[Ffmpeg-cvslog] CVS: ffmpeg/libavcodec snow.c, 1.68, 1.69 avcodec.h, 1.436, 1.437

[Jeff Clagg]

On Wed, Dec 28, 2005 at 04:43:56PM +0100, Michael Niedermayer CVS wrote:
> Update of /cvsroot/ffmpeg/ffmpeg/libavcodec
> In directory mail:/var2/tmp/cvs-serv16368/libavcodec
> 
> Modified Files:
> 	snow.c avcodec.h 
> Log Message:
> iterative overlapped block based motion estimation for snow

I tried this out. The short summary of my results is that the new code
is maybe a couple % faster with no significant change in psnr
(+-0.01dB).

Test methodology: vqscale=4 for pass 1, vbitrate=1200 for pass 2. Tested
with some varying combinations of *cmp and v4mv, first with cvs checkout
from December 22, then with today's CVS. Results are here:
http://snackbox.org/~snacky/testclips/snowtests.txt
(annoying non-wrapped lines warning!)
I'm on a 1.13GHz P3 using GCC 3.3.6.

I also have the relevant clips here:
http://snackbox.org/~snacky/testclips/
The first 400 frames of the source file used:
http://snackbox.org/~snacky/testclips/mushi_cut.avi

There is also a VERY large amount of chroma artifacting in the form of
repeating rainbowish horizontal lines. Seems to be a lot worse when v4mv
is used. I don't recall ever seeing anything like this last time I used
snow - any ideas offhand?