* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License")
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file broadcast_to_with_dma.h
* \brief kernel of broadcast_to with nddma
*/
#ifndef BROADCAST_TO_WITH_NDDMA_H_
#define BROADCAST_TO_WITH_NDDMA_H_
#include "broadcast_to_base.h"
#include "kernel_operator.h"
namespace BrcTo
{
using namespace AscendC;
constexpr uint8_t bufferNum = 1;
constexpr int32_t queDepth = 1;
constexpr MultiCopyConfig copyCfg{false, 0, 0, false};
template <typename T, typename U, uint8_t maxDim = 4>
class BrcToWithNDDMA : public BrcToBase<U>
{
public:
__aicore__ inline BrcToWithNDDMA(){};
__aicore__ inline void Init(GM_ADDR x, GM_ADDR y, const U* tilingDataPtr, TPipe* pipeIn);
__aicore__ inline void Process();
private:
__aicore__ inline void SetDMAParams();
__aicore__ inline void CalcOutSize();
__aicore__ inline void CopyDataInWithDMA(LocalTensor<T>& tensor);
__aicore__ inline void CopyDataOut(LocalTensor<T>& tensor);
private:
const U* tdPtr_;
int64_t blockIdx = 0;
AxesLpInfo lpInfo;
int64_t aBaseIdx = 0;
int64_t bBaseIdx = 0;
TPipe* pipe_;
TQue<QuePosition::VECIN, queDepth> que;
GlobalTensor<T> inGM;
GlobalTensor<T> outGM;
int64_t gmInOffset;
int64_t gmOutOffset;
int64_t inBlockOffset = 0;
int64_t outBlockOffset = 0;
AscendC::MultiCopyLoopInfo<maxDim> copyLpInfo;
AscendC::MultiCopyParams<T, maxDim> mCopyParams;
AscendC::DataCopyExtParams copyParams{1, 0, 0, 0, 0};
uint32_t outLen = sizeof(T);
uint8_t copySwitch = 1;
};
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::Init(GM_ADDR x, GM_ADDR y, const U* tilingDataPtr, TPipe* pipeIn)
{
inGM.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(x));
outGM.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(y));
pipe_ = pipeIn;
tdPtr_ = tilingDataPtr;
pipe_->InitBuffer(que, bufferNum, tdPtr_->tensorSize * sizeof(T));
blockIdx = GetBlockIdx() % tdPtr_->usedCoreCnt;
}
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::SetDMAParams()
{
for (uint8_t i = 0; i < maxDim; i++) {
copyLpInfo.loopSrcStride[i] = tdPtr_->xSrcStride[i];
copyLpInfo.loopDstStride[i] = tdPtr_->xDstStride[i];
copyLpInfo.loopSize[i] = tdPtr_->xSize[i];
}
mCopyParams.loopInfo = copyLpInfo;
}
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::CopyDataInWithDMA(LocalTensor<T>& tensor)
{
if (copySwitch > 0) {
DataCopy<T, maxDim, copyCfg>(tensor, inGM[gmInOffset], mCopyParams);
this->InsertSync(HardEvent::MTE2_MTE3);
copySwitch = copySwitch * uint8_t(tdPtr_->isUNotB);
}
}
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::CalcOutSize()
{
for (uint8_t i = 1; i < maxDim; i++) {
outLen *= tdPtr_->xSize[i];
}
}
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::CopyDataOut(LocalTensor<T>& tensor)
{
int64_t bOffset = 0;
for (int64_t bIdx = lpInfo.bLpBegIdx; bIdx < lpInfo.bLpEndIdx; bIdx++) {
bOffset = this->CalcBAxesOffset(tdPtr_, bBaseIdx + bIdx);
DataCopyPad(outGM[gmOutOffset + bOffset], tensor, copyParams);
}
this->InsertSync(HardEvent::MTE3_MTE2);
}
template <typename T, typename U, uint8_t maxDim>
__aicore__ inline void BrcToWithNDDMA<T, U, maxDim>::Process()
{
this->CalcABBaseIdx(tdPtr_, blockIdx, aBaseIdx, bBaseIdx);
this->CalcAxesLoopInfo(tdPtr_, blockIdx, lpInfo);
this->CalcInBlockOffset(tdPtr_, blockIdx, inBlockOffset);
this->CalcOutBlockOffset(tdPtr_, blockIdx, outBlockOffset);
LocalTensor<T> tensor = que.AllocTensor<T>();
SetDMAParams();
CalcOutSize();
for (int64_t aIdx = 0; aIdx < lpInfo.aLpCnt; aIdx++) {
mCopyParams.loopInfo.loopSize[0] = tdPtr_->uLpUnit;
copyParams.blockLen = tdPtr_->uLpUnit * outLen;
copySwitch = 1;
int64_t aInOffset = this->CalcAInAxesOffset(tdPtr_, aBaseIdx + aIdx);
int64_t aOutOffset = this->CalcAOutAxesOffset(tdPtr_, aBaseIdx + aIdx);
for (int64_t uLpIdx = lpInfo.uLpBegIdx; uLpIdx < lpInfo.uLpEndIdx; uLpIdx++) {
gmInOffset = (aInOffset + inBlockOffset + uLpIdx * tdPtr_->uLpUnit * tdPtr_->uInOffset * tdPtr_->isUNotB);
CopyDataInWithDMA(tensor);
gmOutOffset = (aOutOffset + outBlockOffset + uLpIdx * tdPtr_->uLpUnit * tdPtr_->uOutOffset);
CopyDataOut(tensor);
}
if (lpInfo.uLeft > 0) {
mCopyParams.loopInfo.loopSize[0] = lpInfo.uLeft;
copyParams.blockLen = lpInfo.uLeft * outLen;
gmInOffset =
(aInOffset + inBlockOffset + lpInfo.uLpCnt * tdPtr_->uLpUnit * tdPtr_->uInOffset * tdPtr_->isUNotB);
CopyDataInWithDMA(tensor);
gmOutOffset = (aOutOffset + outBlockOffset + lpInfo.uLpCnt * tdPtr_->uLpUnit * tdPtr_->uOutOffset);
CopyDataOut(tensor);
}
}
que.FreeTensor(tensor);
}
}
#endif