src/backends/neon/workloads/NeonConvertFp16ToFp32Workload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "NeonConvertFp16ToFp32Workload.hpp"
   7 #include <FloatingPointConverter.hpp>
   8
   9 #include <Half.hpp>
  10 #include <backendsCommon/WorkloadUtils.hpp>
  11
  12 namespace armnn
  13 {
  14
  15 NeonConvertFp16ToFp32Workload::NeonConvertFp16ToFp32Workload(const ConvertFp16ToFp32QueueDescriptor& descriptor,
  16                                                              const WorkloadInfo& info)
  17      : Float16ToFloat32Workload<ConvertFp16ToFp32QueueDescriptor>(descriptor, info)
  18 {
  19     this->m_Data.ValidateInputsOutputs("NeonConvertFp16ToFp32Workload", 1, 1);
  20     GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
  21 }
  22
  23 void NeonConvertFp16ToFp32Workload::Execute() const
  24 {
  25     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonConvertFp16ToFp32Workload_Execute");
  26
  27     auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
  28         {
  29             auto input = reinterpret_cast<const Half*>(src);
  30             auto output = reinterpret_cast<float*>(dst);
  31             size_t numElements = size/2; // 2 bytes per fp16
  32             armnnUtils::FloatingPointConverter::ConvertFloat16To32(input, numElements, output);
  33         };
  34
  35     for (const auto& pair : m_TensorHandlePairs)
  36     {
  37         CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
  38     }
  39 }
  40
  41 } //namespace armnn