/*
SLEPc singular value solver: "lanczos"
Method: Golub-Kahan-Lanczos bidiagonalization
Last update: Nov 2006
*/
#include "src/svd/svdimpl.h" /*I "slepcsvd.h" I*/
#include "slepcblaslapack.h"
typedef struct {
PetscTruth oneside;
} SVD_LANCZOS;
#undef __FUNCT__
#define __FUNCT__ "SVDSetUp_LANCZOS"
PetscErrorCode SVDSetUp_LANCZOS(SVD svd)
{
PetscErrorCode ierr;
PetscInt M,N;
PetscFunctionBegin;
ierr = MatGetSize(svd->A,&M,&N);CHKERRQ(ierr);
if (svd->ncv == PETSC_DECIDE)
svd->ncv = PetscMin(PetscMin(M,N),PetscMax(2*svd->nsv,10));
if (svd->max_it == PETSC_DECIDE)
svd->max_it = PetscMax(PetscMin(M,N)/svd->ncv,100);
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDSolve_LANCZOS"
PetscErrorCode SVDSolve_LANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_LANCZOS *lanczos = (SVD_LANCZOS *)svd->data;
PetscReal *alpha,*beta,norm,*work;
PetscScalar *Q,*PT;
PetscInt *perm;
int i,j,k,l,n,zero=0,info;
Vec *V,*U;
PetscTruth conv;
PetscFunctionBegin;
/* allocate working space */
ierr = PetscMalloc(sizeof(PetscReal)*svd->n,&alpha);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscReal)*svd->n,&beta);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n*svd->n,&Q);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n*svd->n,&PT);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscReal)*4*svd->n,&work);CHKERRQ(ierr);
ierr = VecDuplicateVecs(svd->V[0],svd->n+1,&V);CHKERRQ(ierr);
ierr = VecDuplicateVecs(svd->U[0],svd->n,&U);CHKERRQ(ierr);
/* normalize start vector */
ierr = VecCopy(svd->vec_initial,V[0]);CHKERRQ(ierr);
ierr = VecNormalize(V[0],&norm);CHKERRQ(ierr);
while (svd->reason == SVD_CONVERGED_ITERATING) {
svd->its++;
/* inner loop */
for (i=svd->nconv;i<svd->n;i++) {
svd->matvecs++;
ierr = MatMult(svd->A,V[i],U[i]);CHKERRQ(ierr);
if (lanczos->oneside) {
if (i>svd->nconv) { ierr = VecAXPY(U[i],-beta[i-svd->nconv-1],U[i-1]);CHKERRQ(ierr); }
} else {
svd->dots += i;
ierr = IPOrthogonalize(svd->ip,i,PETSC_NULL,U,U[i],PT,alpha+i-svd->nconv,PETSC_NULL);CHKERRQ(ierr);
ierr = VecScale(U[i],1.0/alpha[i-svd->nconv]);CHKERRQ(ierr);
}
svd->matvecs++;
if (svd->AT) {
ierr = MatMult(svd->AT,U[i],V[i+1]);CHKERRQ(ierr);
} else {
ierr = MatMultTranspose(svd->A,U[i],V[i+1]);CHKERRQ(ierr);
}
if (lanczos->oneside) {
svd->dots += i+1;
ierr = VecNormBegin(U[i],NORM_2,alpha+i-svd->nconv);CHKERRQ(ierr);
ierr = VecMDotBegin(V[i+1],i+1,V,PT);CHKERRQ(ierr);
ierr = VecNormEnd(U[i],NORM_2,alpha+i-svd->nconv);CHKERRQ(ierr);
ierr = VecMDotEnd(V[i+1],i+1,V,PT);CHKERRQ(ierr);
ierr = VecScale(U[i],1.0/alpha[i-svd->nconv]);CHKERRQ(ierr);
ierr = VecScale(V[i+1],1.0/alpha[i-svd->nconv]);CHKERRQ(ierr);
for (j=0;j<=i;j++) PT[j] = - PT[j] / alpha[i-svd->nconv];
ierr = VecMAXPY(V[i+1],i+1,PT,V);CHKERRQ(ierr);
ierr = IPOrthogonalizeGS(svd->ip,i+1,PETSC_NULL,V,V[i+1],PT,PETSC_NULL,beta+i-svd->nconv);CHKERRQ(ierr);
ierr = VecScale(V[i+1],1.0/beta[i-svd->nconv]);CHKERRQ(ierr);
} else {
svd->dots += i+1;
ierr = IPOrthogonalize(svd->ip,i+1,PETSC_NULL,V,V[i+1],PT,beta+i-svd->nconv,PETSC_NULL);CHKERRQ(ierr);
ierr = VecScale(V[i+1],1.0/beta[i-svd->nconv]);CHKERRQ(ierr);
}
}
/* compute SVD of bidiagonal matrix */
n = svd->n - svd->nconv;
ierr = PetscMemzero(PT,sizeof(PetscScalar)*n*n);CHKERRQ(ierr);
ierr = PetscMemzero(Q,sizeof(PetscScalar)*n*n);CHKERRQ(ierr);
for (i=0;i<n;i++)
PT[i*n+i] = Q[i*n+i] = 1.0;
LAPACKbdsqr_("U",&n,&n,&n,&zero,alpha,beta,PT,&n,Q,&n,PETSC_NULL,&n,work,&info,1);
/* compute error estimates and converged singular vectors */
k = svd->nconv;
conv = PETSC_TRUE;
for (i=svd->nconv;i<svd->n;i++) {
if (svd->which == SVD_SMALLEST) j = n-i+svd->nconv-1;
else j = i-svd->nconv;
svd->sigma[i] = alpha[j];
svd->errest[i] = PetscAbsReal(Q[j*n+n-1])*beta[n-1] / alpha[j];
if (conv) {
if (svd->errest[i] < svd->tol) {
ierr = VecSet(svd->V[i],0.0);CHKERRQ(ierr);
for (l=0;l<n;l++) {
ierr = VecAXPY(svd->V[i],PT[l*n+j],V[l+svd->nconv]);CHKERRQ(ierr);
}
ierr = VecSet(svd->U[i],0.0);CHKERRQ(ierr);
ierr = VecMAXPY(svd->U[i],n,Q+j*n,U+svd->nconv);CHKERRQ(ierr);
k++;
} else conv = PETSC_FALSE;
}
}
if (svd->its > svd->max_it) svd->reason = SVD_DIVERGED_ITS;
if (k >= svd->nsv) svd->reason = SVD_CONVERGED_TOL;
if (svd->reason == SVD_CONVERGED_ITERATING) {
/* compute restart vector */
if (svd->which == SVD_SMALLEST) j = n-k+svd->nconv-1;
else j = k-svd->nconv;
ierr = VecSet(svd->V[k],0.0);CHKERRQ(ierr);
for (l=0;l<n;l++) {
ierr = VecAXPY(svd->V[k],PT[l*n+j],V[l+svd->nconv]);CHKERRQ(ierr);
}
ierr = VecCopy(svd->V[k],V[k]);CHKERRQ(ierr);
}
/* copy converged singular vectors from temporary space */
for (i=svd->nconv;i<k;i++) {
ierr = VecCopy(svd->V[i],V[i]);CHKERRQ(ierr);
ierr = VecCopy(svd->U[i],U[i]);CHKERRQ(ierr);
}
svd->nconv = k;
SVDMonitor(svd,svd->its,svd->nconv,svd->sigma,svd->errest,svd->n);
}
/* sort singular triplets */
ierr = PetscMalloc(sizeof(PetscInt)*svd->nconv,&perm);CHKERRQ(ierr);
for (i=0;i<svd->nconv;i++) {
alpha[i] = svd->sigma[i];
beta[i] = svd->errest[i];
perm[i] = i;
}
ierr = PetscSortRealWithPermutation(svd->nconv,svd->sigma,perm);CHKERRQ(ierr);
for (i=0;i<svd->nconv;i++) {
if (svd->which == SVD_SMALLEST) j = perm[i];
else j = perm[svd->nconv-i-1];
svd->sigma[i] = alpha[j];
svd->errest[i] = beta[j];
ierr = VecCopy(V[j],svd->V[i]);CHKERRQ(ierr);
ierr = VecCopy(U[j],svd->U[i]);CHKERRQ(ierr);
}
/* free working space */
ierr = VecDestroyVecs(V,svd->n+1);CHKERRQ(ierr);
ierr = VecDestroyVecs(U,svd->n);CHKERRQ(ierr);
ierr = PetscFree(alpha);CHKERRQ(ierr);
ierr = PetscFree(beta);CHKERRQ(ierr);
ierr = PetscFree(Q);CHKERRQ(ierr);
ierr = PetscFree(PT);CHKERRQ(ierr);
ierr = PetscFree(work);CHKERRQ(ierr);
ierr = PetscFree(perm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDSetFromOptions_LANCZOS"
PetscErrorCode SVDSetFromOptions_LANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_LANCZOS *lanczos = (SVD_LANCZOS *)svd->data;
PetscFunctionBegin;
ierr = PetscOptionsBegin(svd->comm,svd->prefix,"LANCZOS Singular Value Solver Options","SVD");CHKERRQ(ierr);
ierr = PetscOptionsTruth("-svd_lanczos_oneside","Lanczos one-side reorthogonalization","SVDLanczosSetOneSideReorthogonalization",PETSC_FALSE,&lanczos->oneside,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
PetscFunctionReturn(0);
}
EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "SVDLanczosSetOneSideReorthogonalization_LANCZOS"
PetscErrorCode SVDLanczosSetOneSideReorthogonalization_LANCZOS(SVD svd,PetscTruth oneside)
{
SVD_LANCZOS *lanczos = (SVD_LANCZOS *)svd->data;
PetscFunctionBegin;
lanczos->oneside = oneside;
PetscFunctionReturn(0);
}
EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "SVDLanczosSetOneSideReorthogonalization"
PetscErrorCode SVDLanczosSetOneSideReorthogonalization(SVD svd,PetscTruth oneside)
{
PetscErrorCode ierr, (*f)(SVD,PetscTruth);
PetscFunctionBegin;
PetscValidHeaderSpecific(svd,SVD_COOKIE,1);
ierr = PetscObjectQueryFunction((PetscObject)svd,"SVDLanczosSetOneSideReorthogonalization_C",(void (**)())&f);CHKERRQ(ierr);
if (f) {
ierr = (*f)(svd,oneside);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDView_LANCZOS"
PetscErrorCode SVDView_LANCZOS(SVD svd,PetscViewer viewer)
{
PetscErrorCode ierr;
SVD_LANCZOS *lanczos = (SVD_LANCZOS *)svd->data;
PetscFunctionBegin;
ierr = PetscViewerASCIIPrintf(viewer,"Lanczos reorthogonalization: %s\n",lanczos->oneside ? "one-side" : "two-side");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "SVDCreate_LANCZOS"
PetscErrorCode SVDCreate_LANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_LANCZOS *lanczos;
PetscFunctionBegin;
ierr = PetscNew(SVD_LANCZOS,&lanczos);CHKERRQ(ierr);
PetscLogObjectMemory(svd,sizeof(SVD_LANCZOS));
svd->data = (void *)lanczos;
svd->ops->setup = SVDSetUp_LANCZOS;
svd->ops->solve = SVDSolve_LANCZOS;
svd->ops->setfromoptions = SVDSetFromOptions_LANCZOS;
svd->ops->view = SVDView_LANCZOS;
lanczos->oneside = PETSC_FALSE;
ierr = PetscObjectComposeFunctionDynamic((PetscObject)svd,"SVDLanczosSetOneSideReorthogonalization_C","SVDLanczosSetOneSideReorthogonalization_LANCZOS",SVDLanczosSetOneSideReorthogonalization_LANCZOS);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
EXTERN_C_END