}
+void CLDeviceBackend::set_kernel_tuning_mode(CLTunerMode tuning_mode)
+{
+ _tuner.set_tuner_mode(tuning_mode);
+}
+
void CLDeviceBackend::initialize_backend()
{
// Setup Scheduler
@@ -118,6 +123,7 @@
}
set_kernel_tuning(ctx.config().use_tuner);
+ set_kernel_tuning_mode(ctx.config().tuner_mode);
// Setup a management backend
if(ctx.memory_management_ctx(Target::CL) == nullptr)
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index b9e3ddc..90c1613 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -40,7 +40,8 @@
/** Target specific information structure used to pass information to the layer templates */
struct CLTargetInfo
{
- using TensorType = arm_compute::ICLTensor;
+ using TensorType = arm_compute::ICLTensor;
+ using TensorConcreteType = CLTensor;
static Target TargetType;
};
@@ -69,6 +70,14 @@
using Subtraction = CLArithmeticSubtraction;
using Multiplication = CLPixelWiseMultiplication;
};
+
+/** Function and tensor types to be used inside a CL fused convolution/batch normalization layer */
+struct CLFusedLayerTypes
+{
+ using ConvolutionLayer = CLConvolutionLayer;
+ using FuseBatchNormalization = CLFuseBatchNormalization;
+};
+
// TODO (isagot01): Remove once we support heterogeneous scheduling at function level
/** Wrapper for the CPP Function in the OpenCL backend **/
class CPPWrapperFunction : public IFunction
@@ -192,6 +201,8 @@
return detail::create_flatten_layer<CLFlattenLayer, CLTargetInfo>(*polymorphic_downcast<FlattenLayerNode *>(node));
case NodeType::FullyConnectedLayer:
return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx);
+ case NodeType::FusedConvolutionBatchNormalizationLayer:
+ return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node));
case NodeType::GenerateProposalsLayer:
return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx);
case NodeType::NormalizationLayer:
@@ -218,6 +229,8 @@
return detail::create_slice_layer<CLSlice, CLTargetInfo>(*polymorphic_downcast<SliceLayerNode *>(node));
case NodeType::SoftmaxLayer:
return detail::create_softmax_layer<CLSoftmaxLayer, CLTargetInfo>(*polymorphic_downcast<SoftmaxLayerNode *>(node), ctx);
+ case NodeType::StackLayer:
+ return detail::create_stack_layer<CLStackLayer, CLTargetInfo>(*polymorphic_downcast<StackLayerNode *>(node));
case NodeType::UpsampleLayer:
return detail::create_upsample_layer<CLUpsampleLayer, CLTargetInfo>(*polymorphic_downcast<UpsampleLayerNode *>(node), ctx);
case NodeType::YOLOLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index 4b71837..cb8dc0a 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -74,6 +74,8 @@
return detail::validate_priorbox_layer<CLPriorBoxLayer>(*polymorphic_downcast<PriorBoxLayerNode *>(node));
case NodeType::ReorgLayer:
return detail::validate_reorg_layer<CLReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
+ case NodeType::ReshapeLayer:
+ return detail::validate_reshape_layer<CLReshapeLayer>(*polymorphic_downcast<ReshapeLayerNode *>(node));
case NodeType::ROIAlignLayer:
return detail::validate_roi_align_layer<CLROIAlignLayer>(*polymorphic_downcast<ROIAlignLayerNode *>(node));
case NodeType::SliceLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index dc987dd..690a311 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -46,7 +46,8 @@
/** Target specific information structure used to pass information to the layer templates */
struct NETargetInfo
{
- using TensorType = arm_compute::ITensor;
+ using TensorType = arm_compute::ITensor;
+ using TensorConcreteType = arm_compute::Tensor;
static Target TargetType;
};
@@ -76,6 +77,13 @@
using Multiplication = NEPixelWiseMultiplication;
};
+/** Function and tensor types to be used inside a NEON fused convolution/batch normalization layer */
+struct NEFusedLayerTypes
+{
+ using ConvolutionLayer = NEConvolutionLayer;
+ using FuseBatchNormalization = NEFuseBatchNormalization;
+};
+
namespace detail
{
// Specialized functions
@@ -135,8 +143,10 @@
<< " Weights QuantInfo: " << weights->info()->quantization_info()
<< " Output QuantInfo: " << output->info()->quantization_info();
}
- ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
- << " Target " << NETargetInfo::TargetType
+ ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+ << node.name()
+ << " Type: " << func_name
+ << " Target: " << NETargetInfo::TargetType
<< " Data Type: " << input->info()->data_type()
<< qss.str()
<< " Input shape: " << input->info()->tensor_shape()
@@ -210,6 +220,8 @@
return detail::create_flatten_layer<NEFlattenLayer, NETargetInfo>(*polymorphic_downcast<FlattenLayerNode *>(node));
case NodeType::FullyConnectedLayer:
return detail::create_fully_connected_layer<NEFullyConnectedLayer, NETargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx);
+ case NodeType::FusedConvolutionBatchNormalizationLayer:
+ return detail::create_fused_convolution_batch_normalization_layer<NEFusedLayerTypes, NETargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node));
case NodeType::NormalizationLayer:
return detail::create_normalization_layer<NENormalizationLayer, NETargetInfo>(*polymorphic_downcast<NormalizationLayerNode *>(node), ctx);
case NodeType::PermuteLayer:
@@ -226,6 +238,8 @@
return detail::create_resize_layer<NEScale, NETargetInfo>(*polymorphic_downcast<ResizeLayerNode *>(node));
case NodeType::SoftmaxLayer:
return detail::create_softmax_layer<NESoftmaxLayer, NETargetInfo>(*polymorphic_downcast<SoftmaxLayerNode *>(node), ctx);
+ case NodeType::StackLayer:
+ return detail::create_stack_layer<NEStackLayer, NETargetInfo>(*polymorphic_downcast<StackLayerNode *>(node));
case NodeType::UpsampleLayer:
return detail::create_upsample_layer<NEUpsampleLayer, NETargetInfo>(*polymorphic_downcast<UpsampleLayerNode *>(node), ctx);
case NodeType::YOLOLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index b0feec5..77f2e7f 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -74,6 +74,8 @@
return detail::validate_priorbox_layer<NEPriorBoxLayer>(*polymorphic_downcast<PriorBoxLayerNode *>(node));
case NodeType::ReorgLayer:
return detail::validate_reorg_layer<NEReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
+ case NodeType::ReshapeLayer:
+ return detail::validate_reshape_layer<NEReshapeLayer>(*polymorphic_downcast<ReshapeLayerNode *>(node));
case NodeType::ROIAlignLayer:
return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : ROIAlignLayer");
case NodeType::SliceLayer:
diff --git a/src/graph/detail/CrossLayerMemoryManagerHelpers.cpp b/src/graph/detail/CrossLayerMemoryManagerHelpers.cpp
index 7fc5ca0..5e31309 100644
--- a/src/graph/detail/CrossLayerMemoryManagerHelpers.cpp
+++ b/src/graph/detail/CrossLayerMemoryManagerHelpers.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -136,7 +136,7 @@
// Then add it to the list of transition buffers
ITensorHandle *tensor_handle = input_edge->tensor()->handle()->parent_handle();
IMemoryGroup *mm_group = get_memory_group_from_handle(ctx, tensor_handle);
- transition_handles.input_handles.push_back(std::make_pair(tensor_handle, mm_group));
+ transition_handles.input_handles.emplace_back(std::make_pair(tensor_handle, mm_group));
}
}
@@ -149,7 +149,7 @@
{
ITensorHandle *tensor_handle = output_tensor->handle()->parent_handle();
IMemoryGroup *mm_group = get_memory_group_from_handle(ctx, tensor_handle);
- transition_handles.output_handles.push_back(std::make_pair(tensor_handle, mm_group));
+ transition_handles.output_handles.emplace_back(std::make_pair(tensor_handle, mm_group));
}
}
diff --git a/src/graph/detail/ExecutionHelpers.cpp b/src/graph/detail/ExecutionHelpers.cpp
index 767154b..900be42 100644
--- a/src/graph/detail/ExecutionHelpers.cpp
+++ b/src/graph/detail/ExecutionHelpers.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -204,10 +204,13 @@
bool call_all_input_node_accessors(ExecutionWorkload &workload)
{
- return !std::any_of(std::begin(workload.inputs), std::end(workload.inputs), [](Tensor * input_tensor)
+ bool is_valid = true;
+ std::for_each(std::begin(workload.inputs), std::end(workload.inputs), [&](Tensor * input_tensor)
{
- return (input_tensor == nullptr) || !input_tensor->call_accessor();
+ bool valid_input = (input_tensor != nullptr) && input_tensor->call_accessor();
+ is_valid = is_valid && valid_input;
});
+ return is_valid;
}
void prepare_all_tasks(ExecutionWorkload &workload)
diff --git a/src/graph/mutators/DepthConcatSubTensorMutator.cpp b/src/graph/mutators/DepthConcatSubTensorMutator.cpp
index a170c4d..7994541 100644
--- a/src/graph/mutators/DepthConcatSubTensorMutator.cpp
+++ b/src/graph/mutators/DepthConcatSubTensorMutator.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -62,18 +62,19 @@
// Get output tensor
auto output_tensor = node->output(0);
- // Check concatenation axis (Sub-tensor optimization is support for concatenation axis >=2)
+ // Check concatenation axis (Sub-tensor optimization is supported for concatenation axis >=2)
auto *concat_node = arm_compute::utils::cast::polymorphic_downcast<ConcatenateLayerNode *>(node);
- if(output_tensor == nullptr || get_dimension_idx(output_tensor->desc(), concat_node->concatenation_axis()) < 2)
+ if(output_tensor == nullptr || get_dimension_idx(output_tensor->desc().layout, concat_node->concatenation_axis()) < 2)
{
continue;
}
- // Check that all tensor have the same target and valid inputs
+ // Check that all tensor have the same target, valid inputs and same quantization info
bool is_valid = std::all_of(node->input_edges().cbegin(), node->input_edges().cend(),
[&](const EdgeID & eid)
{
- return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target);
+ return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target)
+ && (g.edge(eid)->tensor()->desc().quant_info == output_tensor->desc().quant_info);
});
// Create subtensors
diff --git a/src/graph/mutators/GroupedConvolutionMutator.cpp b/src/graph/mutators/GroupedConvolutionMutator.cpp
index d69d2cd..3d53f49 100644
--- a/src/graph/mutators/GroupedConvolutionMutator.cpp
+++ b/src/graph/mutators/GroupedConvolutionMutator.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -47,12 +47,12 @@
// Split input
const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
- const unsigned int input_idx = get_dimension_idx(input_tensor_desc, DataLayoutDimension::CHANNEL);
+ const unsigned int input_idx = get_dimension_idx(input_tensor_desc.layout, DataLayoutDimension::CHANNEL);
NodeID input_split = GraphBuilder::add_split_node(g, params, input, num_groups, input_idx);
// Split weights
const TensorDescriptor weights_tensor_desc = get_tensor_descriptor(g, g.node(weights)->outputs()[0]);
- const unsigned int batch_idx = get_dimension_idx(weights_tensor_desc, DataLayoutDimension::BATCHES);
+ const unsigned int batch_idx = get_dimension_idx(weights_tensor_desc.layout, DataLayoutDimension::BATCHES);
NodeID weights_split = GraphBuilder::add_split_node(g, params, { weights, 0 }, num_groups, batch_idx);
// Split bias
diff --git a/src/graph/mutators/InPlaceOperationMutator.cpp b/src/graph/mutators/InPlaceOperationMutator.cpp
index 31921b3..1c2985d 100644
--- a/src/graph/mutators/InPlaceOperationMutator.cpp
+++ b/src/graph/mutators/InPlaceOperationMutator.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -56,8 +56,8 @@
ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr || new_output_tensor == nullptr);
- // Prevent in-place operation if there is an accessor bound to the in-place tensor
- if(new_output_tensor->accessor() == nullptr)
+ // Prevent in-place operation if there is an accessor bound to the in-place tensor or quantization info are different
+ if(new_output_tensor->accessor() == nullptr || current_output_tensor->desc().quant_info == new_output_tensor->desc().quant_info)
{
ARM_COMPUTE_LOG_GRAPH_VERBOSE("Switching to in-place computation for the node with ID : "
<< node->id() << " and name : " << node->name() << std::endl);
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 9dc02d1..427d7b5 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -23,9 +23,11 @@
*/
#include "arm_compute/graph/mutators/NodeFusionMutator.h"
-#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/GraphBuilder.h"
#include "arm_compute/graph/Logger.h"
#include "arm_compute/graph/Utils.h"
+#include "arm_compute/graph/backends/BackendRegistry.h"
+#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h"
#include "arm_compute/graph/nodes/Nodes.h"
#include "arm_compute/core/utils/misc/Cast.h"
@@ -38,69 +40,156 @@
{
namespace detail
{
+void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge)
+{
+ ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
+
+ auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<ConvolutionLayerNode *>(output_edge->producer());
+ auto *bn_node = arm_compute::utils::cast::polymorphic_downcast<BatchNormalizationLayerNode *>(output_edge->consumer());
+
+ // Not fusing if number of groups is greater than 1
+ if(conv_node->num_groups() > 1)
+ {
+ return;
+ }
+
+ ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing convolution node with ID : " << output_edge->producer_id()
+ << " with BatchNormalization Layer node with ID : " << output_edge->consumer_id() << std::endl);
+
+ // Prevent fusion if fused node has an output accessor
+ if(conv_node->output(0)->accessor() == nullptr)
+ {
+ const Target assigned_target = conv_node->assigned_target();
+
+ // Extract conv inputs
+ const auto conv_input_id = conv_node->input_edge(0)->producer_id();
+ const auto conv_weights_id = conv_node->input_edge(1)->producer_id();
+ const auto out_quant_info = conv_node->output(0)->desc().quant_info;
+ const auto conv_info = conv_node->convolution_info();
+ const auto conv_method = conv_node->convolution_method();
+ const auto num_groups = conv_node->num_groups();
+ const auto act_info = bn_node->fused_activation();
+ FastMathHint fast_math_hint = conv_node->fast_math_hint();
+
+ // Extract bn inputs
+ const auto bn_mean_id = bn_node->input_edge(1)->producer_id();
+ const auto bn_var_id = bn_node->input_edge(2)->producer_id();
+ const auto bn_beta_id = bn_node->input_edge(3)->producer_id();
+ const auto bn_gamma_id = bn_node->input_edge(4)->producer_id();
+ const auto epsilon = bn_node->epsilon();
+
+ // Create the fused node
+ const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint, out_quant_info, act_info);
+
+ if(conv_node->input_edge(2) != nullptr)
+ {
+ auto conv_bias_id = conv_node->input_edge(2)->producer_id();
+ g.add_connection(conv_bias_id, 0, fused_id, 2);
+ }
+
+ // Add connections from the conv/batch_norm inputs to the fused node
+ g.add_connection(conv_input_id, 0, fused_id, 0);
+ g.add_connection(conv_weights_id, 0, fused_id, 1);
+ g.add_connection(bn_mean_id, 0, fused_id, 3);
+ g.add_connection(bn_var_id, 0, fused_id, 4);
+ g.add_connection(bn_beta_id, 0, fused_id, 5);
+ g.add_connection(bn_gamma_id, 0, fused_id, 6);
+
+ auto fused_node = g.node(fused_id);
+ std::vector<NodeIdxPair> bn_driving_nodes = get_driving_nodes(*bn_node);
+
+ // Extract batch normalization node accessor if any
+ auto bn_node_accessor = bn_node->output(0)->extract_accessor();
+ auto bn_node_name = bn_node->name();
+
+ // Remove batch normalization node
+ g.remove_node(bn_node->id());
+
+ // Get driving nodes of batch normalization node
+ for(auto &driving_node : bn_driving_nodes)
+ {
+ g.add_connection(fused_id, 0, driving_node.node_id, driving_node.index);
+ configure_tensor(fused_node->output(0));
+ }
+ // Update fused node outputs
+ fused_node->output(0)->set_accessor(std::move(bn_node_accessor));
+ fused_node->set_assigned_target(assigned_target);
+ fused_node->set_common_node_parameters(NodeParams{ conv_node->name() + "+" + bn_node_name, assigned_target });
+
+ // Remove convolution node
+ g.remove_node(conv_node->id());
+ }
+ else
+ {
+ ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution with batch normalization due to the presence of an output accessor\n");
+ }
+}
+
template <typename N>
-void fuse_node_with_activation(Graph &g,
- const std::set<Activation> &supported_fused_activations,
- std::function<bool(INode &)> const &prec)
+void fuse_node_with_activation(Graph &g, const Edge *output_edge, const std::set<Activation> &supported_fused_activations)
+{
+ ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
+
+ auto *n_node = arm_compute::utils::cast::polymorphic_downcast<N *>(output_edge->producer());
+ auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(output_edge->consumer());
+
+ ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr || n_node->output(0) == nullptr);
+
+ // Check if activation is supported for fusion
+ if(supported_fused_activations.count(act_node->activation_info().activation()) == 0)
+ {
+ return;
+ }
+
+ ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing node with ID : " << output_edge->producer_id()
+ << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl);
+
+ // Prevent fusion if fused node has an output accessor
+ if(n_node->output(0)->accessor() == nullptr)
+ {
+ // Get driving nodes of activation node
+ std::vector<NodeIdxPair> act_driving_nodes = get_driving_nodes(*act_node);
+
+ // Set activation info to fused node
+ n_node->set_fused_activation(act_node->activation_info());
+
+ // Extract activation node accessor if any
+ auto act_node_accessor = act_node->output(0)->extract_accessor();
+
+ // Remove activation node
+ g.remove_node(act_node->id());
+
+ // Update fused node outputs
+ for(auto &driving_node : act_driving_nodes)
+ {
+ g.add_connection(n_node->id(), 0, driving_node.node_id, driving_node.index);
+ }
+
+ // Update accessor to fused node
+ n_node->output(0)->set_accessor(std::move(act_node_accessor));
+ }
+ else
+ {
+ ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of node with activation due to the presence of an output accessor\n");
+ }
+}
+
+template <typename N1, typename N2, typename F, typename... Args>
+void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse_fcn, Args &&... optional_arguments)
{
// Not interested in the order of nodes
for(auto &node : g.nodes())
{
// Check if the node is of type N and not a branching node
- if(node && node->type() == N::node_type && node->output_edges().size() == 1)
+ if(node && node->type() == N1::node_type && node->output_edges().size() == 1)
{
- auto output_edge_id = *node->output_edges().begin();
- auto output_edge = g.edge(output_edge_id);
+ const auto output_edge_id = *node->output_edges().begin();
+ const auto output_edge = g.edge(output_edge_id);
+
// Check if following node is an activation layer node
- if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == NodeType::ActivationLayer))
+ if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == N2::node_type) && prec(*output_edge->producer()))
{
- auto *n_node = arm_compute::utils::cast::polymorphic_downcast<N *>(output_edge->producer());
- auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(output_edge->consumer());
-
- ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr || n_node->output(0) == nullptr);
-
- // Check given precondition
- if(!prec(*n_node))
- {
- continue;
- }
- // Check if activation is supported for fusion
- if(supported_fused_activations.count(act_node->activation_info().activation()) == 0)
- {
- continue;
- }
-
- ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing node with ID : " << output_edge->producer_id()
- << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl);
-
- // Prevent fusion if fused node has an output accessor
- if(n_node->output(0)->accessor() == nullptr)
- {
- // Get driving nodes of activation node
- std::vector<NodeIdxPair> act_driving_nodes = get_driving_nodes(*act_node);
-
- // Set activation info to fused node
- n_node->set_fused_activation(act_node->activation_info());
-
- // Extract activation node accessor if any
- auto act_node_accessor = act_node->output(0)->extract_accessor();
-
- // Remove activation node
- g.remove_node(act_node->id());
-
- // Update fused node outputs
- for(auto &driving_node : act_driving_nodes)
- {
- g.add_connection(n_node->id(), 0, driving_node.node_id, driving_node.index);
- }
-
- // Update accessor to fused node
- n_node->output(0)->set_accessor(std::move(act_node_accessor));
- }
- else
- {
- ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of node with activation due to the presence of an output accessor\n");
- }
+ fuse_fcn(g, output_edge, optional_arguments...);
}
}
}
@@ -118,20 +207,30 @@
const std::set<Activation> supported_fused_activations = { Activation::RELU, Activation::BOUNDED_RELU, Activation::LU_BOUNDED_RELU };
// Preconditions
- auto empty_prec = [](INode & n)
+ auto empty_prec = [](INode &)
{
return true;
};
- auto qs8_prec = [](INode & n)
+ auto qs8_prec = [&g](INode & n)
{
ARM_COMPUTE_ERROR_ON(n.output(0) == nullptr);
- return n.output(0)->desc().data_type == DataType::QASYMM8;
+
+ const auto output_edge_id = *n.output_edges().begin();
+ const auto output_edge = g.edge(output_edge_id);
+ // To perform fusion the two nodes must have same output quantization information
+ const bool same_qinfo = n.output(0)->desc().quant_info == output_edge->producer()->output(0)->desc().quant_info;
+ const bool output_qasymm8 = n.output(0)->desc().data_type == DataType::QASYMM8;
+
+ return (output_qasymm8 && same_qinfo) || !output_qasymm8;
};
// Fusion mutations
- detail::fuse_node_with_activation<BatchNormalizationLayerNode>(g, supported_fused_activations, empty_prec);
- detail::fuse_node_with_activation<ConvolutionLayerNode>(g, supported_fused_activations, empty_prec);
- detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>(g, supported_fused_activations, qs8_prec);
+ detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations);
+ detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations);
+ detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>(g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations);
+
+ // TODO (COMPMID-2055): re-enable once we fuse bias and activations to convolution
+ // detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
}
} // namespace graph
} // namespace arm_compute
diff --git a/src/graph/nodes/ActivationLayerNode.cpp b/src/graph/nodes/ActivationLayerNode.cpp
index 414684c..ada6cf9 100644
--- a/src/graph/nodes/ActivationLayerNode.cpp
+++ b/src/graph/nodes/ActivationLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -30,8 +30,8 @@
{
namespace graph
{
-ActivationLayerNode::ActivationLayerNode(ActivationLayerInfo info)
- : _info(info)
+ActivationLayerNode::ActivationLayerNode(ActivationLayerInfo info, QuantizationInfo out_quant_info)
+ : _info(info), _out_quant_info(out_quant_info)
{
_input_edges.resize(1, EmptyEdgeID);
_outputs.resize(1, NullTensorID);
@@ -62,12 +62,18 @@
const Tensor *src = input(0);
ARM_COMPUTE_ERROR_ON(src == nullptr);
- return src->desc();
+ TensorDescriptor output_info = src->desc();
+ if(!_out_quant_info.empty())
+ {
+ output_info.quant_info = _out_quant_info;
+ }
+
+ return output_info;
}
NodeType ActivationLayerNode::type() const
{
- return NodeType::ActivationLayer;
+ return ActivationLayerNode::node_type;
}
void ActivationLayerNode::accept(INodeVisitor &v)
@@ -75,4 +81,4 @@
v.visit(*this);
}
} // namespace graph
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/graph/nodes/ConcatenateLayerNode.cpp b/src/graph/nodes/ConcatenateLayerNode.cpp
index ade3f6e..5f13b90 100644
--- a/src/graph/nodes/ConcatenateLayerNode.cpp
+++ b/src/graph/nodes/ConcatenateLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -34,8 +34,8 @@
{
namespace graph
{
-ConcatenateLayerNode::ConcatenateLayerNode(unsigned int total_nodes, DataLayoutDimension axis)
- : _total_nodes(total_nodes), _axis(axis), _is_enabled(true)
+ConcatenateLayerNode::ConcatenateLayerNode(unsigned int total_nodes, descriptors::ConcatLayerDescriptor concat_descriptor)
+ : _total_nodes(total_nodes), _concat_descriptor(std::move(concat_descriptor)), _is_enabled(true)
{
_input_edges.resize(_total_nodes, EmptyEdgeID);
_outputs.resize(1, NullTensorID);
@@ -53,7 +53,12 @@
DataLayoutDimension ConcatenateLayerNode::concatenation_axis() const
{
- return _axis;
+ return _concat_descriptor.axis;
+}
+
+QuantizationInfo ConcatenateLayerNode::output_quantization_info() const
+{
+ return _concat_descriptor.output_qinfo;
}
TensorDescriptor ConcatenateLayerNode::compute_output_descriptor(const std::vector<TensorDescriptor> &input_descriptors,
@@ -62,28 +67,18 @@
ARM_COMPUTE_ERROR_ON(input_descriptors.size() == 0);
TensorDescriptor output_descriptor = input_descriptors[0];
- const int axis_idx = get_dimension_idx(output_descriptor, axis);
+ const int axis_idx = get_dimension_idx(output_descriptor.layout, axis);
+ ARM_COMPUTE_ERROR_ON_MSG(axis_idx > 2, "Unsupported concatenation axis!");
// Extract shapes
std::vector<const TensorShape *> shapes;
+ shapes.reserve(input_descriptors.size());
for(auto &input_descriptor : input_descriptors)
{
shapes.emplace_back(&input_descriptor.shape);
}
- // Calculate output shape
- if(axis_idx == 0)
- {
- output_descriptor.shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(shapes);
- }
- else if(axis_idx == 2)
- {
- output_descriptor.shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(shapes);
- }
- else
- {
- ARM_COMPUTE_ERROR("Unsupported concatenation axis!");
- }
+ output_descriptor.shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(shapes, axis_idx);
return output_descriptor;
}
@@ -122,7 +117,11 @@
ARM_COMPUTE_ERROR_ON(t == nullptr);
inputs_descriptors.push_back(t->desc());
}
- output_info = compute_output_descriptor(inputs_descriptors, _axis);
+ output_info = compute_output_descriptor(inputs_descriptors, _concat_descriptor.axis);
+ if(!_concat_descriptor.output_qinfo.empty())
+ {
+ output_info.quant_info = _concat_descriptor.output_qinfo;
+ }
}
return output_info;
@@ -138,4 +137,4 @@
v.visit(*this);
}
} // namespace graph
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/graph/nodes/ConvolutionLayerNode.cpp b/src/graph/nodes/ConvolutionLayerNode.cpp
index 15c7ff6..1c8dcae 100644
--- a/src/graph/nodes/ConvolutionLayerNode.cpp
+++ b/src/graph/nodes/ConvolutionLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -97,10 +97,11 @@
std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), output_width);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), output_height);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
return output_descriptor;
}
diff --git a/src/graph/nodes/DeconvolutionLayerNode.cpp b/src/graph/nodes/DeconvolutionLayerNode.cpp
index e7ccffd..b1a6db7 100644
--- a/src/graph/nodes/DeconvolutionLayerNode.cpp
+++ b/src/graph/nodes/DeconvolutionLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -66,10 +66,11 @@
info.pad().first, info.pad().second,
info.stride().first, info.stride().second);
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), output_width);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), output_height);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
return output_descriptor;
}
diff --git a/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp b/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
index 75ca5f4..cdd9e7b 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -32,8 +32,9 @@
{
namespace graph
{
-DepthwiseConvolutionLayerNode::DepthwiseConvolutionLayerNode(PadStrideInfo info, int depth_multiplier, DepthwiseConvolutionMethod method)
- : _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _fused_activation()
+DepthwiseConvolutionLayerNode::DepthwiseConvolutionLayerNode(PadStrideInfo info, int depth_multiplier, DepthwiseConvolutionMethod method,
+ QuantizationInfo out_quant_info)
+ : _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _out_quant_info(out_quant_info), _fused_activation()
{
_input_edges.resize(3, EmptyEdgeID);
_outputs.resize(1, NullTensorID);
@@ -85,10 +86,11 @@
std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), output_width);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), output_height);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), input_channels * depth_multiplier);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), input_channels * depth_multiplier);
return output_descriptor;
}
@@ -113,7 +115,13 @@
ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
- return compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+ TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+ if(!_out_quant_info.empty())
+ {
+ output_info.quant_info = _out_quant_info;
+ }
+
+ return output_info;
}
NodeType DepthwiseConvolutionLayerNode::type() const
diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
new file mode 100644
index 0000000..c304a6c
--- /dev/null
+++ b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
@@ -0,0 +1,153 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h"
+
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+FusedConvolutionBatchNormalizationNode::FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info,
+ unsigned int num_groups,
+ ConvolutionMethod method,
+ FastMathHint fast_math_hint,
+ QuantizationInfo out_quant_info, ActivationLayerInfo fused_activation)
+ : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(out_quant_info), _fused_activation(fused_activation)
+{
+ _input_edges.resize(7, EmptyEdgeID);
+ _outputs.resize(1, NullTensorID);
+}
+
+void FusedConvolutionBatchNormalizationNode::set_convolution_method(ConvolutionMethod method)
+{
+ _method = method;
+}
+
+float FusedConvolutionBatchNormalizationNode::epsilon() const
+{
+ return _epsilon;
+}
+
+ConvolutionMethod FusedConvolutionBatchNormalizationNode::convolution_method() const
+{
+ return _method;
+}
+
+void FusedConvolutionBatchNormalizationNode::set_fast_math_hint(FastMathHint hint)
+{
+ _fast_math_hint = hint;
+}
+
+FastMathHint FusedConvolutionBatchNormalizationNode::fast_math_hint() const
+{
+ return _fast_math_hint;
+}
+
+PadStrideInfo FusedConvolutionBatchNormalizationNode::convolution_info() const
+{
+ return _info;
+}
+
+unsigned int FusedConvolutionBatchNormalizationNode::num_groups() const
+{
+ return _num_groups;
+}
+
+ActivationLayerInfo FusedConvolutionBatchNormalizationNode::fused_activation() const
+{
+ return _fused_activation;
+}
+
+void FusedConvolutionBatchNormalizationNode::set_fused_activation(ActivationLayerInfo fused_activation)
+{
+ _fused_activation = fused_activation;
+}
+
+TensorDescriptor FusedConvolutionBatchNormalizationNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
+ const TensorDescriptor &weights_descriptor,
+ const PadStrideInfo &info)
+{
+ unsigned int output_width = 0;
+ unsigned int output_height = 0;
+
+ const unsigned int input_width = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH);
+ const unsigned int input_height = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT);
+ const unsigned int kernel_width = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH);
+ const unsigned int kernel_height = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT);
+
+ std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
+
+ const DataLayout data_layout = input_descriptor.layout;
+ TensorDescriptor output_descriptor = input_descriptor;
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
+
+ return output_descriptor;
+}
+
+bool FusedConvolutionBatchNormalizationNode::forward_descriptors()
+{
+ if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (output_id(0) != NullTensorID))
+ {
+ Tensor *dst = output(0);
+ ARM_COMPUTE_ERROR_ON(dst == nullptr);
+ dst->desc() = configure_output(0);
+ return true;
+ }
+ return false;
+}
+
+TensorDescriptor FusedConvolutionBatchNormalizationNode::configure_output(size_t idx) const
+{
+ ARM_COMPUTE_UNUSED(idx);
+ const Tensor *src = input(0);
+ const Tensor *weights = input(1);
+
+ ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
+
+ TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info);
+ if(!_out_quant_info.empty())
+ {
+ output_info.quant_info = _out_quant_info;
+ }
+
+ return output_info;
+}
+
+NodeType FusedConvolutionBatchNormalizationNode::type() const
+{
+ return FusedConvolutionBatchNormalizationNode::node_type;
+}
+
+void FusedConvolutionBatchNormalizationNode::accept(INodeVisitor &v)
+{
+ v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
diff --git a/src/graph/nodes/PoolingLayerNode.cpp b/src/graph/nodes/PoolingLayerNode.cpp
index 26c145a..48b93c9 100644
--- a/src/graph/nodes/PoolingLayerNode.cpp
+++ b/src/graph/nodes/PoolingLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -57,9 +57,10 @@
std::tie(pooled_width, pooled_height) = scaled_dimensions(input_width, input_height, pool_size_x, pool_size_y, info.pad_stride_info());
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), pooled_width);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), pooled_height);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), pooled_width);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), pooled_height);
return output_descriptor;
}
diff --git a/src/graph/nodes/ReorgLayerNode.cpp b/src/graph/nodes/ReorgLayerNode.cpp
index 6b83f6b..21ad451 100644
--- a/src/graph/nodes/ReorgLayerNode.cpp
+++ b/src/graph/nodes/ReorgLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -53,10 +53,11 @@
ARM_COMPUTE_ERROR_ON_MSG((input_width % stride != 0), "The width of the input tensor must be a multiple of stride");
ARM_COMPUTE_ERROR_ON_MSG((input_height % stride != 0), "The height of the input tensor must be a multiple of stride");
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), input_width / stride);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), input_height / stride);
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), input_channel * stride * stride);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), input_width / stride);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), input_height / stride);
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), input_channel * stride * stride);
return output_descriptor;
}
diff --git a/src/graph/nodes/ResizeLayerNode.cpp b/src/graph/nodes/ResizeLayerNode.cpp
index a6aa7bf..a399229 100644
--- a/src/graph/nodes/ResizeLayerNode.cpp
+++ b/src/graph/nodes/ResizeLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -68,9 +68,10 @@
const Tensor *src = input(0);
ARM_COMPUTE_ERROR_ON(src == nullptr);
+ const DataLayout data_layout = src->desc().layout;
TensorDescriptor output_desc = src->desc();
- size_t width_idx = get_dimension_idx(output_desc, DataLayoutDimension::WIDTH);
- size_t height_idx = get_dimension_idx(output_desc, DataLayoutDimension::HEIGHT);
+ size_t width_idx = get_dimension_idx(data_layout, DataLayoutDimension::WIDTH);
+ size_t height_idx = get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT);
output_desc.shape.set(width_idx, static_cast<int>(output_desc.shape[width_idx] * _scale_width));
output_desc.shape.set(height_idx, static_cast<int>(output_desc.shape[height_idx] * _scale_height));
diff --git a/src/graph/nodes/StackLayerNode.cpp b/src/graph/nodes/StackLayerNode.cpp
new file mode 100644
index 0000000..d26498a
--- /dev/null
+++ b/src/graph/nodes/StackLayerNode.cpp
@@ -0,0 +1,115 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/nodes/StackLayerNode.h"
+
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+StackLayerNode::StackLayerNode(unsigned int total_nodes, int axis)
+ : _total_nodes(total_nodes), _axis(axis)
+{
+ _input_edges.resize(_total_nodes, EmptyEdgeID);
+ _outputs.resize(1, NullTensorID);
+}
+
+int StackLayerNode::axis() const
+{
+ return _axis;
+}
+
+TensorDescriptor StackLayerNode::compute_output_descriptor(const std::vector<TensorDescriptor> &input_descriptors,
+ int axis)
+{
+ ARM_COMPUTE_ERROR_ON(input_descriptors.size() == 0);
+
+ TensorDescriptor output_descriptor = input_descriptors[0];
+
+ const TensorInfo input_info(input_descriptors[0].shape, 1, input_descriptors[0].data_type);
+ const unsigned int num_tensors = input_descriptors.size();
+
+ output_descriptor.shape = arm_compute::misc::shape_calculator::compute_stack_shape(input_info, axis, num_tensors);
+
+ return output_descriptor;
+}
+
+bool StackLayerNode::forward_descriptors()
+{
+ if(_outputs[0] != NullTensorID)
+ {
+ Tensor *dst = output(0);
+ ARM_COMPUTE_ERROR_ON(dst == nullptr);
+ dst->desc() = configure_output(0);
+ return true;
+ }
+ return false;
+}
+
+TensorDescriptor StackLayerNode::configure_output(size_t idx) const
+{
+ ARM_COMPUTE_UNUSED(idx);
+ ARM_COMPUTE_ERROR_ON(idx >= _outputs.size());
+
+ // Check if all input tensors are set
+ bool are_all_inputs_set = std::all_of(std::begin(_input_edges), std::end(_input_edges), [](const EdgeID & eid)
+ {
+ return eid != EmptyEdgeID;
+ });
+
+ TensorDescriptor output_info = {};
+
+ if(are_all_inputs_set)
+ {
+ std::vector<TensorDescriptor> inputs_descriptors;
+ for(unsigned int i = 0; i < _input_edges.size(); ++i)
+ {
+ const Tensor *t = _graph->tensor(input_id(i));
+ ARM_COMPUTE_ERROR_ON(t == nullptr);
+ inputs_descriptors.push_back(t->desc());
+ }
+ output_info = compute_output_descriptor(inputs_descriptors, _axis);
+ }
+
+ return output_info;
+}
+
+NodeType StackLayerNode::type() const
+{
+ return NodeType::StackLayer;
+}
+
+void StackLayerNode::accept(INodeVisitor &v)
+{
+ v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
diff --git a/src/graph/nodes/UpsampleLayerNode.cpp b/src/graph/nodes/UpsampleLayerNode.cpp
index bdd39e8..88af122 100644
--- a/src/graph/nodes/UpsampleLayerNode.cpp
+++ b/src/graph/nodes/UpsampleLayerNode.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -54,9 +54,10 @@
const unsigned int input_width = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH);
const unsigned int input_height = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT);
+ const DataLayout data_layout = input_descriptor.layout;
TensorDescriptor output_descriptor = input_descriptor;
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), input_width * info.x());
- output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), input_height * info.y());
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), input_width * info.x());
+ output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), input_height * info.y());
return output_descriptor;
}
diff --git a/src/graph/printers/DotGraphPrinter.cpp b/src/graph/printers/DotGraphPrinter.cpp
index ef156ea..c939de1 100644
--- a/src/graph/printers/DotGraphPrinter.cpp
+++ b/src/graph/printers/DotGraphPrinter.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -77,6 +77,14 @@
_info = ss.str();
}
+void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n)
+{
+ ARM_COMPUTE_UNUSED(n);
+ std::stringstream ss;
+ ss << "FusedConvolutionBatchNormalizationNode";
+ _info = ss.str();
+}
+
void DotGraphVisitor::visit(NormalizationLayerNode &n)
{
std::stringstream ss;