EthosnRefConvolution2dWorkload.cpp

Go to the documentation of this file.

//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "EthosnRefConvolution2dWorkload.hpp"
#include "EthosnRefConvImpl.hpp"
#include "EthosnRefWorkloadUtils.hpp"

using namespace armnn::ethosnref;

namespace armnn
{

static inline std::string Split(const std::string& s, const std::string& delim, std::size_t& pos)
{
    if (pos >= s.length())
    {
        return "";
    }

    std::size_t end = s.find(delim, pos);
    if (end == std::string::npos)
    {
        end = s.length();
    }

    std::string str = s.substr(pos, end - pos);
    pos             = end + 1;

    return str;
}

EthosnRefConvolution2dWorkload::EthosnRefConvolution2dWorkload(
        const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, bool winograd, std::string BlockConfigs)
        : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
{
    m_Weight = std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Weight));

    if (descriptor.m_Parameters.m_BiasEnabled)
    {
        m_Bias = std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Bias));
    }

    m_Winograd = winograd;

    m_IsBlockConfigsWinograd2dCompatible = false;
    std::string part;
    std::size_t pos = 0;
    while (!(part = Split(BlockConfigs, ",", pos)).empty())
    {
        if (part == "16x8")
        {
            m_IsBlockConfigsWinograd2dCompatible = true;
        }
        else if (part == "8x16")
        {
            m_IsBlockConfigsWinograd2dCompatible = true;
        }
        else if (part == "8x8")
        {
            m_IsBlockConfigsWinograd2dCompatible = true;
        }
    }
    if (BlockConfigs == "")
    {
        m_IsBlockConfigsWinograd2dCompatible = true;
    }
}


void EthosnRefConvolution2dWorkload::Execute() const {

    ARMNN_SCOPED_PROFILING_EVENT_ETHOSN("EthosnRefConvolution2dWorkload_Execute");

    const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
    const TensorInfo& weightsInfo = m_Weight->GetTensorInfo();

    bool useWinograd = false;

    if (m_Winograd == true) {
      const armnnUtils::DataLayoutIndexed dataLayoutIndexed(m_Data.m_Parameters.m_DataLayout);
      unsigned int weightsHeight = weightsInfo.GetShape()[dataLayoutIndexed.GetHeightIndex()];
      unsigned int weightsWidth  = weightsInfo.GetShape()[dataLayoutIndexed.GetWidthIndex()];
      unsigned int xStride       = m_Data.m_Parameters.m_StrideX;
      unsigned int yStride       = m_Data.m_Parameters.m_StrideY;

      bool isWinograd1D    = weightsWidth == 1 || weightsHeight == 1;
      bool isBlockSizeOk   = isWinograd1D? true : m_IsBlockConfigsWinograd2dCompatible;

      uint32_t numMultsDirect, numMultsWinograd;

      if ( weightsWidth == 1 || weightsHeight == 1)
      {
          // 1D convolution kernel dim w x 1 or 1 x h
          // numOfMultiplications = 2 * w or 2 * h                   DIRECT
          //                      = 4 * CEIL(W/3) or 4 * CEIL(H/3)   WINOGRAD
          numMultsDirect   = weightsWidth * weightsHeight * 2;
          numMultsWinograd = 4 * DivideRoundUp(weightsWidth * weightsHeight, 3);
      }
      else
      {
          // 2D convolution kernel dim w x h
          // numOfMultiplications = 4 * w * h                    DIRECT
          //                      = 16 * CEIL(W/3) * CEIL(H/3)   WINOGRAD
          numMultsDirect   = weightsWidth * weightsHeight * 4;
          numMultsWinograd = 16 *  DivideRoundUp(weightsWidth, 3) * DivideRoundUp(weightsHeight, 3);
      }
      if (numMultsWinograd < numMultsDirect &&
              isBlockSizeOk &&
              //TODO: FENCHSW-922 Need to revisit then when
              // SW add support for Winogard + Strided
              xStride == 1 && yStride == 1)
          useWinograd = true;
    }

    if (useWinograd == true)
    {
        std::cout << "[EthosNRef] Runnning in Winograd Mode.\n";
    }
    else
    {
        std::cout << "[EthosNRef] Runnning in Direct Mode.\n";
    }

    if (CheckDataType(DataType::QSymmS16, inputInfo.GetDataType(), weightsInfo.GetDataType())) {
        const int16_t* inputData = GetInputTensorData<int16_t>(0, m_Data);
        const int16_t* weightsData = m_Weight->template GetConstTensor<int16_t>();
        const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
        const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);

        if(useWinograd == false)
        {
          EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, int16_t, int16_t, int32_t, int64_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
        else
        {
          EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, int16_t, int16_t, int32_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
    }
    else if (CheckDataType(DataType::QSymmS8, inputInfo.GetDataType(), weightsInfo.GetDataType())) {
        const int8_t* inputData = GetInputTensorData<int8_t>(0, m_Data);
        const int8_t* weightsData = m_Weight->template GetConstTensor<int8_t>();
        const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
        const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);

        if(useWinograd == false)
        {
          EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, int8_t, int8_t, int32_t, int32_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
        else
        {
          EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, int8_t, int8_t, int32_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
    }
    else { // QAsymmU8
        assert(CheckDataType(DataType::QAsymmU8, inputInfo.GetDataType(), weightsInfo.GetDataType()));

        const uint8_t* inputData = GetInputTensorData<uint8_t>(0, m_Data);
        const uint8_t* weightsData = m_Weight->template GetConstTensor<uint8_t>();
        const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
        const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);

        if(useWinograd == false)
        {
          EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, uint8_t, uint8_t, int32_t, int32_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
        else
        {
          EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, uint8_t, uint8_t, int32_t>(
            m_Data,
            inputData, inputInfo.GetQuantizationScale(),  inputInfo.GetQuantizationOffset(),
            weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
            biasData,
            outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
        }
    }
}

} //namespace armnn