[mythtv] revised greedy high motion deinterlacer + yadif

[Markus Schulz]

Am Freitag, 30. November 2007 schrieb Phil Vid:
> On Nov 26, 2007 7:45 AM, Markus Schulz <msc at antzsystem.de> wrote:
> > Am Sonntag, 25. November 2007 schrieb Markus Schulz:
> > > new version:
> > > http://www.tastatur-junkie.de/greedyh_and_yadif_svn14949.diff.gz
> >
> > small fix for trashed lines at top/below.
> > http://www.tastatur-junkie.de/yadif_allocfix.diff
>
> Compiles and runs fine for me, however for 1080i my AMD X2 5200+
> system can't keep up. For a comparison, yadif 2x runs consistently at
> 40%/90% and greedy 2x will stay around 50%/50%.
> Maybe the optimized memcpy will assist, but probably not enough.
> Yadif looks better than greedy for the lower quality stuff though so
> I'll keep using it for that. I found greedy would sometimes cause
> artifacts at the sides of the screen and some shimmering at times.
> This was only noticeable in animated content.

I agree that yadif looks better then greedy (intensive checks with some 
soccer games). 
I've measured the average ProcessFrame duration of all three adaptive 
deinterlacers withd SD material:
Yadif:  ~ 170 fps
Greedy: ~ 350 fps
Kernel: ~ 950 fps

I have no HD material (only material which needs spatial direct mode 
which is not supported thru ffmpeg) to check for performance boost from 
optimized memcpy code.
For SD material it will result in performance loss! (yes, cause too 
small block sizes) and i don't think this will be different for HD 
(i've tried to simulate with test program).

If someone want to try the version with benchmarking and optimized 
memcpy code(aclib from mplayer), let me know. Currently i would prefer 
to release without optimized memcpy code, but perhaps i've made some 
mistakes ;)

-- 
Markus Schulz