For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
SMemTileArrayWithLayout
struct SMemTileArrayWithLayout[dtype: DType, tile_layout: Layout[tile_layout.shape_types, tile_layout.stride_types], num_tiles: Int, alignment: Int = Int(128)]
Array of TileTensor tiles with explicit Layout (preserves swizzle info).
Unlike SMemTileArray2D which uses row_major internally, this type preserves the full layout information from TMA swizzling, enabling .to_layout_tensor() to produce correctly swizzled LayoutTensors.
Example: comptime swizzled_layout = tile_layout_k_majordtype, BM, BK, ... comptime MyArray = SMemTileArrayWithLayout[dtype, swizzled_layout, 4]
var array = MyArray.stack_allocation() var tile = array[0] # Returns TileTensor with swizzled layout var lt = tile.to_layout_tensor() # Correctly swizzled!
Parameters
- dtype (
DType): Tile element data type. - tile_layout (
Layout[tile_layout.shape_types, tile_layout.stride_types]): The full layout including swizzle information. - num_tiles (
Int): Number of tiles in the array. - alignment (
Int): Memory alignment (default 128 for shared memory).
Fields
- ptr (
UnsafePointer[Scalar[dtype], MutUntrackedOrigin, address_space=AddressSpace.SHARED]):
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDeletable,
Movable,
RegisterPassable,
TrivialRegisterPassable
comptime members
num_elements
comptime num_elements = (tile_layout.product() * num_tiles)
Storage
comptime Storage = InlineArray[Scalar[dtype], Int((mul tile_layout.product(), num_tiles))]
storage_size
comptime storage_size = (Int((mul tile_layout.product(), num_tiles)) * size_of[dtype]())
Tile
comptime Tile = TileTensor[dtype, Layout[shape_types, stride_types], MutAnyOrigin, address_space=AddressSpace.SHARED]
tile_size
comptime tile_size = tile_layout.product()
Methods
__init__
def __init__(ref[AddressSpace._value] storage: InlineArray[Scalar[dtype], Int((mul tile_layout.product(), num_tiles))]) -> Self
Initialize from inline storage.
Args:
- storage (
InlineArray[Scalar[dtype], Int((mul tile_layout.product(), num_tiles))]): The inline storage array.
Returns:
Self: A new SMemTileArrayWithLayout pointing to the storage.
def __init__(unsafe_ptr: UnsafePointer[Scalar[dtype], address_space=AddressSpace.SHARED]) -> Self
Initialize with a shared memory pointer.
Args:
- unsafe_ptr (
UnsafePointer[Scalar[dtype], address_space=AddressSpace.SHARED]): Pointer to shared memory storage.
__getitem__
def __getitem__[T: Intable](self, index: T) -> Self.Tile
Get tile at the given index.
Args:
- index (
T): The tile index.
Returns:
Self.Tile: A TileTensor with correct swizzled layout at the given index.
slice
def slice[length: Int](self, start: Int) -> SMemTileArrayWithLayout[dtype, tile_layout, length, alignment]
Get a slice of the array.
Parameters:
- length (
Int): The length of the slice.
Args:
- start (
Int): The starting index.
Returns:
SMemTileArrayWithLayout[dtype, tile_layout, length, alignment]: A new SMemTileArrayWithLayout representing the slice.
stack_allocation
static def stack_allocation() -> Self
Allocate the array on the stack (in shared memory).
Returns:
Self: A new SMemTileArrayWithLayout backed by stack-allocated shared memory.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!