CCL:G: Be a bit cautious...

 Sent to CCL by: Zsolt Zsoldos [zsolt|simbiosys.ca]
 This is a serious problem that I have also been fighting for years --
 so far without any general silver-bullet solution. There is a very
 simple example code to demonstrate how horrible the problem is on my
 earlier blog post I made as response to a similar emal on this list:
 http://www.simbiosys.ca/blog/2008/05/28/floating-point-errors/
 If you compile that sample code on different platform, different
 optimization options, you get different cases giving the wrong result,
 but what is common to all platforms and option is that you have a very
 significant portion of the tested cases fail. Which means, floating
 point errors can wrack havoc on our calculations even at  single
 decimal digit precisions!
 ZZ
 On Fri, Oct 30, 2009 at 6:04 PM, John McKelvey jmmckel^^^gmail.com
 <owner-chemistry[A]ccl.net> wrote:
 > Folks,
 >
 > I have a dilemma: Significantly results depending on the compiler.
 >
 > All copies of source are identical and compiled for real*8 (double
 precision
 > for 32 bits), and all answers are independent of optimization level.  The
 > gfortran is the latest available download version. ( I chose it because it
 > is available across several OS's)  The g77 came with RHEL.
 >
 > The calculation involved reading output from a G03 log file where Mulliken
 > populations were computed from selected printed open shell MO's and printed
 > AO overlaps. [log file numbers had 5-6 significant figuresafter decimal]
 > Again, all calculations done in the special program were done in real*8,
 and
 > results are independent of optimization level.
 >            Compiler                      Answer      Hardware, OS
 > 1) Linux Intel compiler (ifort)            ~16   AMD64 opteron  RHEL4
 > (~=Centos 4)
 > 2) Linux g77 compiler                     ~10             "
 > 3) Windows Digital Corp compiler    ~16    Intel P4  Windows2000
 > 4) Windows gfortran  compiler         ~10       "
 > 5) Linux      gfortran        "               ~10      Xeon
 Centos 5.3
 > (~= RHEL5.3)
 >
 > Thoughts?
 >
 > Thanks!
 >
 > John
 >
 > --
 > John McKelvey
 > 10819 Middleford Pl
 > Ft Wayne, IN 46818
 > 260-489-2160
 > jmmckel%a%gmail.com
 >
 >
 >
 > --
 > John McKelvey
 > 10819 Middleford Pl
 > Ft Wayne, IN 46818
 > 260-489-2160
 > jmmckel%a%gmail.com
 >
 --
 Zsolt Zsoldos