* 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 slice_nddma.h
* \brief
*/
#ifndef SLICE_NDDMA_H
#define SLICE_NDDMA_H
#include "slice_base.h"
namespace Slice {
using namespace AscendC;
template <typename T, typename U, typename V = int8_t>
class SliceNDDMA : public SliceBase<T, U, V> {
public:
__aicore__ inline SliceNDDMA(){};
__aicore__ inline void Init(GM_ADDR x, GM_ADDR begin, GM_ADDR end, GM_ADDR strides, GM_ADDR y,
const SliceNDDMATilingData* tilingData, TPipe* pipeIn);
__aicore__ inline void Process();
private:
__aicore__ inline void ParseNDDMATilingData(
GM_ADDR begin, const SliceNDDMATilingData* tilingData, int64_t blockIdx);
__aicore__ inline void SetLoopInfo(MultiCopyLoopInfo<NDDMA_MAX_DIMS>& loopInfo,
MultiCopyLoopInfo<NDDMA_MAX_DIMS>& loopInfoTail);
__aicore__ inline void SetCopyOutParams(DataCopyExtParams ©OutParamsMain,
DataCopyExtParams ©OutParamsTail);
__aicore__ inline void ProcessPerBlock();
private:
TPipe *pipe_;
TQueBind<TPosition::VECIN, TPosition::VECOUT, 1> vecQue_;
GlobalTensor<T> inputGM_;
GlobalTensor<T> outputGM_;
int64_t blockIdx_ = 0;
static constexpr MultiCopyConfig nddmaConfig_ = {false, 0, 0, false};
int64_t ubSplitLoopsNum_ = 0;
int64_t curCoreLoopsNum_ = 0;
int64_t curCoreRowsOffset_ = 0;
};
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::Init(GM_ADDR x, GM_ADDR begin, GM_ADDR end, GM_ADDR strides,
GM_ADDR y, const SliceNDDMATilingData* tilingData,
TPipe* pipeIn)
{
blockIdx_ = GetBlockIdx();
pipe_ = pipeIn;
inputGM_.SetGlobalBuffer((__gm__ T*)x);
outputGM_.SetGlobalBuffer((__gm__ T*)y);
this->ParseNDDMATilingData(begin, tilingData, blockIdx_);
pipe_->InitBuffer(vecQue_, DOUBLE_BUFFER, this->ubSize_ / DOUBLE_BUFFER);
}
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::ParseNDDMATilingData(GM_ADDR begin, const SliceNDDMATilingData* tilingData,
int64_t blockIdx)
{
this->ParseBaseTilingData(begin, &(tilingData->sliceBaseTilingData), blockIdx);
this->ubOutLoopSteps_ = tilingData->ubOutLoopSteps;
for (int32_t i = 0; i < STRIDED_SLICE_MAX_AXIS_NUM; i++) {
this->nddmaLoopSize_[i] = tilingData->nddmaLoopSize[i];
this->nddmaLoopSrcStride_[i] = tilingData->nddmaLoopSrcStride[i];
this->nddmaLoopDstStride_[i] = tilingData->nddmaLoopDstStride[i];
}
this->nddmaTotalNum_ = tilingData->nddmaTotalNum;
}
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::Process()
{
if (blockIdx_ >= this->realCoreNum_) {
return;
}
this->CalcProcessLoopsNum(curCoreLoopsNum_, ubSplitLoopsNum_, blockIdx_);
this->GetProcessRowsOffset(curCoreRowsOffset_, blockIdx_);
ProcessPerBlock();
}
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::SetLoopInfo(MultiCopyLoopInfo<NDDMA_MAX_DIMS>& loopInfo,
MultiCopyLoopInfo<NDDMA_MAX_DIMS>& loopInfoTail)
{
int64_t inUbDims = this->inputDims_ - this->ubIndex_;
for (int64_t i = 0; i < inUbDims; i++) {
loopInfo.loopSize[i] = this->nddmaLoopSize_[NDDMA_MAX_DIMS - 1 - i];
loopInfo.loopSrcStride[i] = this->nddmaLoopSrcStride_[NDDMA_MAX_DIMS - 1 - i];
loopInfo.loopDstStride[i] = this->nddmaLoopDstStride_[NDDMA_MAX_DIMS - 1 - i];
loopInfo.loopLpSize[i] = 0;
loopInfo.loopRpSize[i] = 0;
loopInfoTail.loopSize[i] = this->nddmaLoopSize_[NDDMA_MAX_DIMS - 1 - i];
loopInfoTail.loopSrcStride[i] = this->nddmaLoopSrcStride_[NDDMA_MAX_DIMS - 1 - i];
loopInfoTail.loopDstStride[i] = this->nddmaLoopDstStride_[NDDMA_MAX_DIMS - 1 - i];
loopInfoTail.loopLpSize[i] = 0;
loopInfoTail.loopRpSize[i] = 0;
}
loopInfoTail.loopSize[inUbDims - 1] = this->ubTailFactor_;
for (int64_t i = inUbDims; i < NDDMA_MAX_DIMS; i++) {
loopInfo.loopSize[i] = 1;
loopInfo.loopSrcStride[i] = 0;
loopInfo.loopDstStride[i] = 0;
loopInfo.loopLpSize[i] = 0;
loopInfo.loopRpSize[i] = 0;
loopInfoTail.loopSize[i] = 1;
loopInfoTail.loopSrcStride[i] = 0;
loopInfoTail.loopDstStride[i] = 0;
loopInfoTail.loopLpSize[i] = 0;
loopInfoTail.loopRpSize[i] = 0;
}
}
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::SetCopyOutParams(DataCopyExtParams ©OutParamsMain,
DataCopyExtParams ©OutParamsTail)
{
copyOutParamsMain.blockCount = 1;
copyOutParamsMain.blockLen = this->nddmaTotalNum_ * sizeof(T);
copyOutParamsMain.dstStride = 0;
copyOutParamsMain.srcStride = 0;
if (this->ubTailFactor_ > 0) {
copyOutParamsTail.blockCount = 1;
copyOutParamsTail.blockLen = this->nddmaTotalNum_ / this->ubFactor_ * this->ubTailFactor_ * sizeof(T);
copyOutParamsTail.dstStride = 0;
copyOutParamsTail.srcStride = 0;
}
}
template <typename T, typename U, typename V>
__aicore__ inline void SliceNDDMA<T, U, V>::ProcessPerBlock()
{
int64_t inputGmAddr = 0;
int64_t outputGmAddr = 0;
int64_t handleRowsNum = 0;
MultiCopyParams<T, NDDMA_MAX_DIMS> paramsMain;
MultiCopyParams<T, NDDMA_MAX_DIMS> paramsTail;
paramsMain.constantValue = 0;
paramsTail.constantValue = 0;
SetLoopInfo(paramsMain.loopInfo, paramsTail.loopInfo);
DataCopyExtParams copyOutParamsMain;
DataCopyExtParams copyOutParamsTail;
SetCopyOutParams(copyOutParamsMain, copyOutParamsTail);
for (int64_t idx = 0; idx < curCoreLoopsNum_; idx++) {
inputGmAddr = this->GetInputGmAddr(curCoreRowsOffset_ + idx * this->rowsOffsetSteps_[this->ubIndex_]);
outputGmAddr = this->GetOutputGmAddr(curCoreRowsOffset_ + idx * this->rowsOffsetSteps_[this->ubIndex_]);
for (int64_t loops = 0; loops < ubSplitLoopsNum_; loops++) {
LocalTensor<T> inputLocal = vecQue_.AllocTensor<T>();
DataCopy<T, NDDMA_MAX_DIMS, nddmaConfig_>(inputLocal,
inputGM_[inputGmAddr + loops * this->ubInLoopSteps_],
paramsMain);
vecQue_.EnQue(inputLocal);
inputLocal = vecQue_.DeQue<T>();
DataCopyPad(outputGM_[outputGmAddr + loops * this->ubOutLoopSteps_], inputLocal, copyOutParamsMain);
vecQue_.FreeTensor(inputLocal);
}
if (this->ubTailFactor_ > 0) {
LocalTensor<T> inputLocal = vecQue_.AllocTensor<T>();
DataCopy<T, NDDMA_MAX_DIMS, nddmaConfig_>(inputLocal,
inputGM_[inputGmAddr + ubSplitLoopsNum_ * this->ubInLoopSteps_], paramsTail);
vecQue_.EnQue(inputLocal);
inputLocal = vecQue_.DeQue<T>();
DataCopyPad(outputGM_[outputGmAddr + ubSplitLoopsNum_ * this->ubOutLoopSteps_],
inputLocal, copyOutParamsTail);
vecQue_.FreeTensor(inputLocal);
}
}
}
}
#endif