mbutrovich commented on PR #12347: URL: https://github.com/apache/arrow-rs/pull/12347#issuecomment-7052651015
Thanks for the feedback and benchmarking, @andygrove! I also have a reply somewhere in the bot noise here: https://github.com/apache/arrow-rs/pull/12347#discussion_r4648646701 > With native, `count_set_bits_offset` itself got slower than on the default build (about 107 µs vs 71 µs at width 4). This is outside the PR, but I compared the benchmark's `count_set_bits_offset` loop built for `x86-64` and for `-C target-cpu=znver4`. At width 4 no row holds a complete 64-bit word, so in both builds a row costs a call to `UnalignedBitChunk::new` and popcounts of the masked first and last words. The znver4 build adds a `vzeroupper` before each call and uses `popcnt` instead of the SSE2 sequence, and `UnalignedBitChunk::new` has no AVX-512 code in either build. I don't see what would make it 1.5x slower, and the bot's Xeon shows no change at `x86-64-v3` (82.9 us against 83.3 us). Does it reproduce when both builds run pinned to the same CCD? If it does, I'll open an issue with a profile. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
