/*
SLEPc singular value solver: "trlanczos"
Method: Golub-Kahan-Lanczos bidiagonalization with thick-restart
Last update: Jun 2007
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
SLEPc - Scalable Library for Eigenvalue Problem Computations
Copyright (c) 2002-2009, Universidad Politecnica de Valencia, Spain
This file is part of SLEPc.
SLEPc is free software: you can redistribute it and/or modify it under the
terms of version 3 of the GNU Lesser General Public License as published by
the Free Software Foundation.
SLEPc is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for
more details.
You should have received a copy of the GNU Lesser General Public License
along with SLEPc. If not, see <http://www.gnu.org/licenses/>.
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
*/
#include "private/svdimpl.h" /*I "slepcsvd.h" I*/
#include "private/ipimpl.h"
#include "slepcblaslapack.h"
typedef struct {
PetscTruth oneside;
} SVD_TRLANCZOS;
#undef __FUNCT__
#define __FUNCT__ "SVDSetUp_TRLANCZOS"
PetscErrorCode SVDSetUp_TRLANCZOS(SVD svd)
{
PetscErrorCode ierr;
PetscInt i,N,nloc;
PetscScalar *pU;
PetscFunctionBegin;
ierr = SVDMatGetSize(svd,PETSC_NULL,&N);CHKERRQ(ierr);
if (svd->ncv) { /* ncv set */
if (svd->ncv<svd->nsv) SETERRQ(1,"The value of ncv must be at least nsv");
}
else if (svd->mpd) { /* mpd set */
svd->ncv = PetscMin(N,svd->nsv+svd->mpd);
}
else { /* neither set: defaults depend on nsv being small or large */
if (svd->nsv<500) svd->ncv = PetscMin(N,PetscMax(2*svd->nsv,10));
else { svd->mpd = 500; svd->ncv = PetscMin(N,svd->nsv+svd->mpd); }
}
if (!svd->mpd) svd->mpd = svd->ncv;
if (svd->ncv>svd->nsv+svd->mpd) SETERRQ(1,"The value of ncv must not be larger than nev+mpd");
if (!svd->max_it)
svd->max_it = PetscMax(N/svd->ncv,100);
if (svd->ncv!=svd->n) {
if (svd->U) {
ierr = VecGetArray(svd->U[0],&pU);CHKERRQ(ierr);
for (i=0;i<svd->n;i++) { ierr = VecDestroy(svd->U[i]); CHKERRQ(ierr); }
ierr = PetscFree(pU);CHKERRQ(ierr);
ierr = PetscFree(svd->U);CHKERRQ(ierr);
}
ierr = PetscMalloc(sizeof(Vec)*svd->ncv,&svd->U);CHKERRQ(ierr);
ierr = SVDMatGetLocalSize(svd,&nloc,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscMalloc(svd->ncv*nloc*sizeof(PetscScalar),&pU);CHKERRQ(ierr);
for (i=0;i<svd->ncv;i++) {
ierr = VecCreateMPIWithArray(((PetscObject)svd)->comm,nloc,PETSC_DECIDE,pU+i*nloc,&svd->U[i]);CHKERRQ(ierr);
}
}
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDOneSideTRLanczosMGS"
static PetscErrorCode SVDOneSideTRLanczosMGS(SVD svd,PetscReal *alpha,PetscReal *beta,PetscScalar* bb,Vec *V,Vec v,Vec* U,PetscInt nconv,PetscInt l,PetscInt n,PetscScalar* work,Vec wv,Vec wu)
{
PetscErrorCode ierr;
PetscReal a,b;
PetscInt i,k=nconv+l;
PetscFunctionBegin;
ierr = SVDMatMult(svd,PETSC_FALSE,V[k],U[k]);CHKERRQ(ierr);
if (l>0) {
ierr = VecSet(wu,0.0);CHKERRQ(ierr);
ierr = VecMAXPY(wu,l,bb,U+nconv);CHKERRQ(ierr);
ierr = VecAXPY(U[k],-1.0,wu);CHKERRQ(ierr);
}
ierr = IPNorm(svd->ip,U[k],&a);CHKERRQ(ierr);
ierr = VecScale(U[k],1.0/a);CHKERRQ(ierr);
alpha[0] = a;
for (i=k+1;i<n;i++) {
ierr = SVDMatMult(svd,PETSC_TRUE,U[i-1],V[i]);CHKERRQ(ierr);
ierr = IPOrthogonalize(svd->ip,i,PETSC_NULL,V,V[i],work,&b,PETSC_NULL,wv,PETSC_NULL);CHKERRQ(ierr);
ierr = VecScale(V[i],1.0/b);CHKERRQ(ierr);
beta[i-k-1] = b;
ierr = SVDMatMult(svd,PETSC_FALSE,V[i],U[i]);CHKERRQ(ierr);
ierr = VecAXPY(U[i],-b,U[i-1]);CHKERRQ(ierr);
ierr = IPNorm(svd->ip,U[i],&a);CHKERRQ(ierr);
ierr = VecScale(U[i],1.0/a);CHKERRQ(ierr);
alpha[i-k] = a;
}
ierr = SVDMatMult(svd,PETSC_TRUE,U[n-1],v);CHKERRQ(ierr);
ierr = IPOrthogonalize(svd->ip,n,PETSC_NULL,V,v,work,&b,PETSC_NULL,wv,PETSC_NULL);CHKERRQ(ierr);
beta[n-k-1] = b;
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDOneSideTRLanczosCGS"
static PetscErrorCode SVDOneSideTRLanczosCGS(SVD svd,PetscReal *alpha,PetscReal *beta,PetscScalar* bb,Vec *V,Vec v,Vec* U,PetscInt nconv,PetscInt l,PetscInt n,PetscScalar* work,Vec wv,Vec wu)
{
PetscErrorCode ierr;
PetscReal a,b,sum,onorm;
PetscScalar dot;
PetscInt i,j,k=nconv+l;
PetscFunctionBegin;
ierr = SVDMatMult(svd,PETSC_FALSE,V[k],U[k]);CHKERRQ(ierr);
if (l>0) {
ierr = VecSet(wu,0.0);CHKERRQ(ierr);
ierr = VecMAXPY(wu,l,bb,U+nconv);CHKERRQ(ierr);
ierr = VecAXPY(U[k],-1.0,wu);CHKERRQ(ierr);
}
for (i=k+1;i<n;i++) {
ierr = SVDMatMult(svd,PETSC_TRUE,U[i-1],V[i]);CHKERRQ(ierr);
ierr = IPNormBegin(svd->ip,U[i-1],&a);CHKERRQ(ierr);
if (svd->ip->orthog_ref == IP_ORTH_REFINE_IFNEEDED) {
ierr = IPInnerProductBegin(svd->ip,V[i],V[i],&dot);CHKERRQ(ierr);
}
ierr = IPMInnerProductBegin(svd->ip,V[i],i,V,work);CHKERRQ(ierr);
ierr = IPNormEnd(svd->ip,U[i-1],&a);CHKERRQ(ierr);
if (svd->ip->orthog_ref == IP_ORTH_REFINE_IFNEEDED) {
ierr = IPInnerProductEnd(svd->ip,V[i],V[i],&dot);CHKERRQ(ierr);
}
ierr = IPMInnerProductEnd(svd->ip,V[i],i,V,work);CHKERRQ(ierr);
ierr = VecScale(U[i-1],1.0/a);CHKERRQ(ierr);
ierr = VecScale(V[i],1.0/a);CHKERRQ(ierr);
for (j=0;j<i;j++) work[j] = - work[j] / a;
ierr = VecMAXPY(V[i],i,work,V);CHKERRQ(ierr);
switch (svd->ip->orthog_ref) {
case IP_ORTH_REFINE_NEVER:
ierr = IPNorm(svd->ip,V[i],&b);CHKERRQ(ierr);
break;
case IP_ORTH_REFINE_ALWAYS:
ierr = IPOrthogonalizeCGS(svd->ip,i,PETSC_NULL,V,V[i],work,PETSC_NULL,&b,wv);CHKERRQ(ierr);
break;
case IP_ORTH_REFINE_IFNEEDED:
onorm = sqrt(PetscRealPart(dot)) / a;
sum = 0.0;
for (j=0;j<i;j++) {
sum += PetscRealPart(work[j] * PetscConj(work[j]));
}
b = PetscRealPart(dot)/(a*a) - sum;
if (b>0.0) b = sqrt(b);
else {
ierr = IPNorm(svd->ip,V[i],&b);CHKERRQ(ierr);
}
if (b < svd->ip->orthog_eta * onorm) {
ierr = IPOrthogonalizeCGS(svd->ip,i,PETSC_NULL,V,V[i],work,PETSC_NULL,&b,wv);CHKERRQ(ierr);
}
break;
}
ierr = VecScale(V[i],1.0/b);CHKERRQ(ierr);
ierr = SVDMatMult(svd,PETSC_FALSE,V[i],U[i]);CHKERRQ(ierr);
ierr = VecAXPY(U[i],-b,U[i-1]);CHKERRQ(ierr);
alpha[i-k-1] = a;
beta[i-k-1] = b;
}
ierr = SVDMatMult(svd,PETSC_TRUE,U[n-1],v);CHKERRQ(ierr);
ierr = IPNormBegin(svd->ip,U[n-1],&a);CHKERRQ(ierr);
if (svd->ip->orthog_ref == IP_ORTH_REFINE_IFNEEDED) {
ierr = IPInnerProductBegin(svd->ip,v,v,&dot);CHKERRQ(ierr);
}
ierr = IPMInnerProductBegin(svd->ip,v,n,V,work);CHKERRQ(ierr);
ierr = IPNormEnd(svd->ip,U[n-1],&a);CHKERRQ(ierr);
if (svd->ip->orthog_ref == IP_ORTH_REFINE_IFNEEDED) {
ierr = IPInnerProductEnd(svd->ip,v,v,&dot);CHKERRQ(ierr);
}
ierr = IPMInnerProductEnd(svd->ip,v,n,V,work);CHKERRQ(ierr);
ierr = VecScale(U[n-1],1.0/a);CHKERRQ(ierr);
ierr = VecScale(v,1.0/a);CHKERRQ(ierr);
for (j=0;j<n;j++) work[j] = - work[j] / a;
ierr = VecMAXPY(v,n,work,V);CHKERRQ(ierr);
switch (svd->ip->orthog_ref) {
case IP_ORTH_REFINE_NEVER:
ierr = IPNorm(svd->ip,v,&b);CHKERRQ(ierr);
break;
case IP_ORTH_REFINE_ALWAYS:
ierr = IPOrthogonalizeCGS(svd->ip,n,PETSC_NULL,V,v,work,PETSC_NULL,&b,wv);CHKERRQ(ierr);
break;
case IP_ORTH_REFINE_IFNEEDED:
onorm = sqrt(PetscRealPart(dot)) / a;
sum = 0.0;
for (j=0;j<i;j++) {
sum += PetscRealPart(work[j] * PetscConj(work[j]));
}
b = PetscRealPart(dot)/(a*a) - sum;
if (b>0.0) b = sqrt(b);
else {
ierr = IPNorm(svd->ip,v,&b);CHKERRQ(ierr);
}
if (b < svd->ip->orthog_eta * onorm) {
ierr = IPOrthogonalizeCGS(svd->ip,n,PETSC_NULL,V,v,work,PETSC_NULL,&b,wv);CHKERRQ(ierr);
}
break;
}
alpha[n-k-1] = a;
beta[n-k-1] = b;
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDSolve_TRLANCZOS"
PetscErrorCode SVDSolve_TRLANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_TRLANCZOS *lanczos = (SVD_TRLANCZOS *)svd->data;
PetscReal *alpha,*beta,norm;
PetscScalar *b,*Q,*PT,*swork;
PetscInt i,j,k,l,m,n,nv;
Vec v,wv,wu;
PetscTruth conv;
IPOrthogonalizationType orthog;
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,&b);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n*svd->n,&Q);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n*svd->n,&PT);CHKERRQ(ierr);
#if !defined(PETSC_USE_COMPLEX)
if (svd->which == SVD_SMALLEST) {
#endif
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n*svd->n,&swork);CHKERRQ(ierr);
#if !defined(PETSC_USE_COMPLEX)
} else {
ierr = PetscMalloc(sizeof(PetscScalar)*svd->n,&swork);CHKERRQ(ierr);
}
#endif
ierr = VecDuplicate(svd->V[0],&v);CHKERRQ(ierr);
ierr = VecDuplicate(svd->V[0],&wv);CHKERRQ(ierr);
ierr = VecDuplicate(svd->U[0],&wu);CHKERRQ(ierr);
ierr = IPGetOrthogonalization(svd->ip,&orthog,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
/* normalize start vector */
ierr = VecCopy(svd->vec_initial,svd->V[0]);CHKERRQ(ierr);
ierr = VecNormalize(svd->V[0],&norm);CHKERRQ(ierr);
l = 0;
while (svd->reason == SVD_CONVERGED_ITERATING) {
svd->its++;
/* inner loop */
nv = PetscMin(svd->nconv+svd->mpd,svd->n);
if (lanczos->oneside) {
if (orthog == IP_MGS_ORTH) {
ierr = SVDOneSideTRLanczosMGS(svd,alpha,beta,b+svd->nconv,svd->V,v,svd->U,svd->nconv,l,nv,swork,wv,wu);CHKERRQ(ierr);
} else {
ierr = SVDOneSideTRLanczosCGS(svd,alpha,beta,b+svd->nconv,svd->V,v,svd->U,svd->nconv,l,nv,swork,wv,wu);CHKERRQ(ierr);
}
} else {
ierr = SVDTwoSideLanczos(svd,alpha,beta,svd->V,v,svd->U,svd->nconv+l,nv,swork,wv,wu);CHKERRQ(ierr);
}
ierr = VecScale(v,1.0/beta[nv-svd->nconv-l-1]);CHKERRQ(ierr);
/* compute SVD of general matrix */
n = nv - svd->nconv;
/* first l columns */
for (j=0;j<l;j++) {
for (i=0;i<j;i++) Q[j*n+i] = 0.0;
Q[j*n+j] = svd->sigma[svd->nconv+j];
for (i=j+1;i<n;i++) Q[j*n+i] = 0.0;
}
/* l+1 column */
for (i=0;i<l;i++) Q[l*n+i] = b[i+svd->nconv];
Q[l*n+l] = alpha[0];
for (i=l+1;i<n;i++) Q[l*n+i] = 0.0;
/* rest of matrix */
for (j=l+1;j<n;j++) {
for (i=0;i<j-1;i++) Q[j*n+i] = 0.0;
Q[j*n+j-1] = beta[j-l-1];
Q[j*n+j] = alpha[j-l];
for (i=j+1;i<n;i++) Q[j*n+i] = 0.0;
}
ierr = SVDDense(n,n,Q,alpha,PETSC_NULL,PT);CHKERRQ(ierr);
/* compute error estimates */
k = 0;
conv = PETSC_TRUE;
for (i=svd->nconv;i<nv;i++) {
if (svd->which == SVD_SMALLEST) j = n-i+svd->nconv-1;
else j = i-svd->nconv;
svd->sigma[i] = alpha[j];
b[i] = Q[j*n+n-1]*beta[n-l-1];
svd->errest[i] = PetscAbsScalar(b[i]);
if (alpha[j] > svd->tol) svd->errest[i] /= alpha[j];
if (conv) {
if (svd->errest[i] < svd->tol) k++;
else conv = PETSC_FALSE;
}
}
/* check convergence and update l */
if (svd->its >= svd->max_it) svd->reason = SVD_DIVERGED_ITS;
if (svd->nconv+k >= svd->nsv) svd->reason = SVD_CONVERGED_TOL;
if (svd->reason != SVD_CONVERGED_ITERATING) l = 0;
else l = PetscMax((nv - svd->nconv - k) / 2,0);
/* compute converged singular vectors and restart vectors*/
#if !defined(PETSC_USE_COMPLEX)
if (svd->which == SVD_SMALLEST) {
#endif
for (i=0;i<k+l;i++) {
if (svd->which == SVD_SMALLEST) j = n-i-1;
else j = i;
for (m=0;m<n;m++) swork[j*n+m] = PT[m*n+j];
}
ierr = SlepcUpdateVectors(n,svd->V+svd->nconv,0,k+l,swork,n,PETSC_FALSE);CHKERRQ(ierr);
for (i=0;i<k+l;i++) {
if (svd->which == SVD_SMALLEST) j = n-i-1;
else j = i;
for (m=0;m<n;m++) swork[j*n+m] = Q[j*n+m];
}
ierr = SlepcUpdateVectors(n,svd->U+svd->nconv,0,k+l,swork,n,PETSC_FALSE);CHKERRQ(ierr);
#if !defined(PETSC_USE_COMPLEX)
} else {
ierr = SlepcUpdateVectors(n,svd->V+svd->nconv,0,k+l,PT,n,PETSC_TRUE);CHKERRQ(ierr);
ierr = SlepcUpdateVectors(n,svd->U+svd->nconv,0,k+l,Q,n,PETSC_FALSE);CHKERRQ(ierr);
}
#endif
/* copy the last vector to be the next initial vector */
if (svd->reason == SVD_CONVERGED_ITERATING) {
ierr = VecCopy(v,svd->V[svd->nconv+k+l]);CHKERRQ(ierr);
}
svd->nconv += k;
SVDMonitor(svd,svd->its,svd->nconv,svd->sigma,svd->errest,nv);
}
/* orthonormalize U columns in one side method */
if (lanczos->oneside) {
for (i=0;i<svd->nconv;i++) {
ierr = IPOrthogonalize(svd->ip,i,PETSC_NULL,svd->U,svd->U[i],PETSC_NULL,&norm,PETSC_NULL,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
ierr = VecScale(svd->U[i],1.0/norm);CHKERRQ(ierr);
}
}
/* free working space */
ierr = VecDestroy(v);CHKERRQ(ierr);
ierr = VecDestroy(wv);CHKERRQ(ierr);
ierr = VecDestroy(wu);CHKERRQ(ierr);
ierr = PetscFree(alpha);CHKERRQ(ierr);
ierr = PetscFree(beta);CHKERRQ(ierr);
ierr = PetscFree(b);CHKERRQ(ierr);
ierr = PetscFree(Q);CHKERRQ(ierr);
ierr = PetscFree(PT);CHKERRQ(ierr);
ierr = PetscFree(swork);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDSetFromOptions_TRLANCZOS"
PetscErrorCode SVDSetFromOptions_TRLANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_TRLANCZOS *lanczos = (SVD_TRLANCZOS *)svd->data;
PetscFunctionBegin;
ierr = PetscOptionsBegin(((PetscObject)svd)->comm,((PetscObject)svd)->prefix,"TRLANCZOS Singular Value Solver Options","SVD");CHKERRQ(ierr);
ierr = PetscOptionsTruth("-svd_trlanczos_oneside","Lanczos one-side reorthogonalization","SVDTRLanczosSetOneSide",PETSC_FALSE,&lanczos->oneside,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
PetscFunctionReturn(0);
}
EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "SVDTRLanczosSetOneSide_TRLANCZOS"
PetscErrorCode SVDTRLanczosSetOneSide_TRLANCZOS(SVD svd,PetscTruth oneside)
{
SVD_TRLANCZOS *lanczos = (SVD_TRLANCZOS *)svd->data;
PetscFunctionBegin;
lanczos->oneside = oneside;
PetscFunctionReturn(0);
}
EXTERN_C_END
#undef __FUNCT__
#define __FUNCT__ "SVDTRLanczosSetOneSide"
/*@
SVDTRLanczosSetOneSide - Indicate if the variant of the Lanczos method
to be used is one-sided or two-sided.
Collective on SVD
Input Parameters:
+ svd - singular value solver
- oneside - boolean flag indicating if the method is one-sided or not
Options Database Key:
. -svd_trlanczos_oneside <boolean> - Indicates the boolean flag
Note:
By default, a two-sided variant is selected, which is sometimes slightly
more robust. However, the one-sided variant is faster because it avoids
the orthogonalization associated to left singular vectors.
Level: advanced
.seealso: SVDLanczosSetOneSide()
@*/
PetscErrorCode SVDTRLanczosSetOneSide(SVD svd,PetscTruth oneside)
{
PetscErrorCode ierr, (*f)(SVD,PetscTruth);
PetscFunctionBegin;
PetscValidHeaderSpecific(svd,SVD_COOKIE,1);
ierr = PetscObjectQueryFunction((PetscObject)svd,"SVDTRLanczosSetOneSide_C",(void (**)())&f);CHKERRQ(ierr);
if (f) {
ierr = (*f)(svd,oneside);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "SVDView_TRLANCZOS"
PetscErrorCode SVDView_TRLANCZOS(SVD svd,PetscViewer viewer)
{
PetscErrorCode ierr;
SVD_TRLANCZOS *lanczos = (SVD_TRLANCZOS *)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_TRLANCZOS"
PetscErrorCode SVDCreate_TRLANCZOS(SVD svd)
{
PetscErrorCode ierr;
SVD_TRLANCZOS *lanczos;
PetscFunctionBegin;
ierr = PetscNew(SVD_TRLANCZOS,&lanczos);CHKERRQ(ierr);
PetscLogObjectMemory(svd,sizeof(SVD_TRLANCZOS));
svd->data = (void *)lanczos;
svd->ops->setup = SVDSetUp_TRLANCZOS;
svd->ops->solve = SVDSolve_TRLANCZOS;
svd->ops->destroy = SVDDestroy_Default;
svd->ops->setfromoptions = SVDSetFromOptions_TRLANCZOS;
svd->ops->view = SVDView_TRLANCZOS;
lanczos->oneside = PETSC_FALSE;
ierr = PetscObjectComposeFunctionDynamic((PetscObject)svd,"SVDTRLanczosSetOneSide_C","SVDTRLanczosSetOneSide_TRLANCZOS",SVDTRLanczosSetOneSide_TRLANCZOS);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
EXTERN_C_END