OK: I can reproduce the reported problems with gfortran 4.6.1 (from macports).
The attached patch passes the test suite: basically, I fixed up the function
declarations in amg1r5.f, and removed the FCREAL8 flag from the Makefile.
(I did not yet do anything about the random() function inside amg1r5.f)
Note that trying to force the compiler to have default double precision, as
opposed to writing the desired code in the source itself, is dangerous. For
example, before removing the FCREAL8 flag, gfortran4.6 was making most arrays
in amg1r5.f as REAL*16 (!), i.e. it was taking a default real8, applying the
"double precision" implicit flag, and coming up with quad precision. I'm
pretty sure this is not what was intended, and might not be what other
compilers would do (including earlier versions of gfortran). Moral, I think,
is not to use such flags.
I'm kicking this over now to the UCI folks for further testing and evaluation.
Even though gcc 4.6 is still listed as "beta" and "prerelease", it seems clear
that people are starting to find it on their machines, so we will need to
issue at bugfix.
...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Jun 20 2011 - 07:00:03 PDT
