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/cl/workloads/ClBatchNormalizationFloatWorkload.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/cl/workloads/ClBatchNormalizationFloatWorkload.cpp')
-rw-r--r-- | src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp | 22 |
1 files changed, 17 insertions, 5 deletions
diff --git a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp b/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp index fa0be85100..68942e2a01 100644 --- a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp +++ b/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp @@ -4,12 +4,16 @@ // #include "ClBatchNormalizationFloatWorkload.hpp" +#include "ClWorkloadUtils.hpp" + #include <cl/ClTensorHandle.hpp> + #include <backendsCommon/CpuTensorHandle.hpp> + #include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/ClLayerSupport.hpp> +#include <aclCommon/ArmComputeUtils.hpp> -#include "ClWorkloadUtils.hpp" +#include <cl/ClLayerSupport.hpp> namespace armnn { @@ -21,7 +25,8 @@ arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input, const TensorInfo& var, const TensorInfo& beta, const TensorInfo& gamma, - const BatchNormalizationDescriptor &desc) + const BatchNormalizationDescriptor& desc, + const ActivationDescriptor* activationDescriptor) { const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input, desc.m_DataLayout); @@ -36,13 +41,17 @@ arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input, const arm_compute::TensorInfo aclGammaInfo = armcomputetensorutils::BuildArmComputeTensorInfo(gamma, desc.m_DataLayout); + const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo( + activationDescriptor); + return arm_compute::CLBatchNormalizationLayer::validate(&aclInputInfo, &aclOutputInfo, &aclMeanInfo, &aclVarInfo, &aclBetaInfo, &aclGammaInfo, - desc.m_Eps); + desc.m_Eps, + activationInfo); } ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload( @@ -70,13 +79,16 @@ ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload( input.info()->set_data_layout(aclDataLayout); output.info()->set_data_layout(aclDataLayout); + const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); + m_Layer.configure(&input, &output, m_Mean.get(), m_Variance.get(), m_Beta.get(), m_Gamma.get(), - m_Data.m_Parameters.m_Eps); + m_Data.m_Parameters.m_Eps, + activationInfo); InitializeArmComputeClTensorData(*m_Mean, m_Data.m_Mean); InitializeArmComputeClTensorData(*m_Variance, m_Data.m_Variance); |