src/armnn/backends/RefWorkloads/RefFullyConnectedUint8Workload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // See LICENSE file in the project root for full license information.
   4 //
   5
   6 #include "RefFullyConnectedUint8Workload.hpp"
   7
   8 #include "FullyConnected.hpp"
   9 #include "RefWorkloadUtils.hpp"
  10
  11 #include "Profiling.hpp"
  12
  13 #include <vector>
  14
  15 namespace armnn
  16 {
  17 RefFullyConnectedUint8Workload::RefFullyConnectedUint8Workload(
  18     const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info)
  19      : Uint8Workload<FullyConnectedQueueDescriptor>(descriptor, info),
  20         m_Weight(std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Weight))),
  21         m_Bias(descriptor.m_Parameters.m_BiasEnabled
  22                ? std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Bias)) : nullptr) {}
  23
  24 void RefFullyConnectedUint8Workload::Execute() const
  25 {
  26     ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefFullyConnectedUint8Workload_Execute");
  27
  28     const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
  29     const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
  30
  31     const uint8_t* weightData = m_Weight->GetConstTensor<uint8_t>();
  32
  33     auto dequant = Dequantize(GetInputTensorDataU8(0, m_Data), inputInfo);
  34
  35     auto weight = Dequantize(weightData, m_Weight->GetTensorInfo());
  36
  37     std::vector<float> results(outputInfo.GetNumElements());
  38
  39     if (m_Data.m_Parameters.m_BiasEnabled)
  40     {
  41         const int32_t* biasData = m_Bias->GetConstTensor<int32_t>();
  42         auto           bias     = Dequantize(biasData, m_Bias->GetTensorInfo());
  43
  44         FullyConnected(dequant.data(),
  45                        results.data(),
  46                        inputInfo,
  47                        outputInfo,
  48                        weight.data(),
  49                        bias.data(),
  50                        m_Data.m_Parameters.m_TransposeWeightMatrix);
  51     }
  52     else
  53     {
  54         FullyConnected(dequant.data(),
  55                        results.data(),
  56                        inputInfo,
  57                        outputInfo,
  58                        weight.data(),
  59                        nullptr,
  60                        m_Data.m_Parameters.m_TransposeWeightMatrix);
  61     }
  62
  63     Quantize(GetOutputTensorDataU8(0, m_Data), results.data(), outputInfo);
  64 }
  65
  66 } //namespace armnn