为AclnnNormalTensorTensor接口增加广播功能-ops-math-AtomGit

已合并

为AclnnNormalTensorTensor接口增加广播功能 #2146

huairuchen创建于 4月9日从已删除 :pr_broadcast 合入到 cann/ops-math： master

为AclnnNormalTensorTensor接口增加广播功能 #2146

已合并

huairuchen创建于 4月9日

从已删除 :pr_broadcast合入到cann/ops-math master

共 2 个文件变更+49-9

Mrandom/stateless_random_normal_v2/op_api/aclnn_normal_out.cpp+33-7

	@@ -16,6 +16,7 @@
16	#include "aclnn_kernels/cast.h"	16	#include "aclnn_kernels/cast.h"
17	#include "conversion/view_copy/op_api/view_copy.h"	17	#include "conversion/view_copy/op_api/view_copy.h"
18	#include "aclnn_kernels/contiguous.h"	18	#include "aclnn_kernels/contiguous.h"
		19	+#include "conversion/broadcast_to/op_api/broadcast_to.h"
19	#include "aclnn/aclnn_base.h"	20	#include "aclnn/aclnn_base.h"
20	#include "opdev/shape_utils.h"	21	#include "opdev/shape_utils.h"
21	#include "aclnn_kernels/common/op_error_check.h"	22	#include "aclnn_kernels/common/op_error_check.h"
	@@ -218,13 +219,38 @@ aclnnStatus CommonLogicGeneralNormal(
218	const aclTensor* addOut = nullptr;	219	const aclTensor* addOut = nullptr;
219		220
220	if(GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510 && self->GetDataType() != DataType::DT_DOUBLE){	221	if(GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510 && self->GetDataType() != DataType::DT_DOUBLE){
221	- // 调用normal_算子kernel function(AI Core算子)	222	+ // V3 kernel 要求 mean/std 参数为 DT_FLOAT
222	- // V3 kernel 要求 mean/std 参数为 DT_FLOAT，与 InplaceNormal 保持一致	223	+ auto meanCasted = l0op::Cast(mean, DataType::DT_FLOAT, uniqueExecutor.get());
223	- auto meanFP32 = l0op::Cast(mean, DataType::DT_FLOAT, uniqueExecutor.get());	224	+ CHECK_RET(meanCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
224	- CHECK_RET(meanFP32 != nullptr, ACLNN_ERR_INNER_NULLPTR);	225	+ auto stdCasted = l0op::Cast(std, DataType::DT_FLOAT, uniqueExecutor.get());
225	- auto stdFP32 = l0op::Cast(std, DataType::DT_FLOAT, uniqueExecutor.get());	226	+ CHECK_RET(stdCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
226	- CHECK_RET(stdFP32 != nullptr, ACLNN_ERR_INNER_NULLPTR);	227	+
227	- addOut = l0op::StatelessRandomNormalV3(self, keyArr, counterArr, meanFP32, stdFP32, uniqueExecutor.get());	228	+ // V3 kernel 不支持广播，需要将 shape 与 out 不一致的多元素 tensor 广播到 out shape
		229	+ // Size()==1 的 tensor（scalar 转换而来）由 V3 kernel 的 scalar 路径原生处理，无需广播
		230	+ // self 需要单独广播以保持原始 dtype（决定 V3 kernel 的计算精度）
		231	+ bool selfNeedBcast = self->GetViewShape() != out->GetViewShape();
		232	+ bool meanNeedBcast = meanCasted->GetViewShape() != out->GetViewShape() && mean->Size() > 1;
		233	+ bool stdNeedBcast = stdCasted->GetViewShape() != out->GetViewShape() && std->Size() > 1;
		234	+ if (selfNeedBcast \|\| meanNeedBcast \|\| stdNeedBcast) {
		235	+ op::FVector<int64_t, op::MAX_DIM_NUM> outDims = op::ToShapeVector(out->GetViewShape());
		236	+ auto outShapeArray = uniqueExecutor.get()->AllocIntArray(outDims.data(), outDims.size());
		237	+ CHECK_RET(outShapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
		238	+
		239	+ if (selfNeedBcast) {
		240	+ auto selfBroadcast = l0op::BroadcastTo(self, outShapeArray, uniqueExecutor.get());
		241	+ CHECK_RET(selfBroadcast != nullptr, ACLNN_ERR_INNER_NULLPTR);
		242	+ self = const_cast<aclTensor*>(selfBroadcast);
		243	+ }
		244	+ if (meanNeedBcast) {
		245	+ meanCasted = l0op::BroadcastTo(meanCasted, outShapeArray, uniqueExecutor.get());
		246	+ CHECK_RET(meanCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
		247	+ }
		248	+ if (stdNeedBcast) {
		249	+ stdCasted = l0op::BroadcastTo(stdCasted, outShapeArray, uniqueExecutor.get());
		250	+ CHECK_RET(stdCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
		251	+ }
		252	+ }
		253	+ addOut = l0op::StatelessRandomNormalV3(self, keyArr, counterArr, meanCasted, stdCasted, uniqueExecutor.get());
228	}	254	}
229	else{	255	else{
230	// 调用normal_算子kernel function(AI Cpu算子)	256	// 调用normal_算子kernel function(AI Cpu算子)

Mrandom/stateless_random_normal_v3/op_host/arch35/stateless_random_normal_v3_tiling_arch35.cpp+16-2

	@@ -93,16 +93,30 @@ ge::graphStatus StatelessRandomNormalV3Tiling::UniqueProcess()
93	{	93	{
94	uint32_t v3KernelMode = 0;	94	uint32_t v3KernelMode = 0;
95		95
		96	+ auto outputShape = context_->GetOutputShape(OUTPUT_IDX_Y);
		97	+ OP_CHECK_NULL_WITH_CONTEXT(context_, outputShape);
		98	+ int64_t outputSize = outputShape->GetStorageShape().GetShapeSize();
		99	+
96	auto meanTensor = context_->GetInputTensor(INPUT_IDX_MEAN);	100	auto meanTensor = context_->GetInputTensor(INPUT_IDX_MEAN);
97	OP_CHECK_NULL_WITH_CONTEXT(context_, meanTensor);	101	OP_CHECK_NULL_WITH_CONTEXT(context_, meanTensor);
98	- if (meanTensor->GetShapeSize() == 1) {	102	+ int64_t meanSize = meanTensor->GetShapeSize();
		103	+ if (meanSize == 1) {
99	v3KernelMode \|= MEAN_SCALAR_FLAG;	104	v3KernelMode \|= MEAN_SCALAR_FLAG;
		105	+ } else if (meanSize != outputSize) {
		106	+ OP_LOGE(context_, "StatelessRandomNormalV3 does not support broadcast for mean, "
		107	+ "mean shapeSize: %ld != output shapeSize: %ld.", meanSize, outputSize);
		108	+ return ge::GRAPH_FAILED;
100	}	109	}
101		110
102	auto stdevTensor = context_->GetInputTensor(INPUT_IDX_STDEV);	111	auto stdevTensor = context_->GetInputTensor(INPUT_IDX_STDEV);
103	OP_CHECK_NULL_WITH_CONTEXT(context_, stdevTensor);	112	OP_CHECK_NULL_WITH_CONTEXT(context_, stdevTensor);
104	- if (stdevTensor->GetShapeSize() == 1) {	113	+ int64_t stdevSize = stdevTensor->GetShapeSize();
		114	+ if (stdevSize == 1) {
105	v3KernelMode \|= STDEV_SCALAR_FLAG;	115	v3KernelMode \|= STDEV_SCALAR_FLAG;
		116	+ } else if (stdevSize != outputSize) {
		117	+ OP_LOGE(context_, "StatelessRandomNormalV3 does not support broadcast for stdev, "
		118	+ "stdev shapeSize: %ld != output shapeSize: %ld.", stdevSize, outputSize);
		119	+ return ge::GRAPH_FAILED;
106	}	120	}
107		121
108	tilingData_.v3KernelMode = v3KernelMode;	122	tilingData_.v3KernelMode = v3KernelMode;