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_types

Target-neutral ContextParams + Context types for the row-wise reduction scaffolder. In its own module so the CPU and GPU backends can both import it without an import cycle through the unified rowwise dispatcher.

ContextParams.target (comptime StaticString) is the dispatcher's switch: "cpu" selects the CPU backend, anything else the GPU backend. The GPU-only runtime fields on Context (split-K plumbing) hold sentinels on CPU and are never read — every access is behind a comptime if is_cpu[params.target]: branch that dead-code-eliminates.

Member-naming convention: members a body reads have plain names (axis, emit_tile_width, simd_width, BLOCK_SIZE, target, accumulator_width); scaffolder-internal members — the tier discriminators and split-K plumbing — are _-prefixed.

comptime values

RowBody

comptime RowBody = def[_p: ContextParams](Coord[**?], mut Context[_p]) -> None & RegisterPassable & ImplicitlyCopyable

Bound for the per-row body threaded through the launch as a VALUE.

The body is a value closure (not a comptime capturing parameter): the GPU backend crosses it into the kernel as a struct field, so its copy-captured state (the fused input_fn / output_fn value closures, axis_size) is serialized across the launch. ImplicitlyCopyable & RegisterPassable lets the kernel struct hold it as a register-passable field; the callable trait is non-capturing (captures ride the value).

Structs

• ​Context: Per-block dispatch bundle. Comptime tier identity and widths flow through params; runtime per-block plumbing (split-K scratch + position) sits in var fields. The body forwards ctx to rowwise.reduce / rowwise.once and treats it as opaque, reading only the plainly-named members (ctx.axis, ctx.accumulator_width, ...). The _-prefixed members (ctx._tier, ctx._partials_base, ...) are scaffolder internals.
• ​ContextParams: Comptime half of Context. Bundled into one template parameter so callers write Context[params] instead of threading eight separate comptime args.
• ​ReduceTier: GPU tier discriminator for the row-wise scaffolder. Mutually exclusive by construction (a single field, not independent bools): Block is the default (also every CPU tier, where only emit_tile_width/simd_width matter). Scaffolder-internal — a body never reads Context._tier.

Functions

• ​tile_alignment: Memory alignment a row-wise body's store — or its input_fn's load — should use for a width-ws tile of dtype on target.