[Wien] accuracy problems

[Sorin Lazar]

Dear prof. Blaha,

Sorry that I wasn't  more explicit about the calculation. Actually the 
tests were performed by a colleague of mine.
He said that he used the test case TiC from the user guide. He did the 
calculation on various machines(Athlon, Opteron, G4 with Absoft) and he 
always got the same results.
By incorrect he means that the calculation doesn't converge in 20 cycles 
(usually needs only 7 cycles on any other machine) the total energy jumps 
by a big amount (+/- 0.1-0.01 Ry) the same for charge distance (jumps) and 
the fermi energy is completely wrong.
He compiled also without optimization and the results are the same.
If you need more details or some files let me know and I will ask him to 
send you what you need.

thanks,
Sorin


At 13:01 2003-12-04 +0100, you wrote:

> > We are seeing some accuracy trouble with Wien2k when combining various
> > compilers and blas+lapack libraries on PowerMac systems.
> > The data from the TiC sample case is as follows:
> >
> >     compiler      library                                        result
> > 1) Absoft         Absoft                                        correct
> > 2) Absoft         vecLib                                        incorrect
> > 3) 
> xlf               veclib                                         incorrect
> > 4) xlf               manually compiled blas/lapack     incorrect
> >
> > The incorrect results are pretty similar for the three wrong cases. From
> > 1-3 the vecLib libraries would seem the culprit. But 4 kills that logic.
> >  From this data, it seems that it is not just the compiler or just the
> > library. 2+3 tells you it is not just the compiler that is giving the
> > error. 3+4 tells you the library is not the sole culprit.
> > Someone who was helpful in finding out the above data suggested that it
> > could be some numerical instability issue in the Wien2k code. The code runs
> > fine of course on many platforms, including PowerMacs using all-Absoft. Has
> > anyone ever seen something similar, perhaps using development-stage
> > versions of the code?
>
>You did not tell what means: "incorrect" ?
>
>Anyway, I would claim it is due to compiler bugs (+veclib bugs?). At least the
>xlf is known for various bugs at some optimization levels.
>
>I'd expect correct behaviour if you explicitly reduce the optimization
>(maybe down to -O0).
>
>                                       P.Blaha
>--------------------------------------------------------------------------
>Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
>Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
>Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
>--------------------------------------------------------------------------
>
>_______________________________________________
>Wien mailing list
>Wien at zeus.theochem.tuwien.ac.at
>http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>
>
>
>
>
>---
>Incoming mail is certified Virus Free.
>Checked by AVG anti-virus system (http://www.grisoft.com).
>Version: 6.0.528 / Virus Database: 324 - Release Date: 2003-10-16
-------------- next part --------------

---
Outgoing mail is certified Virus Free.
Checked by AVG anti-virus system (http://www.grisoft.com).
Version: 6.0.528 / Virus Database: 324 - Release Date: 2003-10-16