[FFmpeg-devel] [NOPATCH] lavfi/mp: drop tinterlace wrapper

[Stefano Sabatini]

On date Wednesday 2012-12-26 01:15:52 +0100, Stefano Sabatini encoded:
> On date Tuesday 2012-12-25 14:03:44 +0100, Michael Niedermayer encoded:
> > On Tue, Dec 25, 2012 at 12:29:47PM +0100, Stefano Sabatini wrote:
> > > On date Wednesday 2012-09-05 10:33:46 +0200, Stefano Sabatini encoded:
> > > > On date Wednesday 2012-09-05 04:31:56 +0200, Michael Niedermayer encoded:
> > > > > On Wed, Sep 05, 2012 at 12:26:01AM +0200, Stefano Sabatini wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > some benchmarks comparing mp=tinterlace and interlace follow:
> > > > > > 0:917992 decicycles in tinterlace, 4096 runs, 0 skips
> > > > > > 0:884563 decicycles in mp=tinterlace, 8192 runs, 0 skips
> > > > > 
> > > > > you should pick lines with similar runs for comparission
> > > > 
> > > > 0:791404 decicycles in tinterlace, 4096 runs, 0 skips
> > > > 0:761499 decicycles in mp=tinterlace, 4096 runs, 0 skips
> > > > 1:10264 decicycles in tinterlace, 4096 runs, 0 skips
> > > > 1:182090 decicycles in mp=tinterlace, 4096 runs, 0 skips
> > > > 2:10917 decicycles in tinterlace, 4096 runs, 0 skips
> > > > 2:184288 decicycles in mp=tinterlace, 4096 runs, 0 skips
> > > > 3:818253 decicycles in tinterlace, 8192 runs, 0 skips
> > > > 3:1407655 decicycles in mp=tinterlace, 8192 runs, 0 skips
> > > > 4:442110 decicycles in tinterlace, 4096 runs, 0 skips
> > > > 4:372288 decicycles in mp=tinterlace, 4096 runs, 0 skips
> > > 
> > > I plan to remove this filter, the difference in performance are mostly
> > > due to the framework code so it is not something which can be
> > > optimized at the filter level, *AND* mp=tinterlace is broken with
> > > regards to timestamps and has less features, so there is little point
> > > in keeping it.
> > > 
> > > I'll drop mp=tinterlace in two days if I read no objections.
> > 
> 
> > code being 10% faster should not be droped.
> > IMHO leave it, so others can compare, and improve either, if they want.
> > 
> > Also if mp can do it fast so can any filter, mp has more code then
> > others to make the interfacing to libmpcodecs possible, any filter
> > could duplicate that  but mp cant loose it so mp is more constrained
> > and should never be faster.
> 
> Given that the two filters have different codepaths they are difficult
> to compare (and the speed gain is somewhat irrelevant for such a
> filter, which works mostly by operating on data/linesizes
> pointers). Also mp=tinterlace is broken and has less features, thus
> believing that someone has interest into using mp=tinterlace or
> comparing it with tinterlace is delusional, while keeping two versions
> of the filter will confuse people.

First note, mp=tinterlace only supports yuv420p and variants formats,
so it doesn't make sense to test other formats, nor with modes which
are only supported in tinterlace. With the other formats tinterlace
is faster since it avoids conversions.

Second observations, due to various issues the null muxer is
duplicating frames, I need to set PTS to NAN (thus reproducing the
same issue of mp=tinterlace) in order to avoid the dup problem.

Updated benchmarks follow.

ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,tinterlace=0,setpts=nan" -f null - -benchmark
bench: utime=9.325s
bench: maxrss=35744kB
ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=0,setpts=nan" -f null - -benchmark
bench: utime=9.329s
bench: maxrss=35740kB

ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=1,setpts=nan" -f null - -benchmark
bench: utime=9.161s
bench: maxrss=19668kB
ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=1,setpts=nan" -f null - -benchmark
bench: utime=8.909s
bench: maxrss=330652kB

ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,tinterlace=2,setpts=nan" -f null - -benchmark
bench: utime=9.121s
bench: maxrss=20656kB
ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=2,setpts=nan" -f null - -benchmark
bench: utime=8.981s
bench: maxrss=330592kB

ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,tinterlace=3,setpts=nan" -f null - -benchmark
bench: utime=9.529s
bench: maxrss=42832kB
ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=3,setpts=nan" -f null - -benchmark
bench: utime=9.597s
bench: maxrss=40780kB

ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,tinterlace=4,setpts=nan" -f null - -benchmark
bench: utime=9.193s
bench: maxrss=19952kB
ffmpeg -f lavfi -i "testsrc=d=200,format=yuv420p,mp=tinterlace=4,setpts=nan" -f null - -benchmark
bench: utime=8.981s
bench: maxrss=330848kB
-- 
FFmpeg = Free and Fierce Majestic Proud Emblematic Genius