AMD Releases Open-Source UVD Video Support

[brosis]

catalyst. (10 Chars)

Thanks!

So why then you suggest

No, to both questions.

when the original question was about radeon(uvd), not catalyst(uvd)?
Isn?t it possible to implement uvd+OpenCL(shaders) now for all codecs on radeon driver?

[curaga]

It's simply not useful or even possible to do OpenCL/shaders for some codecs. They are not parallelizable.

[Nille]

when the original question was about radeon(uvd), not catalyst(uvd)?

Both are using the same Hardware an Firmware. 10bit is not possible with catalyst so i doubt that its possible with the radeon UVD code. The Hardware is not an FPGA with what you can map all possible codecs.

Isn?t it possible to implement uvd+OpenCL(shaders) now for all codecs on radeon driver?

Not maybe impossible but ineffective (a 50$ CPU can handle 1080p high10). Bridgman and Christian K?nig has said something about this Topic in the past.

[brosis]

@Nille, curaga:
So its possible, but developers consider its efficiency to be inadequate.
Which is pretty much blind-sighted, because CPU could do something else. The whole video acceleration is about "offloading". You can?t easily add another "50$" CPU into the system, but you CAN put another GPU into PCIe.

OP should look into HSA, maybe.

[Nille]

You can?t easily add another "50$" CPU into the system, but you CAN put another GPU into PCIe.

And maybe its need an 200$ GPU only for Play an 1080p Stream and consumes 150W/h. Its easier and cheaper to replace the CPU or re-encode the video the asic in the gpu like quiksync or nvidias cuda encoder ( that has nothing todo with cuda its use only the name for marketing and use like everyone else an hardware chip for this )

AMD is for encoding the worst case btw. its impossible for an normal user or developer to get the SDK for use the Hardware encoder in the radeon products.

[brosis]

And maybe its need an 200$ GPU only for Play an 1080p Stream and consumes 150W/h. Its easier and cheaper to replace the CPU or re-encode the video the asic in the gpu like quiksync or nvidias cuda encoder ( that has nothing todo with cuda its use only the name for marketing and use like everyone else an hardware chip for this )

AMD is for encoding the worst case btw. its impossible for an normal user or developer to get the SDK for use the Hardware encoder in the radeon products.

Hardware encoder is only in recent HD7xxx series AFAIK.
Used 57xx+ cost around 50$ now. If PM were right, they would use minimal wattage and there is already OpenCL work done.
I am pretty sure one could easily implement video decode backend for OpenCL stack by replacing existant codecs codepaths and optimizing them.
IMHO its all possible, but I don? t claim anyone is interested in that yet.

[Kano]

That's completely wrong. Better look at:

http://en.wikipedia.org/wiki/Unified_Video_Decoder

It was never emulated via OpenCL. You can do that maybe for very simple codes but that would be never a good idea. xvba crashed very hard in the beginning with my hd 3450, then i got rid of it and got a hd 4550. the hd 5670 has basically a similar uvd part. Only hd 6000+ has got UVD 3.

[Nille]

Used 57xx+ cost around 50$ now.

And why you think its has enough compute capability do decode h264 complete with the ALUs?

If PM were right, they would use minimal wattage and there is already OpenCL work done.

This cards need around 100W/h under Full Hardware usage. And node that UVD is an special ASIC for decoding and has nothing todo with an decoding an video on the ALUs.

I am pretty sure one could easily implement video decode backend for OpenCL stack by replacing existant codecs codepaths and optimizing them.
IMHO its all possible, but I don? t claim anyone is interested in that yet.

Look at VP8 there was an Project that try to Decode VP8 with OpenCL. There is no result.

[curaga]

@Nille, curaga:
So its possible, but developers consider its efficiency to be inadequate.
Which is pretty much blind-sighted, because CPU could do something else. The whole video acceleration is about "offloading". You can?t easily add another "50$" CPU into the system, but you CAN put another GPU into PCIe.

If the codec is not parallelizable (= most of them), the GPU will use its full power and still be _slower_ than your CPU. Would you consider that a worthy offload?

[johnc]

This 10-bit stuff was a dumb idea from the get-go.