PaStiX

PaStiX users documentation. More...

Modules
	Parameters
	This group describes parameters and constants to control the solver.
	Examples
	This group describes how to use the solver with some examples.
	Analyze step
	This group describes all the analyse routines that prepare data structures for the factorization and solve steps.
	Factorization step
	This group describes all the routines of the internal block CSC and the solver matrix structure before applying the numerical factorization step.
	Solve step
	This group describes all the routines to solve the system.
	Refinement step
	This group describes all the routines to apply iterative refinement on the system.
	Schur complement routines

Data Structures
struct	pastix_data_s
	Main PaStiX data structure. More...
struct	pastix_rhs_s
	Main PaStiX RHS structure. More...

Functions
int	pastix_task_analyze (pastix_data_t *pastix_data, const spmatrix_t *spm)
	Perform all the preprocessing steps: ordering, symbolic factorization, reordering, proportionnal mapping, ... More...
int	pastix_task_numfact (pastix_data_t *pastix_data, spmatrix_t *spm)
	Perform all the numerical factorization steps: fill the internal block CSC and the solver matrix structures, then apply the factorization step. More...
int	pastix_task_solve (pastix_data_t *pastix_data, pastix_int_t m, pastix_int_t nrhs, void *b, pastix_int_t ldb)
	Solve the given problem. More...
int	pastix_task_refine (pastix_data_t *pastix_data, pastix_int_t m, pastix_int_t nrhs, void *b, pastix_int_t ldb, void *x, pastix_int_t ldx)
	Perform iterative refinement. More...
int	pastix_task_solve_and_refine (pastix_data_t *pastix_data, pastix_int_t m, pastix_int_t nrhs, void *b, pastix_int_t ldb, void *x, pastix_int_t ldx)
	Performs solve and iterative refinement without unnecessary permutations. More...

Detailed Description

PaStiX users documentation.

---

Data Type Documentation

pastix_data_s

struct pastix_data_s

Main PaStiX data structure.

This structure holds all informations related to the library and problem instance. It stores information from one step to another.

Warning

This structure should not be modified directly by the user.

Definition at line 67 of file pastixdata.h.

Data Fields
pastix_int_t	id	Unique identifier of the pastix instance
pastix_int_t *	iparm	Store integer parameters (input/output)
double *	dparm	Store floating parameters (input/output)
pastix_int_t	steps	Bitmask of the steps performed or not
pastix_scheduler_t	sched	Indicates the scheduler used for the factorization step
PASTIX_Comm	pastix_comm	PaStiX MPI communicator used for the ordering step
PASTIX_Comm	intra_node_comm	PaStiX intra node MPI communicator used for synchronizations
PASTIX_Comm	inter_node_comm	PaStiX inter node MPI communicator used for the factorization
int	procnbr	Total number of MPI processes
int	procnum	Local MPI rank
int	intra_node_procnbr	Number of MPI tasks in intra node communicator
int	intra_node_procnum	Local MPI rank in intra node communicator
int	inter_node_procnbr	Number of MPI tasks in inter node communicator
int	inter_node_procnum	Local MPI rank in inter node communicator
isched_t *	isched	Internal scheduler structure that is always available
void *	parsec	PaRSEC context if available
void *	starpu	StarPU context if available
const spmatrix_t *	csc	Pointer to the user csc structure used as input
pastix_graph_t *	graph	Symmetrized graph of the problem used within ordering and symbolic factorization steps.
pastix_int_t	schur_n	Number of entries for the Schur complement
pastix_int_t *	schur_list	List of entries for the schur complement
pastix_int_t	zeros_n	Number of diagonal entries considered as zeros
pastix_int_t *	zeros_list	List of diagonal entries considered as zeros
pastix_order_t *	ordemesh	Ordering structure
symbol_matrix_t *	symbmtx	Symbol Matrix
pastix_bcsc_t *	bcsc	Csc after reordering grouped by cblk
SolverMatrix *	solvmatr	Solver informations associated to the matrix problem
SolverMatrix *	solvloc	Solver informations associated to the matrix problem - Local
SolverMatrix *	solvglob	Solver informations associated to the matrix problem - Global
pastix_model_t *	cpu_models	CPU model coefficients for the kernels
pastix_model_t *	gpu_models	GPU model coefficients for the kernels
char *	dir_global	Unique directory name to store output files
char *	dir_local	Unique directory name to store output specific to a MPI process
void *	b
void *	x0
int *	bindtab	Former fields that are no longer used for now
void *	schur_tab
pastix_int_t	schur_tab_set
int	cscInternFilled
int	scaling
void *	scalerowtab
void *	iscalerowtab
void *	scalecoltab
void *	iscalecoltab
pastix_int_t	pastix_id

pastix_rhs_s

struct pastix_rhs_s

Main PaStiX RHS structure.

This structure holds all informations about a set of permuted right-hand-sides to be used in the solve/refinement steps.

Warning

This structure should not be modified directly by the user.

Definition at line 150 of file pastixdata.h.

Data Fields
int8_t	allocated	Flag to know if the vector b is allocated internally or not.
pastix_coeftype_t	flttype	Floating type of the vector.
pastix_int_t	m	Local number of rows in the right hand sides.
pastix_int_t	n	Number of columns in the right hand sides.
pastix_int_t	ld	Leading dimension of the right hand side matrix.
void *	b	Right hand sides of size ldb-by-n.
void **	cblkb	Array to store the temporary buffers associated to fanin/recv.
bvec_handle_comm_t *	rhs_comm	Structure which handles the MPI communication (= NULL if PASTIX_WITH_MPI=OFF).
pastix_int_t *	Ploc2Pglob	Array containing the local permuted index corresponding to the global permuted index.

Function Documentation

pastix_task_analyze()

int pastix_task_analyze	(	pastix_data_t *	pastix_data,
		const spmatrix_t *	spm
	)

Perform all the preprocessing steps: ordering, symbolic factorization, reordering, proportionnal mapping, ...

This function combines the calls to all the preprocessing steps. See pastix_subtask_order() for further details on the ordering step, pastix_subtask_symbfact() for further details on the symbolic factotrization step, pastix_subtask_reordering() for further details on the reordering process, and pastix_subtask_blend() for the final analyse step that performs proportionnal mapin and generates the numerical factorization data structures.

Parameters

[in,out]	pastix_data	The pastix_data structure that describes the solver instance.
[in]	spm	The sparse matrix descriptor that describes problem instance.

Return values

PASTIX_SUCCESS	on successful exit
PASTIX_ERR_BADPARAMETER	if one parameter is incorrect.
PASTIX_ERR_OUTOFMEMORY	if one allocation failed.

Definition at line 53 of file pastix_task_analyze.c.

References pastix_data_s::dparm, DPARM_ANALYZE_TIME, pastix_data_s::iparm, IPARM_VERBOSE, PASTIX_ERR_BADPARAMETER, pastix_subtask_blend(), pastix_subtask_order(), pastix_subtask_reordering(), pastix_subtask_symbfact(), PASTIX_SUCCESS, PastixVerboseNot, pastix_data_s::procnum, and pastix_data_s::steps.

pastix_task_numfact()

int pastix_task_numfact	(	pastix_data_t *	pastix_data,
		spmatrix_t *	spm
	)

Perform all the numerical factorization steps: fill the internal block CSC and the solver matrix structures, then apply the factorization step.

The user can call the pastix_task_solve() to obtain the solution.

This routine is affected by the following parameters: IPARM_VERBOSE.

Parameters

[in,out]	pastix_data	The pastix_data structure that describes the solver instance. On exit, the internal block CSC is filled with entries from the spm matrix and the solver matrix structure stores the factorization of the given problem.
[in,out]	spm	The sparse matrix descriptor that describes problem instance. On exit, the spm structure is freed if IPARM_FREE_CSCUSER is true.

Return values

PASTIX_SUCCESS	on successful exit,
PASTIX_ERR_BADPARAMETER	if one parameter is incorrect,
PASTIX_ERR_OUTOFMEMORY	if one allocation failed.

Definition at line 645 of file pastix_task_sopalin.c.

References pastix_data_s::inter_node_procnum, pastix_data_s::iparm, IPARM_FACTORIZATION, IPARM_VERBOSE, PASTIX_ERR_BADPARAMETER, pastix_int_t, pastix_subtask_bcsc2ctab(), pastix_subtask_sopalin(), pastix_subtask_spm2bcsc(), PASTIX_SUCCESS, PastixVerboseNot, and pastix_data_s::steps.

Referenced by pastix().

pastix_task_solve()

int pastix_task_solve	(	pastix_data_t *	pastix_data,
		pastix_int_t	m,
		pastix_int_t	nrhs,
		void *	b,
		pastix_int_t	ldb
	)

Solve the given problem.

This routine is affected by the following parameters: IPARM_VERBOSE, IPARM_FACTORIZATION, IPARM_TRANSPOSE_SOLVE.

Parameters

[in,out]	pastix_data	The pastix_data structure that describes the solver instance.
[in]	m	The local number of rows of the right-and-side vectors. If the spm is centralized or replicated, m must be equal to spm->gNexp, otherwise m must be equal to spm->nexp.
[in]	nrhs	The number of right-and-side vectors.
[in,out]	b	The right-and-side vectors (can be multiple rhs). On exit, the solution is stored in place of the right-hand-side vector.
[in]	ldb	The leading dimension of the right-and-side vectors.

Return values

PASTIX_SUCCESS	on successful exit,
PASTIX_ERR_BADPARAMETER	if one parameter is incorrect.

Definition at line 757 of file pastix_task_solve.c.

References pastix_rhs_s::b, pastix_rhs_s::m, PASTIX_ERR_BADPARAMETER, pastix_int_t, pastix_subtask_applyorder(), pastix_subtask_solve(), PASTIX_SUCCESS, PastixDirBackward, PastixDirForward, pastixRhsFinalize(), pastixRhsInit(), and pastix_data_s::steps.

pastix_task_refine()

int pastix_task_refine	(	pastix_data_t *	pastix_data,
		pastix_int_t	m,
		pastix_int_t	nrhs,
		void *	b,
		pastix_int_t	ldb,
		void *	x,
		pastix_int_t	ldx
	)

Perform iterative refinement.

This routine performs the permutation of x, and b before and after the iterative refinement solution. To prevent extra permuation to happen, see pastix_subtask_refine(). This routine is affected by the following parameters: IPARM_REFINEMENT, DPARM_EPSILON_REFINEMENT

Parameters

[in]	pastix_data	The PaStiX data structure that describes the solver instance.
[in]	m	The size of system to solve, and the number of rows of both matrices b and x.
[in]	nrhs	The number of right hand side members, and the number of columns of b and x.
[in,out]	b	The right hand side matrix of size ldb-by-nrhs. B is noted as inout, as permutation might be performed on the matrix. On exit, the matrix is restored as it was on entry.
[in]	ldb	The leading dimension of the matrix b. ldb >= n.
[in,out]	x	The matrix x of size ldx-by-nrhs. On entry, the initial guess x0 for the refinement step, that may be the solution returned by the solve step or any other initial guess. On exit, contains the final solution after the iterative refinement.
[in]	ldx	The leading dimension of the matrix x. ldx >= n.

Return values

PASTIX_SUCCESS	on successful exit,
PASTIX_ERR_BADPARAMETER	if one parameter is incorrect,

Definition at line 231 of file pastix_task_refine.c.

References pastix_data_s::iparm, IPARM_SCHUR_SOLV_MODE, pastix_int_t, and pastix_data_s::schur_n.

pastix_task_solve_and_refine()

int pastix_task_solve_and_refine	(	pastix_data_t *	pastix_data,
		pastix_int_t	m,
		pastix_int_t	nrhs,
		void *	b,
		pastix_int_t	ldb,
		void *	x,
		pastix_int_t	ldx
	)

Performs solve and iterative refinement without unnecessary permutations.

This routine performs the permutation of x and b before and after the solve and iterative refinement solution. This prevent the permutation of x at the end of the solve and the permutations of x and b at the beginnning of the refinement when the user wants to perfom both solve and refinement. This routine is affected by the following parameters: IPARM_REFINEMENT, DPARM_EPSILON_REFINEMENT, IPARM_VERBOSE, IPARM_FACTORIZATION, IPARM_TRANSPOSE_SOLVE

Parameters

[in]	pastix_data	The PaStiX data structure that describes the solver instance.
[in]	m	The size of system to solve, and the number of rows of both matrices b and x.
[in]	nrhs	The number of right hand side members, and the number of columns of b and x.
[in,out]	b	The right hand side matrix of size ldb-by-nrhs. B is noted as inout, as permutation might be performed on the matrix. On exit, the matrix is restored as it was on entry.
[in]	ldb	The leading dimension of the matrix b. ldb >= n.
[in,out]	x	The matrix x of size ldx-by-nrhs. On entry, a copy of b. On exit, contains the final solution after the iterative refinement.
[in]	ldx	The leading dimension of the matrix x. ldx >= n.

Return values

PASTIX_SUCCESS	on successful exit,
PASTIX_ERR_BADPARAMETER	if one parameter is incorrect,

Definition at line 354 of file pastix_task_refine.c.

References pastix_data_s::iparm, IPARM_SCHUR_SOLV_MODE, PASTIX_ERR_BADPARAMETER, pastix_int_t, pastix_data_s::schur_n, and pastix_data_s::steps.