Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 12 Apr 2015 01:49:43 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: [GSoC] John the Ripper support for PHC finalists

On 2015-04-11 21:14, Solar Designer wrote:
> On Sat, Apr 11, 2015 at 09:05:49PM +0200, Agnieszka Bielec wrote:
>> 2015-04-11 20:02 GMT+02:00 magnum <john.magnum@...hmail.com>:
>>> On 2015-04-11 16:39, Solar Designer wrote:
>>>> magnum, what base format would you suggest for a new split kernel format?
>>>
>>> RAR5 might be a clean base. Many of the other ones support vectorizing
>>> (decided at run-time) so are somewhat more complex (although that might
>>> be interesting for Pomelo too?). I suspect some of Claudio's formats
>>> might be a good start too, I haven't looked much at them as he actively
>>> maintains them himself.
>>
>> you mean only opencl version?
>> or also non-opencl?
> 
> I think magnum meant only OpenCL, because the above was in context of
> the topic of split kernels.  We say "split kernel" to mean an OpenCL
> kernel that performs only a portion of a hash computation (or rather,
> portions of many parallel hash computations, obviously) in order to keep
> its running time per invocation sane.  (And then we invoke such kernels
> multiple times in a loop from within crypt_all() in order to complete
> the computation of final hashes.)  This is needed primarily to prevent
> triggering timeouts and watchdogs in drivers.  It may also improve
> interactive response times of john and of other concurrent GPU-using
> programs.

Exactly. Actually split kernels tend to perform better than monolithic
ones, perhaps because the loop kernel becomes lean and can be optimized
better by the compiler. And on weak devices also because we simply can
use a higher GWS.

The overhead you could expect of calling the loop kernel many times, and
even the kernel overhead of storing/retrieving intermediate results (we
can only use global memory for that but we never have to transfer it
anywhere), seem to be negligible.

magnum

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.