[MPlayer-dev-eng] Using -O4 vs. -Os

[Zoltan Hidvegi]

> Zoltan Hidvegi writes:
>  > For the discussion about using -O4 vs. -Os, I've run sume benchmarks,
>  > on my Athlon XP Thoroughbred 2233 MHz, 194MHz fsb machine, using
>  > gcc-3.3.2 prerelease (debian unstable 3.3.2-0pre5).
> 
> I assume that much of the differences you see is related to cache
> utilization.  Please bear in mind that the code needs to run fast on a
> modest machine also.  At 2GHz you have plenty of horsepower to watch
> all kinds of movies anyway.  IMHO performance should be optimal in the
> <=500MHz (and smaller cache) range where every bit really matters.

That may be the case as long as you do not start applying filters, do
encoding etc.  I have several Athlons, and the performance (encoder
speed) increases pretty linear to the CPU speed, so I do not think
that the cache/memory is the main botleneck.  I haven't benchmarked
mencoder, that should be the next.

The slowest machine I have access to is my laptop, which runs at
533MHz when switches to low-power mode, maybe I can run sum tests on
that.  But even that laptop is more than enough for watching
everything but HDTV streams.  <500MHz machines are obsolete, and you
can buy GHz machines for $100 if you need to (not new, you may have to
buy a few parts).  I've paied $50 for my 2.2GHz CPU, which is of
course only labeled as a 1.46 GHz, but overclocks well.  Slow machines
are still interesting in a low-power PDA formats, but I do not have
access to such gadgets.

And while MPlayer is my favorite movie viewer, I think its main
strength lies in the under-appreciated MEncoder part, which is just so
much better and faster than any other encoder I've seen.  And for
encoding, you need an infinitely fast CPU, and probably even that
would not be fast enough :-).  I like to capture movies via my TV
tuner in real-time in 640x480 resolution, and using denoise3d,
inverse-telecine and perhaps a small gaussian blur followed by mpeg4
vhq and sometimes also v4mv encoding.  It is really quite amazing how
good captured TV programs look this way.  But such a capture uses
almost all of the CPU time on my 2.2GHz machine.  I also have an HDTV
tuner, and processing 1080i mpeg can easily push the limits of my
machine.  And I like to encode HDTV streams to mpeg4 scaling them down
and using denoise3d and high-quality 2-pass encoding, and encoding a 1
hour program can take 5 hours.  So, even though I know that my machine
is top of the line, I still think that my machine is not fast enough,
and I am very much interested in making mencoder faster.

Zoli