FC3 x86_64 math functions slowdown

[Tony Nelson]

At 9:53 PM -0700 7/9/05, Ulrich Drepper wrote:
>Content-Type: multipart/signed; micalg=pgp-sha1;
>	protocol="application/pgp-signature";
>	boundary="------------enigAD37CF810E21D765208A1EDB"
>
>D. Angelis wrote:
>> I have noticed a significant (up to 5 times !!!) slowdown of some math
>> functions in /lib64/libm.so.6 (glibc-2.3-5) shipped with FC3
>> compared to the version (glibc-2.3.3) shipped with SuSe 9.1.
>
>Very simple explanation: the other code uses non-standard math functions
>which are optimized for the processor but don't produce the same
>results.  Remember, I can arbitrarily fast compute wrong results.

That is, please check that the functions you are testing are producing the
same output on each system tested, and show that output.  (Not just the
sums, also the basic functions passed to loop(), with reasonable test
inputs.)

Umm, I think that (real Sun) Java requires the math library to produce the
same result on any conforming platform; that might be useful to you as a
baseline (though trig identities might also be useful).
____________________________________________________________________
TonyN.:'                       <mailto:tonynelson at georgeanelson.com>
      '                              <http://www.georgeanelson.com/>