[gmx-developers] CVS Version - segmentation faults in double precision

[Berk Hess]

Marc Baaden wrote:

>>>>Berk Hess said:
>>>>        
>>>>
> >> I did only --disable-float (plus some prefix stuff).
> >> Are you sure you used the right binary/compilation?
> >> At the top of the md.log file you can now see when the code was compiled.
>
>Yes. Verified.
>I also just recompiled with only the --disable-float option.
>Same result: segmentation fault.
>Something very fishy seems to be going on.
>Any suggestions ?
>
>
>The run ends like this:
>[..]
>Back Off! I just backed up ener.edr to ./#ener.edr.3#
>Steepest Descents:
>   Tolerance (Fmax)   =  5.00000e+02
>   Number of steps    =          200
>s.sh: line 2:   825 Segmentation fault      mdrun_d -v -s em_vide -o em_vide -c after_em_vide -g em_videlog
>
>
>I have put the output of running with -debug here:
>http://www.shaman.ibpc.fr/mdrun_d.log
>
>
>What else could I do to find the origin for this segmentation fault ?
>
>The weird thing as I said is that it runs fine in single precision.
>What kind of stuff would depend on precision ??
>  
>
It could be many things.
Compilation problems for instance.
The first thing to try is check out a fresh CVS tree and compile only a 
double precision
version.
It could depend on the machine type. I am running a 64 bit compilation 
on an AMD64.
On thing to check could be doing:
setenv NOASSEMBLYLOOPS
and then run, if this runs there is a problem with running the assembly 
loops.

You can also use the debugger.
Compile with:
make CFLAGS=-g
and then run mdrun with gdb.

Berk.