Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 17 Nov 2012 11:10:10 +0100
From: magnum <john.magnum@...hmail.com>
To: john-users@...ts.openwall.com
Subject: Re: scalar mode

This drives me mad. Could you check out the latest again, and apply the enclosed patch? I actually do the same for OSX in office2007 due to what I think is an nvidia bug.


Download attachment "bughunt.diff" of type "application/octet-stream" (1084 bytes)



In case this solves the problem, please provide benchmarks from running with and without the --request-vectorize option.

magnum


On 17 Nov, 2012, at 10:13 , Marc Brinkmann <marc.brinkmann@...il.com> wrote:

> I',m sorry, but it did not help in my case :(
> i tried sabayon with recent catalyst 12.10 and my ubuntu 12.04 with
> catalyst 12.8 and the very recent 12.11 beta8
> none of them were working, but 12.11beta even did not work with the version
> <= 13.11.2012
> I guess its AMD's damn driver and will simply use the old build with
> catalyst 12.8.
> 
> Btw. while i tested your recent builds on my Netbook running ION i never
> had any issues, so Nvidia drivers seem to be much more stable ....
> Of course it is slow 120c/s.. but i never runs into segfault...
> 
> 
> 2012/11/16 magnum <john.magnum@...hmail.com>
> 
>> On 16 Nov, 2012, at 23:10 , magnum <john.magnum@...hmail.com> wrote:
>>> Try latest commit. The office2007/2010/2013 and ntlmv2 OpenCL formats
>> will now default to scalar mode on any device, including CPU. You can now
>> add --request-vectorize to your command line and see if vectorizing
>> triggers the bug, or if it gets faster or slower. Office actually gets a
>> tad faster on some nvidias, I've never seen that before.
>> 
>> Correction: It is ntlmv2 that gets faster vectorized on some nvidias. And
>> only the "one salt" case (the opposite would seem more natural to me as
>> we're memory bound).
>> 
>> Vectorizing the Office formats are very bad for nvidia but has shown some
>> 40% boost on Cayman.
>> 
>> CPUs should benefit from vectorizing but I have seen huge regressions in
>> some cases and great speedups in other cases. Maybe the ideal would be to
>> decide at runtime after a quick benchmark. I'm not feeling very inspired
>> right now though. Damn drivers :-(
>> 
>> magnum
>> 


Powered by blists - more mailing lists

Your e-mail address:

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.