BTASolver Class Reference

creates solver class using BTA-GPU for factorising, solving and selectively inverting linear system. More...

#include <BTASolver.h>

Inheritance diagram for BTASolver:

Public Member Functions
	BTASolver (size_t ns_, size_t nt_, size_t nb_, size_t no_, int thread_ID_)
void	symbolic_factorization (SpMat &Q, int &init)
	not used for BTASolver, only in PARDISO
void	factorize (SpMat &Q, double &log_det, double &t_priorLatChol)
	numerical factorisation using block-wise factorisation on GPU.
void	factorize_w_constr (SpMat &Q, const MatrixXd &D, double &log_det, MatrixXd &V)
void	factorize_solve (SpMat &Q, Vect &rhs, Vect &sol, double &log_det, double &t_condLatChol, double &t_condLatSolve)
	factorises and solves matrix in one call
void	factorize_solve_w_constr (SpMat &Q, Vect &rhs, const MatrixXd &Dxy, double &log_det, Vect &sol, MatrixXd &V)
void	selected_inversion (SpMat &Q, Vect &inv_diag)
	selected inversion of the diagonal elements of Q.
void	selected_inversion_w_constr (SpMat &Q, const MatrixXd &D, Vect &inv_diag, MatrixXd &V)
	~BTASolver ()
	class destructor. Frees memory allocated by BTA.

Private Attributes
int	MPI_size
int	MPI_rank
char	processor_name [MPI_MAX_PROCESSOR_NAME]
int	name_len
int	threads_level1
int	thread_ID
unsigned int	nnz
size_t	i
int	GPU_rank
size_t	ns_t
size_t	nt_t
size_t	nb_t
size_t	no_t
size_t	n
SpMat	Q
int *	ia
int *	ja
double *	a
double *	b
double *	x
BTA< double > *	solver
double	dummy_time_1
double	dummy_time_2

Additional Inherited Members
Protected Attributes inherited from Solver
int	init
double	log_det
SpMat	Q
Vect	rhs
Vect	sol

Detailed Description

creates solver class using BTA-GPU for factorising, solving and selectively inverting linear system.

divided into set up, symbolic factorisation, numerical factorisation, numerical factorisation & solve and selected inversion (of the diagonal elements)

Note

in each BTASolver function call factorise, selected_inversion etc. class BTA gets created. Is this the best way to handle things. Potentially merge them somehow? Maybe actually does not take any time.

Member Function Documentation

factorize()

void BTASolver::factorize ( SpMat &  Q, double &  log_det, double &  t_priorLatChol  )

virtual

numerical factorisation using block-wise factorisation on GPU.

Parameters

[in]	Q	precision matrix to be factorised.
[in,out]	log_det	computes log determinant of Q.

Implements Solver.

factorize_solve()

void BTASolver::factorize_solve ( SpMat &  Q, Vect &  rhs, Vect &  sol, double &  log_det, double &  t_condLatChol, double &  t_condLatSolve  )

virtual

factorises and solves matrix in one call

Parameters

[in]	Q	precision matrix.
[in]	rhs	right-hand side of the system.
[in,out]	sol	solution of the system.
[in,out]	log_det	log determinant of Q.

Implements Solver.

factorize_solve_w_constr()

void BTASolver::factorize_solve_w_constr ( SpMat &  Q, Vect &  rhs, const MatrixXd &  Dxy, double &  log_det, Vect &  sol, MatrixXd &  V  )

virtual

Implements Solver.

factorize_w_constr()

void BTASolver::factorize_w_constr ( SpMat &  Q, const MatrixXd &  D, double &  log_det, MatrixXd &  V  )

virtual

Implements Solver.

selected_inversion()

void BTASolver::selected_inversion ( SpMat &  Q, Vect &  inv_diag  )

virtual

selected inversion of the diagonal elements of Q.

Parameters

[in]	Q	precision matrix.
[in,out]	inv_diag	inverse diagonal to hold the solution vector.

Note

is there some way to potentially reuse Cholesky factor that is already on CPU?

Implements Solver.

selected_inversion_w_constr()

void BTASolver::selected_inversion_w_constr ( SpMat &  Q, const MatrixXd &  D, Vect &  inv_diag, MatrixXd &  V  )

virtual

Implements Solver.

symbolic_factorization()

void BTASolver::symbolic_factorization ( SpMat &  Q, int &  init  )

virtual

not used for BTASolver, only in PARDISO

Implements Solver.

Member Data Documentation

a

double* BTASolver::a

private

CSR format. values.

b

double* BTASolver::b

private

right-hand side.

ia

int* BTASolver::ia

private

CSR format. row indices.

ja

int* BTASolver::ja

private

CSR format. col pointers.

nnz

unsigned int BTASolver::nnz

private

number of nonzeros

Q

SpMat BTASolver::Q

private

sparse precision matrix Q. Eigen format.

solver

BTA<double>* BTASolver::solver

private

BTA solver object

x

double* BTASolver::x

private

placeholder for solution.

---

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

• INLA_main/BTASolver.h
• INLA_main/BTASolver.cpp