[Open-graphics] decoding video in real-time

[Nicolas Boulay]

2007/4/17, Attila Kinali <attila at kinali.ch>:
> On Mon, 16 Apr 2007 19:26:15 +0200
> "Nicolas Boulay" <nicolas.boulay at gmail.com> wrote:
>
> > > No, the only reasonable way to do that is to optimize the hardware
> > > for this specific task.
> >
> > I think about a generic cpu to avoid the need to create a good
> > compiler. But you could always add an idct in hardware and the need
> > for a powerfull cpu vanished at the cost of few modification in the
> > codec used (motion compensation need power too ?). Beside that, all of
> > this is very close to all the graphical ressources.
>
> I don't fully understand what you mean here. Could
> you explain it a little bit?

Few years ago, i had profiled ffmpeg. 98% of the cpu was used by the iDCT.

I know that things change a lot for HD, motion compensation need now a
lot of cpu too.

Such iDCT are a 100 ASM line function. It's very localised. So it's
very easy to track and modify a codec to use an hardware accelerator
or a specific taylord instruction.

So for the 2% of generic work you use a generic cpu and for the iDCT a
specific core or a new instruction (new instruction or tightly
connected coprocesseur that share registers with the cpu are great to
minimise communication latency)

> I leave out the rest of your mail, as i don't have any
> knowledge about 3D programming at all.
>
>                         Attila Kinali
>
> --
> Linux ist... wenn man einfache Dinge auch mit einer kryptischen
> post-fix Sprache loesen kann
>                         -- Daniel Hottinger
>