* 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 reduce_tensor_empty.h
* \brief reduce_tensor_empty kernel
*/
#ifndef REDUCE_TENSOR_EMPTY_H
#define REDUCE_TENSOR_EMPTY_H
namespace Ops {
namespace Base {
namespace ReduceOpTmpl {
using namespace Ops::Base;
template <typename T>
class ReduceTensorEmpty {
public:
template <class... Args>
__aicore__ inline explicit ReduceTensorEmpty(const ReduceOpTilingData* tilingData, GlobalTensor<uint8_t>* output,
TPipe* pipeIn, Args... args)
{
tiling_ = tilingData;
bufferSize_ = tiling_->basicBlock == 0 ? bufferSize_ : tiling_->basicBlock;
output_ = output;
pipeIn->InitBuffer(dupTempBuf_, UB_SIZE);
pipeIn->InitBuffer(vecQue_, BUFFER_NUM, bufferSize_);
InitValue(args...);
};
__aicore__ inline void Process();
private:
__aicore__ inline void ProcessPerCore();
template <class T1, class... Args>
__aicore__ inline void InitValue(T1 value, Args... args)
{
constexpr uint32_t UB_SIZE = Ops::Base::GetUbBlockSize();
if constexpr (IsSameType<T, T1>::value) {
value_ = value;
} else {
auto tempBuf = dupTempBuf_.Get<T1>();
Duplicate<T1>(tempBuf, value, UB_SIZE / sizeof(T1));
auto tempCastBuf = tempBuf.template ReinterpretCast<T>();
if constexpr (IsSameType<T, bfloat16_t>::value) {
Cast(tempCastBuf, tempBuf, RoundMode::CAST_RINT, 1);
} else {
Cast(tempCastBuf, tempBuf, RoundMode::CAST_NONE, 1);
}
event_t eventIdVToS = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::V_S));
SetFlag<HardEvent::V_S>(eventIdVToS);
WaitFlag<HardEvent::V_S>(eventIdVToS);
value_ = tempCastBuf.GetValue(0);
}
}
__aicore__ inline void InitValue() {};
private:
constexpr static uint32_t UB_SIZE = Ops::Base::GetUbBlockSize();
constexpr static int32_t BUFFER_NUM = 2;
const ReduceOpTilingData* tiling_;
GlobalTensor<uint8_t>* output_;
TBuf<TPosition::VECCALC> dupTempBuf_;
TQueBind<QuePosition::VECIN, QuePosition::VECOUT, BUFFER_NUM> vecQue_;
int64_t bufferSize_ = 64 * 1024;
uint64_t loopStartIdx_ = 0;
uint64_t loopEndIdx_ = 0;
T value_ = 0;
DataCopyExtParams copyOutParams_{1, 0, 0, 0, 0};
};
template <typename T>
__aicore__ inline void ReduceTensorEmpty<T>::Process()
{
if (tiling_->outSize <= 0) {
return;
}
int64_t blockIdx = GetBlockIdx();
loopStartIdx_ = blockIdx * tiling_->factorACntPerCore;
loopEndIdx_ = loopStartIdx_ + tiling_->factorACntPerCore;
if (unlikely(loopEndIdx_ > tiling_->factorATotalCnt)) {
loopEndIdx_ = tiling_->factorATotalCnt;
}
ProcessPerCore();
}
template <typename T>
__aicore__ inline void ReduceTensorEmpty<T>::ProcessPerCore()
{
int64_t copyElementNum = tiling_->ubFactorA;
for (uint64_t loopIdx = loopStartIdx_; loopIdx < loopEndIdx_; loopIdx++) {
if (loopIdx == tiling_->factorATotalCnt - 1) {
copyElementNum = tiling_->outSize - (tiling_->factorATotalCnt - 1) * tiling_->ubFactorA;
}
auto bindLocalIn = vecQue_.AllocTensor<T>();
copyOutParams_.blockLen = copyElementNum * sizeof(T);
Duplicate<T>(bindLocalIn, value_, copyElementNum);
vecQue_.EnQue<QuePosition::VECCALC, QuePosition::VECOUT>(bindLocalIn);
auto bindLocalOut = vecQue_.DeQue<QuePosition::VECCALC, QuePosition::VECOUT, T>();
GlobalTensor<T> globalTensor;
globalTensor.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(output_[0].GetPhyAddr(0)));
DataCopyPad(globalTensor[loopIdx * tiling_->ubFactorA], bindLocalOut, copyOutParams_);
vecQue_.FreeTensor(bindLocalOut);
}
}
}
}
}
#endif