* 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.
*/
#include "aclnn_glu_backward.h"
#include "glu_grad.h"
#include "sigmoid.h"
#include "level0/split_v.h"
#include "level0/mul.h"
#include "level0/sub.h"
#include "level0/concat.h"
#include "aclnn_kernels/cast.h"
#include "aclnn_kernels/contiguous.h"
#include "opdev/common_types.h"
#include "opdev/data_type_utils.h"
#include "opdev/shape_utils.h"
#include "opdev/format_utils.h"
#include "opdev/op_dfx.h"
#include "opdev/op_executor.h"
#include "opdev/op_log.h"
#include "opdev/tensor_view_utils.h"
#include "opdev/platform.h"
#include "aclnn/aclnn_base.h"
#include "aclnn_kernels/common/op_error_check.h"
#include "op_api/aclnn_util.h"
using namespace op;
#ifdef __cplusplus
extern "C" {
#endif
*
* A2路径 (SplitV + Sigmoid + Mul + Sub + ConcatD):
* self
* |
* gradOut Contiguous(workspace_8)
* \ |-----------------------|
* \ | dim \
* \ | / \ \
* Contiguous(workspace_0) SplitV(workspace_1) SplitV(workspace_1)
* \ / \ / \
* \ Sigmoid(workspace_2) / \
* \ / \ /--------/
* Mul(workspace_3) Mul(workspace_4) /
* | \ \ /
* | \ Sub(workspace_5)
* | \ /
* dim | Mul(workspace_6)
* \ | /
* \ | /
* ConcatD(workspace_7)
* |
* ViewCopy
* |
* out
*
* A5路径 (GluGrad单kernel):
* gradOut self
* | |
* Contiguous(workspace_0) Contiguous(workspace_1)
* \ /
* \ /
* GluGrad(workspace_2)
* |
* ViewCopy
* |
* out
*/
constexpr size_t MAX_DIM_LEN = 8;
constexpr int64_t SPLIT_NUM = 2;
static bool CheckNotNull(const aclTensor *gradOut, const aclTensor *self, const aclTensor *out) {
OP_CHECK_NULL(self, return false);
OP_CHECK_NULL(gradOut, return false);
OP_CHECK_NULL(out, return false);
return true;
}
static const std::initializer_list<op::DataType> ASCEND910_DTYPE_SUPPORT_LIST = {
op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_DOUBLE};
static const std::initializer_list<op::DataType> ASCEND910B_DTYPE_SUPPORT_LIST = {
op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_DOUBLE, op::DataType::DT_BF16};
static const std::initializer_list<op::DataType> ASCEND950_DTYPE_SUPPORT_LIST = {
op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_BF16, op::DataType::DT_DOUBLE};
static const std::initializer_list<DataType>& GetDtypeSupportList() {
auto curArch = GetCurrentPlatformInfo().GetCurNpuArch();
if (Ops::NN::AclnnUtil::IsRegbase(curArch)) {
return ASCEND950_DTYPE_SUPPORT_LIST;
} else if (GetCurrentPlatformInfo().GetSocVersion() >= SocVersion::ASCEND910B &&
GetCurrentPlatformInfo().GetSocVersion() <= SocVersion::ASCEND910E) {
return ASCEND910B_DTYPE_SUPPORT_LIST;
} else {
return ASCEND910_DTYPE_SUPPORT_LIST;
}
}
static bool CheckDtypeValid(const aclTensor *gradOut, const aclTensor *self, const aclTensor *out) {
const auto& supportList = GetDtypeSupportList();
OP_CHECK_DTYPE_NOT_SUPPORT(self, supportList, return false);
OP_CHECK_DTYPE_NOT_MATCH(gradOut, self->GetDataType(), return false);
OP_CHECK_DTYPE_NOT_MATCH(out, self->GetDataType(), return false);
return true;
}
static bool CheckParamsDataAndShape(const aclTensor *gradOut, const aclTensor *self,
int64_t dim, const aclTensor *out) {
OP_CHECK_MAX_DIM(self, MAX_DIM_LEN, return false);
OP_CHECK_MAX_DIM(out, MAX_DIM_LEN, return false);
OP_CHECK_MAX_DIM(gradOut, MAX_DIM_LEN, return false);
OP_CHECK_MIN_DIM(self, 1, return false);
int64_t selfDim = static_cast<int64_t>(self->GetViewShape().GetDimNum());
if (dim < -selfDim || dim >= selfDim) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Dimension out of range (expected to be in range of [%ld, %ld], but got %ld).",
-selfDim, (selfDim - 1), dim);
return false;
}
int64_t positiveDim = dim;
if (dim < 0) {
positiveDim += selfDim;
}
int64_t splitShape = self->GetViewShape().GetDim(positiveDim);
if (splitShape % SPLIT_NUM != 0) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Halving dimension must be even, but dimension %ld is size %ld.", dim, splitShape);
return false;
}
op::Shape gradOutShapeExpect = self->GetViewShape();
int64_t gradOutShapeExpectForDim = gradOutShapeExpect.GetDim(static_cast<size_t>(positiveDim)) / SPLIT_NUM;
gradOutShapeExpect.SetDim(static_cast<size_t>(positiveDim), gradOutShapeExpectForDim);
if (gradOutShapeExpect != gradOut->GetViewShape()) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "The shape of gradOut must be %s, but got %s.",
op::ToString(gradOutShapeExpect).GetString(), op::ToString(gradOut->GetViewShape()).GetString());
return false;
}
OP_CHECK_SHAPE_NOT_EQUAL(out, self, return false);
return true;
}
static aclnnStatus CheckParams(const aclTensor *gradOut, const aclTensor *self, int64_t dim, const aclTensor *out) {
CHECK_RET(CheckNotNull(gradOut, self, out), ACLNN_ERR_PARAM_NULLPTR);
CHECK_RET(CheckDtypeValid(gradOut, self, out), ACLNN_ERR_PARAM_INVALID);
CHECK_RET(CheckParamsDataAndShape(gradOut, self, dim, out), ACLNN_ERR_PARAM_INVALID);
return ACLNN_SUCCESS;
}
static bool IsRegbaseArch() {
auto curArch = GetCurrentPlatformInfo().GetCurNpuArch();
return Ops::NN::AclnnUtil::IsRegbase(curArch);
}
static inline aclIntArray *getDimAndSplitSize(const aclTensor *self, int64_t &positiveDim, int64_t dim,
aclOpExecutor *executor) {
int64_t selfDim = static_cast<int64_t>(self->GetViewShape().GetDimNum());
if (dim < 0) {
positiveDim += selfDim;
}
int64_t splitShape = self->GetViewShape().GetDim(positiveDim) / SPLIT_NUM;
int64_t splitSizeValue[] = {splitShape, splitShape};
return executor->AllocIntArray(splitSizeValue, SPLIT_NUM);
}
static aclnnStatus PrepareInputs(const aclTensor *self, const aclTensor *gradOut, bool needCast,
const aclTensor **selfContiguous, const aclTensor **gradOutContiguous,
aclOpExecutor *executor) {
*selfContiguous = l0op::Contiguous(self, executor);
CHECK_RET(*selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
*gradOutContiguous = l0op::Contiguous(gradOut, executor);
CHECK_RET(*gradOutContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
if (needCast) {
*selfContiguous = l0op::Cast(*selfContiguous, op::DataType::DT_FLOAT, executor);
CHECK_RET(*selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
*gradOutContiguous = l0op::Cast(*gradOutContiguous, op::DataType::DT_FLOAT, executor);
CHECK_RET(*gradOutContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
}
return ACLNN_SUCCESS;
}
static aclnnStatus ComputeGluBackward(const aclTensor *selfContiguous, const aclTensor *gradOutContiguous,
aclIntArray *splitSize, int64_t positiveDim,
const aclTensor **gradA, const aclTensor **gradB,
aclOpExecutor *executor) {
auto splitResult = l0op::SplitV(selfContiguous, splitSize, positiveDim, executor);
CHECK_RET(splitResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
if (splitResult->Size() != static_cast<size_t>(SPLIT_NUM)) {
OP_LOGE(ACLNN_ERR_INNER, "The result of SplitV must be equal 2, but get %zu.", splitResult->Size());
return ACLNN_ERR_INNER;
}
auto splitFirst = (*splitResult)[0];
auto splitSecond = (*splitResult)[1];
splitSecond = l0op::Sigmoid(splitSecond, executor);
CHECK_RET(splitSecond != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto aGrad = l0op::Mul(gradOutContiguous, splitSecond, executor);
CHECK_RET(aGrad != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto sigmoidBA = l0op::Mul(splitFirst, splitSecond, executor);
CHECK_RET(sigmoidBA != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto subResult = l0op::Sub(splitFirst, sigmoidBA, executor);
CHECK_RET(subResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
*gradB = l0op::Mul(subResult, aGrad, executor);
CHECK_RET(*gradB != nullptr, ACLNN_ERR_INNER_NULLPTR);
*gradA = aGrad;
return ACLNN_SUCCESS;
}
static aclnnStatus AssembleOutput(const aclTensor *gradA, const aclTensor *gradB, int64_t positiveDim,
bool needCast, const aclTensor *out, aclOpExecutor *executor) {
op::FVector<const aclTensor*> tensorListVector;
tensorListVector.emplace_back(gradA);
tensorListVector.emplace_back(gradB);
auto tensorList = executor->AllocTensorList(tensorListVector.data(), tensorListVector.size());
auto concatTensor = l0op::ConcatD(tensorList, positiveDim, executor);
CHECK_RET(concatTensor != nullptr, ACLNN_ERR_INNER_NULLPTR);
const aclTensor* gluBackwardOut = concatTensor;
if (needCast) {
gluBackwardOut = l0op::Cast(concatTensor, op::DataType::DT_FLOAT16, executor);
CHECK_RET(gluBackwardOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
}
auto viewCopyResult = l0op::ViewCopy(gluBackwardOut, out, executor);
CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
return ACLNN_SUCCESS;
}
static aclnnStatus ExecGluBackwardA2(const aclTensor *gradOut, const aclTensor *self, int64_t dim,
const aclTensor *out, aclOpExecutor *executor) {
int64_t positiveDim = dim;
auto splitSize = getDimAndSplitSize(self, positiveDim, dim, executor);
bool needCast = (self->GetDataType() == op::DataType::DT_FLOAT16);
const aclTensor *selfContiguous = nullptr;
const aclTensor *gradOutContiguous = nullptr;
auto status = PrepareInputs(self, gradOut, needCast, &selfContiguous, &gradOutContiguous, executor);
CHECK_RET(status == ACLNN_SUCCESS, status);
const aclTensor *gradA = nullptr;
const aclTensor *gradB = nullptr;
status = ComputeGluBackward(selfContiguous, gradOutContiguous, splitSize, positiveDim,
&gradA, &gradB, executor);
CHECK_RET(status == ACLNN_SUCCESS, status);
status = AssembleOutput(gradA, gradB, positiveDim, needCast, out, executor);
CHECK_RET(status == ACLNN_SUCCESS, status);
return ACLNN_SUCCESS;
}
static aclnnStatus ExecGluBackwardA5(const aclTensor *gradOut, const aclTensor *self, int64_t dim,
const aclTensor *out, aclOpExecutor *executor) {
auto selfContiguous = l0op::Contiguous(self, executor);
CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto gradOutContiguous = l0op::Contiguous(gradOut, executor);
CHECK_RET(gradOutContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto result = l0op::GluGrad(gradOutContiguous, selfContiguous, dim, executor);
CHECK_RET(result != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto viewCopyResult = l0op::ViewCopy(result, out, executor);
CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
return ACLNN_SUCCESS;
}
aclnnStatus aclnnGluBackwardGetWorkspaceSize(const aclTensor *gradOut, const aclTensor *self, int64_t dim,
const aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor) {
L2_DFX_PHASE_1(aclnnGluBackward, DFX_IN(gradOut, self, dim), DFX_OUT(out));
auto uniqueExecutor = CREATE_EXECUTOR();
CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
auto ret = CheckParams(gradOut, self, dim, out);
CHECK_RET(ret == ACLNN_SUCCESS, ret);
if (self->IsEmpty()) {
*workspaceSize = 0;
uniqueExecutor.ReleaseTo(executor);
return ACLNN_SUCCESS;
}
aclnnStatus status;
if (IsRegbaseArch() && gradOut->GetDataType() != op::DataType::DT_DOUBLE) {
status = ExecGluBackwardA5(gradOut, self, dim, out, uniqueExecutor.get());
} else {
status = ExecGluBackwardA2(gradOut, self, dim, out, uniqueExecutor.get());
}
CHECK_RET(status == ACLNN_SUCCESS, status);
*workspaceSize = uniqueExecutor->GetWorkspaceSize();
uniqueExecutor.ReleaseTo(executor);
return ACLNN_SUCCESS;
}
aclnnStatus aclnnGluBackward(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream) {
L2_DFX_PHASE_2(aclnnGluBackward);
return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
}
#ifdef __cplusplus
}
#endif
