* This file is part of the OpenBOAT project at Harbin Institute of Technology (HIT)
* and is contributed to the CANN Open Software.
*
* Copyright (c) 2025 AISS Group, Harbin Institute of Technology (HIT).
* All Rights Reserved.
*
* Authors (accounts):
* - Liu Jun <@kbryantttt>
* - Su Tonghua <@sutonghua>
*
* This program is free software: you can redistribute it and/or modify it.
* Licensed under the CANN Open Software License Agreement Version 2.0 (the "License").
* You may not use this file except in compliance with the License.
* See the LICENSE file at the root of the repository for the full text of the License.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTIES OF ANY KIND, EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
*/
* \file onehot.h
* \brief
*/
#ifndef __ONEHOT_H__
#define __ONEHOT_H__
#include "kernel_operator.h"
#include "kernel_tiling/kernel_tiling.h"
#include "onehot_tiling_data.h"
#include "onehot_tiling_key.h"
namespace NsOnehot {
using namespace AscendC;
constexpr int32_t BUFFER_NUM = 2;
template <typename T>
class Onehot {
public:
__aicore__ inline Onehot(){};
__aicore__ inline void Init(GM_ADDR x, GM_ADDR z, const OnehotTilingData* tilingData);
__aicore__ inline void Process();
private:
__aicore__ inline void CopyIn(int32_t progress,int32_t j);
__aicore__ inline void CopyOut(int32_t progress,int32_t j);
__aicore__ inline void Compute(int32_t progress,int32_t j);
private:
AscendC::TPipe pipe;
AscendC::TQue<AscendC::QuePosition::VECIN, BUFFER_NUM> inQueueX;
AscendC::TQue<AscendC::QuePosition::VECOUT, BUFFER_NUM> outQueueZ;
AscendC::GlobalTensor<T> xGm;
AscendC::GlobalTensor<T> zGm;
AscendC::TBuf<AscendC::QuePosition::VECCALC> cmpBuf;
AscendC::TBuf<AscendC::QuePosition::VECCALC> oneBuf;
AscendC::TBuf<AscendC::QuePosition::VECCALC> zBuf;
AscendC::TBuf<AscendC::QuePosition::VECCALC> dBuf;
uint32_t coreDataNum;
uint32_t tileNum;
uint32_t tileDataNum;
uint32_t tailDataNum;
uint32_t processDataNum;
int32_t depth;
int32_t depthAlign;
uint64_t maskCmp;
int repeat = 1;
AscendC::UnaryRepeatParams repeatParams = { 1, 1, 8, 8 };
};
template <typename T>
__aicore__ inline void Onehot<T>::Init(GM_ADDR x, GM_ADDR z, const OnehotTilingData* tilingData)
{
ASSERT(AscendC::GetBlockNum() != 0 && "block dim can not be zero!");
uint32_t coreNum = AscendC::GetBlockIdx();
uint32_t globalBufferIndex = tilingData->bigCoreDataNum * AscendC::GetBlockIdx();
this->tileDataNum = tilingData->tileDataNum;
this->depth = tilingData->depth;
this->depthAlign = ((this->depth + 8 - 1) / 8) * 8;
this->maskCmp = depthAlign;
this->processDataNum = this->tileDataNum;
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) * (AscendC::GetBlockIdx() - tilingData->tailBlockNum);
}
xGm.SetGlobalBuffer((__gm__ T*)x + globalBufferIndex, this->coreDataNum);
zGm.SetGlobalBuffer((__gm__ T*)z + globalBufferIndex*this->depth, this->coreDataNum*this->depth);
pipe.InitBuffer(cmpBuf, 8*depthAlign * sizeof(uint8_t));
pipe.InitBuffer(oneBuf, depthAlign * sizeof(float));
pipe.InitBuffer(zBuf, depthAlign * sizeof(float));
pipe.InitBuffer(dBuf, depthAlign * sizeof(T));
pipe.InitBuffer(inQueueX, BUFFER_NUM, this->tileDataNum * sizeof(T));
pipe.InitBuffer(outQueueZ, BUFFER_NUM, depthAlign * sizeof(T));
}
template <typename T>
__aicore__ inline void Onehot<T>::CopyIn(int32_t progress,int32_t j)
{
AscendC::LocalTensor<T> xLocal = inQueueX.AllocTensor<T>();
AscendC::DataCopy(xLocal, xGm[progress * this->tileDataNum], this->processDataNum);
inQueueX.EnQue(xLocal);
}
template <typename T>
__aicore__ inline void Onehot<T>::CopyOut(int32_t progress,int32_t j)
{
AscendC::LocalTensor<T> zLocal = outQueueZ.DeQue<T>();
AscendC::DataCopyExtParams copyParams{static_cast<uint16_t>(1), static_cast<uint32_t>(this->depth * sizeof(T)), 0, 0, 0};
AscendC::DataCopyPad(zGm[(progress * this->tileDataNum+j)*this->depth], zLocal, copyParams);
outQueueZ.FreeTensor(zLocal);
}
template <typename T>
__aicore__ inline void Onehot<T>::Compute(int32_t progress,int32_t j)
{
AscendC::LocalTensor<T> xLocal = inQueueX.DeQue<T>();
AscendC::LocalTensor<T> zLocal = outQueueZ.AllocTensor<T>();
AscendC::LocalTensor<uint8_t> cmp = cmpBuf.AllocTensor<uint8_t>();
AscendC::LocalTensor<float> zfloat = zBuf.AllocTensor<float>();
AscendC::LocalTensor<float> oneLocal = oneBuf.AllocTensor<float>();
AscendC::LocalTensor<T> dLocal = dBuf.AllocTensor<T>();
AscendC::Duplicate(oneLocal, 1.0f, depthAlign);
AscendC::PipeBarrier<PIPE_V>();
for(int i=0;i<this->depth;i++){
dLocal.SetValue(i, i);
}
for(int i=this->depth;i<depthAlign;i++){
dLocal.SetValue(i, -1);
}
int32_t xVal = xLocal.GetValue(j);
AscendC::CompareScalar(cmp, dLocal, xVal, AscendC::CMPMODE::EQ, maskCmp, repeat, repeatParams);
AscendC::LocalTensor<uint16_t> mask_Dup = cmpBuf.AllocTensor<uint16_t>();
AscendC::Select(zfloat, mask_Dup[0], oneLocal, 0.0f, AscendC::SELMODE::VSEL_TENSOR_SCALAR_MODE, this->depth);
AscendC::Cast(zLocal, zfloat, AscendC::RoundMode::CAST_CEIL,this->depth);
outQueueZ.EnQue<T>(zLocal);
inQueueX.FreeTensor(xLocal);
cmpBuf.FreeTensor(cmp);
oneBuf.FreeTensor(oneLocal);
zBuf.FreeTensor(zfloat);
dBuf.FreeTensor(dLocal);
}
template <typename T>
__aicore__ inline void Onehot<T>::Process()
{
int32_t loopCount = this->tileNum;
for (int32_t i = 0; i < loopCount; i++) {
if (i == this->tileNum - 1) {
this->processDataNum = this->tailDataNum;
}
for (int32_t j = 0; j<this->processDataNum; j++) {
CopyIn(i,j);
Compute(i,j);
CopyOut(i,j);
}
}
}
}
#endif