src/backends/neon/workloads/NeonConvertFp32ToFp16Workload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "NeonConvertFp32ToFp16Workload.hpp"
   7
   8 #include <Half.hpp>
   9 #include <FloatingPointConverter.hpp>
  10 #include <Profiling.hpp>
  11 #include <backendsCommon/WorkloadUtils.hpp>
  12
  13 namespace armnn
  14 {
  15
  16 NeonConvertFp32ToFp16Workload::NeonConvertFp32ToFp16Workload(const ConvertFp32ToFp16QueueDescriptor& descriptor,
  17                                                              const WorkloadInfo& info)
  18     : Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info)
  19 {
  20     this->m_Data.ValidateInputsOutputs("NeonConvertFp32ToFp16Workload", 1, 1);
  21     GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
  22 }
  23
  24 void NeonConvertFp32ToFp16Workload::Execute() const
  25 {
  26     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonConvertFp32ToFp16Workload_Execute");
  27
  28     auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
  29         {
  30             auto input = reinterpret_cast<const float*>(src);
  31             auto output = reinterpret_cast<Half*>(dst);
  32             size_t numElements = size/2; // 2 bytes per fp16
  33             armnnUtils::FloatingPointConverter::ConvertFloat32To16(input, numElements, output);
  34         };
  35
  36     for (const auto& pair : m_TensorHandlePairs)
  37     {
  38         CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
  39     }
  40 }
  41
  42 } //namespace armnn