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

[Berk Hess]

Berk Hess wrote:

>
>> Thanks for these precious suggestions.
>>
>> Running with NOASSEMBLYLOOPS works ... !!
>> ... so what exactly means problem with the assembly loops ?
>> A problem (bug) in Gromacs, or some problem with the assembler
>> used on my machine locally, or .. ?
>>  
>>
> The problem could be in many places.
> What kind of processor are you using?
> I saw you are using a 32 bit compilation.
>
> I could try to run a 32 bit double precision version.

The 32 bit double precision version also crashes on my AMD64 machine.
If I compile only nonbonded.c with -g it does not crash, but the fr->tab14
is corrupted, already immediately when do_nonbonded is entered.
This indicates that there probably is some pointer access problem
in nonbonded.c.
But this is hard to trace when -g compilation does not crash.

We have efence, but this only runs with 32 bit on our old AMD Athlons,
there I get:
Program received signal SIGILL, Illegal instruction.
0x0811fd5e in nb_kernel_ia32_sse2_test_asm ()

Maybe Erik can comment on this?

Berk.