GitLab

The associaten with Cuda memory as Fortran arrays was sometimes
done wrongly , since not the number of array elements but instead
the number of array elements times the sizeof(datatype) was used.
This was a mix-up between the C allocation and the Fortran reshape.

This could lead to memory corruption

Fix a bug in elpa_multiply_a_b

See merge request !35

Remove some debug prints

See merge request !34

In elpa_multiply_a_b I wanted to call mpi_comm_rank but accidentally
called mpi_comm_size. This only affects the GPU initialization where
GPUs are assigned to CPU cores based on their MPI ranks.

Matrix redistribute

See merge request !31

Fixes

See merge request !30

The old code crashes when calling an MPI_Allreduce. The reason is that
part of the input array (t_tmp in this case) is not initialized.

Due to an error, on systems where AVX-2 (FMA) was supported, also
the "pure" AVX kernels used FMA instructions. This has been fixed
now

If a C program needs an underscore to link against a Fortran library
append it, otherwise not

Merge branch 'matrix_redistribute' of https://gitlab.mpcdf.mpg.de/elpa/elpa into matrix_redistribute

Elpa gpu pinned

See merge request !28