Subversion Repositories slepc-dev

/*                      

   Common subroutines for all Krylov-type solvers.

   - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

   SLEPc - Scalable Library for Eigenvalue Problem Computations

   Copyright (c) 2002-2010, 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/epsimpl.h"                /*I "slepceps.h" I*/

#include "private/slepcimpl.h"

#include "slepcblaslapack.h"

#undef __FUNCT__  

#define __FUNCT__ "EPSBasicArnoldi"

/*

   EPSBasicArnoldi - Computes an m-step Arnoldi factorization. The first k

   columns are assumed to be locked and therefore they are not modified. On

   exit, the following relation is satisfied:

                    OP * V - V * H = f * e_m^T

   where the columns of V are the Arnoldi vectors (which are B-orthonormal),

   H is an upper Hessenberg matrix, f is the residual vector and e_m is

   the m-th vector of the canonical basis. The vector f is B-orthogonal to

   the columns of V. On exit, beta contains the B-norm of f and the next

   Arnoldi vector can be computed as v_{m+1} = f / beta.

*/

PetscErrorCode EPSBasicArnoldi(EPS eps,PetscTruth trans,PetscScalar *H,PetscInt ldh,Vec *V,PetscInt k,PetscInt *M,Vec f,PetscReal *beta,PetscTruth *breakdown)

{

  PetscErrorCode ierr;

  PetscInt       j,m = *M;

  PetscReal      norm;

  PetscFunctionBegin;

  for (j=k;j<m-1;j++) {

    if (trans) { ierr = STApplyTranspose(eps->OP,V[j],V[j+1]);CHKERRQ(ierr); }

    else { ierr = STApply(eps->OP,V[j],V[j+1]);CHKERRQ(ierr); }

    ierr = IPOrthogonalize(eps->ip,eps->nds,eps->DS,j+1,PETSC_NULL,V,V[j+1],H+ldh*j,&norm,breakdown);CHKERRQ(ierr);

    H[j+1+ldh*j] = norm;

    if (*breakdown) {

      *M = j+1;

      *beta = norm;

      PetscFunctionReturn(0);

    } else {

      ierr = VecScale(V[j+1],1/norm);CHKERRQ(ierr);

    }

  }

  if (trans) { ierr = STApplyTranspose(eps->OP,V[m-1],f);CHKERRQ(ierr); }

  else { ierr = STApply(eps->OP,V[m-1],f);CHKERRQ(ierr); }

  ierr = IPOrthogonalize(eps->ip,eps->nds,eps->DS,m,PETSC_NULL,V,f,H+ldh*(m-1),beta,PETSC_NULL);CHKERRQ(ierr);

  PetscFunctionReturn(0);

}

#undef __FUNCT__  

#define __FUNCT__ "EPSKrylovConvergence"

/*

   EPSKrylovConvergence - Implements the loop that checks for convergence

   in Krylov methods.

   Input Parameters:

     eps   - the eigensolver; some error estimates are updated in eps->errest

     issym - whether the projected problem is symmetric or not

     kini  - initial value of k (the loop variable)

     nits  - number of iterations of the loop

     S     - Schur form of projected matrix (not referenced if issym)

     lds   - leading dimension of S

     Q     - Schur vectors of projected matrix (eigenvectors if issym)

     V     - set of basis vectors (used only if trueresidual is activated)

     nv    - number of vectors to process (dimension of Q, columns of V)

     beta  - norm of f (the residual vector of the Arnoldi/Lanczos factorization)

     corrf - correction factor for residual estimates (only in harmonic KS)

   Output Parameters:

     kout  - the first index where the convergence test failed

   Workspace:

     work is workspace to store 5*nv scalars, nv booleans and nv reals (only if !issym)

*/

PetscErrorCode EPSKrylovConvergence(EPS eps,PetscTruth issym,PetscInt kini,PetscInt nits,PetscScalar *S,PetscInt lds,PetscScalar *Q,Vec *V,PetscInt nv,PetscReal beta,PetscReal corrf,PetscInt *kout,PetscScalar *work)

{

  PetscErrorCode ierr;

  PetscInt       k,marker;

  PetscScalar    re,im,*Z = work,*work2 = work;

  PetscReal      resnorm;

  PetscTruth     iscomplex,isshift;

  PetscFunctionBegin;

  if (!issym) { Z = work; work2 = work+2*nv; }

  ierr = PetscTypeCompare((PetscObject)eps->OP,STSHIFT,&isshift);CHKERRQ(ierr);

  marker = -1;

  for (k=kini;k<kini+nits;k++) {

    /* eigenvalue */

    re = eps->eigr[k];

    im = eps->eigi[k];

    if (eps->trueres || isshift) {

      ierr = STBackTransform(eps->OP,1,&re,&im);CHKERRQ(ierr);

    }

    iscomplex = PETSC_FALSE;

    if (!issym && k<nv-1 && S[k+1+k*lds] != 0.0) iscomplex = PETSC_TRUE;

    /* residual norm */

    if (issym) {

      resnorm = beta*PetscAbsScalar(Q[(k-kini+1)*nv-1]);

    } else {

      ierr = DenseSelectedEvec(S,lds,Q,Z,k,iscomplex,nv,work2);CHKERRQ(ierr);

      if (iscomplex) resnorm = beta*SlepcAbsEigenvalue(Z[nv-1],Z[2*nv-1]);

      else resnorm = beta*PetscAbsScalar(Z[nv-1]);

    }

    if (eps->trueres) {

      if (issym) Z = Q+(k-kini)*nv;

      ierr = EPSComputeTrueResidual(eps,re,im,Z,V,nv,&resnorm);CHKERRQ(ierr);

    }

    else resnorm *= corrf;

    /* error estimate */

    ierr = (*eps->conv_func)(eps,re,im,resnorm,&eps->errest[k],eps->conv_ctx);CHKERRQ(ierr);

    if (marker==-1 && eps->errest[k] >= eps->tol) marker = k;

    if (iscomplex) { eps->errest[k+1] = eps->errest[k]; k++; }

    if (marker!=-1 && !eps->trackall) break;

  }

  if (marker!=-1) k = marker;

  *kout = k;

  PetscFunctionReturn(0);

}

#undef __FUNCT__  

#define __FUNCT__ "EPSFullLanczos"

/*

   EPSFullLanczos - Computes an m-step Lanczos factorization with full

   reorthogonalization.  At each Lanczos step, the corresponding Lanczos

   vector is orthogonalized with respect to all previous Lanczos vectors.

   This is equivalent to computing an m-step Arnoldi factorization and

   exploting symmetry of the operator.

   The first k columns are assumed to be locked and therefore they are

   not modified. On exit, the following relation is satisfied:

                    OP * V - V * T = f * e_m^T

   where the columns of V are the Lanczos vectors (which are B-orthonormal),

   T is a real symmetric tridiagonal matrix, f is the residual vector and e_m

   is the m-th vector of the canonical basis. The tridiagonal is stored as

   two arrays: alpha contains the diagonal elements, beta the off-diagonal.

   The vector f is B-orthogonal to the columns of V. On exit, the last element

   of beta contains the B-norm of f and the next Lanczos vector can be

   computed as v_{m+1} = f / beta(end).

*/

PetscErrorCode EPSFullLanczos(EPS eps,PetscReal *alpha,PetscReal *beta,Vec *V,PetscInt k,PetscInt *M,Vec f,PetscTruth *breakdown)

{

  PetscErrorCode ierr;

  PetscInt       j,m = *M;

  PetscReal      norm;

  PetscScalar    *hwork,lhwork[100];

  PetscFunctionBegin;

  if (m > 100) {

    ierr = PetscMalloc((eps->nds+m)*sizeof(PetscScalar),&hwork);CHKERRQ(ierr);

  } else {

    hwork = lhwork;

  }

  for (j=k;j<m-1;j++) {

    ierr = STApply(eps->OP,V[j],V[j+1]);CHKERRQ(ierr);

    ierr = IPOrthogonalize(eps->ip,eps->nds,eps->DS,j+1,PETSC_NULL,V,V[j+1],hwork,&norm,breakdown);CHKERRQ(ierr);

    alpha[j-k] = PetscRealPart(hwork[j]);

    beta[j-k] = norm;

    if (*breakdown) {

      *M = j+1;

      if (m > 100) {

        ierr = PetscFree(hwork);CHKERRQ(ierr);

      }

      PetscFunctionReturn(0);

    } else {

      ierr = VecScale(V[j+1],1.0/norm);CHKERRQ(ierr);

    }

  }

  ierr = STApply(eps->OP,V[m-1],f);CHKERRQ(ierr);

  ierr = IPOrthogonalize(eps->ip,eps->nds,eps->DS,m,PETSC_NULL,V,f,hwork,&norm,PETSC_NULL);CHKERRQ(ierr);

  alpha[m-1-k] = PetscRealPart(hwork[m-1]);

  beta[m-1-k] = norm;

  if (m > 100) {

    ierr = PetscFree(hwork);CHKERRQ(ierr);

  }

  PetscFunctionReturn(0);

}