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

rowwise

GPU row-wise reduction scaffolder.

The author writes one body[ctx: Context](row_coords) callable, taking a comptime Context (the dispatch tier) and the row's coords. The body composes a few helpers from this module:

• reduce — tier-aware iteration over the reduce axis, then a cross-thread join of the body's accumulators. Takes a per-tile callback tile_fn[ws, _r] (which closes over input closures + local monoid states) and a variadic of ReduceOp states to pjoin after the loop. Zero states runs iteration only — used by 2-pass algorithms' second pass.
• once — run a closure exactly once per (logical) output row. Picks the canonical writer thread; the body never sees lane_id() / thread_idx.x.
• simd — width-polymorphic SIMD constructor from a per-lane callback; degenerates cleanly at w = 1.

launch is the top-level entry point. It picks the tier from shape and axis, then instantiates the matching kernel template.

No GPU primitives leak into the body: it sees ctx, lambdas, and the helpers above.

Structs

• ​BlockReducer: Reduces a scalar across BLOCK_SIZE threads in a block. Broadcasts the result to every thread.
• ​WarpReducer: Reduces a scalar across all lanes in a warp. Broadcasts the result to every lane.

Functions

• ​launch: Top-level scaffolder. Picks the tier from shape + axis, instantiates the matching kernel, and launches.
• ​once: Runs emit exactly once per (logical) output row.
• ​pjoin: Cross-thread join for a single monoid state, tier-appropriately.
• ​reduce: Drives the tier-aware iteration over the reduce axis, with no monoid state — pure per-tile iteration for map/emit terminals (see the state-carrying overload below for reduce phases).