GitLab

Do not consider basis functions removed due to ill-conditioning.

In the complex case, it is ok to take the square root of negative
numbers. Therefore, there is no need to check whether or not all
the occupation numbers are non-negative.

The code now handles occupation numbers that are all zero.

The code should now be able to handle a special case where the occupation
numbers are like 0, ..., 0, 1, ..., 1, 0, ..., 0.

Before using bisection to find the chemical potential, first check
if one of the interval bounds is already a solution.

* Use the correct version number 6.1.0
* Let CMake (>= 3.8) automatically add C99 compiler flags

This update mainly cleans up some unnecessary use of iso_c_bindings.
No difference should be expected on the ELSI side.

* USE_MPI_MODULE CMake keyword must be set
* ELPA generic kernel currently requires `-fallow-argument-mismatch`

Added proper support for setting ill-conditioning threshold.

We restore the option to skip the calculation of the eigenvectors of
the overlap matrix. Its eigenvectors are only needed if it's found
to be ill-conditioned.

ELSI now checks the user-specified number of poles *before* calling PEXSI.

AVX512-DQ is not available on KNL (Intel Xeon Phi). Only use AVX512-F
which is available on both Xeon Phi and Xeon Scalable.

Default is CPU and double precision.

Density matrix P = C * C^T. The ELPA matrix multiplication routine is GPU
accelerated, so it is faster than pdsyrk/pzherk if GPUs are available.

When using PEXSI, there was an `MPI_Allreduce` in `elsi_get_entropy`,
which means that `elsi_get_entropy` must be called by all MPI ranks.
This all reduce is now performed right after the PEXSI solver finishes
instead of when `elsi_get_entropy` is called.

Fixes #43

Next will move part of density matrix build to GPUs.

And better message in case of errors.

The transformation between generalized and standard eigenproblems makes use
of a matrix multiply routine elpa_multiply_a_b in ELPA. In this subroutine,
an mpi_comm_size call should have been mpi_comm_rank. This only affects the
GPU initialization where GPUs are assigned to CPU cores based on MPI ranks.

Next we will try using GPU in the construction of the density matrix from
the computed occupation numbers and eigenvectors.

Tests show that this new method is always slower than the existing code on
CPU-only platforms. And it does not support GPU acceleration at the moment.
If needed, it can be easily enabled in the future.

* ELPA 2020.05.001
* ELPA no longer compiles with NAG Fortran (GNU extensions)
* C compiler becomes mandatory
* MPI-3 becomes mandatory