src/backends/neon/workloads/NeonConstantWorkload.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "NeonConstantWorkload.hpp"
   7
   8 #include <arm_compute/core/Types.h>
   9 #include <Half.hpp>
  10 #include <aclCommon/ArmComputeTensorUtils.hpp>
  11 #include <neon/NeonTensorHandle.hpp>
  12 #include <backendsCommon/CpuTensorHandle.hpp>
  13 #include <backendsCommon/Workload.hpp>
  14
  15 #include <boost/cast.hpp>
  16
  17 namespace armnn
  18 {
  19
  20 NeonConstantWorkload::NeonConstantWorkload(const ConstantQueueDescriptor& descriptor,
  21                                            const WorkloadInfo& info)
  22     : BaseWorkload<ConstantQueueDescriptor>(descriptor, info)
  23     , m_RanOnce(false)
  24 {
  25 }
  26
  27 void NeonConstantWorkload::Execute() const
  28 {
  29     ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonConstantWorkload_Execute");
  30
  31     using namespace armcomputetensorutils;
  32
  33     // The intermediate tensor held by the corresponding layer output handler can be initialised with the
  34     // given data on the first inference, then reused for subsequent inferences.
  35     // The initialisation cannot happen at workload construction time since the ACL kernel for the next layer
  36     // may not have been configured at the time.
  37     if (!m_RanOnce)
  38     {
  39         const ConstantQueueDescriptor& data = this->m_Data;
  40
  41         BOOST_ASSERT(data.m_LayerOutput != nullptr);
  42         arm_compute::ITensor& output =
  43             boost::polymorphic_downcast<NeonTensorHandle*>(data.m_Outputs[0])->GetTensor();
  44         arm_compute::DataType computeDataType =
  45             boost::polymorphic_downcast<NeonTensorHandle*>(data.m_Outputs[0])->GetDataType();
  46
  47         switch (computeDataType)
  48         {
  49             case arm_compute::DataType::F16:
  50             {
  51                 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<Half>(), output);
  52                 break;
  53             }
  54             case arm_compute::DataType::F32:
  55             {
  56                 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<float>(), output);
  57                 break;
  58             }
  59             case arm_compute::DataType::QASYMM8:
  60             {
  61                 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<uint8_t>(), output);
  62                 break;
  63             }
  64             default:
  65             {
  66                 BOOST_ASSERT_MSG(false, "Unknown data type");
  67                 break;
  68             }
  69         }
  70
  71         m_RanOnce = true;
  72     }
  73 }
  74
  75 } //namespace armnn