Improve proxy generator

[alexfikl]

This takes out another chunk out of #30 that just changes the proxy generator. The bigger changes are

• Separated out a ProxyGenerator, that ignores any QBX expansions, and a QBXProxyGenerator, that makes proxies outside of the QBX expansions.
• Separated the loopy kernels that get centers and radii to make the proxy balls. Also split some inames and things, but that probably requires a careful (and knowledgeable) look.
• removed unused gather_block_interaction_points, which was meant to just make a big array of neighbors + proxies to be evaluated together.
• cleaned up the tests a bit and added some more checks.

Looking at the diff, it looks a bit big, but it should just be moving things around.

[inducer]

That diff is genuinely unhelpful, but the end result looks OK. As for your split_inames: It wouldn't really fly on a GPU, but it'll do for CPU use. :)

[inducer]

Thanks!

[alexfikl]

As for your split_inames: It wouldn't really fly on a GPU, but it'll do for CPU use. :)

I mostly copied that from some stuff I managed to sneak by you a long time ago in sumpy. What would be a better way to optimize those kernels a bit?

[inducer]

Got a whole course on that: https://relate.cs.illinois.edu/course/cs598apk-f18/ :) But I'd be happy to chat on Thursday. Which ones are the expensive ones?

[alexfikl]

Got a whole course on that: relate.cs.illinois.edu/course/cs598apk-f18 :) But I'd be happy to chat on Thursday. Which ones are the expensive ones?

It was just meant to do a bare minimum sort of thing, since I haven't profiled anything. For the moment, all the geometries we've been playing with have like 2-3000 points, so it probably doesn't matter that much anyway.

Thanks! I'll give that a read!