Frido was complaining that the nbl was too slow, so I thought I would
check in my local changes. The big thing is that the fastest
implementation I have uses the CGAL library.
All patches to IMP need to be reviewed before they are checked in.
Obviously, I am not going to do this while I'm on vacation, so unless
somebody else with some understanding of the code volunteers to check
through these patches, they'll have to wait until I'm back in the lab.
I don't think we can use CGAL. It is certainly available for Fedora, but
we build on a lot of other platforms too. And my understanding is that
large chunks of it are QPL licensed, which is certainly incompatible
with the IMP license. If we did use it, we'd have to be very careful to
use only the LGPL parts. So I'd rather we didn't use it at all.
Your timings suggest that the current grid nonbonded list implementation
is not O(N). Why not? It seems like it should be from my understanding
of the code.