Module analyze 

IR pattern classification for ONNX lowering.

Analyzes an entry point’s global variables and function body to classify the computation into a known ONNX-mappable pattern.

Modules

data_type

ONNX-compatible data type constants (matches TensorProto.DataType).

Structs

ChainStep

One step of a KernelPattern::ElementWiseChain: acc = acc <op> operand.

Conv2DShape

Conv2D shape parameters extracted from uniform params.

EmbeddedWeight

Embedded constant weight data extracted from GlobalVariable initializers.

MatMulShape

Symbolic dimension names for matrix multiplication.

PoolShape

Pooling shape parameters.

ScalarBinding

A scalar the host supplies at dispatch time.

TensorBinding

A storage buffer bound as a tensor.

Enums

ActivationOp

Activation function kind.

AnalysisError

Errors during IR pattern analysis.

ChainOperand

One operand of a KernelPattern::ElementWiseChain step.

ElementWiseOp

Element-wise binary operation kind.

KernelPattern

A classified kernel pattern that maps to ONNX operators.

NormType

Normalization type.

PoolKind

Pooling operation kind.

ReduceOp

Reduction operation kind.

TensorRole

Role of a tensor in the computation.

Functions

chain_op_names

The operators a chain lowers to, in order: ["Mul", "Add"], ["Cast", "Mul", "Mul"].

chain_summary

Name a chain by the operators it lowers to, in order: Mul+Add, Cast+Mul+Mul.

classify_entry_point

Classify an entry point into a known ONNX-mappable pattern.

extract_embedded_weights

Extract constant weight data from module globals with initializers.

normalize_axis

Normalize a potentially negative axis to a positive axis.

pattern_op_names

The operator names a pattern will emit, for asking a support matrix about.