[FFmpeg-soc] [soc]: r586 - dirac/libavcodec/dirac.c

[Marco Gerards]

Michael Niedermayer <michaelni at gmx.at> writes:

Hi,

> On Thu, Aug 02, 2007 at 04:05:47PM +0200, Marco Gerards wrote:
> [...]
>> >> +
>> >> +    if (s->picture.reference) {
>> >> +        s->refframes[s->refcnt++] = s->picture;
>> >> +    }
>> >
>> > i think refcnt should be checked somewhere so it cannot become too large
>> 
>> Yes.  I am waiting for the next update of the specification.  It will
>> mention the levels that can be used and the amount of reference frames
>> a certain level can have.
>
> a check against REFFRAME_CNT would do IMHO
>
> theres also the issue that frames can be lost due to packet loss or whatever
> or they can be damaged in which case the retirement info for some frames would
> be lost and we have to guess at some point which frame to kill
> if we do this decission immedeatly when the max for the level/profile is
> reached then its more likely that we would choose the wrong frame
> if we wait a little longer then simply removing the oldest frame is more
> likely really a unused frame

Yes, I thought of that too.  According to the specification the oldest
frame has to be removed when the buffer is full.  But I wonder if that
means that this also has to be done in the case a frame was lost, or
specifically in this situation.

What do you propose?  Make the buffer a bit bigger than what the
specification says and if even that buffer is overflowed, just kill
the oldest frames?

--
Marco