CUDA/OpenCL h.264 DirectShow encoders?

[easy2Bcheesy]

Hi,

Just wondering if there are any CUDA or OpenCL DirectShow encoders out there? I know the quality might be somewhat suspect but I'm interested to know if 720p30 real-time h.264 encoding would be possible using NVIDIA ION or ION 2...

I'd be looking to overcome the quality issue by throwing a lot of bandwidth at the encode, but right now the problem seems to be that there I can't find any DirectShow encoders.

[Sharktooth]

none that i know

[hust_xcl]

try arcsoft mediaconventer. i know it supports cuda. but i am not sure it supports directshow.
btw, it is not free.

[AJ73]

The CUDA SDK 3.2RC includes the CUDA video encoder for H.264 and VC-1. The encoder has limited visibility. It is not visible in the filter graph manager. You can use the filter only by programming a graph via C/C++.

The quality of the encoder can not compete with e.g. x264 but it is very fast (depending on the GPU and the number of shaders). I am able to get > 200fps for a 720x576 video.

The SDK includes not only the DirectShow filter but also the C API for direct access to the encoder.

[Didée]

The depressing result CUDA gave me lately: http://forum.doom9.org/showthread.php?p=1444371#post1444371

[easy2Bcheesy]

@AJ73 - what GPU were you using that could get 720x576 encoding at 200fps? I'm curious about the application of ION or ION2.

[Dark Shikari]

The CUDA SDK 3.2RC includes the CUDA video encoder for H.264 and VC-1. The encoder has limited visibility. It is not visible in the filter graph manager. You can use the filter only by programming a graph via C/C++.

The quality of the encoder can not compete with e.g. x264 but it is very fast (depending on the GPU and the number of shaders). I am able to get > 200fps for a 720x576 video.Only 200fps? Assuming that's on a top-end GPU, that seems rather slow; x264 on my 1.6Ghz Core i7 gets 321fps on "superfast" and 529fps on "ultrafast" on a raw 720x576 input.

[Selur]

(did some tests before adding cuda.exe to Hybrid and posted some benchmarks with the Elephants Dream DVD and posted them in the german doom9 forum: http://forum.gleitz.info/showthread.php?42660-Graka-zum-Encoden-mit-x264-bzw.-Staxrip)
Here both CUDA&x264 deliver around 600fps on my system (i7 875k, 8GB RAM, Windows 7 64bit pro, NVIDIA GTX 470) for DVD input decoded by mencoder and piped to the encoder,... but x264s output looks better ;)
-> I agree 200fps seems a bit slow for DVD input

[easy2Bcheesy]

I'd imagine that the 16 CUDA cores in ION would make for pretty pathetic performance at 720p then.

[kidjan]

I think you're right (http://www.anandtech.com/show/2878/9). Seems like ION really isn't an encoding platform, not enough CUDA cores...

[Sharktooth]

@easy2Bcheesy: i think AMD Bobcat 8 or 18Watts processors (including a powerfull DX11 integrated video core) may be what you need. Products based on them will hit the stores just before xmas. however you need a h.264 encoder based on OpenCL or StreamSDK.

[popper]

when, and what do we expect sandy bridge with its internal Encode/Decode engine ASIC to provide with the high level Intel® Media SDK 2.0, and at what watt, speed, and cost, BTW, any guesses given http://doom10.org/index.php?topic=717.0

i dont really want to be looking at AMD anymore, due to their very poor unusable UVD and the mass of false AMD/ATI 'PR innovation' that went into that since the original X1550 and its original UVD ASIC, all the way upto today's UVD 3.0 (HD6870 etc), and still no real way to use it with bugs in the drivers.

apparently so bad and unresponsive where AMD ,that even beauche the xvba splitted-desktop UVD NDA Signed Dev got bored after telling them about all the bug's he found while developing the xvba framework and them not fixing even simple ones.

[easy2Bcheesy]

I'm unclear what is actually being discussed in the Doom10 thread - is Sandy Bridge going to accelerate x264 or is it simply going to provide a superior architecture for the existing code?

[popper]

I'm unclear what is actually being discussed in the Doom10 thread - is Sandy Bridge going to accelerate x264 or is it simply going to provide a superior architecture for the existing code?

its yet to be proven OC with a real code drop at least to DS etc and actual access to a real sandy bridge in some way to test , but drwho? (aka Francois Piednoel , Senior Performance analyst at Intel) says he's beta patched a current x264 master code-base, to give it a lower level access than the new higher level Intel® Media SDK 2.0 has to the new sandy bridge internal Encode/Decode engine and open up some internal routines for direct x264 use as he understands x264 assembly etc and that "the x264 community would not include a call to a DLL without any control of the encoding process."

he also said in IRC about the x264 patches he's working on right now.

"2010-10-18 05:58:07 < Dark_Shikari> a) What exactly is this hardware?

2010-10-18 05:58:12 < Dark_Shikari> Is it a single encoding ASIC that takes YUV in?
2010-10-18 05:58:27 < Dark_Shikari> Is it a set of asynchronous units, like on the OMAP4?
2010-10-18 05:58:35 < frpiednoel> it does take YUV in ... and few other formats
2010-10-18 05:58:40 < Dark_Shikari> Is it a set of synchronous units, accessible via CPU instructions?
2010-10-18 05:59:18 < frpiednoel> for the moment, we are only supporting windows ... and via DXVA2
2010-10-18 05:59:40 < frpiednoel> I am the one who is trying to get it out as "hardware" ...
"
"
you ll get access to the motion vector field, you ll be able to use Quanti, inverse quanti, hard Zig Zag, and reserve ZigZag
2010-10-18 06:02:39 full hard decoder
2010-10-18 06:02:56 < Dark_Shikari> can it search an arbitrary number of refs?
2010-10-18 06:02:58 < frpiednoel> and the decoder decode in the L3 cache" so far.