[Bf-committers] GPU computing

[Erwin Coumans]

One more recommendation to go for OpenCL.

We are working on accelerating pur Bullet physics sdk using OpenCL,  
and I recommend it for cross platform/vendor support. Our work is  
still early stages, but we try to make sure it works on all  
implementations.
Eventually this work should go into Blender too.

Although OpenCL supports task parallelism, it works best when dealing  
with fine grain data parallelism: 1000 to 10000 small kernels  
processing similar work. Some GPU ray tracer for Blender would be  
cool, perhaps to create soft shadows.

How about accelerating screen space post processing?

Good luck with the effort!
Erwin

Sent from my iPhone

On Nov 24, 2009, at 7:19, Charles Wardlaw <cwardlaw at marchentertainment.com 
 > wrote:

>>
>> Animation is a big part of blender, and faster subsurf/armature
>> systems would be very, very helpful.  To be honest it'd be far more
>> helpful then faster physics or rendering systems, and you'd have a
>> much better shot at success.
>>
>
> I would love to see GPGPU-enabled armature calculations... Although,  
> since
> many of the best rigs require custom python scripts I wonder if Python
> wouldn't be a bottleneck there.
>
> On the renderer: you wouldn't have to rewrite everything.  Not every  
> part of
> the rendering process benefits from multiprocessing anyways, just  
> like not
> every part benefits from the various data structures.  But there've  
> been a
> number of nice papers on using GPGPU processing to accelerate the  
> parts of
> rendering that tend to be the most heavy -- ray collisions,  
> subdivision (as
> you said above), ambient occlusion, or even the generation of point  
> clouds
> for other purposes (AO, GI, FG).  If Blender could generate point  
> clouds
> quickly and that data could be accessed and exported, a lot of  
> studios would
> be very interested.
>
> There's also the idea of accelerating nodes in the compositing or  
> texture
> graphs.  And now that sculpt is multithreaded, I wonder how hard it  
> would be
> to get some of the processing offloaded to OpenCL cores.
>
> Then again, hardware-accelerated subdivision is a serious boon.  At  
> work
> we're using Mach Studio, which is a GPU-based real-time rendering  
> system.
> It subdivides on the fly, on the card, and even with a few million  
> polys in
> a scene you have completely interactive turnarounds.  Less so with
> full-scene AO, but it's still usable.  The last version released  
> something
> akin to the DX11 Tessellator functionality on DX10 cards, and  
> watching it
> generate a million polygons from a bump map and a plane is something  
> to
> behold.
>
> ~ C
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> http://lists.blender.org/mailman/listinfo/bf-committers