The dependency of m_ts_options on atomlist causes a cascade of
further dependencies. This should be fixed.

  mnt: disabled openmp in fft.F
  doc: added more documentation for ts
  bug: fixed spin-box calculation of PDOS in SOC

mnt: disabled openmp in fft.F

Performance of the FFT module was not good under OpenMP. We might come back to this in the future.

See merge request !95

Fixes issue 186

Closes #186

See merge request !98

mnt: changed MPI_ISSend to MPI_ISend

See merge request !97

doc: added more documentation for ts

See merge request !94

mnt: added timers for easier profiling of Transiesta SCF iteration

See merge request !96

Only in a very minor circumstance will the former be better.
The use here is very restricted and we will resort to using
MPI_ISend instead.

One cannot use omp statements in pure routines unless
they have been declared.

Cleaned transiesta/tbtrans output files and memory management

See merge request !34

This enables one to pass a dummy argument and the
fortran std 2008 storage_size will take care of
evaluating its byte size.

It is not necessary for end-users to see
the entire BTD format. Instead we only print
the blocks in a sorted fashion. This is of course
not *true* in the sense that internally the BTD is
very different. However, it may give users a hint with
a much reduced output for large matrices.

Also ensured ts-deallocs does not interfere with regular
Siesta runs.

Updated copy_to_reference.sh script for TS tests.

Also cleaned many empty spaces.

The ts warning for non-bulk electrodes has been removed
as well as the notice about increasing precision. It
seems to just confuse users.

Also, fixed the excess charge print-out which only had E-1
which for initial SCF could be lower than 1.e-10 resulting
in overflow when printing.

Now init_electrode_HS has been moved to electrode_t%prepare_SE which
superseedes the prior implementation.
This means that basically everything that alters the electrode_t is
handled in the ts_electrode.F90 code.

Ran valgrind and fixed memory leaks upon siesta+tbtrans exit.
Also, fixed some real memory leaks which was due to twice reading
the electrode H and S. Certain arrays were just nullified instead
of deallocated; xa, lasto and isc_off.

Removed some output when creating the GF files. They didn't provide
any details not already present in the output.

This was an ancient leftover.

Fixed writing to a variable if that variable
is in the argument list.
One should not do
  write(var, *), 'tsaoeuhs', var

Also fixed a potential variable overflow
for TS.Analyze output.

When TBT.dH was provided a non-existing file, it just
continued without dH, this is not optimal.

Also changed the calculation of ETA which should now be
more precise for the first iteration.

Now there is a utility tscontour which only
reads in the chemical potential and the contours.

This enables one to directly write-out the
contours without reading in anything else.

This is nice for testing subsequent work by using
label.

Fixed delete/deallocations.

Now it is called electrode_t and all methods
are contained for OO paradigm. This cleans the
method quite a bit.

Now we just need to adapt the changes to allow
Elec as arguments (more clear than El).

Should be marginally faster.

Now we have eliminated many warnings related to
temporary array creations. Even though the warnings were
not effective (if using -fcheck=array-temps) they still
impacted performance. I have thus cleaned a large fraction
of them.

Some of the same issues are also updated in transiesta

I have also updated TriMat class for slightly better performance
of index-look ups. I hope this should give better performance.
There is now no if-statements in the index functions (and all
are pure)

I have filed questions related to temporary arrays here:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=48655

to figure out why I get better performance. I can't see
why if there is no code change. Basically I have relied
in many places about changing of dimensions for easier
subroutine handling, but apparently when the ranks change,
gfortran will try and create temporary arrays.

Changed netcdf-parallel writes to be collective

General recommendation is to use collective MPI-IO for
best performance. In this case it also seems easy since
it requires little change.

This is hard to detect for low core counts and requires
testing on larger machines.

I.e. on my small laptop I do not see any performance
increase.

Now tbtrans cleans all memory by calling the
appropriate reset functions.
This meant a lot of added *_reset routines for
the modules to ensure staticly memory is
deallocated.
So far I have no missing deallocates with
valgrind.

There are much edited code, but basically
nothing has changed. Only added
deallocates and ensured memory leaks are
closed.
Signed-off-by: Nick Papior <nickpapior@gmail.com>

Made use of byte_count in m_transiesta and m_tbtrans.
Generally this means removing code (which is good)!

Made only long integers accessible for calculating

Allowed tbtrans to write out BTD matrices (and fixed
it overwriting TS output).

Added padding and worksize arrays for TBtrans memory
estimation.

All CDF files now use a common routine
to store common data. This makes maintenance
more easy.

Added a byte_count module to accummulate
bytes for determining file sizes.
Signed-off-by: Nick Papior <nickpapior@gmail.com>

bug: fixed spin-box calculation of PDOS in SOC

See merge request !83