Date: Thu, 10 May 2012 01:50:58 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: Sayantan: Weekly Report #3 Sayantan - On Thu, May 10, 2012 at 02:46:58AM +0530, SAYANTAN DATTA wrote: > Thanks. Done. (regarding loop unrolling) Thanks. I'll take a look once this is in magnum-jumbo. > On Thu, May 10, 2012 at 12:02 AM, Solar Designer <solar@...nwall.com> wrote: > > > Oh, do you mean that the on-CPU portion of MSCash2, which we're > > executing sequentially with the on-GPU code (not in parallel yet), was > > consuming a substantial portion of total time? > > Yes. It consumed around 20% of the total time. Wow. 20% is excessive and not expected. We need to figure out why this is so - or maybe it is not (that is, part of the speedup might be from some side-effect, not directly from the OpenMP parallelization). You compute two MD4 hashes on host per MSCash2 hash, correct? At 80k c/s, this means you compute 160k MD4s per second, and you say this corresponds to over 20% of total running time (you say you saved 20% by parallelizing this). Thus, the total speed would be less than 800k MD4/s - sounds too low even for the somewhat non-optimal and non-vectorized code that we have there. For the raw-md4 format in a "make generic" build (thus also non-vectorized), I am getting between 5M and 6M c/s on this CPU (also on one core, indeed). (For vectorized code, it's more like 30M c/s on one core.) So there's a factor of 7 difference here, which we don't have an explanation for. > BTW GTX 570 performance has been also improved to 25k c/s with OMP. ...and what was it for the same code revision, but without OpenMP? > Also changed the file name from opencl_MSCASH2_fmt.c to > opencl_mscash2_fmt.c as requested by Lukas. Yes, thanks. That uppercase was weird. Alexander
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.