Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Thu, 25 Aug 2011 08:46:35 +0200
From: Lukas Odzioba <lukas.odzioba@...il.com>
To: john-dev@...ts.openwall.com
Subject: Re: 1.7.8-allcuda-0.3

2011/8/25 Solar Designer <solar@...nwall.com>:
> Lukas -
>
> I've just tried john-1.7.8-allcuda-0.3.diff on my 8800 GTS 512 under
> Yellow Dog Linux with CUDA 3.2.  Here's how it works/fails:
>
> Benchmarking: PHPASSCUDA [PORTABLE-MD5]... DONE
> Raw:    182237 c/s real, 182237 c/s virtual
>
> Benchmarking: CRYPTMD5CUDA saltlen=8,passlen=8 [MD5-based CRYPT]... FAILED (get_hash[0](0))
> Benchmarking: CRYPTSHA256CUDA [SHA256-based CRYPT]... DONE
> Raw:    1556 c/s real, 1556 c/s virtual
>
> Benchmarking: CRYPTSHA512CUDA [SHA512-based CRYPT]... DONE
> Raw:    1033 c/s real, 1033 c/s virtual
>
> The failure for CRYPTMD5CUDA occurs every time, including when I invoke
> this one test separately.
>
> Alexander
>

I suppose it is because to big THREADS value in cryptmd5.h. Could you
please test  192 or 160 in free time (it's not urgent)? For now I'll
try set 192 as default and hopefully this will be highest safe value.

Anyway it seems that It should be done automatically, but as far as I
remeber threads and blocks values must be known during compilation so
each kernel should be compiled for every (most plausible)
threads/blocks configuration, compilation process will take too much
time, and we don't wont that. Maybe kernels should be distributed in
precompiled version... We'll see.

Lukas

Powered by blists - more mailing lists

Your e-mail address:

Powered by Openwall GNU/*/Linux - Powered by OpenVZ