src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "ClBatchNormalizationFloatWorkload.hpp"
   7 #include <cl/ClTensorHandle.hpp>
   8 #include <backendsCommon/CpuTensorHandle.hpp>
   9 #include <aclCommon/ArmComputeTensorUtils.hpp>
  10 #include <cl/ClLayerSupport.hpp>
  11
  12 #include "ClWorkloadUtils.hpp"
  13
  14 namespace armnn
  15 {
  16 using namespace armcomputetensorutils;
  17
  18 arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input,
  19                                                  const TensorInfo& output,
  20                                                  const TensorInfo& mean,
  21                                                  const TensorInfo& var,
  22                                                  const TensorInfo& beta,
  23                                                  const TensorInfo& gamma,
  24                                                  const BatchNormalizationDescriptor &desc)
  25 {
  26     const DataLayout dataLayout = desc.m_DataLayout.GetDataLayout();
  27
  28     const arm_compute::TensorInfo aclInputInfo =
  29           armcomputetensorutils::BuildArmComputeTensorInfo(input, dataLayout);
  30     const arm_compute::TensorInfo aclOutputInfo =
  31           armcomputetensorutils::BuildArmComputeTensorInfo(output, dataLayout);
  32     const arm_compute::TensorInfo aclMeanInfo =
  33           armcomputetensorutils::BuildArmComputeTensorInfo(mean, dataLayout);
  34     const arm_compute::TensorInfo aclVarInfo =
  35           armcomputetensorutils::BuildArmComputeTensorInfo(var, dataLayout);
  36     const arm_compute::TensorInfo aclBetaInfo =
  37           armcomputetensorutils::BuildArmComputeTensorInfo(beta, dataLayout);
  38     const arm_compute::TensorInfo aclGammaInfo =
  39           armcomputetensorutils::BuildArmComputeTensorInfo(gamma, dataLayout);
  40
  41     return arm_compute::CLBatchNormalizationLayer::validate(&aclInputInfo,
  42                                                             &aclOutputInfo,
  43                                                             &aclMeanInfo,
  44                                                             &aclVarInfo,
  45                                                             &aclBetaInfo,
  46                                                             &aclGammaInfo,
  47                                                             desc.m_Eps);
  48 }
  49
  50 ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload(
  51     const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info)
  52     : FloatWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
  53 {
  54     m_Mean = std::make_unique<arm_compute::CLTensor>();
  55     BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
  56
  57     m_Variance = std::make_unique<arm_compute::CLTensor>();
  58     BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
  59
  60     m_Gamma = std::make_unique<arm_compute::CLTensor>();
  61     BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
  62
  63     m_Beta = std::make_unique<arm_compute::CLTensor>();
  64     BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
  65
  66     m_Data.ValidateInputsOutputs("ClBatchNormalizationFloatWorkload", 1, 1);
  67
  68     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
  69     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
  70
  71     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout.GetDataLayout());
  72     input.info()->set_data_layout(aclDataLayout);
  73     output.info()->set_data_layout(aclDataLayout);
  74
  75     m_Layer.configure(&input,
  76                       &output,
  77                       m_Mean.get(),
  78                       m_Variance.get(),
  79                       m_Beta.get(),
  80                       m_Gamma.get(),
  81                       m_Data.m_Parameters.m_Eps);
  82
  83     InitializeArmComputeClTensorData(*m_Mean, m_Data.m_Mean);
  84     InitializeArmComputeClTensorData(*m_Variance, m_Data.m_Variance);
  85     InitializeArmComputeClTensorData(*m_Beta, m_Data.m_Beta);
  86     InitializeArmComputeClTensorData(*m_Gamma, m_Data.m_Gamma);
  87
  88     // Force Compute Library to perform the necessary copying and reshaping, after which
  89     // delete all the input tensors that will no longer be needed
  90     m_Layer.prepare();
  91     FreeUnusedTensors();
  92 }
  93
  94 void ClBatchNormalizationFloatWorkload::Execute() const
  95 {
  96     ARMNN_SCOPED_PROFILING_EVENT_CL("ClBatchNormalizationFloatWorkload_Execute");
  97     RunClFunction(m_Layer, CHECK_LOCATION());
  98 }
  99
 100 void ClBatchNormalizationFloatWorkload::FreeUnusedTensors()
 101 {
 102     FreeTensorIfUnused(m_Mean);
 103     FreeTensorIfUnused(m_Variance);
 104     FreeTensorIfUnused(m_Gamma);
 105     FreeTensorIfUnused(m_Beta);
 106 }
 107
 108 } //namespace armnn