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

pjoin

def pjoin[State: ReduceOp, params: ContextParams, //](mut state: State, mut ctx: Context[params])

CPU: drives the monoid's pjoin body via SerialReducer in the cooperative tier, or merges the cross-worker partials buffer in the split-axis tier.

Three tiers, picked from params:

Cooperative (emit_tile_width == 1, !is_splitk_tier): call state.pjoin(SerialReducer()). No cross-worker combine (one worker per output), but monoids may carry per-state finalization in pjoin (e.g. ReduceSum horizontally reducing its SIMD accumulator into the scalar acc field).
Tiled (emit_tile_width > 1): per-output accumulators already final; each lane is a separate output; skip the cross-participant join.
Split-axis (is_splitk_tier): write this worker's partial to its slot in ctx._partials_base, atomically increment counters_base[row_idx], and — only the worker whose increment brings the counter to blocks_per_row — gather all partials, join them into self.state, and set ctx._is_last_block so rowwise.once gates emission.

Parameters:

Args:

state (State): The body's monoid accumulator. On return, holds the finalized value.
ctx (Context[params]): The dispatch bundle. Mutated in the split-axis tier so rowwise.once knows which worker emits.