* 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_sum.h"
#include "accumulate_nv2.h"
#include "aclnn_kernels/cast.h"
#include "aclnn_kernels/contiguous.h"
#include "aclnn_kernels/reshape.h"
#include "aclnn_kernels/common/op_error_check.h"
#include "conversion/broadcast_to/op_api/broadcast_to.h"
#include "opdev/common_types.h"
#include "opdev/data_type_utils.h"
#include "opdev/format_utils.h"
#include "opdev/op_dfx.h"
#include "opdev/op_executor.h"
#include "opdev/shape_utils.h"
#include "opdev/tensor_view_utils.h"
#include "op_api/op_api_def.h"
using namespace op;
#ifdef __cplusplus
extern "C" {
#endif
static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST = {
op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, op::DataType::DT_INT8, op::DataType::DT_INT32,
op::DataType::DT_UINT8};
static bool CheckNotNull(const aclTensorList *tensors, const aclTensor* out) {
OP_CHECK_NULL(tensors, return false);
for (uint64_t i = 0; i < tensors->Size(); i++) {
if ((*tensors)[i] == nullptr) {
OP_LOGE(ACLNN_ERR_PARAM_NULLPTR, "expected a proper Tensor but got null for tensor %lu.", i);
return false;
}
}
OP_CHECK_NULL(out, return false);
return true;
}
static bool CheckDtypeValid(const aclTensorList *tensors, const aclTensor* out) {
for (uint64_t i = 0; i < tensors->Size(); i++) {
if (!CheckType((*tensors)[i]->GetDataType(), DTYPE_SUPPORT_LIST)) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "tensor %lu not implemented for %s.", i,
op::ToString((*tensors)[i]->GetDataType()).GetString());
return false;
}
}
OP_CHECK_DTYPE_NOT_SUPPORT(out, DTYPE_SUPPORT_LIST, return false);
return true;
}
static bool GetTensorsBroadcastShape(const aclTensorList *tensors, op::Shape &broadcastShape)
{
broadcastShape = (*tensors)[0]->GetViewShape();
for (uint64_t i = 1; i < tensors->Size(); ++i) {
if (!BroadcastInferShape((*tensors)[i]->GetViewShape(), broadcastShape, broadcastShape)) {
return false;
}
}
return true;
}
static bool CheckShape(const aclTensorList *tensors, const aclTensor* out) {
for (uint64_t i = 0; i < tensors->Size(); ++i) {
auto dimNum = (*tensors)[i]->GetViewShape().GetDimNum();
if (dimNum > MAX_SUPPORT_DIMS_NUMS) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Dim of tensor %lu is %zu, can't be greater than %zu.", i, dimNum,
MAX_SUPPORT_DIMS_NUMS);
return false;
}
}
OP_CHECK_MAX_DIM(out, MAX_SUPPORT_DIMS_NUMS, return false);
op::Shape broadcastShape;
if (!GetTensorsBroadcastShape(tensors, broadcastShape)) {
OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Input tensors can not broadcast.");
return false;
}
OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE(out, broadcastShape, return false);
return true;
}
static aclnnStatus CheckParams(const aclTensorList *tensors, const aclTensor* out) {
CHECK_RET(CheckNotNull(tensors, out), ACLNN_ERR_PARAM_NULLPTR);
CHECK_RET(CheckDtypeValid(tensors, out), ACLNN_ERR_PARAM_INVALID);
CHECK_RET(CheckShape(tensors, out), ACLNN_ERR_PARAM_INVALID);
return ACLNN_SUCCESS;
}
static aclnnStatus SplitToSumN(const aclTensorList *tensors, const aclIntArray *broadcastShapeArray,
const aclTensor **sumOut, aclOpExecutor *executor,
bool promoteFp16) {
size_t MAX_TENSOR_SIZE = 16;
op::FVector<const aclTensor *> tensorList;
for (uint64_t i = 0; i < tensors->Size(); i++) {
auto contiguousOut = l0op::Contiguous((*tensors)[i], executor);
CHECK_RET(contiguousOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
auto broadcastOut = l0op::BroadcastTo(contiguousOut, broadcastShapeArray, executor);
CHECK_RET(broadcastOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
if (promoteFp16) {
broadcastOut = l0op::Cast(broadcastOut, op::DataType::DT_FLOAT, executor);
CHECK_RET(broadcastOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
}
tensorList.push_back(broadcastOut);
}
while (tensorList.size() >= MAX_TENSOR_SIZE) {
op::FVector<const aclTensor *> tensorListOnes;
tensorListOnes.assign(tensorList.end() - MAX_TENSOR_SIZE, tensorList.end());
auto onesComputeOut = l0op::AccumulateNV2(executor->AllocTensorList(tensorListOnes.data(), tensorListOnes.size()),
executor);
CHECK_RET(onesComputeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
tensorList.erase(tensorList.end() - MAX_TENSOR_SIZE, tensorList.end());
tensorList.push_back(onesComputeOut);
}
*sumOut = tensorList[0];
if (tensorList.size() > 1) {
*sumOut = l0op::AccumulateNV2(executor->AllocTensorList(tensorList.data(), tensorList.size()), executor);
}
CHECK_RET(sumOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
return ACLNN_SUCCESS;
}
static aclTensorList* SumAdaptInputZeroDimTensor(const aclTensorList *tensors, aclOpExecutor *executor) {
op::FVector<const aclTensor*> fTensorList;
int64_t selfShapeValue[1] = {1};
aclIntArray *selfShape = executor->AllocIntArray(selfShapeValue, 1);
for (uint64_t i = 0; i < tensors->Size(); i++) {
auto oneTensor = (*tensors)[i];
if (oneTensor->GetViewShape().GetDimNum() == 0) {
auto reshapeTensor = l0op::Reshape(oneTensor, selfShape, executor);
CHECK_RET(reshapeTensor != nullptr, nullptr);
fTensorList.push_back(reshapeTensor);
} else {
fTensorList.push_back(oneTensor);
}
}
return executor->AllocTensorList(fTensorList.data(), fTensorList.size());
}
aclnnStatus aclnnSumGetWorkspaceSize(const aclTensorList *tensors, aclTensor* out, uint64_t* workspaceSize,
aclOpExecutor** executor) {
OP_CHECK_COMM_INPUT(workspaceSize, executor);
L2_DFX_PHASE_1(aclnnSum, DFX_IN(tensors), DFX_OUT(out));
auto uniqueExecutor = CREATE_EXECUTOR();
CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
auto ret = CheckParams(tensors, out);
CHECK_RET(ret == ACLNN_SUCCESS, ret);
if ((*tensors)[0]->IsEmpty()) {
*workspaceSize = 0;
uniqueExecutor.ReleaseTo(executor);
return ACLNN_SUCCESS;
}
auto reshapeTensors = SumAdaptInputZeroDimTensor(tensors, uniqueExecutor.get());
CHECK_RET(reshapeTensors != nullptr, ACLNN_ERR_INNER_NULLPTR);
op::Shape broadcastShape = (*reshapeTensors)[0]->GetViewShape();
for (uint64_t i = 1; i < reshapeTensors->Size(); i++) {
BroadcastInferShape((*reshapeTensors)[i]->GetViewShape(), broadcastShape, broadcastShape);
}
op::FVector<int64_t, op::MAX_DIM_NUM> broadcastDims = op::ToShapeVector(broadcastShape);
auto broadcastShapeArray = uniqueExecutor.get()->AllocIntArray(broadcastDims.data(), broadcastDims.size());
CHECK_RET(broadcastShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
bool promoteFp16 = ((*reshapeTensors)[0]->GetDataType() == op::DataType::DT_FLOAT16);
const aclTensor *sumOut = nullptr;
aclnnStatus retSplit = SplitToSumN(reshapeTensors, broadcastShapeArray, &sumOut, uniqueExecutor.get(),
promoteFp16);
CHECK_RET(retSplit == ACLNN_SUCCESS, retSplit);
if (promoteFp16) {
sumOut = l0op::Cast(sumOut, op::DataType::DT_FLOAT16, uniqueExecutor.get());
CHECK_RET(sumOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
}
auto viewCopyResult = l0op::ViewCopy(sumOut, out, uniqueExecutor.get());
CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
*workspaceSize = uniqueExecutor->GetWorkspaceSize();
uniqueExecutor.ReleaseTo(executor);
return ACLNN_SUCCESS;
}
aclnnStatus aclnnSum(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) {
L2_DFX_PHASE_2(aclnnSum);
return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
}
#ifdef __cplusplus
}
#endif
