Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Tue, 31 May 2011 13:13:58 +0400
From: Solar Designer <>
Subject: Re: Lukas's Status Report - #3 of 15


On Mon, May 30, 2011 at 11:07:32PM +0200, ?ukasz Odzioba wrote:
> Done:
> -bought GTX460 1024Mb card
> -started to read ATI stream documentation

Sounds good.

Based on oclhashcat's success with just CUDA and OpenCL, I guess we'll
end up using OpenCL rather than ATI Stream for ATI/AMD cards.  But I am
not familiar with this, so you're to research it.  Using OpenCL will let
you do some testing on your NVidia cards as well (not ATI/AMD specific
benchmarks, though).

> -implemented ~20% of md5-based crypt
> -tried to profile patches in nVidia nSight but it turned out to be
> impossible on windows XP. I need to install windows Vista od Seven,
> but it must wait until end of semester(2weeks).
> -after Solar's sugestions I've tested full unrolled version of
> sha256-based crypt. It is 5-10% faster than partially unrolled. To
> achieve better results I must use only shared memory.

OK. Based on our IRC conversation, full unrolling was needed to use
registers instead of shared memory, but you also have uses of global
memory, which kill performance anyway.

> ToDo:
> -finish md5-based crypt patch


> -generate md5-based crypt test data

I think there's plenty of test data in the test suite by JimF and magnum -
perhaps download it and take a look.

> -try to coascale memory operations in sha256patch

You mean "coalesce".  Yes, although avoiding them altogether (inside the
performance critical loop) would be even better (if possible).



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.