GitLab

Lorenz Huedepohl authored May 24, 2018 and Pavel Kus committed May 24, 2018

Lead to the first configuration be tested twice

time_spent is not initialized. Should be dealt with in a better way

instead, blacs context has to be set beforehand

Added C interface for generalized eigen problem
Extended the functionality of test.c to check generalized eigenproblem,
both with and without already decomposed matrix

added possibility to re-use the already decomposed matrix B in the
generalized problem
Added respective test

To be used like this

   class(elpa_t), pointer      :: e
   class(elpa_autotune_t), pointer :: tune_state

   e => elpa_allocate()
   call e%set(...)
   [...]
   assert_elpa_ok(e%setup())

   tune_state => e%autotune_setup(ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_REAL)

   ! Autotuning loop, continues until all combinations have been tried
   do while (e%autotune_step(tune_state))
     ! Do the steps that are representative of your calculation
     call e%eigenvectors(a, ev, z, error)
   end do

   ! Fix best parameters, and de-allocate the autotune object
   call e%autotune_set_best(tune_state)
   call elpa_autotune_deallocate(tune_state)

This fixes commit 1cd2bb99, which attempted to plug the memory leak due
to unfreed MPI communicators.

The MPI communicators were never freed, this created a memory leak for
every elpa_allocate()/elpa_deallocate() cycle.

The timer object was not properly removed when elpa_destroy() was called
on an ELPA instance, leading to a memory leak

We got reports from a user that there were troubles with certain domain
decompositions. So far the tests only looked at (approximately) square
decompositions in column-major process order.

Now, a new class of tests loops over all possible decompositions
(row * col) for a given number of total tasks.

So far, we can not confirm that there are any problems, all
possibilities work as expected.

This prevented compilation with Intel 2018 Beta and solves issue #54. It
turns out you can only override type-bound procedures that are
accessible in the module where you override them, thus they have to be
private. A "deferred, private" type bound procedure could only be
overriden from the same module (or a submodule thereof).

Many thanks to Igor Vorobtsov from Intel for clarifying this.

It is not quite clear to me why GCC complains here when
WANT_SINGLE_PRECISION_REAL is not defined, as the variables are then not
used _at all_.

Nonetheless, this again prevents build on the SuSE OBS, so fix it to
make the compiler happy

These can be used internally by

  call self%set_private(name, value)

   or

  call self%get_private(name, value)

but are not valid for the self%set()/get() routines, i.e. are not
visible to users of ELPA.

The public "solve" method has been renamed "eigenvectors" since it
computes the eigenvalues and (at least parts of) the eigenvectors

Another routine "eigenvalues" will just compute the eigenvalues