mulugetam opened a new issue, #12091: URL: https://github.com/apache/lucene/issues/12091
### Description Lucene's implementation of ANN relies on a scalar implementation of the vector similarity functions [dot-product,](https://github.com/apache/lucene/blob/4fe8424925ca404d335fa41d261545d3182c22fa/lucene/core/src/java/org/apache/lucene/index/VectorSimilarityFunction.java#L53) [Euclidean distance](https://github.com/apache/lucene/blob/4fe8424925ca404d335fa41d261545d3182c22fa/lucene/core/src/java/org/apache/lucene/index/VectorSimilarityFunction.java#L34), and [cosine](https://github.com/apache/lucene/blob/4fe8424925ca404d335fa41d261545d3182c22fa/lucene/core/src/java/org/apache/lucene/index/VectorSimilarityFunction.java#L71). The vector implementation of these functions is quite straightforward. Below is performance data I got, based on JMH, comparing the vector implementation of the `dot product` and `Euclidean` against the equivalent default (scalar with loop-unrolling) implementation. `dim` is the dimension/length of the `float[]` arrays in test and `score` is the number of dot product/Euclidean distance operations done per second. ``` Benchmark dim Mode Cnt Score Units Gain ---------------------------------------------------------------------------------------------- scalarDotProduct 60 thrpt 12 32031825.541 ± 6151.580 ops/s 1.00 scalarDotProduct 120 thrpt 12 17120537.911 ± 5793.505 ops/s 1.00 scalarDotProduct 480 thrpt 12 4506350.215 ± 1677.755 ops/s 1.00 vectorDotProduct 60 thrpt 12 98862701.038 ± 85554.695 ops/s 3.09 vectorDotProduct 120 thrpt 12 99059913.888 ± 20609.182 ops/s 5.79 vectorDotProduct 480 thrpt 12 220320941.436 ± 173467.603 ops/s 48.89 ``` ``` Benchmark dim Mode Cnt Score Units Gain ---------------------------------------------------------------------------------------------- scalarSquareDistance 60 thrpt 12 25890614.822 ± 7071.413 ops/s 1.00 scalarSquareDistance 120 thrpt 12 12524294.760 ± 3435.882 ops/s 1.00 scalarSquareDistance 480 thrpt 12 3145045.026 ± 409.361 ops/s 1.00 vectorSquareDistance 60 thrpt 12 104317302.765 ± 36895.474 ops/s 4.03 vectorSquareDistance 120 thrpt 12 122083614.889 ± 11821.642 ops/s 9.75 vectorSquareDistance 480 thrpt 12 362229408.898 ± 85439.065 ops/s 115.17 ``` I have also tested the same with [Msokolov's ANN benchmark suite](https://github.com/msokolov/ann-benchmarks) and saw a speedup of more than 2x in indexing (docs/sec) and search performance (QPS). Will do a PR for it soon. Let's discuss this :-) -- 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.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
