COMPMID-1740: Fuse batch normalization with Convolution Layer at graph level
Change-Id: I77ca51c2c72783cc26a099a6a9c3210cdbbe822d
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/797
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index b9e3ddc..7473ff4 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:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index dc987dd..f23845c 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
@@ -210,6 +218,8 @@
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: