[Ffmpeg-devel] Re: [Ffmpeg-cvslog] CVS: ffmpeg/libavcodec dv.c, 1.72, 1.73

[Roman Shaposhnick]

On Thu, Feb 23, 2006 at 09:21:57PM -0500, Dan Maas wrote:
> >   for(a=2; a==2 || vs_total_ac_bits < size[0]; a+=a){
> >        b = blks;
> >        size[0] = 0;
> 
> I could be wrong, but shouldn't that be
> size[0] = (6*5)*4;
> to account for the 4-bit EOB code in each block?

  Right! Good catch.

> > Are you sure that we need this chunk of code ? My concern is -- in case
> > where encoder maxes out on all QNOs being 0s -- there's very little we
> > can to do salvage the situation anyway. Or did you have another idea 
> > when creating this chunk of code ?
> 
> Actually I like that addition... IMHO it is better to zero out the
> high frequencies rather than run out of space and get essentially
> random/undefined results.
> 
> You can see this by trying to encode pure noise. Most blocks run out
> of space before encoding all the coefficients, which creates some odd
> visual patterns. IMHO blurring out the noise is preferable to creating
> structured artifacts.

  Sounds reasonable. I was just curious that this case doesn't seem to
  be covered by the DV spec which makes it sort of an undefined behaviour.
  But then, again, I only have an SMTPE one, so may be it *is* covered
  by the original Sony's SPEC, so if anybosy knows for sure please check
  it out.

  Otherwise I'm totally fine with this plan.   

Thanks,
Roman.