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
TileLoaderTMA
struct TileLoaderTMA[tma_origin: ImmOrigin, dtype: DType, tma_rank: Int, tile_shape: IndexList[tma_rank], desc_shape: IndexList[tma_rank], /, *, BK: Int, cluster_size: Int32, use_partitioned_multicast: Bool]
TMA-based tile loader for hardware-accelerated memory transfers.
This loader uses NVIDIA's Tensor Memory Accelerator (TMA) for efficient 2D tile transfers from global to shared memory, with optional multicast support for multi-block clusters.
Parameters
- tma_origin (
ImmOrigin): Origin type for the TMA operation. - dtype (
DType): Data type of the elements being loaded. - tma_rank (
Int): Rank of the TMA tile (number of dimensions). - tile_shape (
IndexList[tma_rank]): Shape of the complete tile in shared memory. - desc_shape (
IndexList[tma_rank]): Shape described by the TMA descriptor (may be smaller). - BK (
Int): Block size in the K dimension (for coordinate conversion). - cluster_size (
Int32): Number of blocks in the cluster (1 for no clustering). - use_partitioned_multicast (
Bool): Whether to use partitioned multicast loading.
Fields
- tma_op (
TileLoaderTMA[tma_origin, dtype, tma_rank, tile_shape, desc_shape, BK=BK, cluster_size=cluster_size, use_partitioned_multicast=use_partitioned_multicast].TMATensorTilePtr): - rank (
Int): - multicast_mask (
UInt16):
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
RegisterPassable,
TileLoader,
TrivialRegisterPassable
comptime members
TMATensorTilePtr
comptime TMATensorTilePtr = Pointer[TMATensorTile[dtype, tma_rank, tile_shape, desc_shape], tma_origin]
Methods
__init__
def __init__(tma_op: Pointer[TMATensorTile[dtype, tma_rank, tile_shape, desc_shape], tma_origin], rank: Int, multicast_mask: UInt16) -> Self
Initialize the TMA tile loader.
Args:
- tma_op (
Pointer[TMATensorTile[dtype, tma_rank, tile_shape, desc_shape], tma_origin]): Pointer to the TMA tensor descriptor. - rank (
Int): Rank of this block within the cluster. - multicast_mask (
UInt16): Bit mask for multicast targets.
load_tile
def load_tile(self, dst: TileTensor[address_space=AddressSpace.SHARED, linear_idx_type=dst.linear_idx_type], mem_barrier: Pointer[SharedMemBarrier, address_space=AddressSpace.SHARED], _coords: Tuple[Int, Int])
Load a tile using TMA hardware acceleration.
Converts tile indices to element coordinates and initiates a TMA transfer. For clusters, uses multicast to share data across blocks.
Note: Coordinates are converted from (row, col) tile indices to (k_elements, row/col_elements) for TMA's K-major ordering.
Args:
- dst (
TileTensor[address_space=AddressSpace.SHARED, linear_idx_type=dst.linear_idx_type]): Destination tile in shared memory. - mem_barrier (
Pointer[SharedMemBarrier, address_space=AddressSpace.SHARED]): Memory barrier for synchronization. - _coords (
Tuple[Int, Int]): Tile coordinates (row_tile_idx, col_tile_idx).