Re: [Haskell-cafe] Re: [Haskell] Google Summer of Code 2009

On Thu, 12 Feb 2009, Achim Schneider wrote:

Jamie wrote:

...

...

For Theora playback we've found that the largest CPU load comes from colorspace conversion, where the YUV output of the codec needs to be converted to RGB for some targets (like Firefox). That is some fairly straightforward array processing, and would be a good place to start for anyone trying to implement video codecs in Haskell.

That is great idea and a great seed to plant. Wonder if Theora is as good as H.264 in terms of video quality and bandwidth usage?

Theora isn't meant to be a H.264 competitor, but a replacement for flash, wmv and the ilk. I dare to guess that it works decent for low bitrates, especially if you're more interested in detecting shapes than skin pores. I guess you just have to do field tests: encoding video on the fly just isn't what general-purpose CPUs are made for, it's the playing field of processors that take SIMD seriously, i.e. GPUs.

Signers (deaf people or "hearing" people who know sign language) naturally put strong emphasize on smooth video quality (i.e. at least 25 fps with no blurries/fuzzy). I use Mirial Softphone (to me, a best H.323/SIP video softphone so far, run on Windows) and it runs very nicely and perfectly on my Dell 1 GHz Centrind Duo laptop in both H.263 and H.264 (I even tried H.264 at 704x576 at 30 FPS and it works real nice). However Mirial sure did display "CPU overload" message time to time on my Samsung NC10 netbook especially using H.264 and result in pixelization of video frames. So no problem at all with my Dell laptop but kind of borderline on Atom 1.6Ghz netbook. I agree it would surely help offload the CPU work when there is hardware encoder/decoder present in GPU. I see more and more GPU now support H.264 decoder which is nice. Jamie