Actual source code: mumps.c


  2: /*
  3:     Provides an interface to the MUMPS sparse solver
  4: */
  5: #include <petscpkg_version.h>
  6: #include <../src/mat/impls/aij/mpi/mpiaij.h>
  7: #include <../src/mat/impls/sbaij/mpi/mpisbaij.h>
  8: #include <../src/mat/impls/sell/mpi/mpisell.h>

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

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

 46: /* MUMPS uses MUMPS_INT for nonzero indices such as irn/jcn, irn_loc/jcn_loc and uses int64_t for
 47:    number of nonzeros such as nnz, nnz_loc. We typedef MUMPS_INT to PetscMUMPSInt to follow the
 48:    naming convention in PetscMPIInt, PetscBLASInt etc.
 49: */
 50: typedef MUMPS_INT PetscMUMPSInt;

 52: #if PETSC_PKG_MUMPS_VERSION_GE(5,3,0)
 53:   #if defined(MUMPS_INTSIZE64) /* MUMPS_INTSIZE64 is in MUMPS headers if it is built in full 64-bit mode, therefore the macro is more reliable */
 54:     #error "Petsc has not been tested with full 64-bit MUMPS and we choose to error out"
 55:   #endif
 56: #else
 57:   #if defined(INTSIZE64) /* INTSIZE64 is a command line macro one used to build MUMPS in full 64-bit mode */
 58:     #error "Petsc has not been tested with full 64-bit MUMPS and we choose to error out"
 59:   #endif
 60: #endif

 62: #define MPIU_MUMPSINT             MPI_INT
 63: #define PETSC_MUMPS_INT_MAX       2147483647
 64: #define PETSC_MUMPS_INT_MIN       -2147483648

 66: /* Cast PetscInt to PetscMUMPSInt. Usually there is no overflow since <a> is row/col indices or some small integers*/
 67: PETSC_STATIC_INLINE PetscErrorCode PetscMUMPSIntCast(PetscInt a,PetscMUMPSInt *b)
 68: {
 70:   if (PetscDefined(USE_64BIT_INDICES) && PetscUnlikelyDebug(a > PETSC_MUMPS_INT_MAX || a < PETSC_MUMPS_INT_MIN)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"PetscInt too long for PetscMUMPSInt");
 71:   *b = (PetscMUMPSInt)(a);
 72:   return(0);
 73: }

 75: /* Put these utility routines here since they are only used in this file */
 76: PETSC_STATIC_INLINE PetscErrorCode PetscOptionsMUMPSInt_Private(PetscOptionItems *PetscOptionsObject,const char opt[],const char text[],const char man[],PetscMUMPSInt currentvalue,PetscMUMPSInt *value,PetscBool *set,PetscMUMPSInt lb,PetscMUMPSInt ub)
 77: {
 79:   PetscInt       myval;
 80:   PetscBool      myset;
 82:   /* PetscInt's size should be always >= PetscMUMPSInt's. It is safe to call PetscOptionsInt_Private to read a PetscMUMPSInt */
 83:   PetscOptionsInt_Private(PetscOptionsObject,opt,text,man,(PetscInt)currentvalue,&myval,&myset,lb,ub);
 84:   if (myset) {PetscMUMPSIntCast(myval,value);}
 85:   if (set) *set = myset;
 86:   return(0);
 87: }
 88: #define PetscOptionsMUMPSInt(a,b,c,d,e,f) PetscOptionsMUMPSInt_Private(PetscOptionsObject,a,b,c,d,e,f,PETSC_MUMPS_INT_MIN,PETSC_MUMPS_INT_MAX)

 90: /* if using PETSc OpenMP support, we only call MUMPS on master ranks. Before/after the call, we change/restore CPUs the master ranks can run on */
 91: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
 92: #define PetscMUMPS_c(mumps) \
 93:   do { \
 94:     if (mumps->use_petsc_omp_support) { \
 95:       if (mumps->is_omp_master) { \
 96:         PetscOmpCtrlOmpRegionOnMasterBegin(mumps->omp_ctrl); \
 97:         MUMPS_c(&mumps->id); \
 98:         PetscOmpCtrlOmpRegionOnMasterEnd(mumps->omp_ctrl); \
 99:       } \
100:       PetscOmpCtrlBarrier(mumps->omp_ctrl); \
101:       /* Global info is same on all processes so we Bcast it within omp_comm. Local info is specific      \
102:          to processes, so we only Bcast info[1], an error code and leave others (since they do not have   \
103:          an easy translation between omp_comm and petsc_comm). See MUMPS-5.1.2 manual p82.                   \
104:          omp_comm is a small shared memory communicator, hence doing multiple Bcast as shown below is OK. \
105:       */ \
106:       MPI_Bcast(mumps->id.infog, 40,MPIU_MUMPSINT, 0,mumps->omp_comm);\
107:       MPI_Bcast(mumps->id.rinfog,20,MPIU_REAL,     0,mumps->omp_comm);\
108:       MPI_Bcast(mumps->id.info,  1, MPIU_MUMPSINT, 0,mumps->omp_comm);\
109:     } else { \
110:       MUMPS_c(&mumps->id); \
111:     } \
112:   } while (0)
113: #else
114: #define PetscMUMPS_c(mumps) \
115:   do { MUMPS_c(&mumps->id); } while (0)
116: #endif

118: /* declare MumpsScalar */
119: #if defined(PETSC_USE_COMPLEX)
120: #if defined(PETSC_USE_REAL_SINGLE)
121: #define MumpsScalar mumps_complex
122: #else
123: #define MumpsScalar mumps_double_complex
124: #endif
125: #else
126: #define MumpsScalar PetscScalar
127: #endif

129: /* macros s.t. indices match MUMPS documentation */
130: #define ICNTL(I) icntl[(I)-1]
131: #define CNTL(I) cntl[(I)-1]
132: #define INFOG(I) infog[(I)-1]
133: #define INFO(I) info[(I)-1]
134: #define RINFOG(I) rinfog[(I)-1]
135: #define RINFO(I) rinfo[(I)-1]

137: typedef struct Mat_MUMPS Mat_MUMPS;
138: struct Mat_MUMPS {
139: #if defined(PETSC_USE_COMPLEX)
140: #if defined(PETSC_USE_REAL_SINGLE)
141:   CMUMPS_STRUC_C id;
142: #else
143:   ZMUMPS_STRUC_C id;
144: #endif
145: #else
146: #if defined(PETSC_USE_REAL_SINGLE)
147:   SMUMPS_STRUC_C id;
148: #else
149:   DMUMPS_STRUC_C id;
150: #endif
151: #endif

153:   MatStructure   matstruc;
154:   PetscMPIInt    myid,petsc_size;
155:   PetscMUMPSInt  *irn,*jcn;             /* the (i,j,v) triplets passed to mumps. */
156:   PetscScalar    *val,*val_alloc;       /* For some matrices, we can directly access their data array without a buffer. For others, we need a buffer. So comes val_alloc. */
157:   PetscInt64     nnz;                   /* number of nonzeros. The type is called selective 64-bit in mumps */
158:   PetscMUMPSInt  sym;
159:   MPI_Comm       mumps_comm;
160:   PetscMUMPSInt  ICNTL9_pre;            /* check if ICNTL(9) is changed from previous MatSolve */
161:   VecScatter     scat_rhs, scat_sol;    /* used by MatSolve() */
162:   PetscMUMPSInt  ICNTL20;               /* use centralized (0) or distributed (10) dense RHS */
163:   PetscMUMPSInt  lrhs_loc,nloc_rhs,*irhs_loc;
164: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
165:   PetscInt       *rhs_nrow,max_nrhs;
166:   PetscMPIInt    *rhs_recvcounts,*rhs_disps;
167:   PetscScalar    *rhs_loc,*rhs_recvbuf;
168: #endif
169:   Vec            b_seq,x_seq;
170:   PetscInt       ninfo,*info;           /* which INFO to display */
171:   PetscInt       sizeredrhs;
172:   PetscScalar    *schur_sol;
173:   PetscInt       schur_sizesol;
174:   PetscMUMPSInt  *ia_alloc,*ja_alloc;   /* work arrays used for the CSR struct for sparse rhs */
175:   PetscInt64     cur_ilen,cur_jlen;     /* current len of ia_alloc[], ja_alloc[] */
176:   PetscErrorCode (*ConvertToTriples)(Mat,PetscInt,MatReuse,Mat_MUMPS*);

178:   /* stuff used by petsc/mumps OpenMP support*/
179:   PetscBool      use_petsc_omp_support;
180:   PetscOmpCtrl   omp_ctrl;              /* an OpenMP controler that blocked processes will release their CPU (MPI_Barrier does not have this guarantee) */
181:   MPI_Comm       petsc_comm,omp_comm;   /* petsc_comm is petsc matrix's comm */
182:   PetscInt64     *recvcount;            /* a collection of nnz on omp_master */
183:   PetscMPIInt    tag,omp_comm_size;
184:   PetscBool      is_omp_master;         /* is this rank the master of omp_comm */
185:   MPI_Request    *reqs;
186: };

188: /* Cast a 1-based CSR represented by (nrow, ia, ja) of type PetscInt to a CSR of type PetscMUMPSInt.
189:    Here, nrow is number of rows, ia[] is row pointer and ja[] is column indices.
190:  */
191: static PetscErrorCode PetscMUMPSIntCSRCast(Mat_MUMPS *mumps,PetscInt nrow,PetscInt *ia,PetscInt *ja,PetscMUMPSInt **ia_mumps,PetscMUMPSInt **ja_mumps,PetscMUMPSInt *nnz_mumps)
192: {
194:   PetscInt       nnz=ia[nrow]-1; /* mumps uses 1-based indices. Uses PetscInt instead of PetscInt64 since mumps only uses PetscMUMPSInt for rhs */

197: #if defined(PETSC_USE_64BIT_INDICES)
198:   {
199:     PetscInt i;
200:     if (nrow+1 > mumps->cur_ilen) { /* realloc ia_alloc/ja_alloc to fit ia/ja */
201:       PetscFree(mumps->ia_alloc);
202:       PetscMalloc1(nrow+1,&mumps->ia_alloc);
203:       mumps->cur_ilen = nrow+1;
204:     }
205:     if (nnz > mumps->cur_jlen) {
206:       PetscFree(mumps->ja_alloc);
207:       PetscMalloc1(nnz,&mumps->ja_alloc);
208:       mumps->cur_jlen = nnz;
209:     }
210:     for (i=0; i<nrow+1; i++) {PetscMUMPSIntCast(ia[i],&(mumps->ia_alloc[i]));}
211:     for (i=0; i<nnz; i++)    {PetscMUMPSIntCast(ja[i],&(mumps->ja_alloc[i]));}
212:     *ia_mumps = mumps->ia_alloc;
213:     *ja_mumps = mumps->ja_alloc;
214:   }
215: #else
216:   *ia_mumps = ia;
217:   *ja_mumps = ja;
218: #endif
219:   PetscMUMPSIntCast(nnz,nnz_mumps);
220:   return(0);
221: }

223: static PetscErrorCode MatMumpsResetSchur_Private(Mat_MUMPS* mumps)
224: {

228:   PetscFree(mumps->id.listvar_schur);
229:   PetscFree(mumps->id.redrhs);
230:   PetscFree(mumps->schur_sol);
231:   mumps->id.size_schur = 0;
232:   mumps->id.schur_lld  = 0;
233:   mumps->id.ICNTL(19)  = 0;
234:   return(0);
235: }

237: /* solve with rhs in mumps->id.redrhs and return in the same location */
238: static PetscErrorCode MatMumpsSolveSchur_Private(Mat F)
239: {
240:   Mat_MUMPS            *mumps=(Mat_MUMPS*)F->data;
241:   Mat                  S,B,X;
242:   MatFactorSchurStatus schurstatus;
243:   PetscInt             sizesol;
244:   PetscErrorCode       ierr;

247:   MatFactorFactorizeSchurComplement(F);
248:   MatFactorGetSchurComplement(F,&S,&schurstatus);
249:   MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&B);
250:   MatSetType(B,((PetscObject)S)->type_name);
251: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
252:   MatBindToCPU(B,S->boundtocpu);
253: #endif
254:   switch (schurstatus) {
255:   case MAT_FACTOR_SCHUR_FACTORED:
256:     MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&X);
257:     MatSetType(X,((PetscObject)S)->type_name);
258: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
259:     MatBindToCPU(X,S->boundtocpu);
260: #endif
261:     if (!mumps->id.ICNTL(9)) { /* transpose solve */
262:       MatMatSolveTranspose(S,B,X);
263:     } else {
264:       MatMatSolve(S,B,X);
265:     }
266:     break;
267:   case MAT_FACTOR_SCHUR_INVERTED:
268:     sizesol = mumps->id.nrhs*mumps->id.size_schur;
269:     if (!mumps->schur_sol || sizesol > mumps->schur_sizesol) {
270:       PetscFree(mumps->schur_sol);
271:       PetscMalloc1(sizesol,&mumps->schur_sol);
272:       mumps->schur_sizesol = sizesol;
273:     }
274:     MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,mumps->schur_sol,&X);
275:     MatSetType(X,((PetscObject)S)->type_name);
276: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
277:     MatBindToCPU(X,S->boundtocpu);
278: #endif
279:     MatProductCreateWithMat(S,B,NULL,X);
280:     if (!mumps->id.ICNTL(9)) { /* transpose solve */
281:       MatProductSetType(X,MATPRODUCT_AtB);
282:     } else {
283:       MatProductSetType(X,MATPRODUCT_AB);
284:     }
285:     MatProductSetFromOptions(X);
286:     MatProductSymbolic(X);
287:     MatProductNumeric(X);

289:     MatCopy(X,B,SAME_NONZERO_PATTERN);
290:     break;
291:   default:
292:     SETERRQ1(PetscObjectComm((PetscObject)F),PETSC_ERR_SUP,"Unhandled MatFactorSchurStatus %D",F->schur_status);
293:   }
294:   MatFactorRestoreSchurComplement(F,&S,schurstatus);
295:   MatDestroy(&B);
296:   MatDestroy(&X);
297:   return(0);
298: }

300: static PetscErrorCode MatMumpsHandleSchur_Private(Mat F, PetscBool expansion)
301: {
302:   Mat_MUMPS     *mumps=(Mat_MUMPS*)F->data;

306:   if (!mumps->id.ICNTL(19)) { /* do nothing when Schur complement has not been computed */
307:     return(0);
308:   }
309:   if (!expansion) { /* prepare for the condensation step */
310:     PetscInt sizeredrhs = mumps->id.nrhs*mumps->id.size_schur;
311:     /* allocate MUMPS internal array to store reduced right-hand sides */
312:     if (!mumps->id.redrhs || sizeredrhs > mumps->sizeredrhs) {
313:       PetscFree(mumps->id.redrhs);
314:       mumps->id.lredrhs = mumps->id.size_schur;
315:       PetscMalloc1(mumps->id.nrhs*mumps->id.lredrhs,&mumps->id.redrhs);
316:       mumps->sizeredrhs = mumps->id.nrhs*mumps->id.lredrhs;
317:     }
318:     mumps->id.ICNTL(26) = 1; /* condensation phase */
319:   } else { /* prepare for the expansion step */
320:     /* solve Schur complement (this has to be done by the MUMPS user, so basically us) */
321:     MatMumpsSolveSchur_Private(F);
322:     mumps->id.ICNTL(26) = 2; /* expansion phase */
323:     PetscMUMPS_c(mumps);
324:     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));
325:     /* restore defaults */
326:     mumps->id.ICNTL(26) = -1;
327:     /* free MUMPS internal array for redrhs if we have solved for multiple rhs in order to save memory space */
328:     if (mumps->id.nrhs > 1) {
329:       PetscFree(mumps->id.redrhs);
330:       mumps->id.lredrhs = 0;
331:       mumps->sizeredrhs = 0;
332:     }
333:   }
334:   return(0);
335: }

337: /*
338:   MatConvertToTriples_A_B - convert Petsc matrix to triples: row[nz], col[nz], val[nz]

340:   input:
341:     A       - matrix in aij,baij or sbaij format
342:     shift   - 0: C style output triple; 1: Fortran style output triple.
343:     reuse   - MAT_INITIAL_MATRIX: spaces are allocated and values are set for the triple
344:               MAT_REUSE_MATRIX:   only the values in v array are updated
345:   output:
346:     nnz     - dim of r, c, and v (number of local nonzero entries of A)
347:     r, c, v - row and col index, matrix values (matrix triples)

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

353:  */

355: PetscErrorCode MatConvertToTriples_seqaij_seqaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
356: {
357:   const PetscScalar *av;
358:   const PetscInt    *ai,*aj,*ajj,M=A->rmap->n;
359:   PetscInt64        nz,rnz,i,j,k;
360:   PetscErrorCode    ierr;
361:   PetscMUMPSInt     *row,*col;
362:   Mat_SeqAIJ        *aa=(Mat_SeqAIJ*)A->data;

365:   MatSeqAIJGetArrayRead(A,&av);
366:   mumps->val = (PetscScalar*)av;
367:   if (reuse == MAT_INITIAL_MATRIX) {
368:     nz   = aa->nz;
369:     ai   = aa->i;
370:     aj   = aa->j;
371:     PetscMalloc2(nz,&row,nz,&col);
372:     for (i=k=0; i<M; i++) {
373:       rnz = ai[i+1] - ai[i];
374:       ajj = aj + ai[i];
375:       for (j=0; j<rnz; j++) {
376:         PetscMUMPSIntCast(i+shift,&row[k]);
377:         PetscMUMPSIntCast(ajj[j] + shift,&col[k]);
378:         k++;
379:       }
380:     }
381:     mumps->irn = row;
382:     mumps->jcn = col;
383:     mumps->nnz = nz;
384:   }
385:   MatSeqAIJRestoreArrayRead(A,&av);
386:   return(0);
387: }

389: PetscErrorCode MatConvertToTriples_seqsell_seqaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
390: {
392:   PetscInt64     nz,i,j,k,r;
393:   Mat_SeqSELL    *a=(Mat_SeqSELL*)A->data;
394:   PetscMUMPSInt  *row,*col;

397:   mumps->val = a->val;
398:   if (reuse == MAT_INITIAL_MATRIX) {
399:     nz   = a->sliidx[a->totalslices];
400:     PetscMalloc2(nz,&row,nz,&col);
401:     for (i=k=0; i<a->totalslices; i++) {
402:       for (j=a->sliidx[i],r=0; j<a->sliidx[i+1]; j++,r=((r+1)&0x07)) {
403:         PetscMUMPSIntCast(8*i+r+shift,&row[k++]);
404:       }
405:     }
406:     for (i=0;i<nz;i++) {PetscMUMPSIntCast(a->colidx[i]+shift,&col[i]);}
407:     mumps->irn = row;
408:     mumps->jcn = col;
409:     mumps->nnz = nz;
410:   }
411:   return(0);
412: }

414: PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
415: {
416:   Mat_SeqBAIJ    *aa=(Mat_SeqBAIJ*)A->data;
417:   const PetscInt *ai,*aj,*ajj,bs2 = aa->bs2;
418:   PetscInt64     M,nz,idx=0,rnz,i,j,k,m;
419:   PetscInt       bs;
421:   PetscMUMPSInt  *row,*col;

424:   MatGetBlockSize(A,&bs);
425:   M          = A->rmap->N/bs;
426:   mumps->val = aa->a;
427:   if (reuse == MAT_INITIAL_MATRIX) {
428:     ai   = aa->i; aj = aa->j;
429:     nz   = bs2*aa->nz;
430:     PetscMalloc2(nz,&row,nz,&col);
431:     for (i=0; i<M; i++) {
432:       ajj = aj + ai[i];
433:       rnz = ai[i+1] - ai[i];
434:       for (k=0; k<rnz; k++) {
435:         for (j=0; j<bs; j++) {
436:           for (m=0; m<bs; m++) {
437:             PetscMUMPSIntCast(i*bs + m + shift,&row[idx]);
438:             PetscMUMPSIntCast(bs*ajj[k] + j + shift,&col[idx]);
439:             idx++;
440:           }
441:         }
442:       }
443:     }
444:     mumps->irn = row;
445:     mumps->jcn = col;
446:     mumps->nnz = nz;
447:   }
448:   return(0);
449: }

451: PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
452: {
453:   const PetscInt *ai, *aj,*ajj;
454:   PetscInt        bs;
455:   PetscInt64      nz,rnz,i,j,k,m;
456:   PetscErrorCode  ierr;
457:   PetscMUMPSInt   *row,*col;
458:   PetscScalar     *val;
459:   Mat_SeqSBAIJ    *aa=(Mat_SeqSBAIJ*)A->data;
460:   const PetscInt  bs2=aa->bs2,mbs=aa->mbs;
461: #if defined(PETSC_USE_COMPLEX)
462:   PetscBool       hermitian;
463: #endif

466: #if defined(PETSC_USE_COMPLEX)
467:   MatGetOption(A,MAT_HERMITIAN,&hermitian);
468:   if (hermitian) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MUMPS does not support Hermitian symmetric matrices for Choleksy");
469: #endif
470:   ai   = aa->i;
471:   aj   = aa->j;
472:   MatGetBlockSize(A,&bs);
473:   if (reuse == MAT_INITIAL_MATRIX) {
474:     nz   = aa->nz;
475:     PetscMalloc2(bs2*nz,&row,bs2*nz,&col);
476:     if (bs>1) {
477:       PetscMalloc1(bs2*nz,&mumps->val_alloc);
478:       mumps->val = mumps->val_alloc;
479:     } else {
480:       mumps->val = aa->a;
481:     }
482:     mumps->irn = row;
483:     mumps->jcn = col;
484:   } else {
485:     if (bs == 1) mumps->val = aa->a;
486:     row = mumps->irn;
487:     col = mumps->jcn;
488:   }
489:   val = mumps->val;

491:   nz = 0;
492:   if (bs>1) {
493:     for (i=0; i<mbs; i++) {
494:       rnz = ai[i+1] - ai[i];
495:       ajj = aj + ai[i];
496:       for (j=0; j<rnz; j++) {
497:         for (k=0; k<bs; k++) {
498:           for (m=0; m<bs; m++) {
499:             if (ajj[j]>i || k>=m) {
500:               if (reuse == MAT_INITIAL_MATRIX) {
501:                 PetscMUMPSIntCast(i*bs + m + shift,&row[nz]);
502:                 PetscMUMPSIntCast(ajj[j]*bs + k + shift,&col[nz]);
503:               }
504:               val[nz++] = aa->a[(ai[i]+j)*bs2 + m + k*bs];
505:             }
506:           }
507:         }
508:       }
509:     }
510:   } else if (reuse == MAT_INITIAL_MATRIX) {
511:     for (i=0; i<mbs; i++) {
512:       rnz = ai[i+1] - ai[i];
513:       ajj = aj + ai[i];
514:       for (j=0; j<rnz; j++) {
515:         PetscMUMPSIntCast(i+shift,&row[nz]);
516:         PetscMUMPSIntCast(ajj[j] + shift,&col[nz]);
517:         nz++;
518:       }
519:     }
520:     if (nz != aa->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Different numbers of nonzeros %D != %D",nz,aa->nz);
521:   }
522:   if (reuse == MAT_INITIAL_MATRIX) mumps->nnz = nz;
523:   return(0);
524: }

526: PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
527: {
528:   const PetscInt    *ai,*aj,*ajj,*adiag,M=A->rmap->n;
529:   PetscInt64        nz,rnz,i,j;
530:   const PetscScalar *av,*v1;
531:   PetscScalar       *val;
532:   PetscErrorCode    ierr;
533:   PetscMUMPSInt     *row,*col;
534:   Mat_SeqAIJ        *aa=(Mat_SeqAIJ*)A->data;
535:   PetscBool         missing;
536: #if defined(PETSC_USE_COMPLEX)
537:   PetscBool         hermitian;
538: #endif

541: #if defined(PETSC_USE_COMPLEX)
542:   MatGetOption(A,MAT_HERMITIAN,&hermitian);
543:   if (hermitian) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MUMPS does not support Hermitian symmetric matrices for Choleksy");
544: #endif
545:   MatSeqAIJGetArrayRead(A,&av);
546:   ai    = aa->i; aj = aa->j;
547:   adiag = aa->diag;
548:   MatMissingDiagonal_SeqAIJ(A,&missing,NULL);
549:   if (reuse == MAT_INITIAL_MATRIX) {
550:     /* count nz in the upper triangular part of A */
551:     nz = 0;
552:     if (missing) {
553:       for (i=0; i<M; i++) {
554:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
555:           for (j=ai[i];j<ai[i+1];j++) {
556:             if (aj[j] < i) continue;
557:             nz++;
558:           }
559:         } else {
560:           nz += ai[i+1] - adiag[i];
561:         }
562:       }
563:     } else {
564:       for (i=0; i<M; i++) nz += ai[i+1] - adiag[i];
565:     }
566:     PetscMalloc2(nz,&row,nz,&col);
567:     PetscMalloc1(nz,&val);
568:     mumps->nnz = nz;
569:     mumps->irn = row;
570:     mumps->jcn = col;
571:     mumps->val = mumps->val_alloc = val;

573:     nz = 0;
574:     if (missing) {
575:       for (i=0; i<M; i++) {
576:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
577:           for (j=ai[i];j<ai[i+1];j++) {
578:             if (aj[j] < i) continue;
579:             PetscMUMPSIntCast(i+shift,&row[nz]);
580:             PetscMUMPSIntCast(aj[j]+shift,&col[nz]);
581:             val[nz] = av[j];
582:             nz++;
583:           }
584:         } else {
585:           rnz = ai[i+1] - adiag[i];
586:           ajj = aj + adiag[i];
587:           v1  = av + adiag[i];
588:           for (j=0; j<rnz; j++) {
589:             PetscMUMPSIntCast(i+shift,&row[nz]);
590:             PetscMUMPSIntCast(ajj[j] + shift,&col[nz]);
591:             val[nz++] = v1[j];
592:           }
593:         }
594:       }
595:     } else {
596:       for (i=0; i<M; i++) {
597:         rnz = ai[i+1] - adiag[i];
598:         ajj = aj + adiag[i];
599:         v1  = av + adiag[i];
600:         for (j=0; j<rnz; j++) {
601:           PetscMUMPSIntCast(i+shift,&row[nz]);
602:           PetscMUMPSIntCast(ajj[j] + shift,&col[nz]);
603:           val[nz++] = v1[j];
604:         }
605:       }
606:     }
607:   } else {
608:     nz = 0;
609:     val = mumps->val;
610:     if (missing) {
611:       for (i=0; i <M; i++) {
612:         if (PetscUnlikely(adiag[i] >= ai[i+1])) {
613:           for (j=ai[i];j<ai[i+1];j++) {
614:             if (aj[j] < i) continue;
615:             val[nz++] = av[j];
616:           }
617:         } else {
618:           rnz = ai[i+1] - adiag[i];
619:           v1  = av + adiag[i];
620:           for (j=0; j<rnz; j++) {
621:             val[nz++] = v1[j];
622:           }
623:         }
624:       }
625:     } else {
626:       for (i=0; i <M; i++) {
627:         rnz = ai[i+1] - adiag[i];
628:         v1  = av + adiag[i];
629:         for (j=0; j<rnz; j++) {
630:           val[nz++] = v1[j];
631:         }
632:       }
633:     }
634:   }
635:   MatSeqAIJRestoreArrayRead(A,&av);
636:   return(0);
637: }

639: PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
640: {
641:   PetscErrorCode    ierr;
642:   const PetscInt    *ai,*aj,*bi,*bj,*garray,*ajj,*bjj;
643:   PetscInt          bs;
644:   PetscInt64        rstart,nz,i,j,k,m,jj,irow,countA,countB;
645:   PetscMUMPSInt     *row,*col;
646:   const PetscScalar *av,*bv,*v1,*v2;
647:   PetscScalar       *val;
648:   Mat_MPISBAIJ      *mat = (Mat_MPISBAIJ*)A->data;
649:   Mat_SeqSBAIJ      *aa  = (Mat_SeqSBAIJ*)(mat->A)->data;
650:   Mat_SeqBAIJ       *bb  = (Mat_SeqBAIJ*)(mat->B)->data;
651:   const PetscInt    bs2=aa->bs2,mbs=aa->mbs;
652: #if defined(PETSC_USE_COMPLEX)
653:   PetscBool         hermitian;
654: #endif

657: #if defined(PETSC_USE_COMPLEX)
658:   MatGetOption(A,MAT_HERMITIAN,&hermitian);
659:   if (hermitian) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MUMPS does not support Hermitian symmetric matrices for Choleksy");
660: #endif
661:   MatGetBlockSize(A,&bs);
662:   rstart = A->rmap->rstart;
663:   ai = aa->i;
664:   aj = aa->j;
665:   bi = bb->i;
666:   bj = bb->j;
667:   av = aa->a;
668:   bv = bb->a;

670:   garray = mat->garray;

672:   if (reuse == MAT_INITIAL_MATRIX) {
673:     nz   = (aa->nz+bb->nz)*bs2; /* just a conservative estimate */
674:     PetscMalloc2(nz,&row,nz,&col);
675:     PetscMalloc1(nz,&val);
676:     /* can not decide the exact mumps->nnz now because of the SBAIJ */
677:     mumps->irn = row;
678:     mumps->jcn = col;
679:     mumps->val = mumps->val_alloc = val;
680:   } else {
681:     val = mumps->val;
682:   }

684:   jj = 0; irow = rstart;
685:   for (i=0; i<mbs; i++) {
686:     ajj    = aj + ai[i];                 /* ptr to the beginning of this row */
687:     countA = ai[i+1] - ai[i];
688:     countB = bi[i+1] - bi[i];
689:     bjj    = bj + bi[i];
690:     v1     = av + ai[i]*bs2;
691:     v2     = bv + bi[i]*bs2;

693:     if (bs>1) {
694:       /* A-part */
695:       for (j=0; j<countA; j++) {
696:         for (k=0; k<bs; k++) {
697:           for (m=0; m<bs; m++) {
698:             if (rstart + ajj[j]*bs>irow || k>=m) {
699:               if (reuse == MAT_INITIAL_MATRIX) {
700:                 PetscMUMPSIntCast(irow + m + shift,&row[jj]);
701:                 PetscMUMPSIntCast(rstart + ajj[j]*bs + k + shift,&col[jj]);
702:               }
703:               val[jj++] = v1[j*bs2 + m + k*bs];
704:             }
705:           }
706:         }
707:       }

709:       /* B-part */
710:       for (j=0; j < countB; j++) {
711:         for (k=0; k<bs; k++) {
712:           for (m=0; m<bs; m++) {
713:             if (reuse == MAT_INITIAL_MATRIX) {
714:               PetscMUMPSIntCast(irow + m + shift,&row[jj]);
715:               PetscMUMPSIntCast(garray[bjj[j]]*bs + k + shift,&col[jj]);
716:             }
717:             val[jj++] = v2[j*bs2 + m + k*bs];
718:           }
719:         }
720:       }
721:     } else {
722:       /* A-part */
723:       for (j=0; j<countA; j++) {
724:         if (reuse == MAT_INITIAL_MATRIX) {
725:           PetscMUMPSIntCast(irow + shift,&row[jj]);
726:           PetscMUMPSIntCast(rstart + ajj[j] + shift,&col[jj]);
727:         }
728:         val[jj++] = v1[j];
729:       }

731:       /* B-part */
732:       for (j=0; j < countB; j++) {
733:         if (reuse == MAT_INITIAL_MATRIX) {
734:           PetscMUMPSIntCast(irow + shift,&row[jj]);
735:           PetscMUMPSIntCast(garray[bjj[j]] + shift,&col[jj]);
736:         }
737:         val[jj++] = v2[j];
738:       }
739:     }
740:     irow+=bs;
741:   }
742:   mumps->nnz = jj;
743:   return(0);
744: }

746: PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
747: {
748:   const PetscInt    *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
749:   PetscErrorCode    ierr;
750:   PetscInt64        rstart,nz,i,j,jj,irow,countA,countB;
751:   PetscMUMPSInt     *row,*col;
752:   const PetscScalar *av, *bv,*v1,*v2;
753:   PetscScalar       *val;
754:   Mat               Ad,Ao;
755:   Mat_SeqAIJ        *aa;
756:   Mat_SeqAIJ        *bb;

759:   MatMPIAIJGetSeqAIJ(A,&Ad,&Ao,&garray);
760:   MatSeqAIJGetArrayRead(Ad,&av);
761:   MatSeqAIJGetArrayRead(Ao,&bv);

763:   aa = (Mat_SeqAIJ*)(Ad)->data;
764:   bb = (Mat_SeqAIJ*)(Ao)->data;
765:   ai = aa->i;
766:   aj = aa->j;
767:   bi = bb->i;
768:   bj = bb->j;

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

772:   if (reuse == MAT_INITIAL_MATRIX) {
773:     nz   = (PetscInt64)aa->nz + bb->nz; /* make sure the sum won't overflow PetscInt */
774:     PetscMalloc2(nz,&row,nz,&col);
775:     PetscMalloc1(nz,&val);
776:     mumps->nnz = nz;
777:     mumps->irn = row;
778:     mumps->jcn = col;
779:     mumps->val = mumps->val_alloc = val;
780:   } else {
781:     val = mumps->val;
782:   }

784:   jj = 0; irow = rstart;
785:   for (i=0; i<m; i++) {
786:     ajj    = aj + ai[i];                 /* ptr to the beginning of this row */
787:     countA = ai[i+1] - ai[i];
788:     countB = bi[i+1] - bi[i];
789:     bjj    = bj + bi[i];
790:     v1     = av + ai[i];
791:     v2     = bv + bi[i];

793:     /* A-part */
794:     for (j=0; j<countA; j++) {
795:       if (reuse == MAT_INITIAL_MATRIX) {
796:         PetscMUMPSIntCast(irow + shift,&row[jj]);
797:         PetscMUMPSIntCast(rstart + ajj[j] + shift,&col[jj]);
798:       }
799:       val[jj++] = v1[j];
800:     }

802:     /* B-part */
803:     for (j=0; j < countB; j++) {
804:       if (reuse == MAT_INITIAL_MATRIX) {
805:         PetscMUMPSIntCast(irow + shift,&row[jj]);
806:         PetscMUMPSIntCast(garray[bjj[j]] + shift,&col[jj]);
807:       }
808:       val[jj++] = v2[j];
809:     }
810:     irow++;
811:   }
812:   MatSeqAIJRestoreArrayRead(Ad,&av);
813:   MatSeqAIJRestoreArrayRead(Ao,&bv);
814:   return(0);
815: }

817: PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
818: {
819:   Mat_MPIBAIJ       *mat    = (Mat_MPIBAIJ*)A->data;
820:   Mat_SeqBAIJ       *aa     = (Mat_SeqBAIJ*)(mat->A)->data;
821:   Mat_SeqBAIJ       *bb     = (Mat_SeqBAIJ*)(mat->B)->data;
822:   const PetscInt    *ai     = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j,*ajj, *bjj;
823:   const PetscInt    *garray = mat->garray,mbs=mat->mbs,rstart=A->rmap->rstart;
824:   const PetscInt    bs2=mat->bs2;
825:   PetscErrorCode    ierr;
826:   PetscInt          bs;
827:   PetscInt64        nz,i,j,k,n,jj,irow,countA,countB,idx;
828:   PetscMUMPSInt     *row,*col;
829:   const PetscScalar *av=aa->a, *bv=bb->a,*v1,*v2;
830:   PetscScalar       *val;

833:   MatGetBlockSize(A,&bs);
834:   if (reuse == MAT_INITIAL_MATRIX) {
835:     nz   = bs2*(aa->nz + bb->nz);
836:     PetscMalloc2(nz,&row,nz,&col);
837:     PetscMalloc1(nz,&val);
838:     mumps->nnz = nz;
839:     mumps->irn = row;
840:     mumps->jcn = col;
841:     mumps->val = mumps->val_alloc = val;
842:   } else {
843:     val = mumps->val;
844:   }

846:   jj = 0; irow = rstart;
847:   for (i=0; i<mbs; i++) {
848:     countA = ai[i+1] - ai[i];
849:     countB = bi[i+1] - bi[i];
850:     ajj    = aj + ai[i];
851:     bjj    = bj + bi[i];
852:     v1     = av + bs2*ai[i];
853:     v2     = bv + bs2*bi[i];

855:     idx = 0;
856:     /* A-part */
857:     for (k=0; k<countA; k++) {
858:       for (j=0; j<bs; j++) {
859:         for (n=0; n<bs; n++) {
860:           if (reuse == MAT_INITIAL_MATRIX) {
861:             PetscMUMPSIntCast(irow + n + shift,&row[jj]);
862:             PetscMUMPSIntCast(rstart + bs*ajj[k] + j + shift,&col[jj]);
863:           }
864:           val[jj++] = v1[idx++];
865:         }
866:       }
867:     }

869:     idx = 0;
870:     /* B-part */
871:     for (k=0; k<countB; k++) {
872:       for (j=0; j<bs; j++) {
873:         for (n=0; n<bs; n++) {
874:           if (reuse == MAT_INITIAL_MATRIX) {
875:             PetscMUMPSIntCast(irow + n + shift,&row[jj]);
876:             PetscMUMPSIntCast(bs*garray[bjj[k]] + j + shift,&col[jj]);
877:           }
878:           val[jj++] = v2[idx++];
879:         }
880:       }
881:     }
882:     irow += bs;
883:   }
884:   return(0);
885: }

887: PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A,PetscInt shift,MatReuse reuse,Mat_MUMPS *mumps)
888: {
889:   const PetscInt    *ai, *aj,*adiag, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
890:   PetscErrorCode    ierr;
891:   PetscInt64        rstart,nz,nza,nzb,i,j,jj,irow,countA,countB;
892:   PetscMUMPSInt     *row,*col;
893:   const PetscScalar *av, *bv,*v1,*v2;
894:   PetscScalar       *val;
895:   Mat               Ad,Ao;
896:   Mat_SeqAIJ        *aa;
897:   Mat_SeqAIJ        *bb;
898: #if defined(PETSC_USE_COMPLEX)
899:   PetscBool         hermitian;
900: #endif

903: #if defined(PETSC_USE_COMPLEX)
904:   MatGetOption(A,MAT_HERMITIAN,&hermitian);
905:   if (hermitian) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MUMPS does not support Hermitian symmetric matrices for Choleksy");
906: #endif
907:   MatMPIAIJGetSeqAIJ(A,&Ad,&Ao,&garray);
908:   MatSeqAIJGetArrayRead(Ad,&av);
909:   MatSeqAIJGetArrayRead(Ao,&bv);

911:   aa    = (Mat_SeqAIJ*)(Ad)->data;
912:   bb    = (Mat_SeqAIJ*)(Ao)->data;
913:   ai    = aa->i;
914:   aj    = aa->j;
915:   adiag = aa->diag;
916:   bi    = bb->i;
917:   bj    = bb->j;

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

921:   if (reuse == MAT_INITIAL_MATRIX) {
922:     nza = 0;    /* num of upper triangular entries in mat->A, including diagonals */
923:     nzb = 0;    /* num of upper triangular entries in mat->B */
924:     for (i=0; i<m; i++) {
925:       nza   += (ai[i+1] - adiag[i]);
926:       countB = bi[i+1] - bi[i];
927:       bjj    = bj + bi[i];
928:       for (j=0; j<countB; j++) {
929:         if (garray[bjj[j]] > rstart) nzb++;
930:       }
931:     }

933:     nz   = nza + nzb; /* total nz of upper triangular part of mat */
934:     PetscMalloc2(nz,&row,nz,&col);
935:     PetscMalloc1(nz,&val);
936:     mumps->nnz = nz;
937:     mumps->irn = row;
938:     mumps->jcn = col;
939:     mumps->val = mumps->val_alloc = val;
940:   } else {
941:     val = mumps->val;
942:   }

944:   jj = 0; irow = rstart;
945:   for (i=0; i<m; i++) {
946:     ajj    = aj + adiag[i];                 /* ptr to the beginning of the diagonal of this row */
947:     v1     = av + adiag[i];
948:     countA = ai[i+1] - adiag[i];
949:     countB = bi[i+1] - bi[i];
950:     bjj    = bj + bi[i];
951:     v2     = bv + bi[i];

953:     /* A-part */
954:     for (j=0; j<countA; j++) {
955:       if (reuse == MAT_INITIAL_MATRIX) {
956:         PetscMUMPSIntCast(irow + shift,&row[jj]);
957:         PetscMUMPSIntCast(rstart + ajj[j] + shift,&col[jj]);
958:       }
959:       val[jj++] = v1[j];
960:     }

962:     /* B-part */
963:     for (j=0; j < countB; j++) {
964:       if (garray[bjj[j]] > rstart) {
965:         if (reuse == MAT_INITIAL_MATRIX) {
966:           PetscMUMPSIntCast(irow + shift,&row[jj]);
967:           PetscMUMPSIntCast(garray[bjj[j]] + shift,&col[jj]);
968:         }
969:         val[jj++] = v2[j];
970:       }
971:     }
972:     irow++;
973:   }
974:   MatSeqAIJRestoreArrayRead(Ad,&av);
975:   MatSeqAIJRestoreArrayRead(Ao,&bv);
976:   return(0);
977: }

979: PetscErrorCode MatDestroy_MUMPS(Mat A)
980: {
982:   Mat_MUMPS      *mumps=(Mat_MUMPS*)A->data;

985:   PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
986:   VecScatterDestroy(&mumps->scat_rhs);
987:   VecScatterDestroy(&mumps->scat_sol);
988:   VecDestroy(&mumps->b_seq);
989:   VecDestroy(&mumps->x_seq);
990:   PetscFree(mumps->id.perm_in);
991:   PetscFree2(mumps->irn,mumps->jcn);
992:   PetscFree(mumps->val_alloc);
993:   PetscFree(mumps->info);
994:   MatMumpsResetSchur_Private(mumps);
995:   mumps->id.job = JOB_END;
996:   PetscMUMPS_c(mumps);
997:   if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in MatDestroy_MUMPS: INFOG(1)=%d\n",mumps->id.INFOG(1));
998: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
999:   if (mumps->use_petsc_omp_support) {
1000:     PetscOmpCtrlDestroy(&mumps->omp_ctrl);
1001:     PetscFree2(mumps->rhs_loc,mumps->rhs_recvbuf);
1002:     PetscFree3(mumps->rhs_nrow,mumps->rhs_recvcounts,mumps->rhs_disps);
1003:   }
1004: #endif
1005:   PetscFree(mumps->ia_alloc);
1006:   PetscFree(mumps->ja_alloc);
1007:   PetscFree(mumps->recvcount);
1008:   PetscFree(mumps->reqs);
1009:   PetscFree(mumps->irhs_loc);
1010:   if (mumps->mumps_comm != MPI_COMM_NULL) {MPI_Comm_free(&mumps->mumps_comm);}
1011:   PetscFree(A->data);

1013:   /* clear composed functions */
1014:   PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);
1015:   PetscObjectComposeFunction((PetscObject)A,"MatFactorSetSchurIS_C",NULL);
1016:   PetscObjectComposeFunction((PetscObject)A,"MatFactorCreateSchurComplement_C",NULL);
1017:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetIcntl_C",NULL);
1018:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetIcntl_C",NULL);
1019:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetCntl_C",NULL);
1020:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetCntl_C",NULL);
1021:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfo_C",NULL);
1022:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfog_C",NULL);
1023:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfo_C",NULL);
1024:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfog_C",NULL);
1025:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverse_C",NULL);
1026:   PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverseTranspose_C",NULL);
1027:   return(0);
1028: }

1030: /* Set up the distributed RHS info for MUMPS. <nrhs> is the number of RHS. <array> points to start of RHS on the local processor. */
1031: static PetscErrorCode MatMumpsSetUpDistRHSInfo(Mat A,PetscInt nrhs,const PetscScalar *array)
1032: {
1033:   PetscErrorCode     ierr;
1034:   Mat_MUMPS          *mumps=(Mat_MUMPS*)A->data;
1035:   const PetscMPIInt  ompsize=mumps->omp_comm_size;
1036:   PetscInt           i,m,M,rstart;

1039:   MatGetSize(A,&M,NULL);
1040:   MatGetLocalSize(A,&m,NULL);
1041:   if (M > PETSC_MUMPS_INT_MAX) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"PetscInt too long for PetscMUMPSInt");
1042:   if (ompsize == 1) {
1043:     if (!mumps->irhs_loc) {
1044:       mumps->nloc_rhs = m;
1045:       PetscMalloc1(m,&mumps->irhs_loc);
1046:       MatGetOwnershipRange(A,&rstart,NULL);
1047:       for (i=0; i<m; i++) mumps->irhs_loc[i] = rstart+i+1; /* use 1-based indices */
1048:     }
1049:     mumps->id.rhs_loc = (MumpsScalar*)array;
1050:   } else {
1051:   #if defined(PETSC_HAVE_OPENMP_SUPPORT)
1052:     const PetscInt  *ranges;
1053:     PetscMPIInt     j,k,sendcount,*petsc_ranks,*omp_ranks;
1054:     MPI_Group       petsc_group,omp_group;
1055:     PetscScalar     *recvbuf=NULL;

1057:     if (mumps->is_omp_master) {
1058:       /* Lazily initialize the omp stuff for distributed rhs */
1059:       if (!mumps->irhs_loc) {
1060:         PetscMalloc2(ompsize,&omp_ranks,ompsize,&petsc_ranks);
1061:         PetscMalloc3(ompsize,&mumps->rhs_nrow,ompsize,&mumps->rhs_recvcounts,ompsize,&mumps->rhs_disps);
1062:         MPI_Comm_group(mumps->petsc_comm,&petsc_group);
1063:         MPI_Comm_group(mumps->omp_comm,&omp_group);
1064:         for (j=0; j<ompsize; j++) omp_ranks[j] = j;
1065:         MPI_Group_translate_ranks(omp_group,ompsize,omp_ranks,petsc_group,petsc_ranks);

1067:         /* Populate mumps->irhs_loc[], rhs_nrow[] */
1068:         mumps->nloc_rhs = 0;
1069:         MatGetOwnershipRanges(A,&ranges);
1070:         for (j=0; j<ompsize; j++) {
1071:           mumps->rhs_nrow[j] = ranges[petsc_ranks[j]+1] - ranges[petsc_ranks[j]];
1072:           mumps->nloc_rhs   += mumps->rhs_nrow[j];
1073:         }
1074:         PetscMalloc1(mumps->nloc_rhs,&mumps->irhs_loc);
1075:         for (j=k=0; j<ompsize; j++) {
1076:           for (i=ranges[petsc_ranks[j]]; i<ranges[petsc_ranks[j]+1]; i++,k++) mumps->irhs_loc[k] = i+1; /* uses 1-based indices */
1077:         }

1079:         PetscFree2(omp_ranks,petsc_ranks);
1080:         MPI_Group_free(&petsc_group);
1081:         MPI_Group_free(&omp_group);
1082:       }

1084:       /* Realloc buffers when current nrhs is bigger than what we have met */
1085:       if (nrhs > mumps->max_nrhs) {
1086:         PetscFree2(mumps->rhs_loc,mumps->rhs_recvbuf);
1087:         PetscMalloc2(mumps->nloc_rhs*nrhs,&mumps->rhs_loc,mumps->nloc_rhs*nrhs,&mumps->rhs_recvbuf);
1088:         mumps->max_nrhs = nrhs;
1089:       }

1091:       /* Setup recvcounts[], disps[], recvbuf on omp rank 0 for the upcoming MPI_Gatherv */
1092:       for (j=0; j<ompsize; j++) {PetscMPIIntCast(mumps->rhs_nrow[j]*nrhs,&mumps->rhs_recvcounts[j]);}
1093:       mumps->rhs_disps[0] = 0;
1094:       for (j=1; j<ompsize; j++) {
1095:         mumps->rhs_disps[j] = mumps->rhs_disps[j-1] + mumps->rhs_recvcounts[j-1];
1096:         if (mumps->rhs_disps[j] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"PetscMPIInt overflow!");
1097:       }
1098:       recvbuf = (nrhs == 1) ? mumps->rhs_loc : mumps->rhs_recvbuf; /* Directly use rhs_loc[] as recvbuf. Single rhs is common in Ax=b */
1099:     }

1101:     PetscMPIIntCast(m*nrhs,&sendcount);
1102:     MPI_Gatherv(array,sendcount,MPIU_SCALAR,recvbuf,mumps->rhs_recvcounts,mumps->rhs_disps,MPIU_SCALAR,0,mumps->omp_comm);

1104:     if (mumps->is_omp_master) {
1105:       if (nrhs > 1) { /* Copy & re-arrange data from rhs_recvbuf[] to mumps->rhs_loc[] only when there are multiple rhs */
1106:         PetscScalar *dst,*dstbase = mumps->rhs_loc;
1107:         for (j=0; j<ompsize; j++) {
1108:           const PetscScalar *src = mumps->rhs_recvbuf + mumps->rhs_disps[j];
1109:           dst = dstbase;
1110:           for (i=0; i<nrhs; i++) {
1111:             PetscArraycpy(dst,src,mumps->rhs_nrow[j]);
1112:             src += mumps->rhs_nrow[j];
1113:             dst += mumps->nloc_rhs;
1114:           }
1115:           dstbase += mumps->rhs_nrow[j];
1116:         }
1117:       }
1118:       mumps->id.rhs_loc = (MumpsScalar*)mumps->rhs_loc;
1119:     }
1120:   #endif /* PETSC_HAVE_OPENMP_SUPPORT */
1121:   }
1122:   mumps->id.nrhs     = nrhs;
1123:   mumps->id.nloc_rhs = mumps->nloc_rhs;
1124:   mumps->id.lrhs_loc = mumps->nloc_rhs;
1125:   mumps->id.irhs_loc = mumps->irhs_loc;
1126:   return(0);
1127: }

1129: PetscErrorCode MatSolve_MUMPS(Mat A,Vec b,Vec x)
1130: {
1131:   Mat_MUMPS          *mumps=(Mat_MUMPS*)A->data;
1132:   const PetscScalar  *rarray = NULL;
1133:   PetscScalar        *array;
1134:   IS                 is_iden,is_petsc;
1135:   PetscErrorCode     ierr;
1136:   PetscInt           i;
1137:   PetscBool          second_solve = PETSC_FALSE;
1138:   static PetscBool   cite1 = PETSC_FALSE,cite2 = PETSC_FALSE;

1141:   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);
1142:   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);

1144:   if (A->factorerrortype) {
1145:     PetscInfo2(A,"MatSolve is called with singular matrix factor, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1146:     VecSetInf(x);
1147:     return(0);
1148:   }

1150:   mumps->id.nrhs = 1;
1151:   if (mumps->petsc_size > 1) {
1152:     if (mumps->ICNTL20 == 10) {
1153:       mumps->id.ICNTL(20) = 10; /* dense distributed RHS */
1154:       VecGetArrayRead(b,&rarray);
1155:       MatMumpsSetUpDistRHSInfo(A,1,rarray);
1156:     } else {
1157:       mumps->id.ICNTL(20) = 0; /* dense centralized RHS; Scatter b into a sequential rhs vector*/
1158:       VecScatterBegin(mumps->scat_rhs,b,mumps->b_seq,INSERT_VALUES,SCATTER_FORWARD);
1159:       VecScatterEnd(mumps->scat_rhs,b,mumps->b_seq,INSERT_VALUES,SCATTER_FORWARD);
1160:       if (!mumps->myid) {
1161:         VecGetArray(mumps->b_seq,&array);
1162:         mumps->id.rhs = (MumpsScalar*)array;
1163:       }
1164:     }
1165:   } else {  /* petsc_size == 1 */
1166:     mumps->id.ICNTL(20) = 0; /* dense centralized RHS */
1167:     VecCopy(b,x);
1168:     VecGetArray(x,&array);
1169:     mumps->id.rhs = (MumpsScalar*)array;
1170:   }

1172:   /*
1173:      handle condensation step of Schur complement (if any)
1174:      We set by default ICNTL(26) == -1 when Schur indices have been provided by the user.
1175:      According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase
1176:      Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system.
1177:      This requires an extra call to PetscMUMPS_c and the computation of the factors for S
1178:   */
1179:   if (mumps->id.size_schur > 0 && (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2)) {
1180:     if (mumps->petsc_size > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n");
1181:     second_solve = PETSC_TRUE;
1182:     MatMumpsHandleSchur_Private(A,PETSC_FALSE);
1183:   }
1184:   /* solve phase */
1185:   /*-------------*/
1186:   mumps->id.job = JOB_SOLVE;
1187:   PetscMUMPS_c(mumps);
1188:   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));

1190:   /* handle expansion step of Schur complement (if any) */
1191:   if (second_solve) {
1192:     MatMumpsHandleSchur_Private(A,PETSC_TRUE);
1193:   }

1195:   if (mumps->petsc_size > 1) { /* convert mumps distributed solution to petsc mpi x */
1196:     if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) {
1197:       /* when id.ICNTL(9) changes, the contents of lsol_loc may change (not its size, lsol_loc), recreates scat_sol */
1198:       VecScatterDestroy(&mumps->scat_sol);
1199:     }
1200:     if (!mumps->scat_sol) { /* create scatter scat_sol */
1201:       PetscInt *isol2_loc=NULL;
1202:       ISCreateStride(PETSC_COMM_SELF,mumps->id.lsol_loc,0,1,&is_iden); /* from */
1203:       PetscMalloc1(mumps->id.lsol_loc,&isol2_loc);
1204:       for (i=0; i<mumps->id.lsol_loc; i++) isol2_loc[i] = mumps->id.isol_loc[i]-1; /* change Fortran style to C style */
1205:       ISCreateGeneral(PETSC_COMM_SELF,mumps->id.lsol_loc,isol2_loc,PETSC_OWN_POINTER,&is_petsc);  /* to */
1206:       VecScatterCreate(mumps->x_seq,is_iden,x,is_petsc,&mumps->scat_sol);
1207:       ISDestroy(&is_iden);
1208:       ISDestroy(&is_petsc);
1209:       mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */
1210:     }

1212:     VecScatterBegin(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
1213:     VecScatterEnd(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
1214:   }

1216:   if (mumps->petsc_size > 1) {
1217:     if (mumps->ICNTL20 == 10) {
1218:       VecRestoreArrayRead(b,&rarray);
1219:     } else if (!mumps->myid) {
1220:       VecRestoreArray(mumps->b_seq,&array);
1221:     }
1222:   } else {VecRestoreArray(x,&array);}

1224:   PetscLogFlops(2.0*mumps->id.RINFO(3));
1225:   return(0);
1226: }

1228: PetscErrorCode MatSolveTranspose_MUMPS(Mat A,Vec b,Vec x)
1229: {
1230:   Mat_MUMPS      *mumps=(Mat_MUMPS*)A->data;

1234:   mumps->id.ICNTL(9) = 0;
1235:   MatSolve_MUMPS(A,b,x);
1236:   mumps->id.ICNTL(9) = 1;
1237:   return(0);
1238: }

1240: PetscErrorCode MatMatSolve_MUMPS(Mat A,Mat B,Mat X)
1241: {
1242:   PetscErrorCode    ierr;
1243:   Mat               Bt = NULL;
1244:   PetscBool         denseX,denseB,flg,flgT;
1245:   Mat_MUMPS         *mumps=(Mat_MUMPS*)A->data;
1246:   PetscInt          i,nrhs,M;
1247:   PetscScalar       *array;
1248:   const PetscScalar *rbray;
1249:   PetscInt          lsol_loc,nlsol_loc,*idxx,iidx = 0;
1250:   PetscMUMPSInt     *isol_loc,*isol_loc_save;
1251:   PetscScalar       *bray,*sol_loc,*sol_loc_save;
1252:   IS                is_to,is_from;
1253:   PetscInt          k,proc,j,m,myrstart;
1254:   const PetscInt    *rstart;
1255:   Vec               v_mpi,msol_loc;
1256:   VecScatter        scat_sol;
1257:   Vec               b_seq;
1258:   VecScatter        scat_rhs;
1259:   PetscScalar       *aa;
1260:   PetscInt          spnr,*ia,*ja;
1261:   Mat_MPIAIJ        *b = NULL;

1264:   PetscObjectTypeCompareAny((PetscObject)X,&denseX,MATSEQDENSE,MATMPIDENSE,NULL);
1265:   if (!denseX) SETERRQ(PetscObjectComm((PetscObject)X),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix");

1267:   PetscObjectTypeCompareAny((PetscObject)B,&denseB,MATSEQDENSE,MATMPIDENSE,NULL);
1268:   if (denseB) {
1269:     if (B->rmap->n != X->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Matrix B and X must have same row distribution");
1270:     mumps->id.ICNTL(20)= 0; /* dense RHS */
1271:   } else { /* sparse B */
1272:     if (X == B) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_IDN,"X and B must be different matrices");
1273:     PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&flgT);
1274:     if (flgT) { /* input B is transpose of actural RHS matrix,
1275:                  because mumps requires sparse compressed COLUMN storage! See MatMatTransposeSolve_MUMPS() */
1276:       MatTransposeGetMat(B,&Bt);
1277:     } else SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATTRANSPOSEMAT matrix");
1278:     mumps->id.ICNTL(20)= 1; /* sparse RHS */
1279:   }

1281:   MatGetSize(B,&M,&nrhs);
1282:   mumps->id.nrhs = nrhs;
1283:   mumps->id.lrhs = M;
1284:   mumps->id.rhs  = NULL;

1286:   if (mumps->petsc_size == 1) {
1287:     PetscScalar *aa;
1288:     PetscInt    spnr,*ia,*ja;
1289:     PetscBool   second_solve = PETSC_FALSE;

1291:     MatDenseGetArray(X,&array);
1292:     mumps->id.rhs = (MumpsScalar*)array;

1294:     if (denseB) {
1295:       /* copy B to X */
1296:       MatDenseGetArrayRead(B,&rbray);
1297:       PetscArraycpy(array,rbray,M*nrhs);
1298:       MatDenseRestoreArrayRead(B,&rbray);
1299:     } else { /* sparse B */
1300:       MatSeqAIJGetArray(Bt,&aa);
1301:       MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1302:       if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
1303:       PetscMUMPSIntCSRCast(mumps,spnr,ia,ja,&mumps->id.irhs_ptr,&mumps->id.irhs_sparse,&mumps->id.nz_rhs);
1304:       mumps->id.rhs_sparse  = (MumpsScalar*)aa;
1305:     }
1306:     /* handle condensation step of Schur complement (if any) */
1307:     if (mumps->id.size_schur > 0 && (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2)) {
1308:       second_solve = PETSC_TRUE;
1309:       MatMumpsHandleSchur_Private(A,PETSC_FALSE);
1310:     }
1311:     /* solve phase */
1312:     /*-------------*/
1313:     mumps->id.job = JOB_SOLVE;
1314:     PetscMUMPS_c(mumps);
1315:     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));

1317:     /* handle expansion step of Schur complement (if any) */
1318:     if (second_solve) {
1319:       MatMumpsHandleSchur_Private(A,PETSC_TRUE);
1320:     }
1321:     if (!denseB) { /* sparse B */
1322:       MatSeqAIJRestoreArray(Bt,&aa);
1323:       MatRestoreRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1324:       if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure");
1325:     }
1326:     MatDenseRestoreArray(X,&array);
1327:     return(0);
1328:   }

1330:   /*--------- parallel case: MUMPS requires rhs B to be centralized on the host! --------*/
1331:   if (mumps->petsc_size > 1 && mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n");

1333:   /* create msol_loc to hold mumps local solution */
1334:   isol_loc_save = mumps->id.isol_loc; /* save it for MatSolve() */
1335:   sol_loc_save  = (PetscScalar*)mumps->id.sol_loc;

1337:   lsol_loc  = mumps->id.lsol_loc;
1338:   nlsol_loc = nrhs*lsol_loc;     /* length of sol_loc */
1339:   PetscMalloc2(nlsol_loc,&sol_loc,lsol_loc,&isol_loc);
1340:   mumps->id.sol_loc  = (MumpsScalar*)sol_loc;
1341:   mumps->id.isol_loc = isol_loc;

1343:   VecCreateSeqWithArray(PETSC_COMM_SELF,1,nlsol_loc,(PetscScalar*)sol_loc,&msol_loc);

1345:   if (denseB) {
1346:     if (mumps->ICNTL20 == 10) {
1347:       mumps->id.ICNTL(20) = 10; /* dense distributed RHS */
1348:       MatDenseGetArrayRead(B,&rbray);
1349:       MatMumpsSetUpDistRHSInfo(A,nrhs,rbray);
1350:       MatDenseRestoreArrayRead(B,&rbray);
1351:       MatGetLocalSize(B,&m,NULL);
1352:       VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,NULL,&v_mpi);
1353:     } else {
1354:       mumps->id.ICNTL(20) = 0; /* dense centralized RHS */
1355:       /* TODO: Because of non-contiguous indices, the created vecscatter scat_rhs is not done in MPI_Gather, resulting in
1356:         very inefficient communication. An optimization is to use VecScatterCreateToZero to gather B to rank 0. Then on rank
1357:         0, re-arrange B into desired order, which is a local operation.
1358:       */

1360:       /* scatter v_mpi to b_seq because MUMPS before 5.3.0 only supports centralized rhs */
1361:       /* wrap dense rhs matrix B into a vector v_mpi */
1362:       MatGetLocalSize(B,&m,NULL);
1363:       MatDenseGetArray(B,&bray);
1364:       VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);
1365:       MatDenseRestoreArray(B,&bray);

1367:       /* scatter v_mpi to b_seq in proc[0]. MUMPS requires rhs to be centralized on the host! */
1368:       if (!mumps->myid) {
1369:         PetscInt *idx;
1370:         /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B */
1371:         PetscMalloc1(nrhs*M,&idx);
1372:         MatGetOwnershipRanges(B,&rstart);
1373:         k = 0;
1374:         for (proc=0; proc<mumps->petsc_size; proc++) {
1375:           for (j=0; j<nrhs; j++) {
1376:             for (i=rstart[proc]; i<rstart[proc+1]; i++) idx[k++] = j*M + i;
1377:           }
1378:         }

1380:         VecCreateSeq(PETSC_COMM_SELF,nrhs*M,&b_seq);
1381:         ISCreateGeneral(PETSC_COMM_SELF,nrhs*M,idx,PETSC_OWN_POINTER,&is_to);
1382:         ISCreateStride(PETSC_COMM_SELF,nrhs*M,0,1,&is_from);
1383:       } else {
1384:         VecCreateSeq(PETSC_COMM_SELF,0,&b_seq);
1385:         ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_to);
1386:         ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_from);
1387:       }
1388:       VecScatterCreate(v_mpi,is_from,b_seq,is_to,&scat_rhs);
1389:       VecScatterBegin(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);
1390:       ISDestroy(&is_to);
1391:       ISDestroy(&is_from);
1392:       VecScatterEnd(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);

1394:       if (!mumps->myid) { /* define rhs on the host */
1395:         VecGetArray(b_seq,&bray);
1396:         mumps->id.rhs = (MumpsScalar*)bray;
1397:         VecRestoreArray(b_seq,&bray);
1398:       }
1399:     }
1400:   } else { /* sparse B */
1401:     b = (Mat_MPIAIJ*)Bt->data;

1403:     /* wrap dense X into a vector v_mpi */
1404:     MatGetLocalSize(X,&m,NULL);
1405:     MatDenseGetArray(X,&bray);
1406:     VecCreateMPIWithArray(PetscObjectComm((PetscObject)X),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);
1407:     MatDenseRestoreArray(X,&bray);

1409:     if (!mumps->myid) {
1410:       MatSeqAIJGetArray(b->A,&aa);
1411:       MatGetRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1412:       if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
1413:       PetscMUMPSIntCSRCast(mumps,spnr,ia,ja,&mumps->id.irhs_ptr,&mumps->id.irhs_sparse,&mumps->id.nz_rhs);
1414:       mumps->id.rhs_sparse  = (MumpsScalar*)aa;
1415:     } else {
1416:       mumps->id.irhs_ptr    = NULL;
1417:       mumps->id.irhs_sparse = NULL;
1418:       mumps->id.nz_rhs      = 0;
1419:       mumps->id.rhs_sparse  = NULL;
1420:     }
1421:   }

1423:   /* solve phase */
1424:   /*-------------*/
1425:   mumps->id.job = JOB_SOLVE;
1426:   PetscMUMPS_c(mumps);
1427:   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));

1429:   /* scatter mumps distributed solution to petsc vector v_mpi, which shares local arrays with solution matrix X */
1430:   MatDenseGetArray(X,&array);
1431:   VecPlaceArray(v_mpi,array);

1433:   /* create scatter scat_sol */
1434:   MatGetOwnershipRanges(X,&rstart);
1435:   /* iidx: index for scatter mumps solution to petsc X */

1437:   ISCreateStride(PETSC_COMM_SELF,nlsol_loc,0,1,&is_from);
1438:   PetscMalloc1(nlsol_loc,&idxx);
1439:   for (i=0; i<lsol_loc; i++) {
1440:     isol_loc[i] -= 1; /* change Fortran style to C style. isol_loc[i+j*lsol_loc] contains x[isol_loc[i]] in j-th vector */

1442:     for (proc=0; proc<mumps->petsc_size; proc++) {
1443:       if (isol_loc[i] >= rstart[proc] && isol_loc[i] < rstart[proc+1]) {
1444:         myrstart = rstart[proc];
1445:         k        = isol_loc[i] - myrstart;        /* local index on 1st column of petsc vector X */
1446:         iidx     = k + myrstart*nrhs;             /* maps mumps isol_loc[i] to petsc index in X */
1447:         m        = rstart[proc+1] - rstart[proc]; /* rows of X for this proc */
1448:         break;
1449:       }
1450:     }

1452:     for (j=0; j<nrhs; j++) idxx[i+j*lsol_loc] = iidx + j*m;
1453:   }
1454:   ISCreateGeneral(PETSC_COMM_SELF,nlsol_loc,idxx,PETSC_COPY_VALUES,&is_to);
1455:   VecScatterCreate(msol_loc,is_from,v_mpi,is_to,&scat_sol);
1456:   VecScatterBegin(scat_sol,msol_loc,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
1457:   ISDestroy(&is_from);
1458:   ISDestroy(&is_to);
1459:   VecScatterEnd(scat_sol,msol_loc,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
1460:   MatDenseRestoreArray(X,&array);

1462:   /* free spaces */
1463:   mumps->id.sol_loc  = (MumpsScalar*)sol_loc_save;
1464:   mumps->id.isol_loc = isol_loc_save;

1466:   PetscFree2(sol_loc,isol_loc);
1467:   PetscFree(idxx);
1468:   VecDestroy(&msol_loc);
1469:   VecDestroy(&v_mpi);
1470:   if (!denseB) {
1471:     if (!mumps->myid) {
1472:       b = (Mat_MPIAIJ*)Bt->data;
1473:       MatSeqAIJRestoreArray(b->A,&aa);
1474:       MatRestoreRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1475:       if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure");
1476:     }
1477:   } else {
1478:     if (mumps->ICNTL20 == 0) {
1479:       VecDestroy(&b_seq);
1480:       VecScatterDestroy(&scat_rhs);
1481:     }
1482:   }
1483:   VecScatterDestroy(&scat_sol);
1484:   PetscLogFlops(2.0*nrhs*mumps->id.RINFO(3));
1485:   return(0);
1486: }

1488: PetscErrorCode MatMatTransposeSolve_MUMPS(Mat A,Mat Bt,Mat X)
1489: {
1491:   PetscBool      flg;
1492:   Mat            B;

1495:   PetscObjectTypeCompareAny((PetscObject)Bt,&flg,MATSEQAIJ,MATMPIAIJ,NULL);
1496:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Matrix Bt must be MATAIJ matrix");

1498:   /* Create B=Bt^T that uses Bt's data structure */
1499:   MatCreateTranspose(Bt,&B);

1501:   MatMatSolve_MUMPS(A,B,X);
1502:   MatDestroy(&B);
1503:   return(0);
1504: }

1506: #if !defined(PETSC_USE_COMPLEX)
1507: /*
1508:   input:
1509:    F:        numeric factor
1510:   output:
1511:    nneg:     total number of negative pivots
1512:    nzero:    total number of zero pivots
1513:    npos:     (global dimension of F) - nneg - nzero
1514: */
1515: PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F,PetscInt *nneg,PetscInt *nzero,PetscInt *npos)
1516: {
1517:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
1519:   PetscMPIInt    size;

1522:   MPI_Comm_size(PetscObjectComm((PetscObject)F),&size);
1523:   /* 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 */
1524:   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));

1526:   if (nneg) *nneg = mumps->id.INFOG(12);
1527:   if (nzero || npos) {
1528:     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");
1529:     if (nzero) *nzero = mumps->id.INFOG(28);
1530:     if (npos) *npos   = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28));
1531:   }
1532:   return(0);
1533: }
1534: #endif

1536: PetscErrorCode MatMumpsGatherNonzerosOnMaster(MatReuse reuse,Mat_MUMPS *mumps)
1537: {
1539:   PetscInt       i,nreqs;
1540:   PetscMUMPSInt  *irn,*jcn;
1541:   PetscMPIInt    count;
1542:   PetscInt64     totnnz,remain;
1543:   const PetscInt osize=mumps->omp_comm_size;
1544:   PetscScalar    *val;

1547:   if (osize > 1) {
1548:     if (reuse == MAT_INITIAL_MATRIX) {
1549:       /* master first gathers counts of nonzeros to receive */
1550:       if (mumps->is_omp_master) {PetscMalloc1(osize,&mumps->recvcount);}
1551:       MPI_Gather(&mumps->nnz,1,MPIU_INT64,mumps->recvcount,1,MPIU_INT64,0/*master*/,mumps->omp_comm);

1553:       /* Then each computes number of send/recvs */
1554:       if (mumps->is_omp_master) {
1555:         /* Start from 1 since self communication is not done in MPI */
1556:         nreqs = 0;
1557:         for (i=1; i<osize; i++) nreqs += (mumps->recvcount[i]+PETSC_MPI_INT_MAX-1)/PETSC_MPI_INT_MAX;
1558:       } else {
1559:         nreqs = (mumps->nnz+PETSC_MPI_INT_MAX-1)/PETSC_MPI_INT_MAX;
1560:       }
1561:       PetscMalloc1(nreqs*3,&mumps->reqs); /* Triple the requests since we send irn, jcn and val seperately */

1563:       /* The following code is doing a very simple thing: omp_master rank gathers irn/jcn/val from others.
1564:          MPI_Gatherv would be enough if it supports big counts > 2^31-1. Since it does not, and mumps->nnz
1565:          might be a prime number > 2^31-1, we have to slice the message. Note omp_comm_size
1566:          is very small, the current approach should have no extra overhead compared to MPI_Gatherv.
1567:        */
1568:       nreqs = 0; /* counter for actual send/recvs */
1569:       if (mumps->is_omp_master) {
1570:         for (i=0,totnnz=0; i<osize; i++) totnnz += mumps->recvcount[i]; /* totnnz = sum of nnz over omp_comm */
1571:         PetscMalloc2(totnnz,&irn,totnnz,&jcn);
1572:         PetscMalloc1(totnnz,&val);

1574:         /* Self communication */
1575:         PetscArraycpy(irn,mumps->irn,mumps->nnz);
1576:         PetscArraycpy(jcn,mumps->jcn,mumps->nnz);
1577:         PetscArraycpy(val,mumps->val,mumps->nnz);

1579:         /* Replace mumps->irn/jcn etc on master with the newly allocated bigger arrays */
1580:         PetscFree2(mumps->irn,mumps->jcn);
1581:         PetscFree(mumps->val_alloc);
1582:         mumps->nnz = totnnz;
1583:         mumps->irn = irn;
1584:         mumps->jcn = jcn;
1585:         mumps->val = mumps->val_alloc = val;

1587:         irn += mumps->recvcount[0]; /* recvcount[0] is old mumps->nnz on omp rank 0 */
1588:         jcn += mumps->recvcount[0];
1589:         val += mumps->recvcount[0];

1591:         /* Remote communication */
1592:         for (i=1; i<osize; i++) {
1593:           count  = PetscMin(mumps->recvcount[i],PETSC_MPI_INT_MAX);
1594:           remain = mumps->recvcount[i] - count;
1595:           while (count>0) {
1596:             MPI_Irecv(irn,count,MPIU_MUMPSINT,i,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1597:             MPI_Irecv(jcn,count,MPIU_MUMPSINT,i,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1598:             MPI_Irecv(val,count,MPIU_SCALAR,  i,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1599:             irn    += count;
1600:             jcn    += count;
1601:             val    += count;
1602:             count   = PetscMin(remain,PETSC_MPI_INT_MAX);
1603:             remain -= count;
1604:           }
1605:         }
1606:       } else {
1607:         irn    = mumps->irn;
1608:         jcn    = mumps->jcn;
1609:         val    = mumps->val;
1610:         count  = PetscMin(mumps->nnz,PETSC_MPI_INT_MAX);
1611:         remain = mumps->nnz - count;
1612:         while (count>0) {
1613:           MPI_Isend(irn,count,MPIU_MUMPSINT,0,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1614:           MPI_Isend(jcn,count,MPIU_MUMPSINT,0,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1615:           MPI_Isend(val,count,MPIU_SCALAR,  0,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1616:           irn    += count;
1617:           jcn    += count;
1618:           val    += count;
1619:           count   = PetscMin(remain,PETSC_MPI_INT_MAX);
1620:           remain -= count;
1621:         }
1622:       }
1623:     } else {
1624:       nreqs = 0;
1625:       if (mumps->is_omp_master) {
1626:         val = mumps->val + mumps->recvcount[0];
1627:         for (i=1; i<osize; i++) { /* Remote communication only since self data is already in place */
1628:           count  = PetscMin(mumps->recvcount[i],PETSC_MPI_INT_MAX);
1629:           remain = mumps->recvcount[i] - count;
1630:           while (count>0) {
1631:             MPI_Irecv(val,count,MPIU_SCALAR,i,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1632:             val    += count;
1633:             count   = PetscMin(remain,PETSC_MPI_INT_MAX);
1634:             remain -= count;
1635:           }
1636:         }
1637:       } else {
1638:         val    = mumps->val;
1639:         count  = PetscMin(mumps->nnz,PETSC_MPI_INT_MAX);
1640:         remain = mumps->nnz - count;
1641:         while (count>0) {
1642:           MPI_Isend(val,count,MPIU_SCALAR,0,mumps->tag,mumps->omp_comm,&mumps->reqs[nreqs++]);
1643:           val    += count;
1644:           count   = PetscMin(remain,PETSC_MPI_INT_MAX);
1645:           remain -= count;
1646:         }
1647:       }
1648:     }
1649:     MPI_Waitall(nreqs,mumps->reqs,MPI_STATUSES_IGNORE);
1650:     mumps->tag++; /* It is totally fine for above send/recvs to share one mpi tag */
1651:   }
1652:   return(0);
1653: }

1655: PetscErrorCode MatFactorNumeric_MUMPS(Mat F,Mat A,const MatFactorInfo *info)
1656: {
1657:   Mat_MUMPS      *mumps =(Mat_MUMPS*)(F)->data;
1659:   PetscBool      isMPIAIJ;

1662:   if (mumps->id.INFOG(1) < 0 && !(mumps->id.INFOG(1) == -16 && mumps->id.INFOG(1) == 0)) {
1663:     if (mumps->id.INFOG(1) == -6) {
1664:       PetscInfo2(A,"MatFactorNumeric is called with singular matrix structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1665:     }
1666:     PetscInfo2(A,"MatFactorNumeric is called after analysis phase fails, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1667:     return(0);
1668:   }

1670:   (*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, mumps);
1671:   MatMumpsGatherNonzerosOnMaster(MAT_REUSE_MATRIX,mumps);

1673:   /* numerical factorization phase */
1674:   /*-------------------------------*/
1675:   mumps->id.job = JOB_FACTNUMERIC;
1676:   if (!mumps->id.ICNTL(18)) { /* A is centralized */
1677:     if (!mumps->myid) {
1678:       mumps->id.a = (MumpsScalar*)mumps->val;
1679:     }
1680:   } else {
1681:     mumps->id.a_loc = (MumpsScalar*)mumps->val;
1682:   }
1683:   PetscMUMPS_c(mumps);
1684:   if (mumps->id.INFOG(1) < 0) {
1685:     if (A->erroriffailure) {
1686:       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));
1687:     } else {
1688:       if (mumps->id.INFOG(1) == -10) { /* numerically singular matrix */
1689:         PetscInfo2(F,"matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1690:         F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1691:       } else if (mumps->id.INFOG(1) == -13) {
1692:         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));
1693:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1694:       } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10)) {
1695:         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));
1696:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1697:       } else {
1698:         PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1699:         F->factorerrortype = MAT_FACTOR_OTHER;
1700:       }
1701:     }
1702:   }
1703:   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));

1705:   F->assembled    = PETSC_TRUE;
1706:   mumps->matstruc = SAME_NONZERO_PATTERN;
1707:   if (F->schur) { /* reset Schur status to unfactored */
1708: #if defined(PETSC_HAVE_CUDA)
1709:     F->schur->offloadmask = PETSC_OFFLOAD_CPU;
1710: #endif
1711:     if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
1712:       mumps->id.ICNTL(19) = 2;
1713:       MatTranspose(F->schur,MAT_INPLACE_MATRIX,&F->schur);
1714:     }
1715:     MatFactorRestoreSchurComplement(F,NULL,MAT_FACTOR_SCHUR_UNFACTORED);
1716:   }

1718:   /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */
1719:   if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3;

1721:   if (!mumps->is_omp_master) mumps->id.INFO(23) = 0;
1722:   if (mumps->petsc_size > 1) {
1723:     PetscInt    lsol_loc;
1724:     PetscScalar *sol_loc;

1726:     PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&isMPIAIJ);

1728:     /* distributed solution; Create x_seq=sol_loc for repeated use */
1729:     if (mumps->x_seq) {
1730:       VecScatterDestroy(&mumps->scat_sol);
1731:       PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
1732:       VecDestroy(&mumps->x_seq);
1733:     }
1734:     lsol_loc = mumps->id.INFO(23); /* length of sol_loc */
1735:     PetscMalloc2(lsol_loc,&sol_loc,lsol_loc,&mumps->id.isol_loc);
1736:     mumps->id.lsol_loc = lsol_loc;
1737:     mumps->id.sol_loc = (MumpsScalar*)sol_loc;
1738:     VecCreateSeqWithArray(PETSC_COMM_SELF,1,lsol_loc,sol_loc,&mumps->x_seq);
1739:   }
1740:   PetscLogFlops(mumps->id.RINFO(2));
1741:   return(0);
1742: }

1744: /* Sets MUMPS options from the options database */
1745: PetscErrorCode PetscSetMUMPSFromOptions(Mat F, Mat A)
1746: {
1747:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1749:   PetscMUMPSInt  icntl=0;
1750:   PetscInt       info[80],i,ninfo=80;
1751:   PetscBool      flg=PETSC_FALSE;

1754:   PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MUMPS Options","Mat");
1755:   PetscOptionsMUMPSInt("-mat_mumps_icntl_1","ICNTL(1): output stream for error messages","None",mumps->id.ICNTL(1),&icntl,&flg);
1756:   if (flg) mumps->id.ICNTL(1) = icntl;
1757:   PetscOptionsMUMPSInt("-mat_mumps_icntl_2","ICNTL(2): output stream for diagnostic printing, statistics, and warning","None",mumps->id.ICNTL(2),&icntl,&flg);
1758:   if (flg) mumps->id.ICNTL(2) = icntl;
1759:   PetscOptionsMUMPSInt("-mat_mumps_icntl_3","ICNTL(3): output stream for global information, collected on the host","None",mumps->id.ICNTL(3),&icntl,&flg);
1760:   if (flg) mumps->id.ICNTL(3) = icntl;

1762:   PetscOptionsMUMPSInt("-mat_mumps_icntl_4","ICNTL(4): level of printing (0 to 4)","None",mumps->id.ICNTL(4),&icntl,&flg);
1763:   if (flg) mumps->id.ICNTL(4) = icntl;
1764:   if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */

1766:   PetscOptionsMUMPSInt("-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);
1767:   if (flg) mumps->id.ICNTL(6) = icntl;

1769:   PetscOptionsMUMPSInt("-mat_mumps_icntl_7","ICNTL(7): computes a symmetric permutation in sequential analysis. 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto(default)","None",mumps->id.ICNTL(7),&icntl,&flg);
1770:   if (flg) {
1771:     if (icntl == 1 || icntl < 0 || icntl > 7) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Valid values are 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto\n");
1772:     mumps->id.ICNTL(7) = icntl;
1773:   }

1775:   PetscOptionsMUMPSInt("-mat_mumps_icntl_8","ICNTL(8): scaling strategy (-2 to 8 or 77)","None",mumps->id.ICNTL(8),&mumps->id.ICNTL(8),NULL);
1776:   /* 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() */
1777:   PetscOptionsMUMPSInt("-mat_mumps_icntl_10","ICNTL(10): max num of refinements","None",mumps->id.ICNTL(10),&mumps->id.ICNTL(10),NULL);
1778:   PetscOptionsMUMPSInt("-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);
1779:   PetscOptionsMUMPSInt("-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);
1780:   PetscOptionsMUMPSInt("-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);
1781:   PetscOptionsMUMPSInt("-mat_mumps_icntl_14","ICNTL(14): percentage increase in the estimated working space","None",mumps->id.ICNTL(14),&mumps->id.ICNTL(14),NULL);
1782:   PetscOptionsMUMPSInt("-mat_mumps_icntl_19","ICNTL(19): computes the Schur complement","None",mumps->id.ICNTL(19),&mumps->id.ICNTL(19),NULL);
1783:   if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any schur data (if any) */
1784:     MatDestroy(&F->schur);
1785:     MatMumpsResetSchur_Private(mumps);
1786:   }

1788:   /* Two MPICH Fortran MPI_IN_PLACE binding bugs prevented the use of 'mpich + mumps'. One happened with "mpi4py + mpich + mumps",
1789:      and was reported by Firedrake. See https://bitbucket.org/mpi4py/mpi4py/issues/162/mpi4py-initialization-breaks-fortran
1790:      and a petsc-maint mailing list thread with subject 'MUMPS segfaults in parallel because of ...'
1791:      This bug was fixed by https://github.com/pmodels/mpich/pull/4149. But the fix brought a new bug,
1792:      see https://github.com/pmodels/mpich/issues/5589. This bug was fixed by https://github.com/pmodels/mpich/pull/5590.
1793:      In short, we could not use distributed RHS with MPICH until v4.0b1.
1794:    */
1795: #if PETSC_PKG_MUMPS_VERSION_LT(5,3,0) || (defined(PETSC_HAVE_MPICH_NUMVERSION) && (PETSC_HAVE_MPICH_NUMVERSION < 40000101))
1796:   mumps->ICNTL20 = 0;  /* Centralized dense RHS*/
1797: #else
1798:   mumps->ICNTL20 = 10; /* Distributed dense RHS*/
1799: #endif
1800:   PetscOptionsMUMPSInt("-mat_mumps_icntl_20","ICNTL(20): give mumps centralized (0) or distributed (10) dense right-hand sides","None",mumps->ICNTL20,&mumps->ICNTL20,&flg);
1801:   if (flg && mumps->ICNTL20 != 10 && mumps->ICNTL20 != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"ICNTL(20)=%d is not supported by the PETSc/MUMPS interface. Allowed values are 0, 10\n",(int)mumps->ICNTL20);
1802: #if PETSC_PKG_MUMPS_VERSION_LT(5,3,0)
1803:   if (flg && mumps->ICNTL20 == 10) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"ICNTL(20)=10 is not supported before MUMPS-5.3.0\n");
1804: #endif
1805:   /* PetscOptionsMUMPSInt("-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 */

1807:   PetscOptionsMUMPSInt("-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);
1808:   PetscOptionsMUMPSInt("-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);
1809:   PetscOptionsMUMPSInt("-mat_mumps_icntl_24","ICNTL(24): detection of null pivot rows (0 or 1)","None",mumps->id.ICNTL(24),&mumps->id.ICNTL(24),NULL);
1810:   if (mumps->id.ICNTL(24)) {
1811:     mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */
1812:   }

1814:   PetscOptionsMUMPSInt("-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);
1815:   PetscOptionsMUMPSInt("-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);
1816:   PetscOptionsMUMPSInt("-mat_mumps_icntl_27","ICNTL(27): controls the blocking size for multiple right-hand sides","None",mumps->id.ICNTL(27),&mumps->id.ICNTL(27),NULL);
1817:   PetscOptionsMUMPSInt("-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);
1818:   PetscOptionsMUMPSInt("-mat_mumps_icntl_29","ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis","None",mumps->id.ICNTL(29),&mumps->id.ICNTL(29),NULL);
1819:   /* PetscOptionsMUMPSInt("-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); */ /* call MatMumpsGetInverse() directly */
1820:   PetscOptionsMUMPSInt("-mat_mumps_icntl_31","ICNTL(31): indicates which factors may be discarded during factorization","None",mumps->id.ICNTL(31),&mumps->id.ICNTL(31),NULL);
1821:   /* PetscOptionsMUMPSInt("-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 */
1822:   PetscOptionsMUMPSInt("-mat_mumps_icntl_33","ICNTL(33): compute determinant","None",mumps->id.ICNTL(33),&mumps->id.ICNTL(33),NULL);
1823:   PetscOptionsMUMPSInt("-mat_mumps_icntl_35","ICNTL(35): activates Block Low Rank (BLR) based factorization","None",mumps->id.ICNTL(35),&mumps->id.ICNTL(35),NULL);
1824:   PetscOptionsMUMPSInt("-mat_mumps_icntl_36","ICNTL(36): choice of BLR factorization variant","None",mumps->id.ICNTL(36),&mumps->id.ICNTL(36),NULL);
1825:   PetscOptionsMUMPSInt("-mat_mumps_icntl_38","ICNTL(38): estimated compression rate of LU factors with BLR","None",mumps->id.ICNTL(38),&mumps->id.ICNTL(38),NULL);

1827:   PetscOptionsReal("-mat_mumps_cntl_1","CNTL(1): relative pivoting threshold","None",mumps->id.CNTL(1),&mumps->id.CNTL(1),NULL);
1828:   PetscOptionsReal("-mat_mumps_cntl_2","CNTL(2): stopping criterion of refinement","None",mumps->id.CNTL(2),&mumps->id.CNTL(2),NULL);
1829:   PetscOptionsReal("-mat_mumps_cntl_3","CNTL(3): absolute pivoting threshold","None",mumps->id.CNTL(3),&mumps->id.CNTL(3),NULL);
1830:   PetscOptionsReal("-mat_mumps_cntl_4","CNTL(4): value for static pivoting","None",mumps->id.CNTL(4),&mumps->id.CNTL(4),NULL);
1831:   PetscOptionsReal("-mat_mumps_cntl_5","CNTL(5): fixation for null pivots","None",mumps->id.CNTL(5),&mumps->id.CNTL(5),NULL);
1832:   PetscOptionsReal("-mat_mumps_cntl_7","CNTL(7): dropping parameter used during BLR","None",mumps->id.CNTL(7),&mumps->id.CNTL(7),NULL);

1834:   PetscOptionsString("-mat_mumps_ooc_tmpdir", "out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, sizeof(mumps->id.ooc_tmpdir), NULL);

1836:   PetscOptionsIntArray("-mat_mumps_view_info","request INFO local to each processor","",info,&ninfo,NULL);
1837:   if (ninfo) {
1838:     if (ninfo > 80) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"number of INFO %d must <= 80\n",ninfo);
1839:     PetscMalloc1(ninfo,&mumps->info);
1840:     mumps->ninfo = ninfo;
1841:     for (i=0; i<ninfo; i++) {
1842:       if (info[i] < 0 || info[i]>80) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"index of INFO %d must between 1 and 80\n",ninfo);
1843:       else  mumps->info[i] = info[i];
1844:     }
1845:   }

1847:   PetscOptionsEnd();
1848:   return(0);
1849: }

1851: PetscErrorCode PetscInitializeMUMPS(Mat A,Mat_MUMPS *mumps)
1852: {
1854:   PetscInt       nthreads=0;
1855:   MPI_Comm       newcomm=MPI_COMM_NULL;

1858:   mumps->petsc_comm = PetscObjectComm((PetscObject)A);
1859:   MPI_Comm_size(mumps->petsc_comm,&mumps->petsc_size);
1860:   MPI_Comm_rank(mumps->petsc_comm,&mumps->myid);/* "if (!myid)" still works even if mumps_comm is different */

1862:   PetscOptionsHasName(NULL,((PetscObject)A)->prefix,"-mat_mumps_use_omp_threads",&mumps->use_petsc_omp_support);
1863:   if (mumps->use_petsc_omp_support) nthreads = -1; /* -1 will let PetscOmpCtrlCreate() guess a proper value when user did not supply one */
1864:   PetscOptionsGetInt(NULL,((PetscObject)A)->prefix,"-mat_mumps_use_omp_threads",&nthreads,NULL);
1865:   if (mumps->use_petsc_omp_support) {
1866: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
1867:     PetscOmpCtrlCreate(mumps->petsc_comm,nthreads,&mumps->omp_ctrl);
1868:     PetscOmpCtrlGetOmpComms(mumps->omp_ctrl,&mumps->omp_comm,&mumps->mumps_comm,&mumps->is_omp_master);
1869: #else
1870:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP_SYS,"the system does not have PETSc OpenMP support but you added the -%smat_mumps_use_omp_threads option. Configure PETSc with --with-openmp --download-hwloc (or --with-hwloc) to enable it, see more in MATSOLVERMUMPS manual\n",((PetscObject)A)->prefix?((PetscObject)A)->prefix:"");
1871: #endif
1872:   } else {
1873:     mumps->omp_comm      = PETSC_COMM_SELF;
1874:     mumps->mumps_comm    = mumps->petsc_comm;
1875:     mumps->is_omp_master = PETSC_TRUE;
1876:   }
1877:   MPI_Comm_size(mumps->omp_comm,&mumps->omp_comm_size);
1878:   mumps->reqs = NULL;
1879:   mumps->tag  = 0;

1881:   /* It looks like MUMPS does not dup the input comm. Dup a new comm for MUMPS to avoid any tag mismatches. */
1882:   if (mumps->mumps_comm != MPI_COMM_NULL) {
1883:     MPI_Comm_dup(mumps->mumps_comm,&newcomm);
1884:     mumps->mumps_comm = newcomm;
1885:   }

1887:   mumps->id.comm_fortran = MPI_Comm_c2f(mumps->mumps_comm);
1888:   mumps->id.job = JOB_INIT;
1889:   mumps->id.par = 1;  /* host participates factorizaton and solve */
1890:   mumps->id.sym = mumps->sym;

1892:   PetscMUMPS_c(mumps);
1893:   if (mumps->id.INFOG(1) < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in PetscInitializeMUMPS: INFOG(1)=%d\n",mumps->id.INFOG(1));

1895:   /* copy MUMPS default control values from master to slaves. Although slaves do not call MUMPS, they may access these values in code.
1896:      For example, ICNTL(9) is initialized to 1 by MUMPS and slaves check ICNTL(9) in MatSolve_MUMPS.
1897:    */
1898:   MPI_Bcast(mumps->id.icntl,40,MPI_INT,  0,mumps->omp_comm);
1899:   MPI_Bcast(mumps->id.cntl, 15,MPIU_REAL,0,mumps->omp_comm);

1901:   mumps->scat_rhs = NULL;
1902:   mumps->scat_sol = NULL;

1904:   /* set PETSc-MUMPS default options - override MUMPS default */
1905:   mumps->id.ICNTL(3) = 0;
1906:   mumps->id.ICNTL(4) = 0;
1907:   if (mumps->petsc_size == 1) {
1908:     mumps->id.ICNTL(18) = 0;   /* centralized assembled matrix input */
1909:     mumps->id.ICNTL(7)  = 7;   /* automatic choice of ordering done by the package */
1910:   } else {
1911:     mumps->id.ICNTL(18) = 3;   /* distributed assembled matrix input */
1912:     mumps->id.ICNTL(21) = 1;   /* distributed solution */
1913:   }

1915:   /* schur */
1916:   mumps->id.size_schur    = 0;
1917:   mumps->id.listvar_schur = NULL;
1918:   mumps->id.schur         = NULL;
1919:   mumps->sizeredrhs       = 0;
1920:   mumps->schur_sol        = NULL;
1921:   mumps->schur_sizesol    = 0;
1922:   return(0);
1923: }

1925: PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F,Mat A,const MatFactorInfo *info,Mat_MUMPS *mumps)
1926: {

1930:   if (mumps->id.INFOG(1) < 0) {
1931:     if (A->erroriffailure) {
1932:       SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in analysis phase: INFOG(1)=%d\n",mumps->id.INFOG(1));
1933:     } else {
1934:       if (mumps->id.INFOG(1) == -6) {
1935:         PetscInfo2(F,"matrix is singular in structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1936:         F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT;
1937:       } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) {
1938:         PetscInfo2(F,"problem of workspace, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1939:         F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1940:       } else if (mumps->id.INFOG(1) == -16 && mumps->id.INFOG(1) == 0) {
1941:         PetscInfo(F,"Empty matrix\n");
1942:       } else {
1943:         PetscInfo2(F,"Error reported by MUMPS in analysis phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1944:         F->factorerrortype = MAT_FACTOR_OTHER;
1945:       }
1946:     }
1947:   }
1948:   return(0);
1949: }

1951: PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
1952: {
1953:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
1955:   Vec            b;
1956:   const PetscInt M = A->rmap->N;

1959:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

1961:   /* Set MUMPS options from the options database */
1962:   PetscSetMUMPSFromOptions(F,A);

1964:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps);
1965:   MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);

1967:   /* analysis phase */
1968:   /*----------------*/
1969:   mumps->id.job = JOB_FACTSYMBOLIC;
1970:   mumps->id.n   = M;
1971:   switch (mumps->id.ICNTL(18)) {
1972:   case 0:  /* centralized assembled matrix input */
1973:     if (!mumps->myid) {
1974:       mumps->id.nnz = mumps->nnz;
1975:       mumps->id.irn = mumps->irn;
1976:       mumps->id.jcn = mumps->jcn;
1977:       if (mumps->id.ICNTL(6)>1) mumps->id.a = (MumpsScalar*)mumps->val;
1978:       if (r) {
1979:         mumps->id.ICNTL(7) = 1;
1980:         if (!mumps->myid) {
1981:           const PetscInt *idx;
1982:           PetscInt       i;

1984:           PetscMalloc1(M,&mumps->id.perm_in);
1985:           ISGetIndices(r,&idx);
1986:           for (i=0; i<M; i++) {PetscMUMPSIntCast(idx[i]+1,&(mumps->id.perm_in[i]));} /* perm_in[]: start from 1, not 0! */
1987:           ISRestoreIndices(r,&idx);
1988:         }
1989:       }
1990:     }
1991:     break;
1992:   case 3:  /* distributed assembled matrix input (size>1) */
1993:     mumps->id.nnz_loc = mumps->nnz;
1994:     mumps->id.irn_loc = mumps->irn;
1995:     mumps->id.jcn_loc = mumps->jcn;
1996:     if (mumps->id.ICNTL(6)>1) mumps->id.a_loc = (MumpsScalar*)mumps->val;
1997:     if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1998:       MatCreateVecs(A,NULL,&b);
1999:       VecScatterCreateToZero(b,&mumps->scat_rhs,&mumps->b_seq);
2000:       VecDestroy(&b);
2001:     }
2002:     break;
2003:   }
2004:   PetscMUMPS_c(mumps);
2005:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

2007:   F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
2008:   F->ops->solve           = MatSolve_MUMPS;
2009:   F->ops->solvetranspose  = MatSolveTranspose_MUMPS;
2010:   F->ops->matsolve        = MatMatSolve_MUMPS;
2011:   F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS;
2012:   return(0);
2013: }

2015: /* Note the Petsc r and c permutations are ignored */
2016: PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
2017: {
2018:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
2020:   Vec            b;
2021:   const PetscInt M = A->rmap->N;

2024:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

2026:   /* Set MUMPS options from the options database */
2027:   PetscSetMUMPSFromOptions(F,A);

2029:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps);
2030:   MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);

2032:   /* analysis phase */
2033:   /*----------------*/
2034:   mumps->id.job = JOB_FACTSYMBOLIC;
2035:   mumps->id.n   = M;
2036:   switch (mumps->id.ICNTL(18)) {
2037:   case 0:  /* centralized assembled matrix input */
2038:     if (!mumps->myid) {
2039:       mumps->id.nnz = mumps->nnz;
2040:       mumps->id.irn = mumps->irn;
2041:       mumps->id.jcn = mumps->jcn;
2042:       if (mumps->id.ICNTL(6)>1) {
2043:         mumps->id.a = (MumpsScalar*)mumps->val;
2044:       }
2045:     }
2046:     break;
2047:   case 3:  /* distributed assembled matrix input (size>1) */
2048:     mumps->id.nnz_loc = mumps->nnz;
2049:     mumps->id.irn_loc = mumps->irn;
2050:     mumps->id.jcn_loc = mumps->jcn;
2051:     if (mumps->id.ICNTL(6)>1) {
2052:       mumps->id.a_loc = (MumpsScalar*)mumps->val;
2053:     }
2054:     if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
2055:       MatCreateVecs(A,NULL,&b);
2056:       VecScatterCreateToZero(b,&mumps->scat_rhs,&mumps->b_seq);
2057:       VecDestroy(&b);
2058:     }
2059:     break;
2060:   }
2061:   PetscMUMPS_c(mumps);
2062:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

2064:   F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
2065:   F->ops->solve           = MatSolve_MUMPS;
2066:   F->ops->solvetranspose  = MatSolveTranspose_MUMPS;
2067:   return(0);
2068: }

2070: /* Note the Petsc r permutation and factor info are ignored */
2071: PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F,Mat A,IS r,const MatFactorInfo *info)
2072: {
2073:   Mat_MUMPS      *mumps = (Mat_MUMPS*)F->data;
2075:   Vec            b;
2076:   const PetscInt M = A->rmap->N;

2079:   mumps->matstruc = DIFFERENT_NONZERO_PATTERN;

2081:   /* Set MUMPS options from the options database */
2082:   PetscSetMUMPSFromOptions(F,A);

2084:   (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, mumps);
2085:   MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);

2087:   /* analysis phase */
2088:   /*----------------*/
2089:   mumps->id.job = JOB_FACTSYMBOLIC;
2090:   mumps->id.n   = M;
2091:   switch (mumps->id.ICNTL(18)) {
2092:   case 0:  /* centralized assembled matrix input */
2093:     if (!mumps->myid) {
2094:       mumps->id.nnz = mumps->nnz;
2095:       mumps->id.irn = mumps->irn;
2096:       mumps->id.jcn = mumps->jcn;
2097:       if (mumps->id.ICNTL(6)>1) {
2098:         mumps->id.a = (MumpsScalar*)mumps->val;
2099:       }
2100:     }
2101:     break;
2102:   case 3:  /* distributed assembled matrix input (size>1) */
2103:     mumps->id.nnz_loc = mumps->nnz;
2104:     mumps->id.irn_loc = mumps->irn;
2105:     mumps->id.jcn_loc = mumps->jcn;
2106:     if (mumps->id.ICNTL(6)>1) {
2107:       mumps->id.a_loc = (MumpsScalar*)mumps->val;
2108:     }
2109:     if (mumps->ICNTL20 == 0) { /* Centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
2110:       MatCreateVecs(A,NULL,&b);
2111:       VecScatterCreateToZero(b,&mumps->scat_rhs,&mumps->b_seq);
2112:       VecDestroy(&b);
2113:     }
2114:     break;
2115:   }
2116:   PetscMUMPS_c(mumps);
2117:   MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);

2119:   F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS;
2120:   F->ops->solve                 = MatSolve_MUMPS;
2121:   F->ops->solvetranspose        = MatSolve_MUMPS;
2122:   F->ops->matsolve              = MatMatSolve_MUMPS;
2123:   F->ops->mattransposesolve     = MatMatTransposeSolve_MUMPS;
2124: #if defined(PETSC_USE_COMPLEX)
2125:   F->ops->getinertia = NULL;
2126: #else
2127:   F->ops->getinertia = MatGetInertia_SBAIJMUMPS;
2128: #endif
2129:   return(0);
2130: }

2132: PetscErrorCode MatView_MUMPS(Mat A,PetscViewer viewer)
2133: {
2134:   PetscErrorCode    ierr;
2135:   PetscBool         iascii;
2136:   PetscViewerFormat format;
2137:   Mat_MUMPS         *mumps=(Mat_MUMPS*)A->data;

2140:   /* check if matrix is mumps type */
2141:   if (A->ops->solve != MatSolve_MUMPS) return(0);

2143:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
2144:   if (iascii) {
2145:     PetscViewerGetFormat(viewer,&format);
2146:     if (format == PETSC_VIEWER_ASCII_INFO) {
2147:       PetscViewerASCIIPrintf(viewer,"MUMPS run parameters:\n");
2148:       PetscViewerASCIIPrintf(viewer,"  SYM (matrix type):                   %d\n",mumps->id.sym);
2149:       PetscViewerASCIIPrintf(viewer,"  PAR (host participation):            %d\n",mumps->id.par);
2150:       PetscViewerASCIIPrintf(viewer,"  ICNTL(1) (output for error):         %d\n",mumps->id.ICNTL(1));
2151:       PetscViewerASCIIPrintf(viewer,"  ICNTL(2) (output of diagnostic msg): %d\n",mumps->id.ICNTL(2));
2152:       PetscViewerASCIIPrintf(viewer,"  ICNTL(3) (output for global info):   %d\n",mumps->id.ICNTL(3));
2153:       PetscViewerASCIIPrintf(viewer,"  ICNTL(4) (level of printing):        %d\n",mumps->id.ICNTL(4));
2154:       PetscViewerASCIIPrintf(viewer,"  ICNTL(5) (input mat struct):         %d\n",mumps->id.ICNTL(5));
2155:       PetscViewerASCIIPrintf(viewer,"  ICNTL(6) (matrix prescaling):        %d\n",mumps->id.ICNTL(6));
2156:       PetscViewerASCIIPrintf(viewer,"  ICNTL(7) (sequential matrix ordering):%d\n",mumps->id.ICNTL(7));
2157:       PetscViewerASCIIPrintf(viewer,"  ICNTL(8) (scaling strategy):        %d\n",mumps->id.ICNTL(8));
2158:       PetscViewerASCIIPrintf(viewer,"  ICNTL(10) (max num of refinements):  %d\n",mumps->id.ICNTL(10));
2159:       PetscViewerASCIIPrintf(viewer,"  ICNTL(11) (error analysis):          %d\n",mumps->id.ICNTL(11));
2160:       if (mumps->id.ICNTL(11)>0) {
2161:         PetscViewerASCIIPrintf(viewer,"    RINFOG(4) (inf norm of input mat):        %g\n",mumps->id.RINFOG(4));
2162:         PetscViewerASCIIPrintf(viewer,"    RINFOG(5) (inf norm of solution):         %g\n",mumps->id.RINFOG(5));
2163:         PetscViewerASCIIPrintf(viewer,"    RINFOG(6) (inf norm of residual):         %g\n",mumps->id.RINFOG(6));
2164:         PetscViewerASCIIPrintf(viewer,"    RINFOG(7),RINFOG(8) (backward error est): %g, %g\n",mumps->id.RINFOG(7),mumps->id.RINFOG(8));
2165:         PetscViewerASCIIPrintf(viewer,"    RINFOG(9) (error estimate):               %g \n",mumps->id.RINFOG(9));
2166:         PetscViewerASCIIPrintf(viewer,"    RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n",mumps->id.RINFOG(10),mumps->id.RINFOG(11));
2167:       }
2168:       PetscViewerASCIIPrintf(viewer,"  ICNTL(12) (efficiency control):                         %d\n",mumps->id.ICNTL(12));
2169:       PetscViewerASCIIPrintf(viewer,"  ICNTL(13) (sequential factorization of the root node):  %d\n",mumps->id.ICNTL(13));
2170:       PetscViewerASCIIPrintf(viewer,"  ICNTL(14) (percentage of estimated workspace increase): %d\n",mumps->id.ICNTL(14));
2171:       /* ICNTL(15-17) not used */
2172:       PetscViewerASCIIPrintf(viewer,"  ICNTL(18) (input mat struct):                           %d\n",mumps->id.ICNTL(18));
2173:       PetscViewerASCIIPrintf(viewer,"  ICNTL(19) (Schur complement info):                      %d\n",mumps->id.ICNTL(19));
2174:       PetscViewerASCIIPrintf(viewer,"  ICNTL(20) (RHS sparse pattern):                         %d\n",mumps->id.ICNTL(20));
2175:       PetscViewerASCIIPrintf(viewer,"  ICNTL(21) (solution struct):                            %d\n",mumps->id.ICNTL(21));
2176:       PetscViewerASCIIPrintf(viewer,"  ICNTL(22) (in-core/out-of-core facility):               %d\n",mumps->id.ICNTL(22));
2177:       PetscViewerASCIIPrintf(viewer,"  ICNTL(23) (max size of memory can be allocated locally):%d\n",mumps->id.ICNTL(23));

2179:       PetscViewerASCIIPrintf(viewer,"  ICNTL(24) (detection of null pivot rows):               %d\n",mumps->id.ICNTL(24));
2180:       PetscViewerASCIIPrintf(viewer,"  ICNTL(25) (computation of a null space basis):          %d\n",mumps->id.ICNTL(25));
2181:       PetscViewerASCIIPrintf(viewer,"  ICNTL(26) (Schur options for RHS or solution):          %d\n",mumps->id.ICNTL(26));
2182:       PetscViewerASCIIPrintf(viewer,"  ICNTL(27) (blocking size for multiple RHS):             %d\n",mumps->id.ICNTL(27));
2183:       PetscViewerASCIIPrintf(viewer,"  ICNTL(28) (use parallel or sequential ordering):        %d\n",mumps->id.ICNTL(28));
2184:       PetscViewerASCIIPrintf(viewer,"  ICNTL(29) (parallel ordering):                          %d\n",mumps->id.ICNTL(29));

2186:       PetscViewerASCIIPrintf(viewer,"  ICNTL(30) (user-specified set of entries in inv(A)):    %d\n",mumps->id.ICNTL(30));
2187:       PetscViewerASCIIPrintf(viewer,"  ICNTL(31) (factors is discarded in the solve phase):    %d\n",mumps->id.ICNTL(31));
2188:       PetscViewerASCIIPrintf(viewer,"  ICNTL(33) (compute determinant):                        %d\n",mumps->id.ICNTL(33));
2189:       PetscViewerASCIIPrintf(viewer,"  ICNTL(35) (activate BLR based factorization):           %d\n",mumps->id.ICNTL(35));
2190:       PetscViewerASCIIPrintf(viewer,"  ICNTL(36) (choice of BLR factorization variant):        %d\n",mumps->id.ICNTL(36));
2191:       PetscViewerASCIIPrintf(viewer,"  ICNTL(38) (estimated compression rate of LU factors):   %d\n",mumps->id.ICNTL(38));

2193:       PetscViewerASCIIPrintf(viewer,"  CNTL(1) (relative pivoting threshold):      %g \n",mumps->id.CNTL(1));
2194:       PetscViewerASCIIPrintf(viewer,"  CNTL(2) (stopping criterion of refinement): %g \n",mumps->id.CNTL(2));
2195:       PetscViewerASCIIPrintf(viewer,"  CNTL(3) (absolute pivoting threshold):      %g \n",mumps->id.CNTL(3));
2196:       PetscViewerASCIIPrintf(viewer,"  CNTL(4) (value of static pivoting):         %g \n",mumps->id.CNTL(4));
2197:       PetscViewerASCIIPrintf(viewer,"  CNTL(5) (fixation for null pivots):         %g \n",mumps->id.CNTL(5));
2198:       PetscViewerASCIIPrintf(viewer,"  CNTL(7) (dropping parameter for BLR):       %g \n",mumps->id.CNTL(7));

2200:       /* information local to each processor */
2201:       PetscViewerASCIIPrintf(viewer, "  RINFO(1) (local estimated flops for the elimination after analysis): \n");
2202:       PetscViewerASCIIPushSynchronized(viewer);
2203:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %g \n",mumps->myid,mumps->id.RINFO(1));
2204:       PetscViewerFlush(viewer);
2205:       PetscViewerASCIIPrintf(viewer, "  RINFO(2) (local estimated flops for the assembly after factorization): \n");
2206:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d]  %g \n",mumps->myid,mumps->id.RINFO(2));
2207:       PetscViewerFlush(viewer);
2208:       PetscViewerASCIIPrintf(viewer, "  RINFO(3) (local estimated flops for the elimination after factorization): \n");
2209:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d]  %g \n",mumps->myid,mumps->id.RINFO(3));
2210:       PetscViewerFlush(viewer);

2212:       PetscViewerASCIIPrintf(viewer, "  INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization): \n");
2213:       PetscViewerASCIISynchronizedPrintf(viewer,"  [%d] %d\n",mumps->myid,mumps->id.INFO(15));
2214:       PetscViewerFlush(viewer);

2216:       PetscViewerASCIIPrintf(viewer, "  INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization): \n");
2217:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d\n",mumps->myid,mumps->id.INFO(16));
2218:       PetscViewerFlush(viewer);

2220:       PetscViewerASCIIPrintf(viewer, "  INFO(23) (num of pivots eliminated on this processor after factorization): \n");
2221:       PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d\n",mumps->myid,mumps->id.INFO(23));
2222:       PetscViewerFlush(viewer);

2224:       if (mumps->ninfo && mumps->ninfo <= 80) {
2225:         PetscInt i;
2226:         for (i=0; i<mumps->ninfo; i++) {
2227:           PetscViewerASCIIPrintf(viewer, "  INFO(%d): \n",mumps->info[i]);
2228:           PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d\n",mumps->myid,mumps->id.INFO(mumps->info[i]));
2229:           PetscViewerFlush(viewer);
2230:         }
2231:       }
2232:       PetscViewerASCIIPopSynchronized(viewer);

2234:       if (!mumps->myid) { /* information from the host */
2235:         PetscViewerASCIIPrintf(viewer,"  RINFOG(1) (global estimated flops for the elimination after analysis): %g \n",mumps->id.RINFOG(1));
2236:         PetscViewerASCIIPrintf(viewer,"  RINFOG(2) (global estimated flops for the assembly after factorization): %g \n",mumps->id.RINFOG(2));
2237:         PetscViewerASCIIPrintf(viewer,"  RINFOG(3) (global estimated flops for the elimination after factorization): %g \n",mumps->id.RINFOG(3));
2238:         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));

2240:         PetscViewerASCIIPrintf(viewer,"  INFOG(3) (estimated real workspace for factors on all processors after analysis): %d\n",mumps->id.INFOG(3));
2241:         PetscViewerASCIIPrintf(viewer,"  INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d\n",mumps->id.INFOG(4));
2242:         PetscViewerASCIIPrintf(viewer,"  INFOG(5) (estimated maximum front size in the complete tree): %d\n",mumps->id.INFOG(5));
2243:         PetscViewerASCIIPrintf(viewer,"  INFOG(6) (number of nodes in the complete tree): %d\n",mumps->id.INFOG(6));
2244:         PetscViewerASCIIPrintf(viewer,"  INFOG(7) (ordering option effectively used after analysis): %d\n",mumps->id.INFOG(7));
2245:         PetscViewerASCIIPrintf(viewer,"  INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d\n",mumps->id.INFOG(8));
2246:         PetscViewerASCIIPrintf(viewer,"  INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d\n",mumps->id.INFOG(9));
2247:         PetscViewerASCIIPrintf(viewer,"  INFOG(10) (total integer space store the matrix factors after factorization): %d\n",mumps->id.INFOG(10));
2248:         PetscViewerASCIIPrintf(viewer,"  INFOG(11) (order of largest frontal matrix after factorization): %d\n",mumps->id.INFOG(11));
2249:         PetscViewerASCIIPrintf(viewer,"  INFOG(12) (number of off-diagonal pivots): %d\n",mumps->id.INFOG(12));
2250:         PetscViewerASCIIPrintf(viewer,"  INFOG(13) (number of delayed pivots after factorization): %d\n",mumps->id.INFOG(13));
2251:         PetscViewerASCIIPrintf(viewer,"  INFOG(14) (number of memory compress after factorization): %d\n",mumps->id.INFOG(14));
2252:         PetscViewerASCIIPrintf(viewer,"  INFOG(15) (number of steps of iterative refinement after solution): %d\n",mumps->id.INFOG(15));
2253:         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));
2254:         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));
2255:         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));
2256:         PetscViewerASCIIPrintf(viewer,"  INFOG(19) (size of all MUMPS internal data allocated during factorization: sum over all processors): %d\n",mumps->id.INFOG(19));
2257:         PetscViewerASCIIPrintf(viewer,"  INFOG(20) (estimated number of entries in the factors): %d\n",mumps->id.INFOG(20));
2258:         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));
2259:         PetscViewerASCIIPrintf(viewer,"  INFOG(22) (size in MB of memory effectively used during factorization - sum over all processors): %d\n",mumps->id.INFOG(22));
2260:         PetscViewerASCIIPrintf(viewer,"  INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d\n",mumps->id.INFOG(23));
2261:         PetscViewerASCIIPrintf(viewer,"  INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d\n",mumps->id.INFOG(24));
2262:         PetscViewerASCIIPrintf(viewer,"  INFOG(25) (after factorization: number of pivots modified by static pivoting): %d\n",mumps->id.INFOG(25));
2263:         PetscViewerASCIIPrintf(viewer,"  INFOG(28) (after factorization: number of null pivots encountered): %d\n",mumps->id.INFOG(28));
2264:         PetscViewerASCIIPrintf(viewer,"  INFOG(29) (after factorization: effective number of entries in the factors (sum over all processors)): %d\n",mumps->id.INFOG(29));
2265:         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));
2266:         PetscViewerASCIIPrintf(viewer,"  INFOG(32) (after analysis: type of analysis done): %d\n",mumps->id.INFOG(32));
2267:         PetscViewerASCIIPrintf(viewer,"  INFOG(33) (value used for ICNTL(8)): %d\n",mumps->id.INFOG(33));
2268:         PetscViewerASCIIPrintf(viewer,"  INFOG(34) (exponent of the determinant if determinant is requested): %d\n",mumps->id.INFOG(34));
2269:         PetscViewerASCIIPrintf(viewer,"  INFOG(35) (after factorization: number of entries taking into account BLR factor compression - sum over all processors): %d\n",mumps->id.INFOG(35));
2270:         PetscViewerASCIIPrintf(viewer,"  INFOG(36) (after analysis: estimated size of all MUMPS internal data for running BLR in-core - value on the most memory consuming processor): %d\n",mumps->id.INFOG(36));
2271:         PetscViewerASCIIPrintf(viewer,"  INFOG(37) (after analysis: estimated size of all MUMPS internal data for running BLR in-core - sum over all processors): %d\n",mumps->id.INFOG(37));
2272:         PetscViewerASCIIPrintf(viewer,"  INFOG(38) (after analysis: estimated size of all MUMPS internal data for running BLR out-of-core - value on the most memory consuming processor): %d\n",mumps->id.INFOG(38));
2273:         PetscViewerASCIIPrintf(viewer,"  INFOG(39) (after analysis: estimated size of all MUMPS internal data for running BLR out-of-core - sum over all processors): %d\n",mumps->id.INFOG(39));
2274:       }
2275:     }
2276:   }
2277:   return(0);
2278: }

2280: PetscErrorCode MatGetInfo_MUMPS(Mat A,MatInfoType flag,MatInfo *info)
2281: {
2282:   Mat_MUMPS *mumps =(Mat_MUMPS*)A->data;

2285:   info->block_size        = 1.0;
2286:   info->nz_allocated      = mumps->id.INFOG(20);
2287:   info->nz_used           = mumps->id.INFOG(20);
2288:   info->nz_unneeded       = 0.0;
2289:   info->assemblies        = 0.0;
2290:   info->mallocs           = 0.0;
2291:   info->memory            = 0.0;
2292:   info->fill_ratio_given  = 0;
2293:   info->fill_ratio_needed = 0;
2294:   info->factor_mallocs    = 0;
2295:   return(0);
2296: }

2298: /* -------------------------------------------------------------------------------------------*/
2299: PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is)
2300: {
2301:   Mat_MUMPS         *mumps =(Mat_MUMPS*)F->data;
2302:   const PetscScalar *arr;
2303:   const PetscInt    *idxs;
2304:   PetscInt          size,i;
2305:   PetscErrorCode    ierr;

2308:   ISGetLocalSize(is,&size);
2309:   if (mumps->petsc_size > 1) {
2310:     PetscBool ls,gs; /* gs is false if any rank other than root has non-empty IS */

2312:     ls   = mumps->myid ? (size ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE; /* always true on root; false on others if their size != 0 */
2313:     MPI_Allreduce(&ls,&gs,1,MPIU_BOOL,MPI_LAND,mumps->petsc_comm);
2314:     if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MUMPS distributed parallel Schur complements not yet supported from PETSc\n");
2315:   }

2317:   /* Schur complement matrix */
2318:   MatDestroy(&F->schur);
2319:   MatCreateSeqDense(PETSC_COMM_SELF,size,size,NULL,&F->schur);
2320:   MatDenseGetArrayRead(F->schur,&arr);
2321:   mumps->id.schur      = (MumpsScalar*)arr;
2322:   mumps->id.size_schur = size;
2323:   mumps->id.schur_lld  = size;
2324:   MatDenseRestoreArrayRead(F->schur,&arr);
2325:   if (mumps->sym == 1) {
2326:     MatSetOption(F->schur,MAT_SPD,PETSC_TRUE);
2327:   }

2329:   /* MUMPS expects Fortran style indices */
2330:   PetscFree(mumps->id.listvar_schur);
2331:   PetscMalloc1(size,&mumps->id.listvar_schur);
2332:   ISGetIndices(is,&idxs);
2333:   for (i=0; i<size; i++) {PetscMUMPSIntCast(idxs[i]+1,&(mumps->id.listvar_schur[i]));}
2334:   ISRestoreIndices(is,&idxs);
2335:   if (mumps->petsc_size > 1) {
2336:     mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */
2337:   } else {
2338:     if (F->factortype == MAT_FACTOR_LU) {
2339:       mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */
2340:     } else {
2341:       mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */
2342:     }
2343:   }
2344:   /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */
2345:   mumps->id.ICNTL(26) = -1;
2346:   return(0);
2347: }

2349: /* -------------------------------------------------------------------------------------------*/
2350: PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F,Mat* S)
2351: {
2352:   Mat            St;
2353:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
2354:   PetscScalar    *array;
2355: #if defined(PETSC_USE_COMPLEX)
2356:   PetscScalar    im = PetscSqrtScalar((PetscScalar)-1.0);
2357: #endif

2361:   if (!mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ORDER,"Schur complement mode not selected! You should call MatFactorSetSchurIS to enable it");
2362:   MatCreate(PETSC_COMM_SELF,&St);
2363:   MatSetSizes(St,PETSC_DECIDE,PETSC_DECIDE,mumps->id.size_schur,mumps->id.size_schur);
2364:   MatSetType(St,MATDENSE);
2365:   MatSetUp(St);
2366:   MatDenseGetArray(St,&array);
2367:   if (!mumps->sym) { /* MUMPS always return a full matrix */
2368:     if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
2369:       PetscInt i,j,N=mumps->id.size_schur;
2370:       for (i=0;i<N;i++) {
2371:         for (j=0;j<N;j++) {
2372: #if !defined(PETSC_USE_COMPLEX)
2373:           PetscScalar val = mumps->id.schur[i*N+j];
2374: #else
2375:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
2376: #endif
2377:           array[j*N+i] = val;
2378:         }
2379:       }
2380:     } else { /* stored by columns */
2381:       PetscArraycpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur);
2382:     }
2383:   } else { /* either full or lower-triangular (not packed) */
2384:     if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */
2385:       PetscInt i,j,N=mumps->id.size_schur;
2386:       for (i=0;i<N;i++) {
2387:         for (j=i;j<N;j++) {
2388: #if !defined(PETSC_USE_COMPLEX)
2389:           PetscScalar val = mumps->id.schur[i*N+j];
2390: #else
2391:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
2392: #endif
2393:           array[i*N+j] = val;
2394:           array[j*N+i] = val;
2395:         }
2396:       }
2397:     } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */
2398:       PetscArraycpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur);
2399:     } else { /* ICNTL(19) == 1 lower triangular stored by rows */
2400:       PetscInt i,j,N=mumps->id.size_schur;
2401:       for (i=0;i<N;i++) {
2402:         for (j=0;j<i+1;j++) {
2403: #if !defined(PETSC_USE_COMPLEX)
2404:           PetscScalar val = mumps->id.schur[i*N+j];
2405: #else
2406:           PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
2407: #endif
2408:           array[i*N+j] = val;
2409:           array[j*N+i] = val;
2410:         }
2411:       }
2412:     }
2413:   }
2414:   MatDenseRestoreArray(St,&array);
2415:   *S   = St;
2416:   return(0);
2417: }

2419: /* -------------------------------------------------------------------------------------------*/
2420: PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt ival)
2421: {
2423:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2426:   PetscMUMPSIntCast(ival,&mumps->id.ICNTL(icntl));
2427:   return(0);
2428: }

2430: PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt *ival)
2431: {
2432:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2435:   *ival = mumps->id.ICNTL(icntl);
2436:   return(0);
2437: }

2439: /*@
2440:   MatMumpsSetIcntl - Set MUMPS parameter ICNTL()

2442:    Logically Collective on Mat

2444:    Input Parameters:
2445: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2446: .  icntl - index of MUMPS parameter array ICNTL()
2447: -  ival - value of MUMPS ICNTL(icntl)

2449:   Options Database:
2450: .   -mat_mumps_icntl_<icntl> <ival>

2452:    Level: beginner

2454:    References:
2455: .     MUMPS Users' Guide

2457: .seealso: MatGetFactor(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2458:  @*/
2459: PetscErrorCode MatMumpsSetIcntl(Mat F,PetscInt icntl,PetscInt ival)
2460: {

2465:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2468:   PetscTryMethod(F,"MatMumpsSetIcntl_C",(Mat,PetscInt,PetscInt),(F,icntl,ival));
2469:   return(0);
2470: }

2472: /*@
2473:   MatMumpsGetIcntl - Get MUMPS parameter ICNTL()

2475:    Logically Collective on Mat

2477:    Input Parameters:
2478: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2479: -  icntl - index of MUMPS parameter array ICNTL()

2481:   Output Parameter:
2482: .  ival - value of MUMPS ICNTL(icntl)

2484:    Level: beginner

2486:    References:
2487: .     MUMPS Users' Guide

2489: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2490: @*/
2491: PetscErrorCode MatMumpsGetIcntl(Mat F,PetscInt icntl,PetscInt *ival)
2492: {

2497:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2500:   PetscUseMethod(F,"MatMumpsGetIcntl_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2501:   return(0);
2502: }

2504: /* -------------------------------------------------------------------------------------------*/
2505: PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal val)
2506: {
2507:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2510:   mumps->id.CNTL(icntl) = val;
2511:   return(0);
2512: }

2514: PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal *val)
2515: {
2516:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2519:   *val = mumps->id.CNTL(icntl);
2520:   return(0);
2521: }

2523: /*@
2524:   MatMumpsSetCntl - Set MUMPS parameter CNTL()

2526:    Logically Collective on Mat

2528:    Input Parameters:
2529: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2530: .  icntl - index of MUMPS parameter array CNTL()
2531: -  val - value of MUMPS CNTL(icntl)

2533:   Options Database:
2534: .   -mat_mumps_cntl_<icntl> <val>

2536:    Level: beginner

2538:    References:
2539: .     MUMPS Users' Guide

2541: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2542: @*/
2543: PetscErrorCode MatMumpsSetCntl(Mat F,PetscInt icntl,PetscReal val)
2544: {

2549:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2552:   PetscTryMethod(F,"MatMumpsSetCntl_C",(Mat,PetscInt,PetscReal),(F,icntl,val));
2553:   return(0);
2554: }

2556: /*@
2557:   MatMumpsGetCntl - Get MUMPS parameter CNTL()

2559:    Logically Collective on Mat

2561:    Input Parameters:
2562: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2563: -  icntl - index of MUMPS parameter array CNTL()

2565:   Output Parameter:
2566: .  val - value of MUMPS CNTL(icntl)

2568:    Level: beginner

2570:    References:
2571: .      MUMPS Users' Guide

2573: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2574: @*/
2575: PetscErrorCode MatMumpsGetCntl(Mat F,PetscInt icntl,PetscReal *val)
2576: {

2581:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2584:   PetscUseMethod(F,"MatMumpsGetCntl_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2585:   return(0);
2586: }

2588: PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F,PetscInt icntl,PetscInt *info)
2589: {
2590:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2593:   *info = mumps->id.INFO(icntl);
2594:   return(0);
2595: }

2597: PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F,PetscInt icntl,PetscInt *infog)
2598: {
2599:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2602:   *infog = mumps->id.INFOG(icntl);
2603:   return(0);
2604: }

2606: PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfo)
2607: {
2608:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2611:   *rinfo = mumps->id.RINFO(icntl);
2612:   return(0);
2613: }

2615: PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfog)
2616: {
2617:   Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;

2620:   *rinfog = mumps->id.RINFOG(icntl);
2621:   return(0);
2622: }

2624: PetscErrorCode MatMumpsGetInverse_MUMPS(Mat F,Mat spRHS)
2625: {
2627:   Mat            Bt = NULL,Btseq = NULL;
2628:   PetscBool      flg;
2629:   Mat_MUMPS      *mumps =(Mat_MUMPS*)F->data;
2630:   PetscScalar    *aa;
2631:   PetscInt       spnr,*ia,*ja,M,nrhs;

2635:   PetscObjectTypeCompare((PetscObject)spRHS,MATTRANSPOSEMAT,&flg);
2636:   if (flg) {
2637:     MatTransposeGetMat(spRHS,&Bt);
2638:   } else SETERRQ(PetscObjectComm((PetscObject)spRHS),PETSC_ERR_ARG_WRONG,"Matrix spRHS must be type MATTRANSPOSEMAT matrix");

2640:   MatMumpsSetIcntl(F,30,1);

2642:   if (mumps->petsc_size > 1) {
2643:     Mat_MPIAIJ *b = (Mat_MPIAIJ*)Bt->data;
2644:     Btseq = b->A;
2645:   } else {
2646:     Btseq = Bt;
2647:   }

2649:   MatGetSize(spRHS,&M,&nrhs);
2650:   mumps->id.nrhs = nrhs;
2651:   mumps->id.lrhs = M;
2652:   mumps->id.rhs  = NULL;

2654:   if (!mumps->myid) {
2655:     MatSeqAIJGetArray(Btseq,&aa);
2656:     MatGetRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
2657:     if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
2658:     PetscMUMPSIntCSRCast(mumps,spnr,ia,ja,&mumps->id.irhs_ptr,&mumps->id.irhs_sparse,&mumps->id.nz_rhs);
2659:     mumps->id.rhs_sparse  = (MumpsScalar*)aa;
2660:   } else {
2661:     mumps->id.irhs_ptr    = NULL;
2662:     mumps->id.irhs_sparse = NULL;
2663:     mumps->id.nz_rhs      = 0;
2664:     mumps->id.rhs_sparse  = NULL;
2665:   }
2666:   mumps->id.ICNTL(20)   = 1; /* rhs is sparse */
2667:   mumps->id.ICNTL(21)   = 0; /* solution is in assembled centralized format */

2669:   /* solve phase */
2670:   /*-------------*/
2671:   mumps->id.job = JOB_SOLVE;
2672:   PetscMUMPS_c(mumps);
2673:   if (mumps->id.INFOG(1) < 0)
2674:     SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in solve phase: INFOG(1)=%d INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));

2676:   if (!mumps->myid) {
2677:     MatSeqAIJRestoreArray(Btseq,&aa);
2678:     MatRestoreRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
2679:     if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
2680:   }
2681:   return(0);
2682: }

2684: /*@
2685:   MatMumpsGetInverse - Get user-specified set of entries in inverse of A

2687:    Logically Collective on Mat

2689:    Input Parameters:
2690: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2691: -  spRHS - sequential sparse matrix in MATTRANSPOSEMAT format holding specified indices in processor[0]

2693:   Output Parameter:
2694: . spRHS - requested entries of inverse of A

2696:    Level: beginner

2698:    References:
2699: .      MUMPS Users' Guide

2701: .seealso: MatGetFactor(), MatCreateTranspose()
2702: @*/
2703: PetscErrorCode MatMumpsGetInverse(Mat F,Mat spRHS)
2704: {

2709:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2710:   PetscUseMethod(F,"MatMumpsGetInverse_C",(Mat,Mat),(F,spRHS));
2711:   return(0);
2712: }

2714: PetscErrorCode MatMumpsGetInverseTranspose_MUMPS(Mat F,Mat spRHST)
2715: {
2717:   Mat            spRHS;

2720:   MatCreateTranspose(spRHST,&spRHS);
2721:   MatMumpsGetInverse_MUMPS(F,spRHS);
2722:   MatDestroy(&spRHS);
2723:   return(0);
2724: }

2726: /*@
2727:   MatMumpsGetInverseTranspose - Get user-specified set of entries in inverse of matrix A^T

2729:    Logically Collective on Mat

2731:    Input Parameters:
2732: +  F - the factored matrix of A obtained by calling MatGetFactor() from PETSc-MUMPS interface
2733: -  spRHST - sequential sparse matrix in MATAIJ format holding specified indices of A^T in processor[0]

2735:   Output Parameter:
2736: . spRHST - requested entries of inverse of A^T

2738:    Level: beginner

2740:    References:
2741: .      MUMPS Users' Guide

2743: .seealso: MatGetFactor(), MatCreateTranspose(), MatMumpsGetInverse()
2744: @*/
2745: PetscErrorCode MatMumpsGetInverseTranspose(Mat F,Mat spRHST)
2746: {
2748:   PetscBool      flg;

2752:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2753:   PetscObjectTypeCompareAny((PetscObject)spRHST,&flg,MATSEQAIJ,MATMPIAIJ,NULL);
2754:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)spRHST),PETSC_ERR_ARG_WRONG,"Matrix spRHST must be MATAIJ matrix");

2756:   PetscUseMethod(F,"MatMumpsGetInverseTranspose_C",(Mat,Mat),(F,spRHST));
2757:   return(0);
2758: }

2760: /*@
2761:   MatMumpsGetInfo - Get MUMPS parameter INFO()

2763:    Logically Collective on Mat

2765:    Input Parameters:
2766: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2767: -  icntl - index of MUMPS parameter array INFO()

2769:   Output Parameter:
2770: .  ival - value of MUMPS INFO(icntl)

2772:    Level: beginner

2774:    References:
2775: .      MUMPS Users' Guide

2777: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2778: @*/
2779: PetscErrorCode MatMumpsGetInfo(Mat F,PetscInt icntl,PetscInt *ival)
2780: {

2785:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2787:   PetscUseMethod(F,"MatMumpsGetInfo_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2788:   return(0);
2789: }

2791: /*@
2792:   MatMumpsGetInfog - Get MUMPS parameter INFOG()

2794:    Logically Collective on Mat

2796:    Input Parameters:
2797: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2798: -  icntl - index of MUMPS parameter array INFOG()

2800:   Output Parameter:
2801: .  ival - value of MUMPS INFOG(icntl)

2803:    Level: beginner

2805:    References:
2806: .      MUMPS Users' Guide

2808: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2809: @*/
2810: PetscErrorCode MatMumpsGetInfog(Mat F,PetscInt icntl,PetscInt *ival)
2811: {

2816:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2818:   PetscUseMethod(F,"MatMumpsGetInfog_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2819:   return(0);
2820: }

2822: /*@
2823:   MatMumpsGetRinfo - Get MUMPS parameter RINFO()

2825:    Logically Collective on Mat

2827:    Input Parameters:
2828: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2829: -  icntl - index of MUMPS parameter array RINFO()

2831:   Output Parameter:
2832: .  val - value of MUMPS RINFO(icntl)

2834:    Level: beginner

2836:    References:
2837: .       MUMPS Users' Guide

2839: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfog()
2840: @*/
2841: PetscErrorCode MatMumpsGetRinfo(Mat F,PetscInt icntl,PetscReal *val)
2842: {

2847:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2849:   PetscUseMethod(F,"MatMumpsGetRinfo_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2850:   return(0);
2851: }

2853: /*@
2854:   MatMumpsGetRinfog - Get MUMPS parameter RINFOG()

2856:    Logically Collective on Mat

2858:    Input Parameters:
2859: +  F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2860: -  icntl - index of MUMPS parameter array RINFOG()

2862:   Output Parameter:
2863: .  val - value of MUMPS RINFOG(icntl)

2865:    Level: beginner

2867:    References:
2868: .      MUMPS Users' Guide

2870: .seealso: MatGetFactor(), MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo()
2871: @*/
2872: PetscErrorCode MatMumpsGetRinfog(Mat F,PetscInt icntl,PetscReal *val)
2873: {

2878:   if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2880:   PetscUseMethod(F,"MatMumpsGetRinfog_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2881:   return(0);
2882: }

2884: /*MC
2885:   MATSOLVERMUMPS -  A matrix type providing direct solvers (LU and Cholesky) for
2886:   distributed and sequential matrices via the external package MUMPS.

2888:   Works with MATAIJ and MATSBAIJ matrices

2890:   Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch to have PETSc installed with MUMPS

2892:   Use ./configure --with-openmp --download-hwloc (or --with-hwloc) to enable running MUMPS in MPI+OpenMP hybrid mode and non-MUMPS in flat-MPI mode. See details below.

2894:   Use -pc_type cholesky or lu -pc_factor_mat_solver_type mumps to use this direct solver

2896:   Options Database Keys:
2897: +  -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages
2898: .  -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning
2899: .  -mat_mumps_icntl_3 -  ICNTL(3): output stream for global information, collected on the host
2900: .  -mat_mumps_icntl_4 -  ICNTL(4): level of printing (0 to 4)
2901: .  -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)
2902: .  -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis, 0=AMD, 2=AMF, 3=Scotch, 4=PORD, 5=Metis, 6=QAMD, and 7=auto
2903:                         Use -pc_factor_mat_ordering_type <type> to have PETSc perform the ordering (sequential only)
2904: .  -mat_mumps_icntl_8  - ICNTL(8): scaling strategy (-2 to 8 or 77)
2905: .  -mat_mumps_icntl_10  - ICNTL(10): max num of refinements
2906: .  -mat_mumps_icntl_11  - ICNTL(11): statistics related to an error analysis (via -ksp_view)
2907: .  -mat_mumps_icntl_12  - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)
2908: .  -mat_mumps_icntl_13  - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting
2909: .  -mat_mumps_icntl_14  - ICNTL(14): percentage increase in the estimated working space
2910: .  -mat_mumps_icntl_19  - ICNTL(19): computes the Schur complement
2911: .  -mat_mumps_icntl_20  - ICNTL(20): give MUMPS centralized (0) or distributed (10) dense RHS
2912: .  -mat_mumps_icntl_22  - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)
2913: .  -mat_mumps_icntl_23  - ICNTL(23): max size of the working memory (MB) that can allocate per processor
2914: .  -mat_mumps_icntl_24  - ICNTL(24): detection of null pivot rows (0 or 1)
2915: .  -mat_mumps_icntl_25  - ICNTL(25): compute a solution of a deficient matrix and a null space basis
2916: .  -mat_mumps_icntl_26  - ICNTL(26): drives the solution phase if a Schur complement matrix
2917: .  -mat_mumps_icntl_28  - ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering
2918: .  -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis
2919: .  -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A)
2920: .  -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization
2921: .  -mat_mumps_icntl_33 - ICNTL(33): compute determinant
2922: .  -mat_mumps_icntl_35 - ICNTL(35): level of activation of BLR (Block Low-Rank) feature
2923: .  -mat_mumps_icntl_36 - ICNTL(36): controls the choice of BLR factorization variant
2924: .  -mat_mumps_icntl_38 - ICNTL(38): sets the estimated compression rate of LU factors with BLR
2925: .  -mat_mumps_cntl_1  - CNTL(1): relative pivoting threshold
2926: .  -mat_mumps_cntl_2  -  CNTL(2): stopping criterion of refinement
2927: .  -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold
2928: .  -mat_mumps_cntl_4 - CNTL(4): value for static pivoting
2929: .  -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots
2930: .  -mat_mumps_cntl_7 - CNTL(7): precision of the dropping parameter used during BLR factorization
2931: -  -mat_mumps_use_omp_threads [m] - run MUMPS in MPI+OpenMP hybrid mode as if omp_set_num_threads(m) is called before calling MUMPS.
2932:                                    Default might be the number of cores per CPU package (socket) as reported by hwloc and suggested by the MUMPS manual.

2934:   Level: beginner

2936:     Notes:
2937:     MUMPS Cholesky does not handle (complex) Hermitian matrices http://mumps.enseeiht.fr/doc/userguide_5.2.1.pdf so using it will error if the matrix is Hermitian.

2939:     When a MUMPS factorization fails inside a KSP solve, for example with a KSP_DIVERGED_PC_FAILED, one can find the MUMPS information about the failure by calling
2940: $          KSPGetPC(ksp,&pc);
2941: $          PCFactorGetMatrix(pc,&mat);
2942: $          MatMumpsGetInfo(mat,....);
2943: $          MatMumpsGetInfog(mat,....); etc.
2944:            Or you can run with -ksp_error_if_not_converged and the program will be stopped and the information printed in the error message.

2946:   Using MUMPS with 64-bit integers
2947:     MUMPS provides 64-bit integer support in two build modes:
2948:       full 64-bit: here MUMPS is built with C preprocessing flag -DINTSIZE64 and Fortran compiler option -i8, -fdefault-integer-8 or equivalent, and
2949:       requires all dependent libraries MPI, ScaLAPACK, LAPACK and BLAS built the same way with 64-bit integers (for example ILP64 Intel MKL and MPI).

2951:       selective 64-bit: with the default MUMPS build, 64-bit integers have been introduced where needed. In compressed sparse row (CSR) storage of matrices,
2952:       MUMPS stores column indices in 32-bit, but row offsets in 64-bit, so you can have a huge number of non-zeros, but must have less than 2^31 rows and
2953:       columns. This can lead to significant memory and performance gains with respect to a full 64-bit integer MUMPS version. This requires a regular (32-bit
2954:       integer) build of all dependent libraries MPI, ScaLAPACK, LAPACK and BLAS.

2956:     With --download-mumps=1, PETSc always build MUMPS in selective 64-bit mode, which can be used by both --with-64-bit-indices=0/1 variants of PETSc.

2958:   Two modes to run MUMPS/PETSc with OpenMP
2959: $     Set OMP_NUM_THREADS and run with fewer MPI ranks than cores. For example, if you want to have 16 OpenMP
2960: $     threads per rank, then you may use "export OMP_NUM_THREADS=16 && mpirun -n 4 ./test".

2962: $     -mat_mumps_use_omp_threads [m] and run your code with as many MPI ranks as the number of cores. For example,
2963: $     if a compute node has 32 cores and you run on two nodes, you may use "mpirun -n 64 ./test -mat_mumps_use_omp_threads 16"

2965:    To run MUMPS in MPI+OpenMP hybrid mode (i.e., enable multithreading in MUMPS), but still run the non-MUMPS part
2966:    (i.e., PETSc part) of your code in the so-called flat-MPI (aka pure-MPI) mode, you need to configure PETSc with --with-openmp --download-hwloc
2967:    (or --with-hwloc), and have an MPI that supports MPI-3.0's process shared memory (which is usually available). Since MUMPS calls BLAS
2968:    libraries, to really get performance, you should have multithreaded BLAS libraries such as Intel MKL, AMD ACML, Cray libSci or OpenBLAS
2969:    (PETSc will automatically try to utilized a threaded BLAS if --with-openmp is provided).

2971:    If you run your code through a job submission system, there are caveats in MPI rank mapping. We use MPI_Comm_split_type() to obtain MPI
2972:    processes on each compute node. Listing the processes in rank ascending order, we split processes on a node into consecutive groups of
2973:    size m and create a communicator called omp_comm for each group. Rank 0 in an omp_comm is called the master rank, and others in the omp_comm
2974:    are called slave ranks (or slaves). Only master ranks are seen to MUMPS and slaves are not. We will free CPUs assigned to slaves (might be set
2975:    by CPU binding policies in job scripts) and make the CPUs available to the master so that OMP threads spawned by MUMPS can run on the CPUs.
2976:    In a multi-socket compute node, MPI rank mapping is an issue. Still use the above example and suppose your compute node has two sockets,
2977:    if you interleave MPI ranks on the two sockets, in other words, even ranks are placed on socket 0, and odd ranks are on socket 1, and bind
2978:    MPI ranks to cores, then with -mat_mumps_use_omp_threads 16, a master rank (and threads it spawns) will use half cores in socket 0, and half
2979:    cores in socket 1, that definitely hurts locality. On the other hand, if you map MPI ranks consecutively on the two sockets, then the
2980:    problem will not happen. Therefore, when you use -mat_mumps_use_omp_threads, you need to keep an eye on your MPI rank mapping and CPU binding.
2981:    For example, with the Slurm job scheduler, one can use srun --cpu-bind=verbose -m block:block to map consecutive MPI ranks to sockets and
2982:    examine the mapping result.

2984:    PETSc does not control thread binding in MUMPS. So to get best performance, one still has to set OMP_PROC_BIND and OMP_PLACES in job scripts,
2985:    for example, export OMP_PLACES=threads and export OMP_PROC_BIND=spread. One does not need to export OMP_NUM_THREADS=m in job scripts as PETSc
2986:    calls omp_set_num_threads(m) internally before calling MUMPS.

2988:    References:
2989: +   1. - Heroux, Michael A., R. Brightwell, and Michael M. Wolf. "Bi-modal MPI and MPI+ threads computing on scalable multicore systems." IJHPCA (Submitted) (2011).
2990: -   2. - Gutierrez, Samuel K., et al. "Accommodating Thread-Level Heterogeneity in Coupled Parallel Applications." Parallel and Distributed Processing Symposium (IPDPS), 2017 IEEE International. IEEE, 2017.

2992: .seealso: PCFactorSetMatSolverType(), MatSolverType, MatMumpsSetIcntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog(), KSPGetPC(), PCFactorGetMatrix()

2994: M*/

2996: static PetscErrorCode MatFactorGetSolverType_mumps(Mat A,MatSolverType *type)
2997: {
2999:   *type = MATSOLVERMUMPS;
3000:   return(0);
3001: }

3003: /* MatGetFactor for Seq and MPI AIJ matrices */
3004: static PetscErrorCode MatGetFactor_aij_mumps(Mat A,MatFactorType ftype,Mat *F)
3005: {
3006:   Mat            B;
3008:   Mat_MUMPS      *mumps;
3009:   PetscBool      isSeqAIJ;
3010:   PetscMPIInt    size;

3013:  #if defined(PETSC_USE_COMPLEX)
3014:   if (A->hermitian && !A->symmetric && ftype == MAT_FACTOR_CHOLESKY) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Hermitian CHOLESKY Factor is not supported");
3015:  #endif
3016:   /* Create the factorization matrix */
3017:   PetscObjectBaseTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);
3018:   MatCreate(PetscObjectComm((PetscObject)A),&B);
3019:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
3020:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
3021:   MatSetUp(B);

3023:   PetscNewLog(B,&mumps);

3025:   B->ops->view    = MatView_MUMPS;
3026:   B->ops->getinfo = MatGetInfo_MUMPS;

3028:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
3029:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
3030:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
3031:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
3032:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
3033:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
3034:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
3035:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
3036:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
3037:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
3038:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
3039:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
3040:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);

3042:   if (ftype == MAT_FACTOR_LU) {
3043:     B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
3044:     B->factortype            = MAT_FACTOR_LU;
3045:     if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij;
3046:     else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij;
3047:     PetscStrallocpy(MATORDERINGEXTERNAL,(char**)&B->preferredordering[MAT_FACTOR_LU]);
3048:     mumps->sym = 0;
3049:   } else {
3050:     B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
3051:     B->factortype                  = MAT_FACTOR_CHOLESKY;
3052:     if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij;
3053:     else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij;
3054:     PetscStrallocpy(MATORDERINGEXTERNAL,(char**)&B->preferredordering[MAT_FACTOR_CHOLESKY]);
3055: #if defined(PETSC_USE_COMPLEX)
3056:     mumps->sym = 2;
3057: #else
3058:     if (A->spd_set && A->spd) mumps->sym = 1;
3059:     else                      mumps->sym = 2;
3060: #endif
3061:   }

3063:   /* set solvertype */
3064:   PetscFree(B->solvertype);
3065:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
3066:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
3067:   if (size == 1) {
3068:     /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3069:     B->canuseordering = PETSC_TRUE;
3070:   }
3071:   B->ops->destroy = MatDestroy_MUMPS;
3072:   B->data         = (void*)mumps;

3074:   PetscInitializeMUMPS(A,mumps);

3076:   *F = B;
3077:   return(0);
3078: }

3080: /* MatGetFactor for Seq and MPI SBAIJ matrices */
3081: static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A,MatFactorType ftype,Mat *F)
3082: {
3083:   Mat            B;
3085:   Mat_MUMPS      *mumps;
3086:   PetscBool      isSeqSBAIJ;
3087:   PetscMPIInt    size;

3090:  #if defined(PETSC_USE_COMPLEX)
3091:   if (A->hermitian && !A->symmetric) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Hermitian CHOLESKY Factor is not supported");
3092:  #endif
3093:   MatCreate(PetscObjectComm((PetscObject)A),&B);
3094:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
3095:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
3096:   MatSetUp(B);

3098:   PetscNewLog(B,&mumps);
3099:   PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isSeqSBAIJ);
3100:   if (isSeqSBAIJ) {
3101:     mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij;
3102:   } else {
3103:     mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij;
3104:   }

3106:   B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
3107:   B->ops->view                   = MatView_MUMPS;
3108:   B->ops->getinfo                = MatGetInfo_MUMPS;

3110:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
3111:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
3112:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
3113:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
3114:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
3115:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
3116:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
3117:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
3118:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
3119:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
3120:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
3121:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
3122:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);

3124:   B->factortype = MAT_FACTOR_CHOLESKY;
3125: #if defined(PETSC_USE_COMPLEX)
3126:   mumps->sym = 2;
3127: #else
3128:   if (A->spd_set && A->spd) mumps->sym = 1;
3129:   else                      mumps->sym = 2;
3130: #endif

3132:   /* set solvertype */
3133:   PetscFree(B->solvertype);
3134:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
3135:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
3136:   if (size == 1) {
3137:     /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3138:     B->canuseordering = PETSC_TRUE;
3139:   }
3140:   PetscStrallocpy(MATORDERINGEXTERNAL,(char**)&B->preferredordering[MAT_FACTOR_CHOLESKY]);
3141:   B->ops->destroy = MatDestroy_MUMPS;
3142:   B->data         = (void*)mumps;

3144:   PetscInitializeMUMPS(A,mumps);

3146:   *F = B;
3147:   return(0);
3148: }

3150: static PetscErrorCode MatGetFactor_baij_mumps(Mat A,MatFactorType ftype,Mat *F)
3151: {
3152:   Mat            B;
3154:   Mat_MUMPS      *mumps;
3155:   PetscBool      isSeqBAIJ;
3156:   PetscMPIInt    size;

3159:   /* Create the factorization matrix */
3160:   PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isSeqBAIJ);
3161:   MatCreate(PetscObjectComm((PetscObject)A),&B);
3162:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
3163:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
3164:   MatSetUp(B);

3166:   PetscNewLog(B,&mumps);
3167:   if (ftype == MAT_FACTOR_LU) {
3168:     B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS;
3169:     B->factortype            = MAT_FACTOR_LU;
3170:     if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij;
3171:     else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij;
3172:     mumps->sym = 0;
3173:     PetscStrallocpy(MATORDERINGEXTERNAL,(char**)&B->preferredordering[MAT_FACTOR_LU]);
3174:   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead\n");

3176:   B->ops->view        = MatView_MUMPS;
3177:   B->ops->getinfo     = MatGetInfo_MUMPS;

3179:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
3180:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
3181:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
3182:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
3183:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
3184:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
3185:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
3186:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
3187:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
3188:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
3189:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
3190:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
3191:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);

3193:   /* set solvertype */
3194:   PetscFree(B->solvertype);
3195:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
3196:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
3197:   if (size == 1) {
3198:     /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization */
3199:     B->canuseordering = PETSC_TRUE;
3200:   }
3201:   B->ops->destroy = MatDestroy_MUMPS;
3202:   B->data         = (void*)mumps;

3204:   PetscInitializeMUMPS(A,mumps);

3206:   *F = B;
3207:   return(0);
3208: }

3210: /* MatGetFactor for Seq and MPI SELL matrices */
3211: static PetscErrorCode MatGetFactor_sell_mumps(Mat A,MatFactorType ftype,Mat *F)
3212: {
3213:   Mat            B;
3215:   Mat_MUMPS      *mumps;
3216:   PetscBool      isSeqSELL;
3217:   PetscMPIInt    size;

3220:   /* Create the factorization matrix */
3221:   PetscObjectTypeCompare((PetscObject)A,MATSEQSELL,&isSeqSELL);
3222:   MatCreate(PetscObjectComm((PetscObject)A),&B);
3223:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
3224:   PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
3225:   MatSetUp(B);

3227:   PetscNewLog(B,&mumps);

3229:   B->ops->view        = MatView_MUMPS;
3230:   B->ops->getinfo     = MatGetInfo_MUMPS;

3232:   PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
3233:   PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
3234:   PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
3235:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
3236:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
3237:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
3238:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
3239:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
3240:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
3241:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
3242:   PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);

3244:   if (ftype == MAT_FACTOR_LU) {
3245:     B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
3246:     B->factortype            = MAT_FACTOR_LU;
3247:     if (isSeqSELL) mumps->ConvertToTriples = MatConvertToTriples_seqsell_seqaij;
3248:     else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented");
3249:     mumps->sym = 0;
3250:     PetscStrallocpy(MATORDERINGEXTERNAL,(char**)&B->preferredordering[MAT_FACTOR_LU]);
3251:   } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented");

3253:   /* set solvertype */
3254:   PetscFree(B->solvertype);
3255:   PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
3256:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
3257:   if (size == 1) {
3258:     /* MUMPS option -mat_mumps_icntl_7 1 is automatically set if PETSc ordering is passed into symbolic factorization  */
3259:     B->canuseordering = PETSC_TRUE;
3260:   }
3261:   B->ops->destroy = MatDestroy_MUMPS;
3262:   B->data         = (void*)mumps;

3264:   PetscInitializeMUMPS(A,mumps);

3266:   *F = B;
3267:   return(0);
3268: }

3270: PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_MUMPS(void)
3271: {

3275:   MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
3276:   MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
3277:   MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
3278:   MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
3279:   MatSolverTypeRegister(MATSOLVERMUMPS,MATMPISBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
3280:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
3281:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
3282:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
3283:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
3284:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
3285:   MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSELL,MAT_FACTOR_LU,MatGetFactor_sell_mumps);
3286:   return(0);
3287: }