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

RowCache

struct RowCache[params: ContextParams, T: DType, dtype: DType, axis: Int, rank: Int, NCH: Int, W: Int, cols: Int, fuse: Bool, shared: Bool]

A row value cached once by Row.cache and read by later phases. See the module note above for the two backings.

Carries no closures of its own: recompute / load take the owning view's input_fn / compute as fresh value-closure ARGUMENTS (a "forwarding bridge") rather than storing them as typed fields. A field whose type is a trait-bound closure referencing a sibling struct parameter (dtype/T) can't be named generically from another struct's method signature in current Mojo (over: RowCache[..., InputFn, Compute] fails to elaborate with "lacking evidence to prove correctness" — a struct-level, not method-level, limitation); passing the closures as per-call arguments instead keeps RowCache's own type free of them entirely.

Parameters

• ​params (ContextParams): The owning view's comptime tier parameters.
• ​T (DType): The cached value's dtype.
• ​dtype (DType): The primary input dtype (for the recompute fallback).
• ​axis (Int): Reduced axis.
• ​rank (Int): Tensor rank.
• ​NCH (Int): Per-participant chunk count (mirrors Row._NCH).
• ​W (Int): SIMD width (mirrors Row._W).
• ​cols (Int): Reduced-axis length (mirrors Row._cols).
• ​fuse (Bool): Whether the owning view is on a cache-eligible tier.
• ​shared (Bool): Whether the row is also published to shared memory.

Fields

• ​row_il (IndexList[rank]): The row's coords (reduced axis pinned to its base).

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable

Methods

__init__

def __init__(out self, *, copy: Self)

Explicit copy constructor: InlineArray lost ImplicitlyCopyable conformance, so _owned can't be auto-derived and needs .copy().

Args:

• ​copy (Self): The handle to copy.

def __init__(out self, row_il: IndexList[rank])

Builds an empty handle; cache fills the backing.

Args:

• ​row_il (IndexList[rank]): The row's coords.

recompute

def recompute[w: Int, InputFn: def[width: Int, alignment: Int, coord_rank: Int](IndexList[coord_rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, Compute: def[w: Int](SIMD[dtype, w], IndexList[rank]) -> SIMD[T, w] & RegisterPassable & ImplicitlyCopyable](self, coord: IndexList[rank], input_fn: InputFn, compute: Compute) -> SIMD[T, w] where (eq InputFn.dtype, dtype) where (eq Compute.dtype, dtype) where (eq Compute.rank, rank) where (eq Compute.T, T)

Recomputes the value at coord from a fresh global load — the fallback when the row was not staged.

w is given explicitly (recompute[w](...)); InputFn / Compute are inferred from input_fn / compute.

Parameters:

• ​w (Int): SIMD width.
• ​InputFn (def[width: Int, alignment: Int, coord_rank: Int](IndexList[coord_rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable): The value-closure type of input_fn.
• ​Compute (def[w: Int](SIMD[dtype, w], IndexList[rank]) -> SIMD[T, w] & RegisterPassable & ImplicitlyCopyable): The value-closure type of compute.

Args:

• ​coord (IndexList[rank]): The element coord.
• ​input_fn (InputFn): The owning view's primary-input loader, as a value.
• ​compute (Compute): The per-element producer, as a value.

Returns:

SIMD[T, w]: compute(input_fn(coord), coord).

load

def load[w: Int, InputFn: def[width: Int, alignment: Int, coord_rank: Int](IndexList[coord_rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable, Compute: def[w: Int](SIMD[dtype, w], IndexList[rank]) -> SIMD[T, w] & RegisterPassable & ImplicitlyCopyable](self, coord: IndexList[rank], input_fn: InputFn, compute: Compute) -> SIMD[T, w] where (eq InputFn.dtype, dtype) where (eq Compute.dtype, dtype) where (eq Compute.rank, rank) where (eq Compute.T, T)

Reads the cached value at any column coord — from the per-row shmem strip when staged (cross-participant), else the recompute fallback.

w is given explicitly (load[w](...)); InputFn / Compute are inferred from input_fn / compute.

Parameters:

• ​w (Int): SIMD width.
• ​InputFn (def[width: Int, alignment: Int, coord_rank: Int](IndexList[coord_rank]) -> SIMD[dtype, width] & RegisterPassable & ImplicitlyCopyable): The value-closure type of input_fn.
• ​Compute (def[w: Int](SIMD[dtype, w], IndexList[rank]) -> SIMD[T, w] & RegisterPassable & ImplicitlyCopyable): The value-closure type of compute.

Args:

• ​coord (IndexList[rank]): The element coord (any column of this row).
• ​input_fn (InputFn): The owning view's primary-input loader, as a value (fallback path only).
• ​compute (Compute): The per-element producer, as a value (fallback path only).

Returns:

SIMD[T, w]: The cached value at coord.