Mojo struct
TileLoaderTMA
@register_passable(trivial)
struct TileLoaderTMA[tma_origin: ImmutOrigin, dtype: DType, gmem_layout: Layout, desc_layout: Layout, /, *, cta_group: Int]
TMA-based tile loader for SM100.
Wraps a TMA descriptor and multicast mask for efficient tile loading. The load method issues async_multicast_load with proper CTA group handling.
Parameters
- tma_origin (
ImmutOrigin): Origin of the TMA descriptor pointer. - dtype (
DType): Element data type. - gmem_layout (
Layout): Global memory tensor layout. - desc_layout (
Layout): TMA descriptor layout (tile dimensions). - cta_group (
Int): CTA group size (1 or 2 for SM100 2-SM MMA).
Fields
- tma_op (
TileLoaderTMA[tma_origin, dtype, gmem_layout, desc_layout, cta_group=cta_group].TmaOpPtr): - multicast_mask (
UInt16):
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDestructible,
Movable
comptime members
__copyinit__is_trivial
comptime __copyinit__is_trivial = True
__del__is_trivial
comptime __del__is_trivial = True
__moveinit__is_trivial
comptime __moveinit__is_trivial = True
TmaOp
comptime TmaOp = TMATensorTile[dtype, gmem_layout, desc_layout]
TmaOpPtr
comptime TmaOpPtr = Pointer[TileLoaderTMA[tma_origin, dtype, gmem_layout, desc_layout, cta_group=cta_group].TmaOp, tma_origin]
Methods
__init__
__init__(tma_op: Pointer[TileLoaderTMA[tma_origin, dtype, gmem_layout, desc_layout, cta_group=cta_group].TmaOp, tma_origin], multicast_mask: UInt16) -> Self
Initialize the TMA tile loader.
Args:
load
load(self, dest: LayoutTensor[dtype, layout, origin, address_space=AddressSpace.SHARED, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], ref [3] barrier: SharedMemBarrier, k_coord: UInt, row_coord: UInt)
Load a tile using TMA hardware acceleration.
Issues an async multicast load from global memory to shared memory. Coordinates are in element units (not tile units).
Args:
- dest (
LayoutTensor): Destination SMEM tile (already sliced for peer CTA if needed). - barrier (
SharedMemBarrier): Memory barrier for TMA completion signaling. - k_coord (
UInt): K dimension coordinate in global memory (elements). - row_coord (
UInt): Row coordinate (M for A, N for B) in global memory (elements).
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!