[XviD-devel] better size prediction in 2pass2

[Radek Czyz]

Hi everyone,

I just made a simple change and I print total bytes taken by textures in 
1st-pass statsfile. What I get seems very very useful, because I my 
observations show that some scenes don't compress at all - which results 
in horrible overflow soon.

I made 1st pass of such scene and look at this stats output:
i 2 1032 0 0 2901 0
p 2 1032 0 0 3529 14
p 2 1032 0 0 3496 1
p 2 1032 0 0 3544 3
p 2 1032 0 0 3500 0
p 2 1032 0 0 3506 0
p 2 1032 0 0 3503 2
p 2 1032 0 0 3508 6
p 2 1032 0 0 3499 3

The last value is bytes taken by textures, the one before is total 
length. Note that this is *exactly* the frames that boosted overflow by 
2kb each time, because the frames were 3200 at quant 8.

Compare with "normal" scene:
i 2 1032 0 0 16021 12798
b 4 0 1032 0 609 144
b 4 0 1032 0 739 189
p 2 1 998 33 11059 9573
b 4 0 1032 0 730 231
b 4 0 1032 0 872 266
p 2 1 1019 12 10788 9306

... almost all bytes are taken by texture.

I know we wanted to keep the old 2-pass code, but imho working on it is 
worth it - our 2-pass makes big scaling errors too often.

What do you think?

Radek