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
TMemStore
struct TMemStore[dt: DType, cols: Int]
tcgen05.st.32x32b.x{cols}: a warp stores cols columns into its TMEM quadrant, each thread its own lane.
Parameters
- dt (
DType): A 4-byte element type. - cols (
Int): The columns one issue moves, a power of two up to 64.
Implemented traits
AnyType,
Atom,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
RegisterPassable,
StoreAtom,
TrivialRegisterPassable
comptime members
Completion
comptime Completion = TMemStoreWait
tcgen05.wait::st waits for the thread's TMEM stores.
dtype
comptime dtype = dt
The element type the instruction moves.
Engine
comptime Engine = TMemEngine
The engine of the memory the instruction writes.
mem_map
comptime mem_map = tmem_32x32b_map(cols)
Where each index writes memory.
reg_map
comptime reg_map = registers_map(cols, Int(32))
Which register each index reads.
shape
comptime shape = ParameterList()
The instruction's shape, such as (32, 16) for tcgen05.ld.32x32b.x16 or (m, n, k) for an MMA.
shape_values
comptime shape_values = Int(32), cols
One issue covers 32 lanes by cols columns.
Methods
__call__
def __call__[T: DType, LM: TensorLayout, EM: TensorEngine, LR: TensorLayout, ER: TensorEngine](self, dst: MutTileTensor[T, LM, Engine=EM, address_space=dst.address_space, linear_idx_type=dst.linear_idx_type], src: ImmTileTensor[T, LR, Engine=ER, address_space=src.address_space, linear_idx_type=src.linear_idx_type])
Issues the store without waiting for it; TMemStoreWait waits.
Parameters:
- T (
DType): The operands' element type. - LM (
TensorLayout): The TMEM view's layout. - EM (
TensorEngine): The TMEM view's engine. - LR (
TensorLayout): The register view's layout. - ER (
TensorEngine): The register view's engine.
Args:
- dst (
MutTileTensor[T, LM, Engine=EM, address_space=dst.address_space, linear_idx_type=dst.linear_idx_type]): The warp's quadrant, from the base every thread passes. - src (
ImmTileTensor[T, LR, Engine=ER, address_space=src.address_space, linear_idx_type=src.linear_idx_type]): This thread'scolsregisters.