[chemfp] 33% speedups for 4096 bit searches

[Andrew Dalke]

On Nov 9, 2011, at 12:44 AM, Andrew Dalke wrote:

> If I did my math right, that's about 28 million 4096-bit

> comparisons per second, or 35ns per comparison. On a

> single node of a desktop machine.


Nope! I can't count digits. It's 2.8 million per second,
or 350ns per comparison.

This problem scales linearly with fingerprint size, so
the 1024 bit fingerprint should be about 11 million/sec.

There's someone I've been talking to who is good at
this sort of optimization. He uses popcounts for computing
primes and prime pairs. I would love to bring him in
on this.

I'm estimating, based on other benchmarks, that a
several-fold improvement still available.


				Andrew
				dalke at dalkescientific.com