Struct SyncPartialTilewiseLoading

pub struct SyncPartialTilewiseLoading<T: TilingOrder> { /* private fields */ }

Each tile is guaranteed to be loaded entirely by the same plane. Each plane can load multiple tiles, provided the number of planes evenly divides the number of tiles. In this case, a plane loads contiguous tiles following the TilingOrder, until it would otherwise write to the opposite stage. At that point, it continues on the next row or column of the same stage, skipping over the memory region of the other stage.

Only supports RowMajorTilingOrder for Lhs and ColMajorTilingOrder for Rhs

Trait Implementations

impl<T: Clone + TilingOrder> Clone for SyncPartialTilewiseLoading<T>

fn clone(&self) -> SyncPartialTilewiseLoading<T>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: TilingOrder> CubeType for SyncPartialTilewiseLoading<T>

type ExpandType = SyncPartialTilewiseLoadingExpand<T>

fn into_mut(scope: &mut Scope, expand: Self::ExpandType) -> Self::ExpandType

Wrapper around the init method, necessary to type inference.

impl<TO: TilingOrder> LoadMaxRoundPlaneCount for SyncPartialTilewiseLoading<TO>

fn max_round_plane_count( tiling_scheme: &TilingScheme, ident: InputIdent, _line_size: u8, _plane_dim: u32, ) -> u32

Returns the largest number of planes that evenly divides the tasks.

impl<T: TilingOrder> LoadingValidation for SyncPartialTilewiseLoading<T>

fn check<C: GlobalConfig>( config: &C, ident: Ident, ) -> Result<(), InvalidConfigError>

Verify that configs are valid for a loader, otherwise return an error stating why

impl<TO: TilingOrder> SyncPartialLoadingStrategy for SyncPartialTilewiseLoading<TO>

type TilingLayout = ContiguousTilingLayout<TO>

The layout describing how data is tiled across the stage.

type Job<MP: MatmulPrecision> = SyncPartialTilewiseJob

The LoadingJob for this strategy.

fn new_job<MP: MatmulPrecision, G: GlobalConfig>( stage_index: u32, input_ident: InputIdent, config: G, ) -> SyncPartialTilewiseJob

Returns the job with preliminary calculations done.

fn __expand_new_job<MP: MatmulPrecision, G: GlobalConfig>( scope: &mut Scope, stage_index: u32, input_ident: InputIdent, config: G, ) -> <SyncPartialTilewiseJob as CubeType>::ExpandType

impl<T: Copy + TilingOrder> Copy for SyncPartialTilewiseLoading<T>