Merged
modified md files #13602
gitee-duhuipingAdd on 18 days ago
modified md files #13602
Merged
Merged fromgitee-duhuiping/ops-nnmasterintocann/ops-nnmaster
Total: 12 files changed+15-21
| @@ -96,7 +96,7 @@ aclnnStatus aclnnForeachCopy( | |||
| 96 | <td>out(aclTensorList*)</td> | 96 | <td>out(aclTensorList*)</td> |
| 97 | <td>输出</td> | 97 | <td>输出</td> |
| 98 | <td>表示进行内容复制的输出张量列表,对应公式中的`y`。</td> | 98 | <td>表示进行内容复制的输出张量列表,对应公式中的`y`。</td> |
| 99 | - <td><ul><li>支持空Tensor。</li><li>该参数中所有Tensor的数据类型保持一致。</li><li>数据格式与入参`x`的数据格式一致,shape size大于等于入参`x`的shape size。</li><li>数据类型与入参`x`的数据类型存在一定的对应关系:<ul><li>当入参`x`的数据类型为FLOAT32时,输出的数据类型支持FLOAT16或者BFLOAT16。</li><li>当入参`x`的数据类型为FLOAT16时,输出的数据类型支持FLOAT32。</li><li>当输入`x`的数据类型不为FLOAT32或者FLOAT16时,输出的数据类型与输入`x`的数据类型保持一致。</li></ul></li></ul></td> | 99 | + <td><ul><li>支持空Tensor。</li><li>该参数中所有Tensor的数据类型保持一致。</li><li>数据格式与入参`x`的数据格式一致,shape size大于等于入参`x`的shape size。</li><li>数据类型与入参`x`的数据类型存在一定的对应关系:<ul><li>当入参`x`的数据类型为FLOAT32时,输出的数据类型支持FLOAT16或者BFLOAT16。</li><li>当入参`x`的数据类型为FLOAT16时,输出的数据类型支持FLOAT32。</li><li>当入参`x`的数据类型为BFLOAT16时,输出的数据类型支持FLOAT32、BFLOAT16。</li><li>当输入`x`的数据类型不为FLOAT32、FLOAT16、BFLOAT16时,输出的数据类型与输入`x`的数据类型保持一致。</li></ul></li></ul></td> |
| 100 | <td>FLOAT32、FLOAT16、BFLOAT16、INT8、UINT8、INT16、UINT16、INT32、UINT32、INT64、DOUBLE、BOOL</td> | 100 | <td>FLOAT32、FLOAT16、BFLOAT16、INT8、UINT8、INT16、UINT16、INT32、UINT32、INT64、DOUBLE、BOOL</td> |
| 101 | <td>ND</td> | 101 | <td>ND</td> |
| 102 | <td>0-8</td> | 102 | <td>0-8</td> |
| @@ -98,7 +98,7 @@ aclnnStatus aclnnForeachCopy( | |||
| 98 | <td>out (aclTensorList*)</td> | 98 | <td>out (aclTensorList*)</td> |
| 99 | <td>Output</td> | 99 | <td>Output</td> |
| 100 | <td>Output tensor list for content copy, corresponding to `y` in the formula.</td> | 100 | <td>Output tensor list for content copy, corresponding to `y` in the formula.</td> |
| 101 | - <td><ul><li>Supports empty tensors.</li><li>All tensors in this parameter must have the same data type.</li><li>The data format is the same as that of input `x`, and the shape size is greater than or equal to that of input `x`.</li><li>The data type corresponds to the data type of input `x` as follows:<ul><li>If the data type of input `x` is FLOAT32, the output data type can be FLOAT16 or BFLOAT16.</li><li>If the data type of input `x` is FLOAT16, the output data type can be FLOAT32.</li><li>If the data type of input `x` is not FLOAT32 or FLOAT16, the output data type is the same as that of input `x`.</li></ul></li></ul></td> | 101 | + <td><ul><li>Supports empty tensors.</li><li>All tensors in this parameter must have the same data type.</li><li>The data format is the same as that of input `x`, and the shape size is greater than or equal to that of input `x`.</li><li>The data type corresponds to the data type of input `x` as follows:<ul><li>If the data type of input `x` is FLOAT32, the output data type can be FLOAT16 or BFLOAT16.</li><li>If the data type of input `x` is FLOAT16, the output data type can be FLOAT32.</li><li>If the data type of input `x` is BFLOAT16, the output data type can be FLOAT32 or BFLOAT16.</li><li>If the data type of input `x` is not FLOAT32, BFLOAT16 or FLOAT16, the output data type is the same as that of input `x`.</li></ul></li></ul></td> |
| 102 | <td>FLOAT32, FLOAT16, BFLOAT16, INT8, UINT8, INT16, UINT16, INT32, UINT32, INT64, DOUBLE, BOOL</td> | 102 | <td>FLOAT32, FLOAT16, BFLOAT16, INT8, UINT8, INT16, UINT16, INT32, UINT32, INT64, DOUBLE, BOOL</td> |
| 103 | <td>ND</td> | 103 | <td>ND</td> |
| 104 | <td>0-8</td> | 104 | <td>0-8</td> |
| @@ -47,7 +47,7 @@ | |||
| 47 | 47 | ||
| 48 | pergroup量化方式 | 48 | pergroup量化方式 |
| 49 | 49 | ||
| 50 | - 1. 矩阵乘后x''的shape为[M,N1,N2],在计算pergroup量化方式其中的mx_quantize时,需reshape为[M,N1*N2],记为x2 | 50 | + 1. 矩阵乘后x''的shape为[M,N1,N2],在计算pergroup量化方式其中的mx_quantize时,需reshape为[M,N1*N2],记为x2 |
| 51 | 51 | ||
| 52 | 2. 在x2第二维上按照groupsize进行分组,包含元素e0,e1...e31。计算出emax | 52 | 2. 在x2第二维上按照groupsize进行分组,包含元素e0,e1...e31。计算出emax |
| 53 | 53 | ||
| @@ -133,7 +133,7 @@ | |||
| 133 | <tr> | 133 | <tr> |
| 134 | <td>dst_dtype</td> | 134 | <td>dst_dtype</td> |
| 135 | <td>可选属性</td> | 135 | <td>可选属性</td> |
| 136 | - <td><ul><li>用于控制量化方式,当值为DT_FLOAT4_E2M1时表示为pergroup量化方式,当值为DT_INT32时表示pertoken量化方式。</li><li>默认值为DT_INT32。</li></ul></td> | 136 | + <td><ul><li>用于控制量化方式,当值为DT_FLOAT4_E2M1时表示为pergroup量化方式,当值为DT_INT32时表示pertoken量化方式。</li><li>默认值为DT_INT4。</li></ul></td> |
| 137 | <td>INT</td> | 137 | <td>INT</td> |
| 138 | <td>-</td> | 138 | <td>-</td> |
| 139 | </tr> | 139 | </tr> |
| @@ -135,7 +135,7 @@ English | [简体中文](./README.md) | |||
| 135 | <tr> | 135 | <tr> |
| 136 | <td>dst_dtype</td> | 136 | <td>dst_dtype</td> |
| 137 | <td>Optional attribute</td> | 137 | <td>Optional attribute</td> |
| 138 | - <td><ul><li>Controls the quantization mode. The value DT_FLOAT4_E2M1 indicates the pergroup quantization mode, and the value DT_INT32 indicates the pertoken quantization mode.</li><li>The default value is DT_INT32.</li></ul></td> | 138 | + <td><ul><li>Controls the quantization mode. The value DT_FLOAT4_E2M1 indicates the pergroup quantization mode, and the value DT_INT32 indicates the pertoken quantization mode.</li><li>The default value is DT_INT4.</li></ul></td> |
| 139 | <td>INT</td> | 139 | <td>INT</td> |
| 140 | <td>-</td> | 140 | <td>-</td> |
| 141 | </tr> | 141 | </tr> |
| @@ -327,8 +327,8 @@ aclnnStatus aclnnFlatQuantV3( | |||
| 327 | <td>传入参数中的必选输入(x、kroneckerP1、kroneckerP2)、必选输出(out、quantScale)是空指针。</td> | 327 | <td>传入参数中的必选输入(x、kroneckerP1、kroneckerP2)、必选输出(out、quantScale)是空指针。</td> |
| 328 | </tr> | 328 | </tr> |
| 329 | <tr> | 329 | <tr> |
| 330 | - <td rowspan="13">ACLNN_ERR_PARAM_INVALID</td> | 330 | + <td rowspan="12">ACLNN_ERR_PARAM_INVALID</td> |
| 331 | - <td rowspan="13">161002</td> | 331 | + <td rowspan="12">161002</td> |
| 332 | <td>x、kroneckerP1、kroneckerP2、out、quantScale的数据类型或数据格式不在支持的范围之内。</td> | 332 | <td>x、kroneckerP1、kroneckerP2、out、quantScale的数据类型或数据格式不在支持的范围之内。</td> |
| 333 | </tr> | 333 | </tr> |
| 334 | <tr> | 334 | <tr> |
| @@ -361,9 +361,6 @@ aclnnStatus aclnnFlatQuantV3( | |||
| 361 | <tr> | 361 | <tr> |
| 362 | <td>out的数据类型为INT32时,x的shape尾轴不是out的shape尾轴大小的8倍,或者x与out的shape的非尾轴的大小不一致。</td> | 362 | <td>out的数据类型为INT32时,x的shape尾轴不是out的shape尾轴大小的8倍,或者x与out的shape的非尾轴的大小不一致。</td> |
| 363 | </tr> | 363 | </tr> |
| 364 | - <tr> | ||
| 365 | - <td>out的数据类型为INT32时,x的shape尾轴不是out的shape尾轴大小的8倍,或者x与out的shape的非尾轴的大小不一致。</td> | ||
| 366 | - </tr> | ||
| 367 | <tr> | 364 | <tr> |
| 368 | <td>groupList、groupListType的数据类型、维度、shape或参数取值不符合约束。</td> | 365 | <td>groupList、groupListType的数据类型、维度、shape或参数取值不符合约束。</td> |
| 369 | </tr> | 366 | </tr> |
| @@ -329,8 +329,8 @@ aclnnStatus aclnnFlatQuantV3( | |||
| 329 | <td>The required inputs (x, kroneckerP1, kroneckerP2) or required outputs (out, quantScale) in the input parameters are null pointers.</td> | 329 | <td>The required inputs (x, kroneckerP1, kroneckerP2) or required outputs (out, quantScale) in the input parameters are null pointers.</td> |
| 330 | </tr> | 330 | </tr> |
| 331 | <tr> | 331 | <tr> |
| 332 | - <td rowspan="13">ACLNN_ERR_PARAM_INVALID</td> | 332 | + <td rowspan="12">ACLNN_ERR_PARAM_INVALID</td> |
| 333 | - <td rowspan="13">161002</td> | 333 | + <td rowspan="12">161002</td> |
| 334 | <td>The data type or data format of x, kroneckerP1, kroneckerP2, out, or quantScale is not within the supported range.</td> | 334 | <td>The data type or data format of x, kroneckerP1, kroneckerP2, out, or quantScale is not within the supported range.</td> |
| 335 | </tr> | 335 | </tr> |
| 336 | <tr> | 336 | <tr> |
| @@ -363,9 +363,6 @@ aclnnStatus aclnnFlatQuantV3( | |||
| 363 | <tr> | 363 | <tr> |
| 364 | <td>If the data type of out is INT32, the tail axis of the shape of x is not 8 times the tail axis size of the shape of out, or the sizes of the non-tail axes of the shapes of x and out do not match.</td> | 364 | <td>If the data type of out is INT32, the tail axis of the shape of x is not 8 times the tail axis size of the shape of out, or the sizes of the non-tail axes of the shapes of x and out do not match.</td> |
| 365 | </tr> | 365 | </tr> |
| 366 | - <tr> | ||
| 367 | - <td>If the data type of out is INT32, the tail axis of the shape of x is not 8 times the tail axis size of the shape of out, or the sizes of the non-tail axes of the shapes of x and out do not match.</td> | ||
| 368 | - </tr> | ||
| 369 | <tr> | 366 | <tr> |
| 370 | <td>The data type, dimension, shape, or parameter value of groupList or groupListType does not comply with the constraints.</td> | 367 | <td>The data type, dimension, shape, or parameter value of groupList or groupListType does not comply with the constraints.</td> |
| 371 | </tr> | 368 | </tr> |
| @@ -26,7 +26,7 @@ | |||
| 26 | - 接口功能:计算输入self的组归一化结果out,均值meanOut,标准差的倒数rstdOut。 | 26 | - 接口功能:计算输入self的组归一化结果out,均值meanOut,标准差的倒数rstdOut。 |
| 27 | - 计算公式: | 27 | - 计算公式: |
| 28 | 28 | ||
| 29 | - 记$x=self$,$E[x] = \bar{x}$代表$x$的均值,$Var[x] = \frac{1}{n} * \sum_{i=1}^n(x_i - E[x])^2$代表$x$的样本方差,则 | 29 | + 记$x=self$,$E[x] = \bar{x}$代表$x$的均值,$Var[x] = \frac{1}{n} * \sum_{i=1}^n(x_i - E[x])^2$代表$x$的总体方差,则 |
| 30 | 30 | ||
| 31 | $$ | 31 | $$ |
| 32 | out = \frac{x - E[x]}{\sqrt{Var[x] + eps}} * \gamma + \beta | 32 | out = \frac{x - E[x]}{\sqrt{Var[x] + eps}} * \gamma + \beta |
| @@ -28,7 +28,7 @@ | |||
| 28 | - Description: Computes the group normalization result out, mean meanOut, and reciprocal of standard deviation rstdOut for the input self. | 28 | - Description: Computes the group normalization result out, mean meanOut, and reciprocal of standard deviation rstdOut for the input self. |
| 29 | - Formula: | 29 | - Formula: |
| 30 | 30 | ||
| 31 | - Let $x=self$, where $E[x] = \bar{x}$ denotes the mean of $x$, and $Var[x] = \frac{1}{n} * \sum_{i=1}^n(x_i - E[x])^2$ denotes the sample variance of $x$. Then | 31 | + Let $x=self$, where $E[x] = \bar{x}$ denotes the mean of $x$, and $Var[x] = \frac{1}{n} * \sum_{i=1}^n(x_i - E[x])^2$ denotes the population variance of $x$. Then |
| 32 | 32 | ||
| 33 | $$ | 33 | $$ |
| 34 | out = \frac{x - E[x]}{\sqrt{Var[x] + eps}} * \gamma + \beta | 34 | out = \frac{x - E[x]}{\sqrt{Var[x] + eps}} * \gamma + \beta |
| @@ -13,7 +13,7 @@ | |||
| 13 | 13 | ||
| 14 | ## 功能说明 | 14 | ## 功能说明 |
| 15 | 15 | ||
| 16 | -- 算子功能:前向L2Normalize(`y = x/max(sqrt(sum(x^2, dim)), eps)`)的反向算子。三输入`x`(前向输入)、`y`(前向输出,即归一化后的x)、`dy`(上游梯度),单输出`dx`(对x的梯度)。 | 16 | +- 算子功能:前向L2Normalize($y = x/max(sqrt(sum(x^2, dim)), eps)$)的反向算子。三输入`x`(前向输入)、`y`(前向输出,即归一化后的x)、`dy`(上游梯度),单输出`dx`(对x的梯度)。 |
| 17 | 17 | ||
| 18 | - 计算公式(沿属性`dim`指定的归约轴集合归约,keepdims广播回全形状): | 18 | - 计算公式(沿属性`dim`指定的归约轴集合归约,keepdims广播回全形状): |
| 19 | 19 | ||
| @@ -15,7 +15,7 @@ English | [简体中文](./README.md) | |||
| 15 | 15 | ||
| 16 | ## Function | 16 | ## Function |
| 17 | 17 | ||
| 18 | -- Operator function: Backward operator of the forward L2Normalize (`y = x/max(sqrt(sum(x^2, dim)), eps)`). The three inputs are `x` (forward input), `y` (forward output, that is, the normalized x), and `dy` (upstream gradient), and the single output is `dx` (gradient with respect to x). | 18 | +- Operator function: Backward operator of the forward L2Normalize ($y = x/max(sqrt(sum(x^2, dim)), eps)$). The three inputs are `x` (forward input), `y` (forward output, that is, the normalized x), and `dy` (upstream gradient), and the single output is `dx` (gradient with respect to x). |
| 19 | 19 | ||
| 20 | - Formula (reduction is performed along the reduction axis set specified by the attribute `dim`, and keepdims broadcasts back to the full shape): | 20 | - Formula (reduction is performed along the reduction axis set specified by the attribute `dim`, and keepdims broadcasts back to the full shape): |
| 21 | 21 | ||
| @@ -216,7 +216,7 @@ aclnnStatus aclnnSyncBatchNormGatherStats( | |||
| 216 | </table> | 216 | </table> |
| 217 | 217 | ||
| 218 | <!-- npu="A3,910b" id7 --> | 218 | <!-- npu="A3,910b" id7 --> |
| 219 | - - <term>Atlas A3系列产品</term>、<term>Atlas A2系列产品</term>:参数`totalSum`、`totalSquareSum`、`mean`、`variance`、`batchMean`、`batchInvstd`的数据类型不支持BFLOAT16。 | 219 | + - <term>Atlas A3系列产品</term>、<term>Atlas A2系列产品</term>:参数`totalSum`、`totalSquareSum`、`sampleCount`、`mean`、`variance`、`batchMean`、`batchInvstd`的数据类型不支持BFLOAT16。 |
| 220 | <!-- end id7 --> | 220 | <!-- end id7 --> |
| 221 | 221 | ||
| 222 | - **返回值** | 222 | - **返回值** |
| @@ -218,7 +218,7 @@ aclnnStatus aclnnSyncBatchNormGatherStats( | |||
| 218 | </table> | 218 | </table> |
| 219 | 219 | ||
| 220 | <!-- npu="A3,910b" id7 --> | 220 | <!-- npu="A3,910b" id7 --> |
| 221 | - - <term>Atlas A3 products</term>, <term>Atlas A2 products</term>: The data type of the parameters `totalSum`, `totalSquareSum`, `mean`, `variance`, `batchMean`, and `batchInvstd` does not support BFLOAT16. | 221 | + - <term>Atlas A3 products</term>, <term>Atlas A2 products</term>: The data type of the parameters `totalSum`, `totalSquareSum`, `sampleCount`, `mean`, `variance`, `batchMean`, and `batchInvstd` does not support BFLOAT16. |
| 222 | <!-- end id7 --> | 222 | <!-- end id7 --> |
| 223 | 223 | ||
| 224 | - **Return Value** | 224 | - **Return Value** |