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/NeonFullyConnectedWorkload.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/NeonFullyConnectedWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp | 16 |
1 files changed, 10 insertions, 6 deletions
diff --git a/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp b/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp index e808c60c0c..31489a0c32 100644 --- a/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp +++ b/src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp @@ -6,9 +6,12 @@ #include "NeonFullyConnectedWorkload.hpp" #include "NeonWorkloadUtils.hpp" + #include <aclCommon/ArmComputeTensorUtils.hpp> #include <aclCommon/ArmComputeUtils.hpp> + #include <armnn/utility/PolymorphicDowncast.hpp> + #include <backendsCommon/CpuTensorHandle.hpp> #include <arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h> @@ -21,7 +24,8 @@ arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input, const TensorInfo& output, const TensorInfo& weights, const TensorInfo& biases, - const FullyConnectedDescriptor& descriptor) + const FullyConnectedDescriptor& descriptor, + const ActivationDescriptor* activationDescriptor) { const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input); const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output); @@ -36,8 +40,7 @@ arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input, } const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo = - ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor); - + ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor, activationDescriptor); return arm_compute::NEFullyConnectedLayer::validate(&aclInput, &aclWeights, @@ -64,9 +67,10 @@ NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueue BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo()); } - // Construct - arm_compute::FullyConnectedLayerInfo fc_info; - fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix; + const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); + + arm_compute::FullyConnectedLayerInfo fc_info = + ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, activationInfo); auto layer = std::make_unique<arm_compute::NEFullyConnectedLayer>(memoryManager); layer->configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info); |