Prefetch

[RagnarGrootKoerkamp]

This adds the ability to prefetch the cacheline required to answer a rank query.

I'm not quite sure I did it correctly though... should be fixed now
In my evals (https://curiouscoding.nl/posts/quadrank/#evals), the data points (the black plusses) get better when going from for to stream with prefetch.

My wrapping code is here: https://github.com/RagnarGrootKoerkamp/quadrank/blob/master/src/genedex.rs

[feldroop]

Hello @RagnarGrootKoerkamp,
Thank you for the issues and this PR. I will work on this repo in the next couple of days and I'll try to address everything. My goal also is to do some more profiling and optimizing, to get closer to the performance of your implementation. Maybe I'll even be able to add an implementation of the FMD-index.

[feldroop]

Just out of curiousity, did you measure a performance improvement from the inline annotations? I think it's fine to have them, but I assumed that the compiler already does is it, at least for the smaller functions where it makes a significant difference. Or is there another technical reason for the inlining?

[RagnarGrootKoerkamp]

They're not always necessary, but one non-inline function can massively (10-20%) impact perf. Just yesterday I had an add([[u64;4],[u64;4]) function not unlined even though it's pretty trivial.
Specifically if other functions become large it does sometimes happen that smaller functions are not inlined in my experience, so now I just like to put them just to be safe.

Also, inlining constructors and such allows the compiler to see into the actual data structure which can help optimizing the querying as well.