[FFmpeg-devel] inverse weighitng tables for DV seems to be wrong

[Michael Niedermayer]

On Thu, Feb 26, 2009 at 10:30:48AM -0800, Roman V. Shaposhnik wrote:
> Guys,
> 
> while working on a particular discrepancy in DV codec, I've noticed
> that the static inverse weighting tables in DV codec seem to be
> slightly wrong. Here's how I verified it:
> 
>     for (i=1; i<64; i++) {
>         for (level = 0; level < 8192; level++) {
>             int coeff  = (level * dv_weight_88[i]  + (1 << (dv_weight_bits))) >> (dv_weight_bits+1);
>             int nlevel = (2*coeff * dv_iweight_88[i] + (1 << (dv_iweight_bits - 1))) >> dv_iweight_bits;
>             if (level != nlevel)
>                 printf("%03d %d %d %d [%d]\n", FABS(level-nlevel), level, coeff, nlevel, i);
>         }
>     }
> 
> and here's what I've got:
>    
> $ ./test | cut -f1 -d\  | sort -n | uniq -c | awk '{print $0; total+=$1;} END {print total}'
>  182700 001
>   82533 002
>   48647 003
>   30470 004
>   22314 005
>   16054 006
>    9311 007
>    4336 008
>    1557 009
>    1198 010
>    1188 011
>    1142 012
>     854 013
>     516 014
>     190 015
> 403010
> 
> So as you can see we get errors as high as 15. 
> 
> Now comes the question. The best way I know how to fix the situation
> is to get rid of one of the tables altogether and generate it on
> the fly. At least that way ffmpeg itself will be as close to preserving
> material in multiple generations as possible. Now, here's the code:
> 
>     for (i=0; i<64; i++)
>         dv_iweight_88[i] = ((uint64_t)8589934592/dv_weight_88[i] + 1)>>1;
> 
> and here's the result of my verification step:
> 
> $ ./test | cut -f1 -d\  | sort -n | uniq -c
>  309105 001
>    9421 002
> 318526
> 
> As you can see -- almost all of the errors are now in 1 range. 
> 
> But! Is there a way we can do better?
> 
> My division-foo is weak today, so any kind of suggestions will be appreciated.

the best table is not neccessarily the one that is closest to the inverse
but the one that generates the best PSNR at a specific bitrate

our tiny_psnr tool prints the PSNR and dv is constant bitrate so it should not
be hard to find a good table by +-1 of one coeff and trial and error

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

I am the wisest man alive, for I know one thing, and that is that I know
nothing. -- Socrates
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20090226/6632f36d/attachment.pgp>