GitLab

...for non-spinpolarized systems.

The problem was a "double free or corruption" error. It was related to the
ylm array in eigen/hsmt_ab.f90 being allocated on the heap and not
being a stack array. It seems to be a requirement that OpenMP private
variables are stack variables. At least putting ylm on the stack
seems to have fixed the problem.

The problem was that the 2nd dimension of vCoul%pw was set to 2 by the assignment
modified in this commit. This lead to an MPI broadcasting problem as it was only
done for rank 0. At some point one should check whether we actually need more
than the 2 spin dimensions for this potential contribution.

It turns out that the different BLACS descriptors for the matrices have different
context handles. This is not allowed in the pdsygvx call. I now explicitly set
the context handles to the same value.

...Unfortunately this also brings back the never ending execution of some of
the ctest test cases. There has to be another bug.

Note: This seems to have introduced some bugs. ...to be corrected soon.

The remaining problem were the jg and gj arrays that were needed in
other spin dimensions than in the other routines.

...Fe_bct_SOCXML test is still not working

...it is not initialized properly any more.

The test failed because the unrotated input density was not kept for the mixer.
I solved it by introducing an inDenRot variable that is a copy of inDen and then
gets rotated in rhodirgen if that soubroutine is in the program flow. It is only
used as an input to those subroutines that require the rotated input density or
are insensitive to it.

The convolution with the step function has been removed from vmatgen as this
is done in vgen already.

I had to compensate the newly present diagonal term in the td%tuu,... arrays in force/force_a21.F90.