* 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 softmax_grad_base_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/activation/softmax/softmax_grad_base_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"adv_api/activation/softmaxgrad.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_GRAD_BASE_IMPL_H__
#endif
#ifndef IMPL_ACTIVATION_SOFTMAX_SOFTMAX_GRAD_BASE_IMPL_H
#define IMPL_ACTIVATION_SOFTMAX_SOFTMAX_GRAD_BASE_IMPL_H
#if defined(__NPU_ARCH__) && \
(__NPU_ARCH__ == 3510 || __NPU_ARCH__ == 5102 || __NPU_ARCH__ == 3003 || __NPU_ARCH__ == 3113)
#include "regbase/3510/softmax_grad_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 3002
#include "regbase/v300/softmax_grad_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 2201
#include "membase/v220/softmax_grad_impl.h"
#elif defined(__NPU_ARCH__) && __NPU_ARCH__ == 2002
#include "membase/v200/softmax_grad_impl.h"
#endif
#ifdef ASCENDC_CPU_DEBUG
#include "../../api_check/kernel_check/activation/softmax/softmax_grad/softmax_grad_check.h"
#include "../../api_check/kernel_check/activation/softmax/softmax_grad_front/softmax_grad_front_check.h"
#endif
#include "../../api_check/kernel_api_check.h"
namespace AscendC {
template <typename T, bool isBasicBlock = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradFrontImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<float>& workLocal, const SoftMaxTiling& tiling, const SoftMaxShapeInfo& softmaxShapeInfo)
{
CHECK_FUNC_HIGHLEVEL_API(
SoftmaxGradFront, (T, isBasicBlock, isDataFormatNZ),
(dstTensor, gradTensor, srcTensor, workLocal, tiling, softmaxShapeInfo));
ShapeInfo srcShape = srcTensor.GetShapeInfo();
uint32_t elementNumPerBlk = ONE_BLK_SIZE / sizeof(T);
LastAxisShapeND srcNDinfo;
LastAxisShapeND originalSrcShape;
if (softmaxShapeInfo.srcM == 0 || softmaxShapeInfo.srcK == 0) {
srcNDinfo = GetLastAxisShapeND(srcShape);
originalSrcShape = GetLastAxisOriginShapeND(srcShape);
} else {
srcNDinfo = {softmaxShapeInfo.srcM, softmaxShapeInfo.srcK};
originalSrcShape = {softmaxShapeInfo.oriSrcM, softmaxShapeInfo.oriSrcK};
}
if constexpr (isDataFormatNZ) {
if (unlikely(srcNDinfo.k != tiling.srcK || srcNDinfo.m != tiling.srcM)) {
SoftMaxTiling newTiling = tiling;
SoftMaxGradTilingFunc(workLocal.GetSize(), srcNDinfo, newTiling, elementNumPerBlk, true, false, true);
SoftmaxGradFrontNZImpl(dstTensor, gradTensor, srcTensor, workLocal, originalSrcShape, newTiling);
} else {
SoftmaxGradFrontNZImpl(dstTensor, gradTensor, srcTensor, workLocal, originalSrcShape, tiling);
}
} else {
if (unlikely(srcNDinfo.k != tiling.srcK || srcNDinfo.m != tiling.srcM)) {
SoftMaxTiling newTiling = tiling;
SoftMaxGradTilingFunc(workLocal.GetSize(), srcNDinfo, newTiling, elementNumPerBlk, true, isBasicBlock);
SoftmaxGradFrontNDImpl<T, isBasicBlock>(
dstTensor, gradTensor, srcTensor, workLocal, newTiling, originalSrcShape);
} else {
SoftmaxGradFrontNDImpl<T, isBasicBlock>(
dstTensor, gradTensor, srcTensor, workLocal, tiling, originalSrcShape);
}
}
}
template <typename T, bool isBasicBlock = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradFrontImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const SoftMaxTiling& tiling, const SoftMaxShapeInfo& softmaxShapeInfo)
{
LocalTensor<float> workLocal;
PopStackBuffer<float, TPosition::LCM>(workLocal);
SoftmaxGradFrontImpl<T, isBasicBlock, isDataFormatNZ>(
dstTensor, gradTensor, srcTensor, workLocal, tiling, softmaxShapeInfo);
}
template <typename T, bool isBasicBlock = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradFrontImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const SoftMaxTiling& tiling, const SoftMaxShapeInfo& softmaxShapeInfo)
{
auto workLocal = sharedTmpBuffer.ReinterpretCast<float>();
SoftmaxGradFrontImpl<T, isBasicBlock, isDataFormatNZ>(
dstTensor, gradTensor, srcTensor, workLocal, tiling, softmaxShapeInfo);
}
template <typename T, bool isReuseSource = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<float>& workLocal, const SoftMaxTiling& tiling, bool isFront,
const SoftMaxShapeInfo& softmaxShapeInfo)
{
CHECK_FUNC_HIGHLEVEL_API(
SoftmaxGrad, (T, isReuseSource, isDataFormatNZ),
(dstTensor, gradTensor, srcTensor, workLocal, tiling, isFront, softmaxShapeInfo));
ShapeInfo srcShape = srcTensor.GetShapeInfo();
uint32_t elementNumPerBlk = ONE_BLK_SIZE / sizeof(T);
LastAxisShapeND srcNDinfo;
LastAxisShapeND originalSrcShape;
if (softmaxShapeInfo.srcM == 0 || softmaxShapeInfo.srcK == 0) {
srcNDinfo = GetLastAxisShapeND(srcShape);
originalSrcShape = GetLastAxisOriginShapeND(srcShape);
} else {
srcNDinfo = {softmaxShapeInfo.srcM, softmaxShapeInfo.srcK};
originalSrcShape = {softmaxShapeInfo.oriSrcM, softmaxShapeInfo.oriSrcK};
}
if constexpr (isDataFormatNZ) {
if (unlikely(srcNDinfo.k != tiling.srcK || srcNDinfo.m != tiling.srcM)) {
SoftMaxTiling newTiling = tiling;
SoftMaxGradTilingFunc(workLocal.GetSize(), srcNDinfo, newTiling, elementNumPerBlk, isFront, false, true);
SoftmaxGradNZImpl(dstTensor, gradTensor, srcTensor, workLocal, originalSrcShape, newTiling, isFront);
} else {
SoftmaxGradNZImpl(dstTensor, gradTensor, srcTensor, workLocal, originalSrcShape, tiling, isFront);
}
} else {
if (unlikely(srcNDinfo.k != tiling.srcK || srcNDinfo.m != tiling.srcM)) {
SoftMaxTiling newTiling = tiling;
SoftMaxGradTilingFunc(workLocal.GetSize(), srcNDinfo, newTiling, elementNumPerBlk, isFront, false);
SoftmaxGradPostProcess<T>(
dstTensor, gradTensor, srcTensor, workLocal, newTiling, originalSrcShape, isFront);
} else {
SoftmaxGradPostProcess<T>(dstTensor, gradTensor, srcTensor, workLocal, tiling, originalSrcShape, isFront);
}
}
}
template <typename T, bool isReuseSource = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const SoftMaxTiling& tiling, bool isFront, const SoftMaxShapeInfo& softmaxShapeInfo)
{
LocalTensor<float> workLocal;
PopStackBuffer<float, TPosition::LCM>(workLocal);
SoftmaxGradImpl<T, isReuseSource, isDataFormatNZ>(
dstTensor, gradTensor, srcTensor, workLocal, tiling, isFront, softmaxShapeInfo);
}
template <typename T, bool isReuseSource = false, bool isDataFormatNZ = false>
__aicore__ inline void SoftmaxGradImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& gradTensor, const LocalTensor<T>& srcTensor,
const LocalTensor<uint8_t>& sharedTmpBuffer, const SoftMaxTiling& tiling, bool isFront,
const SoftMaxShapeInfo& softmaxShapeInfo)
{
auto workLocal = sharedTmpBuffer.ReinterpretCast<float>();
SoftmaxGradImpl<T, isReuseSource, isDataFormatNZ>(
dstTensor, gradTensor, srcTensor, workLocal, tiling, isFront, softmaxShapeInfo);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_GRAD_BASE_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SOFTMAX_GRAD_BASE_IMPL_H__
#endif
