vasp 4.6 optics on itantium2

[lawson]

I am having problems with the vasp 4.6 (mpi&serial) optics routines
(LOPTICS=.TRUE.,NPAR=1) on our Itantium2 cluster (ifort 9.1 compiler).
When I run the serial version on a Linux/Pentium system (pgf90 compiler),
it works fine. Diamond is my test case and I can reproduce the plots in
Adolph et al, PRB, (2001).

For the Itantium2 version, however, the NABIJ matrix elements in the
OPTIC file are very different than those I get from the Pentium/PGF90
system. The other parts of the OPTIC file seem ok, however. If I change
compiler options for ifort, the values of NABIJ change, but I still get the
wrong answer. Also, if I change the number of processors N=1,2,8, in the
MPI version, I get different results. I have tried OFLAG = -O0 and also
the ifort 7.1 compiler with no luck.

I am using the following settings:
CPP_ = ./preprocess <$*.F | /usr/bin/cpp -P -C -traditional >$*$(SUFFIX)
CPP = $(CPP_) -DMPI -DHOST=\"LinuxIFC\" -DIFC \
-Dkind8 -DNGZhalf -DCACHE_SIZE=4000 -DPGF90 -Davoidalloc \
-DMPI_BLOCK=500 -DPROC_GROUP=8 -Dpro_loop -Duse_collective \
-DRPROMU_DGEMV -DRACCMU_DGEMV
FFLAGS = -FR -lowercase -cm -w95 -tpp2 -safe_cray_ptr
OFLAG=-O3 -unroll0 -ivdep_parallel -fno-alias
LIB = -L../vasp.4.lib -ldmy \
../vasp.4.lib/linpack_double.o $(LAPACK) \
$(BLAS) -lmpi -lg2c
optics.o: optics$(SUFFIX)
$(FC) $(FFLAGS) $(OFLAG) -c optics$(SUFFIX)

Any suggestions about proper compiler options and/or any other fixes
would be most appreciated.
Thanks!

[lawson]

Optics seems to work on my itanitum2 system. The individual
elements of NABIJ are still very different on the different
platforms (for the same vasp parameters, NKPTS, NBANDS, etc),
but plots of EPS and JDOS look the same, so they seem to
be converged in some sense.

[FastForward]

Please mind that the phase is arbitrary for each state, and thus for the NABIJ elements. However the absolute value needs to be the same on all platforms.