Skip to content

Index

Operators API

The 'operators' module provides classes to represent operators in the network graph. These classes all inherit from the AotOperator class, which provides a common interface for all operators. Each operator class is responsible for generating the C code for its specific operation. Built-in classes are registered into RegistryContext.aot_operator_classes via register_default_aot_operators.

Available Operators

  • Add: Element-wise addition
  • ArgMax: Argmax reduction
  • ArgMin: Argmin reduction
  • AssignVariable: Assigns a value to a variable
  • AveragePool: Average pooling operation
  • BatchMatMul: Batch matrix multiplication
  • BatchToSpaceND: Batch to space transformation
  • Comparison: Comparison operations (e.g., equal, not equal, less than)
  • Concatenation: Concatenates tensors along a specified axis
  • Conv: 2D convolution operation
  • DepthwiseConv: Depthwise separable convolution operation
  • DepthToSpace: Rearranges data from depth into blocks of spatial data
  • Dequantize: Dequantizes a tensor
  • EthosU: Ethos-U NPU operator
  • ExpandDims: Expands the shape of a tensor
  • Fill: Fills a tensor with a specified value
  • Gather: Gathers elements along an axis
  • GatherNd: Gathers slices using multi-dimensional indices
  • FullyConnected: Fully connected layer
  • HardSwish: Hard swish activation function
  • LeakyRelu: Leaky ReLU activation function
  • Logistic: Logistic activation function
  • MaxPool: Max pooling operation
  • Maximum: Element-wise maximum
  • Mean: Computes the mean of a tensor along specified axes
  • Minimum: Element-wise minimum
  • Mul: Element-wise multiplication
  • Pack: Packs a list of tensors into a single tensor
  • Pad: Pads a tensor
  • Quantize: Quantizes a tensor
  • ReadVariable: Reads a variable
  • ReduceMax: Reduces a tensor by taking the maximum along specified axes
  • ReduceMin: Reduces a tensor by taking the minimum along specified axes
  • Relu: Rectified Linear Unit activation
  • Reshape: Reshapes a tensor
  • Shape: Returns the shape of a tensor
  • Softmax: Softmax activation
  • SpaceToBatchND: Space to batch transformation
  • SpaceToDepth: Rearranges data from spatial blocks into depth
  • Split: Splits a tensor into multiple tensors along a specified axis
  • Squeeze: Removes dimensions of size 1 from the shape of a tensor
  • StridedSlice: Slices a tensor with strides
  • Sub: Element-wise subtraction
  • Svdf: SVDF operation
  • Tanh: Hyperbolic tangent activation function
  • TransposeConv: Transpose convolution operation
  • Transpose: Transposes a tensor
  • Unpack: Unpacks a tensor into multiple tensors along a specified axis
  • ZerosLike: Generates a tensor of zeros

Copyright 2025 Ambiq. All Rights Reserved.

Classes

Functions

register_default_aot_operators

register_default_aot_operators(registry: Registry[AirOpType | str, type[AotOperator]]) -> None

Register all built-in AOT operators into the provided registry.