Gromacs  2024.4
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
Functions
gmx_arpack.h File Reference
#include "config.h"
+ Include dependency graph for gmx_arpack.h:

Description

Selected routines from ARPACK.

This file contains a subset of ARPACK functions to perform diagonalization and SVD for sparse matrices in Gromacs.

Consult the main ARPACK site for detailed documentation: http://www.caam.rice.edu/software/ARPACK/

Below, we just list the options and any specific differences from ARPACK. The code is essentially the same, but the routines have been made thread-safe by using extra workspace arrays.

Functions

void dsaupd (int *ido, const char *bmat, int *n, const char *which, int *nev, double *tol, double *resid, int *ncv, double *v, int *ldv, int *iparam, int *ipntr, double *workd, int *iwork, double *workl, int *lworkl, int *info)
 Implicitly Restarted Arnoldi Iteration, double precision. More...
 
void dseupd (int *rvec, const char *howmny, int *select, double *d, double *z, int *ldz, double *sigma, const char *bmat, int *n, const char *which, int *nev, double *tol, double *resid, int *ncv, double *v, int *ldv, int *iparam, int *ipntr, double *workd, double *workl, int *lworkl, int *info)
 Get eigenvalues/vectors after Arnoldi iteration, double prec. More...
 
void ssaupd (int *ido, const char *bmat, int *n, const char *which, int *nev, float *tol, float *resid, int *ncv, float *v, int *ldv, int *iparam, int *ipntr, float *workd, int *iwork, float *workl, int *lworkl, int *info)
 Implicitly Restarted Arnoldi Iteration, single precision. More...
 
void sseupd (int *rvec, const char *howmny, int *select, float *d, float *z, int *ldz, float *sigma, const char *bmat, int *n, const char *which, int *nev, float *tol, float *resid, int *ncv, float *v, int *ldv, int *iparam, int *ipntr, float *workd, float *workl, int *lworkl, int *info)
 Get eigenvalues/vectors after Arnoldi iteration, single prec. More...
 

Function Documentation

void dsaupd ( int *  ido,
const char *  bmat,
int *  n,
const char *  which,
int *  nev,
double *  tol,
double *  resid,
int *  ncv,
double *  v,
int *  ldv,
int *  iparam,
int *  ipntr,
double *  workd,
int *  iwork,
double *  workl,
int *  lworkl,
int *  info 
)

Implicitly Restarted Arnoldi Iteration, double precision.

Reverse communication interface for the Implicitly Restarted Arnoldi Iteration. For symmetric problems this reduces to a variant of the Lanczos method. See the ARPACK site for details.

Parameters
idoReverse communication flag. Set to 0 first time. Upon return with ido=-1 or ido=1 you should calculate Y=A*X and recall the routine. Return with ido=2 means Y=B*X should be calculated. ipntr[0] is the pointer in workd for X, ipntr[1] is the index for Y. Return with ido=99 means it finished.
bmat'I' for standard eigenproblem, 'G' for generalized.
nOrder of eigenproblem.
whichWhich eigenvalues to calculate. 'LA' for largest algebraic, 'SA' for smallest algebraic, 'LM' for largest magnitude, 'SM' for smallest magnitude, and finally 'BE' (both ends) to calculate half from each end of the spectrum.
nevNumber of eigenvalues to calculate. 0<nev<n.
tolTolerance. Machine precision of it is 0.
residOptional starting residual vector at input if info=1, otherwise a random one is used. Final residual vector on return.
ncvNumber of columns in matrix v.
vN*NCV matrix. V contain the Lanczos basis vectors.
ldvLeading dimension of v.
iparamInteger array, size 11. Same contents as arpack.
ipntrInteger array, size 11. Points to starting locations in the workd/workl arrays. Same contents as arpack.
workdDouble precision work array, length 3*n+4. Provide the same array for all calls, and don't touch it. IMPORTANT: This is 4 units larger than standard ARPACK!
iworkInteger work array, size 80. Provide the same array for all calls, and don't touch it. IMPORTANT: New argument compared to standard ARPACK!
worklDouble precision work array, length lwork.
lworklLength of the work array workl. Must be at least ncv*(ncv+8)
infoSet info to 0 to use random initial residual vector, or to 1 if you provide a one. On output, info=0 means normal exit, 1 that max number of iterations was reached, and 3 that no shifts could be applied. Negative numbers correspond to errors in the arguments provided.
void dseupd ( int *  rvec,
const char *  howmny,
int *  select,
double *  d,
double *  z,
int *  ldz,
double *  sigma,
const char *  bmat,
int *  n,
const char *  which,
int *  nev,
double *  tol,
double *  resid,
int *  ncv,
double *  v,
int *  ldv,
int *  iparam,
int *  ipntr,
double *  workd,
double *  workl,
int *  lworkl,
int *  info 
)

Get eigenvalues/vectors after Arnoldi iteration, double prec.

See the ARPACK site for details. You must have finished the interative part with dsaupd() before calling this function.

Parameters
rvec1 if you want eigenvectors, 0 if not.
howmny'A' if you want all nvec vectors, 'S' if you provide a subset selection in select[].
selectInteger array, dimension nev. Indices of the eigenvectors to calculate. Fortran code means we start counting on 1. This array must be given even in howmny is 'A'. (Arpack documentation is wrong on this).
dDouble precision array, length nev. Eigenvalues.
zDouble precision array, n*nev. Eigenvectors.
ldzLeading dimension of z. Normally n.
sigmaShift if iparam[6] is 3,4, or 5. Ignored otherwise.
bmatProvide the same argument as you did to dsaupd()
nProvide the same argument as you did to dsaupd()
whichProvide the same argument as you did to dsaupd()
nevProvide the same argument as you did to dsaupd()
tolProvide the same argument as you did to dsaupd()
residProvide the same argument as you did to dsaupd() The array must not be touched between the two function calls!
ncvProvide the same argument as you did to dsaupd()
vProvide the same argument as you did to dsaupd() The array must not be touched between the two function calls!
ldvProvide the same argument as you did to dsaupd()
iparamProvide the same argument as you did to dsaupd() The array must not be touched between the two function calls!
ipntrProvide the same argument as you did to dsaupd() The array must not be touched between the two function calls!
workdProvide the same argument as you did to dsaupd() The array must not be touched between the two function calls!
worklDouble precision work array, length lwork. The array must not be touched between the two function calls!
lworklProvide the same argument as you did to dsaupd()
infoProvide the same argument as you did to dsaupd()
void ssaupd ( int *  ido,
const char *  bmat,
int *  n,
const char *  which,
int *  nev,
float *  tol,
float *  resid,
int *  ncv,
float *  v,
int *  ldv,
int *  iparam,
int *  ipntr,
float *  workd,
int *  iwork,
float *  workl,
int *  lworkl,
int *  info 
)

Implicitly Restarted Arnoldi Iteration, single precision.

Reverse communication interface for the Implicitly Restarted Arnoldi Iteration. For symmetric problems this reduces to a variant of the Lanczos method. See the ARPACK site for details.

Parameters
idoReverse communication flag. Set to 0 first time. Upon return with ido=-1 or ido=1 you should calculate Y=A*X and recall the routine. Return with ido=2 means Y=B*X should be calculated. ipntr[0] is the pointer in workd for X, ipntr[1] is the index for Y. Return with ido=99 means it finished.
bmat'I' for standard eigenproblem, 'G' for generalized.
nOrder of eigenproblem.
whichWhich eigenvalues to calculate. 'LA' for largest algebraic, 'SA' for smallest algebraic, 'LM' for largest magnitude, 'SM' for smallest magnitude, and finally 'BE' (both ends) to calculate half from each end of the spectrum.
nevNumber of eigenvalues to calculate. 0<nev<n.
tolTolerance. Machine precision of it is 0.
residOptional starting residual vector at input if info=1, otherwise a random one is used. Final residual vector on return.
ncvNumber of columns in matrix v.
vN*NCV matrix. V contain the Lanczos basis vectors.
ldvLeading dimension of v.
iparamInteger array, size 11. Same contents as arpack.
ipntrInteger array, size 11. Points to starting locations in the workd/workl arrays. Same contents as arpack.
workdSingle precision work array, length 3*n+4. Provide the same array for all calls, and don't touch it. IMPORTANT: This is 4 units larger than standard ARPACK!
iworkInteger work array, size 80. Provide the same array for all calls, and don't touch it. IMPORTANT: New argument compared to standard ARPACK!
worklSingle precision work array, length lwork.
lworklLength of the work array workl. Must be at least ncv*(ncv+8)
infoSet info to 0 to use random initial residual vector, or to 1 if you provide a one. On output, info=0 means normal exit, 1 that max number of iterations was reached, and 3 that no shifts could be applied. Negative numbers correspond to errors in the arguments provided.
void sseupd ( int *  rvec,
const char *  howmny,
int *  select,
float *  d,
float *  z,
int *  ldz,
float *  sigma,
const char *  bmat,
int *  n,
const char *  which,
int *  nev,
float *  tol,
float *  resid,
int *  ncv,
float *  v,
int *  ldv,
int *  iparam,
int *  ipntr,
float *  workd,
float *  workl,
int *  lworkl,
int *  info 
)

Get eigenvalues/vectors after Arnoldi iteration, single prec.

See the ARPACK site for details. You must have finished the interative part with ssaupd() before calling this function.

Parameters
rvec1 if you want eigenvectors, 0 if not.
howmny'A' if you want all nvec vectors, 'S' if you provide a subset selection in select[].
selectInteger array, dimension nev. Indices of the eigenvectors to calculate. Fortran code means we start counting on 1. This array must be given even in howmny is 'A'. (Arpack documentation is wrong on this).
dSingle precision array, length nev. Eigenvalues.
zSingle precision array, n*nev. Eigenvectors.
ldzLeading dimension of z. Normally n.
sigmaShift if iparam[6] is 3,4, or 5. Ignored otherwise.
bmatProvide the same argument as you did to ssaupd()
nProvide the same argument as you did to ssaupd()
whichProvide the same argument as you did to ssaupd()
nevProvide the same argument as you did to ssaupd()
tolProvide the same argument as you did to ssaupd()
residProvide the same argument as you did to ssaupd() The array must not be touched between the two function calls!
ncvProvide the same argument as you did to ssaupd()
vProvide the same argument as you did to ssaupd() The array must not be touched between the two function calls!
ldvProvide the same argument as you did to ssaupd()
iparamProvide the same argument as you did to ssaupd() The array must not be touched between the two function calls!
ipntrProvide the same argument as you did to ssaupd() The array must not be touched between the two function calls!
workdProvide the same argument as you did to ssaupd() The array must not be touched between the two function calls!
worklSingle precision work array, length lwork. The array must not be touched between the two function calls!
lworklProvide the same argument as you did to ssaupd()
infoProvide the same argument as you did to ssaupd()