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 : #include "aclnn_foreach_minimum_scalar_v2.h"
12 : #include "foreach_minimum_scalar_v2.h"
13 : #include "aclnn_kernels/contiguous.h"
14 : #include "op_api/op_api_def_nn.h"
15 : #include "aclnn_kernels/common/op_error_check.h"
16 : #include "opdev/op_dfx.h"
17 : #include "op_api/aclnn_util.h"
18 : #include "opdev/platform.h"
19 : #include "opdev/make_op_executor.h"
20 :
21 : using namespace op;
22 :
23 : #ifdef __cplusplus
24 : extern "C" {
25 : #endif
26 :
27 : namespace {
28 : const float FLOAT32_MIN_VALUE = -3.4028235e+38f;
29 : const float FLOAT32_MAX_VALUE = 3.4028235e+38f;
30 : const float FLOAT16_MAX_VALUE = 65504.0f;
31 : const float FLOAT16_MIN_VALUE = -65504.0f;
32 : const int32_t INT32_MIN_VAL = -2147483648;
33 : const float BFLOAT16_MAX_VALUE = 3.3895314e+38f;
34 : const float BFLOAT16_MIN_VALUE = -3.3895314e+38f;
35 : const int32_t INT32_MAX_VAL = 2147483647;
36 : } // namespace
37 :
38 : static const std::initializer_list<DataType> ASCEND910BC_TENSOR_DTYPE_DTYPE_SUPPORT_LIST = {
39 : DataType::DT_FLOAT, DataType::DT_FLOAT16, DataType::DT_BF16, DataType::DT_INT32};
40 :
41 : static const std::initializer_list<DataType> FOREACH_SCALAR_FLOAT16_SUPPORT_LIST = {DataType::DT_FLOAT16,
42 : DataType::DT_DOUBLE};
43 :
44 : static const std::initializer_list<DataType> FOREACH_SCALAR_INT_SUPPORT_LIST = {DataType::DT_INT32, DataType::DT_INT64};
45 :
46 : static const std::initializer_list<DataType> FOREACH_SCALAR_FLOAT_SUPPORT_LIST = {DataType::DT_FLOAT,
47 : DataType::DT_DOUBLE};
48 :
49 : static const std::initializer_list<DataType> EMPTY_LIST = {};
50 :
51 : static inline bool CheckNull(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
52 : {
53 : OP_CHECK_NULL(self, return false);
54 : OP_CHECK_NULL(scalar, return false);
55 : OP_CHECK_NULL(out, return false);
56 : return true;
57 : }
58 :
59 : static inline bool CheckFormat(const aclTensorList* self, const aclTensorList* out)
60 : {
61 : for (uint64_t m = 0; m < self->Size(); m++) {
62 : // self格式不能是私有格式
63 : if (IsPrivateFormat((*self)[m]->GetStorageFormat()) || IsPrivateFormat((*out)[m]->GetStorageFormat())) {
64 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Format only support ND, NCHW, NHWC, HWCN, NDHWC, NCDHW.");
65 : return false;
66 : }
67 : }
68 : return true;
69 : }
70 :
71 : static const std::initializer_list<DataType>& GetDtypeSupportList()
72 : {
73 : auto curArch_2 = GetCurrentPlatformInfo().GetCurNpuArch();
74 : if (Ops::NN::AclnnUtil::IsRegbase(curArch_2) || curArch_2 == NpuArch::DAV_2201) {
75 : return ASCEND910BC_TENSOR_DTYPE_DTYPE_SUPPORT_LIST;
76 : } else {
77 : OP_LOGE(ACLNN_ERR_RUNTIME_ERROR, "support for %s is not implemented",
78 : op::ToString(GetCurrentPlatformInfo().GetSocVersion()).GetString());
79 : return EMPTY_LIST;
80 : }
81 : }
82 :
83 : static inline bool CheckDtypeValid(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
84 : {
85 : const auto& dtypeSupportList_2 = GetDtypeSupportList();
86 : if (dtypeSupportList_2.size() == 0) {
87 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "support for %s is not implemented",
88 : op::ToString(GetCurrentPlatformInfo().GetSocVersion()).GetString());
89 : return false;
90 : }
91 : auto selfDtyte = (*self)[0]->GetDataType();
92 : if (self->Size() == 0) {
93 : return true;
94 : }
95 : // checkself input dtype, and check the releation of input and out
96 : OP_CHECK_DTYPE_NOT_SUPPORT((*self)[0], dtypeSupportList_2, return false);
97 : for (uint64_t k = 0; k < self->Size(); k++) {
98 : OP_CHECK_DTYPE_NOT_MATCH((*self)[k], selfDtyte, return false);
99 : }
100 :
101 : for (uint64_t k = 0; k < out->Size(); k++) {
102 : OP_CHECK_DTYPE_NOT_MATCH((*out)[k], selfDtyte, return false);
103 : }
104 :
105 : // check the releation of self and scalar
106 : if (selfDtyte == DataType::DT_FLOAT || selfDtyte == DataType::DT_BF16) {
107 : OP_CHECK_DTYPE_NOT_SUPPORT(scalar, FOREACH_SCALAR_FLOAT_SUPPORT_LIST, return false);
108 : } else if (selfDtyte == DataType::DT_FLOAT16) {
109 : OP_CHECK_DTYPE_NOT_SUPPORT(scalar, FOREACH_SCALAR_FLOAT16_SUPPORT_LIST, return false);
110 : } else {
111 : OP_CHECK_DTYPE_NOT_SUPPORT(scalar, FOREACH_SCALAR_INT_SUPPORT_LIST, return false);
112 : }
113 : return true;
114 : }
115 :
116 : static inline bool CheckShape(const aclTensorList* self, const aclTensorList* out)
117 : {
118 : // tensor 维度检查
119 : for (uint64_t k = 0; k < self->Size(); k++) {
120 : OP_CHECK_MAX_DIM((*self)[k], MAX_SUPPORT_DIMS_NUMS, return false);
121 : }
122 :
123 : // self和out的shape必须一致
124 : for (uint64_t k = 0; k < self->Size(); k++) {
125 : OP_CHECK_SHAPE_NOT_EQUAL((*self)[k], (*out)[k], return false);
126 : }
127 : return true;
128 : }
129 :
130 : static inline aclnnStatus CheckParams(const aclTensorList* self, const aclScalar* scalar, const aclTensorList* out)
131 : {
132 : // 1. 检查参数是否为空指针
133 : CHECK_RET(CheckNull(self, scalar, out), ACLNN_ERR_PARAM_NULLPTR);
134 :
135 : // Check every entry in tensor lists is not null, to avoid null pointer
136 : // dereference in CheckDtypeValid/CheckShape/CheckFormat.
137 0 : for (uint64_t i = 0; i < self->Size(); i++) {
138 0 : if ((*self)[i] == nullptr) {
139 0 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "self[%lu] is null.", i);
140 0 : return ACLNN_ERR_PARAM_INVALID;
141 : }
142 : }
143 0 : for (uint64_t i = 0; i < out->Size(); i++) {
144 0 : if ((*out)[i] == nullptr) {
145 0 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "out[%lu] is null.", i);
146 0 : return ACLNN_ERR_PARAM_INVALID;
147 : }
148 : }
149 :
150 : // 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
151 : CHECK_RET(CheckDtypeValid(self, scalar, out), ACLNN_ERR_PARAM_INVALID);
152 : // 3. 检查shape是否满足约束
153 : CHECK_RET(CheckShape(self, out), ACLNN_ERR_PARAM_INVALID);
154 : // 4. 检查Format是否满足约束
155 : CHECK_RET(CheckFormat(self, out), ACLNN_ERR_PARAM_INVALID);
156 : return ACLNN_SUCCESS;
157 : }
158 :
159 : static inline aclnnStatus CheckScalarValueValid(const aclScalar* scalar, const DataType tensorDtype)
160 : {
161 : if (scalar->GetDataType() == tensorDtype) {
162 : return ACLNN_SUCCESS;
163 : }
164 :
165 : double val = static_cast<double>(scalar->ToDouble());
166 :
167 : switch (tensorDtype) {
168 : case DataType::DT_INT32:
169 : if (val > static_cast<double>(INT32_MAX_VAL) || val < static_cast<double>(INT32_MIN_VAL)) {
170 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Scalar value exceeds int32 range.");
171 : return ACLNN_ERR_PARAM_INVALID;
172 : }
173 : return ACLNN_SUCCESS;
174 : case DataType::DT_FLOAT:
175 : if (val < static_cast<double>(FLOAT32_MIN_VALUE) || val > static_cast<double>(FLOAT32_MAX_VALUE)) {
176 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Scalar value exceeds float32 range.");
177 : return ACLNN_ERR_PARAM_INVALID;
178 : }
179 : return ACLNN_SUCCESS;
180 : case DataType::DT_FLOAT16:
181 : if (val > static_cast<double>(FLOAT16_MAX_VALUE) || val < static_cast<double>(FLOAT16_MIN_VALUE)) {
182 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Scalar value exceeds float16 range.");
183 : return ACLNN_ERR_PARAM_INVALID;
184 : }
185 : return ACLNN_SUCCESS;
186 : case DataType::DT_BF16:
187 : if (val > static_cast<double>(BFLOAT16_MAX_VALUE) || val < static_cast<double>(BFLOAT16_MIN_VALUE)) {
188 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Scalar value exceeds bfloat16 range.");
189 : return ACLNN_ERR_PARAM_INVALID;
190 : }
191 : return ACLNN_SUCCESS;
192 : default:
193 : OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Unsupported tensor data type for scalar value check");
194 : return ACLNN_ERR_PARAM_INVALID;
195 : }
196 : }
197 :
198 : static aclnnStatus ExecForeachMinimumScalarV2GetWorkspaceSize(const aclTensorList* x, const aclScalar* scalar,
199 : const aclTensorList* out, uint64_t* workspaceSize,
200 : aclOpExecutor** executor)
201 : {
202 : // 固定写法，创建OpExecutor
203 : auto uniqueExecutor = CREATE_EXECUTOR();
204 : CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
205 :
206 : // 固定写法，参数检查
207 : auto ret_3 = CheckParams(x, scalar, out);
208 : CHECK_RET(ret_3 == ACLNN_SUCCESS, ret_3);
209 :
210 : // 空Tensorlist处理
211 : if (x->Size() == 0) {
212 : *workspaceSize = 0;
213 : uniqueExecutor.ReleaseTo(executor);
214 : return ACLNN_SUCCESS;
215 : }
216 :
217 : // self如果非连续，需要转连续
218 : std::vector<const aclTensor*> tensorsVec;
219 : for (size_t j = 0; j < x->Size(); ++j) {
220 : auto secondContiguous = l0op::Contiguous((*x)[j], uniqueExecutor.get());
221 : CHECK_RET(secondContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
222 : tensorsVec.push_back(secondContiguous);
223 : }
224 : auto contiguousTensors = uniqueExecutor.get()->AllocTensorList(tensorsVec.data(), tensorsVec.size());
225 : CHECK_RET(contiguousTensors != nullptr, ACLNN_ERR_INNER_NULLPTR);
226 :
227 : // 校验scalar的数值是否满足约束
228 : ret_3 = CheckScalarValueValid(scalar, (*x)[0]->GetDataType());
229 : CHECK_RET(ret_3 == ACLNN_SUCCESS, ret_3);
230 :
231 : // sclar to tensor
232 : const aclTensor* otherTensor;
233 : if ((*x)[0]->GetDataType() == DataType::DT_BF16) {
234 : otherTensor = uniqueExecutor.get()->ConvertToTensor(scalar, DataType::DT_FLOAT);
235 : } else {
236 : otherTensor = uniqueExecutor.get()->ConvertToTensor(scalar, (*x)[0]->GetDataType());
237 : }
238 :
239 : // 调用l0算子ForeachMinimumScalarV2进行计算
240 : auto result = l0op::ForeachMinimumScalarV2(contiguousTensors, otherTensor, out, uniqueExecutor.get());
241 : CHECK_RET(result != nullptr, ACLNN_ERR_INNER_NULLPTR);
242 :
243 : // 固定写法，获取计算过程中需要使用的workspace大小
244 : *workspaceSize = uniqueExecutor->GetWorkspaceSize();
245 : uniqueExecutor.ReleaseTo(executor);
246 : return ACLNN_SUCCESS;
247 : }
248 :
249 : aclnnStatus aclnnForeachMinimumScalarV2GetWorkspaceSize(const aclTensorList* x, const aclScalar* scalar,
250 : aclTensorList* out, uint64_t* workspaceSize,
251 : aclOpExecutor** executor)
252 : {
253 : L2_DFX_PHASE_1(aclnnForeachMinimumScalarV2, DFX_IN(x, scalar), DFX_OUT(out));
254 : return ExecForeachMinimumScalarV2GetWorkspaceSize(x, scalar, out, workspaceSize, executor);
255 : }
256 :
257 : aclnnStatus aclnnForeachMinimumScalarV2(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor,
258 : const aclrtStream stream)
259 : {
260 : L2_DFX_PHASE_2(aclnnForeachMinimumScalarV2);
261 : return CommonOpExecutorRun(workspace, workspaceSize, executor, stream);
262 : }
263 :
264 : #ifdef __cplusplus
265 : }
266 : #endif
|
