[XviD-devel] HUGE ME regressions between beta1 and beta2

Michael Militzer michael at xvid.org
Sat Nov 26 15:03:46 CET 2005


Quoting Radek Czyz <radoslaw at syskin.cjb.net>:

> Now you see how strange it is :)
> 
> It would appear that we've got an impressive quality boost from your 
> patch. Initially you only saw that with higher quants, however *if* I 
> make all tables linear (same formula for low and high quants) then 
> there's a detectable gain in lower quants too.
> 
> I tweaked 8x8 lambda multiplier a bit and got some more.
> 
> If we keep this, I'll be happy to sacrifice some speed.
> 
> What I said above is true for no bframes.
> Example I just did, quant 3, defaults, no bframes:
> old code, size 43,010,048, psnr 44.1543, time 2:03
> new code, size 42,905,600, psnr 44.1544, time 2:07
> 
> Funny, heh? It searches more and actually finds more. I'm happy with that.

Good. That fits my findings. So seems the 8x8/16x16 decision wasn't working
perfectly for DVD input and VHQ=0. Oh btw: I had similar gains with the
pure linear table (with which I started) but then decided to trade the bit
reduction for a smaller lambda - hoping that the perceived quality increase
would be worth the 0.25-0.5% bitrate loss. But not sure here of course - you
need a lot of testing with your own eyes to decide what's actually better.
On the other hand: either way the loss wouldn't be huge...

> With bframes, I get this:
> 
> old code: size 35,866,624, psnr 43.1332, time 2:01
> new code: size 35,934,208, psnr 43.1822, time 1:52
> 
> So new and old are equivalent for psnr (more or less) with a rather 
> major slowdown. Whatever new p-frames got us, b-frames wasted.

Hm, assuming your times are minutes:seconds then it looks as if the new 
code would be faster?? ;)
 
> I'll continue working on it now.
> 
> If I "fix" bframes, then I will highly recommend a simple patch that 
> changes 16x16 table to be linear. It looks like an unexpected win.

Well, the gains at high QPs are _huge_ (>2 dB) so even if XviD would run
100% slower here it would still be worth it. On low QPs a 23% slow-down
(as you mentioned in an earlier mail) isn't justified. But I don't think
that comes from the lambda change itself but rather from some unexpected
side-effects. So that needs to be fixed, sure.

Bye,
Michael



More information about the XviD-devel mailing list