Merged
fix(mse_loss): MseLossOut空tensor输入按reduction语义填充输出 #14080
fix(mse_loss): MseLossOut空tensor输入按reduction语义填充输出 #14080
Merged
佐助大王Add on 11 days ago
Total: 4 files changed+141-2
@@ -247,6 +247,10 @@ aclnnStatus aclnnMseLossOut(
247 247 
248- 确定性计算:248- 确定性计算:
249 - aclnnMseLossOut默认确定性实现。249 - aclnnMseLossOut默认确定性实现。
250+- 空 tensor 输入(元素个数为0)的输出语义:
251+ - reduction为0('none')时,输出为空 tensor,不写入任何元素;
252+ - reduction为1('mean')时,输出填充为 NaN;
253+ - reduction为2('sum')时,输出填充为 0。
250 254 
251## 调用示例255## 调用示例
252 256 
@@ -249,6 +249,10 @@ aclnnStatus aclnnMseLossOut(
249 249 
250- Deterministic computation:250- Deterministic computation:
251 - aclnnMseLossOut defaults to a deterministic implementation.251 - aclnnMseLossOut defaults to a deterministic implementation.
252+- Output semantics for empty tensor inputs (with 0 elements):
253+ - When reduction is 0 ('none'), the output is an empty tensor and no element is written.
254+ - When reduction is 1 ('mean'), the output is filled with NaN.
255+ - When reduction is 2 ('sum'), the output is filled with 0.
252 256 
253## Example257## Example
254 258 
@@ -102,6 +102,31 @@ inline static const std::string& GetReductionStr(int64_t reduction)
102 }102 }
103}103}
104 104 
105+// 空tensor场景按0轴归约语义填充out:mean时0轴ReduceMean结果逐元素为NaN,sum时0轴ReduceSum结果逐元素为0,
106+// 需按out的实际view shape整体填充,out为空tensor时无需写入
107+static aclnnStatus FillOutForEmptyInput(aclTensor* out, float val, aclOpExecutor* executor)
108+{
109+ if (out->GetViewShape().GetShapeSize() == 0) {
110+ return ACLNN_SUCCESS;
111+ }
112+ auto outShape = op::ToShapeVector(out->GetViewShape());
113+ auto dims = executor->ConvertToTensor(outShape.data(), outShape.size(), op::DataType::DT_INT64);
114+ CHECK_RET(dims != nullptr, ACLNN_ERR_INNER_NULLPTR);
115+ auto shapeArray = executor->AllocIntArray(outShape.data(), outShape.size());
116+ CHECK_RET(shapeArray != nullptr, ACLNN_ERR_INNER_NULLPTR);
117+ 
118+ FVector<float> valVector = {val};
119+ auto valTensor = executor->ConvertToTensor(valVector.data(), valVector.size(), out->GetDataType());
120+ CHECK_RET(valTensor != nullptr, ACLNN_ERR_INNER_NULLPTR);
121+ 
122+ auto fillOut = l0op::Fill(dims, valTensor, shapeArray, executor);
123+ CHECK_RET(fillOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
124+ // 固定写法,将填充结果拷贝到输出out上,out可能是非连续的tensor
125+ auto viewCopyResult = l0op::ViewCopy(fillOut, out, executor);
126+ CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
127+ return ACLNN_SUCCESS;
128+}
129+ 
105aclnnStatus aclnnMseLossOutGetWorkspaceSize(const aclTensor* self, const aclTensor* target, int64_t reduction,130aclnnStatus aclnnMseLossOutGetWorkspaceSize(const aclTensor* self, const aclTensor* target, int64_t reduction,
106 aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor)131 aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor)
107{132{
@@ -117,8 +142,14 @@ aclnnStatus aclnnMseLossOutGetWorkspaceSize(const aclTensor* self, const aclTens
117 CHECK_RET(ret == ACLNN_SUCCESS, ret);142 CHECK_RET(ret == ACLNN_SUCCESS, ret);
118 143 
119 if (self->IsEmpty() || target->IsEmpty()) {144 if (self->IsEmpty() || target->IsEmpty()) {
120- // 根据实际支持情况补充145+ // reduction为mean时对0轴ReduceMean结果逐元素为NaN,为sum时对0轴ReduceSum结果逐元素为0,为none时不写out
121- *workspaceSize = 0;146+ if (reduction == REDUCTION_MEAN_NUM) {
147+ ret = FillOutForEmptyInput(out, NAN, uniqueExecutor.get());
148+ } else if (reduction == REDUCTION_SUM_NUM) {
149+ ret = FillOutForEmptyInput(out, 0, uniqueExecutor.get());
150+ }
151+ CHECK_RET(ret == ACLNN_SUCCESS, ret);
152+ *workspaceSize = uniqueExecutor->GetWorkspaceSize();
122 uniqueExecutor.ReleaseTo(executor);153 uniqueExecutor.ReleaseTo(executor);
123 return ACLNN_SUCCESS;154 return ACLNN_SUCCESS;
124 }155 }
@@ -164,6 +164,66 @@ TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_07_float_nd_empty_tensor_sum)
164 ut.TestPrecision();164 ut.TestPrecision();
165}165}
166 166 
167+// 空输入+非空out场景: self/target为[0,4]空tensor, reduction=mean时out按标量广播填充NaN, 验证填充路径真实执行
168+TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_08_float_nd_empty_input_mean_out_not_empty)
169+{
170+ auto selfDesc = TensorDesc({0, 4}, ACL_FLOAT, ACL_FORMAT_ND);
171+ auto targetDesc = TensorDesc({0, 4}, ACL_FLOAT, ACL_FORMAT_ND);
172+ int64_t reduction = 1;
173+ 
174+ auto outDesc = TensorDesc({4}, ACL_FLOAT, ACL_FORMAT_ND).Precision(0.0001, 0.0001);
175+ 
176+ auto ut = OP_API_UT(aclnnMseLossOut, INPUT(selfDesc, targetDesc, reduction), OUTPUT(outDesc));
177+ 
178+ // SAMPLE: only test GetWorkspaceSize
179+ uint64_t workspace_size = 0;
180+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
181+ EXPECT_EQ(aclRet, ACL_SUCCESS);
182+ 
183+ // SAMPLE: precision simulate
184+ ut.TestPrecision();
185+}
186+ 
187+// 空输入+非空out场景: self/target为[0,4]空tensor, reduction=sum时out按标量广播填充0, 验证填充路径真实执行
188+TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_09_float16_nd_empty_input_sum_out_not_empty)
189+{
190+ auto selfDesc = TensorDesc({0, 4}, ACL_FLOAT16, ACL_FORMAT_ND);
191+ auto targetDesc = TensorDesc({0, 4}, ACL_FLOAT16, ACL_FORMAT_ND);
192+ int64_t reduction = 2;
193+ 
194+ auto outDesc = TensorDesc({4}, ACL_FLOAT16, ACL_FORMAT_ND).Precision(0.001, 0.001);
195+ 
196+ auto ut = OP_API_UT(aclnnMseLossOut, INPUT(selfDesc, targetDesc, reduction), OUTPUT(outDesc));
197+ 
198+ // SAMPLE: only test GetWorkspaceSize
199+ uint64_t workspace_size = 0;
200+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
201+ EXPECT_EQ(aclRet, ACL_SUCCESS);
202+ 
203+ // SAMPLE: precision simulate
204+ ut.TestPrecision();
205+}
206+ 
207+// 空输入+非空out场景: reduction=none时空输入不写out, 非空out合法场景补充覆盖
208+TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_10_float_nd_empty_input_none_out_empty)
209+{
210+ auto selfDesc = TensorDesc({0, 4}, ACL_FLOAT, ACL_FORMAT_ND);
211+ auto targetDesc = TensorDesc({0, 4}, ACL_FLOAT, ACL_FORMAT_ND);
212+ int64_t reduction = 0;
213+ 
214+ auto outDesc = TensorDesc({0, 4}, ACL_FLOAT, ACL_FORMAT_ND).Precision(0.0001, 0.0001);
215+ 
216+ auto ut = OP_API_UT(aclnnMseLossOut, INPUT(selfDesc, targetDesc, reduction), OUTPUT(outDesc));
217+ 
218+ // SAMPLE: only test GetWorkspaceSize
219+ uint64_t workspace_size = 0;
220+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
221+ EXPECT_EQ(aclRet, ACL_SUCCESS);
222+ 
223+ // SAMPLE: precision simulate
224+ ut.TestPrecision();
225+}
226+ 
167TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_11_float_hwcn_input_not_contiguous)227TEST_F(l2_mse_loss_out_test, aclnnMseLossOut_11_float_hwcn_input_not_contiguous)
168{228{
169 auto selfDesc = TensorDesc({5, 4}, ACL_FLOAT, ACL_FORMAT_HWCN, {1, 5}, 0, {4, 5}).ValueRange(-2, 2);229 auto selfDesc = TensorDesc({5, 4}, ACL_FLOAT, ACL_FORMAT_HWCN, {1, 5}, 0, {4, 5}).ValueRange(-2, 2);
@@ -368,3 +428,43 @@ TEST_F(l2_mse_loss_out_test, ascend910B2_aclnnMseLossOut_bf16_float_nhwc_mean)
368 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspaceSize);428 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspaceSize);
369 EXPECT_EQ(aclRet, ACLNN_SUCCESS);429 EXPECT_EQ(aclRet, ACLNN_SUCCESS);
370}430}
431+ 
432+// 空输入+非空out场景: self/target为[0,4]空tensor, reduction=mean时out按标量广播填充NaN, bf16覆盖
433+TEST_F(l2_mse_loss_out_test, ascend910B2_aclnnMseLossOut_bf16_nd_empty_input_mean_out_not_empty)
434+{
435+ auto selfDesc = TensorDesc({0, 4}, ACL_BF16, ACL_FORMAT_ND);
436+ auto targetDesc = TensorDesc({0, 4}, ACL_BF16, ACL_FORMAT_ND);
437+ int64_t reduction = 1;
438+ 
439+ auto outDesc = TensorDesc({4}, ACL_BF16, ACL_FORMAT_ND).Precision(0.0001, 0.0001);
440+ 
441+ auto ut = OP_API_UT(aclnnMseLossOut, INPUT(selfDesc, targetDesc, reduction), OUTPUT(outDesc));
442+ 
443+ // SAMPLE: only test GetWorkspaceSize
444+ uint64_t workspace_size = 0;
445+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
446+ EXPECT_EQ(aclRet, ACL_SUCCESS);
447+ 
448+ // SAMPLE: precision simulate
449+ ut.TestPrecision();
450+}
451+ 
452+// 空输入+非空out场景: self/target为[0,4]空tensor, reduction=sum时out按标量广播填充0, bf16覆盖
453+TEST_F(l2_mse_loss_out_test, ascend910B2_aclnnMseLossOut_bf16_nd_empty_input_sum_out_not_empty)
454+{
455+ auto selfDesc = TensorDesc({0, 4}, ACL_BF16, ACL_FORMAT_ND);
456+ auto targetDesc = TensorDesc({0, 4}, ACL_BF16, ACL_FORMAT_ND);
457+ int64_t reduction = 2;
458+ 
459+ auto outDesc = TensorDesc({4}, ACL_BF16, ACL_FORMAT_ND).Precision(0.0001, 0.0001);
460+ 
461+ auto ut = OP_API_UT(aclnnMseLossOut, INPUT(selfDesc, targetDesc, reduction), OUTPUT(outDesc));
462+ 
463+ // SAMPLE: only test GetWorkspaceSize
464+ uint64_t workspace_size = 0;
465+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
466+ EXPECT_EQ(aclRet, ACL_SUCCESS);
467+ 
468+ // SAMPLE: precision simulate
469+ ut.TestPrecision();
470+}