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 function

reduce

def reduce[params: ContextParams, //, TileFn: def[ws: Int, _r: Int](IndexList[_r]) -> None & ImplicitlyCopyable](row_coords: Coord, axis_size: Int, ctx: Context[params], tile_fn: TileFn)

Drives tile_fn over the reduce axis, CPU-side, with no monoid state — pure per-tile iteration for map/emit terminals (see the state-carrying overload below for reduce phases).

Tiers mirror the state-carrying overload; see there for the per-tier walk description.

Parameters:

• ​params (ContextParams): Comptime dispatch parameters.
• ​TileFn (def[ws: Int, _r: Int](IndexList[_r]) -> None & ImplicitlyCopyable): The value-closure type of tile_fn.

Args:

• ​row_coords (Coord): The current row's coords.
• ​axis_size (Int): Length of the reduce axis.
• ​ctx (Context[params]): The dispatch bundle (unused on CPU beyond comptime params reads).
• ​tile_fn (TileFn): Per-tile callback; closes over input/output closures.

def reduce[State: ReduceOp, params: ContextParams, //, TileFn: def[ws: Int, _r: Int](mut State, IndexList[_r]) -> None & ImplicitlyCopyable](row_coords: Coord, axis_size: Int, ctx: Context[params], mut state: State, tile_fn: TileFn) where (eq TileFn.State, State)

Drives tile_fn over the reduce axis, CPU-side.

Two tiers, picked from params.emit_tile_width:

• emit_tile_width == 1: walk axis in simd_width chunks, 8-way unrolled (matches legacy _reduce_along_inner_dimension), then SIMD and scalar tails. tile_fn[ws] is called with ws ∈ {simd_width, 1}; the monoid's width-polymorphic reduce[w] collapses each tile.
• emit_tile_width > 1: scalar axis walk, tile_fn[W] per step. The body holds W parallel accumulators (the monoid M at width W); each step issues one SIMD load on the innermost-non-axis dim, spreading lanes across the accumulators.

tile_fn is a value closure (its copy-captured state — input closures, side-loaded tensors — rides the value); it takes the caller's monoid state as a mut argument rather than capturing it, since a captured accumulator can't be mutated through a value closure (captures are immutable) but an inout argument can.

Parameters:

• ​State (ReduceOp): The monoid type being accumulated.
• ​params (ContextParams): Comptime dispatch parameters.
• ​TileFn (def[ws: Int, _r: Int](mut State, IndexList[_r]) -> None & ImplicitlyCopyable): The value-closure type of tile_fn.

Args:

• ​row_coords (Coord): The current row's coords.
• ​axis_size (Int): Length of the reduce axis.
• ​ctx (Context[params]): The dispatch bundle (unused on CPU beyond comptime params reads).
• ​state (State): The caller's monoid accumulator, threaded through by mut reference and mutated in place by tile_fn.
• ​tile_fn (TileFn): Per-tile callback; closes over input closures and folds each tile into state.