* 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 real_div.h
* \brief
*/
#ifndef REALDIV_H
#define REALDIV_H
#ifndef K_MAX_SHAPE_DIM
#define K_MAX_SHAPE_DIM 8
#endif
#include "kernel_operator.h"
#include "kernel_tiling/kernel_tiling.h"
#include "real_div_tiling_data.h"
#include "real_div_tiling_key.h"
namespace MyRealDiv {
using namespace AscendC;
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
class KernelRealDiv {
public:
__aicore__ inline KernelRealDiv(){};
__aicore__ inline void Init(GM_ADDR x1, GM_ADDR x2, GM_ADDR y, const RealDivTilingData* tilingData, TPipe* pipeIn);
__aicore__ inline void Process();
private:
__aicore__ inline void CopyInAndCompute(int32_t offset);
__aicore__ inline void CopyOut(int32_t offset);
__aicore__ inline void CopyInAndCompute16B(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local, LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset);
__aicore__ inline void CopyInAndComputeInt32(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local, LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset);
__aicore__ inline void CopyInAndComputeBool(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local, LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset);
private:
TPipe* pipe;
TQue<QuePosition::VECIN, BUFFER_NUM> inQueue;
TQue<QuePosition::VECOUT, BUFFER_NUM> outQueueY;
TBuf<QuePosition::VECCALC> castTmp, halfCastTmp;
GlobalTensor<TYPE_X> x1Gm;
GlobalTensor<TYPE_X> x2Gm;
GlobalTensor<TYPE_Y> yGm;
uint32_t coreDataNum;
uint32_t tileNum;
uint32_t tileDataNum;
uint32_t tailDataNum;
uint32_t processDataNum;
};
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::Init(GM_ADDR x1, GM_ADDR x2, GM_ADDR y, const RealDivTilingData* tilingData, TPipe* pipeIn)
{
ASSERT(GetBlockNum() != 0 && "block dim can not be zero!");
uint32_t coreNum = GetBlockIdx();
uint32_t globalBufferIndex = tilingData->bigCoreDataNum * GetBlockIdx();
this->tileDataNum = tilingData->tileDataNum;
this->pipe = pipeIn;
if (coreNum < tilingData->tailBlockNum)
{
this->coreDataNum = tilingData->bigCoreDataNum;
this->tileNum = tilingData->finalBigTileNum;
this->tailDataNum = tilingData->bigTailDataNum;
}
else
{
this->coreDataNum = tilingData->smallCoreDataNum;
this->tileNum = tilingData->finalSmallTileNum;
this->tailDataNum = tilingData->smallTailDataNum;
globalBufferIndex -= (tilingData->bigCoreDataNum - tilingData->smallCoreDataNum) * (GetBlockIdx() - tilingData->tailBlockNum);
}
x1Gm.SetGlobalBuffer((__gm__ TYPE_X *)x1 + globalBufferIndex, this->coreDataNum);
x2Gm.SetGlobalBuffer((__gm__ TYPE_X *)x2 + globalBufferIndex, this->coreDataNum);
yGm.SetGlobalBuffer((__gm__ TYPE_Y *)y + globalBufferIndex, this->coreDataNum);
pipe->InitBuffer(inQueue, BUFFER_NUM, this->tileDataNum * sizeof(TYPE_X) * 2);
pipe->InitBuffer(outQueueY, BUFFER_NUM, this->tileDataNum * sizeof(TYPE_Y));
if constexpr ( IsSameType<TYPE_X, bfloat16_t>::value)
{
pipe->InitBuffer(castTmp, this->tileDataNum * sizeof(float));
}
if constexpr ( IsSameType<TYPE_X, half>::value)
{
#if defined(HIGH_PERFORMANCE) && HIGH_PERFORMANCE == 1
#else
pipe->InitBuffer(castTmp, this->tileDataNum * sizeof(float));
#endif
}
if constexpr ( IsSameType<TYPE_X, bool>::value){
pipe->InitBuffer(castTmp, this->tileDataNum * sizeof(float));
pipe->InitBuffer(halfCastTmp, this->tileDataNum * sizeof(half));
}
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::CopyInAndCompute16B(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local,
LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset)
{
LocalTensor<float> x1LocalFp = castTmp.Get<float>();
LocalTensor<float> x2LocalFp = x1Local.template ReinterpretCast<float>();
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1Local, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2Local, RoundMode::CAST_NONE, this->processDataNum);
Div(x1LocalFp, x1LocalFp, x2LocalFp, this->processDataNum);
Cast(yLocal, x1LocalFp, RoundMode::CAST_RINT, this->processDataNum);
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::CopyInAndComputeInt32(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local,
LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset)
{
LocalTensor<float> x1LocalFp = x1Local.template ReinterpretCast<float>();
LocalTensor<float> x2LocalFp = x2Local.template ReinterpretCast<float>();
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1Local, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2Local, RoundMode::CAST_NONE, this->processDataNum);
Div(x1LocalFp, x1LocalFp, x2LocalFp, this->processDataNum);
Cast(yLocal, x1LocalFp, RoundMode::CAST_TRUNC, this->processDataNum);
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::CopyInAndComputeBool(LocalTensor<TYPE_X>& x1Local, LocalTensor<TYPE_X>& x2Local,
LocalTensor<TYPE_Y>& yLocal, int eventIDMTE2ToV, int offset)
{
LocalTensor<half> xLocalHalf = halfCastTmp.Get<half>();
LocalTensor<float> x2LocalFp = castTmp.Get<float>();
LocalTensor<int8_t> x1LocalInt8 = x1Local.template ReinterpretCast<int8_t>();
LocalTensor<int8_t> x2LocalInt8 = x2Local.template ReinterpretCast<int8_t>();
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(xLocalHalf, x1LocalInt8, RoundMode::CAST_NONE, this->processDataNum);
Cast(yLocal, xLocalHalf, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(xLocalHalf, x2LocalInt8, RoundMode::CAST_NONE, this->processDataNum);
Cast(x2LocalFp, xLocalHalf, RoundMode::CAST_NONE, this->processDataNum);
Div(yLocal, yLocal, x2LocalFp, this->processDataNum);
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::CopyInAndCompute(int32_t offset)
{
LocalTensor<TYPE_X> x1Local = inQueue.template AllocTensor<TYPE_X>();
LocalTensor<TYPE_X> x2Local = x1Local[this->processDataNum];
LocalTensor<TYPE_Y> yLocal = outQueueY.template AllocTensor<TYPE_Y>();
int32_t eventIDMTE2ToV = static_cast<int32_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE2_V));
if constexpr ( IsSameType<TYPE_X, bfloat16_t>::value)
{
CopyInAndCompute16B(x1Local, x2Local, yLocal, eventIDMTE2ToV, offset);
}
else if constexpr ( IsSameType<TYPE_X, int32_t>::value)
{
CopyInAndComputeInt32(x1Local, x2Local, yLocal, eventIDMTE2ToV, offset);
}
else if constexpr ( IsSameType<TYPE_X, half>::value)
{
#if defined(HIGH_PERFORMANCE) && HIGH_PERFORMANCE == 1
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Div(yLocal, x1Local, x2Local, this->processDataNum);
#else
CopyInAndCompute16B(x1Local, x2Local, yLocal, eventIDMTE2ToV, offset);
#endif
}
else if constexpr ( IsSameType<TYPE_X, bool>::value)
{
CopyInAndComputeBool(x1Local, x2Local, yLocal, eventIDMTE2ToV, offset);
}
else
{
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Div(yLocal, x1Local, x2Local, this->processDataNum);
}
outQueueY.template EnQue<TYPE_Y>(yLocal);
inQueue.template FreeTensor(x1Local);
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::CopyOut(int32_t offset)
{
LocalTensor<TYPE_Y> yLocal = outQueueY.template DeQue<TYPE_Y>();
DataCopy(yGm[offset], yLocal, this->processDataNum);
outQueueY.template FreeTensor(yLocal);
}
template <typename TYPE_X, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void KernelRealDiv<TYPE_X, TYPE_Y, BUFFER_NUM>::Process()
{
int32_t loopCount = this->tileNum - 1;
this->processDataNum = this->tileDataNum;
int32_t offset = 0;
for (int32_t i = 0; i < loopCount; i++, offset+=this->tileDataNum)
{
CopyInAndCompute(offset);
CopyOut(offset);
}
this->processDataNum = this->tailDataNum;
CopyInAndCompute(offset);
CopyOut(offset);
}
}
#endif
