Vorbis - Sep 2006 - ov_open_callbacks() performance & memory requirements

Hi guys,

I'm writing a sample playback engine using vorbisfile, but I've hit a
performance problem - hoping someone here can help!

I'm using ov_open_callbacks() to decode memory-based ogg files. It works
OK, but when I call ov_open_callbacks() approximately 300k of memory is
used (presumably to initialise the decoder). It also takes some time to
execute ov_open_callbacks() - the read & seek functions are being called
a huge number of times, even for a tiny 10k ogg file.

My sample engine needs to load a large number of small audio files. This
gives me two problems - first, it's taking too long to initialise all
those files with ov_open_callbacks(), and secondly I'm running out of
memory! For example, 100 x 10k files needs a thumping 29MB of memory.

Is there anything I can do to speed up ov_open_callbacks() or reduce
it's memory requirement?

Many thanks,
Dave.

On 9/14/06, Dave Brown <dave@db-audioware.com>
wrote:
> Hi guys,
>
> I'm writing a sample playback engine using vorbisfile, but I've hit
a
> performance problem - hoping someone here can help!
>
> I'm using ov_open_callbacks() to decode memory-based ogg files. It
works
> OK, but when I call ov_open_callbacks() approximately 300k of memory is
> used (presumably to initialise the decoder). It also takes some time to
> execute ov_open_callbacks() - the read & seek functions are being
called
> a huge number of times, even for a tiny 10k ogg file.
If you have very small files, perhaps you're not interested in being
able to seek around them. Returning failure from your seek callback is
fine, and will disable seeking; that'll probably speed things up
somewhat.

As for memory use: not much you can do with that implementation, the
libvorbis decoder is simply fairly memory-intensive. You could look at
changing to use the tremor-lowmem vorbis decoder, which is
specifically designed to minimise memory use.

Mike

On Thu, Sep 14, 2006 at 03:55:05PM +0100, Dave Brown wrote:
> My sample engine needs to load a large number of small audio files. This
> gives me two problems - first, it's taking too long to initialise all
> those files with ov_open_callbacks(), and secondly I'm running out of
> memory! For example, 100 x 10k files needs a thumping 29MB of memory.
Two other things you can try, in addition to Mike's suggestions: It's 
legal to concatenate ogg files, and libvorbisfile will handle that, so 
you can try concatenating the encoded data and only having one decoder 
context. This should reduce your memory footprint significantly, but 
will add the cpu/latency cost of seeking when you switch clips within a 
decoder context.

Another thing you can try to do is concatentate the *input* files and 
compress them as a single stream (perhaps with some padding at the 
joins). Then you have a single logical ogg stream. Use a custom cue 
table to seek in that for clip selection. This is save the overhead of 
the chain mapping as well as reducing the number of decoder contexts.

These both assume seeking, of course, but are both also applicable to 
using the Tremor decoder Mike mentioned.

HTH,
 -r