Actual source code: mpiaij.c
1: #define PETSCMAT_DLL
3: #include src/mat/impls/aij/mpi/mpiaij.h
4: #include src/inline/spops.h
6: /*
7: Local utility routine that creates a mapping from the global column
8: number to the local number in the off-diagonal part of the local
9: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
10: a slightly higher hash table cost; without it it is not scalable (each processor
11: has an order N integer array but is fast to acess.
12: */
15: PetscErrorCode CreateColmap_MPIAIJ_Private(Mat mat)
16: {
17: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
19: PetscInt n = aij->B->cmap.n,i;
22: #if defined (PETSC_USE_CTABLE)
23: PetscTableCreate(n,&aij->colmap);
24: for (i=0; i<n; i++){
25: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1);
26: }
27: #else
28: PetscMalloc((mat->cmap.N+1)*sizeof(PetscInt),&aij->colmap);
29: PetscLogObjectMemory(mat,mat->cmap.N*sizeof(PetscInt));
30: PetscMemzero(aij->colmap,mat->cmap.N*sizeof(PetscInt));
31: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
32: #endif
33: return(0);
34: }
37: #define CHUNKSIZE 15
38: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \
39: { \
40: if (col <= lastcol1) low1 = 0; else high1 = nrow1; \
41: lastcol1 = col;\
42: while (high1-low1 > 5) { \
43: t = (low1+high1)/2; \
44: if (rp1[t] > col) high1 = t; \
45: else low1 = t; \
46: } \
47: for (_i=low1; _i<high1; _i++) { \
48: if (rp1[_i] > col) break; \
49: if (rp1[_i] == col) { \
50: if (addv == ADD_VALUES) ap1[_i] += value; \
51: else ap1[_i] = value; \
52: goto a_noinsert; \
53: } \
54: } \
55: if (value == 0.0 && ignorezeroentries) goto a_noinsert; \
56: if (nonew == 1) goto a_noinsert; \
57: if (nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
58: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew); \
59: N = nrow1++ - 1; a->nz++; high1++; \
60: /* shift up all the later entries in this row */ \
61: for (ii=N; ii>=_i; ii--) { \
62: rp1[ii+1] = rp1[ii]; \
63: ap1[ii+1] = ap1[ii]; \
64: } \
65: rp1[_i] = col; \
66: ap1[_i] = value; \
67: a_noinsert: ; \
68: ailen[row] = nrow1; \
69: }
72: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \
73: { \
74: if (col <= lastcol2) low2 = 0; else high2 = nrow2; \
75: lastcol2 = col;\
76: while (high2-low2 > 5) { \
77: t = (low2+high2)/2; \
78: if (rp2[t] > col) high2 = t; \
79: else low2 = t; \
80: } \
81: for (_i=low2; _i<high2; _i++) { \
82: if (rp2[_i] > col) break; \
83: if (rp2[_i] == col) { \
84: if (addv == ADD_VALUES) ap2[_i] += value; \
85: else ap2[_i] = value; \
86: goto b_noinsert; \
87: } \
88: } \
89: if (value == 0.0 && ignorezeroentries) goto b_noinsert; \
90: if (nonew == 1) goto b_noinsert; \
91: if (nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
92: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew); \
93: N = nrow2++ - 1; b->nz++; high2++;\
94: /* shift up all the later entries in this row */ \
95: for (ii=N; ii>=_i; ii--) { \
96: rp2[ii+1] = rp2[ii]; \
97: ap2[ii+1] = ap2[ii]; \
98: } \
99: rp2[_i] = col; \
100: ap2[_i] = value; \
101: b_noinsert: ; \
102: bilen[row] = nrow2; \
103: }
107: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
108: {
109: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
110: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
112: PetscInt l,*garray = mat->garray,diag;
115: /* code only works for square matrices A */
117: /* find size of row to the left of the diagonal part */
118: MatGetOwnershipRange(A,&diag,0);
119: row = row - diag;
120: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
121: if (garray[b->j[b->i[row]+l]] > diag) break;
122: }
123: PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));
125: /* diagonal part */
126: PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));
128: /* right of diagonal part */
129: 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));
130: return(0);
131: }
135: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
136: {
137: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
138: PetscScalar value;
140: PetscInt i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
141: PetscInt cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;
142: PetscTruth roworiented = aij->roworiented;
144: /* Some Variables required in the macro */
145: Mat A = aij->A;
146: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
147: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
148: PetscScalar *aa = a->a;
149: PetscTruth ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
150: Mat B = aij->B;
151: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
152: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap.n,am = aij->A->rmap.n;
153: PetscScalar *ba = b->a;
155: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
156: PetscInt nonew = a->nonew;
157: PetscScalar *ap1,*ap2;
160: for (i=0; i<m; i++) {
161: if (im[i] < 0) continue;
162: #if defined(PETSC_USE_DEBUG)
163: if (im[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap.N-1);
164: #endif
165: if (im[i] >= rstart && im[i] < rend) {
166: row = im[i] - rstart;
167: lastcol1 = -1;
168: rp1 = aj + ai[row];
169: ap1 = aa + ai[row];
170: rmax1 = aimax[row];
171: nrow1 = ailen[row];
172: low1 = 0;
173: high1 = nrow1;
174: lastcol2 = -1;
175: rp2 = bj + bi[row];
176: ap2 = ba + bi[row];
177: rmax2 = bimax[row];
178: nrow2 = bilen[row];
179: low2 = 0;
180: high2 = nrow2;
182: for (j=0; j<n; j++) {
183: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
184: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
185: if (in[j] >= cstart && in[j] < cend){
186: col = in[j] - cstart;
187: MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
188: } else if (in[j] < 0) continue;
189: #if defined(PETSC_USE_DEBUG)
190: else if (in[j] >= mat->cmap.N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap.N-1);}
191: #endif
192: else {
193: if (mat->was_assembled) {
194: if (!aij->colmap) {
195: CreateColmap_MPIAIJ_Private(mat);
196: }
197: #if defined (PETSC_USE_CTABLE)
198: PetscTableFind(aij->colmap,in[j]+1,&col);
199: col--;
200: #else
201: col = aij->colmap[in[j]] - 1;
202: #endif
203: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
204: DisAssemble_MPIAIJ(mat);
205: col = in[j];
206: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
207: B = aij->B;
208: b = (Mat_SeqAIJ*)B->data;
209: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
210: rp2 = bj + bi[row];
211: ap2 = ba + bi[row];
212: rmax2 = bimax[row];
213: nrow2 = bilen[row];
214: low2 = 0;
215: high2 = nrow2;
216: bm = aij->B->rmap.n;
217: ba = b->a;
218: }
219: } else col = in[j];
220: MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
221: }
222: }
223: } else {
224: if (!aij->donotstash) {
225: if (roworiented) {
226: if (ignorezeroentries && v[i*n] == 0.0) continue;
227: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
228: } else {
229: if (ignorezeroentries && v[i] == 0.0) continue;
230: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
231: }
232: }
233: }
234: }
235: return(0);
236: }
241: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
242: {
243: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
245: PetscInt i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
246: PetscInt cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;
249: for (i=0; i<m; i++) {
250: if (idxm[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);
251: if (idxm[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap.N-1);
252: if (idxm[i] >= rstart && idxm[i] < rend) {
253: row = idxm[i] - rstart;
254: for (j=0; j<n; j++) {
255: if (idxn[j] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]);
256: if (idxn[j] >= mat->cmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap.N-1);
257: if (idxn[j] >= cstart && idxn[j] < cend){
258: col = idxn[j] - cstart;
259: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
260: } else {
261: if (!aij->colmap) {
262: CreateColmap_MPIAIJ_Private(mat);
263: }
264: #if defined (PETSC_USE_CTABLE)
265: PetscTableFind(aij->colmap,idxn[j]+1,&col);
266: col --;
267: #else
268: col = aij->colmap[idxn[j]] - 1;
269: #endif
270: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
271: else {
272: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
273: }
274: }
275: }
276: } else {
277: SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
278: }
279: }
280: return(0);
281: }
285: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
286: {
287: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
289: PetscInt nstash,reallocs;
290: InsertMode addv;
293: if (aij->donotstash) {
294: return(0);
295: }
297: /* make sure all processors are either in INSERTMODE or ADDMODE */
298: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);
299: if (addv == (ADD_VALUES|INSERT_VALUES)) {
300: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
301: }
302: mat->insertmode = addv; /* in case this processor had no cache */
304: MatStashScatterBegin_Private(&mat->stash,mat->rmap.range);
305: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
306: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
307: return(0);
308: }
312: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
313: {
314: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
315: Mat_SeqAIJ *a=(Mat_SeqAIJ *)aij->A->data;
317: PetscMPIInt n;
318: PetscInt i,j,rstart,ncols,flg;
319: PetscInt *row,*col,other_disassembled;
320: PetscScalar *val;
321: InsertMode addv = mat->insertmode;
323: /* do not use 'b = (Mat_SeqAIJ *)aij->B->data' as B can be reset in disassembly */
325: if (!aij->donotstash) {
326: while (1) {
327: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
328: if (!flg) break;
330: for (i=0; i<n;) {
331: /* Now identify the consecutive vals belonging to the same row */
332: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
333: if (j < n) ncols = j-i;
334: else ncols = n-i;
335: /* Now assemble all these values with a single function call */
336: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
337: i = j;
338: }
339: }
340: MatStashScatterEnd_Private(&mat->stash);
341: }
342: a->compressedrow.use = PETSC_FALSE;
343: MatAssemblyBegin(aij->A,mode);
344: MatAssemblyEnd(aij->A,mode);
346: /* determine if any processor has disassembled, if so we must
347: also disassemble ourselfs, in order that we may reassemble. */
348: /*
349: if nonzero structure of submatrix B cannot change then we know that
350: no processor disassembled thus we can skip this stuff
351: */
352: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
353: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);
354: if (mat->was_assembled && !other_disassembled) {
355: DisAssemble_MPIAIJ(mat);
356: }
357: }
358: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
359: MatSetUpMultiply_MPIAIJ(mat);
360: }
361: MatSetOption(aij->B,MAT_DO_NOT_USE_INODES);
362: ((Mat_SeqAIJ *)aij->B->data)->compressedrow.use = PETSC_TRUE; /* b->compressedrow.use */
363: MatAssemblyBegin(aij->B,mode);
364: MatAssemblyEnd(aij->B,mode);
366: PetscFree(aij->rowvalues);
367: aij->rowvalues = 0;
369: /* used by MatAXPY() */
370: a->xtoy = 0; ((Mat_SeqAIJ *)aij->B->data)->xtoy = 0; /* b->xtoy = 0 */
371: a->XtoY = 0; ((Mat_SeqAIJ *)aij->B->data)->XtoY = 0; /* b->XtoY = 0 */
373: return(0);
374: }
378: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
379: {
380: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
384: MatZeroEntries(l->A);
385: MatZeroEntries(l->B);
386: return(0);
387: }
391: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag)
392: {
393: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
395: PetscMPIInt size = l->size,imdex,n,rank = l->rank,tag = A->tag,lastidx = -1;
396: PetscInt i,*owners = A->rmap.range;
397: PetscInt *nprocs,j,idx,nsends,row;
398: PetscInt nmax,*svalues,*starts,*owner,nrecvs;
399: PetscInt *rvalues,count,base,slen,*source;
400: PetscInt *lens,*lrows,*values,rstart=A->rmap.rstart;
401: MPI_Comm comm = A->comm;
402: MPI_Request *send_waits,*recv_waits;
403: MPI_Status recv_status,*send_status;
404: #if defined(PETSC_DEBUG)
405: PetscTruth found = PETSC_FALSE;
406: #endif
409: /* first count number of contributors to each processor */
410: PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
411: PetscMemzero(nprocs,2*size*sizeof(PetscInt));
412: PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
413: j = 0;
414: for (i=0; i<N; i++) {
415: if (lastidx > (idx = rows[i])) j = 0;
416: lastidx = idx;
417: for (; j<size; j++) {
418: if (idx >= owners[j] && idx < owners[j+1]) {
419: nprocs[2*j]++;
420: nprocs[2*j+1] = 1;
421: owner[i] = j;
422: #if defined(PETSC_DEBUG)
423: found = PETSC_TRUE;
424: #endif
425: break;
426: }
427: }
428: #if defined(PETSC_DEBUG)
429: if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
430: found = PETSC_FALSE;
431: #endif
432: }
433: nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}
435: /* inform other processors of number of messages and max length*/
436: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
438: /* post receives: */
439: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
440: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
441: for (i=0; i<nrecvs; i++) {
442: MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
443: }
445: /* do sends:
446: 1) starts[i] gives the starting index in svalues for stuff going to
447: the ith processor
448: */
449: PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
450: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
451: PetscMalloc((size+1)*sizeof(PetscInt),&starts);
452: starts[0] = 0;
453: for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
454: for (i=0; i<N; i++) {
455: svalues[starts[owner[i]]++] = rows[i];
456: }
458: starts[0] = 0;
459: for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
460: count = 0;
461: for (i=0; i<size; i++) {
462: if (nprocs[2*i+1]) {
463: MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
464: }
465: }
466: PetscFree(starts);
468: base = owners[rank];
470: /* wait on receives */
471: PetscMalloc(2*(nrecvs+1)*sizeof(PetscInt),&lens);
472: source = lens + nrecvs;
473: count = nrecvs; slen = 0;
474: while (count) {
475: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
476: /* unpack receives into our local space */
477: MPI_Get_count(&recv_status,MPIU_INT,&n);
478: source[imdex] = recv_status.MPI_SOURCE;
479: lens[imdex] = n;
480: slen += n;
481: count--;
482: }
483: PetscFree(recv_waits);
484:
485: /* move the data into the send scatter */
486: PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
487: count = 0;
488: for (i=0; i<nrecvs; i++) {
489: values = rvalues + i*nmax;
490: for (j=0; j<lens[i]; j++) {
491: lrows[count++] = values[j] - base;
492: }
493: }
494: PetscFree(rvalues);
495: PetscFree(lens);
496: PetscFree(owner);
497: PetscFree(nprocs);
498:
499: /* actually zap the local rows */
500: /*
501: Zero the required rows. If the "diagonal block" of the matrix
502: is square and the user wishes to set the diagonal we use separate
503: code so that MatSetValues() is not called for each diagonal allocating
504: new memory, thus calling lots of mallocs and slowing things down.
506: Contributed by: Matthew Knepley
507: */
508: /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
509: MatZeroRows(l->B,slen,lrows,0.0);
510: if ((diag != 0.0) && (l->A->rmap.N == l->A->cmap.N)) {
511: MatZeroRows(l->A,slen,lrows,diag);
512: } else if (diag != 0.0) {
513: MatZeroRows(l->A,slen,lrows,0.0);
514: if (((Mat_SeqAIJ*)l->A->data)->nonew) {
515: SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options\n\
516: MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
517: }
518: for (i = 0; i < slen; i++) {
519: row = lrows[i] + rstart;
520: MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
521: }
522: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
523: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
524: } else {
525: MatZeroRows(l->A,slen,lrows,0.0);
526: }
527: PetscFree(lrows);
529: /* wait on sends */
530: if (nsends) {
531: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
532: MPI_Waitall(nsends,send_waits,send_status);
533: PetscFree(send_status);
534: }
535: PetscFree(send_waits);
536: PetscFree(svalues);
538: return(0);
539: }
543: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
544: {
545: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
547: PetscInt nt;
550: VecGetLocalSize(xx,&nt);
551: if (nt != A->cmap.n) {
552: SETERRQ2(PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap.n,nt);
553: }
554: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
555: (*a->A->ops->mult)(a->A,xx,yy);
556: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
557: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
558: return(0);
559: }
563: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
564: {
565: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
569: VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
570: (*a->A->ops->multadd)(a->A,xx,yy,zz);
571: VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
572: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
573: return(0);
574: }
578: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
579: {
580: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
582: PetscTruth merged;
585: VecScatterGetMerged(a->Mvctx,&merged);
586: /* do nondiagonal part */
587: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
588: if (!merged) {
589: /* send it on its way */
590: VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
591: /* do local part */
592: (*a->A->ops->multtranspose)(a->A,xx,yy);
593: /* receive remote parts: note this assumes the values are not actually */
594: /* added in yy until the next line, */
595: VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
596: } else {
597: /* do local part */
598: (*a->A->ops->multtranspose)(a->A,xx,yy);
599: /* send it on its way */
600: VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
601: /* values actually were received in the Begin() but we need to call this nop */
602: VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
603: }
604: return(0);
605: }
610: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscTruth *f)
611: {
612: MPI_Comm comm;
613: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *) Amat->data, *Bij;
614: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
615: IS Me,Notme;
617: PetscInt M,N,first,last,*notme,i;
618: PetscMPIInt size;
622: /* Easy test: symmetric diagonal block */
623: Bij = (Mat_MPIAIJ *) Bmat->data; Bdia = Bij->A;
624: MatIsTranspose(Adia,Bdia,tol,f);
625: if (!*f) return(0);
626: PetscObjectGetComm((PetscObject)Amat,&comm);
627: MPI_Comm_size(comm,&size);
628: if (size == 1) return(0);
630: /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
631: MatGetSize(Amat,&M,&N);
632: MatGetOwnershipRange(Amat,&first,&last);
633: PetscMalloc((N-last+first)*sizeof(PetscInt),¬me);
634: for (i=0; i<first; i++) notme[i] = i;
635: for (i=last; i<M; i++) notme[i-last+first] = i;
636: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,&Notme);
637: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
638: MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
639: Aoff = Aoffs[0];
640: MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
641: Boff = Boffs[0];
642: MatIsTranspose(Aoff,Boff,tol,f);
643: MatDestroyMatrices(1,&Aoffs);
644: MatDestroyMatrices(1,&Boffs);
645: ISDestroy(Me);
646: ISDestroy(Notme);
648: return(0);
649: }
654: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
655: {
656: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
660: /* do nondiagonal part */
661: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
662: /* send it on its way */
663: VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
664: /* do local part */
665: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
666: /* receive remote parts */
667: VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
668: return(0);
669: }
671: /*
672: This only works correctly for square matrices where the subblock A->A is the
673: diagonal block
674: */
677: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
678: {
680: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
683: if (A->rmap.N != A->cmap.N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
684: if (A->rmap.rstart != A->cmap.rstart || A->rmap.rend != A->cmap.rend) {
685: SETERRQ(PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
686: }
687: MatGetDiagonal(a->A,v);
688: return(0);
689: }
693: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
694: {
695: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
699: MatScale(a->A,aa);
700: MatScale(a->B,aa);
701: return(0);
702: }
706: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
707: {
708: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
712: #if defined(PETSC_USE_LOG)
713: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap.N,mat->cmap.N);
714: #endif
715: MatStashDestroy_Private(&mat->stash);
716: MatDestroy(aij->A);
717: MatDestroy(aij->B);
718: #if defined (PETSC_USE_CTABLE)
719: if (aij->colmap) {PetscTableDelete(aij->colmap);}
720: #else
721: PetscFree(aij->colmap);
722: #endif
723: PetscFree(aij->garray);
724: if (aij->lvec) {VecDestroy(aij->lvec);}
725: if (aij->Mvctx) {VecScatterDestroy(aij->Mvctx);}
726: PetscFree(aij->rowvalues);
727: PetscFree(aij);
729: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
730: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
731: PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
732: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C","",PETSC_NULL);
733: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C","",PETSC_NULL);
734: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C","",PETSC_NULL);
735: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);
736: return(0);
737: }
741: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
742: {
743: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
744: Mat_SeqAIJ* A = (Mat_SeqAIJ*)aij->A->data;
745: Mat_SeqAIJ* B = (Mat_SeqAIJ*)aij->B->data;
746: PetscErrorCode ierr;
747: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
748: int fd;
749: PetscInt nz,header[4],*row_lengths,*range=0,rlen,i;
750: PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap.rstart,rnz;
751: PetscScalar *column_values;
754: MPI_Comm_rank(mat->comm,&rank);
755: MPI_Comm_size(mat->comm,&size);
756: nz = A->nz + B->nz;
757: if (!rank) {
758: header[0] = MAT_FILE_COOKIE;
759: header[1] = mat->rmap.N;
760: header[2] = mat->cmap.N;
761: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,mat->comm);
762: PetscViewerBinaryGetDescriptor(viewer,&fd);
763: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
764: /* get largest number of rows any processor has */
765: rlen = mat->rmap.n;
766: range = mat->rmap.range;
767: for (i=1; i<size; i++) {
768: rlen = PetscMax(rlen,range[i+1] - range[i]);
769: }
770: } else {
771: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,mat->comm);
772: rlen = mat->rmap.n;
773: }
775: /* load up the local row counts */
776: PetscMalloc((rlen+1)*sizeof(PetscInt),&row_lengths);
777: for (i=0; i<mat->rmap.n; i++) {
778: row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
779: }
781: /* store the row lengths to the file */
782: if (!rank) {
783: MPI_Status status;
784: PetscBinaryWrite(fd,row_lengths,mat->rmap.n,PETSC_INT,PETSC_TRUE);
785: for (i=1; i<size; i++) {
786: rlen = range[i+1] - range[i];
787: MPI_Recv(row_lengths,rlen,MPIU_INT,i,tag,mat->comm,&status);
788: PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
789: }
790: } else {
791: MPI_Send(row_lengths,mat->rmap.n,MPIU_INT,0,tag,mat->comm);
792: }
793: PetscFree(row_lengths);
795: /* load up the local column indices */
796: nzmax = nz; /* )th processor needs space a largest processor needs */
797: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,mat->comm);
798: PetscMalloc((nzmax+1)*sizeof(PetscInt),&column_indices);
799: cnt = 0;
800: for (i=0; i<mat->rmap.n; i++) {
801: for (j=B->i[i]; j<B->i[i+1]; j++) {
802: if ( (col = garray[B->j[j]]) > cstart) break;
803: column_indices[cnt++] = col;
804: }
805: for (k=A->i[i]; k<A->i[i+1]; k++) {
806: column_indices[cnt++] = A->j[k] + cstart;
807: }
808: for (; j<B->i[i+1]; j++) {
809: column_indices[cnt++] = garray[B->j[j]];
810: }
811: }
812: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
814: /* store the column indices to the file */
815: if (!rank) {
816: MPI_Status status;
817: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
818: for (i=1; i<size; i++) {
819: MPI_Recv(&rnz,1,MPIU_INT,i,tag,mat->comm,&status);
820: if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
821: MPI_Recv(column_indices,rnz,MPIU_INT,i,tag,mat->comm,&status);
822: PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
823: }
824: } else {
825: MPI_Send(&nz,1,MPIU_INT,0,tag,mat->comm);
826: MPI_Send(column_indices,nz,MPIU_INT,0,tag,mat->comm);
827: }
828: PetscFree(column_indices);
830: /* load up the local column values */
831: PetscMalloc((nzmax+1)*sizeof(PetscScalar),&column_values);
832: cnt = 0;
833: for (i=0; i<mat->rmap.n; i++) {
834: for (j=B->i[i]; j<B->i[i+1]; j++) {
835: if ( garray[B->j[j]] > cstart) break;
836: column_values[cnt++] = B->a[j];
837: }
838: for (k=A->i[i]; k<A->i[i+1]; k++) {
839: column_values[cnt++] = A->a[k];
840: }
841: for (; j<B->i[i+1]; j++) {
842: column_values[cnt++] = B->a[j];
843: }
844: }
845: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
847: /* store the column values to the file */
848: if (!rank) {
849: MPI_Status status;
850: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
851: for (i=1; i<size; i++) {
852: MPI_Recv(&rnz,1,MPIU_INT,i,tag,mat->comm,&status);
853: if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
854: MPI_Recv(column_values,rnz,MPIU_SCALAR,i,tag,mat->comm,&status);
855: PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
856: }
857: } else {
858: MPI_Send(&nz,1,MPIU_INT,0,tag,mat->comm);
859: MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,mat->comm);
860: }
861: PetscFree(column_values);
862: return(0);
863: }
867: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
868: {
869: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
870: PetscErrorCode ierr;
871: PetscMPIInt rank = aij->rank,size = aij->size;
872: PetscTruth isdraw,iascii,isbinary;
873: PetscViewer sviewer;
874: PetscViewerFormat format;
877: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
878: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
879: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
880: if (iascii) {
881: PetscViewerGetFormat(viewer,&format);
882: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
883: MatInfo info;
884: PetscTruth inodes;
886: MPI_Comm_rank(mat->comm,&rank);
887: MatGetInfo(mat,MAT_LOCAL,&info);
888: MatInodeGetInodeSizes(aij->A,PETSC_NULL,(PetscInt **)&inodes,PETSC_NULL);
889: if (!inodes) {
890: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
891: rank,mat->rmap.n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
892: } else {
893: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
894: rank,mat->rmap.n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
895: }
896: MatGetInfo(aij->A,MAT_LOCAL,&info);
897: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
898: MatGetInfo(aij->B,MAT_LOCAL,&info);
899: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
900: PetscViewerFlush(viewer);
901: VecScatterView(aij->Mvctx,viewer);
902: return(0);
903: } else if (format == PETSC_VIEWER_ASCII_INFO) {
904: PetscInt inodecount,inodelimit,*inodes;
905: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
906: if (inodes) {
907: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
908: } else {
909: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
910: }
911: return(0);
912: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
913: return(0);
914: }
915: } else if (isbinary) {
916: if (size == 1) {
917: PetscObjectSetName((PetscObject)aij->A,mat->name);
918: MatView(aij->A,viewer);
919: } else {
920: MatView_MPIAIJ_Binary(mat,viewer);
921: }
922: return(0);
923: } else if (isdraw) {
924: PetscDraw draw;
925: PetscTruth isnull;
926: PetscViewerDrawGetDraw(viewer,0,&draw);
927: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
928: }
930: if (size == 1) {
931: PetscObjectSetName((PetscObject)aij->A,mat->name);
932: MatView(aij->A,viewer);
933: } else {
934: /* assemble the entire matrix onto first processor. */
935: Mat A;
936: Mat_SeqAIJ *Aloc;
937: PetscInt M = mat->rmap.N,N = mat->cmap.N,m,*ai,*aj,row,*cols,i,*ct;
938: PetscScalar *a;
940: MatCreate(mat->comm,&A);
941: if (!rank) {
942: MatSetSizes(A,M,N,M,N);
943: } else {
944: MatSetSizes(A,0,0,M,N);
945: }
946: /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
947: MatSetType(A,MATMPIAIJ);
948: MatMPIAIJSetPreallocation(A,0,PETSC_NULL,0,PETSC_NULL);
949: PetscLogObjectParent(mat,A);
951: /* copy over the A part */
952: Aloc = (Mat_SeqAIJ*)aij->A->data;
953: m = aij->A->rmap.n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
954: row = mat->rmap.rstart;
955: for (i=0; i<ai[m]; i++) {aj[i] += mat->cmap.rstart ;}
956: for (i=0; i<m; i++) {
957: MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
958: row++; a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
959: }
960: aj = Aloc->j;
961: for (i=0; i<ai[m]; i++) {aj[i] -= mat->cmap.rstart;}
963: /* copy over the B part */
964: Aloc = (Mat_SeqAIJ*)aij->B->data;
965: m = aij->B->rmap.n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
966: row = mat->rmap.rstart;
967: PetscMalloc((ai[m]+1)*sizeof(PetscInt),&cols);
968: ct = cols;
969: for (i=0; i<ai[m]; i++) {cols[i] = aij->garray[aj[i]];}
970: for (i=0; i<m; i++) {
971: MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
972: row++; a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
973: }
974: PetscFree(ct);
975: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
976: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
977: /*
978: Everyone has to call to draw the matrix since the graphics waits are
979: synchronized across all processors that share the PetscDraw object
980: */
981: PetscViewerGetSingleton(viewer,&sviewer);
982: if (!rank) {
983: PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,mat->name);
984: MatView(((Mat_MPIAIJ*)(A->data))->A,sviewer);
985: }
986: PetscViewerRestoreSingleton(viewer,&sviewer);
987: MatDestroy(A);
988: }
989: return(0);
990: }
994: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
995: {
997: PetscTruth iascii,isdraw,issocket,isbinary;
998:
1000: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
1001: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
1002: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
1003: PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
1004: if (iascii || isdraw || isbinary || issocket) {
1005: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1006: } else {
1007: SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPIAIJ matrices",((PetscObject)viewer)->type_name);
1008: }
1009: return(0);
1010: }
1016: PetscErrorCode MatRelax_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1017: {
1018: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1020: Vec bb1;
1023: if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
1025: VecDuplicate(bb,&bb1);
1027: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
1028: if (flag & SOR_ZERO_INITIAL_GUESS) {
1029: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
1030: its--;
1031: }
1032:
1033: while (its--) {
1034: VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1035: VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1037: /* update rhs: bb1 = bb - B*x */
1038: VecScale(mat->lvec,-1.0);
1039: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1041: /* local sweep */
1042: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
1043:
1044: }
1045: } else if (flag & SOR_LOCAL_FORWARD_SWEEP){
1046: if (flag & SOR_ZERO_INITIAL_GUESS) {
1047: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1048: its--;
1049: }
1050: while (its--) {
1051: VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1052: VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1054: /* update rhs: bb1 = bb - B*x */
1055: VecScale(mat->lvec,-1.0);
1056: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1058: /* local sweep */
1059: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1060:
1061: }
1062: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){
1063: if (flag & SOR_ZERO_INITIAL_GUESS) {
1064: (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1065: its--;
1066: }
1067: while (its--) {
1068: VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1069: VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1071: /* update rhs: bb1 = bb - B*x */
1072: VecScale(mat->lvec,-1.0);
1073: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1075: /* local sweep */
1076: (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1077:
1078: }
1079: } else {
1080: SETERRQ(PETSC_ERR_SUP,"Parallel SOR not supported");
1081: }
1083: VecDestroy(bb1);
1084: return(0);
1085: }
1089: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1090: {
1091: MPI_Comm comm,pcomm;
1092: PetscInt first,local_size,nrows,*rows;
1093: int ntids;
1094: IS crowp,growp,irowp,lrowp,lcolp,icolp;
1098: PetscObjectGetComm((PetscObject)A,&comm);
1099: /* make a collective version of 'rowp' */
1100: PetscObjectGetComm((PetscObject)rowp,&pcomm);
1101: if (pcomm==comm) {
1102: crowp = rowp;
1103: } else {
1104: ISGetSize(rowp,&nrows);
1105: ISGetIndices(rowp,&rows);
1106: ISCreateGeneral(comm,nrows,rows,&crowp);
1107: ISRestoreIndices(rowp,&rows);
1108: }
1109: /* collect the global row permutation and invert it */
1110: ISAllGather(crowp,&growp);
1111: ISSetPermutation(growp);
1112: if (pcomm!=comm) {
1113: ISDestroy(crowp);
1114: }
1115: ISInvertPermutation(growp,PETSC_DECIDE,&irowp);
1116: /* get the local target indices */
1117: MatGetOwnershipRange(A,&first,PETSC_NULL);
1118: MatGetLocalSize(A,&local_size,PETSC_NULL);
1119: ISGetIndices(irowp,&rows);
1120: ISCreateGeneral(MPI_COMM_SELF,local_size,rows+first,&lrowp);
1121: ISRestoreIndices(irowp,&rows);
1122: ISDestroy(irowp);
1123: /* the column permutation is so much easier;
1124: make a local version of 'colp' and invert it */
1125: PetscObjectGetComm((PetscObject)colp,&pcomm);
1126: MPI_Comm_size(pcomm,&ntids);
1127: if (ntids==1) {
1128: lcolp = colp;
1129: } else {
1130: ISGetSize(colp,&nrows);
1131: ISGetIndices(colp,&rows);
1132: ISCreateGeneral(MPI_COMM_SELF,nrows,rows,&lcolp);
1133: }
1134: ISInvertPermutation(lcolp,PETSC_DECIDE,&icolp);
1135: ISSetPermutation(lcolp);
1136: if (ntids>1) {
1137: ISRestoreIndices(colp,&rows);
1138: ISDestroy(lcolp);
1139: }
1140: /* now we just get the submatrix */
1141: MatGetSubMatrix(A,lrowp,icolp,local_size,MAT_INITIAL_MATRIX,B);
1142: /* clean up */
1143: ISDestroy(lrowp);
1144: ISDestroy(icolp);
1145: return(0);
1146: }
1150: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1151: {
1152: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1153: Mat A = mat->A,B = mat->B;
1155: PetscReal isend[5],irecv[5];
1158: info->block_size = 1.0;
1159: MatGetInfo(A,MAT_LOCAL,info);
1160: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1161: isend[3] = info->memory; isend[4] = info->mallocs;
1162: MatGetInfo(B,MAT_LOCAL,info);
1163: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1164: isend[3] += info->memory; isend[4] += info->mallocs;
1165: if (flag == MAT_LOCAL) {
1166: info->nz_used = isend[0];
1167: info->nz_allocated = isend[1];
1168: info->nz_unneeded = isend[2];
1169: info->memory = isend[3];
1170: info->mallocs = isend[4];
1171: } else if (flag == MAT_GLOBAL_MAX) {
1172: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);
1173: info->nz_used = irecv[0];
1174: info->nz_allocated = irecv[1];
1175: info->nz_unneeded = irecv[2];
1176: info->memory = irecv[3];
1177: info->mallocs = irecv[4];
1178: } else if (flag == MAT_GLOBAL_SUM) {
1179: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);
1180: info->nz_used = irecv[0];
1181: info->nz_allocated = irecv[1];
1182: info->nz_unneeded = irecv[2];
1183: info->memory = irecv[3];
1184: info->mallocs = irecv[4];
1185: }
1186: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1187: info->fill_ratio_needed = 0;
1188: info->factor_mallocs = 0;
1189: info->rows_global = (double)matin->rmap.N;
1190: info->columns_global = (double)matin->cmap.N;
1191: info->rows_local = (double)matin->rmap.n;
1192: info->columns_local = (double)matin->cmap.N;
1194: return(0);
1195: }
1199: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op)
1200: {
1201: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1205: switch (op) {
1206: case MAT_NO_NEW_NONZERO_LOCATIONS:
1207: case MAT_YES_NEW_NONZERO_LOCATIONS:
1208: case MAT_COLUMNS_UNSORTED:
1209: case MAT_COLUMNS_SORTED:
1210: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1211: case MAT_KEEP_ZEROED_ROWS:
1212: case MAT_NEW_NONZERO_LOCATION_ERR:
1213: case MAT_USE_INODES:
1214: case MAT_DO_NOT_USE_INODES:
1215: case MAT_IGNORE_ZERO_ENTRIES:
1216: MatSetOption(a->A,op);
1217: MatSetOption(a->B,op);
1218: break;
1219: case MAT_ROW_ORIENTED:
1220: a->roworiented = PETSC_TRUE;
1221: MatSetOption(a->A,op);
1222: MatSetOption(a->B,op);
1223: break;
1224: case MAT_ROWS_SORTED:
1225: case MAT_ROWS_UNSORTED:
1226: case MAT_YES_NEW_DIAGONALS:
1227: PetscInfo(A,"Option ignored\n");
1228: break;
1229: case MAT_COLUMN_ORIENTED:
1230: a->roworiented = PETSC_FALSE;
1231: MatSetOption(a->A,op);
1232: MatSetOption(a->B,op);
1233: break;
1234: case MAT_IGNORE_OFF_PROC_ENTRIES:
1235: a->donotstash = PETSC_TRUE;
1236: break;
1237: case MAT_NO_NEW_DIAGONALS:
1238: SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1239: case MAT_SYMMETRIC:
1240: MatSetOption(a->A,op);
1241: break;
1242: case MAT_STRUCTURALLY_SYMMETRIC:
1243: case MAT_HERMITIAN:
1244: case MAT_SYMMETRY_ETERNAL:
1245: MatSetOption(a->A,op);
1246: break;
1247: case MAT_NOT_SYMMETRIC:
1248: case MAT_NOT_STRUCTURALLY_SYMMETRIC:
1249: case MAT_NOT_HERMITIAN:
1250: case MAT_NOT_SYMMETRY_ETERNAL:
1251: break;
1252: default:
1253: SETERRQ1(PETSC_ERR_SUP,"unknown option %d",op);
1254: }
1255: return(0);
1256: }
1260: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1261: {
1262: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1263: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1265: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap.rstart;
1266: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap.rstart,rend = matin->rmap.rend;
1267: PetscInt *cmap,*idx_p;
1270: if (mat->getrowactive) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1271: mat->getrowactive = PETSC_TRUE;
1273: if (!mat->rowvalues && (idx || v)) {
1274: /*
1275: allocate enough space to hold information from the longest row.
1276: */
1277: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1278: PetscInt max = 1,tmp;
1279: for (i=0; i<matin->rmap.n; i++) {
1280: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1281: if (max < tmp) { max = tmp; }
1282: }
1283: PetscMalloc(max*(sizeof(PetscInt)+sizeof(PetscScalar)),&mat->rowvalues);
1284: mat->rowindices = (PetscInt*)(mat->rowvalues + max);
1285: }
1287: if (row < rstart || row >= rend) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Only local rows")
1288: lrow = row - rstart;
1290: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1291: if (!v) {pvA = 0; pvB = 0;}
1292: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1293: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1294: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1295: nztot = nzA + nzB;
1297: cmap = mat->garray;
1298: if (v || idx) {
1299: if (nztot) {
1300: /* Sort by increasing column numbers, assuming A and B already sorted */
1301: PetscInt imark = -1;
1302: if (v) {
1303: *v = v_p = mat->rowvalues;
1304: for (i=0; i<nzB; i++) {
1305: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1306: else break;
1307: }
1308: imark = i;
1309: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1310: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1311: }
1312: if (idx) {
1313: *idx = idx_p = mat->rowindices;
1314: if (imark > -1) {
1315: for (i=0; i<imark; i++) {
1316: idx_p[i] = cmap[cworkB[i]];
1317: }
1318: } else {
1319: for (i=0; i<nzB; i++) {
1320: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1321: else break;
1322: }
1323: imark = i;
1324: }
1325: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1326: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1327: }
1328: } else {
1329: if (idx) *idx = 0;
1330: if (v) *v = 0;
1331: }
1332: }
1333: *nz = nztot;
1334: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1335: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1336: return(0);
1337: }
1341: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1342: {
1343: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1346: if (!aij->getrowactive) {
1347: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1348: }
1349: aij->getrowactive = PETSC_FALSE;
1350: return(0);
1351: }
1355: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1356: {
1357: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1358: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1360: PetscInt i,j,cstart = mat->cmap.rstart;
1361: PetscReal sum = 0.0;
1362: PetscScalar *v;
1365: if (aij->size == 1) {
1366: MatNorm(aij->A,type,norm);
1367: } else {
1368: if (type == NORM_FROBENIUS) {
1369: v = amat->a;
1370: for (i=0; i<amat->nz; i++) {
1371: #if defined(PETSC_USE_COMPLEX)
1372: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1373: #else
1374: sum += (*v)*(*v); v++;
1375: #endif
1376: }
1377: v = bmat->a;
1378: for (i=0; i<bmat->nz; i++) {
1379: #if defined(PETSC_USE_COMPLEX)
1380: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1381: #else
1382: sum += (*v)*(*v); v++;
1383: #endif
1384: }
1385: MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPI_SUM,mat->comm);
1386: *norm = sqrt(*norm);
1387: } else if (type == NORM_1) { /* max column norm */
1388: PetscReal *tmp,*tmp2;
1389: PetscInt *jj,*garray = aij->garray;
1390: PetscMalloc((mat->cmap.N+1)*sizeof(PetscReal),&tmp);
1391: PetscMalloc((mat->cmap.N+1)*sizeof(PetscReal),&tmp2);
1392: PetscMemzero(tmp,mat->cmap.N*sizeof(PetscReal));
1393: *norm = 0.0;
1394: v = amat->a; jj = amat->j;
1395: for (j=0; j<amat->nz; j++) {
1396: tmp[cstart + *jj++ ] += PetscAbsScalar(*v); v++;
1397: }
1398: v = bmat->a; jj = bmat->j;
1399: for (j=0; j<bmat->nz; j++) {
1400: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1401: }
1402: MPI_Allreduce(tmp,tmp2,mat->cmap.N,MPIU_REAL,MPI_SUM,mat->comm);
1403: for (j=0; j<mat->cmap.N; j++) {
1404: if (tmp2[j] > *norm) *norm = tmp2[j];
1405: }
1406: PetscFree(tmp);
1407: PetscFree(tmp2);
1408: } else if (type == NORM_INFINITY) { /* max row norm */
1409: PetscReal ntemp = 0.0;
1410: for (j=0; j<aij->A->rmap.n; j++) {
1411: v = amat->a + amat->i[j];
1412: sum = 0.0;
1413: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1414: sum += PetscAbsScalar(*v); v++;
1415: }
1416: v = bmat->a + bmat->i[j];
1417: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1418: sum += PetscAbsScalar(*v); v++;
1419: }
1420: if (sum > ntemp) ntemp = sum;
1421: }
1422: MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPI_MAX,mat->comm);
1423: } else {
1424: SETERRQ(PETSC_ERR_SUP,"No support for two norm");
1425: }
1426: }
1427: return(0);
1428: }
1432: PetscErrorCode MatTranspose_MPIAIJ(Mat A,Mat *matout)
1433: {
1434: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1435: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ*)a->A->data;
1437: PetscInt M = A->rmap.N,N = A->cmap.N,m,*ai,*aj,row,*cols,i,*ct;
1438: Mat B;
1439: PetscScalar *array;
1442: if (!matout && M != N) {
1443: SETERRQ(PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1444: }
1446: MatCreate(A->comm,&B);
1447: MatSetSizes(B,A->cmap.n,A->rmap.n,N,M);
1448: MatSetType(B,A->type_name);
1449: MatMPIAIJSetPreallocation(B,0,PETSC_NULL,0,PETSC_NULL);
1451: /* copy over the A part */
1452: Aloc = (Mat_SeqAIJ*)a->A->data;
1453: m = a->A->rmap.n; ai = Aloc->i; aj = Aloc->j; array = Aloc->a;
1454: row = A->rmap.rstart;
1455: for (i=0; i<ai[m]; i++) {aj[i] += A->cmap.rstart ;}
1456: for (i=0; i<m; i++) {
1457: MatSetValues(B,ai[i+1]-ai[i],aj,1,&row,array,INSERT_VALUES);
1458: row++; array += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1459: }
1460: aj = Aloc->j;
1461: for (i=0; i<ai[m]; i++) {aj[i] -= A->cmap.rstart ;}
1463: /* copy over the B part */
1464: Aloc = (Mat_SeqAIJ*)a->B->data;
1465: m = a->B->rmap.n; ai = Aloc->i; aj = Aloc->j; array = Aloc->a;
1466: row = A->rmap.rstart;
1467: PetscMalloc((1+ai[m])*sizeof(PetscInt),&cols);
1468: ct = cols;
1469: for (i=0; i<ai[m]; i++) {cols[i] = a->garray[aj[i]];}
1470: for (i=0; i<m; i++) {
1471: MatSetValues(B,ai[i+1]-ai[i],cols,1,&row,array,INSERT_VALUES);
1472: row++; array += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1473: }
1474: PetscFree(ct);
1475: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1476: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
1477: if (matout) {
1478: *matout = B;
1479: } else {
1480: MatHeaderCopy(A,B);
1481: }
1482: return(0);
1483: }
1487: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
1488: {
1489: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1490: Mat a = aij->A,b = aij->B;
1492: PetscInt s1,s2,s3;
1495: MatGetLocalSize(mat,&s2,&s3);
1496: if (rr) {
1497: VecGetLocalSize(rr,&s1);
1498: if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1499: /* Overlap communication with computation. */
1500: VecScatterBegin(rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD,aij->Mvctx);
1501: }
1502: if (ll) {
1503: VecGetLocalSize(ll,&s1);
1504: if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1505: (*b->ops->diagonalscale)(b,ll,0);
1506: }
1507: /* scale the diagonal block */
1508: (*a->ops->diagonalscale)(a,ll,rr);
1510: if (rr) {
1511: /* Do a scatter end and then right scale the off-diagonal block */
1512: VecScatterEnd(rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD,aij->Mvctx);
1513: (*b->ops->diagonalscale)(b,0,aij->lvec);
1514: }
1515:
1516: return(0);
1517: }
1521: PetscErrorCode MatSetBlockSize_MPIAIJ(Mat A,PetscInt bs)
1522: {
1523: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1527: MatSetBlockSize(a->A,bs);
1528: MatSetBlockSize(a->B,bs);
1529: return(0);
1530: }
1533: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1534: {
1535: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1539: MatSetUnfactored(a->A);
1540: return(0);
1541: }
1545: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscTruth *flag)
1546: {
1547: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
1548: Mat a,b,c,d;
1549: PetscTruth flg;
1553: a = matA->A; b = matA->B;
1554: c = matB->A; d = matB->B;
1556: MatEqual(a,c,&flg);
1557: if (flg) {
1558: MatEqual(b,d,&flg);
1559: }
1560: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);
1561: return(0);
1562: }
1566: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
1567: {
1569: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1570: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
1573: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1574: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1575: /* because of the column compression in the off-processor part of the matrix a->B,
1576: the number of columns in a->B and b->B may be different, hence we cannot call
1577: the MatCopy() directly on the two parts. If need be, we can provide a more
1578: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
1579: then copying the submatrices */
1580: MatCopy_Basic(A,B,str);
1581: } else {
1582: MatCopy(a->A,b->A,str);
1583: MatCopy(a->B,b->B,str);
1584: }
1585: return(0);
1586: }
1590: PetscErrorCode MatSetUpPreallocation_MPIAIJ(Mat A)
1591: {
1595: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1596: return(0);
1597: }
1599: #include petscblaslapack.h
1602: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1603: {
1605: PetscInt i;
1606: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data,*yy = (Mat_MPIAIJ *)Y->data;
1607: PetscBLASInt bnz,one=1;
1608: Mat_SeqAIJ *x,*y;
1611: if (str == SAME_NONZERO_PATTERN) {
1612: PetscScalar alpha = a;
1613: x = (Mat_SeqAIJ *)xx->A->data;
1614: y = (Mat_SeqAIJ *)yy->A->data;
1615: bnz = (PetscBLASInt)x->nz;
1616: BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1617: x = (Mat_SeqAIJ *)xx->B->data;
1618: y = (Mat_SeqAIJ *)yy->B->data;
1619: bnz = (PetscBLASInt)x->nz;
1620: BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1621: } else if (str == SUBSET_NONZERO_PATTERN) {
1622: MatAXPY_SeqAIJ(yy->A,a,xx->A,str);
1624: x = (Mat_SeqAIJ *)xx->B->data;
1625: y = (Mat_SeqAIJ *)yy->B->data;
1626: if (y->xtoy && y->XtoY != xx->B) {
1627: PetscFree(y->xtoy);
1628: MatDestroy(y->XtoY);
1629: }
1630: if (!y->xtoy) { /* get xtoy */
1631: MatAXPYGetxtoy_Private(xx->B->rmap.n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);
1632: y->XtoY = xx->B;
1633: }
1634: for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]);
1635: } else {
1636: MatAXPY_Basic(Y,a,X,str);
1637: }
1638: return(0);
1639: }
1641: EXTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
1645: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
1646: {
1647: #if defined(PETSC_USE_COMPLEX)
1649: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1652: MatConjugate_SeqAIJ(aij->A);
1653: MatConjugate_SeqAIJ(aij->B);
1654: #else
1656: #endif
1657: return(0);
1658: }
1662: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
1663: {
1664: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1668: MatRealPart(a->A);
1669: MatRealPart(a->B);
1670: return(0);
1671: }
1675: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
1676: {
1677: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1681: MatImaginaryPart(a->A);
1682: MatImaginaryPart(a->B);
1683: return(0);
1684: }
1686: #ifdef PETSC_HAVE_PBGL
1687: #include <boost/parallel/mpi/bsp_process_group.hpp>
1688: typedef boost::parallel::mpi::bsp_process_group process_group_type;
1690: #include <boost/graph/distributed/adjacency_list.hpp>
1691: #include <boost/parallel/mpi/bsp_process_group.hpp>
1693: #include <boost/graph/distributed/petsc/interface.hpp>
1694: #include <boost/graph/distributed/ilu_0.hpp>
1696: namespace petsc = boost::distributed::petsc;
1697: using namespace std;
1698: typedef double value_type;
1699: typedef boost::graph::distributed::ilu_elimination_state elimination_state;
1700: typedef boost::adjacency_list<boost::listS,
1701: boost::distributedS<process_group_type, boost::vecS>,
1702: boost::bidirectionalS,
1703: // Vertex properties
1704: boost::no_property,
1705: // Edge properties
1706: boost::property<boost::edge_weight_t, value_type,
1707: boost::property<boost::edge_finished_t, elimination_state> > > graph_type;
1709: typedef boost::graph_traits<graph_type>::vertex_descriptor vertex_type;
1710: typedef boost::graph_traits<graph_type>::edge_descriptor edge_type;
1711: typedef boost::property_map<graph_type, boost::edge_weight_t>::type weight_map_type;
1715: /*
1716: This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
1717: */
1718: PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat A, IS isrow, IS iscol, MatFactorInfo *info, Mat *fact)
1719: {
1720: PetscTruth row_identity, col_identity;
1721: PetscObjectContainer c;
1722: PetscInt m, n, M, N;
1723: PetscErrorCode ierr;
1726: if (info->levels != 0) SETERRQ(PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu");
1727: ISIdentity(isrow, &row_identity);
1728: ISIdentity(iscol, &col_identity);
1729: if (!row_identity || !col_identity) {
1730: SETERRQ(PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU");
1731: }
1733: process_group_type pg;
1734: graph_type* graph_p = new graph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg));
1735: graph_type& graph = *graph_p;
1736: petsc::read_matrix(A, graph, get(boost::edge_weight, graph));
1738: //write_graphviz("petsc_matrix_as_graph.dot", graph, default_writer(), matrix_graph_writer<graph_type>(graph));
1739: boost::property_map<graph_type, boost::edge_finished_t>::type finished = get(boost::edge_finished, graph);
1740: BGL_FORALL_EDGES(e, graph, graph_type)
1741: put(finished, e, boost::graph::distributed::unseen);
1743: ilu_0(graph, get(boost::edge_weight, graph), get(boost::edge_finished, graph));
1745: /* put together the new matrix */
1746: MatCreate(A->comm, fact);
1747: MatGetLocalSize(A, &m, &n);
1748: MatGetSize(A, &M, &N);
1749: MatSetSizes(*fact, m, n, M, N);
1750: MatSetType(*fact, A->type_name);
1751: MatAssemblyBegin(*fact, MAT_FINAL_ASSEMBLY);
1752: MatAssemblyEnd(*fact, MAT_FINAL_ASSEMBLY);
1753: (*fact)->factor = FACTOR_LU;
1755: PetscObjectContainerCreate(A->comm, &c);
1756: PetscObjectContainerSetPointer(c, graph_p);
1757: PetscObjectCompose((PetscObject) (*fact), "graph", (PetscObject) c);
1758: return(0);
1759: }
1763: PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat A, MatFactorInfo *info, Mat *B)
1764: {
1766: return(0);
1767: }
1771: /*
1772: This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
1773: */
1774: PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x)
1775: {
1776: graph_type* graph_p;
1777: PetscObjectContainer c;
1778: PetscErrorCode ierr;
1781: PetscObjectQuery((PetscObject) A, "graph", (PetscObject *) &c);
1782: PetscObjectContainerGetPointer(c, (void **) &graph_p);
1783: VecCopy(b, x);
1784: return(0);
1785: }
1786: #endif
1788: /* -------------------------------------------------------------------*/
1789: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
1790: MatGetRow_MPIAIJ,
1791: MatRestoreRow_MPIAIJ,
1792: MatMult_MPIAIJ,
1793: /* 4*/ MatMultAdd_MPIAIJ,
1794: MatMultTranspose_MPIAIJ,
1795: MatMultTransposeAdd_MPIAIJ,
1796: #ifdef PETSC_HAVE_PBGL
1797: MatSolve_MPIAIJ,
1798: #else
1799: 0,
1800: #endif
1801: 0,
1802: 0,
1803: /*10*/ 0,
1804: 0,
1805: 0,
1806: MatRelax_MPIAIJ,
1807: MatTranspose_MPIAIJ,
1808: /*15*/ MatGetInfo_MPIAIJ,
1809: MatEqual_MPIAIJ,
1810: MatGetDiagonal_MPIAIJ,
1811: MatDiagonalScale_MPIAIJ,
1812: MatNorm_MPIAIJ,
1813: /*20*/ MatAssemblyBegin_MPIAIJ,
1814: MatAssemblyEnd_MPIAIJ,
1815: 0,
1816: MatSetOption_MPIAIJ,
1817: MatZeroEntries_MPIAIJ,
1818: /*25*/ MatZeroRows_MPIAIJ,
1819: 0,
1820: #ifdef PETSC_HAVE_PBGL
1821: MatLUFactorNumeric_MPIAIJ,
1822: #else
1823: 0,
1824: #endif
1825: 0,
1826: 0,
1827: /*30*/ MatSetUpPreallocation_MPIAIJ,
1828: #ifdef PETSC_HAVE_PBGL
1829: MatILUFactorSymbolic_MPIAIJ,
1830: #else
1831: 0,
1832: #endif
1833: 0,
1834: 0,
1835: 0,
1836: /*35*/ MatDuplicate_MPIAIJ,
1837: 0,
1838: 0,
1839: 0,
1840: 0,
1841: /*40*/ MatAXPY_MPIAIJ,
1842: MatGetSubMatrices_MPIAIJ,
1843: MatIncreaseOverlap_MPIAIJ,
1844: MatGetValues_MPIAIJ,
1845: MatCopy_MPIAIJ,
1846: /*45*/ 0,
1847: MatScale_MPIAIJ,
1848: 0,
1849: 0,
1850: 0,
1851: /*50*/ MatSetBlockSize_MPIAIJ,
1852: 0,
1853: 0,
1854: 0,
1855: 0,
1856: /*55*/ MatFDColoringCreate_MPIAIJ,
1857: 0,
1858: MatSetUnfactored_MPIAIJ,
1859: MatPermute_MPIAIJ,
1860: 0,
1861: /*60*/ MatGetSubMatrix_MPIAIJ,
1862: MatDestroy_MPIAIJ,
1863: MatView_MPIAIJ,
1864: 0,
1865: 0,
1866: /*65*/ 0,
1867: 0,
1868: 0,
1869: 0,
1870: 0,
1871: /*70*/ 0,
1872: 0,
1873: MatSetColoring_MPIAIJ,
1874: #if defined(PETSC_HAVE_ADIC)
1875: MatSetValuesAdic_MPIAIJ,
1876: #else
1877: 0,
1878: #endif
1879: MatSetValuesAdifor_MPIAIJ,
1880: /*75*/ 0,
1881: 0,
1882: 0,
1883: 0,
1884: 0,
1885: /*80*/ 0,
1886: 0,
1887: 0,
1888: 0,
1889: /*84*/ MatLoad_MPIAIJ,
1890: 0,
1891: 0,
1892: 0,
1893: 0,
1894: 0,
1895: /*90*/ MatMatMult_MPIAIJ_MPIAIJ,
1896: MatMatMultSymbolic_MPIAIJ_MPIAIJ,
1897: MatMatMultNumeric_MPIAIJ_MPIAIJ,
1898: MatPtAP_Basic,
1899: MatPtAPSymbolic_MPIAIJ,
1900: /*95*/ MatPtAPNumeric_MPIAIJ,
1901: 0,
1902: 0,
1903: 0,
1904: 0,
1905: /*100*/0,
1906: MatPtAPSymbolic_MPIAIJ_MPIAIJ,
1907: MatPtAPNumeric_MPIAIJ_MPIAIJ,
1908: MatConjugate_MPIAIJ,
1909: 0,
1910: /*105*/MatSetValuesRow_MPIAIJ,
1911: MatRealPart_MPIAIJ,
1912: MatImaginaryPart_MPIAIJ};
1914: /* ----------------------------------------------------------------------------------------*/
1919: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
1920: {
1921: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1925: MatStoreValues(aij->A);
1926: MatStoreValues(aij->B);
1927: return(0);
1928: }
1934: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
1935: {
1936: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1940: MatRetrieveValues(aij->A);
1941: MatRetrieveValues(aij->B);
1942: return(0);
1943: }
1946: #include petscpc.h
1950: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
1951: {
1952: Mat_MPIAIJ *b;
1954: PetscInt i;
1957: B->preallocated = PETSC_TRUE;
1958: if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
1959: if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
1960: if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
1961: if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);
1963: B->rmap.bs = B->cmap.bs = 1;
1964: PetscMapInitialize(B->comm,&B->rmap);
1965: PetscMapInitialize(B->comm,&B->cmap);
1966: if (d_nnz) {
1967: for (i=0; i<B->rmap.n; i++) {
1968: if (d_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than 0: local row %D value %D",i,d_nnz[i]);
1969: }
1970: }
1971: if (o_nnz) {
1972: for (i=0; i<B->rmap.n; i++) {
1973: if (o_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than 0: local row %D value %D",i,o_nnz[i]);
1974: }
1975: }
1976: b = (Mat_MPIAIJ*)B->data;
1978: /* Explicitly create 2 MATSEQAIJ matrices. */
1979: MatCreate(PETSC_COMM_SELF,&b->A);
1980: MatSetSizes(b->A,B->rmap.n,B->cmap.n,B->rmap.n,B->cmap.n);
1981: MatSetType(b->A,MATSEQAIJ);
1982: PetscLogObjectParent(B,b->A);
1983: MatCreate(PETSC_COMM_SELF,&b->B);
1984: MatSetSizes(b->B,B->rmap.n,B->cmap.N,B->rmap.n,B->cmap.N);
1985: MatSetType(b->B,MATSEQAIJ);
1986: PetscLogObjectParent(B,b->B);
1988: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
1989: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
1991: return(0);
1992: }
1997: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
1998: {
1999: Mat mat;
2000: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
2004: *newmat = 0;
2005: MatCreate(matin->comm,&mat);
2006: MatSetSizes(mat,matin->rmap.n,matin->cmap.n,matin->rmap.N,matin->cmap.N);
2007: MatSetType(mat,matin->type_name);
2008: PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
2009: a = (Mat_MPIAIJ*)mat->data;
2010:
2011: mat->factor = matin->factor;
2012: mat->rmap.bs = matin->rmap.bs;
2013: mat->assembled = PETSC_TRUE;
2014: mat->insertmode = NOT_SET_VALUES;
2015: mat->preallocated = PETSC_TRUE;
2017: a->size = oldmat->size;
2018: a->rank = oldmat->rank;
2019: a->donotstash = oldmat->donotstash;
2020: a->roworiented = oldmat->roworiented;
2021: a->rowindices = 0;
2022: a->rowvalues = 0;
2023: a->getrowactive = PETSC_FALSE;
2025: PetscMapCopy(mat->comm,&matin->rmap,&mat->rmap);
2026: PetscMapCopy(mat->comm,&matin->cmap,&mat->cmap);
2028: MatStashCreate_Private(matin->comm,1,&mat->stash);
2029: if (oldmat->colmap) {
2030: #if defined (PETSC_USE_CTABLE)
2031: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2032: #else
2033: PetscMalloc((mat->cmap.N)*sizeof(PetscInt),&a->colmap);
2034: PetscLogObjectMemory(mat,(mat->cmap.N)*sizeof(PetscInt));
2035: PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap.N)*sizeof(PetscInt));
2036: #endif
2037: } else a->colmap = 0;
2038: if (oldmat->garray) {
2039: PetscInt len;
2040: len = oldmat->B->cmap.n;
2041: PetscMalloc((len+1)*sizeof(PetscInt),&a->garray);
2042: PetscLogObjectMemory(mat,len*sizeof(PetscInt));
2043: if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
2044: } else a->garray = 0;
2045:
2046: VecDuplicate(oldmat->lvec,&a->lvec);
2047: PetscLogObjectParent(mat,a->lvec);
2048: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2049: PetscLogObjectParent(mat,a->Mvctx);
2050: MatDuplicate(oldmat->A,cpvalues,&a->A);
2051: PetscLogObjectParent(mat,a->A);
2052: MatDuplicate(oldmat->B,cpvalues,&a->B);
2053: PetscLogObjectParent(mat,a->B);
2054: PetscFListDuplicate(matin->qlist,&mat->qlist);
2055: *newmat = mat;
2056: return(0);
2057: }
2059: #include petscsys.h
2063: PetscErrorCode MatLoad_MPIAIJ(PetscViewer viewer, MatType type,Mat *newmat)
2064: {
2065: Mat A;
2066: PetscScalar *vals,*svals;
2067: MPI_Comm comm = ((PetscObject)viewer)->comm;
2068: MPI_Status status;
2070: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag,maxnz;
2071: PetscInt i,nz,j,rstart,rend,mmax;
2072: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
2073: PetscInt *ourlens = PETSC_NULL,*procsnz = PETSC_NULL,*offlens = PETSC_NULL,jj,*mycols,*smycols;
2074: PetscInt cend,cstart,n,*rowners;
2075: int fd;
2078: MPI_Comm_size(comm,&size);
2079: MPI_Comm_rank(comm,&rank);
2080: if (!rank) {
2081: PetscViewerBinaryGetDescriptor(viewer,&fd);
2082: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
2083: if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2084: }
2086: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2087: M = header[1]; N = header[2];
2088: /* determine ownership of all rows */
2089: m = M/size + ((M % size) > rank);
2090: PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
2091: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
2093: /* First process needs enough room for process with most rows */
2094: if (!rank) {
2095: mmax = rowners[1];
2096: for (i=2; i<size; i++) {
2097: mmax = PetscMax(mmax,rowners[i]);
2098: }
2099: } else mmax = m;
2101: rowners[0] = 0;
2102: for (i=2; i<=size; i++) {
2103: mmax = PetscMax(mmax,rowners[i]);
2104: rowners[i] += rowners[i-1];
2105: }
2106: rstart = rowners[rank];
2107: rend = rowners[rank+1];
2109: /* distribute row lengths to all processors */
2110: PetscMalloc2(mmax,PetscInt,&ourlens,mmax,PetscInt,&offlens);
2111: if (!rank) {
2112: PetscBinaryRead(fd,ourlens,m,PETSC_INT);
2113: PetscMalloc(m*sizeof(PetscInt),&rowlengths);
2114: PetscMalloc(size*sizeof(PetscInt),&procsnz);
2115: PetscMemzero(procsnz,size*sizeof(PetscInt));
2116: for (j=0; j<m; j++) {
2117: procsnz[0] += ourlens[j];
2118: }
2119: for (i=1; i<size; i++) {
2120: PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
2121: /* calculate the number of nonzeros on each processor */
2122: for (j=0; j<rowners[i+1]-rowners[i]; j++) {
2123: procsnz[i] += rowlengths[j];
2124: }
2125: MPI_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
2126: }
2127: PetscFree(rowlengths);
2128: } else {
2129: MPI_Recv(ourlens,m,MPIU_INT,0,tag,comm,&status);
2130: }
2132: if (!rank) {
2133: /* determine max buffer needed and allocate it */
2134: maxnz = 0;
2135: for (i=0; i<size; i++) {
2136: maxnz = PetscMax(maxnz,procsnz[i]);
2137: }
2138: PetscMalloc(maxnz*sizeof(PetscInt),&cols);
2140: /* read in my part of the matrix column indices */
2141: nz = procsnz[0];
2142: PetscMalloc(nz*sizeof(PetscInt),&mycols);
2143: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
2145: /* read in every one elses and ship off */
2146: for (i=1; i<size; i++) {
2147: nz = procsnz[i];
2148: PetscBinaryRead(fd,cols,nz,PETSC_INT);
2149: MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
2150: }
2151: PetscFree(cols);
2152: } else {
2153: /* determine buffer space needed for message */
2154: nz = 0;
2155: for (i=0; i<m; i++) {
2156: nz += ourlens[i];
2157: }
2158: PetscMalloc(nz*sizeof(PetscInt),&mycols);
2160: /* receive message of column indices*/
2161: MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
2162: MPI_Get_count(&status,MPIU_INT,&maxnz);
2163: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2164: }
2166: /* determine column ownership if matrix is not square */
2167: if (N != M) {
2168: n = N/size + ((N % size) > rank);
2169: MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
2170: cstart = cend - n;
2171: } else {
2172: cstart = rstart;
2173: cend = rend;
2174: n = cend - cstart;
2175: }
2177: /* loop over local rows, determining number of off diagonal entries */
2178: PetscMemzero(offlens,m*sizeof(PetscInt));
2179: jj = 0;
2180: for (i=0; i<m; i++) {
2181: for (j=0; j<ourlens[i]; j++) {
2182: if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
2183: jj++;
2184: }
2185: }
2187: /* create our matrix */
2188: for (i=0; i<m; i++) {
2189: ourlens[i] -= offlens[i];
2190: }
2191: MatCreate(comm,&A);
2192: MatSetSizes(A,m,n,M,N);
2193: MatSetType(A,type);
2194: MatMPIAIJSetPreallocation(A,0,ourlens,0,offlens);
2196: MatSetOption(A,MAT_COLUMNS_SORTED);
2197: for (i=0; i<m; i++) {
2198: ourlens[i] += offlens[i];
2199: }
2201: if (!rank) {
2202: PetscMalloc((maxnz+1)*sizeof(PetscScalar),&vals);
2204: /* read in my part of the matrix numerical values */
2205: nz = procsnz[0];
2206: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2207:
2208: /* insert into matrix */
2209: jj = rstart;
2210: smycols = mycols;
2211: svals = vals;
2212: for (i=0; i<m; i++) {
2213: MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
2214: smycols += ourlens[i];
2215: svals += ourlens[i];
2216: jj++;
2217: }
2219: /* read in other processors and ship out */
2220: for (i=1; i<size; i++) {
2221: nz = procsnz[i];
2222: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2223: MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);
2224: }
2225: PetscFree(procsnz);
2226: } else {
2227: /* receive numeric values */
2228: PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);
2230: /* receive message of values*/
2231: MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);
2232: MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2233: if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2235: /* insert into matrix */
2236: jj = rstart;
2237: smycols = mycols;
2238: svals = vals;
2239: for (i=0; i<m; i++) {
2240: MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
2241: smycols += ourlens[i];
2242: svals += ourlens[i];
2243: jj++;
2244: }
2245: }
2246: PetscFree2(ourlens,offlens);
2247: PetscFree(vals);
2248: PetscFree(mycols);
2249: PetscFree(rowners);
2251: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2252: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2253: *newmat = A;
2254: return(0);
2255: }
2259: /*
2260: Not great since it makes two copies of the submatrix, first an SeqAIJ
2261: in local and then by concatenating the local matrices the end result.
2262: Writing it directly would be much like MatGetSubMatrices_MPIAIJ()
2263: */
2264: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2265: {
2267: PetscMPIInt rank,size;
2268: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j;
2269: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
2270: Mat *local,M,Mreuse;
2271: PetscScalar *vwork,*aa;
2272: MPI_Comm comm = mat->comm;
2273: Mat_SeqAIJ *aij;
2277: MPI_Comm_rank(comm,&rank);
2278: MPI_Comm_size(comm,&size);
2280: if (call == MAT_REUSE_MATRIX) {
2281: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);
2282: if (!Mreuse) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2283: local = &Mreuse;
2284: MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&local);
2285: } else {
2286: MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&local);
2287: Mreuse = *local;
2288: PetscFree(local);
2289: }
2291: /*
2292: m - number of local rows
2293: n - number of columns (same on all processors)
2294: rstart - first row in new global matrix generated
2295: */
2296: MatGetSize(Mreuse,&m,&n);
2297: if (call == MAT_INITIAL_MATRIX) {
2298: aij = (Mat_SeqAIJ*)(Mreuse)->data;
2299: ii = aij->i;
2300: jj = aij->j;
2302: /*
2303: Determine the number of non-zeros in the diagonal and off-diagonal
2304: portions of the matrix in order to do correct preallocation
2305: */
2307: /* first get start and end of "diagonal" columns */
2308: if (csize == PETSC_DECIDE) {
2309: ISGetSize(isrow,&mglobal);
2310: if (mglobal == n) { /* square matrix */
2311: nlocal = m;
2312: } else {
2313: nlocal = n/size + ((n % size) > rank);
2314: }
2315: } else {
2316: nlocal = csize;
2317: }
2318: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2319: rstart = rend - nlocal;
2320: if (rank == size - 1 && rend != n) {
2321: SETERRQ2(PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
2322: }
2324: /* next, compute all the lengths */
2325: PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
2326: olens = dlens + m;
2327: for (i=0; i<m; i++) {
2328: jend = ii[i+1] - ii[i];
2329: olen = 0;
2330: dlen = 0;
2331: for (j=0; j<jend; j++) {
2332: if (*jj < rstart || *jj >= rend) olen++;
2333: else dlen++;
2334: jj++;
2335: }
2336: olens[i] = olen;
2337: dlens[i] = dlen;
2338: }
2339: MatCreate(comm,&M);
2340: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
2341: MatSetType(M,mat->type_name);
2342: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
2343: PetscFree(dlens);
2344: } else {
2345: PetscInt ml,nl;
2347: M = *newmat;
2348: MatGetLocalSize(M,&ml,&nl);
2349: if (ml != m) SETERRQ(PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2350: MatZeroEntries(M);
2351: /*
2352: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2353: rather than the slower MatSetValues().
2354: */
2355: M->was_assembled = PETSC_TRUE;
2356: M->assembled = PETSC_FALSE;
2357: }
2358: MatGetOwnershipRange(M,&rstart,&rend);
2359: aij = (Mat_SeqAIJ*)(Mreuse)->data;
2360: ii = aij->i;
2361: jj = aij->j;
2362: aa = aij->a;
2363: for (i=0; i<m; i++) {
2364: row = rstart + i;
2365: nz = ii[i+1] - ii[i];
2366: cwork = jj; jj += nz;
2367: vwork = aa; aa += nz;
2368: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2369: }
2371: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2372: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2373: *newmat = M;
2375: /* save submatrix used in processor for next request */
2376: if (call == MAT_INITIAL_MATRIX) {
2377: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2378: PetscObjectDereference((PetscObject)Mreuse);
2379: }
2381: return(0);
2382: }
2387: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
2388: {
2389: PetscInt m,cstart, cend,j,nnz,i,d;
2390: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
2391: const PetscInt *JJ;
2392: PetscScalar *values;
2396: if (Ii[0]) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
2398: B->rmap.bs = B->cmap.bs = 1;
2399: PetscMapInitialize(B->comm,&B->rmap);
2400: PetscMapInitialize(B->comm,&B->cmap);
2401: m = B->rmap.n;
2402: cstart = B->cmap.rstart;
2403: cend = B->cmap.rend;
2404: rstart = B->rmap.rstart;
2406: PetscMalloc((2*m+1)*sizeof(PetscInt),&d_nnz);
2407: o_nnz = d_nnz + m;
2409: for (i=0; i<m; i++) {
2410: nnz = Ii[i+1]- Ii[i];
2411: JJ = J + Ii[i];
2412: nnz_max = PetscMax(nnz_max,nnz);
2413: if (nnz < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
2414: for (j=0; j<nnz; j++) {
2415: if (*JJ >= cstart) break;
2416: JJ++;
2417: }
2418: d = 0;
2419: for (; j<nnz; j++) {
2420: if (*JJ++ >= cend) break;
2421: d++;
2422: }
2423: d_nnz[i] = d;
2424: o_nnz[i] = nnz - d;
2425: }
2426: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2427: PetscFree(d_nnz);
2429: if (v) values = (PetscScalar*)v;
2430: else {
2431: PetscMalloc((nnz_max+1)*sizeof(PetscScalar),&values);
2432: PetscMemzero(values,nnz_max*sizeof(PetscScalar));
2433: }
2435: MatSetOption(B,MAT_COLUMNS_SORTED);
2436: for (i=0; i<m; i++) {
2437: ii = i + rstart;
2438: nnz = Ii[i+1]- Ii[i];
2439: MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
2440: }
2441: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2442: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2443: MatSetOption(B,MAT_COLUMNS_UNSORTED);
2445: if (!v) {
2446: PetscFree(values);
2447: }
2448: return(0);
2449: }
2454: /*@
2455: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
2456: (the default parallel PETSc format).
2458: Collective on MPI_Comm
2460: Input Parameters:
2461: + B - the matrix
2462: . i - the indices into j for the start of each local row (starts with zero)
2463: . j - the column indices for each local row (starts with zero) these must be sorted for each row
2464: - v - optional values in the matrix
2466: Level: developer
2468: Notes: this actually copies the values from i[], j[], and a[] to put them into PETSc's internal
2469: storage format. Thus changing the values in a[] after this call will not effect the matrix values.
2471: .keywords: matrix, aij, compressed row, sparse, parallel
2473: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateMPIAIJ(), MPIAIJ,
2474: MatCreateSeqAIJWithArrays()
2475: @*/
2476: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2477: {
2478: PetscErrorCode ierr,(*f)(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]);
2481: PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",(void (**)(void))&f);
2482: if (f) {
2483: (*f)(B,i,j,v);
2484: }
2485: return(0);
2486: }
2490: /*@C
2491: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
2492: (the default parallel PETSc format). For good matrix assembly performance
2493: the user should preallocate the matrix storage by setting the parameters
2494: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
2495: performance can be increased by more than a factor of 50.
2497: Collective on MPI_Comm
2499: Input Parameters:
2500: + A - the matrix
2501: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
2502: (same value is used for all local rows)
2503: . d_nnz - array containing the number of nonzeros in the various rows of the
2504: DIAGONAL portion of the local submatrix (possibly different for each row)
2505: or PETSC_NULL, if d_nz is used to specify the nonzero structure.
2506: The size of this array is equal to the number of local rows, i.e 'm'.
2507: You must leave room for the diagonal entry even if it is zero.
2508: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
2509: submatrix (same value is used for all local rows).
2510: - o_nnz - array containing the number of nonzeros in the various rows of the
2511: OFF-DIAGONAL portion of the local submatrix (possibly different for
2512: each row) or PETSC_NULL, if o_nz is used to specify the nonzero
2513: structure. The size of this array is equal to the number
2514: of local rows, i.e 'm'.
2516: If the *_nnz parameter is given then the *_nz parameter is ignored
2518: The AIJ format (also called the Yale sparse matrix format or
2519: compressed row storage (CSR)), is fully compatible with standard Fortran 77
2520: storage. The stored row and column indices begin with zero. See the users manual for details.
2522: The parallel matrix is partitioned such that the first m0 rows belong to
2523: process 0, the next m1 rows belong to process 1, the next m2 rows belong
2524: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
2526: The DIAGONAL portion of the local submatrix of a processor can be defined
2527: as the submatrix which is obtained by extraction the part corresponding
2528: to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the
2529: first row that belongs to the processor, and r2 is the last row belonging
2530: to the this processor. This is a square mxm matrix. The remaining portion
2531: of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.
2533: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
2535: Example usage:
2536:
2537: Consider the following 8x8 matrix with 34 non-zero values, that is
2538: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
2539: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
2540: as follows:
2542: .vb
2543: 1 2 0 | 0 3 0 | 0 4
2544: Proc0 0 5 6 | 7 0 0 | 8 0
2545: 9 0 10 | 11 0 0 | 12 0
2546: -------------------------------------
2547: 13 0 14 | 15 16 17 | 0 0
2548: Proc1 0 18 0 | 19 20 21 | 0 0
2549: 0 0 0 | 22 23 0 | 24 0
2550: -------------------------------------
2551: Proc2 25 26 27 | 0 0 28 | 29 0
2552: 30 0 0 | 31 32 33 | 0 34
2553: .ve
2555: This can be represented as a collection of submatrices as:
2557: .vb
2558: A B C
2559: D E F
2560: G H I
2561: .ve
2563: Where the submatrices A,B,C are owned by proc0, D,E,F are
2564: owned by proc1, G,H,I are owned by proc2.
2566: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2567: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2568: The 'M','N' parameters are 8,8, and have the same values on all procs.
2570: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
2571: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
2572: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
2573: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
2574: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
2575: matrix, ans [DF] as another SeqAIJ matrix.
2577: When d_nz, o_nz parameters are specified, d_nz storage elements are
2578: allocated for every row of the local diagonal submatrix, and o_nz
2579: storage locations are allocated for every row of the OFF-DIAGONAL submat.
2580: One way to choose d_nz and o_nz is to use the max nonzerors per local
2581: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
2582: In this case, the values of d_nz,o_nz are:
2583: .vb
2584: proc0 : dnz = 2, o_nz = 2
2585: proc1 : dnz = 3, o_nz = 2
2586: proc2 : dnz = 1, o_nz = 4
2587: .ve
2588: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
2589: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
2590: for proc3. i.e we are using 12+15+10=37 storage locations to store
2591: 34 values.
2593: When d_nnz, o_nnz parameters are specified, the storage is specified
2594: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
2595: In the above case the values for d_nnz,o_nnz are:
2596: .vb
2597: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
2598: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
2599: proc2: d_nnz = [1,1] and o_nnz = [4,4]
2600: .ve
2601: Here the space allocated is sum of all the above values i.e 34, and
2602: hence pre-allocation is perfect.
2604: Level: intermediate
2606: .keywords: matrix, aij, compressed row, sparse, parallel
2608: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateMPIAIJ(), MatMPIAIJSetPreallocationCSR(),
2609: MPIAIJ
2610: @*/
2611: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2612: {
2613: PetscErrorCode ierr,(*f)(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);
2616: PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",(void (**)(void))&f);
2617: if (f) {
2618: (*f)(B,d_nz,d_nnz,o_nz,o_nnz);
2619: }
2620: return(0);
2621: }
2625: /*@C
2626: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
2627: CSR format the local rows.
2629: Collective on MPI_Comm
2631: Input Parameters:
2632: + comm - MPI communicator
2633: . m - number of local rows (Cannot be PETSC_DECIDE)
2634: . n - This value should be the same as the local size used in creating the
2635: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
2636: calculated if N is given) For square matrices n is almost always m.
2637: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
2638: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
2639: . i - row indices
2640: . j - column indices
2641: - a - matrix values
2643: Output Parameter:
2644: . mat - the matrix
2645: Level: intermediate
2647: Notes:
2648: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
2649: thus you CANNOT change the matrix entries by changing the values of a[] after you have
2650: called this routine.
2652: The i and j indices are 0 based
2654: .keywords: matrix, aij, compressed row, sparse, parallel
2656: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
2657: MPIAIJ, MatCreateMPIAIJ()
2658: @*/
2659: PetscErrorCode MatCreateMPIIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
2660: {
2664: if (i[0]) {
2665: SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
2666: }
2667: if (m < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
2668: MatCreate(comm,mat);
2669: MatSetType(*mat,MATMPIAIJ);
2670: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
2671: return(0);
2672: }
2676: /*@C
2677: MatCreateMPIAIJ - Creates a sparse parallel matrix in AIJ format
2678: (the default parallel PETSc format). For good matrix assembly performance
2679: the user should preallocate the matrix storage by setting the parameters
2680: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
2681: performance can be increased by more than a factor of 50.
2683: Collective on MPI_Comm
2685: Input Parameters:
2686: + comm - MPI communicator
2687: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
2688: This value should be the same as the local size used in creating the
2689: y vector for the matrix-vector product y = Ax.
2690: . n - This value should be the same as the local size used in creating the
2691: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
2692: calculated if N is given) For square matrices n is almost always m.
2693: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
2694: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
2695: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
2696: (same value is used for all local rows)
2697: . d_nnz - array containing the number of nonzeros in the various rows of the
2698: DIAGONAL portion of the local submatrix (possibly different for each row)
2699: or PETSC_NULL, if d_nz is used to specify the nonzero structure.
2700: The size of this array is equal to the number of local rows, i.e 'm'.
2701: You must leave room for the diagonal entry even if it is zero.
2702: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
2703: submatrix (same value is used for all local rows).
2704: - o_nnz - array containing the number of nonzeros in the various rows of the
2705: OFF-DIAGONAL portion of the local submatrix (possibly different for
2706: each row) or PETSC_NULL, if o_nz is used to specify the nonzero
2707: structure. The size of this array is equal to the number
2708: of local rows, i.e 'm'.
2710: Output Parameter:
2711: . A - the matrix
2713: Notes:
2714: If the *_nnz parameter is given then the *_nz parameter is ignored
2716: m,n,M,N parameters specify the size of the matrix, and its partitioning across
2717: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
2718: storage requirements for this matrix.
2720: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
2721: processor than it must be used on all processors that share the object for
2722: that argument.
2724: The user MUST specify either the local or global matrix dimensions
2725: (possibly both).
2727: The parallel matrix is partitioned such that the first m0 rows belong to
2728: process 0, the next m1 rows belong to process 1, the next m2 rows belong
2729: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
2731: The DIAGONAL portion of the local submatrix of a processor can be defined
2732: as the submatrix which is obtained by extraction the part corresponding
2733: to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the
2734: first row that belongs to the processor, and r2 is the last row belonging
2735: to the this processor. This is a square mxm matrix. The remaining portion
2736: of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.
2738: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
2740: When calling this routine with a single process communicator, a matrix of
2741: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
2742: type of communicator, use the construction mechanism:
2743: MatCreate(...,&A); MatSetType(A,MPIAIJ); MatMPIAIJSetPreallocation(A,...);
2745: By default, this format uses inodes (identical nodes) when possible.
2746: We search for consecutive rows with the same nonzero structure, thereby
2747: reusing matrix information to achieve increased efficiency.
2749: Options Database Keys:
2750: + -mat_no_inode - Do not use inodes
2751: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
2752: - -mat_aij_oneindex - Internally use indexing starting at 1
2753: rather than 0. Note that when calling MatSetValues(),
2754: the user still MUST index entries starting at 0!
2757: Example usage:
2758:
2759: Consider the following 8x8 matrix with 34 non-zero values, that is
2760: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
2761: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
2762: as follows:
2764: .vb
2765: 1 2 0 | 0 3 0 | 0 4
2766: Proc0 0 5 6 | 7 0 0 | 8 0
2767: 9 0 10 | 11 0 0 | 12 0
2768: -------------------------------------
2769: 13 0 14 | 15 16 17 | 0 0
2770: Proc1 0 18 0 | 19 20 21 | 0 0
2771: 0 0 0 | 22 23 0 | 24 0
2772: -------------------------------------
2773: Proc2 25 26 27 | 0 0 28 | 29 0
2774: 30 0 0 | 31 32 33 | 0 34
2775: .ve
2777: This can be represented as a collection of submatrices as:
2779: .vb
2780: A B C
2781: D E F
2782: G H I
2783: .ve
2785: Where the submatrices A,B,C are owned by proc0, D,E,F are
2786: owned by proc1, G,H,I are owned by proc2.
2788: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2789: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2790: The 'M','N' parameters are 8,8, and have the same values on all procs.
2792: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
2793: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
2794: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
2795: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
2796: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
2797: matrix, ans [DF] as another SeqAIJ matrix.
2799: When d_nz, o_nz parameters are specified, d_nz storage elements are
2800: allocated for every row of the local diagonal submatrix, and o_nz
2801: storage locations are allocated for every row of the OFF-DIAGONAL submat.
2802: One way to choose d_nz and o_nz is to use the max nonzerors per local
2803: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
2804: In this case, the values of d_nz,o_nz are:
2805: .vb
2806: proc0 : dnz = 2, o_nz = 2
2807: proc1 : dnz = 3, o_nz = 2
2808: proc2 : dnz = 1, o_nz = 4
2809: .ve
2810: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
2811: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
2812: for proc3. i.e we are using 12+15+10=37 storage locations to store
2813: 34 values.
2815: When d_nnz, o_nnz parameters are specified, the storage is specified
2816: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
2817: In the above case the values for d_nnz,o_nnz are:
2818: .vb
2819: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
2820: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
2821: proc2: d_nnz = [1,1] and o_nnz = [4,4]
2822: .ve
2823: Here the space allocated is sum of all the above values i.e 34, and
2824: hence pre-allocation is perfect.
2826: Level: intermediate
2828: .keywords: matrix, aij, compressed row, sparse, parallel
2830: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
2831: MPIAIJ, MatCreateMPIAIJWithArrays()
2832: @*/
2833: PetscErrorCode MatCreateMPIAIJ(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)
2834: {
2836: PetscMPIInt size;
2839: MatCreate(comm,A);
2840: MatSetSizes(*A,m,n,M,N);
2841: MPI_Comm_size(comm,&size);
2842: if (size > 1) {
2843: MatSetType(*A,MATMPIAIJ);
2844: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
2845: } else {
2846: MatSetType(*A,MATSEQAIJ);
2847: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
2848: }
2849: return(0);
2850: }
2854: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[])
2855: {
2856: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2859: *Ad = a->A;
2860: *Ao = a->B;
2861: *colmap = a->garray;
2862: return(0);
2863: }
2867: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
2868: {
2870: PetscInt i;
2871: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2874: if (coloring->ctype == IS_COLORING_LOCAL) {
2875: ISColoringValue *allcolors,*colors;
2876: ISColoring ocoloring;
2878: /* set coloring for diagonal portion */
2879: MatSetColoring_SeqAIJ(a->A,coloring);
2881: /* set coloring for off-diagonal portion */
2882: ISAllGatherColors(A->comm,coloring->n,coloring->colors,PETSC_NULL,&allcolors);
2883: PetscMalloc((a->B->cmap.n+1)*sizeof(ISColoringValue),&colors);
2884: for (i=0; i<a->B->cmap.n; i++) {
2885: colors[i] = allcolors[a->garray[i]];
2886: }
2887: PetscFree(allcolors);
2888: ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap.n,colors,&ocoloring);
2889: MatSetColoring_SeqAIJ(a->B,ocoloring);
2890: ISColoringDestroy(ocoloring);
2891: } else if (coloring->ctype == IS_COLORING_GHOSTED) {
2892: ISColoringValue *colors;
2893: PetscInt *larray;
2894: ISColoring ocoloring;
2896: /* set coloring for diagonal portion */
2897: PetscMalloc((a->A->cmap.n+1)*sizeof(PetscInt),&larray);
2898: for (i=0; i<a->A->cmap.n; i++) {
2899: larray[i] = i + A->cmap.rstart;
2900: }
2901: ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->A->cmap.n,larray,PETSC_NULL,larray);
2902: PetscMalloc((a->A->cmap.n+1)*sizeof(ISColoringValue),&colors);
2903: for (i=0; i<a->A->cmap.n; i++) {
2904: colors[i] = coloring->colors[larray[i]];
2905: }
2906: PetscFree(larray);
2907: ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap.n,colors,&ocoloring);
2908: MatSetColoring_SeqAIJ(a->A,ocoloring);
2909: ISColoringDestroy(ocoloring);
2911: /* set coloring for off-diagonal portion */
2912: PetscMalloc((a->B->cmap.n+1)*sizeof(PetscInt),&larray);
2913: ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->B->cmap.n,a->garray,PETSC_NULL,larray);
2914: PetscMalloc((a->B->cmap.n+1)*sizeof(ISColoringValue),&colors);
2915: for (i=0; i<a->B->cmap.n; i++) {
2916: colors[i] = coloring->colors[larray[i]];
2917: }
2918: PetscFree(larray);
2919: ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap.n,colors,&ocoloring);
2920: MatSetColoring_SeqAIJ(a->B,ocoloring);
2921: ISColoringDestroy(ocoloring);
2922: } else {
2923: SETERRQ1(PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);
2924: }
2926: return(0);
2927: }
2929: #if defined(PETSC_HAVE_ADIC)
2932: PetscErrorCode MatSetValuesAdic_MPIAIJ(Mat A,void *advalues)
2933: {
2934: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2938: MatSetValuesAdic_SeqAIJ(a->A,advalues);
2939: MatSetValuesAdic_SeqAIJ(a->B,advalues);
2940: return(0);
2941: }
2942: #endif
2946: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
2947: {
2948: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2952: MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
2953: MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
2954: return(0);
2955: }
2959: /*@C
2960: MatMerge - Creates a single large PETSc matrix by concatinating sequential
2961: matrices from each processor
2963: Collective on MPI_Comm
2965: Input Parameters:
2966: + comm - the communicators the parallel matrix will live on
2967: . inmat - the input sequential matrices
2968: . n - number of local columns (or PETSC_DECIDE)
2969: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
2971: Output Parameter:
2972: . outmat - the parallel matrix generated
2974: Level: advanced
2976: Notes: The number of columns of the matrix in EACH processor MUST be the same.
2978: @*/
2979: PetscErrorCode MatMerge(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
2980: {
2982: PetscInt m,N,i,rstart,nnz,Ii,*dnz,*onz;
2983: PetscInt *indx;
2984: PetscScalar *values;
2987: MatGetSize(inmat,&m,&N);
2988: if (scall == MAT_INITIAL_MATRIX){
2989: /* count nonzeros in each row, for diagonal and off diagonal portion of matrix */
2990: if (n == PETSC_DECIDE){
2991: PetscSplitOwnership(comm,&n,&N);
2992: }
2993: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
2994: rstart -= m;
2996: MatPreallocateInitialize(comm,m,n,dnz,onz);
2997: for (i=0;i<m;i++) {
2998: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
2999: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
3000: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
3001: }
3002: /* This routine will ONLY return MPIAIJ type matrix */
3003: MatCreate(comm,outmat);
3004: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3005: MatSetType(*outmat,MATMPIAIJ);
3006: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
3007: MatPreallocateFinalize(dnz,onz);
3008:
3009: } else if (scall == MAT_REUSE_MATRIX){
3010: MatGetOwnershipRange(*outmat,&rstart,PETSC_NULL);
3011: } else {
3012: SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
3013: }
3015: for (i=0;i<m;i++) {
3016: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3017: Ii = i + rstart;
3018: MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
3019: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3020: }
3021: MatDestroy(inmat);
3022: MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
3023: MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
3025: return(0);
3026: }
3030: PetscErrorCode MatFileSplit(Mat A,char *outfile)
3031: {
3032: PetscErrorCode ierr;
3033: PetscMPIInt rank;
3034: PetscInt m,N,i,rstart,nnz;
3035: size_t len;
3036: const PetscInt *indx;
3037: PetscViewer out;
3038: char *name;
3039: Mat B;
3040: const PetscScalar *values;
3043: MatGetLocalSize(A,&m,0);
3044: MatGetSize(A,0,&N);
3045: /* Should this be the type of the diagonal block of A? */
3046: MatCreate(PETSC_COMM_SELF,&B);
3047: MatSetSizes(B,m,N,m,N);
3048: MatSetType(B,MATSEQAIJ);
3049: MatSeqAIJSetPreallocation(B,0,PETSC_NULL);
3050: MatGetOwnershipRange(A,&rstart,0);
3051: for (i=0;i<m;i++) {
3052: MatGetRow(A,i+rstart,&nnz,&indx,&values);
3053: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
3054: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
3055: }
3056: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3057: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3059: MPI_Comm_rank(A->comm,&rank);
3060: PetscStrlen(outfile,&len);
3061: PetscMalloc((len+5)*sizeof(char),&name);
3062: sprintf(name,"%s.%d",outfile,rank);
3063: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
3064: PetscFree(name);
3065: MatView(B,out);
3066: PetscViewerDestroy(out);
3067: MatDestroy(B);
3068: return(0);
3069: }
3071: EXTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
3074: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
3075: {
3076: PetscErrorCode ierr;
3077: Mat_Merge_SeqsToMPI *merge;
3078: PetscObjectContainer container;
3081: PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);
3082: if (container) {
3083: PetscObjectContainerGetPointer(container,(void **)&merge);
3084: PetscFree(merge->id_r);
3085: PetscFree(merge->len_s);
3086: PetscFree(merge->len_r);
3087: PetscFree(merge->bi);
3088: PetscFree(merge->bj);
3089: PetscFree(merge->buf_ri);
3090: PetscFree(merge->buf_rj);
3091: PetscFree(merge->coi);
3092: PetscFree(merge->coj);
3093: PetscFree(merge->owners_co);
3094: PetscFree(merge->rowmap.range);
3095:
3096: PetscObjectContainerDestroy(container);
3097: PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
3098: }
3099: PetscFree(merge);
3101: MatDestroy_MPIAIJ(A);
3102: return(0);
3103: }
3105: #include src/mat/utils/freespace.h
3106: #include petscbt.h
3107: static PetscEvent logkey_seqstompinum = 0;
3110: /*@C
3111: MatMerge_SeqsToMPI - Creates a MPIAIJ matrix by adding sequential
3112: matrices from each processor
3114: Collective on MPI_Comm
3116: Input Parameters:
3117: + comm - the communicators the parallel matrix will live on
3118: . seqmat - the input sequential matrices
3119: . m - number of local rows (or PETSC_DECIDE)
3120: . n - number of local columns (or PETSC_DECIDE)
3121: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3123: Output Parameter:
3124: . mpimat - the parallel matrix generated
3126: Level: advanced
3128: Notes:
3129: The dimensions of the sequential matrix in each processor MUST be the same.
3130: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
3131: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
3132: @*/
3133: PetscErrorCode MatMerge_SeqsToMPINumeric(Mat seqmat,Mat mpimat)
3134: {
3135: PetscErrorCode ierr;
3136: MPI_Comm comm=mpimat->comm;
3137: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
3138: PetscMPIInt size,rank,taga,*len_s;
3139: PetscInt N=mpimat->cmap.N,i,j,*owners,*ai=a->i,*aj=a->j;
3140: PetscInt proc,m;
3141: PetscInt **buf_ri,**buf_rj;
3142: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
3143: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
3144: MPI_Request *s_waits,*r_waits;
3145: MPI_Status *status;
3146: MatScalar *aa=a->a,**abuf_r,*ba_i;
3147: Mat_Merge_SeqsToMPI *merge;
3148: PetscObjectContainer container;
3149:
3151: if (!logkey_seqstompinum) {
3153: }
3156: MPI_Comm_size(comm,&size);
3157: MPI_Comm_rank(comm,&rank);
3159: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject *)&container);
3160: if (container) {
3161: PetscObjectContainerGetPointer(container,(void **)&merge);
3162: }
3163: bi = merge->bi;
3164: bj = merge->bj;
3165: buf_ri = merge->buf_ri;
3166: buf_rj = merge->buf_rj;
3168: PetscMalloc(size*sizeof(MPI_Status),&status);
3169: owners = merge->rowmap.range;
3170: len_s = merge->len_s;
3172: /* send and recv matrix values */
3173: /*-----------------------------*/
3174: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
3175: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
3177: PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);
3178: for (proc=0,k=0; proc<size; proc++){
3179: if (!len_s[proc]) continue;
3180: i = owners[proc];
3181: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
3182: k++;
3183: }
3185: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
3186: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
3187: PetscFree(status);
3189: PetscFree(s_waits);
3190: PetscFree(r_waits);
3192: /* insert mat values of mpimat */
3193: /*----------------------------*/
3194: PetscMalloc(N*sizeof(MatScalar),&ba_i);
3195: PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
3196: nextrow = buf_ri_k + merge->nrecv;
3197: nextai = nextrow + merge->nrecv;
3199: for (k=0; k<merge->nrecv; k++){
3200: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
3201: nrows = *(buf_ri_k[k]);
3202: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
3203: nextai[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */
3204: }
3206: /* set values of ba */
3207: m = merge->rowmap.n;
3208: for (i=0; i<m; i++) {
3209: arow = owners[rank] + i;
3210: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
3211: bnzi = bi[i+1] - bi[i];
3212: PetscMemzero(ba_i,bnzi*sizeof(MatScalar));
3214: /* add local non-zero vals of this proc's seqmat into ba */
3215: anzi = ai[arow+1] - ai[arow];
3216: aj = a->j + ai[arow];
3217: aa = a->a + ai[arow];
3218: nextaj = 0;
3219: for (j=0; nextaj<anzi; j++){
3220: if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
3221: ba_i[j] += aa[nextaj++];
3222: }
3223: }
3225: /* add received vals into ba */
3226: for (k=0; k<merge->nrecv; k++){ /* k-th received message */
3227: /* i-th row */
3228: if (i == *nextrow[k]) {
3229: anzi = *(nextai[k]+1) - *nextai[k];
3230: aj = buf_rj[k] + *(nextai[k]);
3231: aa = abuf_r[k] + *(nextai[k]);
3232: nextaj = 0;
3233: for (j=0; nextaj<anzi; j++){
3234: if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
3235: ba_i[j] += aa[nextaj++];
3236: }
3237: }
3238: nextrow[k]++; nextai[k]++;
3239: }
3240: }
3241: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
3242: }
3243: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
3244: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
3246: PetscFree(abuf_r);
3247: PetscFree(ba_i);
3248: PetscFree(buf_ri_k);
3250: return(0);
3251: }
3253: static PetscEvent logkey_seqstompisym = 0;
3256: PetscErrorCode MatMerge_SeqsToMPISymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
3257: {
3258: PetscErrorCode ierr;
3259: Mat B_mpi;
3260: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
3261: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
3262: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
3263: PetscInt M=seqmat->rmap.n,N=seqmat->cmap.n,i,*owners,*ai=a->i,*aj=a->j;
3264: PetscInt len,proc,*dnz,*onz;
3265: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
3266: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
3267: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
3268: MPI_Status *status;
3269: PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
3270: PetscBT lnkbt;
3271: Mat_Merge_SeqsToMPI *merge;
3272: PetscObjectContainer container;
3275: if (!logkey_seqstompisym) {
3277: }
3280: /* make sure it is a PETSc comm */
3281: PetscCommDuplicate(comm,&comm,PETSC_NULL);
3282: MPI_Comm_size(comm,&size);
3283: MPI_Comm_rank(comm,&rank);
3284:
3285: PetscNew(Mat_Merge_SeqsToMPI,&merge);
3286: PetscMalloc(size*sizeof(MPI_Status),&status);
3288: /* determine row ownership */
3289: /*---------------------------------------------------------*/
3290: merge->rowmap.n = m;
3291: merge->rowmap.N = M;
3292: merge->rowmap.bs = 1;
3293: PetscMapInitialize(comm,&merge->rowmap);
3294: PetscMalloc(size*sizeof(PetscMPIInt),&len_si);
3295: PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);
3296:
3297: m = merge->rowmap.n;
3298: M = merge->rowmap.N;
3299: owners = merge->rowmap.range;
3301: /* determine the number of messages to send, their lengths */
3302: /*---------------------------------------------------------*/
3303: len_s = merge->len_s;
3305: len = 0; /* length of buf_si[] */
3306: merge->nsend = 0;
3307: for (proc=0; proc<size; proc++){
3308: len_si[proc] = 0;
3309: if (proc == rank){
3310: len_s[proc] = 0;
3311: } else {
3312: len_si[proc] = owners[proc+1] - owners[proc] + 1;
3313: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
3314: }
3315: if (len_s[proc]) {
3316: merge->nsend++;
3317: nrows = 0;
3318: for (i=owners[proc]; i<owners[proc+1]; i++){
3319: if (ai[i+1] > ai[i]) nrows++;
3320: }
3321: len_si[proc] = 2*(nrows+1);
3322: len += len_si[proc];
3323: }
3324: }
3326: /* determine the number and length of messages to receive for ij-structure */
3327: /*-------------------------------------------------------------------------*/
3328: PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);
3329: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
3331: /* post the Irecv of j-structure */
3332: /*-------------------------------*/
3333: PetscCommGetNewTag(comm,&tagj);
3334: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
3336: /* post the Isend of j-structure */
3337: /*--------------------------------*/
3338: PetscMalloc((2*merge->nsend+1)*sizeof(MPI_Request),&si_waits);
3339: sj_waits = si_waits + merge->nsend;
3341: for (proc=0, k=0; proc<size; proc++){
3342: if (!len_s[proc]) continue;
3343: i = owners[proc];
3344: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
3345: k++;
3346: }
3348: /* receives and sends of j-structure are complete */
3349: /*------------------------------------------------*/
3350: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
3351: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
3352:
3353: /* send and recv i-structure */
3354: /*---------------------------*/
3355: PetscCommGetNewTag(comm,&tagi);
3356: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
3357:
3358: PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);
3359: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
3360: for (proc=0,k=0; proc<size; proc++){
3361: if (!len_s[proc]) continue;
3362: /* form outgoing message for i-structure:
3363: buf_si[0]: nrows to be sent
3364: [1:nrows]: row index (global)
3365: [nrows+1:2*nrows+1]: i-structure index
3366: */
3367: /*-------------------------------------------*/
3368: nrows = len_si[proc]/2 - 1;
3369: buf_si_i = buf_si + nrows+1;
3370: buf_si[0] = nrows;
3371: buf_si_i[0] = 0;
3372: nrows = 0;
3373: for (i=owners[proc]; i<owners[proc+1]; i++){
3374: anzi = ai[i+1] - ai[i];
3375: if (anzi) {
3376: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
3377: buf_si[nrows+1] = i-owners[proc]; /* local row index */
3378: nrows++;
3379: }
3380: }
3381: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
3382: k++;
3383: buf_si += len_si[proc];
3384: }
3386: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
3387: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
3389: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
3390: for (i=0; i<merge->nrecv; i++){
3391: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
3392: }
3394: PetscFree(len_si);
3395: PetscFree(len_ri);
3396: PetscFree(rj_waits);
3397: PetscFree(si_waits);
3398: PetscFree(ri_waits);
3399: PetscFree(buf_s);
3400: PetscFree(status);
3402: /* compute a local seq matrix in each processor */
3403: /*----------------------------------------------*/
3404: /* allocate bi array and free space for accumulating nonzero column info */
3405: PetscMalloc((m+1)*sizeof(PetscInt),&bi);
3406: bi[0] = 0;
3408: /* create and initialize a linked list */
3409: nlnk = N+1;
3410: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
3411:
3412: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
3413: len = 0;
3414: len = ai[owners[rank+1]] - ai[owners[rank]];
3415: PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);
3416: current_space = free_space;
3418: /* determine symbolic info for each local row */
3419: PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
3420: nextrow = buf_ri_k + merge->nrecv;
3421: nextai = nextrow + merge->nrecv;
3422: for (k=0; k<merge->nrecv; k++){
3423: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
3424: nrows = *buf_ri_k[k];
3425: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
3426: nextai[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */
3427: }
3429: MatPreallocateInitialize(comm,m,n,dnz,onz);
3430: len = 0;
3431: for (i=0;i<m;i++) {
3432: bnzi = 0;
3433: /* add local non-zero cols of this proc's seqmat into lnk */
3434: arow = owners[rank] + i;
3435: anzi = ai[arow+1] - ai[arow];
3436: aj = a->j + ai[arow];
3437: PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3438: bnzi += nlnk;
3439: /* add received col data into lnk */
3440: for (k=0; k<merge->nrecv; k++){ /* k-th received message */
3441: if (i == *nextrow[k]) { /* i-th row */
3442: anzi = *(nextai[k]+1) - *nextai[k];
3443: aj = buf_rj[k] + *nextai[k];
3444: PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3445: bnzi += nlnk;
3446: nextrow[k]++; nextai[k]++;
3447: }
3448: }
3449: if (len < bnzi) len = bnzi; /* =max(bnzi) */
3451: /* if free space is not available, make more free space */
3452: if (current_space->local_remaining<bnzi) {
3453: PetscFreeSpaceGet(current_space->total_array_size,¤t_space);
3454: nspacedouble++;
3455: }
3456: /* copy data into free space, then initialize lnk */
3457: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
3458: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
3460: current_space->array += bnzi;
3461: current_space->local_used += bnzi;
3462: current_space->local_remaining -= bnzi;
3463:
3464: bi[i+1] = bi[i] + bnzi;
3465: }
3466:
3467: PetscFree(buf_ri_k);
3469: PetscMalloc((bi[m]+1)*sizeof(PetscInt),&bj);
3470: PetscFreeSpaceContiguous(&free_space,bj);
3471: PetscLLDestroy(lnk,lnkbt);
3473: /* create symbolic parallel matrix B_mpi */
3474: /*---------------------------------------*/
3475: MatCreate(comm,&B_mpi);
3476: if (n==PETSC_DECIDE) {
3477: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
3478: } else {
3479: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3480: }
3481: MatSetType(B_mpi,MATMPIAIJ);
3482: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
3483: MatPreallocateFinalize(dnz,onz);
3485: /* B_mpi is not ready for use - assembly will be done by MatMerge_SeqsToMPINumeric() */
3486: B_mpi->assembled = PETSC_FALSE;
3487: B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
3488: merge->bi = bi;
3489: merge->bj = bj;
3490: merge->buf_ri = buf_ri;
3491: merge->buf_rj = buf_rj;
3492: merge->coi = PETSC_NULL;
3493: merge->coj = PETSC_NULL;
3494: merge->owners_co = PETSC_NULL;
3496: /* attach the supporting struct to B_mpi for reuse */
3497: PetscObjectContainerCreate(PETSC_COMM_SELF,&container);
3498: PetscObjectContainerSetPointer(container,merge);
3499: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
3500: *mpimat = B_mpi;
3502: PetscCommDestroy(&comm);
3504: return(0);
3505: }
3507: static PetscEvent logkey_seqstompi = 0;
3510: PetscErrorCode MatMerge_SeqsToMPI(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
3511: {
3512: PetscErrorCode ierr;
3515: if (!logkey_seqstompi) {
3517: }
3519: if (scall == MAT_INITIAL_MATRIX){
3520: MatMerge_SeqsToMPISymbolic(comm,seqmat,m,n,mpimat);
3521: }
3522: MatMerge_SeqsToMPINumeric(seqmat,*mpimat);
3524: return(0);
3525: }
3526: static PetscEvent logkey_getlocalmat = 0;
3529: /*@C
3530: MatGetLocalMat - Creates a SeqAIJ matrix by taking all its local rows
3532: Not Collective
3534: Input Parameters:
3535: + A - the matrix
3536: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3538: Output Parameter:
3539: . A_loc - the local sequential matrix generated
3541: Level: developer
3543: @*/
3544: PetscErrorCode MatGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
3545: {
3546: PetscErrorCode ierr;
3547: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
3548: Mat_SeqAIJ *mat,*a=(Mat_SeqAIJ*)(mpimat->A)->data,*b=(Mat_SeqAIJ*)(mpimat->B)->data;
3549: PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*cmap=mpimat->garray;
3550: PetscScalar *aa=a->a,*ba=b->a,*ca;
3551: PetscInt am=A->rmap.n,i,j,k,cstart=A->cmap.rstart;
3552: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
3555: if (!logkey_getlocalmat) {
3557: }
3559: if (scall == MAT_INITIAL_MATRIX){
3560: PetscMalloc((1+am)*sizeof(PetscInt),&ci);
3561: ci[0] = 0;
3562: for (i=0; i<am; i++){
3563: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
3564: }
3565: PetscMalloc((1+ci[am])*sizeof(PetscInt),&cj);
3566: PetscMalloc((1+ci[am])*sizeof(PetscScalar),&ca);
3567: k = 0;
3568: for (i=0; i<am; i++) {
3569: ncols_o = bi[i+1] - bi[i];
3570: ncols_d = ai[i+1] - ai[i];
3571: /* off-diagonal portion of A */
3572: for (jo=0; jo<ncols_o; jo++) {
3573: col = cmap[*bj];
3574: if (col >= cstart) break;
3575: cj[k] = col; bj++;
3576: ca[k++] = *ba++;
3577: }
3578: /* diagonal portion of A */
3579: for (j=0; j<ncols_d; j++) {
3580: cj[k] = cstart + *aj++;
3581: ca[k++] = *aa++;
3582: }
3583: /* off-diagonal portion of A */
3584: for (j=jo; j<ncols_o; j++) {
3585: cj[k] = cmap[*bj++];
3586: ca[k++] = *ba++;
3587: }
3588: }
3589: /* put together the new matrix */
3590: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap.N,ci,cj,ca,A_loc);
3591: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
3592: /* Since these are PETSc arrays, change flags to free them as necessary. */
3593: mat = (Mat_SeqAIJ*)(*A_loc)->data;
3594: mat->free_a = PETSC_TRUE;
3595: mat->free_ij = PETSC_TRUE;
3596: mat->nonew = 0;
3597: } else if (scall == MAT_REUSE_MATRIX){
3598: mat=(Mat_SeqAIJ*)(*A_loc)->data;
3599: ci = mat->i; cj = mat->j; ca = mat->a;
3600: for (i=0; i<am; i++) {
3601: /* off-diagonal portion of A */
3602: ncols_o = bi[i+1] - bi[i];
3603: for (jo=0; jo<ncols_o; jo++) {
3604: col = cmap[*bj];
3605: if (col >= cstart) break;
3606: *ca++ = *ba++; bj++;
3607: }
3608: /* diagonal portion of A */
3609: ncols_d = ai[i+1] - ai[i];
3610: for (j=0; j<ncols_d; j++) *ca++ = *aa++;
3611: /* off-diagonal portion of A */
3612: for (j=jo; j<ncols_o; j++) {
3613: *ca++ = *ba++; bj++;
3614: }
3615: }
3616: } else {
3617: SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
3618: }
3621: return(0);
3622: }
3624: static PetscEvent logkey_getlocalmatcondensed = 0;
3627: /*@C
3628: MatGetLocalMatCondensed - Creates a SeqAIJ matrix by taking all its local rows and NON-ZERO columns
3630: Not Collective
3632: Input Parameters:
3633: + A - the matrix
3634: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3635: - row, col - index sets of rows and columns to extract (or PETSC_NULL)
3637: Output Parameter:
3638: . A_loc - the local sequential matrix generated
3640: Level: developer
3642: @*/
3643: PetscErrorCode MatGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
3644: {
3645: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
3646: PetscErrorCode ierr;
3647: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
3648: IS isrowa,iscola;
3649: Mat *aloc;
3652: if (!logkey_getlocalmatcondensed) {
3654: }
3656: if (!row){
3657: start = A->rmap.rstart; end = A->rmap.rend;
3658: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
3659: } else {
3660: isrowa = *row;
3661: }
3662: if (!col){
3663: start = A->cmap.rstart;
3664: cmap = a->garray;
3665: nzA = a->A->cmap.n;
3666: nzB = a->B->cmap.n;
3667: PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
3668: ncols = 0;
3669: for (i=0; i<nzB; i++) {
3670: if (cmap[i] < start) idx[ncols++] = cmap[i];
3671: else break;
3672: }
3673: imark = i;
3674: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
3675: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
3676: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&iscola);
3677: PetscFree(idx);
3678: } else {
3679: iscola = *col;
3680: }
3681: if (scall != MAT_INITIAL_MATRIX){
3682: PetscMalloc(sizeof(Mat),&aloc);
3683: aloc[0] = *A_loc;
3684: }
3685: MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
3686: *A_loc = aloc[0];
3687: PetscFree(aloc);
3688: if (!row){
3689: ISDestroy(isrowa);
3690: }
3691: if (!col){
3692: ISDestroy(iscola);
3693: }
3695: return(0);
3696: }
3698: static PetscEvent logkey_GetBrowsOfAcols = 0;
3701: /*@C
3702: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
3704: Collective on Mat
3706: Input Parameters:
3707: + A,B - the matrices in mpiaij format
3708: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3709: - rowb, colb - index sets of rows and columns of B to extract (or PETSC_NULL)
3711: Output Parameter:
3712: + rowb, colb - index sets of rows and columns of B to extract
3713: . brstart - row index of B_seq from which next B->rmap.n rows are taken from B's local rows
3714: - B_seq - the sequential matrix generated
3716: Level: developer
3718: @*/
3719: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,PetscInt *brstart,Mat *B_seq)
3720: {
3721: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
3722: PetscErrorCode ierr;
3723: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
3724: IS isrowb,iscolb;
3725: Mat *bseq;
3726:
3728: if (A->cmap.rstart != B->rmap.rstart || A->cmap.rend != B->rmap.rend){
3729: SETERRQ4(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);
3730: }
3731: if (!logkey_GetBrowsOfAcols) {
3733: }
3735:
3736: if (scall == MAT_INITIAL_MATRIX){
3737: start = A->cmap.rstart;
3738: cmap = a->garray;
3739: nzA = a->A->cmap.n;
3740: nzB = a->B->cmap.n;
3741: PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
3742: ncols = 0;
3743: for (i=0; i<nzB; i++) { /* row < local row index */
3744: if (cmap[i] < start) idx[ncols++] = cmap[i];
3745: else break;
3746: }
3747: imark = i;
3748: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
3749: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
3750: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&isrowb);
3751: PetscFree(idx);
3752: *brstart = imark;
3753: ISCreateStride(PETSC_COMM_SELF,B->cmap.N,0,1,&iscolb);
3754: } else {
3755: if (!rowb || !colb) SETERRQ(PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
3756: isrowb = *rowb; iscolb = *colb;
3757: PetscMalloc(sizeof(Mat),&bseq);
3758: bseq[0] = *B_seq;
3759: }
3760: MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
3761: *B_seq = bseq[0];
3762: PetscFree(bseq);
3763: if (!rowb){
3764: ISDestroy(isrowb);
3765: } else {
3766: *rowb = isrowb;
3767: }
3768: if (!colb){
3769: ISDestroy(iscolb);
3770: } else {
3771: *colb = iscolb;
3772: }
3774: return(0);
3775: }
3777: static PetscEvent logkey_GetBrowsOfAocols = 0;
3780: /*@C
3781: MatGetBrowsOfAoCols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
3782: of the OFF-DIAGONAL portion of local A
3784: Collective on Mat
3786: Input Parameters:
3787: + A,B - the matrices in mpiaij format
3788: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3789: . startsj - starting point in B's sending and receiving j-arrays, saved for MAT_REUSE (or PETSC_NULL)
3790: - bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or PETSC_NULL)
3792: Output Parameter:
3793: + B_oth - the sequential matrix generated
3795: Level: developer
3797: @*/
3798: PetscErrorCode MatGetBrowsOfAoCols(Mat A,Mat B,MatReuse scall,PetscInt **startsj,PetscScalar **bufa_ptr,Mat *B_oth)
3799: {
3800: VecScatter_MPI_General *gen_to,*gen_from;
3801: PetscErrorCode ierr;
3802: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
3803: Mat_SeqAIJ *b_oth;
3804: VecScatter ctx=a->Mvctx;
3805: MPI_Comm comm=ctx->comm;
3806: PetscMPIInt *rprocs,*sprocs,tag=ctx->tag,rank;
3807: PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap.n,row,*b_othi,*b_othj;
3808: PetscScalar *rvalues,*svalues,*b_otha,*bufa,*bufA;
3809: PetscInt i,k,l,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
3810: MPI_Request *rwaits = PETSC_NULL,*swaits = PETSC_NULL;
3811: MPI_Status *sstatus,rstatus;
3812: PetscInt *cols;
3813: PetscScalar *vals;
3814: PetscMPIInt j;
3815:
3817: if (A->cmap.rstart != B->rmap.rstart || A->cmap.rend != B->rmap.rend){
3818: SETERRQ4(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);
3819: }
3820: if (!logkey_GetBrowsOfAocols) {
3822: }
3824: MPI_Comm_rank(comm,&rank);
3826: gen_to = (VecScatter_MPI_General*)ctx->todata;
3827: gen_from = (VecScatter_MPI_General*)ctx->fromdata;
3828: rvalues = gen_from->values; /* holds the length of sending row */
3829: svalues = gen_to->values; /* holds the length of receiving row */
3830: nrecvs = gen_from->n;
3831: nsends = gen_to->n;
3833: PetscMalloc2(nrecvs,MPI_Request,&rwaits,nsends,MPI_Request,&swaits);
3834: srow = gen_to->indices; /* local row index to be sent */
3835: rstarts = gen_from->starts;
3836: sstarts = gen_to->starts;
3837: rprocs = gen_from->procs;
3838: sprocs = gen_to->procs;
3839: sstatus = gen_to->sstatus;
3841: if (!startsj || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
3842: if (scall == MAT_INITIAL_MATRIX){
3843: /* i-array */
3844: /*---------*/
3845: /* post receives */
3846: for (i=0; i<nrecvs; i++){
3847: rowlen = (PetscInt*)rvalues + rstarts[i];
3848: nrows = rstarts[i+1]-rstarts[i];
3849: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
3850: }
3852: /* pack the outgoing message */
3853: PetscMalloc((nsends+nrecvs+3)*sizeof(PetscInt),&sstartsj);
3854: rstartsj = sstartsj + nsends +1;
3855: sstartsj[0] = 0; rstartsj[0] = 0;
3856: len = 0; /* total length of j or a array to be sent */
3857: k = 0;
3858: for (i=0; i<nsends; i++){
3859: rowlen = (PetscInt*)svalues + sstarts[i];
3860: nrows = sstarts[i+1]-sstarts[i]; /* num of rows */
3861: for (j=0; j<nrows; j++) {
3862: row = srow[k] + B->rmap.range[rank]; /* global row idx */
3863: MatGetRow_MPIAIJ(B,row,&rowlen[j],PETSC_NULL,PETSC_NULL); /* rowlength */
3864: len += rowlen[j];
3865: MatRestoreRow_MPIAIJ(B,row,&ncols,PETSC_NULL,PETSC_NULL);
3866: k++;
3867: }
3868: MPI_Isend(rowlen,nrows,MPIU_INT,sprocs[i],tag,comm,swaits+i);
3869: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
3870: }
3871: /* recvs and sends of i-array are completed */
3872: i = nrecvs;
3873: while (i--) {
3874: MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3875: }
3876: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3877: /* allocate buffers for sending j and a arrays */
3878: PetscMalloc((len+1)*sizeof(PetscInt),&bufj);
3879: PetscMalloc((len+1)*sizeof(PetscScalar),&bufa);
3881: /* create i-array of B_oth */
3882: PetscMalloc((aBn+2)*sizeof(PetscInt),&b_othi);
3883: b_othi[0] = 0;
3884: len = 0; /* total length of j or a array to be received */
3885: k = 0;
3886: for (i=0; i<nrecvs; i++){
3887: rowlen = (PetscInt*)rvalues + rstarts[i];
3888: nrows = rstarts[i+1]-rstarts[i];
3889: for (j=0; j<nrows; j++) {
3890: b_othi[k+1] = b_othi[k] + rowlen[j];
3891: len += rowlen[j]; k++;
3892: }
3893: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
3894: }
3896: /* allocate space for j and a arrrays of B_oth */
3897: PetscMalloc((b_othi[aBn]+1)*sizeof(PetscInt),&b_othj);
3898: PetscMalloc((b_othi[aBn]+1)*sizeof(PetscScalar),&b_otha);
3900: /* j-array */
3901: /*---------*/
3902: /* post receives of j-array */
3903: for (i=0; i<nrecvs; i++){
3904: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
3905: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
3906: }
3907: k = 0;
3908: for (i=0; i<nsends; i++){
3909: nrows = sstarts[i+1]-sstarts[i]; /* num of rows */
3910: bufJ = bufj+sstartsj[i];
3911: for (j=0; j<nrows; j++) {
3912: row = srow[k++] + B->rmap.range[rank]; /* global row idx */
3913: MatGetRow_MPIAIJ(B,row,&ncols,&cols,PETSC_NULL);
3914: for (l=0; l<ncols; l++){
3915: *bufJ++ = cols[l];
3916: }
3917: MatRestoreRow_MPIAIJ(B,row,&ncols,&cols,PETSC_NULL);
3918: }
3919: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
3920: }
3922: /* recvs and sends of j-array are completed */
3923: i = nrecvs;
3924: while (i--) {
3925: MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3926: }
3927: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3928: } else if (scall == MAT_REUSE_MATRIX){
3929: sstartsj = *startsj;
3930: rstartsj = sstartsj + nsends +1;
3931: bufa = *bufa_ptr;
3932: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
3933: b_otha = b_oth->a;
3934: } else {
3935: SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
3936: }
3938: /* a-array */
3939: /*---------*/
3940: /* post receives of a-array */
3941: for (i=0; i<nrecvs; i++){
3942: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
3943: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
3944: }
3945: k = 0;
3946: for (i=0; i<nsends; i++){
3947: nrows = sstarts[i+1]-sstarts[i];
3948: bufA = bufa+sstartsj[i];
3949: for (j=0; j<nrows; j++) {
3950: row = srow[k++] + B->rmap.range[rank]; /* global row idx */
3951: MatGetRow_MPIAIJ(B,row,&ncols,PETSC_NULL,&vals);
3952: for (l=0; l<ncols; l++){
3953: *bufA++ = vals[l];
3954: }
3955: MatRestoreRow_MPIAIJ(B,row,&ncols,PETSC_NULL,&vals);
3957: }
3958: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
3959: }
3960: /* recvs and sends of a-array are completed */
3961: i = nrecvs;
3962: while (i--) {
3963: MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3964: }
3965: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3966: PetscFree2(rwaits,swaits);
3968: if (scall == MAT_INITIAL_MATRIX){
3969: /* put together the new matrix */
3970: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap.N,b_othi,b_othj,b_otha,B_oth);
3972: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
3973: /* Since these are PETSc arrays, change flags to free them as necessary. */
3974: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
3975: b_oth->free_a = PETSC_TRUE;
3976: b_oth->free_ij = PETSC_TRUE;
3977: b_oth->nonew = 0;
3979: PetscFree(bufj);
3980: if (!startsj || !bufa_ptr){
3981: PetscFree(sstartsj);
3982: PetscFree(bufa_ptr);
3983: } else {
3984: *startsj = sstartsj;
3985: *bufa_ptr = bufa;
3986: }
3987: }
3989:
3990: return(0);
3991: }
3995: /*@C
3996: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
3998: Not Collective
4000: Input Parameters:
4001: . A - The matrix in mpiaij format
4003: Output Parameter:
4004: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
4005: . colmap - A map from global column index to local index into lvec
4006: - multScatter - A scatter from the argument of a matrix-vector product to lvec
4008: Level: developer
4010: @*/
4011: #if defined (PETSC_USE_CTABLE)
4012: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
4013: #else
4014: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
4015: #endif
4016: {
4017: Mat_MPIAIJ *a;
4024: a = (Mat_MPIAIJ *) A->data;
4025: if (lvec) *lvec = a->lvec;
4026: if (colmap) *colmap = a->colmap;
4027: if (multScatter) *multScatter = a->Mvctx;
4028: return(0);
4029: }
4036: /*MC
4037: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
4039: Options Database Keys:
4040: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
4042: Level: beginner
4044: .seealso: MatCreateMPIAIJ
4045: M*/
4050: PetscErrorCode MatCreate_MPIAIJ(Mat B)
4051: {
4052: Mat_MPIAIJ *b;
4054: PetscMPIInt size;
4057: MPI_Comm_size(B->comm,&size);
4059: PetscNew(Mat_MPIAIJ,&b);
4060: B->data = (void*)b;
4061: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4062: B->factor = 0;
4063: B->rmap.bs = 1;
4064: B->assembled = PETSC_FALSE;
4065: B->mapping = 0;
4067: B->insertmode = NOT_SET_VALUES;
4068: b->size = size;
4069: MPI_Comm_rank(B->comm,&b->rank);
4071: /* build cache for off array entries formed */
4072: MatStashCreate_Private(B->comm,1,&B->stash);
4073: b->donotstash = PETSC_FALSE;
4074: b->colmap = 0;
4075: b->garray = 0;
4076: b->roworiented = PETSC_TRUE;
4078: /* stuff used for matrix vector multiply */
4079: b->lvec = PETSC_NULL;
4080: b->Mvctx = PETSC_NULL;
4082: /* stuff for MatGetRow() */
4083: b->rowindices = 0;
4084: b->rowvalues = 0;
4085: b->getrowactive = PETSC_FALSE;
4088: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
4089: "MatStoreValues_MPIAIJ",
4090: MatStoreValues_MPIAIJ);
4091: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
4092: "MatRetrieveValues_MPIAIJ",
4093: MatRetrieveValues_MPIAIJ);
4094: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
4095: "MatGetDiagonalBlock_MPIAIJ",
4096: MatGetDiagonalBlock_MPIAIJ);
4097: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatIsTranspose_C",
4098: "MatIsTranspose_MPIAIJ",
4099: MatIsTranspose_MPIAIJ);
4100: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocation_C",
4101: "MatMPIAIJSetPreallocation_MPIAIJ",
4102: MatMPIAIJSetPreallocation_MPIAIJ);
4103: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",
4104: "MatMPIAIJSetPreallocationCSR_MPIAIJ",
4105: MatMPIAIJSetPreallocationCSR_MPIAIJ);
4106: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
4107: "MatDiagonalScaleLocal_MPIAIJ",
4108: MatDiagonalScaleLocal_MPIAIJ);
4109: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpicsrperm_C",
4110: "MatConvert_MPIAIJ_MPICSRPERM",
4111: MatConvert_MPIAIJ_MPICSRPERM);
4112: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpicrl_C",
4113: "MatConvert_MPIAIJ_MPICRL",
4114: MatConvert_MPIAIJ_MPICRL);
4115: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
4116: return(0);
4117: }
4120: /*
4121: Special version for direct calls from Fortran
4122: */
4123: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4124: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
4125: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4126: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
4127: #endif
4129: /* Change these macros so can be used in void function */
4130: #undef CHKERRQ
4131: #define CHKERRQ(ierr) CHKERRABORT(mat->comm,ierr)
4132: #undef SETERRQ2
4133: #define SETERRQ2(ierr,b,c,d) CHKERRABORT(mat->comm,ierr)
4134: #undef SETERRQ
4135: #define SETERRQ(ierr,b) CHKERRABORT(mat->comm,ierr)
4140: void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
4141: {
4142: Mat mat = *mmat;
4143: PetscInt m = *mm, n = *mn;
4144: InsertMode addv = *maddv;
4145: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
4146: PetscScalar value;
4149: MatPreallocated(mat);
4150: if (mat->insertmode == NOT_SET_VALUES) {
4151: mat->insertmode = addv;
4152: }
4153: #if defined(PETSC_USE_DEBUG)
4154: else if (mat->insertmode != addv) {
4155: SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
4156: }
4157: #endif
4158: {
4159: PetscInt i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
4160: PetscInt cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;
4161: PetscTruth roworiented = aij->roworiented;
4163: /* Some Variables required in the macro */
4164: Mat A = aij->A;
4165: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4166: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
4167: PetscScalar *aa = a->a;
4168: PetscTruth ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
4169: Mat B = aij->B;
4170: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
4171: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap.n,am = aij->A->rmap.n;
4172: PetscScalar *ba = b->a;
4174: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
4175: PetscInt nonew = a->nonew;
4176: PetscScalar *ap1,*ap2;
4179: for (i=0; i<m; i++) {
4180: if (im[i] < 0) continue;
4181: #if defined(PETSC_USE_DEBUG)
4182: if (im[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap.N-1);
4183: #endif
4184: if (im[i] >= rstart && im[i] < rend) {
4185: row = im[i] - rstart;
4186: lastcol1 = -1;
4187: rp1 = aj + ai[row];
4188: ap1 = aa + ai[row];
4189: rmax1 = aimax[row];
4190: nrow1 = ailen[row];
4191: low1 = 0;
4192: high1 = nrow1;
4193: lastcol2 = -1;
4194: rp2 = bj + bi[row];
4195: ap2 = ba + bi[row];
4196: rmax2 = bimax[row];
4197: nrow2 = bilen[row];
4198: low2 = 0;
4199: high2 = nrow2;
4201: for (j=0; j<n; j++) {
4202: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
4203: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
4204: if (in[j] >= cstart && in[j] < cend){
4205: col = in[j] - cstart;
4206: MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
4207: } else if (in[j] < 0) continue;
4208: #if defined(PETSC_USE_DEBUG)
4209: else if (in[j] >= mat->cmap.N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap.N-1);}
4210: #endif
4211: else {
4212: if (mat->was_assembled) {
4213: if (!aij->colmap) {
4214: CreateColmap_MPIAIJ_Private(mat);
4215: }
4216: #if defined (PETSC_USE_CTABLE)
4217: PetscTableFind(aij->colmap,in[j]+1,&col);
4218: col--;
4219: #else
4220: col = aij->colmap[in[j]] - 1;
4221: #endif
4222: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
4223: DisAssemble_MPIAIJ(mat);
4224: col = in[j];
4225: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
4226: B = aij->B;
4227: b = (Mat_SeqAIJ*)B->data;
4228: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
4229: rp2 = bj + bi[row];
4230: ap2 = ba + bi[row];
4231: rmax2 = bimax[row];
4232: nrow2 = bilen[row];
4233: low2 = 0;
4234: high2 = nrow2;
4235: bm = aij->B->rmap.n;
4236: ba = b->a;
4237: }
4238: } else col = in[j];
4239: MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
4240: }
4241: }
4242: } else {
4243: if (!aij->donotstash) {
4244: if (roworiented) {
4245: if (ignorezeroentries && v[i*n] == 0.0) continue;
4246: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
4247: } else {
4248: if (ignorezeroentries && v[i] == 0.0) continue;
4249: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
4250: }
4251: }
4252: }
4253: }}
4254: PetscFunctionReturnVoid();
4255: }