Line data Source code
1 : /**
2 : * Copyright (c) 2025-2026 Huawei Technologies Co., Ltd.
3 : * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 : * CANN Open Software License Agreement Version 2.0 (the "License").
5 : * Please refer to the License for details. You may not use this file except in compliance with the License.
6 : * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7 : * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8 : * See LICENSE in the root of the software repository for the full text of the License.
9 : */
10 :
11 : /*!
12 : * \file aclnn_foreach_round_off_number_v2.cpp
13 : * \brief
14 : */
15 :
16 : #include "aclnn_foreach_round_off_number_v2.h"
17 : #include "foreach_round_off_number_v2.h"
18 : #include "aclnn_kernels/contiguous.h"
19 : #include "op_api/op_api_def_nn.h"
20 : #include "op_api/aclnn_util.h"
21 : #include "aclnn_kernels/common/op_error_check.h"
22 : #include "opdev/op_dfx.h"
23 : #include "opdev/make_op_executor.h"
24 : #include "opdev/platform.h"
25 :
26 : using namespace op;
27 :
28 : #ifdef __cplusplus
29 : extern "C" {
30 : #endif
31 :
32 : static const std::initializer_list<DataType> ASCEND910BC_TENSOR_DTYPE_DTYPE_SUPPORT_LIST = {
33 : DataType::DT_FLOAT, DataType::DT_FLOAT16, DataType::DT_BF16, DataType::DT_INT32, DataType::DT_INT16};
34 :
35 : static const std::initializer_list<DataType> ASCEND950_TENSOR_DTYPE_DTYPE_SUPPORT_LIST = {
36 : DataType::DT_FLOAT, DataType::DT_FLOAT16, DataType::DT_BF16, DataType::DT_INT32};
37 :
38 : static const std::initializer_list<DataType> FOREACH_ROUND_SCALAR_SUPPORT_LIST = {DataType::DT_INT8,
39 : DataType::DT_INT64};
40 :
41 : static const std::initializer_list<DataType> EMPTY_LIST = {};
42 :
43 : static inline bool CheckNotNull(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
44 : {
45 : OP_CHECK_NULL(self, return false);
46 : OP_CHECK_NULL(scalar, return false);
47 : OP_CHECK_NULL(out, return false);
48 : return true;
49 : }
50 :
51 : static const std::initializer_list<DataType>& GetDtypeSupportList()
52 : {
53 : auto curArch = GetCurrentPlatformInfo().GetCurNpuArch();
54 : if (Ops::NN::AclnnUtil::IsRegbase(curArch)) {
55 : return ASCEND950_TENSOR_DTYPE_DTYPE_SUPPORT_LIST;
56 : } else if (curArch == NpuArch::DAV_2201) {
57 : return ASCEND910BC_TENSOR_DTYPE_DTYPE_SUPPORT_LIST;
58 : } else {
59 : OP_LOGE(ACLNN_ERR_RUNTIME_ERROR, "support for %s is not implemented",
60 : op::ToString(GetCurrentPlatformInfo().GetSocVersion()).GetString());
61 : return EMPTY_LIST;
62 : }
63 : }
64 :
65 : static inline bool CheckFormat(const aclTensorList* self, const aclTensorList* out)
66 : {
67 : for (uint64_t i = 0; i < self->Size(); i++) {
68 : // self格式不能是私有格式
69 : if (IsPrivateFormat((*self)[i]->GetStorageFormat()) || IsPrivateFormat((*out)[i]->GetStorageFormat())) {
70 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Format only support ND, NCHW, NHWC, HWCN, NDHWC, NCDHW.");
71 : return false;
72 : }
73 : }
74 : return true;
75 : }
76 :
77 : static inline bool CheckDtypeValid(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
78 : {
79 : const auto& dtypeSupportList = GetDtypeSupportList();
80 : if (dtypeSupportList.size() == 0) {
81 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "support for %s is not implemented",
82 : op::ToString(GetCurrentPlatformInfo().GetSocVersion()).GetString());
83 : return false;
84 : }
85 : if (self->Size() == 0) {
86 : return true;
87 : }
88 :
89 : // checkself input dtype, and check the releation of input and out
90 : auto selfDtyte = (*self)[0]->GetDataType();
91 : OP_CHECK_DTYPE_NOT_SUPPORT((*self)[0], dtypeSupportList, return false);
92 : for (uint64_t b = 0; b < self->Size(); b++) {
93 : OP_CHECK_DTYPE_NOT_MATCH((*self)[b], selfDtyte, return false);
94 : }
95 :
96 : for (uint64_t b = 0; b < out->Size(); b++) {
97 : OP_CHECK_DTYPE_NOT_MATCH((*out)[b], selfDtyte, return false);
98 : }
99 :
100 : // check the releation of self and scalar
101 : OP_CHECK_DTYPE_NOT_SUPPORT(scalar, FOREACH_ROUND_SCALAR_SUPPORT_LIST, return false);
102 : return true;
103 : }
104 :
105 : static inline bool CheckShape(const aclTensorList* self, const aclTensorList* out)
106 : {
107 : // tensor 维度检查
108 : for (uint64_t i = 0; i < self->Size(); i++) {
109 : OP_CHECK_MAX_DIM((*self)[i], MAX_SUPPORT_DIMS_NUMS, return false);
110 : }
111 :
112 : // self和out的shape必须一致
113 : for (uint64_t i = 0; i < self->Size(); i++) {
114 : OP_CHECK_SHAPE_NOT_EQUAL((*self)[i], (*out)[i], return false);
115 : }
116 : return true;
117 : }
118 :
119 : static inline aclnnStatus CheckParams(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
120 : {
121 : // 1. 检查参数是否为空指针
122 : CHECK_RET(CheckNotNull(self, scalar, out), ACLNN_ERR_PARAM_NULLPTR);
123 :
124 : // Check every entry in tensor lists is not null, to avoid null pointer
125 : // dereference in CheckDtypeValid/CheckShape/CheckFormat.
126 10 : for (uint64_t i = 0; i < self->Size(); i++) {
127 5 : if ((*self)[i] == nullptr) {
128 0 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "self[%lu] is null.", i);
129 0 : return ACLNN_ERR_PARAM_INVALID;
130 : }
131 : }
132 10 : for (uint64_t i = 0; i < out->Size(); i++) {
133 5 : if ((*out)[i] == nullptr) {
134 0 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "out[%lu] is null.", i);
135 0 : return ACLNN_ERR_PARAM_INVALID;
136 : }
137 : }
138 :
139 : // 2. 检查输入的数据类型是否在API支持的数据类型范围之内，需要根据api定义校验
140 : CHECK_RET(CheckDtypeValid(self, scalar, out), ACLNN_ERR_PARAM_INVALID);
141 : // 3. 检查shape是否满足约束
142 : CHECK_RET(CheckShape(self, out), ACLNN_ERR_PARAM_INVALID);
143 : // 4. 检查Format是否满足约束
144 : CHECK_RET(CheckFormat(self, out), ACLNN_ERR_PARAM_INVALID);
145 : return ACLNN_SUCCESS;
146 : }
147 :
148 : static aclnnStatus ExecForeachRoundOffNumberV2GetWorkspaceSize(const aclTensorList* x, const aclScalar* scalar,
149 : const aclTensorList* out, uint64_t* workspaceSize,
150 : aclOpExecutor** executor)
151 : {
152 : // 固定写法，创建OpExecutor
153 : auto uniqueExecutor = CREATE_EXECUTOR();
154 : CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
155 :
156 : // 固定写法，参数检查
157 : auto ret_1 = CheckParams(x, scalar, out);
158 : CHECK_RET(ret_1 == ACLNN_SUCCESS, ret_1);
159 :
160 : // 空Tensorlist处理
161 : if (x->Size() == 0) {
162 : *workspaceSize = 0;
163 : uniqueExecutor.ReleaseTo(executor);
164 : return ACLNN_SUCCESS;
165 : }
166 :
167 : // self如果非连续，需要转连续
168 : std::vector<const aclTensor*> tensorsVec;
169 : for (size_t j = 0; j < x->Size(); ++j) {
170 : auto secondContiguous = l0op::Contiguous((*x)[j], uniqueExecutor.get());
171 : CHECK_RET(secondContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
172 : tensorsVec.push_back(secondContiguous);
173 : }
174 : auto contiguousTensors = uniqueExecutor.get()->AllocTensorList(tensorsVec.data(), tensorsVec.size());
175 : CHECK_RET(contiguousTensors != nullptr, ACLNN_ERR_INNER_NULLPTR);
176 :
177 : // sclar to tensor
178 : const aclTensor* otherTensor = uniqueExecutor.get()->ConvertToTensor(scalar, DataType::DT_INT8);
179 :
180 : // 调用l0算子ForeachRoundOffNumberV2进行计算
181 : auto result = l0op::ForeachRoundOffNumberV2(contiguousTensors, otherTensor, out, uniqueExecutor.get());
182 : CHECK_RET(result != nullptr, ACLNN_ERR_INNER_NULLPTR);
183 :
184 : // 固定写法，获取计算过程中需要使用的workspace大小
185 : *workspaceSize = uniqueExecutor->GetWorkspaceSize();
186 : uniqueExecutor.ReleaseTo(executor);
187 : return ACLNN_SUCCESS;
188 : }
189 :
190 : aclnnStatus aclnnForeachRoundOffNumberV2GetWorkspaceSize(const aclTensorList* x, const aclScalar* roundMode,
191 : aclTensorList* out, uint64_t* workspaceSize,
192 : aclOpExecutor** executor)
193 : {
194 : L2_DFX_PHASE_1(aclnnForeachRoundOffNumberV2, DFX_IN(x, roundMode), DFX_OUT(out));
195 : return ExecForeachRoundOffNumberV2GetWorkspaceSize(x, roundMode, out, workspaceSize, executor);
196 : }
197 :
198 : aclnnStatus aclnnForeachRoundOffNumberV2(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
199 : const aclrtStream stream)
200 : {
201 : L2_DFX_PHASE_2(aclnnForeachRoundOffNumberV2);
202 : return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
203 : }
204 :
205 : #ifdef __cplusplus
206 : }
207 : #endif
|
