Fonctions de vote et de scrutin

Beyond moving data, a warp can answer yes-or-no questions together. Vote intrinsics let every lane share a true-or-false predicate in one cheap step.

The most flexible vote is ballot_sync. It returns a 32-bit number where bit i is 1 exactly when lane i's predicate was true.

That returned integer is a tiny bitmap of the warp. Reading bit i tells you whether lane i voted true, all packed into a single register.

Want how many lanes said yes? Pop-count the ballot result. popc counts the set bits, giving the number of true votes in the warp instantly.

Two shortcuts answer common questions. all_sync returns true only if every lane agreed, while any_sync returns true if at least one lane did.

Use any_sync for early exits: if any lane found a match, the whole warp can react together without scanning lane by lane.

Ballot powers compaction: each lane decides if it keeps its element, votes, and the bitmap tells lanes where to write into a packed output array.

To place its value, a lane counts the true bits below its own position. That prefix count from the ballot is its index in the compacted result.

Sometimes one lane should do shared work. The lowest set bit in the ballot, found with ffs, is a natural leader the warp can all agree on.

Like shuffles, vote intrinsics take an active mask. Inactive lanes contribute a 0 bit, so pass the correct set to avoid counting absent lanes.

Votes cost a single instruction and touch no memory. They let a warp coordinate decisions far cheaper than atomics or shared-memory flags would.

Recall what ballot_sync hands back to each lane.

You polled a warp with ballot and vote: count trues, check all or any, and compact streams cheaply. Next: cooperative groups for flexible sync. ✨