Module block

Expand description

Block collectives. Every thread in the block participates in every call.

Modules§

exclusive_scan: Ordered exclusive scan. Returns the unseeded aggregate to every thread.
inclusive_scan: Ordered inclusive scan of a blocked register tile. Scratch contains threads * items_per_thread elements. valid_items is uniform and positive; invalid output elements are unspecified.
reduce: Ordered tree reduction; the aggregate is broadcast to all block threads. valid_items must be positive. Input is in blocked register order.
scan_with_prefix: Inclusive/exclusive scan with a stateful block prefix callback. The first native subgroup invokes the callback, and thread zero’s result is used. Scratch requires threads * items_per_thread + 1 entries. Valid input is positive and callback state updates occur on the participating threads.