src/backends/cl/workloads/ClConvertFp32ToFp16Workload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "ClConvertFp32ToFp16Workload.hpp"
   7 #include <cl/ClTensorHandle.hpp>
   8
   9 #include "ClWorkloadUtils.hpp"
  10
  11 namespace armnn
  12 {
  13 using namespace armcomputetensorutils;
  14
  15 static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE;
  16
  17 ClConvertFp32ToFp16Workload::ClConvertFp32ToFp16Workload(
  18     const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info) :
  19     Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info)
  20 {
  21     this->m_Data.ValidateInputsOutputs("ClConvertFp32ToFp16Workload", 1, 1);
  22
  23     arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor();
  24     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor();
  25
  26     m_Layer.configure(&input, &output, g_AclConvertPolicy, 0);
  27 }
  28
  29 void ClConvertFp32ToFp16Workload::Execute() const
  30 {
  31     ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvertFp32ToFp16Workload_Execute");
  32     RunClFunction(m_Layer, CHECK_LOCATION());
  33 }
  34
  35 arm_compute::Status ClConvertFp32ToFp16WorkloadValidate(const TensorInfo& input, const TensorInfo& output)
  36 {
  37     if (input.GetDataType() != DataType::Float32)
  38     {
  39         return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Input should be Float32");
  40     }
  41     if (output.GetDataType() != DataType::Float16)
  42     {
  43         return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Output should be Float16");
  44     }
  45
  46     const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
  47     const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
  48
  49     const arm_compute::Status aclStatus = arm_compute::CLDepthConvertLayer::validate(
  50         &aclInputInfo, &aclOutputInfo, g_AclConvertPolicy, 0);
  51
  52     return aclStatus;
  53 }
  54
  55
  56 } //namespace armnn