On Thu, Feb 14, 2019, Stephan Schott wrote:
>OpenMPI 4 worked for me by using gcc 7 and including the
>"--enable-mpi1-compatibility"
>during compilation. gcc 8 is still not a supported compiler for CUDA afaik,
>so you should use gcc 7 with amber anyway.
I tried this: add "--enable-mpi1-compatibility" at about line 129
of AMBERHOME/AmberTools/src/configure_openmpi. (You may also have to adjust
the "mpidirs" variable at line 107 to include openmpi-4.0.0.)
I then used this script with gcc 8.2 to build openmpi. Seems to work
fine. So this looks like a workaround for now. But I hope we can
remove reliance on this flag soon, so as to track upstream development
more easily, because this flag is going away at some point, and because
most package managers won't set this flag.
Thanks to Stephan for this hint. Also, I may have missed gcc8-specific
problem (I'm on OSX 10.14), so please chime in if your experience is
different from mine.
...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Feb 15 2019 - 11:30:03 PST
