2020-11-19, 17:17   #73
bsquared

"Ben"
Feb 2007

3,371 Posts

There have been a few updates to the avx-ecm repo:

1) added algebraic factor removal of Mersenne inputs
2) added checkpoint saving every 1e8 primes in stage 1
3) implemented batch inversion in stage 2, now almost twice as fast

Here's an example of that last item compared to previously
Quote:
 Originally Posted by bsquared Expanding on the example from before, with B1=7M on 2^1277-1: avx-ecm stage 1: 78 sec for 8 curves, 9.7 sec/stg1-curve avx-ecm stage 2: 55 sec for 8 curves, 6.8 sec/stg2-curve
Now:
Code:
B1=7M on 2^1277-1:
avx-ecm stage 1: 78 sec for 8 curves, 9.7 sec/stg1-curve
avx-ecm stage 2: 32 sec for 8 curves, 4.0 sec/stg2-curve
At this point I'm not sure whether it's better to keep the faster stage 2 time or run to higher B2 bounds, keeping timings similar to previously.

I think that Bob's paper says to equalize the stage 1 and stage 2 runtimes, so the best answer is probably the latter. It might be moot if the best-best answer is to still use a hybrid plan with gmp-ecm stage 2. However this might change the crossover point (from pure avx-ecm to hybrid).