handling of cuDSS-based linear solvers More...

#include "base/cs_defs.h"
#include <stdarg.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <math.h>
#include <cudss.h>
#include "bft/bft_error.h"
#include "bft/bft_printf.h"
#include "base/cs_base.h"
#include "base/cs_dispatch.h"
#include "base/cs_log.h"
#include "base/cs_fp_exception.h"
#include "base/cs_halo.h"
#include "base/cs_math.h"
#include "base/cs_mem.h"
#include "alge/cs_matrix.h"
#include "alge/cs_matrix_default.h"
#include "base/cs_timer.h"
#include "alge/cs_sles.h"
#include "alge/cs_sles_cudss.h"

Include dependency graph for cs_sles_cudss.cpp:

Functions
cs_sles_cudss_t *	cs_sles_cudss_define (int f_id, const char *name)
	Define and associate an cuDSS linear system solver for a given field or equation name. More...

cs_sles_cudss_t *	cs_sles_cudss_create (void)
	Create cuDSS linear system solver info and context. More...

void *	cs_sles_cudss_copy (const void *context)
	Create cuDSS linear system solver info and context based on existing info and context. More...

void	cs_sles_cudss_destroy (void **context)
	Destroy cuDSS linear system solver info and context. More...

void	cs_sles_cudss_setup (void context, const char name, const cs_matrix_t *a, int verbosity)
	Setup cuDSS linear equation solver. More...

cs_sles_convergence_state_t	cs_sles_cudss_solve (void context, const char name, const cs_matrix_t a, int verbosity, double precision, double r_norm, int n_iter, double residual, const cs_real_t rhs, cs_real_t vx_ini, cs_real_t vx, size_t aux_size, void *aux_vectors)
	Call cuDSS linear equation solver. More...

void	cs_sles_cudss_free (void *context)
	Free cuDSS linear equation solver setup context. More...

void	cs_sles_cudss_log (const void *context, cs_log_t log_type)
	Log sparse linear equation solver info. More...

void	cs_sles_cudss_library_info (cs_log_t log_type)
	Print information on cuDSS library. More...

Detailed Description

handling of cuDSS-based linear solvers

Function Documentation

◆ cs_sles_cudss_copy()

void * cs_sles_cudss_copy ( const void * context )

Create cuDSS linear system solver info and context based on existing info and context.

Most configuration parameters will be copied from the existing context, though not all, since cuDSS does not provide a comprehensive way to do this.

Parameters

[in] context pointer to reference info and context (actual type: cs_sles_cudss_t *)

Returns: pointer to newly created solver info object. (actual type: cs_sles_cudss_t *)

◆ cs_sles_cudss_create()

cs_sles_cudss_t * cs_sles_cudss_create ( void )

Create cuDSS linear system solver info and context.

In case of rotational periodicity for a block (non-scalar) matrix, the matrix type will be forced to MATSHELL ("shell") regardless of the option used.

Returns: pointer to associated linear system object.

◆ cs_sles_cudss_define()

cs_sles_cudss_t * cs_sles_cudss_define	(	int	f_id,
		const char *	name
	)

Define and associate an cuDSS linear system solver for a given field or equation name.

If this system did not previously exist, it is added to the list of "known" systems. Otherwise, its definition is replaced by the one defined here.

This is a utility function: if finer control is needed, see cs_sles_define and cs_sles_cudss_create.

In case of rotational periodicity for a block (non-scalar) matrix, the matrix type will be forced to MATSHELL ("shell") regardless of the option used.

Note that this function returns a pointer directly to the solver management structure. This may be used to set further options. If needed, cs_sles_find may be used to obtain a pointer to the matching cs_sles_t container.

Parameters

[in]	f_id	associated field id, or < 0
[in]	name	associated name if f_id < 0, or nullptr

Returns: pointer to newly created cuDSS solver info object.

◆ cs_sles_cudss_destroy()

void cs_sles_cudss_destroy ( void ** context )

Destroy cuDSS linear system solver info and context.

Parameters

[in,out] context pointer to cuDSS solver info and context (actual type: cs_sles_cudss_t **)

◆ cs_sles_cudss_free()

void cs_sles_cudss_free ( void * context )

Free cuDSS linear equation solver setup context.

This function frees resolution-related data, such as buffers and preconditioning but does not free the whole context, as info used for logging (especially performance data) is maintained.

Parameters

[in,out] context pointer to cuDSS solver info and context (actual type: cs_sles_cudss_t *)

◆ cs_sles_cudss_library_info()

void cs_sles_cudss_library_info ( cs_log_t log_type )

Print information on cuDSS library.

Parameters

[in] log_type log type

◆ cs_sles_cudss_log()

void cs_sles_cudss_log	(	const void *	context,
		cs_log_t	log_type
	)

Log sparse linear equation solver info.

Parameters

[in]	context	pointer to cuDSS solver info and context (actual type: cs_sles_cudss_t *)
[in]	log_type	log type

◆ cs_sles_cudss_setup()

void cs_sles_cudss_setup	(	void *	context,
		const char *	name,
		const cs_matrix_t *	a,
		int	verbosity
	)

Setup cuDSS linear equation solver.

Parameters

[in,out]	context	pointer to cuDSS solver info and context (actual type: cs_sles_cudss_t *)
[in]	name	pointer to system name
[in]	a	associated matrix
[in]	verbosity	associated verbosity

◆ cs_sles_cudss_solve()

cs_sles_convergence_state_t cs_sles_cudss_solve	(	void *	context,
		const char *	name,
		const cs_matrix_t *	a,
		int	verbosity,
		double	precision,
		double	r_norm,
		int *	n_iter,
		double *	residual,
		const cs_real_t *	rhs,
		cs_real_t *	vx_ini,
		cs_real_t *	vx,
		size_t	aux_size,
		void *	aux_vectors
	)

Call cuDSS linear equation solver.

Warning: The precision, r_norm, and n_iter parameters are ignored here. the matching configuration options should be set earlier, using the cs_sles_cudss_set_config function

Parameters

[in,out]	context	pointer to cuDSS solver info and context (actual type: cs_sles_cudss_t *)
[in]	name	pointer to system name
[in]	a	matrix
[in]	verbosity	associated verbosity
[in]	precision	solver precision
[in]	r_norm	residual normalization
[out]	n_iter	number of "equivalent" iterations
[out]	residual	residual
[in]	rhs	right hand side
[in]	vx_ini	initial system solution (vx if nonzero, nullptr if zero)
[in,out]	vx	system solution
[in]	aux_size	number of elements in aux_vectors (in bytes)
	aux_vectors	optional working area (internal allocation if nullptr)

Returns: convergence state

Functions

Detailed Description

Function Documentation

◆ cs_sles_cudss_copy()

◆ cs_sles_cudss_create()

◆ cs_sles_cudss_define()

◆ cs_sles_cudss_destroy()

◆ cs_sles_cudss_free()

◆ cs_sles_cudss_library_info()

◆ cs_sles_cudss_log()

◆ cs_sles_cudss_setup()

◆ cs_sles_cudss_solve()