StrumpackSparseSolver is the main sequential or multithreaded sparse solver class. More...

#include <StrumpackSparseSolverBase.hpp>

Inheritance diagram for strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >:

Public Member Functions
	StrumpackSparseSolverBase (int argc, char *argv[], bool verbose=true, bool root=true)

	StrumpackSparseSolverBase (bool verbose=true, bool root=true)

virtual	~StrumpackSparseSolverBase ()

ReturnCode	reorder (int nx=1, int ny=1, int nz=1, int components=1, int width=1)

ReturnCode	reorder (const int *p, int base=0)

ReturnCode	factor ()

void	move_to_gpu ()

void	remove_from_gpu ()

ReturnCode	solve (const scalar_t b, scalar_t x, bool use_initial_guess=false)

ReturnCode	solve (const DenseM_t &b, DenseM_t &x, bool use_initial_guess=false)

SPOptions< scalar_t > &	options ()

const SPOptions< scalar_t > &	options () const

void	set_from_options ()

void	set_from_options (int argc, char *argv[])

int	maximum_rank () const

std::size_t	factor_nonzeros () const

std::size_t	factor_memory () const

int	Krylov_iterations () const

void	draw (const std::string &name) const

Detailed Description

template<typename scalar_t, typename integer_t = int>
class strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >

StrumpackSparseSolver is the main sequential or multithreaded sparse solver class.

This is the main interface to STRUMPACK's sparse solver. Use this for a sequential or multithreaded sparse solver. For the fully distributed solver, see StrumpackSparseSolverMPIDist.

Template Parameters

scalar_t	can be: float, double, std::complex<float> or std::complex<double>.
integer_t	defaults to a regular int. If regular int causes 32 bit integer overflows, you should switch to integer_t=int64_t instead. This should be a signed integer type.

See also: StrumpackSparseSolverMPIDist, StrumpackSparseSolverMPI

Constructor & Destructor Documentation

◆ StrumpackSparseSolverBase() [1/2]

template<typename scalar_t , typename integer_t = int>

strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::StrumpackSparseSolverBase	(	int	argc,
		char *	argv[],
		bool	verbose = `true`,
		bool	root = `true`
	)

Constructor of the StrumpackSparseSolver class, taking command line arguments.

Parameters

argc	number of arguments, i.e, number of elements in the argv array
argv	command line arguments. Add -h or –help to have a description printed
verbose	flag to enable/disable output to cout
root	flag to denote whether this process is the root MPI process, only the root will print certain messages to cout

◆ StrumpackSparseSolverBase() [2/2]

template<typename scalar_t , typename integer_t = int>

strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::StrumpackSparseSolverBase	(	bool	verbose = `true`,
		bool	root = `true`
	)

Constructor of the StrumpackSparseSolver class.

Parameters

verbose	flag to enable/disable output to cout
root	flag to denote whether this process is the root MPI process. Only the root will print certain messages

See also: set_from_options

◆ ~StrumpackSparseSolverBase()

template<typename scalar_t , typename integer_t = int>

virtual strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::~StrumpackSparseSolverBase ( )

virtual

(Virtual) destructor of the StrumpackSparseSolver class.

Member Function Documentation

◆ draw()

template<typename scalar_t , typename integer_t = int>

void strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::draw ( const std::string & name ) const

Create a gnuplot script to draw/plot the sparse factors. Only do this for small matrices! It is very slow!

Parameters

name	filename of the generated gnuplot script. Running gnuplot plotname.gnuplot will generate a pdf file.

◆ factor()

template<typename scalar_t , typename integer_t = int>

ReturnCode strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::factor ( )

Perform numerical factorization of the sparse input matrix.

This is the computationally expensive part of the code. However, once the factorization is performed, it can be reused for multiple solves. This requires that a valid matrix has been assigned to this solver. internally this check whether the matrix has been reordered (with a call to reorder), and if not, it will call reorder.

Returns: error code

See also: set_matrix, set_csr_matrix

◆ factor_memory()

template<typename scalar_t , typename integer_t = int>

std::size_t strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::factor_memory ( ) const

Return the amount of memory taken by the sparse factorization factors. This is the fill-in. It is simply computed as factor_nonzeros() * sizeof(scalar_t), so it does not include any overhead from the metadata for the datastructures. This should be called after the factorization. For the StrumpackSparseSolverMPI and StrumpackSparseSolverMPIDist distributed memory solvers, this routine is collective on the MPI communicator.

◆ factor_nonzeros()

template<typename scalar_t , typename integer_t = int>

std::size_t strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::factor_nonzeros ( ) const

Return the number of nonzeros in the (sparse) factors. This is known as the fill-in. This should be called after computing the numerical factorization. For the StrumpackSparseSolverMPI and StrumpackSparseSolverMPIDist distributed memory solvers, this routine is collective on the MPI communicator.

◆ Krylov_iterations()

template<typename scalar_t , typename integer_t = int>

int strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::Krylov_iterations ( ) const

Return the number of iterations performed by the outer (Krylov) iterative solver. Call this after calling the solve routine.

◆ maximum_rank()

template<typename scalar_t , typename integer_t = int>

int strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::maximum_rank ( ) const

Return the maximum rank encountered in any of the HSS matrices used to compress the sparse triangular factors. This should be called after the factorization phase. For the StrumpackSparseSolverMPI and StrumpackSparseSolverMPIDist distributed memory solvers, this routine is collective on the MPI communicator.

◆ options() [1/2]

template<typename scalar_t , typename integer_t = int>

SPOptions<scalar_t>& strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::options ( )

Return the object holding the options for this sparse solver.

◆ options() [2/2]

template<typename scalar_t , typename integer_t = int>

const SPOptions<scalar_t>& strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::options ( ) const

Return the object holding the options for this sparse solver.

◆ reorder() [1/2]

template<typename scalar_t , typename integer_t = int>

ReturnCode strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::reorder	(	const int *	p,
		int	base = `0`
	)

Perform sparse matrix reordering, with a user-supplied permutation vector. Using this will ignore the reordering method selected in the options struct.

Parameters

p	permutation vector, should be of size N, the size of the sparse matrix associated with this solver
base	is the permutation 0 or 1 based?

◆ reorder() [2/2]

template<typename scalar_t , typename integer_t = int>

ReturnCode strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::reorder	(	int	nx = `1`,
		int	ny = `1`,
		int	nz = `1`,
		int	components = `1`,
		int	width = `1`
	)

Compute matrix reorderings for numerical stability and to reduce fill-in.

Start computation of the matrix reorderings. See the relevant options to control the matrix reordering in the manual. A first reordering is the MC64 column permutation for numerical stability. This can be disabled if the matrix has large nonzero diagonal entries. MC64 optionally also performs row and column scaling. Next, a fill-reducing reordering is computed. This is done with the nested dissection algortihms of either (PT-)Scotch, (Par)Metis or a simple geometric nested dissection code which only works on regular meshes.

Parameters

nx	this (optional) parameter is only meaningful when the matrix corresponds to a stencil on a regular mesh. The stencil is assumed to be at most 3 points wide in each dimension and only contain a single degree of freedom per grid point. The nx parameter denotes the number of grid points in the first spatial dimension.
ny	see parameters nx. Parameter ny denotes the number of gridpoints in the second spatial dimension. This should only be set if the mesh is 2 or 3 dimensional.
nz	See parameters nx. Parameter nz denotes the number of gridpoints in the third spatial dimension. This should only be set if the mesh is 3 dimensional.

Returns: error code

See also: SPOptions

◆ set_from_options() [1/2]

template<typename scalar_t , typename integer_t = int>

void strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::set_from_options ( )

Parse the command line options passed in the constructor, and modify the options object accordingly. Run with option -h or –help to see a list of supported options. Or check the SPOptions documentation.

◆ set_from_options() [2/2]

template<typename scalar_t , typename integer_t = int>

void strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::set_from_options	(	int	argc,
		char *	argv[]
	)

Parse the command line options, and modify the options object accordingly. Run with option -h or –help to see a list of supported options. Or check the SPOptions documentation.

Parameters

argc	number of options in argv
argv	list of options

◆ solve() [1/2]

template<typename scalar_t , typename integer_t = int>

ReturnCode strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::solve	(	const DenseM_t &	b,
		DenseM_t &	x,
		bool	use_initial_guess = `false`
	)

Solve a linear system with a single or multiple right-hand sides. Before being able to solve a linear system, the matrix needs to be factored. One can call factor() explicitly, or if this was not yet done, this routine will call factor() internally.

Parameters

b	input, will not be modified. DenseMatrix containgin the right-hand side vector/matrix. Should have N rows, with N the dimension of the input matrix for StrumpackSparseSolver and StrumpackSparseSolverMPI. For StrumpackSparseSolverMPIDist, the number or rows of b should be correspond to the partitioning of the block-row distributed input matrix.
x	Output, pointer to the solution vector. Array should be lenght N, the dimension of the input matrix for StrumpackSparseSolver and StrumpackSparseSolverMPI. For StrumpackSparseSolverMPIDist, the length of b should be correspond the partitioning of the block-row distributed input matrix.
use_initial_guess	set to true if x contains an intial guess to the solution. This is mainly useful when using an iterative solver. If set to false, x should not be set (but should be allocated).

Returns: error code

See also: DenseMatrix, solve(), factor()

◆ solve() [2/2]

template<typename scalar_t , typename integer_t = int>

ReturnCode strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >::solve	(	const scalar_t *	b,
		scalar_t *	x,
		bool	use_initial_guess = `false`
	)

Solve a linear system with a single right-hand side. Before being able to solve a linear system, the matrix needs to be factored. One can call factor() explicitly, or if this was not yet done, this routine will call factor() internally.

Parameters

b	input, will not be modified. Pointer to the right-hand side. Array should be lenght N, the dimension of the input matrix for StrumpackSparseSolver and StrumpackSparseSolverMPI. For StrumpackSparseSolverMPIDist, the length of b should be correspond the partitioning of the block-row distributed input matrix.
x	Output, pointer to the solution vector. Array should be lenght N, the dimension of the input matrix for StrumpackSparseSolver and StrumpackSparseSolverMPI. For StrumpackSparseSolverMPIDist, the length of b should be correspond the partitioning of the block-row distributed input matrix.
use_initial_guess	set to true if x contains an intial guess to the solution. This is mainly useful when using an iterative solver. If set to false, x should not be set (but should be allocated).

Returns: error code

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

/home/pieterg/LBL/STRUMPACK/src/StrumpackSparseSolverBase.hpp

Public Member Functions

Detailed Description

template<typename scalar_t, typename integer_t = int> class strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >

Constructor & Destructor Documentation

◆ StrumpackSparseSolverBase() [1/2]

◆ StrumpackSparseSolverBase() [2/2]

◆ ~StrumpackSparseSolverBase()

Member Function Documentation

◆ draw()

◆ factor()

◆ factor_memory()

◆ factor_nonzeros()

◆ Krylov_iterations()

◆ maximum_rank()

◆ options() [1/2]

◆ options() [2/2]

◆ reorder() [1/2]

◆ reorder() [2/2]

◆ set_from_options() [1/2]

◆ set_from_options() [2/2]

◆ solve() [1/2]

◆ solve() [2/2]

template<typename scalar_t, typename integer_t = int>
class strumpack::StrumpackSparseSolverBase< scalar_t, integer_t >