Actual source code: mpiaij.c
petsc-3.10.0 2018-09-12
3: #include <../src/mat/impls/aij/mpi/mpiaij.h>
4: #include <petsc/private/vecimpl.h>
5: #include <petsc/private/isimpl.h>
6: #include <petscblaslapack.h>
7: #include <petscsf.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: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
48: {
50: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
53: if (mat->A) {
54: MatSetBlockSizes(mat->A,rbs,cbs);
55: MatSetBlockSizes(mat->B,rbs,1);
56: }
57: return(0);
58: }
60: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
61: {
62: PetscErrorCode ierr;
63: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
64: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data;
65: Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data;
66: const PetscInt *ia,*ib;
67: const MatScalar *aa,*bb;
68: PetscInt na,nb,i,j,*rows,cnt=0,n0rows;
69: PetscInt m = M->rmap->n,rstart = M->rmap->rstart;
72: *keptrows = 0;
73: ia = a->i;
74: ib = b->i;
75: for (i=0; i<m; i++) {
76: na = ia[i+1] - ia[i];
77: nb = ib[i+1] - ib[i];
78: if (!na && !nb) {
79: cnt++;
80: goto ok1;
81: }
82: aa = a->a + ia[i];
83: for (j=0; j<na; j++) {
84: if (aa[j] != 0.0) goto ok1;
85: }
86: bb = b->a + ib[i];
87: for (j=0; j <nb; j++) {
88: if (bb[j] != 0.0) goto ok1;
89: }
90: cnt++;
91: ok1:;
92: }
93: MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
94: if (!n0rows) return(0);
95: PetscMalloc1(M->rmap->n-cnt,&rows);
96: cnt = 0;
97: for (i=0; i<m; i++) {
98: na = ia[i+1] - ia[i];
99: nb = ib[i+1] - ib[i];
100: if (!na && !nb) continue;
101: aa = a->a + ia[i];
102: for (j=0; j<na;j++) {
103: if (aa[j] != 0.0) {
104: rows[cnt++] = rstart + i;
105: goto ok2;
106: }
107: }
108: bb = b->a + ib[i];
109: for (j=0; j<nb; j++) {
110: if (bb[j] != 0.0) {
111: rows[cnt++] = rstart + i;
112: goto ok2;
113: }
114: }
115: ok2:;
116: }
117: ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
118: return(0);
119: }
121: PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
122: {
123: PetscErrorCode ierr;
124: Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data;
125: PetscBool cong;
128: MatHasCongruentLayouts(Y,&cong);
129: if (Y->assembled && cong) {
130: MatDiagonalSet(aij->A,D,is);
131: } else {
132: MatDiagonalSet_Default(Y,D,is);
133: }
134: return(0);
135: }
137: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
138: {
139: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data;
141: PetscInt i,rstart,nrows,*rows;
144: *zrows = NULL;
145: MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
146: MatGetOwnershipRange(M,&rstart,NULL);
147: for (i=0; i<nrows; i++) rows[i] += rstart;
148: ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
149: return(0);
150: }
152: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
153: {
155: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
156: PetscInt i,n,*garray = aij->garray;
157: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data;
158: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data;
159: PetscReal *work;
162: MatGetSize(A,NULL,&n);
163: PetscCalloc1(n,&work);
164: if (type == NORM_2) {
165: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
166: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
167: }
168: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
169: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
170: }
171: } else if (type == NORM_1) {
172: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
173: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
174: }
175: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
176: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
177: }
178: } else if (type == NORM_INFINITY) {
179: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
180: work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
181: }
182: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
183: work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
184: }
186: } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
187: if (type == NORM_INFINITY) {
188: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
189: } else {
190: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
191: }
192: PetscFree(work);
193: if (type == NORM_2) {
194: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
195: }
196: return(0);
197: }
199: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
200: {
201: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
202: IS sis,gis;
203: PetscErrorCode ierr;
204: const PetscInt *isis,*igis;
205: PetscInt n,*iis,nsis,ngis,rstart,i;
208: MatFindOffBlockDiagonalEntries(a->A,&sis);
209: MatFindNonzeroRows(a->B,&gis);
210: ISGetSize(gis,&ngis);
211: ISGetSize(sis,&nsis);
212: ISGetIndices(sis,&isis);
213: ISGetIndices(gis,&igis);
215: PetscMalloc1(ngis+nsis,&iis);
216: PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));
217: PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));
218: n = ngis + nsis;
219: PetscSortRemoveDupsInt(&n,iis);
220: MatGetOwnershipRange(A,&rstart,NULL);
221: for (i=0; i<n; i++) iis[i] += rstart;
222: ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);
224: ISRestoreIndices(sis,&isis);
225: ISRestoreIndices(gis,&igis);
226: ISDestroy(&sis);
227: ISDestroy(&gis);
228: return(0);
229: }
231: /*
232: Distributes a SeqAIJ matrix across a set of processes. Code stolen from
233: MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.
235: Only for square matrices
237: Used by a preconditioner, hence PETSC_EXTERN
238: */
239: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
240: {
241: PetscMPIInt rank,size;
242: PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
244: Mat mat;
245: Mat_SeqAIJ *gmata;
246: PetscMPIInt tag;
247: MPI_Status status;
248: PetscBool aij;
249: MatScalar *gmataa,*ao,*ad,*gmataarestore=0;
252: MPI_Comm_rank(comm,&rank);
253: MPI_Comm_size(comm,&size);
254: if (!rank) {
255: PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
256: if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
257: }
258: if (reuse == MAT_INITIAL_MATRIX) {
259: MatCreate(comm,&mat);
260: MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
261: MatGetBlockSizes(gmat,&bses[0],&bses[1]);
262: MPI_Bcast(bses,2,MPIU_INT,0,comm);
263: MatSetBlockSizes(mat,bses[0],bses[1]);
264: MatSetType(mat,MATAIJ);
265: PetscMalloc1(size+1,&rowners);
266: PetscMalloc2(m,&dlens,m,&olens);
267: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
269: rowners[0] = 0;
270: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
271: rstart = rowners[rank];
272: rend = rowners[rank+1];
273: PetscObjectGetNewTag((PetscObject)mat,&tag);
274: if (!rank) {
275: gmata = (Mat_SeqAIJ*) gmat->data;
276: /* send row lengths to all processors */
277: for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
278: for (i=1; i<size; i++) {
279: MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
280: }
281: /* determine number diagonal and off-diagonal counts */
282: PetscMemzero(olens,m*sizeof(PetscInt));
283: PetscCalloc1(m,&ld);
284: jj = 0;
285: for (i=0; i<m; i++) {
286: for (j=0; j<dlens[i]; j++) {
287: if (gmata->j[jj] < rstart) ld[i]++;
288: if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
289: jj++;
290: }
291: }
292: /* send column indices to other processes */
293: for (i=1; i<size; i++) {
294: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
295: MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
296: MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
297: }
299: /* send numerical values to other processes */
300: for (i=1; i<size; i++) {
301: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
302: MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
303: }
304: gmataa = gmata->a;
305: gmataj = gmata->j;
307: } else {
308: /* receive row lengths */
309: MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
310: /* receive column indices */
311: MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
312: PetscMalloc2(nz,&gmataa,nz,&gmataj);
313: MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
314: /* determine number diagonal and off-diagonal counts */
315: PetscMemzero(olens,m*sizeof(PetscInt));
316: PetscCalloc1(m,&ld);
317: jj = 0;
318: for (i=0; i<m; i++) {
319: for (j=0; j<dlens[i]; j++) {
320: if (gmataj[jj] < rstart) ld[i]++;
321: if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
322: jj++;
323: }
324: }
325: /* receive numerical values */
326: PetscMemzero(gmataa,nz*sizeof(PetscScalar));
327: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
328: }
329: /* set preallocation */
330: for (i=0; i<m; i++) {
331: dlens[i] -= olens[i];
332: }
333: MatSeqAIJSetPreallocation(mat,0,dlens);
334: MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
336: for (i=0; i<m; i++) {
337: dlens[i] += olens[i];
338: }
339: cnt = 0;
340: for (i=0; i<m; i++) {
341: row = rstart + i;
342: MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
343: cnt += dlens[i];
344: }
345: if (rank) {
346: PetscFree2(gmataa,gmataj);
347: }
348: PetscFree2(dlens,olens);
349: PetscFree(rowners);
351: ((Mat_MPIAIJ*)(mat->data))->ld = ld;
353: *inmat = mat;
354: } else { /* column indices are already set; only need to move over numerical values from process 0 */
355: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
356: Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
357: mat = *inmat;
358: PetscObjectGetNewTag((PetscObject)mat,&tag);
359: if (!rank) {
360: /* send numerical values to other processes */
361: gmata = (Mat_SeqAIJ*) gmat->data;
362: MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
363: gmataa = gmata->a;
364: for (i=1; i<size; i++) {
365: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
366: MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
367: }
368: nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
369: } else {
370: /* receive numerical values from process 0*/
371: nz = Ad->nz + Ao->nz;
372: PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
373: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
374: }
375: /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
376: ld = ((Mat_MPIAIJ*)(mat->data))->ld;
377: ad = Ad->a;
378: ao = Ao->a;
379: if (mat->rmap->n) {
380: i = 0;
381: nz = ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
382: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
383: }
384: for (i=1; i<mat->rmap->n; i++) {
385: nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
386: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
387: }
388: i--;
389: if (mat->rmap->n) {
390: nz = Ao->i[i+1] - Ao->i[i] - ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));
391: }
392: if (rank) {
393: PetscFree(gmataarestore);
394: }
395: }
396: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
397: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
398: return(0);
399: }
401: /*
402: Local utility routine that creates a mapping from the global column
403: number to the local number in the off-diagonal part of the local
404: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
405: a slightly higher hash table cost; without it it is not scalable (each processor
406: has an order N integer array but is fast to acess.
407: */
408: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
409: {
410: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
412: PetscInt n = aij->B->cmap->n,i;
415: if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
416: #if defined(PETSC_USE_CTABLE)
417: PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
418: for (i=0; i<n; i++) {
419: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
420: }
421: #else
422: PetscCalloc1(mat->cmap->N+1,&aij->colmap);
423: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
424: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
425: #endif
426: return(0);
427: }
429: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \
430: { \
431: if (col <= lastcol1) low1 = 0; \
432: else high1 = nrow1; \
433: lastcol1 = col;\
434: while (high1-low1 > 5) { \
435: t = (low1+high1)/2; \
436: if (rp1[t] > col) high1 = t; \
437: else low1 = t; \
438: } \
439: for (_i=low1; _i<high1; _i++) { \
440: if (rp1[_i] > col) break; \
441: if (rp1[_i] == col) { \
442: if (addv == ADD_VALUES) ap1[_i] += value; \
443: else ap1[_i] = value; \
444: goto a_noinsert; \
445: } \
446: } \
447: if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
448: if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \
449: 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); \
450: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
451: N = nrow1++ - 1; a->nz++; high1++; \
452: /* shift up all the later entries in this row */ \
453: for (ii=N; ii>=_i; ii--) { \
454: rp1[ii+1] = rp1[ii]; \
455: ap1[ii+1] = ap1[ii]; \
456: } \
457: rp1[_i] = col; \
458: ap1[_i] = value; \
459: A->nonzerostate++;\
460: a_noinsert: ; \
461: ailen[row] = nrow1; \
462: }
464: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
465: { \
466: if (col <= lastcol2) low2 = 0; \
467: else high2 = nrow2; \
468: lastcol2 = col; \
469: while (high2-low2 > 5) { \
470: t = (low2+high2)/2; \
471: if (rp2[t] > col) high2 = t; \
472: else low2 = t; \
473: } \
474: for (_i=low2; _i<high2; _i++) { \
475: if (rp2[_i] > col) break; \
476: if (rp2[_i] == col) { \
477: if (addv == ADD_VALUES) ap2[_i] += value; \
478: else ap2[_i] = value; \
479: goto b_noinsert; \
480: } \
481: } \
482: if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
483: if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
484: 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); \
485: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
486: N = nrow2++ - 1; b->nz++; high2++; \
487: /* shift up all the later entries in this row */ \
488: for (ii=N; ii>=_i; ii--) { \
489: rp2[ii+1] = rp2[ii]; \
490: ap2[ii+1] = ap2[ii]; \
491: } \
492: rp2[_i] = col; \
493: ap2[_i] = value; \
494: B->nonzerostate++; \
495: b_noinsert: ; \
496: bilen[row] = nrow2; \
497: }
499: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
500: {
501: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
502: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
504: PetscInt l,*garray = mat->garray,diag;
507: /* code only works for square matrices A */
509: /* find size of row to the left of the diagonal part */
510: MatGetOwnershipRange(A,&diag,0);
511: row = row - diag;
512: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
513: if (garray[b->j[b->i[row]+l]] > diag) break;
514: }
515: PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));
517: /* diagonal part */
518: PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));
520: /* right of diagonal part */
521: PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));
522: return(0);
523: }
525: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
526: {
527: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
528: PetscScalar value;
530: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
531: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
532: PetscBool roworiented = aij->roworiented;
534: /* Some Variables required in the macro */
535: Mat A = aij->A;
536: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
537: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
538: MatScalar *aa = a->a;
539: PetscBool ignorezeroentries = a->ignorezeroentries;
540: Mat B = aij->B;
541: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
542: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
543: MatScalar *ba = b->a;
545: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
546: PetscInt nonew;
547: MatScalar *ap1,*ap2;
550: for (i=0; i<m; i++) {
551: if (im[i] < 0) continue;
552: #if defined(PETSC_USE_DEBUG)
553: if (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);
554: #endif
555: if (im[i] >= rstart && im[i] < rend) {
556: row = im[i] - rstart;
557: lastcol1 = -1;
558: rp1 = aj + ai[row];
559: ap1 = aa + ai[row];
560: rmax1 = aimax[row];
561: nrow1 = ailen[row];
562: low1 = 0;
563: high1 = nrow1;
564: lastcol2 = -1;
565: rp2 = bj + bi[row];
566: ap2 = ba + bi[row];
567: rmax2 = bimax[row];
568: nrow2 = bilen[row];
569: low2 = 0;
570: high2 = nrow2;
572: for (j=0; j<n; j++) {
573: if (roworiented) value = v[i*n+j];
574: else value = v[i+j*m];
575: if (in[j] >= cstart && in[j] < cend) {
576: col = in[j] - cstart;
577: nonew = a->nonew;
578: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
579: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
580: } else if (in[j] < 0) continue;
581: #if defined(PETSC_USE_DEBUG)
582: 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);
583: #endif
584: else {
585: if (mat->was_assembled) {
586: if (!aij->colmap) {
587: MatCreateColmap_MPIAIJ_Private(mat);
588: }
589: #if defined(PETSC_USE_CTABLE)
590: PetscTableFind(aij->colmap,in[j]+1,&col);
591: col--;
592: #else
593: col = aij->colmap[in[j]] - 1;
594: #endif
595: if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
596: MatDisAssemble_MPIAIJ(mat);
597: col = in[j];
598: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
599: B = aij->B;
600: b = (Mat_SeqAIJ*)B->data;
601: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
602: rp2 = bj + bi[row];
603: ap2 = ba + bi[row];
604: rmax2 = bimax[row];
605: nrow2 = bilen[row];
606: low2 = 0;
607: high2 = nrow2;
608: bm = aij->B->rmap->n;
609: ba = b->a;
610: } else if (col < 0) {
611: if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) {
612: 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]);
613: } 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]);
614: }
615: } else col = in[j];
616: nonew = b->nonew;
617: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
618: }
619: }
620: } else {
621: 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]);
622: if (!aij->donotstash) {
623: mat->assembled = PETSC_FALSE;
624: if (roworiented) {
625: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
626: } else {
627: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
628: }
629: }
630: }
631: }
632: return(0);
633: }
635: /*
636: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
637: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
638: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
639: */
640: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[])
641: {
642: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
643: Mat A = aij->A; /* diagonal part of the matrix */
644: Mat B = aij->B; /* offdiagonal part of the matrix */
645: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
646: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
647: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col;
648: PetscInt *ailen = a->ilen,*aj = a->j;
649: PetscInt *bilen = b->ilen,*bj = b->j;
650: PetscInt am = aij->A->rmap->n,j;
651: PetscInt diag_so_far = 0,dnz;
652: PetscInt offd_so_far = 0,onz;
655: /* Iterate over all rows of the matrix */
656: for (j=0; j<am; j++) {
657: dnz = onz = 0;
658: /* Iterate over all non-zero columns of the current row */
659: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
660: /* If column is in the diagonal */
661: if (mat_j[col] >= cstart && mat_j[col] < cend) {
662: aj[diag_so_far++] = mat_j[col] - cstart;
663: dnz++;
664: } else { /* off-diagonal entries */
665: bj[offd_so_far++] = mat_j[col];
666: onz++;
667: }
668: }
669: ailen[j] = dnz;
670: bilen[j] = onz;
671: }
672: return(0);
673: }
675: /*
676: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
677: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
678: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
679: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
680: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
681: */
682: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[])
683: {
684: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
685: Mat A = aij->A; /* diagonal part of the matrix */
686: Mat B = aij->B; /* offdiagonal part of the matrix */
687: Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data;
688: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
689: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
690: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend;
691: PetscInt *ailen = a->ilen,*aj = a->j;
692: PetscInt *bilen = b->ilen,*bj = b->j;
693: PetscInt am = aij->A->rmap->n,j;
694: 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. */
695: PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd;
696: PetscScalar *aa = a->a,*ba = b->a;
699: /* Iterate over all rows of the matrix */
700: for (j=0; j<am; j++) {
701: dnz_row = onz_row = 0;
702: rowstart_offd = full_offd_i[j];
703: rowstart_diag = full_diag_i[j];
704: /* Iterate over all non-zero columns of the current row */
705: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
706: /* If column is in the diagonal */
707: if (mat_j[col] >= cstart && mat_j[col] < cend) {
708: aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
709: aa[rowstart_diag+dnz_row] = mat_a[col];
710: dnz_row++;
711: } else { /* off-diagonal entries */
712: bj[rowstart_offd+onz_row] = mat_j[col];
713: ba[rowstart_offd+onz_row] = mat_a[col];
714: onz_row++;
715: }
716: }
717: ailen[j] = dnz_row;
718: bilen[j] = onz_row;
719: }
720: return(0);
721: }
723: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
724: {
725: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
727: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
728: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
731: for (i=0; i<m; i++) {
732: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
733: 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);
734: if (idxm[i] >= rstart && idxm[i] < rend) {
735: row = idxm[i] - rstart;
736: for (j=0; j<n; j++) {
737: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
738: 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);
739: if (idxn[j] >= cstart && idxn[j] < cend) {
740: col = idxn[j] - cstart;
741: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
742: } else {
743: if (!aij->colmap) {
744: MatCreateColmap_MPIAIJ_Private(mat);
745: }
746: #if defined(PETSC_USE_CTABLE)
747: PetscTableFind(aij->colmap,idxn[j]+1,&col);
748: col--;
749: #else
750: col = aij->colmap[idxn[j]] - 1;
751: #endif
752: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
753: else {
754: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
755: }
756: }
757: }
758: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
759: }
760: return(0);
761: }
763: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);
765: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
766: {
767: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
769: PetscInt nstash,reallocs;
772: if (aij->donotstash || mat->nooffprocentries) return(0);
774: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
775: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
776: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
777: return(0);
778: }
780: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
781: {
782: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
783: Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data;
785: PetscMPIInt n;
786: PetscInt i,j,rstart,ncols,flg;
787: PetscInt *row,*col;
788: PetscBool other_disassembled;
789: PetscScalar *val;
791: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
794: if (!aij->donotstash && !mat->nooffprocentries) {
795: while (1) {
796: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
797: if (!flg) break;
799: for (i=0; i<n; ) {
800: /* Now identify the consecutive vals belonging to the same row */
801: for (j=i,rstart=row[j]; j<n; j++) {
802: if (row[j] != rstart) break;
803: }
804: if (j < n) ncols = j-i;
805: else ncols = n-i;
806: /* Now assemble all these values with a single function call */
807: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);
809: i = j;
810: }
811: }
812: MatStashScatterEnd_Private(&mat->stash);
813: }
814: MatAssemblyBegin(aij->A,mode);
815: MatAssemblyEnd(aij->A,mode);
817: /* determine if any processor has disassembled, if so we must
818: also disassemble ourselfs, in order that we may reassemble. */
819: /*
820: if nonzero structure of submatrix B cannot change then we know that
821: no processor disassembled thus we can skip this stuff
822: */
823: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
824: MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
825: if (mat->was_assembled && !other_disassembled) {
826: MatDisAssemble_MPIAIJ(mat);
827: }
828: }
829: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
830: MatSetUpMultiply_MPIAIJ(mat);
831: }
832: MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
833: MatAssemblyBegin(aij->B,mode);
834: MatAssemblyEnd(aij->B,mode);
836: PetscFree2(aij->rowvalues,aij->rowindices);
838: aij->rowvalues = 0;
840: VecDestroy(&aij->diag);
841: if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
843: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
844: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
845: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
846: MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
847: }
848: return(0);
849: }
851: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
852: {
853: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
857: MatZeroEntries(l->A);
858: MatZeroEntries(l->B);
859: return(0);
860: }
862: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
863: {
864: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
865: PetscInt *lrows;
866: PetscInt r, len;
867: PetscBool cong;
871: /* get locally owned rows */
872: MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);
873: /* fix right hand side if needed */
874: if (x && b) {
875: const PetscScalar *xx;
876: PetscScalar *bb;
878: VecGetArrayRead(x, &xx);
879: VecGetArray(b, &bb);
880: for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
881: VecRestoreArrayRead(x, &xx);
882: VecRestoreArray(b, &bb);
883: }
884: /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/
885: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
886: MatHasCongruentLayouts(A,&cong);
887: if ((diag != 0.0) && cong) {
888: MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
889: } else if (diag != 0.0) {
890: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
891: if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
892: for (r = 0; r < len; ++r) {
893: const PetscInt row = lrows[r] + A->rmap->rstart;
894: MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
895: }
896: MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);
897: MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);
898: } else {
899: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
900: }
901: PetscFree(lrows);
903: /* only change matrix nonzero state if pattern was allowed to be changed */
904: if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) {
905: PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
906: MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
907: }
908: return(0);
909: }
911: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
912: {
913: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
914: PetscErrorCode ierr;
915: PetscMPIInt n = A->rmap->n;
916: PetscInt i,j,r,m,p = 0,len = 0;
917: PetscInt *lrows,*owners = A->rmap->range;
918: PetscSFNode *rrows;
919: PetscSF sf;
920: const PetscScalar *xx;
921: PetscScalar *bb,*mask;
922: Vec xmask,lmask;
923: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data;
924: const PetscInt *aj, *ii,*ridx;
925: PetscScalar *aa;
928: /* Create SF where leaves are input rows and roots are owned rows */
929: PetscMalloc1(n, &lrows);
930: for (r = 0; r < n; ++r) lrows[r] = -1;
931: PetscMalloc1(N, &rrows);
932: for (r = 0; r < N; ++r) {
933: const PetscInt idx = rows[r];
934: 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);
935: if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
936: PetscLayoutFindOwner(A->rmap,idx,&p);
937: }
938: rrows[r].rank = p;
939: rrows[r].index = rows[r] - owners[p];
940: }
941: PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
942: PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
943: /* Collect flags for rows to be zeroed */
944: PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
945: PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
946: PetscSFDestroy(&sf);
947: /* Compress and put in row numbers */
948: for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
949: /* zero diagonal part of matrix */
950: MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
951: /* handle off diagonal part of matrix */
952: MatCreateVecs(A,&xmask,NULL);
953: VecDuplicate(l->lvec,&lmask);
954: VecGetArray(xmask,&bb);
955: for (i=0; i<len; i++) bb[lrows[i]] = 1;
956: VecRestoreArray(xmask,&bb);
957: VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
958: VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
959: VecDestroy(&xmask);
960: if (x) {
961: VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
962: VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
963: VecGetArrayRead(l->lvec,&xx);
964: VecGetArray(b,&bb);
965: }
966: VecGetArray(lmask,&mask);
967: /* remove zeroed rows of off diagonal matrix */
968: ii = aij->i;
969: for (i=0; i<len; i++) {
970: PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));
971: }
972: /* loop over all elements of off process part of matrix zeroing removed columns*/
973: if (aij->compressedrow.use) {
974: m = aij->compressedrow.nrows;
975: ii = aij->compressedrow.i;
976: ridx = aij->compressedrow.rindex;
977: for (i=0; i<m; i++) {
978: n = ii[i+1] - ii[i];
979: aj = aij->j + ii[i];
980: aa = aij->a + ii[i];
982: for (j=0; j<n; j++) {
983: if (PetscAbsScalar(mask[*aj])) {
984: if (b) bb[*ridx] -= *aa*xx[*aj];
985: *aa = 0.0;
986: }
987: aa++;
988: aj++;
989: }
990: ridx++;
991: }
992: } else { /* do not use compressed row format */
993: m = l->B->rmap->n;
994: for (i=0; i<m; i++) {
995: n = ii[i+1] - ii[i];
996: aj = aij->j + ii[i];
997: aa = aij->a + ii[i];
998: for (j=0; j<n; j++) {
999: if (PetscAbsScalar(mask[*aj])) {
1000: if (b) bb[i] -= *aa*xx[*aj];
1001: *aa = 0.0;
1002: }
1003: aa++;
1004: aj++;
1005: }
1006: }
1007: }
1008: if (x) {
1009: VecRestoreArray(b,&bb);
1010: VecRestoreArrayRead(l->lvec,&xx);
1011: }
1012: VecRestoreArray(lmask,&mask);
1013: VecDestroy(&lmask);
1014: PetscFree(lrows);
1016: /* only change matrix nonzero state if pattern was allowed to be changed */
1017: if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
1018: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1019: MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1020: }
1021: return(0);
1022: }
1024: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1025: {
1026: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1028: PetscInt nt;
1029: VecScatter Mvctx = a->Mvctx;
1032: VecGetLocalSize(xx,&nt);
1033: 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);
1035: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1036: (*a->A->ops->mult)(a->A,xx,yy);
1037: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1038: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1039: return(0);
1040: }
1042: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1043: {
1044: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1048: MatMultDiagonalBlock(a->A,bb,xx);
1049: return(0);
1050: }
1052: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1053: {
1054: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1056: VecScatter Mvctx = a->Mvctx;
1059: if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1;
1060: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1061: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1062: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1063: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1064: return(0);
1065: }
1067: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1068: {
1069: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1071: PetscBool merged;
1074: VecScatterGetMerged(a->Mvctx,&merged);
1075: /* do nondiagonal part */
1076: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1077: if (!merged) {
1078: /* send it on its way */
1079: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1080: /* do local part */
1081: (*a->A->ops->multtranspose)(a->A,xx,yy);
1082: /* receive remote parts: note this assumes the values are not actually */
1083: /* added in yy until the next line, */
1084: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1085: } else {
1086: /* do local part */
1087: (*a->A->ops->multtranspose)(a->A,xx,yy);
1088: /* send it on its way */
1089: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1090: /* values actually were received in the Begin() but we need to call this nop */
1091: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1092: }
1093: return(0);
1094: }
1096: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f)
1097: {
1098: MPI_Comm comm;
1099: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1100: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1101: IS Me,Notme;
1103: PetscInt M,N,first,last,*notme,i;
1104: PetscMPIInt size;
1107: /* Easy test: symmetric diagonal block */
1108: Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1109: MatIsTranspose(Adia,Bdia,tol,f);
1110: if (!*f) return(0);
1111: PetscObjectGetComm((PetscObject)Amat,&comm);
1112: MPI_Comm_size(comm,&size);
1113: if (size == 1) return(0);
1115: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1116: MatGetSize(Amat,&M,&N);
1117: MatGetOwnershipRange(Amat,&first,&last);
1118: PetscMalloc1(N-last+first,¬me);
1119: for (i=0; i<first; i++) notme[i] = i;
1120: for (i=last; i<M; i++) notme[i-last+first] = i;
1121: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1122: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1123: MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1124: Aoff = Aoffs[0];
1125: MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1126: Boff = Boffs[0];
1127: MatIsTranspose(Aoff,Boff,tol,f);
1128: MatDestroyMatrices(1,&Aoffs);
1129: MatDestroyMatrices(1,&Boffs);
1130: ISDestroy(&Me);
1131: ISDestroy(&Notme);
1132: PetscFree(notme);
1133: return(0);
1134: }
1136: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f)
1137: {
1141: MatIsTranspose_MPIAIJ(A,A,tol,f);
1142: return(0);
1143: }
1145: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1146: {
1147: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1151: /* do nondiagonal part */
1152: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1153: /* send it on its way */
1154: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1155: /* do local part */
1156: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1157: /* receive remote parts */
1158: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1159: return(0);
1160: }
1162: /*
1163: This only works correctly for square matrices where the subblock A->A is the
1164: diagonal block
1165: */
1166: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1167: {
1169: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1172: if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1173: 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");
1174: MatGetDiagonal(a->A,v);
1175: return(0);
1176: }
1178: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1179: {
1180: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1184: MatScale(a->A,aa);
1185: MatScale(a->B,aa);
1186: return(0);
1187: }
1189: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1190: {
1191: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1195: #if defined(PETSC_USE_LOG)
1196: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1197: #endif
1198: MatStashDestroy_Private(&mat->stash);
1199: VecDestroy(&aij->diag);
1200: MatDestroy(&aij->A);
1201: MatDestroy(&aij->B);
1202: #if defined(PETSC_USE_CTABLE)
1203: PetscTableDestroy(&aij->colmap);
1204: #else
1205: PetscFree(aij->colmap);
1206: #endif
1207: PetscFree(aij->garray);
1208: VecDestroy(&aij->lvec);
1209: VecScatterDestroy(&aij->Mvctx);
1210: if (aij->Mvctx_mpi1) {VecScatterDestroy(&aij->Mvctx_mpi1);}
1211: PetscFree2(aij->rowvalues,aij->rowindices);
1212: PetscFree(aij->ld);
1213: PetscFree(mat->data);
1215: PetscObjectChangeTypeName((PetscObject)mat,0);
1216: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1217: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1218: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1219: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1220: PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);
1221: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1222: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1223: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1224: #if defined(PETSC_HAVE_ELEMENTAL)
1225: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1226: #endif
1227: #if defined(PETSC_HAVE_HYPRE)
1228: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);
1229: PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);
1230: #endif
1231: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);
1232: PetscObjectComposeFunction((PetscObject)mat,"MatPtAP_is_mpiaij_C",NULL);
1233: return(0);
1234: }
1236: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1237: {
1238: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1239: Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data;
1240: Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data;
1242: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
1243: int fd;
1244: PetscInt nz,header[4],*row_lengths,*range=0,rlen,i;
1245: PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0;
1246: PetscScalar *column_values;
1247: PetscInt message_count,flowcontrolcount;
1248: FILE *file;
1251: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1252: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1253: nz = A->nz + B->nz;
1254: PetscViewerBinaryGetDescriptor(viewer,&fd);
1255: if (!rank) {
1256: header[0] = MAT_FILE_CLASSID;
1257: header[1] = mat->rmap->N;
1258: header[2] = mat->cmap->N;
1260: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1261: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1262: /* get largest number of rows any processor has */
1263: rlen = mat->rmap->n;
1264: range = mat->rmap->range;
1265: for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]);
1266: } else {
1267: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1268: rlen = mat->rmap->n;
1269: }
1271: /* load up the local row counts */
1272: PetscMalloc1(rlen+1,&row_lengths);
1273: for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1275: /* store the row lengths to the file */
1276: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1277: if (!rank) {
1278: PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1279: for (i=1; i<size; i++) {
1280: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1281: rlen = range[i+1] - range[i];
1282: MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1283: PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1284: }
1285: PetscViewerFlowControlEndMaster(viewer,&message_count);
1286: } else {
1287: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1288: MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1289: PetscViewerFlowControlEndWorker(viewer,&message_count);
1290: }
1291: PetscFree(row_lengths);
1293: /* load up the local column indices */
1294: nzmax = nz; /* th processor needs space a largest processor needs */
1295: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1296: PetscMalloc1(nzmax+1,&column_indices);
1297: cnt = 0;
1298: for (i=0; i<mat->rmap->n; i++) {
1299: for (j=B->i[i]; j<B->i[i+1]; j++) {
1300: if ((col = garray[B->j[j]]) > cstart) break;
1301: column_indices[cnt++] = col;
1302: }
1303: for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart;
1304: for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]];
1305: }
1306: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1308: /* store the column indices to the file */
1309: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1310: if (!rank) {
1311: MPI_Status status;
1312: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1313: for (i=1; i<size; i++) {
1314: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1315: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1316: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1317: MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1318: PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1319: }
1320: PetscViewerFlowControlEndMaster(viewer,&message_count);
1321: } else {
1322: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1323: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1324: MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1325: PetscViewerFlowControlEndWorker(viewer,&message_count);
1326: }
1327: PetscFree(column_indices);
1329: /* load up the local column values */
1330: PetscMalloc1(nzmax+1,&column_values);
1331: cnt = 0;
1332: for (i=0; i<mat->rmap->n; i++) {
1333: for (j=B->i[i]; j<B->i[i+1]; j++) {
1334: if (garray[B->j[j]] > cstart) break;
1335: column_values[cnt++] = B->a[j];
1336: }
1337: for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k];
1338: for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j];
1339: }
1340: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1342: /* store the column values to the file */
1343: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1344: if (!rank) {
1345: MPI_Status status;
1346: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1347: for (i=1; i<size; i++) {
1348: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1349: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1350: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1351: MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));
1352: PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1353: }
1354: PetscViewerFlowControlEndMaster(viewer,&message_count);
1355: } else {
1356: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1357: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1358: MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1359: PetscViewerFlowControlEndWorker(viewer,&message_count);
1360: }
1361: PetscFree(column_values);
1363: PetscViewerBinaryGetInfoPointer(viewer,&file);
1364: if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs));
1365: return(0);
1366: }
1368: #include <petscdraw.h>
1369: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1370: {
1371: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1372: PetscErrorCode ierr;
1373: PetscMPIInt rank = aij->rank,size = aij->size;
1374: PetscBool isdraw,iascii,isbinary;
1375: PetscViewer sviewer;
1376: PetscViewerFormat format;
1379: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1380: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1381: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1382: if (iascii) {
1383: PetscViewerGetFormat(viewer,&format);
1384: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1385: PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz;
1386: PetscMalloc1(size,&nz);
1387: MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
1388: for (i=0; i<(PetscInt)size; i++) {
1389: nmax = PetscMax(nmax,nz[i]);
1390: nmin = PetscMin(nmin,nz[i]);
1391: navg += nz[i];
1392: }
1393: PetscFree(nz);
1394: navg = navg/size;
1395: PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);
1396: return(0);
1397: }
1398: PetscViewerGetFormat(viewer,&format);
1399: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1400: MatInfo info;
1401: PetscBool inodes;
1403: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1404: MatGetInfo(mat,MAT_LOCAL,&info);
1405: MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1406: PetscViewerASCIIPushSynchronized(viewer);
1407: if (!inodes) {
1408: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n",
1409: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1410: } else {
1411: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n",
1412: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1413: }
1414: MatGetInfo(aij->A,MAT_LOCAL,&info);
1415: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1416: MatGetInfo(aij->B,MAT_LOCAL,&info);
1417: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1418: PetscViewerFlush(viewer);
1419: PetscViewerASCIIPopSynchronized(viewer);
1420: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1421: VecScatterView(aij->Mvctx,viewer);
1422: return(0);
1423: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1424: PetscInt inodecount,inodelimit,*inodes;
1425: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1426: if (inodes) {
1427: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1428: } else {
1429: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1430: }
1431: return(0);
1432: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1433: return(0);
1434: }
1435: } else if (isbinary) {
1436: if (size == 1) {
1437: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1438: MatView(aij->A,viewer);
1439: } else {
1440: MatView_MPIAIJ_Binary(mat,viewer);
1441: }
1442: return(0);
1443: } else if (isdraw) {
1444: PetscDraw draw;
1445: PetscBool isnull;
1446: PetscViewerDrawGetDraw(viewer,0,&draw);
1447: PetscDrawIsNull(draw,&isnull);
1448: if (isnull) return(0);
1449: }
1451: {
1452: /* assemble the entire matrix onto first processor. */
1453: Mat A;
1454: Mat_SeqAIJ *Aloc;
1455: PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct;
1456: MatScalar *a;
1458: MatCreate(PetscObjectComm((PetscObject)mat),&A);
1459: if (!rank) {
1460: MatSetSizes(A,M,N,M,N);
1461: } else {
1462: MatSetSizes(A,0,0,M,N);
1463: }
1464: /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
1465: MatSetType(A,MATMPIAIJ);
1466: MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);
1467: MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1468: PetscLogObjectParent((PetscObject)mat,(PetscObject)A);
1470: /* copy over the A part */
1471: Aloc = (Mat_SeqAIJ*)aij->A->data;
1472: m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1473: row = mat->rmap->rstart;
1474: for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart;
1475: for (i=0; i<m; i++) {
1476: MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
1477: row++;
1478: a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1479: }
1480: aj = Aloc->j;
1481: for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart;
1483: /* copy over the B part */
1484: Aloc = (Mat_SeqAIJ*)aij->B->data;
1485: m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1486: row = mat->rmap->rstart;
1487: PetscMalloc1(ai[m]+1,&cols);
1488: ct = cols;
1489: for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]];
1490: for (i=0; i<m; i++) {
1491: MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
1492: row++;
1493: a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1494: }
1495: PetscFree(ct);
1496: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1497: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1498: /*
1499: Everyone has to call to draw the matrix since the graphics waits are
1500: synchronized across all processors that share the PetscDraw object
1501: */
1502: PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1503: if (!rank) {
1504: PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);
1505: MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);
1506: }
1507: PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1508: PetscViewerFlush(viewer);
1509: MatDestroy(&A);
1510: }
1511: return(0);
1512: }
1514: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1515: {
1517: PetscBool iascii,isdraw,issocket,isbinary;
1520: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1521: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1522: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1523: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1524: if (iascii || isdraw || isbinary || issocket) {
1525: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1526: }
1527: return(0);
1528: }
1530: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1531: {
1532: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1534: Vec bb1 = 0;
1535: PetscBool hasop;
1538: if (flag == SOR_APPLY_UPPER) {
1539: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1540: return(0);
1541: }
1543: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1544: VecDuplicate(bb,&bb1);
1545: }
1547: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1548: if (flag & SOR_ZERO_INITIAL_GUESS) {
1549: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1550: its--;
1551: }
1553: while (its--) {
1554: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1555: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1557: /* update rhs: bb1 = bb - B*x */
1558: VecScale(mat->lvec,-1.0);
1559: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1561: /* local sweep */
1562: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1563: }
1564: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1565: if (flag & SOR_ZERO_INITIAL_GUESS) {
1566: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1567: its--;
1568: }
1569: while (its--) {
1570: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1571: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1573: /* update rhs: bb1 = bb - B*x */
1574: VecScale(mat->lvec,-1.0);
1575: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1577: /* local sweep */
1578: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1579: }
1580: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1581: if (flag & SOR_ZERO_INITIAL_GUESS) {
1582: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1583: its--;
1584: }
1585: while (its--) {
1586: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1587: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1589: /* update rhs: bb1 = bb - B*x */
1590: VecScale(mat->lvec,-1.0);
1591: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1593: /* local sweep */
1594: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1595: }
1596: } else if (flag & SOR_EISENSTAT) {
1597: Vec xx1;
1599: VecDuplicate(bb,&xx1);
1600: (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);
1602: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1603: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1604: if (!mat->diag) {
1605: MatCreateVecs(matin,&mat->diag,NULL);
1606: MatGetDiagonal(matin,mat->diag);
1607: }
1608: MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1609: if (hasop) {
1610: MatMultDiagonalBlock(matin,xx,bb1);
1611: } else {
1612: VecPointwiseMult(bb1,mat->diag,xx);
1613: }
1614: VecAYPX(bb1,(omega-2.0)/omega,bb);
1616: MatMultAdd(mat->B,mat->lvec,bb1,bb1);
1618: /* local sweep */
1619: (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1620: VecAXPY(xx,1.0,xx1);
1621: VecDestroy(&xx1);
1622: } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");
1624: VecDestroy(&bb1);
1626: matin->factorerrortype = mat->A->factorerrortype;
1627: return(0);
1628: }
1630: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1631: {
1632: Mat aA,aB,Aperm;
1633: const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1634: PetscScalar *aa,*ba;
1635: PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1636: PetscSF rowsf,sf;
1637: IS parcolp = NULL;
1638: PetscBool done;
1642: MatGetLocalSize(A,&m,&n);
1643: ISGetIndices(rowp,&rwant);
1644: ISGetIndices(colp,&cwant);
1645: PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);
1647: /* Invert row permutation to find out where my rows should go */
1648: PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1649: PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1650: PetscSFSetFromOptions(rowsf);
1651: for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1652: PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1653: PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1655: /* Invert column permutation to find out where my columns should go */
1656: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1657: PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1658: PetscSFSetFromOptions(sf);
1659: for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1660: PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1661: PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1662: PetscSFDestroy(&sf);
1664: ISRestoreIndices(rowp,&rwant);
1665: ISRestoreIndices(colp,&cwant);
1666: MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);
1668: /* Find out where my gcols should go */
1669: MatGetSize(aB,NULL,&ng);
1670: PetscMalloc1(ng,&gcdest);
1671: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1672: PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1673: PetscSFSetFromOptions(sf);
1674: PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1675: PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1676: PetscSFDestroy(&sf);
1678: PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1679: MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1680: MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1681: for (i=0; i<m; i++) {
1682: PetscInt row = rdest[i],rowner;
1683: PetscLayoutFindOwner(A->rmap,row,&rowner);
1684: for (j=ai[i]; j<ai[i+1]; j++) {
1685: PetscInt cowner,col = cdest[aj[j]];
1686: PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1687: if (rowner == cowner) dnnz[i]++;
1688: else onnz[i]++;
1689: }
1690: for (j=bi[i]; j<bi[i+1]; j++) {
1691: PetscInt cowner,col = gcdest[bj[j]];
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: PetscReal 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_REAL,MPIU_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_REAL,MPIU_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: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1819: break;
1820: case MAT_IGNORE_OFF_PROC_ENTRIES:
1821: a->donotstash = flg;
1822: break;
1823: case MAT_SPD:
1824: A->spd_set = PETSC_TRUE;
1825: A->spd = flg;
1826: if (flg) {
1827: A->symmetric = PETSC_TRUE;
1828: A->structurally_symmetric = PETSC_TRUE;
1829: A->symmetric_set = PETSC_TRUE;
1830: A->structurally_symmetric_set = PETSC_TRUE;
1831: }
1832: break;
1833: case MAT_SYMMETRIC:
1834: MatCheckPreallocated(A,1);
1835: MatSetOption(a->A,op,flg);
1836: break;
1837: case MAT_STRUCTURALLY_SYMMETRIC:
1838: MatCheckPreallocated(A,1);
1839: MatSetOption(a->A,op,flg);
1840: break;
1841: case MAT_HERMITIAN:
1842: MatCheckPreallocated(A,1);
1843: MatSetOption(a->A,op,flg);
1844: break;
1845: case MAT_SYMMETRY_ETERNAL:
1846: MatCheckPreallocated(A,1);
1847: MatSetOption(a->A,op,flg);
1848: break;
1849: case MAT_SUBMAT_SINGLEIS:
1850: A->submat_singleis = flg;
1851: break;
1852: case MAT_STRUCTURE_ONLY:
1853: /* The option is handled directly by MatSetOption() */
1854: break;
1855: default:
1856: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1857: }
1858: return(0);
1859: }
1861: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1862: {
1863: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1864: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1866: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1867: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1868: PetscInt *cmap,*idx_p;
1871: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1872: mat->getrowactive = PETSC_TRUE;
1874: if (!mat->rowvalues && (idx || v)) {
1875: /*
1876: allocate enough space to hold information from the longest row.
1877: */
1878: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1879: PetscInt max = 1,tmp;
1880: for (i=0; i<matin->rmap->n; i++) {
1881: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1882: if (max < tmp) max = tmp;
1883: }
1884: PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1885: }
1887: if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1888: lrow = row - rstart;
1890: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1891: if (!v) {pvA = 0; pvB = 0;}
1892: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1893: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1894: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1895: nztot = nzA + nzB;
1897: cmap = mat->garray;
1898: if (v || idx) {
1899: if (nztot) {
1900: /* Sort by increasing column numbers, assuming A and B already sorted */
1901: PetscInt imark = -1;
1902: if (v) {
1903: *v = v_p = mat->rowvalues;
1904: for (i=0; i<nzB; i++) {
1905: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1906: else break;
1907: }
1908: imark = i;
1909: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1910: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1911: }
1912: if (idx) {
1913: *idx = idx_p = mat->rowindices;
1914: if (imark > -1) {
1915: for (i=0; i<imark; i++) {
1916: idx_p[i] = cmap[cworkB[i]];
1917: }
1918: } else {
1919: for (i=0; i<nzB; i++) {
1920: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1921: else break;
1922: }
1923: imark = i;
1924: }
1925: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1926: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1927: }
1928: } else {
1929: if (idx) *idx = 0;
1930: if (v) *v = 0;
1931: }
1932: }
1933: *nz = nztot;
1934: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1935: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1936: return(0);
1937: }
1939: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1940: {
1941: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1944: if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1945: aij->getrowactive = PETSC_FALSE;
1946: return(0);
1947: }
1949: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1950: {
1951: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1952: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1954: PetscInt i,j,cstart = mat->cmap->rstart;
1955: PetscReal sum = 0.0;
1956: MatScalar *v;
1959: if (aij->size == 1) {
1960: MatNorm(aij->A,type,norm);
1961: } else {
1962: if (type == NORM_FROBENIUS) {
1963: v = amat->a;
1964: for (i=0; i<amat->nz; i++) {
1965: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1966: }
1967: v = bmat->a;
1968: for (i=0; i<bmat->nz; i++) {
1969: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1970: }
1971: MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1972: *norm = PetscSqrtReal(*norm);
1973: PetscLogFlops(2*amat->nz+2*bmat->nz);
1974: } else if (type == NORM_1) { /* max column norm */
1975: PetscReal *tmp,*tmp2;
1976: PetscInt *jj,*garray = aij->garray;
1977: PetscCalloc1(mat->cmap->N+1,&tmp);
1978: PetscMalloc1(mat->cmap->N+1,&tmp2);
1979: *norm = 0.0;
1980: v = amat->a; jj = amat->j;
1981: for (j=0; j<amat->nz; j++) {
1982: tmp[cstart + *jj++] += PetscAbsScalar(*v); v++;
1983: }
1984: v = bmat->a; jj = bmat->j;
1985: for (j=0; j<bmat->nz; j++) {
1986: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1987: }
1988: MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1989: for (j=0; j<mat->cmap->N; j++) {
1990: if (tmp2[j] > *norm) *norm = tmp2[j];
1991: }
1992: PetscFree(tmp);
1993: PetscFree(tmp2);
1994: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1995: } else if (type == NORM_INFINITY) { /* max row norm */
1996: PetscReal ntemp = 0.0;
1997: for (j=0; j<aij->A->rmap->n; j++) {
1998: v = amat->a + amat->i[j];
1999: sum = 0.0;
2000: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
2001: sum += PetscAbsScalar(*v); v++;
2002: }
2003: v = bmat->a + bmat->i[j];
2004: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
2005: sum += PetscAbsScalar(*v); v++;
2006: }
2007: if (sum > ntemp) ntemp = sum;
2008: }
2009: MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
2010: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
2011: } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
2012: }
2013: return(0);
2014: }
2016: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
2017: {
2018: Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b;
2019: Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag;
2020: PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,*B_diag_ilen,*B_diag_i,i,ncol,A_diag_ncol;
2022: Mat B,A_diag,*B_diag;
2023: MatScalar *array;
2026: ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
2027: ai = Aloc->i; aj = Aloc->j;
2028: bi = Bloc->i; bj = Bloc->j;
2029: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2030: PetscInt *d_nnz,*g_nnz,*o_nnz;
2031: PetscSFNode *oloc;
2032: PETSC_UNUSED PetscSF sf;
2034: PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
2035: /* compute d_nnz for preallocation */
2036: PetscMemzero(d_nnz,na*sizeof(PetscInt));
2037: for (i=0; i<ai[ma]; i++) {
2038: d_nnz[aj[i]]++;
2039: }
2040: /* compute local off-diagonal contributions */
2041: PetscMemzero(g_nnz,nb*sizeof(PetscInt));
2042: for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
2043: /* map those to global */
2044: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2045: PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2046: PetscSFSetFromOptions(sf);
2047: PetscMemzero(o_nnz,na*sizeof(PetscInt));
2048: PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2049: PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2050: PetscSFDestroy(&sf);
2052: MatCreate(PetscObjectComm((PetscObject)A),&B);
2053: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2054: MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2055: MatSetType(B,((PetscObject)A)->type_name);
2056: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2057: PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2058: } else {
2059: B = *matout;
2060: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2061: }
2063: b = (Mat_MPIAIJ*)B->data;
2064: A_diag = a->A;
2065: B_diag = &b->A;
2066: sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data;
2067: A_diag_ncol = A_diag->cmap->N;
2068: B_diag_ilen = sub_B_diag->ilen;
2069: B_diag_i = sub_B_diag->i;
2071: /* Set ilen for diagonal of B */
2072: for (i=0; i<A_diag_ncol; i++) {
2073: B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i];
2074: }
2076: /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
2077: very quickly (=without using MatSetValues), because all writes are local. */
2078: MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);
2080: /* copy over the B part */
2081: PetscCalloc1(bi[mb],&cols);
2082: array = Bloc->a;
2083: row = A->rmap->rstart;
2084: for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2085: cols_tmp = cols;
2086: for (i=0; i<mb; i++) {
2087: ncol = bi[i+1]-bi[i];
2088: MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2089: row++;
2090: array += ncol; cols_tmp += ncol;
2091: }
2092: PetscFree(cols);
2094: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2095: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2096: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2097: *matout = B;
2098: } else {
2099: MatHeaderMerge(A,&B);
2100: }
2101: return(0);
2102: }
2104: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2105: {
2106: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2107: Mat a = aij->A,b = aij->B;
2109: PetscInt s1,s2,s3;
2112: MatGetLocalSize(mat,&s2,&s3);
2113: if (rr) {
2114: VecGetLocalSize(rr,&s1);
2115: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2116: /* Overlap communication with computation. */
2117: VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2118: }
2119: if (ll) {
2120: VecGetLocalSize(ll,&s1);
2121: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2122: (*b->ops->diagonalscale)(b,ll,0);
2123: }
2124: /* scale the diagonal block */
2125: (*a->ops->diagonalscale)(a,ll,rr);
2127: if (rr) {
2128: /* Do a scatter end and then right scale the off-diagonal block */
2129: VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2130: (*b->ops->diagonalscale)(b,0,aij->lvec);
2131: }
2132: return(0);
2133: }
2135: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2136: {
2137: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2141: MatSetUnfactored(a->A);
2142: return(0);
2143: }
2145: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag)
2146: {
2147: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2148: Mat a,b,c,d;
2149: PetscBool flg;
2153: a = matA->A; b = matA->B;
2154: c = matB->A; d = matB->B;
2156: MatEqual(a,c,&flg);
2157: if (flg) {
2158: MatEqual(b,d,&flg);
2159: }
2160: MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2161: return(0);
2162: }
2164: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2165: {
2167: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2168: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
2171: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2172: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2173: /* because of the column compression in the off-processor part of the matrix a->B,
2174: the number of columns in a->B and b->B may be different, hence we cannot call
2175: the MatCopy() directly on the two parts. If need be, we can provide a more
2176: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2177: then copying the submatrices */
2178: MatCopy_Basic(A,B,str);
2179: } else {
2180: MatCopy(a->A,b->A,str);
2181: MatCopy(a->B,b->B,str);
2182: }
2183: PetscObjectStateIncrease((PetscObject)B);
2184: return(0);
2185: }
2187: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2188: {
2192: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2193: return(0);
2194: }
2196: /*
2197: Computes the number of nonzeros per row needed for preallocation when X and Y
2198: have different nonzero structure.
2199: */
2200: 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)
2201: {
2202: PetscInt i,j,k,nzx,nzy;
2205: /* Set the number of nonzeros in the new matrix */
2206: for (i=0; i<m; i++) {
2207: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2208: nzx = xi[i+1] - xi[i];
2209: nzy = yi[i+1] - yi[i];
2210: nnz[i] = 0;
2211: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2212: for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2213: if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */
2214: nnz[i]++;
2215: }
2216: for (; k<nzy; k++) nnz[i]++;
2217: }
2218: return(0);
2219: }
2221: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2222: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2223: {
2225: PetscInt m = Y->rmap->N;
2226: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2227: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2230: MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2231: return(0);
2232: }
2234: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2235: {
2237: Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2238: PetscBLASInt bnz,one=1;
2239: Mat_SeqAIJ *x,*y;
2242: if (str == SAME_NONZERO_PATTERN) {
2243: PetscScalar alpha = a;
2244: x = (Mat_SeqAIJ*)xx->A->data;
2245: PetscBLASIntCast(x->nz,&bnz);
2246: y = (Mat_SeqAIJ*)yy->A->data;
2247: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2248: x = (Mat_SeqAIJ*)xx->B->data;
2249: y = (Mat_SeqAIJ*)yy->B->data;
2250: PetscBLASIntCast(x->nz,&bnz);
2251: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2252: PetscObjectStateIncrease((PetscObject)Y);
2253: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2254: MatAXPY_Basic(Y,a,X,str);
2255: } else {
2256: Mat B;
2257: PetscInt *nnz_d,*nnz_o;
2258: PetscMalloc1(yy->A->rmap->N,&nnz_d);
2259: PetscMalloc1(yy->B->rmap->N,&nnz_o);
2260: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2261: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2262: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2263: MatSetBlockSizesFromMats(B,Y,Y);
2264: MatSetType(B,MATMPIAIJ);
2265: MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2266: MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2267: MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2268: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2269: MatHeaderReplace(Y,&B);
2270: PetscFree(nnz_d);
2271: PetscFree(nnz_o);
2272: }
2273: return(0);
2274: }
2276: extern PetscErrorCode MatConjugate_SeqAIJ(Mat);
2278: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2279: {
2280: #if defined(PETSC_USE_COMPLEX)
2282: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2285: MatConjugate_SeqAIJ(aij->A);
2286: MatConjugate_SeqAIJ(aij->B);
2287: #else
2289: #endif
2290: return(0);
2291: }
2293: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2294: {
2295: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2299: MatRealPart(a->A);
2300: MatRealPart(a->B);
2301: return(0);
2302: }
2304: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2305: {
2306: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2310: MatImaginaryPart(a->A);
2311: MatImaginaryPart(a->B);
2312: return(0);
2313: }
2315: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2316: {
2317: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2319: PetscInt i,*idxb = 0;
2320: PetscScalar *va,*vb;
2321: Vec vtmp;
2324: MatGetRowMaxAbs(a->A,v,idx);
2325: VecGetArray(v,&va);
2326: if (idx) {
2327: for (i=0; i<A->rmap->n; i++) {
2328: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2329: }
2330: }
2332: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2333: if (idx) {
2334: PetscMalloc1(A->rmap->n,&idxb);
2335: }
2336: MatGetRowMaxAbs(a->B,vtmp,idxb);
2337: VecGetArray(vtmp,&vb);
2339: for (i=0; i<A->rmap->n; i++) {
2340: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2341: va[i] = vb[i];
2342: if (idx) idx[i] = a->garray[idxb[i]];
2343: }
2344: }
2346: VecRestoreArray(v,&va);
2347: VecRestoreArray(vtmp,&vb);
2348: PetscFree(idxb);
2349: VecDestroy(&vtmp);
2350: return(0);
2351: }
2353: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2354: {
2355: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2357: PetscInt i,*idxb = 0;
2358: PetscScalar *va,*vb;
2359: Vec vtmp;
2362: MatGetRowMinAbs(a->A,v,idx);
2363: VecGetArray(v,&va);
2364: if (idx) {
2365: for (i=0; i<A->cmap->n; i++) {
2366: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2367: }
2368: }
2370: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2371: if (idx) {
2372: PetscMalloc1(A->rmap->n,&idxb);
2373: }
2374: MatGetRowMinAbs(a->B,vtmp,idxb);
2375: VecGetArray(vtmp,&vb);
2377: for (i=0; i<A->rmap->n; i++) {
2378: if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2379: va[i] = vb[i];
2380: if (idx) idx[i] = a->garray[idxb[i]];
2381: }
2382: }
2384: VecRestoreArray(v,&va);
2385: VecRestoreArray(vtmp,&vb);
2386: PetscFree(idxb);
2387: VecDestroy(&vtmp);
2388: return(0);
2389: }
2391: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2392: {
2393: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2394: PetscInt n = A->rmap->n;
2395: PetscInt cstart = A->cmap->rstart;
2396: PetscInt *cmap = mat->garray;
2397: PetscInt *diagIdx, *offdiagIdx;
2398: Vec diagV, offdiagV;
2399: PetscScalar *a, *diagA, *offdiagA;
2400: PetscInt r;
2404: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2405: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);
2406: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);
2407: MatGetRowMin(mat->A, diagV, diagIdx);
2408: MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2409: VecGetArray(v, &a);
2410: VecGetArray(diagV, &diagA);
2411: VecGetArray(offdiagV, &offdiagA);
2412: for (r = 0; r < n; ++r) {
2413: if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2414: a[r] = diagA[r];
2415: idx[r] = cstart + diagIdx[r];
2416: } else {
2417: a[r] = offdiagA[r];
2418: idx[r] = cmap[offdiagIdx[r]];
2419: }
2420: }
2421: VecRestoreArray(v, &a);
2422: VecRestoreArray(diagV, &diagA);
2423: VecRestoreArray(offdiagV, &offdiagA);
2424: VecDestroy(&diagV);
2425: VecDestroy(&offdiagV);
2426: PetscFree2(diagIdx, offdiagIdx);
2427: return(0);
2428: }
2430: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2431: {
2432: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2433: PetscInt n = A->rmap->n;
2434: PetscInt cstart = A->cmap->rstart;
2435: PetscInt *cmap = mat->garray;
2436: PetscInt *diagIdx, *offdiagIdx;
2437: Vec diagV, offdiagV;
2438: PetscScalar *a, *diagA, *offdiagA;
2439: PetscInt r;
2443: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2444: VecCreateSeq(PETSC_COMM_SELF, n, &diagV);
2445: VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);
2446: MatGetRowMax(mat->A, diagV, diagIdx);
2447: MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2448: VecGetArray(v, &a);
2449: VecGetArray(diagV, &diagA);
2450: VecGetArray(offdiagV, &offdiagA);
2451: for (r = 0; r < n; ++r) {
2452: if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2453: a[r] = diagA[r];
2454: idx[r] = cstart + diagIdx[r];
2455: } else {
2456: a[r] = offdiagA[r];
2457: idx[r] = cmap[offdiagIdx[r]];
2458: }
2459: }
2460: VecRestoreArray(v, &a);
2461: VecRestoreArray(diagV, &diagA);
2462: VecRestoreArray(offdiagV, &offdiagA);
2463: VecDestroy(&diagV);
2464: VecDestroy(&offdiagV);
2465: PetscFree2(diagIdx, offdiagIdx);
2466: return(0);
2467: }
2469: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2470: {
2472: Mat *dummy;
2475: MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2476: *newmat = *dummy;
2477: PetscFree(dummy);
2478: return(0);
2479: }
2481: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2482: {
2483: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data;
2487: MatInvertBlockDiagonal(a->A,values);
2488: A->factorerrortype = a->A->factorerrortype;
2489: return(0);
2490: }
2492: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2493: {
2495: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data;
2498: MatSetRandom(aij->A,rctx);
2499: MatSetRandom(aij->B,rctx);
2500: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2501: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2502: return(0);
2503: }
2505: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2506: {
2508: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2509: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2510: return(0);
2511: }
2513: /*@
2514: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2516: Collective on Mat
2518: Input Parameters:
2519: + A - the matrix
2520: - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)
2522: Level: advanced
2524: @*/
2525: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2526: {
2527: PetscErrorCode ierr;
2530: PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2531: return(0);
2532: }
2534: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2535: {
2536: PetscErrorCode ierr;
2537: PetscBool sc = PETSC_FALSE,flg;
2540: PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2541: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2542: PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2543: if (flg) {
2544: MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2545: }
2546: PetscOptionsTail();
2547: return(0);
2548: }
2550: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2551: {
2553: Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data;
2554: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data;
2557: if (!Y->preallocated) {
2558: MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2559: } else if (!aij->nz) {
2560: PetscInt nonew = aij->nonew;
2561: MatSeqAIJSetPreallocation(maij->A,1,NULL);
2562: aij->nonew = nonew;
2563: }
2564: MatShift_Basic(Y,a);
2565: return(0);
2566: }
2568: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d)
2569: {
2570: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2574: if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2575: MatMissingDiagonal(a->A,missing,d);
2576: if (d) {
2577: PetscInt rstart;
2578: MatGetOwnershipRange(A,&rstart,NULL);
2579: *d += rstart;
2581: }
2582: return(0);
2583: }
2585: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
2586: {
2587: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2591: MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);
2592: return(0);
2593: }
2595: /* -------------------------------------------------------------------*/
2596: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2597: MatGetRow_MPIAIJ,
2598: MatRestoreRow_MPIAIJ,
2599: MatMult_MPIAIJ,
2600: /* 4*/ MatMultAdd_MPIAIJ,
2601: MatMultTranspose_MPIAIJ,
2602: MatMultTransposeAdd_MPIAIJ,
2603: 0,
2604: 0,
2605: 0,
2606: /*10*/ 0,
2607: 0,
2608: 0,
2609: MatSOR_MPIAIJ,
2610: MatTranspose_MPIAIJ,
2611: /*15*/ MatGetInfo_MPIAIJ,
2612: MatEqual_MPIAIJ,
2613: MatGetDiagonal_MPIAIJ,
2614: MatDiagonalScale_MPIAIJ,
2615: MatNorm_MPIAIJ,
2616: /*20*/ MatAssemblyBegin_MPIAIJ,
2617: MatAssemblyEnd_MPIAIJ,
2618: MatSetOption_MPIAIJ,
2619: MatZeroEntries_MPIAIJ,
2620: /*24*/ MatZeroRows_MPIAIJ,
2621: 0,
2622: 0,
2623: 0,
2624: 0,
2625: /*29*/ MatSetUp_MPIAIJ,
2626: 0,
2627: 0,
2628: MatGetDiagonalBlock_MPIAIJ,
2629: 0,
2630: /*34*/ MatDuplicate_MPIAIJ,
2631: 0,
2632: 0,
2633: 0,
2634: 0,
2635: /*39*/ MatAXPY_MPIAIJ,
2636: MatCreateSubMatrices_MPIAIJ,
2637: MatIncreaseOverlap_MPIAIJ,
2638: MatGetValues_MPIAIJ,
2639: MatCopy_MPIAIJ,
2640: /*44*/ MatGetRowMax_MPIAIJ,
2641: MatScale_MPIAIJ,
2642: MatShift_MPIAIJ,
2643: MatDiagonalSet_MPIAIJ,
2644: MatZeroRowsColumns_MPIAIJ,
2645: /*49*/ MatSetRandom_MPIAIJ,
2646: 0,
2647: 0,
2648: 0,
2649: 0,
2650: /*54*/ MatFDColoringCreate_MPIXAIJ,
2651: 0,
2652: MatSetUnfactored_MPIAIJ,
2653: MatPermute_MPIAIJ,
2654: 0,
2655: /*59*/ MatCreateSubMatrix_MPIAIJ,
2656: MatDestroy_MPIAIJ,
2657: MatView_MPIAIJ,
2658: 0,
2659: MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ,
2660: /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ,
2661: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2662: 0,
2663: 0,
2664: 0,
2665: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2666: MatGetRowMinAbs_MPIAIJ,
2667: 0,
2668: 0,
2669: 0,
2670: 0,
2671: /*75*/ MatFDColoringApply_AIJ,
2672: MatSetFromOptions_MPIAIJ,
2673: 0,
2674: 0,
2675: MatFindZeroDiagonals_MPIAIJ,
2676: /*80*/ 0,
2677: 0,
2678: 0,
2679: /*83*/ MatLoad_MPIAIJ,
2680: MatIsSymmetric_MPIAIJ,
2681: 0,
2682: 0,
2683: 0,
2684: 0,
2685: /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
2686: MatMatMultSymbolic_MPIAIJ_MPIAIJ,
2687: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2688: MatPtAP_MPIAIJ_MPIAIJ,
2689: MatPtAPSymbolic_MPIAIJ_MPIAIJ,
2690: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2691: 0,
2692: 0,
2693: 0,
2694: 0,
2695: /*99*/ 0,
2696: 0,
2697: 0,
2698: MatConjugate_MPIAIJ,
2699: 0,
2700: /*104*/MatSetValuesRow_MPIAIJ,
2701: MatRealPart_MPIAIJ,
2702: MatImaginaryPart_MPIAIJ,
2703: 0,
2704: 0,
2705: /*109*/0,
2706: 0,
2707: MatGetRowMin_MPIAIJ,
2708: 0,
2709: MatMissingDiagonal_MPIAIJ,
2710: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2711: 0,
2712: MatGetGhosts_MPIAIJ,
2713: 0,
2714: 0,
2715: /*119*/0,
2716: 0,
2717: 0,
2718: 0,
2719: MatGetMultiProcBlock_MPIAIJ,
2720: /*124*/MatFindNonzeroRows_MPIAIJ,
2721: MatGetColumnNorms_MPIAIJ,
2722: MatInvertBlockDiagonal_MPIAIJ,
2723: MatInvertVariableBlockDiagonal_MPIAIJ,
2724: MatCreateSubMatricesMPI_MPIAIJ,
2725: /*129*/0,
2726: MatTransposeMatMult_MPIAIJ_MPIAIJ,
2727: MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
2728: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2729: 0,
2730: /*134*/0,
2731: 0,
2732: MatRARt_MPIAIJ_MPIAIJ,
2733: 0,
2734: 0,
2735: /*139*/MatSetBlockSizes_MPIAIJ,
2736: 0,
2737: 0,
2738: MatFDColoringSetUp_MPIXAIJ,
2739: MatFindOffBlockDiagonalEntries_MPIAIJ,
2740: /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ
2741: };
2743: /* ----------------------------------------------------------------------------------------*/
2745: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2746: {
2747: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2751: MatStoreValues(aij->A);
2752: MatStoreValues(aij->B);
2753: return(0);
2754: }
2756: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2757: {
2758: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2762: MatRetrieveValues(aij->A);
2763: MatRetrieveValues(aij->B);
2764: return(0);
2765: }
2767: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2768: {
2769: Mat_MPIAIJ *b;
2773: PetscLayoutSetUp(B->rmap);
2774: PetscLayoutSetUp(B->cmap);
2775: b = (Mat_MPIAIJ*)B->data;
2777: #if defined(PETSC_USE_CTABLE)
2778: PetscTableDestroy(&b->colmap);
2779: #else
2780: PetscFree(b->colmap);
2781: #endif
2782: PetscFree(b->garray);
2783: VecDestroy(&b->lvec);
2784: VecScatterDestroy(&b->Mvctx);
2786: /* Because the B will have been resized we simply destroy it and create a new one each time */
2787: MatDestroy(&b->B);
2788: MatCreate(PETSC_COMM_SELF,&b->B);
2789: MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
2790: MatSetBlockSizesFromMats(b->B,B,B);
2791: MatSetType(b->B,MATSEQAIJ);
2792: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2794: if (!B->preallocated) {
2795: MatCreate(PETSC_COMM_SELF,&b->A);
2796: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2797: MatSetBlockSizesFromMats(b->A,B,B);
2798: MatSetType(b->A,MATSEQAIJ);
2799: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2800: }
2802: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2803: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
2804: B->preallocated = PETSC_TRUE;
2805: B->was_assembled = PETSC_FALSE;
2806: B->assembled = PETSC_FALSE;;
2807: return(0);
2808: }
2810: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2811: {
2812: Mat_MPIAIJ *b;
2817: PetscLayoutSetUp(B->rmap);
2818: PetscLayoutSetUp(B->cmap);
2819: b = (Mat_MPIAIJ*)B->data;
2821: #if defined(PETSC_USE_CTABLE)
2822: PetscTableDestroy(&b->colmap);
2823: #else
2824: PetscFree(b->colmap);
2825: #endif
2826: PetscFree(b->garray);
2827: VecDestroy(&b->lvec);
2828: VecScatterDestroy(&b->Mvctx);
2830: MatResetPreallocation(b->A);
2831: MatResetPreallocation(b->B);
2832: B->preallocated = PETSC_TRUE;
2833: B->was_assembled = PETSC_FALSE;
2834: B->assembled = PETSC_FALSE;
2835: return(0);
2836: }
2838: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2839: {
2840: Mat mat;
2841: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
2845: *newmat = 0;
2846: MatCreate(PetscObjectComm((PetscObject)matin),&mat);
2847: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
2848: MatSetBlockSizesFromMats(mat,matin,matin);
2849: MatSetType(mat,((PetscObject)matin)->type_name);
2850: PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
2851: a = (Mat_MPIAIJ*)mat->data;
2853: mat->factortype = matin->factortype;
2854: mat->assembled = PETSC_TRUE;
2855: mat->insertmode = NOT_SET_VALUES;
2856: mat->preallocated = PETSC_TRUE;
2858: a->size = oldmat->size;
2859: a->rank = oldmat->rank;
2860: a->donotstash = oldmat->donotstash;
2861: a->roworiented = oldmat->roworiented;
2862: a->rowindices = 0;
2863: a->rowvalues = 0;
2864: a->getrowactive = PETSC_FALSE;
2866: PetscLayoutReference(matin->rmap,&mat->rmap);
2867: PetscLayoutReference(matin->cmap,&mat->cmap);
2869: if (oldmat->colmap) {
2870: #if defined(PETSC_USE_CTABLE)
2871: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2872: #else
2873: PetscMalloc1(mat->cmap->N,&a->colmap);
2874: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
2875: PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));
2876: #endif
2877: } else a->colmap = 0;
2878: if (oldmat->garray) {
2879: PetscInt len;
2880: len = oldmat->B->cmap->n;
2881: PetscMalloc1(len+1,&a->garray);
2882: PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
2883: if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
2884: } else a->garray = 0;
2886: VecDuplicate(oldmat->lvec,&a->lvec);
2887: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
2888: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2889: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
2891: if (oldmat->Mvctx_mpi1) {
2892: VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);
2893: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);
2894: }
2896: MatDuplicate(oldmat->A,cpvalues,&a->A);
2897: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
2898: MatDuplicate(oldmat->B,cpvalues,&a->B);
2899: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
2900: PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2901: *newmat = mat;
2902: return(0);
2903: }
2905: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
2906: {
2907: PetscScalar *vals,*svals;
2908: MPI_Comm comm;
2910: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
2911: PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0;
2912: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
2913: PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols;
2914: PetscInt cend,cstart,n,*rowners;
2915: int fd;
2916: PetscInt bs = newMat->rmap->bs;
2919: /* force binary viewer to load .info file if it has not yet done so */
2920: PetscViewerSetUp(viewer);
2921: PetscObjectGetComm((PetscObject)viewer,&comm);
2922: MPI_Comm_size(comm,&size);
2923: MPI_Comm_rank(comm,&rank);
2924: PetscViewerBinaryGetDescriptor(viewer,&fd);
2925: if (!rank) {
2926: PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
2927: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2928: if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ");
2929: }
2931: PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");
2932: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
2933: PetscOptionsEnd();
2934: if (bs < 0) bs = 1;
2936: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2937: M = header[1]; N = header[2];
2939: /* If global sizes are set, check if they are consistent with that given in the file */
2940: if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M);
2941: if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N);
2943: /* determine ownership of all (block) rows */
2944: if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs);
2945: if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */
2946: else m = newMat->rmap->n; /* Set by user */
2948: PetscMalloc1(size+1,&rowners);
2949: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
2951: /* First process needs enough room for process with most rows */
2952: if (!rank) {
2953: mmax = rowners[1];
2954: for (i=2; i<=size; i++) {
2955: mmax = PetscMax(mmax, rowners[i]);
2956: }
2957: } else mmax = -1; /* unused, but compilers complain */
2959: rowners[0] = 0;
2960: for (i=2; i<=size; i++) {
2961: rowners[i] += rowners[i-1];
2962: }
2963: rstart = rowners[rank];
2964: rend = rowners[rank+1];
2966: /* distribute row lengths to all processors */
2967: PetscMalloc2(m,&ourlens,m,&offlens);
2968: if (!rank) {
2969: PetscBinaryRead(fd,ourlens,m,PETSC_INT);
2970: PetscMalloc1(mmax,&rowlengths);
2971: PetscCalloc1(size,&procsnz);
2972: for (j=0; j<m; j++) {
2973: procsnz[0] += ourlens[j];
2974: }
2975: for (i=1; i<size; i++) {
2976: PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
2977: /* calculate the number of nonzeros on each processor */
2978: for (j=0; j<rowners[i+1]-rowners[i]; j++) {
2979: procsnz[i] += rowlengths[j];
2980: }
2981: MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
2982: }
2983: PetscFree(rowlengths);
2984: } else {
2985: MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
2986: }
2988: if (!rank) {
2989: /* determine max buffer needed and allocate it */
2990: maxnz = 0;
2991: for (i=0; i<size; i++) {
2992: maxnz = PetscMax(maxnz,procsnz[i]);
2993: }
2994: PetscMalloc1(maxnz,&cols);
2996: /* read in my part of the matrix column indices */
2997: nz = procsnz[0];
2998: PetscMalloc1(nz,&mycols);
2999: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3001: /* read in every one elses and ship off */
3002: for (i=1; i<size; i++) {
3003: nz = procsnz[i];
3004: PetscBinaryRead(fd,cols,nz,PETSC_INT);
3005: MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
3006: }
3007: PetscFree(cols);
3008: } else {
3009: /* determine buffer space needed for message */
3010: nz = 0;
3011: for (i=0; i<m; i++) {
3012: nz += ourlens[i];
3013: }
3014: PetscMalloc1(nz,&mycols);
3016: /* receive message of column indices*/
3017: MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
3018: }
3020: /* determine column ownership if matrix is not square */
3021: if (N != M) {
3022: if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank);
3023: else n = newMat->cmap->n;
3024: MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
3025: cstart = cend - n;
3026: } else {
3027: cstart = rstart;
3028: cend = rend;
3029: n = cend - cstart;
3030: }
3032: /* loop over local rows, determining number of off diagonal entries */
3033: PetscMemzero(offlens,m*sizeof(PetscInt));
3034: jj = 0;
3035: for (i=0; i<m; i++) {
3036: for (j=0; j<ourlens[i]; j++) {
3037: if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
3038: jj++;
3039: }
3040: }
3042: for (i=0; i<m; i++) {
3043: ourlens[i] -= offlens[i];
3044: }
3045: MatSetSizes(newMat,m,n,M,N);
3047: if (bs > 1) {MatSetBlockSize(newMat,bs);}
3049: MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);
3051: for (i=0; i<m; i++) {
3052: ourlens[i] += offlens[i];
3053: }
3055: if (!rank) {
3056: PetscMalloc1(maxnz+1,&vals);
3058: /* read in my part of the matrix numerical values */
3059: nz = procsnz[0];
3060: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3062: /* insert into matrix */
3063: jj = rstart;
3064: smycols = mycols;
3065: svals = vals;
3066: for (i=0; i<m; i++) {
3067: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3068: smycols += ourlens[i];
3069: svals += ourlens[i];
3070: jj++;
3071: }
3073: /* read in other processors and ship out */
3074: for (i=1; i<size; i++) {
3075: nz = procsnz[i];
3076: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3077: MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3078: }
3079: PetscFree(procsnz);
3080: } else {
3081: /* receive numeric values */
3082: PetscMalloc1(nz+1,&vals);
3084: /* receive message of values*/
3085: MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);
3087: /* insert into matrix */
3088: jj = rstart;
3089: smycols = mycols;
3090: svals = vals;
3091: for (i=0; i<m; i++) {
3092: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3093: smycols += ourlens[i];
3094: svals += ourlens[i];
3095: jj++;
3096: }
3097: }
3098: PetscFree2(ourlens,offlens);
3099: PetscFree(vals);
3100: PetscFree(mycols);
3101: PetscFree(rowners);
3102: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3103: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3104: return(0);
3105: }
3107: /* Not scalable because of ISAllGather() unless getting all columns. */
3108: PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq)
3109: {
3111: IS iscol_local;
3112: PetscBool isstride;
3113: PetscMPIInt lisstride=0,gisstride;
3116: /* check if we are grabbing all columns*/
3117: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);
3119: if (isstride) {
3120: PetscInt start,len,mstart,mlen;
3121: ISStrideGetInfo(iscol,&start,NULL);
3122: ISGetLocalSize(iscol,&len);
3123: MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
3124: if (mstart == start && mlen-mstart == len) lisstride = 1;
3125: }
3127: MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3128: if (gisstride) {
3129: PetscInt N;
3130: MatGetSize(mat,NULL,&N);
3131: ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);
3132: ISSetIdentity(iscol_local);
3133: PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3134: } else {
3135: PetscInt cbs;
3136: ISGetBlockSize(iscol,&cbs);
3137: ISAllGather(iscol,&iscol_local);
3138: ISSetBlockSize(iscol_local,cbs);
3139: }
3141: *isseq = iscol_local;
3142: return(0);
3143: }
3145: /*
3146: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3147: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3149: Input Parameters:
3150: mat - matrix
3151: isrow - parallel row index set; its local indices are a subset of local columns of mat,
3152: i.e., mat->rstart <= isrow[i] < mat->rend
3153: iscol - parallel column index set; its local indices are a subset of local columns of mat,
3154: i.e., mat->cstart <= iscol[i] < mat->cend
3155: Output Parameter:
3156: isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A
3157: iscol_o - sequential column index set for retrieving mat->B
3158: garray - column map; garray[i] indicates global location of iscol_o[i] in iscol
3159: */
3160: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[])
3161: {
3163: Vec x,cmap;
3164: const PetscInt *is_idx;
3165: PetscScalar *xarray,*cmaparray;
3166: PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count;
3167: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3168: Mat B=a->B;
3169: Vec lvec=a->lvec,lcmap;
3170: PetscInt i,cstart,cend,Bn=B->cmap->N;
3171: MPI_Comm comm;
3172: VecScatter Mvctx=a->Mvctx;
3175: PetscObjectGetComm((PetscObject)mat,&comm);
3176: ISGetLocalSize(iscol,&ncols);
3178: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3179: MatCreateVecs(mat,&x,NULL);
3180: VecSet(x,-1.0);
3181: VecDuplicate(x,&cmap);
3182: VecSet(cmap,-1.0);
3184: /* Get start indices */
3185: MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);
3186: isstart -= ncols;
3187: MatGetOwnershipRangeColumn(mat,&cstart,&cend);
3189: ISGetIndices(iscol,&is_idx);
3190: VecGetArray(x,&xarray);
3191: VecGetArray(cmap,&cmaparray);
3192: PetscMalloc1(ncols,&idx);
3193: for (i=0; i<ncols; i++) {
3194: xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i];
3195: cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */
3196: idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */
3197: }
3198: VecRestoreArray(x,&xarray);
3199: VecRestoreArray(cmap,&cmaparray);
3200: ISRestoreIndices(iscol,&is_idx);
3202: /* Get iscol_d */
3203: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);
3204: ISGetBlockSize(iscol,&i);
3205: ISSetBlockSize(*iscol_d,i);
3207: /* Get isrow_d */
3208: ISGetLocalSize(isrow,&m);
3209: rstart = mat->rmap->rstart;
3210: PetscMalloc1(m,&idx);
3211: ISGetIndices(isrow,&is_idx);
3212: for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart;
3213: ISRestoreIndices(isrow,&is_idx);
3215: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);
3216: ISGetBlockSize(isrow,&i);
3217: ISSetBlockSize(*isrow_d,i);
3219: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3220: VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3221: VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3223: VecDuplicate(lvec,&lcmap);
3225: VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3226: VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3228: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3229: /* off-process column indices */
3230: count = 0;
3231: PetscMalloc1(Bn,&idx);
3232: PetscMalloc1(Bn,&cmap1);
3234: VecGetArray(lvec,&xarray);
3235: VecGetArray(lcmap,&cmaparray);
3236: for (i=0; i<Bn; i++) {
3237: if (PetscRealPart(xarray[i]) > -1.0) {
3238: idx[count] = i; /* local column index in off-diagonal part B */
3239: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3240: count++;
3241: }
3242: }
3243: VecRestoreArray(lvec,&xarray);
3244: VecRestoreArray(lcmap,&cmaparray);
3246: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);
3247: /* cannot ensure iscol_o has same blocksize as iscol! */
3249: PetscFree(idx);
3250: *garray = cmap1;
3252: VecDestroy(&x);
3253: VecDestroy(&cmap);
3254: VecDestroy(&lcmap);
3255: return(0);
3256: }
3258: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3259: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat)
3260: {
3262: Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub;
3263: Mat M = NULL;
3264: MPI_Comm comm;
3265: IS iscol_d,isrow_d,iscol_o;
3266: Mat Asub = NULL,Bsub = NULL;
3267: PetscInt n;
3270: PetscObjectGetComm((PetscObject)mat,&comm);
3272: if (call == MAT_REUSE_MATRIX) {
3273: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3274: PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);
3275: if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse");
3277: PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);
3278: if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse");
3280: PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);
3281: if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse");
3283: /* Update diagonal and off-diagonal portions of submat */
3284: asub = (Mat_MPIAIJ*)(*submat)->data;
3285: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);
3286: ISGetLocalSize(iscol_o,&n);
3287: if (n) {
3288: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);
3289: }
3290: MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3291: MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);
3293: } else { /* call == MAT_INITIAL_MATRIX) */
3294: const PetscInt *garray;
3295: PetscInt BsubN;
3297: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3298: ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);
3300: /* Create local submatrices Asub and Bsub */
3301: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);
3302: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);
3304: /* Create submatrix M */
3305: MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);
3307: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3308: asub = (Mat_MPIAIJ*)M->data;
3310: ISGetLocalSize(iscol_o,&BsubN);
3311: n = asub->B->cmap->N;
3312: if (BsubN > n) {
3313: /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */
3314: const PetscInt *idx;
3315: PetscInt i,j,*idx_new,*subgarray = asub->garray;
3316: PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);
3318: PetscMalloc1(n,&idx_new);
3319: j = 0;
3320: ISGetIndices(iscol_o,&idx);
3321: for (i=0; i<n; i++) {
3322: if (j >= BsubN) break;
3323: while (subgarray[i] > garray[j]) j++;
3325: if (subgarray[i] == garray[j]) {
3326: idx_new[i] = idx[j++];
3327: } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]);
3328: }
3329: ISRestoreIndices(iscol_o,&idx);
3331: ISDestroy(&iscol_o);
3332: ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);
3334: } else if (BsubN < n) {
3335: SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N);
3336: }
3338: PetscFree(garray);
3339: *submat = M;
3341: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3342: PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);
3343: ISDestroy(&isrow_d);
3345: PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);
3346: ISDestroy(&iscol_d);
3348: PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);
3349: ISDestroy(&iscol_o);
3350: }
3351: return(0);
3352: }
3354: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3355: {
3357: IS iscol_local=NULL,isrow_d;
3358: PetscInt csize;
3359: PetscInt n,i,j,start,end;
3360: PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2];
3361: MPI_Comm comm;
3364: /* If isrow has same processor distribution as mat,
3365: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3366: if (call == MAT_REUSE_MATRIX) {
3367: PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);
3368: if (isrow_d) {
3369: sameRowDist = PETSC_TRUE;
3370: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3371: } else {
3372: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);
3373: if (iscol_local) {
3374: sameRowDist = PETSC_TRUE;
3375: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3376: }
3377: }
3378: } else {
3379: /* Check if isrow has same processor distribution as mat */
3380: sameDist[0] = PETSC_FALSE;
3381: ISGetLocalSize(isrow,&n);
3382: if (!n) {
3383: sameDist[0] = PETSC_TRUE;
3384: } else {
3385: ISGetMinMax(isrow,&i,&j);
3386: MatGetOwnershipRange(mat,&start,&end);
3387: if (i >= start && j < end) {
3388: sameDist[0] = PETSC_TRUE;
3389: }
3390: }
3392: /* Check if iscol has same processor distribution as mat */
3393: sameDist[1] = PETSC_FALSE;
3394: ISGetLocalSize(iscol,&n);
3395: if (!n) {
3396: sameDist[1] = PETSC_TRUE;
3397: } else {
3398: ISGetMinMax(iscol,&i,&j);
3399: MatGetOwnershipRangeColumn(mat,&start,&end);
3400: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3401: }
3403: PetscObjectGetComm((PetscObject)mat,&comm);
3404: MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);
3405: sameRowDist = tsameDist[0];
3406: }
3408: if (sameRowDist) {
3409: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3410: /* isrow and iscol have same processor distribution as mat */
3411: MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);
3412: return(0);
3413: } else { /* sameRowDist */
3414: /* isrow has same processor distribution as mat */
3415: if (call == MAT_INITIAL_MATRIX) {
3416: PetscBool sorted;
3417: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3418: ISGetLocalSize(iscol_local,&n); /* local size of iscol_local = global columns of newmat */
3419: ISGetSize(iscol,&i);
3420: if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i);
3422: ISSorted(iscol_local,&sorted);
3423: if (sorted) {
3424: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3425: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);
3426: return(0);
3427: }
3428: } else { /* call == MAT_REUSE_MATRIX */
3429: IS iscol_sub;
3430: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3431: if (iscol_sub) {
3432: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);
3433: return(0);
3434: }
3435: }
3436: }
3437: }
3439: /* General case: iscol -> iscol_local which has global size of iscol */
3440: if (call == MAT_REUSE_MATRIX) {
3441: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3442: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3443: } else {
3444: if (!iscol_local) {
3445: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3446: }
3447: }
3449: ISGetLocalSize(iscol,&csize);
3450: MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);
3452: if (call == MAT_INITIAL_MATRIX) {
3453: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3454: ISDestroy(&iscol_local);
3455: }
3456: return(0);
3457: }
3459: /*@C
3460: MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal"
3461: and "off-diagonal" part of the matrix in CSR format.
3463: Collective on MPI_Comm
3465: Input Parameters:
3466: + comm - MPI communicator
3467: . A - "diagonal" portion of matrix
3468: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3469: - garray - global index of B columns
3471: Output Parameter:
3472: . mat - the matrix, with input A as its local diagonal matrix
3473: Level: advanced
3475: Notes:
3476: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3477: A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore.
3479: .seealso: MatCreateMPIAIJWithSplitArrays()
3480: @*/
3481: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat)
3482: {
3484: Mat_MPIAIJ *maij;
3485: Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew;
3486: PetscInt *oi=b->i,*oj=b->j,i,nz,col;
3487: PetscScalar *oa=b->a;
3488: Mat Bnew;
3489: PetscInt m,n,N;
3492: MatCreate(comm,mat);
3493: MatGetSize(A,&m,&n);
3494: if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N);
3495: 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);
3496: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3497: /* 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); */
3499: /* Get global columns of mat */
3500: MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);
3502: MatSetSizes(*mat,m,n,PETSC_DECIDE,N);
3503: MatSetType(*mat,MATMPIAIJ);
3504: MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);
3505: maij = (Mat_MPIAIJ*)(*mat)->data;
3507: (*mat)->preallocated = PETSC_TRUE;
3509: PetscLayoutSetUp((*mat)->rmap);
3510: PetscLayoutSetUp((*mat)->cmap);
3512: /* Set A as diagonal portion of *mat */
3513: maij->A = A;
3515: nz = oi[m];
3516: for (i=0; i<nz; i++) {
3517: col = oj[i];
3518: oj[i] = garray[col];
3519: }
3521: /* Set Bnew as off-diagonal portion of *mat */
3522: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);
3523: bnew = (Mat_SeqAIJ*)Bnew->data;
3524: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3525: maij->B = Bnew;
3527: if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N);
3529: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3530: b->free_a = PETSC_FALSE;
3531: b->free_ij = PETSC_FALSE;
3532: MatDestroy(&B);
3534: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3535: bnew->free_a = PETSC_TRUE;
3536: bnew->free_ij = PETSC_TRUE;
3538: /* condense columns of maij->B */
3539: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
3540: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3541: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3542: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
3543: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3544: return(0);
3545: }
3547: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*);
3549: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat)
3550: {
3552: PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs;
3553: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3554: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3555: Mat M,Msub,B=a->B;
3556: MatScalar *aa;
3557: Mat_SeqAIJ *aij;
3558: PetscInt *garray = a->garray,*colsub,Ncols;
3559: PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend;
3560: IS iscol_sub,iscmap;
3561: const PetscInt *is_idx,*cmap;
3562: PetscBool allcolumns=PETSC_FALSE;
3563: MPI_Comm comm;
3566: PetscObjectGetComm((PetscObject)mat,&comm);
3568: if (call == MAT_REUSE_MATRIX) {
3569: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3570: if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse");
3571: ISGetLocalSize(iscol_sub,&count);
3573: PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);
3574: if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse");
3576: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);
3577: if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3579: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);
3581: } else { /* call == MAT_INITIAL_MATRIX) */
3582: PetscBool flg;
3584: ISGetLocalSize(iscol,&n);
3585: ISGetSize(iscol,&Ncols);
3587: /* (1) iscol -> nonscalable iscol_local */
3588: /* Check for special case: each processor gets entire matrix columns */
3589: ISIdentity(iscol_local,&flg);
3590: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3591: if (allcolumns) {
3592: iscol_sub = iscol_local;
3593: PetscObjectReference((PetscObject)iscol_local);
3594: ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);
3596: } else {
3597: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3598: PetscInt *idx,*cmap1,k;
3599: PetscMalloc1(Ncols,&idx);
3600: PetscMalloc1(Ncols,&cmap1);
3601: ISGetIndices(iscol_local,&is_idx);
3602: count = 0;
3603: k = 0;
3604: for (i=0; i<Ncols; i++) {
3605: j = is_idx[i];
3606: if (j >= cstart && j < cend) {
3607: /* diagonal part of mat */
3608: idx[count] = j;
3609: cmap1[count++] = i; /* column index in submat */
3610: } else if (Bn) {
3611: /* off-diagonal part of mat */
3612: if (j == garray[k]) {
3613: idx[count] = j;
3614: cmap1[count++] = i; /* column index in submat */
3615: } else if (j > garray[k]) {
3616: while (j > garray[k] && k < Bn-1) k++;
3617: if (j == garray[k]) {
3618: idx[count] = j;
3619: cmap1[count++] = i; /* column index in submat */
3620: }
3621: }
3622: }
3623: }
3624: ISRestoreIndices(iscol_local,&is_idx);
3626: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);
3627: ISGetBlockSize(iscol,&cbs);
3628: ISSetBlockSize(iscol_sub,cbs);
3630: ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);
3631: }
3633: /* (3) Create sequential Msub */
3634: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);
3635: }
3637: ISGetLocalSize(iscol_sub,&count);
3638: aij = (Mat_SeqAIJ*)(Msub)->data;
3639: ii = aij->i;
3640: ISGetIndices(iscmap,&cmap);
3642: /*
3643: m - number of local rows
3644: Ncols - number of columns (same on all processors)
3645: rstart - first row in new global matrix generated
3646: */
3647: MatGetSize(Msub,&m,NULL);
3649: if (call == MAT_INITIAL_MATRIX) {
3650: /* (4) Create parallel newmat */
3651: PetscMPIInt rank,size;
3652: PetscInt csize;
3654: MPI_Comm_size(comm,&size);
3655: MPI_Comm_rank(comm,&rank);
3657: /*
3658: Determine the number of non-zeros in the diagonal and off-diagonal
3659: portions of the matrix in order to do correct preallocation
3660: */
3662: /* first get start and end of "diagonal" columns */
3663: ISGetLocalSize(iscol,&csize);
3664: if (csize == PETSC_DECIDE) {
3665: ISGetSize(isrow,&mglobal);
3666: if (mglobal == Ncols) { /* square matrix */
3667: nlocal = m;
3668: } else {
3669: nlocal = Ncols/size + ((Ncols % size) > rank);
3670: }
3671: } else {
3672: nlocal = csize;
3673: }
3674: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3675: rstart = rend - nlocal;
3676: 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);
3678: /* next, compute all the lengths */
3679: jj = aij->j;
3680: PetscMalloc1(2*m+1,&dlens);
3681: olens = dlens + m;
3682: for (i=0; i<m; i++) {
3683: jend = ii[i+1] - ii[i];
3684: olen = 0;
3685: dlen = 0;
3686: for (j=0; j<jend; j++) {
3687: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3688: else dlen++;
3689: jj++;
3690: }
3691: olens[i] = olen;
3692: dlens[i] = dlen;
3693: }
3695: ISGetBlockSize(isrow,&bs);
3696: ISGetBlockSize(iscol,&cbs);
3698: MatCreate(comm,&M);
3699: MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);
3700: MatSetBlockSizes(M,bs,cbs);
3701: MatSetType(M,((PetscObject)mat)->type_name);
3702: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3703: PetscFree(dlens);
3705: } else { /* call == MAT_REUSE_MATRIX */
3706: M = *newmat;
3707: MatGetLocalSize(M,&i,NULL);
3708: if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3709: MatZeroEntries(M);
3710: /*
3711: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3712: rather than the slower MatSetValues().
3713: */
3714: M->was_assembled = PETSC_TRUE;
3715: M->assembled = PETSC_FALSE;
3716: }
3718: /* (5) Set values of Msub to *newmat */
3719: PetscMalloc1(count,&colsub);
3720: MatGetOwnershipRange(M,&rstart,NULL);
3722: jj = aij->j;
3723: aa = aij->a;
3724: for (i=0; i<m; i++) {
3725: row = rstart + i;
3726: nz = ii[i+1] - ii[i];
3727: for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]];
3728: MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);
3729: jj += nz; aa += nz;
3730: }
3731: ISRestoreIndices(iscmap,&cmap);
3733: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3734: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3736: PetscFree(colsub);
3738: /* save Msub, iscol_sub and iscmap used in processor for next request */
3739: if (call == MAT_INITIAL_MATRIX) {
3740: *newmat = M;
3741: PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);
3742: MatDestroy(&Msub);
3744: PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);
3745: ISDestroy(&iscol_sub);
3747: PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);
3748: ISDestroy(&iscmap);
3750: if (iscol_local) {
3751: PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);
3752: ISDestroy(&iscol_local);
3753: }
3754: }
3755: return(0);
3756: }
3758: /*
3759: Not great since it makes two copies of the submatrix, first an SeqAIJ
3760: in local and then by concatenating the local matrices the end result.
3761: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3763: Note: This requires a sequential iscol with all indices.
3764: */
3765: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3766: {
3768: PetscMPIInt rank,size;
3769: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3770: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3771: Mat M,Mreuse;
3772: MatScalar *aa,*vwork;
3773: MPI_Comm comm;
3774: Mat_SeqAIJ *aij;
3775: PetscBool colflag,allcolumns=PETSC_FALSE;
3778: PetscObjectGetComm((PetscObject)mat,&comm);
3779: MPI_Comm_rank(comm,&rank);
3780: MPI_Comm_size(comm,&size);
3782: /* Check for special case: each processor gets entire matrix columns */
3783: ISIdentity(iscol,&colflag);
3784: ISGetLocalSize(iscol,&n);
3785: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3787: if (call == MAT_REUSE_MATRIX) {
3788: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3789: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3790: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);
3791: } else {
3792: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);
3793: }
3795: /*
3796: m - number of local rows
3797: n - number of columns (same on all processors)
3798: rstart - first row in new global matrix generated
3799: */
3800: MatGetSize(Mreuse,&m,&n);
3801: MatGetBlockSizes(Mreuse,&bs,&cbs);
3802: if (call == MAT_INITIAL_MATRIX) {
3803: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3804: ii = aij->i;
3805: jj = aij->j;
3807: /*
3808: Determine the number of non-zeros in the diagonal and off-diagonal
3809: portions of the matrix in order to do correct preallocation
3810: */
3812: /* first get start and end of "diagonal" columns */
3813: if (csize == PETSC_DECIDE) {
3814: ISGetSize(isrow,&mglobal);
3815: if (mglobal == n) { /* square matrix */
3816: nlocal = m;
3817: } else {
3818: nlocal = n/size + ((n % size) > rank);
3819: }
3820: } else {
3821: nlocal = csize;
3822: }
3823: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3824: rstart = rend - nlocal;
3825: 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);
3827: /* next, compute all the lengths */
3828: PetscMalloc1(2*m+1,&dlens);
3829: olens = dlens + m;
3830: for (i=0; i<m; i++) {
3831: jend = ii[i+1] - ii[i];
3832: olen = 0;
3833: dlen = 0;
3834: for (j=0; j<jend; j++) {
3835: if (*jj < rstart || *jj >= rend) olen++;
3836: else dlen++;
3837: jj++;
3838: }
3839: olens[i] = olen;
3840: dlens[i] = dlen;
3841: }
3842: MatCreate(comm,&M);
3843: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3844: MatSetBlockSizes(M,bs,cbs);
3845: MatSetType(M,((PetscObject)mat)->type_name);
3846: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3847: PetscFree(dlens);
3848: } else {
3849: PetscInt ml,nl;
3851: M = *newmat;
3852: MatGetLocalSize(M,&ml,&nl);
3853: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3854: MatZeroEntries(M);
3855: /*
3856: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3857: rather than the slower MatSetValues().
3858: */
3859: M->was_assembled = PETSC_TRUE;
3860: M->assembled = PETSC_FALSE;
3861: }
3862: MatGetOwnershipRange(M,&rstart,&rend);
3863: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3864: ii = aij->i;
3865: jj = aij->j;
3866: aa = aij->a;
3867: for (i=0; i<m; i++) {
3868: row = rstart + i;
3869: nz = ii[i+1] - ii[i];
3870: cwork = jj; jj += nz;
3871: vwork = aa; aa += nz;
3872: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3873: }
3875: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3876: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3877: *newmat = M;
3879: /* save submatrix used in processor for next request */
3880: if (call == MAT_INITIAL_MATRIX) {
3881: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3882: MatDestroy(&Mreuse);
3883: }
3884: return(0);
3885: }
3887: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3888: {
3889: PetscInt m,cstart, cend,j,nnz,i,d;
3890: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3891: const PetscInt *JJ;
3892: PetscScalar *values;
3894: PetscBool nooffprocentries;
3897: if (Ii && Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
3899: PetscLayoutSetUp(B->rmap);
3900: PetscLayoutSetUp(B->cmap);
3901: m = B->rmap->n;
3902: cstart = B->cmap->rstart;
3903: cend = B->cmap->rend;
3904: rstart = B->rmap->rstart;
3906: PetscMalloc2(m,&d_nnz,m,&o_nnz);
3908: #if defined(PETSC_USE_DEBUG)
3909: for (i=0; i<m; i++) {
3910: nnz = Ii[i+1]- Ii[i];
3911: JJ = J + Ii[i];
3912: if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3913: if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]);
3914: 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);
3915: }
3916: #endif
3918: for (i=0; i<m; i++) {
3919: nnz = Ii[i+1]- Ii[i];
3920: JJ = J + Ii[i];
3921: nnz_max = PetscMax(nnz_max,nnz);
3922: d = 0;
3923: for (j=0; j<nnz; j++) {
3924: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3925: }
3926: d_nnz[i] = d;
3927: o_nnz[i] = nnz - d;
3928: }
3929: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3930: PetscFree2(d_nnz,o_nnz);
3932: if (v) values = (PetscScalar*)v;
3933: else {
3934: PetscCalloc1(nnz_max+1,&values);
3935: }
3937: for (i=0; i<m; i++) {
3938: ii = i + rstart;
3939: nnz = Ii[i+1]- Ii[i];
3940: MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
3941: }
3942: nooffprocentries = B->nooffprocentries;
3943: B->nooffprocentries = PETSC_TRUE;
3944: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3945: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3946: B->nooffprocentries = nooffprocentries;
3948: if (!v) {
3949: PetscFree(values);
3950: }
3951: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3952: return(0);
3953: }
3955: /*@
3956: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3957: (the default parallel PETSc format).
3959: Collective on MPI_Comm
3961: Input Parameters:
3962: + B - the matrix
3963: . i - the indices into j for the start of each local row (starts with zero)
3964: . j - the column indices for each local row (starts with zero)
3965: - v - optional values in the matrix
3967: Level: developer
3969: Notes:
3970: The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc;
3971: thus you CANNOT change the matrix entries by changing the values of v[] after you have
3972: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
3974: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
3976: The format which is used for the sparse matrix input, is equivalent to a
3977: row-major ordering.. i.e for the following matrix, the input data expected is
3978: as shown
3980: $ 1 0 0
3981: $ 2 0 3 P0
3982: $ -------
3983: $ 4 5 6 P1
3984: $
3985: $ Process0 [P0]: rows_owned=[0,1]
3986: $ i = {0,1,3} [size = nrow+1 = 2+1]
3987: $ j = {0,0,2} [size = 3]
3988: $ v = {1,2,3} [size = 3]
3989: $
3990: $ Process1 [P1]: rows_owned=[2]
3991: $ i = {0,3} [size = nrow+1 = 1+1]
3992: $ j = {0,1,2} [size = 3]
3993: $ v = {4,5,6} [size = 3]
3995: .keywords: matrix, aij, compressed row, sparse, parallel
3997: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ,
3998: MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
3999: @*/
4000: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
4001: {
4005: PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4006: return(0);
4007: }
4009: /*@C
4010: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
4011: (the default parallel PETSc format). For good matrix assembly performance
4012: the user should preallocate the matrix storage by setting the parameters
4013: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4014: performance can be increased by more than a factor of 50.
4016: Collective on MPI_Comm
4018: Input Parameters:
4019: + B - the matrix
4020: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4021: (same value is used for all local rows)
4022: . d_nnz - array containing the number of nonzeros in the various rows of the
4023: DIAGONAL portion of the local submatrix (possibly different for each row)
4024: or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
4025: The size of this array is equal to the number of local rows, i.e 'm'.
4026: For matrices that will be factored, you must leave room for (and set)
4027: the diagonal entry even if it is zero.
4028: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4029: submatrix (same value is used for all local rows).
4030: - o_nnz - array containing the number of nonzeros in the various rows of the
4031: OFF-DIAGONAL portion of the local submatrix (possibly different for
4032: each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
4033: structure. The size of this array is equal to the number
4034: of local rows, i.e 'm'.
4036: If the *_nnz parameter is given then the *_nz parameter is ignored
4038: The AIJ format (also called the Yale sparse matrix format or
4039: compressed row storage (CSR)), is fully compatible with standard Fortran 77
4040: storage. The stored row and column indices begin with zero.
4041: See Users-Manual: ch_mat for details.
4043: The parallel matrix is partitioned such that the first m0 rows belong to
4044: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4045: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4047: The DIAGONAL portion of the local submatrix of a processor can be defined
4048: as the submatrix which is obtained by extraction the part corresponding to
4049: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4050: first row that belongs to the processor, r2 is the last row belonging to
4051: the this processor, and c1-c2 is range of indices of the local part of a
4052: vector suitable for applying the matrix to. This is an mxn matrix. In the
4053: common case of a square matrix, the row and column ranges are the same and
4054: the DIAGONAL part is also square. The remaining portion of the local
4055: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4057: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4059: You can call MatGetInfo() to get information on how effective the preallocation was;
4060: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4061: You can also run with the option -info and look for messages with the string
4062: malloc in them to see if additional memory allocation was needed.
4064: Example usage:
4066: Consider the following 8x8 matrix with 34 non-zero values, that is
4067: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4068: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4069: as follows:
4071: .vb
4072: 1 2 0 | 0 3 0 | 0 4
4073: Proc0 0 5 6 | 7 0 0 | 8 0
4074: 9 0 10 | 11 0 0 | 12 0
4075: -------------------------------------
4076: 13 0 14 | 15 16 17 | 0 0
4077: Proc1 0 18 0 | 19 20 21 | 0 0
4078: 0 0 0 | 22 23 0 | 24 0
4079: -------------------------------------
4080: Proc2 25 26 27 | 0 0 28 | 29 0
4081: 30 0 0 | 31 32 33 | 0 34
4082: .ve
4084: This can be represented as a collection of submatrices as:
4086: .vb
4087: A B C
4088: D E F
4089: G H I
4090: .ve
4092: Where the submatrices A,B,C are owned by proc0, D,E,F are
4093: owned by proc1, G,H,I are owned by proc2.
4095: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4096: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4097: The 'M','N' parameters are 8,8, and have the same values on all procs.
4099: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4100: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4101: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4102: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4103: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4104: matrix, ans [DF] as another SeqAIJ matrix.
4106: When d_nz, o_nz parameters are specified, d_nz storage elements are
4107: allocated for every row of the local diagonal submatrix, and o_nz
4108: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4109: One way to choose d_nz and o_nz is to use the max nonzerors per local
4110: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4111: In this case, the values of d_nz,o_nz are:
4112: .vb
4113: proc0 : dnz = 2, o_nz = 2
4114: proc1 : dnz = 3, o_nz = 2
4115: proc2 : dnz = 1, o_nz = 4
4116: .ve
4117: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4118: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4119: for proc3. i.e we are using 12+15+10=37 storage locations to store
4120: 34 values.
4122: When d_nnz, o_nnz parameters are specified, the storage is specified
4123: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4124: In the above case the values for d_nnz,o_nnz are:
4125: .vb
4126: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4127: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4128: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4129: .ve
4130: Here the space allocated is sum of all the above values i.e 34, and
4131: hence pre-allocation is perfect.
4133: Level: intermediate
4135: .keywords: matrix, aij, compressed row, sparse, parallel
4137: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
4138: MATMPIAIJ, MatGetInfo(), PetscSplitOwnership()
4139: @*/
4140: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
4141: {
4147: PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
4148: return(0);
4149: }
4151: /*@
4152: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
4153: CSR format the local rows.
4155: Collective on MPI_Comm
4157: Input Parameters:
4158: + comm - MPI communicator
4159: . m - number of local rows (Cannot be PETSC_DECIDE)
4160: . n - This value should be the same as the local size used in creating the
4161: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4162: calculated if N is given) For square matrices n is almost always m.
4163: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4164: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4165: . i - row indices
4166: . j - column indices
4167: - a - matrix values
4169: Output Parameter:
4170: . mat - the matrix
4172: Level: intermediate
4174: Notes:
4175: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4176: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4177: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4179: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4181: The format which is used for the sparse matrix input, is equivalent to a
4182: row-major ordering.. i.e for the following matrix, the input data expected is
4183: as shown
4185: $ 1 0 0
4186: $ 2 0 3 P0
4187: $ -------
4188: $ 4 5 6 P1
4189: $
4190: $ Process0 [P0]: rows_owned=[0,1]
4191: $ i = {0,1,3} [size = nrow+1 = 2+1]
4192: $ j = {0,0,2} [size = 3]
4193: $ v = {1,2,3} [size = 3]
4194: $
4195: $ Process1 [P1]: rows_owned=[2]
4196: $ i = {0,3} [size = nrow+1 = 1+1]
4197: $ j = {0,1,2} [size = 3]
4198: $ v = {4,5,6} [size = 3]
4200: .keywords: matrix, aij, compressed row, sparse, parallel
4202: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4203: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
4204: @*/
4205: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4206: {
4210: if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4211: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4212: MatCreate(comm,mat);
4213: MatSetSizes(*mat,m,n,M,N);
4214: /* MatSetBlockSizes(M,bs,cbs); */
4215: MatSetType(*mat,MATMPIAIJ);
4216: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4217: return(0);
4218: }
4220: /*@C
4221: MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4222: (the default parallel PETSc format). For good matrix assembly performance
4223: the user should preallocate the matrix storage by setting the parameters
4224: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4225: performance can be increased by more than a factor of 50.
4227: Collective on MPI_Comm
4229: Input Parameters:
4230: + comm - MPI communicator
4231: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4232: This value should be the same as the local size used in creating the
4233: y vector for the matrix-vector product y = Ax.
4234: . n - This value should be the same as the local size used in creating the
4235: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4236: calculated if N is given) For square matrices n is almost always m.
4237: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4238: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4239: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4240: (same value is used for all local rows)
4241: . d_nnz - array containing the number of nonzeros in the various rows of the
4242: DIAGONAL portion of the local submatrix (possibly different for each row)
4243: or NULL, if d_nz is used to specify the nonzero structure.
4244: The size of this array is equal to the number of local rows, i.e 'm'.
4245: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4246: submatrix (same value is used for all local rows).
4247: - o_nnz - array containing the number of nonzeros in the various rows of the
4248: OFF-DIAGONAL portion of the local submatrix (possibly different for
4249: each row) or NULL, if o_nz is used to specify the nonzero
4250: structure. The size of this array is equal to the number
4251: of local rows, i.e 'm'.
4253: Output Parameter:
4254: . A - the matrix
4256: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4257: MatXXXXSetPreallocation() paradgm instead of this routine directly.
4258: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
4260: Notes:
4261: If the *_nnz parameter is given then the *_nz parameter is ignored
4263: m,n,M,N parameters specify the size of the matrix, and its partitioning across
4264: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4265: storage requirements for this matrix.
4267: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
4268: processor than it must be used on all processors that share the object for
4269: that argument.
4271: The user MUST specify either the local or global matrix dimensions
4272: (possibly both).
4274: The parallel matrix is partitioned across processors such that the
4275: first m0 rows belong to process 0, the next m1 rows belong to
4276: process 1, the next m2 rows belong to process 2 etc.. where
4277: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4278: values corresponding to [m x N] submatrix.
4280: The columns are logically partitioned with the n0 columns belonging
4281: to 0th partition, the next n1 columns belonging to the next
4282: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4284: The DIAGONAL portion of the local submatrix on any given processor
4285: is the submatrix corresponding to the rows and columns m,n
4286: corresponding to the given processor. i.e diagonal matrix on
4287: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4288: etc. The remaining portion of the local submatrix [m x (N-n)]
4289: constitute the OFF-DIAGONAL portion. The example below better
4290: illustrates this concept.
4292: For a square global matrix we define each processor's diagonal portion
4293: to be its local rows and the corresponding columns (a square submatrix);
4294: each processor's off-diagonal portion encompasses the remainder of the
4295: local matrix (a rectangular submatrix).
4297: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4299: When calling this routine with a single process communicator, a matrix of
4300: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
4301: type of communicator, use the construction mechanism
4302: .vb
4303: MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4304: .ve
4306: $ MatCreate(...,&A);
4307: $ MatSetType(A,MATMPIAIJ);
4308: $ MatSetSizes(A, m,n,M,N);
4309: $ MatMPIAIJSetPreallocation(A,...);
4311: By default, this format uses inodes (identical nodes) when possible.
4312: We search for consecutive rows with the same nonzero structure, thereby
4313: reusing matrix information to achieve increased efficiency.
4315: Options Database Keys:
4316: + -mat_no_inode - Do not use inodes
4317: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4321: Example usage:
4323: Consider the following 8x8 matrix with 34 non-zero values, that is
4324: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4325: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4326: as follows
4328: .vb
4329: 1 2 0 | 0 3 0 | 0 4
4330: Proc0 0 5 6 | 7 0 0 | 8 0
4331: 9 0 10 | 11 0 0 | 12 0
4332: -------------------------------------
4333: 13 0 14 | 15 16 17 | 0 0
4334: Proc1 0 18 0 | 19 20 21 | 0 0
4335: 0 0 0 | 22 23 0 | 24 0
4336: -------------------------------------
4337: Proc2 25 26 27 | 0 0 28 | 29 0
4338: 30 0 0 | 31 32 33 | 0 34
4339: .ve
4341: This can be represented as a collection of submatrices as
4343: .vb
4344: A B C
4345: D E F
4346: G H I
4347: .ve
4349: Where the submatrices A,B,C are owned by proc0, D,E,F are
4350: owned by proc1, G,H,I are owned by proc2.
4352: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4353: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4354: The 'M','N' parameters are 8,8, and have the same values on all procs.
4356: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4357: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4358: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4359: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4360: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4361: matrix, ans [DF] as another SeqAIJ matrix.
4363: When d_nz, o_nz parameters are specified, d_nz storage elements are
4364: allocated for every row of the local diagonal submatrix, and o_nz
4365: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4366: One way to choose d_nz and o_nz is to use the max nonzerors per local
4367: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4368: In this case, the values of d_nz,o_nz are
4369: .vb
4370: proc0 : dnz = 2, o_nz = 2
4371: proc1 : dnz = 3, o_nz = 2
4372: proc2 : dnz = 1, o_nz = 4
4373: .ve
4374: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4375: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4376: for proc3. i.e we are using 12+15+10=37 storage locations to store
4377: 34 values.
4379: When d_nnz, o_nnz parameters are specified, the storage is specified
4380: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4381: In the above case the values for d_nnz,o_nnz are
4382: .vb
4383: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4384: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4385: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4386: .ve
4387: Here the space allocated is sum of all the above values i.e 34, and
4388: hence pre-allocation is perfect.
4390: Level: intermediate
4392: .keywords: matrix, aij, compressed row, sparse, parallel
4394: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4395: MATMPIAIJ, MatCreateMPIAIJWithArrays()
4396: @*/
4397: 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)
4398: {
4400: PetscMPIInt size;
4403: MatCreate(comm,A);
4404: MatSetSizes(*A,m,n,M,N);
4405: MPI_Comm_size(comm,&size);
4406: if (size > 1) {
4407: MatSetType(*A,MATMPIAIJ);
4408: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4409: } else {
4410: MatSetType(*A,MATSEQAIJ);
4411: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4412: }
4413: return(0);
4414: }
4416: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
4417: {
4418: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4419: PetscBool flg;
4423: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);
4424: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input");
4425: if (Ad) *Ad = a->A;
4426: if (Ao) *Ao = a->B;
4427: if (colmap) *colmap = a->garray;
4428: return(0);
4429: }
4431: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4432: {
4434: PetscInt m,N,i,rstart,nnz,Ii;
4435: PetscInt *indx;
4436: PetscScalar *values;
4439: MatGetSize(inmat,&m,&N);
4440: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4441: PetscInt *dnz,*onz,sum,bs,cbs;
4443: if (n == PETSC_DECIDE) {
4444: PetscSplitOwnership(comm,&n,&N);
4445: }
4446: /* Check sum(n) = N */
4447: MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4448: if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N);
4450: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4451: rstart -= m;
4453: MatPreallocateInitialize(comm,m,n,dnz,onz);
4454: for (i=0; i<m; i++) {
4455: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4456: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4457: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4458: }
4460: MatCreate(comm,outmat);
4461: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4462: MatGetBlockSizes(inmat,&bs,&cbs);
4463: MatSetBlockSizes(*outmat,bs,cbs);
4464: MatSetType(*outmat,MATAIJ);
4465: MatSeqAIJSetPreallocation(*outmat,0,dnz);
4466: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4467: MatPreallocateFinalize(dnz,onz);
4468: }
4470: /* numeric phase */
4471: MatGetOwnershipRange(*outmat,&rstart,NULL);
4472: for (i=0; i<m; i++) {
4473: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4474: Ii = i + rstart;
4475: MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4476: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4477: }
4478: MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
4479: MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
4480: return(0);
4481: }
4483: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4484: {
4485: PetscErrorCode ierr;
4486: PetscMPIInt rank;
4487: PetscInt m,N,i,rstart,nnz;
4488: size_t len;
4489: const PetscInt *indx;
4490: PetscViewer out;
4491: char *name;
4492: Mat B;
4493: const PetscScalar *values;
4496: MatGetLocalSize(A,&m,0);
4497: MatGetSize(A,0,&N);
4498: /* Should this be the type of the diagonal block of A? */
4499: MatCreate(PETSC_COMM_SELF,&B);
4500: MatSetSizes(B,m,N,m,N);
4501: MatSetBlockSizesFromMats(B,A,A);
4502: MatSetType(B,MATSEQAIJ);
4503: MatSeqAIJSetPreallocation(B,0,NULL);
4504: MatGetOwnershipRange(A,&rstart,0);
4505: for (i=0; i<m; i++) {
4506: MatGetRow(A,i+rstart,&nnz,&indx,&values);
4507: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4508: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4509: }
4510: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4511: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4513: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
4514: PetscStrlen(outfile,&len);
4515: PetscMalloc1(len+5,&name);
4516: sprintf(name,"%s.%d",outfile,rank);
4517: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4518: PetscFree(name);
4519: MatView(B,out);
4520: PetscViewerDestroy(&out);
4521: MatDestroy(&B);
4522: return(0);
4523: }
4525: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4526: {
4527: PetscErrorCode ierr;
4528: Mat_Merge_SeqsToMPI *merge;
4529: PetscContainer container;
4532: PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);
4533: if (container) {
4534: PetscContainerGetPointer(container,(void**)&merge);
4535: PetscFree(merge->id_r);
4536: PetscFree(merge->len_s);
4537: PetscFree(merge->len_r);
4538: PetscFree(merge->bi);
4539: PetscFree(merge->bj);
4540: PetscFree(merge->buf_ri[0]);
4541: PetscFree(merge->buf_ri);
4542: PetscFree(merge->buf_rj[0]);
4543: PetscFree(merge->buf_rj);
4544: PetscFree(merge->coi);
4545: PetscFree(merge->coj);
4546: PetscFree(merge->owners_co);
4547: PetscLayoutDestroy(&merge->rowmap);
4548: PetscFree(merge);
4549: PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4550: }
4551: MatDestroy_MPIAIJ(A);
4552: return(0);
4553: }
4555: #include <../src/mat/utils/freespace.h>
4556: #include <petscbt.h>
4558: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4559: {
4560: PetscErrorCode ierr;
4561: MPI_Comm comm;
4562: Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data;
4563: PetscMPIInt size,rank,taga,*len_s;
4564: PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4565: PetscInt proc,m;
4566: PetscInt **buf_ri,**buf_rj;
4567: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4568: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
4569: MPI_Request *s_waits,*r_waits;
4570: MPI_Status *status;
4571: MatScalar *aa=a->a;
4572: MatScalar **abuf_r,*ba_i;
4573: Mat_Merge_SeqsToMPI *merge;
4574: PetscContainer container;
4577: PetscObjectGetComm((PetscObject)mpimat,&comm);
4578: PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);
4580: MPI_Comm_size(comm,&size);
4581: MPI_Comm_rank(comm,&rank);
4583: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4584: PetscContainerGetPointer(container,(void**)&merge);
4586: bi = merge->bi;
4587: bj = merge->bj;
4588: buf_ri = merge->buf_ri;
4589: buf_rj = merge->buf_rj;
4591: PetscMalloc1(size,&status);
4592: owners = merge->rowmap->range;
4593: len_s = merge->len_s;
4595: /* send and recv matrix values */
4596: /*-----------------------------*/
4597: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4598: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
4600: PetscMalloc1(merge->nsend+1,&s_waits);
4601: for (proc=0,k=0; proc<size; proc++) {
4602: if (!len_s[proc]) continue;
4603: i = owners[proc];
4604: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4605: k++;
4606: }
4608: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4609: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4610: PetscFree(status);
4612: PetscFree(s_waits);
4613: PetscFree(r_waits);
4615: /* insert mat values of mpimat */
4616: /*----------------------------*/
4617: PetscMalloc1(N,&ba_i);
4618: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4620: for (k=0; k<merge->nrecv; k++) {
4621: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4622: nrows = *(buf_ri_k[k]);
4623: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
4624: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4625: }
4627: /* set values of ba */
4628: m = merge->rowmap->n;
4629: for (i=0; i<m; i++) {
4630: arow = owners[rank] + i;
4631: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
4632: bnzi = bi[i+1] - bi[i];
4633: PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));
4635: /* add local non-zero vals of this proc's seqmat into ba */
4636: anzi = ai[arow+1] - ai[arow];
4637: aj = a->j + ai[arow];
4638: aa = a->a + ai[arow];
4639: nextaj = 0;
4640: for (j=0; nextaj<anzi; j++) {
4641: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4642: ba_i[j] += aa[nextaj++];
4643: }
4644: }
4646: /* add received vals into ba */
4647: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4648: /* i-th row */
4649: if (i == *nextrow[k]) {
4650: anzi = *(nextai[k]+1) - *nextai[k];
4651: aj = buf_rj[k] + *(nextai[k]);
4652: aa = abuf_r[k] + *(nextai[k]);
4653: nextaj = 0;
4654: for (j=0; nextaj<anzi; j++) {
4655: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4656: ba_i[j] += aa[nextaj++];
4657: }
4658: }
4659: nextrow[k]++; nextai[k]++;
4660: }
4661: }
4662: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4663: }
4664: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4665: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
4667: PetscFree(abuf_r[0]);
4668: PetscFree(abuf_r);
4669: PetscFree(ba_i);
4670: PetscFree3(buf_ri_k,nextrow,nextai);
4671: PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4672: return(0);
4673: }
4675: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4676: {
4677: PetscErrorCode ierr;
4678: Mat B_mpi;
4679: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4680: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4681: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
4682: PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4683: PetscInt len,proc,*dnz,*onz,bs,cbs;
4684: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4685: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4686: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
4687: MPI_Status *status;
4688: PetscFreeSpaceList free_space=NULL,current_space=NULL;
4689: PetscBT lnkbt;
4690: Mat_Merge_SeqsToMPI *merge;
4691: PetscContainer container;
4694: PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);
4696: /* make sure it is a PETSc comm */
4697: PetscCommDuplicate(comm,&comm,NULL);
4698: MPI_Comm_size(comm,&size);
4699: MPI_Comm_rank(comm,&rank);
4701: PetscNew(&merge);
4702: PetscMalloc1(size,&status);
4704: /* determine row ownership */
4705: /*---------------------------------------------------------*/
4706: PetscLayoutCreate(comm,&merge->rowmap);
4707: PetscLayoutSetLocalSize(merge->rowmap,m);
4708: PetscLayoutSetSize(merge->rowmap,M);
4709: PetscLayoutSetBlockSize(merge->rowmap,1);
4710: PetscLayoutSetUp(merge->rowmap);
4711: PetscMalloc1(size,&len_si);
4712: PetscMalloc1(size,&merge->len_s);
4714: m = merge->rowmap->n;
4715: owners = merge->rowmap->range;
4717: /* determine the number of messages to send, their lengths */
4718: /*---------------------------------------------------------*/
4719: len_s = merge->len_s;
4721: len = 0; /* length of buf_si[] */
4722: merge->nsend = 0;
4723: for (proc=0; proc<size; proc++) {
4724: len_si[proc] = 0;
4725: if (proc == rank) {
4726: len_s[proc] = 0;
4727: } else {
4728: len_si[proc] = owners[proc+1] - owners[proc] + 1;
4729: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4730: }
4731: if (len_s[proc]) {
4732: merge->nsend++;
4733: nrows = 0;
4734: for (i=owners[proc]; i<owners[proc+1]; i++) {
4735: if (ai[i+1] > ai[i]) nrows++;
4736: }
4737: len_si[proc] = 2*(nrows+1);
4738: len += len_si[proc];
4739: }
4740: }
4742: /* determine the number and length of messages to receive for ij-structure */
4743: /*-------------------------------------------------------------------------*/
4744: PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4745: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
4747: /* post the Irecv of j-structure */
4748: /*-------------------------------*/
4749: PetscCommGetNewTag(comm,&tagj);
4750: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
4752: /* post the Isend of j-structure */
4753: /*--------------------------------*/
4754: PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);
4756: for (proc=0, k=0; proc<size; proc++) {
4757: if (!len_s[proc]) continue;
4758: i = owners[proc];
4759: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4760: k++;
4761: }
4763: /* receives and sends of j-structure are complete */
4764: /*------------------------------------------------*/
4765: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4766: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
4768: /* send and recv i-structure */
4769: /*---------------------------*/
4770: PetscCommGetNewTag(comm,&tagi);
4771: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
4773: PetscMalloc1(len+1,&buf_s);
4774: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4775: for (proc=0,k=0; proc<size; proc++) {
4776: if (!len_s[proc]) continue;
4777: /* form outgoing message for i-structure:
4778: buf_si[0]: nrows to be sent
4779: [1:nrows]: row index (global)
4780: [nrows+1:2*nrows+1]: i-structure index
4781: */
4782: /*-------------------------------------------*/
4783: nrows = len_si[proc]/2 - 1;
4784: buf_si_i = buf_si + nrows+1;
4785: buf_si[0] = nrows;
4786: buf_si_i[0] = 0;
4787: nrows = 0;
4788: for (i=owners[proc]; i<owners[proc+1]; i++) {
4789: anzi = ai[i+1] - ai[i];
4790: if (anzi) {
4791: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4792: buf_si[nrows+1] = i-owners[proc]; /* local row index */
4793: nrows++;
4794: }
4795: }
4796: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4797: k++;
4798: buf_si += len_si[proc];
4799: }
4801: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4802: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
4804: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4805: for (i=0; i<merge->nrecv; i++) {
4806: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4807: }
4809: PetscFree(len_si);
4810: PetscFree(len_ri);
4811: PetscFree(rj_waits);
4812: PetscFree2(si_waits,sj_waits);
4813: PetscFree(ri_waits);
4814: PetscFree(buf_s);
4815: PetscFree(status);
4817: /* compute a local seq matrix in each processor */
4818: /*----------------------------------------------*/
4819: /* allocate bi array and free space for accumulating nonzero column info */
4820: PetscMalloc1(m+1,&bi);
4821: bi[0] = 0;
4823: /* create and initialize a linked list */
4824: nlnk = N+1;
4825: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
4827: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4828: len = ai[owners[rank+1]] - ai[owners[rank]];
4829: PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);
4831: current_space = free_space;
4833: /* determine symbolic info for each local row */
4834: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4836: for (k=0; k<merge->nrecv; k++) {
4837: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4838: nrows = *buf_ri_k[k];
4839: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4840: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4841: }
4843: MatPreallocateInitialize(comm,m,n,dnz,onz);
4844: len = 0;
4845: for (i=0; i<m; i++) {
4846: bnzi = 0;
4847: /* add local non-zero cols of this proc's seqmat into lnk */
4848: arow = owners[rank] + i;
4849: anzi = ai[arow+1] - ai[arow];
4850: aj = a->j + ai[arow];
4851: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4852: bnzi += nlnk;
4853: /* add received col data into lnk */
4854: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4855: if (i == *nextrow[k]) { /* i-th row */
4856: anzi = *(nextai[k]+1) - *nextai[k];
4857: aj = buf_rj[k] + *nextai[k];
4858: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4859: bnzi += nlnk;
4860: nextrow[k]++; nextai[k]++;
4861: }
4862: }
4863: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4865: /* if free space is not available, make more free space */
4866: if (current_space->local_remaining<bnzi) {
4867: PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);
4868: nspacedouble++;
4869: }
4870: /* copy data into free space, then initialize lnk */
4871: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4872: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
4874: current_space->array += bnzi;
4875: current_space->local_used += bnzi;
4876: current_space->local_remaining -= bnzi;
4878: bi[i+1] = bi[i] + bnzi;
4879: }
4881: PetscFree3(buf_ri_k,nextrow,nextai);
4883: PetscMalloc1(bi[m]+1,&bj);
4884: PetscFreeSpaceContiguous(&free_space,bj);
4885: PetscLLDestroy(lnk,lnkbt);
4887: /* create symbolic parallel matrix B_mpi */
4888: /*---------------------------------------*/
4889: MatGetBlockSizes(seqmat,&bs,&cbs);
4890: MatCreate(comm,&B_mpi);
4891: if (n==PETSC_DECIDE) {
4892: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4893: } else {
4894: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4895: }
4896: MatSetBlockSizes(B_mpi,bs,cbs);
4897: MatSetType(B_mpi,MATMPIAIJ);
4898: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4899: MatPreallocateFinalize(dnz,onz);
4900: MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
4902: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4903: B_mpi->assembled = PETSC_FALSE;
4904: B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
4905: merge->bi = bi;
4906: merge->bj = bj;
4907: merge->buf_ri = buf_ri;
4908: merge->buf_rj = buf_rj;
4909: merge->coi = NULL;
4910: merge->coj = NULL;
4911: merge->owners_co = NULL;
4913: PetscCommDestroy(&comm);
4915: /* attach the supporting struct to B_mpi for reuse */
4916: PetscContainerCreate(PETSC_COMM_SELF,&container);
4917: PetscContainerSetPointer(container,merge);
4918: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4919: PetscContainerDestroy(&container);
4920: *mpimat = B_mpi;
4922: PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4923: return(0);
4924: }
4926: /*@C
4927: MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential
4928: matrices from each processor
4930: Collective on MPI_Comm
4932: Input Parameters:
4933: + comm - the communicators the parallel matrix will live on
4934: . seqmat - the input sequential matrices
4935: . m - number of local rows (or PETSC_DECIDE)
4936: . n - number of local columns (or PETSC_DECIDE)
4937: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4939: Output Parameter:
4940: . mpimat - the parallel matrix generated
4942: Level: advanced
4944: Notes:
4945: The dimensions of the sequential matrix in each processor MUST be the same.
4946: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4947: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4948: @*/
4949: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4950: {
4952: PetscMPIInt size;
4955: MPI_Comm_size(comm,&size);
4956: if (size == 1) {
4957: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4958: if (scall == MAT_INITIAL_MATRIX) {
4959: MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4960: } else {
4961: MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4962: }
4963: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4964: return(0);
4965: }
4966: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4967: if (scall == MAT_INITIAL_MATRIX) {
4968: MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4969: }
4970: MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4971: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4972: return(0);
4973: }
4975: /*@
4976: MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with
4977: mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4978: with MatGetSize()
4980: Not Collective
4982: Input Parameters:
4983: + A - the matrix
4984: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4986: Output Parameter:
4987: . A_loc - the local sequential matrix generated
4989: Level: developer
4991: .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed()
4993: @*/
4994: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4995: {
4997: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
4998: Mat_SeqAIJ *mat,*a,*b;
4999: PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
5000: MatScalar *aa,*ba,*cam;
5001: PetscScalar *ca;
5002: PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
5003: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
5004: PetscBool match;
5005: MPI_Comm comm;
5006: PetscMPIInt size;
5009: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5010: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5011: PetscObjectGetComm((PetscObject)A,&comm);
5012: MPI_Comm_size(comm,&size);
5013: if (size == 1 && scall == MAT_REUSE_MATRIX) return(0);
5015: PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
5016: a = (Mat_SeqAIJ*)(mpimat->A)->data;
5017: b = (Mat_SeqAIJ*)(mpimat->B)->data;
5018: ai = a->i; aj = a->j; bi = b->i; bj = b->j;
5019: aa = a->a; ba = b->a;
5020: if (scall == MAT_INITIAL_MATRIX) {
5021: if (size == 1) {
5022: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);
5023: return(0);
5024: }
5026: PetscMalloc1(1+am,&ci);
5027: ci[0] = 0;
5028: for (i=0; i<am; i++) {
5029: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
5030: }
5031: PetscMalloc1(1+ci[am],&cj);
5032: PetscMalloc1(1+ci[am],&ca);
5033: k = 0;
5034: for (i=0; i<am; i++) {
5035: ncols_o = bi[i+1] - bi[i];
5036: ncols_d = ai[i+1] - ai[i];
5037: /* off-diagonal portion of A */
5038: for (jo=0; jo<ncols_o; jo++) {
5039: col = cmap[*bj];
5040: if (col >= cstart) break;
5041: cj[k] = col; bj++;
5042: ca[k++] = *ba++;
5043: }
5044: /* diagonal portion of A */
5045: for (j=0; j<ncols_d; j++) {
5046: cj[k] = cstart + *aj++;
5047: ca[k++] = *aa++;
5048: }
5049: /* off-diagonal portion of A */
5050: for (j=jo; j<ncols_o; j++) {
5051: cj[k] = cmap[*bj++];
5052: ca[k++] = *ba++;
5053: }
5054: }
5055: /* put together the new matrix */
5056: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
5057: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5058: /* Since these are PETSc arrays, change flags to free them as necessary. */
5059: mat = (Mat_SeqAIJ*)(*A_loc)->data;
5060: mat->free_a = PETSC_TRUE;
5061: mat->free_ij = PETSC_TRUE;
5062: mat->nonew = 0;
5063: } else if (scall == MAT_REUSE_MATRIX) {
5064: mat=(Mat_SeqAIJ*)(*A_loc)->data;
5065: ci = mat->i; cj = mat->j; cam = mat->a;
5066: for (i=0; i<am; i++) {
5067: /* off-diagonal portion of A */
5068: ncols_o = bi[i+1] - bi[i];
5069: for (jo=0; jo<ncols_o; jo++) {
5070: col = cmap[*bj];
5071: if (col >= cstart) break;
5072: *cam++ = *ba++; bj++;
5073: }
5074: /* diagonal portion of A */
5075: ncols_d = ai[i+1] - ai[i];
5076: for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5077: /* off-diagonal portion of A */
5078: for (j=jo; j<ncols_o; j++) {
5079: *cam++ = *ba++; bj++;
5080: }
5081: }
5082: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5083: PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5084: return(0);
5085: }
5087: /*@C
5088: MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns
5090: Not Collective
5092: Input Parameters:
5093: + A - the matrix
5094: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5095: - row, col - index sets of rows and columns to extract (or NULL)
5097: Output Parameter:
5098: . A_loc - the local sequential matrix generated
5100: Level: developer
5102: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()
5104: @*/
5105: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5106: {
5107: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5109: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5110: IS isrowa,iscola;
5111: Mat *aloc;
5112: PetscBool match;
5115: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5116: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5117: PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5118: if (!row) {
5119: start = A->rmap->rstart; end = A->rmap->rend;
5120: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5121: } else {
5122: isrowa = *row;
5123: }
5124: if (!col) {
5125: start = A->cmap->rstart;
5126: cmap = a->garray;
5127: nzA = a->A->cmap->n;
5128: nzB = a->B->cmap->n;
5129: PetscMalloc1(nzA+nzB, &idx);
5130: ncols = 0;
5131: for (i=0; i<nzB; i++) {
5132: if (cmap[i] < start) idx[ncols++] = cmap[i];
5133: else break;
5134: }
5135: imark = i;
5136: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5137: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5138: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5139: } else {
5140: iscola = *col;
5141: }
5142: if (scall != MAT_INITIAL_MATRIX) {
5143: PetscMalloc1(1,&aloc);
5144: aloc[0] = *A_loc;
5145: }
5146: MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5147: if (!col) { /* attach global id of condensed columns */
5148: PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);
5149: }
5150: *A_loc = aloc[0];
5151: PetscFree(aloc);
5152: if (!row) {
5153: ISDestroy(&isrowa);
5154: }
5155: if (!col) {
5156: ISDestroy(&iscola);
5157: }
5158: PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5159: return(0);
5160: }
5162: /*@C
5163: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5165: Collective on Mat
5167: Input Parameters:
5168: + A,B - the matrices in mpiaij format
5169: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5170: - rowb, colb - index sets of rows and columns of B to extract (or NULL)
5172: Output Parameter:
5173: + rowb, colb - index sets of rows and columns of B to extract
5174: - B_seq - the sequential matrix generated
5176: Level: developer
5178: @*/
5179: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5180: {
5181: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5183: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5184: IS isrowb,iscolb;
5185: Mat *bseq=NULL;
5188: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5189: 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);
5190: }
5191: PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5193: if (scall == MAT_INITIAL_MATRIX) {
5194: start = A->cmap->rstart;
5195: cmap = a->garray;
5196: nzA = a->A->cmap->n;
5197: nzB = a->B->cmap->n;
5198: PetscMalloc1(nzA+nzB, &idx);
5199: ncols = 0;
5200: for (i=0; i<nzB; i++) { /* row < local row index */
5201: if (cmap[i] < start) idx[ncols++] = cmap[i];
5202: else break;
5203: }
5204: imark = i;
5205: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
5206: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5207: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5208: ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5209: } else {
5210: if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5211: isrowb = *rowb; iscolb = *colb;
5212: PetscMalloc1(1,&bseq);
5213: bseq[0] = *B_seq;
5214: }
5215: MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5216: *B_seq = bseq[0];
5217: PetscFree(bseq);
5218: if (!rowb) {
5219: ISDestroy(&isrowb);
5220: } else {
5221: *rowb = isrowb;
5222: }
5223: if (!colb) {
5224: ISDestroy(&iscolb);
5225: } else {
5226: *colb = iscolb;
5227: }
5228: PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5229: return(0);
5230: }
5232: #include <petsc/private/vecscatterimpl.h>
5233: /*
5234: MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5235: of the OFF-DIAGONAL portion of local A
5237: Collective on Mat
5239: Input Parameters:
5240: + A,B - the matrices in mpiaij format
5241: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5243: Output Parameter:
5244: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5245: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5246: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5247: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5249: Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product
5250: for this matrix. This is not desirable..
5252: Level: developer
5254: */
5255: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5256: {
5257: VecScatter_MPI_General *gen_to,*gen_from;
5258: PetscErrorCode ierr;
5259: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5260: Mat_SeqAIJ *b_oth;
5261: VecScatter ctx;
5262: MPI_Comm comm;
5263: PetscMPIInt *rprocs,*sprocs,tag,rank;
5264: PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj;
5265: PetscInt *rvalues,*svalues,*cols,sbs,rbs;
5266: PetscScalar *b_otha,*bufa,*bufA,*vals;
5267: PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
5268: MPI_Request *rwaits = NULL,*swaits = NULL;
5269: MPI_Status *sstatus,rstatus;
5270: PetscMPIInt jj,size;
5271: VecScatterType type;
5272: PetscBool mpi1;
5275: PetscObjectGetComm((PetscObject)A,&comm);
5276: MPI_Comm_size(comm,&size);
5278: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5279: 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);
5280: }
5281: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5282: MPI_Comm_rank(comm,&rank);
5284: if (size == 1) {
5285: startsj_s = NULL;
5286: bufa_ptr = NULL;
5287: *B_oth = NULL;
5288: return(0);
5289: }
5291: ctx = a->Mvctx;
5292: VecScatterGetType(ctx,&type);
5293: PetscStrcmp(type,"mpi1",&mpi1);
5294: if (!mpi1) {
5295: /* a->Mvctx is not type MPI1 which is not implemented for Mat-Mat ops,
5296: thus create a->Mvctx_mpi1 */
5297: if (!a->Mvctx_mpi1) {
5298: a->Mvctx_mpi1_flg = PETSC_TRUE;
5299: MatSetUpMultiply_MPIAIJ(A);
5300: }
5301: ctx = a->Mvctx_mpi1;
5302: }
5303: tag = ((PetscObject)ctx)->tag;
5305: gen_to = (VecScatter_MPI_General*)ctx->todata;
5306: gen_from = (VecScatter_MPI_General*)ctx->fromdata;
5307: nrecvs = gen_from->n;
5308: nsends = gen_to->n;
5310: PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);
5311: srow = gen_to->indices; /* local row index to be sent */
5312: sstarts = gen_to->starts;
5313: sprocs = gen_to->procs;
5314: sstatus = gen_to->sstatus;
5315: sbs = gen_to->bs;
5316: rstarts = gen_from->starts;
5317: rprocs = gen_from->procs;
5318: rbs = gen_from->bs;
5320: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5321: if (scall == MAT_INITIAL_MATRIX) {
5322: /* i-array */
5323: /*---------*/
5324: /* post receives */
5325: PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);
5326: for (i=0; i<nrecvs; i++) {
5327: rowlen = rvalues + rstarts[i]*rbs;
5328: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5329: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5330: }
5332: /* pack the outgoing message */
5333: PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);
5335: sstartsj[0] = 0;
5336: rstartsj[0] = 0;
5337: len = 0; /* total length of j or a array to be sent */
5338: k = 0;
5339: PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);
5340: for (i=0; i<nsends; i++) {
5341: rowlen = svalues + sstarts[i]*sbs;
5342: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5343: for (j=0; j<nrows; j++) {
5344: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5345: for (l=0; l<sbs; l++) {
5346: MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */
5348: rowlen[j*sbs+l] = ncols;
5350: len += ncols;
5351: MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
5352: }
5353: k++;
5354: }
5355: MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5357: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5358: }
5359: /* recvs and sends of i-array are completed */
5360: i = nrecvs;
5361: while (i--) {
5362: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5363: }
5364: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5365: PetscFree(svalues);
5367: /* allocate buffers for sending j and a arrays */
5368: PetscMalloc1(len+1,&bufj);
5369: PetscMalloc1(len+1,&bufa);
5371: /* create i-array of B_oth */
5372: PetscMalloc1(aBn+2,&b_othi);
5374: b_othi[0] = 0;
5375: len = 0; /* total length of j or a array to be received */
5376: k = 0;
5377: for (i=0; i<nrecvs; i++) {
5378: rowlen = rvalues + rstarts[i]*rbs;
5379: nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */
5380: for (j=0; j<nrows; j++) {
5381: b_othi[k+1] = b_othi[k] + rowlen[j];
5382: PetscIntSumError(rowlen[j],len,&len);
5383: k++;
5384: }
5385: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5386: }
5387: PetscFree(rvalues);
5389: /* allocate space for j and a arrrays of B_oth */
5390: PetscMalloc1(b_othi[aBn]+1,&b_othj);
5391: PetscMalloc1(b_othi[aBn]+1,&b_otha);
5393: /* j-array */
5394: /*---------*/
5395: /* post receives of j-array */
5396: for (i=0; i<nrecvs; i++) {
5397: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5398: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5399: }
5401: /* pack the outgoing message j-array */
5402: k = 0;
5403: for (i=0; i<nsends; i++) {
5404: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5405: bufJ = bufj+sstartsj[i];
5406: for (j=0; j<nrows; j++) {
5407: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5408: for (ll=0; ll<sbs; ll++) {
5409: MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5410: for (l=0; l<ncols; l++) {
5411: *bufJ++ = cols[l];
5412: }
5413: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5414: }
5415: }
5416: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5417: }
5419: /* recvs and sends of j-array are completed */
5420: i = nrecvs;
5421: while (i--) {
5422: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5423: }
5424: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5425: } else if (scall == MAT_REUSE_MATRIX) {
5426: sstartsj = *startsj_s;
5427: rstartsj = *startsj_r;
5428: bufa = *bufa_ptr;
5429: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5430: b_otha = b_oth->a;
5431: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5433: /* a-array */
5434: /*---------*/
5435: /* post receives of a-array */
5436: for (i=0; i<nrecvs; i++) {
5437: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5438: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5439: }
5441: /* pack the outgoing message a-array */
5442: k = 0;
5443: for (i=0; i<nsends; i++) {
5444: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5445: bufA = bufa+sstartsj[i];
5446: for (j=0; j<nrows; j++) {
5447: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5448: for (ll=0; ll<sbs; ll++) {
5449: MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5450: for (l=0; l<ncols; l++) {
5451: *bufA++ = vals[l];
5452: }
5453: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5454: }
5455: }
5456: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5457: }
5458: /* recvs and sends of a-array are completed */
5459: i = nrecvs;
5460: while (i--) {
5461: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5462: }
5463: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5464: PetscFree2(rwaits,swaits);
5466: if (scall == MAT_INITIAL_MATRIX) {
5467: /* put together the new matrix */
5468: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);
5470: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5471: /* Since these are PETSc arrays, change flags to free them as necessary. */
5472: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5473: b_oth->free_a = PETSC_TRUE;
5474: b_oth->free_ij = PETSC_TRUE;
5475: b_oth->nonew = 0;
5477: PetscFree(bufj);
5478: if (!startsj_s || !bufa_ptr) {
5479: PetscFree2(sstartsj,rstartsj);
5480: PetscFree(bufa_ptr);
5481: } else {
5482: *startsj_s = sstartsj;
5483: *startsj_r = rstartsj;
5484: *bufa_ptr = bufa;
5485: }
5486: }
5487: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5488: return(0);
5489: }
5491: /*@C
5492: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
5494: Not Collective
5496: Input Parameters:
5497: . A - The matrix in mpiaij format
5499: Output Parameter:
5500: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5501: . colmap - A map from global column index to local index into lvec
5502: - multScatter - A scatter from the argument of a matrix-vector product to lvec
5504: Level: developer
5506: @*/
5507: #if defined(PETSC_USE_CTABLE)
5508: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5509: #else
5510: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5511: #endif
5512: {
5513: Mat_MPIAIJ *a;
5520: a = (Mat_MPIAIJ*) A->data;
5521: if (lvec) *lvec = a->lvec;
5522: if (colmap) *colmap = a->colmap;
5523: if (multScatter) *multScatter = a->Mvctx;
5524: return(0);
5525: }
5527: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
5528: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
5529: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*);
5530: #if defined(PETSC_HAVE_MKL_SPARSE)
5531: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*);
5532: #endif
5533: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
5534: #if defined(PETSC_HAVE_ELEMENTAL)
5535: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
5536: #endif
5537: #if defined(PETSC_HAVE_HYPRE)
5538: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
5539: PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
5540: #endif
5541: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
5542: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*);
5543: PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
5545: /*
5546: Computes (B'*A')' since computing B*A directly is untenable
5548: n p p
5549: ( ) ( ) ( )
5550: m ( A ) * n ( B ) = m ( C )
5551: ( ) ( ) ( )
5553: */
5554: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5555: {
5557: Mat At,Bt,Ct;
5560: MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5561: MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5562: MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5563: MatDestroy(&At);
5564: MatDestroy(&Bt);
5565: MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5566: MatDestroy(&Ct);
5567: return(0);
5568: }
5570: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
5571: {
5573: PetscInt m=A->rmap->n,n=B->cmap->n;
5574: Mat Cmat;
5577: 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);
5578: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
5579: MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5580: MatSetBlockSizesFromMats(Cmat,A,B);
5581: MatSetType(Cmat,MATMPIDENSE);
5582: MatMPIDenseSetPreallocation(Cmat,NULL);
5583: MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5584: MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
5586: Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
5588: *C = Cmat;
5589: return(0);
5590: }
5592: /* ----------------------------------------------------------------*/
5593: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
5594: {
5598: if (scall == MAT_INITIAL_MATRIX) {
5599: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
5600: MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
5601: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
5602: }
5603: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
5604: MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
5605: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
5606: return(0);
5607: }
5609: /*MC
5610: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
5612: Options Database Keys:
5613: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
5615: Level: beginner
5617: .seealso: MatCreateAIJ()
5618: M*/
5620: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
5621: {
5622: Mat_MPIAIJ *b;
5624: PetscMPIInt size;
5627: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
5629: PetscNewLog(B,&b);
5630: B->data = (void*)b;
5631: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5632: B->assembled = PETSC_FALSE;
5633: B->insertmode = NOT_SET_VALUES;
5634: b->size = size;
5636: MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
5638: /* build cache for off array entries formed */
5639: MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);
5641: b->donotstash = PETSC_FALSE;
5642: b->colmap = 0;
5643: b->garray = 0;
5644: b->roworiented = PETSC_TRUE;
5646: /* stuff used for matrix vector multiply */
5647: b->lvec = NULL;
5648: b->Mvctx = NULL;
5650: /* stuff for MatGetRow() */
5651: b->rowindices = 0;
5652: b->rowvalues = 0;
5653: b->getrowactive = PETSC_FALSE;
5655: /* flexible pointer used in CUSP/CUSPARSE classes */
5656: b->spptr = NULL;
5658: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
5659: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
5660: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
5661: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
5662: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
5663: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);
5664: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
5665: PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
5666: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
5667: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);
5668: #if defined(PETSC_HAVE_MKL_SPARSE)
5669: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);
5670: #endif
5671: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
5672: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
5673: #if defined(PETSC_HAVE_ELEMENTAL)
5674: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
5675: #endif
5676: #if defined(PETSC_HAVE_HYPRE)
5677: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);
5678: #endif
5679: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);
5680: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);
5681: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);
5682: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);
5683: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);
5684: #if defined(PETSC_HAVE_HYPRE)
5685: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
5686: #endif
5687: PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_mpiaij_C",MatPtAP_IS_XAIJ);
5688: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5689: return(0);
5690: }
5692: /*@C
5693: MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5694: and "off-diagonal" part of the matrix in CSR format.
5696: Collective on MPI_Comm
5698: Input Parameters:
5699: + comm - MPI communicator
5700: . m - number of local rows (Cannot be PETSC_DECIDE)
5701: . n - This value should be the same as the local size used in creating the
5702: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5703: calculated if N is given) For square matrices n is almost always m.
5704: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5705: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5706: . i - row indices for "diagonal" portion of matrix
5707: . j - column indices
5708: . a - matrix values
5709: . oi - row indices for "off-diagonal" portion of matrix
5710: . oj - column indices
5711: - oa - matrix values
5713: Output Parameter:
5714: . mat - the matrix
5716: Level: advanced
5718: Notes:
5719: The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5720: must free the arrays once the matrix has been destroyed and not before.
5722: The i and j indices are 0 based
5724: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix
5726: This sets local rows and cannot be used to set off-processor values.
5728: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
5729: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
5730: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
5731: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
5732: keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
5733: communication if it is known that only local entries will be set.
5735: .keywords: matrix, aij, compressed row, sparse, parallel
5737: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5738: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5739: @*/
5740: 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)
5741: {
5743: Mat_MPIAIJ *maij;
5746: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5747: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5748: if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5749: MatCreate(comm,mat);
5750: MatSetSizes(*mat,m,n,M,N);
5751: MatSetType(*mat,MATMPIAIJ);
5752: maij = (Mat_MPIAIJ*) (*mat)->data;
5754: (*mat)->preallocated = PETSC_TRUE;
5756: PetscLayoutSetUp((*mat)->rmap);
5757: PetscLayoutSetUp((*mat)->cmap);
5759: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5760: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);
5762: MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5763: MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5764: MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5765: MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);
5767: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
5768: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5769: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5770: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
5771: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
5772: return(0);
5773: }
5775: /*
5776: Special version for direct calls from Fortran
5777: */
5778: #include <petsc/private/fortranimpl.h>
5780: /* Change these macros so can be used in void function */
5781: #undef CHKERRQ
5782: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
5783: #undef SETERRQ2
5784: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5785: #undef SETERRQ3
5786: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5787: #undef SETERRQ
5788: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)
5790: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5791: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
5792: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5793: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
5794: #else
5795: #endif
5796: PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
5797: {
5798: Mat mat = *mmat;
5799: PetscInt m = *mm, n = *mn;
5800: InsertMode addv = *maddv;
5801: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
5802: PetscScalar value;
5805: MatCheckPreallocated(mat,1);
5806: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
5808: #if defined(PETSC_USE_DEBUG)
5809: else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
5810: #endif
5811: {
5812: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
5813: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
5814: PetscBool roworiented = aij->roworiented;
5816: /* Some Variables required in the macro */
5817: Mat A = aij->A;
5818: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
5819: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
5820: MatScalar *aa = a->a;
5821: PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
5822: Mat B = aij->B;
5823: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
5824: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
5825: MatScalar *ba = b->a;
5827: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
5828: PetscInt nonew = a->nonew;
5829: MatScalar *ap1,*ap2;
5832: for (i=0; i<m; i++) {
5833: if (im[i] < 0) continue;
5834: #if defined(PETSC_USE_DEBUG)
5835: if (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);
5836: #endif
5837: if (im[i] >= rstart && im[i] < rend) {
5838: row = im[i] - rstart;
5839: lastcol1 = -1;
5840: rp1 = aj + ai[row];
5841: ap1 = aa + ai[row];
5842: rmax1 = aimax[row];
5843: nrow1 = ailen[row];
5844: low1 = 0;
5845: high1 = nrow1;
5846: lastcol2 = -1;
5847: rp2 = bj + bi[row];
5848: ap2 = ba + bi[row];
5849: rmax2 = bimax[row];
5850: nrow2 = bilen[row];
5851: low2 = 0;
5852: high2 = nrow2;
5854: for (j=0; j<n; j++) {
5855: if (roworiented) value = v[i*n+j];
5856: else value = v[i+j*m];
5857: if (in[j] >= cstart && in[j] < cend) {
5858: col = in[j] - cstart;
5859: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
5860: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
5861: } else if (in[j] < 0) continue;
5862: #if defined(PETSC_USE_DEBUG)
5863: /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */
5864: 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);}
5865: #endif
5866: else {
5867: if (mat->was_assembled) {
5868: if (!aij->colmap) {
5869: MatCreateColmap_MPIAIJ_Private(mat);
5870: }
5871: #if defined(PETSC_USE_CTABLE)
5872: PetscTableFind(aij->colmap,in[j]+1,&col);
5873: col--;
5874: #else
5875: col = aij->colmap[in[j]] - 1;
5876: #endif
5877: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
5878: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
5879: MatDisAssemble_MPIAIJ(mat);
5880: col = in[j];
5881: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
5882: B = aij->B;
5883: b = (Mat_SeqAIJ*)B->data;
5884: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
5885: rp2 = bj + bi[row];
5886: ap2 = ba + bi[row];
5887: rmax2 = bimax[row];
5888: nrow2 = bilen[row];
5889: low2 = 0;
5890: high2 = nrow2;
5891: bm = aij->B->rmap->n;
5892: ba = b->a;
5893: }
5894: } else col = in[j];
5895: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
5896: }
5897: }
5898: } else if (!aij->donotstash) {
5899: if (roworiented) {
5900: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5901: } else {
5902: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5903: }
5904: }
5905: }
5906: }
5907: PetscFunctionReturnVoid();
5908: }