Re: Bug in 'fft'

[Thomas Walter <walter at pctc dot chemie dot uni-erlangen dot de>]

>>>>> "Brian" == Brian Gough <bjg@network-theory.co.uk> writes:

    Brian> Thanks for the bug report.  The "n mod 7 = 0" behaviour isolates the
    Brian> problem to c_pass7.c (the n=7 butterfly).

    Brian> I think it may be a problem with gcc-2.95.1 compiling that file.  The
    Brian> code is working ok with other compilers.  I don't know if you tried
    Brian> compiling without -O, but if that makes the problem go away it would
    Brian> support this hypothesis.

    Brian> Perhaps someone on the list might know if there are potential problems
    Brian> with that version of gcc. 

First, sorry for not telling my compile flags.
They are:   '-O2 -mpentium -malign-double'
And I use the stuff from the CVS repository, last update today
morning.

After changing from '-O2' to '-O' the problem goes away.

I tried the same procedure with the stuff in 'sum' but stil error
reports.  Especially the test for 
	  correct digits of asymptotic series
is independent of the optimizing level.

I also added 'gsl_ieee_env_setup ();' to the 'test.c' and played with
      1) make check
      2) GSL_IEEE_MODE="double-precision;round-to-nearest;" make check

To my surprise this switches the PASS/FAIL state of
  1) PASS: correct digits for exp(10)
     FAIL: correct digits for -log(2)

  2) FAIL: correct digits for exp(10)
     PASS: correct digits for -log(2)

Then I switched to 'ieee-utils':
I think there may be problems in the ieee-utils stuff for linux.  The
example program from the docs about suming up works only if I use
     GSL_IEEE_MODE="double-precision;round-to-nearest;"
the default mode  'extended-precision' does some 100 evaluations and
ends with a floating point exception and a core dump.

Some light about these problems are welcome, thanks.

Bye
Thomas

-- 
Latest news:
Einbruch bei Microsoft, Bill Gates hat Windows offen gelassen.

----------------------------------------------
Dipl. Phys. Thomas Walter
Inst. f. Physiklische Chemie II
Egerlandstr. 3				Tel.: ++9131-85 27326 / 27330
91058 Erlangen, Germany			email: walter@pctc.chemie.uni-erlangen.de