GitLab

Previously we checked ortonormality of eigenvectors by comparing
the matrix S^T * S to identity matrix. The new feature is also
checking just the diagonal of the matrix S^T * S. By that we get
the information, how far are the eigenvectors from having length 1.
If it turns out to be important, we could try to normalize them
at the end of elpa, which is simple (in contrast with enforcing
better orthogonality).

for mpi_allreduce in test_check_correctness

of _GEMM and P_GEMM calls

The functions in elpa_utilities are considered "ELPA internal",
i.e. the should not be accessible by the users and thus not
be part of the API.

Should be done in a more systematic way. In this case, in certain
configuration 1stage_analytic test pased, while 2stage_analytic tests
failed due to error larger then tolerance

Introducing new test in which matrix and its eigendecomposition is
known and thus can be easily created and checked directly, without the
need to use scalapack or any other communication (apart from reducing
error).

The test is based on the fact, that if L_A and S_A are eigenvalues and
eigenvectors of matrix A, respectively, and L_B and S_B eigenvalues and
eigenvectors of B, then kron(L_A, L_B) and kron (S_A, S_B) are
eigenvalues and eigenvectors of kron(A, B).
Since it is easy to know exact eigendecomposition of a small matrix (e.g.
2x2), and kron operator has very simple structure, we can construct
arbitrarily large matrix and its eigendecomposition. We only have to
select small matrices such that the resulting matrix has unique and
ordered eigenvalues, so that the checking of the result is than easy.
Each element of matrix, eigenvector matrix and eigenvalue vector can
be quickly computed independently, just using its global coordinates.

The test is currently limited to matrices of size 2^n, but by
storing eigendecompositions of more small matrices (e.g. 3x3 and 5x5) we
could construct any matrix of size 2^n*3^m*5^o, which would probably be
sufficient, since most often used sizes (150, 1000, 5000, 2000, 60000)
are of this form.

The routines "eigenvalues" and "solve_tridiagonal" are now
also tested directly with the new API (and not only via the
legacy API -> new API mapping)

Todo: adapt test generator script to contain some logic

In case of GPU build it is NOT sufficient to loop over all kernels:

For the GPU kernels one must ALSO set that GPUs should be used

In the current form, the assert macro is not usable, since it might
generate too long lines depending on the lenght of __FILE__ string.
This is a temporary fix to allow setting up tests on Minsky

the test failed, since THIS_REAL_ELPA_KERNEL_API has not been specified
as it is in other simmilar tests. This fixed the issue, but probably the
behavior of elpa_solve_evp_real_double should be changed not to check
value of THIS_REAL_ELPA_KERNEL_API for 1stage test

Another blocker for the SuSE build server: Even though the variable is
later not actually used by the function, a recent GCC complains about
the use of an uninitialized value for that variable.

This prevents build on the OBS, as they use compiliation flags that
detect such behaviour.

Error message there:

 ../test/C/driver/legacy_interface/legacy_real_driver_c_version.c:193:12: warning: 'THIS_REAL_ELPA_KERNEL_API' is used uninitialized in this function [-Wuninitialized]
     success = elpa_solve_evp_real_double(na, nev, a, na_rows, ev, z, na_rows, nblk, na_cols, mpi_comm_rows, mpi_comm_cols, my_mpi_comm_world, THIS_REAL_ELPA_KERNEL_API, useQr, useGPU,

 [...]

 I: Program is using uninitialized variables.
    Note the difference between "is used" and "may be used"
 W: elpa uninitialized-variable ../test/C/driver/legacy_interface/legacy_real_driver_c_version.c:193

Most of the test programs for the new interface were all essentially
copy&pasted from test.F90, now all of those directly use test.F90 with
appropriate preprocessor flags

- Remove all use of ELPA internal modules, the test programs now
  only use the public ELPA API. This is now enforced, by hiding the
  private modules

- Prefix all test internal modules with "test_" to make it really
  clear that these modules are not to be used by users.

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