Merged
Revert "modulate_grad无原型,仅支持aclnn,删除README.md"(回撤 #9177) #12478
babeieeAdd on Oct 21
Revert "modulate_grad无原型,仅支持aclnn,删除README.md"(回撤 #9177) #12478
Merged
babeieeAdd on Oct 21
Total: 5 files changed+303-89
@@ -5798,6 +5798,16 @@
5798 <td>AI Core</td>5798 <td>AI Core</td>
5799 <td>GRUCell中三个门中matmul后剩余计算的反向传播,计算正向输出三个门激活前的值gates、前向workspace([r,z,n,hx,hn]五平面)的梯度,以及输入hx与偏置b的梯度。</td>5799 <td>GRUCell中三个门中matmul后剩余计算的反向传播,计算正向输出三个门激活前的值gates、前向workspace([r,z,n,hx,hn]五平面)的梯度,以及输入hx与偏置b的梯度。</td>
5800 </tr>5800 </tr>
5801+ <tr>
5802+ <td>vfusion</td>
5803+ <td><a href="../../vfusion/modulate_grad/README.md">modulate_grad</a></td>
5804+ <td>✓</td>
5805+ <td>✓</td>
5806+ <td>✓</td>
5807+ <td>✗</td>
5808+ <td>AI Core</td>
5809+ <td>Modulate算子的反向传播,根据输出梯度计算输入x、scale和shift的梯度。</td>
5810+ </tr>
5801 <tr>5811 <tr>
5802 <td>vfusion</td>5812 <td>vfusion</td>
5803 <td><a href="../../vfusion/multi_scale_deformable_attention_grad/README.md">multi_scale_deformable_attention_grad</a></td>5813 <td><a href="../../vfusion/multi_scale_deformable_attention_grad/README.md">multi_scale_deformable_attention_grad</a></td>
@@ -0,0 +1,113 @@
1+# ModulateGrad
2+ 
3+## 产品支持情况
4+ 
5+| 产品 | 是否支持 |
6+| :----------------------------------------------------------- | :------: |
7+| <term>Ascend 950PR&950DT系列产品</term> | √ |
8+| <term>Atlas A3系列产品</term> | √ |
9+| <term>Atlas A2系列产品</term> | √ |
10+| <term>Atlas 200I/500 A2推理产品</term> | × |
11+| <term>Atlas推理系列产品</term> | × |
12+| <term>Atlas训练系列产品</term> | × |
13+ 
14+## 功能说明
15+ 
16+- 算子功能:完成Modulate反向传播中参数的计算,进行梯度更新。
17+- 计算公式:
18+ 
19+ 设输入self的shape为[B, L, D]计算公式如下:
20+ 公式:
21+ 
22+ $$
23+ \begin{cases}
24+ \text{grad\_x} = \text{grad\_output} \odot \left(1 + \text{scale}^{\uparrow L}\right) \\
25+ \text{grad\_scale} = \sum_{l=1}^{L} (\text{grad\_output} \odot \text{input})_{b,l,d} \\
26+ \text{grad\_shift} = \sum_{l=1}^{L} \text{grad\_output}_{b,l,d}
27+ \end{cases}
28+ $$
29+ 
30+ 符号说明:
31+ - $\odot$: 表示逐元素乘法;
32+ - $\sum_{l=1}^{L}$: 求和操作,沿序列维度$L$(即dim=1)进行
33+ - $b,l,d$:下标,表示张量的维度索引(通常为Batch,Length,Dimension)
34+ - $\text{scale}^{\uparrow L}$:表示将scale张量在序列维度 $L$ 上进行广播(扩展)
35+ 
36+## 参数说明
37+ 
38+<table style="undefined;table-layout: fixed; width: 820px"><colgroup>
39+ <col style="width: 100px">
40+ <col style="width: 150px">
41+ <col style="width: 190px">
42+ <col style="width: 260px">
43+ <col style="width: 120px">
44+ </colgroup>
45+ <thead>
46+ <tr>
47+ <th>参数名</th>
48+ <th>输入/输出/属性</th>
49+ <th>描述</th>
50+ <th>数据类型</th>
51+ <th>数据格式</th>
52+ </tr></thead>
53+ <tbody>
54+ <tr>
55+ <td>grad_output</td>
56+ <td>输入</td>
57+ <td>表示传入的特征张量,公式中的grad_output。</td>
58+ <td>FLOAT、FLOAT16、BFLOAT16</td>
59+ <td>ND</td>
60+ </tr>
61+ <tr>
62+ <td>input</td>
63+ <td>输入</td>
64+ <td>表示传入的特征张量,公式中的input。</td>
65+ <td>FLOAT、FLOAT16、BFLOAT16</td>
66+ <td>ND</td>
67+ </tr>
68+ <tr>
69+ <td>scale</td>
70+ <td>输入</td>
71+ <td>可选参数,表示缩放的系数,公式中的scale。</td>
72+ <td>FLOAT、FLOAT16、BFLOAT16</td>
73+ <td>ND</td>
74+ </tr>
75+ <tr>
76+ <td>shift</td>
77+ <td>输入</td>
78+ <td>可选参数,表示平移的系数,公式中的shift。</td>
79+ <td>FLOAT、FLOAT16、BFLOAT16</td>
80+ <td>ND</td>
81+ </tr>
82+ <tr>
83+ <td>grad_input</td>
84+ <td>输出</td>
85+ <td>表示缩放后的输出张量,公式中的grad_input。</td>
86+ <td>FLOAT、FLOAT16、BFLOAT16</td>
87+ <td>ND</td>
88+ </tr>
89+ <tr>
90+ <td>grad_scale</td>
91+ <td>输出</td>
92+ <td>表示缩放后的输出张量,公式中的grad_scale。</td>
93+ <td>FLOAT、FLOAT16、BFLOAT16</td>
94+ <td>ND</td>
95+ </tr>
96+ <tr>
97+ <td>grad_shift</td>
98+ <td>输出</td>
99+ <td>表示缩放后的输出张量,公式中的grad_shift。</td>
100+ <td>FLOAT、FLOAT16、BFLOAT16</td>
101+ <td>ND</td>
102+ </tr>
103+ </tbody></table>
104+ 
105+## 约束说明
106+ 
107+无
108+ 
109+## 调用说明
110+ 
111+| 调用方式 | 调用样例 | 说明 |
112+|--------------|------------------------------------------------------------------------|----------------------------------------------------------------|
113+| aclnn调用 | [test_aclnn_modulatebackward](./examples/test_aclnn_modulatebackward.cpp) | 通过[aclnnModulateBackward](../../vfusion/modulate_grad/docs/aclnnModulateBackward.md)接口方式调用ModulateGrad算子。 |
@@ -196,26 +196,41 @@ aclnnStatus aclnnModulateBackward(
196 </colgroup>196 </colgroup>
197 <thead>197 <thead>
198 <tr>198 <tr>
199- <th>返回值</th>199+ <th>返回码</th>
200- <th>错误码</th>200+ <th>码值</th>
201 <th>描述</th>201 <th>描述</th>
202 </tr></thead>202 </tr></thead>
203 <tbody>203 <tbody>
204 <tr>204 <tr>
205 <td>ACLNN_ERR_PARAM_NULLPTR</td>205 <td>ACLNN_ERR_PARAM_NULLPTR</td>
206 <td>161001</td>206 <td>161001</td>
207- <td>传入的grad_output或input等入参是空指针。</td>207+ <td>grad_output、input或grad_input为空指针。</td>
208 </tr>208 </tr>
209 <tr>209 <tr>
210- <td rowspan="3">ACLNN_ERR_PARAM_INVALID</td>210+ <td rowspan="8">ACLNN_ERR_PARAM_INVALID</td>
211- <td rowspan="3">161002</td>211+ <td rowspan="8">161002</td>
212- <td>grad_output、input、scale、shift的数据类型和数据格式不在支持的范围之内。</td>212+ <td>grad_output、input、scale、shift、grad_input、grad_scale或grad_shift的数据类型不在支持范围内。</td>
213 </tr>213 </tr>
214 <tr>214 <tr>
215- <td>grad_output、input、scale、shift之间的shape不满足约束。</td>215+ <td>input、scale、shift、grad_input、grad_scale或grad_shift的数据类型与grad_output的数据类型不一致。</td>
216 </tr>216 </tr>
217 <tr>217 <tr>
218- <td>input为空tensor,且scale或shift不为空tensor。</td>218+ <td>grad_output、input、scale、shift、grad_input、grad_scale或grad_shift的数据格式不是ND。</td>
219+ </tr>
220+ <tr>
221+ <td>grad_output或input不是3维。</td>
222+ </tr>
223+ <tr>
224+ <td>grad_output、input与grad_input的shape不一致。</td>
225+ </tr>
226+ <tr>
227+ <td>scale或shift不是2维,或者shape与grad_output不满足约束。</td>
228+ </tr>
229+ <tr>
230+ <td>grad_scale与scale的shape不一致,grad_shift与shift的shape不一致,或者提供grad_scale但未提供scale、提供grad_shift但未提供shift。</td>
231+ </tr>
232+ <tr>
233+ <td>grad_output为空Tensor时,scale或shift不是空Tensor。</td>
219 </tr>234 </tr>
220 </tbody>235 </tbody>
221 </table>236 </table>
@@ -257,38 +272,6 @@ aclnnStatus aclnnModulateBackward(
257 272 
258 aclnnStatus:返回状态码,具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。273 aclnnStatus:返回状态码,具体参见[aclnn返回码](../../../docs/zh/context/aclnn_return_code.md)。
259 274 
260- 第一段接口完成入参校验,出现以下场景时报错:
261- <table style="undefined;table-layout: fixed; width: 1030px"><colgroup>
262- <col style="width: 250px">
263- <col style="width: 130px">
264- <col style="width: 650px">
265- </colgroup>
266- <thead>
267- <tr>
268- <th>返回值</th>
269- <th>错误码</th>
270- <th>描述</th>
271- </tr></thead>
272- <tbody>
273- <tr>
274- <td>ACLNN_ERR_PARAM_NULLPTR</td>
275- <td>161001</td>
276- <td>传入的grad_output或input是空指针。</td>
277- </tr>
278- <tr>
279- <td rowspan="3">ACLNN_ERR_PARAM_INVALID</td>
280- <td rowspan="3">161002</td>
281- <td>输入和输出的数据类型和数据格式不在支持的范围之内。</td>
282- </tr>
283- <tr>
284- <td>input、scale、shift之间的shape不满足约束。</td>
285- </tr>
286- <tr>
287- <td>输入维度超过3维</td>
288- </tr>
289- </tbody>
290- </table>
291- 
292## 约束说明275## 约束说明
293 276 
294- 确定性计算:277- 确定性计算:
@@ -297,10 +280,15 @@ aclnnStatus aclnnModulateBackward(
297- scale和shift是二维向量,第一维需要和input的第一维shape相同,第二维需要和input的第三维shape相同。280- scale和shift是二维向量,第一维需要和input的第一维shape相同,第二维需要和input的第三维shape相同。
298- 输入grad_output的shape需要和输入input的shape保持一致,均为3维 [B, L, D]。281- 输入grad_output的shape需要和输入input的shape保持一致,均为3维 [B, L, D]。
299- 数据格式仅支持ND;各tensor的dtype必须一致(FLOAT16/FLOAT/BFLOAT16)。282- 数据格式仅支持ND;各tensor的dtype必须一致(FLOAT16/FLOAT/BFLOAT16)。
300-- scale和shift为可选输入:未提供scale/shift时,对应的grad_scale/grad_shift不输出。283+- scale和shift为可选输入:未提供scale/shift时,对应的grad_scale/grad_shift不输出;未提供scale时,grad_input = grad_output。
301- 支持空Tensor(B/L/D任一为0):直接返回对应shape的空输出;此时scale和shift必须为空或缺失。284- 支持空Tensor(B/L/D任一为0):直接返回对应shape的空输出;此时scale和shift必须为空或缺失。
302- 支持动态shape(含-1维度):tiling在运行期按实际shape计算多核与UB切分,无需重新编译;rank固定为3。285- 支持动态shape(含-1维度):tiling在运行期按实际shape计算多核与UB切分,无需重新编译;rank固定为3。
303 286 
287+ 
288+- 精度说明:
289+ - kernel 内部以 fp32 累加归约(grad_scale = Σ input·grad_output, grad_shift = Σ grad_output),结果经 RNE 舍回输出 dtype。
290+ - 测试参考(golden)以 fp64 计算后同样经 RNE 舍回输出 dtype,与 kernel 舍入路径一致。
291+ 
304## 调用示例292## 调用示例
305 293 
306示例代码如下,仅供参考,具体编译和执行过程请参考[编译与运行样例](../../../docs/zh/context/compile_and_run_sample.md)。294示例代码如下,仅供参考,具体编译和执行过程请参考[编译与运行样例](../../../docs/zh/context/compile_and_run_sample.md)。
@@ -127,62 +127,77 @@ static bool CheckDtypeConsistency(const aclTensor* grad_output, const aclTensor*
127 return true;127 return true;
128}128}
129 129 
130-static bool CheckMaxDimension(const aclTensor* grad_output, const aclTensor* input, const aclTensor* scale,130+static bool CheckFormat(const aclTensor* grad_output, const aclTensor* input, const aclTensor* scale,
131- const aclTensor* shift)131+ const aclTensor* shift, const aclTensor* grad_input, const aclTensor* grad_scale,
132+ const aclTensor* grad_shift)
132{133{
133- op::Shape gradoutputShape = grad_output->GetViewShape();134+ if (grad_output->GetStorageFormat() != Format::FORMAT_ND || input->GetStorageFormat() != Format::FORMAT_ND ||
134- size_t gradoutputDimNum = gradoutputShape.GetDimNum();135+ grad_input->GetStorageFormat() != Format::FORMAT_ND ||
135- op::Shape inputShape = input->GetViewShape();136+ (scale != nullptr && scale->GetStorageFormat() != Format::FORMAT_ND) ||
136- size_t inputDimNum = inputShape.GetDimNum();137+ (shift != nullptr && shift->GetStorageFormat() != Format::FORMAT_ND) ||
137- if (gradoutputDimNum != MAX_SUPPORT_DIM || inputDimNum != MAX_SUPPORT_DIM) {138+ (grad_scale != nullptr && grad_scale->GetStorageFormat() != Format::FORMAT_ND) ||
138- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "The dimension of self must be 3");139+ (grad_shift != nullptr && grad_shift->GetStorageFormat() != Format::FORMAT_ND)) {
140+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "All input and output tensors only support ND format");
139 return false;141 return false;
140 }142 }
141- if ((grad_output->IsEmpty() || input->IsEmpty()) && ((CheckNotNullForScaleAndShift(scale) && !scale->IsEmpty()) ||
142- (CheckNotNullForScaleAndShift(shift) && !shift->IsEmpty()))) {
143- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "when grad_output or input is empty, scale and shift must be empty");
144- return false;
145- }
146- 
147 return true;143 return true;
148}144}
149 145 
150-static bool CheckDimension(const aclTensor* grad_output, const aclTensor* input, const aclTensor* scale,146+static bool CheckShape(const aclTensor* grad_output, const aclTensor* input, const aclTensor* scale,
151- const aclTensor* shift)147+ const aclTensor* shift, const aclTensor* grad_input, const aclTensor* grad_scale,
148+ const aclTensor* grad_shift)
152{149{
153- auto gradoutput_Shape = grad_output->GetViewShape();150+ const auto gradOutputShape = grad_output->GetViewShape();
154- // 检查Format151+ if (gradOutputShape.GetDimNum() != MAX_SUPPORT_DIM || input->GetViewShape().GetDimNum() != MAX_SUPPORT_DIM) {
155- if (input->GetStorageFormat() != Format::FORMAT_ND || grad_output->GetStorageFormat() != Format::FORMAT_ND ||152+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "grad_output and input must be 3-dimensional tensors");
156- (CheckNotNullForScaleAndShift(scale) && scale->GetStorageFormat() != Format::FORMAT_ND) ||
157- (CheckNotNullForScaleAndShift(shift) && shift->GetStorageFormat() != Format::FORMAT_ND)) {
158- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Format only support ND, got input format %d, grad_output format %d",
159- static_cast<int32_t>(input->GetStorageFormat()), static_cast<int32_t>(grad_output->GetStorageFormat()));
160 return false;153 return false;
161 }154 }
162- if (CheckNotNullForScaleAndShift(scale)) {155+ if (gradOutputShape != input->GetViewShape() || gradOutputShape != grad_input->GetViewShape()) {
163- auto scale_Shape = scale->GetViewShape();156+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "grad_output, input and grad_input must have the same shape");
164- if (scale_Shape.GetDimNum() != SCALE_DIM) {157+ return false;
165- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "scale dim is not equal 2");158+ }
159+ 
160+ if (scale == nullptr && grad_scale != nullptr) {
161+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "grad_scale requires scale to be provided");
162+ return false;
163+ }
164+ if (scale != nullptr) {
165+ const auto scaleShape = scale->GetViewShape();
166+ if (scaleShape.GetDimNum() != SCALE_DIM) {
167+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "scale must be a 2-dimensional tensor");
166 return false;168 return false;
167 }169 }
168- if (gradoutput_Shape.GetDim(0) != scale_Shape.GetDim(FIRST_DIM) ||170+ if (gradOutputShape.GetDim(FIRST_DIM) != scaleShape.GetDim(FIRST_DIM) ||
169- gradoutput_Shape.GetDim(LAST_DIM) != scale_Shape.GetDim(MIDDLE_DIM)) {171+ gradOutputShape.GetDim(LAST_DIM) != scaleShape.GetDim(MIDDLE_DIM) ||
170- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "gradoutput_Shape dim0 or dim 2 not match scale dim 0 or dim 1");172+ (grad_scale != nullptr && scaleShape != grad_scale->GetViewShape())) {
173+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "scale or grad_scale shape does not match grad_output");
171 return false;174 return false;
172 }175 }
173 }176 }
174- if (CheckNotNullForScaleAndShift(shift)) {177+ 
175- auto shift_Shape = shift->GetViewShape();178+ if (shift == nullptr && grad_shift != nullptr) {
176- if (shift_Shape.GetDimNum() != SCALE_DIM) {179+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "grad_shift requires shift to be provided");
177- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "shift dim is not equal 2");180+ return false;
181+ }
182+ if (shift != nullptr) {
183+ const auto shiftShape = shift->GetViewShape();
184+ if (shiftShape.GetDimNum() != SCALE_DIM) {
185+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "shift must be a 2-dimensional tensor");
178 return false;186 return false;
179 }187 }
180- if (gradoutput_Shape.GetDim(FIRST_DIM) != shift_Shape.GetDim(FIRST_DIM) ||188+ if (gradOutputShape.GetDim(FIRST_DIM) != shiftShape.GetDim(FIRST_DIM) ||
181- gradoutput_Shape.GetDim(LAST_DIM) != shift_Shape.GetDim(MIDDLE_DIM)) {189+ gradOutputShape.GetDim(LAST_DIM) != shiftShape.GetDim(MIDDLE_DIM) ||
182- OP_LOGE(ACLNN_ERR_PARAM_INVALID, "gradoutput_Shape dim0 or dim 2 not match shift dim 0 or dim 1");190+ (grad_shift != nullptr && shiftShape != grad_shift->GetViewShape())) {
191+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "shift or grad_shift shape does not match grad_output");
183 return false;192 return false;
184 }193 }
185 }194 }
195+ 
196+ if (grad_output->IsEmpty() &&
197+ ((scale != nullptr && !scale->IsEmpty()) || (shift != nullptr && !shift->IsEmpty()))) {
198+ OP_LOGE(ACLNN_ERR_PARAM_INVALID, "when grad_output is empty, scale and shift must be empty or absent");
199+ return false;
200+ }
186 return true;201 return true;
187}202}
188 203 
@@ -198,10 +213,12 @@ static aclnnStatus CheckParams(const aclTensor* grad_output, const aclTensor* in
198 // 3.检查数据类型一致性213 // 3.检查数据类型一致性
199 CHECK_RET(CheckDtypeConsistency(grad_output, input, scale, shift, grad_input, grad_scale, grad_shift),214 CHECK_RET(CheckDtypeConsistency(grad_output, input, scale, shift, grad_input, grad_scale, grad_shift),
200 ACLNN_ERR_PARAM_INVALID);215 ACLNN_ERR_PARAM_INVALID);
201- // 4.检查维度是否超过8维 在这里不会超过8维216+ // 4.检查数据格式
202- CHECK_RET(CheckMaxDimension(grad_output, input, scale, shift), ACLNN_ERR_PARAM_INVALID);217+ CHECK_RET(CheckFormat(grad_output, input, scale, shift, grad_input, grad_scale, grad_shift),
203- // 5.检查维度是否匹配218+ ACLNN_ERR_PARAM_INVALID);
204- CHECK_RET(CheckDimension(grad_output, input, scale, shift), ACLNN_ERR_PARAM_INVALID);219+ // 5.检查shape及可选输入输出关系
220+ CHECK_RET(CheckShape(grad_output, input, scale, shift, grad_input, grad_scale, grad_shift),
221+ ACLNN_ERR_PARAM_INVALID);
205 222 
206 return ACLNN_SUCCESS;223 return ACLNN_SUCCESS;
207}224}
@@ -365,7 +365,7 @@ TEST_F(l2_modulate_backward_test, ascend910B_case_18)
365 EXPECT_EQ(aclRet, ACL_SUCCESS);365 EXPECT_EQ(aclRet, ACL_SUCCESS);
366}366}
367 367 
368-// grad_output 为空 tensor 且 scale/shift 传 nullptr 时,CheckMaxDimension 不允许解引用空指针368+// grad_output、input和grad_input的shape必须一致
369TEST_F(l2_modulate_backward_test, ascend910B_case_19)369TEST_F(l2_modulate_backward_test, ascend910B_case_19)
370{370{
371 auto grad_output_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);371 auto grad_output_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
@@ -378,18 +378,17 @@ TEST_F(l2_modulate_backward_test, ascend910B_case_19)
378 // SAMPLE: only test GetWorkspaceSize378 // SAMPLE: only test GetWorkspaceSize
379 uint64_t workspace_size = 0;379 uint64_t workspace_size = 0;
380 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);380 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
381- EXPECT_EQ(aclRet, ACL_SUCCESS);381+ EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
382- EXPECT_EQ(workspace_size, 0);
383}382}
384 383 
385-// grad_output 为空 tensor 但 scale/shift 非空时,CheckMaxDimension 的拦截逻辑需保持有效384+// 输入为空 tensor 但 scale/shift 非空时,软件校验需拦截
386TEST_F(l2_modulate_backward_test, ascend910B_case_20)385TEST_F(l2_modulate_backward_test, ascend910B_case_20)
387{386{
388- auto grad_output_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);387+ auto grad_output_desc = TensorDesc({32, 0, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
389- auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);388+ auto input_desc = TensorDesc({32, 0, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
390 auto scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);389 auto scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
391 auto shift_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);390 auto shift_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
392- auto grad_input_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);391+ auto grad_input_desc = TensorDesc({32, 0, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
393 auto grad_scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);392 auto grad_scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
394 auto grad_shift_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);393 auto grad_shift_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
395 394 
@@ -400,3 +399,90 @@ TEST_F(l2_modulate_backward_test, ascend910B_case_20)
400 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);399 aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
401 EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);400 EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
402}401}
402+ 
403+TEST_F(l2_modulate_backward_test, ascend910B_case_21)
404+{
405+ auto grad_output_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
406+ auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
407+ auto grad_input_desc = TensorDesc({32, 8, 512}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
408+ 
409+ auto ut = OP_API_UT(aclnnModulateBackward,
410+ INPUT(grad_output_desc, input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr),
411+ OUTPUT(grad_input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr));
412+ uint64_t workspace_size = 0;
413+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
414+ EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
415+}
416+ 
417+TEST_F(l2_modulate_backward_test, ascend910B_case_22)
418+{
419+ auto grad_output_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
420+ auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
421+ auto scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
422+ auto grad_input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
423+ auto grad_scale_desc = TensorDesc({32, 512}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
424+ 
425+ auto ut = OP_API_UT(aclnnModulateBackward, INPUT(grad_output_desc, input_desc, scale_desc, (aclTensor*)nullptr),
426+ OUTPUT(grad_input_desc, grad_scale_desc, (aclTensor*)nullptr));
427+ uint64_t workspace_size = 0;
428+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
429+ EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
430+}
431+ 
432+TEST_F(l2_modulate_backward_test, ascend910B_case_23)
433+{
434+ auto grad_output_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
435+ auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
436+ auto grad_input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
437+ auto grad_scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
438+ 
439+ auto ut = OP_API_UT(aclnnModulateBackward,
440+ INPUT(grad_output_desc, input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr),
441+ OUTPUT(grad_input_desc, grad_scale_desc, (aclTensor*)nullptr));
442+ uint64_t workspace_size = 0;
443+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
444+ EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
445+}
446+ 
447+TEST_F(l2_modulate_backward_test, ascend910B_case_24)
448+{
449+ auto grad_output_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
450+ auto input_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
451+ auto grad_input_desc = TensorDesc({0, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
452+ 
453+ auto ut = OP_API_UT(aclnnModulateBackward,
454+ INPUT(grad_output_desc, input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr),
455+ OUTPUT(grad_input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr));
456+ uint64_t workspace_size = 0;
457+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
458+ EXPECT_EQ(aclRet, ACL_SUCCESS);
459+ EXPECT_EQ(workspace_size, 0);
460+}
461+ 
462+TEST_F(l2_modulate_backward_test, ascend910B_case_25)
463+{
464+ auto grad_output_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
465+ auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
466+ auto grad_input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_NCHW).ValueRange(-10, 10);
467+ 
468+ auto ut = OP_API_UT(aclnnModulateBackward,
469+ INPUT(grad_output_desc, input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr),
470+ OUTPUT(grad_input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr));
471+ uint64_t workspace_size = 0;
472+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
473+ EXPECT_EQ(aclRet, ACLNN_ERR_PARAM_INVALID);
474+}
475+ 
476+TEST_F(l2_modulate_backward_test, ascend910B_case_26)
477+{
478+ auto grad_output_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
479+ auto input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
480+ auto scale_desc = TensorDesc({32, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
481+ auto grad_input_desc = TensorDesc({32, 8, 1024}, ACL_FLOAT16, ACL_FORMAT_ND).ValueRange(-10, 10);
482+ 
483+ auto ut = OP_API_UT(aclnnModulateBackward, INPUT(grad_output_desc, input_desc, scale_desc, (aclTensor*)nullptr),
484+ OUTPUT(grad_input_desc, (aclTensor*)nullptr, (aclTensor*)nullptr));
485+ uint64_t workspace_size = 0;
486+ aclnnStatus aclRet = ut.TestGetWorkspaceSize(&workspace_size);
487+ EXPECT_EQ(aclRet, ACL_SUCCESS);
488+}