Using SCALAPACKFX

Before you can use the SCALAPACKFX routines, you need basically the following two steps.

1. Import the module libscalapackfx_module in your routines.

2. Initialize a grid for the BLACS-communcation using type(blacsgrid).

Below you find an example draft for reading in a matrix from a file, distributing it across the processes and finally diagonalizing it:

program test_diag
  use libscalapackfx_module

  integer, parameter :: nprow = 4, npcol = 4   ! process rows/cols
  integer, parameter :: bsize = 64             ! block size
  integer, parameter :: nn = 1000              ! matrix size

  type(blacsgrid) :: mygrid  ! BLACS grid descriptor
  real(dp), allocatable :: aa(:,:), bb(:,:), eigvecs(:,:), eigvals(:)
  integer :: desc(DLEN_)    ! matrix descriptor
  integer :: mloc, nloc     ! nr. of local rows/columns of the matrices
  type(linecomm) :: distr               ! distributes matrix when read
  real(dp), allocatable :: iobuffer(:)  ! buffer during read

  ! Initialize your BLACS grid
  call mygrid%gridinit(nprow, npcol)

  ! Allocate the local part of the distributed matrices
  call scalafx_getdescriptor(mygrid, nn, nn, bsize, bsize, desc)
  call scalafx_getlocalshape(mygrid, desc, mloc, nloc)
  allocate(aa(mloc, nloc))
  allocate(bb(mloc, nloc))
  allocate(eigvecs(mloc, nloc))
  allocate(eigvals(nn))
  ...
  ! Here comes the code which distributes your matrix
  ! You can use the linecomm type to distribute it if you read it from file
  if (mygrid%lead) then
      allocate(iobuffer(nn))
  end if
  call distr%init(desc, "c")
  do icol = 1, nn
    if (mygrid%lead) then
      read(12, *) iobuffer(:)
      call distr%setline_lead(mygrid, icol, iobuffer, aa)
    else
      call distr%setline_follow(mygrid, icol, mtxloc)
    end if
  end do
  ...
  ! Get eigenvalues (on all nodes) and eigenvectors (distributed)
  call psygvd(aa, desc, bb, desc, eigvals, eigvecs, desc, jobz="V", uplo="L")
  ...
end program test_diag

Have a look at the test folder in the source tree for further examples.