theora dev - Oct 2006 - Encoding - CPU usage ...

Hi all.
I would really like to use Theora for the commercial product I work on...
It's a Linux based DVR. Actually we use ffmpeg to produce MPeg4 short 
clips from 4 BNC channels in CIF resolution.
I made a small module for our project that uses Theora... but I see that 
it gets a lot of CPU :(

So I'm here to ask for any kind tips to reduce the CPU usage.
- Hardware... do you think there are better CPUs for Theora? AMD64?
- CPU optimisations... they are enabled by default, right?
- Encoding parameters... now I use:

    ...
    theora_info tTheoraInfo;

    tTheoraInfo.colorspace           = OC_CS_ITU_REC_470BG;     /* 
OC_CS_UNSPECIFIED ?     */
    tTheoraInfo.pixelformat          = OC_PF_420;
    tTheoraInfo.target_bitrate       = 0;
    tTheoraInfo.quality              = 15;

    tTheoraInfo.dropframes_p                 = 0;
    tTheoraInfo.quick_p                      = 1;
    tTheoraInfo.keyframe_auto_p              = 1;
    tTheoraInfo.keyframe_frequency           = 64;
    tTheoraInfo.keyframe_frequency_force     = 64;
    tTheoraInfo.keyframe_data_target_bitrate = 
tTheoraInfo.target_bitrate * 1.5;
    tTheoraInfo.keyframe_auto_threshold      = 80;
    tTheoraInfo.keyframe_mindistance         = 8;
    tTheoraInfo.noise_sensitivity            = 1;
    ...

- Other parameters...? Other ideas...?
- Perhaps some options I could disable that consume CPU to get also less 
quality if it needs...

I'm studying the Theora's examples and ffmpeg2theora to see if I can 
optimize my code... but I don't know libogg / libtheora so I can't 
understand well some functions. Is there an API reference also for 
libogg... ?

Thank you in advance,
-Mat-

On Wed, Oct 04, 2006 at 12:55:58PM +0200, Mat wrote:
> So I'm here to ask for any kind tips to reduce the CPU usage.
> - Hardware... do you think there are better CPUs for Theora? AMD64?
What optimization we have should work with both x86 and x86_64 cpus. You 
do need a fast machine for SD res, but CIF shouldn't be a problem. By
"4
BNC" do you mean capturing 4 separate CIF feeds (4x2.5 Mpixel/second) 
simultanteously?
> - CPU optimisations... they are enabled by default, right?
Should be, if you're using the latest release. alpha6 had a bug where 
they were not. If you 'make debug' from svn it should tell you what
it's
using.
> - Encoding parameters... now I use:
These looks like they're appropriate. Reducing the framerate will help a 
lot, but that may not be acceptable for a DVR.
> I'm studying the Theora's examples and ffmpeg2theora to see if I
can
> optimize my code... but I don't know libogg / libtheora so I can't 
> understand well some functions. Is there an API reference also for 
> libogg... ?
http://xiph.org/ogg/doc/libogg/

It is unlikely to be a bottleneck, however.

HTH,
 -r

On Wed, 4 Oct 2006, Ralph Giles wrote:
> It is unlikely to be a bottleneck, however.
The main bottleneck for theora (live) broadcasts is currently the bitrate
instability with fast motion. I wonder if this *can* be fixed.


Stefan de Konink

Now i write application to capture and encode video stream from V4L2
device to Theora/Ogg file (384x288 25fps). And i have low performance
problems on my athlon processor (2Ghz). Maybe i must change theora to
some faster codec (mpeg in ffmpeg?)


But maybe there is something which i can do to make better performance?
Example, i have code to convert RGB buffer to YUV buffer, but this is
can take lot CPU ticks.
(function ogg_encode_frame in
http://linpvr.googlecode.com/svn/trunk/src/ogg_enc.c )

Any suggestions ???


PS. Maybe someone know about applications like Analog VCR for v4l ?