[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 13 Jun 2012 22:07:09 +0530
From: SAYANTAN DATTA <std2048@...il.com>
To: john-dev@...ts.openwall.com
Subject: Re: Sayantan: Weekly Report #8
On Wed, Jun 13, 2012 at 1:30 PM, Lukas Odzioba <lukas.odzioba@...il.com>wrote:
> 2012/6/13 SAYANTAN DATTA <std2048@...il.com>:
> > Accomplishments:None except some minor bug fixes.
> > (I spent most of my time experimenting with different optimizations for
> > blowfish kernel which unfortunately didn't result in much improvement in
> > performance.)
> Would you like to share info about what type of improvements you've
> tried and how they affected speed?
> Have you tried any profilers?
>
I have tried using LDS and global memory for storing the S_boxes. The LDS
implementation is faster by around 50-60%. The LDSBankConflicts parameter
suggested by the profiler is around 8.3 which I think is low enough. I also
made an attempt to furthur minimize LDSBankConflicts which wasn't much
helpful because the performance gains due to slightly lowered LDS bank
conflicts was nullified by the complex indexing arithmatic. So a simple
access pattern seems to be most suitable. However the global memory version
of the kernel hugely benifitted from the minimized channel conflicts(around
2x performance gain). Still the global memory version was slower than the
LDS version. Since my implementation was mostly limited by LDS so I tried
to increase the VGPRs used by the kernel. To do this I manually unrolled a
few loops which increased the VGPR counts to maximum value 256(maximum
possible value on 7970) and also started using scratch registers. As a
result the performance dropped by 10%. This was clearly an unexpected
behavior. I also tried #pragma unroll but it caused wrong calcualation of
hashes on 7970.
>
> > Priorites: Please suggest something.
> Alexander should suggest something. I might guess that after blowfish
> we might want to have bcrypt now, but ask Alexander first.
>
> 1) We've got site like this:
> http://openwall.info/wiki/john/GPU
> It would be great if you could add info (table + bench results) about
> your patches to it.
>
> 2) Add your format to doc/README.opencl
> 3) Format your code properly (bf_kernel.cl)
>
Okay, I'll do that.
Content of type "text/html" skipped
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.
