Yesterday I experimented a little bit with the upcoming WavPack 4.40 (svn revision 33) and my MMX patch for the extra modes. Running a few speed tests in the new high mode (switch -hx3) on a Pentium III Coppermine with Linux and GCC 4.1.1 showed, that the encoding time decreases by about 18%. So maybe some testers here will find that useful.
If you're interested in x86 binaries for GNU/Linux or Mac OS X, you can download the archive 'wp-4.40s33-bin.tar.bz2' from http://base91.sourceforge.net/download/wavpack/ - there you also find the patch 'wp-4.40s33-jh_mmx.diff.gz'. This patch is nothing really new; beside small adjustments to the new version, it also contains some modifications to the update_weight macros, which can lead to a very minor speedup in all modes.
Next target for me is to write something similar with Altivec on PowerPC, but please don't expect that too soon. Unfortunately, I do not have very much spare time recently. *sigh*
Kind regards,
Jo.
Full Version: speed optimizations for WavPack 4.4x
Not bad, not bad... 
Thank you very much.
OS: WinXP SP2 AMD 64 3200+ / 2 Go RAM
Sample : wav - 44100 Hz - 2 ch - 0:48:26.45
Compiler: gcc 4.1.1 + CFLAGS (-march=k8 -O3 -mmmx -pipe -fomit-frame-pointer -DOPT_MMX)
Values: encoding time (in seconds)
Thank you very much.
OS: WinXP SP2 AMD 64 3200+ / 2 Go RAM
Sample : wav - 44100 Hz - 2 ch - 0:48:26.45
Compiler: gcc 4.1.1 + CFLAGS (-march=k8 -O3 -mmmx -pipe -fomit-frame-pointer -DOPT_MMX)
Values: encoding time (in seconds)
QUOTE(kurtnoise @ Nov 13 2006, 12:01) 
Compiler: gcc 4.1.1 + CFLAGS (-march=k8 -O3 -pipe -fomit-frame-pointer)
This doesn't enable the MMX optimizations. You should build with
CODE
make CFLAGS='-pipe -march=k8 -mmmx -O3 -fomit-frame-pointer -DOPT_MMX'
Sorry, I forgot to mention that again.
ok, I stand corrected. Previous post updated. I thought it was automatic...
It's a pity that you removed the old chart, because it showed the effect of the modified update_weight macros. I'm sorry, but your new test results appear to be a little bit strange 
- my patch should not affect the normal mode that much, but should have a higher impact on the extra modes. How did you measure the encoding time?
- my patch should not affect the normal mode that much, but should have a higher impact on the extra modes. How did you measure the encoding time?
QUOTE(he-jo @ Nov 13 2006, 16:46)
It's a pity that you removed the old chart, because it showed the effect of the modified update_weight macros.
The old one...
QUOTE(he-jo @ Nov 13 2006, 16:46)
How did you measure the encoding time?
by using values from the wavpack encoder...
My measures on Athlon XP 2200+ under Ubuntu 6.10 Linux.
Wavpack 4.40.0 r37 compiled by GCC 4.1.1.
First column is patched version w/o -DOPT_MMX, second - with.
Wavpack 4.40.0 r37 compiled by GCC 4.1.1.
First column is patched version w/o -DOPT_MMX, second - with.
CODE
| Time, sec
Mode |-----------------------
| NO_MMX | MMX
--------+-----------+-----------
-f | 4.01 | 3.75
-f -x | 7.73 | 7.24
-f -x1 | 7.56 | 7.19
-f -x2 | 10.73 | 10.00
-f -x3 | 18.38 | 16.86
default | 6.07 | 4.77
-x | 10.58 | 9.29
-x1 | 10.42 | 9.27
-x2 | 16.44 | 14.34
-x3 | 31.93 | 27.02
-h | 6.85 | 6.73
-h -x | 15.20 | 13.02
-h -x1 | 15.02 | 13.02
-h -x2 | 25.38 | 21.30
-h -x3 | 51.47 | 42.51
-hh | 9.29 | 8.73
-hh -x | 20.51 | 16.99
-hh -x1 | 20.42 | 17.07
-hh -x2 | 36.19 | 29.39
-hh -x3 | 76.38 | 60.92
Mode |-----------------------
| NO_MMX | MMX
--------+-----------+-----------
-f | 4.01 | 3.75
-f -x | 7.73 | 7.24
-f -x1 | 7.56 | 7.19
-f -x2 | 10.73 | 10.00
-f -x3 | 18.38 | 16.86
default | 6.07 | 4.77
-x | 10.58 | 9.29
-x1 | 10.42 | 9.27
-x2 | 16.44 | 14.34
-x3 | 31.93 | 27.02
-h | 6.85 | 6.73
-h -x | 15.20 | 13.02
-h -x1 | 15.02 | 13.02
-h -x2 | 25.38 | 21.30
-h -x3 | 51.47 | 42.51
-hh | 9.29 | 8.73
-hh -x | 20.51 | 16.99
-hh -x1 | 20.42 | 17.07
-hh -x2 | 36.19 | 29.39
-hh -x3 | 76.38 | 60.92
QUOTE(kurtnoise)
by using values from the wavpack encoder...
Ok 
But according to my experience, these values are not very exact and vary quite a lot between several runs. With Windows it would probably be the best to use the commandline tool timer, with Unix the command 'time' - like
CODE
timer wavpack -hx3 music.wav
time wavpack -hx3 music.wav
time wavpack -hx3 music.wav
Some Altivec modifications for the extra stereo modes are available now. I was just able to test on a G5. Unfortunately they don't give a real speed up there, when encoding 16 bit files. But I measured a speedup of ~16% when encoding 24 bit and 32 bit tracks. I would really appreciate it, if someone would run some tests on a G4.
Get the patch in wp-4.40beta-jh_altivec.diff.gz and binaries for Mac OS X in wp-4.40beta-ppc.tar.bz2 from http://base91.sourceforge.net/download/wavpack/.
Many thanks in advance,
Jo.
Get the patch in wp-4.40beta-jh_altivec.diff.gz and binaries for Mac OS X in wp-4.40beta-ppc.tar.bz2 from http://base91.sourceforge.net/download/wavpack/.
Many thanks in advance,
Jo.
This is a "lo-fi" version of our main content. To view the full version with more information, formatting and images, please click here.