How many GFLOPs are actually in use to decode hd level VC-1/mpeg-4?

[nAo]

But only on predicate situations (like this)?

No, I was not even thinking about predicates..

 

[pc999]

No, I was not even thinking about predicates..

Thanks, that is very interesting, so I guess it is only tied to dev effort, I hope is not very hard to do a good job with this.

 

[Mariner]

The question is, how efficient can software decoders become?

CoreAVC claim their codec has the following requirements:

CoreAVC™ for Windows @ 1080p video at 24-30 frames per second
2.8 GHz Pentium 4 or faster processor
At least 1GB of RAM
256MB or greater video card
Windows 98, 2000 or XP

CoreAVC™ for Windows @ 720p video at 24-30 frames per second
2.2 GHz Pentium 4 or faster processor
At least 512MB of RAM
128MB or greater video card
Windows 98, 2000 or XP


I understand that hardware requirements are a lot lower than other available codecs. A quick scan of their page says that they intend to support hardware acceleration in the future so the figures above are software only.

 

[randycat99]

Thank you, all, for replies to my topic.

I guess I learned something today about video decoding not being FLOPs bound! Thanks for the info. I am intrigued by the scale and type of processing that these mpeg-4 (and variants) are demanding to pull off the highest level hd video. So this bears the question, is it that these codecs genuinely demand tremendous processing resources to do the job, or is it the architecture of today's forefront processors (P4, G5, A64, Centrino) are not ideally equipped to do the kind of processing required (hence they are running hard with what they got, but not necessarily running efficiently)?

What sort of hardware conjuring would be a good candidate for mpeg-4 acceleration? Is it a return to sheer amounts of integer processing? An increase-fold of branching capacity with superscaler width? Does the parallelism of SIMD configurations help (assuming integer operations are similarly supported), or has this regressed back to a sequential nature? Can you just keep throwing more cores at the job?

Is there also a component that these codecs are still in an early implementation state with lots of optimization still left on the table to enable the same job to be done using increasingly less actual computational resources?

 

[Simon F]

The SPE's ability to perfectly prefetch with no hit at all the correct instructions in advance via branch hints might be a big win, even over complex OOOE cores.

Errr.. how is that possible? Surely that would contradict some theory or another?

 

[DemoCoder]

VC-1 runs much faster than MPEG-4/AVC.

 

[Fafalada]

But only on predicate situations (like this)?

Predication is used to replace branching where applicable, nAo was talking about speed of actual branching.

Errr.. how is that possible? Surely that would contradict some theory or another?

Why?
Fact is that dynamic hints have the benefit of domain knowledge no hw predictor could ever dream to have. In some situations that means penalty-free branching.

CoreAVC claim their codec has the following requirements:

While we're just throwing random unrelated implementation requirements around, I could mention PSP decodes 480P H.264 streams with time to spare on a paltry 222mhz R4000.

 

[deathkiller]

VC-1 runs much faster than MPEG-4/AVC.

Not in my PC, I hate that game trailers come in 720p60 wmv(VC-1), I don't have problems with 720p24-30 in any format.

 

[mestaice]

Interesting thread.

I found some IBM benchs :

IMO, it's more about the functions than brute power. In fact Terranex use SIMD for their restauration machines. I think the software will be very important like usual..


ps: sorry for english

 

[pc999]

Predication is used to replace branching where applicable, nAo was talking about speed of actual branching.


Why?
Fact is that dynamic hints have the benefit of domain knowledge no hw predictor could ever dream to have. In some situations that means penalty-free branching.

This is very interesting as must people (me too) thought it will be otherwise but it seems that it will be a hell of a work for the dev.

 

[cho]

the idct should(can) be do by the FPU .

 

[Simon F]

Why?
Fact is that dynamic hints have the benefit of domain knowledge no hw predictor could ever dream to have. In some situations that means penalty-free branching.

"In some situations"... that makes a difference. nAo (?) had originally said something like "always" and I'm sure that would probably be equivalent to solving the halting problem!

While we're just throwing random unrelated implementation requirements around, I could mention PSP decodes 480P H.264 streams with time to spare on a paltry 222mhz R4000.

There's H.264 and there's H.246... there are so many options. Now if you instead had said, say, HD resolution with 50Mb/s input and all the bells and whistles, then you need something a little beefier!

 

[nAo]

"In some situations"... that makes a difference. nAo (?) had originally said something like "always" and I'm sure that would probably be equivalent to solving the halting problem!

Maybe my english was more crappy than ever yesterday but I did not mean/imply/write that SPEs can *always* do that

 

[Mmmkay]

There's H.264 and there's H.246... there are so many options. Now if you instead had said, say, HD resolution with 50Mb/s input and all the bells and whistles, then you need something a little beefier!

It supports up to H.264 AVC MP@L3. The resolution is 720x480 and has a max bitrate of 10Mbps.

 

[mestaice]

There's H.264 and there's H.246... there are so many options. Now if you instead had said, say, HD resolution with 50Mb/s input and all the bells and whistles, then you need something a little beefier!

Do you mean 4:4:4 profile?

What kind of quality can we expect with Blu-ray on Ps3 compare to a dedicated player (i mean the first available). Depend only on their software? Sony claim a high quality with cell decoding and RSX processing. That's finally THE question..?

 

[Fruitfrenzy]

It supports up to H.264 AVC MP@L3. The resolution is 720x480 and has a max bitrate of 10Mbps.

Are you sure about that? The screen resolution on a PSP is way below 720x480 and it seemed to support quite a restricted set of options when you are encoding video yourself for memorystick playback.

 

[Mmmkay]

Are you sure about that? The screen resolution on a PSP is way below 720x480 and it seemed to support quite a restricted set of options when you are encoding video yourself for memorystick playback.

Yes I'm quite sure of that. People have ripped the contents of UMD movies to find that the .mps files have a resolution of 720x480.

Memory stick playback is completely different to what is supported by UMD's. Originally the PSP could only playback Mpeg-4 ASP content, but that subsequently got upgraded to AVC in a later firmware. Even so, it's still more restrictive than UMD's support for MP@L3.

[edit]It looks like memory stick playback is restricted to AVC MP@L1.3

 

[Titanio]

Maybe my english was more crappy than ever yesterday but I did not mean/imply/write that SPEs can *always* do that

Your "might be a win" had you covered I didn't read that to mean in every case it'd be that way, at all.

 

[Bohdy]

I've always assumed that the PSP had dedicated AVC decoding logic to be able to decode UMD's.

 

[NANOTEC]

I've always assumed that the PSP had dedicated AVC decoding logic to be able to decode UMD's.

Same here.