Struct TileScorer 

pub struct TileScorer {
    pub tile: usize,
    pub active: usize,
}

A reusable tiled scorer over a fixed decoder. Holds the tile width so the router can score a whole minibatch with one object, never materialising an N×K block.

Fields

tile: usize

active: usize

Implementations

impl TileScorer

pub fn new(active: usize, tile: usize) -> Self

pub fn route_row( &self, row: ArrayView1<'_, f32>, decoder: ArrayView2<'_, f32>, ) -> Vec<(u32, f32)>

Top-active atoms for row against decoder.

pub fn route_minibatch( &self, rows: ArrayView2<'_, f32>, decoder: ArrayView2<'_, f32>, ) -> Vec<Vec<(u32, f32)>>

Top-active atoms for every row of a minibatch rows (B × P) against decoder (K × P), scored a column tile at a time via a batched GEMM.

This is the implementation that delivers the module’s promise: the score block formed at any instant is B × tile (peak rows × tile, independent of K), and each tile is a single (B × P)·(P × tile) matrix multiply rather than B × tile scalar dot loops. The online top-s selector sees the atoms in the same global order as Self::route_row (tile 0 first, ascending atom index). The GEMM contracts the same P terms but matrixmultiply may accumulate them in a blocked order, so the per-atom scores agree with the row-at-a-time path only to f32 rounding; where two atoms tie within that rounding the two paths may select different members of the tie (interchangeable for the reconstruction, which is why the fit stays minibatch-invariant rather than bit-identical). Returns one (atom, score) shortlist per row, in row order.

Trait Implementations

impl Clone for TileScorer

fn clone(&self) -> TileScorer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for TileScorer

impl Debug for TileScorer

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

Formats the value using the given formatter.