Star1283
1397
代码
代码
Issues95
Pull Requests174
Actions
项目讨论
Wiki
分析
Star1283
1397
已合并
aclnnRandom支持非连续 #4422
wushuai创建于 17 天前
aclnnRandom支持非连续 #4422
已合并
wushuai创建于 17 天前
从wu-shuai2580/ops-math_07160804合入到cann/ops-mathmaster
共 2 个文件变更+45-25
Mrandom/dsa_random_uniform/docs/aclnnInplaceRandomTensor.md+7-3
@@ -85,7 +85,7 @@ aclnnStatus aclnnInplaceRandomTensor(
85 <td>BFLOAT16、FLOAT16、FLOAT32、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、COMPLEX64、COMPLEX128</td>85 <td>BFLOAT16、FLOAT16、FLOAT32、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、COMPLEX64、COMPLEX128</td>
86 <td>ND</td>86 <td>ND</td>
87 <td>0-8</td>87 <td>0-8</td>
88- <td>×</td>88+ <td>√</td>
89 </tr>89 </tr>
90 <tr>90 <tr>
91 <td>from</td>91 <td>from</td>
@@ -159,6 +159,10 @@ aclnnStatus aclnnInplaceRandomTensor(
159 </tr>159 </tr>
160 </tbody></table>160 </tbody></table>
161 161 
162+<!-- npu="950" id8 -->
163+ - <term>Ascend 950PR/Ascend 950DT</term>：数据类型不支持COMPLEX64、COMPLEX128。
164+ <!-- end id8 -->
165+ 
162- **返回值：**166- **返回值：**
163 167 
164 aclnnStatus：返回状态码，具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。168 aclnnStatus：返回状态码，具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。
@@ -191,7 +195,7 @@ aclnnStatus aclnnInplaceRandomTensor(
191 <td>from大于等于to。</td>195 <td>from大于等于to。</td>
192 </tr>196 </tr>
193 <tr>197 <tr>
194- <td>to-1超过selfRef数据类型的取值范围。</td>198+ <td>selfRef的shape超过8维。</td>
195 </tr>199 </tr>
196 </tbody>200 </tbody>
197 </table>201 </table>
@@ -246,7 +250,7 @@ aclnnStatus aclnnInplaceRandomTensor(
246 250 
247<!-- npu="950" id7 -->251<!-- npu="950" id7 -->
248- <term>Ascend 950PR/Ascend 950DT</term>：252- <term>Ascend 950PR/Ascend 950DT</term>：
249- - offset必须为4的倍数。253+ - 通过aclnnSetPytorchRandom设置对标PyTorch模式时，offset必须为4的倍数。
250 - from必须小于to，且from/to-1指定范围不可超出self数据类型的范围。254 - from必须小于to，且from/to-1指定范围不可超出self数据类型的范围。
251 255 
252<!-- end id7 -->256<!-- end id7 -->
Mrandom/dsa_random_uniform/op_host/op_api/aclnn_random.cpp+38-22
@@ -136,6 +136,11 @@ static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST = {
136 op::DataType::DT_INT16, op::DataType::DT_INT8, op::DataType::DT_UINT8, op::DataType::DT_BOOL,136 op::DataType::DT_INT16, op::DataType::DT_INT8, op::DataType::DT_UINT8, op::DataType::DT_BOOL,
137 op::DataType::DT_DOUBLE, op::DataType::DT_BF16, op::DataType::DT_COMPLEX64, op::DataType::DT_COMPLEX128};137 op::DataType::DT_DOUBLE, op::DataType::DT_BF16, op::DataType::DT_COMPLEX64, op::DataType::DT_COMPLEX128};
138 138 
139+static const std::initializer_list<op::DataType> REGBASE_DTYPE_SUPPORT_LIST = {
140+ op::DataType::DT_FLOAT, op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_FLOAT16,
141+ op::DataType::DT_INT16, op::DataType::DT_INT8, op::DataType::DT_UINT8, op::DataType::DT_BOOL,
142+ op::DataType::DT_DOUBLE, op::DataType::DT_BF16};
143+ 
139static const std::initializer_list<op::DataType> INT_DTYPE_LIST = {op::DataType::DT_INT32, op::DataType::DT_INT64,144static const std::initializer_list<op::DataType> INT_DTYPE_LIST = {op::DataType::DT_INT32, op::DataType::DT_INT64,
140 op::DataType::DT_INT16, op::DataType::DT_INT8,145 op::DataType::DT_INT16, op::DataType::DT_INT8,
141 op::DataType::DT_UINT8, op::DataType::DT_BOOL};146 op::DataType::DT_UINT8, op::DataType::DT_BOOL};
@@ -150,7 +155,12 @@ static bool CheckNotNull(const aclTensor* selfRef)
150 155 
151static bool CheckDtypeValid(const aclTensor* selfRef)156static bool CheckDtypeValid(const aclTensor* selfRef)
152{157{
153- OP_CHECK_DTYPE_NOT_SUPPORT(selfRef, DTYPE_SUPPORT_LIST, return false);158+ if (IsRegBase()) {
159+ OP_CHECK_DTYPE_NOT_SUPPORT(selfRef, REGBASE_DTYPE_SUPPORT_LIST, return false);
160+ } else {
161+ OP_CHECK_DTYPE_NOT_SUPPORT(selfRef, DTYPE_SUPPORT_LIST, return false);
162+ }
163+ 
154 return true;164 return true;
155}165}
156 166 
@@ -345,10 +355,10 @@ aclnnStatus aclnnInplaceRandomGetWorkspaceSize(const aclTensor* selfRef, int64_t
345 int64_t offset, uint64_t* workspaceSize, aclOpExecutor** executor)355 int64_t offset, uint64_t* workspaceSize, aclOpExecutor** executor)
346{356{
347 L2_DFX_PHASE_1(aclnnInplaceRandom, DFX_IN(selfRef, from, to, seed, offset), DFX_OUT(selfRef));357 L2_DFX_PHASE_1(aclnnInplaceRandom, DFX_IN(selfRef, from, to, seed, offset), DFX_OUT(selfRef));
358+ auto out = const_cast<aclTensor*>(selfRef);
348 auto ret = CheckParams(selfRef);359 auto ret = CheckParams(selfRef);
349 CHECK_RET(ret == ACLNN_SUCCESS, ret);360 CHECK_RET(ret == ACLNN_SUCCESS, ret);
350 CHECK_RET(CheckFromToRange(from, to), ACLNN_ERR_PARAM_INVALID);361 CHECK_RET(CheckFromToRange(from, to), ACLNN_ERR_PARAM_INVALID);
351- 
352 auto uniqueExecutor = CREATE_EXECUTOR();362 auto uniqueExecutor = CREATE_EXECUTOR();
353 CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);363 CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
354 364 
@@ -357,16 +367,18 @@ aclnnStatus aclnnInplaceRandomGetWorkspaceSize(const aclTensor* selfRef, int64_t
357 uniqueExecutor.ReleaseTo(executor);367 uniqueExecutor.ReleaseTo(executor);
358 return ACLNN_SUCCESS;368 return ACLNN_SUCCESS;
359 }369 }
370+ auto selfContiguous = l0op::Contiguous(selfRef, uniqueExecutor.get());
371+ CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_PARAM_NULLPTR);
360 372 
361 // 更新from、to，将from、to移动到下一个最接近且不会超出[from, to)范围的值373 // 更新from、to，将from、to移动到下一个最接近且不会超出[from, to)范围的值
362- ret = updateFrom(from, selfRef->GetDataType());374+ ret = updateFrom(from, selfContiguous->GetDataType());
363 CHECK_RET(ret == ACLNN_SUCCESS, ret);375 CHECK_RET(ret == ACLNN_SUCCESS, ret);
364- ret = updateTo(to, selfRef->GetDataType());376+ ret = updateTo(to, selfContiguous->GetDataType());
365 CHECK_RET(ret == ACLNN_SUCCESS, ret);377 CHECK_RET(ret == ACLNN_SUCCESS, ret);
366 378 
367 const aclTensor* computeOut = nullptr;379 const aclTensor* computeOut = nullptr;
368 if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_2201) {380 if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_2201) {
369- auto inputShape = op::ToShapeVector(selfRef->GetViewShape());381+ auto inputShape = op::ToShapeVector(selfContiguous->GetViewShape());
370 auto inputShapeArray = uniqueExecutor.get()->AllocIntArray(inputShape.data(), inputShape.size());382 auto inputShapeArray = uniqueExecutor.get()->AllocIntArray(inputShape.data(), inputShape.size());
371 CHECK_RET(inputShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);383 CHECK_RET(inputShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
372 auto low = uniqueExecutor.get()->AllocScalar(static_cast<float>(from));384 auto low = uniqueExecutor.get()->AllocScalar(static_cast<float>(from));
@@ -375,20 +387,20 @@ aclnnStatus aclnnInplaceRandomGetWorkspaceSize(const aclTensor* selfRef, int64_t
375 CHECK_RET(high != nullptr, ACLNN_ERR_INNER_NULLPTR);387 CHECK_RET(high != nullptr, ACLNN_ERR_INNER_NULLPTR);
376 computeOut = l0op::DSARandomUniform(inputShapeArray, seed, offset, low, high, uniqueExecutor.get());388 computeOut = l0op::DSARandomUniform(inputShapeArray, seed, offset, low, high, uniqueExecutor.get());
377 } else {389 } else {
378- computeOut = randomDavidPath(selfRef, seed, offset, from, to, uniqueExecutor.get());390+ computeOut = randomDavidPath(selfContiguous, seed, offset, from, to, uniqueExecutor.get());
379 }391 }
380 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);392 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
381 393 
382- if (selfRef->GetDataType() == op::DataType::DT_BOOL) {394+ if (selfContiguous->GetDataType() == op::DataType::DT_BOOL) {
383 int64_t decimals = 0;395 int64_t decimals = 0;
384 computeOut = l0op::RoundDecimals(computeOut, decimals, uniqueExecutor.get());396 computeOut = l0op::RoundDecimals(computeOut, decimals, uniqueExecutor.get());
385 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);397 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
386 }398 }
387 399 
388- auto castResult = CastProcess(selfRef, computeOut, uniqueExecutor.get());400+ auto castResult = CastProcess(out, computeOut, uniqueExecutor.get());
389 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);401 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
390 402 
391- auto viewCopyResult = l0op::ViewCopy(castResult, selfRef, uniqueExecutor.get());403+ auto viewCopyResult = l0op::ViewCopy(castResult, out, uniqueExecutor.get());
392 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);404 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
393 405 
394 *workspaceSize = uniqueExecutor->GetWorkspaceSize();406 *workspaceSize = uniqueExecutor->GetWorkspaceSize();
@@ -409,7 +421,7 @@ aclnnStatus aclnnInplaceRandomTensorGetWorkspaceSize(const aclTensor* selfRef, i
409{421{
410 L2_DFX_PHASE_1(aclnnInplaceRandomTensor, DFX_IN(selfRef, from, to, seedTensor, offsetTensor, offset),422 L2_DFX_PHASE_1(aclnnInplaceRandomTensor, DFX_IN(selfRef, from, to, seedTensor, offsetTensor, offset),
411 DFX_OUT(selfRef));423 DFX_OUT(selfRef));
412- 424+ auto out = const_cast<aclTensor*>(selfRef);
413 auto ret = CheckParams(selfRef);425 auto ret = CheckParams(selfRef);
414 CHECK_RET(ret == ACLNN_SUCCESS, ret);426 CHECK_RET(ret == ACLNN_SUCCESS, ret);
415 CHECK_RET(CheckFromToRange(from, to), ACLNN_ERR_PARAM_INVALID);427 CHECK_RET(CheckFromToRange(from, to), ACLNN_ERR_PARAM_INVALID);
@@ -421,18 +433,20 @@ aclnnStatus aclnnInplaceRandomTensorGetWorkspaceSize(const aclTensor* selfRef, i
421 uniqueExecutor.ReleaseTo(executor);433 uniqueExecutor.ReleaseTo(executor);
422 return ACLNN_SUCCESS;434 return ACLNN_SUCCESS;
423 }435 }
436+ auto selfContiguous = l0op::Contiguous(selfRef, uniqueExecutor.get());
437+ CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_PARAM_NULLPTR);
424 438 
425 // 更新from、to，将from、to移动到下一个最接近且不会超出[from, to)范围的值439 // 更新from、to，将from、to移动到下一个最接近且不会超出[from, to)范围的值
426- ret = updateFrom(from, selfRef->GetDataType());440+ ret = updateFrom(from, selfContiguous->GetDataType());
427 CHECK_RET(ret == ACLNN_SUCCESS, ret);441 CHECK_RET(ret == ACLNN_SUCCESS, ret);
428- ret = updateTo(to, selfRef->GetDataType());442+ ret = updateTo(to, selfContiguous->GetDataType());
429 CHECK_RET(ret == ACLNN_SUCCESS, ret);443 CHECK_RET(ret == ACLNN_SUCCESS, ret);
430 444 
431 const aclTensor* computeOut = nullptr;445 const aclTensor* computeOut = nullptr;
432 if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_2201) {446 if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_2201) {
433 auto concatTensor = ProcessOffsetTensor(offsetTensor, offset, uniqueExecutor.get());447 auto concatTensor = ProcessOffsetTensor(offsetTensor, offset, uniqueExecutor.get());
434 CHECK_RET(concatTensor != nullptr, ACLNN_ERR_INNER_NULLPTR);448 CHECK_RET(concatTensor != nullptr, ACLNN_ERR_INNER_NULLPTR);
435- auto inputShape = op::ToShapeVector(selfRef->GetViewShape());449+ auto inputShape = op::ToShapeVector(selfContiguous->GetViewShape());
436 auto inputShapeArray = uniqueExecutor.get()->AllocIntArray(inputShape.data(), inputShape.size());450 auto inputShapeArray = uniqueExecutor.get()->AllocIntArray(inputShape.data(), inputShape.size());
437 CHECK_RET(inputShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);451 CHECK_RET(inputShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
438 auto low = uniqueExecutor.get()->AllocScalar(static_cast<float>(from));452 auto low = uniqueExecutor.get()->AllocScalar(static_cast<float>(from));
@@ -443,20 +457,21 @@ aclnnStatus aclnnInplaceRandomTensorGetWorkspaceSize(const aclTensor* selfRef, i
443 computeOut = l0op::DSARandomUniformTensor(inputShapeArray, seedTensor, concatTensor, low, high,457 computeOut = l0op::DSARandomUniformTensor(inputShapeArray, seedTensor, concatTensor, low, high,
444 uniqueExecutor.get());458 uniqueExecutor.get());
445 } else {459 } else {
446- computeOut = randomTensorDavidPath(selfRef, seedTensor, offsetTensor, offset, from, to, uniqueExecutor.get());460+ computeOut = randomTensorDavidPath(selfContiguous, seedTensor, offsetTensor, offset, from, to,
461+ uniqueExecutor.get());
447 }462 }
448 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);463 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
449 464 
450- if (selfRef->GetDataType() == op::DataType::DT_BOOL) {465+ if (out->GetDataType() == op::DataType::DT_BOOL) {
451 int64_t decimals = 0;466 int64_t decimals = 0;
452 computeOut = l0op::RoundDecimals(computeOut, decimals, uniqueExecutor.get());467 computeOut = l0op::RoundDecimals(computeOut, decimals, uniqueExecutor.get());
453 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);468 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
454 }469 }
455 470 
456- auto castResult = CastProcess(selfRef, computeOut, uniqueExecutor.get());471+ auto castResult = CastProcess(out, computeOut, uniqueExecutor.get());
457 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);472 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
458 473 
459- auto viewCopyResult = l0op::ViewCopy(castResult, selfRef, uniqueExecutor.get());474+ auto viewCopyResult = l0op::ViewCopy(castResult, out, uniqueExecutor.get());
460 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);475 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
461 476 
462 *workspaceSize = uniqueExecutor->GetWorkspaceSize();477 *workspaceSize = uniqueExecutor->GetWorkspaceSize();
@@ -476,6 +491,7 @@ aclnnStatus aclnnInplaceRandomWithoutFromToGetWorkspaceSize(const aclTensor* sel
476 uint64_t* workspaceSize, aclOpExecutor** executor)491 uint64_t* workspaceSize, aclOpExecutor** executor)
477{492{
478 L2_DFX_PHASE_1(aclnnInplaceRandomWithoutFromTo, DFX_IN(selfRef, seed, offset), DFX_OUT(selfRef));493 L2_DFX_PHASE_1(aclnnInplaceRandomWithoutFromTo, DFX_IN(selfRef, seed, offset), DFX_OUT(selfRef));
494+ auto out = const_cast<aclTensor*>(selfRef);
479 auto ret = CheckParams(selfRef);495 auto ret = CheckParams(selfRef);
480 CHECK_RET(ret == ACLNN_SUCCESS, ret);496 CHECK_RET(ret == ACLNN_SUCCESS, ret);
481 auto uniqueExecutor = CREATE_EXECUTOR();497 auto uniqueExecutor = CREATE_EXECUTOR();
@@ -502,10 +518,10 @@ aclnnStatus aclnnInplaceRandomWithoutFromToGetWorkspaceSize(const aclTensor* sel
502 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);518 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
503 }519 }
504 520 
505- auto castResult = CastProcess(selfContiguous, computeOut, uniqueExecutor.get());521+ auto castResult = CastProcess(out, computeOut, uniqueExecutor.get());
506 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);522 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
507 523 
508- auto viewCopyResult = l0op::ViewCopy(castResult, selfRef, uniqueExecutor.get());524+ auto viewCopyResult = l0op::ViewCopy(castResult, out, uniqueExecutor.get());
509 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);525 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
510 526 
511 *workspaceSize = uniqueExecutor->GetWorkspaceSize();527 *workspaceSize = uniqueExecutor->GetWorkspaceSize();
@@ -527,7 +543,7 @@ aclnnStatus aclnnInplaceRandomWithoutFromToTensorGetWorkspaceSize(const aclTenso
527{543{
528 L2_DFX_PHASE_1(aclnnInplaceRandomWithoutFromToTensor, DFX_IN(selfRef, seedTensor, offsetTensor, offset),544 L2_DFX_PHASE_1(aclnnInplaceRandomWithoutFromToTensor, DFX_IN(selfRef, seedTensor, offsetTensor, offset),
529 DFX_OUT(selfRef));545 DFX_OUT(selfRef));
530- 546+ auto out = const_cast<aclTensor*>(selfRef);
531 auto ret = CheckParams(selfRef);547 auto ret = CheckParams(selfRef);
532 CHECK_RET(ret == ACLNN_SUCCESS, ret);548 CHECK_RET(ret == ACLNN_SUCCESS, ret);
533 auto uniqueExecutor = CREATE_EXECUTOR();549 auto uniqueExecutor = CREATE_EXECUTOR();
@@ -571,10 +587,10 @@ aclnnStatus aclnnInplaceRandomWithoutFromToTensorGetWorkspaceSize(const aclTenso
571 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);587 CHECK_RET(computeOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
572 }588 }
573 589 
574- auto castResult = CastProcess(selfContiguous, computeOut, uniqueExecutor.get());590+ auto castResult = CastProcess(out, computeOut, uniqueExecutor.get());
575 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);591 CHECK_RET(castResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
576 592 
577- auto viewCopyResult = l0op::ViewCopy(castResult, selfRef, uniqueExecutor.get());593+ auto viewCopyResult = l0op::ViewCopy(castResult, out, uniqueExecutor.get());
578 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);594 CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
579 595 
580 *workspaceSize = uniqueExecutor->GetWorkspaceSize();596 *workspaceSize = uniqueExecutor->GetWorkspaceSize();