Hardware Supported Operators
Table below shows the list of operators supports base on ONNX operators.
The operators NPU supports
| Node | 520 | 720 | 530 | 630 |
|---|---|---|---|---|
| Add | Y | Y | Y | Y |
| AveragePool1 | Y | Y | Y | Y |
| BatchNormalization | Y | Y | Y | Y |
| BitShift | N | N | N | N |
| Cast | N | N | N | N |
| Clip2 | Y | Y | Y | Y |
| Concat | Y | Y | Y | Y |
| Conv | Y | Y | Y | Y |
| ConvTranspose3 | N | Y | Y | Y |
| DepthToSpace4 | N | N | Y | Y |
| Div | N | N | N | N |
| Dropout | Y | Y | Y | Y |
| Elu | N | N | Y | Y |
| Erf | N | N | Y | Y |
| Exp | N | N | Y | Y |
| Expand5 | N | N | Y | Y |
| Flatten6 | Y | Y | Y | Y |
| Floor | N | N | N | N |
| Gather | N | N | N | N |
| GatherElements | N | N | N | N |
| GatherND | N | N | N | N |
| Gemm | Y | Y | Y | Y |
| GlobalAveragePool7 | Y | Y | Y | Y |
| GlobalLpPool | N | N | N | N |
| GlobalMaxPool | Y | Y | Y | Y |
| GRU | N | Y | Y | Y |
| Hardmax | N | N | N | N |
| HardSigmoid | Y | Y | Y | Y |
| InstanceNormalization | N | N | N | N |
| LeakyRelu | Y | Y | Y | Y |
| LpNormalization | N | N | N | N |
| LRN | N | N | N | N |
| LSTM | N | Y | Y | Y |
| MatMul | N | Y | Y | Y |
| MaxPool8 | Y | Y | Y | Y |
| MaxRoiPool | N | Y | Y | Y |
| MaxUnpool | N | N | N | N |
| Mean | Y | Y | Y | Y |
| Min | N | N | N | N |
| Mod | N | N | N | N |
| Mul | N | Y | Y | Y |
| Multinomial | N | N | N | N |
| Neg | Y | Y | Y | Y |
| NonMaxSuppression | N | N | N | N |
| NonZero | N | N | N | N |
| Not | N | N | N | N |
| OneHot | N | N | N | N |
| Or | N | N | N | N |
| Pad9 | Y | Y | Y | Y |
| Pow10 | N | Y | Y | Y |
| PRelu | Y | Y | Y | Y |
| RandomUniformLike | N | N | N | N |
| ReduceLogSum | N | N | N | N |
| ReduceLogSumExp | N | N | Y | Y |
| ReduceMax | N | N | N | N |
| ReduceMean11 | Y | Y | Y | Y |
| ReduceSum | Y | Y | Y | Y |
| ReduceSumSquare | N | Y | Y | Y |
| Relu | Y | Y | Y | Y |
| Reshape | N | Y | Y | Y |
| Resize12 | N | Y | Y | Y |
| RNN | N | Y | Y | Y |
| RoiAlign | N | Y | Y | Y |
| Selu | N | N | N | N |
| Shrink | N | N | N | N |
| Sigmoid | N | Y | Y | Y |
| Slice | N | Y | Y | Y |
| Softmax | N | N | Y | Y |
| Softplus | N | N | Y | Y |
| Softsign | N | N | Y | Y |
| SpaceToDepth13 | N | N | Y | Y |
| Split | Y | Y | Y | Y |
| Squeeze | N | N | N | N |
| Sub | Y | Y | Y | Y |
| Sum | Y | Y | Y | Y |
| Tanh | N | Y | Y | Y |
| ThresholdedRelu | N | N | N | N |
| Tile | N | N | N | N |
| Transpose | N | Y | Y | Y |
| Unsqueeze | N | N | N | N |
| Upsample14 | N | Y | Y | Y |
Notes:
- For AveragePool kernel size, 520 and 720 support square kernel up to 3x3. while 530 and 630 also support non-square kernel up to 3x3.
- All hardware only support Clip with min set to 0.
- 720 only supports ConvTranspose with stride set to 2.
- 530 and 630 only support DepthToSpace with blocksize set to 2 or 4.
- 530 and 630 only support Expand on channel or column and row.
- 520 and 720 only support Flatten before Gemm.
- 520 only supports GlobalAveragePool on input within 256 pixels while other hardware support up 16384 pixels.
- For MaxPool kernel size, 520 and 720 support square kernel up to 3x3. while 530 and 630 also support non-square kernel up to 3x3.
- NPUs only support constant pad mode and constant value set to 0.
- NPUs only support power set to 2.
- NPUs only support ReduceMean nodes that behave the same as GlobalAveragePool. And it has the same limitation, too.
- 720 only supports Resize nodes which work as upsampleing. Same limitation as the Upsample.
- 530 and 630 only support SpaceToDepth with blocksize set to 2 or 4.
- 720 only supports Upsample with mode set to bilinear or nearest.