2. NPU Operators support list(AX620E)

This section introduces the NPU supports for the ONNX operator in AX630C AX620Q.

• Supported ONNX opset_version >= 11. For detailed operator description, please refer to onnx Operators.

• Some of the supported operators do not have standard ONNX definitions yet. If such operators are included in the model, please consult technical support.

Note

"Not supported yet": Indicates that the current version of the operator implementation does not support it, but the NPU can theoretically support it, and subsequent versions may support it.

"Unlimited": Indicates that the current operator implementation can support it. Since the test may not necessarily cover the entire parameter space, if something unexpected happens, you can give us feedback and we will treat it as a BUG and fix it as soon as possible.

"Not supported": Indicates that the implementation of this attribute cannot be supported.

Operators	Attrs limitation
Abs	Unlimited
Add	Unlimited
ArgMax	- axis: Unlimited - keepdims: Unlimited - select_last_index: Only supports setting to 0
ArgMin	- axis: Unlimited - keepdims: Unlimited - select_last_index: Only supports setting to 0
AveragePool	- auto_pad: Only supports NOTSET - ceil_mode: Unlimited - count_include_pad: Only supports setting to 1 - kernel_shape: Unlimited - pads: Unlimited - strides: Unlimited
BatchNormalization	- epsilon: Unlimited - momentum: Not supported - training_mode: Not supported
Cast	to:uint8/int8/uint16/int16/uint32/int32/float32
Ceil	Unlimited
Clip	- min: Unlimited - max: Unlimited
Concat	axis: Unlimited
Constant	Unlimited
ConstantOfShape	Unlimited
Conv	- auto_pad: Only supports NOTSET - dilations: Unlimited - group: Unlimited - kernel_shape: Unlimited - pads: Unlimited - strides: Unlimited - note: The efficiency would be lower when using DepthWise/Group Conv and dilation not equals to 1
ConvTranspose	- auto_pad: Only supports NOTSET - dilations: currently only setting to 1 - group: Unlimited - kernel_shape: Unlimited - output_shape: Not supported yet - pads: Unlimited - strides: Unlimited - note: The efficiency would be lower in DepthWise ConvTranspose. - output_padding: output_padding_h <= pads_bottom, output_padding_w <= pads_right
DepthToSpace	- blocksize: Unlimited - mode: currently Only supports DCR
Div	Unlimited
Elu	Unlimited
Equal	Unlimited
Erf	Unlimited
Exp	Unlimited
Expand	Unlimited
Flatten	Unlimited
Gather	- axis: Unlimited - indices: currently Only supports 1 dimension
Gelu	Unlimited
Gemm	- alpha: Not supported yet - beta: Not supported yet - transA: Unlimited - transB: Unlimited
GlobalAveragePool	Unlimited
GlobalMaxPool	Unlimited
Greater	Unlimited
GreaterOrEqual	Unlimited
GridSample	Unlimited
HardSigmoid	Unlimited
HardSwish	Unlimited
Identity	Unlimited
InstanceNormalization	epsilon:Unlimited
LayerNormalization	axis Only supports -1 (i.e. the last dimension)
Less	Unlimited
LessOrEqual	Unlimited
LpNormalization	- axis currently Only supports -1 - (i.e. the last dimension) - p only supports 1 or 2
LSTM	- activation_alpha: Not supported yet - activation_beta: Not supported yet - activations: Not supported yet - clip: Not supported yet - hidden_size: Unlimited - input_forget: Not supported yet - layout: Only supports setting to 0 - B: Unlimited - sequence_lens: Not supported - initial_h: Unlimited - initial_c: Unlimited - P: Not supported yet - direction: Supports "bidirectional", "reverse", "forward"
LeakyRelu	Unlimited
MatMul	Unlimited
Max	Unlimited
Min	Unlimited
Mish	Unlimited
MaxPool	- auto_pad: Only supports setting to NOTSET - ceil_mode: Unlimited - dilations: Only supports 为1 - kernel_shape: Unlimited - pads: Unlimited - storage_order: Only supports setting to 0 - strides: Unlimited
Mul	Unlimited
PRelu	When 4D tensor is input, the channel dimension is in the second dimension, and slope shape currently Only supports (channel,) or (1, channel, 1, 1)
Pad	- pads: Unlimited - constant_value: Unlimited - mode: Only supports constant - axes: Not supported yet
Pow	not suppors elemwise's calculation, exponent only supports initializer form and is a scalar。
ReduceL2	- axes: Unlimited - keepdims: Unlimited - noop_with_empty_axes: This parameter is not supported yet
ReduceMax	- axes: Unlimited - keepdims: Unlimited - noop_with_empty_axes: This parameter is not supported yet
ReduceMean	- axes: Unlimited - keepdims: Unlimited - noop_with_empty_axes: This parameter is not supported yet
ReduceSum	- axes: Unlimited - keepdims: Unlimited - noop_with_empty_axes: This parameter is not supported yet
Relu	Unlimited
Reshape	shape: Unlimited
Resize	- mode: supports "nearest"、"linear" - scales: Unlimited - nearest_mode: Only supports setting to round_prefer_ceil
Sigmoid	Unlimited
Slice	- starts: Unlimited - ends: Unlimited - axes: Unlimited - steps: Unlimited
SpatialTransformer	The interpolation method is "bilinear", The boundary processing method is "constant" (value = 0)
Split	- axis: Unlimited - num_outputs: Unlimited
Sqrt	Unlimited
Silu	Unlimited
Sin	Unlimited
Swish	Unlimited
Squeeze	- axes: Unlimited
Softmax	- axis: Unlimited
Softplus	Unlimited
SpaceToDepth	blocksize: Unlimited
Sub	Unlimited
Tanh	Unlimited
Topk	Unlimited
Transpose	- perm: Unlimited
Unsqueeze	- axes: Unlimited
Where	Unlimited