commit acce504ec4aebe5e5da470c1cfc3cee401ff11f3
author giuros01 <giuseppe.rossini@arm.com> Thu Feb 21 17:32:34 2019 +0000
committer Giuseppe Rossini <giuseppe.rossini@arm.com> Wed Mar 13 10:31:18 2019 +0000
tree bff9107fe7facf4be68140380192ee1ea049d05d
parent ba5e096b8b2a9f777695844746ec3ff1ef90ade8

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>

src/graph/backends/CL/CLFunctionsFactory.cpp

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: