Vector efficient math operations for KNN distance functions

[Tradunsky]

Is your feature request related to a problem? Please describe.
Exact-KNN distance functions use sub-optimal vector math operations. This can contribute to extra time on high vector dimension distance compute.

Describe the solution you'd like
Do you think it is possible to use Vector API for math operations in the distance functions?
For example, the following cosine similarity that iterates over each element in both vectors to multiply numbers:

k-NN/src/main/java/org/opensearch/knn/plugin/script/KNNScoringUtil.java

Line 96 in 120c4ce

public static float cosinesimilOptimized(float[] queryVector, float[] inputVector, float normQueryVector) {

could apply vector optimal fused multiply-add instead.

If for the usage of Java Vector API, supported java version is a concern, it could also start with Math.fma().

Describe alternatives you've considered
ANN would be an option if pre-filtering supported.

Additional context
Much appreciate OpenSearch work on KNN, as it gets bigger demand with ML to relay on vector supporting DBs on distributed computations.

[dblock]

Where would we want to configure which distance function to use? Index level options?

[vamshin]

@dblock this will be applicable to all https://opensearch.org/docs/latest/search-plugins/knn/knn-score-script/ functions and will not be part of index options. This will be optimization to our existing vector calculations for bruteforce/exact knn search

[vamshin]

Looks like we need OpenSearch core to move to open jdk 19 to use Math.fma(). Will prioritize once we have this version in core

[Tradunsky]

Linking here for visibility:
apache/lucene#12091

This is also implemented in Elasticsearch and dropped latencies by at least 50% in my case.

[jmazanec15]

Addressed in #1699