[Bf-committers] OpenGL accelerated Nodes

[Brecht Van Lommel]

Hi,

There may be differences in the results depending on the graphics 
card, but I don't think that is a big issue, rather something you 
would fix as problems come up. It will mostly show in corner cases 
anyway, like very small or big numbers, inf/nan, hard to test in advance.

A tricky thing about accelerating a few nodes is that it won't 
necessarily make things faster, since moving data to and from the GPU 
has a lot of latency, by the time you send something to the GPU, have 
it compute, and get it back it may be much faster to just do it on the 
CPU. So you need a good integration with the existing node execution 
system to make sure that there is the least GPU-CPU transfer possible, 
and that as much as possible stays on the GPU.

Not sure which nodes would be desired to have on the GPU, I guess 
people would just want nodes in general to run faster. I think rgb 
curves, mix and blur are used most often. In my experience, most of 
the work would be coding the system, if that is done well, adding more 
nodes should not be much work.

Brecht.

Bob Holcomb wrote:
> My 2 cents:
> 
> I was under the impression that the differences between floating point 
> numbers representation on most graphics cards and on the cpu was 
> different enough to cause artifacts in rendering?  Would this only be 
> for preview or for actual rendering?  I do know that some of the latest 
> video cards (my nvidia 8800 GTX for example) does support the same IEEE 
> 754 floating point standard, for representation, however I'm not sure 
> that the process of the mathematics is exactly the same, so it may 
> return slightly different results (I could run some quick tests for 
> those interested).  If there is a desire for accelerated nodes, which 
> ones specifically?  Some will be harder than others to put on GPU due to 
> instruction limitations (again, depending on GPU).  This would be a good 
> use case for a plugin-in system, which should encompass more than just 
> nodes.
> 
> Cheers,
> 
> Bob
> _______________________________________________
> Bf-committers mailing list
> Bf-committers at blender.org
> http://lists.blender.org/mailman/listinfo/bf-committers
> 
>