For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

SourceTileStorage

struct SourceTileStorage[src_type: DType, src_shape: IndexList[Int(2)], num_epi_load_stages: Int]

Storage for source tensor C tiles (residual/skip connection input).

Used by the epilogue load warp to pre-fetch source tensor C via TMA, enabling overlap with MMA computation for residual operations like D = Conv(A,B) + betaC or D = MatMul(A,B) + betaC.

Parameters

• ​src_type (DType): Data type for source tiles (same as output type).
• ​src_shape (IndexList[Int(2)]): Tile shape for source tiles as (OutputM, OutputN).
• ​num_epi_load_stages (Int): Number of epilogue load pipeline stages.

Fields

• ​src_tiles_storage (SourceTileStorage[src_type, src_shape, num_epi_load_stages].SrcTileArray.Storage):

Implemented traits

AnyType, Deinitable, Movable

comptime members

src_tile_layout

comptime src_tile_layout = Layout.row_major(src_shape[Int(0)], src_shape[Int(1)])

SrcTileArray

comptime SrcTileArray = SMemTileArray2DRowMajor[src_type, src_shape[Int(0)], src_shape[Int(1)], num_epi_load_stages]

Methods

src_tiles

def src_tiles(ref[AddressSpace._value] self) -> Self.SrcTileArray

Get source tile array accessor (TileTensor-based).

Returns:

Self.SrcTileArray