rmuir commented on PR #12311: URL: https://github.com/apache/lucene/pull/12311#issuecomment-1560143502
my issues with maxvectorsize/preferred was a jshell luser issue. it really works, so its easy for me to simulate 64/128 bit vectors on my 256-bit machine. i am still fine with disabling 64-bit vectors, but i am not sure we need to do it. We could add `-XX:MaxVectorSize=8, -XX:MaxVectorSize=16, -XX:MaxVectorSize=32, ...` to jenkins randomization if we want to ensure correctness? the 64-bit vectors (via `-XX:MaxVectorSize=8`) don't cause trappy behavior for me. For our integer functions, 64-bit only vectors are already skipped by the check mentioned above (we don't code up an algorithm for them), so there's no performance regression but also no gain. 64-bit only vectors still give some speedup (e.g. 2x) for floats on intel. And I'm not worried about correctness issues given we have an easy way to test. The bigger trap is what happens when you disable AVX: `-XX:UseAVX=0`. Supplying this flag gives incredibly trappy performance. IntVector.SPECIES_PREFERRED is still set to 128-bits, but there is no real vectorization for the integer functions and they run incredibly slow. floating point functions are still much faster, and work fine. I'm not sure what all UseAVX=0 is doing so maybe I need to spin up a QEMU that isn't configured well, "hiding" the AVX cpu flags, to see if it behaves like `-XX:UseAVX=0`. But this is definitely the kind of situation i'd like to avoid, and we don't detect it. So we are currently guarding against the wrong thing. -- 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: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
