* 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.
*/
* \file mhc_pre_backward_infershape.cpp
* \brief
*/
#include <map>
#include <string>
#include <sstream>
#include <initializer_list>
#include "exe_graph/runtime/infer_shape_context.h"
#include "exe_graph/runtime/shape.h"
#include "exe_graph/runtime/storage_shape.h"
#include "register/op_impl_registry.h"
#include "log/log.h"
#include "err/ops_err.h"
using namespace gert;
using namespace ge;
namespace ops {
const constexpr int64_t GRAD_H_IN_INDEX = 3;
const constexpr int64_t GRAD_H_POST_INDEX = 4;
const constexpr int64_t GRAD_H_RES_INDEX = 5;
const constexpr int64_t OUT_GRAD_X_INDEX = 0;
const constexpr int64_t OUT_GRAD_PHI_INDEX = 1;
const constexpr int64_t OUT_GRAD_ALPHA_INDEX = 2;
const constexpr int64_t OUT_GRAD_BIAS_INDEX = 3;
const constexpr int64_t OUT_GRAD_GAMMA_INDEX = 4;
const constexpr int64_t BSD_DIM_NUM = 3;
const constexpr int64_t BSNN_DIM_NUM = 4;
const constexpr int64_t TD_DIM_NUM = 2;
const constexpr int64_t TN_DIM_NUM = 2;
const constexpr int64_t TNN_DIM_NUM = 3;
const constexpr int64_t TND_DIM_NUM = 3;
const constexpr int64_t GRAD_ALPHA_DIM_SIZE = 3;
const constexpr int64_t INDEX_0 = 0;
const constexpr int64_t INDEX_1 = 1;
const constexpr int64_t INDEX_2 = 2;
const constexpr int64_t INDEX_T = 0;
const constexpr int64_t INDEX_N = 1;
const constexpr int64_t INDEX_D_TND = 1;
ge::graphStatus GetInputShapes(InferShapeContext *context, const gert::Shape *&gradHInShape,
const gert::Shape *&gradHPostShape, const gert::Shape *&gradHResShape)
{
gradHInShape = context->GetDynamicInputShape(GRAD_H_IN_INDEX, 0);
OP_CHECK_NULL_WITH_CONTEXT(context, gradHInShape);
gradHPostShape = context->GetDynamicInputShape(GRAD_H_POST_INDEX, 0);
OP_CHECK_NULL_WITH_CONTEXT(context, gradHPostShape);
gradHResShape = context->GetDynamicInputShape(GRAD_H_RES_INDEX, 0);
OP_CHECK_NULL_WITH_CONTEXT(context, gradHResShape);
return GRAPH_SUCCESS;
}
ge::graphStatus ValidateInputDims(int64_t gradHInDimNum, int64_t gradHPostDimNum, int64_t gradHResDimNum)
{
if ((gradHInDimNum != BSD_DIM_NUM && gradHInDimNum != TD_DIM_NUM) ||
(gradHPostDimNum != BSD_DIM_NUM && gradHPostDimNum != TN_DIM_NUM) ||
(gradHResDimNum != BSNN_DIM_NUM && gradHResDimNum != TNN_DIM_NUM)) {
return GRAPH_FAILED;
}
if (gradHInDimNum != gradHPostDimNum) {
return GRAPH_FAILED;
}
return GRAPH_SUCCESS;
}
ge::graphStatus InferBSDFormat(const gert::Shape *gradHInShape, const gert::Shape *gradHPostShape,
gert::Shape *gradXShape, uint64_t &numsResidual, uint64_t &dimen)
{
uint64_t batch = gradHInShape->GetDim(INDEX_0);
uint64_t sequence = gradHInShape->GetDim(INDEX_1);
dimen = gradHInShape->GetDim(INDEX_2);
numsResidual = gradHPostShape->GetDim(INDEX_2);
gradXShape->SetDimNum(BSNN_DIM_NUM);
gradXShape->SetDim(0, batch);
gradXShape->SetDim(1, sequence);
gradXShape->SetDim(2, numsResidual);
gradXShape->SetDim(3, dimen);
return GRAPH_SUCCESS;
}
ge::graphStatus InferTNDFormat(const gert::Shape *gradHInShape, const gert::Shape *gradHPostShape,
gert::Shape *gradXShape, uint64_t &numsResidual, uint64_t &dimen)
{
uint64_t t = gradHInShape->GetDim(INDEX_T);
dimen = gradHInShape->GetDim(INDEX_D_TND);
numsResidual = gradHPostShape->GetDim(INDEX_N);
gradXShape->SetDimNum(TND_DIM_NUM);
gradXShape->SetDim(0, t);
gradXShape->SetDim(1, numsResidual);
gradXShape->SetDim(2, dimen);
return GRAPH_SUCCESS;
}
ge::graphStatus InferOutputShapes(gert::Shape *gradPhiShape, gert::Shape *gradAlphaShape,
gert::Shape *gradBiasShape, gert::Shape *gradGammaShape, uint64_t numsResidual,
uint64_t dimen)
{
uint64_t n2Plus2n = (2 * numsResidual) + (numsResidual * numsResidual);
gradPhiShape->SetDimNum(2);
gradPhiShape->SetDim(0, n2Plus2n);
gradPhiShape->SetDim(1, numsResidual * dimen);
gradAlphaShape->SetDimNum(1);
gradAlphaShape->SetDim(0, GRAD_ALPHA_DIM_SIZE);
gradBiasShape->SetDimNum(1);
gradBiasShape->SetDim(0, n2Plus2n);
if (gradGammaShape != nullptr) {
gradGammaShape->SetDimNum(2);
gradGammaShape->SetDim(0, numsResidual);
gradGammaShape->SetDim(1, dimen);
}
return GRAPH_SUCCESS;
}
static ge::graphStatus InferShape4mHCPreGrad(InferShapeContext *context)
{
OP_LOGD(context->GetNodeName(), "Begin to do InferShape MhcPreBackward");
const gert::Shape *gradHInShape = nullptr;
const gert::Shape *gradHPostShape = nullptr;
const gert::Shape *gradHResShape = nullptr;
auto ret = GetInputShapes(context, gradHInShape, gradHPostShape, gradHResShape);
if (ret != GRAPH_SUCCESS) {
OP_LOGE(context->GetNodeName(), "input shapes are invalid");
return ret;
}
auto gradXShape = context->GetOutputShape(OUT_GRAD_X_INDEX);
auto gradPhiShape = context->GetOutputShape(OUT_GRAD_PHI_INDEX);
auto gradAlphaShape = context->GetOutputShape(OUT_GRAD_ALPHA_INDEX);
auto gradBiasShape = context->GetOutputShape(OUT_GRAD_BIAS_INDEX);
auto gradGammaShape = context->GetOutputShape(OUT_GRAD_GAMMA_INDEX);
auto gradHInDimNum = gradHInShape->GetDimNum();
auto gradHPostDimNum = gradHPostShape->GetDimNum();
auto gradHResDimNum = gradHResShape->GetDimNum();
ret = ValidateInputDims(gradHInDimNum, gradHPostDimNum, gradHResDimNum);
if (ret != GRAPH_SUCCESS) {
OP_LOGE(context->GetNodeName(), "input dims invalid for MhcPreBackward");
return ret;
}
uint64_t numsResidual = 0;
uint64_t dimen = 0;
if (gradHInDimNum == BSD_DIM_NUM) {
ret = InferBSDFormat(gradHInShape, gradHPostShape, gradXShape, numsResidual, dimen);
} else if (gradHInDimNum == TD_DIM_NUM) {
ret = InferTNDFormat(gradHInShape, gradHPostShape, gradXShape, numsResidual, dimen);
}
if (ret != GRAPH_SUCCESS) {
return ret;
}
ret = InferOutputShapes(gradPhiShape, gradAlphaShape, gradBiasShape, gradGammaShape,
numsResidual, dimen);
if (ret != GRAPH_SUCCESS) {
return ret;
}
OP_LOGD(context->GetNodeName(), "End to do InferShape MhcPreBackward");
return GRAPH_SUCCESS;
}
static graphStatus InferDataType4mHCPreGrad(gert::InferDataTypeContext *context)
{
auto xGradType = context->GetInputDataType(GRAD_H_IN_INDEX);
context->SetOutputDataType(OUT_GRAD_X_INDEX, xGradType);
context->SetOutputDataType(OUT_GRAD_PHI_INDEX, DataType::DT_FLOAT);
context->SetOutputDataType(OUT_GRAD_ALPHA_INDEX, DataType::DT_FLOAT);
context->SetOutputDataType(OUT_GRAD_BIAS_INDEX, DataType::DT_FLOAT);
context->SetOutputDataType(OUT_GRAD_GAMMA_INDEX, DataType::DT_FLOAT);
return GRAPH_SUCCESS;
}
IMPL_OP_INFERSHAPE(MhcPreBackward).InferShape(InferShape4mHCPreGrad).InferDataType(InferDataType4mHCPreGrad);
}