[Ffmpeg-devel] Parallelizing the h.264 Decoder

[Luca Barbato]

Philip Peter wrote:
> 
> Luca Barbato schrieb:
>> Keep in mind that you don't have infinite SPEs and every time you want
>> them doing something different you'll pay some time to feed the new
>> code, more on this later.
>>   
> That is one reason why the decoding of a complete frame on one SPE would
> be impractical.

Right

> I don't have any hard data, but i'm pretty sure, that the code for the
> whole decoding process wouldn't leave much room for actual data to decode.

Right again

> My current approach is to have a few cores dedicated to the cabac
> decoding and a few dedicated to the rest.

keep in in mind that few would be at most <3, say 2...

>   
> I'm currently working on this, but the optimization is a bit complex,
> since i try to minimize branching.
> As soon as i have some results i will post the data.

Great =)

> Is there already optimized CABAC code in the current svn version, or are
> you refering to the CoreAVC decoder?

svn version, check the main files since the x86 asm is in some ifdefs
there...

> I'm also fairly sceptical on how to implement this in the current
> structure.

eh...

> However since the current processor development seems to be
> going towards multiple cores, it might be important to consider this not
> only in the context of h.264, but for the ffmpeg structure in general.

yes, but it must no cause large overhead on uniprocessors =/

lu

-- 

Luca Barbato

Gentoo/linux Gentoo/PPC
http://dev.gentoo.org/~lu_zero