Bug #263

segfault on entry to ia64_single kernel

Added by Mark Abraham over 11 years ago. Updated almost 10 years ago.

Erik Lindahl
Target version:
Affected version - extra info:
Affected version:


The CVS version is doing something strange with the single-precision kernels on ia64. After correcting the last bug I posted about checking for a valid file point for the logfile, I found a further problem.

Using a selection of Intel compilers from versions 9, 10 and 11, and gcc 4.0, I get a segfault in the first non-neighbor-search kernel function to be called - 310 in my case. With NOASSEMBLYLOOPS set, the run goes fine. Stepping through with Totalview on four MPI processes, I can see that on program entry, the generic kernel table is fine, but static kernellist_ia64_single does not contain the addresses of the nb_kernelxxx_ia64_single functions. Later, when kernelptr in do_nonbonded gets such a pointer there's a segfault. I'm at a loss to understand how the static table can get mangled in this way.


#1 Updated by Mark Abraham over 11 years ago

Hmm I get the same issue with 4.0.2 and 4.0 (both bugfixed as above). On the theory that this is a linking issue, I am continuing to try to simplify things for the linker.

#2 Updated by Rossen Apostolov about 10 years ago

Mark, do you still have that issue with the current git master?

#3 Updated by Mark Abraham about 10 years ago

I no longer have access to such a machine, so am not sure whether it's still an issue.

I'm fairly sure that I never resolved this satisfactorily, and shifted my focus off the then-CVS version. It could easily have been a compiler/linker issue, rather than a GROMACS issue, as there were numerous threads on gmx-users complaining about that kind of machine (some SGI Altix) with various Intel compilers.

How about we resolve this as "WORKSFORME"?

#4 Updated by Rossen Apostolov almost 10 years ago

Closing the bug as WONTFIX. The assembly loops for ia64 will be deprecated in future releases. See also

Also available in: Atom PDF