Module tile 

Tile module.

Modules

allocate_bounce_tile

allocate_unit_tile

allocate_whitebox_fragment

cmma_allocate_acc

cmma_allocate_lhs

cmma_allocate_rhs

interleaved_allocate_acc

interleaved_allocate_lhs

interleaved_allocate_rhs

mask_layout_absolute_pos

mask_layout_num_units_per_row

matmul

Per-flavor tile matmul compute: *_execute, *_load_*, *_write_to_shared, plus the fragment readers/writers for each flavor. Tile data and matmul configs live alongside the corresponding data structures in [crate::tile::data].

mma_allocate_acc

mma_allocate_lhs

mma_allocate_rhs

mma_register_vector_sizes

planevec_allocate_acc

planevec_allocate_lhs

planevec_allocate_rhs

register_allocate_acc

register_allocate_lhs

register_allocate_rhs

rowwise

Row-wise tile compute. The Tile<E, Plane, ReadWrite> impl in [dispatch] routes per-variant row_max, row_sum, exp_diff, and rowwise_scale to the right backend; [reducer] handles the cross-unit plane reduction used by the WhiteboxFragment and BounceTile arms.

softmax_init_state

whitebox_fragment_absolute_pos

whitebox_fragment_zero_slice

Structs

BounceConfig

Comptime configuration for BounceTile.

BounceTile

BounceTileExpand

CmmaMatmul

CmmaTile

CmmaTileExpand

Cube

Filled

Tile is a single value that gets filled in everywhere

FilledExpand

InterleavedMatmul

InterleavedTile

InterleavedTileExpand

MmaIOConfig

MmaMatmul

MmaTile

Single MMA tile carrier. The role (Lhs / Rhs / Acc) lives inside MmaFragment because each role’s fragment uses a different inner vector size (NL / NR / NA); the outer carrier holds the shared comptime metadata.

MmaTileExpand

Plane

PlaneVecMatInnerProduct

PlaneVecTile

PlaneVecTileExpand

RegisterMatmul

RegisterTile

RegisterTileExpand

RowWise

Contains one value per row of a fragment for which the unit contributes

RowWiseExpand

ScopeMarker

Zero-sized comptime marker used to carry a Scope generic through [Tile].

ScopeMarkerExpand

SharedTile

V-erased view over a shared-memory tile. Wraps a StridedTile and hides its vectorization from the type system. The underlying slice is downcast to a scalar Slice<E, IO> only at the Rust type level — the runtime vector_size on the cubecl slice is preserved, so projecting back via view::<V>() is a pure retype with no metadata change. V must match the original V the tile was wrapped with.

SharedTileExpand

Strided

Tile is a slice of memory with a stride

StridedExpand

StridedTile

Tile with a linear major dimension, and a strided minor dimension. Basic tile kind supported by all stage matmuls.

StridedTileExpand

Unit

UnitTile

UnitTileExpand

UnitTileLayout

Value

Wrapper over val to make enum work

ValueExpand

WhiteboxFragment

Assumes:

WhiteboxFragmentExpand

WhiteboxFragmentLayout

__NA

__NL

__NPlaneVec

__NR

Enums

InnerLayout

LoadMethod

MaskLayout

Layout of an attention-style mask fragment across the units of a plane. Purely comptime — all variants carry only comptime data.

MmaFragment

MmaFragmentExpand

ProductType

Execution mode for the RegisterMatmul

ScopeKind

SoftmaxKind

Comptime descriptor for the row-shape used by online softmax. Determines how many rows per unit each running-state vector holds.

StoreMethod

Tile

TileExpand

Constants

FULLY_MASKED_ROW_THRESHOLD

Any value smaller than this is considered numerically zero (used for fully-masked rows or tiny contributions). Value chosen to be above f16 smallest normal (~6.1e-5).

LOGIT_MASKED

Logits below this are considered masked (effectively -inf). Value chosen to fit within f16 range (~-65,504 max).

Traits

Mask

Minimal mask abstraction used by row-wise tile operations. Returns true when the element at local_pos should be treated as masked (i.e. driven to -inf by Tile::scale_and_mask).

MaskExpand

Minimal mask abstraction used by row-wise tile operations. Returns true when the element at local_pos should be treated as masked (i.e. driven to -inf by Tile::scale_and_mask).

TileKind

Kind (family) of the tiles returned by a stage and ingested by a tile matmul reader

TileScope

Identifies which compute primitive executes a tile matmul.

Functions

allocate_bounce_tile

Wraps a freshly built CmmaTile in a Tile::Bounce. Panics at expansion time unless Sc = Plane.

allocate_unit_tile

Allocates a Tile::Unit. The variant is valid in any scope — each unit just holds its own row-major copy of the tile.

allocate_whitebox_fragment

Allocates a Tile::WhiteboxFragment for the given scope. Panics at expansion time unless Sc = Plane.

assert_plane_scope

Comptime assertion that a tile-scope generic resolves to Plane.

assert_unit_scope

Comptime assertion that a tile-scope generic resolves to Unit.

cmma_allocate_acc

cmma_allocate_lhs

cmma_allocate_rhs

interleaved_allocate_acc

interleaved_allocate_lhs

interleaved_allocate_rhs

mask_layout_absolute_pos

Maps a per-unit (row, col) to its absolute position within the tile.

mask_layout_num_units_per_row

Returns how many units in a plane participate in the same row.

mma_allocate_acc

mma_allocate_lhs

mma_allocate_rhs

mma_register_vector_sizes

planevec_allocate_acc

planevec_allocate_lhs

planevec_allocate_rhs

register_allocate_acc

register_allocate_lhs

register_allocate_rhs

softmax_init_state

Initial running state (m, l) for the online softmax over a single tile row.

whitebox_fragment_absolute_pos

Maps a per-unit (row, col) to its absolute position within the tile described by layout.

whitebox_fragment_zero_slice

Zeroes a slice giving responsibility to units following layout.

Type Aliases

NA

NL

NPlaneVec

NR