Actual source code: cgne.c

  1: #define PETSCKSP_DLL

  3: /*
  4:        cgctx.h defines the simple data structured used to store information
  5:     related to the type of matrix (e.g. complex symmetric) being solved and
  6:     data used during the optional Lanczo process used to compute eigenvalues
  7: */
 8:  #include ../src/ksp/ksp/impls/cg/cgctx.h
  9: EXTERN PetscErrorCode KSPComputeExtremeSingularValues_CG(KSP,PetscReal *,PetscReal *);
 10: EXTERN PetscErrorCode KSPComputeEigenvalues_CG(KSP,PetscInt,PetscReal *,PetscReal *,PetscInt *);


 13: /*
 14:      KSPSetUp_CGNE - Sets up the workspace needed by the CGNE method. 

 16:      IDENTICAL TO THE CG ONE EXCEPT for one extra work vector!
 17: */
 20: PetscErrorCode KSPSetUp_CGNE(KSP ksp)
 21: {
 22:   KSP_CG         *cgP = (KSP_CG*)ksp->data;
 24:   PetscInt       maxit = ksp->max_it;

 27:   /* 
 28:        This implementation of CGNE only handles left preconditioning
 29:      so generate an error otherwise.
 30:   */
 31:   if (ksp->pc_side == PC_RIGHT) {
 32:     SETERRQ(PETSC_ERR_SUP,"No right preconditioning for KSPCGNE");
 33:   } else if (ksp->pc_side == PC_SYMMETRIC) {
 34:     SETERRQ(PETSC_ERR_SUP,"No symmetric preconditioning for KSPCGNE");
 35:   }

 37:   /* get work vectors needed by CGNE */
 38:   KSPDefaultGetWork(ksp,4);

 40:   /*
 41:      If user requested computations of eigenvalues then allocate work
 42:      work space needed
 43:   */
 44:   if (ksp->calc_sings) {
 45:     /* get space to store tridiagonal matrix for Lanczos */
 46:     PetscMalloc(2*(maxit+1)*sizeof(PetscScalar),&cgP->e);
 47:     PetscLogObjectMemory(ksp,2*(maxit+1)*sizeof(PetscScalar));
 48:     cgP->d                         = cgP->e + maxit + 1;
 49:     PetscMalloc(2*(maxit+1)*sizeof(PetscReal),&cgP->ee);
 50:     PetscLogObjectMemory(ksp,2*(maxit+1)*sizeof(PetscScalar));
 51:     cgP->dd                        = cgP->ee + maxit + 1;
 52:     ksp->ops->computeextremesingularvalues = KSPComputeExtremeSingularValues_CG;
 53:     ksp->ops->computeeigenvalues           = KSPComputeEigenvalues_CG;
 54:   }
 55:   return(0);
 56: }

 58: /*
 59:        KSPSolve_CGNE - This routine actually applies the conjugate gradient 
 60:     method

 62:    Input Parameter:
 63: .     ksp - the Krylov space object that was set to use conjugate gradient, by, for 
 64:             example, KSPCreate(MPI_Comm,KSP *ksp); KSPSetType(ksp,KSPCG);


 67:     Virtually identical to the KSPSolve_CG, it should definitely reuse the same code.

 69: */
 72: PetscErrorCode  KSPSolve_CGNE(KSP ksp)
 73: {
 75:   PetscInt       i,stored_max_it,eigs;
 76:   PetscScalar    dpi,a = 1.0,beta,betaold = 1.0,b = 0,*e = 0,*d = 0;
 77:   PetscReal      dp = 0.0;
 78:   Vec            X,B,Z,R,P,T;
 79:   KSP_CG         *cg;
 80:   Mat            Amat,Pmat;
 81:   MatStructure   pflag;
 82:   PetscTruth     diagonalscale,transpose_pc;

 85:   PCDiagonalScale(ksp->pc,&diagonalscale);
 86:   if (diagonalscale) SETERRQ1(PETSC_ERR_SUP,"Krylov method %s does not support diagonal scaling",((PetscObject)ksp)->type_name);
 87:   PCApplyTransposeExists(ksp->pc,&transpose_pc);

 89:   cg            = (KSP_CG*)ksp->data;
 90:   eigs          = ksp->calc_sings;
 91:   stored_max_it = ksp->max_it;
 92:   X             = ksp->vec_sol;
 93:   B             = ksp->vec_rhs;
 94:   R             = ksp->work[0];
 95:   Z             = ksp->work[1];
 96:   P             = ksp->work[2];
 97:   T             = ksp->work[3];

 99: #if !defined(PETSC_USE_COMPLEX)
100: #define VecXDot(x,y,a) VecDot(x,y,a)
101: #else
102: #define VecXDot(x,y,a) (((cg->type) == (KSP_CG_HERMITIAN)) ? VecDot(x,y,a) : VecTDot(x,y,a))
103: #endif

105:   if (eigs) {e = cg->e; d = cg->d; e[0] = 0.0; }
106:   PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);

108:   ksp->its = 0;
109:   MatMultTranspose(Amat,B,T);
110:   if (!ksp->guess_zero) {
111:     KSP_MatMult(ksp,Amat,X,P);
112:     KSP_MatMultTranspose(ksp,Amat,P,R);
113:     VecAYPX(R,-1.0,T);
114:   } else {
115:     VecCopy(T,R);              /*     r <- b (x is 0) */
116:   }
117:   KSP_PCApply(ksp,R,T);
118:   if (transpose_pc) {
119:     KSP_PCApplyTranspose(ksp,T,Z);
120:   } else {
121:     KSP_PCApply(ksp,T,Z);
122:   }

124:   if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
125:     VecNorm(Z,NORM_2,&dp); /*    dp <- z'*z       */
126:   } else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
127:     VecNorm(R,NORM_2,&dp); /*    dp <- r'*r       */
128:   } else if (ksp->normtype == KSP_NORM_NATURAL) {
129:     VecXDot(Z,R,&beta);
130:     dp = sqrt(PetscAbsScalar(beta));
131:   } else dp = 0.0;
132:   KSPLogResidualHistory(ksp,dp);
133:   KSPMonitor(ksp,0,dp);                              /* call any registered monitor routines */
134:   ksp->rnorm = dp;
135:   (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);      /* test for convergence */
136:   if (ksp->reason) return(0);

138:   i = 0;
139:   do {
140:      ksp->its = i+1;
141:      VecXDot(Z,R,&beta);     /*     beta <- r'z     */
142:      if (beta == 0.0) {
143:        ksp->reason = KSP_CONVERGED_ATOL;
144:        PetscInfo(ksp,"converged due to beta = 0\n");
145:        break;
146: #if !defined(PETSC_USE_COMPLEX)
147:      } else if (beta < 0.0) {
148:        ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
149:        PetscInfo(ksp,"diverging due to indefinite preconditioner\n");
150:        break;
151: #endif
152:      }
153:      if (!i) {
154:        VecCopy(Z,P);         /*     p <- z          */
155:        b = 0.0;
156:      } else {
157:        b = beta/betaold;
158:        if (eigs) {
159:          if (ksp->max_it != stored_max_it) {
160:            SETERRQ(PETSC_ERR_SUP,"Can not change maxit AND calculate eigenvalues");
161:          }
162:          e[i] = sqrt(PetscAbsScalar(b))/a;
163:        }
164:        VecAYPX(P,b,Z);    /*     p <- z + b* p   */
165:      }
166:      betaold = beta;
167:      MatMult(Amat,P,T);
168:      MatMultTranspose(Amat,T,Z);
169:      VecXDot(P,Z,&dpi);      /*     dpi <- z'p      */
170:      a = beta/dpi;                                 /*     a = beta/p'z    */
171:      if (eigs) {
172:        d[i] = sqrt(PetscAbsScalar(b))*e[i] + 1.0/a;
173:      }
174:      VecAXPY(X,a,P);          /*     x <- x + ap     */
175:      VecAXPY(R,-a,Z);                      /*     r <- r - az     */
176:      if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
177:        KSP_PCApply(ksp,R,T);
178:        if (transpose_pc) {
179:          KSP_PCApplyTranspose(ksp,T,Z);
180:        } else {
181:          KSP_PCApply(ksp,T,Z);
182:        }
183:        VecNorm(Z,NORM_2,&dp);              /*    dp <- z'*z       */
184:      } else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
185:        VecNorm(R,NORM_2,&dp);
186:      } else if (ksp->normtype == KSP_NORM_NATURAL) {
187:        dp = sqrt(PetscAbsScalar(beta));
188:      } else {
189:        dp = 0.0;
190:      }
191:      ksp->rnorm = dp;
192:      KSPLogResidualHistory(ksp,dp);
193:      KSPMonitor(ksp,i+1,dp);
194:      (*ksp->converged)(ksp,i+1,dp,&ksp->reason,ksp->cnvP);
195:      if (ksp->reason) break;
196:      if (ksp->normtype != KSP_NORM_PRECONDITIONED) {
197:        if (transpose_pc) {
198:          KSP_PCApplyTranspose(ksp,T,Z);
199:        } else {
200:          KSP_PCApply(ksp,T,Z);
201:        }
202:      }
203:      i++;
204:   } while (i<ksp->max_it);
205:   if (i >= ksp->max_it) {
206:     ksp->reason = KSP_DIVERGED_ITS;
207:   }
208:   return(0);
209: }

211: /*
212:     KSPCreate_CGNE - Creates the data structure for the Krylov method CGNE and sets the 
213:        function pointers for all the routines it needs to call (KSPSolve_CGNE() etc)

216: */

218: /*MC
219:      KSPCGNE - Applies the preconditioned conjugate gradient method to the normal equations
220:           without explicitly forming A^t*A

222:    Options Database Keys:
223: .   -ksp_cg_type <Hermitian or symmetric - (for complex matrices only) indicates the matrix is Hermitian or symmetric


226:    Level: beginner

228:    Notes: eigenvalue computation routines will return information about the
229:           spectrum of A^t*A, rather than A.

231:    This is NOT a different algorithm then used with KSPCG, it merely uses that algorithm with the 
232:    matrix defined by A^t*A and preconditioner defined by B^t*B where B is the preconditioner for A.

234:    This method requires that one be apply to apply the transpose of the preconditioner and operator
235:    as well as the operator and preconditioner. If the transpose of the preconditioner is not available then
236:    the preconditioner is used in its place so one ends up preconditioning A'A with B B. Seems odd?

238:    This object is subclassed off of KSPCG

240: .seealso:  KSPCreate(), KSPSetType(), KSPType (for list of available types), KSP,
241:            KSPCGSetType(), KSPBICG

243: M*/


255: PetscErrorCode  KSPCreate_CGNE(KSP ksp)
256: {
258:   KSP_CG         *cg;

261:   PetscNewLog(ksp,KSP_CG,&cg);
262: #if !defined(PETSC_USE_COMPLEX)
263:   cg->type                       = KSP_CG_SYMMETRIC;
264: #else
265:   cg->type                       = KSP_CG_HERMITIAN;
266: #endif
267:   ksp->data                      = (void*)cg;
268:   ksp->pc_side                   = PC_LEFT;

270:   /*
271:        Sets the functions that are associated with this data structure 
272:        (in C++ this is the same as defining virtual functions)
273:   */
274:   ksp->ops->setup                = KSPSetUp_CGNE;
275:   ksp->ops->solve                = KSPSolve_CGNE;
276:   ksp->ops->destroy              = KSPDestroy_CG;
277:   ksp->ops->view                 = KSPView_CG;
278:   ksp->ops->setfromoptions       = KSPSetFromOptions_CG;
279:   ksp->ops->buildsolution        = KSPDefaultBuildSolution;
280:   ksp->ops->buildresidual        = KSPDefaultBuildResidual;

282:   /*
283:       Attach the function KSPCGSetType_CGNE() to this object. The routine 
284:       KSPCGSetType() checks for this attached function and calls it if it finds
285:       it. (Sort of like a dynamic member function that can be added at run time
286:   */
287:   PetscObjectComposeFunctionDynamic((PetscObject)ksp,"KSPCGSetType_C","KSPCGSetType_CG",KSPCGSetType_CG);
288:   return(0);
289: }