src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "NeonFullyConnectedWorkload.hpp"
   7
   8 #include <aclCommon/ArmComputeTensorUtils.hpp>
   9 #include <aclCommon/ArmComputeUtils.hpp>
  10 #include <backendsCommon/CpuTensorHandle.hpp>
  11
  12 namespace armnn
  13 {
  14 using namespace armcomputetensorutils;
  15
  16 arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
  17                                                        const TensorInfo& output,
  18                                                        const TensorInfo& weights,
  19                                                        const TensorInfo& biases,
  20                                                        const FullyConnectedDescriptor& descriptor)
  21 {
  22     const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input);
  23     const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output);
  24     const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights);
  25
  26     arm_compute::TensorInfo aclBiases;
  27     arm_compute::TensorInfo *optionalAclBiases = nullptr;
  28     if (descriptor.m_BiasEnabled)
  29     {
  30         aclBiases  = BuildArmComputeTensorInfo(biases);
  31         optionalAclBiases = &aclBiases;
  32     }
  33
  34     const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo =
  35         ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor);
  36
  37
  38     return arm_compute::NEFullyConnectedLayer::validate(&aclInput,
  39                                                         &aclWeights,
  40                                                         optionalAclBiases,
  41                                                         &aclOutput,
  42                                                         fullyConnectedLayerInfo);
  43 }
  44
  45 NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
  46     const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
  47     : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
  48     , m_FullyConnectedLayer(memoryManager)
  49 {
  50     m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
  51
  52     arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
  53     arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
  54
  55     m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
  56     BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());
  57
  58     if (m_Data.m_Parameters.m_BiasEnabled)
  59     {
  60         m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
  61         BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
  62     }
  63
  64     // Construct
  65     arm_compute::FullyConnectedLayerInfo fc_info;
  66     fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix;
  67     m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
  68
  69     // Allocate
  70     if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QuantisedAsymm8)
  71     {
  72         InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
  73     }
  74     else
  75     {
  76         InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
  77     }
  78
  79     if (m_BiasesTensor)
  80     {
  81         if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
  82         {
  83             InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
  84         }
  85         else
  86         {
  87             InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
  88         }
  89     }
  90
  91     // Force Compute Library to perform the necessary copying and reshaping, after which
  92     // delete all the input tensors that will no longer be needed
  93     m_FullyConnectedLayer.prepare();
  94     FreeUnusedTensors();
  95 }
  96
  97 void NeonFullyConnectedWorkload::Execute() const
  98 {
  99     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedWorkload_Execute");
 100     m_FullyConnectedLayer.run();
 101 }
 102
 103 void NeonFullyConnectedWorkload::FreeUnusedTensors()
 104 {
 105     FreeTensorIfUnused(m_WeightsTensor);
 106     FreeTensorIfUnused(m_BiasesTensor);
 107 }
 108
 109 } //namespace armnn