src/backends/cl/workloads/ClConvolution2dWorkload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "ClConvolution2dWorkload.hpp"
   7
   8 #include "ClWorkloadUtils.hpp"
   9
  10 #include <cl/ClLayerSupport.hpp>
  11 #include <cl/ClTensorHandle.hpp>
  12 #include <cl/ClLayerSupport.hpp>
  13 #include <aclCommon/ArmComputeUtils.hpp>
  14 #include <aclCommon/ArmComputeTensorUtils.hpp>
  15 #include <backendsCommon/CpuTensorHandle.hpp>
  16
  17 #include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h>
  18
  19 namespace armnn
  20 {
  21 using namespace armcomputetensorutils;
  22
  23 arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
  24                                                     const TensorInfo& output,
  25                                                     const Convolution2dDescriptor& descriptor,
  26                                                     const TensorInfo& weights,
  27                                                     const Optional<TensorInfo>& biases)
  28 {
  29     const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
  30     const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
  31     const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
  32
  33     arm_compute::TensorInfo aclBiasesInfo;
  34     arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
  35
  36     if (descriptor.m_BiasEnabled)
  37     {
  38         BOOST_ASSERT(biases.has_value());
  39
  40         aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
  41         optionalAclBiasesInfo = &aclBiasesInfo;
  42     }
  43
  44     arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
  45
  46     return arm_compute::CLConvolutionLayer::validate(&aclInputInfo,
  47                                                      &aclWeightsInfo,
  48                                                      optionalAclBiasesInfo,
  49                                                      &aclOutputInfo,
  50                                                      layerInfo);
  51 }
  52
  53 ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor,
  54     const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
  55     : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
  56     , m_ConvolutionLayer(memoryManager)
  57 {
  58     // todo: check tensor shapes match.
  59     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
  60
  61     m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
  62     BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
  63
  64     arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
  65                                              m_Data.m_Parameters.m_StrideY,
  66                                              m_Data.m_Parameters.m_PadLeft,
  67                                              m_Data.m_Parameters.m_PadRight,
  68                                              m_Data.m_Parameters.m_PadTop,
  69                                              m_Data.m_Parameters.m_PadBottom,
  70                                              arm_compute::DimensionRoundingType::FLOOR);
  71
  72     if (m_Data.m_Parameters.m_BiasEnabled)
  73     {
  74         m_BiasTensor = std::make_unique<arm_compute::CLTensor>();
  75         BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
  76     }
  77
  78     m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1);
  79
  80     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
  81     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
  82
  83     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
  84     input.info()->set_data_layout(aclDataLayout);
  85     output.info()->set_data_layout(aclDataLayout);
  86
  87     m_ConvolutionLayer.configure(&input,
  88                                  m_KernelTensor.get(),
  89                                  m_BiasTensor.get(),
  90                                  &output,
  91                                  padStrideInfo);
  92
  93     InitializeArmComputeClTensorData(*m_KernelTensor, m_Data.m_Weight);
  94
  95     if (m_BiasTensor)
  96     {
  97         InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias);
  98     }
  99
 100     // Force Compute Library to perform the necessary copying and reshaping, after which
 101     // delete all the input tensors that will no longer be needed
 102     m_ConvolutionLayer.prepare();
 103     FreeUnusedTensors();
 104 }
 105
 106 void ClConvolution2dWorkload::Execute() const
 107 {
 108     ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dWorkload_Execute");
 109     RunClFunction(m_ConvolutionLayer, CHECK_LOCATION());
 110 }
 111
 112 void ClConvolution2dWorkload::FreeUnusedTensors()
 113 {
 114     FreeTensorIfUnused(m_KernelTensor);
 115     FreeTensorIfUnused(m_BiasTensor);
 116 }
 117
 118 } //namespace armnn