Actual source code: mumps.c

petsc-master 2017-11-21
Report Typos and Errors

  2: /*
  3:     Provides an interface to the MUMPS sparse solver
  4: */

  6:  #include <../src/mat/impls/aij/mpi/mpiaij.h>
  7:  #include <../src/mat/impls/sbaij/mpi/mpisbaij.h>

  9: EXTERN_C_BEGIN
 10: #if defined(PETSC_USE_COMPLEX)
 11: #if defined(PETSC_USE_REAL_SINGLE)
 12: #include <cmumps_c.h>
 13: #else
 14: #include <zmumps_c.h>
 15: #endif
 16: #else
 17: #if defined(PETSC_USE_REAL_SINGLE)
 18: #include <smumps_c.h>
 19: #else
 20: #include <dmumps_c.h>
 21: #endif
 22: #endif
 23: EXTERN_C_END
 24: #define JOB_INIT -1
 25: #define JOB_FACTSYMBOLIC 1
 26: #define JOB_FACTNUMERIC 2
 27: #define JOB_SOLVE 3
 28: #define JOB_END -2

 30: /* calls to MUMPS */
 31: #if defined(PETSC_USE_COMPLEX)
 32: #if defined(PETSC_USE_REAL_SINGLE)
 33: #define PetscMUMPS_c cmumps_c
 34: #else
 35: #define PetscMUMPS_c zmumps_c
 36: #endif
 37: #else
 38: #if defined(PETSC_USE_REAL_SINGLE)
 39: #define PetscMUMPS_c smumps_c
 40: #else
 41: #define PetscMUMPS_c dmumps_c
 42: #endif
 43: #endif

 45: /* declare MumpsScalar */
 46: #if defined(PETSC_USE_COMPLEX)
 47: #if defined(PETSC_USE_REAL_SINGLE)
 48: #define MumpsScalar mumps_complex
 49: #else
 50: #define MumpsScalar mumps_double_complex
 51: #endif
 52: #else
 53: #define MumpsScalar PetscScalar
 54: #endif

 56: /* macros s.t. indices match MUMPS documentation */
 57: #define ICNTL(I) icntl[(I)-1]
 58: #define CNTL(I) cntl[(I)-1]
 59: #define INFOG(I) infog[(I)-1]
 60: #define INFO(I) info[(I)-1]
 61: #define RINFOG(I) rinfog[(I)-1]
 62: #define RINFO(I) rinfo[(I)-1]

 64: typedef struct {
 65: #if defined(PETSC_USE_COMPLEX)
 66: #if defined(PETSC_USE_REAL_SINGLE)
 67:   CMUMPS_STRUC_C id;
 68: #else
 69:   ZMUMPS_STRUC_C id;
 70: #endif
 71: #else
 72: #if defined(PETSC_USE_REAL_SINGLE)
 73:   SMUMPS_STRUC_C id;
 74: #else
 75:   DMUMPS_STRUC_C id;
 76: #endif
 77: #endif

 79:   MatStructure matstruc;
 80:   PetscMPIInt  myid,size;
 81:   PetscInt     *irn,*jcn,nz,sym;
 82:   PetscScalar  *val;
 83:   MPI_Comm     comm_mumps;
 84:   PetscBool    isAIJ;
 85:   PetscInt     ICNTL9_pre;           /* check if ICNTL(9) is changed from previous MatSolve */
 86:   VecScatter   scat_rhs, scat_sol;   /* used by MatSolve() */
 87:   Vec          b_seq,x_seq;
 88:   PetscInt     ninfo,*info;          /* display INFO */
 89:   PetscInt     sizeredrhs;
 90:   PetscScalar  *schur_sol;
 91:   PetscInt     schur_sizesol;

 93:   PetscErrorCode (*ConvertToTriples)(Mat, int, MatReuse, int*, int**, int**, PetscScalar**);
 94: } Mat_MUMPS;

 96: extern PetscErrorCode MatDuplicate_MUMPS(Mat,MatDuplicateOption,Mat*);

 98: static PetscErrorCode MatMumpsResetSchur_Private(Mat_MUMPS* mumps)
 99: {

103:   PetscFree2(mumps->id.listvar_schur,mumps->id.schur);
104:   PetscFree(mumps->id.redrhs);
105:   PetscFree(mumps->schur_sol);
106:   mumps->id.size_schur = 0;
107:   mumps->id.schur_lld  = 0;
108:   mumps->id.ICNTL(19)  = 0;
109:   return(0);
110: }

112: /* solve with rhs in mumps->id.redrhs and return in the same location */
113: static PetscErrorCode MatMumpsSolveSchur_Private(Mat F)
114: {
115:   Mat_MUMPS            *mumps=(Mat_MUMPS*)F->data;
116:   Mat                  S,B,X;
117:   MatFactorSchurStatus schurstatus;
118:   PetscInt             sizesol;
119:   PetscErrorCode       ierr;

122:   MatFactorFactorizeSchurComplement(F);
123:   MatFactorGetSchurComplement(F,&S,&schurstatus);
124:   MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&B);
125:   switch (schurstatus) {
126:   case MAT_FACTOR_SCHUR_FACTORED:
127:     MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&X);
128:     if (!mumps->id.ICNTL(9)) { /* transpose solve */
129:       MatMatSolveTranspose(S,B,X);
130:     } else {
131:       MatMatSolve(S,B,X);
132:     }
133:     break;
134:   case MAT_FACTOR_SCHUR_INVERTED:
135:     sizesol = mumps->id.nrhs*mumps->id.size_schur;
136:     if (!mumps->schur_sol || sizesol > mumps->schur_sizesol) {
137:       PetscFree(mumps->schur_sol);
138:       PetscMalloc1(sizesol,&mumps->schur_sol);
139:       mumps->schur_sizesol = sizesol;
140:     }
141:     MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,mumps->schur_sol,&X);
142:     if (!mumps->id.ICNTL(9)) { /* transpose solve */
143:       MatTransposeMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);
144:     } else {
145:       MatMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);
146:     }
147:     MatCopy(X,B,SAME_NONZERO_PATTERN);
148:     break;
149:   default:
150:     SETERRQ1(PetscObjectComm((PetscObject)F),PETSC_ERR_SUP,"Unhandled MatFactorSchurStatus %D",F->schur_status);
151:     break;
152:   }
153:   MatFactorRestoreSchurComplement(F,&S,schurstatus);
154:   MatDestroy(&B);
155:   MatDestroy(&X);
156:   return(0);
157: }

159: static PetscErrorCode MatMumpsHandleSchur_Private(Mat F, PetscBool expansion)
160: {
161:   Mat_MUMPS     *mumps=(Mat_MUMPS*)F->data;

165:   if (!mumps->id.ICNTL(19)) { /* do nothing when Schur complement has not been computed */
166:     return(0);
167:   }
168:   if (!expansion) { /* prepare for the condensation step */
169:     PetscInt sizeredrhs = mumps->id.nrhs*mumps->id.size_schur;
170:     /* allocate MUMPS internal array to store reduced right-hand sides */
171:     if (!mumps->id.redrhs || sizeredrhs > mumps->sizeredrhs) {
172:       PetscFree(mumps->id.redrhs);
173:       mumps->id.lredrhs = mumps->id.size_schur;
174:       PetscMalloc1(mumps->id.nrhs*mumps->id.lredrhs,&mumps->id.redrhs);
175:       mumps->sizeredrhs = mumps->id.nrhs*mumps->id.lredrhs;
176:     }
177:     mumps->id.ICNTL(26) = 1; /* condensation phase */
178:   } else { /* prepare for the expansion step */
179:     /* solve Schur complement (this has to be done by the MUMPS user, so basically us) */
180:     MatMumpsSolveSchur_Private(F);
181:     mumps->id.ICNTL(26) = 2; /* expansion phase */
182:     PetscMUMPS_c(&mumps->id);
183:     if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1));
184:     /* restore defaults */
185:     mumps->id.ICNTL(26) = -1;
186:     /* free MUMPS internal array for redrhs if we have solved for multiple rhs in order to save memory space */
187:     if (mumps->id.nrhs > 1) {
188:       PetscFree(mumps->id.redrhs);
189:       mumps->id.lredrhs = 0;
190:       mumps->sizeredrhs = 0;
191:     }
192:   }
193:   return(0);
194: }

196: /*
197:   MatConvertToTriples_A_B - convert Petsc matrix to triples: row[nz], col[nz], val[nz] 

199:   input:
200:     A       - matrix in aij,baij or sbaij (bs=1) format
201:     shift   - 0: C style output triple; 1: Fortran style output triple.
202:     reuse   - MAT_INITIAL_MATRIX: spaces are allocated and values are set for the triple
203:               MAT_REUSE_MATRIX:   only the values in v array are updated
204:   output:
205:     nnz     - dim of r, c, and v (number of local nonzero entries of A)
206:     r, c, v - row and col index, matrix values (matrix triples)

208:   The returned values r, c, and sometimes v are obtained in a single PetscMalloc(). Then in MatDestroy_MUMPS() it is
209:   freed with PetscFree((mumps->irn);  This is not ideal code, the fact that v is ONLY sometimes part of mumps->irn means
210:   that the PetscMalloc() cannot easily be replaced with a PetscMalloc3(). 

212:  */

214: PetscErrorCode MatConvertToTriples_seqaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
215: {
216:   const PetscInt *ai,*aj,*ajj,M=A->rmap->n;
217:   PetscInt       nz,rnz,i,j;
219:   PetscInt       *row,*col;
220:   Mat_SeqAIJ     *aa=(Mat_SeqAIJ*)A->data;

223:   *v=aa->a;
224:   if (reuse == MAT_INITIAL_MATRIX) {
225:     nz   = aa->nz;
226:     ai   = aa->i;
227:     aj   = aa->j;
228:     *nnz = nz;
229:     PetscMalloc1(2*nz, &row);
230:     col  = row + nz;

232:     nz = 0;
233:     for (i=0; i<M; i++) {
234:       rnz = ai[i+1] - ai[i];
235:       ajj = aj + ai[i];
236:       for (j=0; j<rnz; j++) {
237:         row[nz] = i+shift; col[nz++] = ajj[j] + shift;
238:       }
239:     }
240:     *r = row; *c = col;
241:   }
242:   return(0);
243: }

245: PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
246: {
247:   Mat_SeqBAIJ    *aa=(Mat_SeqBAIJ*)A->data;
248:   const PetscInt *ai,*aj,*ajj,bs2 = aa->bs2;
249:   PetscInt       bs,M,nz,idx=0,rnz,i,j,k,m;
251:   PetscInt       *row,*col;

254:   MatGetBlockSize(A,&bs);
255:   M = A->rmap->N/bs;
256:   *v = aa->a;
257:   if (reuse == MAT_INITIAL_MATRIX) {
258:     ai   = aa->i; aj = aa->j;
259:     nz   = bs2*aa->nz;
260:     *nnz = nz;
261:     PetscMalloc1(2*nz, &row);
262:     col  = row + nz;

264:     for (i=0; i<M; i++) {
265:       ajj = aj + ai[i];
266:       rnz = ai[i+1] - ai[i];
267:       for (k=0; k<rnz; k++) {
268:         for (j=0; j<bs; j++) {
269:           for (m=0; m<bs; m++) {
270:             row[idx]   = i*bs + m + shift;
271:             col[idx++] = bs*(ajj[k]) + j + shift;
272:           }
273:         }
274:       }
275:     }
276:     *r = row; *c = col;
277:   }
278:   return(0);
279: }

281: PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
282: {
283:   const PetscInt *ai, *aj,*ajj,M=A->rmap->n;
284:   PetscInt       nz,rnz,i,j;
286:   PetscInt       *row,*col;
287:   Mat_SeqSBAIJ   *aa=(Mat_SeqSBAIJ*)A->data;

290:   *v = aa->a;
291:   if (reuse == MAT_INITIAL_MATRIX) {
292:     nz   = aa->nz;
293:     ai   = aa->i;
294:     aj   = aa->j;
295:     *v   = aa->a;
296:     *nnz = nz;
297:     PetscMalloc1(2*nz, &row);
298:     col  = row + nz;

300:     nz = 0;
301:     for (i=0; i<M; i++) {
302:       rnz = ai[i+1] - ai[i];
303:       ajj = aj + ai[i];
304:       for (j=0; j<rnz; j++) {
305:         row[nz] = i+shift; col[nz++] = ajj[j] + shift;
306:       }
307:     }
308:     *r = row; *c = col;
309:   }
310:   return(0);
311: }

313: PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
314: {
315:   const PetscInt    *ai,*aj,*ajj,*adiag,M=A->rmap->n;
316:   PetscInt          nz,rnz,i,j;
317:   const PetscScalar *av,*v1;
318:   PetscScalar       *val;
319:   PetscErrorCode    ierr;
320:   PetscInt          *row,*col;
321:   Mat_SeqAIJ        *aa=(Mat_SeqAIJ*)A->data;
322:   PetscBool         missing;

325:   ai   =aa->i; aj=aa->j;av=aa->a;
326:   adiag=aa->diag;
327:   MatMissingDiagonal_SeqAIJ(A,&missing,&i);
328:   if (reuse == MAT_INITIAL_MATRIX) {
329:     /* count nz in the uppper triangular part of A */
330:     nz = 0;
331:     if (missing) {
332:       for (i=0; i<M; i++) {
333:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
334:           for (j=ai[i];j<ai[i+1];j++) {
335:             if (aj[j] < i) continue;
336:             nz++;
337:           }
338:         } else {
339:           nz += ai[i+1] - adiag[i];
340:         }
341:       }
342:     } else {
343:       for (i=0; i<M; i++) nz += ai[i+1] - adiag[i];
344:     }
345:     *nnz = nz;

347:     PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
348:     col  = row + nz;
349:     val  = (PetscScalar*)(col + nz);

351:     nz = 0;
352:     if (missing) {
353:       for (i=0; i<M; i++) {
354:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
355:           for (j=ai[i];j<ai[i+1];j++) {
356:             if (aj[j] < i) continue;
357:             row[nz] = i+shift;
358:             col[nz] = aj[j]+shift;
359:             val[nz] = av[j];
360:             nz++;
361:           }
362:         } else {
363:           rnz = ai[i+1] - adiag[i];
364:           ajj = aj + adiag[i];
365:           v1  = av + adiag[i];
366:           for (j=0; j<rnz; j++) {
367:             row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j];
368:           }
369:         }
370:       }
371:     } else {
372:       for (i=0; i<M; i++) {
373:         rnz = ai[i+1] - adiag[i];
374:         ajj = aj + adiag[i];
375:         v1  = av + adiag[i];
376:         for (j=0; j<rnz; j++) {
377:           row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j];
378:         }
379:       }
380:     }
381:     *r = row; *c = col; *v = val;
382:   } else {
383:     nz = 0; val = *v;
384:     if (missing) {
385:       for (i=0; i <M; i++) {
386:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
387:           for (j=ai[i];j<ai[i+1];j++) {
388:             if (aj[j] < i) continue;
389:             val[nz++] = av[j];
390:           }
391:         } else {
392:           rnz = ai[i+1] - adiag[i];
393:           v1  = av + adiag[i];
394:           for (j=0; j<rnz; j++) {
395:             val[nz++] = v1[j];
396:           }
397:         }
398:       }
399:     } else {
400:       for (i=0; i <M; i++) {
401:         rnz = ai[i+1] - adiag[i];
402:         v1  = av + adiag[i];
403:         for (j=0; j<rnz; j++) {
404:           val[nz++] = v1[j];
405:         }
406:       }
407:     }
408:   }
409:   return(0);
410: }

412: PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
413: {
414:   const PetscInt    *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
415:   PetscErrorCode    ierr;
416:   PetscInt          rstart,nz,i,j,jj,irow,countA,countB;
417:   PetscInt          *row,*col;
418:   const PetscScalar *av, *bv,*v1,*v2;
419:   PetscScalar       *val;
420:   Mat_MPISBAIJ      *mat = (Mat_MPISBAIJ*)A->data;
421:   Mat_SeqSBAIJ      *aa  = (Mat_SeqSBAIJ*)(mat->A)->data;
422:   Mat_SeqBAIJ       *bb  = (Mat_SeqBAIJ*)(mat->B)->data;

425:   ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart;
426:   av=aa->a; bv=bb->a;

428:   garray = mat->garray;

430:   if (reuse == MAT_INITIAL_MATRIX) {
431:     nz   = aa->nz + bb->nz;
432:     *nnz = nz;
433:     PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
434:     col  = row + nz;
435:     val  = (PetscScalar*)(col + nz);

437:     *r = row; *c = col; *v = val;
438:   } else {
439:     row = *r; col = *c; val = *v;
440:   }

442:   jj = 0; irow = rstart;
443:   for (i=0; i<m; i++) {
444:     ajj    = aj + ai[i];                 /* ptr to the beginning of this row */
445:     countA = ai[i+1] - ai[i];
446:     countB = bi[i+1] - bi[i];
447:     bjj    = bj + bi[i];
448:     v1     = av + ai[i];
449:     v2     = bv + bi[i];

451:     /* A-part */
452:     for (j=0; j<countA; j++) {
453:       if (reuse == MAT_INITIAL_MATRIX) {
454:         row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
455:       }
456:       val[jj++] = v1[j];
457:     }

459:     /* B-part */
460:     for (j=0; j < countB; j++) {
461:       if (reuse == MAT_INITIAL_MATRIX) {
462:         row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
463:       }
464:       val[jj++] = v2[j];
465:     }
466:     irow++;
467:   }
468:   return(0);
469: }

471: PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
472: {
473:   const PetscInt    *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
474:   PetscErrorCode    ierr;
475:   PetscInt          rstart,nz,i,j,jj,irow,countA,countB;
476:   PetscInt          *row,*col;
477:   const PetscScalar *av, *bv,*v1,*v2;
478:   PetscScalar       *val;
479:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ*)A->data;
480:   Mat_SeqAIJ        *aa  = (Mat_SeqAIJ*)(mat->A)->data;
481:   Mat_SeqAIJ        *bb  = (Mat_SeqAIJ*)(mat->B)->data;

484:   ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart;
485:   av=aa->a; bv=bb->a;

487:   garray = mat->garray;

489:   if (reuse == MAT_INITIAL_MATRIX) {
490:     nz   = aa->nz + bb->nz;
491:     *nnz = nz;
492:     PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
493:     col  = row + nz;
494:     val  = (PetscScalar*)(col + nz);

496:     *r = row; *c = col; *v = val;
497:   } else {
498:     row = *r; col = *c; val = *v;
499:   }

501:   jj = 0; irow = rstart;
502:   for (i=0; i<m; i++) {
503:     ajj    = aj + ai[i];                 /* ptr to the beginning of this row */
504:     countA = ai[i+1] - ai[i];
505:     countB = bi[i+1] - bi[i];
506:     bjj    = bj + bi[i];
507:     v1     = av + ai[i];
508:     v2     = bv + bi[i];

510:     /* A-part */
511:     for (j=0; j<countA; j++) {
512:       if (reuse == MAT_INITIAL_MATRIX) {
513:         row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
514:       }
515:       val[jj++] = v1[j];
516:     }

518:     /* B-part */
519:     for (j=0; j < countB; j++) {
520:       if (reuse == MAT_INITIAL_MATRIX) {
521:         row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
522:       }
523:       val[jj++] = v2[j];
524:     }
525:     irow++;
526:   }
527:   return(0);
528: }

530: PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
531: {
532:   Mat_MPIBAIJ       *mat    = (Mat_MPIBAIJ*)A->data;
533:   Mat_SeqBAIJ       *aa     = (Mat_SeqBAIJ*)(mat->A)->data;
534:   Mat_SeqBAIJ       *bb     = (Mat_SeqBAIJ*)(mat->B)->data;
535:   const PetscInt    *ai     = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j,*ajj, *bjj;
536:   const PetscInt    *garray = mat->garray,mbs=mat->mbs,rstart=A->rmap->rstart;
537:   const PetscInt    bs2=mat->bs2;
538:   PetscErrorCode    ierr;
539:   PetscInt          bs,nz,i,j,k,n,jj,irow,countA,countB,idx;
540:   PetscInt          *row,*col;
541:   const PetscScalar *av=aa->a, *bv=bb->a,*v1,*v2;
542:   PetscScalar       *val;

545:   MatGetBlockSize(A,&bs);
546:   if (reuse == MAT_INITIAL_MATRIX) {
547:     nz   = bs2*(aa->nz + bb->nz);
548:     *nnz = nz;
549:     PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
550:     col  = row + nz;
551:     val  = (PetscScalar*)(col + nz);

553:     *r = row; *c = col; *v = val;
554:   } else {
555:     row = *r; col = *c; val = *v;
556:   }

558:   jj = 0; irow = rstart;
559:   for (i=0; i<mbs; i++) {
560:     countA = ai[i+1] - ai[i];
561:     countB = bi[i+1] - bi[i];
562:     ajj    = aj + ai[i];
563:     bjj    = bj + bi[i];
564:     v1     = av + bs2*ai[i];
565:     v2     = bv + bs2*bi[i];

567:     idx = 0;
568:     /* A-part */
569:     for (k=0; k<countA; k++) {
570:       for (j=0; j<bs; j++) {
571:         for (n=0; n<bs; n++) {
572:           if (reuse == MAT_INITIAL_MATRIX) {
573:             row[jj] = irow + n + shift;
574:             col[jj] = rstart + bs*ajj[k] + j + shift;
575:           }
576:           val[jj++] = v1[idx++];
577:         }
578:       }
579:     }

581:     idx = 0;
582:     /* B-part */
583:     for (k=0; k<countB; k++) {
584:       for (j=0; j<bs; j++) {
585:         for (n=0; n<bs; n++) {
586:           if (reuse == MAT_INITIAL_MATRIX) {
587:             row[jj] = irow + n + shift;
588:             col[jj] = bs*garray[bjj[k]] + j + shift;
589:           }
590:           val[jj++] = v2[idx++];
591:         }
592:       }
593:     }
594:     irow += bs;
595:   }
596:   return(0);
597: }

599: PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
600: {
601:   const PetscInt    *ai, *aj,*adiag, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
602:   PetscErrorCode    ierr;
603:   PetscInt          rstart,nz,nza,nzb,i,j,jj,irow,countA,countB;
604:   PetscInt          *row,*col;
605:   const PetscScalar *av, *bv,*v1,*v2;
606:   PetscScalar       *val;
607:   Mat_MPIAIJ        *mat =  (Mat_MPIAIJ*)A->data;
608:   Mat_SeqAIJ        *aa  =(Mat_SeqAIJ*)(mat->A)->data;
609:   Mat_SeqAIJ        *bb  =(Mat_SeqAIJ*)(mat->B)->data;

612:   ai=aa->i; aj=aa->j; adiag=aa->diag;
613:   bi=bb->i; bj=bb->j; garray = mat->garray;
614:   av=aa->a; bv=bb->a;

616:   rstart = A->rmap->rstart;

618:   if (reuse == MAT_INITIAL_MATRIX) {
619:     nza = 0;    /* num of upper triangular entries in mat->A, including diagonals */
620:     nzb = 0;    /* num of upper triangular entries in mat->B */
621:     for (i=0; i<m; i++) {
622:       nza   += (ai[i+1] - adiag[i]);
623:       countB = bi[i+1] - bi[i];
624:       bjj    = bj + bi[i];
625:       for (j=0; j<countB; j++) {
626:         if (garray[bjj[j]] > rstart) nzb++;
627:       }
628:     }

630:     nz   = nza + nzb; /* total nz of upper triangular part of mat */
631:     *nnz = nz;
632:     PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
633:     col  = row + nz;
634:     val  = (PetscScalar*)(col + nz);

636:     *r = row; *c = col; *v = val;
637:   } else {
638:     row = *r; col = *c; val = *v;
639:   }

641:   jj = 0; irow = rstart;
642:   for (i=0; i<m; i++) {
643:     ajj    = aj + adiag[i];                 /* ptr to the beginning of the diagonal of this row */
644:     v1     = av + adiag[i];
645:     countA = ai[i+1] - adiag[i];
646:     countB = bi[i+1] - bi[i];
647:     bjj    = bj + bi[i];
648:     v2     = bv + bi[i];

650:     /* A-part */
651:     for (j=0; j<countA; j++) {
652:       if (reuse == MAT_INITIAL_MATRIX) {
653:         row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
654:       }
655:       val[jj++] = v1[j];
656:     }

658:     /* B-part */
659:     for (j=0; j < countB; j++) {
660:       if (garray[bjj[j]] > rstart) {
661:         if (reuse == MAT_INITIAL_MATRIX) {
662:           row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
663:         }
664:         val[jj++] = v2[j];
665:       }
666:     }
667:     irow++;
668:   }
669:   return(0);
670: }

672: PetscErrorCode MatDestroy_MUMPS(Mat A)
673: {
674:   Mat_MUMPS      *mumps=(Mat_MUMPS*)A->data;

678:   PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
679:   VecScatterDestroy(&mumps->scat_rhs);
680:   VecScatterDestroy(&mumps->scat_sol);
681:   VecDestroy(&mumps->b_seq);
682:   VecDestroy(&mumps->x_seq);
683:   PetscFree(mumps->id.perm_in);
684:   PetscFree(mumps->irn);
685:   PetscFree(mumps->info);
686:   MatMumpsResetSchur_Private(mumps);
687:   mumps->id.job = JOB_END;
688:   PetscMUMPS_c(&mumps->id);
689:   MPI_Comm_free(&mumps->comm_mumps);
690:   PetscFree(A->data);

692:   /* clear composed functions */
693:   PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverPackage_C",NULL);
694:   PetscObjectComposeFunction((PetscObject)A,"MatFactorSetSchurIS_C",NULL);
695:   PetscObjectComposeFunction((PetscObject)A,"MatFactorCreateSchurComplement_C",NULL);
696:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetIcntl_C",NULL);
697:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetIcntl_C",NULL);
698:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetCntl_C",NULL);
699:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetCntl_C",NULL);
700:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfo_C",NULL);
701:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfog_C",NULL);
702:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfo_C",NULL);
703:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfog_C",NULL);
704:   return(0);
705: }

707: PetscErrorCode MatSolve_MUMPS(Mat A,Vec b,Vec x)
708: {
709:   Mat_MUMPS        *mumps=(Mat_MUMPS*)A->data;
710:   PetscScalar      *array;
711:   Vec              b_seq;
712:   IS               is_iden,is_petsc;
713:   PetscErrorCode   ierr;
714:   PetscInt         i;
715:   PetscBool        second_solve = PETSC_FALSE;
716:   static PetscBool cite1 = PETSC_FALSE,cite2 = PETSC_FALSE;

719:   PetscCitationsRegister("@article{MUMPS01,\n  author = {P.~R. Amestoy and I.~S. Duff and J.-Y. L'Excellent and J. Koster},\n  title = {A fully asynchronous multifrontal solver using distributed dynamic scheduling},\n  journal = {SIAM Journal on Matrix Analysis and Applications},\n  volume = {23},\n  number = {1},\n  pages = {15--41},\n  year = {2001}\n}\n",&cite1);
720:   PetscCitationsRegister("@article{MUMPS02,\n  author = {P.~R. Amestoy and A. Guermouche and J.-Y. L'Excellent and S. Pralet},\n  title = {Hybrid scheduling for the parallel solution of linear systems},\n  journal = {Parallel Computing},\n  volume = {32},\n  number = {2},\n  pages = {136--156},\n  year = {2006}\n}\n",&cite2);

722:   if (A->factorerrortype) {
723:     PetscInfo2(A,"MatSolve is called with singular matrix factor, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
724:     VecSetInf(x);
725:     return(0);
726:   }

728:   mumps->id.nrhs = 1;
729:   b_seq          = mumps->b_seq;
730:   if (mumps->size > 1) {
731:     /* MUMPS only supports centralized rhs. Scatter b into a seqential rhs vector */
732:     VecScatterBegin(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);
733:     VecScatterEnd(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);
734:     if (!mumps->myid) {VecGetArray(b_seq,&array);}
735:   } else {  /* size == 1 */
736:     VecCopy(b,x);
737:     VecGetArray(x,&array);
738:   }
739:   if (!mumps->myid) { /* define rhs on the host */
740:     mumps->id.nrhs = 1;
741:     mumps->id.rhs = (MumpsScalar*)array;
742:   }

744:   /*
745:      handle condensation step of Schur complement (if any)
746:      We set by default ICNTL(26) == -1 when Schur indices have been provided by the user.
747:      According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase
748:      Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system.
749:      This requires an extra call to PetscMUMPS_c and the computation of the factors for S
750:   */
751:   if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) {
752:     if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n");
753:     second_solve = PETSC_TRUE;
754:     MatMumpsHandleSchur_Private(A,PETSC_FALSE);
755:   }
756:   /* solve phase */
757:   /*-------------*/
758:   mumps->id.job = JOB_SOLVE;
759:   PetscMUMPS_c(&mumps->id);
760:   if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1));

762:   /* handle expansion step of Schur complement (if any) */
763:   if (second_solve) {
764:     MatMumpsHandleSchur_Private(A,PETSC_TRUE);
765:   }

767:   if (mumps->size > 1) { /* convert mumps distributed solution to petsc mpi x */
768:     if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) {
769:       /* when id.ICNTL(9) changes, the contents of lsol_loc may change (not its size, lsol_loc), recreates scat_sol */
770:       VecScatterDestroy(&mumps->scat_sol);
771:     }
772:     if (!mumps->scat_sol) { /* create scatter scat_sol */
773:       ISCreateStride(PETSC_COMM_SELF,mumps->id.lsol_loc,0,1,&is_iden); /* from */
774:       for (i=0; i<mumps->id.lsol_loc; i++) {
775:         mumps->id.isol_loc[i] -= 1; /* change Fortran style to C style */
776:       }
777:       ISCreateGeneral(PETSC_COMM_SELF,mumps->id.lsol_loc,mumps->id.isol_loc,PETSC_COPY_VALUES,&is_petsc);  /* to */
778:       VecScatterCreate(mumps->x_seq,is_iden,x,is_petsc,&mumps->scat_sol);
779:       ISDestroy(&is_iden);
780:       ISDestroy(&is_petsc);

782:       mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */
783:     }

785:     VecScatterBegin(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
786:     VecScatterEnd(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
787:   }
788:   return(0);
789: }

791: PetscErrorCode MatSolveTranspose_MUMPS(Mat A,Vec b,Vec x)
792: {
793:   Mat_MUMPS      *mumps=(Mat_MUMPS*)A->data;

797:   mumps->id.ICNTL(9) = 0;
798:   MatSolve_MUMPS(A,b,x);
799:   mumps->id.ICNTL(9) = 1;
800:   return(0);
801: }

803: PetscErrorCode MatMatSolve_MUMPS(Mat A,Mat B,Mat X)
804: {
806:   Mat            Bt = NULL;
807:   PetscBool      flg, flgT;
808:   Mat_MUMPS      *mumps=(Mat_MUMPS*)A->data;
809:   PetscInt       i,nrhs,M;
810:   PetscScalar    *array,*bray;

813:   PetscObjectTypeCompareAny((PetscObject)B,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
814:   PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&flgT);
815:   if (flgT) {
816:     if (mumps->size > 1) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix");
817:     MatTransposeGetMat(B,&Bt);
818:   } else {
819:     if (!flg) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATDENSE matrix");
820:   }
821:   PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
822:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)X),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix");
823:   if (B->rmap->n != X->rmap->n) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B and X must have same row distribution");

825:   MatGetSize(B,&M,&nrhs);
826:   mumps->id.nrhs = nrhs;
827:   mumps->id.lrhs = M;

829:   if (mumps->size == 1) {
830:     PetscScalar *aa;
831:     PetscInt    spnr,*ia,*ja;
832:     PetscBool   second_solve = PETSC_FALSE;

834:     /* copy B to X */
835:     MatDenseGetArray(X,&array);
836:     mumps->id.rhs = (MumpsScalar*)array;
837:     if (!Bt) {
838:       MatDenseGetArray(B,&bray);
839:       PetscMemcpy(array,bray,M*nrhs*sizeof(PetscScalar));
840:       MatDenseRestoreArray(B,&bray);
841:     } else {
842:       PetscBool done;

844:       MatSeqAIJGetArray(Bt,&aa);
845:       MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);
846:       if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
847:       mumps->id.irhs_ptr    = ia;
848:       mumps->id.irhs_sparse = ja;
849:       mumps->id.nz_rhs      = ia[spnr] - 1;
850:       mumps->id.rhs_sparse  = (MumpsScalar*)aa;
851:       mumps->id.ICNTL(20)   = 1;
852:     }
853:     /* handle condensation step of Schur complement (if any) */
854:     if (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2) {
855:       second_solve = PETSC_TRUE;
856:       MatMumpsHandleSchur_Private(A,PETSC_FALSE);
857:     }
858:     /* solve phase */
859:     /*-------------*/
860:     mumps->id.job = JOB_SOLVE;
861:     PetscMUMPS_c(&mumps->id);
862:     if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1));

864:     /* handle expansion step of Schur complement (if any) */
865:     if (second_solve) {
866:       MatMumpsHandleSchur_Private(A,PETSC_TRUE);
867:     }
868:     if (Bt) {
869:       PetscBool done;

871:       MatSeqAIJRestoreArray(Bt,&aa);
872:       MatRestoreRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&done);
873:       if (!done) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure");
874:       mumps->id.ICNTL(20) = 0;
875:     }
876:     MatDenseRestoreArray(X,&array);
877:   } else {  /*--------- parallel case --------*/
878:     PetscInt       lsol_loc,nlsol_loc,*isol_loc,*idx,*iidx,*idxx,*isol_loc_save;
879:     MumpsScalar    *sol_loc,*sol_loc_save;
880:     IS             is_to,is_from;
881:     PetscInt       k,proc,j,m;
882:     const PetscInt *rstart;
883:     Vec            v_mpi,b_seq,x_seq;
884:     VecScatter     scat_rhs,scat_sol;

886:     if (mumps->size > 1 && mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n");

888:     /* create x_seq to hold local solution */
889:     isol_loc_save = mumps->id.isol_loc; /* save it for MatSovle() */
890:     sol_loc_save  = mumps->id.sol_loc;

892:     lsol_loc  = mumps->id.INFO(23);
893:     nlsol_loc = nrhs*lsol_loc;     /* length of sol_loc */
894:     PetscMalloc2(nlsol_loc,&sol_loc,nlsol_loc,&isol_loc);
895:     mumps->id.sol_loc = (MumpsScalar*)sol_loc;
896:     mumps->id.isol_loc = isol_loc;

898:     VecCreateSeqWithArray(PETSC_COMM_SELF,1,nlsol_loc,(PetscScalar*)sol_loc,&x_seq);

900:     /* copy rhs matrix B into vector v_mpi */
901:     MatGetLocalSize(B,&m,NULL);
902:     MatDenseGetArray(B,&bray);
903:     VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);
904:     MatDenseRestoreArray(B,&bray);

906:     /* scatter v_mpi to b_seq because MUMPS only supports centralized rhs */
907:     /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B;
908:       iidx: inverse of idx, will be used by scattering xx_seq -> X       */
909:     PetscMalloc2(nrhs*M,&idx,nrhs*M,&iidx);
910:     MatGetOwnershipRanges(B,&rstart);
911:     k = 0;
912:     for (proc=0; proc<mumps->size; proc++){
913:       for (j=0; j<nrhs; j++){
914:         for (i=rstart[proc]; i<rstart[proc+1]; i++){
915:           iidx[j*M + i] = k;
916:           idx[k++]      = j*M + i;
917:         }
918:       }
919:     }

921:     if (!mumps->myid) {
922:       VecCreateSeq(PETSC_COMM_SELF,nrhs*M,&b_seq);
923:       ISCreateGeneral(PETSC_COMM_SELF,nrhs*M,idx,PETSC_COPY_VALUES,&is_to);
924:       ISCreateStride(PETSC_COMM_SELF,nrhs*M,0,1,&is_from);
925:     } else {
926:       VecCreateSeq(PETSC_COMM_SELF,0,&b_seq);
927:       ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_to);
928:       ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_from);
929:     }
930:     VecScatterCreate(v_mpi,is_from,b_seq,is_to,&scat_rhs);
931:     VecScatterBegin(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);
932:     ISDestroy(&is_to);
933:     ISDestroy(&is_from);
934:     VecScatterEnd(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);

936:     if (!mumps->myid) { /* define rhs on the host */
937:       VecGetArray(b_seq,&bray);
938:       mumps->id.rhs = (MumpsScalar*)bray;
939:       VecRestoreArray(b_seq,&bray);
940:     }

942:     /* solve phase */
943:     /*-------------*/
944:     mumps->id.job = JOB_SOLVE;
945:     PetscMUMPS_c(&mumps->id);
946:     if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d\n",mumps->id.INFOG(1));

948:     /* scatter mumps distributed solution to petsc vector v_mpi, which shares local arrays with solution matrix X */
949:     MatDenseGetArray(X,&array);
950:     VecPlaceArray(v_mpi,array);
951: 
952:     /* create scatter scat_sol */
953:     PetscMalloc1(nlsol_loc,&idxx);
954:     ISCreateStride(PETSC_COMM_SELF,nlsol_loc,0,1,&is_from);
955:     for (i=0; i<lsol_loc; i++) {
956:       isol_loc[i] -= 1; /* change Fortran style to C style */
957:       idxx[i] = iidx[isol_loc[i]];
958:       for (j=1; j<nrhs; j++){
959:         idxx[j*lsol_loc+i] = iidx[isol_loc[i]+j*M];
960:       }
961:     }
962:     ISCreateGeneral(PETSC_COMM_SELF,nlsol_loc,idxx,PETSC_COPY_VALUES,&is_to);
963:     VecScatterCreate(x_seq,is_from,v_mpi,is_to,&scat_sol);
964:     VecScatterBegin(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
965:     ISDestroy(&is_from);
966:     ISDestroy(&is_to);
967:     VecScatterEnd(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
968:     MatDenseRestoreArray(X,&array);

970:     /* free spaces */
971:     mumps->id.sol_loc = sol_loc_save;
972:     mumps->id.isol_loc = isol_loc_save;

974:     PetscFree2(sol_loc,isol_loc);
975:     PetscFree2(idx,iidx);
976:     PetscFree(idxx);
977:     VecDestroy(&x_seq);
978:     VecDestroy(&v_mpi);
979:     VecDestroy(&b_seq);
980:     VecScatterDestroy(&scat_rhs);
981:     VecScatterDestroy(&scat_sol);
982:   }
983:   return(0);
984: }

986: #if !defined(PETSC_USE_COMPLEX)
987: /*
988:   input:
989:    F:        numeric factor
990:   output:
991:    nneg:     total number of negative pivots
992:    nzero:    total number of zero pivots
993:    npos:     (global dimension of F) - nneg - nzero
994: */
995: PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F,int *nneg,int *nzero,int *npos)
996: {
997:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
999:   PetscMPIInt    size;

1002:   MPI_Comm_size(PetscObjectComm((PetscObject)F),&size);
1003:   /* MUMPS 4.3.1 calls ScaLAPACK when ICNTL(13)=0 (default), which does not offer the possibility to compute the inertia of a dense matrix. Set ICNTL(13)=1 to skip ScaLAPACK */
1004:   if (size > 1 && mumps->id.ICNTL(13) != 1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"ICNTL(13)=%d. -mat_mumps_icntl_13 must be set as 1 for correct global matrix inertia\n",mumps->id.INFOG(13));

1006:   if (nneg) *nneg = mumps->id.INFOG(12);
1007:   if (nzero || npos) {
1008:     if (mumps->id.ICNTL(24) != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"-mat_mumps_icntl_24 must be set as 1 for null pivot row detection");
1009:     if (nzero) *nzero = mumps->id.INFOG(28);
1010:     if (npos) *npos   = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28));
1011:   }
1012:   return(0);
1013: }
1014: #endif

1016: PetscErrorCode MatFactorNumeric_MUMPS(Mat F,Mat A,const MatFactorInfo *info)
1017: {
1018:   Mat_MUMPS      *mumps =(Mat_MUMPS*)(F)->data;
1020:   PetscBool      isMPIAIJ;

1023:   if (mumps->id.INFOG(1) < 0) {
1024:     if (mumps->id.INFOG(1) == -6) {
1025:       PetscInfo2(A,"MatFactorNumeric is called with singular matrix structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1026:     }
1027:     PetscInfo2(A,"MatFactorNumeric is called after analysis phase fails, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1028:     return(0);
1029:   }

1031:   (*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);

1033:   /* numerical factorization phase */
1034:   /*-------------------------------*/
1035:   mumps->id.job = JOB_FACTNUMERIC;
1036:   if (!mumps->id.ICNTL(18)) { /* A is centralized */
1037:     if (!mumps->myid) {
1038:       mumps->id.a = (MumpsScalar*)mumps->val;
1039:     }
1040:   } else {
1041:     mumps->id.a_loc = (MumpsScalar*)mumps->val;
1042:   }
1043:   PetscMUMPS_c(&mumps->id);
1044:   if (mumps->id.INFOG(1) < 0) {
1045:     if (A->erroriffailure) {
1046:       SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1047:     } else {
1048:       if (mumps->id.INFOG(1) == -10) { /* numerically singular matrix */
1049:         PetscInfo2(F,"matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1050:         F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1051:       } else if (mumps->id.INFOG(1) == -13) {
1052:         PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, cannot allocate required memory %d megabytes\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1053:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1054:       } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10) ) {
1055:         PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d, problem with workarray \n",mumps->id.INFOG(1),mumps->id.INFO(2));
1056:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1057:       } else {
1058:         PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1059:         F->factorerrortype = MAT_FACTOR_OTHER;
1060:       }
1061:     }
1062:   }
1063:   if (!mumps->myid && mumps->id.ICNTL(16) > 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"  mumps->id.ICNTL(16):=%d\n",mumps->id.INFOG(16));

1065:   F->assembled    = PETSC_TRUE;
1066:   mumps->matstruc = SAME_NONZERO_PATTERN;
1067:   if (F->schur) { /* reset Schur status to unfactored */
1068:     if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
1069:       mumps->id.ICNTL(19) = 2;
1070:       MatTranspose(F->schur,MAT_INPLACE_MATRIX,&F->schur);
1071:     }
1072:     MatFactorRestoreSchurComplement(F,NULL,MAT_FACTOR_SCHUR_UNFACTORED);
1073:   }

1075:   /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */
1076:   if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3;

1078:   if (mumps->size > 1) {
1079:     PetscInt    lsol_loc;
1080:     PetscScalar *sol_loc;

1082:     PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&isMPIAIJ);

1084:     /* distributed solution; Create x_seq=sol_loc for repeated use */
1085:     if (mumps->x_seq) {
1086:       VecScatterDestroy(&mumps->scat_sol);
1087:       PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
1088:       VecDestroy(&mumps->x_seq);
1089:     }
1090:     lsol_loc = mumps->id.INFO(23); /* length of sol_loc */
1091:     PetscMalloc2(lsol_loc,&sol_loc,lsol_loc,&mumps->id.isol_loc);
1092:     mumps->id.lsol_loc = lsol_loc;
1093:     mumps->id.sol_loc = (MumpsScalar*)sol_loc;
1094:     VecCreateSeqWithArray(PETSC_COMM_SELF,1,lsol_loc,sol_loc,&mumps->x_seq);
1095:   }
1096:   return(0);
1097: }

1099: /* Sets MUMPS options from the options database */
1100: PetscErrorCode PetscSetMUMPSFromOptions(Mat F, Mat A)
1101: {
1102:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1104:   PetscInt       icntl,info[40],i,ninfo=40;
1105:   PetscBool      flg;

1108:   PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MUMPS Options","Mat");
1109:   PetscOptionsInt("-mat_mumps_icntl_1","ICNTL(1): output stream for error messages","None",mumps->id.ICNTL(1),&icntl,&flg);
1110:   if (flg) mumps->id.ICNTL(1) = icntl;
1111:   PetscOptionsInt("-mat_mumps_icntl_2","ICNTL(2): output stream for diagnostic printing, statistics, and warning","None",mumps->id.ICNTL(2),&icntl,&flg);
1112:   if (flg) mumps->id.ICNTL(2) = icntl;
1113:   PetscOptionsInt("-mat_mumps_icntl_3","ICNTL(3): output stream for global information, collected on the host","None",mumps->id.ICNTL(3),&icntl,&flg);
1114:   if (flg) mumps->id.ICNTL(3) = icntl;

1116:   PetscOptionsInt("-mat_mumps_icntl_4","ICNTL(4): level of printing (0 to 4)","None",mumps->id.ICNTL(4),&icntl,&flg);
1117:   if (flg) mumps->id.ICNTL(4) = icntl;
1118:   if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */

1120:   PetscOptionsInt("-mat_mumps_icntl_6","ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)","None",mumps->id.ICNTL(6),&icntl,&flg);
1121:   if (flg) mumps->id.ICNTL(6) = icntl;

1123:   PetscOptionsInt("-mat_mumps_icntl_7","ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis","None",mumps->id.ICNTL(7),&icntl,&flg);
1124:   if (flg) {
1125:     if (icntl== 1 && mumps->size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"pivot order be set by the user in PERM_IN -- not supported by the PETSc/MUMPS interface\n");
1126:     else mumps->id.ICNTL(7) = icntl;
1127:   }

1129:   PetscOptionsInt("-mat_mumps_icntl_8","ICNTL(8): scaling strategy (-2 to 8 or 77)","None",mumps->id.ICNTL(8),&mumps->id.ICNTL(8),NULL);
1130:   /* PetscOptionsInt("-mat_mumps_icntl_9","ICNTL(9): computes the solution using A or A^T","None",mumps->id.ICNTL(9),&mumps->id.ICNTL(9),NULL); handled by MatSolveTranspose_MUMPS() */
1131:   PetscOptionsInt("-mat_mumps_icntl_10","ICNTL(10): max num of refinements","None",mumps->id.ICNTL(10),&mumps->id.ICNTL(10),NULL);
1132:   PetscOptionsInt("-mat_mumps_icntl_11","ICNTL(11): statistics related to an error analysis (via -ksp_view)","None",mumps->id.ICNTL(11),&mumps->id.ICNTL(11),NULL);
1133:   PetscOptionsInt("-mat_mumps_icntl_12","ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)","None",mumps->id.ICNTL(12),&mumps->id.ICNTL(12),NULL);
1134:   PetscOptionsInt("-mat_mumps_icntl_13","ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting","None",mumps->id.ICNTL(13),&mumps->id.ICNTL(13),NULL);
1135:   PetscOptionsInt("-mat_mumps_icntl_14","ICNTL(14): percentage increase in the estimated working space","None",mumps->id.ICNTL(14),&mumps->id.ICNTL(14),NULL);
1136:   PetscOptionsInt("-mat_mumps_icntl_19","ICNTL(19): computes the Schur complement","None",mumps->id.ICNTL(19),&mumps->id.ICNTL(19),NULL);
1137:   if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any schur data (if any) */
1138:     MatDestroy(&F->schur);
1139:     MatMumpsResetSchur_Private(mumps);
1140:   }
1141:   /* PetscOptionsInt("-mat_mumps_icntl_20","ICNTL(20): the format (dense or sparse) of the right-hand sides","None",mumps->id.ICNTL(20),&mumps->id.ICNTL(20),NULL); -- sparse rhs is not supported in PETSc API */
1142:   /* PetscOptionsInt("-mat_mumps_icntl_21","ICNTL(21): the distribution (centralized or distributed) of the solution vectors","None",mumps->id.ICNTL(21),&mumps->id.ICNTL(21),NULL); we only use distributed solution vector */

1144:   PetscOptionsInt("-mat_mumps_icntl_22","ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)","None",mumps->id.ICNTL(22),&mumps->id.ICNTL(22),NULL);
1145:   PetscOptionsInt("-mat_mumps_icntl_23","ICNTL(23): max size of the working memory (MB) that can allocate per processor","None",mumps->id.ICNTL(23),&mumps->id.ICNTL(23),NULL);
1146:   PetscOptionsInt("-mat_mumps_icntl_24","ICNTL(24): detection of null pivot rows (0 or 1)","None",mumps->id.ICNTL(24),&mumps->id.ICNTL(24),NULL);
1147:   if (mumps->id.ICNTL(24)) {
1148:     mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */
1149:   }

1151:   PetscOptionsInt("-mat_mumps_icntl_25","ICNTL(25): computes a solution of a deficient matrix and a null space basis","None",mumps->id.ICNTL(25),&mumps->id.ICNTL(25),NULL);
1152:   PetscOptionsInt("-mat_mumps_icntl_26","ICNTL(26): drives the solution phase if a Schur complement matrix","None",mumps->id.ICNTL(26),&mumps->id.ICNTL(26),NULL);
1153:   PetscOptionsInt("-mat_mumps_icntl_27","ICNTL(27): the blocking size for multiple right-hand sides","None",mumps->id.ICNTL(27),&mumps->id.ICNTL(27),NULL);
1154:   PetscOptionsInt("-mat_mumps_icntl_28","ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering","None",mumps->id.ICNTL(28),&mumps->id.ICNTL(28),NULL);
1155:   PetscOptionsInt("-mat_mumps_icntl_29","ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis","None",mumps->id.ICNTL(29),&mumps->id.ICNTL(29),NULL);
1156:   PetscOptionsInt("-mat_mumps_icntl_30","ICNTL(30): compute user-specified set of entries in inv(A)","None",mumps->id.ICNTL(30),&mumps->id.ICNTL(30),NULL);
1157:   PetscOptionsInt("-mat_mumps_icntl_31","ICNTL(31): indicates which factors may be discarded during factorization","None",mumps->id.ICNTL(31),&mumps->id.ICNTL(31),NULL);
1158:   /* PetscOptionsInt("-mat_mumps_icntl_32","ICNTL(32): performs the forward elemination of the right-hand sides during factorization","None",mumps->id.ICNTL(32),&mumps->id.ICNTL(32),NULL);  -- not supported by PETSc API */
1159:   PetscOptionsInt("-mat_mumps_icntl_33","ICNTL(33): compute determinant","None",mumps->id.ICNTL(33),&mumps->id.ICNTL(33),NULL);
1160:   PetscOptionsInt("-mat_mumps_icntl_35","ICNTL(35): activates Block Lock Rank (BLR) based factorization","None",mumps->id.ICNTL(35),&mumps->id.ICNTL(35),NULL);

1162:   PetscOptionsReal("-mat_mumps_cntl_1","CNTL(1): relative pivoting threshold","None",mumps->id.CNTL(1),&mumps->id.CNTL(1),NULL);
1163:   PetscOptionsReal("-mat_mumps_cntl_2","CNTL(2): stopping criterion of refinement","None",mumps->id.CNTL(2),&mumps->id.CNTL(2),NULL);
1164:   PetscOptionsReal("-mat_mumps_cntl_3","CNTL(3): absolute pivoting threshold","None",mumps->id.CNTL(3),&mumps->id.CNTL(3),NULL);
1165:   PetscOptionsReal("-mat_mumps_cntl_4","CNTL(4): value for static pivoting","None",mumps->id.CNTL(4),&mumps->id.CNTL(4),NULL);
1166:   PetscOptionsReal("-mat_mumps_cntl_5","CNTL(5): fixation for null pivots","None",mumps->id.CNTL(5),&mumps->id.CNTL(5),NULL);
1167:   PetscOptionsReal("-mat_mumps_cntl_7","CNTL(7): dropping parameter used during BLR","None",mumps->id.CNTL(7),&mumps->id.CNTL(7),NULL);

1169:   PetscOptionsString("-mat_mumps_ooc_tmpdir", "out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, 256, NULL);

1171:   PetscOptionsIntArray("-mat_mumps_view_info","request INFO local to each processor","",info,&ninfo,NULL);
1172:   if (ninfo) {
1173:     if (ninfo > 40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"number of INFO %d must <= 40\n",ninfo);
1174:     PetscMalloc1(ninfo,&mumps->info);
1175:     mumps->ninfo = ninfo;
1176:     for (i=0; i<ninfo; i++) {
1177:       if (info[i] < 0 || info[i]>40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"index of INFO %d must between 1 and 40\n",ninfo);
1178:       else  mumps->info[i] = info[i];
1179:     }
1180:   }

1182:   PetscOptionsEnd();
1183:   return(0);
1184: }

1186: PetscErrorCode PetscInitializeMUMPS(Mat A,Mat_MUMPS *mumps)
1187: {

1191:   MPI_Comm_rank(PetscObjectComm((PetscObject)A), &mumps->myid);
1192:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&mumps->size);
1193:   MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(mumps->comm_mumps));

1195:   mumps->id.comm_fortran = MPI_Comm_c2f(mumps->comm_mumps);

1197:   mumps->id.job = JOB_INIT;
1198:   mumps->id.par = 1;  /* host participates factorizaton and solve */
1199:   mumps->id.sym = mumps->sym;
1200:   PetscMUMPS_c(&mumps->id);

1202:   mumps->scat_rhs     = NULL;
1203:   mumps->scat_sol     = NULL;

1205:   /* set PETSc-MUMPS default options - override MUMPS default */
1206:   mumps->id.ICNTL(3) = 0;
1207:   mumps->id.ICNTL(4) = 0;
1208:   if (mumps->size == 1) {
1209:     mumps->id.ICNTL(18) = 0;   /* centralized assembled matrix input */
1210:   } else {
1211:     mumps->id.ICNTL(18) = 3;   /* distributed assembled matrix input */
1212:     mumps->id.ICNTL(20) = 0;   /* rhs is in dense format */
1213:     mumps->id.ICNTL(21) = 1;   /* distributed solution */
1214:   }

1216:   /* schur */
1217:   mumps->id.size_schur      = 0;
1218:   mumps->id.listvar_schur   = NULL;
1219:   mumps->id.schur           = NULL;
1220:   mumps->sizeredrhs         = 0;
1221:   mumps->schur_sol          = NULL;
1222:   mumps->schur_sizesol      = 0;
1223:   return(0);
1224: }

1226: PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F,Mat A,const MatFactorInfo *info,Mat_MUMPS *mumps)
1227: {

1231:   if (mumps->id.INFOG(1) < 0) {
1232:     if (A->erroriffailure) {
1233:       SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in analysis phase: INFOG(1)=%d\n",mumps->id.INFOG(1));
1234:     } else {
1235:       if (mumps->id.INFOG(1) == -6) {
1236:         PetscInfo2(F,"matrix is singular in structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1237:         F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT;
1238:       } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) {
1239:         PetscInfo2(F,"problem of workspace, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1240:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1241:       } else {
1242:         PetscInfo2(F,"Error reported by MUMPS in analysis phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1243:         F->factorerrortype = MAT_FACTOR_OTHER;
1244:       }
1245:     }
1246:   }
1247:   return(0);
1248: }

1250: /* Note Petsc r(=c) permutation is used when mumps->id.ICNTL(7)==1 with centralized assembled matrix input; otherwise r and c are ignored */
1251: PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
1252: {
1253:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1255:   Vec            b;
1256:   IS             is_iden;
1257:   const PetscInt M = A->rmap->N;

1260:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

1262:   /* Set MUMPS options from the options database */
1263:   PetscSetMUMPSFromOptions(F,A);

1265:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);

1267:   /* analysis phase */
1268:   /*----------------*/
1269:   mumps->id.job = JOB_FACTSYMBOLIC;
1270:   mumps->id.n   = M;
1271:   switch (mumps->id.ICNTL(18)) {
1272:   case 0:  /* centralized assembled matrix input */
1273:     if (!mumps->myid) {
1274:       mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1275:       if (mumps->id.ICNTL(6)>1) {
1276:         mumps->id.a = (MumpsScalar*)mumps->val;
1277:       }
1278:       if (mumps->id.ICNTL(7) == 1) { /* use user-provide matrix ordering - assuming r = c ordering */
1279:         /*
1280:         PetscBool      flag;
1281:         ISEqual(r,c,&flag);
1282:         if (!flag) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"row_perm != col_perm");
1283:         ISView(r,PETSC_VIEWER_STDOUT_SELF);
1284:          */
1285:         if (!mumps->myid) {
1286:           const PetscInt *idx;
1287:           PetscInt       i,*perm_in;

1289:           PetscMalloc1(M,&perm_in);
1290:           ISGetIndices(r,&idx);

1292:           mumps->id.perm_in = perm_in;
1293:           for (i=0; i<M; i++) perm_in[i] = idx[i]+1; /* perm_in[]: start from 1, not 0! */
1294:           ISRestoreIndices(r,&idx);
1295:         }
1296:       }
1297:     }
1298:     break;
1299:   case 3:  /* distributed assembled matrix input (size>1) */
1300:     mumps->id.nz_loc = mumps->nz;
1301:     mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1302:     if (mumps->id.ICNTL(6)>1) {
1303:       mumps->id.a_loc = (MumpsScalar*)mumps->val;
1304:     }
1305:     /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1306:     if (!mumps->myid) {
1307:       VecCreateSeq(PETSC_COMM_SELF,A->rmap->N,&mumps->b_seq);
1308:       ISCreateStride(PETSC_COMM_SELF,A->rmap->N,0,1,&is_iden);
1309:     } else {
1310:       VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1311:       ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1312:     }
1313:     MatCreateVecs(A,NULL,&b);
1314:     VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1315:     ISDestroy(&is_iden);
1316:     VecDestroy(&b);
1317:     break;
1318:   }
1319:   PetscMUMPS_c(&mumps->id);
1320:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

1322:   F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
1323:   F->ops->solve           = MatSolve_MUMPS;
1324:   F->ops->solvetranspose  = MatSolveTranspose_MUMPS;
1325:   F->ops->matsolve        = MatMatSolve_MUMPS;
1326:   return(0);
1327: }

1329: /* Note the Petsc r and c permutations are ignored */
1330: PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
1331: {
1332:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1334:   Vec            b;
1335:   IS             is_iden;
1336:   const PetscInt M = A->rmap->N;

1339:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

1341:   /* Set MUMPS options from the options database */
1342:   PetscSetMUMPSFromOptions(F,A);

1344:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);

1346:   /* analysis phase */
1347:   /*----------------*/
1348:   mumps->id.job = JOB_FACTSYMBOLIC;
1349:   mumps->id.n   = M;
1350:   switch (mumps->id.ICNTL(18)) {
1351:   case 0:  /* centralized assembled matrix input */
1352:     if (!mumps->myid) {
1353:       mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1354:       if (mumps->id.ICNTL(6)>1) {
1355:         mumps->id.a = (MumpsScalar*)mumps->val;
1356:       }
1357:     }
1358:     break;
1359:   case 3:  /* distributed assembled matrix input (size>1) */
1360:     mumps->id.nz_loc = mumps->nz;
1361:     mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1362:     if (mumps->id.ICNTL(6)>1) {
1363:       mumps->id.a_loc = (MumpsScalar*)mumps->val;
1364:     }
1365:     /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1366:     if (!mumps->myid) {
1367:       VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);
1368:       ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);
1369:     } else {
1370:       VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1371:       ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1372:     }
1373:     MatCreateVecs(A,NULL,&b);
1374:     VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1375:     ISDestroy(&is_iden);
1376:     VecDestroy(&b);
1377:     break;
1378:   }
1379:   PetscMUMPS_c(&mumps->id);
1380:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

1382:   F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
1383:   F->ops->solve           = MatSolve_MUMPS;
1384:   F->ops->solvetranspose  = MatSolveTranspose_MUMPS;
1385:   return(0);
1386: }

1388: /* Note the Petsc r permutation and factor info are ignored */
1389: PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F,Mat A,IS r,const MatFactorInfo *info)
1390: {
1391:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1393:   Vec            b;
1394:   IS             is_iden;
1395:   const PetscInt M = A->rmap->N;

1398:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

1400:   /* Set MUMPS options from the options database */
1401:   PetscSetMUMPSFromOptions(F,A);

1403:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);

1405:   /* analysis phase */
1406:   /*----------------*/
1407:   mumps->id.job = JOB_FACTSYMBOLIC;
1408:   mumps->id.n   = M;
1409:   switch (mumps->id.ICNTL(18)) {
1410:   case 0:  /* centralized assembled matrix input */
1411:     if (!mumps->myid) {
1412:       mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1413:       if (mumps->id.ICNTL(6)>1) {
1414:         mumps->id.a = (MumpsScalar*)mumps->val;
1415:       }
1416:     }
1417:     break;
1418:   case 3:  /* distributed assembled matrix input (size>1) */
1419:     mumps->id.nz_loc = mumps->nz;
1420:     mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1421:     if (mumps->id.ICNTL(6)>1) {
1422:       mumps->id.a_loc = (MumpsScalar*)mumps->val;
1423:     }
1424:     /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1425:     if (!mumps->myid) {
1426:       VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);
1427:       ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);
1428:     } else {
1429:       VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1430:       ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1431:     }
1432:     MatCreateVecs(A,NULL,&b);
1433:     VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1434:     ISDestroy(&is_iden);
1435:     VecDestroy(&b);
1436:     break;
1437:   }
1438:   PetscMUMPS_c(&mumps->id);
1439:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

1441:   F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS;
1442:   F->ops->solve                 = MatSolve_MUMPS;
1443:   F->ops->solvetranspose        = MatSolve_MUMPS;
1444:   F->ops->matsolve              = MatMatSolve_MUMPS;
1445: #if defined(PETSC_USE_COMPLEX)
1446:   F->ops->getinertia = NULL;
1447: #else
1448:   F->ops->getinertia = MatGetInertia_SBAIJMUMPS;
1449: #endif
1450:   return(0);
1451: }

1453: PetscErrorCode MatView_MUMPS(Mat A,PetscViewer viewer)
1454: {
1455:   PetscErrorCode    ierr;
1456:   PetscBool         iascii;
1457:   PetscViewerFormat format;
1458:   Mat_MUMPS         *mumps=(Mat_MUMPS*)A->data;

1461:   /* check if matrix is mumps type */
1462:   if (A->ops->solve != MatSolve_MUMPS) return(0);

1464:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1465:   if (iascii) {
1466:     PetscViewerGetFormat(viewer,&format);
1467:     if (format == PETSC_VIEWER_ASCII_INFO) {
1468:       PetscViewerASCIIPrintf(viewer,"MUMPS run parameters:\n");
1469:       PetscViewerASCIIPrintf(viewer,"  SYM (matrix type):                   %d \n",mumps->id.sym);
1470:       PetscViewerASCIIPrintf(viewer,"  PAR (host participation):            %d \n",mumps->id.par);
1471:       PetscViewerASCIIPrintf(viewer,"  ICNTL(1) (output for error):         %d \n",mumps->id.ICNTL(1));
1472:       PetscViewerASCIIPrintf(viewer,"  ICNTL(2) (output of diagnostic msg): %d \n",mumps->id.ICNTL(2));
1473:       PetscViewerASCIIPrintf(viewer,"  ICNTL(3) (output for global info):   %d \n",mumps->id.ICNTL(3));
1474:       PetscViewerASCIIPrintf(viewer,"  ICNTL(4) (level of printing):        %d \n",mumps->id.ICNTL(4));
1475:       PetscViewerASCIIPrintf(viewer,"  ICNTL(5) (input mat struct):         %d \n",mumps->id.ICNTL(5));
1476:       PetscViewerASCIIPrintf(viewer,"  ICNTL(6) (matrix prescaling):        %d \n",mumps->id.ICNTL(6));
1477:       PetscViewerASCIIPrintf(viewer,"  ICNTL(7) (sequential matrix ordering):%d \n",mumps->id.ICNTL(7));
1478:       PetscViewerASCIIPrintf(viewer,"  ICNTL(8) (scaling strategy):        %d \n",mumps->id.ICNTL(8));
1479:       PetscViewerASCIIPrintf(viewer,"  ICNTL(10) (max num of refinements):  %d \n",mumps->id.ICNTL(10));
1480:       PetscViewerASCIIPrintf(viewer,"  ICNTL(11) (error analysis):          %d \n",mumps->id.ICNTL(11));
1481:       if (mumps->id.ICNTL(11)>0) {
1482:         PetscViewerASCIIPrintf(viewer,"    RINFOG(4) (inf norm of input mat):        %g\n",mumps->id.RINFOG(4));
1483:         PetscViewerASCIIPrintf(viewer,"    RINFOG(5) (inf norm of solution):         %g\n",mumps->id.RINFOG(5));
1484:         PetscViewerASCIIPrintf(viewer,"    RINFOG(6) (inf norm of residual):         %g\n",mumps->id.RINFOG(6));
1485:         PetscViewerASCIIPrintf(viewer,"    RINFOG(7),RINFOG(8) (backward error est): %g, %g\n",mumps->id.RINFOG(7),mumps->id.RINFOG(8));
1486:         PetscViewerASCIIPrintf(viewer,"    RINFOG(9) (error estimate):               %g \n",mumps->id.RINFOG(9));
1487:         PetscViewerASCIIPrintf(viewer,"    RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n",mumps->id.RINFOG(10),mumps->id.RINFOG(11));
1488:       }
1489:       PetscViewerASCIIPrintf(viewer,"  ICNTL(12) (efficiency control):                         %d \n",mumps->id.ICNTL(12));
1490:       PetscViewerASCIIPrintf(viewer,"  ICNTL(13) (efficiency control):                         %d \n",mumps->id.ICNTL(13));
1491:       PetscViewerASCIIPrintf(viewer,"  ICNTL(14) (percentage of estimated workspace increase): %d \n",mumps->id.ICNTL(14));
1492:       /* ICNTL(15-17) not used */
1493:       PetscViewerASCIIPrintf(viewer,"  ICNTL(18) (input mat struct):                           %d \n",mumps->id.ICNTL(18));
1494:       PetscViewerASCIIPrintf(viewer,"  ICNTL(19) (Schur complement info):                       %d \n",mumps->id.ICNTL(19));
1495:       PetscViewerASCIIPrintf(viewer,"  ICNTL(20) (rhs sparse pattern):                         %d \n",mumps->id.ICNTL(20));
1496:       PetscViewerASCIIPrintf(viewer,"  ICNTL(21) (solution struct):                            %d \n",mumps->id.ICNTL(21));
1497:       PetscViewerASCIIPrintf(viewer,"  ICNTL(22) (in-core/out-of-core facility):               %d \n",mumps->id.ICNTL(22));
1498:       PetscViewerASCIIPrintf(viewer,"  ICNTL(23) (max size of memory can be allocated locally):%d \n",mumps->id.ICNTL(23));

1500:       PetscViewerASCIIPrintf(viewer,"  ICNTL(24) (detection of null pivot rows):               %d \n",mumps->id.ICNTL(24));
1501:       PetscViewerASCIIPrintf(viewer,"  ICNTL(25) (computation of a null space basis):          %d \n",mumps->id.ICNTL(25));
1502:       PetscViewerASCIIPrintf(viewer,"  ICNTL(26) (Schur options for rhs or solution):          %d \n",mumps->id.ICNTL(26));
1503:       PetscViewerASCIIPrintf(viewer,"  ICNTL(27) (experimental parameter):                     %d \n",mumps->id.ICNTL(27));
1504:       PetscViewerASCIIPrintf(viewer,"  ICNTL(28) (use parallel or sequential ordering):        %d \n",mumps->id.ICNTL(28));
1505:       PetscViewerASCIIPrintf(viewer,"  ICNTL(29) (parallel ordering):                          %d \n",mumps->id.ICNTL(29));

1507:       PetscViewerASCIIPrintf(viewer,"  ICNTL(30) (user-specified set of entries in inv(A)):    %d \n",mumps->id.ICNTL(30));
1508:       PetscViewerASCIIPrintf(viewer,"  ICNTL(31) (factors is discarded in the solve phase):    %d \n",mumps->id.ICNTL(31));
1509:       PetscViewerASCIIPrintf(viewer,"  ICNTL(33) (compute determinant):                        %d \n",mumps->id.ICNTL(33));
1510:       PetscViewerASCIIPrintf(viewer,"  ICNTL(35) (activate BLR based factorization):           %d \n",mumps->id.ICNTL(35));

1512:       PetscViewerASCIIPrintf(viewer,"  CNTL(1) (relative pivoting threshold):      %g \n",mumps->id.CNTL(1));
1513:       PetscViewerASCIIPrintf(viewer,"  CNTL(2) (stopping criterion of refinement): %g \n",mumps->id.CNTL(2));
1514:       PetscViewerASCIIPrintf(viewer,"  CNTL(3) (absolute pivoting threshold):      %g \n",mumps->id.CNTL(3));
1515:       PetscViewerASCIIPrintf(viewer,"  CNTL(4) (value of static pivoting):         %g \n",mumps->id.CNTL(4));
1516:       PetscViewerASCIIPrintf(viewer,"  CNTL(5) (fixation for null pivots):         %g \n",mumps->id.CNTL(5));
1517:       PetscViewerASCIIPrintf(viewer,"  CNTL(7) (dropping parameter for BLR):       %g \n",mumps->id.CNTL(7));

1519:       /* infomation local to each processor */
1520:       PetscViewerASCIIPrintf(viewer, "  RINFO(1) (local estimated flops for the elimination after analysis): \n");
1521:       PetscViewerASCIIPushSynchronized(viewer);
1522:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %g \n",mumps->myid,mumps->id.RINFO(1));
1523:       PetscViewerFlush(viewer);
1524:       PetscViewerASCIIPrintf(viewer, "  RINFO(2) (local estimated flops for the assembly after factorization): \n");
1525:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d]  %g \n",mumps->myid,mumps->id.RINFO(2));
1526:       PetscViewerFlush(viewer);
1527:       PetscViewerASCIIPrintf(viewer, "  RINFO(3) (local estimated flops for the elimination after factorization): \n");
1528:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d]  %g \n",mumps->myid,mumps->id.RINFO(3));
1529:       PetscViewerFlush(viewer);

1531:       PetscViewerASCIIPrintf(viewer, "  INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization): \n");
1532:       PetscViewerASCIISynchronizedPrintf(viewer,"  [%d] %d \n",mumps->myid,mumps->id.INFO(15));
1533:       PetscViewerFlush(viewer);

1535:       PetscViewerASCIIPrintf(viewer, "  INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization): \n");
1536:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d \n",mumps->myid,mumps->id.INFO(16));
1537:       PetscViewerFlush(viewer);

1539:       PetscViewerASCIIPrintf(viewer, "  INFO(23) (num of pivots eliminated on this processor after factorization): \n");
1540:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d \n",mumps->myid,mumps->id.INFO(23));
1541:       PetscViewerFlush(viewer);

1543:       if (mumps->ninfo && mumps->ninfo <= 40){
1544:         PetscInt i;
1545:         for (i=0; i<mumps->ninfo; i++){
1546:           PetscViewerASCIIPrintf(viewer, "  INFO(%d): \n",mumps->info[i]);
1547:           PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d \n",mumps->myid,mumps->id.INFO(mumps->info[i]));
1548:           PetscViewerFlush(viewer);
1549:         }
1550:       }


1553:       PetscViewerASCIIPopSynchronized(viewer);

1555:       if (!mumps->myid) { /* information from the host */
1556:         PetscViewerASCIIPrintf(viewer,"  RINFOG(1) (global estimated flops for the elimination after analysis): %g \n",mumps->id.RINFOG(1));
1557:         PetscViewerASCIIPrintf(viewer,"  RINFOG(2) (global estimated flops for the assembly after factorization): %g \n",mumps->id.RINFOG(2));
1558:         PetscViewerASCIIPrintf(viewer,"  RINFOG(3) (global estimated flops for the elimination after factorization): %g \n",mumps->id.RINFOG(3));
1559:         PetscViewerASCIIPrintf(viewer,"  (RINFOG(12) RINFOG(13))*2^INFOG(34) (determinant): (%g,%g)*(2^%d)\n",mumps->id.RINFOG(12),mumps->id.RINFOG(13),mumps->id.INFOG(34));

1561:         PetscViewerASCIIPrintf(viewer,"  INFOG(3) (estimated real workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(3));
1562:         PetscViewerASCIIPrintf(viewer,"  INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(4));
1563:         PetscViewerASCIIPrintf(viewer,"  INFOG(5) (estimated maximum front size in the complete tree): %d \n",mumps->id.INFOG(5));
1564:         PetscViewerASCIIPrintf(viewer,"  INFOG(6) (number of nodes in the complete tree): %d \n",mumps->id.INFOG(6));
1565:         PetscViewerASCIIPrintf(viewer,"  INFOG(7) (ordering option effectively use after analysis): %d \n",mumps->id.INFOG(7));
1566:         PetscViewerASCIIPrintf(viewer,"  INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d \n",mumps->id.INFOG(8));
1567:         PetscViewerASCIIPrintf(viewer,"  INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d \n",mumps->id.INFOG(9));
1568:         PetscViewerASCIIPrintf(viewer,"  INFOG(10) (total integer space store the matrix factors after factorization): %d \n",mumps->id.INFOG(10));
1569:         PetscViewerASCIIPrintf(viewer,"  INFOG(11) (order of largest frontal matrix after factorization): %d \n",mumps->id.INFOG(11));
1570:         PetscViewerASCIIPrintf(viewer,"  INFOG(12) (number of off-diagonal pivots): %d \n",mumps->id.INFOG(12));
1571:         PetscViewerASCIIPrintf(viewer,"  INFOG(13) (number of delayed pivots after factorization): %d \n",mumps->id.INFOG(13));
1572:         PetscViewerASCIIPrintf(viewer,"  INFOG(14) (number of memory compress after factorization): %d \n",mumps->id.INFOG(14));
1573:         PetscViewerASCIIPrintf(viewer,"  INFOG(15) (number of steps of iterative refinement after solution): %d \n",mumps->id.INFOG(15));
1574:         PetscViewerASCIIPrintf(viewer,"  INFOG(16) (estimated size (in MB) of all MUMPS internal data for factorization after analysis: value on the most memory consuming processor): %d \n",mumps->id.INFOG(16));
1575:         PetscViewerASCIIPrintf(viewer,"  INFOG(17) (estimated size of all MUMPS internal data for factorization after analysis: sum over all processors): %d \n",mumps->id.INFOG(17));
1576:         PetscViewerASCIIPrintf(viewer,"  INFOG(18) (size of all MUMPS internal data allocated during factorization: value on the most memory consuming processor): %d \n",mumps->id.INFOG(18));
1577:         PetscViewerASCIIPrintf(viewer,"  INFOG(19) (size of all MUMPS internal data allocated during factorization: sum over all processors): %d \n",mumps->id.INFOG(19));
1578:         PetscViewerASCIIPrintf(viewer,"  INFOG(20) (estimated number of entries in the factors): %d \n",mumps->id.INFOG(20));
1579:         PetscViewerASCIIPrintf(viewer,"  INFOG(21) (size in MB of memory effectively used during factorization - value on the most memory consuming processor): %d \n",mumps->id.INFOG(21));
1580:         PetscViewerASCIIPrintf(viewer,"  INFOG(22) (size in MB of memory effectively used during factorization - sum over all processors): %d \n",mumps->id.INFOG(22));
1581:         PetscViewerASCIIPrintf(viewer,"  INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d \n",mumps->id.INFOG(23));
1582:         PetscViewerASCIIPrintf(viewer,"  INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d \n",mumps->id.INFOG(24));
1583:         PetscViewerASCIIPrintf(viewer,"  INFOG(25) (after factorization: number of pivots modified by static pivoting): %d \n",mumps->id.INFOG(25));
1584:         PetscViewerASCIIPrintf(viewer,"  INFOG(28) (after factorization: number of null pivots encountered): %d\n",mumps->id.INFOG(28));
1585:         PetscViewerASCIIPrintf(viewer,"  INFOG(29) (after factorization: effective number of entries in the factors (sum over all processors)): %d\n",mumps->id.INFOG(29));
1586:         PetscViewerASCIIPrintf(viewer,"  INFOG(30, 31) (after solution: size in Mbytes of memory used during solution phase): %d, %d\n",mumps->id.INFOG(30),mumps->id.INFOG(31));
1587:         PetscViewerASCIIPrintf(viewer,"  INFOG(32) (after analysis: type of analysis done): %d\n",mumps->id.INFOG(32));
1588:         PetscViewerASCIIPrintf(viewer,"  INFOG(33) (value used for ICNTL(8)): %d\n",mumps->id.INFOG(33));
1589:         PetscViewerASCIIPrintf(viewer,"  INFOG(34) (exponent of the determinant if determinant is requested): %d\n",mumps->id.INFOG(34));
1590:       }
1591:     }
1592:   }
1593:   return(0);
1594: }

1596: PetscErrorCode MatGetInfo_MUMPS(Mat A,MatInfoType flag,MatInfo *info)
1597: {
1598:   Mat_MUMPS *mumps =(Mat_MUMPS*)A->data;

1601:   info->block_size        = 1.0;
1602:   info->nz_allocated      = mumps->id.INFOG(20);
1603:   info->nz_used           = mumps->id.INFOG(20);
1604:   info->nz_unneeded       = 0.0;
1605:   info->assemblies        = 0.0;
1606:   info->mallocs           = 0.0;
1607:   info->memory            = 0.0;
1608:   info->fill_ratio_given  = 0;
1609:   info->fill_ratio_needed = 0;
1610:   info->factor_mallocs    = 0;
1611:   return(0);
1612: }

1614: /* -------------------------------------------------------------------------------------------*/
1615: PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is)
1616: {
1617:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
1618:   const PetscInt *idxs;
1619:   PetscInt       size,i;

1623:   ISGetLocalSize(is,&size);
1624:   if (mumps->size > 1) {
1625:     PetscBool ls,gs;

1627:     ls   = mumps->myid ? (size ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE;
1628:     MPI_Allreduce(&ls,&gs,1,MPIU_BOOL,MPI_LAND,mumps->comm_mumps);
1629:     if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MUMPS distributed parallel Schur complements not yet supported from PETSc\n");
1630:   }
1631:   if (mumps->id.size_schur != size) {
1632:     PetscFree2(mumps->id.listvar_schur,mumps->id.schur);
1633:     mumps->id.size_schur = size;
1634:     mumps->id.schur_lld  = size;
1635:     PetscMalloc2(size,&mumps->id.listvar_schur,size*size,&mumps->id.schur);
1636:   }

1638:   /* Schur complement matrix */
1639:   MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.size_schur,(PetscScalar*)mumps->id.schur,&F->schur);
1640:   if (mumps->sym == 1) {
1641:     MatSetOption(F->schur,MAT_SPD,PETSC_TRUE);
1642:   }

1644:   /* MUMPS expects Fortran style indices */
1645:   ISGetIndices(is,&idxs);
1646:   PetscMemcpy(mumps->id.listvar_schur,idxs,size*sizeof(PetscInt));
1647:   for (i=0;i<size;i++) mumps->id.listvar_schur[i]++;
1648:   ISRestoreIndices(is,&idxs);
1649:   if (mumps->size > 1) {
1650:     mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */
1651:   } else {
1652:     if (F->factortype == MAT_FACTOR_LU) {
1653:       mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */
1654:     } else {
1655:       mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */
1656:     }
1657:   }
1658:   /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */
1659:   mumps->id.ICNTL(26) = -1;
1660:   return(0);
1661: }

1663: /* -------------------------------------------------------------------------------------------*/
1664: PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F,Mat* S)
1665: {
1666:   Mat            St;
1667:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
1668:   PetscScalar    *array;
1669: #if defined(PETSC_USE_COMPLEX)
1670:   PetscScalar    im = PetscSqrtScalar((PetscScalar)-1.0);
1671: #endif

1675:   if (!mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ORDER,"Schur complement mode not selected! You should call MatFactorSetSchurIS to enable it");
1676:   MatCreate(PETSC_COMM_SELF,&St);
1677:   MatSetSizes(St,PETSC_DECIDE,PETSC_DECIDE,mumps->id.size_schur,mumps->id.size_schur);
1678:   MatSetType(St,MATDENSE);
1679:   MatSetUp(St);
1680:   MatDenseGetArray(St,&array);
1681:   if (!mumps->sym) { /* MUMPS always return a full matrix */
1682:     if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
1683:       PetscInt i,j,N=mumps->id.size_schur;
1684:       for (i=0;i<N;i++) {
1685:         for (j=0;j<N;j++) {
1686: #if !defined(PETSC_USE_COMPLEX)
1687:           PetscScalar val = mumps->id.schur[i*N+j];
1688: #else
1689:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1690: #endif
1691:           array[j*N+i] = val;
1692:         }
1693:       }
1694:     } else { /* stored by columns */
1695:       PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));
1696:     }
1697:   } else { /* either full or lower-triangular (not packed) */
1698:     if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */
1699:       PetscInt i,j,N=mumps->id.size_schur;
1700:       for (i=0;i<N;i++) {
1701:         for (j=i;j<N;j++) {
1702: #if !defined(PETSC_USE_COMPLEX)
1703:           PetscScalar val = mumps->id.schur[i*N+j];
1704: #else
1705:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1706: #endif
1707:           array[i*N+j] = val;
1708:           array[j*N+i] = val;
1709:         }
1710:       }
1711:     } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */
1712:       PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));
1713:     } else { /* ICNTL(19) == 1 lower triangular stored by rows */
1714:       PetscInt i,j,N=mumps->id.size_schur;
1715:       for (i=0;i<N;i++) {
1716:         for (j=0;j<i+1;j++) {
1717: #if !defined(PETSC_USE_COMPLEX)
1718:           PetscScalar val = mumps->id.schur[i*N+j];
1719: #else
1720:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1721: #endif
1722:           array[i*N+j] = val;
1723:           array[j*N+i] = val;
1724:         }
1725:       }
1726:     }
1727:   }
1728:   MatDenseRestoreArray(St,&array);
1729:   *S   = St;
1730:   return(0);
1731: }

1733: /* -------------------------------------------------------------------------------------------*/
1734: PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt ival)
1735: {
1736:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1739:   mumps->id.ICNTL(icntl) = ival;
1740:   return(0);
1741: }

1743: PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt *ival)
1744: {
1745:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1748:   *ival = mumps->id.ICNTL(icntl);
1749:   return(0);
1750: }

1752: /*@
1753:   MatMumpsSetIcntl - Set MUMPS parameter ICNTL()

1755:    Logically Collective on Mat

1757:    Input Parameters:
1758: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1759: .  icntl - index of MUMPS parameter array ICNTL()
1760: -  ival - value of MUMPS ICNTL(icntl)

1762:   Options Database:
1763: .   -mat_mumps_icntl_<icntl> <ival>

1765:    Level: beginner

1767:    References:
1768: .     MUMPS Users' Guide

1770: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1771:  @*/
1772: PetscErrorCode MatMumpsSetIcntl(Mat F,PetscInt icntl,PetscInt ival)
1773: {
1775: 
1778:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1781:   PetscTryMethod(F,"MatMumpsSetIcntl_C",(Mat,PetscInt,PetscInt),(F,icntl,ival));
1782:   return(0);
1783: }

1785: /*@
1786:   MatMumpsGetIcntl - Get MUMPS parameter ICNTL()

1788:    Logically Collective on Mat

1790:    Input Parameters:
1791: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1792: -  icntl - index of MUMPS parameter array ICNTL()

1794:   Output Parameter:
1795: .  ival - value of MUMPS ICNTL(icntl)

1797:    Level: beginner

1799:    References:
1800: .     MUMPS Users' Guide

1802: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1803: @*/
1804: PetscErrorCode MatMumpsGetIcntl(Mat F,PetscInt icntl,PetscInt *ival)
1805: {

1810:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1813:   PetscUseMethod(F,"MatMumpsGetIcntl_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
1814:   return(0);
1815: }

1817: /* -------------------------------------------------------------------------------------------*/
1818: PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal val)
1819: {
1820:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1823:   mumps->id.CNTL(icntl) = val;
1824:   return(0);
1825: }

1827: PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal *val)
1828: {
1829:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1832:   *val = mumps->id.CNTL(icntl);
1833:   return(0);
1834: }

1836: /*@
1837:   MatMumpsSetCntl - Set MUMPS parameter CNTL()

1839:    Logically Collective on Mat

1841:    Input Parameters:
1842: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1843: .  icntl - index of MUMPS parameter array CNTL()
1844: -  val - value of MUMPS CNTL(icntl)

1846:   Options Database:
1847: .   -mat_mumps_cntl_<icntl> <val>

1849:    Level: beginner

1851:    References:
1852: .     MUMPS Users' Guide

1854: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1855: @*/
1856: PetscErrorCode MatMumpsSetCntl(Mat F,PetscInt icntl,PetscReal val)
1857: {

1862:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1865:   PetscTryMethod(F,"MatMumpsSetCntl_C",(Mat,PetscInt,PetscReal),(F,icntl,val));
1866:   return(0);
1867: }

1869: /*@
1870:   MatMumpsGetCntl - Get MUMPS parameter CNTL()

1872:    Logically Collective on Mat

1874:    Input Parameters:
1875: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1876: -  icntl - index of MUMPS parameter array CNTL()

1878:   Output Parameter:
1879: .  val - value of MUMPS CNTL(icntl)

1881:    Level: beginner

1883:    References:
1884: .      MUMPS Users' Guide

1886: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1887: @*/
1888: PetscErrorCode MatMumpsGetCntl(Mat F,PetscInt icntl,PetscReal *val)
1889: {

1894:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1897:   PetscUseMethod(F,"MatMumpsGetCntl_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
1898:   return(0);
1899: }

1901: PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F,PetscInt icntl,PetscInt *info)
1902: {
1903:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1906:   *info = mumps->id.INFO(icntl);
1907:   return(0);
1908: }

1910: PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F,PetscInt icntl,PetscInt *infog)
1911: {
1912:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1915:   *infog = mumps->id.INFOG(icntl);
1916:   return(0);
1917: }

1919: PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfo)
1920: {
1921:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1924:   *rinfo = mumps->id.RINFO(icntl);
1925:   return(0);
1926: }

1928: PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfog)
1929: {
1930:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

1933:   *rinfog = mumps->id.RINFOG(icntl);
1934:   return(0);
1935: }

1937: /*@
1938:   MatMumpsGetInfo - Get MUMPS parameter INFO()

1940:    Logically Collective on Mat

1942:    Input Parameters:
1943: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1944: -  icntl - index of MUMPS parameter array INFO()

1946:   Output Parameter:
1947: .  ival - value of MUMPS INFO(icntl)

1949:    Level: beginner

1951:    References:
1952: .      MUMPS Users' Guide

1954: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1955: @*/
1956: PetscErrorCode MatMumpsGetInfo(Mat F,PetscInt icntl,PetscInt *ival)
1957: {

1962:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1964:   PetscUseMethod(F,"MatMumpsGetInfo_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
1965:   return(0);
1966: }

1968: /*@
1969:   MatMumpsGetInfog - Get MUMPS parameter INFOG()

1971:    Logically Collective on Mat

1973:    Input Parameters:
1974: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1975: -  icntl - index of MUMPS parameter array INFOG()

1977:   Output Parameter:
1978: .  ival - value of MUMPS INFOG(icntl)

1980:    Level: beginner

1982:    References:
1983: .      MUMPS Users' Guide

1985: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1986: @*/
1987: PetscErrorCode MatMumpsGetInfog(Mat F,PetscInt icntl,PetscInt *ival)
1988: {

1993:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1995:   PetscUseMethod(F,"MatMumpsGetInfog_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
1996:   return(0);
1997: }

1999: /*@
2000:   MatMumpsGetRinfo - Get MUMPS parameter RINFO()

2002:    Logically Collective on Mat

2004:    Input Parameters:
2005: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2006: -  icntl - index of MUMPS parameter array RINFO()

2008:   Output Parameter:
2009: .  val - value of MUMPS RINFO(icntl)

2011:    Level: beginner

2013:    References:
2014: .       MUMPS Users' Guide

2016: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2017: @*/
2018: PetscErrorCode MatMumpsGetRinfo(Mat F,PetscInt icntl,PetscReal *val)
2019: {

2024:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2026:   PetscUseMethod(F,"MatMumpsGetRinfo_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2027:   return(0);
2028: }

2030: /*@
2031:   MatMumpsGetRinfog - Get MUMPS parameter RINFOG()

2033:    Logically Collective on Mat

2035:    Input Parameters:
2036: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2037: -  icntl - index of MUMPS parameter array RINFOG()

2039:   Output Parameter:
2040: .  val - value of MUMPS RINFOG(icntl)

2042:    Level: beginner

2044:    References:
2045: .      MUMPS Users' Guide

2047: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2048: @*/
2049: PetscErrorCode MatMumpsGetRinfog(Mat F,PetscInt icntl,PetscReal *val)
2050: {

2055:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2057:   PetscUseMethod(F,"MatMumpsGetRinfog_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2058:   return(0);
2059: }

2061: /*MC
2062:   MATSOLVERMUMPS -  A matrix type providing direct solvers (LU and Cholesky) for
2063:   distributed and sequential matrices via the external package MUMPS.

2065:   Works with MATAIJ and MATSBAIJ matrices

2067:   Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch  to have PETSc installed with MUMPS

2069:   Use -pc_type cholesky or lu -pc_factor_mat_solver_package mumps to us this direct solver

2071:   Options Database Keys:
2072: +  -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages 
2073: .  -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning 
2074: .  -mat_mumps_icntl_3 -  ICNTL(3): output stream for global information, collected on the host 
2075: .  -mat_mumps_icntl_4 -  ICNTL(4): level of printing (0 to 4) 
2076: .  -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7) 
2077: .  -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis 
2078: .  -mat_mumps_icntl_8  - ICNTL(8): scaling strategy (-2 to 8 or 77) 
2079: .  -mat_mumps_icntl_10  - ICNTL(10): max num of refinements 
2080: .  -mat_mumps_icntl_11  - ICNTL(11): statistics related to an error analysis (via -ksp_view) 
2081: .  -mat_mumps_icntl_12  - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3) 
2082: .  -mat_mumps_icntl_13  - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting 
2083: .  -mat_mumps_icntl_14  - ICNTL(14): percentage increase in the estimated working space 
2084: .  -mat_mumps_icntl_19  - ICNTL(19): computes the Schur complement 
2085: .  -mat_mumps_icntl_22  - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1) 
2086: .  -mat_mumps_icntl_23  - ICNTL(23): max size of the working memory (MB) that can allocate per processor 
2087: .  -mat_mumps_icntl_24  - ICNTL(24): detection of null pivot rows (0 or 1) 
2088: .  -mat_mumps_icntl_25  - ICNTL(25): compute a solution of a deficient matrix and a null space basis 
2089: .  -mat_mumps_icntl_26  - ICNTL(26): drives the solution phase if a Schur complement matrix 
2090: .  -mat_mumps_icntl_28  - ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering 
2091: .  -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis 
2092: .  -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A) 
2093: .  -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization 
2094: .  -mat_mumps_icntl_33 - ICNTL(33): compute determinant 
2095: .  -mat_mumps_cntl_1  - CNTL(1): relative pivoting threshold 
2096: .  -mat_mumps_cntl_2  -  CNTL(2): stopping criterion of refinement 
2097: .  -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold 
2098: .  -mat_mumps_cntl_4 - CNTL(4): value for static pivoting 
2099: -  -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots 

2101:   Level: beginner

2103:     Notes: When a MUMPS factorization fails inside a KSP solve, for example with a KSP_DIVERGED_PCSETUP_FAILED, one can find the MUMPS information about the failure by calling 
2104: $          KSPGetPC(ksp,&pc);
2105: $          PCFactorGetMatrix(pc,&mat);
2106: $          MatMumpsGetInfo(mat,....);
2107: $          MatMumpsGetInfog(mat,....); etc.
2108:            Or you can run with -ksp_error_if_not_converged and the program will be stopped and the information printed in the error message.

2110: .seealso: PCFactorSetMatSolverPackage(), MatSolverPackage, MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog(), KSPGetPC(), PCGetFactor(), PCFactorGetMatrix()

2112: M*/

2114: static PetscErrorCode MatFactorGetSolverPackage_mumps(Mat A,const MatSolverPackage *type)
2115: {
2117:   *type = MATSOLVERMUMPS;
2118:   return(0);
2119: }

2121: /* MatGetFactor for Seq and MPI AIJ matrices */
2122: static PetscErrorCode MatGetFactor_aij_mumps(Mat A,MatFactorType ftype,Mat *F)
2123: {
2124:   Mat            B;
2126:   Mat_MUMPS      *mumps;
2127:   PetscBool      isSeqAIJ;

2130:   /* Create the factorization matrix */
2131:   PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);
2132:   MatCreate(PetscObjectComm((PetscObject)A),&B);
2133:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2134:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2135:   MatSetUp(B);

2137:   PetscNewLog(B,&mumps);

2139:   B->ops->view        = MatView_MUMPS;
2140:   B->ops->getinfo     = MatGetInfo_MUMPS;

2142:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_mumps);
2143:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2144:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2145:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2146:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2147:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2148:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2149:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2150:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2151:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2152:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);

2154:   if (ftype == MAT_FACTOR_LU) {
2155:     B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
2156:     B->factortype            = MAT_FACTOR_LU;
2157:     if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij;
2158:     else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij;
2159:     mumps->sym = 0;
2160:   } else {
2161:     B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
2162:     B->factortype                  = MAT_FACTOR_CHOLESKY;
2163:     if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij;
2164:     else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij;
2165: #if defined(PETSC_USE_COMPLEX)
2166:     mumps->sym = 2;
2167: #else
2168:     if (A->spd_set && A->spd) mumps->sym = 1;
2169:     else                      mumps->sym = 2;
2170: #endif
2171:   }

2173:   /* set solvertype */
2174:   PetscFree(B->solvertype);
2175:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);

2177:   mumps->isAIJ    = PETSC_TRUE;
2178:   B->ops->destroy = MatDestroy_MUMPS;
2179:   B->data        = (void*)mumps;

2181:   PetscInitializeMUMPS(A,mumps);

2183:   *F = B;
2184:   return(0);
2185: }

2187: /* MatGetFactor for Seq and MPI SBAIJ matrices */
2188: static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A,MatFactorType ftype,Mat *F)
2189: {
2190:   Mat            B;
2192:   Mat_MUMPS      *mumps;
2193:   PetscBool      isSeqSBAIJ;

2196:   if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with MUMPS LU, use AIJ matrix");
2197:   if (A->rmap->bs > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with block size > 1 with MUMPS Cholesky, use AIJ matrix instead");
2198:   PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isSeqSBAIJ);
2199:   /* Create the factorization matrix */
2200:   MatCreate(PetscObjectComm((PetscObject)A),&B);
2201:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2202:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2203:   MatSetUp(B);

2205:   PetscNewLog(B,&mumps);
2206:   if (isSeqSBAIJ) {
2207:     mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij;
2208:   } else {
2209:     mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij;
2210:   }

2212:   B->ops->getinfo                = MatGetInfo_External;
2213:   B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
2214:   B->ops->view                   = MatView_MUMPS;

2216:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_mumps);
2217:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2218:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2219:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2220:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2221:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2222:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2223:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2224:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2225:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2226:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);

2228:   B->factortype = MAT_FACTOR_CHOLESKY;
2229: #if defined(PETSC_USE_COMPLEX)
2230:   mumps->sym = 2;
2231: #else
2232:   if (A->spd_set && A->spd) mumps->sym = 1;
2233:   else                      mumps->sym = 2;
2234: #endif

2236:   /* set solvertype */
2237:   PetscFree(B->solvertype);
2238:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);

2240:   mumps->isAIJ    = PETSC_FALSE;
2241:   B->ops->destroy = MatDestroy_MUMPS;
2242:   B->data        = (void*)mumps;

2244:   PetscInitializeMUMPS(A,mumps);

2246:   *F = B;
2247:   return(0);
2248: }

2250: static PetscErrorCode MatGetFactor_baij_mumps(Mat A,MatFactorType ftype,Mat *F)
2251: {
2252:   Mat            B;
2254:   Mat_MUMPS      *mumps;
2255:   PetscBool      isSeqBAIJ;

2258:   /* Create the factorization matrix */
2259:   PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isSeqBAIJ);
2260:   MatCreate(PetscObjectComm((PetscObject)A),&B);
2261:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2262:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2263:   MatSetUp(B);

2265:   PetscNewLog(B,&mumps);
2266:   if (ftype == MAT_FACTOR_LU) {
2267:     B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS;
2268:     B->factortype            = MAT_FACTOR_LU;
2269:     if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij;
2270:     else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij;
2271:     mumps->sym = 0;
2272:   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead\n");

2274:   B->ops->getinfo     = MatGetInfo_External;
2275:   B->ops->view        = MatView_MUMPS;

2277:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_mumps);
2278:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2279:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2280:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2281:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2282:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2283:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2284:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2285:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2286:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2287:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);

2289:   /* set solvertype */
2290:   PetscFree(B->solvertype);
2291:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);

2293:   mumps->isAIJ    = PETSC_TRUE;
2294:   B->ops->destroy = MatDestroy_MUMPS;
2295:   B->data        = (void*)mumps;

2297:   PetscInitializeMUMPS(A,mumps);

2299:   *F = B;
2300:   return(0);
2301: }

2303: PETSC_EXTERN PetscErrorCode MatSolverPackageRegister_MUMPS(void)
2304: {

2308:   MatSolverPackageRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
2309:   MatSolverPackageRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
2310:   MatSolverPackageRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
2311:   MatSolverPackageRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
2312:   MatSolverPackageRegister(MATSOLVERMUMPS,MATMPISBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
2313:   MatSolverPackageRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
2314:   MatSolverPackageRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
2315:   MatSolverPackageRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
2316:   MatSolverPackageRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
2317:   MatSolverPackageRegister(MATSOLVERMUMPS,MATSEQSBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
2318:   return(0);
2319: }