* 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 pows_bf16.h
* \brief
*/
#ifndef Pows_BFLOAT16_ALIGN_H
#define Pows_BFLOAT16_ALIGN_H
#include "pows_base.h"
namespace Pows {
using namespace AscendC;
template <typename T>
class PowsBfp16 : public PowsBase<T>
{
public:
__aicore__ inline PowsBfp16(){};
__aicore__ inline void Init(GM_ADDR x1, GM_ADDR x2, GM_ADDR y, GM_ADDR workspace, const PowsTilingData* tilingData);
__aicore__ inline void Process();
constexpr static int32_t bufferNum = 2;
private:
__aicore__ inline void CopyInX(const int64_t& index, const int64_t& dataLength);
__aicore__ inline void ComputePows(const int64_t& dataLength);
__aicore__ inline void CopyOut(const int64_t& index, const int64_t& dataLength);
__aicore__ inline void ProcessPerCore();
__aicore__ inline void ProcessLastCore();
private:
TPipe pipe;
TQue<QuePosition::VECIN, bufferNum> inQueueX1;
TQue<QuePosition::VECOUT, 1> outQueue;
TBuf<QuePosition::VECCALC> tempBuf1;
TBuf<QuePosition::VECCALC> tempBuf2;
};
template <typename T>
__aicore__ inline void PowsBfp16<T>::Init(
GM_ADDR x, GM_ADDR y, GM_ADDR gelu, GM_ADDR workspace, const PowsTilingData* tilingData)
{
this->BaseInit(x, y, gelu, tilingData);
*(reinterpret_cast<T*>(&(this->scalar)) + 1) = this->x2Gm.GetValue(0);
pipe.InitBuffer(inQueueX1, bufferNum, this->m_tilingData.bufSize * sizeof(T));
pipe.InitBuffer(outQueue, 1, this->m_tilingData.bufSize * sizeof(T));
pipe.InitBuffer(tempBuf1, this->m_tilingData.bufSize * sizeof(float));
pipe.InitBuffer(tempBuf2, this->m_tilingData.bufSize * sizeof(float));
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::Process()
{
if (this->blockIdx >= this->m_tilingData.realCoreNum) {
return;
}
if (this->isLastCore) {
ProcessLastCore();
} else {
ProcessPerCore();
}
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::ProcessPerCore()
{
for (int64_t idx = 0; idx < this->m_tilingData.mainCoreLoopNum; idx++) {
CopyInX(idx, this->m_tilingData.dataLength);
ComputePows(this->m_tilingData.dataLength);
CopyOut(idx, this->m_tilingData.dataLength);
}
if (this->m_tilingData.mainCoreTailLength > 0) {
CopyInX(this->m_tilingData.mainCoreLoopNum, this->m_tilingData.mainCoreTailLength);
ComputePows(this->m_tilingData.mainCoreTailLength);
CopyOut(this->m_tilingData.mainCoreLoopNum, this->m_tilingData.mainCoreTailLength);
}
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::ProcessLastCore()
{
for (int64_t idx = 0; idx < this->m_tilingData.tailCoreLoopNum; idx++) {
CopyInX(idx, this->m_tilingData.dataLength);
ComputePows(this->m_tilingData.dataLength);
CopyOut(idx, this->m_tilingData.dataLength);
}
if (this->m_tilingData.tailCoreTailLength > 0) {
CopyInX(this->m_tilingData.tailCoreLoopNum, this->m_tilingData.tailCoreTailLength);
ComputePows(this->m_tilingData.tailCoreTailLength);
CopyOut(this->m_tilingData.tailCoreLoopNum, this->m_tilingData.tailCoreTailLength);
}
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::CopyInX(const int64_t& index, const int64_t& dataLength)
{
LocalTensor<T> ubX1 = inQueueX1.AllocTensor<T>();
this->CopyInXBase(index, dataLength, ubX1);
inQueueX1.EnQue(ubX1);
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::ComputePows(const int64_t& dataLength)
{
LocalTensor<T> ubX1 = inQueueX1.DeQue<T>();
LocalTensor<float> ubx_fp32 = tempBuf1.Get<float>();
Cast(ubx_fp32, ubX1, RoundMode::CAST_NONE, dataLength);
inQueueX1.FreeTensor(ubX1);
LocalTensor<float> result_fp32 = tempBuf2.Get<float>();
this->ComputePowsBase(ubx_fp32, result_fp32, dataLength);
LocalTensor<T> out = outQueue.AllocTensor<T>();
Cast(out, result_fp32, RoundMode::CAST_RINT, dataLength);
outQueue.EnQue(out);
}
template <typename T>
__aicore__ inline void PowsBfp16<T>::CopyOut(const int64_t& index, const int64_t& dataLength)
{
LocalTensor<T> outLocal = outQueue.DeQue<T>();
this->CopyOutBase(index, dataLength, outLocal);
outQueue.FreeTensor(outLocal);
}
}
#endif