Struct Net32 

pub struct Net32 {
    pub sizes: Vec<usize>,
    pub w: Vec<Vec<i32>>,
    pub b: Vec<Vec<i32>>,
}

A network prepared for 32-bit inference.

Fields

sizes: Vec<usize>

Layer sizes (input = NFL).

w: Vec<Vec<i32>>

Clamped Q8 weights per layer (out × in).

b: Vec<Vec<i32>>

Clamped Q16 biases.

Implementations

impl Net32

pub fn new(net: &Net) -> Self

From an M1b network (its first NFL inputs).

pub fn infer(&self, x: &[i32; 37]) -> [i32; 5]

Integer forward pass.

pub fn graph(&self) -> (IntGraph, Vec<IId>)

The forward pass as a tang-expr integer graph: params are the NFL clamped inputs, outputs the NO mixture parameters. Zero weights are dropped (integer addition is associative, so no bit changes). This one graph is lowered to the GPU kernel’s MLP and evaluated as the CPU reference (Net32::program).

pub fn program(&self) -> Program

Net32::graph compiled for Net32::infer_ir.

pub fn infer_ir(&self, p: &mut Program, x: &[i32; 37]) -> [i32; 5]

Net32::infer through the integer-graph interpreter (bit-identical).

pub fn flat(&self) -> Vec<u32>

Flattened [nl, sizes…, w₀, b₀, w₁, b₁, …] for the GPU.

Trait Implementations

impl Clone for Net32

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Net32

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

Formats the value using the given formatter.