mikemccand commented on issue #15189: URL: https://github.com/apache/lucene/issues/15189#issuecomment-3312331029
Thanks @Tony-X -- your talk at [Community over Code NA](https://communityovercode.org/schedule/#) (in Minneapolis last week/weekend, at the very last slot (Sunday 5 PM)) was awesome! I managed to record it and will try to process/share soon. I think auto-vectorization is/was somehow related here? @jpountz did tons of work long ago to structure the autogen'd (using Python! yay!) java code "just right", so that C2 hotspot compiler reliably auto-vectorizes. And the resulting performance impact was awesome ... I think that maybe we changed the blocks to `int[]` during that? But I'm not sure ... There are two paths in Java to access vectorized (SIMD) CPU instructions. One is hoping C2 hotspot compiler automagically compiles down to the right instructions for the current CPU. The other is Panama's (still incubating, resulting in the mrjar infra in Lucene's sources and build tooling, since Panama's APIs change from one JDK to the next) explicit but abstract vector APIs. I think the postings decode was done before we had the option of Panama (integrated into Lucene). I wonder if switching to Panama's APIs for postings decode would somehow be better? I don't really understand all the tradeoffs... -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
