Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 2 Nov 2014 13:02:48 -0900
From: Royce Williams <royce@...ho.org>
To: john-dev <john-dev@...ts.openwall.com>
Subject: Re: descrypt speed

On Sun, Nov 2, 2014 at 12:19 PM, magnum <john.magnum@...hmail.com> wrote:

> On 2014-11-02 18:59, Royce Williams wrote:
>
>> On Thu, Oct 30, 2014 at 9:33 PM, magnum <john.magnum@...hmail.com> wrote:
>>
>>> On 2014-10-31 06:02, Royce Williams wrote:
>>>
>>>> On a GTX970, shouldn't this be sm_52?
>>>>
>>>
>>> You can force this by editing NVCC_FLAGS in Makefile. Add something like
>>> "-arch sm_50" (or 52). But I doubt it will make much difference and it
>>> will
>>> only affect CUDA formats.
>>>
>>
>> In my system with both an sm_20 and an sm_50 card, when running solely
>> descrypt-opencl (not CUDA), the ptxas info shows that sm_50 is involved in
>> some way.  Is this cosmetic?
>>
>
> OpenCL compiles a suitable (different) kernel for each and you do not have
> to configure anything.
>

What's giving me pause is that without changing anything on either system,
descrypt-opencl is appropriately using sm_20 and sm_50 on my heterogeneous
system, but is only using sm_20 on my GTX750 system.  Previously, the
latter system was happily using sm_52.  I am not sure what changed.


> You can configure CUDA for compiling several archs at once, see "nvcc
> --help". It something like "-gencode arch=compute_20,code=sm_20 -gencode
> arch=compute_50,code=sm_50" (added to NVCC_FLAGS instead of just -arch
> sm_xx). The one most suitable of them will be picked at runtime.


Interesting -- I'll try that.

Royce

Content of type "text/html" skipped

Powered by blists - more mailing lists

Your e-mail address:

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