src/backends/cl/workloads/ClFullyConnectedWorkload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "ClFullyConnectedWorkload.hpp"
   7 #include <cl/ClTensorHandle.hpp>
   8 #include <backendsCommon/CpuTensorHandle.hpp>
   9 #include <aclCommon/ArmComputeTensorUtils.hpp>
  10 #include <aclCommon/ArmComputeUtils.hpp>
  11 #include <cl/ClLayerSupport.hpp>
  12
  13 #include "ClWorkloadUtils.hpp"
  14
  15 namespace armnn
  16 {
  17 using namespace armcomputetensorutils;
  18
  19 arm_compute::Status ClFullyConnectedWorkloadValidate(const TensorInfo& input,
  20                                                      const TensorInfo& output,
  21                                                      const TensorInfo& weights,
  22                                                      const TensorInfo& biases,
  23                                                      const FullyConnectedDescriptor& descriptor)
  24 {
  25     const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
  26     const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
  27     const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
  28
  29     arm_compute::TensorInfo aclBiases;
  30     arm_compute::TensorInfo *optionalAclBiases = nullptr;
  31     if (descriptor.m_BiasEnabled)
  32     {
  33         aclBiases  = BuildArmComputeTensorInfo(biases);
  34         optionalAclBiases = &aclBiases;
  35     }
  36
  37     const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
  38         ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor);
  39
  40     return arm_compute::CLFullyConnectedLayer::validate(&aclInput,
  41                                                         &aclWeights,
  42                                                         optionalAclBiases,
  43                                                         &aclOutput,
  44                                                         fullyConnectedLayerInfo);
  45 }
  46
  47 ClFullyConnectedWorkload::ClFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
  48     const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
  49     : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
  50     , m_FullyConnectedLayer(memoryManager)
  51 {
  52     m_WeightsTensor = std::make_unique<arm_compute::CLTensor>();
  53     BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
  54
  55     if (m_Data.m_Parameters.m_BiasEnabled)
  56     {
  57         m_BiasesTensor = std::make_unique<arm_compute::CLTensor>();
  58         BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
  59     }
  60
  61     m_Data.ValidateInputsOutputs("ClFullyConnectedWorkload", 1, 1);
  62
  63     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
  64     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
  65
  66     // Construct
  67     arm_compute::FullyConnectedLayerInfo fc_info;
  68     fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix;
  69     m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
  70
  71     InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight);
  72
  73     if (m_BiasesTensor)
  74     {
  75         InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias);
  76     }
  77
  78     // Force Compute Library to perform the necessary copying and reshaping, after which
  79     // delete all the input tensors that will no longer be needed
  80     m_FullyConnectedLayer.prepare();
  81     FreeUnusedTensors();
  82 }
  83
  84 void ClFullyConnectedWorkload::Execute() const
  85 {
  86     ARMNN_SCOPED_PROFILING_EVENT_CL("ClFullyConnectedWorkload_Execute");
  87     RunClFunction(m_FullyConnectedLayer, CHECK_LOCATION());
  88 }
  89
  90 void ClFullyConnectedWorkload::FreeUnusedTensors()
  91 {
  92     FreeTensorIfUnused(m_WeightsTensor);
  93     FreeTensorIfUnused(m_BiasesTensor);
  94 }
  95
  96 } //namespace armnn