Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 16 Aug 2012 23:56:46 +0200
From: Lukas Odzioba <>
Subject: Re: OpenCL implementation of GPG format

2012/8/14 Dhiru Kholia <>:
> Hi,
> I have written an OpenCL implementation of GPG format. Currently, the
> speedup is very low (9x). The compiler complains about register
> spilling. According to
> speedups upto 100x can be achieved.

I love all those "scientists" who publish awesome results but do not
bother to publish any code, did they?

> Can someone on the list take a look and try to troubleshoot the
> performance issues? Thanks!

This is quite typical for all formats which use those magical
update,init,final functions taken from cpu code. This do not work well
on gpu at all. To achieve reasonable performance code needs a lot of
work, step by step removing those functions and replace them with
something smaller, with less memory operations, that do not use 8bit
types, fitted to password and salt max lengths, that can be vectorized
and so on. It is not an easy task but it is possible, maybe even then
it could outperform those claimed 100x.


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.