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 module

tile_tensor

TileTensor type for structured memory access with compile-time layout information.

comptime values

ImmTileTensor

comptime ImmTileTensor[dtype: DType, LayoutType: TensorLayout, origin: ImmOrigin, *, Engine: TensorEngine = DefaultEngine, address_space: AddressSpace = .GENERIC, linear_idx_type: DType = _get_index_type[LayoutType](address_space)] = TileTensor[dtype, LayoutType, origin, Engine=Engine, address_space=address_space, linear_idx_type=linear_idx_type]

Names a TileTensor over immutable storage.

TileTensor infers its mut parameter from origin, so a declaration that must accept only read-only views has to bind and constrain mut itself. This alias restricts origin to ImmOrigin instead, leaving the mutability spelled in the type. See MutTileTensor for the mutable counterpart.

Parameters

• ​dtype (DType): The data type of tensor elements (e.g., DType.float32).
• ​LayoutType (TensorLayout): A type implementing TensorLayout that defines the tensor's shape and stride structure.
• ​origin (ImmOrigin): The immutable origin of the underlying pointer.
• ​Engine (TensorEngine): A type implementing TensorEngine that supplies the storage handle and the load/store/offset operations acting on it. Defaults to DefaultEngine[element_width=1].
• ​address_space (AddressSpace): Memory address space (GENERIC, SHARED, CONSTANT, etc.). Defaults to GENERIC.
• ​linear_idx_type (DType): Integer type for memory indexing. Defaults to int32 for shared/constant memory, int64 otherwise.

LTToTTLayout

comptime LTToTTLayout[lt_layout: Layout] = Layout[TypeList[#kgen.param_list.tabulate(len[IntTuple](product_each(lt_layout.shape)), [idx: __mlir_type.index] ComptimeInt[SIMD[IntTuple](product_each(lt_layout.shape)[Int(idx)])] if (xor identical(SIMD[IntTuple](product_each(lt_layout.shape)[Int(idx)]), -1), True) else Int64)](), TypeList[#kgen.param_list.tabulate(len[IntTuple](product_each(lt_layout.stride)), [idx: __mlir_type.index] ComptimeInt[SIMD[IntTuple](product_each(lt_layout.stride)[Int(idx)])] if (xor identical(SIMD[IntTuple](product_each(lt_layout.stride)[Int(idx)]), -1), True) else Int64)]()]

Derive a TileTensor Layout from a legacy Layout.

Known dimensions become ComptimeInt, UNKNOWN_VALUE dimensions become Scalar. Hierarchical layouts (e.g. from tile_to_shape) are collapsed via product_each so each mode becomes a single value.

Parameters

• ​lt_layout (Layout): The legacy Layout to convert.

MutTileTensor

comptime MutTileTensor[dtype: DType, LayoutType: TensorLayout, origin: MutOrigin, *, Engine: TensorEngine = DefaultEngine, address_space: AddressSpace = .GENERIC, linear_idx_type: DType = _get_index_type[LayoutType](address_space)] = TileTensor[dtype, LayoutType, origin, Engine=Engine, address_space=address_space, linear_idx_type=linear_idx_type]

Names a TileTensor over mutable storage.

The mutable counterpart of ImmTileTensor: origin is restricted to MutOrigin, so the view is writable without the declaration having to bind TileTensor's inferred mut parameter.

Parameters

• ​dtype (DType): The data type of tensor elements (e.g., DType.float32).
• ​LayoutType (TensorLayout): A type implementing TensorLayout that defines the tensor's shape and stride structure.
• ​origin (MutOrigin): The mutable origin of the underlying pointer.
• ​Engine (TensorEngine): A type implementing TensorEngine that supplies the storage handle and the load/store/offset operations acting on it. Defaults to DefaultEngine[element_width=1].
• ​address_space (AddressSpace): Memory address space (GENERIC, SHARED, CONSTANT, etc.). Defaults to GENERIC.
• ​linear_idx_type (DType): Integer type for memory indexing. Defaults to int32 for shared/constant memory, int64 otherwise.

Structs

• ​NullableTileTensor: A TileTensor variant whose pointer may be absent (null).
• ​TileTensor: A tensor type with trait-based layouts supporting nested and hierarchical indexing.

Functions

• ​flatten_leading: Merge the first two dimensions of a rank-3 TileTensor: (A, B, C) -> (A*B, C).
• ​lt_to_tt: Convert a LayoutTensor to a TileTensor.
• ​lt_to_tt_idx: Like lt_to_tt but with an explicit linear_idx_type override.
• ​stack_allocation: Allocate a TileTensor on the stack with the given layout.