On Fri, 2006-05-19 at 20:20 -0700, Otis Gospodnetic wrote: > On Mon, 2006-05-08 at 18:34 -0700, Otis Gospodnetic wrote: > > > about the possibility of eliminating floating point calculations from > > > Lucene's scoring > > > Did you look in to this? > > > No, not yet. I did start looking at how exactly scoring works > (debugger, step through), but then had to switch to benchmarking > Lucene on a set of different machines/OSs/JVMs. Short answer: no. > Damn are those scoring classes hard to follow...
I gave it a glance. The floats originate from Similarity and IndexReader.norms[]. Eliminating them leaves a term count score, or so. Perhaps it is possible to run on int instead if there is a fast sqrt(int) out there. I can think of multiple problems when implementing an int norm value. An plenty of ugly solutions that might not be that fast at all. I also started looking from Hits.gatherMoreHits() -> TopDocs to find out how to bypass the whole scoring mechanism, but it is all very coupled and, as you say, hard to follow. --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
