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

TileTensorIter

struct TileTensorIter[mut: Bool, //, dtype: DType, LayoutType: TensorLayout, origin: Origin[mut=mut], /, *, address_space: AddressSpace = .GENERIC, circular: Bool = False, masked: Bool = False, linear_idx_type: DType = _get_index_type[LayoutType](address_space)]

Iterates over equally shaped tiles of a buffer.

Every tile shares LayoutType; the iterator advances by stride elements from one tile to the next. A circular iterator wraps back to the start of the buffer once the offset reaches bound.

Parameters

• ​mut (Bool): Whether the iterator allows mutation of the underlying data.
• ​dtype (DType): The data type of the tensor elements.
• ​LayoutType (TensorLayout): The layout of each tile the iterator yields.
• ​origin (Origin[mut=mut]): The origin of the underlying pointer.
• ​address_space (AddressSpace): The memory address space of the buffer.
• ​circular (Bool): Whether iteration wraps around at bound.
• ​masked (Bool): Whether the tiles carry a runtime valid-row count in num_valid_rows. The iterator itself never clips the tile layout; consumers that copy a masked tile pass the count as the copy's explicit bound.
• ​linear_idx_type (DType): Signed integer type of offsets into the buffer.

Fields

• ​ptr (Pointer[Scalar[dtype], origin, address_space=address_space]): Pointer to the start of the buffer being iterated.
• ​tile_layout (LayoutType): Layout shared by every tile.
• ​offset (Scalar[linear_idx_type]): Element offset of the current tile from ptr.
• ​stride (Scalar[linear_idx_type]): Element distance between consecutive tiles.
• ​bound (Scalar[linear_idx_type]): Buffer size in elements; the wrap-around point of a circular iterator.
• ​num_valid_rows (Int): Valid rows of each tile when masked; unused otherwise.

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable, RegisterPassable, TrivialRegisterPassable

comptime members

BitcastType

comptime BitcastType[target_dtype: DType] = TileTensorIter[target_dtype, LayoutType, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]

The iterator type produced by bitcast.

Parameters

• ​target_dtype (DType): The element type the buffer is viewed as.

linear_uint_type

comptime linear_uint_type = Scalar[linear_idx_type]

The integer type of offset, stride and bound.

ReshapeType

comptime ReshapeType[NewLayout: TensorLayout] = TileTensorIter[dtype, NewLayout, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]

The iterator type produced by reshape.

Parameters

• ​NewLayout (TensorLayout): The layout of each tile of the result.

TileType

comptime TileType = TileTensor[dtype, LayoutType, origin, address_space=address_space, linear_idx_type=linear_idx_type]

The TileTensor type produced by dereferencing the iterator.

Methods

__init__

def __init__(ptr: Pointer[Scalar[dtype], origin, address_space=address_space], tile_layout: LayoutType, *, bound: Int, stride: Int, offset: Int = Int(0), num_valid_rows: Int = Int(0)) -> Self

Creates an iterator over a buffer.

Args:

• ​ptr (Pointer[Scalar[dtype], origin, address_space=address_space]): Pointer to the start of the buffer.
• ​tile_layout (LayoutType): Layout of every tile.
• ​bound (Int): Buffer size in elements; the wrap-around point when circular.
• ​stride (Int): Element distance between consecutive tiles.
• ​offset (Int): Element offset of the first tile.
• ​num_valid_rows (Int): Valid rows of each tile; only meaningful when masked.

def __init__(ptr: Pointer[Scalar[dtype], origin, address_space=address_space], tile_layout: LayoutType, *, bound: Int, num_valid_rows: Int = Int(0)) -> Self where LayoutType.all_dims_known

Creates an iterator whose tiles are laid out back to back.

Args:

• ​ptr (Pointer[Scalar[dtype], origin, address_space=address_space]): Pointer to the start of the buffer.
• ​tile_layout (LayoutType): Layout of every tile; its element count is the stride.
• ​bound (Int): Buffer size in elements; the wrap-around point when circular.
• ​num_valid_rows (Int): Valid rows of each tile; only meaningful when masked.

__getitem__

def __getitem__(self) -> Self.TileType

Returns the tile at the current position.

Returns:

Self.TileType: A TileTensor view of the current tile.

__iadd__

def __iadd__(mut self, rhs: Int)

Advances by rhs tiles.

A circular iterator wraps with a modulo, so rhs may exceed the number of tiles in the buffer.

Args:

• ​rhs (Int): The number of tiles to advance.

get

def get(self) -> Self.TileType

Returns the tile at the current position.

Returns:

Self.TileType: A TileTensor view of the current tile.

next

def next(self, rhs: Int = Int(1)) -> Self

Returns an iterator rhs tiles ahead.

Args:

• ​rhs (Int): The number of tiles to advance.

Returns:

Self: A copy of this iterator advanced by rhs tiles.

next_unsafe

def next_unsafe(self, rhs: Int = Int(1)) -> Self

Returns an iterator rhs tiles ahead, wrapping at most once.

Cheaper than next for a circular iterator: the wrap is a single conditional subtract, so rhs * stride must be less than bound.

Args:

• ​rhs (Int): The number of tiles to advance.

Returns:

Self: A copy of this iterator advanced by rhs tiles.

bitcast

def bitcast[target_dtype: DType](self) -> TileTensorIter[target_dtype, LayoutType, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]

Reinterprets the buffer as target_dtype, keeping the tile layout, offset, stride and bound.

Parameters:

• ​target_dtype (DType): The element type to view the buffer as.

Returns:

TileTensorIter[target_dtype, LayoutType, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]: An iterator over the same memory and position.

reshape

def reshape[NewLayout: TensorLayout](self, new_layout: NewLayout) -> TileTensorIter[dtype, NewLayout, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]

Views each tile under new_layout, keeping the offset, stride and bound.

Parameters:

• ​NewLayout (TensorLayout): The layout of each tile of the result.

Args:

• ​new_layout (NewLayout): The layout of each tile of the result.

Returns:

TileTensorIter[dtype, NewLayout, origin, address_space=address_space, circular=circular, masked=masked, linear_idx_type=linear_idx_type]: An iterator over the same memory and position.