diff options
| author | Mike Kelly <mike.kelly@arm.com> | 2020-11-12 10:58:48 +0000 |
|---|---|---|
| committer | Jim Flynn <jim.flynn@arm.com> | 2020-11-13 14:25:30 +0000 |
| commit | 07810fc2fcdd34db74222d90cc73ef12a88e7b78 (patch) | |
| tree | 8becef8453674822d079815b06ae37310b97d2cf /src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp | |
| parent | 8502adeafbbb1db0acefa62560d93453e38dcadb (diff) | |
| download | armnn-07810fc2fcdd34db74222d90cc73ef12a88e7b78.tar.gz | |
IVGCVSW-5328-5329 Fuse Activation
* Added Fused Activation Optimization to both CL and Neon backends.
* Added Fused Activation support to all the CL and Neon workloads
that support it.
* Changed ProfilingTest network to be a Convolution layer
followed by an Abs layer rather than an Activation layer.
* Added IBackendInternal::OptimizeSubgraphView function that can accept a
ModelOptions.
* Network will now call OptimizeSubgraphView passing in the ModelOptions.
Signed-off-by: Keith Davis <keith.davis@arm.com>
Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ib536ac3cbafc7d9b35c139ad9a65b7735262cd9d
Diffstat (limited to 'src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp')
| -rw-r--r-- | src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp | 25 |
1 files changed, 16 insertions, 9 deletions
diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp index a9a3c75bfd..db6bcc3ecb 100644 --- a/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp +++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionWorkload.cpp @@ -10,6 +10,7 @@ #include <armnnUtils/DataLayoutIndexed.hpp> #include <aclCommon/ArmComputeTensorUtils.hpp> +#include <aclCommon/ArmComputeUtils.hpp> #include <neon/NeonLayerSupport.hpp> @@ -29,7 +30,8 @@ arm_compute::Status NeonDepthwiseConvolutionWorkloadValidate(const TensorInfo& i const TensorInfo& output, const DepthwiseConvolution2dDescriptor& descriptor, const TensorInfo& weights, - const Optional<TensorInfo>& biases) + const Optional<TensorInfo>& biases, + const ActivationDescriptor* activationDescriptor) { const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); @@ -59,13 +61,16 @@ arm_compute::Status NeonDepthwiseConvolutionWorkloadValidate(const TensorInfo& i const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D( descriptor.m_DilationX,descriptor.m_DilationY); + const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo( + activationDescriptor); + return arm_compute::NEDepthwiseConvolutionLayer::validate(&aclInputInfo, &aclWeightsInfo, optionalAclBiasesInfo, &aclOutputInfo, aclPadStrideInfo, aclDepthMultiplier, - arm_compute::ActivationLayerInfo(), + activationInfo, aclDilationInfo); } @@ -116,16 +121,18 @@ NeonDepthwiseConvolutionWorkload::NeonDepthwiseConvolutionWorkload( arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters); + const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); + m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>(); static_cast<arm_compute::NEDepthwiseConvolutionLayer*>( m_pDepthwiseConvolutionLayer.get())->configure(&input, - m_KernelTensor.get(), - m_BiasTensor.get(), - &output, - padStrideInfo, - depthMultiplier, - arm_compute::ActivationLayerInfo(), - aclDilationInfo); + m_KernelTensor.get(), + m_BiasTensor.get(), + &output, + padStrideInfo, + depthMultiplier, + activationInfo, + aclDilationInfo); ARMNN_ASSERT(m_pDepthwiseConvolutionLayer); |