GitLab

The threshold was used at several places to decide whether two values
are identical, on which side of a plane a point lies, and so on. The
strict value before lead to problems as some points were not considered
to be identical due to numerical reasons, although they were meant to
be identical. The larger threshold now solves the problems.

However, if more problems like this arise it should be checked whether
the threshold has to have different values at different places in the
code. Also additional code may be required to treat certain cases in
a special way.

The Broyden mixing for the density matrix was broken because
in the routine main/mix.F90 one of the two calls to brysh1 was wrong.
In both calls the old density matrix was passed to brysh1, but in
one of the calls this should be the new density matrix. This is
corrected now.

An error occured because an FFT mesh size had too large prime factors.
This lead to too small array sizes in math/cfft.F. Gustav had a
version of this routine in which this error was already fixed. I
pasted his changes to math/cfft.F.

The debug version of fleur coplained that some variable
was not initialized. This was due to a code path invoked by
a bad conventional inp file. Nevertheless I added the
missing initialization.

Added debug option to configure script
TODO:
ELPA version on JURECA can not be used. Interface of ELPA is changed :-(

nbsz determines the size of the region in reciprocal space considered
to construct the IBZ. Up to now this variable was set to 3 in julia.f.
But this value is not necessarily save. We now calculate a save but
also conservative value for this now local valiable nbszLocal in
brzone2.f90.

Also in this commit: Added build.* to .gitignore.

This commit and the 2 before made init/brzone2.f90 faster by
introducing or tightening constraints to filter out possible
boundary planes of the IBZ at an early stage of the algorithm.

I hope I didn't go too far.

Since the routine brzone featured many bugs that occured in special
cases I now reprogrammed the functionality of the routine in the
new subroutine brzone2. The new algorithm is simpler and should
be less vulnerable to the mentioned special cases. It also is slightly
slower. The new routine takes about 2 seconds.

By commenting the new routine out and the old one in both algorithms
can directly be interchanged.

configure.sh script is doing part of the job now
more details should be found in README (later :-) )

This looks a little bit nicer.

I put this in in a slightly dirty way as I use a GO TO. But I
think it is ok.

Also in this commit: Introduction of a larger threshold in vec_for_lo
as it was done by Christoph Friedrich in the old Fleur version
I was comparing with.

to stop SCF iterations if the remaining time is not sufficient

This lead to wrong symmetry operations

...and a small initialization in init/inped.F90.

...since we don't have these executables any more.

The problem was due to a wrong mpi irank being passed to juDFT_end.