Actual source code: sbaijspooles.c

  1: /* 
  2:    Provides an interface to the Spooles serial sparse solver
  3: */

 5:  #include src/mat/impls/aij/seq/spooles/spooles.h

  9: PetscErrorCode MatDestroy_SeqSBAIJSpooles(Mat A)
 10: {
 12: 
 14:   /* SeqSBAIJ_Spooles isn't really the matrix that USES spooles, */
 15:   /* rather it is a factory class for creating a symmetric matrix that can */
 16:   /* invoke Spooles' sequential cholesky solver. */
 17:   /* As a result, we don't have to clean up the stuff set by spooles */
 18:   /* as in MatDestroy_SeqAIJ_Spooles. */
 19:   MatConvert_Spooles_Base(A,MATSEQSBAIJ,&A);
 20:   (*A->ops->destroy)(A);
 21:   return(0);
 22: }

 26: PetscErrorCode MatAssemblyEnd_SeqSBAIJSpooles(Mat A,MatAssemblyType mode) \
 27: {
 29:   PetscInt       bs;
 30:   Mat_Spooles    *lu=(Mat_Spooles *)(A->spptr);

 33:   (*lu->MatAssemblyEnd)(A,mode);
 34:   MatGetBlockSize(A,&bs);
 35:   if (bs > 1) SETERRQ1(PETSC_ERR_SUP,"Block size %D not supported by Spooles",bs);
 36:   lu->MatCholeskyFactorSymbolic  = A->ops->choleskyfactorsymbolic;
 37:   A->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqSBAIJSpooles;
 38:   return(0);
 39: }

 41: /* 
 42:   input:
 43:    F:                 numeric factor
 44:   output:
 45:    nneg, nzero, npos: matrix inertia 
 46: */

 50: PetscErrorCode MatGetInertia_SeqSBAIJSpooles(Mat F,int *nneg,int *nzero,int *npos)
 51: {
 52:   Mat_Spooles *lu = (Mat_Spooles*)F->spptr;
 53:   int         neg,zero,pos;

 56:   FrontMtx_inertia(lu->frontmtx, &neg, &zero, &pos);
 57:   if(nneg)  *nneg  = neg;
 58:   if(nzero) *nzero = zero;
 59:   if(npos)  *npos  = pos;
 60:   return(0);
 61: }

 63: /* Note the Petsc r permutation is ignored */
 66: PetscErrorCode MatCholeskyFactorSymbolic_SeqSBAIJSpooles(Mat A,IS r,MatFactorInfo *info,Mat *F)
 67: {
 68:   Mat         B;
 69:   Mat_Spooles *lu;
 71:   int m=A->m,n=A->n;

 74:   /* Create the factorization matrix */
 75:   MatCreate(A->comm,m,n,m,n,&B);
 76:   MatSetType(B,A->type_name);
 77:   MatSeqSBAIJSetPreallocation(B,1,PETSC_NULL,PETSC_NULL);

 79:   B->ops->choleskyfactornumeric  = MatFactorNumeric_SeqAIJSpooles;
 80:   B->ops->getinertia             = MatGetInertia_SeqSBAIJSpooles;
 81:   B->factor                      = FACTOR_CHOLESKY;

 83:   lu                        = (Mat_Spooles *)(B->spptr);
 84:   lu->options.pivotingflag  = SPOOLES_NO_PIVOTING;
 85:   lu->options.symflag       = SPOOLES_SYMMETRIC;   /* default */
 86:   lu->flg                   = DIFFERENT_NONZERO_PATTERN;
 87:   lu->options.useQR         = PETSC_FALSE;

 89:   *F = B;
 90:   return(0);
 91: }

 96: PetscErrorCode MatConvert_SeqSBAIJ_SeqSBAIJSpooles(Mat A,const MatType type,Mat *newmat) {
 97:   /* This routine is only called to convert a MATSEQSBAIJ matrix */
 98:   /* to a MATSEQSBAIJSPOOLES matrix, so we will ignore 'MatType type'. */
100:   Mat         B=*newmat;
101:   Mat_Spooles *lu;

104:   if (B != A) {
105:     /* This routine is inherited, so we know the type is correct. */
106:     MatDuplicate(A,MAT_COPY_VALUES,&B);
107:   }

109:   PetscNew(Mat_Spooles,&lu);
110:   B->spptr                       = (void*)lu;

112:   lu->basetype                   = MATSEQSBAIJ;
113:   lu->CleanUpSpooles             = PETSC_FALSE;
114:   lu->MatDuplicate               = A->ops->duplicate;
115:   lu->MatCholeskyFactorSymbolic  = A->ops->choleskyfactorsymbolic;
116:   lu->MatLUFactorSymbolic        = A->ops->lufactorsymbolic;
117:   lu->MatView                    = A->ops->view;
118:   lu->MatAssemblyEnd             = A->ops->assemblyend;
119:   lu->MatDestroy                 = A->ops->destroy;

121:   B->ops->duplicate              = MatDuplicate_Spooles;
122:   B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqSBAIJSpooles;
123:   B->ops->assemblyend            = MatAssemblyEnd_SeqSBAIJSpooles;
124:   B->ops->destroy                = MatDestroy_SeqSBAIJSpooles;
125:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_seqsbaijspooles_seqsbaij_C",
126:                                            "MatConvert_Spooles_Base",MatConvert_Spooles_Base);
127:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_seqsbaij_seqsbaijspooles_C",
128:                                            "MatConvert_SeqSBAIJ_SeqSBAIJSpooles",MatConvert_SeqSBAIJ_SeqSBAIJSpooles);
129:   PetscObjectChangeTypeName((PetscObject)B,MATSEQSBAIJSPOOLES);
130:   *newmat = B;
131:   return(0);
132: }

135: /*MC
136:   MATSEQSBAIJSPOOLES - MATSEQSBAIJSPOOLES = "seqsbaijspooles" - A matrix type providing direct solvers (Cholesky) for sequential symmetric
137:   matrices via the external package Spooles.

139:   If Spooles is installed (see the manual for
140:   instructions on how to declare the existence of external packages),
141:   a matrix type can be constructed which invokes Spooles solvers.
142:   After calling MatCreate(...,A), simply call MatSetType(A,MATSEQSBAIJSPOOLES).
143:   This matrix type is only supported for double precision real.

145:   This matrix inherits from MATSEQSBAIJ.  As a result, MatSeqSBAIJSetPreallocation is 
146:   supported for this matrix type.  One can also call MatConvert for an inplace conversion to or from 
147:   the MATSEQSBAIJ type without data copy.

149:   Options Database Keys:
150: + -mat_type seqsbaijspooles - sets the matrix type to seqsbaijspooles during calls to MatSetFromOptions()
151: . -mat_spooles_tau <tau> - upper bound on the magnitude of the largest element in L or U
152: . -mat_spooles_seed <seed> - random number seed used for ordering
153: . -mat_spooles_msglvl <msglvl> - message output level
154: . -mat_spooles_ordering <BestOfNDandMS,MMD,MS,ND> - ordering used
155: . -mat_spooles_maxdomainsize <n> - maximum subgraph size used by Spooles orderings
156: . -mat_spooles_maxzeros <n> - maximum number of zeros inside a supernode
157: . -mat_spooles_maxsize <n> - maximum size of a supernode
158: . -mat_spooles_FrontMtxInfo <true,fase> - print Spooles information about the computed factorization
159: . -mat_spooles_symmetryflag <0,1,2> - 0: SPOOLES_SYMMETRIC, 1: SPOOLES_HERMITIAN, 2: SPOOLES_NONSYMMETRIC
160: . -mat_spooles_patchAndGoFlag <0,1,2> - 0: no patch, 1: use PatchAndGo strategy 1, 2: use PatchAndGo strategy 2
161: . -mat_spooles_toosmall <dt> - drop tolerance for PatchAndGo strategy 1
162: . -mat_spooles_storeids <bool integer> - if nonzero, stores row and col numbers where patches were applied in an IV object
163: . -mat_spooles_fudge <delta> - fudge factor for rescaling diagonals with PatchAndGo strategy 2
164: - -mat_spooles_storevalues <bool integer> - if nonzero and PatchAndGo strategy 2 is used, store change in diagonal value in a DV object

166:    Level: beginner

168: .seealso: MATMPISBAIJSPOOLES, MATSEQAIJSPOOLES, MATMPIAIJSPOOLES, PCCHOLESKY
169: M*/

174: PetscErrorCode MatCreate_SeqSBAIJSpooles(Mat A)
175: {

179:   /* Change type name before calling MatSetType to force proper construction of SeqSBAIJ */
180:   /*   and SeqSBAIJSpooles types */
181:   PetscObjectChangeTypeName((PetscObject)A,MATSEQSBAIJSPOOLES);
182:   MatSetType(A,MATSEQSBAIJ);
183:   MatConvert_SeqSBAIJ_SeqSBAIJSpooles(A,MATSEQSBAIJSPOOLES,&A);
184:   return(0);
185: }

188: /*MC
189:   MATSBAIJSPOOLES - MATSBAIJSPOOLES = "sbaijspooles" - A matrix type providing direct solvers (Cholesky) for sequential and parallel symmetric matrices via the external package Spooles.

191:   If Spooles is installed (see the manual for
192:   instructions on how to declare the existence of external packages),
193:   a matrix type can be constructed which invokes Spooles solvers.
194:   After calling MatCreate(...,A), simply call MatSetType(A,MATSBAIJSPOOLES).
195:   This matrix type is only supported for double precision real.

197:   This matrix inherits from MATSBAIJ.  As a result, MatSeqSBAIJSetPreallocation and MatMPISBAIJSetPreallocation are 
198:   supported for this matrix type.  One can also call MatConvert for an inplace conversion to or from 
199:   the MATSBAIJ type without data copy.

201:   Options Database Keys:
202: + -mat_type sbaijspooles - sets the matrix type to sbaijspooles during calls to MatSetFromOptions()
203: . -mat_spooles_tau <tau> - upper bound on the magnitude of the largest element in L or U
204: . -mat_spooles_seed <seed> - random number seed used for ordering
205: . -mat_spooles_msglvl <msglvl> - message output level
206: . -mat_spooles_ordering <BestOfNDandMS,MMD,MS,ND> - ordering used
207: . -mat_spooles_maxdomainsize <n> - maximum subgraph size used by Spooles orderings
208: . -mat_spooles_maxzeros <n> - maximum number of zeros inside a supernode
209: . -mat_spooles_maxsize <n> - maximum size of a supernode
210: . -mat_spooles_FrontMtxInfo <true,fase> - print Spooles information about the computed factorization
211: . -mat_spooles_symmetryflag <0,1,2> - 0: SPOOLES_SYMMETRIC, 1: SPOOLES_HERMITIAN, 2: SPOOLES_NONSYMMETRIC
212: . -mat_spooles_patchAndGoFlag <0,1,2> - 0: no patch, 1: use PatchAndGo strategy 1, 2: use PatchAndGo strategy 2
213: . -mat_spooles_toosmall <dt> - drop tolerance for PatchAndGo strategy 1
214: . -mat_spooles_storeids <bool integer> - if nonzero, stores row and col numbers where patches were applied in an IV object
215: . -mat_spooles_fudge <delta> - fudge factor for rescaling diagonals with PatchAndGo strategy 2
216: - -mat_spooles_storevalues <bool integer> - if nonzero and PatchAndGo strategy 2 is used, store change in diagonal value in a DV object

218:    Level: beginner

220: .seealso: MATMPISBAIJSPOOLES, MATSEQAIJSPOOLES, MATMPIAIJSPOOLES, PCCHOLESKY
221: M*/

226: PetscErrorCode MatCreate_SBAIJSpooles(Mat A)
227: {
229:   int size;

232:   /* Change type name before calling MatSetType to force proper construction of SeqSBAIJSpooles or MPISBAIJSpooles */
233:   PetscObjectChangeTypeName((PetscObject)A,MATSBAIJSPOOLES);
234:   MPI_Comm_size(A->comm,&size);
235:   if (size == 1) {
236:     MatSetType(A,MATSEQSBAIJ);
237:     MatConvert_SeqSBAIJ_SeqSBAIJSpooles(A,MATSEQSBAIJSPOOLES,&A);
238:   } else {
239:     MatSetType(A,MATMPISBAIJ);
240:     MatConvert_MPISBAIJ_MPISBAIJSpooles(A,MATMPISBAIJSPOOLES,&A);
241:   }
242: 
243:   return(0);
244: }