* Copyright (c) 2026 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.
*/
* NOTE: Portions of this code were AI-generated and have been
* technically reviewed for functional accuracy and security
*/
* \file inv.h
* \brief Inv kernel class definition (arch35)
*
* Inv(x) = 1 / x
*
* Precision strategy:
* - float32: Div(1.0, x) directly (no Reciprocal -- precision requirement)
* - float16/bfloat16: Cast to float32 -> Div(1.0, x) -> Cast back
*
* Template parameter T (mapped by TilingKey):
* - float: TilingKey 0: direct computation in float32
* - half: TilingKey 1: cast to fp32 -> Div -> cast back to fp16
* - bfloat16_t: TilingKey 2: cast to fp32 -> Div -> cast back to bf16
*
* Buffer layout (single buffer):
* inputQueue(1 buf): ubFactor * sizeof(T)
* outputQueue(1 buf): ubFactor * sizeof(T)
* tmpBuf1_: ubFactor * sizeof(float) -- xFloat32 intermediate
* tmpBuf2_: ubFactor * sizeof(float) -- ones vector
*/
#ifndef INV_H
#define INV_H
#include "kernel_operator.h"
#include "kernel_tiling/kernel_tiling.h"
#include "inv_tiling_data.h"
#include "inv_tiling_key.h"
namespace NsInv {
using AscendC::TPipe;
using AscendC::TQue;
using AscendC::TBuf;
using AscendC::QuePosition;
using AscendC::GlobalTensor;
using AscendC::LocalTensor;
using AscendC::DataCopyParams;
using AscendC::DataCopyPad;
using AscendC::DataCopyPadParams;
using AscendC::RoundMode;
using AscendC::GetBlockIdx;
using AscendC::Cast;
using AscendC::Duplicate;
using AscendC::Div;
template <typename T>
class Inv {
public:
__aicore__ inline Inv() {}
__aicore__ inline void Init(GM_ADDR self, GM_ADDR out, const InvTilingData* tilingData);
__aicore__ inline void Process();
private:
__aicore__ inline void CopyIn(int64_t gmOffset, int64_t currentNum);
__aicore__ inline void Compute(int64_t currentNum);
__aicore__ inline void CopyOut(int64_t gmOffset, int64_t currentNum);
__aicore__ inline void ComputeFloat32(LocalTensor<float>& xLocal,
LocalTensor<float>& yLocal,
int64_t alignedNum);
template <typename SrcT>
__aicore__ inline void ComputeWithCast(LocalTensor<SrcT>& xLocal,
LocalTensor<SrcT>& yLocal,
int64_t currentNum,
int64_t alignedNum);
private:
TPipe pipe;
TQue<QuePosition::VECIN, 1> inputQueue;
TQue<QuePosition::VECOUT, 1> outputQueue;
TBuf<QuePosition::VECCALC> tmpBuf1_;
TBuf<QuePosition::VECCALC> tmpBuf2_;
GlobalTensor<T> selfGM_;
GlobalTensor<T> outGM_;
int64_t blockOffset_ = 0;
int64_t blockLen_ = 0;
int64_t ubFactor_ = 0;
};
template <typename T>
__aicore__ inline void Inv<T>::Init(GM_ADDR self, GM_ADDR out, const InvTilingData* tilingData)
{
ubFactor_ = tilingData->ubFactor;
if (tilingData->totalElements == 0 || tilingData->blockFactor == 0) {
blockOffset_ = 0;
blockLen_ = 0;
return;
}
blockOffset_ = tilingData->blockFactor * static_cast<int64_t>(GetBlockIdx());
int64_t remaining = tilingData->totalElements - blockOffset_;
if (remaining <= 0) {
blockLen_ = 0;
return;
}
blockLen_ = (remaining > tilingData->blockFactor) ? tilingData->blockFactor : remaining;
selfGM_.SetGlobalBuffer((__gm__ T*)self + blockOffset_, blockLen_);
outGM_.SetGlobalBuffer((__gm__ T*)out + blockOffset_, blockLen_);
pipe.InitBuffer(inputQueue, 1, ubFactor_ * sizeof(T));
pipe.InitBuffer(outputQueue, 1, ubFactor_ * sizeof(T));
pipe.InitBuffer(tmpBuf1_, ubFactor_ * sizeof(float));
pipe.InitBuffer(tmpBuf2_, ubFactor_ * sizeof(float));
}
template <typename T>
__aicore__ inline void Inv<T>::CopyIn(int64_t gmOffset, int64_t currentNum)
{
LocalTensor<T> xLocal = inputQueue.template AllocTensor<T>();
DataCopyParams copyParams;
copyParams.blockCount = 1;
copyParams.blockLen = currentNum * sizeof(T);
copyParams.srcStride = 0;
copyParams.dstStride = 0;
DataCopyPad(xLocal, selfGM_[gmOffset], copyParams, {false, 0, 0, 0});
inputQueue.EnQue(xLocal);
}
template <typename T>
__aicore__ inline void Inv<T>::CopyOut(int64_t gmOffset, int64_t currentNum)
{
LocalTensor<T> yLocal = outputQueue.template DeQue<T>();
DataCopyParams copyParams;
copyParams.blockCount = 1;
copyParams.blockLen = currentNum * sizeof(T);
copyParams.srcStride = 0;
copyParams.dstStride = 0;
DataCopyPad(outGM_[gmOffset], yLocal, copyParams);
outputQueue.FreeTensor(yLocal);
}
template <typename T>
__aicore__ inline void Inv<T>::ComputeFloat32(LocalTensor<float>& xLocal,
LocalTensor<float>& yLocal,
int64_t alignedNum)
{
LocalTensor<float> ones = tmpBuf2_.template Get<float>();
Duplicate(ones, 1.0f, static_cast<int32_t>(alignedNum));
Div(yLocal, ones, xLocal, static_cast<int32_t>(alignedNum));
}
template <typename T>
template <typename SrcT>
__aicore__ inline void Inv<T>::ComputeWithCast(LocalTensor<SrcT>& xLocal,
LocalTensor<SrcT>& yLocal,
int64_t currentNum,
int64_t alignedNum)
{
LocalTensor<float> xFloat = tmpBuf1_.template Get<float>();
LocalTensor<float> ones = tmpBuf2_.template Get<float>();
Cast(xFloat, xLocal, RoundMode::CAST_NONE, static_cast<uint32_t>(alignedNum));
Duplicate(ones, 1.0f, static_cast<int32_t>(alignedNum));
Div(xFloat, ones, xFloat, static_cast<int32_t>(alignedNum));
Cast(yLocal, xFloat, RoundMode::CAST_ROUND, static_cast<uint32_t>(alignedNum));
}
template <typename T>
__aicore__ inline void Inv<T>::Compute(int64_t currentNum)
{
LocalTensor<T> xLocal = inputQueue.template DeQue<T>();
LocalTensor<T> yLocal = outputQueue.template AllocTensor<T>();
constexpr int64_t floatBlock = 32 / sizeof(float);
constexpr int64_t typeBlock = 32 / sizeof(T);
constexpr int64_t alignBlock = (floatBlock > typeBlock) ? floatBlock : typeBlock;
int64_t alignedNum = ((currentNum + alignBlock - 1) / alignBlock) * alignBlock;
if constexpr (std::is_same_v<T, float>) {
ComputeFloat32(xLocal, yLocal, alignedNum);
} else {
ComputeWithCast(xLocal, yLocal, currentNum, alignedNum);
}
outputQueue.template EnQue<T>(yLocal);
inputQueue.FreeTensor(xLocal);
}
template <typename T>
__aicore__ inline void Inv<T>::Process()
{
if (blockLen_ <= 0) {
return;
}
int64_t loopCount = (blockLen_ + ubFactor_ - 1) / ubFactor_;
for (int64_t i = 0; i < loopCount; i++) {
int64_t gmOffset = i * ubFactor_;
int64_t currentNum = (i == (loopCount - 1)) ? (blockLen_ - gmOffset) : ubFactor_;
CopyIn(gmOffset, currentNum);
Compute(currentNum);
CopyOut(gmOffset, currentNum);
}
}
}
#endif