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
Merged frombabeiee/ops-nnrevert-pr-9177intocann/ops-nnmaster
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Format | 151 | + 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 |
| 369 | TEST_F(l2_modulate_backward_test, ascend910B_case_19) | 369 | TEST_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 GetWorkspaceSize | 378 | // 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 |
| 386 | TEST_F(l2_modulate_backward_test, ascend910B_case_20) | 385 | TEST_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 | +} | ||