diff options
author | David Beck <david.beck@arm.com> | 2018-09-24 15:59:27 +0100 |
---|---|---|
committer | Matthew Bentham <matthew.bentham@arm.com> | 2018-10-10 16:16:57 +0100 |
commit | 0dbe0ee25312b728d77383d11c465156e64ae757 (patch) | |
tree | af37a9802e3ad551e1bf63f7636508cde7a41643 /src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp | |
parent | b4540bef0b0327683fe8e63f727c1212800dc2a9 (diff) | |
download | armnn-0dbe0ee25312b728d77383d11c465156e64ae757.tar.gz |
IVGCVSW-1899 : Neon backend folder structure
armnn:149855
Change-Id: I26e8cf83422a65049386a5ebdb6d0001627aefaa
Diffstat (limited to 'src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp | 93 |
1 files changed, 93 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp b/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp new file mode 100644 index 0000000000..742a768b94 --- /dev/null +++ b/src/backends/neon/workloads/NeonDepthwiseConvolutionFloatWorkload.cpp @@ -0,0 +1,93 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "NeonDepthwiseConvolutionFloatWorkload.hpp" +#include <backends/neon/NeonLayerSupport.hpp> +#include <backends/CpuTensorHandle.hpp> +#include <backends/aclCommon/ArmComputeTensorUtils.hpp> + +namespace armnn +{ +using namespace armcomputetensorutils; + +NeonDepthwiseConvolutionFloatWorkload::NeonDepthwiseConvolutionFloatWorkload( + const DepthwiseConvolution2dQueueDescriptor& descriptor, + const WorkloadInfo& info) + : FloatWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info) +{ + const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo(); + + m_KernelTensor = std::make_unique<arm_compute::Tensor>(); + BuildArmComputeTensor(*m_KernelTensor, weightInfo, descriptor.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(), descriptor.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); + + m_Data.ValidateInputsOutputs("NeonDepthwiseConvolutionFloatWorkload", 1, 1); + + arm_compute::ITensor& input = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); + + bool use3x3Optimisation = weightInfo.GetShape()[3] == 3 && weightInfo.GetShape()[2] == 3; + if (use3x3Optimisation) + { + m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>(); + static_cast<arm_compute::NEDepthwiseConvolutionLayer3x3*>( + m_pDepthwiseConvolutionLayer.get())->configure(&input, + m_KernelTensor.get(), + m_BiasTensor.get(), + &output, + padStrideInfo); + } + else + { + 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); + } + + BOOST_ASSERT(m_pDepthwiseConvolutionLayer); + + InitializeArmComputeTensorDataForFloatTypes(*m_KernelTensor, m_Data.m_Weight); + + if (m_BiasTensor) + { + InitializeArmComputeTensorDataForFloatTypes(*m_BiasTensor, m_Data.m_Bias); + } + + m_pDepthwiseConvolutionLayer->prepare(); + FreeUnusedTensors(); +} + +void NeonDepthwiseConvolutionFloatWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDepthwiseConvolutionFloatWorkload_Execute"); + BOOST_ASSERT(m_pDepthwiseConvolutionLayer); + + m_pDepthwiseConvolutionLayer->run(); +} + +void NeonDepthwiseConvolutionFloatWorkload::FreeUnusedTensors() +{ + FreeTensorIfUnused(m_KernelTensor); + FreeTensorIfUnused(m_BiasTensor); +} + +} //namespace armnn |