[NUT-devel] CVS: main/DOCS/tech mpcf.txt,1.145,1.146

[Michael Niedermayer]

Hi

On Sun, Mar 12, 2006 at 05:15:04PM -0500, Rich Felker wrote:
> On Sun, Mar 12, 2006 at 10:53:03PM +0100, Michael Niedermayer wrote:
> > > > @@ -106,7 +106,7 @@
> > > >  t (v coded universal timestamp)
> > > >      tmp                                 v
> > > >      stream_id= tmp % stream_count
> > > > -    value= (tmp / stream_count) * stream[ stream_id ].timebase
> > > > +    value= (tmp / stream_count) * timebase[stream_id]
> > > 
> > > BTW an idea... instead of having timebase stored in the stream header,
> > > what if we stored a list of timebases in the main header, then the
> > > stream header just had an index into that list? That was if we had 20
> > > different streams but only 2 distinct timebases, the universal
> > > timestamps would still be compact. It would also allow additional
> > > non-stream timebases to be stored for use in chapter extents.
> > 
> > ok
> 
> Hmm, one thing that needs consideration though is if there's a limit
> on the number of timebases. Up until now we never worried about limits
> on the number of streams since we assumed a demuxer could just ignore
> streams it's not interested in. However due to universal timestamp,
> the demuxer needs to be aware of all timebases in the file. Should we
> address this, or just assume that a super-limited demuxer can reread
> the list from the main header if it needs to?
> 
> Sorry for bringing up a nasty issue -- I just realized it.

the problem was already there before the "universal timestamps" as they are
just a cosmetical change, syncpoints already contained stream+timestamp in
stream timebase ...

id at least add a note like
muxers should have mercy with weak demuxers and not create more timebases
then needed
muxers should not round timestamps, but store exact ones
there MUST not be 2 identical timebases in a file

it doesnt solve the problem but should reduce it at least, we could
certainly solve it but iam not sure if its a good idea, i rather tend
toward that a complete fix would have too many dissadvantages

for example one way to fix it is to:
universal_timebase_count MUST be < CONSTANT

universal_timestamp
    tmp                     v
    stream_id= tmp % (universal_timebase_count+1)
    tmp /= (universal_timebase_count+1)
    if(stream_id < universal_timebase_count)
        timestamp= tmp * timebase[ stream_id ]
    else{
        den                 v
        timestamp= tmp/den
    }

or something similar

[...]
-- 
Michael