Repository navigation
bit_util::compress: dispatch per block of words for sparse masks #11543
Description
Follow-up to #11322. Sparse words pay the dispatch and loop exit per word; on the Neoverse-V2 bot run, filter_bits random 1/16 went from 651.3 to 695.8 us, and AArch64 has no pext path. Idea: popcount 4 to 8 words at once and dispatch per block.
Activity
JonSnow1807 commentedon Nov 2, 2026
I would like to take this if you have not started: block-wise popcount over 4-8 words with the compress strategy chosen per block, keeping the per-word Packer, measured on an M2 Pro (AArch64, no pext) with the filter_bits benches before and after. Fine if you prefer to do it yourself.