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 function
tmem_16x256b
def tmem_16x256b(block: TileTensor[Engine=TMemEngine, address_space=block.address_space, linear_idx_type=block.linear_idx_type]) -> TileTensor[block.dtype, Layout[TypeList[Coord[ComptimeInt[Int(2)], ComptimeInt[Int(2)]], ComptimeInt[(block.LayoutType.static_shape[Int(1)] // Int(8))]](), TypeList[Coord[ComptimeInt[Int(128)], ComptimeInt[Int(8)]], ComptimeInt[Int(1024)]]()], block.origin, Engine=TMemEngine, address_space=block.address_space]
Returns this thread's 16x256b fragment view of a (16, N) lane block.
The view keeps the block's storage, the collective base the instruction
takes, and lays TMem16x256bLayout[N] over it. Copying it into an
N / 2-element register tile issues tcgen05.ld.16x256b, copying a
register tile into it issues tcgen05.st.16x256b, and the registers
follow the layout's flat order.
Args:
- block (
TileTensor[Engine=TMemEngine, address_space=block.address_space, linear_idx_type=block.linear_idx_type]): A(16, N)block of a warp's(32, N)TMEM tile, such aswarp_tile.tile[16, N](coord[half, 0]).
Returns:
TileTensor[block.dtype, Layout[TypeList[Coord[ComptimeInt[Int(2)], ComptimeInt[Int(2)]], ComptimeInt[(block.LayoutType.static_shape[Int(1)] // Int(8))]](), TypeList[Coord[ComptimeInt[Int(128)], ComptimeInt[Int(8)]], ComptimeInt[Int(1024)]]()], block.origin, Engine=TMemEngine, address_space=block.address_space]: The fragment view over the block's storage.