Star916
1833
代码
代码
Issues230
Pull Requests324
Actions
项目讨论
Wiki
分析
Star916
1833
已合并
修改了文档中的中英文错误、标点符号错误等 #3915
向芷萍创建于 4月17日
修改了文档中的中英文错误、标点符号错误等 #3915
已合并
向芷萍创建于 4月17日
从Christina_YU/ops-nn9.0.0合入到cann/ops-nn9.0.0
共 13 个文件变更+64-64
Mconv/convolution_backward/docs/aclnnConvTbcBackward.md+1-1
@@ -107,7 +107,7 @@ aclnnStatus aclnnConvTbcBackward(
107 <td>输入</td>107 <td>输入</td>
108 <td>公式中的输出张量y对L的梯度，表示卷积反向的输入。</td>108 <td>公式中的输出张量y对L的梯度，表示卷积反向的输入。</td>
109 <td>109 <td>
110- <ul><li>支持空Tensor。</li><li>shape为(N,C<sub>out</sub>,H<sub>out</sub>)。</li><li>数据类型与 weight 的数据类型需满足数据类型推导规则（参见<a href="../../../docs/zh/context/互推导关系.md">互推导关系</a>)。</li></ul>110+ <ul><li>支持空Tensor。</li><li>shape为(N,C<sub>out</sub>,H<sub>out</sub>)。</li><li>数据类型与 weight 的数据类型需满足数据类型推导规则（参见<a href="../../../docs/zh/context/互推导关系.md">互推导关系</a>）。</li></ul>
111 </td>111 </td>
112 <td>FLOAT、FLOAT16、BFLOAT16</td>112 <td>FLOAT、FLOAT16、BFLOAT16</td>
113 <td>ND、NCL</td>113 <td>ND、NCL</td>
Mconv/convolution_backward/docs/aclnnConvolutionBackward.md+1-1
@@ -117,7 +117,7 @@ aclnnStatus aclnnConvolutionBackward(
117 <td>输出张量y对L的梯度。</td>117 <td>输出张量y对L的梯度。</td>
118 <td> 118 <td>
119 <ul><li>支持空Tensor。</li>119 <ul><li>支持空Tensor。</li>
120- <li>数据类型与input、weight满足数据类型推导规则（参见<a href="../../../docs/zh/context/互推导关系.md" target="_blank">互推关系</a>和<a href="#约束说明" target="_blank">约束说明</a>）。</li>120+ <li>数据类型与input、weight满足数据类型推导规则（参见<a href="../../../docs/zh/context/互推导关系.md" target="_blank">互推导关系</a>和<a href="#约束说明" target="_blank">约束说明</a>）。</li>
121 <li>shape不支持broadcast，要求和input、weight满足卷积输入输出shape的推导关系。</li>121 <li>shape不支持broadcast，要求和input、weight满足卷积输入输出shape的推导关系。</li>
122 <li>数据格式需要与input、gradInput一致。</li></ul>122 <li>数据格式需要与input、gradInput一致。</li></ul>
123 </td>123 </td>
Mdocs/zh/op_api_list.md+7-7
@@ -6,17 +6,17 @@
6 6 
7调用算子API时，需引用依赖的头文件和库文件，一般头文件默认在```${INSTALL_DIR}/include/aclnnop```，库文件默认在```${INSTALL_DIR}/lib64```，具体文件如下：7调用算子API时，需引用依赖的头文件和库文件，一般头文件默认在```${INSTALL_DIR}/include/aclnnop```，库文件默认在```${INSTALL_DIR}/lib64```，具体文件如下：
8 8 
9-- 依赖的头文件：①方式1 （推荐）：引用算子总头文件aclnn\_ops\_\$\{ops\_project\}.h。②方式2：按需引用单算子API头文件aclnn\_\*.h。9+- 依赖的头文件：①方式1 （推荐）：引用算子总头文件```aclnn_ops_${ops_project}.h```。②方式2：按需引用单算子API头文件```aclnn_*.h```。
10-- 依赖的库文件：按需引用算子总库文件libopapi\_\$\{ops\_project\}.so。10+- 依赖的库文件：按需引用算子总库文件```libopapi_${ops_project}.so```。
11 11 
12-其中${INSTALL_DIR}表示CANN安装后文件路径；\$\{ops\_project\}表示算子仓（如math、nn、cv、transformer），请配置为实际算子仓名。12+其中```${INSTALL_DIR}```表示CANN安装后文件路径；```${ops_project}```表示算子仓（如math、nn、cv、transformer），请配置为实际算子仓名。
13 13 
14## 接口列表14## 接口列表
15 15 
16> **确定性简介**：16> **确定性简介**：
17>17>
18> - 配置说明：因CANN或NPU型号不同等原因，可能无法保证同一个算子多次运行结果一致。在相同条件下（平台、设备、版本号和其他随机性参数等），部分算子接口可通过`aclrtCtxSetSysParamOpt`（参见[《acl API（C）》](https://hiascend.com/document/redirect/CannCommunityCppApi)）开启确定性算法，使多次运行结果一致。18> - 配置说明：因CANN或NPU型号不同等原因，可能无法保证同一个算子多次运行结果一致。在相同条件下（平台、设备、版本号和其他随机性参数等），部分算子接口可通过`aclrtCtxSetSysParamOpt`（参见[《acl API（C）》](https://hiascend.com/document/redirect/CannCommunityCppApi)）开启确定性算法，使多次运行结果一致。
19-> - 性能说明：同一个算子采用确定性计算通常比非确定性慢，因此模型单次运行性能可能会下降。但在实验、调试调测等需要保证多次运行结果相同来定位问题的场景，确定性计算可以提升效率。19+> - 性能说明：同一个算子采用确定性计算通常比非确定性慢，因此模型单次运行性能可能会下降。但在实验、调试和调测等需要保证多次运行结果相同来定位问题的场景，确定性计算可以提升效率。
20> - 线程说明：同一线程中只能设置一次确定性状态，多次设置以最后一次有效设置为准。有效设置是指设置确定性状态后，真正执行了一次算子任务下发。如果仅设置，没有算子下发，只能是确定性变量开启但未下发给算子，因此不执行算子。20> - 线程说明：同一线程中只能设置一次确定性状态，多次设置以最后一次有效设置为准。有效设置是指设置确定性状态后，真正执行了一次算子任务下发。如果仅设置，没有算子下发，只能是确定性变量开启但未下发给算子，因此不执行算子。
21> 解决方案：暂不推荐一个线程多次设置确定性。该问题在二进制开启和关闭情况下均存在，在后续版本中会解决该问题。21> 解决方案：暂不推荐一个线程多次设置确定性。该问题在二进制开启和关闭情况下均存在，在后续版本中会解决该问题。
22> - 符号说明：表中 “ - ” 符号表示该接口暂不支持当前列产品。22> - 符号说明：表中 “ - ” 符号表示该接口暂不支持当前列产品。
@@ -191,7 +191,7 @@
191| [aclnnForeachTan](../../foreach/foreach_tan/docs/aclnnForeachTan.md) | 对输入张量列表的每个张量进行正切函数运算。 | 默认确定性实现 | |191| [aclnnForeachTan](../../foreach/foreach_tan/docs/aclnnForeachTan.md) | 对输入张量列表的每个张量进行正切函数运算。 | 默认确定性实现 | |
192| [aclnnForeachTanh](../../foreach/foreach_tanh/docs/aclnnForeachTanh.md) | 对输入张量列表的每个张量进行双曲正切函数运算。 | 默认确定性实现 | |192| [aclnnForeachTanh](../../foreach/foreach_tanh/docs/aclnnForeachTanh.md) | 对输入张量列表的每个张量进行双曲正切函数运算。 | 默认确定性实现 | |
193| [aclnnForeachZeroInplace](../../foreach/foreach_zero_inplace/docs/aclnnForeachZeroInplace.md) | 原地更新输入张量列表，输入张量列表的每个张量置为0。 | 默认确定性实现 | |193| [aclnnForeachZeroInplace](../../foreach/foreach_zero_inplace/docs/aclnnForeachZeroInplace.md) | 原地更新输入张量列表，输入张量列表的每个张量置为0。 | 默认确定性实现 | |
194-| [aclnnFusedLinearOnlineMaxSum](../../matmul/fused_linear_online_max_sum/docs/aclnnFusedLinearOnlineMaxSum.md) | 功能等价Megatron的matmul与fused\_vocab\_parallel\_cross\_entropy的实现，支持vocabulary\_size维度切卡融合matmul与celoss。 | 默认确定性实现 | 默认确定性实现 |194+| [aclnnFusedLinearOnlineMaxSum](../../matmul/fused_linear_online_max_sum/docs/aclnnFusedLinearOnlineMaxSum.md) | 功能等价Megatron的matmul与fused\_vocab\_parallel\_cross\_entropy的实现，支持vocabulary\_size维度切卡融合matmul与cross-entropy loss。 | 默认确定性实现 | 默认确定性实现 |
195| [aclnnFusedLinearCrossEntropyLossGrad](../../matmul/fused_linear_cross_entropy_loss_grad/docs/aclnnFusedLinearCrossEntropyLossGrad.md) | 是词汇表并行场景下交叉熵损失计算模块中的一部分，解决超大规模词汇表下的显存和计算效率问题，当前部分为梯度计算实现，用于计算叶子节点`input`和`weight`的梯度。 | 默认确定性实现 | - |195| [aclnnFusedLinearCrossEntropyLossGrad](../../matmul/fused_linear_cross_entropy_loss_grad/docs/aclnnFusedLinearCrossEntropyLossGrad.md) | 是词汇表并行场景下交叉熵损失计算模块中的一部分，解决超大规模词汇表下的显存和计算效率问题，当前部分为梯度计算实现，用于计算叶子节点`input`和`weight`的梯度。 | 默认确定性实现 | - |
196| [aclnnFusedMatmul](../../matmul/fused_mat_mul/docs/aclnnFusedMatmul.md) | 矩阵乘与通用向量计算融合。 | 默认确定性实现 | 默认确定性实现 |196| [aclnnFusedMatmul](../../matmul/fused_mat_mul/docs/aclnnFusedMatmul.md) | 矩阵乘与通用向量计算融合。 | 默认确定性实现 | 默认确定性实现 |
197| [aclnnFusedQuantMatmul](../../matmul/fused_quant_mat_mul/docs/aclnnFusedQuantMatmul.md) | 量化矩阵乘与通用向量计算融合。 | 默认确定性实现 | - |197| [aclnnFusedQuantMatmul](../../matmul/fused_quant_mat_mul/docs/aclnnFusedQuantMatmul.md) | 量化矩阵乘与通用向量计算融合。 | 默认确定性实现 | - |
@@ -270,9 +270,9 @@
270| [aclnnMaxPool2dWithMask](../../pooling/max_pool3d_with_argmax_v2/docs/aclnnMaxPool2dWithMask.md) | 对于输入信号的输入通道，提供2维最大池化（max pooling）操作，输出池化后的值out和索引indices（采用mask语义计算得出）。 | 默认确定性实现 | - |270| [aclnnMaxPool2dWithMask](../../pooling/max_pool3d_with_argmax_v2/docs/aclnnMaxPool2dWithMask.md) | 对于输入信号的输入通道，提供2维最大池化（max pooling）操作，输出池化后的值out和索引indices（采用mask语义计算得出）。 | 默认确定性实现 | - |
271| [aclnnMaxPool2dWithMaskBackward](../../pooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool2dWithMaskBackward.md) | 正向最大池化aclnnMaxPool2dWithMask的反向传播。 | 默认非确定性实现，支持配置开启。 | - |271| [aclnnMaxPool2dWithMaskBackward](../../pooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool2dWithMaskBackward.md) | 正向最大池化aclnnMaxPool2dWithMask的反向传播。 | 默认非确定性实现，支持配置开启。 | - |
272| [aclnnMaxPool3dWithArgmax](../../pooling/max_pool3d_with_argmax_v2/docs/aclnnMaxPool3dWithArgmax.md) | 对于输入信号的输入通道，提供3维最大池化（max pooling）操作，输出池化后的值out和索引indices。 | 默认确定性实现 | 默认确定性实现 |272| [aclnnMaxPool3dWithArgmax](../../pooling/max_pool3d_with_argmax_v2/docs/aclnnMaxPool3dWithArgmax.md) | 对于输入信号的输入通道，提供3维最大池化（max pooling）操作，输出池化后的值out和索引indices。 | 默认确定性实现 | 默认确定性实现 |
273-| [aclnnMaxPool3dWithArgmaxBackWard](../../pooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool3dWithArgmaxBackward.md) | 正向最大池化aclnnMaxPool3dWithArgmax的反向传播，将梯度回填到每个窗口最大值的坐标处，相同坐标处累加。 | 默认非确定性实现，支持配置开启。 | 默认确定性实现 |273+| [aclnnMaxPool3dWithArgmaxBackward](../../pooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool3dWithArgmaxBackward.md) | 正向最大池化aclnnMaxPool3dWithArgmax的反向传播，将梯度回填到每个窗口最大值的坐标处，相同坐标处累加。 | 默认非确定性实现，支持配置开启。 | 默认确定性实现 |
274| [aclnnMaxUnpool2dBackward](../../index/gather_elements/docs/aclnnMaxUnpool2dBackward.md) | MaxPool2d的逆运算aclnnMaxUnpool2d的反向传播，根据indices索引在out中填入gradOutput的元素值。 | 默认确定性实现 | 默认非确定性实现，支持配置开启 |274| [aclnnMaxUnpool2dBackward](../../index/gather_elements/docs/aclnnMaxUnpool2dBackward.md) | MaxPool2d的逆运算aclnnMaxUnpool2d的反向传播，根据indices索引在out中填入gradOutput的元素值。 | 默认确定性实现 | 默认非确定性实现，支持配置开启 |
275-| [aclnnMaxUnpool3dBackward](../../index/gather_elements/docs/aclnnMaxUnpool3dBackward.md) | axPool3d的逆运算aclnnMaxUnpool3d的反向传播，根据indices索引在out中填入gradOutput的元素值。 | 默认确定性实现 | 默认非确定性实现，支持配置开启 |275+| [aclnnMaxUnpool3dBackward](../../index/gather_elements/docs/aclnnMaxUnpool3dBackward.md) | MaxPool3d的逆运算aclnnMaxUnpool3d的反向传播，根据indices索引在out中填入gradOutput的元素值。 | 默认确定性实现 | 默认非确定性实现，支持配置开启 |
276| [aclnnMedian](../../index/gather_v2/docs/aclnnMedian.md) | 返回所有元素的中位数。 | 默认确定性实现 | 默认确定性实现 |276| [aclnnMedian](../../index/gather_v2/docs/aclnnMedian.md) | 返回所有元素的中位数。 | 默认确定性实现 | 默认确定性实现 |
277| [aclnnMm](../../matmul/mat_mul_v3/docs/aclnnMm.md) | 完成2维张量self与张量mat2的矩阵乘计算。 | 默认确定性实现 | 默认确定性实现 |277| [aclnnMm](../../matmul/mat_mul_v3/docs/aclnnMm.md) | 完成2维张量self与张量mat2的矩阵乘计算。 | 默认确定性实现 | 默认确定性实现 |
278| [aclnnMish&aclnnInplaceMish](../../activation/mish/docs/aclnnMish&aclnnInplaceMish.md) | 一个自正则化的非单调神经网络激活函数。 | 默认确定性实现 | 默认确定性实现 |278| [aclnnMish&aclnnInplaceMish](../../activation/mish/docs/aclnnMish&aclnnInplaceMish.md) | 一个自正则化的非单调神经网络激活函数。 | 默认确定性实现 | 默认确定性实现 |
Mpooling/avg_pool3_d/docs/aclnnAvgPool3d.md+1-1
@@ -106,7 +106,7 @@ aclnnStatus aclnnAvgPool3d(
106 <td>-</td>106 <td>-</td>
107 </tr>107 </tr>
108 <tr>108 <tr>
109- <td>stride</td>109+ <td>strides</td>
110 <td>输入</td>110 <td>输入</td>
111 <td>池化操作的步长，公式中的strides。</td>111 <td>池化操作的步长，公式中的strides。</td>
112 <td>长度为0（数值与kernelSize数值保持一致）或者1(SD = SH = SW)或者3(SD, SH, SW)，长度为1或3时数值必须大于0。</td>112 <td>长度为0（数值与kernelSize数值保持一致）或者1(SD = SH = SW)或者3(SD, SH, SW)，长度为1或3时数值必须大于0。</td>
Mpooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool2dWithIndicesBackward.md+1-1
@@ -254,7 +254,7 @@ aclnnStatus aclnnMaxPool2dWithIndicesBackward(
254 <td>padding的元素个数不等于1或2</td>254 <td>padding的元素个数不等于1或2</td>
255 </tr>255 </tr>
256 <tr>256 <tr>
257- <td>padding的数值中存在小于0或者大于kernelSize</td>257+ <td>padding的数值中存在小于0或者大于kernelSize的数值</td>
258 </tr>258 </tr>
259 <tr>259 <tr>
260 <td>dilation的元素数值不符合入参要求。</td>260 <td>dilation的元素数值不符合入参要求。</td>
Mpooling/max_pool3d_grad_with_argmax/docs/aclnnMaxPool2dWithMaskBackward.md+1-1
@@ -229,7 +229,7 @@ aclnnStatus aclnnMaxPool2dWithMaskBackward(
229 <td>padding的长度不等于1或2。</td>229 <td>padding的长度不等于1或2。</td>
230 </tr>230 </tr>
231 <tr>231 <tr>
232- <td>padding的数值中存在小于0或者大于kernelSize</td>232+ <td>padding的数值中存在小于0或者大于kernelSize的数值</td>
233 </tr>233 </tr>
234 <tr>234 <tr>
235 <td>dilation的数值不等于1。</td>235 <td>dilation的数值不等于1。</td>
Mpooling/max_pool3d_with_argmax_v2/docs/aclnnMaxPool2dWithIndices.md+1-1
@@ -240,7 +240,7 @@ aclnnStatus aclnnMaxPool2dWithIndices(
240 <td>padding的长度不等于1或2。</td>240 <td>padding的长度不等于1或2。</td>
241 </tr>241 </tr>
242 <tr>242 <tr>
243- <td>padding的数值中存在小于0或者大于kernelSize</td>243+ <td>padding的数值中存在小于0或者大于kernelSize/2</td>
244 </tr>244 </tr>
245 <tr>245 <tr>
246 <td>dilation的长度不等于1或2。</td>246 <td>dilation的长度不等于1或2。</td>
Mquant/fake_quant_affine_cachemask/docs/aclnnFakeQuantPerTensorAffineCachemask.md+1-1
@@ -367,7 +367,7 @@ int main() {
367 std::vector<char> maskHostData{1};367 std::vector<char> maskHostData{1};
368 int64_t quantMin = 1;368 int64_t quantMin = 1;
369 int64_t quantMax = 3;369 int64_t quantMax = 3;
370- float fakeQuantEnabled;370+ float fakeQuantEnabled = 1.0f;
371 // 创建 aclTensor371 // 创建 aclTensor
372 ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);372 ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
373 CHECK_RET(ret == ACL_SUCCESS, return ret);373 CHECK_RET(ret == ACL_SUCCESS, return ret);
Mquant/fake_quant_affine_cachemask/examples/test_aclnn_fake_quant_per_tensor_affine_cachemask.cpp+6-6
@@ -68,13 +68,13 @@ int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>&
68}68}
69 69 
70int main() {70int main() {
71- // 1. （固定写法）device/stream初始化，参考acl API71+ // 1. （固定写法）device/stream初始化，参考acl API手册
72 // 根据自己的实际device填写deviceId72 // 根据自己的实际device填写deviceId
73 int32_t deviceId = 0;73 int32_t deviceId = 0;
74 aclrtStream stream;74 aclrtStream stream;
75 auto ret = Init(deviceId, &stream);75 auto ret = Init(deviceId, &stream);
76 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);76 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
77- 77+
78 // 2. 构造输入与输出，需要根据API的接口自定义构造78 // 2. 构造输入与输出，需要根据API的接口自定义构造
79 std::vector<int64_t> selfShape = {1};79 std::vector<int64_t> selfShape = {1};
80 std::vector<int64_t> scaleShape = {1};80 std::vector<int64_t> scaleShape = {1};
@@ -98,7 +98,7 @@ int main() {
98 std::vector<char> maskHostData{1};98 std::vector<char> maskHostData{1};
99 int64_t quantMin = 1;99 int64_t quantMin = 1;
100 int64_t quantMax = 3;100 int64_t quantMax = 3;
101- float fakeQuantEnbled;101+ float fakeQuantEnabled = 1.0f;
102 // 创建 aclTensor102 // 创建 aclTensor
103 ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);103 ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
104 CHECK_RET(ret == ACL_SUCCESS, return ret);104 CHECK_RET(ret == ACL_SUCCESS, return ret);
@@ -115,7 +115,7 @@ int main() {
115 uint64_t workspaceSize = 0;115 uint64_t workspaceSize = 0;
116 aclOpExecutor* executor;116 aclOpExecutor* executor;
117 // 调用aclnnEye第一段接口117 // 调用aclnnEye第一段接口
118- ret = aclnnFakeQuantPerTensorAffineCachemaskGetWorkspaceSize(self, scale, zeroPoint, fakeQuantEnbled, quantMin, quantMax, out, mask, &workspaceSize, &executor);118+ ret = aclnnFakeQuantPerTensorAffineCachemaskGetWorkspaceSize(self, scale, zeroPoint, fakeQuantEnabled, quantMin, quantMax, out, mask, &workspaceSize, &executor);
119 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFakeQuantPerTensorAffineCachemaskGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);119 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFakeQuantPerTensorAffineCachemaskGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
120 // 根据第一段接口计算出的workspaceSize申请device内存120 // 根据第一段接口计算出的workspaceSize申请device内存
121 void* workspaceAddr = nullptr;121 void* workspaceAddr = nullptr;
@@ -126,11 +126,11 @@ int main() {
126 // 调用aclnnFakeQuantPerTensorAffineCachemask第二段接口126 // 调用aclnnFakeQuantPerTensorAffineCachemask第二段接口
127 ret = aclnnFakeQuantPerTensorAffineCachemask(workspaceAddr, workspaceSize, executor, stream);127 ret = aclnnFakeQuantPerTensorAffineCachemask(workspaceAddr, workspaceSize, executor, stream);
128 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFakeQuantPerTensorAffineCachemask failed. ERROR: %d\n", ret); return ret);128 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFakeQuantPerTensorAffineCachemask failed. ERROR: %d\n", ret); return ret);
129- 129+
130 // 4. （固定写法）同步等待任务执行结束130 // 4. （固定写法）同步等待任务执行结束
131 ret = aclrtSynchronizeStream(stream);131 ret = aclrtSynchronizeStream(stream);
132 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);132 CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
133- 133+
134 // 5. 获取输出的值，将device侧内存上的结果拷贝至host侧，需要根据具体API的接口定义修改134 // 5. 获取输出的值，将device侧内存上的结果拷贝至host侧，需要根据具体API的接口定义修改
135 auto size = GetShapeSize(outShape);135 auto size = GetShapeSize(outShape);
136 std::vector<float> resultData(size, 0);136 std::vector<float> resultData(size, 0);
Mrnn/dynamic_rnn/docs/aclnnLSTM.md+14-14
@@ -47,12 +47,12 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
47 const aclTensorList *params,47 const aclTensorList *params,
48 const aclTensorList *hx,48 const aclTensorList *hx,
49 const aclTensor *batchSizes,49 const aclTensor *batchSizes,
50- bool has_biases,50+ bool hasBias,
51 int64_t numLayers,51 int64_t numLayers,
52- double droupout,52+ double dropout,
53 bool train,53 bool train,
54 bool bidirectional, 54 bool bidirectional,
55- bool batch_first,55+ bool batchFirst,
56 aclTensor *output,56 aclTensor *output,
57 aclTensor *hy,57 aclTensor *hy,
58 aclTensor *cy,58 aclTensor *cy,
@@ -108,12 +108,12 @@ aclnnStatus aclnnLSTM(
108 <td>108 <td>
109 <ul>109 <ul>
110 <li><strong>若batchSizes传入空指针：</strong>110 <li><strong>若batchSizes传入空指针：</strong>
111- <br>shape格式根据batch_first参数区分：111+ <br>shape格式根据batchFirst参数区分：
112 <ul>112 <ul>
113- <li>batch_first=False：(time_step, batch_size, input_size)</li>113+ <li>batchFirst=False：(time_step, batch_size, input_size)</li>
114- <li>batch_first=True：(batch_size, time_step, input_size)</li>114+ <li>batchFirst=True：(batch_size, time_step, input_size)</li>
115 </ul>115 </ul>
116- 说明：batch_first表示batch维度是否在第一维；time_step为时间维度；batch_size为每个时刻处理的样本数；input_size为输入特征数。116+ 说明：batchFirst表示batch维度是否在第一维；time_step为时间维度；batch_size为每个时刻处理的样本数；input_size为输入特征数。
117 </li>117 </li>
118 <li><strong>若传入有效batchSizes：</strong>118 <li><strong>若传入有效batchSizes：</strong>
119 <br>shape格式：(time_step * batch_size, input_size)119 <br>shape格式：(time_step * batch_size, input_size)
@@ -201,7 +201,7 @@ aclnnStatus aclnnLSTM(
201 <td>√</td>201 <td>√</td>
202 </tr>202 </tr>
203 <tr>203 <tr>
204- <td>droupout</td>204+ <td>dropout</td>
205 <td>输入</td>205 <td>输入</td>
206 <td>表示随机掩码的概率。</td>206 <td>表示随机掩码的概率。</td>
207 <td>当前不支持该功能</td>207 <td>当前不支持该功能</td>
@@ -221,7 +221,7 @@ aclnnStatus aclnnLSTM(
221 <td>√</td>221 <td>√</td>
222 </tr>222 </tr>
223 <tr>223 <tr>
224- <td>bidirection</td>224+ <td>bidirectional</td>
225 <td>输入</td>225 <td>输入</td>
226 <td>表示是否是双向。</td>226 <td>表示是否是双向。</td>
227 <td>/</td>227 <td>/</td>
@@ -244,7 +244,7 @@ aclnnStatus aclnnLSTM(
244 <td>output</td>244 <td>output</td>
245 <td>输出</td>245 <td>输出</td>
246 <td>表示LSTM运算中最后一层每个时间步的输出结果。</td>246 <td>表示LSTM运算中最后一层每个时间步的输出结果。</td>
247- <td><ul><li>若batchSizes传入空指针：<br>当batch_first=False时shape支持三维（time_step, batch_size, D * hidden_size），否则支持三维（batch_size, time_step, D * hidden_size）。</li><li>若传入有效batchSizes：<br>shape应为(time_step, batch_size, D * hidden_size)。</li></ul></td>247+ <td><ul><li>若batchSizes传入空指针：<br>当batchFirst=False时shape支持三维（time_step, batch_size, D * hidden_size），否则支持三维（batch_size, time_step, D * hidden_size）。</li><li>若传入有效batchSizes：<br>shape应为(time_step, batch_size, D * hidden_size)。</li></ul></td>
248 <td>FLOAT16、FLOAT32</td>248 <td>FLOAT16、FLOAT32</td>
249 <td>ND</td>249 <td>ND</td>
250 <td>3</td>250 <td>3</td>
@@ -254,7 +254,7 @@ aclnnStatus aclnnLSTM(
254 <td>hy</td>254 <td>hy</td>
255 <td>输出</td>255 <td>输出</td>
256 <td>表示进行LSTM运算中每层最后一个时间步的隐藏层（公式（7）的输出）。</td>256 <td>表示进行LSTM运算中每层最后一个时间步的隐藏层（公式（7）的输出）。</td>
257- <td>shape支持三维（D * num_layers, batch_size, hidden_size</td>257+ <td>shape支持三维（D * num_layers, batch_size, hidden_size）</td>
258 <td>FLOAT16、FLOAT32</td>258 <td>FLOAT16、FLOAT32</td>
259 <td>ND</td>259 <td>ND</td>
260 <td>3</td>260 <td>3</td>
@@ -264,7 +264,7 @@ aclnnStatus aclnnLSTM(
264 <td>cy</td>264 <td>cy</td>
265 <td>输出</td>265 <td>输出</td>
266 <td>表示进行LSTM运算中每层最后一个时间步的Cell状态（公式（5）的输出）。</td>266 <td>表示进行LSTM运算中每层最后一个时间步的Cell状态（公式（5）的输出）。</td>
267- <td>shape支持三维（D * num_layers, batch_size, hidden_size</td>267+ <td>shape支持三维（D * num_layers, batch_size, hidden_size）</td>
268 <td>FLOAT16、FLOAT32</td>268 <td>FLOAT16、FLOAT32</td>
269 <td>ND</td>269 <td>ND</td>
270 <td>3</td>270 <td>3</td>
@@ -274,7 +274,7 @@ aclnnStatus aclnnLSTM(
274 <td>hy</td>274 <td>hy</td>
275 <td>输出</td>275 <td>输出</td>
276 <td>表示进行LSTM运算中每层最后一个时间步的隐藏层（公式（7）的输出）。</td>276 <td>表示进行LSTM运算中每层最后一个时间步的隐藏层（公式（7）的输出）。</td>
277- <td>shape支持三维（D * num_layers, batch_size, hidden_size</td>277+ <td>shape支持三维（D * num_layers, batch_size, hidden_size）</td>
278 <td>FLOAT16、FLOAT32</td>278 <td>FLOAT16、FLOAT32</td>
279 <td>ND</td>279 <td>ND</td>
280 <td>3</td>280 <td>3</td>
@@ -284,7 +284,7 @@ aclnnStatus aclnnLSTM(
284 <td>cy</td>284 <td>cy</td>
285 <td>输出</td>285 <td>输出</td>
286 <td>表示进行LSTM运算中每层最后一个时间步的Cell状态（公式（5）的输出）。</td>286 <td>表示进行LSTM运算中每层最后一个时间步的Cell状态（公式（5）的输出）。</td>
287- <td>shape支持三维（D * num_layers, batch_size, hidden_size</td>287+ <td>shape支持三维（D * num_layers, batch_size, hidden_size）</td>
288 <td>FLOAT16、FLOAT32</td>288 <td>FLOAT16、FLOAT32</td>
289 <td>ND</td>289 <td>ND</td>
290 <td>3</td>290 <td>3</td>
Mrnn/dynamic_rnn/op_host/op_api/aclnn_lstm.cpp+26-26
@@ -42,7 +42,7 @@ struct LstmDataParamsIn {
42 const aclTensorList *hx;42 const aclTensorList *hx;
43 const aclTensor *batchSizes;43 const aclTensor *batchSizes;
44 int64_t numLayers;44 int64_t numLayers;
45- bool has_biases;45+ bool hasBias;
46 bool train;46 bool train;
47 bool bidirectional;47 bool bidirectional;
48};48};
@@ -123,10 +123,10 @@ auto nullptrInner = std::tuple<aclTensor*, aclTensor*, aclTensor*, aclTensor*, a
123 123 
124std::tuple<const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *> LstmSingleLayerDirec(124std::tuple<const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *, const aclTensor *> LstmSingleLayerDirec(
125 const aclTensor * input, const aclTensorList * params, const aclTensorList * hx, aclTensor *yOutDirec, aclTensor *iOutDirec, aclTensor *jOutDirec, aclTensor *fOutDirec, aclTensor *oOutDirec, aclTensor *hOutDirec, aclTensor *cOutDirec, aclTensor *tanhCOutDirec, 125 const aclTensor * input, const aclTensorList * params, const aclTensorList * hx, aclTensor *yOutDirec, aclTensor *iOutDirec, aclTensor *jOutDirec, aclTensor *fOutDirec, aclTensor *oOutDirec, aclTensor *hOutDirec, aclTensor *cOutDirec, aclTensor *tanhCOutDirec,
126- const char *direction, bool bidirectional, bool train, int64_t num_layers, bool has_biases, aclOpExecutor* executor)126+ const char *direction, bool bidirectional, bool train, int64_t num_layers, bool hasBias, aclOpExecutor* executor)
127{127{
128 auto oneLayerParams = bidirectional == true ? 4 : 2;128 auto oneLayerParams = bidirectional == true ? 4 : 2;
129- oneLayerParams = has_biases == true ? oneLayerParams * 2 : oneLayerParams;129+ oneLayerParams = hasBias == true ? oneLayerParams * 2 : oneLayerParams;
130 auto weightStart = strcmp(direction, "UNIDIRECTIONAL") == 0 ? 0 : oneLayerParams / 2;130 auto weightStart = strcmp(direction, "UNIDIRECTIONAL") == 0 ? 0 : oneLayerParams / 2;
131 auto paramsOffsets = oneLayerParams * num_layers + weightStart;131 auto paramsOffsets = oneLayerParams * num_layers + weightStart;
132 op::FVector<const aclTensor*> weightConcatList;132 op::FVector<const aclTensor*> weightConcatList;
@@ -144,7 +144,7 @@ std::tuple<const aclTensor *, const aclTensor *, const aclTensor *, const aclTen
144 OP_CHECK_NULL(weightTrans, return nullptrInner);144 OP_CHECK_NULL(weightTrans, return nullptrInner);
145 145 
146 const aclTensor * bias = nullptr;146 const aclTensor * bias = nullptr;
147- if (has_biases) {147+ if (hasBias) {
148 bias = l0op::Add((*params)[paramsOffsets + 2], (*params)[paramsOffsets + 3], executor);148 bias = l0op::Add((*params)[paramsOffsets + 2], (*params)[paramsOffsets + 3], executor);
149 OP_CHECK_NULL(bias, return nullptrInner);149 OP_CHECK_NULL(bias, return nullptrInner);
150 } else {150 } else {
@@ -307,10 +307,10 @@ static inline bool CheckDtypeValid(const aclTensor *input, const aclTensorList
307 return true;307 return true;
308}308}
309 309 
310-static bool CheckDimsSize(const aclTensorList *params, const aclTensorList *hx, bool has_biases, int64_t numLayers, bool train, bool bidirectional, aclTensorList *iOut, aclTensorList *jOut, aclTensorList *fOut, aclTensorList *oOut, 310+static bool CheckDimsSize(const aclTensorList *params, const aclTensorList *hx, bool hasBias, int64_t numLayers, bool train, bool bidirectional, aclTensorList *iOut, aclTensorList *jOut, aclTensorList *fOut, aclTensorList *oOut,
311 aclTensorList *hOut, aclTensorList *cOut, aclTensorList *tanhCOut) {311 aclTensorList *hOut, aclTensorList *cOut, aclTensorList *tanhCOut) {
312 uint64_t dScale = bidirectional == true ? 2 : 1;312 uint64_t dScale = bidirectional == true ? 2 : 1;
313- uint64_t bScale = has_biases == true ? 2 : 1;313+ uint64_t bScale = hasBias == true ? 2 : 1;
314 uint64_t output_nums = dScale * numLayers;314 uint64_t output_nums = dScale * numLayers;
315 uint64_t param_nums = 2 * bScale * dScale * numLayers;315 uint64_t param_nums = 2 * bScale * dScale * numLayers;
316 316 
@@ -355,11 +355,11 @@ static bool CheckDimsSize(const aclTensorList *params, const aclTensorList *hx,
355 return true;355 return true;
356}356}
357 357 
358-static bool CheckDims(const aclTensor *input, const aclTensorList *params, const aclTensorList *hx, bool has_biases, int64_t numLayers, bool train, bool bidirectional, 358+static bool CheckDims(const aclTensor *input, const aclTensorList *params, const aclTensorList *hx, bool hasBias, int64_t numLayers, bool train, bool bidirectional,
359 aclTensor *output, aclTensor *hy, aclTensor *cy, aclTensorList *iOut, aclTensorList *jOut, aclTensorList *fOut, aclTensorList *oOut, 359 aclTensor *output, aclTensor *hy, aclTensor *cy, aclTensorList *iOut, aclTensorList *jOut, aclTensorList *fOut, aclTensorList *oOut,
360 aclTensorList *hOut, aclTensorList *cOut, aclTensorList *tanhCOut) {360 aclTensorList *hOut, aclTensorList *cOut, aclTensorList *tanhCOut) {
361 OP_CHECK_WRONG_DIMENSION(input, INPUT_DIMS, return false);361 OP_CHECK_WRONG_DIMENSION(input, INPUT_DIMS, return false);
362- uint64_t bScale = has_biases == true ? 2 : 1;362+ uint64_t bScale = hasBias == true ? 2 : 1;
363 uint64_t dScale = bidirectional == true ? 2 : 1;363 uint64_t dScale = bidirectional == true ? 2 : 1;
364 uint64_t oneLayerParams = 2 * bScale * dScale;364 uint64_t oneLayerParams = 2 * bScale * dScale;
365 for (uint64_t i = 0; i < (uint64_t)numLayers; i++) {365 for (uint64_t i = 0; i < (uint64_t)numLayers; i++) {
@@ -367,7 +367,7 @@ static bool CheckDims(const aclTensor *input, const aclTensorList *params, cons
367 uint64_t offsets = i * oneLayerParams + j * oneLayerParams / 2;367 uint64_t offsets = i * oneLayerParams + j * oneLayerParams / 2;
368 OP_CHECK_WRONG_DIMENSION((*params)[offsets], WEIGHT_DIMS, return false);368 OP_CHECK_WRONG_DIMENSION((*params)[offsets], WEIGHT_DIMS, return false);
369 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 1], WEIGHT_DIMS, return false);369 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 1], WEIGHT_DIMS, return false);
370- if (has_biases) {370+ if (hasBias) {
371 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 2], BIAS_DIMS, return false);371 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 2], BIAS_DIMS, return false);
372 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 3], BIAS_DIMS, return false);372 OP_CHECK_WRONG_DIMENSION((*params)[offsets + 3], BIAS_DIMS, return false);
373 }373 }
@@ -640,7 +640,7 @@ static aclnnStatus CheckDimsAndListLength(const LstmDataParamsIn& inputs, const
640 int64_t currOffset = group * info.groupLen;640 int64_t currOffset = group * info.groupLen;
641 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_0], INDEX_2, return ACLNN_ERR_PARAM_INVALID);641 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_0], INDEX_2, return ACLNN_ERR_PARAM_INVALID);
642 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_1], INDEX_2, return ACLNN_ERR_PARAM_INVALID);642 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_1], INDEX_2, return ACLNN_ERR_PARAM_INVALID);
643- if (inputs.has_biases) {643+ if (inputs.hasBias) {
644 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_2], INDEX_1, return ACLNN_ERR_PARAM_INVALID);644 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_2], INDEX_1, return ACLNN_ERR_PARAM_INVALID);
645 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_3], INDEX_1, return ACLNN_ERR_PARAM_INVALID);645 OP_CHECK_WRONG_DIMENSION((*inputs.params)[currOffset + INDEX_3], INDEX_1, return ACLNN_ERR_PARAM_INVALID);
646 }646 }
@@ -716,7 +716,7 @@ static aclnnStatus CheckShapes(const LstmDataParamsIn& inputs, const LstmDataPar
716 return ACLNN_ERR_PARAM_INVALID716 return ACLNN_ERR_PARAM_INVALID
717 );717 );
718 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_1], weightHhShape, return ACLNN_ERR_PARAM_INVALID);718 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_1], weightHhShape, return ACLNN_ERR_PARAM_INVALID);
719- if (inputs.has_biases) {719+ if (inputs.hasBias) {
720 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_2], biasShape, return ACLNN_ERR_PARAM_INVALID);720 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_2], biasShape, return ACLNN_ERR_PARAM_INVALID);
721 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_3], biasShape, return ACLNN_ERR_PARAM_INVALID);721 OP_CHECK_SHAPE_NOT_EQUAL_WITH_EXPECTED_SIZE((*inputs.params)[currOffset + INDEX_3], biasShape, return ACLNN_ERR_PARAM_INVALID);
722 }722 }
@@ -807,7 +807,7 @@ static aclnnStatus CheckParamsValid(const LstmDataParamsIn& inputs, const LstmDa
807 807 
808 info.L = inputs.numLayers;808 info.L = inputs.numLayers;
809 info.D = (inputs.bidirectional) ? INDEX_2 : INDEX_1;809 info.D = (inputs.bidirectional) ? INDEX_2 : INDEX_1;
810- info.groupLen = (inputs.has_biases) ? INDEX_4 : INDEX_2;810+ info.groupLen = (inputs.hasBias) ? INDEX_4 : INDEX_2;
811 info.LD = info.L * info.D;811 info.LD = info.L * info.D;
812 812 
813 // list长度与tensor dim校验813 // list长度与tensor dim校验
@@ -992,7 +992,7 @@ static aclnnStatus LstmDataProcessParams(
992 CHECK_RET(baseIn.weight != nullptr, ACLNN_ERR_INNER_NULLPTR);992 CHECK_RET(baseIn.weight != nullptr, ACLNN_ERR_INNER_NULLPTR);
993 993 
994 // bias。add994 // bias。add
995- if (inputs.has_biases) {995+ if (inputs.hasBias) {
996 baseIn.bias = l0op::Add(996 baseIn.bias = l0op::Add(
997 (*inputs.params)[currOffset + INDEX_2],997 (*inputs.params)[currOffset + INDEX_2],
998 (*inputs.params)[currOffset + INDEX_3],998 (*inputs.params)[currOffset + INDEX_3],
@@ -1213,12 +1213,12 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1213 const aclTensorList *params,1213 const aclTensorList *params,
1214 const aclTensorList *hx,1214 const aclTensorList *hx,
1215 const aclTensor *batchSizes,1215 const aclTensor *batchSizes,
1216- bool has_biases,1216+ bool hasBias,
1217 int64_t numLayers,1217 int64_t numLayers,
1218- double droupout,1218+ double dropout,
1219 bool train,1219 bool train,
1220 bool bidirectional,1220 bool bidirectional,
1221- bool batch_first,1221+ bool batchFirst,
1222 aclTensor *output,1222 aclTensor *output,
1223 aclTensor *hy,1223 aclTensor *hy,
1224 aclTensor *cy,1224 aclTensor *cy,
@@ -1232,12 +1232,12 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1232 uint64_t *workspaceSize,1232 uint64_t *workspaceSize,
1233 aclOpExecutor **executor){1233 aclOpExecutor **executor){
1234 OP_CHECK_COMM_INPUT(workspaceSize, executor);1234 OP_CHECK_COMM_INPUT(workspaceSize, executor);
1235- L2_DFX_PHASE_1(aclnnLSTM, DFX_IN(input, params, hx, batchSizes, has_biases, numLayers, droupout, train, bidirectional, batch_first), 1235+ L2_DFX_PHASE_1(aclnnLSTM, DFX_IN(input, params, hx, batchSizes, hasBias, numLayers, dropout, train, bidirectional, batchFirst),
1236 DFX_OUT(output, hy, cy, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut));1236 DFX_OUT(output, hy, cy, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut));
1237 1237 
1238 // 判断是否进入data模式1238 // 判断是否进入data模式
1239 if (batchSizes) {1239 if (batchSizes) {
1240- LstmDataParamsIn inputs = {input, params, hx, batchSizes, numLayers, has_biases, train, bidirectional};1240+ LstmDataParamsIn inputs = {input, params, hx, batchSizes, numLayers, hasBias, train, bidirectional};
1241 LstmDataParamsOut outputs = {output, hy, cy, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut};1241 LstmDataParamsOut outputs = {output, hy, cy, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut};
1242 return LstmDataGetWorkspaceSize(inputs, outputs, workspaceSize, executor);1242 return LstmDataGetWorkspaceSize(inputs, outputs, workspaceSize, executor);
1243 }1243 }
@@ -1254,7 +1254,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1254 }1254 }
1255 1255 
1256 // 固定写法，参数检查1256 // 固定写法，参数检查
1257- auto ret = CheckParams(input, params, hx, has_biases, numLayers, train, bidirectional, batch_first, output, hy, cy,1257+ auto ret = CheckParams(input, params, hx, hasBias, numLayers, train, bidirectional, batchFirst, output, hy, cy,
1258 iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut);1258 iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut);
1259 CHECK_RET(ret == ACLNN_SUCCESS, ret);1259 CHECK_RET(ret == ACLNN_SUCCESS, ret);
1260 1260 
@@ -1275,7 +1275,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1275 1275 
1276 // 输入batchFirst转换1276 // 输入batchFirst转换
1277 auto curInput = inputContiguous;1277 auto curInput = inputContiguous;
1278- if (batch_first == true) {1278+ if (batchFirst == true) {
1279 std::vector<int64_t> perm={1, 0, 2};1279 std::vector<int64_t> perm={1, 0, 2};
1280 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);1280 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);
1281 curInput = l0op::Transpose(inputContiguous, valuePerm, uniqueExecutor.get());1281 curInput = l0op::Transpose(inputContiguous, valuePerm, uniqueExecutor.get());
@@ -1309,7 +1309,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1309 CHECK_RET(tanhCOutForward != nullptr, ACLNN_ERR_INNER_NULLPTR);1309 CHECK_RET(tanhCOutForward != nullptr, ACLNN_ERR_INNER_NULLPTR);
1310 1310 
1311 auto layerResultForward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutForward, iOutForward, jOutForward, fOutForward, oOutForward, hOutForward, cOutForward, tanhCOutForward, 1311 auto layerResultForward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutForward, iOutForward, jOutForward, fOutForward, oOutForward, hOutForward, cOutForward, tanhCOutForward,
1312- "UNIDIRECTIONAL", bidirectional, train, i, has_biases, uniqueExecutor.get());1312+ "UNIDIRECTIONAL", bidirectional, train, i, hasBias, uniqueExecutor.get());
1313 1313
1314 ProcessViewCopy(layerResultForward, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut, i, bidirectional, "UNIDIRECTIONAL", uniqueExecutor.get());1314 ProcessViewCopy(layerResultForward, iOut, jOut, fOut, oOut, hOut, cOut, tanhCOut, i, bidirectional, "UNIDIRECTIONAL", uniqueExecutor.get());
1315 ProcessOutputHC(layerResultForward, hyVector, cyVector, "UNIDIRECTIONAL", uniqueExecutor.get());1315 ProcessOutputHC(layerResultForward, hyVector, cyVector, "UNIDIRECTIONAL", uniqueExecutor.get());
@@ -1333,7 +1333,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1333 CHECK_RET(tanhCOutBackward != nullptr, ACLNN_ERR_INNER_NULLPTR);1333 CHECK_RET(tanhCOutBackward != nullptr, ACLNN_ERR_INNER_NULLPTR);
1334 1334 
1335 auto layerResultBackward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutBackward, iOutBackward, jOutBackward, fOutBackward, oOutBackward, hOutBackward, cOutBackward, tanhCOutBackward, 1335 auto layerResultBackward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutBackward, iOutBackward, jOutBackward, fOutBackward, oOutBackward, hOutBackward, cOutBackward, tanhCOutBackward,
1336- "REDIRECTIONAL", bidirectional, train, i, has_biases, uniqueExecutor.get());1336+ "REDIRECTIONAL", bidirectional, train, i, hasBias, uniqueExecutor.get());
1337 // ConcatInput1337 // ConcatInput
1338 op::FVector<const aclTensor*> inputConcat;1338 op::FVector<const aclTensor*> inputConcat;
1339 inputConcat.emplace_back(std::get<0>(layerResultForward));1339 inputConcat.emplace_back(std::get<0>(layerResultForward));
@@ -1348,7 +1348,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1348 }1348 }
1349 1349 
1350 auto outputY = curInput;1350 auto outputY = curInput;
1351- if (batch_first) {1351+ if (batchFirst) {
1352 std::vector<int64_t> perm={1, 0, 2};1352 std::vector<int64_t> perm={1, 0, 2};
1353 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);1353 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);
1354 outputY = l0op::Transpose(curInput, valuePerm, uniqueExecutor.get());1354 outputY = l0op::Transpose(curInput, valuePerm, uniqueExecutor.get());
@@ -1396,7 +1396,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1396 CHECK_RET(cOutForward != nullptr, ACLNN_ERR_INNER_NULLPTR);1396 CHECK_RET(cOutForward != nullptr, ACLNN_ERR_INNER_NULLPTR);
1397 1397 
1398 auto layerResultForward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutForward, iOutForward, jOutForward, fOutForward, oOutForward, hOutForward, cOutForward, tanhCOutForward, 1398 auto layerResultForward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutForward, iOutForward, jOutForward, fOutForward, oOutForward, hOutForward, cOutForward, tanhCOutForward,
1399- "UNIDIRECTIONAL", bidirectional, train, i, has_biases, uniqueExecutor.get());1399+ "UNIDIRECTIONAL", bidirectional, train, i, hasBias, uniqueExecutor.get());
1400 ProcessOutputHC(layerResultForward, hyVector, cyVector, "UNIDIRECTIONAL", uniqueExecutor.get());1400 ProcessOutputHC(layerResultForward, hyVector, cyVector, "UNIDIRECTIONAL", uniqueExecutor.get());
1401 1401
1402 if (bidirectional == true) {1402 if (bidirectional == true) {
@@ -1408,7 +1408,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1408 CHECK_RET(yOutBackward != nullptr, ACLNN_ERR_INNER_NULLPTR);1408 CHECK_RET(yOutBackward != nullptr, ACLNN_ERR_INNER_NULLPTR);
1409 1409 
1410 auto layerResultBackward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutBackward, iOutBackward, jOutBackward, fOutBackward, oOutBackward, hOutBackward, cOutBackward, tanhCOutBackward, 1410 auto layerResultBackward = LstmSingleLayerDirec(curInput, paramsContiguous, hxContiguous, yOutBackward, iOutBackward, jOutBackward, fOutBackward, oOutBackward, hOutBackward, cOutBackward, tanhCOutBackward,
1411- "REDIRECTIONAL", bidirectional, train, i, has_biases, uniqueExecutor.get());1411+ "REDIRECTIONAL", bidirectional, train, i, hasBias, uniqueExecutor.get());
1412 // ConcatInput1412 // ConcatInput
1413 op::FVector<const aclTensor*> inputConcat;1413 op::FVector<const aclTensor*> inputConcat;
1414 inputConcat.emplace_back(std::get<0>(layerResultForward));1414 inputConcat.emplace_back(std::get<0>(layerResultForward));
@@ -1421,7 +1421,7 @@ aclnnStatus aclnnLSTMGetWorkspaceSize(
1421 }1421 }
1422 }1422 }
1423 auto outputY = curInput;1423 auto outputY = curInput;
1424- if (batch_first) {1424+ if (batchFirst) {
1425 std::vector<int64_t> perm={1, 0, 2};1425 std::vector<int64_t> perm={1, 0, 2};
1426 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);1426 auto valuePerm = uniqueExecutor.get()->AllocIntArray(perm.data(), 3);
1427 outputY = l0op::Transpose(curInput, valuePerm, uniqueExecutor.get());1427 outputY = l0op::Transpose(curInput, valuePerm, uniqueExecutor.get());
Mrnn/dynamic_rnn/op_host/op_api/aclnn_lstm.h+3-3
@@ -22,12 +22,12 @@ ACLNN_API aclnnStatus aclnnLSTMGetWorkspaceSize(
22 const aclTensorList *params,22 const aclTensorList *params,
23 const aclTensorList *hx,23 const aclTensorList *hx,
24 const aclTensor *batchSizes,24 const aclTensor *batchSizes,
25- bool has_biases,25+ bool hasBias,
26 int64_t numLayers,26 int64_t numLayers,
27- double droupout,27+ double dropout,
28 bool train,28 bool train,
29 bool bidirectional, 29 bool bidirectional,
30- bool batch_first,30+ bool batchFirst,
31 aclTensor *output,31 aclTensor *output,
32 aclTensor *hy,32 aclTensor *hy,
33 aclTensor *cy,33 aclTensor *cy,
Mrnn/single_layer_lstm_grad/docs/aclnnLstmBackward.md+1-1
@@ -388,7 +388,7 @@ aclnnStatus aclnnLstmBackward(
388 <tr>388 <tr>
389 <td>o</td>389 <td>o</td>
390 <td>输入</td>390 <td>输入</td>
391- <td>LSTM正向中每层输出门的激活值。对于公式中的o。</td>391+ <td>LSTM正向中每层输出门的激活值。对应公式中的o。</td>
392 <td><ul><li>列表长度为 D * num_layers。</li><li>多层双向时tensor间按先双向后多层排布。</li><li>数据类型与input一致。</li></ul></td>392 <td><ul><li>列表长度为 D * num_layers。</li><li>多层双向时tensor间按先双向后多层排布。</li><li>数据类型与input一致。</li></ul></td>
393 <td>FLOAT32、FLOAT16</td>393 <td>FLOAT32、FLOAT16</td>
394 <td>ND</td>394 <td>ND</td>