LUCENE-9071: Speed up BM25 scores. (#1043)

[IcanDoItL]

https://github.com/apache/lucene/blob/0345fcabb3ceb71fb39c1aa77502bae220f259ca/lucene/core/src/java/org/apache/lucene/search/similarities/BM25Similarity.java#L246,How to understand monotonicity?, freq * 1/norm is guaranteed to be monotonic? but My experimental results are the same as freq / (freq + norm).
the code is
@test
public void test() {
int length = 2;
float freq = 1.0f;
float[] cache = {16777216.0f, 16777217.0f};
float[] inverseCache = new float[length];
for (int i = 0; i < length; i++) {
inverseCache[i] = 1f / cache[i];
}
for (int i = 0; i < length; i++) {
System.out.println(freq / (freq + cache[i]));
}
for (int i = 0; i < length; i++) {
System.out.println(freq * inverseCache[i]);
}
}

[jpountz]

The problem is not with norm but with freq in that case. Try to compute BM25 scores with a few freqs and you'll find cases when you get a lower score evern though freq is greater by one ulp.

[IcanDoItL]

@jpountz thanks

Try to compute BM25 scores with a few freqs and you'll find cases when you get a lower score evern though freq is greater by one ulp.

I didn't find during my testing. Can you give me an example

[jpountz]

If you apply the change that you suggested and run TestBM25Similarity, you should get failures that will give you examples.