[theora] recommendations for web video bit rates

Matthew Gregan kinetik at flim.org
Sun Mar 14 22:00:03 PDT 2010


At 2010-03-14T00:23:11-0500, Denver Gingerich wrote:
> The 100fps figure for theorarm that Gregory mentioned seems to be
> coming from here:
> 
> https://wiki.mozilla.org/Platform/2009-11-24#Layout_Update
> 
> I'd be interested to know where this came from.  Has someone got
> theorarm to compile and run on the N900?  If so, I'd like to try it
> out.

That number came from some decode-only benchmarks I ran when testing
libtheora, theorarm, and leonora (the C64x+ DSP port).  That doesn't include
the time required to do colour conversion, painting, or Vorbis decoding and
playback.  The test player included in theorarm's source doesn't have any
display code suitable for the N900, so some work is necessary to use it for
playback.

I couldn't remember what video I used back then, so I just re-ran the
benchmark on my N900.  Using a copy Big Buck Bunny encoded at ~500kbps from
the 640x360 lossless PNGs at media.xiph.org, testtheora (from theorarm 004)
decodes 14315 frames in 132271ms, which averages 108 frames per second.  For
reference, dump_video -f from the libtheora examples decodes the same video
in 189320ms (76 fps), and using leonora with dump_video gives 187100ms/77
fps (and CPU > 97% idle).

Regarding the Layers stuff for Firefox, I think we'll have something
sufficient to allow fullscreen OpenGL accelerated video playback (with
colour conversion and scaling offloaded to the GPU) on the N900 in the next
couple of months or so.  Based on early testing with a standalone program
using the same decode/display strategy, we should be able to play back
800x480 (i.e. native resolution) videos at 24 fps and still have 50-60% idle
CPU (of which 10-15% is PulseAudio).

Cheers,
-mjg
-- 
Matthew Gregan                     |/
                                  /|                    kinetik at flim.org


More information about the theora mailing list