src/armnn/backends/NeonWorkloads/NeonConvertFp32ToFp16Workload.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 "NeonConvertFp32ToFp16Workload.hpp"
   7
   8 #include "Half.hpp"
   9 #include "FloatingPointConverter.hpp"
  10
  11 #include "Profiling.hpp"
  12 #include "backends/WorkloadUtils.hpp"
  13
  14 namespace armnn
  15 {
  16
  17 NeonConvertFp32ToFp16Workload::NeonConvertFp32ToFp16Workload(const ConvertFp32ToFp16QueueDescriptor& descriptor,
  18                                                              const WorkloadInfo& info)
  19     : Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info)
  20 {
  21     this->m_Data.ValidateInputsOutputs("NeonConvertFp32ToFp16Workload", 1, 1);
  22     GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
  23 }
  24
  25 void NeonConvertFp32ToFp16Workload::Execute() const
  26 {
  27     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonConvertFp32ToFp16Workload_Execute");
  28
  29     auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
  30         {
  31             auto input = reinterpret_cast<const float*>(src);
  32             auto output = reinterpret_cast<Half*>(dst);
  33             size_t numElements = size/2; // 2 bytes per fp16
  34             armnnUtils::FloatingPointConverter::ConvertFloat32To16(input, numElements, output);
  35         };
  36
  37     for (const auto& pair : m_TensorHandlePairs)
  38     {
  39         CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
  40     }
  41 }
  42
  43 } //namespace armnn