Struct UnitTiling 

pub struct UnitTiling { /* private fields */ }

A forced register tiling, expressed as the tiles each unit’s partition holds along (M, N, K).

The tile (4, 4, 4) and stage (16, 8, 1) are held constant, so the stage a preset asks for is (64*m) x (32*n) x (4*k) output/reduction elements. A device whose cube holds fewer than 128 units gets the largest power-of-two stage it holds instead. A preset that overflows shared memory is trimmed by UnitSmemBudget, the same budget the matmul unit selector fits its tilings to.

Implementations

impl UnitTiling

pub const NARROW: Self

128 x 32 x 32: few output channels, as in a 32-wide dense layer.

pub const BALANCED: Self

128 x 64 x 32: a moderate output width.

pub const WIDE: Self

128 x 128 x 32: wide outputs and the wide weight-gradient matrices a large channel count produces (taps * C_in), e.g. 128–512 channels.

pub const DEEP: Self

128 x 64 x 64: a long reduction, as a large C_in * taps contraction gives on a short feature map (the short-spatial/long-channel regime).

pub const TALL: Self

256 x 64 x 32: many output rows, as a large spatial map times batch gives in forward and data-gradient passes.

Trait Implementations

impl Clone for UnitTiling

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for UnitTiling

impl Debug for UnitTiling

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

Formats the value using the given formatter.

impl Eq for UnitTiling

impl PartialEq for UnitTiling

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for UnitTiling