Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 5 Jul 2013 13:36:27 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: bug: GPU use in CPU-only formats

On 5 Jul, 2013, at 8:52 , Sayantan Datta <std2048@...il.com> wrote:
> On Friday 05 July 2013 05:23 AM, magnum wrote:
>> On 5 Jul, 2013, at 1:24 , magnum <john.magnum@...hmail.com> wrote:
>>> I'm looking into some weirdness with mscash2-opencl when using multiple devices, I may have to tweak something.
>> That problem was not introduced now so I'll leave it to Sayantan. It only happens with --test=0:
>> 
>> $ ../run/john -t=0 -form:mscash2-opencl -dev=gpu
>> Warning: doing quick benchmarking - the performance numbers will be inaccurate
>> Device 0: GeForce GTX 570
>> Device 1: Tahiti (AMD Radeon HD 7900 Series)
>> Optimal Work Group Size:256
>> Kernel Execution Speed (Higher is better):1.565550
>> Optimal Global Work Size:253952
>> WARNING: Device 0 is too slow and might cause degradation in performance.
>> WARNING: Devices are NOT arranged in DECREASING order of performance and might cause degradation in performance.
>> Benchmarking: mscash2-opencl, M$ Cache Hash 2 (DCC2) [PBKDF2-SHA1 OpenCL]... OpenCL error (CL_INVALID_WORK_GROUP_SIZE) in file (common_opencl_pbkdf2.c) at line (353) - (Enque Kernel Failed)
> 
> What does the -t=0 option does ? It seems like 570 isn't being tested at all.

For CPU formats, --test=0 (0 second test) [mostly] just self-tests - the benchmark figure is incorrect but all self-tests are completed.

For OpenCL formats I emulate this by internally setting LWS=7 and GWS=49 whenever --test=0 is used. This way, the OpenCL formats test a whole lot quicker (and does not auto-tune, so we save even more time) with the same caveat that the benchmark figure will be low.

For mscash2 I realize you try to balance LWS/GWS for heterogenous devices so the current behaviour is not really a bug. Either you could add some handling of this special case (should be easy), or just ignore it and document.

Ideally, a "./john -t=0 -form:opencl" should test all OpenCL formats pretty quickly (but not properly benchmark them).

magnum

Powered by blists - more mailing lists

Your e-mail address:

Powered by Openwall GNU/*/Linux - Powered by OpenVZ