> making the most of the instruction set can certainly help and might cloud
> comparisons of PII vs. PIII if they are not both one or the other. I am
I'm running an athlon kernel myself. I am compiling with gcc 2.96 and
have used the -mathlon model (generates athlon binaries, supposedly). I
have also used it to compile some other stuff, and do a little amount of
benchmarks with various codes.
In my (limited) use it would seem that the -mathlon is actually somewhat
slower than -mpentiumpro on Athlon processors. This is noticeable in
programs like the simple chess engine benchmark (tscp) and of course your
mileage may vary. Supposedly, compiling with i686 (-mpentiumpro) is
desireable. On the other hand, simply compiling the kernel 'for' a
particular CPU set involves setting one option or two in the kernel's
configuration, and doesn't mean that the kernel is optimized for that
particular CPU.
Also, gcc 2.96 with some codes (particularly parts of the KDE source
tree) has compilation problems when some of the optimization options are
enabled -- in particular the -fscheduleinsns options. (I get some message
about not being able to find a spill register.) These don't happen if
-mpentiumpro is used.
Also, rpm --rebuild with a --target seems to choose some decent additional
options and pass them along to gcc.
> under the impression people should make sure that their kernel is
> optimized for their processor, and if it is not learn how to rebuild if
> they are concerned with performance.
It would also help if fundamental libraries were optimized too. Obviously
the kernel is important, but I'm hot real sure how much an optimized kernel
helps.
>
> a CPUs strengths. I could imagine a PII outperforming a PIII particularly
> in places where a pipeline stage changed (I can't even begin to say I know
> how the PIII operates at this level) and the code was not compiled for
I am not all that clear either, but it seems that PIIIs (and prohably PIIs)
suffer from things like partial register stalls where certain code will just
cause the processor to 'hang' until a resource becomes free. The Athlon, as
far as I know, doesn't suffer from this penalty, and doesn't suffer from the
penalties, so far as I know, of mixing MMX and floating point instructions
that tend to slow down previous CPU generations. And if there might be people
with simple original Pentiums, that chip is pretty hard to optimize for,
although the newer compilers can generate supposedly pretty good code these
days.
More background (in case there is interest) can be found at AMD's web
site, tomshardware, athlonlinux.org, for instance.
> that change in the architecture. Can sheer clock-speed brute-force you
> Albion
for a sec, I thought your first name was "Athlon" :)
------------------------------------------------------------------------
David E. Fox Thanks for letting me
[EMAIL PROTECTED] change magnetic patterns
[EMAIL PROTECTED] on your hard disk.
-----------------------------------------------------------------------
