OpenCL 1.0 Specification Released!

[StefanHamminga]

Thanks for clearing that up Drag! I was wondering because I suppose their goal is to enable many projects to make use of the GPU, but to me that only makes sense when you can actually efficiently share the GPU amongst processes.

Another question: Could OpenCL implementations replace (most of) the 'proprietary' video offloading extensions like UVD and PureVideo?

[drag]

Another question: Could OpenCL implementations replace (most of) the 'proprietary' video offloading extensions like UVD and PureVideo?

I hope so, but I don't know. The comparison is between two APIs where one fairly generic vs having something that is specialized for video.

So I suppose, unless the OpenCL stack is done very well, that the UVD/Purevideo will always have a edge on performance. But with OpenCL it's going to be much easier to deal with and you should see substantial gains versus just CPU-based processing.

Plus I expect that those video-specific APIs are geared towards very specific media types.. like h264 or whatever. Were as OpenCL could be applied to most anything, including things like Dirac. I really don't know all the details or whatnot.. not a programmer. So it's just guessing.

[Zhick]

Another question: Could OpenCL implementations replace (most of) the 'proprietary' video offloading extensions like UVD and PureVideo?

My guess is that Video-acceleration will be implemented directly on top of gallium3d (or better: already is implemented), at least for the free drivers. This will propably also be faster, since OpenCl for the free drivers (I suspect) will also be implemented on top of gallium3d. And it doesn't take a X developer to figure out that Hardware->Gallium3d->VideoDecoding will most likely be faster than Hardware->Gallium3d->OpenCl->VideoDecoding.
And the proprietary drivers will most likely naturally continue to use their proprietary extensions...

[marcheu]

When doing video over gallium directly, we also gain the ability to add in specific video decoding pipeline stages. So yeah, it makes more sense that way.

As for doing OpenCL on top of gallium, it seems possible, but I'm not 100% sure yet, I have only read parts of the specification.

[deanjo]

FYI anandtech has a interesting write up on openCL.

http://www.anandtech.com/video/showdoc.aspx?i=3488&p=1