[CPMD-list] Compiling cpmd v3.7 on IBM p690
Salomon Billeter
ZRLSRB at ch.ibm.com
Tue Feb 11 18:59:54 CET 2003
Dear Bernd,
> 1) I can only link when I add -qsmp=omp to the link flags, otherwise I
> get undefined symbols like
>
> ld: 0711-317 ERROR: Undefined symbol: ._xlsmpInitializeRTE
> ld: 0711-317 ERROR: Undefined symbol: ._xlsmpParallelDoSetup_TPO
> ld: 0711-317 ERROR: Undefined symbol: ._xlsmpParRegionSetup_TPO
>
> By the way, linking smp programs require the -qsmp flag as mentioned in
> the compiler user's guide. Maybe your xlf.cfg is different from ours?
the -qsmp=omp,noauto flag must have got lost during copy-paste. It was and
needs to be present for linking, you are right.
> 2) The executable still crashes, even when I use your makefile template
> with xlf/8:
> ...
> EXCHANGE ENERGY
> [GGA: J.P. PERDEW ET AL. PHYS. REV. B 46, 6671 (1992)]
> CORRELATION ENERGY
> [GGA: J.P. PERDEW ET AL. PHYS. REV. B 46, 6671 (1992)]
>
>
> ERROR: 0031-250 task 0: Segmentation fault
Do you have a coredump? Where does it crash? Please make sure that IRAT is
2 (irat.inc) if you are using my settings (32 bit integers).
> 3) What is the reason behind your suggestion of qarch=pwr3 on a pwr4
> architecture when using xlf/8, which is the first xlf that is able to
> optimize for pwr4?
For production, I am using pwr4. These settings were used for tests which
ran on both pwr3 and pwr4 machines. I am sorry, I should have mentioned
this. The executables tuned for pwr4 returned consistent results.
> 4) Is it possible that I send you a sample input including PPs that
> fails when I compile with -DFFT_ESSL, but works when I use -DFFT_DEFAULT
> (it is a bit too much text to write this inline)?
> By the way, I already posted this problem 10. Dec. 2002 to the cpmd
> mailing list, when I used v3.5.3. There I also explained the background
> and gave some hints that point to a bug in FFT from ESSL. Up to now I
> got no response from IBM.
No problem, please send them. For minimum interference with our
calculations, may I ask you to keep the system as small as possible?
Thanks.
Best regards,
Salomon
More information about the CPMD-list
mailing list