[gmx-users] known bug? compiling gromacs 3.2.1 with gcc 3.3.5 using -march=athlon-whatever gives strange mdrun results on AMD athlon systems

[Sabuj Pattanayek]

Hi,

Has anyone had problems with strange mdrun results after running a 
simulation with gromacs 3.2.1 compiled using -march=athlon-xp or 
-march=athlon-mp on athlon-xp and athlon-mp systems? For example here 
are some sections from some log files (naming is like this: 
energyMinimization_systemType_gccMarchCompileFlag):

this output is "correct" (from em-celeronM_pentium4.log):

Steepest Descents converged to Fmax < 1000 in 37 steps
Potential Energy  = -3.9011406e+04
Maximum force     =  9.3552039e+02 on atom 88
Norm of force     =  6.2108813e+03
 NODE (s)   Real (s)      (%)
Time:     23.000     23.000    100.0

"incorrect" output (from em-xp2800_athlon-xp.log)

Steepest Descents converged to machine precision in 15 steps,
but did not reach the requested Fmax < 1000.
Potential Energy  =            nan
Maximum force     =  2.2911844e+04 on atom 836
Norm of force     =  9.9849359e+04
NODE (s)   Real (s)      (%)
Time:      6.000      6.000    100.0

...compiling gromacs 3.2.1 on athlon systems with -march=i686 fixes this 
problem (from em-xp2800_i686.log):

Steepest Descents converged to Fmax < 1000 in 37 steps
Potential Energy  = -3.9011410e+04
Maximum force     =  9.3552039e+02 on atom 88
Norm of force     =  6.2109053e+03
NODE (s)   Real (s)      (%)
Time:     14.000     14.000    100.0

I found a similar post here 
http://www.gromacs.org/pipermail/gmx-users/2003-May/005614.html and 
under "Which machine do you recommend?" at 
http://www.gromacs.org/benchmarks/single.php it says athlon bugs were 
fixed in gromacs 3.2.1? Is this related to the same problem or is this a 
gcc bug? Is there even any loss of performance when compiling with i686 
rather than athlon-xp or -mp? I tried to compare with a pentium4 3.2 GHz 
machine with gromacs compiled using -march=pentium4 but the same 
simulation takes 50-60 seconds (slower than the celeron M 1.3GHz)! The 
compilations were done using gcc version 3.3.5, lam-mpi 7.0.6, fftw 
3.0.1, and using single precision.

Thanks,
Sabuj Pattanayek