Distributed memory implementation of the HSS (Hierarchically Semi-Separable) matrix format. More...

#include <HSSMatrixMPI.hpp>

Inheritance diagram for strumpack::HSS::HSSMatrixMPI< scalar_t >:

Collaboration diagram for strumpack::HSS::HSSMatrixMPI< scalar_t >:

Public Member Functions
	HSSMatrixMPI (const DistM_t &A, const opts_t &opts)

	HSSMatrixMPI (const structured::ClusterTree &t, const DistM_t &A, const opts_t &opts)

	HSSMatrixMPI (const structured::ClusterTree &t, const BLACSGrid *g, const opts_t &opts)

	HSSMatrixMPI (std::size_t m, std::size_t n, const BLACSGrid *Agrid, const dmult_t &Amult, const delem_t &Aelem, const opts_t &opts)

	HSSMatrixMPI (std::size_t m, std::size_t n, const BLACSGrid *Agrid, const dmult_t &Amult, const delem_blocks_t &Aelem, const opts_t &opts)

	HSSMatrixMPI (const structured::ClusterTree &t, const BLACSGrid *Agrid, const dmult_t &Amult, const delem_t &Aelem, const opts_t &opts)

	HSSMatrixMPI (kernel::Kernel< real_t > &K, const BLACSGrid *Agrid, const opts_t &opts)

	HSSMatrixMPI (const HSSMatrixMPI< scalar_t > &other)

	HSSMatrixMPI (HSSMatrixMPI< scalar_t > &&other)=default

HSSMatrixMPI< scalar_t > &	operator= (const HSSMatrixMPI< scalar_t > &other)

HSSMatrixMPI< scalar_t > &	operator= (HSSMatrixMPI< scalar_t > &&other)=default

std::unique_ptr< HSSMatrixBase< scalar_t > >	clone () const override

const HSSMatrixBase< scalar_t > *	child (int c) const

HSSMatrixBase< scalar_t > *	child (int c)

const BLACSGrid *	grid () const override

const BLACSGrid *	grid (const BLACSGrid *grid) const override

const BLACSGrid *	grid_local () const override

const MPIComm &	Comm () const

MPI_Comm	comm () const

int	Ptotal () const override

int	Pactive () const override

void	compress (const DistM_t &A, const opts_t &opts)

void	compress (const dmult_t &Amult, const delem_t &Aelem, const opts_t &opts)

void	compress (const dmult_t &Amult, const delem_blocks_t &Aelem, const opts_t &opts)

void	compress (const kernel::Kernel< real_t > &K, const opts_t &opts)

void	factor () override

void	partial_factor ()

void	solve (DistM_t &b) const override

void	forward_solve (WorkSolveMPI< scalar_t > &w, const DistM_t &b, bool partial) const override

void	backward_solve (WorkSolveMPI< scalar_t > &w, DistM_t &x) const override

DistM_t	apply (const DistM_t &b) const

DistM_t	applyC (const DistM_t &b) const

void	mult (Trans op, const DistM_t &x, DistM_t &y) const override

scalar_t	get (std::size_t i, std::size_t j) const

DistM_t	extract (const std::vector< std::size_t > &I, const std::vector< std::size_t > &J, const BLACSGrid *Bgrid) const

std::vector< DistM_t >	extract (const std::vector< std::vector< std::size_t > > &I, const std::vector< std::vector< std::size_t > > &J, const BLACSGrid *Bgrid) const

void	extract_add (const std::vector< std::size_t > &I, const std::vector< std::size_t > &J, DistM_t &B) const

void	extract_add (const std::vector< std::vector< std::size_t > > &I, const std::vector< std::vector< std::size_t > > &J, std::vector< DistM_t > &B) const

void	Schur_update (DistM_t &Theta, DistM_t &Vhat, DistM_t &DUB01, DistM_t &Phi) const

void	Schur_product_direct (const DistM_t &Theta, const DistM_t &Vhat, const DistM_t &DUB01, const DistM_t &Phi, const DistM_t &_ThetaVhatC, const DistM_t &VhatCPhiC, const DistM_t &R, DistM_t &Sr, DistM_t &Sc) const

std::size_t	max_rank () const

std::size_t	total_memory () const

std::size_t	total_nonzeros () const

std::size_t	total_factor_nonzeros () const

std::size_t	max_levels () const

std::size_t	rank () const override

std::size_t	memory () const override

std::size_t	nonzeros () const override

std::size_t	factor_nonzeros () const override

std::size_t	levels () const override

void	print_info (std::ostream &out=std::cout, std::size_t roff=0, std::size_t coff=0) const override

DistM_t	dense () const

void	shift (scalar_t sigma) override

const TreeLocalRanges &	tree_ranges () const

void	to_block_row (const DistM_t &A, DenseM_t &sub_A, DistM_t &leaf_A) const override

void	allocate_block_row (int d, DenseM_t &sub_A, DistM_t &leaf_A) const override

void	from_block_row (DistM_t &A, const DenseM_t &sub_A, const DistM_t &leaf_A, const BLACSGrid *lgrid) const override

void	delete_trailing_block () override

void	reset () override

Public Member Functions inherited from strumpack::HSS::HSSMatrixBase< scalar_t >
	HSSMatrixBase (std::size_t m, std::size_t n, bool active)

virtual	~HSSMatrixBase ()=default

	HSSMatrixBase (const HSSMatrixBase< scalar_t > &other)

HSSMatrixBase< scalar_t > &	operator= (const HSSMatrixBase< scalar_t > &other)

	HSSMatrixBase (HSSMatrixBase &&h)=default

HSSMatrixBase &	operator= (HSSMatrixBase &&h)=default

virtual std::unique_ptr< HSSMatrixBase< scalar_t > >	clone () const =0

std::pair< std::size_t, std::size_t >	dims () const

std::size_t	rows () const override

std::size_t	cols () const override

bool	leaf () const

virtual std::size_t	factor_nonzeros () const

const HSSMatrixBase< scalar_t > &	child (int c) const

HSSMatrixBase< scalar_t > &	child (int c)

bool	is_compressed () const

bool	is_untouched () const

bool	active () const

virtual std::size_t	levels () const =0

virtual void	print_info (std::ostream &out=std::cout, std::size_t roff=0, std::size_t coff=0) const =0

void	set_openmp_task_depth (int depth)

virtual void	shift (scalar_t sigma) override=0

virtual void	draw (std::ostream &of, std::size_t rlo, std::size_t clo) const

virtual void	forward_solve (WorkSolveMPI< scalar_t > &w, const DistM_t &b, bool partial) const

virtual void	backward_solve (WorkSolveMPI< scalar_t > &w, DistM_t &x) const

virtual const BLACSGrid *	grid () const

virtual const BLACSGrid *	grid (const BLACSGrid *local_grid) const

virtual const BLACSGrid *	grid_local () const

virtual int	Ptotal () const

virtual int	Pactive () const

virtual void	to_block_row (const DistM_t &A, DenseM_t &sub_A, DistM_t &leaf_A) const

virtual void	allocate_block_row (int d, DenseM_t &sub_A, DistM_t &leaf_A) const

virtual void	from_block_row (DistM_t &A, const DenseM_t &sub_A, const DistM_t &leaf_A, const BLACSGrid *lg) const

Public Member Functions inherited from strumpack::structured::StructuredMatrix< scalar_t >
virtual	~StructuredMatrix ()=default

virtual std::size_t	rows () const =0

virtual std::size_t	cols () const =0

virtual std::size_t	memory () const =0

virtual std::size_t	nonzeros () const =0

virtual std::size_t	rank () const =0

virtual std::size_t	local_rows () const

virtual std::size_t	begin_row () const

virtual std::size_t	end_row () const

virtual const std::vector< int > &	dist () const

virtual const std::vector< int > &	rdist () const

virtual const std::vector< int > &	cdist () const

virtual void	mult (Trans op, const DenseMatrix< scalar_t > &x, DenseMatrix< scalar_t > &y) const

void	mult (Trans op, int m, const scalar_t x, int ldx, scalar_t y, int ldy) const

virtual void	mult (Trans op, const DistributedMatrix< scalar_t > &x, DistributedMatrix< scalar_t > &y) const

virtual void	factor ()

virtual void	solve (DenseMatrix< scalar_t > &b) const

virtual void	solve (int nrhs, scalar_t *b, int ldb) const

virtual void	solve (DistributedMatrix< scalar_t > &b) const

virtual void	shift (scalar_t s)

Friends
class	DistSamples< scalar_t >

template<typename T >
void	apply_HSS (Trans ta, const HSSMatrixMPI< T > &a, const DistributedMatrix< T > &b, T beta, DistributedMatrix< T > &c)

Detailed Description

template<typename scalar_t>
class strumpack::HSS::HSSMatrixMPI< scalar_t >

Distributed memory implementation of the HSS (Hierarchically Semi-Separable) matrix format.

This is for non-symmetric matrices, but can be used with symmetric matrices as well. This class inherits from StructuredMatrix.

Template Parameters

scalar_t Can be float, double, std:complex<float> or std::complex<double>.

See also: HSSMatrix, structured::StructuredMatrix

Member Function Documentation

◆ allocate_block_row()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::allocate_block_row	(	int	d,
		DenseM_t &	sub_A,
		DistM_t &	leaf_A
	)		const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ backward_solve()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::backward_solve	(	WorkSolveMPI< scalar_t > &	w,
		DistM_t &	x
	)		const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ clone()

template<typename scalar_t >

std::unique_ptr< HSSMatrixBase< scalar_t > > strumpack::HSS::HSSMatrixMPI< scalar_t >::clone ( ) const

overridevirtual

Clone this HSS matrix. TODO delete this??

Returns: std::unique_ptr to a clone of this HSS matrix.

Implements strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ factor()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::factor ( )

overridevirtual

Compute a factorization (or the inverse) of this matrix, to be used later for solving linear systems. The actual type of factorization depends on the StructuredMatrix::Type of this matrix.

Reimplemented from strumpack::structured::StructuredMatrix< scalar_t >.

◆ factor_nonzeros()

template<typename scalar_t >

std::size_t strumpack::HSS::HSSMatrixMPI< scalar_t >::factor_nonzeros ( ) const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ forward_solve()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::forward_solve	(	WorkSolveMPI< scalar_t > &	w,
		const DistM_t &	b,
		bool	partial
	)		const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ from_block_row()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::from_block_row	(	DistM_t &	A,
		const DenseM_t &	sub_A,
		const DistM_t &	leaf_A,
		const BLACSGrid *	lgrid
	)		const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ grid() [1/2]

template<typename scalar_t >

const BLACSGrid * strumpack::HSS::HSSMatrixMPI< scalar_t >::grid ( ) const

inlineoverridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ grid() [2/2]

template<typename scalar_t >

const BLACSGrid * strumpack::HSS::HSSMatrixMPI< scalar_t >::grid ( const BLACSGrid * grid ) const

inlineoverridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ grid_local()

template<typename scalar_t >

const BLACSGrid * strumpack::HSS::HSSMatrixMPI< scalar_t >::grid_local ( ) const

inlineoverridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ levels()

template<typename scalar_t >

std::size_t strumpack::HSS::HSSMatrixMPI< scalar_t >::levels ( ) const

overridevirtual

Return the number of levels in the HSS matrix.

Returns: Number of HSS levels (>= 1).

Implements strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ memory()

template<typename scalar_t >

std::size_t strumpack::HSS::HSSMatrixMPI< scalar_t >::memory ( ) const

overridevirtual

Return the total amount of memory used by this matrix, in bytes.

Returns: Memory usage in bytes.

See also: nonzeros

Implements strumpack::structured::StructuredMatrix< scalar_t >.

◆ mult()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::mult	(	Trans	op,
		const DistM_t &	x,
		DistM_t &	y
	)		const

overridevirtual

Multiply the StructuredMatrix (A) with a dense matrix: y = op(A)*x. x and y are 2d block cyclic.

Parameters

op	take transpose/conjugate or not
x	matrix, x.rows() == op(A).cols()
y	matrix, y.cols() == x.cols(), y.rows() == A.rows()

Reimplemented from strumpack::structured::StructuredMatrix< scalar_t >.

◆ nonzeros()

template<typename scalar_t >

std::size_t strumpack::HSS::HSSMatrixMPI< scalar_t >::nonzeros ( ) const

overridevirtual

Return the total number of nonzeros stored by this matrix.

Returns: Nonzeros in the matrix representation.

See also: memory

Implements strumpack::structured::StructuredMatrix< scalar_t >.

◆ Pactive()

template<typename scalar_t >

int strumpack::HSS::HSSMatrixMPI< scalar_t >::Pactive ( ) const

inlineoverridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ print_info()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::print_info	(	std::ostream &	out = `std::cout`,
		std::size_t	roff = `0`,
		std::size_t	coff = `0`
	)		const

overridevirtual

Print info about this HSS matrix, such as tree info, ranks, etc.

Parameters

out	Stream to print to, defaults to std::cout
roff	Row offset of top left corner, defaults to 0. This is used to recursively print the tree, you can leave this at the default.
coff	Column offset of top left corner, defaults to 0. This is used to recursively print the tree, you can leave this at the default.

Implements strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ Ptotal()

template<typename scalar_t >

int strumpack::HSS::HSSMatrixMPI< scalar_t >::Ptotal ( ) const

inlineoverridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ rank()

template<typename scalar_t >

std::size_t strumpack::HSS::HSSMatrixMPI< scalar_t >::rank ( ) const

overridevirtual

Return the maximum rank of this matrix over all low-rank compressed blocks.

Returns: Maximum rank.

Implements strumpack::structured::StructuredMatrix< scalar_t >.

◆ shift()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::shift ( scalar_t sigma )

overridevirtual

Apply a shift to the diagonal of this matrix. Ie, this += sigma * I, with I the identity matrix. Call this after compression.

Parameters

sigma Shift to be applied to the diagonal.

Implements strumpack::HSS::HSSMatrixBase< scalar_t >.

◆ solve()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::solve ( DistM_t & b ) const

overridevirtual

Solve a linear system A*x=b, with this StructuredMatrix (A). This solve is done in-place.

Parameters

b	right-hand side, b.rows() == A.cols(), will be overwritten with the solution x.

Reimplemented from strumpack::structured::StructuredMatrix< scalar_t >.

◆ to_block_row()

template<typename scalar_t >

void strumpack::HSS::HSSMatrixMPI< scalar_t >::to_block_row	(	const DistM_t &	A,
		DenseM_t &	sub_A,
		DistM_t &	leaf_A
	)		const

overridevirtual

Reimplemented from strumpack::HSS::HSSMatrixBase< scalar_t >.

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

/home/pieterg/LBL/STRUMPACK/src/HSS/HSSMatrixMPI.hpp

Public Member Functions

Friends

Detailed Description

Member Function Documentation

◆ allocate_block_row()

◆ backward_solve()

◆ clone()

◆ factor()

◆ factor_nonzeros()

◆ forward_solve()

◆ from_block_row()

◆ grid() [1/2]

◆ grid() [2/2]

◆ grid_local()

◆ levels()

◆ memory()

◆ mult()

◆ nonzeros()

◆ Pactive()

◆ print_info()

◆ Ptotal()

◆ rank()

◆ shift()

◆ solve()

◆ to_block_row()