Solves Hermitian eigenvalue problem by the divide and conquer algorithm. More...

Private Member Functions
subroutine	scalafx_pheevd_complex (aa, desca, ww, zz, descz, jobz, uplo, ia, ja, iz, jz, work, rwork, iwork, allocfix, info)
	Solves Hermitian eigenvalue problem by the divide and conquer algorithm.

subroutine	scalafx_pheevd_dcomplex (aa, desca, ww, zz, descz, jobz, uplo, ia, ja, iz, jz, work, rwork, iwork, allocfix, info)
	Solves Hermitian eigenvalue problem by the divide and conquer algorithm.

Detailed Description

Solves Hermitian eigenvalue problem by the divide and conquer algorithm.

Member Function/Subroutine Documentation

subroutine scalapackfx_module::scalafx_pheevd::scalafx_pheevd_complex	(	complex(sp), dimension(:,:), intent(inout)	aa,
		integer, dimension(dlen_), intent(in)	desca,
		real(sp), dimension(:), intent(out)	ww,
		complex(sp), dimension(:,:), intent(out)	zz,
		integer, dimension(dlen_), intent(in)	descz,
		character, intent(in), optional	jobz,
		character, intent(in), optional	uplo,
		integer, intent(in), optional	ia,
		integer, intent(in), optional	ja,
		integer, intent(in), optional	iz,
		integer, intent(in), optional	jz,
		complex(sp), dimension(:), intent(inout), optional, allocatable	work,
		real(sp), dimension(:), intent(inout), optional, allocatable	rwork,
		integer, dimension(:), intent(inout), optional, allocatable	iwork,
		logical, intent(in), optional	allocfix,
		integer, intent(out), optional	info
	)

private

Solves Hermitian eigenvalue problem by the divide and conquer algorithm.

Parameters

aa	Matrix to diagonalize (A).
desca	Descriptor of matrix A.
ww	Eigenvalues on exit.
zz	Eigenvectors on exit (Z).
descz	Descriptor of the eigenvector matrix.
jobz	Job type (default: "V")
uplo	Upper or lower diagonal matrix (default: "L")
ia	First row of the submatrix A (default: 1)
ja	First column of the submatrix A (default: 1)
iz	First row of the submatrix Z (default: 1)
jz	First column of the submatrix Z (default: 1)
work	Complex working array (if not specified, allocated automatically)
rwork	Real working array (if not specified, allocated automatically)
iwork	Integer working array (if not specified, allocated automatically)
allocfix	If yes, the routine tries to enlarge the workspace size as returned by the appropriate p?syevd() routine by some empirical values.
info	Info flag. If not specified and SCALAPACK calls returns nozero, subroutine stops.

Note

Unfortunately, SCALAPACK seems to return the wrong real work space sizes for many cases. This routine (when allocfix had been set to true) tries to improve on it by two ways:

It queries also the QR routine (p?heev) for workspace size and takes this, if bigger than returned by p?heevd. That should ensure that the pzunmtr() routine does not encounter any difficulties.

It additionally enlarges the real workspace size by the amount of memory needed by the pdlasrt() routine, to make sure this would not fail either.

Those fixes are empirical, may lead to oversized workspace allocations and probably would not even fix the allocation problem, but are the best we could find so far.

See Also: SCALAPACK documentation (routine p?heevd).

subroutine scalapackfx_module::scalafx_pheevd::scalafx_pheevd_dcomplex	(	complex(dp), dimension(:,:), intent(inout)	aa,
		integer, dimension(dlen_), intent(in)	desca,
		real(dp), dimension(:), intent(out)	ww,
		complex(dp), dimension(:,:), intent(out)	zz,
		integer, dimension(dlen_), intent(in)	descz,
		character, intent(in), optional	jobz,
		character, intent(in), optional	uplo,
		integer, intent(in), optional	ia,
		integer, intent(in), optional	ja,
		integer, intent(in), optional	iz,
		integer, intent(in), optional	jz,
		complex(dp), dimension(:), intent(inout), optional, allocatable	work,
		real(dp), dimension(:), intent(inout), optional, allocatable	rwork,
		integer, dimension(:), intent(inout), optional, allocatable	iwork,
		logical, intent(in), optional	allocfix,
		integer, intent(out), optional	info
	)

private

Solves Hermitian eigenvalue problem by the divide and conquer algorithm.

Parameters

aa	Matrix to diagonalize (A).
desca	Descriptor of matrix A.
ww	Eigenvalues on exit.
zz	Eigenvectors on exit (Z).
descz	Descriptor of the eigenvector matrix.
jobz	Job type (default: "V")
uplo	Upper or lower diagonal matrix (default: "L")
ia	First row of the submatrix A (default: 1)
ja	First column of the submatrix A (default: 1)
iz	First row of the submatrix Z (default: 1)
jz	First column of the submatrix Z (default: 1)
work	Complex working array (if not specified, allocated automatically)
rwork	Real working array (if not specified, allocated automatically)
iwork	Integer working array (if not specified, allocated automatically)
allocfix	If yes, the routine tries to enlarge the workspace size as returned by the appropriate p?syevd() routine by some empirical values.
info	Info flag. If not specified and SCALAPACK calls returns nozero, subroutine stops.

Note

Unfortunately, SCALAPACK seems to return the wrong real work space sizes for many cases. This routine (when allocfix had been set to true) tries to improve on it by two ways:

It queries also the QR routine (p?heev) for workspace size and takes this, if bigger than returned by p?heevd. That should ensure that the pzunmtr() routine does not encounter any difficulties.

It additionally enlarges the real workspace size by the amount of memory needed by the pdlasrt() routine, to make sure this would not fail either.

Those fixes are empirical, may lead to oversized workspace allocations and probably would not even fix the allocation problem, but are the best we could find so far.

See Also: SCALAPACK documentation (routine p?heevd).

The documentation for this interface was generated from the following file:

scalapackfx.F90

Private Member Functions

Detailed Description

Member Function/Subroutine Documentation