[dvd::rip] Results of Linux vs. Windows benchmarking

Hello,

I did some benchmarking under Linux (transcode/dvd::rip) and Windows
(AviSynth/VirtualDub) to see where the performance differences I perceive
come from.


Testbed
=======

Source material:  Frames 0-2000 of "Revolution 909" video clip by Daft Punk,
PAL (720x576), interlaced, ripped from DVD.  This is fairly hard to encode,
with many dark scenes at the beginning, light scenes later, a few close ups,
lots of motion/cuts.

When resizing, I resized down to 640x480.

Hardware:  Athlon XP 2800+, 512 MB RAM (dual channel), NForce2 chipset (Asus
A7N8X), IDE drive

Software:
  * Linux: transcode 0.6.12, xvid-1.0.0beta3 (compiled by myself), dvd::rip
  * Windows: Nic's Build of XviD-1.0.0beta from March 2003, recent versions
             of AviSynth/VirtualDub (didn't write them down, but DVX
             requires the latest builds usually), DVX as frontend


Settings
========

  1. XviD:

     a) Quant MPEG, Motion "6 - Ultra High", VHQ "3 - Medium Search",
        Chroma ME, HighQ AC (there's no such setting with the Windows XviD
        build I use), Trellis Quant, 1 B-Frame, rest defaults

        I call this "HQ XviD"

     b) Motion "4 - High", VHQ "0 - Disabled", rest as above

        I call this "LQ XviD"

     My namings of the presets are just for the sake of brevity here, there
     might be "more optimal settings" for the different scenarios.

  2. transcode:

     a) no resize, no deinterlace
        ("No Processing")

     b) "Fast Resize" (-B option), "Zoom to Full Frame" (-I 3 option)
        ("Fast Processing")

     c) "HQ Resize" (-Z option), "Zoom to Full Frame"
        ("HQ Processing")

  3. AviSynth/VirtualDub

     I've used DVDx here and tried the following setup:

     "Deinterlace" [i.e. FieldDeinterlace() in AviSynth] and "Sharp" resizing
     [i.e. LanczosResize() in AviSynth]

     Haven't done other settings, as I mainly wanted to see the difference
     speed of the above two operations (resize/deinterlace) compared to
     transcode.


Results
=======

Numbers are: FPS for Pass 1 / Pass 2 (all figures are rounded to the nearest
half)

Linux
-----
         No Processing  Fast Resize     HQ Resize

HQ XviD  27,5 / 10      16 / 8,5        12 / 7

LQ XviD  28,5 / 35      17 / 17         12 / 12,5

Additionally I tried Motion 5 and Motion 6 with VHQ 1, which resulted in
28/17,5 and 27/16,5 respectively (no preprocessing here).  So apparently VHQ
is the biggest consumer of CPU cycles.  At first sight I could see any
difference in the encoding, probably there is hardly any.

With Motion 6, VHQ 1 and "HQ Resize" I got 12 / 9,5

Windows
-------
         HQ Resize

HQ XviD  24 / 11

LQ XviD  24 / 26

I also ran some short test without preprocessing at all to see if XviD is
any faster/slower under Windows.  Short answer: No, performance is more or
less the same -- the few FPS in difference I got can probably accounted to
the toolchains in question.


Conclusions
===========

a) transcode's routines for resizing and deinterlacing eat up a lot more CPU
   than AviSynth's.  Especially at Pass 1 this is apparent -- I'd like to
   know why (wrong setup?)

b) The higher your preprocessing quality, the less important become your
   XviD parameters, since at best quality, deinterlacing/resizing dominate
   the CPU.  So IMHO it's not much harm to go all the way in xvid4conf if
   one has to deinterlace anyway -- percentage-wise the loss of time is not
   that big anymore then.

c) Lanczos resize takes really long with transcode, and I didn't see any
   difference in the resulting picture, so I wouldn't go for it unless
   absolutely necessary (maybe to get more flexibility in target size).  Joern
   has said that already in dvd::rip's documentation and I guess he was
   right :-).

d) Probably it's worthwhile to port AviSynth's resizing/deinterlacing over
   into transcode.  The AviSynth website states that VirtualDub's routines are
   even better optimized (they explicitely talk about resizing there even). 
   Unfortunately I'm not able to do that I'm afraid...

e) XviD seems to perform very similarly under Windows and Linux with the
   same codec setup on both platforms.


I hope this is useful information for somebody, maybe someone wants to test
how much difference deinterlacing makes in the process.  I didn't test that
since one has no choice in that matter anyway.  Either you do deinterlace or
not.  OTOH it would be interesting to see how much better (worse?) AviSynth
performs here that transcode.

Greetings,
   Frank

____________________________________________________
http://lists.exit1.org/mailman/listinfo/dvdrip-users
How To Ask Questions The Smart Way:
http://catb.org/~esr/faqs/smart-questions.html