* 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 split_all.h
* \brief
*/
#ifndef SPLIT_ALL_TRIANGULATOR
#define SPLIT_ALL_TRIANGULATOR
#include "kernel_operator.h"
namespace Triangulator {
using namespace AscendC;
constexpr int32_t OUTPUT_ZERO_MODE = 0;
constexpr int32_t OUTPUT_INPUT_MODE = 1;
template <typename T, const int32_t MODE>
class SplitAll {
public:
__aicore__ inline SplitAll(){};
__aicore__ inline void Init(
GM_ADDR x, GM_ADDR y, __tiling_data_ptr__ TriangulatorTilingData* tilingDataPtr, TPipe* pipeIn);
__aicore__ inline void ParseTilingData(__tiling_data_ptr__ TriangulatorTilingData* tilingDataPtr);
__aicore__ inline void Process();
private:
__aicore__ inline void ProcessPerLoop(
LocalTensor<T>& srcLocal, TEventID eventId, int64_t loopNum, int64_t loopBlockLength);
private:
TPipe* pipe_;
TBuf<QuePosition::VECCALC> ubBuf1_;
TBuf<QuePosition::VECCALC> ubBuf2_;
GlobalTensor<T> dstGm_;
GlobalTensor<T> srcGm_;
int64_t bufferSize_;
int64_t usedCoreNum_;
int64_t normalCoreProcessNum_;
int64_t tailCoreProcessNum_;
int64_t curCoreProcessNum_;
int64_t curUbLoopCount_;
int64_t normalLoopBlockNum_;
int64_t tailLoopBlockNum_;
uint32_t blockIdx_;
bool pingpong_;
};
template <typename T, const int32_t MODE>
__aicore__ inline void SplitAll<T, MODE>::ParseTilingData(__tiling_data_ptr__ TriangulatorTilingData* tilingDataPtr)
{
usedCoreNum_ = tilingDataPtr->usedCoreNum;
normalCoreProcessNum_ = tilingDataPtr->normalCoreProcessNum;
tailCoreProcessNum_ = tilingDataPtr->tailCoreProcessNum;
if constexpr (MODE == OUTPUT_INPUT_MODE) {
bufferSize_ = tilingDataPtr->bufferSize;
} else {
bufferSize_ = tilingDataPtr->bufferSize * 2;
}
}
template <typename T, const int32_t MODE>
__aicore__ inline void SplitAll<T, MODE>::Init(
GM_ADDR x, GM_ADDR y, __tiling_data_ptr__ TriangulatorTilingData* tilingDataPtr, TPipe* pipeIn)
{
ParseTilingData(tilingDataPtr);
blockIdx_ = GetBlockIdx();
pingpong_ = true;
if (blockIdx_ >= usedCoreNum_) {
return;
}
int64_t burstLen = bufferSize_ / sizeof(T);
curCoreProcessNum_ = blockIdx_ < usedCoreNum_ - 1 ? normalCoreProcessNum_ : tailCoreProcessNum_;
curUbLoopCount_ = (curCoreProcessNum_ + burstLen - 1) / burstLen;
normalLoopBlockNum_ = (curCoreProcessNum_ >= burstLen) ? burstLen : curCoreProcessNum_;
tailLoopBlockNum_ = curCoreProcessNum_ - (curUbLoopCount_ - 1) * normalLoopBlockNum_;
int64_t intraCoreOffset = blockIdx_ * normalCoreProcessNum_;
srcGm_.SetGlobalBuffer((__gm__ T*)x + intraCoreOffset);
dstGm_.SetGlobalBuffer((__gm__ T*)y + intraCoreOffset);
pipe_ = pipeIn;
pipe_->InitBuffer(ubBuf1_, bufferSize_);
if constexpr (MODE == OUTPUT_INPUT_MODE) {
pipe_->InitBuffer(ubBuf2_, bufferSize_);
}
}
template <typename T, const int32_t MODE>
__aicore__ inline void SplitAll<T, MODE>::Process()
{
if (blockIdx_ >= usedCoreNum_) {
return;
}
LocalTensor<T> srcLocal1 = ubBuf1_.Get<T>();
LocalTensor<T> srcLocal2;
if constexpr (MODE == OUTPUT_INPUT_MODE) {
srcLocal2 = ubBuf2_.Get<T>();
} else {
srcLocal2 = srcLocal1;
}
TEventID eventFirst = 0;
TEventID eventSecond = 0;
if constexpr (MODE == OUTPUT_INPUT_MODE) {
eventFirst = GetTPipePtr()->AllocEventID<HardEvent::MTE3_MTE2>();
eventSecond = GetTPipePtr()->AllocEventID<HardEvent::MTE3_MTE2>();
SetFlag<HardEvent::MTE3_MTE2>(eventFirst);
SetFlag<HardEvent::MTE3_MTE2>(eventSecond);
} else {
eventFirst = GetTPipePtr()->AllocEventID<HardEvent::V_MTE3>();
Duplicate(srcLocal1, T(0), bufferSize_ / sizeof(T));
SetFlag<HardEvent::V_MTE3>(eventFirst);
WaitFlag<HardEvent::V_MTE3>(eventFirst);
GetTPipePtr()->ReleaseEventID<HardEvent::V_MTE3>(eventFirst);
}
for (int64_t loopNum = 0; loopNum < curUbLoopCount_ - 1; loopNum++) {
if (pingpong_) {
ProcessPerLoop(srcLocal1, eventFirst, loopNum, normalLoopBlockNum_);
} else {
ProcessPerLoop(srcLocal2, eventSecond, loopNum, normalLoopBlockNum_);
}
pingpong_ = !pingpong_;
}
if (pingpong_) {
ProcessPerLoop(srcLocal1, eventFirst, curUbLoopCount_ - 1, tailLoopBlockNum_);
} else {
ProcessPerLoop(srcLocal2, eventSecond, curUbLoopCount_ - 1, tailLoopBlockNum_);
}
if constexpr (MODE == OUTPUT_INPUT_MODE) {
WaitFlag<HardEvent::MTE3_MTE2>(eventFirst);
WaitFlag<HardEvent::MTE3_MTE2>(eventSecond);
GetTPipePtr()->ReleaseEventID<HardEvent::MTE3_MTE2>(eventFirst);
GetTPipePtr()->ReleaseEventID<HardEvent::MTE3_MTE2>(eventSecond);
} else {
PipeBarrier<PIPE_MTE3>();
}
}
template <typename T, const int32_t MODE>
__aicore__ inline void SplitAll<T, MODE>::ProcessPerLoop(
LocalTensor<T>& srcLocal, TEventID eventId, int64_t loopNum, int64_t loopBlockLength)
{
DataCopyExtParams copyParams = {
static_cast<uint16_t>(1), static_cast<uint32_t>(loopBlockLength * sizeof(T)), static_cast<uint32_t>(0),
static_cast<uint32_t>(0), static_cast<uint32_t>(0)};
DataCopyPadExtParams<T> padParams = {false, static_cast<uint8_t>(0), static_cast<uint8_t>(0), static_cast<T>(0)};
if constexpr (MODE == OUTPUT_INPUT_MODE) {
WaitFlag<HardEvent::MTE3_MTE2>(eventId);
DataCopyPad(srcLocal, srcGm_[normalLoopBlockNum_ * loopNum], copyParams, padParams);
SetFlag<HardEvent::MTE2_MTE3>(eventId);
WaitFlag<HardEvent::MTE2_MTE3>(eventId);
}
DataCopyPad(dstGm_[normalLoopBlockNum_ * loopNum], srcLocal, copyParams);
if constexpr (MODE == OUTPUT_INPUT_MODE) {
SetFlag<HardEvent::MTE3_MTE2>(eventId);
}
}
}
#endif