[gmx-users] Who uses Gromacs on Power4 ?

Marc Baaden baaden at smplinux.de
Wed Jun 2 12:49:36 CEST 2004

Hi Erik,

thanks for your comments. There are some pecularities which make me wonder
a bit whether it is really a problem in the actual simulation code, or rather
a typo somewhere in a more general file (makefile or so).
This is because:

- as I said same machine no problem with gmx 3.1.4, only with 3.2.1

- the error occurs at the very beginning, before even the first step
  is calculated .. this is why I am a bit sceptical about the vectorized
  sqrt being responsible. But it seems indeed the only part that actually
  changed in fnbf.c. I will try it.

- I don't think it is a global variable problem, as gmx 3.1.4 uses the
  same variable names for _buf1/_buf2 in this routine and runs fine

I'll keep on searching .. thanks for your help,

NB: I still don't know whether anybody successfully uses gmx 3.2.1 on
    power3 or power4 ..

>>> Erik Lindahl said:
 >> Hi Marc,
 >> Power3 and Power4 are virtually identical, so this is more likely an 
 >> issue with a new version of the IBM compilers.
 >> First, you can try and disable the vectorized sqrt with the option 
 >> --disable-vectorized-sqrt. I have no idea if there is any performance 
 >> hit or gain on power4.
 >> Second, it doesn't look like a code problem. It's pretty 
 >> straightforward, and all variables have been initialized. We also check 
 >> if the pointers are null and use malloc instead of realloc in that 
 >> case.
 >> In theory there could also be global variables from the OS or compiler 
 >> called _buf1/_buf2, although you would probably have seen lots of error 
 >> messages about it. Anyway - it's a long shot, but it could be worth 
 >> trying to rename them to _gmxbuf1/_gmxbuf2.
 >> Cheers,
 >> Erik
 Dr. Marc Baaden  - Institut de Biologie Physico-Chimique, Paris
 mailto:baaden at smplinux.de      -      http://www.marc-baaden.de
 FAX: +49 697912 39550  -  Tel: +33 15841 5176 ou +33 609 843217

More information about the gromacs.org_gmx-users mailing list