src/armnn/backends/ClWorkloads/ClConvertFp32ToFp16Workload.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 "ClConvertFp32ToFp16Workload.hpp"
   7 #include "backends/ClTensorHandle.hpp"
   8
   9 namespace armnn
  10 {
  11 using namespace armcomputetensorutils;
  12
  13 static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE;
  14
  15 ClConvertFp32ToFp16Workload::ClConvertFp32ToFp16Workload(
  16     const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info) :
  17     Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info)
  18 {
  19     this->m_Data.ValidateInputsOutputs("ClConvertFp32ToFp16Workload", 1, 1);
  20
  21     arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor();
  22     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor();
  23
  24     m_Layer.configure(&input, &output, g_AclConvertPolicy, 0);
  25 }
  26
  27 void ClConvertFp32ToFp16Workload::Execute() const
  28 {
  29     ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvertFp32ToFp16Workload_Execute");
  30     m_Layer.run();
  31 }
  32
  33 arm_compute::Status ClConvertFp32ToFp16WorkloadValidate(const TensorInfo& input,
  34                                                         const TensorInfo& output,
  35                                                         std::string* reasonIfUnsupported)
  36 {
  37     if (input.GetDataType() != DataType::Float32)
  38     {
  39         *reasonIfUnsupported = "Input should be Float32";
  40         return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, *reasonIfUnsupported);
  41     }
  42     if (output.GetDataType() != DataType::Float16)
  43     {
  44         *reasonIfUnsupported = "Output should be Float16";
  45         return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, *reasonIfUnsupported);
  46     }
  47
  48     const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
  49     const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
  50
  51     const arm_compute::Status aclStatus = arm_compute::CLDepthConvertLayer::validate(
  52         &aclInputInfo, &aclOutputInfo, g_AclConvertPolicy, 0);
  53
  54     const bool supported = (aclStatus.error_code() == arm_compute::ErrorCode::OK);
  55     if (!supported && reasonIfUnsupported)
  56     {
  57         *reasonIfUnsupported = aclStatus.error_description();
  58     }
  59
  60     return aclStatus;
  61 }
  62
  63
  64 } //namespace armnn