DXVA + VMR9 vs. MadVR vs. CUDA

Let's discuss the best method for GPU-accelerated H264 decoding

Sure. OpenCL. I'm afraid CUDA is obsolete from a software point of view. Both nVidia and AMD have exposed APIs for direct access to their video decoding hardware via OpenCL.

>>1,2
And for reference:

http://developer.amd.com/gpu/AMDAPPSDK/assets/OpenVideo_Decode_API.PDF

http://developer.download.nvidia.com/compute/cuda/sdk/website/C/src/cudaDecodeGL/doc/nvcuvid.pdf

Remember, OpenCL > DirectCompute > CUDA.

>>3
The other thing cool about OpenCL is that there is an OpenGL interop layer, so you can display the results of a decoded video frame directly to an output adapter, or use it as a texture to be rendered to a set of polygons, all on the GPU without having to copy anything to the system's main memory.

See: http://www.khronos.org/registry/cl/extensions/khr/cl_khr_gl_sharing.txt

Meanwhile, at JCT-VC:

Current indications are that the new standard could provide 2x better video compression performance (i.e. around half the bitrate for a similar quality level) at the expense of significantly higher computational complexity, compared with H.264/AVC.

http://www.vcodex.com/h265.html

I'm not sure how I feel about this. I mean, 2TB HDDs are around $70, but even mid-range CPUs and GPUs can cost hundreds of dollars each. Is this really moving in the right direction?

Yay, more patented shit. Fuck you, Americans.

>>3
Remember, OpenCL > DirectCompute > CUDA.

Why do you say that? I had the impression that CUDA was better able to take advantage of the hardware while OpenCL's only real advantage was portability across cards. Given that CUDA support has been extended to older cards and OpenCL apparently hasn't portability is hardly meaningful.

Ok, forget that nonsense. Is there anything about OpenCL that I'm missing? It seems like the technological chump here.

>>7
OpenCL works on DX10 hardware. OpenCL is an industry standard that works on AMD, Intel, nVidia, Imagination Tech (PowerVR), and various other GPUs, as well as CPUs due to it's heterogeneous nature. You can run OpenCL compute kernels easily on multiple GPU and CPU cores simultaneously on the same system. CUDA can't do this. I'd say OpenCL takes better advantage of hardware.

CUDA is to OpenCL as Glide was to OpenGL. And we all know what happened to Glide at the turn of the century. Proprietary standards never win out.

nVidia
PowerVR
Bring them closed source drivers.

>>7,9
>Imagination Tech (PowerVR)
PowerVR GPUs are the GPUs powering 90% of mobile devices. The OpenCL Embeddeed Profile capable PowerVR equipped mobile phones are coming out later this year, which includes the iPhone 5. OpenCL will pretty much become the defacto compute API.

Also, there's a big movement currently underway among game developers to move away from 3D-specific APIs like Direct3D and OpenGL and to build custom ``software'' renderers using compute languages, like OpenCL and DirectCompute. For example, DICE is using DirectCompute to implement Battlefield 3's tile-based fully-deferred shading renderer.

No one is bothering with CUDA becomes it's proprietary and only works on nVidia GPUS, and AMD's HD5000/HD6000 are the most popular DX11 GPUs currently (they have 75% of the DX11 market).

http://www.qotpa.com/2011/03/25/is-this-the-beginning-of-the-end-for-directx/

>>9
If you're interested AMD publishes all of the information for their newer GPUs including register mappings, shader instruction sets and so on. Unfortunately, the X.org developers are too slow, retarded and backwards to do anything with it.

>>11
I wasn't complaining about AMD, I was complaining about nVidia and PowerVR.

>>11
It ain't easy writing GFX drivers and microcode. Such effort extends over multiple years. I know nothing about this area so I can't really contribute effort here.

Anyone who thinks CUDA is better than OpenCL has fallen for nVidia's marketing propaganda and false advertisements.

>>8,10
So... you're saying portability is the reason? That doesn't really work for me. Intel is not even worth considering, and as nVidia supports CUDA on much *older* cards at this time I wouldn't bet against CUDA being supported on more existing cards than OpenCL. Meanwhile my fairly recent AMD/ATI mobile card doesn't support OpenCL so the portability argument just doesn't register with me.

What I am concerned with is capability. CUDA seems to have lower level capabilities and in theory [C for CUDA/etc] could be optimized better as a result. The way OpenCL works... it just doesn't seem to have the ability to take full advantage of the hardware. It's been a while since I've read over the documentation, but I got the impression it was obviously suboptimal.

Anyway I am not personally concerned with how id Software optimizes Rage or whatever. It makes sense for them to use the common groung--people will buy AMD/ATI cards no matter what I think of their drivers. I'm looking at this for my own personal use, more something like FASTRA who is probably using nVidia primarily because they were the only sane choice at the time for GPGPU, but possibly they continue to go that way partly because CUDA has a greater performance potential than OpenCL. That's what I'm trying to investigate.

>>9,14
Please stop acting like a mentally challenged individual.