* 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 axpy_v2.h
* \brief
*/
#ifndef __AXPY_V2_H__
#define __AXPY_V2_H__
#include "kernel_operator.h"
#include "kernel_tiling/kernel_tiling.h"
#include "axpy_v2_tiling_data.h"
#include "axpy_v2_tiling_key.h"
namespace NsAxpyV2 {
using namespace AscendC;
template <typename TYPE_X1, typename TYPE_X2, typename TYPE_ALPHA, typename TYPE_Y, uint64_t BUFFER_NUM>
class AxpyV2 {
public:
__aicore__ inline AxpyV2(){};
__aicore__ inline void Init(GM_ADDR x1, GM_ADDR x2, GM_ADDR alpha, GM_ADDR y, const AxpyV2TilingData* tilingData, TPipe* pipeIn);
__aicore__ inline void InitAlphaLocal();
__aicore__ inline void Process();
private:
__aicore__ inline void TotalStage(int32_t offset);
private:
TPipe* pipe;
TQueBind<TPosition::VECIN, TPosition::VECOUT, BUFFER_NUM> queBind;
TBuf<TPosition::VECCALC> alphaBuf;
LocalTensor<float> alphaLocalFp;
LocalTensor<int32_t> alphaLocalInt;
GlobalTensor<TYPE_X1> x1Gm;
GlobalTensor<TYPE_X2> x2Gm;
GlobalTensor<TYPE_ALPHA> alphaGm;
GlobalTensor<TYPE_Y> yGm;
TYPE_ALPHA alphaValue;
float alphaValueFp;
uint32_t coreDataNum;
uint32_t tileNum;
uint32_t tileDataNum;
uint32_t tailDataNum;
uint32_t processDataNum;
};
template <typename TYPE_X1, typename TYPE_X2, typename TYPE_ALPHA, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void AxpyV2<TYPE_X1, TYPE_X2, TYPE_ALPHA, TYPE_Y, BUFFER_NUM>::Init(GM_ADDR x1, GM_ADDR x2, GM_ADDR alpha, GM_ADDR y, const AxpyV2TilingData* 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_X1 *)x1 + globalBufferIndex, this->coreDataNum);
x2Gm.SetGlobalBuffer((__gm__ TYPE_X2 *)x2 + globalBufferIndex, this->coreDataNum);
alphaGm.SetGlobalBuffer((__gm__ TYPE_ALPHA *)alpha);
alphaValue = alphaGm.GetValue(0);
yGm.SetGlobalBuffer((__gm__ TYPE_Y *)y + globalBufferIndex, this->coreDataNum);
pipe->InitBuffer(queBind, BUFFER_NUM, this->tileDataNum * sizeof(float) * 2);
pipe->InitBuffer(alphaBuf, this->tileDataNum * sizeof(float));
InitAlphaLocal();
}
template <typename TYPE_X1, typename TYPE_X2, typename TYPE_ALPHA, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void AxpyV2<TYPE_X1, TYPE_X2, TYPE_ALPHA, TYPE_Y, BUFFER_NUM>::InitAlphaLocal()
{
if constexpr ( IsSameType<TYPE_X1, int32_t>::value && IsSameType<TYPE_X2, int32_t>::value && IsSameType<TYPE_ALPHA, int32_t>::value )
{
alphaLocalInt = alphaBuf.Get<int32_t>();
Duplicate(alphaLocalInt, alphaValue, this->tileDataNum);
}
else
{
alphaLocalFp = alphaBuf.Get<float>();
if constexpr ( IsSameType<TYPE_ALPHA, bfloat16_t>::value )
{
LocalTensor<bfloat16_t> tmp = alphaLocalFp.template ReinterpretCast<bfloat16_t>();
Duplicate(tmp[this->tileDataNum], alphaValue, this->tileDataNum);
Cast(alphaLocalFp, tmp[this->tileDataNum], RoundMode::CAST_NONE, this->tileDataNum);
}
else if constexpr ( IsSameType<TYPE_ALPHA, half>::value )
{
LocalTensor<half> tmp = alphaLocalFp.template ReinterpretCast<half>();
Duplicate(tmp[this->tileDataNum], alphaValue, this->tileDataNum);
Cast(alphaLocalFp, tmp[this->tileDataNum], RoundMode::CAST_NONE, this->tileDataNum);
}
else if constexpr ( IsSameType<TYPE_ALPHA, float>::value )
{
Duplicate(alphaLocalFp, alphaValue, this->tileDataNum);
}
else if constexpr ( IsSameType<TYPE_ALPHA, int32_t>::value )
{
LocalTensor<int32_t> tmp = alphaLocalFp.template ReinterpretCast<int32_t>();
Duplicate(tmp, alphaValue, this->tileDataNum);
Cast(alphaLocalFp, tmp, RoundMode::CAST_NONE, this->tileDataNum);
}
}
}
template <typename TYPE_X1, typename TYPE_X2, typename TYPE_ALPHA, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void AxpyV2<TYPE_X1, TYPE_X2, TYPE_ALPHA, TYPE_Y, BUFFER_NUM>::TotalStage(int32_t offset)
{
LocalTensor<float> x1LocalFp = queBind.template AllocTensor<float>();
LocalTensor<float> x2LocalFp = x1LocalFp[this->processDataNum];
LocalTensor<TYPE_X1> x1Local = x1LocalFp.template ReinterpretCast<TYPE_X1>();
LocalTensor<TYPE_X2> x2Local = x2LocalFp.template ReinterpretCast<TYPE_X2>();
LocalTensor<TYPE_Y> yLocal = x1LocalFp.template ReinterpretCast<TYPE_Y>();
int32_t eventIDMTE2ToV = static_cast<int32_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE2_V));
int32_t eventIDVToMTE3 = static_cast<int32_t>(GetTPipePtr()->FetchEventID(HardEvent::V_MTE3));
if constexpr ( IsSameType<TYPE_X1, int32_t>::value && IsSameType<TYPE_X2, int32_t>::value && IsSameType<TYPE_ALPHA, int32_t>::value )
{
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Mul(x2Local, x2Local, alphaLocalInt, this->processDataNum);
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Add(x1Local, x1Local, x2Local, this->processDataNum);
if constexpr ( IsSameType<TYPE_Y, float>::value)
{
Cast(yLocal, x1Local, RoundMode::CAST_RINT, this->processDataNum);
}
if constexpr ( IsSameType<TYPE_Y, half>::value || IsSameType<TYPE_Y, bfloat16_t>::value)
{
Cast(x2LocalFp, x1Local, RoundMode::CAST_NONE, this->processDataNum);
Cast(yLocal, x2LocalFp, RoundMode::CAST_RINT, this->processDataNum);
}
}
else
{
if constexpr ( IsSameType<TYPE_X1, int32_t>::value && IsSameType<TYPE_X2, int32_t>::value )
{
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);
}
if constexpr ( IsSameType<TYPE_X1, float>::value && IsSameType<TYPE_X2, float>::value )
{
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
}
if constexpr ( (IsSameType<TYPE_X1, half>::value || IsSameType<TYPE_X1, bfloat16_t>::value) &&
(IsSameType<TYPE_X2, half>::value || IsSameType<TYPE_X2, bfloat16_t>::value ) )
{
LocalTensor<TYPE_X1> x1LocalLast = x1Local[this->processDataNum];
LocalTensor<TYPE_X2> x2LocalLast = x2Local[this->processDataNum];
DataCopy(x1LocalLast, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1LocalLast, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x2LocalLast, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2LocalLast, RoundMode::CAST_NONE, this->processDataNum);
}
if constexpr ( (IsSameType<TYPE_X1, half>::value || IsSameType<TYPE_X1, bfloat16_t>::value) &&
IsSameType<TYPE_X2, int32_t>::value )
{
LocalTensor<TYPE_X1> x1LocalLast = x1Local[this->processDataNum];
DataCopy(x1LocalLast, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1LocalLast, 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);
}
if constexpr ( (IsSameType<TYPE_X1, half>::value || IsSameType<TYPE_X1, bfloat16_t>::value) &&
IsSameType<TYPE_X2, float>::value )
{
LocalTensor<TYPE_X1> x1LocalLast = x1Local[this->processDataNum];
DataCopy(x1LocalLast, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1LocalLast, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
}
if constexpr ( IsSameType<TYPE_X1, float>::value &&
(IsSameType<TYPE_X2, half>::value || IsSameType<TYPE_X2, bfloat16_t>::value) )
{
LocalTensor<TYPE_X2> x2LocalLast = x2Local[this->processDataNum];
DataCopy(x2LocalLast, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2LocalLast, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
}
if constexpr ( IsSameType<TYPE_X1, int32_t>::value &&
(IsSameType<TYPE_X2, half>::value || IsSameType<TYPE_X2, bfloat16_t>::value) )
{
LocalTensor<TYPE_X2> x2LocalLast = x2Local[this->processDataNum];
DataCopy(x2LocalLast, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2LocalLast, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x1LocalFp, x1Local, RoundMode::CAST_NONE, this->processDataNum);
}
if constexpr ( IsSameType<TYPE_X1, int32_t>::value && IsSameType<TYPE_X2, float>::value )
{
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);
}
if constexpr ( IsSameType<TYPE_X1, float>::value && IsSameType<TYPE_X2, int32_t>::value )
{
DataCopy(x2Local, x2Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
Cast(x2LocalFp, x2Local, RoundMode::CAST_NONE, this->processDataNum);
DataCopy(x1Local, x1Gm[offset], this->processDataNum);
SetFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
WaitFlag<HardEvent::MTE2_V>(eventIDMTE2ToV);
}
MulAddDst(x1LocalFp, x2LocalFp, alphaLocalFp, this->processDataNum);
if constexpr ( IsSameType<TYPE_Y, int32_t>::value)
{
Cast(yLocal, x1LocalFp, RoundMode::CAST_TRUNC, this->processDataNum);
}
if constexpr ( IsSameType<TYPE_Y, half>::value || IsSameType<TYPE_Y, bfloat16_t>::value)
{
Cast(yLocal, x1LocalFp, RoundMode::CAST_RINT, this->processDataNum);
}
}
SetFlag<HardEvent::V_MTE3>(eventIDVToMTE3);
WaitFlag<HardEvent::V_MTE3>(eventIDVToMTE3);
DataCopy(yGm[offset], yLocal, this->processDataNum);
queBind.template FreeTensor(x1LocalFp);
}
template <typename TYPE_X1, typename TYPE_X2, typename TYPE_ALPHA, typename TYPE_Y, uint64_t BUFFER_NUM>
__aicore__ inline void AxpyV2<TYPE_X1, TYPE_X2, TYPE_ALPHA, 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)
{
TotalStage(offset);
}
this->processDataNum = this->tailDataNum;
TotalStage(offset);
}
}
#endif
