Guillaume Cottenceau wrote:
>
> Karl Mitchell <[EMAIL PROTECTED]> writes:
>
> [...]
>
> > > better with -O1 than with -O2 or superior!??
> >
> > Faster with -O1 than -O2 or -O3 under almost all criteria for speed. I can't
> > remember the source though and the figures I've got (below) aren't a good
> > comparison. If you replace -O1 with -O3 on test B (below) you'll see what I mean.
> >
> > By the way, I would urge against using gcc -ffast-math, as it can cause actual
> > errors in floating point routines. I am a numerical modeller and I found for one of
> > my floating point iterative routines the cumulative error was a few percent in one
> > case. This could have been more to do with the version of egcs I was using, but I
> > doubt it. Admittedly this doesn't matter for all programs, but perhaps something
> > like 'octave' would suffer?
>
> I don't know precisely why this has been chosen. I think this is because,
> this is default parameters, and packagers are supposed to override them
> when needed.
>
> [...]
>
> > A) CFLAGS = -s -static -Wall
> > B) CFLAGS = -s -static -Wall -O1
> > C) CFLAGS = -s -static -O3 -fomit-frame-pointer -Wall -mpentiumpro
> > -march=pentiumpro -fforce-addr -fforce-mem -malign-loops=2
> > -malign-functions=4 -malign-jumps=2 -funroll-loops
> > -fexpensive-optimisations -malign-double -fschedule-insns2
> > -mwide-multiply
> >
> > Based on K6-233 Index
> >
> > Mem Index Integer index FP Index
> > A 1.349 1.014 1.824
> > B 3.605 2.284 5.003
> > C 3.609 3.245 7.175
>
> I don't understand.
>
> You did not explain the numbers:
>
> I suppose than B is faster than A because there is no optim for A. So
> given this, I interpret your numbers are "speed index" (opposed to "time
> consumed"), e.g. the greater the faster.
>
> So, For all three things, C is faster than B.
>
> Why do you say -O1 is faster than -O2, then? Is it something to do with
> the other parameters? (so why did you put other parameters for C..!?)
>
> --
> Guillaume Cottenceau -- Distribution Developer for MandrakeSoft
> http://www.mandrakesoft.com/~gc/
Probably he is referring to an article appeared on cpureview showing
that sometimes on K7 -O1 would produce better results than -O2 or more.
Those index seems the Byte index benchmark.
Definitively I think we should do better benchmark for K7, Duron
maybe using ssbench benchmark
http://nastol.astro.lu.se/~stefans/bench.html
Volunteers?
I remember that we've already choosen to avoid -ffast-math for mathematical
packages (gnuplot, octave, etc.). Maybe we should explicitely use it into
the spec files, sort of
CFLAGS="$RPM_OPT_FLAGS -fno-fast-math"?
Bye.
Giuseppe.
