[libdv-dev] Re: libdv and the open source DV CODEC.

[Charles 'B. K. <kr...@ac...>]

David A Newman <da...@ap...> writes:

> Hello,

> I was researching the availablity of software  DV CODECs when I came
> across   the  Quasar DV  CODEC  on  sourceforge.net   (your email is
> connected  with this project).  Is this  still an active project and
> are  you still involved?   I hope so,  but if  not  I would still be
> interested   in discussing what  you   found  when undertaking  this
> project.

The project is still active, although the majority of effort happened
in the April to June period.  Since then most of the work has come
from a few outside contributors.  (Yay!)  Part of the reason I have
had less time to contribute is that I've gone through a major career
change.  I've recently left OGI (and suspended my Phd) to join a
startup.  We're called Digital Mercury.

> A little background.  I work with Applied Magic
> (www.applied-magic.com) and we make stand-alone real-time video
> editing applicances.  It is a fun business involving a lot of custom
> software and hardware.  We currently uses a hardware DV CODEC, which
> is the one costly part of the system that we have little control
> over (everything else was developed at Applied Magic, OS and
> graphics processors included).  We are currently researching a
> software approach to DV, to find out how much CPU performance is
> required to do real-time DV (and hopefully reduce cost).  I am
> interest for any input you might have, or if you can direct me to
> someone who worked in these areas.

> Thank you.   

> David A Newman
> Chief Technology Officer
> Applied Magic Inc
> (760)931-6417 x358
> da...@ap...
> www.appled-magic.com

That's reasonably easy to answer.  The Canopus software codec was, and
still is my benchmark for what can be done in software.  They can do
real time decode on a 300MHz range PII.  (www.canopuscorp.com)

I think with time, libdv could reach that goal.  But I have a lot of
respect for their level of performance.  I think right now libdv is
around a factor of 2 slower than the Canopus decoder.  Libdv already
includes some use of MMX, some critical sections hand converted to
scalar assembly, and use of graphics card overlay for YUV color
conversion.  Given that, I seriously doubt if you could do much better
than the Canopus people have done on IA32 processors.  

Hope this helps.

-- Buck