jpountz commented on PR #14203: URL: https://github.com/apache/lucene/pull/14203#issuecomment-2726015514
Thanks for running benchmarks. So it looks like the JVM doesn't think these shorter loops (with step 128) are worth unrolling? This makes me wonder how something like that performs on your AVX-512 CPU. I think you had something similar in one of your previous iterations. On my machine it's on par with the current version. ```java private void readInts24(IndexInput in, int count, int[] docIDs) throws IOException { if (count == BKDConfig.DEFAULT_MAX_POINTS_IN_LEAF_NODE) { // Same format, but enabling the JVM to specialize the decoding logic for the default number // of points per node proved to help on benchmarks doReadInts24(in, 512, docIDs); } else { doReadInts24(in, count, docIDs); } } private void doReadInts24(IndexInput in, int count, int[] docIDs) throws IOException { // Read the first (count - count % 4) values int quarter = count >> 2; int numBytes = quarter * 3; in.readInts(scratch, 0, numBytes); for (int i = 0; i < numBytes; ++i) { docIDs[i] = scratch[i] >>> 8; scratch[i] &= 0xFF; } for (int i = 0; i < quarter; ++i) { docIDs[numBytes + i] = scratch[i] | (scratch[quarter + i] << 8) | (scratch[2 * quarter + i] << 16); } // Now read the remaining 0, 1, 2 or 3 values for (int i = quarter << 2; i < count; ++i) { docIDs[i] = (in.readShort() & 0xFFFF) | (in.readByte() & 0xFF) << 16; } } ``` -- 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