Blender 2.67 Nears With New Features
Phoronix: Blender 2.67 Nears With New Features
A new release candidate build of version 2.67 for the Blender modelling software is now available. The Blender 2.67 update will be introducing several new features...
This is a pretty nice release, as far as new features. SSS has been sorely needed in Cycles, though I still only use Cycles on a limited basis at the moment because I made the terrible mistake of buying an AMD card just before Cycles was released. So I don't get any of the benefits that come with using Cycles with Cuda and I think I may die of old age before the Khronos group gets OpenCL sorted for use with Cycles.
I am much more excited about Freestyle finally making it in though, been following the development of that for years now. In many situations I prefer an NPR style and Freestyle makes that a lot easier.
Except that Nvidia's OpenCL implementation couldn't compile Cycles shaders either. The problem is with the OpenCL spec in this case. AMD can't do much about it without breaking compatibility with the OpenCL spec, which no one wants. Until an updated version of OpenCL is made available that can handle the Cycles use case we won't be seeing any useful work on this front.
Originally Posted by Braffe
[Edit: I looked into it some more, apparently OpenCL for both Nvidia and AMD cards can compile the Cycles kernel if you have enough memory (and time) available. Nvidia requires a fair bit less at about 6.5+ GB of RAM while AMD needs over 31 GB. Part of that is hardware differences where AMD cards inline everything, but neither case is really acceptable, which just comes back to how OpenCL is specified in the first place.]
Last edited by IanS; 04-28-2013 at 06:41 AM.
From my ignorant point of view... how come there are many commercial programs (Photoshop, password crackers, video editors, etc) that use OpenCL with no issues ? Is it really a limitation of OpenCL, or a lack of knowledge by the main Cycles programmer (Brecht van Lommel) ?
Originally Posted by IanS
Haha, AMD and Nvidia should really send their blob compiler guys to some "how to do compilers (and software!) right" courses.
The problem is the AMD OpenCL implementation. It seems like the compiler will inline all function calls, which makes running any somewhat more complex kernels like Cycles impossible. AMD are aware of the issue and working to fix it afaik.
Originally Posted by wargames