jpountz commented on issue #1043: LUCENE-9071: Speed up BM25 scores. URL: https://github.com/apache/lucene-solr/pull/1043#issuecomment-562194793 I understand it's surprising that this approach worked, but I've been able to reproduce it consistently with luceneutil, including by swapping the base and patch in order to avoid any bias due to which competitor runs first. I also confirmed it via microbenchmarks. I wonder whether there is something special about my CPU. I've been exploring other ideas and found an approach that yields almost exactly the same speedups but is simpler. The way it works is by rewriting the BM25 formula from `weight * (freq / (freq + norm))` to `weight - weight / (1 + freq * (1/norm))`. This formula doesn't need double promotion in order to guarantee monotonicity on `freq` or `norm`.
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
