Actual source code: ispai.c
2: /*
3: 3/99 Modified by Stephen Barnard to support SPAI version 3.0
4: */
6: /*
7: Provides an interface to the SPAI Sparse Approximate Inverse Preconditioner
8: Code written by Stephen Barnard.
10: Note: there is some BAD memory bleeding below!
12: This code needs work
14: 1) get rid of all memory bleeding
15: 2) fix PETSc/interface so that it gets if the matrix is symmetric from the matrix
16: rather than having the sp flag for PC_SPAI
17: 3) fix to set the block size based on the matrix block size
19: */
20: #define PETSC_SKIP_COMPLEX /* since spai uses I which conflicts with some complex implementations */
22: #include <petsc/private/pcimpl.h>
23: #include <../src/ksp/pc/impls/spai/petscspai.h>
25: /*
26: These are the SPAI include files
27: */
28: EXTERN_C_BEGIN
29: #define SPAI_USE_MPI /* required for setting SPAI_Comm correctly in basics.h */
30: #include <spai.h>
31: #include <matrix.h>
32: EXTERN_C_END
34: extern PetscErrorCode ConvertMatToMatrix(MPI_Comm,Mat,Mat,matrix**);
35: extern PetscErrorCode ConvertMatrixToMat(MPI_Comm,matrix*,Mat*);
36: extern PetscErrorCode ConvertVectorToVec(MPI_Comm,vector *v,Vec *Pv);
37: extern PetscErrorCode MM_to_PETSC(char*,char*,char*);
39: typedef struct {
41: matrix *B; /* matrix in SPAI format */
42: matrix *BT; /* transpose of matrix in SPAI format */
43: matrix *M; /* the approximate inverse in SPAI format */
45: Mat PM; /* the approximate inverse PETSc format */
47: double epsilon; /* tolerance */
48: int nbsteps; /* max number of "improvement" steps per line */
49: int max; /* max dimensions of is_I, q, etc. */
50: int maxnew; /* max number of new entries per step */
51: int block_size; /* constant block size */
52: int cache_size; /* one of (1,2,3,4,5,6) indicting size of cache */
53: int verbose; /* SPAI prints timing and statistics */
55: int sp; /* symmetric nonzero pattern */
56: MPI_Comm comm_spai; /* communicator to be used with spai */
57: } PC_SPAI;
59: /**********************************************************************/
61: static PetscErrorCode PCSetUp_SPAI(PC pc)
62: {
63: PC_SPAI *ispai = (PC_SPAI*)pc->data;
65: Mat AT;
68: init_SPAI();
70: if (ispai->sp) {
71: ConvertMatToMatrix(ispai->comm_spai,pc->pmat,pc->pmat,&ispai->B);
72: } else {
73: /* Use the transpose to get the column nonzero structure. */
74: MatTranspose(pc->pmat,MAT_INITIAL_MATRIX,&AT);
75: ConvertMatToMatrix(ispai->comm_spai,pc->pmat,AT,&ispai->B);
76: MatDestroy(&AT);
77: }
79: /* Destroy the transpose */
80: /* Don't know how to do it. PETSc developers? */
82: /* construct SPAI preconditioner */
83: /* FILE *messages */ /* file for warning messages */
84: /* double epsilon */ /* tolerance */
85: /* int nbsteps */ /* max number of "improvement" steps per line */
86: /* int max */ /* max dimensions of is_I, q, etc. */
87: /* int maxnew */ /* max number of new entries per step */
88: /* int block_size */ /* block_size == 1 specifies scalar elments
89: block_size == n specifies nxn constant-block elements
90: block_size == 0 specifies variable-block elements */
91: /* int cache_size */ /* one of (1,2,3,4,5,6) indicting size of cache. cache_size == 0 indicates no caching */
92: /* int verbose */ /* verbose == 0 specifies that SPAI is silent
93: verbose == 1 prints timing and matrix statistics */
95: bspai(ispai->B,&ispai->M,
96: stdout,
97: ispai->epsilon,
98: ispai->nbsteps,
99: ispai->max,
100: ispai->maxnew,
101: ispai->block_size,
102: ispai->cache_size,
103: ispai->verbose);
105: ConvertMatrixToMat(PetscObjectComm((PetscObject)pc),ispai->M,&ispai->PM);
107: /* free the SPAI matrices */
108: sp_free_matrix(ispai->B);
109: sp_free_matrix(ispai->M);
110: return(0);
111: }
113: /**********************************************************************/
115: static PetscErrorCode PCApply_SPAI(PC pc,Vec xx,Vec y)
116: {
117: PC_SPAI *ispai = (PC_SPAI*)pc->data;
121: /* Now using PETSc's multiply */
122: MatMult(ispai->PM,xx,y);
123: return(0);
124: }
126: static PetscErrorCode PCMatApply_SPAI(PC pc,Mat X,Mat Y)
127: {
128: PC_SPAI *ispai = (PC_SPAI*)pc->data;
132: /* Now using PETSc's multiply */
133: MatMatMult(ispai->PM,X,MAT_REUSE_MATRIX,PETSC_DEFAULT,&Y);
134: return(0);
135: }
137: /**********************************************************************/
139: static PetscErrorCode PCDestroy_SPAI(PC pc)
140: {
142: PC_SPAI *ispai = (PC_SPAI*)pc->data;
145: MatDestroy(&ispai->PM);
146: MPI_Comm_free(&(ispai->comm_spai));
147: PetscFree(pc->data);
148: return(0);
149: }
151: /**********************************************************************/
153: static PetscErrorCode PCView_SPAI(PC pc,PetscViewer viewer)
154: {
155: PC_SPAI *ispai = (PC_SPAI*)pc->data;
157: PetscBool iascii;
160: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
161: if (iascii) {
162: PetscViewerASCIIPrintf(viewer," epsilon %g\n", (double)ispai->epsilon);
163: PetscViewerASCIIPrintf(viewer," nbsteps %d\n", ispai->nbsteps);
164: PetscViewerASCIIPrintf(viewer," max %d\n", ispai->max);
165: PetscViewerASCIIPrintf(viewer," maxnew %d\n", ispai->maxnew);
166: PetscViewerASCIIPrintf(viewer," block_size %d\n",ispai->block_size);
167: PetscViewerASCIIPrintf(viewer," cache_size %d\n",ispai->cache_size);
168: PetscViewerASCIIPrintf(viewer," verbose %d\n", ispai->verbose);
169: PetscViewerASCIIPrintf(viewer," sp %d\n", ispai->sp);
170: }
171: return(0);
172: }
174: static PetscErrorCode PCSPAISetEpsilon_SPAI(PC pc,double epsilon1)
175: {
176: PC_SPAI *ispai = (PC_SPAI*)pc->data;
179: ispai->epsilon = epsilon1;
180: return(0);
181: }
183: /**********************************************************************/
185: static PetscErrorCode PCSPAISetNBSteps_SPAI(PC pc,int nbsteps1)
186: {
187: PC_SPAI *ispai = (PC_SPAI*)pc->data;
190: ispai->nbsteps = nbsteps1;
191: return(0);
192: }
194: /**********************************************************************/
196: /* added 1/7/99 g.h. */
197: static PetscErrorCode PCSPAISetMax_SPAI(PC pc,int max1)
198: {
199: PC_SPAI *ispai = (PC_SPAI*)pc->data;
202: ispai->max = max1;
203: return(0);
204: }
206: /**********************************************************************/
208: static PetscErrorCode PCSPAISetMaxNew_SPAI(PC pc,int maxnew1)
209: {
210: PC_SPAI *ispai = (PC_SPAI*)pc->data;
213: ispai->maxnew = maxnew1;
214: return(0);
215: }
217: /**********************************************************************/
219: static PetscErrorCode PCSPAISetBlockSize_SPAI(PC pc,int block_size1)
220: {
221: PC_SPAI *ispai = (PC_SPAI*)pc->data;
224: ispai->block_size = block_size1;
225: return(0);
226: }
228: /**********************************************************************/
230: static PetscErrorCode PCSPAISetCacheSize_SPAI(PC pc,int cache_size)
231: {
232: PC_SPAI *ispai = (PC_SPAI*)pc->data;
235: ispai->cache_size = cache_size;
236: return(0);
237: }
239: /**********************************************************************/
241: static PetscErrorCode PCSPAISetVerbose_SPAI(PC pc,int verbose)
242: {
243: PC_SPAI *ispai = (PC_SPAI*)pc->data;
246: ispai->verbose = verbose;
247: return(0);
248: }
250: /**********************************************************************/
252: static PetscErrorCode PCSPAISetSp_SPAI(PC pc,int sp)
253: {
254: PC_SPAI *ispai = (PC_SPAI*)pc->data;
257: ispai->sp = sp;
258: return(0);
259: }
261: /* -------------------------------------------------------------------*/
263: /*@
264: PCSPAISetEpsilon -- Set the tolerance for the SPAI preconditioner
266: Input Parameters:
267: + pc - the preconditioner
268: - eps - epsilon (default .4)
270: Notes:
271: Espilon must be between 0 and 1. It controls the
272: quality of the approximation of M to the inverse of
273: A. Higher values of epsilon lead to more work, more
274: fill, and usually better preconditioners. In many
275: cases the best choice of epsilon is the one that
276: divides the total solution time equally between the
277: preconditioner and the solver.
279: Level: intermediate
281: .seealso: PCSPAI, PCSetType()
282: @*/
283: PetscErrorCode PCSPAISetEpsilon(PC pc,double epsilon1)
284: {
288: PetscTryMethod(pc,"PCSPAISetEpsilon_C",(PC,double),(pc,epsilon1));
289: return(0);
290: }
292: /**********************************************************************/
294: /*@
295: PCSPAISetNBSteps - set maximum number of improvement steps per row in
296: the SPAI preconditioner
298: Input Parameters:
299: + pc - the preconditioner
300: - n - number of steps (default 5)
302: Notes:
303: SPAI constructs to approximation to every column of
304: the exact inverse of A in a series of improvement
305: steps. The quality of the approximation is determined
306: by epsilon. If an approximation achieving an accuracy
307: of epsilon is not obtained after ns steps, SPAI simply
308: uses the best approximation constructed so far.
310: Level: intermediate
312: .seealso: PCSPAI, PCSetType(), PCSPAISetMaxNew()
313: @*/
314: PetscErrorCode PCSPAISetNBSteps(PC pc,int nbsteps1)
315: {
319: PetscTryMethod(pc,"PCSPAISetNBSteps_C",(PC,int),(pc,nbsteps1));
320: return(0);
321: }
323: /**********************************************************************/
325: /* added 1/7/99 g.h. */
326: /*@
327: PCSPAISetMax - set the size of various working buffers in
328: the SPAI preconditioner
330: Input Parameters:
331: + pc - the preconditioner
332: - n - size (default is 5000)
334: Level: intermediate
336: .seealso: PCSPAI, PCSetType()
337: @*/
338: PetscErrorCode PCSPAISetMax(PC pc,int max1)
339: {
343: PetscTryMethod(pc,"PCSPAISetMax_C",(PC,int),(pc,max1));
344: return(0);
345: }
347: /**********************************************************************/
349: /*@
350: PCSPAISetMaxNew - set maximum number of new nonzero candidates per step
351: in SPAI preconditioner
353: Input Parameters:
354: + pc - the preconditioner
355: - n - maximum number (default 5)
357: Level: intermediate
359: .seealso: PCSPAI, PCSetType(), PCSPAISetNBSteps()
360: @*/
361: PetscErrorCode PCSPAISetMaxNew(PC pc,int maxnew1)
362: {
366: PetscTryMethod(pc,"PCSPAISetMaxNew_C",(PC,int),(pc,maxnew1));
367: return(0);
368: }
370: /**********************************************************************/
372: /*@
373: PCSPAISetBlockSize - set the block size for the SPAI preconditioner
375: Input Parameters:
376: + pc - the preconditioner
377: - n - block size (default 1)
379: Notes:
380: A block
381: size of 1 treats A as a matrix of scalar elements. A
382: block size of s > 1 treats A as a matrix of sxs
383: blocks. A block size of 0 treats A as a matrix with
384: variable sized blocks, which are determined by
385: searching for dense square diagonal blocks in A.
386: This can be very effective for finite-element
387: matrices.
389: SPAI will convert A to block form, use a block
390: version of the preconditioner algorithm, and then
391: convert the result back to scalar form.
393: In many cases the a block-size parameter other than 1
394: can lead to very significant improvement in
395: performance.
397: Level: intermediate
399: .seealso: PCSPAI, PCSetType()
400: @*/
401: PetscErrorCode PCSPAISetBlockSize(PC pc,int block_size1)
402: {
406: PetscTryMethod(pc,"PCSPAISetBlockSize_C",(PC,int),(pc,block_size1));
407: return(0);
408: }
410: /**********************************************************************/
412: /*@
413: PCSPAISetCacheSize - specify cache size in the SPAI preconditioner
415: Input Parameters:
416: + pc - the preconditioner
417: - n - cache size {0,1,2,3,4,5} (default 5)
419: Notes:
420: SPAI uses a hash table to cache messages and avoid
421: redundant communication. If suggest always using
422: 5. This parameter is irrelevant in the serial
423: version.
425: Level: intermediate
427: .seealso: PCSPAI, PCSetType()
428: @*/
429: PetscErrorCode PCSPAISetCacheSize(PC pc,int cache_size)
430: {
434: PetscTryMethod(pc,"PCSPAISetCacheSize_C",(PC,int),(pc,cache_size));
435: return(0);
436: }
438: /**********************************************************************/
440: /*@
441: PCSPAISetVerbose - verbosity level for the SPAI preconditioner
443: Input Parameters:
444: + pc - the preconditioner
445: - n - level (default 1)
447: Notes:
448: print parameters, timings and matrix statistics
450: Level: intermediate
452: .seealso: PCSPAI, PCSetType()
453: @*/
454: PetscErrorCode PCSPAISetVerbose(PC pc,int verbose)
455: {
459: PetscTryMethod(pc,"PCSPAISetVerbose_C",(PC,int),(pc,verbose));
460: return(0);
461: }
463: /**********************************************************************/
465: /*@
466: PCSPAISetSp - specify a symmetric matrix sparsity pattern in the SPAI preconditioner
468: Input Parameters:
469: + pc - the preconditioner
470: - n - 0 or 1
472: Notes:
473: If A has a symmetric nonzero pattern use -sp 1 to
474: improve performance by eliminating some communication
475: in the parallel version. Even if A does not have a
476: symmetric nonzero pattern -sp 1 may well lead to good
477: results, but the code will not follow the published
478: SPAI algorithm exactly.
480: Level: intermediate
482: .seealso: PCSPAI, PCSetType()
483: @*/
484: PetscErrorCode PCSPAISetSp(PC pc,int sp)
485: {
489: PetscTryMethod(pc,"PCSPAISetSp_C",(PC,int),(pc,sp));
490: return(0);
491: }
493: /**********************************************************************/
495: /**********************************************************************/
497: static PetscErrorCode PCSetFromOptions_SPAI(PetscOptionItems *PetscOptionsObject,PC pc)
498: {
499: PC_SPAI *ispai = (PC_SPAI*)pc->data;
501: int nbsteps1,max1,maxnew1,block_size1,cache_size,verbose,sp;
502: double epsilon1;
503: PetscBool flg;
506: PetscOptionsHead(PetscOptionsObject,"SPAI options");
507: PetscOptionsReal("-pc_spai_epsilon","","PCSPAISetEpsilon",ispai->epsilon,&epsilon1,&flg);
508: if (flg) {
509: PCSPAISetEpsilon(pc,epsilon1);
510: }
511: PetscOptionsInt("-pc_spai_nbsteps","","PCSPAISetNBSteps",ispai->nbsteps,&nbsteps1,&flg);
512: if (flg) {
513: PCSPAISetNBSteps(pc,nbsteps1);
514: }
515: /* added 1/7/99 g.h. */
516: PetscOptionsInt("-pc_spai_max","","PCSPAISetMax",ispai->max,&max1,&flg);
517: if (flg) {
518: PCSPAISetMax(pc,max1);
519: }
520: PetscOptionsInt("-pc_spai_maxnew","","PCSPAISetMaxNew",ispai->maxnew,&maxnew1,&flg);
521: if (flg) {
522: PCSPAISetMaxNew(pc,maxnew1);
523: }
524: PetscOptionsInt("-pc_spai_block_size","","PCSPAISetBlockSize",ispai->block_size,&block_size1,&flg);
525: if (flg) {
526: PCSPAISetBlockSize(pc,block_size1);
527: }
528: PetscOptionsInt("-pc_spai_cache_size","","PCSPAISetCacheSize",ispai->cache_size,&cache_size,&flg);
529: if (flg) {
530: PCSPAISetCacheSize(pc,cache_size);
531: }
532: PetscOptionsInt("-pc_spai_verbose","","PCSPAISetVerbose",ispai->verbose,&verbose,&flg);
533: if (flg) {
534: PCSPAISetVerbose(pc,verbose);
535: }
536: PetscOptionsInt("-pc_spai_sp","","PCSPAISetSp",ispai->sp,&sp,&flg);
537: if (flg) {
538: PCSPAISetSp(pc,sp);
539: }
540: PetscOptionsTail();
541: return(0);
542: }
544: /**********************************************************************/
546: /*MC
547: PCSPAI - Use the Sparse Approximate Inverse method of Grote and Barnard
548: as a preconditioner (SIAM J. Sci. Comput.; vol 18, nr 3)
550: Options Database Keys:
551: + -pc_spai_epsilon <eps> - set tolerance
552: . -pc_spai_nbstep <n> - set nbsteps
553: . -pc_spai_max <m> - set max
554: . -pc_spai_max_new <m> - set maxnew
555: . -pc_spai_block_size <n> - set block size
556: . -pc_spai_cache_size <n> - set cache size
557: . -pc_spai_sp <m> - set sp
558: - -pc_spai_set_verbose <true,false> - verbose output
560: Notes:
561: This only works with AIJ matrices.
563: Level: beginner
565: .seealso: PCCreate(), PCSetType(), PCType (for list of available types), PC,
566: PCSPAISetEpsilon(), PCSPAISetMax(), PCSPAISetMaxNew(), PCSPAISetBlockSize(),
567: PCSPAISetVerbose(), PCSPAISetSp()
568: M*/
570: PETSC_EXTERN PetscErrorCode PCCreate_SPAI(PC pc)
571: {
572: PC_SPAI *ispai;
576: PetscNewLog(pc,&ispai);
577: pc->data = ispai;
579: pc->ops->destroy = PCDestroy_SPAI;
580: pc->ops->apply = PCApply_SPAI;
581: pc->ops->matapply = PCMatApply_SPAI;
582: pc->ops->applyrichardson = 0;
583: pc->ops->setup = PCSetUp_SPAI;
584: pc->ops->view = PCView_SPAI;
585: pc->ops->setfromoptions = PCSetFromOptions_SPAI;
587: ispai->epsilon = .4;
588: ispai->nbsteps = 5;
589: ispai->max = 5000;
590: ispai->maxnew = 5;
591: ispai->block_size = 1;
592: ispai->cache_size = 5;
593: ispai->verbose = 0;
595: ispai->sp = 1;
596: MPI_Comm_dup(PetscObjectComm((PetscObject)pc),&(ispai->comm_spai));
598: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetEpsilon_C",PCSPAISetEpsilon_SPAI);
599: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetNBSteps_C",PCSPAISetNBSteps_SPAI);
600: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetMax_C",PCSPAISetMax_SPAI);
601: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetMaxNew_C",PCSPAISetMaxNew_SPAI);
602: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetBlockSize_C",PCSPAISetBlockSize_SPAI);
603: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetCacheSize_C",PCSPAISetCacheSize_SPAI);
604: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetVerbose_C",PCSPAISetVerbose_SPAI);
605: PetscObjectComposeFunction((PetscObject)pc,"PCSPAISetSp_C",PCSPAISetSp_SPAI);
606: return(0);
607: }
609: /**********************************************************************/
611: /*
612: Converts from a PETSc matrix to an SPAI matrix
613: */
614: PetscErrorCode ConvertMatToMatrix(MPI_Comm comm, Mat A,Mat AT,matrix **B)
615: {
616: matrix *M;
617: int i,j,col;
618: int row_indx;
619: int len,pe,local_indx,start_indx;
620: int *mapping;
621: PetscErrorCode ierr;
622: const int *cols;
623: const double *vals;
624: int n,mnl,nnl,nz,rstart,rend;
625: PetscMPIInt size,rank;
626: struct compressed_lines *rows;
629: MPI_Comm_size(comm,&size);
630: MPI_Comm_rank(comm,&rank);
631: MatGetSize(A,&n,&n);
632: MatGetLocalSize(A,&mnl,&nnl);
634: /*
635: not sure why a barrier is required. commenting out
636: MPI_Barrier(comm);
637: */
639: M = new_matrix((SPAI_Comm)comm);
641: M->n = n;
642: M->bs = 1;
643: M->max_block_size = 1;
645: M->mnls = (int*)malloc(sizeof(int)*size);
646: M->start_indices = (int*)malloc(sizeof(int)*size);
647: M->pe = (int*)malloc(sizeof(int)*n);
648: M->block_sizes = (int*)malloc(sizeof(int)*n);
649: for (i=0; i<n; i++) M->block_sizes[i] = 1;
651: MPI_Allgather(&mnl,1,MPI_INT,M->mnls,1,MPI_INT,comm);
653: M->start_indices[0] = 0;
654: for (i=1; i<size; i++) M->start_indices[i] = M->start_indices[i-1] + M->mnls[i-1];
656: M->mnl = M->mnls[M->myid];
657: M->my_start_index = M->start_indices[M->myid];
659: for (i=0; i<size; i++) {
660: start_indx = M->start_indices[i];
661: for (j=0; j<M->mnls[i]; j++) M->pe[start_indx+j] = i;
662: }
664: if (AT) {
665: M->lines = new_compressed_lines(M->mnls[rank],1);
666: } else {
667: M->lines = new_compressed_lines(M->mnls[rank],0);
668: }
670: rows = M->lines;
672: /* Determine the mapping from global indices to pointers */
673: PetscMalloc1(M->n,&mapping);
674: pe = 0;
675: local_indx = 0;
676: for (i=0; i<M->n; i++) {
677: if (local_indx >= M->mnls[pe]) {
678: pe++;
679: local_indx = 0;
680: }
681: mapping[i] = local_indx + M->start_indices[pe];
682: local_indx++;
683: }
685: /*********************************************************/
686: /************** Set up the row structure *****************/
687: /*********************************************************/
689: MatGetOwnershipRange(A,&rstart,&rend);
690: for (i=rstart; i<rend; i++) {
691: row_indx = i - rstart;
692: MatGetRow(A,i,&nz,&cols,&vals);
693: /* allocate buffers */
694: rows->ptrs[row_indx] = (int*)malloc(nz*sizeof(int));
695: rows->A[row_indx] = (double*)malloc(nz*sizeof(double));
696: /* copy the matrix */
697: for (j=0; j<nz; j++) {
698: col = cols[j];
699: len = rows->len[row_indx]++;
701: rows->ptrs[row_indx][len] = mapping[col];
702: rows->A[row_indx][len] = vals[j];
703: }
704: rows->slen[row_indx] = rows->len[row_indx];
706: MatRestoreRow(A,i,&nz,&cols,&vals);
707: }
709: /************************************************************/
710: /************** Set up the column structure *****************/
711: /*********************************************************/
713: if (AT) {
715: for (i=rstart; i<rend; i++) {
716: row_indx = i - rstart;
717: MatGetRow(AT,i,&nz,&cols,&vals);
718: /* allocate buffers */
719: rows->rptrs[row_indx] = (int*)malloc(nz*sizeof(int));
720: /* copy the matrix (i.e., the structure) */
721: for (j=0; j<nz; j++) {
722: col = cols[j];
723: len = rows->rlen[row_indx]++;
725: rows->rptrs[row_indx][len] = mapping[col];
726: }
727: MatRestoreRow(AT,i,&nz,&cols,&vals);
728: }
729: }
731: PetscFree(mapping);
733: order_pointers(M);
734: M->maxnz = calc_maxnz(M);
735: *B = M;
736: return(0);
737: }
739: /**********************************************************************/
741: /*
742: Converts from an SPAI matrix B to a PETSc matrix PB.
743: This assumes that the SPAI matrix B is stored in
744: COMPRESSED-ROW format.
745: */
746: PetscErrorCode ConvertMatrixToMat(MPI_Comm comm,matrix *B,Mat *PB)
747: {
748: PetscMPIInt size,rank;
750: int m,n,M,N;
751: int d_nz,o_nz;
752: int *d_nnz,*o_nnz;
753: int i,k,global_row,global_col,first_diag_col,last_diag_col;
754: PetscScalar val;
757: MPI_Comm_size(comm,&size);
758: MPI_Comm_rank(comm,&rank);
760: m = n = B->mnls[rank];
761: d_nz = o_nz = 0;
763: /* Determine preallocation for MatCreateAIJ */
764: PetscMalloc1(m,&d_nnz);
765: PetscMalloc1(m,&o_nnz);
766: for (i=0; i<m; i++) d_nnz[i] = o_nnz[i] = 0;
767: first_diag_col = B->start_indices[rank];
768: last_diag_col = first_diag_col + B->mnls[rank];
769: for (i=0; i<B->mnls[rank]; i++) {
770: for (k=0; k<B->lines->len[i]; k++) {
771: global_col = B->lines->ptrs[i][k];
772: if ((global_col >= first_diag_col) && (global_col < last_diag_col)) d_nnz[i]++;
773: else o_nnz[i]++;
774: }
775: }
777: M = N = B->n;
778: /* Here we only know how to create AIJ format */
779: MatCreate(comm,PB);
780: MatSetSizes(*PB,m,n,M,N);
781: MatSetType(*PB,MATAIJ);
782: MatSeqAIJSetPreallocation(*PB,d_nz,d_nnz);
783: MatMPIAIJSetPreallocation(*PB,d_nz,d_nnz,o_nz,o_nnz);
785: for (i=0; i<B->mnls[rank]; i++) {
786: global_row = B->start_indices[rank]+i;
787: for (k=0; k<B->lines->len[i]; k++) {
788: global_col = B->lines->ptrs[i][k];
790: val = B->lines->A[i][k];
791: MatSetValues(*PB,1,&global_row,1,&global_col,&val,ADD_VALUES);
792: }
793: }
795: PetscFree(d_nnz);
796: PetscFree(o_nnz);
798: MatAssemblyBegin(*PB,MAT_FINAL_ASSEMBLY);
799: MatAssemblyEnd(*PB,MAT_FINAL_ASSEMBLY);
800: return(0);
801: }
803: /**********************************************************************/
805: /*
806: Converts from an SPAI vector v to a PETSc vec Pv.
807: */
808: PetscErrorCode ConvertVectorToVec(MPI_Comm comm,vector *v,Vec *Pv)
809: {
811: PetscMPIInt size,rank;
812: int m,M,i,*mnls,*start_indices,*global_indices;
815: MPI_Comm_size(comm,&size);
816: MPI_Comm_rank(comm,&rank);
818: m = v->mnl;
819: M = v->n;
821: VecCreateMPI(comm,m,M,Pv);
823: PetscMalloc1(size,&mnls);
824: MPI_Allgather(&v->mnl,1,MPI_INT,mnls,1,MPI_INT,comm);
826: PetscMalloc1(size,&start_indices);
828: start_indices[0] = 0;
829: for (i=1; i<size; i++) start_indices[i] = start_indices[i-1] +mnls[i-1];
831: PetscMalloc1(v->mnl,&global_indices);
832: for (i=0; i<v->mnl; i++) global_indices[i] = start_indices[rank] + i;
834: PetscFree(mnls);
835: PetscFree(start_indices);
837: VecSetValues(*Pv,v->mnl,global_indices,v->v,INSERT_VALUES);
838: VecAssemblyBegin(*Pv);
839: VecAssemblyEnd(*Pv);
841: PetscFree(global_indices);
842: return(0);
843: }