已合并
fix: enable double path for remainder scalar tensor #4018
zhang-song-rui创建于 7月13日
fix: enable double path for remainder scalar tensor #4018
已合并
从已删除 :master合入到cann/ops-mathmaster
共 2 个文件变更+111-82
| @@ -30,18 +30,18 @@ using namespace op; | |||||||||
| 30 | extern "C" { | 30 | extern "C" { | ||||||
| 31 | 31 | ||||||||
| 32 | 32 | ||||||||
| 33 | - static const std::initializer_list<op::DataType> ASCEND910_DTYPE_DTYPE_SUPPORT_LIST = { | 33 | +static const std::initializer_list<op::DataType> ASCEND910_DTYPE_DTYPE_SUPPORT_LIST = { | ||||||
| 34 | - op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, | 34 | + op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, | ||||||
| 35 | - op::DataType::DT_DOUBLE}; | 35 | + op::DataType::DT_DOUBLE}; | ||||||
| 36 | - static const std::initializer_list<op::DataType> ASCEND910B_DTYPE_DTYPE_SUPPORT_LIST = { | 36 | +static const std::initializer_list<op::DataType> ASCEND910B_DTYPE_DTYPE_SUPPORT_LIST = { | ||||||
| 37 | - op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, | 37 | + op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, | ||||||
| 38 | - op::DataType::DT_FLOAT, op::DataType::DT_DOUBLE, op::DataType::DT_BF16}; | 38 | + op::DataType::DT_FLOAT, op::DataType::DT_DOUBLE, op::DataType::DT_BF16}; | ||||||
| 39 | - static const std::initializer_list<op::DataType> ASCEND310P_DTYPE_DTYPE_SUPPORT_LIST = { | 39 | +static const std::initializer_list<op::DataType> ASCEND310P_DTYPE_DTYPE_SUPPORT_LIST = { | ||||||
| 40 | - op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, | 40 | + op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, | ||||||
| 41 | - op::DataType::DT_DOUBLE}; | 41 | + op::DataType::DT_DOUBLE}; | ||||||
| 42 | - static const std::initializer_list<DataType> emptyDtypes = {}; | 42 | +static const std::initializer_list<DataType> emptyDtypes = {}; | ||||||
| 43 | - static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST_COMPLEX = { | 43 | +static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST_COMPLEX = {op::DataType::DT_COMPLEX64, | ||||||
| 44 | - op::DataType::DT_COMPLEX64, op::DataType::DT_COMPLEX128}; | 44 | + op::DataType::DT_COMPLEX128}; | ||||||
| 45 | 45 | ||||||||
| 46 | static const std::initializer_list<DataType>& GetDtypeSupportList(NpuArch npuArch, SocVersion socVersion) | 46 | static const std::initializer_list<DataType>& GetDtypeSupportList(NpuArch npuArch, SocVersion socVersion) | ||||||
| 47 | { | 47 | { | ||||||
| @@ -175,9 +175,8 @@ static bool CheckPromoteType(const op::DataType selfDtype, const op::DataType ot | |||||||||
| 175 | // 检查self和other能否做数据类型推导 | 175 | // 检查self和other能否做数据类型推导 | ||||||
| 176 | auto promoteType = op::PromoteType(selfDtype, otherDtype); | 176 | auto promoteType = op::PromoteType(selfDtype, otherDtype); | ||||||
| 177 | if (promoteType == DataType::DT_UNDEFINED) { | 177 | if (promoteType == DataType::DT_UNDEFINED) { | ||||||
| 178 | - OP_LOGE( | 178 | + OP_LOGE(ACLNN_ERR_PARAM_INVALID, "self dtype %s and other dtype %s can not promote dtype.", | ||||||
| 179 | - ACLNN_ERR_PARAM_INVALID, "self dtype %s and other dtype %s can not promote dtype.", | 179 | + op::ToString(selfDtype).GetString(), op::ToString(otherDtype).GetString()); | ||||||
| 180 | - op::ToString(selfDtype).GetString(), op::ToString(otherDtype).GetString()); | ||||||||
| 181 | return false; | 180 | return false; | ||||||
| 182 | } | 181 | } | ||||||
| 183 | 182 | ||||||||
| @@ -191,9 +190,8 @@ static bool CheckPromoteType(const op::DataType selfDtype, const op::DataType ot | |||||||||
| 191 | return false; | 190 | return false; | ||||||
| 192 | } | 191 | } | ||||||
| 193 | if (!CheckType(promoteType, DTYPE_SUPPORT_LIST)) { | 192 | if (!CheckType(promoteType, DTYPE_SUPPORT_LIST)) { | ||||||
| 194 | - OP_LOGE( | 193 | + OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Promote type %s should be in dtype support list [%s].", | ||||||
| 195 | - ACLNN_ERR_PARAM_INVALID, "Promote type %s should be in dtype support list [%s].", | 194 | + op::ToString(promoteType).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||
| 196 | - op::ToString(promoteType).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||||
| 197 | return false; | 195 | return false; | ||||||
| 198 | } | 196 | } | ||||||
| 199 | 197 | ||||||||
| @@ -201,8 +199,8 @@ static bool CheckPromoteType(const op::DataType selfDtype, const op::DataType ot | |||||||||
| 201 | } | 199 | } | ||||||
| 202 | 200 | ||||||||
| 203 | // 1. self和other没有complex 2. self能cast成castDtype 3. castDtype为算子支持的数据类型 4. castDtype能cast成out | 201 | // 1. self和other没有complex 2. self能cast成castDtype 3. castDtype为算子支持的数据类型 4. castDtype能cast成out | ||||||
| 204 | -static bool CheckPromoteTypeTensorScalar( | 202 | +static bool CheckPromoteTypeTensorScalar(const op::DataType selfDtype, const op::DataType otherDtype, | ||||||
| 205 | - const op::DataType selfDtype, const op::DataType otherDtype, const op::DataType outDtype) | 203 | + const op::DataType outDtype) | ||||||
| 206 | { | 204 | { | ||||||
| 207 | // 检查self和other没有为complex | 205 | // 检查self和other没有为complex | ||||||
| 208 | if (CheckType(selfDtype, DTYPE_SUPPORT_LIST_COMPLEX) || CheckType(otherDtype, DTYPE_SUPPORT_LIST_COMPLEX)) { | 206 | if (CheckType(selfDtype, DTYPE_SUPPORT_LIST_COMPLEX) || CheckType(otherDtype, DTYPE_SUPPORT_LIST_COMPLEX)) { | ||||||
| @@ -222,9 +220,8 @@ static bool CheckPromoteTypeTensorScalar( | |||||||||
| 222 | return false; | 220 | return false; | ||||||
| 223 | } | 221 | } | ||||||
| 224 | if (!CheckType(castDtype, DTYPE_SUPPORT_LIST)) { | 222 | if (!CheckType(castDtype, DTYPE_SUPPORT_LIST)) { | ||||||
| 225 | - OP_LOGE( | 223 | + OP_LOGE(ACLNN_ERR_PARAM_INVALID, "expected dtype %s should be in dtype support list [%s].", | ||||||
| 226 | - ACLNN_ERR_PARAM_INVALID, "expected dtype %s should be in dtype support list [%s].", | 224 | + op::ToString(castDtype).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||
| 227 | - op::ToString(castDtype).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||||
| 228 | return false; | 225 | return false; | ||||||
| 229 | } | 226 | } | ||||||
| 230 | 227 | ||||||||
| @@ -235,8 +232,8 @@ static bool CheckPromoteTypeTensorScalar( | |||||||||
| 235 | } | 232 | } | ||||||
| 236 | 233 | ||||||||
| 237 | // 1. self和other没有complex 2. other能cast成outDtype 3. outDtype为算子支持的数据类型 | 234 | // 1. self和other没有complex 2. other能cast成outDtype 3. outDtype为算子支持的数据类型 | ||||||
| 238 | -static bool CheckPromoteTypeScalarTensor( | 235 | +static bool CheckPromoteTypeScalarTensor(const op::DataType selfDtype, const op::DataType otherDtype, | ||||||
| 239 | - const op::DataType selfDtype, const op::DataType otherDtype, const op::DataType outDtype) | 236 | + const op::DataType outDtype) | ||||||
| 240 | { | 237 | { | ||||||
| 241 | // 检查self和other没有为complex | 238 | // 检查self和other没有为complex | ||||||
| 242 | if (CheckType(selfDtype, DTYPE_SUPPORT_LIST_COMPLEX) || CheckType(otherDtype, DTYPE_SUPPORT_LIST_COMPLEX)) { | 239 | if (CheckType(selfDtype, DTYPE_SUPPORT_LIST_COMPLEX) || CheckType(otherDtype, DTYPE_SUPPORT_LIST_COMPLEX)) { | ||||||
| @@ -255,9 +252,8 @@ static bool CheckPromoteTypeScalarTensor( | |||||||||
| 255 | return false; | 252 | return false; | ||||||
| 256 | } | 253 | } | ||||||
| 257 | if (!CheckType(outDtype, DTYPE_SUPPORT_LIST)) { | 254 | if (!CheckType(outDtype, DTYPE_SUPPORT_LIST)) { | ||||||
| 258 | - OP_LOGE( | 255 | + OP_LOGE(ACLNN_ERR_PARAM_INVALID, "out dtype %s should be in dtype support list [%s].", | ||||||
| 259 | - ACLNN_ERR_PARAM_INVALID, "out dtype %s should be in dtype support list [%s].", | 256 | + op::ToString(outDtype).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||
| 260 | - op::ToString(outDtype).GetString(), op::ToString(DTYPE_SUPPORT_LIST).GetString()); | ||||||||
| 261 | return false; | 257 | return false; | ||||||
| 262 | } | 258 | } | ||||||
| 263 | 259 | ||||||||
| @@ -311,10 +307,9 @@ static bool CheckBroadcastShape(const aclTensor* self, const aclTensor* other, c | |||||||||
| 311 | return true; | 307 | return true; | ||||||
| 312 | } | 308 | } | ||||||
| 313 | 309 | ||||||||
| 314 | - OP_LOGE( | 310 | + OP_LOGE(ACLNN_ERR_PARAM_INVALID, | ||||||
| 315 | - ACLNN_ERR_PARAM_INVALID, | 311 | + "expected consistent tensor shape for the broadcast shape and out, but got %s and %s respectively.", | ||||||
| 316 | - "expected consistent tensor shape for the broadcast shape and out, but got %s and %s respectively.", | 312 | + op::ToString(broadcastShape).GetString(), op::ToString(out->GetViewShape()).GetString()); | ||||||
| 317 | - op::ToString(broadcastShape).GetString(), op::ToString(out->GetViewShape()).GetString()); | ||||||||
| 318 | return false; | 313 | return false; | ||||||
| 319 | } | 314 | } | ||||||
| 320 | 315 | ||||||||
| @@ -372,8 +367,8 @@ static aclIntArray* GetTensorShape(const aclTensor* self, aclOpExecutor* executo | |||||||||
| 372 | } | 367 | } | ||||||
| 373 | 368 | ||||||||
| 374 | // broadcast成对应shape | 369 | // broadcast成对应shape | ||||||
| 375 | -static const aclTensor* BroadcastTensor( | 370 | +static const aclTensor* BroadcastTensor(const aclTensor* x, const aclTensor* out, const aclIntArray* broadcastShape, | ||||||
| 376 | - const aclTensor* x, const aclTensor* out, const aclIntArray* broadcastShape, aclOpExecutor* executor) | 371 | + aclOpExecutor* executor) | ||||||
| 377 | { | 372 | { | ||||||
| 378 | // 涉及3维->4维，4维->5维，因此都reformat成ND | 373 | // 涉及3维->4维，4维->5维，因此都reformat成ND | ||||||
| 379 | x = l0op::ReFormat(x, op::Format::FORMAT_ND); | 374 | x = l0op::ReFormat(x, op::Format::FORMAT_ND); | ||||||
| @@ -396,9 +391,8 @@ static const aclTensor* BroadcastTensor( | |||||||||
| 396 | return x; | 391 | return x; | ||||||
| 397 | } | 392 | } | ||||||
| 398 | 393 | ||||||||
| 399 | -static aclnnStatus RemainderMainProcess( | 394 | +static aclnnStatus RemainderMainProcess(const aclTensor* selfContiguous, const aclTensor* otherContiguous, | ||||||
| 400 | - const aclTensor* selfContiguous, const aclTensor* otherContiguous, const aclTensor* out, bool needUnsqueeze, | 395 | + const aclTensor* out, bool needUnsqueeze, aclOpExecutor* executor) | ||||||
| 401 | - aclOpExecutor* executor) | ||||||||
| 402 | { | 396 | { | ||||||
| 403 | auto remainderOut = l0op::FloorMod(selfContiguous, otherContiguous, executor); | 397 | auto remainderOut = l0op::FloorMod(selfContiguous, otherContiguous, executor); | ||||||
| 404 | CHECK_RET(remainderOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | 398 | CHECK_RET(remainderOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| @@ -461,9 +455,8 @@ static aclnnStatus CheckParamsTensorScalar(const aclTensor* self, const aclScala | |||||||||
| 461 | // 2. self和out的shape一致 | 455 | // 2. self和out的shape一致 | ||||||
| 462 | OP_CHECK_SHAPE_NOT_EQUAL(self, out, return ACLNN_ERR_PARAM_INVALID); | 456 | OP_CHECK_SHAPE_NOT_EQUAL(self, out, return ACLNN_ERR_PARAM_INVALID); | ||||||
| 463 | // 3. self和other没有complex + self能cast成castDtype + castDtype为算子支持的数据类型 + castDtype能cast成out | 457 | // 3. self和other没有complex + self能cast成castDtype + castDtype为算子支持的数据类型 + castDtype能cast成out | ||||||
| 464 | - CHECK_RET( | 458 | + CHECK_RET(CheckPromoteTypeTensorScalar(self->GetDataType(), other->GetDataType(), out->GetDataType()), | ||||||
| 465 | - CheckPromoteTypeTensorScalar(self->GetDataType(), other->GetDataType(), out->GetDataType()), | 459 | + ACLNN_ERR_PARAM_INVALID); | ||||||
| 466 | - ACLNN_ERR_PARAM_INVALID); | ||||||||
| 467 | // 4. 维度数不能超过8维 | 460 | // 4. 维度数不能超过8维 | ||||||
| 468 | CHECK_RET(CheckTensorDimSize(self), ACLNN_ERR_PARAM_INVALID); | 461 | CHECK_RET(CheckTensorDimSize(self), ACLNN_ERR_PARAM_INVALID); | ||||||
| 469 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | 462 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | ||||||
| @@ -479,9 +472,8 @@ static aclnnStatus CheckParamsInplaceTensorScalar(const aclTensor* self, const a | |||||||||
| 479 | // 2. self和out的shape一致 | 472 | // 2. self和out的shape一致 | ||||||
| 480 | OP_CHECK_SHAPE_NOT_EQUAL(self, out, return ACLNN_ERR_PARAM_INVALID); | 473 | OP_CHECK_SHAPE_NOT_EQUAL(self, out, return ACLNN_ERR_PARAM_INVALID); | ||||||
| 481 | // 3. self和other没有complex + self能cast成castDtype + castDtype为算子支持的数据类型 + castDtype能cast成out | 474 | // 3. self和other没有complex + self能cast成castDtype + castDtype为算子支持的数据类型 + castDtype能cast成out | ||||||
| 482 | - CHECK_RET( | 475 | + CHECK_RET(CheckPromoteTypeTensorScalar(self->GetDataType(), other->GetDataType(), out->GetDataType()), | ||||||
| 483 | - CheckPromoteTypeTensorScalar(self->GetDataType(), other->GetDataType(), out->GetDataType()), | 476 | + ACLNN_ERR_PARAM_INVALID); | ||||||
| 484 | - ACLNN_ERR_PARAM_INVALID); | ||||||||
| 485 | // 4. 维度数不能超过8维 | 477 | // 4. 维度数不能超过8维 | ||||||
| 486 | CHECK_RET(CheckTensorDimSize(self), ACLNN_ERR_PARAM_INVALID); | 478 | CHECK_RET(CheckTensorDimSize(self), ACLNN_ERR_PARAM_INVALID); | ||||||
| 487 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | 479 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | ||||||
| @@ -499,9 +491,8 @@ static aclnnStatus CheckParamsScalarTensor(const aclScalar* self, const aclTenso | |||||||||
| 499 | // 2. other和out的shape一致 | 491 | // 2. other和out的shape一致 | ||||||
| 500 | OP_CHECK_SHAPE_NOT_EQUAL(other, out, return ACLNN_ERR_PARAM_INVALID); | 492 | OP_CHECK_SHAPE_NOT_EQUAL(other, out, return ACLNN_ERR_PARAM_INVALID); | ||||||
| 501 | // 3. self和other没有complex + other能cast成outDtype + outDtype为算子支持的数据类型 | 493 | // 3. self和other没有complex + other能cast成outDtype + outDtype为算子支持的数据类型 | ||||||
| 502 | - CHECK_RET( | 494 | + CHECK_RET(CheckPromoteTypeScalarTensor(self->GetDataType(), other->GetDataType(), out->GetDataType()), | ||||||
| 503 | - CheckPromoteTypeScalarTensor(self->GetDataType(), other->GetDataType(), out->GetDataType()), | 495 | + ACLNN_ERR_PARAM_INVALID); | ||||||
| 504 | - ACLNN_ERR_PARAM_INVALID); | ||||||||
| 505 | // 4. 维度数不能超过8维 | 496 | // 4. 维度数不能超过8维 | ||||||
| 506 | CHECK_RET(CheckTensorDimSize(other), ACLNN_ERR_PARAM_INVALID); | 497 | CHECK_RET(CheckTensorDimSize(other), ACLNN_ERR_PARAM_INVALID); | ||||||
| 507 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | 498 | CHECK_RET(CheckTensorDimSize(out), ACLNN_ERR_PARAM_INVALID); | ||||||
| @@ -510,8 +501,8 @@ static aclnnStatus CheckParamsScalarTensor(const aclScalar* self, const aclTenso | |||||||||
| 510 | } | 501 | } | ||||||
| 511 | 502 | ||||||||
| 512 | // Tensor self, Tensor other | 503 | // Tensor self, Tensor other | ||||||
| 513 | -aclnnStatus ExecRemainderTensorTensorGetWorkspaceSize( | 504 | +aclnnStatus ExecRemainderTensorTensorGetWorkspaceSize(const aclTensor* self, const aclTensor* other, aclTensor* out, | ||||||
| 514 | - const aclTensor* self, const aclTensor* other, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor) | 505 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 515 | { | 506 | { | ||||||
| 516 | auto uniqueExecutor = CREATE_EXECUTOR(); | 507 | auto uniqueExecutor = CREATE_EXECUTOR(); | ||||||
| 517 | CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR); | 508 | CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR); | ||||||
| @@ -557,8 +548,8 @@ aclnnStatus ExecRemainderTensorTensorGetWorkspaceSize( | |||||||||
| 557 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 548 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 558 | 549 | ||||||||
| 559 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | 550 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | ||||||
| 560 | - auto remainderRes = | 551 | + auto remainderRes = RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, | ||||||
| 561 | - RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, uniqueExecutor.get()); | 552 | + uniqueExecutor.get()); | ||||||
| 562 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | 553 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | ||||||
| 563 | } | 554 | } | ||||||
| 564 | 555 | ||||||||
| @@ -569,8 +560,8 @@ aclnnStatus ExecRemainderTensorTensorGetWorkspaceSize( | |||||||||
| 569 | } | 560 | } | ||||||
| 570 | 561 | ||||||||
| 571 | // Tensor self, Scalar other | 562 | // Tensor self, Scalar other | ||||||
| 572 | -aclnnStatus ExecRemainderTensorScalarGetWorkspaceSize( | 563 | +aclnnStatus ExecRemainderTensorScalarGetWorkspaceSize(const aclTensor* self, const aclScalar* other, aclTensor* out, | ||||||
| 573 | - const aclTensor* self, const aclScalar* other, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor) | 564 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 574 | { | 565 | { | ||||||
| 575 | auto uniqueExecutor = CREATE_EXECUTOR(); | 566 | auto uniqueExecutor = CREATE_EXECUTOR(); | ||||||
| 576 | CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR); | 567 | CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR); | ||||||
| @@ -589,8 +580,8 @@ aclnnStatus ExecRemainderTensorScalarGetWorkspaceSize( | |||||||||
| 589 | auto selfCasted = l0op::Cast(selfContiguous, castDtype, uniqueExecutor.get()); | 580 | auto selfCasted = l0op::Cast(selfContiguous, castDtype, uniqueExecutor.get()); | ||||||
| 590 | CHECK_RET(selfCasted != nullptr, ACLNN_ERR_INNER_NULLPTR); | 581 | CHECK_RET(selfCasted != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 591 | 582 | ||||||||
| 592 | - auto floorModOpOut = | 583 | + auto floorModOpOut = l0op::FloorMod(selfCasted, uniqueExecutor.get()->ConvertToTensor(other, castDtype), | ||||||
| 593 | - l0op::FloorMod(selfCasted, uniqueExecutor.get()->ConvertToTensor(other, castDtype), uniqueExecutor.get()); | 584 | + uniqueExecutor.get()); | ||||||
| 594 | CHECK_RET(floorModOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | 585 | CHECK_RET(floorModOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 595 | 586 | ||||||||
| 596 | auto castOut = l0op::Cast(floorModOpOut, out->GetDataType(), uniqueExecutor.get()); | 587 | auto castOut = l0op::Cast(floorModOpOut, out->GetDataType(), uniqueExecutor.get()); | ||||||
| @@ -607,8 +598,8 @@ aclnnStatus ExecRemainderTensorScalarGetWorkspaceSize( | |||||||||
| 607 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 598 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 608 | 599 | ||||||||
| 609 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | 600 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | ||||||
| 610 | - auto remainderRes = | 601 | + auto remainderRes = RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, | ||||||
| 611 | - RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, uniqueExecutor.get()); | 602 | + uniqueExecutor.get()); | ||||||
| 612 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | 603 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | ||||||
| 613 | } | 604 | } | ||||||
| 614 | 605 | ||||||||
| @@ -619,8 +610,8 @@ aclnnStatus ExecRemainderTensorScalarGetWorkspaceSize( | |||||||||
| 619 | } | 610 | } | ||||||
| 620 | 611 | ||||||||
| 621 | // 非inplace | 612 | // 非inplace | ||||||
| 622 | -aclnnStatus aclnnRemainderTensorTensorGetWorkspaceSize( | 613 | +aclnnStatus aclnnRemainderTensorTensorGetWorkspaceSize(const aclTensor* self, const aclTensor* other, aclTensor* out, | ||||||
| 623 | - const aclTensor* self, const aclTensor* other, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor) | 614 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 624 | { | 615 | { | ||||||
| 625 | L2_DFX_PHASE_1(aclnnRemainderTensorTensor, DFX_IN(self, other), DFX_OUT(out)); | 616 | L2_DFX_PHASE_1(aclnnRemainderTensorTensor, DFX_IN(self, other), DFX_OUT(out)); | ||||||
| 626 | auto ret = CheckParamsTensorTensor(self, other, out); | 617 | auto ret = CheckParamsTensorTensor(self, other, out); | ||||||
| @@ -628,8 +619,8 @@ aclnnStatus aclnnRemainderTensorTensorGetWorkspaceSize( | |||||||||
| 628 | return ExecRemainderTensorTensorGetWorkspaceSize(self, other, out, workspaceSize, executor); | 619 | return ExecRemainderTensorTensorGetWorkspaceSize(self, other, out, workspaceSize, executor); | ||||||
| 629 | } | 620 | } | ||||||
| 630 | 621 | ||||||||
| 631 | -aclnnStatus aclnnRemainderTensorScalarGetWorkspaceSize( | 622 | +aclnnStatus aclnnRemainderTensorScalarGetWorkspaceSize(const aclTensor* self, const aclScalar* other, aclTensor* out, | ||||||
| 632 | - const aclTensor* self, const aclScalar* other, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor) | 623 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 633 | { | 624 | { | ||||||
| 634 | L2_DFX_PHASE_1(aclnnRemainderTensorScalar, DFX_IN(self, other), DFX_OUT(out)); | 625 | L2_DFX_PHASE_1(aclnnRemainderTensorScalar, DFX_IN(self, other), DFX_OUT(out)); | ||||||
| 635 | auto ret = CheckParamsTensorScalar(self, other, out); | 626 | auto ret = CheckParamsTensorScalar(self, other, out); | ||||||
| @@ -638,8 +629,8 @@ aclnnStatus aclnnRemainderTensorScalarGetWorkspaceSize( | |||||||||
| 638 | } | 629 | } | ||||||
| 639 | 630 | ||||||||
| 640 | // Scalar self, Tensor other | 631 | // Scalar self, Tensor other | ||||||
| 641 | -aclnnStatus aclnnRemainderScalarTensorGetWorkspaceSize( | 632 | +aclnnStatus aclnnRemainderScalarTensorGetWorkspaceSize(const aclScalar* self, const aclTensor* other, aclTensor* out, | ||||||
| 642 | - const aclScalar* self, const aclTensor* other, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor) | 633 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 643 | { | 634 | { | ||||||
| 644 | L2_DFX_PHASE_1(aclnnRemainderScalarTensor, DFX_IN(self, other), DFX_OUT(out)); | 635 | L2_DFX_PHASE_1(aclnnRemainderScalarTensor, DFX_IN(self, other), DFX_OUT(out)); | ||||||
| 645 | auto uniqueExecutor = CREATE_EXECUTOR(); | 636 | auto uniqueExecutor = CREATE_EXECUTOR(); | ||||||
| @@ -655,16 +646,15 @@ aclnnStatus aclnnRemainderScalarTensorGetWorkspaceSize( | |||||||||
| 655 | return ACLNN_SUCCESS; | 646 | return ACLNN_SUCCESS; | ||||||
| 656 | } | 647 | } | ||||||
| 657 | auto npuArch = op::GetCurrentPlatformInfo().GetCurNpuArch(); | 648 | auto npuArch = op::GetCurrentPlatformInfo().GetCurNpuArch(); | ||||||
| 658 | - if (IsRegBase(npuArch) && | 649 | + if (IsRegBase(npuArch)) { | ||||||
F | |||||||||
| 659 | - PromoteTypeScalarV35(other->GetDataType(), self->GetDataType()) != op::DataType::DT_DOUBLE) { | ||||||||
| 660 | auto castDtype = PromoteTypeScalarV35(other->GetDataType(), self->GetDataType()); | 650 | auto castDtype = PromoteTypeScalarV35(other->GetDataType(), self->GetDataType()); | ||||||
| 661 | auto otherContiguous = l0op::Contiguous(other, uniqueExecutor.get()); | 651 | auto otherContiguous = l0op::Contiguous(other, uniqueExecutor.get()); | ||||||
| 662 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 652 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 663 | auto otherCasted = l0op::Cast(otherContiguous, castDtype, uniqueExecutor.get()); | 653 | auto otherCasted = l0op::Cast(otherContiguous, castDtype, uniqueExecutor.get()); | ||||||
| 664 | CHECK_RET(otherCasted != nullptr, ACLNN_ERR_INNER_NULLPTR); | 654 | CHECK_RET(otherCasted != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 665 | 655 | ||||||||
| 666 | - auto floorModOpOut = | 656 | + auto floorModOpOut = l0op::FloorMod(uniqueExecutor.get()->ConvertToTensor(self, castDtype), otherCasted, | ||||||
| 667 | - l0op::FloorMod(uniqueExecutor.get()->ConvertToTensor(self, castDtype), otherCasted, uniqueExecutor.get()); | 657 | + uniqueExecutor.get()); | ||||||
| 668 | CHECK_RET(floorModOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | 658 | CHECK_RET(floorModOpOut != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
 🟠 High Priority 该变更移除了 但 对比同文件中另外两个类似函数 触发条件：在 RegBase（arch35）平台上，当 建议：恢复对 DT_DOUBLE 的排除条件，保持与同文件中其他 RegBase FloorMod 路径一致的守卫逻辑。如果确实需要启用 double 路径，则应先在 arch35 的 FloorMod tiling 和 kernel DAG 中补齐 DT_DOUBLE 支持，再移除该守卫。 改动建议
  | |||||||||
| 669 | 659 | ||||||||
| 670 | auto castOut = l0op::Cast(floorModOpOut, out->GetDataType(), uniqueExecutor.get()); | 660 | auto castOut = l0op::Cast(floorModOpOut, out->GetDataType(), uniqueExecutor.get()); | ||||||
| @@ -681,8 +671,8 @@ aclnnStatus aclnnRemainderScalarTensorGetWorkspaceSize( | |||||||||
| 681 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 671 | CHECK_RET(otherContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | ||||||
| 682 | 672 | ||||||||
| 683 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | 673 | bool needUnsqueeze = (GetTensorDimNum(out) == 0); | ||||||
| 684 | - auto remainderRes = | 674 | + auto remainderRes = RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, | ||||||
| 685 | - RemainderMainProcess(selfContiguous, otherContiguous, out, needUnsqueeze, uniqueExecutor.get()); | 675 | + uniqueExecutor.get()); | ||||||
| 686 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | 676 | CHECK_RET(remainderRes == ACLNN_SUCCESS, remainderRes); | ||||||
| 687 | } | 677 | } | ||||||
| 688 | 678 | ||||||||
| @@ -693,8 +683,8 @@ aclnnStatus aclnnRemainderScalarTensorGetWorkspaceSize( | |||||||||
| 693 | } | 683 | } | ||||||
| 694 | 684 | ||||||||
| 695 | // inplace | 685 | // inplace | ||||||
| 696 | -aclnnStatus aclnnInplaceRemainderTensorTensorGetWorkspaceSize( | 686 | +aclnnStatus aclnnInplaceRemainderTensorTensorGetWorkspaceSize(aclTensor* selfRef, const aclTensor* other, | ||||||
| 697 | - aclTensor* selfRef, const aclTensor* other, uint64_t* workspaceSize, aclOpExecutor** executor) | 687 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 698 | { | 688 | { | ||||||
| 699 | L2_DFX_PHASE_1(aclnnInplaceRemainderTensorTensor, DFX_IN(selfRef, other), DFX_OUT(selfRef)); | 689 | L2_DFX_PHASE_1(aclnnInplaceRemainderTensorTensor, DFX_IN(selfRef, other), DFX_OUT(selfRef)); | ||||||
| 700 | auto out = const_cast<aclTensor*>(selfRef); | 690 | auto out = const_cast<aclTensor*>(selfRef); | ||||||
| @@ -703,8 +693,8 @@ aclnnStatus aclnnInplaceRemainderTensorTensorGetWorkspaceSize( | |||||||||
| 703 | return ExecRemainderTensorTensorGetWorkspaceSize(selfRef, other, out, workspaceSize, executor); | 693 | return ExecRemainderTensorTensorGetWorkspaceSize(selfRef, other, out, workspaceSize, executor); | ||||||
| 704 | } | 694 | } | ||||||
| 705 | 695 | ||||||||
| 706 | -aclnnStatus aclnnInplaceRemainderTensorScalarGetWorkspaceSize( | 696 | +aclnnStatus aclnnInplaceRemainderTensorScalarGetWorkspaceSize(aclTensor* selfRef, const aclScalar* other, | ||||||
| 707 | - aclTensor* selfRef, const aclScalar* other, uint64_t* workspaceSize, aclOpExecutor** executor) | 697 | + uint64_t* workspaceSize, aclOpExecutor** executor) | ||||||
| 708 | { | 698 | { | ||||||
| 709 | L2_DFX_PHASE_1(aclnnInplaceRemainderTensorScalar, DFX_IN(selfRef, other), DFX_OUT(selfRef)); | 699 | L2_DFX_PHASE_1(aclnnInplaceRemainderTensorScalar, DFX_IN(selfRef, other), DFX_OUT(selfRef)); | ||||||
| 710 | auto out = const_cast<aclTensor*>(selfRef); | 700 | auto out = const_cast<aclTensor*>(selfRef); | ||||||
| @@ -714,40 +704,40 @@ aclnnStatus aclnnInplaceRemainderTensorScalarGetWorkspaceSize( | |||||||||
| 714 | } | 704 | } | ||||||
| 715 | 705 | ||||||||
| 716 | // Tensor self, Tensor other | 706 | // Tensor self, Tensor other | ||||||
| 717 | -aclnnStatus aclnnRemainderTensorTensor( | 707 | +aclnnStatus aclnnRemainderTensorTensor(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, | ||||||
| 718 | - void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) | 708 | + aclrtStream stream) | ||||||
| 719 | { | 709 | { | ||||||
| 720 | L2_DFX_PHASE_2(aclnnRemainderTensorTensor); | 710 | L2_DFX_PHASE_2(aclnnRemainderTensorTensor); | ||||||
| 721 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 711 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | ||||||
| 722 | } | 712 | } | ||||||
| 723 | 713 | ||||||||
| 724 | // Tensor self, Scalar other | 714 | // Tensor self, Scalar other | ||||||
| 725 | -aclnnStatus aclnnRemainderTensorScalar( | 715 | +aclnnStatus aclnnRemainderTensorScalar(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, | ||||||
| 726 | - void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) | 716 | + aclrtStream stream) | ||||||
| 727 | { | 717 | { | ||||||
| 728 | L2_DFX_PHASE_2(aclnnRemainderTensorScalar); | 718 | L2_DFX_PHASE_2(aclnnRemainderTensorScalar); | ||||||
| 729 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 719 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | ||||||
| 730 | } | 720 | } | ||||||
| 731 | 721 | ||||||||
| 732 | // Scalar self, Tensor other | 722 | // Scalar self, Tensor other | ||||||
| 733 | -aclnnStatus aclnnRemainderScalarTensor( | 723 | +aclnnStatus aclnnRemainderScalarTensor(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, | ||||||
| 734 | - void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) | 724 | + aclrtStream stream) | ||||||
| 735 | { | 725 | { | ||||||
| 736 | L2_DFX_PHASE_2(aclnnRemainderScalarTensor); | 726 | L2_DFX_PHASE_2(aclnnRemainderScalarTensor); | ||||||
| 737 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 727 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | ||||||
| 738 | } | 728 | } | ||||||
| 739 | 729 | ||||||||
| 740 | // Tensor self, Tensor other | 730 | // Tensor self, Tensor other | ||||||
| 741 | -aclnnStatus aclnnInplaceRemainderTensorTensor( | 731 | +aclnnStatus aclnnInplaceRemainderTensorTensor(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, | ||||||
| 742 | - void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) | 732 | + aclrtStream stream) | ||||||
| 743 | { | 733 | { | ||||||
| 744 | L2_DFX_PHASE_2(aclnnInplaceRemainderTensorTensor); | 734 | L2_DFX_PHASE_2(aclnnInplaceRemainderTensorTensor); | ||||||
| 745 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 735 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | ||||||
| 746 | } | 736 | } | ||||||
| 747 | 737 | ||||||||
| 748 | // Tensor self, Scalar other | 738 | // Tensor self, Scalar other | ||||||
| 749 | -aclnnStatus aclnnInplaceRemainderTensorScalar( | 739 | +aclnnStatus aclnnInplaceRemainderTensorScalar(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, | ||||||
| 750 | - void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream) | 740 | + aclrtStream stream) | ||||||
| 751 | { | 741 | { | ||||||
| 752 | L2_DFX_PHASE_2(aclnnInplaceRemainderTensorScalar); | 742 | L2_DFX_PHASE_2(aclnnInplaceRemainderTensorScalar); | ||||||
| 753 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 743 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | ||||||
| @@ -0,0 +1,39 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 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 | + | ||
| 13 | + | ||
| 14 | + | ||
| 15 | + | ||
| 16 | + | ||
| 17 | + | ||
| 18 | + | ||
| 19 | + | ||
| 20 | + | ||
| 21 | +using namespace op; | ||
| 22 | + | ||
| 23 | +class l2_remainder_scalar_tensor_ascend950_test : public testing::Test { | ||
| 24 | +protected: | ||
| 25 | + void TearDown() override { SetPlatformNpuArch(NpuArch::DAV_2201); } | ||
| 26 | +}; | ||
| 27 | + | ||
| 28 | +TEST_F(l2_remainder_scalar_tensor_ascend950_test, double_scalar_int32_tensor_to_float16) | ||
| 29 | +{ | ||
| 30 | + SetPlatformNpuArch(NpuArch::DAV_3510); | ||
| 31 | + | ||
| 32 | + auto self = ScalarDesc(2049.0); | ||
| 33 | + auto other = TensorDesc({1}, ACL_INT32, ACL_FORMAT_ND).Value(std::vector<int32_t>{2}); | ||
| 34 | + auto out = TensorDesc({1}, ACL_FLOAT16, ACL_FORMAT_ND).Precision(0.0, 0.0); | ||
| 35 | + auto ut = OP_API_UT(aclnnRemainderScalarTensor, INPUT(self, other), OUTPUT(out)); | ||
| 36 | + | ||
| 37 | + uint64_t workspaceSize = 0; | ||
| 38 | + EXPECT_EQ(ut.TestGetWorkspaceSize(&workspaceSize), ACL_SUCCESS); | ||
| 39 | +} | ||
直接删除之后没有看到别的地方有对double处理