c
2.0
The PRIMME SVDS interface is composed of the following functions. To solve real and complex singular value problems call respectively:
not text
- int :c
dprimme_svds <dprimme_svds>
(double *svals, double *svecs, double *resNorms, primme_svds_params *primme_svds)
- int :c
zprimme_svds <zprimme_svds>
(double *svals, :cPRIMME_COMPLEX_DOUBLE
*svecs, double *resNorms, primme_svds_params *primme_svds)
text
int dprimme_svds(double *svals, double *svecs, double *resNorms,
primme_svds_params *primme);
int zprimme_svds(double *svals, PRIMME_COMPLEX_DOUBLE *svecs, double *resNorms,
primme_svds_params *primme);
There are more versions for matrix problems working in other precisions:
Precision | Real | Complex |
---|---|---|
half | :chprimme_svds :chsprimme_svds |
:ckprimme_svds :cksprimme_svds |
single | :csprimme_svds |
:ccprimme_svds |
double | :cdprimme_svds |
:czprimme_svds |
Other useful functions:
not text
void :cprimme_svds_initialize <primme_svds_initialize>
(primme_svds_params *primme_svds) int :cprimme_svds_set_method <primme_svds_set_method>
(primme_svds_preset_method method, primme_preset_method methodStage1, primme_preset_method methodStage2, primme_svds_params *primme_svds) void :cprimme_svds_display_params <primme_svds_display_params>
(primme_svds_params primme_svds) void :cprimme_svds_free <primme_svds_free>
(primme_svds_params *primme_svds)
text
void primme_svds_initialize(primme_svds_params *primme_svds);
int primme_svds_set_method(primme_svds_preset_method method,
primme_preset_method methodStage1, primme_preset_method methodStage2,
primme_svds_params *primme_svds);
void primme_svds_display_params(primme_svds_params primme_svds);
void primme_svds_Free(primme_svds_params *primme_svds);
PRIMME SVDS stores its data on the structure :cprimme_svds_params
. See svds-guide-params
for an introduction about its fields.
To use PRIMME SVDS, follow these basic steps.
Include:
#include "primme.h" /* header file is required to run primme */
Initialize a PRIMME SVDS parameters structure for default settings:
not text
:c
primme_svds_params
primme_svds; :cprimme_svds_initialize <primme_svds_initialize>
(&primme_svds);text
primme_svds_params primme_svds; primme_svds_initialize(&primme_svds);
Set problem parameters (see also
svds-guide-params
), and, optionally, set one of the :cpreset methods <primme_svds_preset_method>
:not text
primme_svds. = matrixMatvec; /* MV product */ primme_svds. = 1000; /* set the matrix dimensions */ primme_svds. = 100; primme_svds. = 10; /* Number of singular values */ :c
primmesvds_set_method <primme_svds_set_method>
(primme_svds_hybrid, PRIMME_DEFAULT_METHOD, PRIMME_DEFAULT_METHOD, &primme_svds); ...text
primme_svds.matrixMatvec = matrixMatvec; /* MV product */ primme_svds.m = 1000; /* set problem dimension */ primme_svds.n = 100; primme_svds.numSvals = 10; /* Number of wanted singular values */ primme_svds_set_method(primme_svds_hybrid, PRIMME_DEFAULT_METHOD, PRIMME_DEFAULT_METHOD, &primme_svds); ...
Then to solve a real singular value problem call:
not text
ret = :c
dprimme_svds <dprimme_svds>
(svals, svecs, resNorms, &primme_svds);text
ret = dprimme_svds(svals, svecs, resNorms, &primme_svds);
The previous is the double precision call. There is available calls for complex double, single and complex single; check :c
zprimme_svds
, :csprimme_svds
and :ccprimme_svds
.To solve complex singular value problems call:
not text
ret = :c
zprimme_svds <zprimme_svds>
(svals, svecs, resNorms, &primme_svds);text
ret = zprimme_svds(svals, svecs, resNorms, &primme_svds);
The call arguments are:
- svals, array to return the found singular values;
- svecs, array to return the found left and right singular vectors;
- resNorms, array to return the residual norms of the found triplets; and
- ret, returned error code.
To free the work arrays in PRIMME SVDS:
not text
:c
primme_svds_free <primme_svds_Free>
(&primme_svds);text
primme_svds_free(&primme_svds);
PRIMME SVDS stores the data on the structure :cprimme_svds_params
, which has the next fields:
not text
PRIMME_INT
, number of rows of the matrix.PRIMME_INT
, number of columns of the matrix.void (*
)(...)
, matrix-vector product.int
, how many singular triplets to find.primme_svds_target
, which singular values to find.double
, tolerance of the residual norm of converged triplets.For parallel programs
int
, number of processesint
, rank of this processPRIMME_INT
, number of rows stored in this processPRIMME_INT
, number of columns stored in this processvoid (*
)(...)
, sum reduction among processesAccelerate the convergence
void (*
)(...)
, preconditioner-vector product.int
, initial vectors as approximate solutions.int
int
int
User data
void *
void *
void *
void *
void *
Advanced options
int
, for targeting interior singular values.double *
int
, orthogonal constrains to the singular vectors.int
PRIMME_INT
PRIMME_INT
[4]
double
int
FILE *
primme_svds_operator
primme_svds_operator
void (*
)(...)
, custom convergence criterion.void (*
)(...)
, custom convergence history.primme_op_datatype
primme_op_datatype
primme_op_datatype
primme_op_datatype
primme_op_datatype
text
/* Basic */
PRIMME_INT m; // number of rows of the matrix
PRIMME_INT n; // number of columns of the matrix
void (*matrixMatvec)(...); // matrix-vector product
int numSvals; // how many singular triplets to find
primme_svds_target target; // which singular values to find
double eps; // tolerance of the converged triplets
/* For parallel programs */
int numProcs; // number of processes
int procID; // rank of this process
PRIMME_INT mLocal; // number of rows stored in this process
PRIMME_INT nLocal; // number of columns stored in this process
void (*globalSumReal)(...); // sum reduction among processes
/* Accelerate the convergence */
void (*applyPreconditioner)(...); // preconditioner-vector product
int initSize; // initial vectors as approximate solutions
int maxBasisSize;
int minRestartSize;
int maxBlockSize;
/* User data */
void *commInfo;
void *matrix;
void *preconditioner;
void *convtest;
void *monitor;
/* Advanced options */
int numTargetShifts; // for targeting interior values
double *targetShifts;
int numOrthoConst; // orthogonal constrains to the vectors
int locking;
PRIMME_INT maxMatvecs;
PRIMME_INT iseed[4];
double aNorm;
int printLevel;
FILE * outputFile;
primme_svds_operator method;
primme_svds_operator methodStage2;
primme_params primme;
primme_params primmeStage2;
void (*convTestFun)(...); // custom convergence criterion
void (*monitorFun)(...); // custom convergence history
primme_op_datatype matrixMatvec_type;
primme_op_datatype applyPreconditioner_type;
primme_op_datatype globalSumReal_type;
primme_op_datatype broadcastReal_type;
primme_op_datatype internalPrecision;
PRIMME SVDS requires the user to set at least the matrix dimensions ( x ) and the matrix-vector product (), as they define the problem to be solved. For parallel programs, , , and are also required.
In addition, most users would want to specify how many singular triplets to find, and provide a preconditioner (if available).
It is useful to have set all these before calling :cprimme_svds_set_method
. Also, if users have a preference on , , etc, they should also provide them into :cprimme_svds_params
prior to the :cprimme_svds_set_method
call. This helps :cprimme_svds_set_method
make the right choice on other parameters. It is sometimes useful to check the actual parameters that PRIMME SVDS is going to use (before calling it) or used (on return) by printing them with :cprimme_svds_display_params
.
The next enumerations and functions are declared in primme.h
.