Actual source code: mpiaij.c
petsc-3.14.0 2020-09-29
1: #include <../src/mat/impls/aij/mpi/mpiaij.h>
2: #include <petsc/private/vecimpl.h>
3: #include <petsc/private/vecscatterimpl.h>
4: #include <petsc/private/isimpl.h>
5: #include <petscblaslapack.h>
6: #include <petscsf.h>
7: #include <petsc/private/hashmapi.h>
9: /*MC
10: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
12: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
13: and MATMPIAIJ otherwise. As a result, for single process communicators,
14: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
15: for communicators controlling multiple processes. It is recommended that you call both of
16: the above preallocation routines for simplicity.
18: Options Database Keys:
19: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
21: Developer Notes:
22: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
23: enough exist.
25: Level: beginner
27: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ
28: M*/
30: /*MC
31: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
33: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
34: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
35: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
36: for communicators controlling multiple processes. It is recommended that you call both of
37: the above preallocation routines for simplicity.
39: Options Database Keys:
40: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
42: Level: beginner
44: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
45: M*/
47: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A,PetscBool flg)
48: {
49: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
53: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL)
54: A->boundtocpu = flg;
55: #endif
56: if (a->A) {
57: MatBindToCPU(a->A,flg);
58: }
59: if (a->B) {
60: MatBindToCPU(a->B,flg);
61: }
62: return(0);
63: }
66: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
67: {
69: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
72: if (mat->A) {
73: MatSetBlockSizes(mat->A,rbs,cbs);
74: MatSetBlockSizes(mat->B,rbs,1);
75: }
76: return(0);
77: }
79: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
80: {
81: PetscErrorCode ierr;
82: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
83: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data;
84: Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data;
85: const PetscInt *ia,*ib;
86: const MatScalar *aa,*bb;
87: PetscInt na,nb,i,j,*rows,cnt=0,n0rows;
88: PetscInt m = M->rmap->n,rstart = M->rmap->rstart;
91: *keptrows = NULL;
92: ia = a->i;
93: ib = b->i;
94: for (i=0; i<m; i++) {
95: na = ia[i+1] - ia[i];
96: nb = ib[i+1] - ib[i];
97: if (!na && !nb) {
98: cnt++;
99: goto ok1;
100: }
101: aa = a->a + ia[i];
102: for (j=0; j<na; j++) {
103: if (aa[j] != 0.0) goto ok1;
104: }
105: bb = b->a + ib[i];
106: for (j=0; j <nb; j++) {
107: if (bb[j] != 0.0) goto ok1;
108: }
109: cnt++;
110: ok1:;
111: }
112: MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
113: if (!n0rows) return(0);
114: PetscMalloc1(M->rmap->n-cnt,&rows);
115: cnt = 0;
116: for (i=0; i<m; i++) {
117: na = ia[i+1] - ia[i];
118: nb = ib[i+1] - ib[i];
119: if (!na && !nb) continue;
120: aa = a->a + ia[i];
121: for (j=0; j<na;j++) {
122: if (aa[j] != 0.0) {
123: rows[cnt++] = rstart + i;
124: goto ok2;
125: }
126: }
127: bb = b->a + ib[i];
128: for (j=0; j<nb; j++) {
129: if (bb[j] != 0.0) {
130: rows[cnt++] = rstart + i;
131: goto ok2;
132: }
133: }
134: ok2:;
135: }
136: ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
137: return(0);
138: }
140: PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
141: {
142: PetscErrorCode ierr;
143: Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data;
144: PetscBool cong;
147: MatHasCongruentLayouts(Y,&cong);
148: if (Y->assembled && cong) {
149: MatDiagonalSet(aij->A,D,is);
150: } else {
151: MatDiagonalSet_Default(Y,D,is);
152: }
153: return(0);
154: }
156: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
157: {
158: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data;
160: PetscInt i,rstart,nrows,*rows;
163: *zrows = NULL;
164: MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
165: MatGetOwnershipRange(M,&rstart,NULL);
166: for (i=0; i<nrows; i++) rows[i] += rstart;
167: ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
168: return(0);
169: }
171: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
172: {
174: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
175: PetscInt i,n,*garray = aij->garray;
176: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data;
177: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data;
178: PetscReal *work;
181: MatGetSize(A,NULL,&n);
182: PetscCalloc1(n,&work);
183: if (type == NORM_2) {
184: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
185: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
186: }
187: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
188: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
189: }
190: } else if (type == NORM_1) {
191: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
192: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
193: }
194: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
195: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
196: }
197: } else if (type == NORM_INFINITY) {
198: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
199: work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
200: }
201: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
202: work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
203: }
205: } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
206: if (type == NORM_INFINITY) {
207: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
208: } else {
209: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
210: }
211: PetscFree(work);
212: if (type == NORM_2) {
213: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
214: }
215: return(0);
216: }
218: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
219: {
220: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
221: IS sis,gis;
222: PetscErrorCode ierr;
223: const PetscInt *isis,*igis;
224: PetscInt n,*iis,nsis,ngis,rstart,i;
227: MatFindOffBlockDiagonalEntries(a->A,&sis);
228: MatFindNonzeroRows(a->B,&gis);
229: ISGetSize(gis,&ngis);
230: ISGetSize(sis,&nsis);
231: ISGetIndices(sis,&isis);
232: ISGetIndices(gis,&igis);
234: PetscMalloc1(ngis+nsis,&iis);
235: PetscArraycpy(iis,igis,ngis);
236: PetscArraycpy(iis+ngis,isis,nsis);
237: n = ngis + nsis;
238: PetscSortRemoveDupsInt(&n,iis);
239: MatGetOwnershipRange(A,&rstart,NULL);
240: for (i=0; i<n; i++) iis[i] += rstart;
241: ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);
243: ISRestoreIndices(sis,&isis);
244: ISRestoreIndices(gis,&igis);
245: ISDestroy(&sis);
246: ISDestroy(&gis);
247: return(0);
248: }
250: /*
251: Distributes a SeqAIJ matrix across a set of processes. Code stolen from
252: MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.
254: Only for square matrices
256: Used by a preconditioner, hence PETSC_EXTERN
257: */
258: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
259: {
260: PetscMPIInt rank,size;
261: PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
263: Mat mat;
264: Mat_SeqAIJ *gmata;
265: PetscMPIInt tag;
266: MPI_Status status;
267: PetscBool aij;
268: MatScalar *gmataa,*ao,*ad,*gmataarestore=NULL;
271: MPI_Comm_rank(comm,&rank);
272: MPI_Comm_size(comm,&size);
273: if (!rank) {
274: PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
275: if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
276: }
277: if (reuse == MAT_INITIAL_MATRIX) {
278: MatCreate(comm,&mat);
279: MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
280: MatGetBlockSizes(gmat,&bses[0],&bses[1]);
281: MPI_Bcast(bses,2,MPIU_INT,0,comm);
282: MatSetBlockSizes(mat,bses[0],bses[1]);
283: MatSetType(mat,MATAIJ);
284: PetscMalloc1(size+1,&rowners);
285: PetscMalloc2(m,&dlens,m,&olens);
286: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
288: rowners[0] = 0;
289: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
290: rstart = rowners[rank];
291: rend = rowners[rank+1];
292: PetscObjectGetNewTag((PetscObject)mat,&tag);
293: if (!rank) {
294: gmata = (Mat_SeqAIJ*) gmat->data;
295: /* send row lengths to all processors */
296: for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
297: for (i=1; i<size; i++) {
298: MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
299: }
300: /* determine number diagonal and off-diagonal counts */
301: PetscArrayzero(olens,m);
302: PetscCalloc1(m,&ld);
303: jj = 0;
304: for (i=0; i<m; i++) {
305: for (j=0; j<dlens[i]; j++) {
306: if (gmata->j[jj] < rstart) ld[i]++;
307: if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
308: jj++;
309: }
310: }
311: /* send column indices to other processes */
312: for (i=1; i<size; i++) {
313: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
314: MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
315: MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
316: }
318: /* send numerical values to other processes */
319: for (i=1; i<size; i++) {
320: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
321: MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
322: }
323: gmataa = gmata->a;
324: gmataj = gmata->j;
326: } else {
327: /* receive row lengths */
328: MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
329: /* receive column indices */
330: MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
331: PetscMalloc2(nz,&gmataa,nz,&gmataj);
332: MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
333: /* determine number diagonal and off-diagonal counts */
334: PetscArrayzero(olens,m);
335: PetscCalloc1(m,&ld);
336: jj = 0;
337: for (i=0; i<m; i++) {
338: for (j=0; j<dlens[i]; j++) {
339: if (gmataj[jj] < rstart) ld[i]++;
340: if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
341: jj++;
342: }
343: }
344: /* receive numerical values */
345: PetscArrayzero(gmataa,nz);
346: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
347: }
348: /* set preallocation */
349: for (i=0; i<m; i++) {
350: dlens[i] -= olens[i];
351: }
352: MatSeqAIJSetPreallocation(mat,0,dlens);
353: MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
355: for (i=0; i<m; i++) {
356: dlens[i] += olens[i];
357: }
358: cnt = 0;
359: for (i=0; i<m; i++) {
360: row = rstart + i;
361: MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
362: cnt += dlens[i];
363: }
364: if (rank) {
365: PetscFree2(gmataa,gmataj);
366: }
367: PetscFree2(dlens,olens);
368: PetscFree(rowners);
370: ((Mat_MPIAIJ*)(mat->data))->ld = ld;
372: *inmat = mat;
373: } else { /* column indices are already set; only need to move over numerical values from process 0 */
374: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
375: Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
376: mat = *inmat;
377: PetscObjectGetNewTag((PetscObject)mat,&tag);
378: if (!rank) {
379: /* send numerical values to other processes */
380: gmata = (Mat_SeqAIJ*) gmat->data;
381: MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
382: gmataa = gmata->a;
383: for (i=1; i<size; i++) {
384: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
385: MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
386: }
387: nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
388: } else {
389: /* receive numerical values from process 0*/
390: nz = Ad->nz + Ao->nz;
391: PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
392: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
393: }
394: /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
395: ld = ((Mat_MPIAIJ*)(mat->data))->ld;
396: ad = Ad->a;
397: ao = Ao->a;
398: if (mat->rmap->n) {
399: i = 0;
400: nz = ld[i]; PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
401: nz = Ad->i[i+1] - Ad->i[i]; PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
402: }
403: for (i=1; i<mat->rmap->n; i++) {
404: nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
405: nz = Ad->i[i+1] - Ad->i[i]; PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
406: }
407: i--;
408: if (mat->rmap->n) {
409: nz = Ao->i[i+1] - Ao->i[i] - ld[i]; PetscArraycpy(ao,gmataa,nz);
410: }
411: if (rank) {
412: PetscFree(gmataarestore);
413: }
414: }
415: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
416: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
417: return(0);
418: }
420: /*
421: Local utility routine that creates a mapping from the global column
422: number to the local number in the off-diagonal part of the local
423: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
424: a slightly higher hash table cost; without it it is not scalable (each processor
425: has an order N integer array but is fast to acess.
426: */
427: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
428: {
429: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
431: PetscInt n = aij->B->cmap->n,i;
434: if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
435: #if defined(PETSC_USE_CTABLE)
436: PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
437: for (i=0; i<n; i++) {
438: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
439: }
440: #else
441: PetscCalloc1(mat->cmap->N+1,&aij->colmap);
442: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
443: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
444: #endif
445: return(0);
446: }
448: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \
449: { \
450: if (col <= lastcol1) low1 = 0; \
451: else high1 = nrow1; \
452: lastcol1 = col;\
453: while (high1-low1 > 5) { \
454: t = (low1+high1)/2; \
455: if (rp1[t] > col) high1 = t; \
456: else low1 = t; \
457: } \
458: for (_i=low1; _i<high1; _i++) { \
459: if (rp1[_i] > col) break; \
460: if (rp1[_i] == col) { \
461: if (addv == ADD_VALUES) { \
462: ap1[_i] += value; \
463: /* Not sure LogFlops will slow dow the code or not */ \
464: (void)PetscLogFlops(1.0); \
465: } \
466: else ap1[_i] = value; \
467: inserted = PETSC_TRUE; \
468: goto a_noinsert; \
469: } \
470: } \
471: if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
472: if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \
473: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
474: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
475: N = nrow1++ - 1; a->nz++; high1++; \
476: /* shift up all the later entries in this row */ \
477: PetscArraymove(rp1+_i+1,rp1+_i,N-_i+1);\
478: PetscArraymove(ap1+_i+1,ap1+_i,N-_i+1);\
479: rp1[_i] = col; \
480: ap1[_i] = value; \
481: A->nonzerostate++;\
482: a_noinsert: ; \
483: ailen[row] = nrow1; \
484: }
486: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
487: { \
488: if (col <= lastcol2) low2 = 0; \
489: else high2 = nrow2; \
490: lastcol2 = col; \
491: while (high2-low2 > 5) { \
492: t = (low2+high2)/2; \
493: if (rp2[t] > col) high2 = t; \
494: else low2 = t; \
495: } \
496: for (_i=low2; _i<high2; _i++) { \
497: if (rp2[_i] > col) break; \
498: if (rp2[_i] == col) { \
499: if (addv == ADD_VALUES) { \
500: ap2[_i] += value; \
501: (void)PetscLogFlops(1.0); \
502: } \
503: else ap2[_i] = value; \
504: inserted = PETSC_TRUE; \
505: goto b_noinsert; \
506: } \
507: } \
508: if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
509: if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
510: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
511: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
512: N = nrow2++ - 1; b->nz++; high2++; \
513: /* shift up all the later entries in this row */ \
514: PetscArraymove(rp2+_i+1,rp2+_i,N-_i+1);\
515: PetscArraymove(ap2+_i+1,ap2+_i,N-_i+1);\
516: rp2[_i] = col; \
517: ap2[_i] = value; \
518: B->nonzerostate++; \
519: b_noinsert: ; \
520: bilen[row] = nrow2; \
521: }
523: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
524: {
525: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
526: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
528: PetscInt l,*garray = mat->garray,diag;
531: /* code only works for square matrices A */
533: /* find size of row to the left of the diagonal part */
534: MatGetOwnershipRange(A,&diag,NULL);
535: row = row - diag;
536: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
537: if (garray[b->j[b->i[row]+l]] > diag) break;
538: }
539: PetscArraycpy(b->a+b->i[row],v,l);
541: /* diagonal part */
542: PetscArraycpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row]));
544: /* right of diagonal part */
545: PetscArraycpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],b->i[row+1]-b->i[row]-l);
546: #if defined(PETSC_HAVE_DEVICE)
547: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && (l || (a->i[row+1]-a->i[row]) || (b->i[row+1]-b->i[row]-l))) A->offloadmask = PETSC_OFFLOAD_CPU;
548: #endif
549: return(0);
550: }
552: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
553: {
554: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
555: PetscScalar value = 0.0;
557: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
558: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
559: PetscBool roworiented = aij->roworiented;
561: /* Some Variables required in the macro */
562: Mat A = aij->A;
563: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
564: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
565: MatScalar *aa = a->a;
566: PetscBool ignorezeroentries = a->ignorezeroentries;
567: Mat B = aij->B;
568: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
569: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
570: MatScalar *ba = b->a;
571: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
572: * cannot use "#if defined" inside a macro. */
573: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
575: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
576: PetscInt nonew;
577: MatScalar *ap1,*ap2;
580: for (i=0; i<m; i++) {
581: if (im[i] < 0) continue;
582: if (PetscUnlikely(im[i] >= mat->rmap->N)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
583: if (im[i] >= rstart && im[i] < rend) {
584: row = im[i] - rstart;
585: lastcol1 = -1;
586: rp1 = aj + ai[row];
587: ap1 = aa + ai[row];
588: rmax1 = aimax[row];
589: nrow1 = ailen[row];
590: low1 = 0;
591: high1 = nrow1;
592: lastcol2 = -1;
593: rp2 = bj + bi[row];
594: ap2 = ba + bi[row];
595: rmax2 = bimax[row];
596: nrow2 = bilen[row];
597: low2 = 0;
598: high2 = nrow2;
600: for (j=0; j<n; j++) {
601: if (v) value = roworiented ? v[i*n+j] : v[i+j*m];
602: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
603: if (in[j] >= cstart && in[j] < cend) {
604: col = in[j] - cstart;
605: nonew = a->nonew;
606: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
607: #if defined(PETSC_HAVE_DEVICE)
608: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
609: #endif
610: } else if (in[j] < 0) continue;
611: else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
612: else {
613: if (mat->was_assembled) {
614: if (!aij->colmap) {
615: MatCreateColmap_MPIAIJ_Private(mat);
616: }
617: #if defined(PETSC_USE_CTABLE)
618: PetscTableFind(aij->colmap,in[j]+1,&col);
619: col--;
620: #else
621: col = aij->colmap[in[j]] - 1;
622: #endif
623: if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
624: MatDisAssemble_MPIAIJ(mat);
625: col = in[j];
626: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
627: B = aij->B;
628: b = (Mat_SeqAIJ*)B->data;
629: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
630: rp2 = bj + bi[row];
631: ap2 = ba + bi[row];
632: rmax2 = bimax[row];
633: nrow2 = bilen[row];
634: low2 = 0;
635: high2 = nrow2;
636: bm = aij->B->rmap->n;
637: ba = b->a;
638: inserted = PETSC_FALSE;
639: } else if (col < 0) {
640: if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) {
641: PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);
642: } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]);
643: }
644: } else col = in[j];
645: nonew = b->nonew;
646: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
647: #if defined(PETSC_HAVE_DEVICE)
648: if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
649: #endif
650: }
651: }
652: } else {
653: if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
654: if (!aij->donotstash) {
655: mat->assembled = PETSC_FALSE;
656: if (roworiented) {
657: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
658: } else {
659: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
660: }
661: }
662: }
663: }
664: return(0);
665: }
667: /*
668: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
669: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
670: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
671: */
672: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[])
673: {
674: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
675: Mat A = aij->A; /* diagonal part of the matrix */
676: Mat B = aij->B; /* offdiagonal part of the matrix */
677: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
678: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
679: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col;
680: PetscInt *ailen = a->ilen,*aj = a->j;
681: PetscInt *bilen = b->ilen,*bj = b->j;
682: PetscInt am = aij->A->rmap->n,j;
683: PetscInt diag_so_far = 0,dnz;
684: PetscInt offd_so_far = 0,onz;
687: /* Iterate over all rows of the matrix */
688: for (j=0; j<am; j++) {
689: dnz = onz = 0;
690: /* Iterate over all non-zero columns of the current row */
691: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
692: /* If column is in the diagonal */
693: if (mat_j[col] >= cstart && mat_j[col] < cend) {
694: aj[diag_so_far++] = mat_j[col] - cstart;
695: dnz++;
696: } else { /* off-diagonal entries */
697: bj[offd_so_far++] = mat_j[col];
698: onz++;
699: }
700: }
701: ailen[j] = dnz;
702: bilen[j] = onz;
703: }
704: return(0);
705: }
707: /*
708: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
709: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
710: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
711: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
712: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
713: */
714: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[])
715: {
716: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
717: Mat A = aij->A; /* diagonal part of the matrix */
718: Mat B = aij->B; /* offdiagonal part of the matrix */
719: Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data;
720: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
721: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
722: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend;
723: PetscInt *ailen = a->ilen,*aj = a->j;
724: PetscInt *bilen = b->ilen,*bj = b->j;
725: PetscInt am = aij->A->rmap->n,j;
726: PetscInt *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */
727: PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd;
728: PetscScalar *aa = a->a,*ba = b->a;
731: /* Iterate over all rows of the matrix */
732: for (j=0; j<am; j++) {
733: dnz_row = onz_row = 0;
734: rowstart_offd = full_offd_i[j];
735: rowstart_diag = full_diag_i[j];
736: /* Iterate over all non-zero columns of the current row */
737: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
738: /* If column is in the diagonal */
739: if (mat_j[col] >= cstart && mat_j[col] < cend) {
740: aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
741: aa[rowstart_diag+dnz_row] = mat_a[col];
742: dnz_row++;
743: } else { /* off-diagonal entries */
744: bj[rowstart_offd+onz_row] = mat_j[col];
745: ba[rowstart_offd+onz_row] = mat_a[col];
746: onz_row++;
747: }
748: }
749: ailen[j] = dnz_row;
750: bilen[j] = onz_row;
751: }
752: return(0);
753: }
755: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
756: {
757: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
759: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
760: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
763: for (i=0; i<m; i++) {
764: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
765: if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
766: if (idxm[i] >= rstart && idxm[i] < rend) {
767: row = idxm[i] - rstart;
768: for (j=0; j<n; j++) {
769: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
770: if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
771: if (idxn[j] >= cstart && idxn[j] < cend) {
772: col = idxn[j] - cstart;
773: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
774: } else {
775: if (!aij->colmap) {
776: MatCreateColmap_MPIAIJ_Private(mat);
777: }
778: #if defined(PETSC_USE_CTABLE)
779: PetscTableFind(aij->colmap,idxn[j]+1,&col);
780: col--;
781: #else
782: col = aij->colmap[idxn[j]] - 1;
783: #endif
784: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
785: else {
786: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
787: }
788: }
789: }
790: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
791: }
792: return(0);
793: }
795: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);
797: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
798: {
799: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
801: PetscInt nstash,reallocs;
804: if (aij->donotstash || mat->nooffprocentries) return(0);
806: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
807: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
808: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
809: return(0);
810: }
812: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
813: {
814: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
815: Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data;
817: PetscMPIInt n;
818: PetscInt i,j,rstart,ncols,flg;
819: PetscInt *row,*col;
820: PetscBool other_disassembled;
821: PetscScalar *val;
823: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
826: if (!aij->donotstash && !mat->nooffprocentries) {
827: while (1) {
828: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
829: if (!flg) break;
831: for (i=0; i<n;) {
832: /* Now identify the consecutive vals belonging to the same row */
833: for (j=i,rstart=row[j]; j<n; j++) {
834: if (row[j] != rstart) break;
835: }
836: if (j < n) ncols = j-i;
837: else ncols = n-i;
838: /* Now assemble all these values with a single function call */
839: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);
840: i = j;
841: }
842: }
843: MatStashScatterEnd_Private(&mat->stash);
844: }
845: #if defined(PETSC_HAVE_DEVICE)
846: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
847: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
848: if (mat->boundtocpu) {
849: MatBindToCPU(aij->A,PETSC_TRUE);
850: MatBindToCPU(aij->B,PETSC_TRUE);
851: }
852: #endif
853: MatAssemblyBegin(aij->A,mode);
854: MatAssemblyEnd(aij->A,mode);
856: /* determine if any processor has disassembled, if so we must
857: also disassemble ourself, in order that we may reassemble. */
858: /*
859: if nonzero structure of submatrix B cannot change then we know that
860: no processor disassembled thus we can skip this stuff
861: */
862: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
863: MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
864: if (mat->was_assembled && !other_disassembled) {
865: #if defined(PETSC_HAVE_DEVICE)
866: aij->B->offloadmask = PETSC_OFFLOAD_BOTH; /* do not copy on the GPU when assembling inside MatDisAssemble_MPIAIJ */
867: #endif
868: MatDisAssemble_MPIAIJ(mat);
869: }
870: }
871: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
872: MatSetUpMultiply_MPIAIJ(mat);
873: }
874: MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
875: #if defined(PETSC_HAVE_DEVICE)
876: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
877: #endif
878: MatAssemblyBegin(aij->B,mode);
879: MatAssemblyEnd(aij->B,mode);
881: PetscFree2(aij->rowvalues,aij->rowindices);
883: aij->rowvalues = NULL;
885: VecDestroy(&aij->diag);
886: if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
888: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
889: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
890: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
891: MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
892: }
893: #if defined(PETSC_HAVE_DEVICE)
894: mat->offloadmask = PETSC_OFFLOAD_BOTH;
895: #endif
896: return(0);
897: }
899: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
900: {
901: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
905: MatZeroEntries(l->A);
906: MatZeroEntries(l->B);
907: return(0);
908: }
910: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
911: {
912: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
913: PetscObjectState sA, sB;
914: PetscInt *lrows;
915: PetscInt r, len;
916: PetscBool cong, lch, gch;
917: PetscErrorCode ierr;
920: /* get locally owned rows */
921: MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);
922: MatHasCongruentLayouts(A,&cong);
923: /* fix right hand side if needed */
924: if (x && b) {
925: const PetscScalar *xx;
926: PetscScalar *bb;
928: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
929: VecGetArrayRead(x, &xx);
930: VecGetArray(b, &bb);
931: for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
932: VecRestoreArrayRead(x, &xx);
933: VecRestoreArray(b, &bb);
934: }
936: sA = mat->A->nonzerostate;
937: sB = mat->B->nonzerostate;
939: if (diag != 0.0 && cong) {
940: MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
941: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
942: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
943: Mat_SeqAIJ *aijA = (Mat_SeqAIJ*)mat->A->data;
944: Mat_SeqAIJ *aijB = (Mat_SeqAIJ*)mat->B->data;
945: PetscInt nnwA, nnwB;
946: PetscBool nnzA, nnzB;
948: nnwA = aijA->nonew;
949: nnwB = aijB->nonew;
950: nnzA = aijA->keepnonzeropattern;
951: nnzB = aijB->keepnonzeropattern;
952: if (!nnzA) {
953: PetscInfo(mat->A,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n");
954: aijA->nonew = 0;
955: }
956: if (!nnzB) {
957: PetscInfo(mat->B,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n");
958: aijB->nonew = 0;
959: }
960: /* Must zero here before the next loop */
961: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
962: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
963: for (r = 0; r < len; ++r) {
964: const PetscInt row = lrows[r] + A->rmap->rstart;
965: if (row >= A->cmap->N) continue;
966: MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
967: }
968: aijA->nonew = nnwA;
969: aijB->nonew = nnwB;
970: } else {
971: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
972: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
973: }
974: PetscFree(lrows);
975: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
976: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
978: /* reduce nonzerostate */
979: lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
980: MPIU_Allreduce(&lch,&gch,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)A));
981: if (gch) A->nonzerostate++;
982: return(0);
983: }
985: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
986: {
987: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
988: PetscErrorCode ierr;
989: PetscMPIInt n = A->rmap->n;
990: PetscInt i,j,r,m,len = 0;
991: PetscInt *lrows,*owners = A->rmap->range;
992: PetscMPIInt p = 0;
993: PetscSFNode *rrows;
994: PetscSF sf;
995: const PetscScalar *xx;
996: PetscScalar *bb,*mask;
997: Vec xmask,lmask;
998: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data;
999: const PetscInt *aj, *ii,*ridx;
1000: PetscScalar *aa;
1003: /* Create SF where leaves are input rows and roots are owned rows */
1004: PetscMalloc1(n, &lrows);
1005: for (r = 0; r < n; ++r) lrows[r] = -1;
1006: PetscMalloc1(N, &rrows);
1007: for (r = 0; r < N; ++r) {
1008: const PetscInt idx = rows[r];
1009: if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
1010: if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
1011: PetscLayoutFindOwner(A->rmap,idx,&p);
1012: }
1013: rrows[r].rank = p;
1014: rrows[r].index = rows[r] - owners[p];
1015: }
1016: PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1017: PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1018: /* Collect flags for rows to be zeroed */
1019: PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1020: PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1021: PetscSFDestroy(&sf);
1022: /* Compress and put in row numbers */
1023: for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1024: /* zero diagonal part of matrix */
1025: MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
1026: /* handle off diagonal part of matrix */
1027: MatCreateVecs(A,&xmask,NULL);
1028: VecDuplicate(l->lvec,&lmask);
1029: VecGetArray(xmask,&bb);
1030: for (i=0; i<len; i++) bb[lrows[i]] = 1;
1031: VecRestoreArray(xmask,&bb);
1032: VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1033: VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1034: VecDestroy(&xmask);
1035: if (x && b) { /* this code is buggy when the row and column layout don't match */
1036: PetscBool cong;
1038: MatHasCongruentLayouts(A,&cong);
1039: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
1040: VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1041: VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1042: VecGetArrayRead(l->lvec,&xx);
1043: VecGetArray(b,&bb);
1044: }
1045: VecGetArray(lmask,&mask);
1046: /* remove zeroed rows of off diagonal matrix */
1047: ii = aij->i;
1048: for (i=0; i<len; i++) {
1049: PetscArrayzero(aij->a + ii[lrows[i]],ii[lrows[i]+1] - ii[lrows[i]]);
1050: }
1051: /* loop over all elements of off process part of matrix zeroing removed columns*/
1052: if (aij->compressedrow.use) {
1053: m = aij->compressedrow.nrows;
1054: ii = aij->compressedrow.i;
1055: ridx = aij->compressedrow.rindex;
1056: for (i=0; i<m; i++) {
1057: n = ii[i+1] - ii[i];
1058: aj = aij->j + ii[i];
1059: aa = aij->a + ii[i];
1061: for (j=0; j<n; j++) {
1062: if (PetscAbsScalar(mask[*aj])) {
1063: if (b) bb[*ridx] -= *aa*xx[*aj];
1064: *aa = 0.0;
1065: }
1066: aa++;
1067: aj++;
1068: }
1069: ridx++;
1070: }
1071: } else { /* do not use compressed row format */
1072: m = l->B->rmap->n;
1073: for (i=0; i<m; i++) {
1074: n = ii[i+1] - ii[i];
1075: aj = aij->j + ii[i];
1076: aa = aij->a + ii[i];
1077: for (j=0; j<n; j++) {
1078: if (PetscAbsScalar(mask[*aj])) {
1079: if (b) bb[i] -= *aa*xx[*aj];
1080: *aa = 0.0;
1081: }
1082: aa++;
1083: aj++;
1084: }
1085: }
1086: }
1087: if (x && b) {
1088: VecRestoreArray(b,&bb);
1089: VecRestoreArrayRead(l->lvec,&xx);
1090: }
1091: VecRestoreArray(lmask,&mask);
1092: VecDestroy(&lmask);
1093: PetscFree(lrows);
1095: /* only change matrix nonzero state if pattern was allowed to be changed */
1096: if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
1097: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1098: MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1099: }
1100: return(0);
1101: }
1103: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1104: {
1105: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1107: PetscInt nt;
1108: VecScatter Mvctx = a->Mvctx;
1111: VecGetLocalSize(xx,&nt);
1112: if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
1113: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1114: (*a->A->ops->mult)(a->A,xx,yy);
1115: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1116: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1117: return(0);
1118: }
1120: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1121: {
1122: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1126: MatMultDiagonalBlock(a->A,bb,xx);
1127: return(0);
1128: }
1130: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1131: {
1132: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1134: VecScatter Mvctx = a->Mvctx;
1137: if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1;
1138: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1139: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1140: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1141: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1142: return(0);
1143: }
1145: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1146: {
1147: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1151: /* do nondiagonal part */
1152: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1153: /* do local part */
1154: (*a->A->ops->multtranspose)(a->A,xx,yy);
1155: /* add partial results together */
1156: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1157: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1158: return(0);
1159: }
1161: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f)
1162: {
1163: MPI_Comm comm;
1164: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1165: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1166: IS Me,Notme;
1168: PetscInt M,N,first,last,*notme,i;
1169: PetscBool lf;
1170: PetscMPIInt size;
1173: /* Easy test: symmetric diagonal block */
1174: Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1175: MatIsTranspose(Adia,Bdia,tol,&lf);
1176: MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));
1177: if (!*f) return(0);
1178: PetscObjectGetComm((PetscObject)Amat,&comm);
1179: MPI_Comm_size(comm,&size);
1180: if (size == 1) return(0);
1182: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1183: MatGetSize(Amat,&M,&N);
1184: MatGetOwnershipRange(Amat,&first,&last);
1185: PetscMalloc1(N-last+first,¬me);
1186: for (i=0; i<first; i++) notme[i] = i;
1187: for (i=last; i<M; i++) notme[i-last+first] = i;
1188: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1189: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1190: MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1191: Aoff = Aoffs[0];
1192: MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1193: Boff = Boffs[0];
1194: MatIsTranspose(Aoff,Boff,tol,f);
1195: MatDestroyMatrices(1,&Aoffs);
1196: MatDestroyMatrices(1,&Boffs);
1197: ISDestroy(&Me);
1198: ISDestroy(&Notme);
1199: PetscFree(notme);
1200: return(0);
1201: }
1203: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f)
1204: {
1208: MatIsTranspose_MPIAIJ(A,A,tol,f);
1209: return(0);
1210: }
1212: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1213: {
1214: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1218: /* do nondiagonal part */
1219: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1220: /* do local part */
1221: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1222: /* add partial results together */
1223: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1224: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1225: return(0);
1226: }
1228: /*
1229: This only works correctly for square matrices where the subblock A->A is the
1230: diagonal block
1231: */
1232: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1233: {
1235: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1238: if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1239: if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1240: MatGetDiagonal(a->A,v);
1241: return(0);
1242: }
1244: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1245: {
1246: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1250: MatScale(a->A,aa);
1251: MatScale(a->B,aa);
1252: return(0);
1253: }
1255: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1256: {
1257: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1261: #if defined(PETSC_USE_LOG)
1262: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1263: #endif
1264: MatStashDestroy_Private(&mat->stash);
1265: VecDestroy(&aij->diag);
1266: MatDestroy(&aij->A);
1267: MatDestroy(&aij->B);
1268: #if defined(PETSC_USE_CTABLE)
1269: PetscTableDestroy(&aij->colmap);
1270: #else
1271: PetscFree(aij->colmap);
1272: #endif
1273: PetscFree(aij->garray);
1274: VecDestroy(&aij->lvec);
1275: VecScatterDestroy(&aij->Mvctx);
1276: if (aij->Mvctx_mpi1) {VecScatterDestroy(&aij->Mvctx_mpi1);}
1277: PetscFree2(aij->rowvalues,aij->rowindices);
1278: PetscFree(aij->ld);
1279: PetscFree(mat->data);
1281: /* may be created by MatCreateMPIAIJSumSeqAIJSymbolic */
1282: PetscObjectCompose((PetscObject)mat,"MatMergeSeqsToMPI",NULL);
1284: PetscObjectChangeTypeName((PetscObject)mat,NULL);
1285: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1286: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1287: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1288: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1289: PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);
1290: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1291: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1292: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpibaij_C",NULL);
1293: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1294: #if defined(PETSC_HAVE_ELEMENTAL)
1295: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1296: #endif
1297: #if defined(PETSC_HAVE_SCALAPACK)
1298: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_scalapack_C",NULL);
1299: #endif
1300: #if defined(PETSC_HAVE_HYPRE)
1301: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);
1302: PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_transpose_mpiaij_mpiaij_C",NULL);
1303: #endif
1304: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);
1305: PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_is_mpiaij_C",NULL);
1306: PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_mpiaij_mpiaij_C",NULL);
1307: return(0);
1308: }
1310: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1311: {
1312: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1313: Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data;
1314: Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data;
1315: const PetscInt *garray = aij->garray;
1316: PetscInt header[4],M,N,m,rs,cs,nz,cnt,i,ja,jb;
1317: PetscInt *rowlens;
1318: PetscInt *colidxs;
1319: PetscScalar *matvals;
1320: PetscErrorCode ierr;
1323: PetscViewerSetUp(viewer);
1325: M = mat->rmap->N;
1326: N = mat->cmap->N;
1327: m = mat->rmap->n;
1328: rs = mat->rmap->rstart;
1329: cs = mat->cmap->rstart;
1330: nz = A->nz + B->nz;
1332: /* write matrix header */
1333: header[0] = MAT_FILE_CLASSID;
1334: header[1] = M; header[2] = N; header[3] = nz;
1335: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1336: PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);
1338: /* fill in and store row lengths */
1339: PetscMalloc1(m,&rowlens);
1340: for (i=0; i<m; i++) rowlens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1341: PetscViewerBinaryWriteAll(viewer,rowlens,m,rs,M,PETSC_INT);
1342: PetscFree(rowlens);
1344: /* fill in and store column indices */
1345: PetscMalloc1(nz,&colidxs);
1346: for (cnt=0, i=0; i<m; i++) {
1347: for (jb=B->i[i]; jb<B->i[i+1]; jb++) {
1348: if (garray[B->j[jb]] > cs) break;
1349: colidxs[cnt++] = garray[B->j[jb]];
1350: }
1351: for (ja=A->i[i]; ja<A->i[i+1]; ja++)
1352: colidxs[cnt++] = A->j[ja] + cs;
1353: for (; jb<B->i[i+1]; jb++)
1354: colidxs[cnt++] = garray[B->j[jb]];
1355: }
1356: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1357: PetscViewerBinaryWriteAll(viewer,colidxs,nz,PETSC_DETERMINE,PETSC_DETERMINE,PETSC_INT);
1358: PetscFree(colidxs);
1360: /* fill in and store nonzero values */
1361: PetscMalloc1(nz,&matvals);
1362: for (cnt=0, i=0; i<m; i++) {
1363: for (jb=B->i[i]; jb<B->i[i+1]; jb++) {
1364: if (garray[B->j[jb]] > cs) break;
1365: matvals[cnt++] = B->a[jb];
1366: }
1367: for (ja=A->i[i]; ja<A->i[i+1]; ja++)
1368: matvals[cnt++] = A->a[ja];
1369: for (; jb<B->i[i+1]; jb++)
1370: matvals[cnt++] = B->a[jb];
1371: }
1372: if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1373: PetscViewerBinaryWriteAll(viewer,matvals,nz,PETSC_DETERMINE,PETSC_DETERMINE,PETSC_SCALAR);
1374: PetscFree(matvals);
1376: /* write block size option to the viewer's .info file */
1377: MatView_Binary_BlockSizes(mat,viewer);
1378: return(0);
1379: }
1381: #include <petscdraw.h>
1382: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1383: {
1384: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1385: PetscErrorCode ierr;
1386: PetscMPIInt rank = aij->rank,size = aij->size;
1387: PetscBool isdraw,iascii,isbinary;
1388: PetscViewer sviewer;
1389: PetscViewerFormat format;
1392: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1393: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1394: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1395: if (iascii) {
1396: PetscViewerGetFormat(viewer,&format);
1397: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1398: PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz;
1399: PetscMalloc1(size,&nz);
1400: MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
1401: for (i=0; i<(PetscInt)size; i++) {
1402: nmax = PetscMax(nmax,nz[i]);
1403: nmin = PetscMin(nmin,nz[i]);
1404: navg += nz[i];
1405: }
1406: PetscFree(nz);
1407: navg = navg/size;
1408: PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);
1409: return(0);
1410: }
1411: PetscViewerGetFormat(viewer,&format);
1412: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1413: MatInfo info;
1414: PetscBool inodes;
1416: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1417: MatGetInfo(mat,MAT_LOCAL,&info);
1418: MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1419: PetscViewerASCIIPushSynchronized(viewer);
1420: if (!inodes) {
1421: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n",
1422: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1423: } else {
1424: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n",
1425: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1426: }
1427: MatGetInfo(aij->A,MAT_LOCAL,&info);
1428: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1429: MatGetInfo(aij->B,MAT_LOCAL,&info);
1430: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1431: PetscViewerFlush(viewer);
1432: PetscViewerASCIIPopSynchronized(viewer);
1433: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1434: VecScatterView(aij->Mvctx,viewer);
1435: return(0);
1436: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1437: PetscInt inodecount,inodelimit,*inodes;
1438: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1439: if (inodes) {
1440: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1441: } else {
1442: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1443: }
1444: return(0);
1445: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1446: return(0);
1447: }
1448: } else if (isbinary) {
1449: if (size == 1) {
1450: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1451: MatView(aij->A,viewer);
1452: } else {
1453: MatView_MPIAIJ_Binary(mat,viewer);
1454: }
1455: return(0);
1456: } else if (iascii && size == 1) {
1457: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1458: MatView(aij->A,viewer);
1459: return(0);
1460: } else if (isdraw) {
1461: PetscDraw draw;
1462: PetscBool isnull;
1463: PetscViewerDrawGetDraw(viewer,0,&draw);
1464: PetscDrawIsNull(draw,&isnull);
1465: if (isnull) return(0);
1466: }
1468: { /* assemble the entire matrix onto first processor */
1469: Mat A = NULL, Av;
1470: IS isrow,iscol;
1472: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1473: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1474: MatCreateSubMatrix(mat,isrow,iscol,MAT_INITIAL_MATRIX,&A);
1475: MatMPIAIJGetSeqAIJ(A,&Av,NULL,NULL);
1476: /* The commented code uses MatCreateSubMatrices instead */
1477: /*
1478: Mat *AA, A = NULL, Av;
1479: IS isrow,iscol;
1481: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1482: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1483: MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA);
1484: if (!rank) {
1485: PetscObjectReference((PetscObject)AA[0]);
1486: A = AA[0];
1487: Av = AA[0];
1488: }
1489: MatDestroySubMatrices(1,&AA);
1490: */
1491: ISDestroy(&iscol);
1492: ISDestroy(&isrow);
1493: /*
1494: Everyone has to call to draw the matrix since the graphics waits are
1495: synchronized across all processors that share the PetscDraw object
1496: */
1497: PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1498: if (!rank) {
1499: if (((PetscObject)mat)->name) {
1500: PetscObjectSetName((PetscObject)Av,((PetscObject)mat)->name);
1501: }
1502: MatView_SeqAIJ(Av,sviewer);
1503: }
1504: PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1505: PetscViewerFlush(viewer);
1506: MatDestroy(&A);
1507: }
1508: return(0);
1509: }
1511: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1512: {
1514: PetscBool iascii,isdraw,issocket,isbinary;
1517: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1518: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1519: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1520: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1521: if (iascii || isdraw || isbinary || issocket) {
1522: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1523: }
1524: return(0);
1525: }
1527: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1528: {
1529: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1531: Vec bb1 = NULL;
1532: PetscBool hasop;
1535: if (flag == SOR_APPLY_UPPER) {
1536: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1537: return(0);
1538: }
1540: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1541: VecDuplicate(bb,&bb1);
1542: }
1544: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1545: if (flag & SOR_ZERO_INITIAL_GUESS) {
1546: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1547: its--;
1548: }
1550: while (its--) {
1551: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1552: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1554: /* update rhs: bb1 = bb - B*x */
1555: VecScale(mat->lvec,-1.0);
1556: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1558: /* local sweep */
1559: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1560: }
1561: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1562: if (flag & SOR_ZERO_INITIAL_GUESS) {
1563: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1564: its--;
1565: }
1566: while (its--) {
1567: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1568: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1570: /* update rhs: bb1 = bb - B*x */
1571: VecScale(mat->lvec,-1.0);
1572: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1574: /* local sweep */
1575: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1576: }
1577: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1578: if (flag & SOR_ZERO_INITIAL_GUESS) {
1579: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1580: its--;
1581: }
1582: while (its--) {
1583: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1584: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1586: /* update rhs: bb1 = bb - B*x */
1587: VecScale(mat->lvec,-1.0);
1588: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1590: /* local sweep */
1591: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1592: }
1593: } else if (flag & SOR_EISENSTAT) {
1594: Vec xx1;
1596: VecDuplicate(bb,&xx1);
1597: (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);
1599: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1600: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1601: if (!mat->diag) {
1602: MatCreateVecs(matin,&mat->diag,NULL);
1603: MatGetDiagonal(matin,mat->diag);
1604: }
1605: MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1606: if (hasop) {
1607: MatMultDiagonalBlock(matin,xx,bb1);
1608: } else {
1609: VecPointwiseMult(bb1,mat->diag,xx);
1610: }
1611: VecAYPX(bb1,(omega-2.0)/omega,bb);
1613: MatMultAdd(mat->B,mat->lvec,bb1,bb1);
1615: /* local sweep */
1616: (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1617: VecAXPY(xx,1.0,xx1);
1618: VecDestroy(&xx1);
1619: } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");
1621: VecDestroy(&bb1);
1623: matin->factorerrortype = mat->A->factorerrortype;
1624: return(0);
1625: }
1627: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1628: {
1629: Mat aA,aB,Aperm;
1630: const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1631: PetscScalar *aa,*ba;
1632: PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1633: PetscSF rowsf,sf;
1634: IS parcolp = NULL;
1635: PetscBool done;
1639: MatGetLocalSize(A,&m,&n);
1640: ISGetIndices(rowp,&rwant);
1641: ISGetIndices(colp,&cwant);
1642: PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);
1644: /* Invert row permutation to find out where my rows should go */
1645: PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1646: PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1647: PetscSFSetFromOptions(rowsf);
1648: for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1649: PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1650: PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1652: /* Invert column permutation to find out where my columns should go */
1653: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1654: PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1655: PetscSFSetFromOptions(sf);
1656: for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1657: PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1658: PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1659: PetscSFDestroy(&sf);
1661: ISRestoreIndices(rowp,&rwant);
1662: ISRestoreIndices(colp,&cwant);
1663: MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);
1665: /* Find out where my gcols should go */
1666: MatGetSize(aB,NULL,&ng);
1667: PetscMalloc1(ng,&gcdest);
1668: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1669: PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1670: PetscSFSetFromOptions(sf);
1671: PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1672: PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1673: PetscSFDestroy(&sf);
1675: PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1676: MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1677: MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1678: for (i=0; i<m; i++) {
1679: PetscInt row = rdest[i];
1680: PetscMPIInt rowner;
1681: PetscLayoutFindOwner(A->rmap,row,&rowner);
1682: for (j=ai[i]; j<ai[i+1]; j++) {
1683: PetscInt col = cdest[aj[j]];
1684: PetscMPIInt cowner;
1685: PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1686: if (rowner == cowner) dnnz[i]++;
1687: else onnz[i]++;
1688: }
1689: for (j=bi[i]; j<bi[i+1]; j++) {
1690: PetscInt col = gcdest[bj[j]];
1691: PetscMPIInt cowner;
1692: PetscLayoutFindOwner(A->cmap,col,&cowner);
1693: if (rowner == cowner) dnnz[i]++;
1694: else onnz[i]++;
1695: }
1696: }
1697: PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);
1698: PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);
1699: PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);
1700: PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);
1701: PetscSFDestroy(&rowsf);
1703: MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);
1704: MatSeqAIJGetArray(aA,&aa);
1705: MatSeqAIJGetArray(aB,&ba);
1706: for (i=0; i<m; i++) {
1707: PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */
1708: PetscInt j0,rowlen;
1709: rowlen = ai[i+1] - ai[i];
1710: for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */
1711: for (; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]];
1712: MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);
1713: }
1714: rowlen = bi[i+1] - bi[i];
1715: for (j0=j=0; j<rowlen; j0=j) {
1716: for (; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]];
1717: MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);
1718: }
1719: }
1720: MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);
1721: MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);
1722: MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1723: MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1724: MatSeqAIJRestoreArray(aA,&aa);
1725: MatSeqAIJRestoreArray(aB,&ba);
1726: PetscFree4(dnnz,onnz,tdnnz,tonnz);
1727: PetscFree3(work,rdest,cdest);
1728: PetscFree(gcdest);
1729: if (parcolp) {ISDestroy(&colp);}
1730: *B = Aperm;
1731: return(0);
1732: }
1734: PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
1735: {
1736: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1740: MatGetSize(aij->B,NULL,nghosts);
1741: if (ghosts) *ghosts = aij->garray;
1742: return(0);
1743: }
1745: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1746: {
1747: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1748: Mat A = mat->A,B = mat->B;
1750: PetscLogDouble isend[5],irecv[5];
1753: info->block_size = 1.0;
1754: MatGetInfo(A,MAT_LOCAL,info);
1756: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1757: isend[3] = info->memory; isend[4] = info->mallocs;
1759: MatGetInfo(B,MAT_LOCAL,info);
1761: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1762: isend[3] += info->memory; isend[4] += info->mallocs;
1763: if (flag == MAT_LOCAL) {
1764: info->nz_used = isend[0];
1765: info->nz_allocated = isend[1];
1766: info->nz_unneeded = isend[2];
1767: info->memory = isend[3];
1768: info->mallocs = isend[4];
1769: } else if (flag == MAT_GLOBAL_MAX) {
1770: MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_MAX,PetscObjectComm((PetscObject)matin));
1772: info->nz_used = irecv[0];
1773: info->nz_allocated = irecv[1];
1774: info->nz_unneeded = irecv[2];
1775: info->memory = irecv[3];
1776: info->mallocs = irecv[4];
1777: } else if (flag == MAT_GLOBAL_SUM) {
1778: MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_SUM,PetscObjectComm((PetscObject)matin));
1780: info->nz_used = irecv[0];
1781: info->nz_allocated = irecv[1];
1782: info->nz_unneeded = irecv[2];
1783: info->memory = irecv[3];
1784: info->mallocs = irecv[4];
1785: }
1786: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1787: info->fill_ratio_needed = 0;
1788: info->factor_mallocs = 0;
1789: return(0);
1790: }
1792: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1793: {
1794: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1798: switch (op) {
1799: case MAT_NEW_NONZERO_LOCATIONS:
1800: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1801: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1802: case MAT_KEEP_NONZERO_PATTERN:
1803: case MAT_NEW_NONZERO_LOCATION_ERR:
1804: case MAT_USE_INODES:
1805: case MAT_IGNORE_ZERO_ENTRIES:
1806: MatCheckPreallocated(A,1);
1807: MatSetOption(a->A,op,flg);
1808: MatSetOption(a->B,op,flg);
1809: break;
1810: case MAT_ROW_ORIENTED:
1811: MatCheckPreallocated(A,1);
1812: a->roworiented = flg;
1814: MatSetOption(a->A,op,flg);
1815: MatSetOption(a->B,op,flg);
1816: break;
1817: case MAT_NEW_DIAGONALS:
1818: case MAT_SORTED_FULL:
1819: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1820: break;
1821: case MAT_IGNORE_OFF_PROC_ENTRIES:
1822: a->donotstash = flg;
1823: break;
1824: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1825: case MAT_SPD:
1826: case MAT_SYMMETRIC:
1827: case MAT_STRUCTURALLY_SYMMETRIC:
1828: case MAT_HERMITIAN:
1829: case MAT_SYMMETRY_ETERNAL:
1830: break;
1831: case MAT_SUBMAT_SINGLEIS:
1832: A->submat_singleis = flg;
1833: break;
1834: case MAT_STRUCTURE_ONLY:
1835: /* The option is handled directly by MatSetOption() */
1836: break;
1837: default:
1838: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1839: }
1840: return(0);
1841: }
1843: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1844: {
1845: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1846: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1848: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1849: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1850: PetscInt *cmap,*idx_p;
1853: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1854: mat->getrowactive = PETSC_TRUE;
1856: if (!mat->rowvalues && (idx || v)) {
1857: /*
1858: allocate enough space to hold information from the longest row.
1859: */
1860: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1861: PetscInt max = 1,tmp;
1862: for (i=0; i<matin->rmap->n; i++) {
1863: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1864: if (max < tmp) max = tmp;
1865: }
1866: PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1867: }
1869: if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1870: lrow = row - rstart;
1872: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1873: if (!v) {pvA = NULL; pvB = NULL;}
1874: if (!idx) {pcA = NULL; if (!v) pcB = NULL;}
1875: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1876: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1877: nztot = nzA + nzB;
1879: cmap = mat->garray;
1880: if (v || idx) {
1881: if (nztot) {
1882: /* Sort by increasing column numbers, assuming A and B already sorted */
1883: PetscInt imark = -1;
1884: if (v) {
1885: *v = v_p = mat->rowvalues;
1886: for (i=0; i<nzB; i++) {
1887: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1888: else break;
1889: }
1890: imark = i;
1891: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1892: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1893: }
1894: if (idx) {
1895: *idx = idx_p = mat->rowindices;
1896: if (imark > -1) {
1897: for (i=0; i<imark; i++) {
1898: idx_p[i] = cmap[cworkB[i]];
1899: }
1900: } else {
1901: for (i=0; i<nzB; i++) {
1902: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1903: else break;
1904: }
1905: imark = i;
1906: }
1907: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1908: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1909: }
1910: } else {
1911: if (idx) *idx = NULL;
1912: if (v) *v = NULL;
1913: }
1914: }
1915: *nz = nztot;
1916: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1917: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1918: return(0);
1919: }
1921: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1922: {
1923: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1926: if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1927: aij->getrowactive = PETSC_FALSE;
1928: return(0);
1929: }
1931: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1932: {
1933: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1934: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1936: PetscInt i,j,cstart = mat->cmap->rstart;
1937: PetscReal sum = 0.0;
1938: MatScalar *v;
1941: if (aij->size == 1) {
1942: MatNorm(aij->A,type,norm);
1943: } else {
1944: if (type == NORM_FROBENIUS) {
1945: v = amat->a;
1946: for (i=0; i<amat->nz; i++) {
1947: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1948: }
1949: v = bmat->a;
1950: for (i=0; i<bmat->nz; i++) {
1951: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1952: }
1953: MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1954: *norm = PetscSqrtReal(*norm);
1955: PetscLogFlops(2.0*amat->nz+2.0*bmat->nz);
1956: } else if (type == NORM_1) { /* max column norm */
1957: PetscReal *tmp,*tmp2;
1958: PetscInt *jj,*garray = aij->garray;
1959: PetscCalloc1(mat->cmap->N+1,&tmp);
1960: PetscMalloc1(mat->cmap->N+1,&tmp2);
1961: *norm = 0.0;
1962: v = amat->a; jj = amat->j;
1963: for (j=0; j<amat->nz; j++) {
1964: tmp[cstart + *jj++] += PetscAbsScalar(*v); v++;
1965: }
1966: v = bmat->a; jj = bmat->j;
1967: for (j=0; j<bmat->nz; j++) {
1968: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1969: }
1970: MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1971: for (j=0; j<mat->cmap->N; j++) {
1972: if (tmp2[j] > *norm) *norm = tmp2[j];
1973: }
1974: PetscFree(tmp);
1975: PetscFree(tmp2);
1976: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1977: } else if (type == NORM_INFINITY) { /* max row norm */
1978: PetscReal ntemp = 0.0;
1979: for (j=0; j<aij->A->rmap->n; j++) {
1980: v = amat->a + amat->i[j];
1981: sum = 0.0;
1982: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1983: sum += PetscAbsScalar(*v); v++;
1984: }
1985: v = bmat->a + bmat->i[j];
1986: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1987: sum += PetscAbsScalar(*v); v++;
1988: }
1989: if (sum > ntemp) ntemp = sum;
1990: }
1991: MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
1992: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1993: } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
1994: }
1995: return(0);
1996: }
1998: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
1999: {
2000: Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b;
2001: Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag;
2002: PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,row,*cols,*cols_tmp,*B_diag_ilen,i,ncol,A_diag_ncol;
2003: const PetscInt *ai,*aj,*bi,*bj,*B_diag_i;
2004: PetscErrorCode ierr;
2005: Mat B,A_diag,*B_diag;
2006: const MatScalar *array;
2009: ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
2010: ai = Aloc->i; aj = Aloc->j;
2011: bi = Bloc->i; bj = Bloc->j;
2012: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2013: PetscInt *d_nnz,*g_nnz,*o_nnz;
2014: PetscSFNode *oloc;
2015: PETSC_UNUSED PetscSF sf;
2017: PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
2018: /* compute d_nnz for preallocation */
2019: PetscArrayzero(d_nnz,na);
2020: for (i=0; i<ai[ma]; i++) {
2021: d_nnz[aj[i]]++;
2022: }
2023: /* compute local off-diagonal contributions */
2024: PetscArrayzero(g_nnz,nb);
2025: for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
2026: /* map those to global */
2027: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2028: PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2029: PetscSFSetFromOptions(sf);
2030: PetscArrayzero(o_nnz,na);
2031: PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2032: PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2033: PetscSFDestroy(&sf);
2035: MatCreate(PetscObjectComm((PetscObject)A),&B);
2036: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2037: MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2038: MatSetType(B,((PetscObject)A)->type_name);
2039: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2040: PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2041: } else {
2042: B = *matout;
2043: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2044: }
2046: b = (Mat_MPIAIJ*)B->data;
2047: A_diag = a->A;
2048: B_diag = &b->A;
2049: sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data;
2050: A_diag_ncol = A_diag->cmap->N;
2051: B_diag_ilen = sub_B_diag->ilen;
2052: B_diag_i = sub_B_diag->i;
2054: /* Set ilen for diagonal of B */
2055: for (i=0; i<A_diag_ncol; i++) {
2056: B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i];
2057: }
2059: /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
2060: very quickly (=without using MatSetValues), because all writes are local. */
2061: MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);
2063: /* copy over the B part */
2064: PetscMalloc1(bi[mb],&cols);
2065: array = Bloc->a;
2066: row = A->rmap->rstart;
2067: for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2068: cols_tmp = cols;
2069: for (i=0; i<mb; i++) {
2070: ncol = bi[i+1]-bi[i];
2071: MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2072: row++;
2073: array += ncol; cols_tmp += ncol;
2074: }
2075: PetscFree(cols);
2077: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2078: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2079: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2080: *matout = B;
2081: } else {
2082: MatHeaderMerge(A,&B);
2083: }
2084: return(0);
2085: }
2087: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2088: {
2089: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2090: Mat a = aij->A,b = aij->B;
2092: PetscInt s1,s2,s3;
2095: MatGetLocalSize(mat,&s2,&s3);
2096: if (rr) {
2097: VecGetLocalSize(rr,&s1);
2098: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2099: /* Overlap communication with computation. */
2100: VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2101: }
2102: if (ll) {
2103: VecGetLocalSize(ll,&s1);
2104: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2105: (*b->ops->diagonalscale)(b,ll,NULL);
2106: }
2107: /* scale the diagonal block */
2108: (*a->ops->diagonalscale)(a,ll,rr);
2110: if (rr) {
2111: /* Do a scatter end and then right scale the off-diagonal block */
2112: VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2113: (*b->ops->diagonalscale)(b,NULL,aij->lvec);
2114: }
2115: return(0);
2116: }
2118: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2119: {
2120: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2124: MatSetUnfactored(a->A);
2125: return(0);
2126: }
2128: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag)
2129: {
2130: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2131: Mat a,b,c,d;
2132: PetscBool flg;
2136: a = matA->A; b = matA->B;
2137: c = matB->A; d = matB->B;
2139: MatEqual(a,c,&flg);
2140: if (flg) {
2141: MatEqual(b,d,&flg);
2142: }
2143: MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2144: return(0);
2145: }
2147: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2148: {
2150: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2151: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
2154: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2155: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2156: /* because of the column compression in the off-processor part of the matrix a->B,
2157: the number of columns in a->B and b->B may be different, hence we cannot call
2158: the MatCopy() directly on the two parts. If need be, we can provide a more
2159: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2160: then copying the submatrices */
2161: MatCopy_Basic(A,B,str);
2162: } else {
2163: MatCopy(a->A,b->A,str);
2164: MatCopy(a->B,b->B,str);
2165: }
2166: PetscObjectStateIncrease((PetscObject)B);
2167: return(0);
2168: }
2170: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2171: {
2175: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,NULL,PETSC_DEFAULT,NULL);
2176: return(0);
2177: }
2179: /*
2180: Computes the number of nonzeros per row needed for preallocation when X and Y
2181: have different nonzero structure.
2182: */
2183: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz)
2184: {
2185: PetscInt i,j,k,nzx,nzy;
2188: /* Set the number of nonzeros in the new matrix */
2189: for (i=0; i<m; i++) {
2190: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2191: nzx = xi[i+1] - xi[i];
2192: nzy = yi[i+1] - yi[i];
2193: nnz[i] = 0;
2194: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2195: for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2196: if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */
2197: nnz[i]++;
2198: }
2199: for (; k<nzy; k++) nnz[i]++;
2200: }
2201: return(0);
2202: }
2204: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2205: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2206: {
2208: PetscInt m = Y->rmap->N;
2209: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2210: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2213: MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2214: return(0);
2215: }
2217: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2218: {
2220: Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2221: PetscBLASInt bnz,one=1;
2222: Mat_SeqAIJ *x,*y;
2225: if (str == SAME_NONZERO_PATTERN) {
2226: PetscScalar alpha = a;
2227: x = (Mat_SeqAIJ*)xx->A->data;
2228: PetscBLASIntCast(x->nz,&bnz);
2229: y = (Mat_SeqAIJ*)yy->A->data;
2230: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2231: x = (Mat_SeqAIJ*)xx->B->data;
2232: y = (Mat_SeqAIJ*)yy->B->data;
2233: PetscBLASIntCast(x->nz,&bnz);
2234: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2235: PetscObjectStateIncrease((PetscObject)Y);
2236: /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU
2237: will be updated */
2238: #if defined(PETSC_HAVE_DEVICE)
2239: if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) {
2240: Y->offloadmask = PETSC_OFFLOAD_CPU;
2241: }
2242: #endif
2243: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2244: MatAXPY_Basic(Y,a,X,str);
2245: } else {
2246: Mat B;
2247: PetscInt *nnz_d,*nnz_o;
2248: PetscMalloc1(yy->A->rmap->N,&nnz_d);
2249: PetscMalloc1(yy->B->rmap->N,&nnz_o);
2250: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2251: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2252: MatSetLayouts(B,Y->rmap,Y->cmap);
2253: MatSetType(B,MATMPIAIJ);
2254: MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2255: MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2256: MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2257: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2258: MatHeaderReplace(Y,&B);
2259: PetscFree(nnz_d);
2260: PetscFree(nnz_o);
2261: }
2262: return(0);
2263: }
2265: extern PetscErrorCode MatConjugate_SeqAIJ(Mat);
2267: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2268: {
2269: #if defined(PETSC_USE_COMPLEX)
2271: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2274: MatConjugate_SeqAIJ(aij->A);
2275: MatConjugate_SeqAIJ(aij->B);
2276: #else
2278: #endif
2279: return(0);
2280: }
2282: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2283: {
2284: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2288: MatRealPart(a->A);
2289: MatRealPart(a->B);
2290: return(0);
2291: }
2293: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2294: {
2295: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2299: MatImaginaryPart(a->A);
2300: MatImaginaryPart(a->B);
2301: return(0);
2302: }
2304: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A,Vec v,PetscInt idx[])
2305: {
2306: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2307: PetscErrorCode ierr;
2308: PetscInt i,*idxb = NULL,m = A->rmap->n;
2309: PetscScalar *va,*vv;
2310: Vec vB,vA;
2311: const PetscScalar *vb;
2314: VecCreateSeq(PETSC_COMM_SELF,m,&vA);
2315: MatGetRowMaxAbs(a->A,vA,idx);
2317: VecGetArrayWrite(vA,&va);
2318: if (idx) {
2319: for (i=0; i<m; i++) {
2320: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2321: }
2322: }
2324: VecCreateSeq(PETSC_COMM_SELF,m,&vB);
2325: PetscMalloc1(m,&idxb);
2326: MatGetRowMaxAbs(a->B,vB,idxb);
2328: VecGetArrayWrite(v,&vv);
2329: VecGetArrayRead(vB,&vb);
2330: for (i=0; i<m; i++) {
2331: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2332: vv[i] = vb[i];
2333: if (idx) idx[i] = a->garray[idxb[i]];
2334: } else {
2335: vv[i] = va[i];
2336: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]])
2337: idx[i] = a->garray[idxb[i]];
2338: }
2339: }
2340: VecRestoreArrayWrite(vA,&vv);
2341: VecRestoreArrayWrite(vA,&va);
2342: VecRestoreArrayRead(vB,&vb);
2343: PetscFree(idxb);
2344: VecDestroy(&vA);
2345: VecDestroy(&vB);
2346: return(0);
2347: }
2349: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2350: {
2351: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2352: PetscInt m = A->rmap->n,n = A->cmap->n;
2353: PetscInt cstart = A->cmap->rstart,cend = A->cmap->rend;
2354: PetscInt *cmap = mat->garray;
2355: PetscInt *diagIdx, *offdiagIdx;
2356: Vec diagV, offdiagV;
2357: PetscScalar *a, *diagA, *offdiagA, *ba;
2358: PetscInt r,j,col,ncols,*bi,*bj;
2360: Mat B = mat->B;
2361: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2364: /* When a process holds entire A and other processes have no entry */
2365: if (A->cmap->N == n) {
2366: VecGetArrayWrite(v,&diagA);
2367: VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,diagA,&diagV);
2368: MatGetRowMinAbs(mat->A,diagV,idx);
2369: VecDestroy(&diagV);
2370: VecRestoreArrayWrite(v,&diagA);
2371: return(0);
2372: } else if (n == 0) {
2373: if (m) {
2374: VecGetArrayWrite(v,&a);
2375: for (r = 0; r < m; r++) {a[r] = 0.0; if (idx) idx[r] = -1;}
2376: VecRestoreArrayWrite(v,&a);
2377: }
2378: return(0);
2379: }
2381: PetscMalloc2(m,&diagIdx,m,&offdiagIdx);
2382: VecCreateSeq(PETSC_COMM_SELF, m, &diagV);
2383: VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV);
2384: MatGetRowMinAbs(mat->A, diagV, diagIdx);
2386: /* Get offdiagIdx[] for implicit 0.0 */
2387: ba = b->a;
2388: bi = b->i;
2389: bj = b->j;
2390: VecGetArrayWrite(offdiagV, &offdiagA);
2391: for (r = 0; r < m; r++) {
2392: ncols = bi[r+1] - bi[r];
2393: if (ncols == A->cmap->N - n) { /* Brow is dense */
2394: offdiagA[r] = *ba; offdiagIdx[r] = cmap[0];
2395: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2396: offdiagA[r] = 0.0;
2398: /* Find first hole in the cmap */
2399: for (j=0; j<ncols; j++) {
2400: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2401: if (col > j && j < cstart) {
2402: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2403: break;
2404: } else if (col > j + n && j >= cstart) {
2405: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2406: break;
2407: }
2408: }
2409: if (j == ncols && ncols < A->cmap->N - n) {
2410: /* a hole is outside compressed Bcols */
2411: if (ncols == 0) {
2412: if (cstart) {
2413: offdiagIdx[r] = 0;
2414: } else offdiagIdx[r] = cend;
2415: } else { /* ncols > 0 */
2416: offdiagIdx[r] = cmap[ncols-1] + 1;
2417: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2418: }
2419: }
2420: }
2422: for (j=0; j<ncols; j++) {
2423: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {offdiagA[r] = *ba; offdiagIdx[r] = cmap[*bj];}
2424: ba++; bj++;
2425: }
2426: }
2428: VecGetArrayWrite(v, &a);
2429: VecGetArrayRead(diagV, (const PetscScalar**)&diagA);
2430: for (r = 0; r < m; ++r) {
2431: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2432: a[r] = diagA[r];
2433: if (idx) idx[r] = cstart + diagIdx[r];
2434: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2435: a[r] = diagA[r];
2436: if (idx) {
2437: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2438: idx[r] = cstart + diagIdx[r];
2439: } else idx[r] = offdiagIdx[r];
2440: }
2441: } else {
2442: a[r] = offdiagA[r];
2443: if (idx) idx[r] = offdiagIdx[r];
2444: }
2445: }
2446: VecRestoreArrayWrite(v, &a);
2447: VecRestoreArrayRead(diagV, (const PetscScalar**)&diagA);
2448: VecRestoreArrayWrite(offdiagV, &offdiagA);
2449: VecDestroy(&diagV);
2450: VecDestroy(&offdiagV);
2451: PetscFree2(diagIdx, offdiagIdx);
2452: return(0);
2453: }
2455: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A,Vec v,PetscInt idx[])
2456: {
2457: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2458: PetscInt m = A->rmap->n,n = A->cmap->n;
2459: PetscInt cstart = A->cmap->rstart,cend = A->cmap->rend;
2460: PetscInt *cmap = mat->garray;
2461: PetscInt *diagIdx, *offdiagIdx;
2462: Vec diagV, offdiagV;
2463: PetscScalar *a, *diagA, *offdiagA, *ba;
2464: PetscInt r,j,col,ncols,*bi,*bj;
2466: Mat B = mat->B;
2467: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2470: /* When a process holds entire A and other processes have no entry */
2471: if (A->cmap->N == n) {
2472: VecGetArrayWrite(v,&diagA);
2473: VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,diagA,&diagV);
2474: MatGetRowMin(mat->A,diagV,idx);
2475: VecDestroy(&diagV);
2476: VecRestoreArrayWrite(v,&diagA);
2477: return(0);
2478: } else if (n == 0) {
2479: if (m) {
2480: VecGetArrayWrite(v,&a);
2481: for (r = 0; r < m; r++) {a[r] = PETSC_MAX_REAL; if (idx) idx[r] = -1;}
2482: VecRestoreArrayWrite(v,&a);
2483: }
2484: return(0);
2485: }
2487: PetscCalloc2(m,&diagIdx,m,&offdiagIdx);
2488: VecCreateSeq(PETSC_COMM_SELF, m, &diagV);
2489: VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV);
2490: MatGetRowMin(mat->A, diagV, diagIdx);
2492: /* Get offdiagIdx[] for implicit 0.0 */
2493: ba = b->a;
2494: bi = b->i;
2495: bj = b->j;
2496: VecGetArrayWrite(offdiagV, &offdiagA);
2497: for (r = 0; r < m; r++) {
2498: ncols = bi[r+1] - bi[r];
2499: if (ncols == A->cmap->N - n) { /* Brow is dense */
2500: offdiagA[r] = *ba; offdiagIdx[r] = cmap[0];
2501: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2502: offdiagA[r] = 0.0;
2504: /* Find first hole in the cmap */
2505: for (j=0; j<ncols; j++) {
2506: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2507: if (col > j && j < cstart) {
2508: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2509: break;
2510: } else if (col > j + n && j >= cstart) {
2511: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2512: break;
2513: }
2514: }
2515: if (j == ncols && ncols < A->cmap->N - n) {
2516: /* a hole is outside compressed Bcols */
2517: if (ncols == 0) {
2518: if (cstart) {
2519: offdiagIdx[r] = 0;
2520: } else offdiagIdx[r] = cend;
2521: } else { /* ncols > 0 */
2522: offdiagIdx[r] = cmap[ncols-1] + 1;
2523: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2524: }
2525: }
2526: }
2528: for (j=0; j<ncols; j++) {
2529: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {offdiagA[r] = *ba; offdiagIdx[r] = cmap[*bj];}
2530: ba++; bj++;
2531: }
2532: }
2534: VecGetArrayWrite(v, &a);
2535: VecGetArrayRead(diagV, (const PetscScalar**)&diagA);
2536: for (r = 0; r < m; ++r) {
2537: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2538: a[r] = diagA[r];
2539: if (idx) idx[r] = cstart + diagIdx[r];
2540: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2541: a[r] = diagA[r];
2542: if (idx) {
2543: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2544: idx[r] = cstart + diagIdx[r];
2545: } else idx[r] = offdiagIdx[r];
2546: }
2547: } else {
2548: a[r] = offdiagA[r];
2549: if (idx) idx[r] = offdiagIdx[r];
2550: }
2551: }
2552: VecRestoreArrayWrite(v, &a);
2553: VecRestoreArrayRead(diagV, (const PetscScalar**)&diagA);
2554: VecRestoreArrayWrite(offdiagV, &offdiagA);
2555: VecDestroy(&diagV);
2556: VecDestroy(&offdiagV);
2557: PetscFree2(diagIdx, offdiagIdx);
2558: return(0);
2559: }
2561: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A,Vec v,PetscInt idx[])
2562: {
2563: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
2564: PetscInt m = A->rmap->n,n = A->cmap->n;
2565: PetscInt cstart = A->cmap->rstart,cend = A->cmap->rend;
2566: PetscInt *cmap = mat->garray;
2567: PetscInt *diagIdx, *offdiagIdx;
2568: Vec diagV, offdiagV;
2569: PetscScalar *a, *diagA, *offdiagA, *ba;
2570: PetscInt r,j,col,ncols,*bi,*bj;
2572: Mat B = mat->B;
2573: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2576: /* When a process holds entire A and other processes have no entry */
2577: if (A->cmap->N == n) {
2578: VecGetArrayWrite(v,&diagA);
2579: VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,diagA,&diagV);
2580: MatGetRowMax(mat->A,diagV,idx);
2581: VecDestroy(&diagV);
2582: VecRestoreArrayWrite(v,&diagA);
2583: return(0);
2584: } else if (n == 0) {
2585: if (m) {
2586: VecGetArrayWrite(v,&a);
2587: for (r = 0; r < m; r++) {a[r] = PETSC_MIN_REAL; if (idx) idx[r] = -1;}
2588: VecRestoreArrayWrite(v,&a);
2589: }
2590: return(0);
2591: }
2593: PetscMalloc2(m,&diagIdx,m,&offdiagIdx);
2594: VecCreateSeq(PETSC_COMM_SELF, m, &diagV);
2595: VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV);
2596: MatGetRowMax(mat->A, diagV, diagIdx);
2598: /* Get offdiagIdx[] for implicit 0.0 */
2599: ba = b->a;
2600: bi = b->i;
2601: bj = b->j;
2602: VecGetArrayWrite(offdiagV, &offdiagA);
2603: for (r = 0; r < m; r++) {
2604: ncols = bi[r+1] - bi[r];
2605: if (ncols == A->cmap->N - n) { /* Brow is dense */
2606: offdiagA[r] = *ba; offdiagIdx[r] = cmap[0];
2607: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2608: offdiagA[r] = 0.0;
2610: /* Find first hole in the cmap */
2611: for (j=0; j<ncols; j++) {
2612: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2613: if (col > j && j < cstart) {
2614: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2615: break;
2616: } else if (col > j + n && j >= cstart) {
2617: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2618: break;
2619: }
2620: }
2621: if (j == ncols && ncols < A->cmap->N - n) {
2622: /* a hole is outside compressed Bcols */
2623: if (ncols == 0) {
2624: if (cstart) {
2625: offdiagIdx[r] = 0;
2626: } else offdiagIdx[r] = cend;
2627: } else { /* ncols > 0 */
2628: offdiagIdx[r] = cmap[ncols-1] + 1;
2629: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2630: }
2631: }
2632: }
2634: for (j=0; j<ncols; j++) {
2635: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {offdiagA[r] = *ba; offdiagIdx[r] = cmap[*bj];}
2636: ba++; bj++;
2637: }
2638: }
2640: VecGetArrayWrite(v, &a);
2641: VecGetArrayRead(diagV,(const PetscScalar**)&diagA);
2642: for (r = 0; r < m; ++r) {
2643: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2644: a[r] = diagA[r];
2645: if (idx) idx[r] = cstart + diagIdx[r];
2646: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2647: a[r] = diagA[r];
2648: if (idx) {
2649: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2650: idx[r] = cstart + diagIdx[r];
2651: } else idx[r] = offdiagIdx[r];
2652: }
2653: } else {
2654: a[r] = offdiagA[r];
2655: if (idx) idx[r] = offdiagIdx[r];
2656: }
2657: }
2658: VecRestoreArrayWrite(v, &a);
2659: VecRestoreArrayRead(diagV, (const PetscScalar**)&diagA);
2660: VecRestoreArrayWrite(offdiagV,&offdiagA);
2661: VecDestroy(&diagV);
2662: VecDestroy(&offdiagV);
2663: PetscFree2(diagIdx, offdiagIdx);
2664: return(0);
2665: }
2667: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2668: {
2670: Mat *dummy;
2673: MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2674: *newmat = *dummy;
2675: PetscFree(dummy);
2676: return(0);
2677: }
2679: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2680: {
2681: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data;
2685: MatInvertBlockDiagonal(a->A,values);
2686: A->factorerrortype = a->A->factorerrortype;
2687: return(0);
2688: }
2690: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2691: {
2693: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data;
2696: if (!x->assembled && !x->preallocated) SETERRQ(PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2697: MatSetRandom(aij->A,rctx);
2698: if (x->assembled) {
2699: MatSetRandom(aij->B,rctx);
2700: } else {
2701: MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B,x->cmap->rstart,x->cmap->rend,rctx);
2702: }
2703: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2704: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2705: return(0);
2706: }
2708: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2709: {
2711: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2712: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2713: return(0);
2714: }
2716: /*@
2717: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2719: Collective on Mat
2721: Input Parameters:
2722: + A - the matrix
2723: - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)
2725: Level: advanced
2727: @*/
2728: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2729: {
2730: PetscErrorCode ierr;
2733: PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2734: return(0);
2735: }
2737: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2738: {
2739: PetscErrorCode ierr;
2740: PetscBool sc = PETSC_FALSE,flg;
2743: PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2744: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2745: PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2746: if (flg) {
2747: MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2748: }
2749: PetscOptionsTail();
2750: return(0);
2751: }
2753: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2754: {
2756: Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data;
2757: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data;
2760: if (!Y->preallocated) {
2761: MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2762: } else if (!aij->nz) {
2763: PetscInt nonew = aij->nonew;
2764: MatSeqAIJSetPreallocation(maij->A,1,NULL);
2765: aij->nonew = nonew;
2766: }
2767: MatShift_Basic(Y,a);
2768: return(0);
2769: }
2771: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d)
2772: {
2773: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2777: if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2778: MatMissingDiagonal(a->A,missing,d);
2779: if (d) {
2780: PetscInt rstart;
2781: MatGetOwnershipRange(A,&rstart,NULL);
2782: *d += rstart;
2784: }
2785: return(0);
2786: }
2788: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
2789: {
2790: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2794: MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);
2795: return(0);
2796: }
2798: /* -------------------------------------------------------------------*/
2799: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2800: MatGetRow_MPIAIJ,
2801: MatRestoreRow_MPIAIJ,
2802: MatMult_MPIAIJ,
2803: /* 4*/ MatMultAdd_MPIAIJ,
2804: MatMultTranspose_MPIAIJ,
2805: MatMultTransposeAdd_MPIAIJ,
2806: NULL,
2807: NULL,
2808: NULL,
2809: /*10*/ NULL,
2810: NULL,
2811: NULL,
2812: MatSOR_MPIAIJ,
2813: MatTranspose_MPIAIJ,
2814: /*15*/ MatGetInfo_MPIAIJ,
2815: MatEqual_MPIAIJ,
2816: MatGetDiagonal_MPIAIJ,
2817: MatDiagonalScale_MPIAIJ,
2818: MatNorm_MPIAIJ,
2819: /*20*/ MatAssemblyBegin_MPIAIJ,
2820: MatAssemblyEnd_MPIAIJ,
2821: MatSetOption_MPIAIJ,
2822: MatZeroEntries_MPIAIJ,
2823: /*24*/ MatZeroRows_MPIAIJ,
2824: NULL,
2825: NULL,
2826: NULL,
2827: NULL,
2828: /*29*/ MatSetUp_MPIAIJ,
2829: NULL,
2830: NULL,
2831: MatGetDiagonalBlock_MPIAIJ,
2832: NULL,
2833: /*34*/ MatDuplicate_MPIAIJ,
2834: NULL,
2835: NULL,
2836: NULL,
2837: NULL,
2838: /*39*/ MatAXPY_MPIAIJ,
2839: MatCreateSubMatrices_MPIAIJ,
2840: MatIncreaseOverlap_MPIAIJ,
2841: MatGetValues_MPIAIJ,
2842: MatCopy_MPIAIJ,
2843: /*44*/ MatGetRowMax_MPIAIJ,
2844: MatScale_MPIAIJ,
2845: MatShift_MPIAIJ,
2846: MatDiagonalSet_MPIAIJ,
2847: MatZeroRowsColumns_MPIAIJ,
2848: /*49*/ MatSetRandom_MPIAIJ,
2849: NULL,
2850: NULL,
2851: NULL,
2852: NULL,
2853: /*54*/ MatFDColoringCreate_MPIXAIJ,
2854: NULL,
2855: MatSetUnfactored_MPIAIJ,
2856: MatPermute_MPIAIJ,
2857: NULL,
2858: /*59*/ MatCreateSubMatrix_MPIAIJ,
2859: MatDestroy_MPIAIJ,
2860: MatView_MPIAIJ,
2861: NULL,
2862: NULL,
2863: /*64*/ NULL,
2864: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2865: NULL,
2866: NULL,
2867: NULL,
2868: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2869: MatGetRowMinAbs_MPIAIJ,
2870: NULL,
2871: NULL,
2872: NULL,
2873: NULL,
2874: /*75*/ MatFDColoringApply_AIJ,
2875: MatSetFromOptions_MPIAIJ,
2876: NULL,
2877: NULL,
2878: MatFindZeroDiagonals_MPIAIJ,
2879: /*80*/ NULL,
2880: NULL,
2881: NULL,
2882: /*83*/ MatLoad_MPIAIJ,
2883: MatIsSymmetric_MPIAIJ,
2884: NULL,
2885: NULL,
2886: NULL,
2887: NULL,
2888: /*89*/ NULL,
2889: NULL,
2890: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2891: NULL,
2892: NULL,
2893: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2894: NULL,
2895: NULL,
2896: NULL,
2897: MatBindToCPU_MPIAIJ,
2898: /*99*/ MatProductSetFromOptions_MPIAIJ,
2899: NULL,
2900: NULL,
2901: MatConjugate_MPIAIJ,
2902: NULL,
2903: /*104*/MatSetValuesRow_MPIAIJ,
2904: MatRealPart_MPIAIJ,
2905: MatImaginaryPart_MPIAIJ,
2906: NULL,
2907: NULL,
2908: /*109*/NULL,
2909: NULL,
2910: MatGetRowMin_MPIAIJ,
2911: NULL,
2912: MatMissingDiagonal_MPIAIJ,
2913: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2914: NULL,
2915: MatGetGhosts_MPIAIJ,
2916: NULL,
2917: NULL,
2918: /*119*/NULL,
2919: NULL,
2920: NULL,
2921: NULL,
2922: MatGetMultiProcBlock_MPIAIJ,
2923: /*124*/MatFindNonzeroRows_MPIAIJ,
2924: MatGetColumnNorms_MPIAIJ,
2925: MatInvertBlockDiagonal_MPIAIJ,
2926: MatInvertVariableBlockDiagonal_MPIAIJ,
2927: MatCreateSubMatricesMPI_MPIAIJ,
2928: /*129*/NULL,
2929: NULL,
2930: NULL,
2931: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2932: NULL,
2933: /*134*/NULL,
2934: NULL,
2935: NULL,
2936: NULL,
2937: NULL,
2938: /*139*/MatSetBlockSizes_MPIAIJ,
2939: NULL,
2940: NULL,
2941: MatFDColoringSetUp_MPIXAIJ,
2942: MatFindOffBlockDiagonalEntries_MPIAIJ,
2943: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2944: /*145*/NULL,
2945: NULL,
2946: NULL
2947: };
2949: /* ----------------------------------------------------------------------------------------*/
2951: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2952: {
2953: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2957: MatStoreValues(aij->A);
2958: MatStoreValues(aij->B);
2959: return(0);
2960: }
2962: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2963: {
2964: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2968: MatRetrieveValues(aij->A);
2969: MatRetrieveValues(aij->B);
2970: return(0);
2971: }
2973: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2974: {
2975: Mat_MPIAIJ *b;
2977: PetscMPIInt size;
2980: PetscLayoutSetUp(B->rmap);
2981: PetscLayoutSetUp(B->cmap);
2982: b = (Mat_MPIAIJ*)B->data;
2984: #if defined(PETSC_USE_CTABLE)
2985: PetscTableDestroy(&b->colmap);
2986: #else
2987: PetscFree(b->colmap);
2988: #endif
2989: PetscFree(b->garray);
2990: VecDestroy(&b->lvec);
2991: VecScatterDestroy(&b->Mvctx);
2993: /* Because the B will have been resized we simply destroy it and create a new one each time */
2994: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
2995: MatDestroy(&b->B);
2996: MatCreate(PETSC_COMM_SELF,&b->B);
2997: MatSetSizes(b->B,B->rmap->n,size > 1 ? B->cmap->N : 0,B->rmap->n,size > 1 ? B->cmap->N : 0);
2998: MatSetBlockSizesFromMats(b->B,B,B);
2999: MatSetType(b->B,MATSEQAIJ);
3000: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
3002: if (!B->preallocated) {
3003: MatCreate(PETSC_COMM_SELF,&b->A);
3004: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
3005: MatSetBlockSizesFromMats(b->A,B,B);
3006: MatSetType(b->A,MATSEQAIJ);
3007: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
3008: }
3010: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
3011: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
3012: B->preallocated = PETSC_TRUE;
3013: B->was_assembled = PETSC_FALSE;
3014: B->assembled = PETSC_FALSE;
3015: return(0);
3016: }
3018: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
3019: {
3020: Mat_MPIAIJ *b;
3025: PetscLayoutSetUp(B->rmap);
3026: PetscLayoutSetUp(B->cmap);
3027: b = (Mat_MPIAIJ*)B->data;
3029: #if defined(PETSC_USE_CTABLE)
3030: PetscTableDestroy(&b->colmap);
3031: #else
3032: PetscFree(b->colmap);
3033: #endif
3034: PetscFree(b->garray);
3035: VecDestroy(&b->lvec);
3036: VecScatterDestroy(&b->Mvctx);
3038: MatResetPreallocation(b->A);
3039: MatResetPreallocation(b->B);
3040: B->preallocated = PETSC_TRUE;
3041: B->was_assembled = PETSC_FALSE;
3042: B->assembled = PETSC_FALSE;
3043: return(0);
3044: }
3046: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3047: {
3048: Mat mat;
3049: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
3053: *newmat = NULL;
3054: MatCreate(PetscObjectComm((PetscObject)matin),&mat);
3055: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3056: MatSetBlockSizesFromMats(mat,matin,matin);
3057: MatSetType(mat,((PetscObject)matin)->type_name);
3058: a = (Mat_MPIAIJ*)mat->data;
3060: mat->factortype = matin->factortype;
3061: mat->assembled = matin->assembled;
3062: mat->insertmode = NOT_SET_VALUES;
3063: mat->preallocated = matin->preallocated;
3065: a->size = oldmat->size;
3066: a->rank = oldmat->rank;
3067: a->donotstash = oldmat->donotstash;
3068: a->roworiented = oldmat->roworiented;
3069: a->rowindices = NULL;
3070: a->rowvalues = NULL;
3071: a->getrowactive = PETSC_FALSE;
3073: PetscLayoutReference(matin->rmap,&mat->rmap);
3074: PetscLayoutReference(matin->cmap,&mat->cmap);
3076: if (oldmat->colmap) {
3077: #if defined(PETSC_USE_CTABLE)
3078: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3079: #else
3080: PetscMalloc1(mat->cmap->N,&a->colmap);
3081: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
3082: PetscArraycpy(a->colmap,oldmat->colmap,mat->cmap->N);
3083: #endif
3084: } else a->colmap = NULL;
3085: if (oldmat->garray) {
3086: PetscInt len;
3087: len = oldmat->B->cmap->n;
3088: PetscMalloc1(len+1,&a->garray);
3089: PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
3090: if (len) { PetscArraycpy(a->garray,oldmat->garray,len); }
3091: } else a->garray = NULL;
3093: /* It may happen MatDuplicate is called with a non-assembled matrix
3094: In fact, MatDuplicate only requires the matrix to be preallocated
3095: This may happen inside a DMCreateMatrix_Shell */
3096: if (oldmat->lvec) {
3097: VecDuplicate(oldmat->lvec,&a->lvec);
3098: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
3099: }
3100: if (oldmat->Mvctx) {
3101: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3102: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
3103: }
3104: if (oldmat->Mvctx_mpi1) {
3105: VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);
3106: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);
3107: }
3109: MatDuplicate(oldmat->A,cpvalues,&a->A);
3110: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
3111: MatDuplicate(oldmat->B,cpvalues,&a->B);
3112: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
3113: PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3114: *newmat = mat;
3115: return(0);
3116: }
3118: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3119: {
3120: PetscBool isbinary, ishdf5;
3126: /* force binary viewer to load .info file if it has not yet done so */
3127: PetscViewerSetUp(viewer);
3128: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
3129: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);
3130: if (isbinary) {
3131: MatLoad_MPIAIJ_Binary(newMat,viewer);
3132: } else if (ishdf5) {
3133: #if defined(PETSC_HAVE_HDF5)
3134: MatLoad_AIJ_HDF5(newMat,viewer);
3135: #else
3136: SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3137: #endif
3138: } else {
3139: SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name);
3140: }
3141: return(0);
3142: }
3144: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3145: {
3146: PetscInt header[4],M,N,m,nz,rows,cols,sum,i;
3147: PetscInt *rowidxs,*colidxs;
3148: PetscScalar *matvals;
3152: PetscViewerSetUp(viewer);
3154: /* read in matrix header */
3155: PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);
3156: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file");
3157: M = header[1]; N = header[2]; nz = header[3];
3158: if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M);
3159: if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N);
3160: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as MPIAIJ");
3162: /* set block sizes from the viewer's .info file */
3163: MatLoad_Binary_BlockSizes(mat,viewer);
3164: /* set global sizes if not set already */
3165: if (mat->rmap->N < 0) mat->rmap->N = M;
3166: if (mat->cmap->N < 0) mat->cmap->N = N;
3167: PetscLayoutSetUp(mat->rmap);
3168: PetscLayoutSetUp(mat->cmap);
3170: /* check if the matrix sizes are correct */
3171: MatGetSize(mat,&rows,&cols);
3172: if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols);
3174: /* read in row lengths and build row indices */
3175: MatGetLocalSize(mat,&m,NULL);
3176: PetscMalloc1(m+1,&rowidxs);
3177: PetscViewerBinaryReadAll(viewer,rowidxs+1,m,PETSC_DECIDE,M,PETSC_INT);
3178: rowidxs[0] = 0; for (i=0; i<m; i++) rowidxs[i+1] += rowidxs[i];
3179: MPIU_Allreduce(&rowidxs[m],&sum,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)viewer));
3180: if (sum != nz) SETERRQ2(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Inconsistent matrix data in file: nonzeros = %D, sum-row-lengths = %D\n",nz,sum);
3181: /* read in column indices and matrix values */
3182: PetscMalloc2(rowidxs[m],&colidxs,rowidxs[m],&matvals);
3183: PetscViewerBinaryReadAll(viewer,colidxs,rowidxs[m],PETSC_DETERMINE,PETSC_DETERMINE,PETSC_INT);
3184: PetscViewerBinaryReadAll(viewer,matvals,rowidxs[m],PETSC_DETERMINE,PETSC_DETERMINE,PETSC_SCALAR);
3185: /* store matrix indices and values */
3186: MatMPIAIJSetPreallocationCSR(mat,rowidxs,colidxs,matvals);
3187: PetscFree(rowidxs);
3188: PetscFree2(colidxs,matvals);
3189: return(0);
3190: }
3192: /* Not scalable because of ISAllGather() unless getting all columns. */
3193: PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq)
3194: {
3196: IS iscol_local;
3197: PetscBool isstride;
3198: PetscMPIInt lisstride=0,gisstride;
3201: /* check if we are grabbing all columns*/
3202: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);
3204: if (isstride) {
3205: PetscInt start,len,mstart,mlen;
3206: ISStrideGetInfo(iscol,&start,NULL);
3207: ISGetLocalSize(iscol,&len);
3208: MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
3209: if (mstart == start && mlen-mstart == len) lisstride = 1;
3210: }
3212: MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3213: if (gisstride) {
3214: PetscInt N;
3215: MatGetSize(mat,NULL,&N);
3216: ISCreateStride(PETSC_COMM_SELF,N,0,1,&iscol_local);
3217: ISSetIdentity(iscol_local);
3218: PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3219: } else {
3220: PetscInt cbs;
3221: ISGetBlockSize(iscol,&cbs);
3222: ISAllGather(iscol,&iscol_local);
3223: ISSetBlockSize(iscol_local,cbs);
3224: }
3226: *isseq = iscol_local;
3227: return(0);
3228: }
3230: /*
3231: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3232: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3234: Input Parameters:
3235: mat - matrix
3236: isrow - parallel row index set; its local indices are a subset of local columns of mat,
3237: i.e., mat->rstart <= isrow[i] < mat->rend
3238: iscol - parallel column index set; its local indices are a subset of local columns of mat,
3239: i.e., mat->cstart <= iscol[i] < mat->cend
3240: Output Parameter:
3241: isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A
3242: iscol_o - sequential column index set for retrieving mat->B
3243: garray - column map; garray[i] indicates global location of iscol_o[i] in iscol
3244: */
3245: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[])
3246: {
3248: Vec x,cmap;
3249: const PetscInt *is_idx;
3250: PetscScalar *xarray,*cmaparray;
3251: PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count;
3252: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3253: Mat B=a->B;
3254: Vec lvec=a->lvec,lcmap;
3255: PetscInt i,cstart,cend,Bn=B->cmap->N;
3256: MPI_Comm comm;
3257: VecScatter Mvctx=a->Mvctx;
3260: PetscObjectGetComm((PetscObject)mat,&comm);
3261: ISGetLocalSize(iscol,&ncols);
3263: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3264: MatCreateVecs(mat,&x,NULL);
3265: VecSet(x,-1.0);
3266: VecDuplicate(x,&cmap);
3267: VecSet(cmap,-1.0);
3269: /* Get start indices */
3270: MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);
3271: isstart -= ncols;
3272: MatGetOwnershipRangeColumn(mat,&cstart,&cend);
3274: ISGetIndices(iscol,&is_idx);
3275: VecGetArray(x,&xarray);
3276: VecGetArray(cmap,&cmaparray);
3277: PetscMalloc1(ncols,&idx);
3278: for (i=0; i<ncols; i++) {
3279: xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i];
3280: cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */
3281: idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */
3282: }
3283: VecRestoreArray(x,&xarray);
3284: VecRestoreArray(cmap,&cmaparray);
3285: ISRestoreIndices(iscol,&is_idx);
3287: /* Get iscol_d */
3288: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);
3289: ISGetBlockSize(iscol,&i);
3290: ISSetBlockSize(*iscol_d,i);
3292: /* Get isrow_d */
3293: ISGetLocalSize(isrow,&m);
3294: rstart = mat->rmap->rstart;
3295: PetscMalloc1(m,&idx);
3296: ISGetIndices(isrow,&is_idx);
3297: for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart;
3298: ISRestoreIndices(isrow,&is_idx);
3300: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);
3301: ISGetBlockSize(isrow,&i);
3302: ISSetBlockSize(*isrow_d,i);
3304: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3305: VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3306: VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3308: VecDuplicate(lvec,&lcmap);
3310: VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3311: VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3313: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3314: /* off-process column indices */
3315: count = 0;
3316: PetscMalloc1(Bn,&idx);
3317: PetscMalloc1(Bn,&cmap1);
3319: VecGetArray(lvec,&xarray);
3320: VecGetArray(lcmap,&cmaparray);
3321: for (i=0; i<Bn; i++) {
3322: if (PetscRealPart(xarray[i]) > -1.0) {
3323: idx[count] = i; /* local column index in off-diagonal part B */
3324: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3325: count++;
3326: }
3327: }
3328: VecRestoreArray(lvec,&xarray);
3329: VecRestoreArray(lcmap,&cmaparray);
3331: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);
3332: /* cannot ensure iscol_o has same blocksize as iscol! */
3334: PetscFree(idx);
3335: *garray = cmap1;
3337: VecDestroy(&x);
3338: VecDestroy(&cmap);
3339: VecDestroy(&lcmap);
3340: return(0);
3341: }
3343: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3344: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat)
3345: {
3347: Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub;
3348: Mat M = NULL;
3349: MPI_Comm comm;
3350: IS iscol_d,isrow_d,iscol_o;
3351: Mat Asub = NULL,Bsub = NULL;
3352: PetscInt n;
3355: PetscObjectGetComm((PetscObject)mat,&comm);
3357: if (call == MAT_REUSE_MATRIX) {
3358: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3359: PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);
3360: if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse");
3362: PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);
3363: if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse");
3365: PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);
3366: if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse");
3368: /* Update diagonal and off-diagonal portions of submat */
3369: asub = (Mat_MPIAIJ*)(*submat)->data;
3370: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);
3371: ISGetLocalSize(iscol_o,&n);
3372: if (n) {
3373: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);
3374: }
3375: MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3376: MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);
3378: } else { /* call == MAT_INITIAL_MATRIX) */
3379: const PetscInt *garray;
3380: PetscInt BsubN;
3382: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3383: ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);
3385: /* Create local submatrices Asub and Bsub */
3386: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);
3387: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);
3389: /* Create submatrix M */
3390: MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);
3392: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3393: asub = (Mat_MPIAIJ*)M->data;
3395: ISGetLocalSize(iscol_o,&BsubN);
3396: n = asub->B->cmap->N;
3397: if (BsubN > n) {
3398: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3399: const PetscInt *idx;
3400: PetscInt i,j,*idx_new,*subgarray = asub->garray;
3401: PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);
3403: PetscMalloc1(n,&idx_new);
3404: j = 0;
3405: ISGetIndices(iscol_o,&idx);
3406: for (i=0; i<n; i++) {
3407: if (j >= BsubN) break;
3408: while (subgarray[i] > garray[j]) j++;
3410: if (subgarray[i] == garray[j]) {
3411: idx_new[i] = idx[j++];
3412: } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]);
3413: }
3414: ISRestoreIndices(iscol_o,&idx);
3416: ISDestroy(&iscol_o);
3417: ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);
3419: } else if (BsubN < n) {
3420: SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N);
3421: }
3423: PetscFree(garray);
3424: *submat = M;
3426: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3427: PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);
3428: ISDestroy(&isrow_d);
3430: PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);
3431: ISDestroy(&iscol_d);
3433: PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);
3434: ISDestroy(&iscol_o);
3435: }
3436: return(0);
3437: }
3439: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3440: {
3442: IS iscol_local=NULL,isrow_d;
3443: PetscInt csize;
3444: PetscInt n,i,j,start,end;
3445: PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2];
3446: MPI_Comm comm;
3449: /* If isrow has same processor distribution as mat,
3450: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3451: if (call == MAT_REUSE_MATRIX) {
3452: PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);
3453: if (isrow_d) {
3454: sameRowDist = PETSC_TRUE;
3455: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3456: } else {
3457: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);
3458: if (iscol_local) {
3459: sameRowDist = PETSC_TRUE;
3460: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3461: }
3462: }
3463: } else {
3464: /* Check if isrow has same processor distribution as mat */
3465: sameDist[0] = PETSC_FALSE;
3466: ISGetLocalSize(isrow,&n);
3467: if (!n) {
3468: sameDist[0] = PETSC_TRUE;
3469: } else {
3470: ISGetMinMax(isrow,&i,&j);
3471: MatGetOwnershipRange(mat,&start,&end);
3472: if (i >= start && j < end) {
3473: sameDist[0] = PETSC_TRUE;
3474: }
3475: }
3477: /* Check if iscol has same processor distribution as mat */
3478: sameDist[1] = PETSC_FALSE;
3479: ISGetLocalSize(iscol,&n);
3480: if (!n) {
3481: sameDist[1] = PETSC_TRUE;
3482: } else {
3483: ISGetMinMax(iscol,&i,&j);
3484: MatGetOwnershipRangeColumn(mat,&start,&end);
3485: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3486: }
3488: PetscObjectGetComm((PetscObject)mat,&comm);
3489: MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);
3490: sameRowDist = tsameDist[0];
3491: }
3493: if (sameRowDist) {
3494: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3495: /* isrow and iscol have same processor distribution as mat */
3496: MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);
3497: return(0);
3498: } else { /* sameRowDist */
3499: /* isrow has same processor distribution as mat */
3500: if (call == MAT_INITIAL_MATRIX) {
3501: PetscBool sorted;
3502: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3503: ISGetLocalSize(iscol_local,&n); /* local size of iscol_local = global columns of newmat */
3504: ISGetSize(iscol,&i);
3505: if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i);
3507: ISSorted(iscol_local,&sorted);
3508: if (sorted) {
3509: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3510: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);
3511: return(0);
3512: }
3513: } else { /* call == MAT_REUSE_MATRIX */
3514: IS iscol_sub;
3515: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3516: if (iscol_sub) {
3517: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);
3518: return(0);
3519: }
3520: }
3521: }
3522: }
3524: /* General case: iscol -> iscol_local which has global size of iscol */
3525: if (call == MAT_REUSE_MATRIX) {
3526: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3527: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3528: } else {
3529: if (!iscol_local) {
3530: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3531: }
3532: }
3534: ISGetLocalSize(iscol,&csize);
3535: MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);
3537: if (call == MAT_INITIAL_MATRIX) {
3538: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3539: ISDestroy(&iscol_local);
3540: }
3541: return(0);
3542: }
3544: /*@C
3545: MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal"
3546: and "off-diagonal" part of the matrix in CSR format.
3548: Collective
3550: Input Parameters:
3551: + comm - MPI communicator
3552: . A - "diagonal" portion of matrix
3553: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3554: - garray - global index of B columns
3556: Output Parameter:
3557: . mat - the matrix, with input A as its local diagonal matrix
3558: Level: advanced
3560: Notes:
3561: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3562: A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore.
3564: .seealso: MatCreateMPIAIJWithSplitArrays()
3565: @*/
3566: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat)
3567: {
3569: Mat_MPIAIJ *maij;
3570: Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew;
3571: PetscInt *oi=b->i,*oj=b->j,i,nz,col;
3572: PetscScalar *oa=b->a;
3573: Mat Bnew;
3574: PetscInt m,n,N;
3577: MatCreate(comm,mat);
3578: MatGetSize(A,&m,&n);
3579: if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N);
3580: if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs);
3581: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3582: /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */
3584: /* Get global columns of mat */
3585: MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);
3587: MatSetSizes(*mat,m,n,PETSC_DECIDE,N);
3588: MatSetType(*mat,MATMPIAIJ);
3589: MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);
3590: maij = (Mat_MPIAIJ*)(*mat)->data;
3592: (*mat)->preallocated = PETSC_TRUE;
3594: PetscLayoutSetUp((*mat)->rmap);
3595: PetscLayoutSetUp((*mat)->cmap);
3597: /* Set A as diagonal portion of *mat */
3598: maij->A = A;
3600: nz = oi[m];
3601: for (i=0; i<nz; i++) {
3602: col = oj[i];
3603: oj[i] = garray[col];
3604: }
3606: /* Set Bnew as off-diagonal portion of *mat */
3607: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);
3608: bnew = (Mat_SeqAIJ*)Bnew->data;
3609: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3610: maij->B = Bnew;
3612: if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N);
3614: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3615: b->free_a = PETSC_FALSE;
3616: b->free_ij = PETSC_FALSE;
3617: MatDestroy(&B);
3619: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3620: bnew->free_a = PETSC_TRUE;
3621: bnew->free_ij = PETSC_TRUE;
3623: /* condense columns of maij->B */
3624: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
3625: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3626: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3627: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
3628: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3629: return(0);
3630: }
3632: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*);
3634: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat)
3635: {
3637: PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs;
3638: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3639: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3640: Mat M,Msub,B=a->B;
3641: MatScalar *aa;
3642: Mat_SeqAIJ *aij;
3643: PetscInt *garray = a->garray,*colsub,Ncols;
3644: PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend;
3645: IS iscol_sub,iscmap;
3646: const PetscInt *is_idx,*cmap;
3647: PetscBool allcolumns=PETSC_FALSE;
3648: MPI_Comm comm;
3651: PetscObjectGetComm((PetscObject)mat,&comm);
3653: if (call == MAT_REUSE_MATRIX) {
3654: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3655: if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse");
3656: ISGetLocalSize(iscol_sub,&count);
3658: PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);
3659: if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse");
3661: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);
3662: if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3664: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);
3666: } else { /* call == MAT_INITIAL_MATRIX) */
3667: PetscBool flg;
3669: ISGetLocalSize(iscol,&n);
3670: ISGetSize(iscol,&Ncols);
3672: /* (1) iscol -> nonscalable iscol_local */
3673: /* Check for special case: each processor gets entire matrix columns */
3674: ISIdentity(iscol_local,&flg);
3675: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3676: MPIU_Allreduce(MPI_IN_PLACE,&allcolumns,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));
3677: if (allcolumns) {
3678: iscol_sub = iscol_local;
3679: PetscObjectReference((PetscObject)iscol_local);
3680: ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);
3682: } else {
3683: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3684: PetscInt *idx,*cmap1,k;
3685: PetscMalloc1(Ncols,&idx);
3686: PetscMalloc1(Ncols,&cmap1);
3687: ISGetIndices(iscol_local,&is_idx);
3688: count = 0;
3689: k = 0;
3690: for (i=0; i<Ncols; i++) {
3691: j = is_idx[i];
3692: if (j >= cstart && j < cend) {
3693: /* diagonal part of mat */
3694: idx[count] = j;
3695: cmap1[count++] = i; /* column index in submat */
3696: } else if (Bn) {
3697: /* off-diagonal part of mat */
3698: if (j == garray[k]) {
3699: idx[count] = j;
3700: cmap1[count++] = i; /* column index in submat */
3701: } else if (j > garray[k]) {
3702: while (j > garray[k] && k < Bn-1) k++;
3703: if (j == garray[k]) {
3704: idx[count] = j;
3705: cmap1[count++] = i; /* column index in submat */
3706: }
3707: }
3708: }
3709: }
3710: ISRestoreIndices(iscol_local,&is_idx);
3712: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);
3713: ISGetBlockSize(iscol,&cbs);
3714: ISSetBlockSize(iscol_sub,cbs);
3716: ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);
3717: }
3719: /* (3) Create sequential Msub */
3720: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);
3721: }
3723: ISGetLocalSize(iscol_sub,&count);
3724: aij = (Mat_SeqAIJ*)(Msub)->data;
3725: ii = aij->i;
3726: ISGetIndices(iscmap,&cmap);
3728: /*
3729: m - number of local rows
3730: Ncols - number of columns (same on all processors)
3731: rstart - first row in new global matrix generated
3732: */
3733: MatGetSize(Msub,&m,NULL);
3735: if (call == MAT_INITIAL_MATRIX) {
3736: /* (4) Create parallel newmat */
3737: PetscMPIInt rank,size;
3738: PetscInt csize;
3740: MPI_Comm_size(comm,&size);
3741: MPI_Comm_rank(comm,&rank);
3743: /*
3744: Determine the number of non-zeros in the diagonal and off-diagonal
3745: portions of the matrix in order to do correct preallocation
3746: */
3748: /* first get start and end of "diagonal" columns */
3749: ISGetLocalSize(iscol,&csize);
3750: if (csize == PETSC_DECIDE) {
3751: ISGetSize(isrow,&mglobal);
3752: if (mglobal == Ncols) { /* square matrix */
3753: nlocal = m;
3754: } else {
3755: nlocal = Ncols/size + ((Ncols % size) > rank);
3756: }
3757: } else {
3758: nlocal = csize;
3759: }
3760: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3761: rstart = rend - nlocal;
3762: if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols);
3764: /* next, compute all the lengths */
3765: jj = aij->j;
3766: PetscMalloc1(2*m+1,&dlens);
3767: olens = dlens + m;
3768: for (i=0; i<m; i++) {
3769: jend = ii[i+1] - ii[i];
3770: olen = 0;
3771: dlen = 0;
3772: for (j=0; j<jend; j++) {
3773: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3774: else dlen++;
3775: jj++;
3776: }
3777: olens[i] = olen;
3778: dlens[i] = dlen;
3779: }
3781: ISGetBlockSize(isrow,&bs);
3782: ISGetBlockSize(iscol,&cbs);
3784: MatCreate(comm,&M);
3785: MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);
3786: MatSetBlockSizes(M,bs,cbs);
3787: MatSetType(M,((PetscObject)mat)->type_name);
3788: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3789: PetscFree(dlens);
3791: } else { /* call == MAT_REUSE_MATRIX */
3792: M = *newmat;
3793: MatGetLocalSize(M,&i,NULL);
3794: if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3795: MatZeroEntries(M);
3796: /*
3797: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3798: rather than the slower MatSetValues().
3799: */
3800: M->was_assembled = PETSC_TRUE;
3801: M->assembled = PETSC_FALSE;
3802: }
3804: /* (5) Set values of Msub to *newmat */
3805: PetscMalloc1(count,&colsub);
3806: MatGetOwnershipRange(M,&rstart,NULL);
3808: jj = aij->j;
3809: aa = aij->a;
3810: for (i=0; i<m; i++) {
3811: row = rstart + i;
3812: nz = ii[i+1] - ii[i];
3813: for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]];
3814: MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);
3815: jj += nz; aa += nz;
3816: }
3817: ISRestoreIndices(iscmap,&cmap);
3819: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3820: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3822: PetscFree(colsub);
3824: /* save Msub, iscol_sub and iscmap used in processor for next request */
3825: if (call == MAT_INITIAL_MATRIX) {
3826: *newmat = M;
3827: PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);
3828: MatDestroy(&Msub);
3830: PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);
3831: ISDestroy(&iscol_sub);
3833: PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);
3834: ISDestroy(&iscmap);
3836: if (iscol_local) {
3837: PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);
3838: ISDestroy(&iscol_local);
3839: }
3840: }
3841: return(0);
3842: }
3844: /*
3845: Not great since it makes two copies of the submatrix, first an SeqAIJ
3846: in local and then by concatenating the local matrices the end result.
3847: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3849: Note: This requires a sequential iscol with all indices.
3850: */
3851: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3852: {
3854: PetscMPIInt rank,size;
3855: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3856: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3857: Mat M,Mreuse;
3858: MatScalar *aa,*vwork;
3859: MPI_Comm comm;
3860: Mat_SeqAIJ *aij;
3861: PetscBool colflag,allcolumns=PETSC_FALSE;
3864: PetscObjectGetComm((PetscObject)mat,&comm);
3865: MPI_Comm_rank(comm,&rank);
3866: MPI_Comm_size(comm,&size);
3868: /* Check for special case: each processor gets entire matrix columns */
3869: ISIdentity(iscol,&colflag);
3870: ISGetLocalSize(iscol,&n);
3871: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3872: MPIU_Allreduce(MPI_IN_PLACE,&allcolumns,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));
3874: if (call == MAT_REUSE_MATRIX) {
3875: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3876: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3877: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);
3878: } else {
3879: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);
3880: }
3882: /*
3883: m - number of local rows
3884: n - number of columns (same on all processors)
3885: rstart - first row in new global matrix generated
3886: */
3887: MatGetSize(Mreuse,&m,&n);
3888: MatGetBlockSizes(Mreuse,&bs,&cbs);
3889: if (call == MAT_INITIAL_MATRIX) {
3890: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3891: ii = aij->i;
3892: jj = aij->j;
3894: /*
3895: Determine the number of non-zeros in the diagonal and off-diagonal
3896: portions of the matrix in order to do correct preallocation
3897: */
3899: /* first get start and end of "diagonal" columns */
3900: if (csize == PETSC_DECIDE) {
3901: ISGetSize(isrow,&mglobal);
3902: if (mglobal == n) { /* square matrix */
3903: nlocal = m;
3904: } else {
3905: nlocal = n/size + ((n % size) > rank);
3906: }
3907: } else {
3908: nlocal = csize;
3909: }
3910: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3911: rstart = rend - nlocal;
3912: if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
3914: /* next, compute all the lengths */
3915: PetscMalloc1(2*m+1,&dlens);
3916: olens = dlens + m;
3917: for (i=0; i<m; i++) {
3918: jend = ii[i+1] - ii[i];
3919: olen = 0;
3920: dlen = 0;
3921: for (j=0; j<jend; j++) {
3922: if (*jj < rstart || *jj >= rend) olen++;
3923: else dlen++;
3924: jj++;
3925: }
3926: olens[i] = olen;
3927: dlens[i] = dlen;
3928: }
3929: MatCreate(comm,&M);
3930: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3931: MatSetBlockSizes(M,bs,cbs);
3932: MatSetType(M,((PetscObject)mat)->type_name);
3933: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3934: PetscFree(dlens);
3935: } else {
3936: PetscInt ml,nl;
3938: M = *newmat;
3939: MatGetLocalSize(M,&ml,&nl);
3940: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3941: MatZeroEntries(M);
3942: /*
3943: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3944: rather than the slower MatSetValues().
3945: */
3946: M->was_assembled = PETSC_TRUE;
3947: M->assembled = PETSC_FALSE;
3948: }
3949: MatGetOwnershipRange(M,&rstart,&rend);
3950: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3951: ii = aij->i;
3952: jj = aij->j;
3953: aa = aij->a;
3954: for (i=0; i<m; i++) {
3955: row = rstart + i;
3956: nz = ii[i+1] - ii[i];
3957: cwork = jj; jj += nz;
3958: vwork = aa; aa += nz;
3959: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3960: }
3962: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3963: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3964: *newmat = M;
3966: /* save submatrix used in processor for next request */
3967: if (call == MAT_INITIAL_MATRIX) {
3968: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3969: MatDestroy(&Mreuse);
3970: }
3971: return(0);
3972: }
3974: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3975: {
3976: PetscInt m,cstart, cend,j,nnz,i,d;
3977: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3978: const PetscInt *JJ;
3980: PetscBool nooffprocentries;
3983: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
3985: PetscLayoutSetUp(B->rmap);
3986: PetscLayoutSetUp(B->cmap);
3987: m = B->rmap->n;
3988: cstart = B->cmap->rstart;
3989: cend = B->cmap->rend;
3990: rstart = B->rmap->rstart;
3992: PetscCalloc2(m,&d_nnz,m,&o_nnz);
3994: if (PetscDefined(USE_DEBUG)) {
3995: for (i=0; i<m; i++) {
3996: nnz = Ii[i+1]- Ii[i];
3997: JJ = J + Ii[i];
3998: if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3999: if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]);
4000: if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
4001: }
4002: }
4004: for (i=0; i<m; i++) {
4005: nnz = Ii[i+1]- Ii[i];
4006: JJ = J + Ii[i];
4007: nnz_max = PetscMax(nnz_max,nnz);
4008: d = 0;
4009: for (j=0; j<nnz; j++) {
4010: if (cstart <= JJ[j] && JJ[j] < cend) d++;
4011: }
4012: d_nnz[i] = d;
4013: o_nnz[i] = nnz - d;
4014: }
4015: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
4016: PetscFree2(d_nnz,o_nnz);
4018: for (i=0; i<m; i++) {
4019: ii = i + rstart;
4020: MatSetValues_MPIAIJ(B,1,&ii,Ii[i+1] - Ii[i],J+Ii[i], v ? v + Ii[i] : NULL,INSERT_VALUES);
4021: }
4022: nooffprocentries = B->nooffprocentries;
4023: B->nooffprocentries = PETSC_TRUE;
4024: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4025: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4026: B->nooffprocentries = nooffprocentries;
4028: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
4029: return(0);
4030: }
4032: /*@
4033: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
4034: (the default parallel PETSc format).
4036: Collective
4038: Input Parameters:
4039: + B - the matrix
4040: . i - the indices into j for the start of each local row (starts with zero)
4041: . j - the column indices for each local row (starts with zero)
4042: - v - optional values in the matrix
4044: Level: developer
4046: Notes:
4047: The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc;
4048: thus you CANNOT change the matrix entries by changing the values of v[] after you have
4049: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4051: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4053: The format which is used for the sparse matrix input, is equivalent to a
4054: row-major ordering.. i.e for the following matrix, the input data expected is
4055: as shown
4057: $ 1 0 0
4058: $ 2 0 3 P0
4059: $ -------
4060: $ 4 5 6 P1
4061: $
4062: $ Process0 [P0]: rows_owned=[0,1]
4063: $ i = {0,1,3} [size = nrow+1 = 2+1]
4064: $ j = {0,0,2} [size = 3]
4065: $ v = {1,2,3} [size = 3]
4066: $
4067: $ Process1 [P1]: rows_owned=[2]
4068: $ i = {0,3} [size = nrow+1 = 1+1]
4069: $ j = {0,1,2} [size = 3]
4070: $ v = {4,5,6} [size = 3]
4072: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ,
4073: MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
4074: @*/
4075: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
4076: {
4080: PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4081: return(0);
4082: }
4084: /*@C
4085: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
4086: (the default parallel PETSc format). For good matrix assembly performance
4087: the user should preallocate the matrix storage by setting the parameters
4088: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4089: performance can be increased by more than a factor of 50.
4091: Collective
4093: Input Parameters:
4094: + B - the matrix
4095: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4096: (same value is used for all local rows)
4097: . d_nnz - array containing the number of nonzeros in the various rows of the
4098: DIAGONAL portion of the local submatrix (possibly different for each row)
4099: or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
4100: The size of this array is equal to the number of local rows, i.e 'm'.
4101: For matrices that will be factored, you must leave room for (and set)
4102: the diagonal entry even if it is zero.
4103: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4104: submatrix (same value is used for all local rows).
4105: - o_nnz - array containing the number of nonzeros in the various rows of the
4106: OFF-DIAGONAL portion of the local submatrix (possibly different for
4107: each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
4108: structure. The size of this array is equal to the number
4109: of local rows, i.e 'm'.
4111: If the *_nnz parameter is given then the *_nz parameter is ignored
4113: The AIJ format (also called the Yale sparse matrix format or
4114: compressed row storage (CSR)), is fully compatible with standard Fortran 77
4115: storage. The stored row and column indices begin with zero.
4116: See Users-Manual: ch_mat for details.
4118: The parallel matrix is partitioned such that the first m0 rows belong to
4119: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4120: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4122: The DIAGONAL portion of the local submatrix of a processor can be defined
4123: as the submatrix which is obtained by extraction the part corresponding to
4124: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4125: first row that belongs to the processor, r2 is the last row belonging to
4126: the this processor, and c1-c2 is range of indices of the local part of a
4127: vector suitable for applying the matrix to. This is an mxn matrix. In the
4128: common case of a square matrix, the row and column ranges are the same and
4129: the DIAGONAL part is also square. The remaining portion of the local
4130: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4132: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4134: You can call MatGetInfo() to get information on how effective the preallocation was;
4135: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4136: You can also run with the option -info and look for messages with the string
4137: malloc in them to see if additional memory allocation was needed.
4139: Example usage:
4141: Consider the following 8x8 matrix with 34 non-zero values, that is
4142: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4143: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4144: as follows:
4146: .vb
4147: 1 2 0 | 0 3 0 | 0 4
4148: Proc0 0 5 6 | 7 0 0 | 8 0
4149: 9 0 10 | 11 0 0 | 12 0
4150: -------------------------------------
4151: 13 0 14 | 15 16 17 | 0 0
4152: Proc1 0 18 0 | 19 20 21 | 0 0
4153: 0 0 0 | 22 23 0 | 24 0
4154: -------------------------------------
4155: Proc2 25 26 27 | 0 0 28 | 29 0
4156: 30 0 0 | 31 32 33 | 0 34
4157: .ve
4159: This can be represented as a collection of submatrices as:
4161: .vb
4162: A B C
4163: D E F
4164: G H I
4165: .ve
4167: Where the submatrices A,B,C are owned by proc0, D,E,F are
4168: owned by proc1, G,H,I are owned by proc2.
4170: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4171: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4172: The 'M','N' parameters are 8,8, and have the same values on all procs.
4174: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4175: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4176: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4177: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4178: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4179: matrix, ans [DF] as another SeqAIJ matrix.
4181: When d_nz, o_nz parameters are specified, d_nz storage elements are
4182: allocated for every row of the local diagonal submatrix, and o_nz
4183: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4184: One way to choose d_nz and o_nz is to use the max nonzerors per local
4185: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4186: In this case, the values of d_nz,o_nz are:
4187: .vb
4188: proc0 : dnz = 2, o_nz = 2
4189: proc1 : dnz = 3, o_nz = 2
4190: proc2 : dnz = 1, o_nz = 4
4191: .ve
4192: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4193: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4194: for proc3. i.e we are using 12+15+10=37 storage locations to store
4195: 34 values.
4197: When d_nnz, o_nnz parameters are specified, the storage is specified
4198: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4199: In the above case the values for d_nnz,o_nnz are:
4200: .vb
4201: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4202: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4203: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4204: .ve
4205: Here the space allocated is sum of all the above values i.e 34, and
4206: hence pre-allocation is perfect.
4208: Level: intermediate
4210: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
4211: MATMPIAIJ, MatGetInfo(), PetscSplitOwnership()
4212: @*/
4213: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
4214: {
4220: PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
4221: return(0);
4222: }
4224: /*@
4225: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
4226: CSR format for the local rows.
4228: Collective
4230: Input Parameters:
4231: + comm - MPI communicator
4232: . m - number of local rows (Cannot be PETSC_DECIDE)
4233: . n - This value should be the same as the local size used in creating the
4234: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4235: calculated if N is given) For square matrices n is almost always m.
4236: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4237: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4238: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4239: . j - column indices
4240: - a - matrix values
4242: Output Parameter:
4243: . mat - the matrix
4245: Level: intermediate
4247: Notes:
4248: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4249: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4250: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4252: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4254: The format which is used for the sparse matrix input, is equivalent to a
4255: row-major ordering.. i.e for the following matrix, the input data expected is
4256: as shown
4258: Once you have created the matrix you can update it with new numerical values using MatUpdateMPIAIJWithArrays
4260: $ 1 0 0
4261: $ 2 0 3 P0
4262: $ -------
4263: $ 4 5 6 P1
4264: $
4265: $ Process0 [P0]: rows_owned=[0,1]
4266: $ i = {0,1,3} [size = nrow+1 = 2+1]
4267: $ j = {0,0,2} [size = 3]
4268: $ v = {1,2,3} [size = 3]
4269: $
4270: $ Process1 [P1]: rows_owned=[2]
4271: $ i = {0,3} [size = nrow+1 = 1+1]
4272: $ j = {0,1,2} [size = 3]
4273: $ v = {4,5,6} [size = 3]
4275: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4276: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4277: @*/
4278: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4279: {
4283: if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4284: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4285: MatCreate(comm,mat);
4286: MatSetSizes(*mat,m,n,M,N);
4287: /* MatSetBlockSizes(M,bs,cbs); */
4288: MatSetType(*mat,MATMPIAIJ);
4289: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4290: return(0);
4291: }
4293: /*@
4294: MatUpdateMPIAIJWithArrays - updates a MPI AIJ matrix using arrays that contain in standard
4295: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical
4297: Collective
4299: Input Parameters:
4300: + mat - the matrix
4301: . m - number of local rows (Cannot be PETSC_DECIDE)
4302: . n - This value should be the same as the local size used in creating the
4303: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4304: calculated if N is given) For square matrices n is almost always m.
4305: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4306: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4307: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4308: . J - column indices
4309: - v - matrix values
4311: Level: intermediate
4313: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4314: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4315: @*/
4316: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
4317: {
4319: PetscInt cstart,nnz,i,j;
4320: PetscInt *ld;
4321: PetscBool nooffprocentries;
4322: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*)mat->data;
4323: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)Aij->A->data, *Ao = (Mat_SeqAIJ*)Aij->B->data;
4324: PetscScalar *ad = Ad->a, *ao = Ao->a;
4325: const PetscInt *Adi = Ad->i;
4326: PetscInt ldi,Iii,md;
4329: if (Ii[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4330: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4331: if (m != mat->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4332: if (n != mat->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4334: cstart = mat->cmap->rstart;
4335: if (!Aij->ld) {
4336: /* count number of entries below block diagonal */
4337: PetscCalloc1(m,&ld);
4338: Aij->ld = ld;
4339: for (i=0; i<m; i++) {
4340: nnz = Ii[i+1]- Ii[i];
4341: j = 0;
4342: while (J[j] < cstart && j < nnz) {j++;}
4343: J += nnz;
4344: ld[i] = j;
4345: }
4346: } else {
4347: ld = Aij->ld;
4348: }
4350: for (i=0; i<m; i++) {
4351: nnz = Ii[i+1]- Ii[i];
4352: Iii = Ii[i];
4353: ldi = ld[i];
4354: md = Adi[i+1]-Adi[i];
4355: PetscArraycpy(ao,v + Iii,ldi);
4356: PetscArraycpy(ad,v + Iii + ldi,md);
4357: PetscArraycpy(ao + ldi,v + Iii + ldi + md,nnz - ldi - md);
4358: ad += md;
4359: ao += nnz - md;
4360: }
4361: nooffprocentries = mat->nooffprocentries;
4362: mat->nooffprocentries = PETSC_TRUE;
4363: PetscObjectStateIncrease((PetscObject)Aij->A);
4364: PetscObjectStateIncrease((PetscObject)Aij->B);
4365: PetscObjectStateIncrease((PetscObject)mat);
4366: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
4367: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
4368: mat->nooffprocentries = nooffprocentries;
4369: return(0);
4370: }
4372: /*@C
4373: MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4374: (the default parallel PETSc format). For good matrix assembly performance
4375: the user should preallocate the matrix storage by setting the parameters
4376: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4377: performance can be increased by more than a factor of 50.
4379: Collective
4381: Input Parameters:
4382: + comm - MPI communicator
4383: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4384: This value should be the same as the local size used in creating the
4385: y vector for the matrix-vector product y = Ax.
4386: . n - This value should be the same as the local size used in creating the
4387: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4388: calculated if N is given) For square matrices n is almost always m.
4389: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4390: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4391: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4392: (same value is used for all local rows)
4393: . d_nnz - array containing the number of nonzeros in the various rows of the
4394: DIAGONAL portion of the local submatrix (possibly different for each row)
4395: or NULL, if d_nz is used to specify the nonzero structure.
4396: The size of this array is equal to the number of local rows, i.e 'm'.
4397: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4398: submatrix (same value is used for all local rows).
4399: - o_nnz - array containing the number of nonzeros in the various rows of the
4400: OFF-DIAGONAL portion of the local submatrix (possibly different for
4401: each row) or NULL, if o_nz is used to specify the nonzero
4402: structure. The size of this array is equal to the number
4403: of local rows, i.e 'm'.
4405: Output Parameter:
4406: . A - the matrix
4408: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4409: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4410: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
4412: Notes:
4413: If the *_nnz parameter is given then the *_nz parameter is ignored
4415: m,n,M,N parameters specify the size of the matrix, and its partitioning across
4416: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4417: storage requirements for this matrix.
4419: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
4420: processor than it must be used on all processors that share the object for
4421: that argument.
4423: The user MUST specify either the local or global matrix dimensions
4424: (possibly both).
4426: The parallel matrix is partitioned across processors such that the
4427: first m0 rows belong to process 0, the next m1 rows belong to
4428: process 1, the next m2 rows belong to process 2 etc.. where
4429: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4430: values corresponding to [m x N] submatrix.
4432: The columns are logically partitioned with the n0 columns belonging
4433: to 0th partition, the next n1 columns belonging to the next
4434: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4436: The DIAGONAL portion of the local submatrix on any given processor
4437: is the submatrix corresponding to the rows and columns m,n
4438: corresponding to the given processor. i.e diagonal matrix on
4439: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4440: etc. The remaining portion of the local submatrix [m x (N-n)]
4441: constitute the OFF-DIAGONAL portion. The example below better
4442: illustrates this concept.
4444: For a square global matrix we define each processor's diagonal portion
4445: to be its local rows and the corresponding columns (a square submatrix);
4446: each processor's off-diagonal portion encompasses the remainder of the
4447: local matrix (a rectangular submatrix).
4449: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4451: When calling this routine with a single process communicator, a matrix of
4452: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
4453: type of communicator, use the construction mechanism
4454: .vb
4455: MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4456: .ve
4458: $ MatCreate(...,&A);
4459: $ MatSetType(A,MATMPIAIJ);
4460: $ MatSetSizes(A, m,n,M,N);
4461: $ MatMPIAIJSetPreallocation(A,...);
4463: By default, this format uses inodes (identical nodes) when possible.
4464: We search for consecutive rows with the same nonzero structure, thereby
4465: reusing matrix information to achieve increased efficiency.
4467: Options Database Keys:
4468: + -mat_no_inode - Do not use inodes
4469: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4473: Example usage:
4475: Consider the following 8x8 matrix with 34 non-zero values, that is
4476: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4477: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4478: as follows
4480: .vb
4481: 1 2 0 | 0 3 0 | 0 4
4482: Proc0 0 5 6 | 7 0 0 | 8 0
4483: 9 0 10 | 11 0 0 | 12 0
4484: -------------------------------------
4485: 13 0 14 | 15 16 17 | 0 0
4486: Proc1 0 18 0 | 19 20 21 | 0 0
4487: 0 0 0 | 22 23 0 | 24 0
4488: -------------------------------------
4489: Proc2 25 26 27 | 0 0 28 | 29 0
4490: 30 0 0 | 31 32 33 | 0 34
4491: .ve
4493: This can be represented as a collection of submatrices as
4495: .vb
4496: A B C
4497: D E F
4498: G H I
4499: .ve
4501: Where the submatrices A,B,C are owned by proc0, D,E,F are
4502: owned by proc1, G,H,I are owned by proc2.
4504: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4505: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4506: The 'M','N' parameters are 8,8, and have the same values on all procs.
4508: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4509: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4510: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4511: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4512: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4513: matrix, ans [DF] as another SeqAIJ matrix.
4515: When d_nz, o_nz parameters are specified, d_nz storage elements are
4516: allocated for every row of the local diagonal submatrix, and o_nz
4517: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4518: One way to choose d_nz and o_nz is to use the max nonzerors per local
4519: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4520: In this case, the values of d_nz,o_nz are
4521: .vb
4522: proc0 : dnz = 2, o_nz = 2
4523: proc1 : dnz = 3, o_nz = 2
4524: proc2 : dnz = 1, o_nz = 4
4525: .ve
4526: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4527: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4528: for proc3. i.e we are using 12+15+10=37 storage locations to store
4529: 34 values.
4531: When d_nnz, o_nnz parameters are specified, the storage is specified
4532: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4533: In the above case the values for d_nnz,o_nnz are
4534: .vb
4535: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4536: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4537: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4538: .ve
4539: Here the space allocated is sum of all the above values i.e 34, and
4540: hence pre-allocation is perfect.
4542: Level: intermediate
4544: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4545: MATMPIAIJ, MatCreateMPIAIJWithArrays()
4546: @*/
4547: PetscErrorCode MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
4548: {
4550: PetscMPIInt size;
4553: MatCreate(comm,A);
4554: MatSetSizes(*A,m,n,M,N);
4555: MPI_Comm_size(comm,&size);
4556: if (size > 1) {
4557: MatSetType(*A,MATMPIAIJ);
4558: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4559: } else {
4560: MatSetType(*A,MATSEQAIJ);
4561: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4562: }
4563: return(0);
4564: }
4566: /*@C
4567: MatMPIAIJGetSeqAIJ - Returns the local piece of this distributed matrix
4569: Not collective
4571: Input Parameter:
4572: . A - The MPIAIJ matrix
4574: Output Parameters:
4575: + Ad - The local diagonal block as a SeqAIJ matrix
4576: . Ao - The local off-diagonal block as a SeqAIJ matrix
4577: - colmap - An array mapping local column numbers of Ao to global column numbers of the parallel matrix
4579: Note: The rows in Ad and Ao are in [0, Nr), where Nr is the number of local rows on this process. The columns
4580: in Ad are in [0, Nc) where Nc is the number of local columns. The columns are Ao are in [0, Nco), where Nco is
4581: the number of nonzero columns in the local off-diagonal piece of the matrix A. The array colmap maps these
4582: local column numbers to global column numbers in the original matrix.
4584: Level: intermediate
4586: .seealso: MatMPIAIJGetLocalMat(), MatMPIAIJGetLocalMatCondensed(), MatCreateAIJ(), MATMPIAIJ, MATSEQAIJ
4587: @*/
4588: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
4589: {
4590: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4591: PetscBool flg;
4595: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);
4596: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input");
4597: if (Ad) *Ad = a->A;
4598: if (Ao) *Ao = a->B;
4599: if (colmap) *colmap = a->garray;
4600: return(0);
4601: }
4603: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4604: {
4606: PetscInt m,N,i,rstart,nnz,Ii;
4607: PetscInt *indx;
4608: PetscScalar *values;
4611: MatGetSize(inmat,&m,&N);
4612: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4613: PetscInt *dnz,*onz,sum,bs,cbs;
4615: if (n == PETSC_DECIDE) {
4616: PetscSplitOwnership(comm,&n,&N);
4617: }
4618: /* Check sum(n) = N */
4619: MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4620: if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N);
4622: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4623: rstart -= m;
4625: MatPreallocateInitialize(comm,m,n,dnz,onz);
4626: for (i=0; i<m; i++) {
4627: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4628: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4629: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4630: }
4632: MatCreate(comm,outmat);
4633: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4634: MatGetBlockSizes(inmat,&bs,&cbs);
4635: MatSetBlockSizes(*outmat,bs,cbs);
4636: MatSetType(*outmat,MATAIJ);
4637: MatSeqAIJSetPreallocation(*outmat,0,dnz);
4638: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4639: MatPreallocateFinalize(dnz,onz);
4640: }
4642: /* numeric phase */
4643: MatGetOwnershipRange(*outmat,&rstart,NULL);
4644: for (i=0; i<m; i++) {
4645: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4646: Ii = i + rstart;
4647: MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4648: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4649: }
4650: MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
4651: MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
4652: return(0);
4653: }
4655: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4656: {
4657: PetscErrorCode ierr;
4658: PetscMPIInt rank;
4659: PetscInt m,N,i,rstart,nnz;
4660: size_t len;
4661: const PetscInt *indx;
4662: PetscViewer out;
4663: char *name;
4664: Mat B;
4665: const PetscScalar *values;
4668: MatGetLocalSize(A,&m,NULL);
4669: MatGetSize(A,NULL,&N);
4670: /* Should this be the type of the diagonal block of A? */
4671: MatCreate(PETSC_COMM_SELF,&B);
4672: MatSetSizes(B,m,N,m,N);
4673: MatSetBlockSizesFromMats(B,A,A);
4674: MatSetType(B,MATSEQAIJ);
4675: MatSeqAIJSetPreallocation(B,0,NULL);
4676: MatGetOwnershipRange(A,&rstart,NULL);
4677: for (i=0; i<m; i++) {
4678: MatGetRow(A,i+rstart,&nnz,&indx,&values);
4679: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4680: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4681: }
4682: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4683: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4685: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
4686: PetscStrlen(outfile,&len);
4687: PetscMalloc1(len+6,&name);
4688: PetscSNPrintf(name,len+6,"%s.%d",outfile,rank);
4689: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4690: PetscFree(name);
4691: MatView(B,out);
4692: PetscViewerDestroy(&out);
4693: MatDestroy(&B);
4694: return(0);
4695: }
4697: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4698: {
4699: PetscErrorCode ierr;
4700: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4703: if (!merge) return(0);
4704: PetscFree(merge->id_r);
4705: PetscFree(merge->len_s);
4706: PetscFree(merge->len_r);
4707: PetscFree(merge->bi);
4708: PetscFree(merge->bj);
4709: PetscFree(merge->buf_ri[0]);
4710: PetscFree(merge->buf_ri);
4711: PetscFree(merge->buf_rj[0]);
4712: PetscFree(merge->buf_rj);
4713: PetscFree(merge->coi);
4714: PetscFree(merge->coj);
4715: PetscFree(merge->owners_co);
4716: PetscLayoutDestroy(&merge->rowmap);
4717: PetscFree(merge);
4718: return(0);
4719: }
4721: #include <../src/mat/utils/freespace.h>
4722: #include <petscbt.h>
4724: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4725: {
4726: PetscErrorCode ierr;
4727: MPI_Comm comm;
4728: Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data;
4729: PetscMPIInt size,rank,taga,*len_s;
4730: PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4731: PetscInt proc,m;
4732: PetscInt **buf_ri,**buf_rj;
4733: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4734: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
4735: MPI_Request *s_waits,*r_waits;
4736: MPI_Status *status;
4737: MatScalar *aa=a->a;
4738: MatScalar **abuf_r,*ba_i;
4739: Mat_Merge_SeqsToMPI *merge;
4740: PetscContainer container;
4743: PetscObjectGetComm((PetscObject)mpimat,&comm);
4744: PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);
4746: MPI_Comm_size(comm,&size);
4747: MPI_Comm_rank(comm,&rank);
4749: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4750: if (!container) SETERRQ(PetscObjectComm((PetscObject)mpimat),PETSC_ERR_PLIB,"Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4751: PetscContainerGetPointer(container,(void**)&merge);
4753: bi = merge->bi;
4754: bj = merge->bj;
4755: buf_ri = merge->buf_ri;
4756: buf_rj = merge->buf_rj;
4758: PetscMalloc1(size,&status);
4759: owners = merge->rowmap->range;
4760: len_s = merge->len_s;
4762: /* send and recv matrix values */
4763: /*-----------------------------*/
4764: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4765: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
4767: PetscMalloc1(merge->nsend+1,&s_waits);
4768: for (proc=0,k=0; proc<size; proc++) {
4769: if (!len_s[proc]) continue;
4770: i = owners[proc];
4771: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4772: k++;
4773: }
4775: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4776: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4777: PetscFree(status);
4779: PetscFree(s_waits);
4780: PetscFree(r_waits);
4782: /* insert mat values of mpimat */
4783: /*----------------------------*/
4784: PetscMalloc1(N,&ba_i);
4785: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4787: for (k=0; k<merge->nrecv; k++) {
4788: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4789: nrows = *(buf_ri_k[k]);
4790: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
4791: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4792: }
4794: /* set values of ba */
4795: m = merge->rowmap->n;
4796: for (i=0; i<m; i++) {
4797: arow = owners[rank] + i;
4798: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
4799: bnzi = bi[i+1] - bi[i];
4800: PetscArrayzero(ba_i,bnzi);
4802: /* add local non-zero vals of this proc's seqmat into ba */
4803: anzi = ai[arow+1] - ai[arow];
4804: aj = a->j + ai[arow];
4805: aa = a->a + ai[arow];
4806: nextaj = 0;
4807: for (j=0; nextaj<anzi; j++) {
4808: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4809: ba_i[j] += aa[nextaj++];
4810: }
4811: }
4813: /* add received vals into ba */
4814: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4815: /* i-th row */
4816: if (i == *nextrow[k]) {
4817: anzi = *(nextai[k]+1) - *nextai[k];
4818: aj = buf_rj[k] + *(nextai[k]);
4819: aa = abuf_r[k] + *(nextai[k]);
4820: nextaj = 0;
4821: for (j=0; nextaj<anzi; j++) {
4822: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4823: ba_i[j] += aa[nextaj++];
4824: }
4825: }
4826: nextrow[k]++; nextai[k]++;
4827: }
4828: }
4829: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4830: }
4831: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4832: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
4834: PetscFree(abuf_r[0]);
4835: PetscFree(abuf_r);
4836: PetscFree(ba_i);
4837: PetscFree3(buf_ri_k,nextrow,nextai);
4838: PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4839: return(0);
4840: }
4842: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4843: {
4844: PetscErrorCode ierr;
4845: Mat B_mpi;
4846: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4847: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4848: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
4849: PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4850: PetscInt len,proc,*dnz,*onz,bs,cbs;
4851: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4852: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4853: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
4854: MPI_Status *status;
4855: PetscFreeSpaceList free_space=NULL,current_space=NULL;
4856: PetscBT lnkbt;
4857: Mat_Merge_SeqsToMPI *merge;
4858: PetscContainer container;
4861: PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);
4863: /* make sure it is a PETSc comm */
4864: PetscCommDuplicate(comm,&comm,NULL);
4865: MPI_Comm_size(comm,&size);
4866: MPI_Comm_rank(comm,&rank);
4868: PetscNew(&merge);
4869: PetscMalloc1(size,&status);
4871: /* determine row ownership */
4872: /*---------------------------------------------------------*/
4873: PetscLayoutCreate(comm,&merge->rowmap);
4874: PetscLayoutSetLocalSize(merge->rowmap,m);
4875: PetscLayoutSetSize(merge->rowmap,M);
4876: PetscLayoutSetBlockSize(merge->rowmap,1);
4877: PetscLayoutSetUp(merge->rowmap);
4878: PetscMalloc1(size,&len_si);
4879: PetscMalloc1(size,&merge->len_s);
4881: m = merge->rowmap->n;
4882: owners = merge->rowmap->range;
4884: /* determine the number of messages to send, their lengths */
4885: /*---------------------------------------------------------*/
4886: len_s = merge->len_s;
4888: len = 0; /* length of buf_si[] */
4889: merge->nsend = 0;
4890: for (proc=0; proc<size; proc++) {
4891: len_si[proc] = 0;
4892: if (proc == rank) {
4893: len_s[proc] = 0;
4894: } else {
4895: len_si[proc] = owners[proc+1] - owners[proc] + 1;
4896: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4897: }
4898: if (len_s[proc]) {
4899: merge->nsend++;
4900: nrows = 0;
4901: for (i=owners[proc]; i<owners[proc+1]; i++) {
4902: if (ai[i+1] > ai[i]) nrows++;
4903: }
4904: len_si[proc] = 2*(nrows+1);
4905: len += len_si[proc];
4906: }
4907: }
4909: /* determine the number and length of messages to receive for ij-structure */
4910: /*-------------------------------------------------------------------------*/
4911: PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4912: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
4914: /* post the Irecv of j-structure */
4915: /*-------------------------------*/
4916: PetscCommGetNewTag(comm,&tagj);
4917: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
4919: /* post the Isend of j-structure */
4920: /*--------------------------------*/
4921: PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);
4923: for (proc=0, k=0; proc<size; proc++) {
4924: if (!len_s[proc]) continue;
4925: i = owners[proc];
4926: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4927: k++;
4928: }
4930: /* receives and sends of j-structure are complete */
4931: /*------------------------------------------------*/
4932: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4933: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
4935: /* send and recv i-structure */
4936: /*---------------------------*/
4937: PetscCommGetNewTag(comm,&tagi);
4938: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
4940: PetscMalloc1(len+1,&buf_s);
4941: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4942: for (proc=0,k=0; proc<size; proc++) {
4943: if (!len_s[proc]) continue;
4944: /* form outgoing message for i-structure:
4945: buf_si[0]: nrows to be sent
4946: [1:nrows]: row index (global)
4947: [nrows+1:2*nrows+1]: i-structure index
4948: */
4949: /*-------------------------------------------*/
4950: nrows = len_si[proc]/2 - 1;
4951: buf_si_i = buf_si + nrows+1;
4952: buf_si[0] = nrows;
4953: buf_si_i[0] = 0;
4954: nrows = 0;
4955: for (i=owners[proc]; i<owners[proc+1]; i++) {
4956: anzi = ai[i+1] - ai[i];
4957: if (anzi) {
4958: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4959: buf_si[nrows+1] = i-owners[proc]; /* local row index */
4960: nrows++;
4961: }
4962: }
4963: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4964: k++;
4965: buf_si += len_si[proc];
4966: }
4968: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4969: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
4971: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4972: for (i=0; i<merge->nrecv; i++) {
4973: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4974: }
4976: PetscFree(len_si);
4977: PetscFree(len_ri);
4978: PetscFree(rj_waits);
4979: PetscFree2(si_waits,sj_waits);
4980: PetscFree(ri_waits);
4981: PetscFree(buf_s);
4982: PetscFree(status);
4984: /* compute a local seq matrix in each processor */
4985: /*----------------------------------------------*/
4986: /* allocate bi array and free space for accumulating nonzero column info */
4987: PetscMalloc1(m+1,&bi);
4988: bi[0] = 0;
4990: /* create and initialize a linked list */
4991: nlnk = N+1;
4992: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
4994: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4995: len = ai[owners[rank+1]] - ai[owners[rank]];
4996: PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);
4998: current_space = free_space;
5000: /* determine symbolic info for each local row */
5001: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
5003: for (k=0; k<merge->nrecv; k++) {
5004: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
5005: nrows = *buf_ri_k[k];
5006: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
5007: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
5008: }
5010: MatPreallocateInitialize(comm,m,n,dnz,onz);
5011: len = 0;
5012: for (i=0; i<m; i++) {
5013: bnzi = 0;
5014: /* add local non-zero cols of this proc's seqmat into lnk */
5015: arow = owners[rank] + i;
5016: anzi = ai[arow+1] - ai[arow];
5017: aj = a->j + ai[arow];
5018: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
5019: bnzi += nlnk;
5020: /* add received col data into lnk */
5021: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
5022: if (i == *nextrow[k]) { /* i-th row */
5023: anzi = *(nextai[k]+1) - *nextai[k];
5024: aj = buf_rj[k] + *nextai[k];
5025: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
5026: bnzi += nlnk;
5027: nextrow[k]++; nextai[k]++;
5028: }
5029: }
5030: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5032: /* if free space is not available, make more free space */
5033: if (current_space->local_remaining<bnzi) {
5034: PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);
5035: nspacedouble++;
5036: }
5037: /* copy data into free space, then initialize lnk */
5038: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
5039: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
5041: current_space->array += bnzi;
5042: current_space->local_used += bnzi;
5043: current_space->local_remaining -= bnzi;
5045: bi[i+1] = bi[i] + bnzi;
5046: }
5048: PetscFree3(buf_ri_k,nextrow,nextai);
5050: PetscMalloc1(bi[m]+1,&bj);
5051: PetscFreeSpaceContiguous(&free_space,bj);
5052: PetscLLDestroy(lnk,lnkbt);
5054: /* create symbolic parallel matrix B_mpi */
5055: /*---------------------------------------*/
5056: MatGetBlockSizes(seqmat,&bs,&cbs);
5057: MatCreate(comm,&B_mpi);
5058: if (n==PETSC_DECIDE) {
5059: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
5060: } else {
5061: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5062: }
5063: MatSetBlockSizes(B_mpi,bs,cbs);
5064: MatSetType(B_mpi,MATMPIAIJ);
5065: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
5066: MatPreallocateFinalize(dnz,onz);
5067: MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
5069: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5070: B_mpi->assembled = PETSC_FALSE;
5071: merge->bi = bi;
5072: merge->bj = bj;
5073: merge->buf_ri = buf_ri;
5074: merge->buf_rj = buf_rj;
5075: merge->coi = NULL;
5076: merge->coj = NULL;
5077: merge->owners_co = NULL;
5079: PetscCommDestroy(&comm);
5081: /* attach the supporting struct to B_mpi for reuse */
5082: PetscContainerCreate(PETSC_COMM_SELF,&container);
5083: PetscContainerSetPointer(container,merge);
5084: PetscContainerSetUserDestroy(container,MatDestroy_MPIAIJ_SeqsToMPI);
5085: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
5086: PetscContainerDestroy(&container);
5087: *mpimat = B_mpi;
5089: PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
5090: return(0);
5091: }
5093: /*@C
5094: MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential
5095: matrices from each processor
5097: Collective
5099: Input Parameters:
5100: + comm - the communicators the parallel matrix will live on
5101: . seqmat - the input sequential matrices
5102: . m - number of local rows (or PETSC_DECIDE)
5103: . n - number of local columns (or PETSC_DECIDE)
5104: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5106: Output Parameter:
5107: . mpimat - the parallel matrix generated
5109: Level: advanced
5111: Notes:
5112: The dimensions of the sequential matrix in each processor MUST be the same.
5113: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5114: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
5115: @*/
5116: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
5117: {
5119: PetscMPIInt size;
5122: MPI_Comm_size(comm,&size);
5123: if (size == 1) {
5124: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
5125: if (scall == MAT_INITIAL_MATRIX) {
5126: MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
5127: } else {
5128: MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
5129: }
5130: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
5131: return(0);
5132: }
5133: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
5134: if (scall == MAT_INITIAL_MATRIX) {
5135: MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
5136: }
5137: MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
5138: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
5139: return(0);
5140: }
5142: /*@
5143: MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with
5144: mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
5145: with MatGetSize()
5147: Not Collective
5149: Input Parameters:
5150: + A - the matrix
5151: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5153: Output Parameter:
5154: . A_loc - the local sequential matrix generated
5156: Level: developer
5158: Notes:
5159: When the communicator associated with A has size 1 and MAT_INITIAL_MATRIX is requested, the matrix returned is the diagonal part of A.
5160: If MAT_REUSE_MATRIX is requested with comm size 1, MatCopy(Adiag,*A_loc,SAME_NONZERO_PATTERN) is called.
5161: This means that one can preallocate the proper sequential matrix first and then call this routine with MAT_REUSE_MATRIX to safely
5162: modify the values of the returned A_loc.
5164: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed()
5166: @*/
5167: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
5168: {
5170: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
5171: Mat_SeqAIJ *mat,*a,*b;
5172: PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
5173: MatScalar *aa,*ba,*cam;
5174: PetscScalar *ca;
5175: PetscMPIInt size;
5176: PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
5177: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
5178: PetscBool match;
5181: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);
5182: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5183: MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
5184: if (size == 1) {
5185: if (scall == MAT_INITIAL_MATRIX) {
5186: PetscObjectReference((PetscObject)mpimat->A);
5187: *A_loc = mpimat->A;
5188: } else if (scall == MAT_REUSE_MATRIX) {
5189: MatCopy(mpimat->A,*A_loc,SAME_NONZERO_PATTERN);
5190: }
5191: return(0);
5192: }
5194: PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
5195: a = (Mat_SeqAIJ*)(mpimat->A)->data;
5196: b = (Mat_SeqAIJ*)(mpimat->B)->data;
5197: ai = a->i; aj = a->j; bi = b->i; bj = b->j;
5198: aa = a->a; ba = b->a;
5199: if (scall == MAT_INITIAL_MATRIX) {
5200: PetscMalloc1(1+am,&ci);
5201: ci[0] = 0;
5202: for (i=0; i<am; i++) {
5203: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
5204: }
5205: PetscMalloc1(1+ci[am],&cj);
5206: PetscMalloc1(1+ci[am],&ca);
5207: k = 0;
5208: for (i=0; i<am; i++) {
5209: ncols_o = bi[i+1] - bi[i];
5210: ncols_d = ai[i+1] - ai[i];
5211: /* off-diagonal portion of A */
5212: for (jo=0; jo<ncols_o; jo++) {
5213: col = cmap[*bj];
5214: if (col >= cstart) break;
5215: cj[k] = col; bj++;
5216: ca[k++] = *ba++;
5217: }
5218: /* diagonal portion of A */
5219: for (j=0; j<ncols_d; j++) {
5220: cj[k] = cstart + *aj++;
5221: ca[k++] = *aa++;
5222: }
5223: /* off-diagonal portion of A */
5224: for (j=jo; j<ncols_o; j++) {
5225: cj[k] = cmap[*bj++];
5226: ca[k++] = *ba++;
5227: }
5228: }
5229: /* put together the new matrix */
5230: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
5231: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5232: /* Since these are PETSc arrays, change flags to free them as necessary. */
5233: mat = (Mat_SeqAIJ*)(*A_loc)->data;
5234: mat->free_a = PETSC_TRUE;
5235: mat->free_ij = PETSC_TRUE;
5236: mat->nonew = 0;
5237: } else if (scall == MAT_REUSE_MATRIX) {
5238: mat=(Mat_SeqAIJ*)(*A_loc)->data;
5239: ci = mat->i; cj = mat->j; cam = mat->a;
5240: for (i=0; i<am; i++) {
5241: /* off-diagonal portion of A */
5242: ncols_o = bi[i+1] - bi[i];
5243: for (jo=0; jo<ncols_o; jo++) {
5244: col = cmap[*bj];
5245: if (col >= cstart) break;
5246: *cam++ = *ba++; bj++;
5247: }
5248: /* diagonal portion of A */
5249: ncols_d = ai[i+1] - ai[i];
5250: for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5251: /* off-diagonal portion of A */
5252: for (j=jo; j<ncols_o; j++) {
5253: *cam++ = *ba++; bj++;
5254: }
5255: }
5256: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5257: PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5258: return(0);
5259: }
5261: /*@C
5262: MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns
5264: Not Collective
5266: Input Parameters:
5267: + A - the matrix
5268: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5269: - row, col - index sets of rows and columns to extract (or NULL)
5271: Output Parameter:
5272: . A_loc - the local sequential matrix generated
5274: Level: developer
5276: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()
5278: @*/
5279: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5280: {
5281: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5283: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5284: IS isrowa,iscola;
5285: Mat *aloc;
5286: PetscBool match;
5289: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5290: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5291: PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5292: if (!row) {
5293: start = A->rmap->rstart; end = A->rmap->rend;
5294: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5295: } else {
5296: isrowa = *row;
5297: }
5298: if (!col) {
5299: start = A->cmap->rstart;
5300: cmap = a->garray;
5301: nzA = a->A->cmap->n;
5302: nzB = a->B->cmap->n;
5303: PetscMalloc1(nzA+nzB, &idx);
5304: ncols = 0;
5305: for (i=0; i<nzB; i++) {
5306: if (cmap[i] < start) idx[ncols++] = cmap[i];
5307: else break;
5308: }
5309: imark = i;
5310: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5311: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5312: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5313: } else {
5314: iscola = *col;
5315: }
5316: if (scall != MAT_INITIAL_MATRIX) {
5317: PetscMalloc1(1,&aloc);
5318: aloc[0] = *A_loc;
5319: }
5320: MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5321: if (!col) { /* attach global id of condensed columns */
5322: PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);
5323: }
5324: *A_loc = aloc[0];
5325: PetscFree(aloc);
5326: if (!row) {
5327: ISDestroy(&isrowa);
5328: }
5329: if (!col) {
5330: ISDestroy(&iscola);
5331: }
5332: PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5333: return(0);
5334: }
5336: /*
5337: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5338: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5339: * on a global size.
5340: * */
5341: PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P,IS rows,Mat *P_oth)
5342: {
5343: Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data;
5344: Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data,*p_oth;
5345: PetscInt plocalsize,nrows,*ilocal,*oilocal,i,lidx,*nrcols,*nlcols,ncol;
5346: PetscMPIInt owner;
5347: PetscSFNode *iremote,*oiremote;
5348: const PetscInt *lrowindices;
5349: PetscErrorCode ierr;
5350: PetscSF sf,osf;
5351: PetscInt pcstart,*roffsets,*loffsets,*pnnz,j;
5352: PetscInt ontotalcols,dntotalcols,ntotalcols,nout;
5353: MPI_Comm comm;
5354: ISLocalToGlobalMapping mapping;
5357: PetscObjectGetComm((PetscObject)P,&comm);
5358: /* plocalsize is the number of roots
5359: * nrows is the number of leaves
5360: * */
5361: MatGetLocalSize(P,&plocalsize,NULL);
5362: ISGetLocalSize(rows,&nrows);
5363: PetscCalloc1(nrows,&iremote);
5364: ISGetIndices(rows,&lrowindices);
5365: for (i=0;i<nrows;i++) {
5366: /* Find a remote index and an owner for a row
5367: * The row could be local or remote
5368: * */
5369: owner = 0;
5370: lidx = 0;
5371: PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,&lidx);
5372: iremote[i].index = lidx;
5373: iremote[i].rank = owner;
5374: }
5375: /* Create SF to communicate how many nonzero columns for each row */
5376: PetscSFCreate(comm,&sf);
5377: /* SF will figure out the number of nonzero colunms for each row, and their
5378: * offsets
5379: * */
5380: PetscSFSetGraph(sf,plocalsize,nrows,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5381: PetscSFSetFromOptions(sf);
5382: PetscSFSetUp(sf);
5384: PetscCalloc1(2*(plocalsize+1),&roffsets);
5385: PetscCalloc1(2*plocalsize,&nrcols);
5386: PetscCalloc1(nrows,&pnnz);
5387: roffsets[0] = 0;
5388: roffsets[1] = 0;
5389: for (i=0;i<plocalsize;i++) {
5390: /* diag */
5391: nrcols[i*2+0] = pd->i[i+1] - pd->i[i];
5392: /* off diag */
5393: nrcols[i*2+1] = po->i[i+1] - po->i[i];
5394: /* compute offsets so that we relative location for each row */
5395: roffsets[(i+1)*2+0] = roffsets[i*2+0] + nrcols[i*2+0];
5396: roffsets[(i+1)*2+1] = roffsets[i*2+1] + nrcols[i*2+1];
5397: }
5398: PetscCalloc1(2*nrows,&nlcols);
5399: PetscCalloc1(2*nrows,&loffsets);
5400: /* 'r' means root, and 'l' means leaf */
5401: PetscSFBcastBegin(sf,MPIU_2INT,nrcols,nlcols);
5402: PetscSFBcastBegin(sf,MPIU_2INT,roffsets,loffsets);
5403: PetscSFBcastEnd(sf,MPIU_2INT,nrcols,nlcols);
5404: PetscSFBcastEnd(sf,MPIU_2INT,roffsets,loffsets);
5405: PetscSFDestroy(&sf);
5406: PetscFree(roffsets);
5407: PetscFree(nrcols);
5408: dntotalcols = 0;
5409: ontotalcols = 0;
5410: ncol = 0;
5411: for (i=0;i<nrows;i++) {
5412: pnnz[i] = nlcols[i*2+0] + nlcols[i*2+1];
5413: ncol = PetscMax(pnnz[i],ncol);
5414: /* diag */
5415: dntotalcols += nlcols[i*2+0];
5416: /* off diag */
5417: ontotalcols += nlcols[i*2+1];
5418: }
5419: /* We do not need to figure the right number of columns
5420: * since all the calculations will be done by going through the raw data
5421: * */
5422: MatCreateSeqAIJ(PETSC_COMM_SELF,nrows,ncol,0,pnnz,P_oth);
5423: MatSetUp(*P_oth);
5424: PetscFree(pnnz);
5425: p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5426: /* diag */
5427: PetscCalloc1(dntotalcols,&iremote);
5428: /* off diag */
5429: PetscCalloc1(ontotalcols,&oiremote);
5430: /* diag */
5431: PetscCalloc1(dntotalcols,&ilocal);
5432: /* off diag */
5433: PetscCalloc1(ontotalcols,&oilocal);
5434: dntotalcols = 0;
5435: ontotalcols = 0;
5436: ntotalcols = 0;
5437: for (i=0;i<nrows;i++) {
5438: owner = 0;
5439: PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,NULL);
5440: /* Set iremote for diag matrix */
5441: for (j=0;j<nlcols[i*2+0];j++) {
5442: iremote[dntotalcols].index = loffsets[i*2+0] + j;
5443: iremote[dntotalcols].rank = owner;
5444: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5445: ilocal[dntotalcols++] = ntotalcols++;
5446: }
5447: /* off diag */
5448: for (j=0;j<nlcols[i*2+1];j++) {
5449: oiremote[ontotalcols].index = loffsets[i*2+1] + j;
5450: oiremote[ontotalcols].rank = owner;
5451: oilocal[ontotalcols++] = ntotalcols++;
5452: }
5453: }
5454: ISRestoreIndices(rows,&lrowindices);
5455: PetscFree(loffsets);
5456: PetscFree(nlcols);
5457: PetscSFCreate(comm,&sf);
5458: /* P serves as roots and P_oth is leaves
5459: * Diag matrix
5460: * */
5461: PetscSFSetGraph(sf,pd->i[plocalsize],dntotalcols,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5462: PetscSFSetFromOptions(sf);
5463: PetscSFSetUp(sf);
5465: PetscSFCreate(comm,&osf);
5466: /* Off diag */
5467: PetscSFSetGraph(osf,po->i[plocalsize],ontotalcols,oilocal,PETSC_OWN_POINTER,oiremote,PETSC_OWN_POINTER);
5468: PetscSFSetFromOptions(osf);
5469: PetscSFSetUp(osf);
5470: /* We operate on the matrix internal data for saving memory */
5471: PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5472: PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5473: MatGetOwnershipRangeColumn(P,&pcstart,NULL);
5474: /* Convert to global indices for diag matrix */
5475: for (i=0;i<pd->i[plocalsize];i++) pd->j[i] += pcstart;
5476: PetscSFBcastBegin(sf,MPIU_INT,pd->j,p_oth->j);
5477: /* We want P_oth store global indices */
5478: ISLocalToGlobalMappingCreate(comm,1,p->B->cmap->n,p->garray,PETSC_COPY_VALUES,&mapping);
5479: /* Use memory scalable approach */
5480: ISLocalToGlobalMappingSetType(mapping,ISLOCALTOGLOBALMAPPINGHASH);
5481: ISLocalToGlobalMappingApply(mapping,po->i[plocalsize],po->j,po->j);
5482: PetscSFBcastBegin(osf,MPIU_INT,po->j,p_oth->j);
5483: PetscSFBcastEnd(sf,MPIU_INT,pd->j,p_oth->j);
5484: /* Convert back to local indices */
5485: for (i=0;i<pd->i[plocalsize];i++) pd->j[i] -= pcstart;
5486: PetscSFBcastEnd(osf,MPIU_INT,po->j,p_oth->j);
5487: nout = 0;
5488: ISGlobalToLocalMappingApply(mapping,IS_GTOLM_DROP,po->i[plocalsize],po->j,&nout,po->j);
5489: if (nout != po->i[plocalsize]) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP,"n %D does not equal to nout %D \n",po->i[plocalsize],nout);
5490: ISLocalToGlobalMappingDestroy(&mapping);
5491: /* Exchange values */
5492: PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5493: PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5494: /* Stop PETSc from shrinking memory */
5495: for (i=0;i<nrows;i++) p_oth->ilen[i] = p_oth->imax[i];
5496: MatAssemblyBegin(*P_oth,MAT_FINAL_ASSEMBLY);
5497: MatAssemblyEnd(*P_oth,MAT_FINAL_ASSEMBLY);
5498: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5499: PetscObjectCompose((PetscObject)*P_oth,"diagsf",(PetscObject)sf);
5500: PetscObjectCompose((PetscObject)*P_oth,"offdiagsf",(PetscObject)osf);
5501: PetscSFDestroy(&sf);
5502: PetscSFDestroy(&osf);
5503: return(0);
5504: }
5506: /*
5507: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5508: * This supports MPIAIJ and MAIJ
5509: * */
5510: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A,Mat P,PetscInt dof,MatReuse reuse,Mat *P_oth)
5511: {
5512: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data;
5513: Mat_SeqAIJ *p_oth;
5514: Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data;
5515: IS rows,map;
5516: PetscHMapI hamp;
5517: PetscInt i,htsize,*rowindices,off,*mapping,key,count;
5518: MPI_Comm comm;
5519: PetscSF sf,osf;
5520: PetscBool has;
5521: PetscErrorCode ierr;
5524: PetscObjectGetComm((PetscObject)A,&comm);
5525: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,P,0,0);
5526: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5527: * and then create a submatrix (that often is an overlapping matrix)
5528: * */
5529: if (reuse == MAT_INITIAL_MATRIX) {
5530: /* Use a hash table to figure out unique keys */
5531: PetscHMapICreate(&hamp);
5532: PetscHMapIResize(hamp,a->B->cmap->n);
5533: PetscCalloc1(a->B->cmap->n,&mapping);
5534: count = 0;
5535: /* Assume that a->g is sorted, otherwise the following does not make sense */
5536: for (i=0;i<a->B->cmap->n;i++) {
5537: key = a->garray[i]/dof;
5538: PetscHMapIHas(hamp,key,&has);
5539: if (!has) {
5540: mapping[i] = count;
5541: PetscHMapISet(hamp,key,count++);
5542: } else {
5543: /* Current 'i' has the same value the previous step */
5544: mapping[i] = count-1;
5545: }
5546: }
5547: ISCreateGeneral(comm,a->B->cmap->n,mapping,PETSC_OWN_POINTER,&map);
5548: PetscHMapIGetSize(hamp,&htsize);
5549: if (htsize!=count) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP," Size of hash map %D is inconsistent with count %D \n",htsize,count);
5550: PetscCalloc1(htsize,&rowindices);
5551: off = 0;
5552: PetscHMapIGetKeys(hamp,&off,rowindices);
5553: PetscHMapIDestroy(&hamp);
5554: PetscSortInt(htsize,rowindices);
5555: ISCreateGeneral(comm,htsize,rowindices,PETSC_OWN_POINTER,&rows);
5556: /* In case, the matrix was already created but users want to recreate the matrix */
5557: MatDestroy(P_oth);
5558: MatCreateSeqSubMatrixWithRows_Private(P,rows,P_oth);
5559: PetscObjectCompose((PetscObject)*P_oth,"aoffdiagtopothmapping",(PetscObject)map);
5560: ISDestroy(&map);
5561: ISDestroy(&rows);
5562: } else if (reuse == MAT_REUSE_MATRIX) {
5563: /* If matrix was already created, we simply update values using SF objects
5564: * that as attached to the matrix ealier.
5565: * */
5566: PetscObjectQuery((PetscObject)*P_oth,"diagsf",(PetscObject*)&sf);
5567: PetscObjectQuery((PetscObject)*P_oth,"offdiagsf",(PetscObject*)&osf);
5568: if (!sf || !osf) SETERRQ(comm,PETSC_ERR_ARG_NULL,"Matrix is not initialized yet");
5569: p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5570: /* Update values in place */
5571: PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5572: PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5573: PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5574: PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5575: } else SETERRQ(comm,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown reuse type");
5576: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,P,0,0);
5577: return(0);
5578: }
5580: /*@C
5581: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5583: Collective on Mat
5585: Input Parameters:
5586: + A,B - the matrices in mpiaij format
5587: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5588: - rowb, colb - index sets of rows and columns of B to extract (or NULL)
5590: Output Parameter:
5591: + rowb, colb - index sets of rows and columns of B to extract
5592: - B_seq - the sequential matrix generated
5594: Level: developer
5596: @*/
5597: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5598: {
5599: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5601: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5602: IS isrowb,iscolb;
5603: Mat *bseq=NULL;
5606: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5607: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5608: }
5609: PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5611: if (scall == MAT_INITIAL_MATRIX) {
5612: start = A->cmap->rstart;
5613: cmap = a->garray;
5614: nzA = a->A->cmap->n;
5615: nzB = a->B->cmap->n;
5616: PetscMalloc1(nzA+nzB, &idx);
5617: ncols = 0;
5618: for (i=0; i<nzB; i++) { /* row < local row index */
5619: if (cmap[i] < start) idx[ncols++] = cmap[i];
5620: else break;
5621: }
5622: imark = i;
5623: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
5624: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5625: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5626: ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5627: } else {
5628: if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5629: isrowb = *rowb; iscolb = *colb;
5630: PetscMalloc1(1,&bseq);
5631: bseq[0] = *B_seq;
5632: }
5633: MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5634: *B_seq = bseq[0];
5635: PetscFree(bseq);
5636: if (!rowb) {
5637: ISDestroy(&isrowb);
5638: } else {
5639: *rowb = isrowb;
5640: }
5641: if (!colb) {
5642: ISDestroy(&iscolb);
5643: } else {
5644: *colb = iscolb;
5645: }
5646: PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5647: return(0);
5648: }
5650: /*
5651: MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5652: of the OFF-DIAGONAL portion of local A
5654: Collective on Mat
5656: Input Parameters:
5657: + A,B - the matrices in mpiaij format
5658: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5660: Output Parameter:
5661: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5662: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5663: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5664: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5666: Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product
5667: for this matrix. This is not desirable..
5669: Level: developer
5671: */
5672: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5673: {
5674: PetscErrorCode ierr;
5675: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5676: Mat_SeqAIJ *b_oth;
5677: VecScatter ctx;
5678: MPI_Comm comm;
5679: const PetscMPIInt *rprocs,*sprocs;
5680: const PetscInt *srow,*rstarts,*sstarts;
5681: PetscInt *rowlen,*bufj,*bufJ,ncols = 0,aBn=a->B->cmap->n,row,*b_othi,*b_othj,*rvalues=NULL,*svalues=NULL,*cols,sbs,rbs;
5682: PetscInt i,j,k=0,l,ll,nrecvs,nsends,nrows,*rstartsj = NULL,*sstartsj,len;
5683: PetscScalar *b_otha,*bufa,*bufA,*vals = NULL;
5684: MPI_Request *rwaits = NULL,*swaits = NULL;
5685: MPI_Status rstatus;
5686: PetscMPIInt jj,size,tag,rank,nsends_mpi,nrecvs_mpi;
5689: PetscObjectGetComm((PetscObject)A,&comm);
5690: MPI_Comm_size(comm,&size);
5692: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5693: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5694: }
5695: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5696: MPI_Comm_rank(comm,&rank);
5698: if (size == 1) {
5699: startsj_s = NULL;
5700: bufa_ptr = NULL;
5701: *B_oth = NULL;
5702: return(0);
5703: }
5705: ctx = a->Mvctx;
5706: tag = ((PetscObject)ctx)->tag;
5708: if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use");
5709: VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&srow,&sprocs,&sbs);
5710: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5711: VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL/*indices not needed*/,&rprocs,&rbs);
5712: PetscMPIIntCast(nsends,&nsends_mpi);
5713: PetscMPIIntCast(nrecvs,&nrecvs_mpi);
5714: PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);
5716: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5717: if (scall == MAT_INITIAL_MATRIX) {
5718: /* i-array */
5719: /*---------*/
5720: /* post receives */
5721: if (nrecvs) {PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);} /* rstarts can be NULL when nrecvs=0 */
5722: for (i=0; i<nrecvs; i++) {
5723: rowlen = rvalues + rstarts[i]*rbs;
5724: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5725: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5726: }
5728: /* pack the outgoing message */
5729: PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);
5731: sstartsj[0] = 0;
5732: rstartsj[0] = 0;
5733: len = 0; /* total length of j or a array to be sent */
5734: if (nsends) {
5735: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5736: PetscMalloc1(sbs*(sstarts[nsends]-sstarts[0]),&svalues);
5737: }
5738: for (i=0; i<nsends; i++) {
5739: rowlen = svalues + (sstarts[i]-sstarts[0])*sbs;
5740: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5741: for (j=0; j<nrows; j++) {
5742: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5743: for (l=0; l<sbs; l++) {
5744: MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */
5746: rowlen[j*sbs+l] = ncols;
5748: len += ncols;
5749: MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
5750: }
5751: k++;
5752: }
5753: MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5755: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5756: }
5757: /* recvs and sends of i-array are completed */
5758: i = nrecvs;
5759: while (i--) {
5760: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5761: }
5762: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5763: PetscFree(svalues);
5765: /* allocate buffers for sending j and a arrays */
5766: PetscMalloc1(len+1,&bufj);
5767: PetscMalloc1(len+1,&bufa);
5769: /* create i-array of B_oth */
5770: PetscMalloc1(aBn+2,&b_othi);
5772: b_othi[0] = 0;
5773: len = 0; /* total length of j or a array to be received */
5774: k = 0;
5775: for (i=0; i<nrecvs; i++) {
5776: rowlen = rvalues + (rstarts[i]-rstarts[0])*rbs;
5777: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of rows to be received */
5778: for (j=0; j<nrows; j++) {
5779: b_othi[k+1] = b_othi[k] + rowlen[j];
5780: PetscIntSumError(rowlen[j],len,&len);
5781: k++;
5782: }
5783: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5784: }
5785: PetscFree(rvalues);
5787: /* allocate space for j and a arrrays of B_oth */
5788: PetscMalloc1(b_othi[aBn]+1,&b_othj);
5789: PetscMalloc1(b_othi[aBn]+1,&b_otha);
5791: /* j-array */
5792: /*---------*/
5793: /* post receives of j-array */
5794: for (i=0; i<nrecvs; i++) {
5795: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5796: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5797: }
5799: /* pack the outgoing message j-array */
5800: if (nsends) k = sstarts[0];
5801: for (i=0; i<nsends; i++) {
5802: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5803: bufJ = bufj+sstartsj[i];
5804: for (j=0; j<nrows; j++) {
5805: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5806: for (ll=0; ll<sbs; ll++) {
5807: MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5808: for (l=0; l<ncols; l++) {
5809: *bufJ++ = cols[l];
5810: }
5811: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5812: }
5813: }
5814: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5815: }
5817: /* recvs and sends of j-array are completed */
5818: i = nrecvs;
5819: while (i--) {
5820: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5821: }
5822: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5823: } else if (scall == MAT_REUSE_MATRIX) {
5824: sstartsj = *startsj_s;
5825: rstartsj = *startsj_r;
5826: bufa = *bufa_ptr;
5827: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5828: b_otha = b_oth->a;
5829: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5831: /* a-array */
5832: /*---------*/
5833: /* post receives of a-array */
5834: for (i=0; i<nrecvs; i++) {
5835: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5836: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5837: }
5839: /* pack the outgoing message a-array */
5840: if (nsends) k = sstarts[0];
5841: for (i=0; i<nsends; i++) {
5842: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5843: bufA = bufa+sstartsj[i];
5844: for (j=0; j<nrows; j++) {
5845: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5846: for (ll=0; ll<sbs; ll++) {
5847: MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5848: for (l=0; l<ncols; l++) {
5849: *bufA++ = vals[l];
5850: }
5851: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5852: }
5853: }
5854: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5855: }
5856: /* recvs and sends of a-array are completed */
5857: i = nrecvs;
5858: while (i--) {
5859: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5860: }
5861: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5862: PetscFree2(rwaits,swaits);
5864: if (scall == MAT_INITIAL_MATRIX) {
5865: /* put together the new matrix */
5866: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);
5868: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5869: /* Since these are PETSc arrays, change flags to free them as necessary. */
5870: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5871: b_oth->free_a = PETSC_TRUE;
5872: b_oth->free_ij = PETSC_TRUE;
5873: b_oth->nonew = 0;
5875: PetscFree(bufj);
5876: if (!startsj_s || !bufa_ptr) {
5877: PetscFree2(sstartsj,rstartsj);
5878: PetscFree(bufa_ptr);
5879: } else {
5880: *startsj_s = sstartsj;
5881: *startsj_r = rstartsj;
5882: *bufa_ptr = bufa;
5883: }
5884: }
5886: VecScatterRestoreRemote_Private(ctx,PETSC_TRUE,&nsends,&sstarts,&srow,&sprocs,&sbs);
5887: VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE,&nrecvs,&rstarts,NULL,&rprocs,&rbs);
5888: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5889: return(0);
5890: }
5892: /*@C
5893: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
5895: Not Collective
5897: Input Parameters:
5898: . A - The matrix in mpiaij format
5900: Output Parameter:
5901: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5902: . colmap - A map from global column index to local index into lvec
5903: - multScatter - A scatter from the argument of a matrix-vector product to lvec
5905: Level: developer
5907: @*/
5908: #if defined(PETSC_USE_CTABLE)
5909: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5910: #else
5911: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5912: #endif
5913: {
5914: Mat_MPIAIJ *a;
5921: a = (Mat_MPIAIJ*) A->data;
5922: if (lvec) *lvec = a->lvec;
5923: if (colmap) *colmap = a->colmap;
5924: if (multScatter) *multScatter = a->Mvctx;
5925: return(0);
5926: }
5928: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
5929: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
5930: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*);
5931: #if defined(PETSC_HAVE_MKL_SPARSE)
5932: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*);
5933: #endif
5934: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat,MatType,MatReuse,Mat*);
5935: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
5936: #if defined(PETSC_HAVE_ELEMENTAL)
5937: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
5938: #endif
5939: #if defined(PETSC_HAVE_SCALAPACK)
5940: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*);
5941: #endif
5942: #if defined(PETSC_HAVE_HYPRE)
5943: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
5944: #endif
5945: #if defined(PETSC_HAVE_CUDA)
5946: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat,MatType,MatReuse,Mat*);
5947: #endif
5948: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*);
5949: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
5950: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
5952: /*
5953: Computes (B'*A')' since computing B*A directly is untenable
5955: n p p
5956: [ ] [ ] [ ]
5957: m [ A ] * n [ B ] = m [ C ]
5958: [ ] [ ] [ ]
5960: */
5961: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5962: {
5964: Mat At,Bt,Ct;
5967: MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5968: MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5969: MatMatMult(Bt,At,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&Ct);
5970: MatDestroy(&At);
5971: MatDestroy(&Bt);
5972: MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5973: MatDestroy(&Ct);
5974: return(0);
5975: }
5977: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat C)
5978: {
5980: PetscBool cisdense;
5983: if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n);
5984: MatSetSizes(C,A->rmap->n,B->cmap->n,A->rmap->N,B->cmap->N);
5985: MatSetBlockSizesFromMats(C,A,B);
5986: PetscObjectTypeCompareAny((PetscObject)C,&cisdense,MATMPIDENSE,MATMPIDENSECUDA,"");
5987: if (!cisdense) {
5988: MatSetType(C,((PetscObject)A)->type_name);
5989: }
5990: MatSetUp(C);
5992: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
5993: return(0);
5994: }
5996: /* ----------------------------------------------------------------*/
5997: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
5998: {
5999: Mat_Product *product = C->product;
6000: Mat A = product->A,B=product->B;
6003: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend)
6004: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
6006: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6007: C->ops->productsymbolic = MatProductSymbolic_AB;
6008: return(0);
6009: }
6011: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6012: {
6014: Mat_Product *product = C->product;
6017: if (product->type == MATPRODUCT_AB) {
6018: MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C);
6019: }
6020: return(0);
6021: }
6022: /* ----------------------------------------------------------------*/
6024: /*MC
6025: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6027: Options Database Keys:
6028: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
6030: Level: beginner
6032: Notes:
6033: MatSetValues() may be called for this matrix type with a NULL argument for the numerical values,
6034: in this case the values associated with the rows and columns one passes in are set to zero
6035: in the matrix
6037: MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
6038: space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored
6040: .seealso: MatCreateAIJ()
6041: M*/
6043: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6044: {
6045: Mat_MPIAIJ *b;
6047: PetscMPIInt size;
6050: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
6052: PetscNewLog(B,&b);
6053: B->data = (void*)b;
6054: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
6055: B->assembled = PETSC_FALSE;
6056: B->insertmode = NOT_SET_VALUES;
6057: b->size = size;
6059: MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
6061: /* build cache for off array entries formed */
6062: MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);
6064: b->donotstash = PETSC_FALSE;
6065: b->colmap = NULL;
6066: b->garray = NULL;
6067: b->roworiented = PETSC_TRUE;
6069: /* stuff used for matrix vector multiply */
6070: b->lvec = NULL;
6071: b->Mvctx = NULL;
6073: /* stuff for MatGetRow() */
6074: b->rowindices = NULL;
6075: b->rowvalues = NULL;
6076: b->getrowactive = PETSC_FALSE;
6078: /* flexible pointer used in CUSP/CUSPARSE classes */
6079: b->spptr = NULL;
6081: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
6082: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
6083: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
6084: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
6085: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
6086: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);
6087: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
6088: PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
6089: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
6090: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);
6091: #if defined(PETSC_HAVE_MKL_SPARSE)
6092: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);
6093: #endif
6094: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
6095: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpibaij_C",MatConvert_MPIAIJ_MPIBAIJ);
6096: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
6097: #if defined(PETSC_HAVE_ELEMENTAL)
6098: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
6099: #endif
6100: #if defined(PETSC_HAVE_SCALAPACK)
6101: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_scalapack_C",MatConvert_AIJ_ScaLAPACK);
6102: #endif
6103: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);
6104: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);
6105: #if defined(PETSC_HAVE_HYPRE)
6106: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);
6107: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_transpose_mpiaij_mpiaij_C",MatProductSetFromOptions_Transpose_AIJ_AIJ);
6108: #endif
6109: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_is_mpiaij_C",MatProductSetFromOptions_IS_XAIJ);
6110: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_mpiaij_mpiaij_C",MatProductSetFromOptions_MPIAIJ);
6111: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
6112: return(0);
6113: }
6115: /*@C
6116: MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
6117: and "off-diagonal" part of the matrix in CSR format.
6119: Collective
6121: Input Parameters:
6122: + comm - MPI communicator
6123: . m - number of local rows (Cannot be PETSC_DECIDE)
6124: . n - This value should be the same as the local size used in creating the
6125: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
6126: calculated if N is given) For square matrices n is almost always m.
6127: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
6128: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
6129: . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
6130: . j - column indices
6131: . a - matrix values
6132: . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix
6133: . oj - column indices
6134: - oa - matrix values
6136: Output Parameter:
6137: . mat - the matrix
6139: Level: advanced
6141: Notes:
6142: The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6143: must free the arrays once the matrix has been destroyed and not before.
6145: The i and j indices are 0 based
6147: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix
6149: This sets local rows and cannot be used to set off-processor values.
6151: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6152: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6153: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6154: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6155: keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
6156: communication if it is known that only local entries will be set.
6158: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
6159: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
6160: @*/
6161: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
6162: {
6164: Mat_MPIAIJ *maij;
6167: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
6168: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
6169: if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
6170: MatCreate(comm,mat);
6171: MatSetSizes(*mat,m,n,M,N);
6172: MatSetType(*mat,MATMPIAIJ);
6173: maij = (Mat_MPIAIJ*) (*mat)->data;
6175: (*mat)->preallocated = PETSC_TRUE;
6177: PetscLayoutSetUp((*mat)->rmap);
6178: PetscLayoutSetUp((*mat)->cmap);
6180: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
6181: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);
6183: MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
6184: MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
6185: MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
6186: MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);
6188: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
6189: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
6190: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
6191: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
6192: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
6193: return(0);
6194: }
6196: /*
6197: Special version for direct calls from Fortran
6198: */
6199: #include <petsc/private/fortranimpl.h>
6201: /* Change these macros so can be used in void function */
6202: #undef CHKERRQ
6203: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
6204: #undef SETERRQ2
6205: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
6206: #undef SETERRQ3
6207: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
6208: #undef SETERRQ
6209: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)
6211: #if defined(PETSC_HAVE_FORTRAN_CAPS)
6212: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
6213: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
6214: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
6215: #else
6216: #endif
6217: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
6218: {
6219: Mat mat = *mmat;
6220: PetscInt m = *mm, n = *mn;
6221: InsertMode addv = *maddv;
6222: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
6223: PetscScalar value;
6226: MatCheckPreallocated(mat,1);
6227: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
6228: else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
6229: {
6230: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
6231: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
6232: PetscBool roworiented = aij->roworiented;
6234: /* Some Variables required in the macro */
6235: Mat A = aij->A;
6236: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
6237: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
6238: MatScalar *aa = a->a;
6239: PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
6240: Mat B = aij->B;
6241: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
6242: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
6243: MatScalar *ba = b->a;
6244: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
6245: * cannot use "#if defined" inside a macro. */
6246: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
6248: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
6249: PetscInt nonew = a->nonew;
6250: MatScalar *ap1,*ap2;
6253: for (i=0; i<m; i++) {
6254: if (im[i] < 0) continue;
6255: if (PetscUnlikelyDebug(im[i] >= mat->rmap->N)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
6256: if (im[i] >= rstart && im[i] < rend) {
6257: row = im[i] - rstart;
6258: lastcol1 = -1;
6259: rp1 = aj + ai[row];
6260: ap1 = aa + ai[row];
6261: rmax1 = aimax[row];
6262: nrow1 = ailen[row];
6263: low1 = 0;
6264: high1 = nrow1;
6265: lastcol2 = -1;
6266: rp2 = bj + bi[row];
6267: ap2 = ba + bi[row];
6268: rmax2 = bimax[row];
6269: nrow2 = bilen[row];
6270: low2 = 0;
6271: high2 = nrow2;
6273: for (j=0; j<n; j++) {
6274: if (roworiented) value = v[i*n+j];
6275: else value = v[i+j*m];
6276: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
6277: if (in[j] >= cstart && in[j] < cend) {
6278: col = in[j] - cstart;
6279: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
6280: #if defined(PETSC_HAVE_DEVICE)
6281: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
6282: #endif
6283: } else if (in[j] < 0) continue;
6284: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
6285: /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */
6286: SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
6287: } else {
6288: if (mat->was_assembled) {
6289: if (!aij->colmap) {
6290: MatCreateColmap_MPIAIJ_Private(mat);
6291: }
6292: #if defined(PETSC_USE_CTABLE)
6293: PetscTableFind(aij->colmap,in[j]+1,&col);
6294: col--;
6295: #else
6296: col = aij->colmap[in[j]] - 1;
6297: #endif
6298: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
6299: MatDisAssemble_MPIAIJ(mat);
6300: col = in[j];
6301: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
6302: B = aij->B;
6303: b = (Mat_SeqAIJ*)B->data;
6304: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
6305: rp2 = bj + bi[row];
6306: ap2 = ba + bi[row];
6307: rmax2 = bimax[row];
6308: nrow2 = bilen[row];
6309: low2 = 0;
6310: high2 = nrow2;
6311: bm = aij->B->rmap->n;
6312: ba = b->a;
6313: inserted = PETSC_FALSE;
6314: }
6315: } else col = in[j];
6316: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
6317: #if defined(PETSC_HAVE_DEVICE)
6318: if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
6319: #endif
6320: }
6321: }
6322: } else if (!aij->donotstash) {
6323: if (roworiented) {
6324: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6325: } else {
6326: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6327: }
6328: }
6329: }
6330: }
6331: PetscFunctionReturnVoid();
6332: }