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/ClMultiplicationWorkload.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/ClMultiplicationWorkload.cpp')
-rw-r--r-- | src/backends/cl/workloads/ClMultiplicationWorkload.cpp | 20 |
1 files changed, 16 insertions, 4 deletions
diff --git a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp b/src/backends/cl/workloads/ClMultiplicationWorkload.cpp index e9b75c3f10..46a1c4bc59 100644 --- a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp +++ b/src/backends/cl/workloads/ClMultiplicationWorkload.cpp @@ -4,8 +4,12 @@ // #include "ClMultiplicationWorkload.hpp" -#include <cl/ClTensorHandle.hpp> + +#include <aclCommon/ArmComputeUtils.hpp> #include <backendsCommon/CpuTensorHandle.hpp> + +#include <cl/ClTensorHandle.hpp> + #include "ClWorkloadUtils.hpp" namespace armnn @@ -13,7 +17,8 @@ namespace armnn arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, const TensorInfo& input1, - const TensorInfo& output) + const TensorInfo& output, + const ActivationDescriptor* activationDescriptor) { const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input0); const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1); @@ -23,6 +28,9 @@ arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, arm_compute::ConvertPolicy::SATURATE : arm_compute::ConvertPolicy::WRAP; + const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo( + activationDescriptor); + // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it, // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be // ignored for F32 tensors. @@ -31,7 +39,8 @@ arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, &aclOutput, 1.0f, convertPolicy, - arm_compute::RoundingPolicy::TO_ZERO); + arm_compute::RoundingPolicy::TO_ZERO, + activationInfo); } @@ -50,13 +59,16 @@ ClMultiplicationWorkload::ClMultiplicationWorkload(const MultiplicationQueueDesc arm_compute::ConvertPolicy::SATURATE : arm_compute::ConvertPolicy::WRAP; + const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); + // Construct m_PixelWiseMultiplication.configure(&input0, &input1, &output, 1.0f, convertPolicy, - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); + arm_compute::RoundingPolicy::TO_NEAREST_EVEN, + activationInfo); } void ClMultiplicationWorkload::Execute() const |