[Bf-committers] Tiled Compositing Render: Progress?

Jeroen Bakker j.bakker at atmind.nl
Sat Dec 11 16:14:05 CET 2010

On 12/11/2010 09:25 AM, Troy Sobotka wrote:
> I know in the past I have been aware of more than a few folks that are
> quite excited about the possibility of GPU enhanced calculations, but
> when talking with industry folks, the concerns are with consistency of
> delivery.
> Does OpenCL suffer from the same inconsistencies?
There is a real difference between GPU enhanced and OpenCL or Cuda for 
that hand. The later two will create identical results as a CPU based. 
We tested this already for some nodes and they are really binary 
> Am I to assume that tiled software driven rendering is to happen
> alongside of this so that the artist pipeline is able to harness the
> power of GPU calculations for more immediate feedback of looks and
> then software based for final consistency of rendering across all
> platforms?
Tiled based compositing is also needed for OpenCL. In an OpenCL 
situation you have devices with fixed memory. In complex compositing 
systems you will need to find a way to divide the work in smaller pieces 
to fit in this memory. IMO you have two scenarios:

a. the artist is working on the system. and should have direct feedback 
on that what he is doing
  - only calculate pixels that are visible (currently all pixels are 
  - usage of OpenCL to make better usage of available processors

b. the artist is rendering the 'final' image.
  - here all pixels are calculated.
  - but still OpenCL can be used to decrease rendertimes.

But always both can create the same (visual) result.
Not all renderfarms and user systems bennefit from OpenCL, so there 
always need to be a CPU fallback.

>> 3. Getting communication, planning and finance aligned. When these are
>> controlled, I have more time to work on the compositor and the faster
>> the realization can be done.
> I'd think that tile based compositor software rendering is very high
> up most people's priority list. As such, group sourced funding might
> be a feasible approach.
> Thanks for taking the time to respond,
Thanks for the tip!
Jeroen Bakker.

More information about the Bf-committers mailing list