Hi,
On Tue, 8 Aug 2006, David A. Case wrote:
> On Tue, Aug 08, 2006, Scott Brozell wrote:
>
> > > Second, edit config.h (in $AMBERHOME/src) and substitute mpif90 (or
> > > the executable wrapper ) for "g95". This may help find the correct
> > > libraries to link in.
>
> > I am not convinced that this user's problem is due to our configure.
> > In effect configure implements your second suggestion.
>
> I don't see how configure "implements my second suggestion"(?) For options
> hpf90 and xlf90_suse, the "fc" (fortran compiler) variable is set to
> mpif90, but not for other input options.
For the other options there is something like this:
loadlib=`$MPI_HOME/bin/mpif90 -show | perl -p -e
which is suppose to in effect put all the flags in the wrapper into our
load command.
> One problem is that configure processes the mpi-related options first, and
> then the compiler-related options. This probably needs to be turned around.
Maybe, but sometimes the last options are the actually used options;
so putting our flags first may deactivate them.
> Basically, there are at least two problems that may require this change:
>
> 1. The latest openmpi implementation has an mpif.h file that itself
> has an 'include' statement inside it. Without using the mpif90 wrapper
> script, the code doesn't know how to resolve this interior include.
I havent installed openmpi yet, but I have several reasons to do it.
> 2. (Apparently) the "mpif90 -show" command for MPICH2 is not including
> -lfmpich as a library that is needed at link time, but it is in fact
> needed. (cf. line 314 of configure).
This is where Im especially skeptical (without trustworthy evidence
from a source better than an unknown reflector poster).
mpich2 is also on my todo list, but sometimes I move at relativistic
speeds on other projects, so from the Amber frame I am slowwwwwwwwwwww.
:)
> Biting the bullet and using the compiler wrappers provided by the MPI
> distributions is probably the simplest way to provide a fix here, but I am
> certainly open to alternative suggestions.
I agree, but there may be issues down that path; I dont remember
why we did it the way we did. First problems 1 and 2 should be verified.
Scott
Received on Sun Aug 13 2006 - 06:07:14 PDT
