希姆计算TensorTurbo ONNX算子支持说明
版本历史
| 版本 | 作者 | 日期 | 说明 |
|---|---|---|---|
| V1.4.0 | 希姆计算 | 2022-04-29 | - 新增支持ONNX算子：Cast、CumSum 、Elu、Gather、Log、Logsoftmax、Mod、SpaceToDepth、Softplus、Scatter、ScatterND - 新增支持TensorTurbo算子层额外的算子：argsort、fast_exp、sort |
| V1.3.0 | 希姆计算 | 2022-03-31 | 初始对外版本。 |
ONNX算子
基于ONNX opset=9，统计出TensorTurbo目前支持的算子列表如下：
| 算子 | 说明 |
|---|---|
| Abs | |
| Add | |
| And | |
| ArgMax | |
| ArgMin | |
| AveragePool | |
| BatchNormalization | |
| Cast | |
| Clip | |
| Concat | |
| Conv | 支持group conv |
| ConvTranspose | |
| Cosh | |
| CumSum | |
| DepthToSpace | |
| Div | |
| Equal | |
| Elu | |
| Erf | |
| Exp | |
| Expand | |
| Flatten | 可在图层替换成reshape |
| GatherElements | |
| GlobalAveragePool | |
| Greater | |
| Gather | |
| Identity | 可在图层替换成npu.copy |
| LeakyRelu | |
| Less | |
| Log | |
| LogSoftmax | |
| MatMul | |
| Max | |
| MaxPool | |
| Mean | |
| Min | |
| Mul | |
| Mod | |
| Neg | |
| Nonzero | 可以在图层替换为npu.not_equal |
| Not | |
| OneHot | |
| Or | |
| Pow | 支持square（n=2）和cube（n=3） |
| PRelu | |
| Reciprocal | |
| ReduceMax | |
| ReduceMean | |
| ReduceMin | |
| ReduceSum | |
| Relu | |
| Reshape | |
| Sigmoid | |
| Sign | |
| Sinh | |
| Slice | |
| Softmax | |
| Split | |
| Sqrt | |
| Squeeze | |
| Sub | |
| Sum | |
| SpaceToDepth | |
| Softplus | |
| Scatter | |
| ScatterND | |
| Tanh | |
| Tile | |
| TopK | |
| Transpose | |
| Upsample | 可以在图层替换为npu.resize2d |
| Unsqueeze | 可以在图层替换为reshape |
| Where | |
| Xor |
TensorTurbo算子层额外支持的算子
| 算子 |
|---|
| argsort |
| broadcast_to |
| depthwise_conv2d |
| dgelu |
| full |
| fast_exp |
| greater_equal |
| image.resize2d |
| layer_norm |
| less_equal |
| List diff |
| mish |
| non_max_suppression |
| not_equal |
| opencv_magnitude |
| relu6 |
| roi_align |
| roi_pool |
| swish |
| sort |
| take |