[Bf-committers] [Bf-blender-cvs] SVN commit: /data/svn/bf-blender [43736] trunk/blender/intern/cycles: Cycles: disable environment importance sampling code for CUDA cards with

Brecht Van Lommel brechtvanlommel at pandora.be
Fri Jan 27 13:56:10 CET 2012


The compiler flag may be related, I couldn't find a similar flag for
4.1 that does the same. With 4.0 there was some sort of limit on the
kernel size past which it would not do some optimizations, this limit
is what the flag modified. I don't know if 4.1 has a similar limit and
if that is the cause of the problem, or if it's something completely
different.

In any case, they switched to a completely different compiler backend,
the idea that application performance would be between 1-2% slower and
10% faster is not reliable information, you just can't make such
assurances when switching to a completely different codebase.

Brecht.

2012/1/27 Miika Hämäläinen <miika.hamalainen at kolumbus.fi>:
> CUDA 4.1 performance is as bad as RC2 here...
>
> I tested CUDA versions 4.1, 4.1 RC2 and 4.0 on same Blender revision.
> (GPU: GF 560ti):
> * v4.1 :    1 min 11 sec
> * v4.1 RC2 :    1 min 11 sec
> * v4.0 :     21 sec
> Performance: 70% slower!
>
> According to "CUDA 4.1 Overview Webinar" video from CUDA website many
> applications got up to 10% better performance and a few applications got
> 1-2% lower performance. In case of any performance drop he suggests to
> file a bug report. Now, if Blender really has 70% lower performance that
> is very suspicious.
>
> Could the reason be a deprecated compiler flag that Blender is still
> using? Like from release notes:
> "* For target architecture sm_2x, a new compiler component 'cicc' is
> used instead of 'nvopencc'. "
>
>
> 26.1.2012 22:10, Thomas Dinges wrote:
>> Downloading that atm... the RC2 of the 4.1 Toolkit was really slower,
>> lets see how the final one does!
>>
>> Am 26.01.2012 21:03, schrieb Antony Riakiotakis:
>>> I see the 4.1 toolkit got released today. Anyone else checked how it
>>> performs and if it can handle the compilation on this one?
>>> _______________________________________________
>>> Bf-committers mailing list
>>> Bf-committers at blender.org
>>> http://lists.blender.org/mailman/listinfo/bf-committers
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> http://lists.blender.org/mailman/listinfo/bf-committers


More information about the Bf-committers mailing list