SpaceMITExecutionProvider加速算子

• 本章节罗列SpaceMITExecutionProvider支持的加速算子及其在capability判定阶段的限制
• ONNX-OP描述参考
• ONNX-Contrib-OP描述参考

• SpaceMITExecutionProvider加速算子
  • Dense
    • Conv
    • ConvTranspose
    • Gemm
    • MatMul
  • QDQ
    • DynamicQuantizeMatMul
    • MatMulInteger
    • DynamicQuantizeLinear
    • QuantizeLinear
    • DequantizeLinear
  • Pool
    • AveragePool
    • GlobalAveragePool
    • MaxPool
    • GlobalMaxPool
  • Reduce
    • ReduceMean
    • ReduceMax
    • ReduceSum
    • ArgMax
    • ArgMin
  • Math
    • Add
    • Sub
    • Sum
    • Mul
    • Div
    • Pow
    • Sqrt
    • Abs
    • Neg
    • Log
    • Reciprocal
    • Sin
    • Cos
    • Tan
    • Sinh
    • Cosh
    • Floor
    • Ceil
  • Activation
    • Sigmoid
    • Swish
    • HardSigmoid
    • HardSwish
    • Tanh
    • LeakyRelu
    • Clip
    • Relu
    • PRelu
    • Elu
    • Gelu
    • Celu
    • Selu
    • Softplus
    • Softsign
    • Erf
    • Softmax
    • LogSoftmax
  • Tensor
    • Cast
    • Concat
    • Split
    • Transpose
    • Unsqueeze
    • Squeeze
    • Reshape
    • Flatten
    • Gather
    • GatherND
    • ScatterND
    • Slice
    • Resize
    • Where
    • Pad
    • Tile
    • GridSample
  • Norm
    • LayerNormalization
    • InstanceNormalization
    • BatchNormalization
    • RMSNormalization
    • GroupNormalization
  • Compare
    • Equal
    • Greater
    • GreaterOrEqual
    • Less
    • LessOrEqual
  • Transformer
    • RotaryEmbedding
    • Attention
  • Custom operator
    • YoloDecode

Dense

Conv

• Domain: ai.onnx
• Opset: 11
• Attributes: kernel_shape需存在；W需为常量，或由无上游输入边的DequantizeLinear节点提供
• Type: T：tensor(float) | tensor(float16)
• Notes: kernel_shape维度数不超过3，支持1D、2D、3D

ConvTranspose

• Domain: ai.onnx
• Opset: 11
• Attributes: kernel_shape需存在；W需为常量，或由无上游输入边的DequantizeLinear节点提供
• Type: T：tensor(float) | tensor(float16)
• Notes: kernel_shape维度数不超过2，支持1D、2D

Gemm

• Domain: ai.onnx
• Opset: 13
• Attributes: transA==0，alpha==1.0，beta==1.0
• Type: T：tensor(float) | tensor(float16)
• Notes: 支持QDQ量化格式，A非对称pertensor，B对称perchannel，当B为非常量时，B需为非对称pertensor

MatMul

• Domain: ai.onnx
• Opset: 13
• Attributes: 无额外属性限制
• Type: T：tensor(float) | tensor(float16)
• Notes: 支持QDQ量化格式，A非对称pertensor，B对称perchannel，当B为非常量时，B需为非对称pertensor

QDQ

DynamicQuantizeMatMul

• Domain: com.microsoft
• Opset: 1（参考ONNX-Contrib）
• Attributes:
• Type: T1：tensor(float)
• Type: T2：tensor(float)

MatMulInteger

• Domain: ai.onnx
• Opset: 10
• Attributes:
• Type: T1：tensor(int8) | tensor(uint8)
• Type: T2：tensor(int32)

DynamicQuantizeLinear

• Domain: ai.onnx
• Opset: 11
• Attributes:
• Type: T1：tensor(float)
• Type: T2：tensor(int8) | tensor(uint8)

QuantizeLinear

• Domain: ai.onnx
• Opset: 19
• Attributes:
• Type: T1：tensor(float)
• Type: T2：tensor(int8) | tensor(uint8)

DequantizeLinear

• Domain: ai.onnx
• Opset: 19
• Attributes:
• Type: T1：tensor(int8) | tensor(uint8) | tensor(int32)
• Type: T2：tensor(float)

Pool

AveragePool

• Domain: ai.onnx
• Opset: 22
• Attributes: 若count_include_pad!=1，则pads必须全为0；若存在kernel_shape，其维度数不超过2
• Type: T：tensor(float) | tensor(float16)

GlobalAveragePool

• Domain: ai.onnx
• Opset: 1
• Attributes: 若存在kernel_shape，其维度数不超过2
• Type: T：tensor(float) | tensor(float16)

MaxPool

• Domain: ai.onnx
• Opset: 12
• Attributes: 若存在kernel_shape，其维度数不超过2
• Type: T：tensor(float) | tensor(float16)

GlobalMaxPool

• Domain: ai.onnx
• Opset: 1
• Attributes: 若存在kernel_shape，其维度数不超过2
• Type: T：tensor(float) | tensor(float16)

Reduce

ReduceMean

• Domain: ai.onnx
• Opset: 18
• Attributes: 除第一个输入外，其余输入需为常量initializer
• Type: T：tensor(float) | tensor(float16)

ReduceMax

• Domain: ai.onnx
• Opset: 20
• Attributes: 除第一个输入外，其余输入需为常量initializer
• Type: T：tensor(float) | tensor(float16)

ReduceSum

• Domain: ai.onnx
• Opset: 13
• Attributes: 除第一个输入外，其余输入需为常量initializer
• Type: T：tensor(float) | tensor(float16)

ArgMax

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int8)
• Type: T2：tensor(int64)

ArgMin

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int8)
• Type: T2：tensor(int64)

Math

Add

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Sub

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Sum

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Mul

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Div

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Pow

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Sqrt

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Abs

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Neg

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Log

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Reciprocal

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Sin

• Domain: ai.onnx
• Opset: 7
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Cos

• Domain: ai.onnx
• Opset: 7
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Tan

• Domain: ai.onnx
• Opset: 7
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Sinh

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Cosh

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Floor

• Domain: ai.onnx
• Opset: 6
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Ceil

• Domain: ai.onnx
• Opset: 6
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Activation

Sigmoid

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Swish

• Domain: ai.onnx
• Opset: 24
• Attributes:
• Type: T：tensor(float) | tensor(float16)

HardSigmoid

• Domain: ai.onnx
• Opset: 22
• Attributes:
• Type: T：tensor(float) | tensor(float16)

HardSwish

• Domain: ai.onnx
• Opset: 22
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Tanh

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

LeakyRelu

• Domain: ai.onnx
• Opset: 16
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Clip

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Relu

• Domain: ai.onnx
• Opset: 14
• Attributes:
• Type: T：tensor(float) | tensor(float16)

PRelu

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Elu

• Domain: ai.onnx
• Opset: 22
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Gelu

• Domain: ai.onnx
• Opset: 20
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Celu

• Domain: ai.onnx
• Opset: 12
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Selu

• Domain: ai.onnx
• Opset: 6
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Softplus

• Domain: ai.onnx
• Opset: 1
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Softsign

• Domain: ai.onnx
• Opset: 1
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Erf

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Softmax

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

LogSoftmax

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Tensor

Cast

• Domain: ai.onnx
• Opset: 24
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Concat

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Split

• Domain: ai.onnx
• Opset: 18
• Attributes: 除第一个输入外，其余输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Transpose

• Domain: ai.onnx
• Opset: 24
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int8)

Unsqueeze

• Domain: ai.onnx
• Opset: 24
• Attributes: axes输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Squeeze

• Domain: ai.onnx
• Opset: 24
• Attributes: axes输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Reshape

• Domain: ai.onnx
• Opset: 24
• Attributes: shape输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Flatten

• Domain: ai.onnx
• Opset: 24
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Gather

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

GatherND

• Domain: ai.onnx
• Opset: 13
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

ScatterND

• Domain: ai.onnx
• Opset: 16
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Slice

• Domain: ai.onnx
• Opset: 13
• Attributes: 除第一个输入外，其余输入需为常量initializer；仅支持opset >= 10
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Resize

• Domain: ai.onnx
• Opset: 19
• Attributes: coordinate_transformation_mode仅支持asymmetric、half_pixel；mode仅支持nearest、linear
• Type: T：tensor(float) | tensor(float16) | tensor(int8)

Where

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

Pad

• Domain: ai.onnx
• Opset: 21
• Attributes: pads输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(int8) | tensor(uint8)

Tile

• Domain: ai.onnx
• Opset: 6
• Attributes: repeats输入需为常量initializer
• Type: T：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)

GridSample

• Domain: ai.onnx
• Opset: 16
• Attributes: mode仅支持bilinear、nearest；padding_mode仅支持zeros、border
• Type: T1：tensor(float) | tensor(float16)

Norm

LayerNormalization

• Domain: ai.onnx
• Opset: 17
• Attributes: capability阶段无额外常量限制
• Type: T：tensor(float) | tensor(float16)

InstanceNormalization

• Domain: ai.onnx
• Opset: 6
• Attributes:
• Type: T：tensor(float) | tensor(float16)

BatchNormalization

• Domain: ai.onnx
• Opset: 15
• Attributes: capability阶段无额外常量限制
• Type: T：tensor(float) | tensor(float16)

RMSNormalization

• Domain: ai.onnx
• Opset: 23
• Attributes:
• Type: T：tensor(float) | tensor(float16)

GroupNormalization

• Domain: ai.onnx
• Opset: 21
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Compare

Equal

• Domain: ai.onnx
• Opset: 11
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)
• Type: T2：tensor(uint8) | tensor(bool)

Greater

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)
• Type: T2：tensor(uint8) | tensor(bool)

GreaterOrEqual

• Domain: ai.onnx
• Opset: 12
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)
• Type: T2：tensor(uint8) | tensor(bool)

Less

• Domain: ai.onnx
• Opset: 9
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)
• Type: T2：tensor(uint8) | tensor(bool)

LessOrEqual

• Domain: ai.onnx
• Opset: 12
• Attributes:
• Type: T1：tensor(float) | tensor(float16) | tensor(int32) | tensor(uint32) | tensor(int8) | tensor(uint8) | tensor(bool)
• Type: T2：tensor(uint8) | tensor(bool)

Transformer

RotaryEmbedding

• Domain: ai.onnx
• Opset: 23
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Attention

• Domain: ai.onnx
• Opset: 23
• Attributes:
• Type: T：tensor(float) | tensor(float16)

Contrib Function operator

YoloDecode

• Domain: spacemit_functions.YoloDecode
• Opset:
• Attributes:
• Type: T：tensor(float) | tensor(float16)