[Ffmpeg-devel] MPEG-1/2 decoding benchmarks

[Michael Niedermayer]

Hi

On Monday 25 July 2005 21:40, Michael Niedermayer wrote:
> Hi
>
> On Monday 25 July 2005 20:50, Diego Biurrun wrote:
> > On Wed, Apr 27, 2005 at 01:41:50AM +0200, Diego Biurrun wrote:
> > > I have repeated my MPEG-1/2 decoding benchmarks with the latest
> > > revisions of mpeg12.c.
> > >
> > > I tested revision 1.230 - 1.233 of mpeg12.c, 1.234 had no functional
> > > changes.  Log messages:
> > >
> > > 1.230 avoid UPDATE_CACHE() in GET_RL_VLC() if not needed
> > > 1.231 optimizing non_intra ac coeff decode
> > > 1.232 mpeg2_fast_decode_block_intra()
> > > 1.233 avoid clearing block[] if its not used at all
> > >
> > > I tested on two different systems, my desktop and notebook
> > >
> > > desktop:
> > > 500MHz AMD K6-III+
> > > 512MB SD-RAM 133
> > > Debian testing (up to date)
> > > gcc 2.95.4 (Debian)
> > >
> > > notebook:
> > > 1500MHz G4 with AltiVec
> > > 512MB PC2700 (333MHz) DDR SDRAM
> > > Debian testing (up to date)
> > > gcc 3.3.5 (Debian)
> > >
> > >
> > > RESULTS:
> > >
> > > K6-III: 1.231 gave a slight speedup (2.3%) in some cases, 1.232 a
> > > slight performance degradation (1.7%) in some cases, no changes in
> > > 1.233.
> > >
> > > G4: 1.231 gave a slight speedup (1.2%) in some cases, 1.232 a slight
> > > improvement (1.2%) in one case, almost no changes in 1.233.
> >
> > Should revision 1.232 maybe be reverted?  It does not seem to have
> > helped performance, at least on i386, which is the main target.  If more
> > details are necessary, I can redo the benchmarks.  I distrust my PPC
> > results a bit now.  Since my PowerBook is from February, the hardware
> > was not yet completely supported when I did the benchmarks in April.
>
> if its slower (almost) everywhere then it should be reverted
> though it should be benchmarked again (at least with gcc 4.0 & 3.something
> & 2.95) and benchmarking should use START/STOP_TIMER in
> avcodec_decode_video, i dont trust mplayers -benchmark

gcc 4.0.1 duron 800mhz matrixbench
cvs
91807280 dezicycles in avcodec_decode_video, 4096 runs, 0 skips
92171617 dezicycles in avcodec_decode_video, 4096 runs, 0 skips
91826789 dezicycles in avcodec_decode_video, 4096 runs, 0 skips

cvs up -j 1.232 -j 1.231 mpeg12.c
93358586 dezicycles in avcodec_decode_video, 4096 runs, 0 skips
93240137 dezicycles in avcodec_decode_video, 4096 runs, 0 skips
92810858 dezicycles in avcodec_decode_video, 4096 runs, 0 skips

[...]
-- 
Michael