[MPlayer-dev-eng] non-free and undistributable files in tremor/

[Diego Biurrun]

On Wed, Aug 02, 2006 at 11:21:16AM +0200, Diego Biurrun wrote:
> On Sun, Jul 30, 2006 at 06:12:28AM -0400, Rich Felker wrote:
> > On Sun, Jul 30, 2006 at 09:47:19AM +0200, Michael Niedermayer wrote:
> > > 
> > > On Sun, Jul 30, 2006 at 02:51:50AM +0200, Diego Biurrun wrote:
> > > [...]
> > > > The tremor subdirectory weighs in at 440k, libvorbisidec.a at 125k.
> > > > FFmpeg can now decode Vorbis natively.  Time to run some benchmarks I
> > > > would say.
> > > 
> > > yes, and some profiling / benchmarking the indivdual parts and then
> > > someone should take the faster parts and clean them up and port them
> > > to lavc
> > 
> > imo the faster part is using integers, which iirc the lavc
> > implementation does not do...
> 
> Then again, CPUs today are not necessarily faster at doing integer
> compared to float computations.  But let's not restart this silly topic.
> 
> I sat down and did some benchmarks on my K6-III 500, Debian stable
> system with all services except getty instances stopped, running Linux
> 2.4.32 and MPlayer r19262.  I used the following two samples to test,
> the first is 10 minutes long, the second 70 minutes
> 
> ftp://ftp.openbsd.org/pub/OpenBSD/songs/song37.ogg
> http://jdolan.dyndns.org/jaydolan/choons/Jay%20Dolan%20-%20Live%20(6-17-04).ogg
> 
> I compared the internal tremor decoder with the FFmpeg Vorbis decoder
> and internal tremor with --enable-tremor-low.  Below are the results.

I repeated the benchmarks after the last round of updates and compiled
gcc 4.1.1 myself to get those extra 3dnow optimizations as suggested.  I
used MPlayer r19309.

The results, best of five runs each are very encouraging:

mplayer -quiet -benchmark -afm ffmpeg -ao pcm:fast:file=/dev/null song37.ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A:  22.187s Sys:   0.099s =   22.286s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.5556% Sys:  0.4444% = 100.0000%

--enable-tremor-low:
mplayer -quiet -benchmark -ao pcm:fast:file=/dev/null song37.ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A:  30.603s Sys:   0.099s =   30.702s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.6773% Sys:  0.3227% = 100.0000%

mplayer -quiet -benchmark -ao pcm:fast:file=/dev/null song37.ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A:  37.964s Sys:   0.108s =   38.072s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.7166% Sys:  0.2834% = 100.0000%


mplayer -quiet -benchmark -afm ffmpeg -ao pcm:fast:file=/dev/null Jay Dolan - Live (6-17-04).ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A: 139.718s Sys:   0.757s =  140.476s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.4609% Sys:  0.5391% = 100.0000%

--enable-tremor-low:
mplayer -quiet -benchmark -ao pcm:fast:file=/dev/null Jay Dolan - Live (6-17-04).ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A: 185.520s Sys:   0.667s =  186.187s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.6417% Sys:  0.3583% = 100.0000%

mplayer -quiet -benchmark -ao pcm:fast:file=/dev/null Jay Dolan - Live (6-17-04).ogg
BENCHMARKs: VC:   0.000s VO:   0.000s A: 234.753s Sys:   0.705s =  235.458s
BENCHMARK%: VC:  0.0000% VO:  0.0000% A: 99.7006% Sys:  0.2994% = 100.0000%

ffvorbis is way faster.  I vote for removing tremor, there appear to be
no circumstances under which it is faster and many where it is
considerably slower.

Diego