Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Thu, 19 Sep 2013 21:59:49 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: CUDA

On 19 sep 2013, at 20:19, Lukas Odzioba <lukas.odzioba@...il.com> wrote:
> 2013/9/19 magnum <john.magnum@...hmail.com>:
>> I now tried this on my (sm_30) CUDA 5 Macbook. CUDA compilation obviously takes longer as it builds three different versions of each kernel. Unfortunately, and as seen in tests long ago (just changing from sm_10 to sm_20/30) some formats actually get much slower from this.
> 
> Sorry that I didn't have time to take a look at this. Do we know which
> code is used in runtime? I guess highest supported by gpu, this is not
> always the best option, but still better than default sm_10 which we
> use now just for compatibility with older gpus.

It seems like the highest supported is used, yes.

Some complications on Well: When using "-gencode arch=compute_10,code=sm_10 -gencode arch=compute_20,code=sm_21" it fails on the Titan (which is sm_35):

Benchmarking: md5crypt-cuda, crypt(3) $1$ [MD5 CUDA]... invalid device symbol in cryptmd5.cu at line 225

If I hack that format out, next format get a similar error. If I instead take the sm_20 stuff out, I also get the same error. So our original "-arch sm_10" is somehow different from "-gencode arch=compute_10,code=sm_10". The former works, the latter doesn't.

Adding "-gencode arch=compute_30,code=sm_30" instead, gives higher lever problems:

...
Benchmarking: sha512crypt-cuda, crypt(3) $6$ (rounds=5000) [SHA512 CUDA]... FAILED (get_hash[0](0))
...
Benchmarking: mscash2-cuda, M$ Cache Hash 2 (DCC2) [PBKDF2-SHA1 CUDA]... FAILED (get_hash[0](0))
...
2 out of 12 tests have FAILED

Reverting to "-arch sm_10", all is fine.

Changing to "-arch sm_20" or "-arch sm_30" and even "-arch sm_35", gets us the same as "-gencode arch=compute_30,code=sm_30", that is 2 formats fail self-test.

So... the only thing that works is sm_10. Is this a release driver?

magnum

Powered by blists - more mailing lists

Your e-mail address:

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