[XviD-devel] Small report from my amd64 merging effort

[Edouard Gomez]

Edouard Gomez (ed.gomez at free.fr) wrote:
> I'll merge into my public tree at [1] pretty soon, and if everything
> seems right... it'll actually make its way to the CVS a bit after.

Ok some teasing for people who will soon test it:

* mencoder+libavcodec as of 2004-12-07
* xvidcore head patch 103 from my tree (probably equivalent to cvs head)
  + mostly all relevant functions when dealing with options i used for
  the encoding test.
* options used:
  max_bframes=1:bquant_ratio=150:bquant_offset=100:bf_threshold=0:\
  chroma_me:vhq=4:bvhq=0:trellis
* Bi-AMD Opteron(tm) Processor 240 (but i don't think the mencoder
  task migrates CPUs, the load stayed around 50% as mencoder is
  single threaded). For reference, this CPU is clocked at 1.4GHz
* Debian GNU/Linux x86_64 (no 32bit compatibility mode used),
  2.6.9-ac kernel

There are two runs, they're configured this way:
./configure --prefix=${HOME}/opt/xvidcore --disable-assembly
./configure --prefix=${HOME}/opt/xvidcore

C code only:
$ time ./twopass-tests.sh
===============================================================================
Two pass test name: underworld-trailer
Using cached first pass stats (hash - ebffd716144b645ec3013b9db5fb1b4d)

Results:
xvid:     Min PSNR y : 31.73 dB, u : 38.62 dB, v : 44.60 dB, in frame 2581
xvid: Average PSNR y : 39.32 dB, u : 43.08 dB, v : 45.47 dB, for 3534 frames
xvid:     Max PSNR y : 99.99 dB, u : 99.99 dB, v : 99.99 dB, in frame 1336
Filesize: 15339702 bytes
===============================================================================

real    9m26.308s
user    9m26.033s
sys     0m0.245s

Optimized run:
$ time ./twopass-tests.sh
===============================================================================
Two pass test name: underworld-trailer
Using cached first pass stats (hash - ebffd716144b645ec3013b9db5fb1b4d)

Results:
xvid:     Min PSNR y : 31.75 dB, u : 38.59 dB, v : 44.42 dB, in frame 2581
xvid: Average PSNR y : 39.32 dB, u : 43.08 dB, v : 45.47 dB, for 3534 frames
xvid:     Max PSNR y : 99.99 dB, u : 99.99 dB, v : 99.99 dB, in frame 1336
Filesize: 15336108 bytes
===============================================================================

real    3m18.491s
user    3m18.281s
sys     0m0.218s

For reference on my AthlonXP 2200+ (clocked 1.8GHz, no oclocking), Debian
i386, 2.6.10-rc3-ck1 (the task scheduler is different, and can influence
the bench).   Mostly the  same softwares. The  set of functions  used in
XviD differ  (i use  3dnow functions,  and i decided  not to  port these
functions in the x86_64 port)

===============================================================================
Two pass test name: underworld-trailer
Using cached first pass stats (hash - ebffd716144b645ec3013b9db5fb1b4d)

Results:
xvid:     Min PSNR y : 31,73 dB, u : 38,62 dB, v : 44,45 dB, in frame 2581
xvid: Average PSNR y : 39,33 dB, u : 43,09 dB, v : 45,48 dB, for 3534 frames
xvid:     Max PSNR y : 99,99 dB, u : 99,99 dB, v : 99,99 dB, in frame 1336
Filesize: 15329386 bytes
===============================================================================

real    4m35.253s
user    4m28.634s
sys     0m0.569s

To sump up:
 - x86_64 non optimized: 6.2fps
 - x86_64 optimized for this encoding option set: 17.8fps
 - ia32 optimized: 12.8fps

So stay tuned ;-)

-- 
Edouard Gomez