Struct ShapeInference 

pub struct ShapeInference;

Implementations

impl ShapeInference

pub fn infer_output_shape( op: &TensorOp, inputs: &[&TensorTypeInfo], attrs: Option<&Attributes>, ) -> Result<Vec<TensorTypeInfo>, String>

pub fn compute_flops( op: &TensorOp, inputs: &[&TensorTypeInfo], attrs: Option<&Attributes>, ) -> Option<u64>

pub fn compute_memory_bytes( op: &TensorOp, inputs: &[&TensorTypeInfo], attrs: Option<&Attributes>, ) -> Option<u64>

Sums every input’s bytes plus the output’s bytes — memory traffic includes writing the result, not just reading the operands. Every op family used to only get this for MatMul/SparseMatMul; every other op (Conv*, Attention, Norm, pooling, activations, …) silently omitted the output entirely, understating real traffic (e.g. by ~46% for a typical Conv2D).

Trait Implementations

impl Debug for ShapeInference

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.