Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 02 Aug 2011 23:06:12 +0200
From: Samuele Giovanni Tonon <samu@...uxasylum.net>
To: john-dev@...ts.openwall.com
Subject: Re: Lukas's Status Report - #12 of 15

On 08/01/11 20:51, Lukas Odzioba wrote:
> Last week I was working on:
> -phpass OpenCL patch - not finished
> -md5crypt OpenCL patch - finished rev0 290710 c/s
> -sha256crypt OpenCL patch -finished rev0 7027 c/s
> 
> This week I will be working on:
> -test md5crypt,sha256crypt opencl patches
> -sha512crypt OpenCL patch:
> -phpass OpenCL patch:
> -pbkdf2 OpenCL patch:
> 
> Additional info:
> With phpass OpenCL I've got weird bug. It's not working at all for
> passwords length equal 4.
> Within few hours I will upload md5crypt and sha256 OpenCL patches to wiki.
> I've purchased new power supply, and I'm looking forward to add
> multigpu support to previous patches.
> I haven't tested OpenCL patches on AMD cards because of lack of that hardware.
> 
> I would appreciate if someone could find free time to test them and
> modify Makefile accordingly.

i can help you on that:
you need to change Makefile accordingly to this patch i'm applying,
beware that lib include depends strictly on arch: people with 32bit
should change that line accordingly (x86 instead of x86_64) .


it seems there's a problem with path:

john-1.7.8/src$../run/john -test --format=CRYPTMD5OPENCL
Benchmarking: CRYPTMD5OPENCL saltlen=8,passlen=8 [MD5-based CRYPT]...
OpenCL Device: Juniper
Error while building program [-11]
BUILD LOG:
 /tmp/OCLyCdIrz.cl(1): catastrophic error: cannot open source file
          "../src/cryptmd5opencl_fmt.h"
  #include "../src/cryptmd5opencl_fmt.h"
                                        ^

1 catastrophic error detected in the compilation of "/tmp/OCLyCdIrz.cl".
Compilation terminated.

but
john-1.7.8/src$ls -la ../src/cryptmd5opencl_fmt.h
-rw-r--r-- 1 samu samu 1039 Aug  2 22:20 ../src/cryptmd5opencl_fmt.h

exist.
program doesn't start even if you run from "run" dir, i suspect it's
some path issues with amd opencl.

i circumvented the problem by statically adding the .h inside the .cl
file but i got this next problem:
./john -test --format=CRYPTMD5OPENCL
Benchmarking: CRYPTMD5OPENCL saltlen=8,passlen=8 [MD5-based CRYPT]...
kp=./../src/cryptmd5.cl

OpenCL Device: Juniper
Error (-55) in file (cryptmd5opencl_fmt.c) at line (368)
which is an CL_INVALID_WORK_ITEM_SIZE .

local_work_size is setted up at 384 (THREADS define)
i forced threads to be 256 (which is max work_size for my hd 5750):

./john -test --format=CRYPTMD5OPENCL
Benchmarking: CRYPTMD5OPENCL saltlen=8,passlen=8 [MD5-based CRYPT]...

OpenCL Device: Juniper
DONE
Raw:    41041 c/s real, 43916 c/s virtual

then i added *32 in the define of KEYS_PER_CRYPT because i'm evil

#define KEYS_PER_CRYPT          BLOCKS*THREADS*32

and rerun test:

$./john -test --format=CRYPTMD5OPENCL
Benchmarking: CRYPTMD5OPENCL saltlen=8,passlen=8 [MD5-based CRYPT]...
OpenCL Device: Juniper
DONE
Raw:    72282 c/s real, 146176 c/s virtual

then i added *64
#define KEYS_PER_CRYPT          BLOCKS*THREADS*64

and got
./john -test --format=CRYPTMD5OPENCL
Benchmarking: CRYPTMD5OPENCL saltlen=8,passlen=8 [MD5-based CRYPT]...
OpenCL Device: Juniper
DONE
Raw:    71364 c/s real, 436906 c/s virtual


:-)
hope this will help you.

about getting the correct work_item_size just wait next patch i'll
upload about rawsha1 and nsldaps, i took some code to alain to try to
detect the best work_size.

Cheers
Samuele

View attachment "opencl_amd.patch" of type "text/x-patch" (1235 bytes)

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.