Struct StridedTile 

pub struct StridedTile {
    pub shape: Tile2,
    pub strides: Strides2,
}

Strided tile: each row stride is configurable. Lets a kernel read a non-contiguous view (e.g. last-axis Narrow output) with the same TileIO interface as a contiguous tile.

Fields

shape: Tile2

strides: Strides2

Trait Implementations

impl Clone for StridedTile

fn clone(&self) -> StridedTile

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for StridedTile

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

Formats the value using the given formatter.

impl TileIO for StridedTile

unsafe fn address(&self, base: *const f32, c: Coord2) -> *const f32

Compute the byte address for a coordinate. Used by load / store / prefetch so impls only need to define the address arithmetic once. SAFETY: caller checks bounds.

unsafe fn load(&self, base: *const f32, c: Coord2) -> f32

Load a tile element by (row, col). SAFETY: caller ensures the address is valid for read.

unsafe fn store(&self, base: *mut f32, c: Coord2, v: f32)

Store an element by (row, col). SAFETY: caller ensures the address is valid for write.

unsafe fn prefetch(&self, base: *const f32, c: Coord2)

Hint to the prefetcher. On aarch64 issues a single prfm pldl1keep (load into L1, retain). Elsewhere a no-op. SAFETY: caller ensures the address is in a valid mapping.

impl Copy for StridedTile