Date: Thu, 05 Jul 2012 15:19:15 -0300 From: Claudio André <claudioandre.br@...il.com> To: john-dev@...ts.openwall.com Subject: Re: [private] common find_best_workgroup() Not sure it deserves such attention. If you get rid of find_best_workgroup, everybody can play with something similar just trying: LWS=32 LWS=64 LWS=128 Maybe 2 options on CPU (1 and 8). It is to much work to deal with a few (32 or 64 or 128) options. No? Em 05-07-2012 14:56, magnum escreveu: > On 2012-07-05 16:23, magnum wrote: >> I have already committed an experimental fix. Seems to work fine but >> still testing (I also changed all formats except Claudio's so they use >> the shared function) > I reverted this change (using a shared find_best_workgroup() ) for > opencl_xsha512_fmt.c because it had a "special" loop in its local one: > it runs each size 10 times and sums the exec time. And its performance > got worse (selecting a lower LWS) with the shared one. > > This gave me this idea (to-do) for the shared one: > > 1. perform a warm-up run of crypt_all() before the loop, but check the > exec time. > 2. From that exec time, chose a suitable number of loops (targeting a > minimum sum of exec times) > 3. Do wot myrice did, using that number of loops. > > This might greatly reduce the "randomness" I have experienced with > find_best(). And when this is implemented, opencl_xsha512_fmt.c can get > on this train too again :) > > magnum >
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.