diff options
author | Nattapat Chaimanowong <nattapat.chaimanowong@arm.com> | 2018-10-15 15:07:34 +0100 |
---|---|---|
committer | Matthew Bentham <matthew.bentham@arm.com> | 2018-10-22 16:57:54 +0100 |
commit | 974b65fea888cc000a5164d5b56d9ed016391151 (patch) | |
tree | 406f1793d434e1b3ecdae997916361b18c6c6c69 /src/backends/neon/workloads/NeonConvolution2dWorkload.cpp | |
parent | 39fedf04d226360a8c77ca1ca3e2528a709101b5 (diff) | |
download | armnn-974b65fea888cc000a5164d5b56d9ed016391151.tar.gz |
IVGCVSW-1951 Remove type templating from NeonConvolution2dWorkload
Change-Id: Id3d8137d60b14e93863ad5b6db752a7f6ef9b7fb
Diffstat (limited to 'src/backends/neon/workloads/NeonConvolution2dWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonConvolution2dWorkload.cpp | 134 |
1 files changed, 134 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp b/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp new file mode 100644 index 0000000000..c26cdea92b --- /dev/null +++ b/src/backends/neon/workloads/NeonConvolution2dWorkload.cpp @@ -0,0 +1,134 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "NeonConvolution2dWorkload.hpp" + +#include <backends/CpuTensorHandle.hpp> +#include <backends/aclCommon/ArmComputeTensorUtils.hpp> +#include <backends/neon/NeonLayerSupport.hpp> + +#include <armnn/Types.hpp> +#include <armnnUtils/Half.hpp> + +namespace armnn +{ + +using namespace armcomputetensorutils; + +arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const Optional<TensorInfo>& biases) +{ + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); + const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); + const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); + + arm_compute::TensorInfo aclBiasesInfo; + arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; + + if (descriptor.m_BiasEnabled) + { + BOOST_ASSERT(biases.has_value()); + + aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); + optionalAclBiasesInfo = &aclBiasesInfo; + } + + arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor); + + return arm_compute::NEConvolutionLayer::validate(&aclInputInfo, + &aclWeightsInfo, + optionalAclBiasesInfo, + &aclOutputInfo, + layerInfo); +} + +NeonConvolution2dWorkload::NeonConvolution2dWorkload( + const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) + : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info) +{ + using arm_compute::NEDirectConvolutionLayer; + + m_Data.ValidateInputsOutputs("NeonConvolution2dWorkload", 1, 1); + + // todo: check tensor shapes match. + + arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); + + arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); + input.info()->set_data_layout(aclDataLayout); + output.info()->set_data_layout(aclDataLayout); + + m_KernelTensor = std::make_unique<arm_compute::Tensor>(); + BuildArmComputeTensor(*m_KernelTensor, m_Data.m_Weight->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); + + if (m_Data.m_Parameters.m_BiasEnabled) + { + m_BiasTensor = std::make_unique<arm_compute::Tensor>(); + BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); + } + + arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX, + m_Data.m_Parameters.m_StrideY, + m_Data.m_Parameters.m_PadLeft, + m_Data.m_Parameters.m_PadRight, + m_Data.m_Parameters.m_PadTop, + m_Data.m_Parameters.m_PadBottom, + arm_compute::DimensionRoundingType::FLOOR); + + const bool preferDirectConvolution = + IsNeonDirectConvolutionPreferred(m_Data.m_Weight->GetTensorInfo(), + m_Data.m_Parameters); + + if (preferDirectConvolution) + { + auto directConvolutionLayer = std::make_unique<arm_compute::NEDirectConvolutionLayer>(memoryManager); + directConvolutionLayer->configure(&input, + m_KernelTensor.get(), + m_BiasTensor.get(), + &output, + padStrideInfo); + m_ConvolutionLayer.reset(directConvolutionLayer.release()); + } + else + { + auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager); + convolutionLayer->configure(&input, + m_KernelTensor.get(), + m_BiasTensor.get(), + &output, + padStrideInfo); + m_ConvolutionLayer.reset(convolutionLayer.release()); + } + BOOST_ASSERT(m_ConvolutionLayer); + + InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight); + + if (m_Data.m_Parameters.m_BiasEnabled) + { + InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias); + } + + m_ConvolutionLayer->prepare(); + FreeUnusedTensors(); +} + +void NeonConvolution2dWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonConvolution2dWorkload_Execute"); + m_ConvolutionLayer->run(); +} + +void NeonConvolution2dWorkload::FreeUnusedTensors() +{ + FreeTensorIfUnused(m_KernelTensor); + FreeTensorIfUnused(m_BiasTensor); +} + +} //namespace armnn |