[LUCENE-10453] Speed up VectorUtil#squareDistance

• Type: Task
• Status: Reopened
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 9.1
• Labels: vector-based-search
• Lucene Fields: New
• Assignee: Unassigned
• Reporter: Adrien Grand
• Created: 03/Mar/22 08:31
• Updated: 27/Sep/22 09:16
• Resolved: 05/Mar/22 17:34

Description

VectorUtil#squareDistance is used in conjunction with VectorSimilarityFunction#EUCLIDEAN.

It didn't get as much love as dot products (LUCENE-9837) yet there seems to be room for improvement. I wrote a quick JMH benchmark to run some comparisons: https://github.com/jpountz/vector-similarity-benchmarks.

While it's not as fast as using the vector API (which makes squareDistance computations more than 2x faster), we can get a ~25% speedup by unrolling the loop in a similar way to what dot product does.

Issue Links

• links to: GitHub Pull Request #725