arm_compute v19.05
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: