Actual source code: mumps.c
petsc-3.10.5 2019-03-28
2: /*
3: Provides an interface to the MUMPS sparse solver
4: */
6: #include <../src/mat/impls/aij/mpi/mpiaij.h>
7: #include <../src/mat/impls/sbaij/mpi/mpisbaij.h>
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: /* 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 */
47: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
48: #define PetscMUMPS_c(mumps) \
49: do { \
50: if (mumps->use_petsc_omp_support) { \
51: if (mumps->is_omp_master) { \
52: PetscOmpCtrlOmpRegionOnMasterBegin(mumps->omp_ctrl); \
53: MUMPS_c(&mumps->id); \
54: PetscOmpCtrlOmpRegionOnMasterEnd(mumps->omp_ctrl); \
55: } \
56: PetscOmpCtrlBarrier(mumps->omp_ctrl); \
57: } else { \
58: MUMPS_c(&mumps->id); \
59: } \
60: } while(0)
61: #else
62: #define PetscMUMPS_c(mumps) \
63: do { MUMPS_c(&mumps->id); } while (0)
64: #endif
66: /* declare MumpsScalar */
67: #if defined(PETSC_USE_COMPLEX)
68: #if defined(PETSC_USE_REAL_SINGLE)
69: #define MumpsScalar mumps_complex
70: #else
71: #define MumpsScalar mumps_double_complex
72: #endif
73: #else
74: #define MumpsScalar PetscScalar
75: #endif
77: /* macros s.t. indices match MUMPS documentation */
78: #define ICNTL(I) icntl[(I)-1]
79: #define CNTL(I) cntl[(I)-1]
80: #define INFOG(I) infog[(I)-1]
81: #define INFO(I) info[(I)-1]
82: #define RINFOG(I) rinfog[(I)-1]
83: #define RINFO(I) rinfo[(I)-1]
85: typedef struct {
86: #if defined(PETSC_USE_COMPLEX)
87: #if defined(PETSC_USE_REAL_SINGLE)
88: CMUMPS_STRUC_C id;
89: #else
90: ZMUMPS_STRUC_C id;
91: #endif
92: #else
93: #if defined(PETSC_USE_REAL_SINGLE)
94: SMUMPS_STRUC_C id;
95: #else
96: DMUMPS_STRUC_C id;
97: #endif
98: #endif
100: MatStructure matstruc;
101: PetscMPIInt myid,petsc_size;
102: PetscInt *irn,*jcn,nz,sym;
103: PetscScalar *val;
104: MPI_Comm mumps_comm;
105: PetscInt ICNTL9_pre; /* check if ICNTL(9) is changed from previous MatSolve */
106: VecScatter scat_rhs, scat_sol; /* used by MatSolve() */
107: Vec b_seq,x_seq;
108: PetscInt ninfo,*info; /* display INFO */
109: PetscInt sizeredrhs;
110: PetscScalar *schur_sol;
111: PetscInt schur_sizesol;
113: PetscBool use_petsc_omp_support;
114: PetscOmpCtrl omp_ctrl; /* an OpenMP controler that blocked processes will release their CPU (MPI_Barrier does not have this guarantee) */
115: MPI_Comm petsc_comm,omp_comm; /* petsc_comm is petsc matrix's comm */
116: PetscMPIInt mpinz; /* on master rank, nz = sum(mpinz) over omp_comm; on other ranks, mpinz = nz*/
117: PetscMPIInt omp_comm_size;
118: PetscBool is_omp_master; /* is this rank the master of omp_comm */
119: PetscMPIInt *recvcount,*displs;
121: PetscErrorCode (*ConvertToTriples)(Mat, int, MatReuse, int*, int**, int**, PetscScalar**);
122: } Mat_MUMPS;
124: extern PetscErrorCode MatDuplicate_MUMPS(Mat,MatDuplicateOption,Mat*);
126: static PetscErrorCode MatMumpsResetSchur_Private(Mat_MUMPS* mumps)
127: {
131: PetscFree2(mumps->id.listvar_schur,mumps->id.schur);
132: PetscFree(mumps->id.redrhs);
133: PetscFree(mumps->schur_sol);
134: mumps->id.size_schur = 0;
135: mumps->id.schur_lld = 0;
136: mumps->id.ICNTL(19) = 0;
137: return(0);
138: }
140: /* solve with rhs in mumps->id.redrhs and return in the same location */
141: static PetscErrorCode MatMumpsSolveSchur_Private(Mat F)
142: {
143: Mat_MUMPS *mumps=(Mat_MUMPS*)F->data;
144: Mat S,B,X;
145: MatFactorSchurStatus schurstatus;
146: PetscInt sizesol;
147: PetscErrorCode ierr;
150: MatFactorFactorizeSchurComplement(F);
151: MatFactorGetSchurComplement(F,&S,&schurstatus);
152: MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&B);
153: switch (schurstatus) {
154: case MAT_FACTOR_SCHUR_FACTORED:
155: MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,(PetscScalar*)mumps->id.redrhs,&X);
156: if (!mumps->id.ICNTL(9)) { /* transpose solve */
157: MatMatSolveTranspose(S,B,X);
158: } else {
159: MatMatSolve(S,B,X);
160: }
161: break;
162: case MAT_FACTOR_SCHUR_INVERTED:
163: sizesol = mumps->id.nrhs*mumps->id.size_schur;
164: if (!mumps->schur_sol || sizesol > mumps->schur_sizesol) {
165: PetscFree(mumps->schur_sol);
166: PetscMalloc1(sizesol,&mumps->schur_sol);
167: mumps->schur_sizesol = sizesol;
168: }
169: MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.nrhs,mumps->schur_sol,&X);
170: if (!mumps->id.ICNTL(9)) { /* transpose solve */
171: MatTransposeMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);
172: } else {
173: MatMatMult(S,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&X);
174: }
175: MatCopy(X,B,SAME_NONZERO_PATTERN);
176: break;
177: default:
178: SETERRQ1(PetscObjectComm((PetscObject)F),PETSC_ERR_SUP,"Unhandled MatFactorSchurStatus %D",F->schur_status);
179: break;
180: }
181: MatFactorRestoreSchurComplement(F,&S,schurstatus);
182: MatDestroy(&B);
183: MatDestroy(&X);
184: return(0);
185: }
187: static PetscErrorCode MatMumpsHandleSchur_Private(Mat F, PetscBool expansion)
188: {
189: Mat_MUMPS *mumps=(Mat_MUMPS*)F->data;
193: if (!mumps->id.ICNTL(19)) { /* do nothing when Schur complement has not been computed */
194: return(0);
195: }
196: if (!expansion) { /* prepare for the condensation step */
197: PetscInt sizeredrhs = mumps->id.nrhs*mumps->id.size_schur;
198: /* allocate MUMPS internal array to store reduced right-hand sides */
199: if (!mumps->id.redrhs || sizeredrhs > mumps->sizeredrhs) {
200: PetscFree(mumps->id.redrhs);
201: mumps->id.lredrhs = mumps->id.size_schur;
202: PetscMalloc1(mumps->id.nrhs*mumps->id.lredrhs,&mumps->id.redrhs);
203: mumps->sizeredrhs = mumps->id.nrhs*mumps->id.lredrhs;
204: }
205: mumps->id.ICNTL(26) = 1; /* condensation phase */
206: } else { /* prepare for the expansion step */
207: /* solve Schur complement (this has to be done by the MUMPS user, so basically us) */
208: MatMumpsSolveSchur_Private(F);
209: mumps->id.ICNTL(26) = 2; /* expansion phase */
210: PetscMUMPS_c(mumps);
211: 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));
212: /* restore defaults */
213: mumps->id.ICNTL(26) = -1;
214: /* free MUMPS internal array for redrhs if we have solved for multiple rhs in order to save memory space */
215: if (mumps->id.nrhs > 1) {
216: PetscFree(mumps->id.redrhs);
217: mumps->id.lredrhs = 0;
218: mumps->sizeredrhs = 0;
219: }
220: }
221: return(0);
222: }
224: /*
225: MatConvertToTriples_A_B - convert Petsc matrix to triples: row[nz], col[nz], val[nz]
227: input:
228: A - matrix in aij,baij or sbaij (bs=1) format
229: shift - 0: C style output triple; 1: Fortran style output triple.
230: reuse - MAT_INITIAL_MATRIX: spaces are allocated and values are set for the triple
231: MAT_REUSE_MATRIX: only the values in v array are updated
232: output:
233: nnz - dim of r, c, and v (number of local nonzero entries of A)
234: r, c, v - row and col index, matrix values (matrix triples)
236: The returned values r, c, and sometimes v are obtained in a single PetscMalloc(). Then in MatDestroy_MUMPS() it is
237: freed with PetscFree(mumps->irn); This is not ideal code, the fact that v is ONLY sometimes part of mumps->irn means
238: that the PetscMalloc() cannot easily be replaced with a PetscMalloc3().
240: */
242: PetscErrorCode MatConvertToTriples_seqaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
243: {
244: const PetscInt *ai,*aj,*ajj,M=A->rmap->n;
245: PetscInt nz,rnz,i,j;
247: PetscInt *row,*col;
248: Mat_SeqAIJ *aa=(Mat_SeqAIJ*)A->data;
251: *v=aa->a;
252: if (reuse == MAT_INITIAL_MATRIX) {
253: nz = aa->nz;
254: ai = aa->i;
255: aj = aa->j;
256: *nnz = nz;
257: PetscMalloc1(2*nz, &row);
258: col = row + nz;
260: nz = 0;
261: for (i=0; i<M; i++) {
262: rnz = ai[i+1] - ai[i];
263: ajj = aj + ai[i];
264: for (j=0; j<rnz; j++) {
265: row[nz] = i+shift; col[nz++] = ajj[j] + shift;
266: }
267: }
268: *r = row; *c = col;
269: }
270: return(0);
271: }
273: PetscErrorCode MatConvertToTriples_seqsell_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
274: {
275: Mat_SeqSELL *a=(Mat_SeqSELL*)A->data;
276: PetscInt *ptr;
279: *v = a->val;
280: if (reuse == MAT_INITIAL_MATRIX) {
281: PetscInt nz,i,j,row;
284: nz = a->sliidx[a->totalslices];
285: *nnz = nz;
286: PetscMalloc1(2*nz, &ptr);
287: *r = ptr;
288: *c = ptr + nz;
290: for (i=0; i<a->totalslices; i++) {
291: for (j=a->sliidx[i],row=0; j<a->sliidx[i+1]; j++,row=((row+1)&0x07)) {
292: *ptr++ = 8*i + row + shift;
293: }
294: }
295: for (i=0;i<nz;i++) *ptr++ = a->colidx[i] + shift;
296: }
297: return(0);
298: }
300: PetscErrorCode MatConvertToTriples_seqbaij_seqaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
301: {
302: Mat_SeqBAIJ *aa=(Mat_SeqBAIJ*)A->data;
303: const PetscInt *ai,*aj,*ajj,bs2 = aa->bs2;
304: PetscInt bs,M,nz,idx=0,rnz,i,j,k,m;
306: PetscInt *row,*col;
309: MatGetBlockSize(A,&bs);
310: M = A->rmap->N/bs;
311: *v = aa->a;
312: if (reuse == MAT_INITIAL_MATRIX) {
313: ai = aa->i; aj = aa->j;
314: nz = bs2*aa->nz;
315: *nnz = nz;
316: PetscMalloc1(2*nz, &row);
317: col = row + nz;
319: for (i=0; i<M; i++) {
320: ajj = aj + ai[i];
321: rnz = ai[i+1] - ai[i];
322: for (k=0; k<rnz; k++) {
323: for (j=0; j<bs; j++) {
324: for (m=0; m<bs; m++) {
325: row[idx] = i*bs + m + shift;
326: col[idx++] = bs*(ajj[k]) + j + shift;
327: }
328: }
329: }
330: }
331: *r = row; *c = col;
332: }
333: return(0);
334: }
336: PetscErrorCode MatConvertToTriples_seqsbaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
337: {
338: const PetscInt *ai, *aj,*ajj,M=A->rmap->n;
339: PetscInt nz,rnz,i,j;
341: PetscInt *row,*col;
342: Mat_SeqSBAIJ *aa=(Mat_SeqSBAIJ*)A->data;
345: *v = aa->a;
346: if (reuse == MAT_INITIAL_MATRIX) {
347: nz = aa->nz;
348: ai = aa->i;
349: aj = aa->j;
350: *v = aa->a;
351: *nnz = nz;
352: PetscMalloc1(2*nz, &row);
353: col = row + nz;
355: nz = 0;
356: for (i=0; i<M; i++) {
357: rnz = ai[i+1] - ai[i];
358: ajj = aj + ai[i];
359: for (j=0; j<rnz; j++) {
360: row[nz] = i+shift; col[nz++] = ajj[j] + shift;
361: }
362: }
363: *r = row; *c = col;
364: }
365: return(0);
366: }
368: PetscErrorCode MatConvertToTriples_seqaij_seqsbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
369: {
370: const PetscInt *ai,*aj,*ajj,*adiag,M=A->rmap->n;
371: PetscInt nz,rnz,i,j;
372: const PetscScalar *av,*v1;
373: PetscScalar *val;
374: PetscErrorCode ierr;
375: PetscInt *row,*col;
376: Mat_SeqAIJ *aa=(Mat_SeqAIJ*)A->data;
377: PetscBool missing;
380: ai = aa->i; aj = aa->j; av = aa->a;
381: adiag = aa->diag;
382: MatMissingDiagonal_SeqAIJ(A,&missing,&i);
383: if (reuse == MAT_INITIAL_MATRIX) {
384: /* count nz in the upper triangular part of A */
385: nz = 0;
386: if (missing) {
387: for (i=0; i<M; i++) {
388: if (PetscUnlikely(adiag[i] >= ai[i+1])) {
389: for (j=ai[i];j<ai[i+1];j++) {
390: if (aj[j] < i) continue;
391: nz++;
392: }
393: } else {
394: nz += ai[i+1] - adiag[i];
395: }
396: }
397: } else {
398: for (i=0; i<M; i++) nz += ai[i+1] - adiag[i];
399: }
400: *nnz = nz;
402: PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
403: col = row + nz;
404: val = (PetscScalar*)(col + nz);
406: nz = 0;
407: if (missing) {
408: for (i=0; i<M; i++) {
409: if (PetscUnlikely(adiag[i] >= ai[i+1])) {
410: for (j=ai[i];j<ai[i+1];j++) {
411: if (aj[j] < i) continue;
412: row[nz] = i+shift;
413: col[nz] = aj[j]+shift;
414: val[nz] = av[j];
415: nz++;
416: }
417: } else {
418: rnz = ai[i+1] - adiag[i];
419: ajj = aj + adiag[i];
420: v1 = av + adiag[i];
421: for (j=0; j<rnz; j++) {
422: row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j];
423: }
424: }
425: }
426: } else {
427: for (i=0; i<M; i++) {
428: rnz = ai[i+1] - adiag[i];
429: ajj = aj + adiag[i];
430: v1 = av + adiag[i];
431: for (j=0; j<rnz; j++) {
432: row[nz] = i+shift; col[nz] = ajj[j] + shift; val[nz++] = v1[j];
433: }
434: }
435: }
436: *r = row; *c = col; *v = val;
437: } else {
438: nz = 0; val = *v;
439: if (missing) {
440: for (i=0; i <M; i++) {
441: if (PetscUnlikely(adiag[i] >= ai[i+1])) {
442: for (j=ai[i];j<ai[i+1];j++) {
443: if (aj[j] < i) continue;
444: val[nz++] = av[j];
445: }
446: } else {
447: rnz = ai[i+1] - adiag[i];
448: v1 = av + adiag[i];
449: for (j=0; j<rnz; j++) {
450: val[nz++] = v1[j];
451: }
452: }
453: }
454: } else {
455: for (i=0; i <M; i++) {
456: rnz = ai[i+1] - adiag[i];
457: v1 = av + adiag[i];
458: for (j=0; j<rnz; j++) {
459: val[nz++] = v1[j];
460: }
461: }
462: }
463: }
464: return(0);
465: }
467: PetscErrorCode MatConvertToTriples_mpisbaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
468: {
469: const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
470: PetscErrorCode ierr;
471: PetscInt rstart,nz,i,j,jj,irow,countA,countB;
472: PetscInt *row,*col;
473: const PetscScalar *av, *bv,*v1,*v2;
474: PetscScalar *val;
475: Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)A->data;
476: Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ*)(mat->A)->data;
477: Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data;
480: ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart;
481: av=aa->a; bv=bb->a;
483: garray = mat->garray;
485: if (reuse == MAT_INITIAL_MATRIX) {
486: nz = aa->nz + bb->nz;
487: *nnz = nz;
488: PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
489: col = row + nz;
490: val = (PetscScalar*)(col + nz);
492: *r = row; *c = col; *v = val;
493: } else {
494: row = *r; col = *c; val = *v;
495: }
497: jj = 0; irow = rstart;
498: for (i=0; i<m; i++) {
499: ajj = aj + ai[i]; /* ptr to the beginning of this row */
500: countA = ai[i+1] - ai[i];
501: countB = bi[i+1] - bi[i];
502: bjj = bj + bi[i];
503: v1 = av + ai[i];
504: v2 = bv + bi[i];
506: /* A-part */
507: for (j=0; j<countA; j++) {
508: if (reuse == MAT_INITIAL_MATRIX) {
509: row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
510: }
511: val[jj++] = v1[j];
512: }
514: /* B-part */
515: for (j=0; j < countB; j++) {
516: if (reuse == MAT_INITIAL_MATRIX) {
517: row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
518: }
519: val[jj++] = v2[j];
520: }
521: irow++;
522: }
523: return(0);
524: }
526: PetscErrorCode MatConvertToTriples_mpiaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
527: {
528: const PetscInt *ai, *aj, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
529: PetscErrorCode ierr;
530: PetscInt rstart,nz,i,j,jj,irow,countA,countB;
531: PetscInt *row,*col;
532: const PetscScalar *av, *bv,*v1,*v2;
533: PetscScalar *val;
534: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
535: Mat_SeqAIJ *aa = (Mat_SeqAIJ*)(mat->A)->data;
536: Mat_SeqAIJ *bb = (Mat_SeqAIJ*)(mat->B)->data;
539: ai=aa->i; aj=aa->j; bi=bb->i; bj=bb->j; rstart= A->rmap->rstart;
540: av=aa->a; bv=bb->a;
542: garray = mat->garray;
544: if (reuse == MAT_INITIAL_MATRIX) {
545: nz = aa->nz + bb->nz;
546: *nnz = nz;
547: PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
548: col = row + nz;
549: val = (PetscScalar*)(col + nz);
551: *r = row; *c = col; *v = val;
552: } else {
553: row = *r; col = *c; val = *v;
554: }
556: jj = 0; irow = rstart;
557: for (i=0; i<m; i++) {
558: ajj = aj + ai[i]; /* ptr to the beginning of this row */
559: countA = ai[i+1] - ai[i];
560: countB = bi[i+1] - bi[i];
561: bjj = bj + bi[i];
562: v1 = av + ai[i];
563: v2 = bv + bi[i];
565: /* A-part */
566: for (j=0; j<countA; j++) {
567: if (reuse == MAT_INITIAL_MATRIX) {
568: row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
569: }
570: val[jj++] = v1[j];
571: }
573: /* B-part */
574: for (j=0; j < countB; j++) {
575: if (reuse == MAT_INITIAL_MATRIX) {
576: row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
577: }
578: val[jj++] = v2[j];
579: }
580: irow++;
581: }
582: return(0);
583: }
585: PetscErrorCode MatConvertToTriples_mpibaij_mpiaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
586: {
587: Mat_MPIBAIJ *mat = (Mat_MPIBAIJ*)A->data;
588: Mat_SeqBAIJ *aa = (Mat_SeqBAIJ*)(mat->A)->data;
589: Mat_SeqBAIJ *bb = (Mat_SeqBAIJ*)(mat->B)->data;
590: const PetscInt *ai = aa->i, *bi = bb->i, *aj = aa->j, *bj = bb->j,*ajj, *bjj;
591: const PetscInt *garray = mat->garray,mbs=mat->mbs,rstart=A->rmap->rstart;
592: const PetscInt bs2=mat->bs2;
593: PetscErrorCode ierr;
594: PetscInt bs,nz,i,j,k,n,jj,irow,countA,countB,idx;
595: PetscInt *row,*col;
596: const PetscScalar *av=aa->a, *bv=bb->a,*v1,*v2;
597: PetscScalar *val;
600: MatGetBlockSize(A,&bs);
601: if (reuse == MAT_INITIAL_MATRIX) {
602: nz = bs2*(aa->nz + bb->nz);
603: *nnz = nz;
604: PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
605: col = row + nz;
606: val = (PetscScalar*)(col + nz);
608: *r = row; *c = col; *v = val;
609: } else {
610: row = *r; col = *c; val = *v;
611: }
613: jj = 0; irow = rstart;
614: for (i=0; i<mbs; i++) {
615: countA = ai[i+1] - ai[i];
616: countB = bi[i+1] - bi[i];
617: ajj = aj + ai[i];
618: bjj = bj + bi[i];
619: v1 = av + bs2*ai[i];
620: v2 = bv + bs2*bi[i];
622: idx = 0;
623: /* A-part */
624: for (k=0; k<countA; k++) {
625: for (j=0; j<bs; j++) {
626: for (n=0; n<bs; n++) {
627: if (reuse == MAT_INITIAL_MATRIX) {
628: row[jj] = irow + n + shift;
629: col[jj] = rstart + bs*ajj[k] + j + shift;
630: }
631: val[jj++] = v1[idx++];
632: }
633: }
634: }
636: idx = 0;
637: /* B-part */
638: for (k=0; k<countB; k++) {
639: for (j=0; j<bs; j++) {
640: for (n=0; n<bs; n++) {
641: if (reuse == MAT_INITIAL_MATRIX) {
642: row[jj] = irow + n + shift;
643: col[jj] = bs*garray[bjj[k]] + j + shift;
644: }
645: val[jj++] = v2[idx++];
646: }
647: }
648: }
649: irow += bs;
650: }
651: return(0);
652: }
654: PetscErrorCode MatConvertToTriples_mpiaij_mpisbaij(Mat A,int shift,MatReuse reuse,int *nnz,int **r, int **c, PetscScalar **v)
655: {
656: const PetscInt *ai, *aj,*adiag, *bi, *bj,*garray,m=A->rmap->n,*ajj,*bjj;
657: PetscErrorCode ierr;
658: PetscInt rstart,nz,nza,nzb,i,j,jj,irow,countA,countB;
659: PetscInt *row,*col;
660: const PetscScalar *av, *bv,*v1,*v2;
661: PetscScalar *val;
662: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
663: Mat_SeqAIJ *aa =(Mat_SeqAIJ*)(mat->A)->data;
664: Mat_SeqAIJ *bb =(Mat_SeqAIJ*)(mat->B)->data;
667: ai=aa->i; aj=aa->j; adiag=aa->diag;
668: bi=bb->i; bj=bb->j; garray = mat->garray;
669: av=aa->a; bv=bb->a;
671: rstart = A->rmap->rstart;
673: if (reuse == MAT_INITIAL_MATRIX) {
674: nza = 0; /* num of upper triangular entries in mat->A, including diagonals */
675: nzb = 0; /* num of upper triangular entries in mat->B */
676: for (i=0; i<m; i++) {
677: nza += (ai[i+1] - adiag[i]);
678: countB = bi[i+1] - bi[i];
679: bjj = bj + bi[i];
680: for (j=0; j<countB; j++) {
681: if (garray[bjj[j]] > rstart) nzb++;
682: }
683: }
685: nz = nza + nzb; /* total nz of upper triangular part of mat */
686: *nnz = nz;
687: PetscMalloc((2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar)), &row);
688: col = row + nz;
689: val = (PetscScalar*)(col + nz);
691: *r = row; *c = col; *v = val;
692: } else {
693: row = *r; col = *c; val = *v;
694: }
696: jj = 0; irow = rstart;
697: for (i=0; i<m; i++) {
698: ajj = aj + adiag[i]; /* ptr to the beginning of the diagonal of this row */
699: v1 = av + adiag[i];
700: countA = ai[i+1] - adiag[i];
701: countB = bi[i+1] - bi[i];
702: bjj = bj + bi[i];
703: v2 = bv + bi[i];
705: /* A-part */
706: for (j=0; j<countA; j++) {
707: if (reuse == MAT_INITIAL_MATRIX) {
708: row[jj] = irow + shift; col[jj] = rstart + ajj[j] + shift;
709: }
710: val[jj++] = v1[j];
711: }
713: /* B-part */
714: for (j=0; j < countB; j++) {
715: if (garray[bjj[j]] > rstart) {
716: if (reuse == MAT_INITIAL_MATRIX) {
717: row[jj] = irow + shift; col[jj] = garray[bjj[j]] + shift;
718: }
719: val[jj++] = v2[j];
720: }
721: }
722: irow++;
723: }
724: return(0);
725: }
727: PetscErrorCode MatDestroy_MUMPS(Mat A)
728: {
729: Mat_MUMPS *mumps=(Mat_MUMPS*)A->data;
733: PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
734: VecScatterDestroy(&mumps->scat_rhs);
735: VecScatterDestroy(&mumps->scat_sol);
736: VecDestroy(&mumps->b_seq);
737: VecDestroy(&mumps->x_seq);
738: PetscFree(mumps->id.perm_in);
739: PetscFree(mumps->irn);
740: PetscFree(mumps->info);
741: MatMumpsResetSchur_Private(mumps);
742: mumps->id.job = JOB_END;
743: PetscMUMPS_c(mumps);
744: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
745: if (mumps->use_petsc_omp_support) { PetscOmpCtrlDestroy(&mumps->omp_ctrl); }
746: #endif
747: PetscFree2(mumps->recvcount,mumps->displs);
748: PetscFree(A->data);
750: /* clear composed functions */
751: PetscObjectComposeFunction((PetscObject)A,"MatFactorGetSolverType_C",NULL);
752: PetscObjectComposeFunction((PetscObject)A,"MatFactorSetSchurIS_C",NULL);
753: PetscObjectComposeFunction((PetscObject)A,"MatFactorCreateSchurComplement_C",NULL);
754: PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetIcntl_C",NULL);
755: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetIcntl_C",NULL);
756: PetscObjectComposeFunction((PetscObject)A,"MatMumpsSetCntl_C",NULL);
757: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetCntl_C",NULL);
758: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfo_C",NULL);
759: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInfog_C",NULL);
760: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfo_C",NULL);
761: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetRinfog_C",NULL);
762: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverse_C",NULL);
763: PetscObjectComposeFunction((PetscObject)A,"MatMumpsGetInverseTranspose_C",NULL);
764: return(0);
765: }
767: PetscErrorCode MatSolve_MUMPS(Mat A,Vec b,Vec x)
768: {
769: Mat_MUMPS *mumps=(Mat_MUMPS*)A->data;
770: PetscScalar *array;
771: Vec b_seq;
772: IS is_iden,is_petsc;
773: PetscErrorCode ierr;
774: PetscInt i;
775: PetscBool second_solve = PETSC_FALSE;
776: static PetscBool cite1 = PETSC_FALSE,cite2 = PETSC_FALSE;
779: 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);
780: 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);
782: if (A->factorerrortype) {
783: PetscInfo2(A,"MatSolve is called with singular matrix factor, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
784: VecSetInf(x);
785: return(0);
786: }
788: mumps->id.ICNTL(20)= 0; /* dense RHS */
789: mumps->id.nrhs = 1;
790: b_seq = mumps->b_seq;
791: if (mumps->petsc_size > 1) {
792: /* MUMPS only supports centralized rhs. Scatter b into a seqential rhs vector */
793: VecScatterBegin(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);
794: VecScatterEnd(mumps->scat_rhs,b,b_seq,INSERT_VALUES,SCATTER_FORWARD);
795: if (!mumps->myid) {VecGetArray(b_seq,&array);}
796: } else { /* petsc_size == 1 */
797: VecCopy(b,x);
798: VecGetArray(x,&array);
799: }
800: if (!mumps->myid) { /* define rhs on the host */
801: mumps->id.nrhs = 1;
802: mumps->id.rhs = (MumpsScalar*)array;
803: }
805: /*
806: handle condensation step of Schur complement (if any)
807: We set by default ICNTL(26) == -1 when Schur indices have been provided by the user.
808: According to MUMPS (5.0.0) manual, any value should be harmful during the factorization phase
809: Unless the user provides a valid value for ICNTL(26), MatSolve and MatMatSolve routines solve the full system.
810: This requires an extra call to PetscMUMPS_c and the computation of the factors for S
811: */
812: if (mumps->id.size_schur > 0 && (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2)) {
813: if (mumps->petsc_size > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Parallel Schur complements not yet supported from PETSc\n");
814: second_solve = PETSC_TRUE;
815: MatMumpsHandleSchur_Private(A,PETSC_FALSE);
816: }
817: /* solve phase */
818: /*-------------*/
819: mumps->id.job = JOB_SOLVE;
820: PetscMUMPS_c(mumps);
821: 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));
823: /* handle expansion step of Schur complement (if any) */
824: if (second_solve) {
825: MatMumpsHandleSchur_Private(A,PETSC_TRUE);
826: }
828: if (mumps->petsc_size > 1) { /* convert mumps distributed solution to petsc mpi x */
829: if (mumps->scat_sol && mumps->ICNTL9_pre != mumps->id.ICNTL(9)) {
830: /* when id.ICNTL(9) changes, the contents of lsol_loc may change (not its size, lsol_loc), recreates scat_sol */
831: VecScatterDestroy(&mumps->scat_sol);
832: }
833: if (!mumps->scat_sol) { /* create scatter scat_sol */
834: ISCreateStride(PETSC_COMM_SELF,mumps->id.lsol_loc,0,1,&is_iden); /* from */
835: for (i=0; i<mumps->id.lsol_loc; i++) {
836: mumps->id.isol_loc[i] -= 1; /* change Fortran style to C style */
837: }
838: ISCreateGeneral(PETSC_COMM_SELF,mumps->id.lsol_loc,mumps->id.isol_loc,PETSC_COPY_VALUES,&is_petsc); /* to */
839: VecScatterCreate(mumps->x_seq,is_iden,x,is_petsc,&mumps->scat_sol);
840: ISDestroy(&is_iden);
841: ISDestroy(&is_petsc);
843: mumps->ICNTL9_pre = mumps->id.ICNTL(9); /* save current value of id.ICNTL(9) */
844: }
846: VecScatterBegin(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
847: VecScatterEnd(mumps->scat_sol,mumps->x_seq,x,INSERT_VALUES,SCATTER_FORWARD);
848: }
849: PetscLogFlops(2.0*mumps->id.RINFO(3));
850: return(0);
851: }
853: PetscErrorCode MatSolveTranspose_MUMPS(Mat A,Vec b,Vec x)
854: {
855: Mat_MUMPS *mumps=(Mat_MUMPS*)A->data;
859: mumps->id.ICNTL(9) = 0;
860: MatSolve_MUMPS(A,b,x);
861: mumps->id.ICNTL(9) = 1;
862: return(0);
863: }
865: PetscErrorCode MatMatSolve_MUMPS(Mat A,Mat B,Mat X)
866: {
868: Mat Bt = NULL;
869: PetscBool flg, flgT;
870: Mat_MUMPS *mumps=(Mat_MUMPS*)A->data;
871: PetscInt i,nrhs,M;
872: PetscScalar *array,*bray;
873: PetscInt lsol_loc,nlsol_loc,*isol_loc,*idx,*iidx,*idxx,*isol_loc_save;
874: MumpsScalar *sol_loc,*sol_loc_save;
875: IS is_to,is_from;
876: PetscInt k,proc,j,m;
877: const PetscInt *rstart;
878: Vec v_mpi,b_seq,x_seq;
879: VecScatter scat_rhs,scat_sol;
880: PetscScalar *aa;
881: PetscInt spnr,*ia,*ja;
882: Mat_MPIAIJ *b = NULL;
885: PetscObjectTypeCompareAny((PetscObject)X,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
886: if (!flg) SETERRQ(PetscObjectComm((PetscObject)X),PETSC_ERR_ARG_WRONG,"Matrix X must be MATDENSE matrix");
888: PetscObjectTypeCompareAny((PetscObject)B,&flg,MATSEQDENSE,MATMPIDENSE,NULL);
889: if (flg) { /* dense B */
890: if (B->rmap->n != X->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Matrix B and X must have same row distribution");
891: mumps->id.ICNTL(20)= 0; /* dense RHS */
892: } else { /* sparse B */
893: PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&flgT);
894: if (flgT) { /* input B is transpose of actural RHS matrix,
895: because mumps requires sparse compressed COLUMN storage! See MatMatTransposeSolve_MUMPS() */
896: MatTransposeGetMat(B,&Bt);
897: } else SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_WRONG,"Matrix B must be MATTRANSPOSEMAT matrix");
898: mumps->id.ICNTL(20)= 1; /* sparse RHS */
899: }
901: MatGetSize(B,&M,&nrhs);
902: mumps->id.nrhs = nrhs;
903: mumps->id.lrhs = M;
904: mumps->id.rhs = NULL;
906: if (mumps->petsc_size == 1) {
907: PetscScalar *aa;
908: PetscInt spnr,*ia,*ja;
909: PetscBool second_solve = PETSC_FALSE;
911: MatDenseGetArray(X,&array);
912: mumps->id.rhs = (MumpsScalar*)array;
914: if (!Bt) { /* dense B */
915: /* copy B to X */
916: MatDenseGetArray(B,&bray);
917: PetscMemcpy(array,bray,M*nrhs*sizeof(PetscScalar));
918: MatDenseRestoreArray(B,&bray);
919: } else { /* sparse B */
920: MatSeqAIJGetArray(Bt,&aa);
921: MatGetRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
922: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
923: /* mumps requires ia and ja start at 1! */
924: mumps->id.irhs_ptr = ia;
925: mumps->id.irhs_sparse = ja;
926: mumps->id.nz_rhs = ia[spnr] - 1;
927: mumps->id.rhs_sparse = (MumpsScalar*)aa;
928: }
929: /* handle condensation step of Schur complement (if any) */
930: if (mumps->id.size_schur > 0 && (mumps->id.ICNTL(26) < 0 || mumps->id.ICNTL(26) > 2)) {
931: second_solve = PETSC_TRUE;
932: MatMumpsHandleSchur_Private(A,PETSC_FALSE);
933: }
934: /* solve phase */
935: /*-------------*/
936: mumps->id.job = JOB_SOLVE;
937: PetscMUMPS_c(mumps);
938: 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));
940: /* handle expansion step of Schur complement (if any) */
941: if (second_solve) {
942: MatMumpsHandleSchur_Private(A,PETSC_TRUE);
943: }
944: if (Bt) { /* sparse B */
945: MatSeqAIJRestoreArray(Bt,&aa);
946: MatRestoreRowIJ(Bt,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
947: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure");
948: }
949: MatDenseRestoreArray(X,&array);
950: return(0);
951: }
953: /*--------- parallel case: MUMPS requires rhs B to be centralized on the host! --------*/
954: 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");
956: /* create x_seq to hold mumps local solution */
957: isol_loc_save = mumps->id.isol_loc; /* save it for MatSovle() */
958: sol_loc_save = mumps->id.sol_loc;
960: lsol_loc = mumps->id.lsol_loc;
961: nlsol_loc = nrhs*lsol_loc; /* length of sol_loc */
962: PetscMalloc2(nlsol_loc,&sol_loc,lsol_loc,&isol_loc);
963: mumps->id.sol_loc = (MumpsScalar*)sol_loc;
964: mumps->id.isol_loc = isol_loc;
966: VecCreateSeqWithArray(PETSC_COMM_SELF,1,nlsol_loc,(PetscScalar*)sol_loc,&x_seq);
968: /* scatter v_mpi to b_seq because MUMPS only supports centralized rhs */
969: /* idx: maps from k-th index of v_mpi to (i,j)-th global entry of B;
970: iidx: inverse of idx, will be used by scattering mumps x_seq -> petsc X */
971: PetscMalloc1(nrhs*M,&idx);
973: if (!Bt) { /* dense B */
974: /* wrap dense rhs matrix B into a vector v_mpi */
975: MatGetLocalSize(B,&m,NULL);
976: MatDenseGetArray(B,&bray);
977: VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);
978: MatDenseRestoreArray(B,&bray);
980: /* scatter v_mpi to b_seq in proc[0]. MUMPS requires rhs to be centralized on the host! */
981: if (!mumps->myid) {
982: MatGetOwnershipRanges(B,&rstart);
983: k = 0;
984: for (proc=0; proc<mumps->petsc_size; proc++){
985: for (j=0; j<nrhs; j++){
986: for (i=rstart[proc]; i<rstart[proc+1]; i++){
987: idx[k++] = j*M + i;
988: }
989: }
990: }
992: VecCreateSeq(PETSC_COMM_SELF,nrhs*M,&b_seq);
993: ISCreateGeneral(PETSC_COMM_SELF,nrhs*M,idx,PETSC_COPY_VALUES,&is_to);
994: ISCreateStride(PETSC_COMM_SELF,nrhs*M,0,1,&is_from);
995: } else {
996: VecCreateSeq(PETSC_COMM_SELF,0,&b_seq);
997: ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_to);
998: ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_from);
999: }
1000: VecScatterCreate(v_mpi,is_from,b_seq,is_to,&scat_rhs);
1001: VecScatterBegin(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);
1002: ISDestroy(&is_to);
1003: ISDestroy(&is_from);
1004: VecScatterEnd(scat_rhs,v_mpi,b_seq,INSERT_VALUES,SCATTER_FORWARD);
1006: if (!mumps->myid) { /* define rhs on the host */
1007: VecGetArray(b_seq,&bray);
1008: mumps->id.rhs = (MumpsScalar*)bray;
1009: VecRestoreArray(b_seq,&bray);
1010: }
1012: } else { /* sparse B */
1013: b = (Mat_MPIAIJ*)Bt->data;
1015: /* wrap dense X into a vector v_mpi */
1016: MatGetLocalSize(X,&m,NULL);
1017: MatDenseGetArray(X,&bray);
1018: VecCreateMPIWithArray(PetscObjectComm((PetscObject)X),1,nrhs*m,nrhs*M,(const PetscScalar*)bray,&v_mpi);
1019: MatDenseRestoreArray(X,&bray);
1021: if (!mumps->myid) {
1022: MatSeqAIJGetArray(b->A,&aa);
1023: MatGetRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1024: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
1025: /* mumps requires ia and ja start at 1! */
1026: mumps->id.irhs_ptr = ia;
1027: mumps->id.irhs_sparse = ja;
1028: mumps->id.nz_rhs = ia[spnr] - 1;
1029: mumps->id.rhs_sparse = (MumpsScalar*)aa;
1030: } else {
1031: mumps->id.irhs_ptr = NULL;
1032: mumps->id.irhs_sparse = NULL;
1033: mumps->id.nz_rhs = 0;
1034: mumps->id.rhs_sparse = NULL;
1035: }
1036: }
1038: /* solve phase */
1039: /*-------------*/
1040: mumps->id.job = JOB_SOLVE;
1041: PetscMUMPS_c(mumps);
1042: 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));
1044: /* scatter mumps distributed solution to petsc vector v_mpi, which shares local arrays with solution matrix X */
1045: MatDenseGetArray(X,&array);
1046: VecPlaceArray(v_mpi,array);
1048: /* create scatter scat_sol */
1049: MatGetOwnershipRanges(X,&rstart);
1050: /* iidx: inverse of idx computed above, used for scattering mumps x_seq to petsc X */
1051: iidx = idx;
1052: k = 0;
1053: for (proc=0; proc<mumps->petsc_size; proc++){
1054: for (j=0; j<nrhs; j++){
1055: for (i=rstart[proc]; i<rstart[proc+1]; i++) iidx[j*M + i] = k++;
1056: }
1057: }
1059: PetscMalloc1(nlsol_loc,&idxx);
1060: ISCreateStride(PETSC_COMM_SELF,nlsol_loc,0,1,&is_from);
1061: for (i=0; i<lsol_loc; i++) {
1062: isol_loc[i] -= 1; /* change Fortran style to C style */
1063: idxx[i] = iidx[isol_loc[i]];
1064: for (j=1; j<nrhs; j++){
1065: idxx[j*lsol_loc+i] = iidx[isol_loc[i]+j*M];
1066: }
1067: }
1068: ISCreateGeneral(PETSC_COMM_SELF,nlsol_loc,idxx,PETSC_COPY_VALUES,&is_to);
1069: VecScatterCreate(x_seq,is_from,v_mpi,is_to,&scat_sol);
1070: VecScatterBegin(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
1071: ISDestroy(&is_from);
1072: ISDestroy(&is_to);
1073: VecScatterEnd(scat_sol,x_seq,v_mpi,INSERT_VALUES,SCATTER_FORWARD);
1074: MatDenseRestoreArray(X,&array);
1076: /* free spaces */
1077: mumps->id.sol_loc = sol_loc_save;
1078: mumps->id.isol_loc = isol_loc_save;
1080: PetscFree2(sol_loc,isol_loc);
1081: PetscFree(idx);
1082: PetscFree(idxx);
1083: VecDestroy(&x_seq);
1084: VecDestroy(&v_mpi);
1085: if (Bt) {
1086: if (!mumps->myid) {
1087: b = (Mat_MPIAIJ*)Bt->data;
1088: MatSeqAIJRestoreArray(b->A,&aa);
1089: MatRestoreRowIJ(b->A,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
1090: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot restore IJ structure");
1091: }
1092: } else {
1093: VecDestroy(&b_seq);
1094: VecScatterDestroy(&scat_rhs);
1095: }
1096: VecScatterDestroy(&scat_sol);
1097: PetscLogFlops(2.0*nrhs*mumps->id.RINFO(3));
1098: return(0);
1099: }
1101: PetscErrorCode MatMatTransposeSolve_MUMPS(Mat A,Mat Bt,Mat X)
1102: {
1104: PetscBool flg;
1105: Mat B;
1108: PetscObjectTypeCompareAny((PetscObject)Bt,&flg,MATSEQAIJ,MATMPIAIJ,NULL);
1109: if (!flg) SETERRQ(PetscObjectComm((PetscObject)Bt),PETSC_ERR_ARG_WRONG,"Matrix Bt must be MATAIJ matrix");
1111: /* Create B=Bt^T that uses Bt's data structure */
1112: MatCreateTranspose(Bt,&B);
1114: MatMatSolve_MUMPS(A,B,X);
1115: MatDestroy(&B);
1116: return(0);
1117: }
1119: #if !defined(PETSC_USE_COMPLEX)
1120: /*
1121: input:
1122: F: numeric factor
1123: output:
1124: nneg: total number of negative pivots
1125: nzero: total number of zero pivots
1126: npos: (global dimension of F) - nneg - nzero
1127: */
1128: PetscErrorCode MatGetInertia_SBAIJMUMPS(Mat F,int *nneg,int *nzero,int *npos)
1129: {
1130: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
1132: PetscMPIInt size;
1135: MPI_Comm_size(PetscObjectComm((PetscObject)F),&size);
1136: /* 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 */
1137: 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));
1139: if (nneg) *nneg = mumps->id.INFOG(12);
1140: if (nzero || npos) {
1141: 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");
1142: if (nzero) *nzero = mumps->id.INFOG(28);
1143: if (npos) *npos = F->rmap->N - (mumps->id.INFOG(12) + mumps->id.INFOG(28));
1144: }
1145: return(0);
1146: }
1147: #endif
1149: PetscErrorCode MatMumpsGatherNonzerosOnMaster(MatReuse reuse,Mat_MUMPS *mumps)
1150: {
1152: PetscInt i,nz=0,*irn,*jcn=0;
1153: PetscScalar *val=0;
1154: PetscMPIInt mpinz,*recvcount=NULL,*displs=NULL;
1157: if (mumps->omp_comm_size > 1) {
1158: if (reuse == MAT_INITIAL_MATRIX) {
1159: /* master first gathers counts of nonzeros to receive */
1160: if (mumps->is_omp_master) { PetscMalloc2(mumps->omp_comm_size,&recvcount,mumps->omp_comm_size,&displs); }
1161: PetscMPIIntCast(mumps->nz,&mpinz);
1162: MPI_Gather(&mpinz,1,MPI_INT,recvcount,1,MPI_INT,0/*root*/,mumps->omp_comm);
1164: /* master allocates memory to receive nonzeros */
1165: if (mumps->is_omp_master) {
1166: displs[0] = 0;
1167: for (i=1; i<mumps->omp_comm_size; i++) displs[i] = displs[i-1] + recvcount[i-1];
1168: nz = displs[mumps->omp_comm_size-1] + recvcount[mumps->omp_comm_size-1];
1169: PetscMalloc(2*nz*sizeof(PetscInt)+nz*sizeof(PetscScalar),&irn);
1170: jcn = irn + nz;
1171: val = (PetscScalar*)(jcn + nz);
1172: }
1174: /* save the gatherv plan */
1175: mumps->mpinz = mpinz; /* used as send count */
1176: mumps->recvcount = recvcount;
1177: mumps->displs = displs;
1179: /* master gathers nonzeros */
1180: MPI_Gatherv(mumps->irn,mpinz,MPIU_INT,irn,mumps->recvcount,mumps->displs,MPIU_INT,0/*root*/,mumps->omp_comm);
1181: MPI_Gatherv(mumps->jcn,mpinz,MPIU_INT,jcn,mumps->recvcount,mumps->displs,MPIU_INT,0/*root*/,mumps->omp_comm);
1182: MPI_Gatherv(mumps->val,mpinz,MPIU_SCALAR,val,mumps->recvcount,mumps->displs,MPIU_SCALAR,0/*root*/,mumps->omp_comm);
1184: /* master frees its row/col/val and replaces them with bigger arrays */
1185: if (mumps->is_omp_master) {
1186: PetscFree(mumps->irn); /* irn/jcn/val are allocated together so free only irn */
1187: mumps->nz = nz; /* it is a sum of mpinz over omp_comm */
1188: mumps->irn = irn;
1189: mumps->jcn = jcn;
1190: mumps->val = val;
1191: }
1192: } else {
1193: MPI_Gatherv((mumps->is_omp_master?MPI_IN_PLACE:mumps->val),mumps->mpinz,MPIU_SCALAR,mumps->val,mumps->recvcount,mumps->displs,MPIU_SCALAR,0/*root*/,mumps->omp_comm);
1194: }
1195: }
1196: return(0);
1197: }
1199: PetscErrorCode MatFactorNumeric_MUMPS(Mat F,Mat A,const MatFactorInfo *info)
1200: {
1201: Mat_MUMPS *mumps =(Mat_MUMPS*)(F)->data;
1203: PetscBool isMPIAIJ;
1206: if (mumps->id.INFOG(1) < 0) {
1207: if (mumps->id.INFOG(1) == -6) {
1208: PetscInfo2(A,"MatFactorNumeric is called with singular matrix structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1209: }
1210: PetscInfo2(A,"MatFactorNumeric is called after analysis phase fails, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1211: return(0);
1212: }
1214: (*mumps->ConvertToTriples)(A, 1, MAT_REUSE_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);
1215: MatMumpsGatherNonzerosOnMaster(MAT_REUSE_MATRIX,mumps);
1217: /* numerical factorization phase */
1218: /*-------------------------------*/
1219: mumps->id.job = JOB_FACTNUMERIC;
1220: if (!mumps->id.ICNTL(18)) { /* A is centralized */
1221: if (!mumps->myid) {
1222: mumps->id.a = (MumpsScalar*)mumps->val;
1223: }
1224: } else {
1225: mumps->id.a_loc = (MumpsScalar*)mumps->val;
1226: }
1227: PetscMUMPS_c(mumps);
1228: if (mumps->id.INFOG(1) < 0) {
1229: if (A->erroriffailure) {
1230: 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));
1231: } else {
1232: if (mumps->id.INFOG(1) == -10) { /* numerically singular matrix */
1233: PetscInfo2(F,"matrix is numerically singular, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1234: F->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1235: } else if (mumps->id.INFOG(1) == -13) {
1236: 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));
1237: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1238: } else if (mumps->id.INFOG(1) == -8 || mumps->id.INFOG(1) == -9 || (-16 < mumps->id.INFOG(1) && mumps->id.INFOG(1) < -10) ) {
1239: 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));
1240: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1241: } else {
1242: PetscInfo2(F,"MUMPS in numerical factorization phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1243: F->factorerrortype = MAT_FACTOR_OTHER;
1244: }
1245: }
1246: }
1247: 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));
1249: F->assembled = PETSC_TRUE;
1250: mumps->matstruc = SAME_NONZERO_PATTERN;
1251: if (F->schur) { /* reset Schur status to unfactored */
1252: if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
1253: mumps->id.ICNTL(19) = 2;
1254: MatTranspose(F->schur,MAT_INPLACE_MATRIX,&F->schur);
1255: }
1256: MatFactorRestoreSchurComplement(F,NULL,MAT_FACTOR_SCHUR_UNFACTORED);
1257: }
1259: /* just to be sure that ICNTL(19) value returned by a call from MatMumpsGetIcntl is always consistent */
1260: if (!mumps->sym && mumps->id.ICNTL(19) && mumps->id.ICNTL(19) != 1) mumps->id.ICNTL(19) = 3;
1262: if (!mumps->is_omp_master) mumps->id.INFO(23) = 0;
1263: if (mumps->petsc_size > 1) {
1264: PetscInt lsol_loc;
1265: PetscScalar *sol_loc;
1267: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&isMPIAIJ);
1269: /* distributed solution; Create x_seq=sol_loc for repeated use */
1270: if (mumps->x_seq) {
1271: VecScatterDestroy(&mumps->scat_sol);
1272: PetscFree2(mumps->id.sol_loc,mumps->id.isol_loc);
1273: VecDestroy(&mumps->x_seq);
1274: }
1275: lsol_loc = mumps->id.INFO(23); /* length of sol_loc */
1276: PetscMalloc2(lsol_loc,&sol_loc,lsol_loc,&mumps->id.isol_loc);
1277: mumps->id.lsol_loc = lsol_loc;
1278: mumps->id.sol_loc = (MumpsScalar*)sol_loc;
1279: VecCreateSeqWithArray(PETSC_COMM_SELF,1,lsol_loc,sol_loc,&mumps->x_seq);
1280: }
1281: PetscLogFlops(mumps->id.RINFO(2));
1282: return(0);
1283: }
1285: /* Sets MUMPS options from the options database */
1286: PetscErrorCode PetscSetMUMPSFromOptions(Mat F, Mat A)
1287: {
1288: Mat_MUMPS *mumps = (Mat_MUMPS*)F->data;
1290: PetscInt icntl,info[40],i,ninfo=40;
1291: PetscBool flg;
1294: PetscOptionsBegin(PetscObjectComm((PetscObject)A),((PetscObject)A)->prefix,"MUMPS Options","Mat");
1295: PetscOptionsInt("-mat_mumps_icntl_1","ICNTL(1): output stream for error messages","None",mumps->id.ICNTL(1),&icntl,&flg);
1296: if (flg) mumps->id.ICNTL(1) = icntl;
1297: PetscOptionsInt("-mat_mumps_icntl_2","ICNTL(2): output stream for diagnostic printing, statistics, and warning","None",mumps->id.ICNTL(2),&icntl,&flg);
1298: if (flg) mumps->id.ICNTL(2) = icntl;
1299: PetscOptionsInt("-mat_mumps_icntl_3","ICNTL(3): output stream for global information, collected on the host","None",mumps->id.ICNTL(3),&icntl,&flg);
1300: if (flg) mumps->id.ICNTL(3) = icntl;
1302: PetscOptionsInt("-mat_mumps_icntl_4","ICNTL(4): level of printing (0 to 4)","None",mumps->id.ICNTL(4),&icntl,&flg);
1303: if (flg) mumps->id.ICNTL(4) = icntl;
1304: if (mumps->id.ICNTL(4) || PetscLogPrintInfo) mumps->id.ICNTL(3) = 6; /* resume MUMPS default id.ICNTL(3) = 6 */
1306: PetscOptionsInt("-mat_mumps_icntl_6","ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)","None",mumps->id.ICNTL(6),&icntl,&flg);
1307: if (flg) mumps->id.ICNTL(6) = icntl;
1309: PetscOptionsInt("-mat_mumps_icntl_7","ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis","None",mumps->id.ICNTL(7),&icntl,&flg);
1310: if (flg) {
1311: if (icntl== 1 && mumps->petsc_size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"pivot order be set by the user in PERM_IN -- not supported by the PETSc/MUMPS interface\n");
1312: else mumps->id.ICNTL(7) = icntl;
1313: }
1315: PetscOptionsInt("-mat_mumps_icntl_8","ICNTL(8): scaling strategy (-2 to 8 or 77)","None",mumps->id.ICNTL(8),&mumps->id.ICNTL(8),NULL);
1316: /* 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() */
1317: PetscOptionsInt("-mat_mumps_icntl_10","ICNTL(10): max num of refinements","None",mumps->id.ICNTL(10),&mumps->id.ICNTL(10),NULL);
1318: PetscOptionsInt("-mat_mumps_icntl_11","ICNTL(11): statistics related to an error analysis (via -ksp_view)","None",mumps->id.ICNTL(11),&mumps->id.ICNTL(11),NULL);
1319: PetscOptionsInt("-mat_mumps_icntl_12","ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)","None",mumps->id.ICNTL(12),&mumps->id.ICNTL(12),NULL);
1320: PetscOptionsInt("-mat_mumps_icntl_13","ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting","None",mumps->id.ICNTL(13),&mumps->id.ICNTL(13),NULL);
1321: PetscOptionsInt("-mat_mumps_icntl_14","ICNTL(14): percentage increase in the estimated working space","None",mumps->id.ICNTL(14),&mumps->id.ICNTL(14),NULL);
1322: PetscOptionsInt("-mat_mumps_icntl_19","ICNTL(19): computes the Schur complement","None",mumps->id.ICNTL(19),&mumps->id.ICNTL(19),NULL);
1323: if (mumps->id.ICNTL(19) <= 0 || mumps->id.ICNTL(19) > 3) { /* reset any schur data (if any) */
1324: MatDestroy(&F->schur);
1325: MatMumpsResetSchur_Private(mumps);
1326: }
1327: /* PetscOptionsInt("-mat_mumps_icntl_20","ICNTL(20): the format (dense or sparse) of the right-hand sides","None",mumps->id.ICNTL(20),&mumps->id.ICNTL(20),NULL); -- sparse rhs is not supported in PETSc API */
1328: /* PetscOptionsInt("-mat_mumps_icntl_21","ICNTL(21): the distribution (centralized or distributed) of the solution vectors","None",mumps->id.ICNTL(21),&mumps->id.ICNTL(21),NULL); we only use distributed solution vector */
1330: PetscOptionsInt("-mat_mumps_icntl_22","ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)","None",mumps->id.ICNTL(22),&mumps->id.ICNTL(22),NULL);
1331: PetscOptionsInt("-mat_mumps_icntl_23","ICNTL(23): max size of the working memory (MB) that can allocate per processor","None",mumps->id.ICNTL(23),&mumps->id.ICNTL(23),NULL);
1332: PetscOptionsInt("-mat_mumps_icntl_24","ICNTL(24): detection of null pivot rows (0 or 1)","None",mumps->id.ICNTL(24),&mumps->id.ICNTL(24),NULL);
1333: if (mumps->id.ICNTL(24)) {
1334: mumps->id.ICNTL(13) = 1; /* turn-off ScaLAPACK to help with the correct detection of null pivots */
1335: }
1337: PetscOptionsInt("-mat_mumps_icntl_25","ICNTL(25): computes a solution of a deficient matrix and a null space basis","None",mumps->id.ICNTL(25),&mumps->id.ICNTL(25),NULL);
1338: PetscOptionsInt("-mat_mumps_icntl_26","ICNTL(26): drives the solution phase if a Schur complement matrix","None",mumps->id.ICNTL(26),&mumps->id.ICNTL(26),NULL);
1339: PetscOptionsInt("-mat_mumps_icntl_27","ICNTL(27): the blocking size for multiple right-hand sides","None",mumps->id.ICNTL(27),&mumps->id.ICNTL(27),NULL);
1340: PetscOptionsInt("-mat_mumps_icntl_28","ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering","None",mumps->id.ICNTL(28),&mumps->id.ICNTL(28),NULL);
1341: PetscOptionsInt("-mat_mumps_icntl_29","ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis","None",mumps->id.ICNTL(29),&mumps->id.ICNTL(29),NULL);
1342: /* PetscOptionsInt("-mat_mumps_icntl_30","ICNTL(30): compute user-specified set of entries in inv(A)","None",mumps->id.ICNTL(30),&mumps->id.ICNTL(30),NULL); */ /* call MatMumpsGetInverse() directly */
1343: PetscOptionsInt("-mat_mumps_icntl_31","ICNTL(31): indicates which factors may be discarded during factorization","None",mumps->id.ICNTL(31),&mumps->id.ICNTL(31),NULL);
1344: /* PetscOptionsInt("-mat_mumps_icntl_32","ICNTL(32): performs the forward elemination of the right-hand sides during factorization","None",mumps->id.ICNTL(32),&mumps->id.ICNTL(32),NULL); -- not supported by PETSc API */
1345: PetscOptionsInt("-mat_mumps_icntl_33","ICNTL(33): compute determinant","None",mumps->id.ICNTL(33),&mumps->id.ICNTL(33),NULL);
1346: PetscOptionsInt("-mat_mumps_icntl_35","ICNTL(35): activates Block Lock Rank (BLR) based factorization","None",mumps->id.ICNTL(35),&mumps->id.ICNTL(35),NULL);
1348: PetscOptionsReal("-mat_mumps_cntl_1","CNTL(1): relative pivoting threshold","None",mumps->id.CNTL(1),&mumps->id.CNTL(1),NULL);
1349: PetscOptionsReal("-mat_mumps_cntl_2","CNTL(2): stopping criterion of refinement","None",mumps->id.CNTL(2),&mumps->id.CNTL(2),NULL);
1350: PetscOptionsReal("-mat_mumps_cntl_3","CNTL(3): absolute pivoting threshold","None",mumps->id.CNTL(3),&mumps->id.CNTL(3),NULL);
1351: PetscOptionsReal("-mat_mumps_cntl_4","CNTL(4): value for static pivoting","None",mumps->id.CNTL(4),&mumps->id.CNTL(4),NULL);
1352: PetscOptionsReal("-mat_mumps_cntl_5","CNTL(5): fixation for null pivots","None",mumps->id.CNTL(5),&mumps->id.CNTL(5),NULL);
1353: PetscOptionsReal("-mat_mumps_cntl_7","CNTL(7): dropping parameter used during BLR","None",mumps->id.CNTL(7),&mumps->id.CNTL(7),NULL);
1355: PetscOptionsString("-mat_mumps_ooc_tmpdir", "out of core directory", "None", mumps->id.ooc_tmpdir, mumps->id.ooc_tmpdir, 256, NULL);
1357: PetscOptionsIntArray("-mat_mumps_view_info","request INFO local to each processor","",info,&ninfo,NULL);
1358: if (ninfo) {
1359: if (ninfo > 40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"number of INFO %d must <= 40\n",ninfo);
1360: PetscMalloc1(ninfo,&mumps->info);
1361: mumps->ninfo = ninfo;
1362: for (i=0; i<ninfo; i++) {
1363: if (info[i] < 0 || info[i]>40) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_USER,"index of INFO %d must between 1 and 40\n",ninfo);
1364: else mumps->info[i] = info[i];
1365: }
1366: }
1368: PetscOptionsEnd();
1369: return(0);
1370: }
1372: PetscErrorCode PetscInitializeMUMPS(Mat A,Mat_MUMPS *mumps)
1373: {
1375: PetscInt nthreads=0;
1378: mumps->petsc_comm = PetscObjectComm((PetscObject)A);
1379: MPI_Comm_size(mumps->petsc_comm,&mumps->petsc_size);
1380: MPI_Comm_rank(mumps->petsc_comm,&mumps->myid); /* so that code like "if (!myid)" still works even if mumps_comm is different */
1382: PetscOptionsHasName(NULL,NULL,"-mat_mumps_use_omp_threads",&mumps->use_petsc_omp_support);
1383: if (mumps->use_petsc_omp_support) nthreads = -1; /* -1 will let PetscOmpCtrlCreate() guess a proper value when user did not supply one */
1384: PetscOptionsGetInt(NULL,NULL,"-mat_mumps_use_omp_threads",&nthreads,NULL);
1385: if (mumps->use_petsc_omp_support) {
1386: #if defined(PETSC_HAVE_OPENMP_SUPPORT)
1387: PetscOmpCtrlCreate(mumps->petsc_comm,nthreads,&mumps->omp_ctrl);
1388: PetscOmpCtrlGetOmpComms(mumps->omp_ctrl,&mumps->omp_comm,&mumps->mumps_comm,&mumps->is_omp_master);
1389: #else
1390: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP_SYS,"the system does not have PETSc OpenMP support but you added the -mat_mumps_use_omp_threads option. Configure PETSc with --with-openmp --download-hwloc (or --with-hwloc) to enable it, see more in MATSOLVERMUMPS manual\n");
1391: #endif
1392: } else {
1393: mumps->omp_comm = PETSC_COMM_SELF;
1394: mumps->mumps_comm = mumps->petsc_comm;
1395: mumps->is_omp_master = PETSC_TRUE;
1396: }
1397: MPI_Comm_size(mumps->omp_comm,&mumps->omp_comm_size);
1399: mumps->id.comm_fortran = MPI_Comm_c2f(mumps->mumps_comm);
1400: mumps->id.job = JOB_INIT;
1401: mumps->id.par = 1; /* host participates factorizaton and solve */
1402: mumps->id.sym = mumps->sym;
1404: PetscMUMPS_c(mumps);
1406: /* copy MUMPS default control values from master to slaves. Although slaves do not call MUMPS, they may access these values in code.
1407: For example, ICNTL(9) is initialized to 1 by MUMPS and slaves check ICNTL(9) in MatSolve_MUMPS.
1408: */
1409: MPI_Bcast(mumps->id.icntl,40,MPIU_INT, 0,mumps->omp_comm); /* see MUMPS-5.1.2 Manual Section 9 */
1410: MPI_Bcast(mumps->id.cntl, 15,MPIU_REAL,0,mumps->omp_comm);
1412: mumps->scat_rhs = NULL;
1413: mumps->scat_sol = NULL;
1415: /* set PETSc-MUMPS default options - override MUMPS default */
1416: mumps->id.ICNTL(3) = 0;
1417: mumps->id.ICNTL(4) = 0;
1418: if (mumps->petsc_size == 1) {
1419: mumps->id.ICNTL(18) = 0; /* centralized assembled matrix input */
1420: } else {
1421: mumps->id.ICNTL(18) = 3; /* distributed assembled matrix input */
1422: mumps->id.ICNTL(20) = 0; /* rhs is in dense format */
1423: mumps->id.ICNTL(21) = 1; /* distributed solution */
1424: }
1426: /* schur */
1427: mumps->id.size_schur = 0;
1428: mumps->id.listvar_schur = NULL;
1429: mumps->id.schur = NULL;
1430: mumps->sizeredrhs = 0;
1431: mumps->schur_sol = NULL;
1432: mumps->schur_sizesol = 0;
1433: return(0);
1434: }
1436: PetscErrorCode MatFactorSymbolic_MUMPS_ReportIfError(Mat F,Mat A,const MatFactorInfo *info,Mat_MUMPS *mumps)
1437: {
1441: MPI_Bcast(mumps->id.infog, 40,MPIU_INT, 0,mumps->omp_comm); /* see MUMPS-5.1.2 manual p82 */
1442: MPI_Bcast(mumps->id.rinfog,20,MPIU_REAL,0,mumps->omp_comm);
1443: if (mumps->id.INFOG(1) < 0) {
1444: if (A->erroriffailure) {
1445: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error reported by MUMPS in analysis phase: INFOG(1)=%d\n",mumps->id.INFOG(1));
1446: } else {
1447: if (mumps->id.INFOG(1) == -6) {
1448: PetscInfo2(F,"matrix is singular in structure, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1449: F->factorerrortype = MAT_FACTOR_STRUCT_ZEROPIVOT;
1450: } else if (mumps->id.INFOG(1) == -5 || mumps->id.INFOG(1) == -7) {
1451: PetscInfo2(F,"problem of workspace, INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1452: F->factorerrortype = MAT_FACTOR_OUTMEMORY;
1453: } else {
1454: PetscInfo2(F,"Error reported by MUMPS in analysis phase: INFOG(1)=%d, INFO(2)=%d\n",mumps->id.INFOG(1),mumps->id.INFO(2));
1455: F->factorerrortype = MAT_FACTOR_OTHER;
1456: }
1457: }
1458: }
1459: return(0);
1460: }
1462: /* Note Petsc r(=c) permutation is used when mumps->id.ICNTL(7)==1 with centralized assembled matrix input; otherwise r and c are ignored */
1463: PetscErrorCode MatLUFactorSymbolic_AIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
1464: {
1465: Mat_MUMPS *mumps = (Mat_MUMPS*)F->data;
1467: Vec b;
1468: IS is_iden;
1469: const PetscInt M = A->rmap->N;
1472: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
1474: /* Set MUMPS options from the options database */
1475: PetscSetMUMPSFromOptions(F,A);
1477: (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);
1478: MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);
1480: /* analysis phase */
1481: /*----------------*/
1482: mumps->id.job = JOB_FACTSYMBOLIC;
1483: mumps->id.n = M;
1484: switch (mumps->id.ICNTL(18)) {
1485: case 0: /* centralized assembled matrix input */
1486: if (!mumps->myid) {
1487: mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1488: if (mumps->id.ICNTL(6)>1) {
1489: mumps->id.a = (MumpsScalar*)mumps->val;
1490: }
1491: if (mumps->id.ICNTL(7) == 1) { /* use user-provide matrix ordering - assuming r = c ordering */
1492: /*
1493: PetscBool flag;
1494: ISEqual(r,c,&flag);
1495: if (!flag) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"row_perm != col_perm");
1496: ISView(r,PETSC_VIEWER_STDOUT_SELF);
1497: */
1498: if (!mumps->myid) {
1499: const PetscInt *idx;
1500: PetscInt i,*perm_in;
1502: PetscMalloc1(M,&perm_in);
1503: ISGetIndices(r,&idx);
1505: mumps->id.perm_in = perm_in;
1506: for (i=0; i<M; i++) perm_in[i] = idx[i]+1; /* perm_in[]: start from 1, not 0! */
1507: ISRestoreIndices(r,&idx);
1508: }
1509: }
1510: }
1511: break;
1512: case 3: /* distributed assembled matrix input (size>1) */
1513: mumps->id.nz_loc = mumps->nz;
1514: mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1515: if (mumps->id.ICNTL(6)>1) {
1516: mumps->id.a_loc = (MumpsScalar*)mumps->val;
1517: }
1518: /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1519: if (!mumps->myid) {
1520: VecCreateSeq(PETSC_COMM_SELF,A->rmap->N,&mumps->b_seq);
1521: ISCreateStride(PETSC_COMM_SELF,A->rmap->N,0,1,&is_iden);
1522: } else {
1523: VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1524: ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1525: }
1526: MatCreateVecs(A,NULL,&b);
1527: VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1528: ISDestroy(&is_iden);
1529: VecDestroy(&b);
1530: break;
1531: }
1532: PetscMUMPS_c(mumps);
1533: MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);
1535: F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
1536: F->ops->solve = MatSolve_MUMPS;
1537: F->ops->solvetranspose = MatSolveTranspose_MUMPS;
1538: F->ops->matsolve = MatMatSolve_MUMPS;
1539: F->ops->mattransposesolve = MatMatTransposeSolve_MUMPS;
1540: return(0);
1541: }
1543: /* Note the Petsc r and c permutations are ignored */
1544: PetscErrorCode MatLUFactorSymbolic_BAIJMUMPS(Mat F,Mat A,IS r,IS c,const MatFactorInfo *info)
1545: {
1546: Mat_MUMPS *mumps = (Mat_MUMPS*)F->data;
1548: Vec b;
1549: IS is_iden;
1550: const PetscInt M = A->rmap->N;
1553: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
1555: /* Set MUMPS options from the options database */
1556: PetscSetMUMPSFromOptions(F,A);
1558: (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);
1559: MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);
1561: /* analysis phase */
1562: /*----------------*/
1563: mumps->id.job = JOB_FACTSYMBOLIC;
1564: mumps->id.n = M;
1565: switch (mumps->id.ICNTL(18)) {
1566: case 0: /* centralized assembled matrix input */
1567: if (!mumps->myid) {
1568: mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1569: if (mumps->id.ICNTL(6)>1) {
1570: mumps->id.a = (MumpsScalar*)mumps->val;
1571: }
1572: }
1573: break;
1574: case 3: /* distributed assembled matrix input (size>1) */
1575: mumps->id.nz_loc = mumps->nz;
1576: mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1577: if (mumps->id.ICNTL(6)>1) {
1578: mumps->id.a_loc = (MumpsScalar*)mumps->val;
1579: }
1580: /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1581: if (!mumps->myid) {
1582: VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);
1583: ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);
1584: } else {
1585: VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1586: ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1587: }
1588: MatCreateVecs(A,NULL,&b);
1589: VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1590: ISDestroy(&is_iden);
1591: VecDestroy(&b);
1592: break;
1593: }
1594: PetscMUMPS_c(mumps);
1595: MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);
1597: F->ops->lufactornumeric = MatFactorNumeric_MUMPS;
1598: F->ops->solve = MatSolve_MUMPS;
1599: F->ops->solvetranspose = MatSolveTranspose_MUMPS;
1600: return(0);
1601: }
1603: /* Note the Petsc r permutation and factor info are ignored */
1604: PetscErrorCode MatCholeskyFactorSymbolic_MUMPS(Mat F,Mat A,IS r,const MatFactorInfo *info)
1605: {
1606: Mat_MUMPS *mumps = (Mat_MUMPS*)F->data;
1608: Vec b;
1609: IS is_iden;
1610: const PetscInt M = A->rmap->N;
1613: mumps->matstruc = DIFFERENT_NONZERO_PATTERN;
1615: /* Set MUMPS options from the options database */
1616: PetscSetMUMPSFromOptions(F,A);
1618: (*mumps->ConvertToTriples)(A, 1, MAT_INITIAL_MATRIX, &mumps->nz, &mumps->irn, &mumps->jcn, &mumps->val);
1619: MatMumpsGatherNonzerosOnMaster(MAT_INITIAL_MATRIX,mumps);
1621: /* analysis phase */
1622: /*----------------*/
1623: mumps->id.job = JOB_FACTSYMBOLIC;
1624: mumps->id.n = M;
1625: switch (mumps->id.ICNTL(18)) {
1626: case 0: /* centralized assembled matrix input */
1627: if (!mumps->myid) {
1628: mumps->id.nz =mumps->nz; mumps->id.irn=mumps->irn; mumps->id.jcn=mumps->jcn;
1629: if (mumps->id.ICNTL(6)>1) {
1630: mumps->id.a = (MumpsScalar*)mumps->val;
1631: }
1632: }
1633: break;
1634: case 3: /* distributed assembled matrix input (size>1) */
1635: mumps->id.nz_loc = mumps->nz;
1636: mumps->id.irn_loc=mumps->irn; mumps->id.jcn_loc=mumps->jcn;
1637: if (mumps->id.ICNTL(6)>1) {
1638: mumps->id.a_loc = (MumpsScalar*)mumps->val;
1639: }
1640: /* MUMPS only supports centralized rhs. Create scatter scat_rhs for repeated use in MatSolve() */
1641: if (!mumps->myid) {
1642: VecCreateSeq(PETSC_COMM_SELF,A->cmap->N,&mumps->b_seq);
1643: ISCreateStride(PETSC_COMM_SELF,A->cmap->N,0,1,&is_iden);
1644: } else {
1645: VecCreateSeq(PETSC_COMM_SELF,0,&mumps->b_seq);
1646: ISCreateStride(PETSC_COMM_SELF,0,0,1,&is_iden);
1647: }
1648: MatCreateVecs(A,NULL,&b);
1649: VecScatterCreate(b,is_iden,mumps->b_seq,is_iden,&mumps->scat_rhs);
1650: ISDestroy(&is_iden);
1651: VecDestroy(&b);
1652: break;
1653: }
1654: PetscMUMPS_c(mumps);
1655: MatFactorSymbolic_MUMPS_ReportIfError(F,A,info,mumps);
1657: F->ops->choleskyfactornumeric = MatFactorNumeric_MUMPS;
1658: F->ops->solve = MatSolve_MUMPS;
1659: F->ops->solvetranspose = MatSolve_MUMPS;
1660: F->ops->matsolve = MatMatSolve_MUMPS;
1661: #if defined(PETSC_USE_COMPLEX)
1662: F->ops->getinertia = NULL;
1663: #else
1664: F->ops->getinertia = MatGetInertia_SBAIJMUMPS;
1665: #endif
1666: return(0);
1667: }
1669: PetscErrorCode MatView_MUMPS(Mat A,PetscViewer viewer)
1670: {
1671: PetscErrorCode ierr;
1672: PetscBool iascii;
1673: PetscViewerFormat format;
1674: Mat_MUMPS *mumps=(Mat_MUMPS*)A->data;
1677: /* check if matrix is mumps type */
1678: if (A->ops->solve != MatSolve_MUMPS) return(0);
1680: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1681: if (iascii) {
1682: PetscViewerGetFormat(viewer,&format);
1683: if (format == PETSC_VIEWER_ASCII_INFO) {
1684: PetscViewerASCIIPrintf(viewer,"MUMPS run parameters:\n");
1685: PetscViewerASCIIPrintf(viewer," SYM (matrix type): %d \n",mumps->id.sym);
1686: PetscViewerASCIIPrintf(viewer," PAR (host participation): %d \n",mumps->id.par);
1687: PetscViewerASCIIPrintf(viewer," ICNTL(1) (output for error): %d \n",mumps->id.ICNTL(1));
1688: PetscViewerASCIIPrintf(viewer," ICNTL(2) (output of diagnostic msg): %d \n",mumps->id.ICNTL(2));
1689: PetscViewerASCIIPrintf(viewer," ICNTL(3) (output for global info): %d \n",mumps->id.ICNTL(3));
1690: PetscViewerASCIIPrintf(viewer," ICNTL(4) (level of printing): %d \n",mumps->id.ICNTL(4));
1691: PetscViewerASCIIPrintf(viewer," ICNTL(5) (input mat struct): %d \n",mumps->id.ICNTL(5));
1692: PetscViewerASCIIPrintf(viewer," ICNTL(6) (matrix prescaling): %d \n",mumps->id.ICNTL(6));
1693: PetscViewerASCIIPrintf(viewer," ICNTL(7) (sequential matrix ordering):%d \n",mumps->id.ICNTL(7));
1694: PetscViewerASCIIPrintf(viewer," ICNTL(8) (scaling strategy): %d \n",mumps->id.ICNTL(8));
1695: PetscViewerASCIIPrintf(viewer," ICNTL(10) (max num of refinements): %d \n",mumps->id.ICNTL(10));
1696: PetscViewerASCIIPrintf(viewer," ICNTL(11) (error analysis): %d \n",mumps->id.ICNTL(11));
1697: if (mumps->id.ICNTL(11)>0) {
1698: PetscViewerASCIIPrintf(viewer," RINFOG(4) (inf norm of input mat): %g\n",mumps->id.RINFOG(4));
1699: PetscViewerASCIIPrintf(viewer," RINFOG(5) (inf norm of solution): %g\n",mumps->id.RINFOG(5));
1700: PetscViewerASCIIPrintf(viewer," RINFOG(6) (inf norm of residual): %g\n",mumps->id.RINFOG(6));
1701: PetscViewerASCIIPrintf(viewer," RINFOG(7),RINFOG(8) (backward error est): %g, %g\n",mumps->id.RINFOG(7),mumps->id.RINFOG(8));
1702: PetscViewerASCIIPrintf(viewer," RINFOG(9) (error estimate): %g \n",mumps->id.RINFOG(9));
1703: PetscViewerASCIIPrintf(viewer," RINFOG(10),RINFOG(11)(condition numbers): %g, %g\n",mumps->id.RINFOG(10),mumps->id.RINFOG(11));
1704: }
1705: PetscViewerASCIIPrintf(viewer," ICNTL(12) (efficiency control): %d \n",mumps->id.ICNTL(12));
1706: PetscViewerASCIIPrintf(viewer," ICNTL(13) (efficiency control): %d \n",mumps->id.ICNTL(13));
1707: PetscViewerASCIIPrintf(viewer," ICNTL(14) (percentage of estimated workspace increase): %d \n",mumps->id.ICNTL(14));
1708: /* ICNTL(15-17) not used */
1709: PetscViewerASCIIPrintf(viewer," ICNTL(18) (input mat struct): %d \n",mumps->id.ICNTL(18));
1710: PetscViewerASCIIPrintf(viewer," ICNTL(19) (Schur complement info): %d \n",mumps->id.ICNTL(19));
1711: PetscViewerASCIIPrintf(viewer," ICNTL(20) (rhs sparse pattern): %d \n",mumps->id.ICNTL(20));
1712: PetscViewerASCIIPrintf(viewer," ICNTL(21) (solution struct): %d \n",mumps->id.ICNTL(21));
1713: PetscViewerASCIIPrintf(viewer," ICNTL(22) (in-core/out-of-core facility): %d \n",mumps->id.ICNTL(22));
1714: PetscViewerASCIIPrintf(viewer," ICNTL(23) (max size of memory can be allocated locally):%d \n",mumps->id.ICNTL(23));
1716: PetscViewerASCIIPrintf(viewer," ICNTL(24) (detection of null pivot rows): %d \n",mumps->id.ICNTL(24));
1717: PetscViewerASCIIPrintf(viewer," ICNTL(25) (computation of a null space basis): %d \n",mumps->id.ICNTL(25));
1718: PetscViewerASCIIPrintf(viewer," ICNTL(26) (Schur options for rhs or solution): %d \n",mumps->id.ICNTL(26));
1719: PetscViewerASCIIPrintf(viewer," ICNTL(27) (experimental parameter): %d \n",mumps->id.ICNTL(27));
1720: PetscViewerASCIIPrintf(viewer," ICNTL(28) (use parallel or sequential ordering): %d \n",mumps->id.ICNTL(28));
1721: PetscViewerASCIIPrintf(viewer," ICNTL(29) (parallel ordering): %d \n",mumps->id.ICNTL(29));
1723: PetscViewerASCIIPrintf(viewer," ICNTL(30) (user-specified set of entries in inv(A)): %d \n",mumps->id.ICNTL(30));
1724: PetscViewerASCIIPrintf(viewer," ICNTL(31) (factors is discarded in the solve phase): %d \n",mumps->id.ICNTL(31));
1725: PetscViewerASCIIPrintf(viewer," ICNTL(33) (compute determinant): %d \n",mumps->id.ICNTL(33));
1726: PetscViewerASCIIPrintf(viewer," ICNTL(35) (activate BLR based factorization): %d \n",mumps->id.ICNTL(35));
1728: PetscViewerASCIIPrintf(viewer," CNTL(1) (relative pivoting threshold): %g \n",mumps->id.CNTL(1));
1729: PetscViewerASCIIPrintf(viewer," CNTL(2) (stopping criterion of refinement): %g \n",mumps->id.CNTL(2));
1730: PetscViewerASCIIPrintf(viewer," CNTL(3) (absolute pivoting threshold): %g \n",mumps->id.CNTL(3));
1731: PetscViewerASCIIPrintf(viewer," CNTL(4) (value of static pivoting): %g \n",mumps->id.CNTL(4));
1732: PetscViewerASCIIPrintf(viewer," CNTL(5) (fixation for null pivots): %g \n",mumps->id.CNTL(5));
1733: PetscViewerASCIIPrintf(viewer," CNTL(7) (dropping parameter for BLR): %g \n",mumps->id.CNTL(7));
1735: /* infomation local to each processor */
1736: PetscViewerASCIIPrintf(viewer, " RINFO(1) (local estimated flops for the elimination after analysis): \n");
1737: PetscViewerASCIIPushSynchronized(viewer);
1738: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(1));
1739: PetscViewerFlush(viewer);
1740: PetscViewerASCIIPrintf(viewer, " RINFO(2) (local estimated flops for the assembly after factorization): \n");
1741: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(2));
1742: PetscViewerFlush(viewer);
1743: PetscViewerASCIIPrintf(viewer, " RINFO(3) (local estimated flops for the elimination after factorization): \n");
1744: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %g \n",mumps->myid,mumps->id.RINFO(3));
1745: PetscViewerFlush(viewer);
1747: PetscViewerASCIIPrintf(viewer, " INFO(15) (estimated size of (in MB) MUMPS internal data for running numerical factorization): \n");
1748: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(15));
1749: PetscViewerFlush(viewer);
1751: PetscViewerASCIIPrintf(viewer, " INFO(16) (size of (in MB) MUMPS internal data used during numerical factorization): \n");
1752: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(16));
1753: PetscViewerFlush(viewer);
1755: PetscViewerASCIIPrintf(viewer, " INFO(23) (num of pivots eliminated on this processor after factorization): \n");
1756: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(23));
1757: PetscViewerFlush(viewer);
1759: if (mumps->ninfo && mumps->ninfo <= 40){
1760: PetscInt i;
1761: for (i=0; i<mumps->ninfo; i++){
1762: PetscViewerASCIIPrintf(viewer, " INFO(%d): \n",mumps->info[i]);
1763: PetscViewerASCIISynchronizedPrintf(viewer," [%d] %d \n",mumps->myid,mumps->id.INFO(mumps->info[i]));
1764: PetscViewerFlush(viewer);
1765: }
1766: }
1767: PetscViewerASCIIPopSynchronized(viewer);
1769: if (!mumps->myid) { /* information from the host */
1770: PetscViewerASCIIPrintf(viewer," RINFOG(1) (global estimated flops for the elimination after analysis): %g \n",mumps->id.RINFOG(1));
1771: PetscViewerASCIIPrintf(viewer," RINFOG(2) (global estimated flops for the assembly after factorization): %g \n",mumps->id.RINFOG(2));
1772: PetscViewerASCIIPrintf(viewer," RINFOG(3) (global estimated flops for the elimination after factorization): %g \n",mumps->id.RINFOG(3));
1773: 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));
1775: PetscViewerASCIIPrintf(viewer," INFOG(3) (estimated real workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(3));
1776: PetscViewerASCIIPrintf(viewer," INFOG(4) (estimated integer workspace for factors on all processors after analysis): %d \n",mumps->id.INFOG(4));
1777: PetscViewerASCIIPrintf(viewer," INFOG(5) (estimated maximum front size in the complete tree): %d \n",mumps->id.INFOG(5));
1778: PetscViewerASCIIPrintf(viewer," INFOG(6) (number of nodes in the complete tree): %d \n",mumps->id.INFOG(6));
1779: PetscViewerASCIIPrintf(viewer," INFOG(7) (ordering option effectively use after analysis): %d \n",mumps->id.INFOG(7));
1780: PetscViewerASCIIPrintf(viewer," INFOG(8) (structural symmetry in percent of the permuted matrix after analysis): %d \n",mumps->id.INFOG(8));
1781: PetscViewerASCIIPrintf(viewer," INFOG(9) (total real/complex workspace to store the matrix factors after factorization): %d \n",mumps->id.INFOG(9));
1782: PetscViewerASCIIPrintf(viewer," INFOG(10) (total integer space store the matrix factors after factorization): %d \n",mumps->id.INFOG(10));
1783: PetscViewerASCIIPrintf(viewer," INFOG(11) (order of largest frontal matrix after factorization): %d \n",mumps->id.INFOG(11));
1784: PetscViewerASCIIPrintf(viewer," INFOG(12) (number of off-diagonal pivots): %d \n",mumps->id.INFOG(12));
1785: PetscViewerASCIIPrintf(viewer," INFOG(13) (number of delayed pivots after factorization): %d \n",mumps->id.INFOG(13));
1786: PetscViewerASCIIPrintf(viewer," INFOG(14) (number of memory compress after factorization): %d \n",mumps->id.INFOG(14));
1787: PetscViewerASCIIPrintf(viewer," INFOG(15) (number of steps of iterative refinement after solution): %d \n",mumps->id.INFOG(15));
1788: 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));
1789: 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));
1790: 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));
1791: PetscViewerASCIIPrintf(viewer," INFOG(19) (size of all MUMPS internal data allocated during factorization: sum over all processors): %d \n",mumps->id.INFOG(19));
1792: PetscViewerASCIIPrintf(viewer," INFOG(20) (estimated number of entries in the factors): %d \n",mumps->id.INFOG(20));
1793: 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));
1794: PetscViewerASCIIPrintf(viewer," INFOG(22) (size in MB of memory effectively used during factorization - sum over all processors): %d \n",mumps->id.INFOG(22));
1795: PetscViewerASCIIPrintf(viewer," INFOG(23) (after analysis: value of ICNTL(6) effectively used): %d \n",mumps->id.INFOG(23));
1796: PetscViewerASCIIPrintf(viewer," INFOG(24) (after analysis: value of ICNTL(12) effectively used): %d \n",mumps->id.INFOG(24));
1797: PetscViewerASCIIPrintf(viewer," INFOG(25) (after factorization: number of pivots modified by static pivoting): %d \n",mumps->id.INFOG(25));
1798: PetscViewerASCIIPrintf(viewer," INFOG(28) (after factorization: number of null pivots encountered): %d\n",mumps->id.INFOG(28));
1799: PetscViewerASCIIPrintf(viewer," INFOG(29) (after factorization: effective number of entries in the factors (sum over all processors)): %d\n",mumps->id.INFOG(29));
1800: 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));
1801: PetscViewerASCIIPrintf(viewer," INFOG(32) (after analysis: type of analysis done): %d\n",mumps->id.INFOG(32));
1802: PetscViewerASCIIPrintf(viewer," INFOG(33) (value used for ICNTL(8)): %d\n",mumps->id.INFOG(33));
1803: PetscViewerASCIIPrintf(viewer," INFOG(34) (exponent of the determinant if determinant is requested): %d\n",mumps->id.INFOG(34));
1804: }
1805: }
1806: }
1807: return(0);
1808: }
1810: PetscErrorCode MatGetInfo_MUMPS(Mat A,MatInfoType flag,MatInfo *info)
1811: {
1812: Mat_MUMPS *mumps =(Mat_MUMPS*)A->data;
1815: info->block_size = 1.0;
1816: info->nz_allocated = mumps->id.INFOG(20);
1817: info->nz_used = mumps->id.INFOG(20);
1818: info->nz_unneeded = 0.0;
1819: info->assemblies = 0.0;
1820: info->mallocs = 0.0;
1821: info->memory = 0.0;
1822: info->fill_ratio_given = 0;
1823: info->fill_ratio_needed = 0;
1824: info->factor_mallocs = 0;
1825: return(0);
1826: }
1828: /* -------------------------------------------------------------------------------------------*/
1829: PetscErrorCode MatFactorSetSchurIS_MUMPS(Mat F, IS is)
1830: {
1831: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
1832: const PetscInt *idxs;
1833: PetscInt size,i;
1837: ISGetLocalSize(is,&size);
1838: if (mumps->petsc_size > 1) {
1839: PetscBool ls,gs; /* gs is false if any rank other than root has non-empty IS */
1841: ls = mumps->myid ? (size ? PETSC_FALSE : PETSC_TRUE) : PETSC_TRUE; /* always true on root; false on others if their size != 0 */
1842: MPI_Allreduce(&ls,&gs,1,MPIU_BOOL,MPI_LAND,mumps->petsc_comm);
1843: if (!gs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MUMPS distributed parallel Schur complements not yet supported from PETSc\n");
1844: }
1845: if (mumps->id.size_schur != size) {
1846: PetscFree2(mumps->id.listvar_schur,mumps->id.schur);
1847: mumps->id.size_schur = size;
1848: mumps->id.schur_lld = size;
1849: PetscMalloc2(size,&mumps->id.listvar_schur,size*size,&mumps->id.schur);
1850: }
1852: /* Schur complement matrix */
1853: MatCreateSeqDense(PETSC_COMM_SELF,mumps->id.size_schur,mumps->id.size_schur,(PetscScalar*)mumps->id.schur,&F->schur);
1854: if (mumps->sym == 1) {
1855: MatSetOption(F->schur,MAT_SPD,PETSC_TRUE);
1856: }
1858: /* MUMPS expects Fortran style indices */
1859: ISGetIndices(is,&idxs);
1860: PetscMemcpy(mumps->id.listvar_schur,idxs,size*sizeof(PetscInt));
1861: for (i=0;i<size;i++) mumps->id.listvar_schur[i]++;
1862: ISRestoreIndices(is,&idxs);
1863: if (mumps->petsc_size > 1) {
1864: mumps->id.ICNTL(19) = 1; /* MUMPS returns Schur centralized on the host */
1865: } else {
1866: if (F->factortype == MAT_FACTOR_LU) {
1867: mumps->id.ICNTL(19) = 3; /* MUMPS returns full matrix */
1868: } else {
1869: mumps->id.ICNTL(19) = 2; /* MUMPS returns lower triangular part */
1870: }
1871: }
1872: /* set a special value of ICNTL (not handled my MUMPS) to be used in the solve phase by PETSc */
1873: mumps->id.ICNTL(26) = -1;
1874: return(0);
1875: }
1877: /* -------------------------------------------------------------------------------------------*/
1878: PetscErrorCode MatFactorCreateSchurComplement_MUMPS(Mat F,Mat* S)
1879: {
1880: Mat St;
1881: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
1882: PetscScalar *array;
1883: #if defined(PETSC_USE_COMPLEX)
1884: PetscScalar im = PetscSqrtScalar((PetscScalar)-1.0);
1885: #endif
1889: if (!mumps->id.ICNTL(19)) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ORDER,"Schur complement mode not selected! You should call MatFactorSetSchurIS to enable it");
1890: MatCreate(PETSC_COMM_SELF,&St);
1891: MatSetSizes(St,PETSC_DECIDE,PETSC_DECIDE,mumps->id.size_schur,mumps->id.size_schur);
1892: MatSetType(St,MATDENSE);
1893: MatSetUp(St);
1894: MatDenseGetArray(St,&array);
1895: if (!mumps->sym) { /* MUMPS always return a full matrix */
1896: if (mumps->id.ICNTL(19) == 1) { /* stored by rows */
1897: PetscInt i,j,N=mumps->id.size_schur;
1898: for (i=0;i<N;i++) {
1899: for (j=0;j<N;j++) {
1900: #if !defined(PETSC_USE_COMPLEX)
1901: PetscScalar val = mumps->id.schur[i*N+j];
1902: #else
1903: PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1904: #endif
1905: array[j*N+i] = val;
1906: }
1907: }
1908: } else { /* stored by columns */
1909: PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));
1910: }
1911: } else { /* either full or lower-triangular (not packed) */
1912: if (mumps->id.ICNTL(19) == 2) { /* lower triangular stored by columns */
1913: PetscInt i,j,N=mumps->id.size_schur;
1914: for (i=0;i<N;i++) {
1915: for (j=i;j<N;j++) {
1916: #if !defined(PETSC_USE_COMPLEX)
1917: PetscScalar val = mumps->id.schur[i*N+j];
1918: #else
1919: PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1920: #endif
1921: array[i*N+j] = val;
1922: array[j*N+i] = val;
1923: }
1924: }
1925: } else if (mumps->id.ICNTL(19) == 3) { /* full matrix */
1926: PetscMemcpy(array,mumps->id.schur,mumps->id.size_schur*mumps->id.size_schur*sizeof(PetscScalar));
1927: } else { /* ICNTL(19) == 1 lower triangular stored by rows */
1928: PetscInt i,j,N=mumps->id.size_schur;
1929: for (i=0;i<N;i++) {
1930: for (j=0;j<i+1;j++) {
1931: #if !defined(PETSC_USE_COMPLEX)
1932: PetscScalar val = mumps->id.schur[i*N+j];
1933: #else
1934: PetscScalar val = mumps->id.schur[i*N+j].r + im*mumps->id.schur[i*N+j].i;
1935: #endif
1936: array[i*N+j] = val;
1937: array[j*N+i] = val;
1938: }
1939: }
1940: }
1941: }
1942: MatDenseRestoreArray(St,&array);
1943: *S = St;
1944: return(0);
1945: }
1947: /* -------------------------------------------------------------------------------------------*/
1948: PetscErrorCode MatMumpsSetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt ival)
1949: {
1950: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
1953: mumps->id.ICNTL(icntl) = ival;
1954: return(0);
1955: }
1957: PetscErrorCode MatMumpsGetIcntl_MUMPS(Mat F,PetscInt icntl,PetscInt *ival)
1958: {
1959: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
1962: *ival = mumps->id.ICNTL(icntl);
1963: return(0);
1964: }
1966: /*@
1967: MatMumpsSetIcntl - Set MUMPS parameter ICNTL()
1969: Logically Collective on Mat
1971: Input Parameters:
1972: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
1973: . icntl - index of MUMPS parameter array ICNTL()
1974: - ival - value of MUMPS ICNTL(icntl)
1976: Options Database:
1977: . -mat_mumps_icntl_<icntl> <ival>
1979: Level: beginner
1981: References:
1982: . MUMPS Users' Guide
1984: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
1985: @*/
1986: PetscErrorCode MatMumpsSetIcntl(Mat F,PetscInt icntl,PetscInt ival)
1987: {
1992: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
1995: PetscTryMethod(F,"MatMumpsSetIcntl_C",(Mat,PetscInt,PetscInt),(F,icntl,ival));
1996: return(0);
1997: }
1999: /*@
2000: MatMumpsGetIcntl - Get MUMPS parameter ICNTL()
2002: Logically Collective on Mat
2004: Input Parameters:
2005: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2006: - icntl - index of MUMPS parameter array ICNTL()
2008: Output Parameter:
2009: . ival - value of MUMPS ICNTL(icntl)
2011: Level: beginner
2013: References:
2014: . MUMPS Users' Guide
2016: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2017: @*/
2018: PetscErrorCode MatMumpsGetIcntl(Mat F,PetscInt icntl,PetscInt *ival)
2019: {
2024: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2027: PetscUseMethod(F,"MatMumpsGetIcntl_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2028: return(0);
2029: }
2031: /* -------------------------------------------------------------------------------------------*/
2032: PetscErrorCode MatMumpsSetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal val)
2033: {
2034: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2037: mumps->id.CNTL(icntl) = val;
2038: return(0);
2039: }
2041: PetscErrorCode MatMumpsGetCntl_MUMPS(Mat F,PetscInt icntl,PetscReal *val)
2042: {
2043: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2046: *val = mumps->id.CNTL(icntl);
2047: return(0);
2048: }
2050: /*@
2051: MatMumpsSetCntl - Set MUMPS parameter CNTL()
2053: Logically Collective on Mat
2055: Input Parameters:
2056: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2057: . icntl - index of MUMPS parameter array CNTL()
2058: - val - value of MUMPS CNTL(icntl)
2060: Options Database:
2061: . -mat_mumps_cntl_<icntl> <val>
2063: Level: beginner
2065: References:
2066: . MUMPS Users' Guide
2068: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2069: @*/
2070: PetscErrorCode MatMumpsSetCntl(Mat F,PetscInt icntl,PetscReal val)
2071: {
2076: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2079: PetscTryMethod(F,"MatMumpsSetCntl_C",(Mat,PetscInt,PetscReal),(F,icntl,val));
2080: return(0);
2081: }
2083: /*@
2084: MatMumpsGetCntl - Get MUMPS parameter CNTL()
2086: Logically Collective on Mat
2088: Input Parameters:
2089: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2090: - icntl - index of MUMPS parameter array CNTL()
2092: Output Parameter:
2093: . val - value of MUMPS CNTL(icntl)
2095: Level: beginner
2097: References:
2098: . MUMPS Users' Guide
2100: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2101: @*/
2102: PetscErrorCode MatMumpsGetCntl(Mat F,PetscInt icntl,PetscReal *val)
2103: {
2108: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2111: PetscUseMethod(F,"MatMumpsGetCntl_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2112: return(0);
2113: }
2115: PetscErrorCode MatMumpsGetInfo_MUMPS(Mat F,PetscInt icntl,PetscInt *info)
2116: {
2117: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2120: *info = mumps->id.INFO(icntl);
2121: return(0);
2122: }
2124: PetscErrorCode MatMumpsGetInfog_MUMPS(Mat F,PetscInt icntl,PetscInt *infog)
2125: {
2126: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2129: *infog = mumps->id.INFOG(icntl);
2130: return(0);
2131: }
2133: PetscErrorCode MatMumpsGetRinfo_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfo)
2134: {
2135: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2138: *rinfo = mumps->id.RINFO(icntl);
2139: return(0);
2140: }
2142: PetscErrorCode MatMumpsGetRinfog_MUMPS(Mat F,PetscInt icntl,PetscReal *rinfog)
2143: {
2144: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2147: *rinfog = mumps->id.RINFOG(icntl);
2148: return(0);
2149: }
2151: PetscErrorCode MatMumpsGetInverse_MUMPS(Mat F,Mat spRHS)
2152: {
2154: Mat Bt = NULL,Btseq = NULL;
2155: PetscBool flg;
2156: Mat_MUMPS *mumps =(Mat_MUMPS*)F->data;
2157: PetscScalar *aa;
2158: PetscInt spnr,*ia,*ja;
2162: PetscObjectTypeCompare((PetscObject)spRHS,MATTRANSPOSEMAT,&flg);
2163: if (flg) {
2164: MatTransposeGetMat(spRHS,&Bt);
2165: } else SETERRQ(PetscObjectComm((PetscObject)spRHS),PETSC_ERR_ARG_WRONG,"Matrix spRHS must be type MATTRANSPOSEMAT matrix");
2167: MatMumpsSetIcntl(F,30,1);
2169: if (mumps->petsc_size > 1) {
2170: Mat_MPIAIJ *b = (Mat_MPIAIJ*)Bt->data;
2171: Btseq = b->A;
2172: } else {
2173: Btseq = Bt;
2174: }
2176: if (!mumps->myid) {
2177: MatSeqAIJGetArray(Btseq,&aa);
2178: MatGetRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
2179: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
2181: mumps->id.irhs_ptr = ia;
2182: mumps->id.irhs_sparse = ja;
2183: mumps->id.nz_rhs = ia[spnr] - 1;
2184: mumps->id.rhs_sparse = (MumpsScalar*)aa;
2185: } else {
2186: mumps->id.irhs_ptr = NULL;
2187: mumps->id.irhs_sparse = NULL;
2188: mumps->id.nz_rhs = 0;
2189: mumps->id.rhs_sparse = NULL;
2190: }
2191: mumps->id.ICNTL(20) = 1; /* rhs is sparse */
2192: mumps->id.ICNTL(21) = 0; /* solution is in assembled centralized format */
2194: /* solve phase */
2195: /*-------------*/
2196: mumps->id.job = JOB_SOLVE;
2197: PetscMUMPS_c(mumps);
2198: if (mumps->id.INFOG(1) < 0)
2199: 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));
2201: if (!mumps->myid) {
2202: MatSeqAIJRestoreArray(Btseq,&aa);
2203: MatRestoreRowIJ(Btseq,1,PETSC_FALSE,PETSC_FALSE,&spnr,(const PetscInt**)&ia,(const PetscInt**)&ja,&flg);
2204: if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Cannot get IJ structure");
2205: }
2206: return(0);
2207: }
2209: /*@
2210: MatMumpsGetInverse - Get user-specified set of entries in inverse of A
2212: Logically Collective on Mat
2214: Input Parameters:
2215: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2216: - spRHS - sequential sparse matrix in MATTRANSPOSEMAT format holding specified indices in processor[0]
2218: Output Parameter:
2219: . spRHS - requested entries of inverse of A
2221: Level: beginner
2223: References:
2224: . MUMPS Users' Guide
2226: .seealso: MatGetFactor(), MatCreateTranspose()
2227: @*/
2228: PetscErrorCode MatMumpsGetInverse(Mat F,Mat spRHS)
2229: {
2234: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2235: PetscUseMethod(F,"MatMumpsGetInverse_C",(Mat,Mat),(F,spRHS));
2236: return(0);
2237: }
2239: PetscErrorCode MatMumpsGetInverseTranspose_MUMPS(Mat F,Mat spRHST)
2240: {
2242: Mat spRHS;
2245: MatCreateTranspose(spRHST,&spRHS);
2246: MatMumpsGetInverse_MUMPS(F,spRHS);
2247: MatDestroy(&spRHS);
2248: return(0);
2249: }
2251: /*@
2252: MatMumpsGetInverseTranspose - Get user-specified set of entries in inverse of matrix A^T
2254: Logically Collective on Mat
2256: Input Parameters:
2257: + F - the factored matrix of A obtained by calling MatGetFactor() from PETSc-MUMPS interface
2258: - spRHST - sequential sparse matrix in MATAIJ format holding specified indices of A^T in processor[0]
2260: Output Parameter:
2261: . spRHST - requested entries of inverse of A^T
2263: Level: beginner
2265: References:
2266: . MUMPS Users' Guide
2268: .seealso: MatGetFactor(), MatCreateTranspose(), MatMumpsGetInverse()
2269: @*/
2270: PetscErrorCode MatMumpsGetInverseTranspose(Mat F,Mat spRHST)
2271: {
2273: PetscBool flg;
2277: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2278: PetscObjectTypeCompareAny((PetscObject)spRHST,&flg,MATSEQAIJ,MATMPIAIJ,NULL);
2279: if (!flg) SETERRQ(PetscObjectComm((PetscObject)spRHST),PETSC_ERR_ARG_WRONG,"Matrix spRHST must be MATAIJ matrix");
2281: PetscUseMethod(F,"MatMumpsGetInverseTranspose_C",(Mat,Mat),(F,spRHST));
2282: return(0);
2283: }
2285: /*@
2286: MatMumpsGetInfo - Get MUMPS parameter INFO()
2288: Logically Collective on Mat
2290: Input Parameters:
2291: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2292: - icntl - index of MUMPS parameter array INFO()
2294: Output Parameter:
2295: . ival - value of MUMPS INFO(icntl)
2297: Level: beginner
2299: References:
2300: . MUMPS Users' Guide
2302: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2303: @*/
2304: PetscErrorCode MatMumpsGetInfo(Mat F,PetscInt icntl,PetscInt *ival)
2305: {
2310: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2312: PetscUseMethod(F,"MatMumpsGetInfo_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2313: return(0);
2314: }
2316: /*@
2317: MatMumpsGetInfog - Get MUMPS parameter INFOG()
2319: Logically Collective on Mat
2321: Input Parameters:
2322: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2323: - icntl - index of MUMPS parameter array INFOG()
2325: Output Parameter:
2326: . ival - value of MUMPS INFOG(icntl)
2328: Level: beginner
2330: References:
2331: . MUMPS Users' Guide
2333: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2334: @*/
2335: PetscErrorCode MatMumpsGetInfog(Mat F,PetscInt icntl,PetscInt *ival)
2336: {
2341: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2343: PetscUseMethod(F,"MatMumpsGetInfog_C",(Mat,PetscInt,PetscInt*),(F,icntl,ival));
2344: return(0);
2345: }
2347: /*@
2348: MatMumpsGetRinfo - Get MUMPS parameter RINFO()
2350: Logically Collective on Mat
2352: Input Parameters:
2353: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2354: - icntl - index of MUMPS parameter array RINFO()
2356: Output Parameter:
2357: . val - value of MUMPS RINFO(icntl)
2359: Level: beginner
2361: References:
2362: . MUMPS Users' Guide
2364: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2365: @*/
2366: PetscErrorCode MatMumpsGetRinfo(Mat F,PetscInt icntl,PetscReal *val)
2367: {
2372: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2374: PetscUseMethod(F,"MatMumpsGetRinfo_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2375: return(0);
2376: }
2378: /*@
2379: MatMumpsGetRinfog - Get MUMPS parameter RINFOG()
2381: Logically Collective on Mat
2383: Input Parameters:
2384: + F - the factored matrix obtained by calling MatGetFactor() from PETSc-MUMPS interface
2385: - icntl - index of MUMPS parameter array RINFOG()
2387: Output Parameter:
2388: . val - value of MUMPS RINFOG(icntl)
2390: Level: beginner
2392: References:
2393: . MUMPS Users' Guide
2395: .seealso: MatGetFactor(), MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog()
2396: @*/
2397: PetscErrorCode MatMumpsGetRinfog(Mat F,PetscInt icntl,PetscReal *val)
2398: {
2403: if (!F->factortype) SETERRQ(PetscObjectComm((PetscObject)F),PETSC_ERR_ARG_WRONGSTATE,"Only for factored matrix");
2405: PetscUseMethod(F,"MatMumpsGetRinfog_C",(Mat,PetscInt,PetscReal*),(F,icntl,val));
2406: return(0);
2407: }
2409: /*MC
2410: MATSOLVERMUMPS - A matrix type providing direct solvers (LU and Cholesky) for
2411: distributed and sequential matrices via the external package MUMPS.
2413: Works with MATAIJ and MATSBAIJ matrices
2415: Use ./configure --download-mumps --download-scalapack --download-parmetis --download-metis --download-ptscotch to have PETSc installed with MUMPS
2417: 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.
2419: Use -pc_type cholesky or lu -pc_factor_mat_solver_type mumps to use this direct solver
2421: Options Database Keys:
2422: + -mat_mumps_icntl_1 - ICNTL(1): output stream for error messages
2423: . -mat_mumps_icntl_2 - ICNTL(2): output stream for diagnostic printing, statistics, and warning
2424: . -mat_mumps_icntl_3 - ICNTL(3): output stream for global information, collected on the host
2425: . -mat_mumps_icntl_4 - ICNTL(4): level of printing (0 to 4)
2426: . -mat_mumps_icntl_6 - ICNTL(6): permutes to a zero-free diagonal and/or scale the matrix (0 to 7)
2427: . -mat_mumps_icntl_7 - ICNTL(7): computes a symmetric permutation in sequential analysis (0 to 7). 3=Scotch, 4=PORD, 5=Metis
2428: . -mat_mumps_icntl_8 - ICNTL(8): scaling strategy (-2 to 8 or 77)
2429: . -mat_mumps_icntl_10 - ICNTL(10): max num of refinements
2430: . -mat_mumps_icntl_11 - ICNTL(11): statistics related to an error analysis (via -ksp_view)
2431: . -mat_mumps_icntl_12 - ICNTL(12): an ordering strategy for symmetric matrices (0 to 3)
2432: . -mat_mumps_icntl_13 - ICNTL(13): parallelism of the root node (enable ScaLAPACK) and its splitting
2433: . -mat_mumps_icntl_14 - ICNTL(14): percentage increase in the estimated working space
2434: . -mat_mumps_icntl_19 - ICNTL(19): computes the Schur complement
2435: . -mat_mumps_icntl_22 - ICNTL(22): in-core/out-of-core factorization and solve (0 or 1)
2436: . -mat_mumps_icntl_23 - ICNTL(23): max size of the working memory (MB) that can allocate per processor
2437: . -mat_mumps_icntl_24 - ICNTL(24): detection of null pivot rows (0 or 1)
2438: . -mat_mumps_icntl_25 - ICNTL(25): compute a solution of a deficient matrix and a null space basis
2439: . -mat_mumps_icntl_26 - ICNTL(26): drives the solution phase if a Schur complement matrix
2440: . -mat_mumps_icntl_28 - ICNTL(28): use 1 for sequential analysis and ictnl(7) ordering, or 2 for parallel analysis and ictnl(29) ordering
2441: . -mat_mumps_icntl_29 - ICNTL(29): parallel ordering 1 = ptscotch, 2 = parmetis
2442: . -mat_mumps_icntl_30 - ICNTL(30): compute user-specified set of entries in inv(A)
2443: . -mat_mumps_icntl_31 - ICNTL(31): indicates which factors may be discarded during factorization
2444: . -mat_mumps_icntl_33 - ICNTL(33): compute determinant
2445: . -mat_mumps_cntl_1 - CNTL(1): relative pivoting threshold
2446: . -mat_mumps_cntl_2 - CNTL(2): stopping criterion of refinement
2447: . -mat_mumps_cntl_3 - CNTL(3): absolute pivoting threshold
2448: . -mat_mumps_cntl_4 - CNTL(4): value for static pivoting
2449: . -mat_mumps_cntl_5 - CNTL(5): fixation for null pivots
2450: - -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.
2451: Default might be the number of cores per CPU package (socket) as reported by hwloc and suggested by the MUMPS manual.
2453: Level: beginner
2455: Notes:
2456: When a MUMPS factorization fails inside a KSP solve, for example with a KSP_DIVERGED_PCSETUP_FAILED, one can find the MUMPS information about the failure by calling
2457: $ KSPGetPC(ksp,&pc);
2458: $ PCFactorGetMatrix(pc,&mat);
2459: $ MatMumpsGetInfo(mat,....);
2460: $ MatMumpsGetInfog(mat,....); etc.
2461: Or you can run with -ksp_error_if_not_converged and the program will be stopped and the information printed in the error message.
2463: If you want to run MUMPS in MPI+OpenMP hybrid mode (i.e., enable multithreading in MUMPS), but still want to run the non-MUMPS part
2464: (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
2465: (or --with-hwloc), and have an MPI that supports MPI-3.0's process shared memory (which is usually available). Since MUMPS calls BLAS
2466: libraries, to really get performance, you should have multithreaded BLAS libraries such as Intel MKL, AMD ACML, Cray libSci or open sourced
2467: OpenBLAS (PETSc has configure options to install/specify them). With these conditions met, you can run your program as before but with
2468: an extra option -mat_mumps_use_omp_threads [m]. It works as if we set OMP_NUM_THREADS=m to MUMPS, with m defaults to the number of cores
2469: per CPU socket (or package, in hwloc term), or number of PETSc MPI processes on a node, whichever is smaller.
2471: By flat-MPI or pure-MPI mode, it means you run your code with as many MPI ranks as the number of cores. For example,
2472: if a compute node has 32 cores and you run on two nodes, you may use "mpirun -n 64 ./test". To run MPI+OpenMP hybrid MUMPS,
2473: the tranditional way is to set OMP_NUM_THREADS and run with fewer MPI ranks than cores. For example, if you want to have 16 OpenMP
2474: threads per rank, then you may use "export OMP_NUM_THREADS=16 && mpirun -n 4 ./test". The problem of this approach is that the non-MUMPS
2475: part of your code is run with fewer cores and CPUs are wasted. "-mat_mumps_use_omp_threads [m]" provides an alternative such that
2476: you can stil run your code with as many MPI ranks as the number of cores, but have MUMPS run in MPI+OpenMP hybrid mode. In our example,
2477: you can use "mpirun -n 64 ./test -mat_mumps_use_omp_threads 16".
2479: If you run your code through a job submission system, there are caveats in MPI rank mapping. We use MPI_Comm_split_type to get MPI
2480: processes on each compute node. Listing the processes in rank ascending order, we split processes on a node into consecutive groups of
2481: 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
2482: 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
2483: 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.
2484: 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,
2485: 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
2486: 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
2487: cores in socket 1, that definitely hurts locality. On the other hand, if you map MPI ranks consecutively on the two sockets, then the
2488: 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.
2489: For example, with the Slurm job scheduler, one can use srun --cpu-bind=verbsoe -m block:block to map consecutive MPI ranks to sockets and
2490: examine the mapping result.
2492: 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,
2493: 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
2494: calls omp_set_num_threads(m) internally before calling MUMPS.
2496: References:
2497: + 1. - Heroux, Michael A., R. Brightwell, and Michael M. Wolf. "Bi-modal MPI and MPI+ threads computing on scalable multicore systems." IJHPCA (Submitted) (2011).
2498: - 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.
2500: .seealso: PCFactorSetMatSolverType(), MatSolverType, MatMumpsSetICntl(), MatMumpsGetIcntl(), MatMumpsSetCntl(), MatMumpsGetCntl(), MatMumpsGetInfo(), MatMumpsGetInfog(), MatMumpsGetRinfo(), MatMumpsGetRinfog(), KSPGetPC(), PCGetFactor(), PCFactorGetMatrix()
2502: M*/
2504: static PetscErrorCode MatFactorGetSolverType_mumps(Mat A,MatSolverType *type)
2505: {
2507: *type = MATSOLVERMUMPS;
2508: return(0);
2509: }
2511: /* MatGetFactor for Seq and MPI AIJ matrices */
2512: static PetscErrorCode MatGetFactor_aij_mumps(Mat A,MatFactorType ftype,Mat *F)
2513: {
2514: Mat B;
2516: Mat_MUMPS *mumps;
2517: PetscBool isSeqAIJ;
2520: /* Create the factorization matrix */
2521: PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isSeqAIJ);
2522: MatCreate(PetscObjectComm((PetscObject)A),&B);
2523: MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2524: PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2525: MatSetUp(B);
2527: PetscNewLog(B,&mumps);
2529: B->ops->view = MatView_MUMPS;
2530: B->ops->getinfo = MatGetInfo_MUMPS;
2532: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
2533: PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2534: PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2535: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2536: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2537: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2538: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2539: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2540: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2541: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2542: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
2543: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
2544: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);
2546: if (ftype == MAT_FACTOR_LU) {
2547: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
2548: B->factortype = MAT_FACTOR_LU;
2549: if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqaij;
2550: else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpiaij;
2551: mumps->sym = 0;
2552: } else {
2553: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
2554: B->factortype = MAT_FACTOR_CHOLESKY;
2555: if (isSeqAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqaij_seqsbaij;
2556: else mumps->ConvertToTriples = MatConvertToTriples_mpiaij_mpisbaij;
2557: #if defined(PETSC_USE_COMPLEX)
2558: mumps->sym = 2;
2559: #else
2560: if (A->spd_set && A->spd) mumps->sym = 1;
2561: else mumps->sym = 2;
2562: #endif
2563: }
2565: /* set solvertype */
2566: PetscFree(B->solvertype);
2567: PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
2569: B->ops->destroy = MatDestroy_MUMPS;
2570: B->data = (void*)mumps;
2572: PetscInitializeMUMPS(A,mumps);
2574: *F = B;
2575: return(0);
2576: }
2578: /* MatGetFactor for Seq and MPI SBAIJ matrices */
2579: static PetscErrorCode MatGetFactor_sbaij_mumps(Mat A,MatFactorType ftype,Mat *F)
2580: {
2581: Mat B;
2583: Mat_MUMPS *mumps;
2584: PetscBool isSeqSBAIJ;
2587: if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with MUMPS LU, use AIJ matrix");
2588: if (A->rmap->bs > 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Cannot use PETSc SBAIJ matrices with block size > 1 with MUMPS Cholesky, use AIJ matrix instead");
2589: PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isSeqSBAIJ);
2590: /* Create the factorization matrix */
2591: MatCreate(PetscObjectComm((PetscObject)A),&B);
2592: MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2593: PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2594: MatSetUp(B);
2596: PetscNewLog(B,&mumps);
2597: if (isSeqSBAIJ) {
2598: mumps->ConvertToTriples = MatConvertToTriples_seqsbaij_seqsbaij;
2599: } else {
2600: mumps->ConvertToTriples = MatConvertToTriples_mpisbaij_mpisbaij;
2601: }
2603: B->ops->getinfo = MatGetInfo_External;
2604: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_MUMPS;
2605: B->ops->view = MatView_MUMPS;
2607: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
2608: PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2609: PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2610: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2611: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2612: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2613: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2614: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2615: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2616: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2617: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
2618: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
2619: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);
2621: B->factortype = MAT_FACTOR_CHOLESKY;
2622: #if defined(PETSC_USE_COMPLEX)
2623: mumps->sym = 2;
2624: #else
2625: if (A->spd_set && A->spd) mumps->sym = 1;
2626: else mumps->sym = 2;
2627: #endif
2629: /* set solvertype */
2630: PetscFree(B->solvertype);
2631: PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
2633: B->ops->destroy = MatDestroy_MUMPS;
2634: B->data = (void*)mumps;
2636: PetscInitializeMUMPS(A,mumps);
2638: *F = B;
2639: return(0);
2640: }
2642: static PetscErrorCode MatGetFactor_baij_mumps(Mat A,MatFactorType ftype,Mat *F)
2643: {
2644: Mat B;
2646: Mat_MUMPS *mumps;
2647: PetscBool isSeqBAIJ;
2650: /* Create the factorization matrix */
2651: PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isSeqBAIJ);
2652: MatCreate(PetscObjectComm((PetscObject)A),&B);
2653: MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2654: PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2655: MatSetUp(B);
2657: PetscNewLog(B,&mumps);
2658: if (ftype == MAT_FACTOR_LU) {
2659: B->ops->lufactorsymbolic = MatLUFactorSymbolic_BAIJMUMPS;
2660: B->factortype = MAT_FACTOR_LU;
2661: if (isSeqBAIJ) mumps->ConvertToTriples = MatConvertToTriples_seqbaij_seqaij;
2662: else mumps->ConvertToTriples = MatConvertToTriples_mpibaij_mpiaij;
2663: mumps->sym = 0;
2664: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot use PETSc BAIJ matrices with MUMPS Cholesky, use SBAIJ or AIJ matrix instead\n");
2666: B->ops->getinfo = MatGetInfo_External;
2667: B->ops->view = MatView_MUMPS;
2669: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
2670: PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2671: PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2672: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2673: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2674: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2675: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2676: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2677: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2678: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2679: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
2680: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverse_C",MatMumpsGetInverse_MUMPS);
2681: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInverseTranspose_C",MatMumpsGetInverseTranspose_MUMPS);
2683: /* set solvertype */
2684: PetscFree(B->solvertype);
2685: PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
2687: B->ops->destroy = MatDestroy_MUMPS;
2688: B->data = (void*)mumps;
2690: PetscInitializeMUMPS(A,mumps);
2692: *F = B;
2693: return(0);
2694: }
2696: /* MatGetFactor for Seq and MPI SELL matrices */
2697: static PetscErrorCode MatGetFactor_sell_mumps(Mat A,MatFactorType ftype,Mat *F)
2698: {
2699: Mat B;
2701: Mat_MUMPS *mumps;
2702: PetscBool isSeqSELL;
2705: /* Create the factorization matrix */
2706: PetscObjectTypeCompare((PetscObject)A,MATSEQSELL,&isSeqSELL);
2707: MatCreate(PetscObjectComm((PetscObject)A),&B);
2708: MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2709: PetscStrallocpy("mumps",&((PetscObject)B)->type_name);
2710: MatSetUp(B);
2712: PetscNewLog(B,&mumps);
2714: B->ops->view = MatView_MUMPS;
2715: B->ops->getinfo = MatGetInfo_MUMPS;
2717: PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverType_C",MatFactorGetSolverType_mumps);
2718: PetscObjectComposeFunction((PetscObject)B,"MatFactorSetSchurIS_C",MatFactorSetSchurIS_MUMPS);
2719: PetscObjectComposeFunction((PetscObject)B,"MatFactorCreateSchurComplement_C",MatFactorCreateSchurComplement_MUMPS);
2720: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetIcntl_C",MatMumpsSetIcntl_MUMPS);
2721: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetIcntl_C",MatMumpsGetIcntl_MUMPS);
2722: PetscObjectComposeFunction((PetscObject)B,"MatMumpsSetCntl_C",MatMumpsSetCntl_MUMPS);
2723: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetCntl_C",MatMumpsGetCntl_MUMPS);
2724: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfo_C",MatMumpsGetInfo_MUMPS);
2725: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetInfog_C",MatMumpsGetInfog_MUMPS);
2726: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfo_C",MatMumpsGetRinfo_MUMPS);
2727: PetscObjectComposeFunction((PetscObject)B,"MatMumpsGetRinfog_C",MatMumpsGetRinfog_MUMPS);
2729: if (ftype == MAT_FACTOR_LU) {
2730: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMUMPS;
2731: B->factortype = MAT_FACTOR_LU;
2732: if (isSeqSELL) mumps->ConvertToTriples = MatConvertToTriples_seqsell_seqaij;
2733: else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented");
2734: mumps->sym = 0;
2735: } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"To be implemented");
2737: /* set solvertype */
2738: PetscFree(B->solvertype);
2739: PetscStrallocpy(MATSOLVERMUMPS,&B->solvertype);
2741: B->ops->destroy = MatDestroy_MUMPS;
2742: B->data = (void*)mumps;
2744: PetscInitializeMUMPS(A,mumps);
2746: *F = B;
2747: return(0);
2748: }
2750: PETSC_EXTERN PetscErrorCode MatSolverTypeRegister_MUMPS(void)
2751: {
2755: MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
2756: MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
2757: MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
2758: MatSolverTypeRegister(MATSOLVERMUMPS,MATMPIBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
2759: MatSolverTypeRegister(MATSOLVERMUMPS,MATMPISBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
2760: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_LU,MatGetFactor_aij_mumps);
2761: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_aij_mumps);
2762: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_LU,MatGetFactor_baij_mumps);
2763: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_baij_mumps);
2764: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSBAIJ,MAT_FACTOR_CHOLESKY,MatGetFactor_sbaij_mumps);
2765: MatSolverTypeRegister(MATSOLVERMUMPS,MATSEQSELL,MAT_FACTOR_LU,MatGetFactor_sell_mumps);
2766: return(0);
2767: }