Refactor SearchResult to avoid NodeScore[] allocations about jvector HOT 2 CLOSED

That said, profiling says that the NodeScore[] copy is roughly 0.1% as expensive as reading the vectors and computing the full precision scores. So I'm going to close this as premature optimization.

from jvector.

... It's actually not obvious to me what the best solution is here.

1. Implement a custom sort that sorts the int[] and vector[] in place while preserving correspondance.
2. Implement a custom radix sort based on score * Int.MAX_VALUE.
3. Allocate an indices array and sort that with a custom comparator. Don't need to move the original data around, just go through the indices array for getNode, getScore.
4. Go up a level, int[] and VectorFloat are the wrong level of abstraction, instead we should inject the corrected scores into the NodeQueue and let the caller pull out just the ones needed instead of sorting the whole set

I like #4 a lot but it's harder to fit native code into.

from jvector.