Actual source code: mpiaij.c
petsc-3.3-p3 2012-08-29
2: #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/
3: #include <petscblaslapack.h>
5: /*MC
6: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
8: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
9: and MATMPIAIJ otherwise. As a result, for single process communicators,
10: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
11: for communicators controlling multiple processes. It is recommended that you call both of
12: the above preallocation routines for simplicity.
14: Options Database Keys:
15: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
17: Developer Notes: Subclasses include MATAIJCUSP, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when
18: enough exist.
20: Level: beginner
22: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
23: M*/
25: /*MC
26: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
28: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
29: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
30: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
31: for communicators controlling multiple processes. It is recommended that you call both of
32: the above preallocation routines for simplicity.
34: Options Database Keys:
35: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
37: Level: beginner
39: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
40: M*/
44: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
45: {
46: PetscErrorCode ierr;
47: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
48: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data;
49: Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data;
50: const PetscInt *ia,*ib;
51: const MatScalar *aa,*bb;
52: PetscInt na,nb,i,j,*rows,cnt=0,n0rows;
53: PetscInt m = M->rmap->n,rstart = M->rmap->rstart;
56: *keptrows = 0;
57: ia = a->i;
58: ib = b->i;
59: for (i=0; i<m; i++) {
60: na = ia[i+1] - ia[i];
61: nb = ib[i+1] - ib[i];
62: if (!na && !nb) {
63: cnt++;
64: goto ok1;
65: }
66: aa = a->a + ia[i];
67: for (j=0; j<na; j++) {
68: if (aa[j] != 0.0) goto ok1;
69: }
70: bb = b->a + ib[i];
71: for (j=0; j <nb; j++) {
72: if (bb[j] != 0.0) goto ok1;
73: }
74: cnt++;
75: ok1:;
76: }
77: MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,((PetscObject)M)->comm);
78: if (!n0rows) return(0);
79: PetscMalloc((M->rmap->n-cnt)*sizeof(PetscInt),&rows);
80: cnt = 0;
81: for (i=0; i<m; i++) {
82: na = ia[i+1] - ia[i];
83: nb = ib[i+1] - ib[i];
84: if (!na && !nb) continue;
85: aa = a->a + ia[i];
86: for(j=0; j<na;j++) {
87: if (aa[j] != 0.0) {
88: rows[cnt++] = rstart + i;
89: goto ok2;
90: }
91: }
92: bb = b->a + ib[i];
93: for (j=0; j<nb; j++) {
94: if (bb[j] != 0.0) {
95: rows[cnt++] = rstart + i;
96: goto ok2;
97: }
98: }
99: ok2:;
100: }
101: ISCreateGeneral(((PetscObject)M)->comm,cnt,rows,PETSC_OWN_POINTER,keptrows);
102: return(0);
103: }
107: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
108: {
109: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data;
111: PetscInt i,rstart,nrows,*rows;
114: *zrows = PETSC_NULL;
115: MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
116: MatGetOwnershipRange(M,&rstart,PETSC_NULL);
117: for (i=0; i<nrows; i++) rows[i] += rstart;
118: ISCreateGeneral(((PetscObject)M)->comm,nrows,rows,PETSC_OWN_POINTER,zrows);
119: return(0);
120: }
124: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
125: {
127: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
128: PetscInt i,n,*garray = aij->garray;
129: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data;
130: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data;
131: PetscReal *work;
134: MatGetSize(A,PETSC_NULL,&n);
135: PetscMalloc(n*sizeof(PetscReal),&work);
136: PetscMemzero(work,n*sizeof(PetscReal));
137: if (type == NORM_2) {
138: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
139: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
140: }
141: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
142: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
143: }
144: } else if (type == NORM_1) {
145: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
146: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
147: }
148: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
149: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
150: }
151: } else if (type == NORM_INFINITY) {
152: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
153: work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
154: }
155: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
156: work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
157: }
159: } else SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_WRONG,"Unknown NormType");
160: if (type == NORM_INFINITY) {
161: MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,A->hdr.comm);
162: } else {
163: MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,A->hdr.comm);
164: }
165: PetscFree(work);
166: if (type == NORM_2) {
167: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
168: }
169: return(0);
170: }
174: /*
175: Distributes a SeqAIJ matrix across a set of processes. Code stolen from
176: MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.
178: Only for square matrices
179: */
180: PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
181: {
182: PetscMPIInt rank,size;
183: PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz,*gmataj,cnt,row,*ld;
185: Mat mat;
186: Mat_SeqAIJ *gmata;
187: PetscMPIInt tag;
188: MPI_Status status;
189: PetscBool aij;
190: MatScalar *gmataa,*ao,*ad,*gmataarestore=0;
193: CHKMEMQ;
194: MPI_Comm_rank(comm,&rank);
195: MPI_Comm_size(comm,&size);
196: if (!rank) {
197: PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
198: if (!aij) SETERRQ1(((PetscObject)gmat)->comm,PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
199: }
200: if (reuse == MAT_INITIAL_MATRIX) {
201: MatCreate(comm,&mat);
202: MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
203: MatSetBlockSizes(mat,gmat->rmap->bs,gmat->cmap->bs);
204: MatSetType(mat,MATAIJ);
205: PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
206: PetscMalloc2(m,PetscInt,&dlens,m,PetscInt,&olens);
207: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
208: rowners[0] = 0;
209: for (i=2; i<=size; i++) {
210: rowners[i] += rowners[i-1];
211: }
212: rstart = rowners[rank];
213: rend = rowners[rank+1];
214: PetscObjectGetNewTag((PetscObject)mat,&tag);
215: if (!rank) {
216: gmata = (Mat_SeqAIJ*) gmat->data;
217: /* send row lengths to all processors */
218: for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
219: for (i=1; i<size; i++) {
220: MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
221: }
222: /* determine number diagonal and off-diagonal counts */
223: PetscMemzero(olens,m*sizeof(PetscInt));
224: PetscMalloc(m*sizeof(PetscInt),&ld);
225: PetscMemzero(ld,m*sizeof(PetscInt));
226: jj = 0;
227: for (i=0; i<m; i++) {
228: for (j=0; j<dlens[i]; j++) {
229: if (gmata->j[jj] < rstart) ld[i]++;
230: if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
231: jj++;
232: }
233: }
234: /* send column indices to other processes */
235: for (i=1; i<size; i++) {
236: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
237: MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
238: MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
239: }
241: /* send numerical values to other processes */
242: for (i=1; i<size; i++) {
243: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
244: MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
245: }
246: gmataa = gmata->a;
247: gmataj = gmata->j;
249: } else {
250: /* receive row lengths */
251: MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
252: /* receive column indices */
253: MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
254: PetscMalloc2(nz,PetscScalar,&gmataa,nz,PetscInt,&gmataj);
255: MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
256: /* determine number diagonal and off-diagonal counts */
257: PetscMemzero(olens,m*sizeof(PetscInt));
258: PetscMalloc(m*sizeof(PetscInt),&ld);
259: PetscMemzero(ld,m*sizeof(PetscInt));
260: jj = 0;
261: for (i=0; i<m; i++) {
262: for (j=0; j<dlens[i]; j++) {
263: if (gmataj[jj] < rstart) ld[i]++;
264: if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
265: jj++;
266: }
267: }
268: /* receive numerical values */
269: PetscMemzero(gmataa,nz*sizeof(PetscScalar));
270: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
271: }
272: /* set preallocation */
273: for (i=0; i<m; i++) {
274: dlens[i] -= olens[i];
275: }
276: MatSeqAIJSetPreallocation(mat,0,dlens);
277: MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
278:
279: for (i=0; i<m; i++) {
280: dlens[i] += olens[i];
281: }
282: cnt = 0;
283: for (i=0; i<m; i++) {
284: row = rstart + i;
285: MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
286: cnt += dlens[i];
287: }
288: if (rank) {
289: PetscFree2(gmataa,gmataj);
290: }
291: PetscFree2(dlens,olens);
292: PetscFree(rowners);
293: ((Mat_MPIAIJ*)(mat->data))->ld = ld;
294: *inmat = mat;
295: } else { /* column indices are already set; only need to move over numerical values from process 0 */
296: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
297: Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
298: mat = *inmat;
299: PetscObjectGetNewTag((PetscObject)mat,&tag);
300: if (!rank) {
301: /* send numerical values to other processes */
302: gmata = (Mat_SeqAIJ*) gmat->data;
303: MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
304: gmataa = gmata->a;
305: for (i=1; i<size; i++) {
306: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
307: MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
308: }
309: nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
310: } else {
311: /* receive numerical values from process 0*/
312: nz = Ad->nz + Ao->nz;
313: PetscMalloc(nz*sizeof(PetscScalar),&gmataa); gmataarestore = gmataa;
314: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
315: }
316: /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
317: ld = ((Mat_MPIAIJ*)(mat->data))->ld;
318: ad = Ad->a;
319: ao = Ao->a;
320: if (mat->rmap->n) {
321: i = 0;
322: nz = ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
323: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
324: }
325: for (i=1; i<mat->rmap->n; i++) {
326: nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
327: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
328: }
329: i--;
330: if (mat->rmap->n) {
331: nz = Ao->i[i+1] - Ao->i[i] - ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
332: }
333: if (rank) {
334: PetscFree(gmataarestore);
335: }
336: }
337: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
338: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
339: CHKMEMQ;
340: return(0);
341: }
343: /*
344: Local utility routine that creates a mapping from the global column
345: number to the local number in the off-diagonal part of the local
346: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
347: a slightly higher hash table cost; without it it is not scalable (each processor
348: has an order N integer array but is fast to acess.
349: */
352: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
353: {
354: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
356: PetscInt n = aij->B->cmap->n,i;
359: if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
360: #if defined (PETSC_USE_CTABLE)
361: PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
362: for (i=0; i<n; i++){
363: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
364: }
365: #else
366: PetscMalloc((mat->cmap->N+1)*sizeof(PetscInt),&aij->colmap);
367: PetscLogObjectMemory(mat,mat->cmap->N*sizeof(PetscInt));
368: PetscMemzero(aij->colmap,mat->cmap->N*sizeof(PetscInt));
369: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
370: #endif
371: return(0);
372: }
374: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \
375: { \
376: if (col <= lastcol1) low1 = 0; else high1 = nrow1; \
377: lastcol1 = col;\
378: while (high1-low1 > 5) { \
379: t = (low1+high1)/2; \
380: if (rp1[t] > col) high1 = t; \
381: else low1 = t; \
382: } \
383: for (_i=low1; _i<high1; _i++) { \
384: if (rp1[_i] > col) break; \
385: if (rp1[_i] == col) { \
386: if (addv == ADD_VALUES) ap1[_i] += value; \
387: else ap1[_i] = value; \
388: goto a_noinsert; \
389: } \
390: } \
391: if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
392: if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \
393: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
394: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
395: N = nrow1++ - 1; a->nz++; high1++; \
396: /* shift up all the later entries in this row */ \
397: for (ii=N; ii>=_i; ii--) { \
398: rp1[ii+1] = rp1[ii]; \
399: ap1[ii+1] = ap1[ii]; \
400: } \
401: rp1[_i] = col; \
402: ap1[_i] = value; \
403: a_noinsert: ; \
404: ailen[row] = nrow1; \
405: }
408: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \
409: { \
410: if (col <= lastcol2) low2 = 0; else high2 = nrow2; \
411: lastcol2 = col;\
412: while (high2-low2 > 5) { \
413: t = (low2+high2)/2; \
414: if (rp2[t] > col) high2 = t; \
415: else low2 = t; \
416: } \
417: for (_i=low2; _i<high2; _i++) { \
418: if (rp2[_i] > col) break; \
419: if (rp2[_i] == col) { \
420: if (addv == ADD_VALUES) ap2[_i] += value; \
421: else ap2[_i] = value; \
422: goto b_noinsert; \
423: } \
424: } \
425: if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
426: if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
427: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
428: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
429: N = nrow2++ - 1; b->nz++; high2++; \
430: /* shift up all the later entries in this row */ \
431: for (ii=N; ii>=_i; ii--) { \
432: rp2[ii+1] = rp2[ii]; \
433: ap2[ii+1] = ap2[ii]; \
434: } \
435: rp2[_i] = col; \
436: ap2[_i] = value; \
437: b_noinsert: ; \
438: bilen[row] = nrow2; \
439: }
443: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
444: {
445: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
446: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
448: PetscInt l,*garray = mat->garray,diag;
451: /* code only works for square matrices A */
453: /* find size of row to the left of the diagonal part */
454: MatGetOwnershipRange(A,&diag,0);
455: row = row - diag;
456: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
457: if (garray[b->j[b->i[row]+l]] > diag) break;
458: }
459: PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));
461: /* diagonal part */
462: PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));
464: /* right of diagonal part */
465: 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));
466: return(0);
467: }
471: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
472: {
473: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
474: PetscScalar value;
476: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
477: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
478: PetscBool roworiented = aij->roworiented;
480: /* Some Variables required in the macro */
481: Mat A = aij->A;
482: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
483: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
484: MatScalar *aa = a->a;
485: PetscBool ignorezeroentries = a->ignorezeroentries;
486: Mat B = aij->B;
487: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
488: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
489: MatScalar *ba = b->a;
491: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
492: PetscInt nonew;
493: MatScalar *ap1,*ap2;
497: for (i=0; i<m; i++) {
498: if (im[i] < 0) continue;
499: #if defined(PETSC_USE_DEBUG)
500: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
501: #endif
502: if (im[i] >= rstart && im[i] < rend) {
503: row = im[i] - rstart;
504: lastcol1 = -1;
505: rp1 = aj + ai[row];
506: ap1 = aa + ai[row];
507: rmax1 = aimax[row];
508: nrow1 = ailen[row];
509: low1 = 0;
510: high1 = nrow1;
511: lastcol2 = -1;
512: rp2 = bj + bi[row];
513: ap2 = ba + bi[row];
514: rmax2 = bimax[row];
515: nrow2 = bilen[row];
516: low2 = 0;
517: high2 = nrow2;
519: for (j=0; j<n; j++) {
520: if (v) {if (roworiented) value = v[i*n+j]; else value = v[i+j*m];} else value = 0.0;
521: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
522: if (in[j] >= cstart && in[j] < cend){
523: col = in[j] - cstart;
524: nonew = a->nonew;
525: MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
526: } else if (in[j] < 0) continue;
527: #if defined(PETSC_USE_DEBUG)
528: else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
529: #endif
530: else {
531: if (mat->was_assembled) {
532: if (!aij->colmap) {
533: MatCreateColmap_MPIAIJ_Private(mat);
534: }
535: #if defined (PETSC_USE_CTABLE)
536: PetscTableFind(aij->colmap,in[j]+1,&col);
537: col--;
538: #else
539: col = aij->colmap[in[j]] - 1;
540: #endif
541: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
542: MatDisAssemble_MPIAIJ(mat);
543: col = in[j];
544: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
545: B = aij->B;
546: b = (Mat_SeqAIJ*)B->data;
547: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
548: rp2 = bj + bi[row];
549: ap2 = ba + bi[row];
550: rmax2 = bimax[row];
551: nrow2 = bilen[row];
552: low2 = 0;
553: high2 = nrow2;
554: bm = aij->B->rmap->n;
555: ba = b->a;
556: }
557: } else col = in[j];
558: nonew = b->nonew;
559: MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
560: }
561: }
562: } else {
563: if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
564: if (!aij->donotstash) {
565: mat->assembled = PETSC_FALSE;
566: if (roworiented) {
567: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
568: } else {
569: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
570: }
571: }
572: }
573: }
574: return(0);
575: }
579: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
580: {
581: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
583: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
584: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
587: for (i=0; i<m; i++) {
588: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
589: if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
590: if (idxm[i] >= rstart && idxm[i] < rend) {
591: row = idxm[i] - rstart;
592: for (j=0; j<n; j++) {
593: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
594: if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
595: if (idxn[j] >= cstart && idxn[j] < cend){
596: col = idxn[j] - cstart;
597: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
598: } else {
599: if (!aij->colmap) {
600: MatCreateColmap_MPIAIJ_Private(mat);
601: }
602: #if defined (PETSC_USE_CTABLE)
603: PetscTableFind(aij->colmap,idxn[j]+1,&col);
604: col --;
605: #else
606: col = aij->colmap[idxn[j]] - 1;
607: #endif
608: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
609: else {
610: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
611: }
612: }
613: }
614: } else {
615: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
616: }
617: }
618: return(0);
619: }
621: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);
625: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
626: {
627: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
629: PetscInt nstash,reallocs;
630: InsertMode addv;
633: if (aij->donotstash || mat->nooffprocentries) {
634: return(0);
635: }
637: /* make sure all processors are either in INSERTMODE or ADDMODE */
638: MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,((PetscObject)mat)->comm);
639: if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
640: mat->insertmode = addv; /* in case this processor had no cache */
642: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
643: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
644: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
645: return(0);
646: }
650: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
651: {
652: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
653: Mat_SeqAIJ *a=(Mat_SeqAIJ *)aij->A->data;
655: PetscMPIInt n;
656: PetscInt i,j,rstart,ncols,flg;
657: PetscInt *row,*col;
658: PetscBool other_disassembled;
659: PetscScalar *val;
660: InsertMode addv = mat->insertmode;
662: /* do not use 'b = (Mat_SeqAIJ *)aij->B->data' as B can be reset in disassembly */
664: if (!aij->donotstash && !mat->nooffprocentries) {
665: while (1) {
666: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
667: if (!flg) break;
669: for (i=0; i<n;) {
670: /* Now identify the consecutive vals belonging to the same row */
671: for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
672: if (j < n) ncols = j-i;
673: else ncols = n-i;
674: /* Now assemble all these values with a single function call */
675: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
676: i = j;
677: }
678: }
679: MatStashScatterEnd_Private(&mat->stash);
680: }
681: MatAssemblyBegin(aij->A,mode);
682: MatAssemblyEnd(aij->A,mode);
684: /* determine if any processor has disassembled, if so we must
685: also disassemble ourselfs, in order that we may reassemble. */
686: /*
687: if nonzero structure of submatrix B cannot change then we know that
688: no processor disassembled thus we can skip this stuff
689: */
690: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
691: MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,((PetscObject)mat)->comm);
692: if (mat->was_assembled && !other_disassembled) {
693: MatDisAssemble_MPIAIJ(mat);
694: }
695: }
696: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
697: MatSetUpMultiply_MPIAIJ(mat);
698: }
699: MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
700: MatSetOption(aij->B,MAT_CHECK_COMPRESSED_ROW,PETSC_FALSE);
701: MatAssemblyBegin(aij->B,mode);
702: MatAssemblyEnd(aij->B,mode);
704: PetscFree2(aij->rowvalues,aij->rowindices);
705: aij->rowvalues = 0;
707: /* used by MatAXPY() */
708: a->xtoy = 0; ((Mat_SeqAIJ *)aij->B->data)->xtoy = 0; /* b->xtoy = 0 */
709: a->XtoY = 0; ((Mat_SeqAIJ *)aij->B->data)->XtoY = 0; /* b->XtoY = 0 */
711: VecDestroy(&aij->diag);
712: if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
713: return(0);
714: }
718: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
719: {
720: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
724: MatZeroEntries(l->A);
725: MatZeroEntries(l->B);
726: return(0);
727: }
731: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
732: {
733: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
734: PetscErrorCode ierr;
735: PetscMPIInt size = l->size,imdex,n,rank = l->rank,tag = ((PetscObject)A)->tag,lastidx = -1;
736: PetscInt i,*owners = A->rmap->range;
737: PetscInt *nprocs,j,idx,nsends,row;
738: PetscInt nmax,*svalues,*starts,*owner,nrecvs;
739: PetscInt *rvalues,count,base,slen,*source;
740: PetscInt *lens,*lrows,*values,rstart=A->rmap->rstart;
741: MPI_Comm comm = ((PetscObject)A)->comm;
742: MPI_Request *send_waits,*recv_waits;
743: MPI_Status recv_status,*send_status;
744: const PetscScalar *xx;
745: PetscScalar *bb;
746: #if defined(PETSC_DEBUG)
747: PetscBool found = PETSC_FALSE;
748: #endif
751: /* first count number of contributors to each processor */
752: PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
753: PetscMemzero(nprocs,2*size*sizeof(PetscInt));
754: PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
755: j = 0;
756: for (i=0; i<N; i++) {
757: if (lastidx > (idx = rows[i])) j = 0;
758: lastidx = idx;
759: for (; j<size; j++) {
760: if (idx >= owners[j] && idx < owners[j+1]) {
761: nprocs[2*j]++;
762: nprocs[2*j+1] = 1;
763: owner[i] = j;
764: #if defined(PETSC_DEBUG)
765: found = PETSC_TRUE;
766: #endif
767: break;
768: }
769: }
770: #if defined(PETSC_DEBUG)
771: if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
772: found = PETSC_FALSE;
773: #endif
774: }
775: nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}
777: if (A->nooffproczerorows) {
778: if (nsends > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"You called MatSetOption(,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE) but set an off process zero row");
779: nrecvs = nsends;
780: nmax = N;
781: } else {
782: /* inform other processors of number of messages and max length*/
783: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
784: }
786: /* post receives: */
787: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
788: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
789: for (i=0; i<nrecvs; i++) {
790: MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
791: }
793: /* do sends:
794: 1) starts[i] gives the starting index in svalues for stuff going to
795: the ith processor
796: */
797: PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
798: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
799: PetscMalloc((size+1)*sizeof(PetscInt),&starts);
800: starts[0] = 0;
801: for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
802: for (i=0; i<N; i++) {
803: svalues[starts[owner[i]]++] = rows[i];
804: }
806: starts[0] = 0;
807: for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
808: count = 0;
809: for (i=0; i<size; i++) {
810: if (nprocs[2*i+1]) {
811: MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
812: }
813: }
814: PetscFree(starts);
816: base = owners[rank];
818: /* wait on receives */
819: PetscMalloc2(nrecvs,PetscInt,&lens,nrecvs,PetscInt,&source);
820: count = nrecvs; slen = 0;
821: while (count) {
822: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
823: /* unpack receives into our local space */
824: MPI_Get_count(&recv_status,MPIU_INT,&n);
825: source[imdex] = recv_status.MPI_SOURCE;
826: lens[imdex] = n;
827: slen += n;
828: count--;
829: }
830: PetscFree(recv_waits);
831:
832: /* move the data into the send scatter */
833: PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
834: count = 0;
835: for (i=0; i<nrecvs; i++) {
836: values = rvalues + i*nmax;
837: for (j=0; j<lens[i]; j++) {
838: lrows[count++] = values[j] - base;
839: }
840: }
841: PetscFree(rvalues);
842: PetscFree2(lens,source);
843: PetscFree(owner);
844: PetscFree(nprocs);
845:
846: /* fix right hand side if needed */
847: if (x && b) {
848: VecGetArrayRead(x,&xx);
849: VecGetArray(b,&bb);
850: for (i=0; i<slen; i++) {
851: bb[lrows[i]] = diag*xx[lrows[i]];
852: }
853: VecRestoreArrayRead(x,&xx);
854: VecRestoreArray(b,&bb);
855: }
856: /*
857: Zero the required rows. If the "diagonal block" of the matrix
858: is square and the user wishes to set the diagonal we use separate
859: code so that MatSetValues() is not called for each diagonal allocating
860: new memory, thus calling lots of mallocs and slowing things down.
862: */
863: /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
864: MatZeroRows(l->B,slen,lrows,0.0,0,0);
865: if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) {
866: MatZeroRows(l->A,slen,lrows,diag,0,0);
867: } else if (diag != 0.0) {
868: MatZeroRows(l->A,slen,lrows,0.0,0,0);
869: if (((Mat_SeqAIJ*)l->A->data)->nonew) {
870: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options\n\
871: MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
872: }
873: for (i = 0; i < slen; i++) {
874: row = lrows[i] + rstart;
875: MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
876: }
877: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
878: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
879: } else {
880: MatZeroRows(l->A,slen,lrows,0.0,0,0);
881: }
882: PetscFree(lrows);
884: /* wait on sends */
885: if (nsends) {
886: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
887: MPI_Waitall(nsends,send_waits,send_status);
888: PetscFree(send_status);
889: }
890: PetscFree(send_waits);
891: PetscFree(svalues);
892: return(0);
893: }
897: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
898: {
899: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
900: PetscErrorCode ierr;
901: PetscMPIInt size = l->size,imdex,n,rank = l->rank,tag = ((PetscObject)A)->tag,lastidx = -1;
902: PetscInt i,*owners = A->rmap->range;
903: PetscInt *nprocs,j,idx,nsends;
904: PetscInt nmax,*svalues,*starts,*owner,nrecvs;
905: PetscInt *rvalues,count,base,slen,*source;
906: PetscInt *lens,*lrows,*values,m;
907: MPI_Comm comm = ((PetscObject)A)->comm;
908: MPI_Request *send_waits,*recv_waits;
909: MPI_Status recv_status,*send_status;
910: const PetscScalar *xx;
911: PetscScalar *bb,*mask;
912: Vec xmask,lmask;
913: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data;
914: const PetscInt *aj, *ii,*ridx;
915: PetscScalar *aa;
916: #if defined(PETSC_DEBUG)
917: PetscBool found = PETSC_FALSE;
918: #endif
921: /* first count number of contributors to each processor */
922: PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
923: PetscMemzero(nprocs,2*size*sizeof(PetscInt));
924: PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
925: j = 0;
926: for (i=0; i<N; i++) {
927: if (lastidx > (idx = rows[i])) j = 0;
928: lastidx = idx;
929: for (; j<size; j++) {
930: if (idx >= owners[j] && idx < owners[j+1]) {
931: nprocs[2*j]++;
932: nprocs[2*j+1] = 1;
933: owner[i] = j;
934: #if defined(PETSC_DEBUG)
935: found = PETSC_TRUE;
936: #endif
937: break;
938: }
939: }
940: #if defined(PETSC_DEBUG)
941: if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
942: found = PETSC_FALSE;
943: #endif
944: }
945: nsends = 0; for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}
947: /* inform other processors of number of messages and max length*/
948: PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
950: /* post receives: */
951: PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
952: PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
953: for (i=0; i<nrecvs; i++) {
954: MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
955: }
957: /* do sends:
958: 1) starts[i] gives the starting index in svalues for stuff going to
959: the ith processor
960: */
961: PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
962: PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
963: PetscMalloc((size+1)*sizeof(PetscInt),&starts);
964: starts[0] = 0;
965: for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
966: for (i=0; i<N; i++) {
967: svalues[starts[owner[i]]++] = rows[i];
968: }
970: starts[0] = 0;
971: for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
972: count = 0;
973: for (i=0; i<size; i++) {
974: if (nprocs[2*i+1]) {
975: MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
976: }
977: }
978: PetscFree(starts);
980: base = owners[rank];
982: /* wait on receives */
983: PetscMalloc2(nrecvs,PetscInt,&lens,nrecvs,PetscInt,&source);
984: count = nrecvs; slen = 0;
985: while (count) {
986: MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
987: /* unpack receives into our local space */
988: MPI_Get_count(&recv_status,MPIU_INT,&n);
989: source[imdex] = recv_status.MPI_SOURCE;
990: lens[imdex] = n;
991: slen += n;
992: count--;
993: }
994: PetscFree(recv_waits);
995:
996: /* move the data into the send scatter */
997: PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
998: count = 0;
999: for (i=0; i<nrecvs; i++) {
1000: values = rvalues + i*nmax;
1001: for (j=0; j<lens[i]; j++) {
1002: lrows[count++] = values[j] - base;
1003: }
1004: }
1005: PetscFree(rvalues);
1006: PetscFree2(lens,source);
1007: PetscFree(owner);
1008: PetscFree(nprocs);
1009: /* lrows are the local rows to be zeroed, slen is the number of local rows */
1011: /* zero diagonal part of matrix */
1012: MatZeroRowsColumns(l->A,slen,lrows,diag,x,b);
1013:
1014: /* handle off diagonal part of matrix */
1015: MatGetVecs(A,&xmask,PETSC_NULL);
1016: VecDuplicate(l->lvec,&lmask);
1017: VecGetArray(xmask,&bb);
1018: for (i=0; i<slen; i++) {
1019: bb[lrows[i]] = 1;
1020: }
1021: VecRestoreArray(xmask,&bb);
1022: VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1023: VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1024: VecDestroy(&xmask);
1025: if (x) {
1026: VecScatterBegin(l->Mvctx,x,l->lvec,ADD_VALUES,SCATTER_FORWARD);
1027: VecScatterEnd(l->Mvctx,x,l->lvec,ADD_VALUES,SCATTER_FORWARD);
1028: VecGetArrayRead(l->lvec,&xx);
1029: VecGetArray(b,&bb);
1030: }
1031: VecGetArray(lmask,&mask);
1033: /* remove zeroed rows of off diagonal matrix */
1034: ii = aij->i;
1035: for (i=0; i<slen; i++) {
1036: PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));
1037: }
1039: /* loop over all elements of off process part of matrix zeroing removed columns*/
1040: if (aij->compressedrow.use){
1041: m = aij->compressedrow.nrows;
1042: ii = aij->compressedrow.i;
1043: ridx = aij->compressedrow.rindex;
1044: for (i=0; i<m; i++){
1045: n = ii[i+1] - ii[i];
1046: aj = aij->j + ii[i];
1047: aa = aij->a + ii[i];
1049: for (j=0; j<n; j++) {
1050: if (PetscAbsScalar(mask[*aj])) {
1051: if (b) bb[*ridx] -= *aa*xx[*aj];
1052: *aa = 0.0;
1053: }
1054: aa++;
1055: aj++;
1056: }
1057: ridx++;
1058: }
1059: } else { /* do not use compressed row format */
1060: m = l->B->rmap->n;
1061: for (i=0; i<m; i++) {
1062: n = ii[i+1] - ii[i];
1063: aj = aij->j + ii[i];
1064: aa = aij->a + ii[i];
1065: for (j=0; j<n; j++) {
1066: if (PetscAbsScalar(mask[*aj])) {
1067: if (b) bb[i] -= *aa*xx[*aj];
1068: *aa = 0.0;
1069: }
1070: aa++;
1071: aj++;
1072: }
1073: }
1074: }
1075: if (x) {
1076: VecRestoreArray(b,&bb);
1077: VecRestoreArrayRead(l->lvec,&xx);
1078: }
1079: VecRestoreArray(lmask,&mask);
1080: VecDestroy(&lmask);
1081: PetscFree(lrows);
1083: /* wait on sends */
1084: if (nsends) {
1085: PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
1086: MPI_Waitall(nsends,send_waits,send_status);
1087: PetscFree(send_status);
1088: }
1089: PetscFree(send_waits);
1090: PetscFree(svalues);
1092: return(0);
1093: }
1097: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1098: {
1099: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1101: PetscInt nt;
1104: VecGetLocalSize(xx,&nt);
1105: if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
1106: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1107: (*a->A->ops->mult)(a->A,xx,yy);
1108: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1109: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1110: return(0);
1111: }
1115: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1116: {
1117: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1121: MatMultDiagonalBlock(a->A,bb,xx);
1122: return(0);
1123: }
1127: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1128: {
1129: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1133: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1134: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1135: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1136: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1137: return(0);
1138: }
1142: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1143: {
1144: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1146: PetscBool merged;
1149: VecScatterGetMerged(a->Mvctx,&merged);
1150: /* do nondiagonal part */
1151: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1152: if (!merged) {
1153: /* send it on its way */
1154: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1155: /* do local part */
1156: (*a->A->ops->multtranspose)(a->A,xx,yy);
1157: /* receive remote parts: note this assumes the values are not actually */
1158: /* added in yy until the next line, */
1159: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1160: } else {
1161: /* do local part */
1162: (*a->A->ops->multtranspose)(a->A,xx,yy);
1163: /* send it on its way */
1164: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1165: /* values actually were received in the Begin() but we need to call this nop */
1166: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1167: }
1168: return(0);
1169: }
1171: EXTERN_C_BEGIN
1174: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f)
1175: {
1176: MPI_Comm comm;
1177: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *) Amat->data, *Bij;
1178: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1179: IS Me,Notme;
1181: PetscInt M,N,first,last,*notme,i;
1182: PetscMPIInt size;
1186: /* Easy test: symmetric diagonal block */
1187: Bij = (Mat_MPIAIJ *) Bmat->data; Bdia = Bij->A;
1188: MatIsTranspose(Adia,Bdia,tol,f);
1189: if (!*f) return(0);
1190: PetscObjectGetComm((PetscObject)Amat,&comm);
1191: MPI_Comm_size(comm,&size);
1192: if (size == 1) return(0);
1194: /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
1195: MatGetSize(Amat,&M,&N);
1196: MatGetOwnershipRange(Amat,&first,&last);
1197: PetscMalloc((N-last+first)*sizeof(PetscInt),¬me);
1198: for (i=0; i<first; i++) notme[i] = i;
1199: for (i=last; i<M; i++) notme[i-last+first] = i;
1200: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1201: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1202: MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1203: Aoff = Aoffs[0];
1204: MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1205: Boff = Boffs[0];
1206: MatIsTranspose(Aoff,Boff,tol,f);
1207: MatDestroyMatrices(1,&Aoffs);
1208: MatDestroyMatrices(1,&Boffs);
1209: ISDestroy(&Me);
1210: ISDestroy(&Notme);
1211: PetscFree(notme);
1212: return(0);
1213: }
1214: EXTERN_C_END
1218: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1219: {
1220: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1224: /* do nondiagonal part */
1225: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1226: /* send it on its way */
1227: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1228: /* do local part */
1229: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1230: /* receive remote parts */
1231: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1232: return(0);
1233: }
1235: /*
1236: This only works correctly for square matrices where the subblock A->A is the
1237: diagonal block
1238: */
1241: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1242: {
1244: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1247: if (A->rmap->N != A->cmap->N) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1248: if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1249: MatGetDiagonal(a->A,v);
1250: return(0);
1251: }
1255: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1256: {
1257: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1261: MatScale(a->A,aa);
1262: MatScale(a->B,aa);
1263: return(0);
1264: }
1268: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1269: {
1270: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1274: #if defined(PETSC_USE_LOG)
1275: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1276: #endif
1277: MatStashDestroy_Private(&mat->stash);
1278: VecDestroy(&aij->diag);
1279: MatDestroy(&aij->A);
1280: MatDestroy(&aij->B);
1281: #if defined (PETSC_USE_CTABLE)
1282: PetscTableDestroy(&aij->colmap);
1283: #else
1284: PetscFree(aij->colmap);
1285: #endif
1286: PetscFree(aij->garray);
1287: VecDestroy(&aij->lvec);
1288: VecScatterDestroy(&aij->Mvctx);
1289: PetscFree2(aij->rowvalues,aij->rowindices);
1290: PetscFree(aij->ld);
1291: PetscFree(mat->data);
1293: PetscObjectChangeTypeName((PetscObject)mat,0);
1294: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
1295: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
1296: PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
1297: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C","",PETSC_NULL);
1298: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C","",PETSC_NULL);
1299: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C","",PETSC_NULL);
1300: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);
1301: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C","",PETSC_NULL);
1302: return(0);
1303: }
1307: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1308: {
1309: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1310: Mat_SeqAIJ* A = (Mat_SeqAIJ*)aij->A->data;
1311: Mat_SeqAIJ* B = (Mat_SeqAIJ*)aij->B->data;
1312: PetscErrorCode ierr;
1313: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
1314: int fd;
1315: PetscInt nz,header[4],*row_lengths,*range=0,rlen,i;
1316: PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz;
1317: PetscScalar *column_values;
1318: PetscInt message_count,flowcontrolcount;
1321: MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
1322: MPI_Comm_size(((PetscObject)mat)->comm,&size);
1323: nz = A->nz + B->nz;
1324: if (!rank) {
1325: header[0] = MAT_FILE_CLASSID;
1326: header[1] = mat->rmap->N;
1327: header[2] = mat->cmap->N;
1328: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
1329: PetscViewerBinaryGetDescriptor(viewer,&fd);
1330: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1331: /* get largest number of rows any processor has */
1332: rlen = mat->rmap->n;
1333: range = mat->rmap->range;
1334: for (i=1; i<size; i++) {
1335: rlen = PetscMax(rlen,range[i+1] - range[i]);
1336: }
1337: } else {
1338: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
1339: rlen = mat->rmap->n;
1340: }
1342: /* load up the local row counts */
1343: PetscMalloc((rlen+1)*sizeof(PetscInt),&row_lengths);
1344: for (i=0; i<mat->rmap->n; i++) {
1345: row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1346: }
1348: /* store the row lengths to the file */
1349: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1350: if (!rank) {
1351: PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1352: for (i=1; i<size; i++) {
1353: PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1354: rlen = range[i+1] - range[i];
1355: MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,((PetscObject)mat)->comm);
1356: PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1357: }
1358: PetscViewerFlowControlEndMaster(viewer,message_count);
1359: } else {
1360: PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1361: MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1362: PetscViewerFlowControlEndWorker(viewer,message_count);
1363: }
1364: PetscFree(row_lengths);
1366: /* load up the local column indices */
1367: nzmax = nz; /* )th processor needs space a largest processor needs */
1368: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,((PetscObject)mat)->comm);
1369: PetscMalloc((nzmax+1)*sizeof(PetscInt),&column_indices);
1370: cnt = 0;
1371: for (i=0; i<mat->rmap->n; i++) {
1372: for (j=B->i[i]; j<B->i[i+1]; j++) {
1373: if ( (col = garray[B->j[j]]) > cstart) break;
1374: column_indices[cnt++] = col;
1375: }
1376: for (k=A->i[i]; k<A->i[i+1]; k++) {
1377: column_indices[cnt++] = A->j[k] + cstart;
1378: }
1379: for (; j<B->i[i+1]; j++) {
1380: column_indices[cnt++] = garray[B->j[j]];
1381: }
1382: }
1383: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1385: /* store the column indices to the file */
1386: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1387: if (!rank) {
1388: MPI_Status status;
1389: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1390: for (i=1; i<size; i++) {
1391: PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1392: MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
1393: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1394: MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,((PetscObject)mat)->comm);
1395: PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1396: }
1397: PetscViewerFlowControlEndMaster(viewer,message_count);
1398: } else {
1399: PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1400: MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1401: MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1402: PetscViewerFlowControlEndWorker(viewer,message_count);
1403: }
1404: PetscFree(column_indices);
1406: /* load up the local column values */
1407: PetscMalloc((nzmax+1)*sizeof(PetscScalar),&column_values);
1408: cnt = 0;
1409: for (i=0; i<mat->rmap->n; i++) {
1410: for (j=B->i[i]; j<B->i[i+1]; j++) {
1411: if ( garray[B->j[j]] > cstart) break;
1412: column_values[cnt++] = B->a[j];
1413: }
1414: for (k=A->i[i]; k<A->i[i+1]; k++) {
1415: column_values[cnt++] = A->a[k];
1416: }
1417: for (; j<B->i[i+1]; j++) {
1418: column_values[cnt++] = B->a[j];
1419: }
1420: }
1421: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1423: /* store the column values to the file */
1424: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1425: if (!rank) {
1426: MPI_Status status;
1427: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1428: for (i=1; i<size; i++) {
1429: PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1430: MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
1431: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1432: MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,((PetscObject)mat)->comm);
1433: PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1434: }
1435: PetscViewerFlowControlEndMaster(viewer,message_count);
1436: } else {
1437: PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1438: MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1439: MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,((PetscObject)mat)->comm);
1440: PetscViewerFlowControlEndWorker(viewer,message_count);
1441: }
1442: PetscFree(column_values);
1443: return(0);
1444: }
1448: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1449: {
1450: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1451: PetscErrorCode ierr;
1452: PetscMPIInt rank = aij->rank,size = aij->size;
1453: PetscBool isdraw,iascii,isbinary;
1454: PetscViewer sviewer;
1455: PetscViewerFormat format;
1458: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1459: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1460: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1461: if (iascii) {
1462: PetscViewerGetFormat(viewer,&format);
1463: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1464: MatInfo info;
1465: PetscBool inodes;
1467: MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
1468: MatGetInfo(mat,MAT_LOCAL,&info);
1469: MatInodeGetInodeSizes(aij->A,PETSC_NULL,(PetscInt **)&inodes,PETSC_NULL);
1470: PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
1471: if (!inodes) {
1472: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
1473: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1474: } else {
1475: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
1476: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1477: }
1478: MatGetInfo(aij->A,MAT_LOCAL,&info);
1479: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1480: MatGetInfo(aij->B,MAT_LOCAL,&info);
1481: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1482: PetscViewerFlush(viewer);
1483: PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
1484: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1485: VecScatterView(aij->Mvctx,viewer);
1486: return(0);
1487: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1488: PetscInt inodecount,inodelimit,*inodes;
1489: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1490: if (inodes) {
1491: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1492: } else {
1493: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1494: }
1495: return(0);
1496: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1497: return(0);
1498: }
1499: } else if (isbinary) {
1500: if (size == 1) {
1501: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1502: MatView(aij->A,viewer);
1503: } else {
1504: MatView_MPIAIJ_Binary(mat,viewer);
1505: }
1506: return(0);
1507: } else if (isdraw) {
1508: PetscDraw draw;
1509: PetscBool isnull;
1510: PetscViewerDrawGetDraw(viewer,0,&draw);
1511: PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
1512: }
1514: if (size == 1) {
1515: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1516: MatView(aij->A,viewer);
1517: } else {
1518: /* assemble the entire matrix onto first processor. */
1519: Mat A;
1520: Mat_SeqAIJ *Aloc;
1521: PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct;
1522: MatScalar *a;
1524: if (mat->rmap->N > 1024) {
1525: PetscBool flg = PETSC_FALSE;
1527: PetscOptionsGetBool(((PetscObject) mat)->prefix, "-mat_ascii_output_large", &flg,PETSC_NULL);
1528: if (!flg) {
1529: SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_OUTOFRANGE,"ASCII matrix output not allowed for matrices with more than 1024 rows, use binary format instead.\nYou can override this restriction using -mat_ascii_output_large.");
1530: }
1531: }
1533: MatCreate(((PetscObject)mat)->comm,&A);
1534: if (!rank) {
1535: MatSetSizes(A,M,N,M,N);
1536: } else {
1537: MatSetSizes(A,0,0,M,N);
1538: }
1539: /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
1540: MatSetType(A,MATMPIAIJ);
1541: MatMPIAIJSetPreallocation(A,0,PETSC_NULL,0,PETSC_NULL);
1542: MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1543: PetscLogObjectParent(mat,A);
1545: /* copy over the A part */
1546: Aloc = (Mat_SeqAIJ*)aij->A->data;
1547: m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1548: row = mat->rmap->rstart;
1549: for (i=0; i<ai[m]; i++) {aj[i] += mat->cmap->rstart ;}
1550: for (i=0; i<m; i++) {
1551: MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
1552: row++; a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1553: }
1554: aj = Aloc->j;
1555: for (i=0; i<ai[m]; i++) {aj[i] -= mat->cmap->rstart;}
1557: /* copy over the B part */
1558: Aloc = (Mat_SeqAIJ*)aij->B->data;
1559: m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1560: row = mat->rmap->rstart;
1561: PetscMalloc((ai[m]+1)*sizeof(PetscInt),&cols);
1562: ct = cols;
1563: for (i=0; i<ai[m]; i++) {cols[i] = aij->garray[aj[i]];}
1564: for (i=0; i<m; i++) {
1565: MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
1566: row++; a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1567: }
1568: PetscFree(ct);
1569: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1570: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1571: /*
1572: Everyone has to call to draw the matrix since the graphics waits are
1573: synchronized across all processors that share the PetscDraw object
1574: */
1575: PetscViewerGetSingleton(viewer,&sviewer);
1576: if (!rank) {
1577: PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);
1578: /* Set the type name to MATMPIAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqAIJ_ASCII()*/
1579: PetscStrcpy(((PetscObject)((Mat_MPIAIJ*)(A->data))->A)->type_name,MATMPIAIJ);
1580: MatView(((Mat_MPIAIJ*)(A->data))->A,sviewer);
1581: }
1582: PetscViewerRestoreSingleton(viewer,&sviewer);
1583: MatDestroy(&A);
1584: }
1585: return(0);
1586: }
1590: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1591: {
1593: PetscBool iascii,isdraw,issocket,isbinary;
1594:
1596: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1597: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1598: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1599: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1600: if (iascii || isdraw || isbinary || issocket) {
1601: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1602: } else {
1603: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Viewer type %s not supported by MPIAIJ matrices",((PetscObject)viewer)->type_name);
1604: }
1605: return(0);
1606: }
1610: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1611: {
1612: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1614: Vec bb1 = 0;
1615: PetscBool hasop;
1618: if (flag == SOR_APPLY_UPPER) {
1619: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1620: return(0);
1621: }
1623: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1624: VecDuplicate(bb,&bb1);
1625: }
1627: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
1628: if (flag & SOR_ZERO_INITIAL_GUESS) {
1629: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1630: its--;
1631: }
1632:
1633: while (its--) {
1634: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1635: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1637: /* update rhs: bb1 = bb - B*x */
1638: VecScale(mat->lvec,-1.0);
1639: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1641: /* local sweep */
1642: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1643: }
1644: } else if (flag & SOR_LOCAL_FORWARD_SWEEP){
1645: if (flag & SOR_ZERO_INITIAL_GUESS) {
1646: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1647: its--;
1648: }
1649: while (its--) {
1650: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1651: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1653: /* update rhs: bb1 = bb - B*x */
1654: VecScale(mat->lvec,-1.0);
1655: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1657: /* local sweep */
1658: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1659: }
1660: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){
1661: if (flag & SOR_ZERO_INITIAL_GUESS) {
1662: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1663: its--;
1664: }
1665: while (its--) {
1666: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1667: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1669: /* update rhs: bb1 = bb - B*x */
1670: VecScale(mat->lvec,-1.0);
1671: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1673: /* local sweep */
1674: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1675: }
1676: } else if (flag & SOR_EISENSTAT) {
1677: Vec xx1;
1679: VecDuplicate(bb,&xx1);
1680: (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);
1682: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1683: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1684: if (!mat->diag) {
1685: MatGetVecs(matin,&mat->diag,PETSC_NULL);
1686: MatGetDiagonal(matin,mat->diag);
1687: }
1688: MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1689: if (hasop) {
1690: MatMultDiagonalBlock(matin,xx,bb1);
1691: } else {
1692: VecPointwiseMult(bb1,mat->diag,xx);
1693: }
1694: VecAYPX(bb1,(omega-2.0)/omega,bb);
1696: MatMultAdd(mat->B,mat->lvec,bb1,bb1);
1698: /* local sweep */
1699: (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1700: VecAXPY(xx,1.0,xx1);
1701: VecDestroy(&xx1);
1702: } else SETERRQ(((PetscObject)matin)->comm,PETSC_ERR_SUP,"Parallel SOR not supported");
1704: VecDestroy(&bb1);
1705: return(0);
1706: }
1710: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1711: {
1712: MPI_Comm comm;
1713: PetscInt first,local_rowsize,local_colsize;
1714: const PetscInt *rows;
1715: IS crowp,growp,irowp,lrowp,lcolp,icolp;
1719: PetscObjectGetComm((PetscObject)A,&comm);
1720: /* make a collective version of 'rowp', this is to be tolerant of users who pass serial index sets */
1721: ISOnComm(rowp,comm,PETSC_USE_POINTER,&crowp);
1722: /* collect the global row permutation and invert it */
1723: ISAllGather(crowp,&growp);
1724: ISSetPermutation(growp);
1725: ISDestroy(&crowp);
1726: ISInvertPermutation(growp,PETSC_DECIDE,&irowp);
1727: ISDestroy(&growp);
1728: /* get the local target indices */
1729: MatGetOwnershipRange(A,&first,PETSC_NULL);
1730: MatGetLocalSize(A,&local_rowsize,&local_colsize);
1731: ISGetIndices(irowp,&rows);
1732: ISCreateGeneral(PETSC_COMM_SELF,local_rowsize,rows+first,PETSC_COPY_VALUES,&lrowp);
1733: ISRestoreIndices(irowp,&rows);
1734: ISDestroy(&irowp);
1735: /* the column permutation is so much easier;
1736: make a local version of 'colp' and invert it */
1737: ISOnComm(colp,PETSC_COMM_SELF,PETSC_USE_POINTER,&lcolp);
1738: ISSetPermutation(lcolp);
1739: ISInvertPermutation(lcolp,PETSC_DECIDE,&icolp);
1740: ISDestroy(&lcolp);
1741: /* now we just get the submatrix */
1742: MatGetSubMatrix_MPIAIJ_Private(A,lrowp,icolp,local_colsize,MAT_INITIAL_MATRIX,B);
1743: /* clean up */
1744: ISDestroy(&lrowp);
1745: ISDestroy(&icolp);
1746: return(0);
1747: }
1751: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1752: {
1753: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1754: Mat A = mat->A,B = mat->B;
1756: PetscReal isend[5],irecv[5];
1759: info->block_size = 1.0;
1760: MatGetInfo(A,MAT_LOCAL,info);
1761: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1762: isend[3] = info->memory; isend[4] = info->mallocs;
1763: MatGetInfo(B,MAT_LOCAL,info);
1764: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1765: isend[3] += info->memory; isend[4] += info->mallocs;
1766: if (flag == MAT_LOCAL) {
1767: info->nz_used = isend[0];
1768: info->nz_allocated = isend[1];
1769: info->nz_unneeded = isend[2];
1770: info->memory = isend[3];
1771: info->mallocs = isend[4];
1772: } else if (flag == MAT_GLOBAL_MAX) {
1773: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,((PetscObject)matin)->comm);
1774: info->nz_used = irecv[0];
1775: info->nz_allocated = irecv[1];
1776: info->nz_unneeded = irecv[2];
1777: info->memory = irecv[3];
1778: info->mallocs = irecv[4];
1779: } else if (flag == MAT_GLOBAL_SUM) {
1780: MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,((PetscObject)matin)->comm);
1781: info->nz_used = irecv[0];
1782: info->nz_allocated = irecv[1];
1783: info->nz_unneeded = irecv[2];
1784: info->memory = irecv[3];
1785: info->mallocs = irecv[4];
1786: }
1787: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1788: info->fill_ratio_needed = 0;
1789: info->factor_mallocs = 0;
1791: return(0);
1792: }
1796: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1797: {
1798: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1802: switch (op) {
1803: case MAT_NEW_NONZERO_LOCATIONS:
1804: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1805: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1806: case MAT_KEEP_NONZERO_PATTERN:
1807: case MAT_NEW_NONZERO_LOCATION_ERR:
1808: case MAT_USE_INODES:
1809: case MAT_IGNORE_ZERO_ENTRIES:
1810: MatCheckPreallocated(A,1);
1811: MatSetOption(a->A,op,flg);
1812: MatSetOption(a->B,op,flg);
1813: break;
1814: case MAT_ROW_ORIENTED:
1815: a->roworiented = flg;
1816: MatSetOption(a->A,op,flg);
1817: MatSetOption(a->B,op,flg);
1818: break;
1819: case MAT_NEW_DIAGONALS:
1820: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1821: break;
1822: case MAT_IGNORE_OFF_PROC_ENTRIES:
1823: a->donotstash = flg;
1824: break;
1825: case MAT_SPD:
1826: A->spd_set = PETSC_TRUE;
1827: A->spd = flg;
1828: if (flg) {
1829: A->symmetric = PETSC_TRUE;
1830: A->structurally_symmetric = PETSC_TRUE;
1831: A->symmetric_set = PETSC_TRUE;
1832: A->structurally_symmetric_set = PETSC_TRUE;
1833: }
1834: break;
1835: case MAT_SYMMETRIC:
1836: MatSetOption(a->A,op,flg);
1837: break;
1838: case MAT_STRUCTURALLY_SYMMETRIC:
1839: MatSetOption(a->A,op,flg);
1840: break;
1841: case MAT_HERMITIAN:
1842: MatSetOption(a->A,op,flg);
1843: break;
1844: case MAT_SYMMETRY_ETERNAL:
1845: MatSetOption(a->A,op,flg);
1846: break;
1847: default:
1848: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1849: }
1850: return(0);
1851: }
1855: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1856: {
1857: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1858: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1860: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1861: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1862: PetscInt *cmap,*idx_p;
1865: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1866: mat->getrowactive = PETSC_TRUE;
1868: if (!mat->rowvalues && (idx || v)) {
1869: /*
1870: allocate enough space to hold information from the longest row.
1871: */
1872: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1873: PetscInt max = 1,tmp;
1874: for (i=0; i<matin->rmap->n; i++) {
1875: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1876: if (max < tmp) { max = tmp; }
1877: }
1878: PetscMalloc2(max,PetscScalar,&mat->rowvalues,max,PetscInt,&mat->rowindices);
1879: }
1881: if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1882: lrow = row - rstart;
1884: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1885: if (!v) {pvA = 0; pvB = 0;}
1886: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1887: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1888: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1889: nztot = nzA + nzB;
1891: cmap = mat->garray;
1892: if (v || idx) {
1893: if (nztot) {
1894: /* Sort by increasing column numbers, assuming A and B already sorted */
1895: PetscInt imark = -1;
1896: if (v) {
1897: *v = v_p = mat->rowvalues;
1898: for (i=0; i<nzB; i++) {
1899: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1900: else break;
1901: }
1902: imark = i;
1903: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1904: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1905: }
1906: if (idx) {
1907: *idx = idx_p = mat->rowindices;
1908: if (imark > -1) {
1909: for (i=0; i<imark; i++) {
1910: idx_p[i] = cmap[cworkB[i]];
1911: }
1912: } else {
1913: for (i=0; i<nzB; i++) {
1914: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1915: else break;
1916: }
1917: imark = i;
1918: }
1919: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1920: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1921: }
1922: } else {
1923: if (idx) *idx = 0;
1924: if (v) *v = 0;
1925: }
1926: }
1927: *nz = nztot;
1928: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1929: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1930: return(0);
1931: }
1935: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1936: {
1937: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1940: if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1941: aij->getrowactive = PETSC_FALSE;
1942: return(0);
1943: }
1947: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1948: {
1949: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1950: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1952: PetscInt i,j,cstart = mat->cmap->rstart;
1953: PetscReal sum = 0.0;
1954: MatScalar *v;
1957: if (aij->size == 1) {
1958: MatNorm(aij->A,type,norm);
1959: } else {
1960: if (type == NORM_FROBENIUS) {
1961: v = amat->a;
1962: for (i=0; i<amat->nz; i++) {
1963: #if defined(PETSC_USE_COMPLEX)
1964: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1965: #else
1966: sum += (*v)*(*v); v++;
1967: #endif
1968: }
1969: v = bmat->a;
1970: for (i=0; i<bmat->nz; i++) {
1971: #if defined(PETSC_USE_COMPLEX)
1972: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1973: #else
1974: sum += (*v)*(*v); v++;
1975: #endif
1976: }
1977: MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);
1978: *norm = PetscSqrtReal(*norm);
1979: } else if (type == NORM_1) { /* max column norm */
1980: PetscReal *tmp,*tmp2;
1981: PetscInt *jj,*garray = aij->garray;
1982: PetscMalloc((mat->cmap->N+1)*sizeof(PetscReal),&tmp);
1983: PetscMalloc((mat->cmap->N+1)*sizeof(PetscReal),&tmp2);
1984: PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));
1985: *norm = 0.0;
1986: v = amat->a; jj = amat->j;
1987: for (j=0; j<amat->nz; j++) {
1988: tmp[cstart + *jj++ ] += PetscAbsScalar(*v); v++;
1989: }
1990: v = bmat->a; jj = bmat->j;
1991: for (j=0; j<bmat->nz; j++) {
1992: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1993: }
1994: MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);
1995: for (j=0; j<mat->cmap->N; j++) {
1996: if (tmp2[j] > *norm) *norm = tmp2[j];
1997: }
1998: PetscFree(tmp);
1999: PetscFree(tmp2);
2000: } else if (type == NORM_INFINITY) { /* max row norm */
2001: PetscReal ntemp = 0.0;
2002: for (j=0; j<aij->A->rmap->n; j++) {
2003: v = amat->a + amat->i[j];
2004: sum = 0.0;
2005: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
2006: sum += PetscAbsScalar(*v); v++;
2007: }
2008: v = bmat->a + bmat->i[j];
2009: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
2010: sum += PetscAbsScalar(*v); v++;
2011: }
2012: if (sum > ntemp) ntemp = sum;
2013: }
2014: MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,((PetscObject)mat)->comm);
2015: } else {
2016: SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_SUP,"No support for two norm");
2017: }
2018: }
2019: return(0);
2020: }
2024: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
2025: {
2026: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2027: Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data;
2029: PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i,*d_nnz;
2030: PetscInt cstart=A->cmap->rstart,ncol;
2031: Mat B;
2032: MatScalar *array;
2035: if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
2037: ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n;
2038: ai = Aloc->i; aj = Aloc->j;
2039: bi = Bloc->i; bj = Bloc->j;
2040: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2041: /* compute d_nnz for preallocation; o_nnz is approximated by d_nnz to avoid communication */
2042: PetscMalloc((1+na)*sizeof(PetscInt),&d_nnz);
2043: PetscMemzero(d_nnz,(1+na)*sizeof(PetscInt));
2044: for (i=0; i<ai[ma]; i++){
2045: d_nnz[aj[i]] ++;
2046: aj[i] += cstart; /* global col index to be used by MatSetValues() */
2047: }
2049: MatCreate(((PetscObject)A)->comm,&B);
2050: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2051: MatSetBlockSizes(B,A->cmap->bs,A->rmap->bs);
2052: MatSetType(B,((PetscObject)A)->type_name);
2053: MatMPIAIJSetPreallocation(B,0,d_nnz,0,d_nnz);
2054: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
2055: PetscFree(d_nnz);
2056: } else {
2057: B = *matout;
2058: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2059: for (i=0; i<ai[ma]; i++){
2060: aj[i] += cstart; /* global col index to be used by MatSetValues() */
2061: }
2062: }
2064: /* copy over the A part */
2065: array = Aloc->a;
2066: row = A->rmap->rstart;
2067: for (i=0; i<ma; i++) {
2068: ncol = ai[i+1]-ai[i];
2069: MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);
2070: row++; array += ncol; aj += ncol;
2071: }
2072: aj = Aloc->j;
2073: for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */
2075: /* copy over the B part */
2076: PetscMalloc(bi[mb]*sizeof(PetscInt),&cols);
2077: PetscMemzero(cols,bi[mb]*sizeof(PetscInt));
2078: array = Bloc->a;
2079: row = A->rmap->rstart;
2080: for (i=0; i<bi[mb]; i++) {cols[i] = a->garray[bj[i]];}
2081: cols_tmp = cols;
2082: for (i=0; i<mb; i++) {
2083: ncol = bi[i+1]-bi[i];
2084: MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2085: row++; array += ncol; cols_tmp += ncol;
2086: }
2087: PetscFree(cols);
2088:
2089: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2090: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2091: if (reuse == MAT_INITIAL_MATRIX || *matout != A) {
2092: *matout = B;
2093: } else {
2094: MatHeaderMerge(A,B);
2095: }
2096: return(0);
2097: }
2101: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2102: {
2103: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2104: Mat a = aij->A,b = aij->B;
2106: PetscInt s1,s2,s3;
2109: MatGetLocalSize(mat,&s2,&s3);
2110: if (rr) {
2111: VecGetLocalSize(rr,&s1);
2112: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2113: /* Overlap communication with computation. */
2114: VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2115: }
2116: if (ll) {
2117: VecGetLocalSize(ll,&s1);
2118: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2119: (*b->ops->diagonalscale)(b,ll,0);
2120: }
2121: /* scale the diagonal block */
2122: (*a->ops->diagonalscale)(a,ll,rr);
2124: if (rr) {
2125: /* Do a scatter end and then right scale the off-diagonal block */
2126: VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2127: (*b->ops->diagonalscale)(b,0,aij->lvec);
2128: }
2129:
2130: return(0);
2131: }
2135: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2136: {
2137: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2141: MatSetUnfactored(a->A);
2142: return(0);
2143: }
2147: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag)
2148: {
2149: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2150: Mat a,b,c,d;
2151: PetscBool flg;
2155: a = matA->A; b = matA->B;
2156: c = matB->A; d = matB->B;
2158: MatEqual(a,c,&flg);
2159: if (flg) {
2160: MatEqual(b,d,&flg);
2161: }
2162: MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);
2163: return(0);
2164: }
2168: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2169: {
2171: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2172: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2175: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2176: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2177: /* because of the column compression in the off-processor part of the matrix a->B,
2178: the number of columns in a->B and b->B may be different, hence we cannot call
2179: the MatCopy() directly on the two parts. If need be, we can provide a more
2180: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2181: then copying the submatrices */
2182: MatCopy_Basic(A,B,str);
2183: } else {
2184: MatCopy(a->A,b->A,str);
2185: MatCopy(a->B,b->B,str);
2186: }
2187: return(0);
2188: }
2192: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2193: {
2197: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2198: return(0);
2199: }
2203: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2204: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt* nnz)
2205: {
2206: PetscInt i,m=Y->rmap->N;
2207: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2208: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2209: const PetscInt *xi = x->i,*yi = y->i;
2212: /* Set the number of nonzeros in the new matrix */
2213: for(i=0; i<m; i++) {
2214: PetscInt j,k,nzx = xi[i+1] - xi[i],nzy = yi[i+1] - yi[i];
2215: const PetscInt *xj = x->j+xi[i],*yj = y->j+yi[i];
2216: nnz[i] = 0;
2217: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2218: for (; k<nzy && yltog[yj[k]]<xltog[xj[j]]; k++) nnz[i]++; /* Catch up to X */
2219: if (k<nzy && yltog[yj[k]]==xltog[xj[j]]) k++; /* Skip duplicate */
2220: nnz[i]++;
2221: }
2222: for (; k<nzy; k++) nnz[i]++;
2223: }
2224: return(0);
2225: }
2229: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2230: {
2232: PetscInt i;
2233: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data,*yy = (Mat_MPIAIJ *)Y->data;
2234: PetscBLASInt bnz,one=1;
2235: Mat_SeqAIJ *x,*y;
2238: if (str == SAME_NONZERO_PATTERN) {
2239: PetscScalar alpha = a;
2240: x = (Mat_SeqAIJ *)xx->A->data;
2241: y = (Mat_SeqAIJ *)yy->A->data;
2242: bnz = PetscBLASIntCast(x->nz);
2243: BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
2244: x = (Mat_SeqAIJ *)xx->B->data;
2245: y = (Mat_SeqAIJ *)yy->B->data;
2246: bnz = PetscBLASIntCast(x->nz);
2247: BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
2248: } else if (str == SUBSET_NONZERO_PATTERN) {
2249: MatAXPY_SeqAIJ(yy->A,a,xx->A,str);
2251: x = (Mat_SeqAIJ *)xx->B->data;
2252: y = (Mat_SeqAIJ *)yy->B->data;
2253: if (y->xtoy && y->XtoY != xx->B) {
2254: PetscFree(y->xtoy);
2255: MatDestroy(&y->XtoY);
2256: }
2257: if (!y->xtoy) { /* get xtoy */
2258: MatAXPYGetxtoy_Private(xx->B->rmap->n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);
2259: y->XtoY = xx->B;
2260: PetscObjectReference((PetscObject)xx->B);
2261: }
2262: for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]);
2263: } else {
2264: Mat B;
2265: PetscInt *nnz_d,*nnz_o;
2266: PetscMalloc(yy->A->rmap->N*sizeof(PetscInt),&nnz_d);
2267: PetscMalloc(yy->B->rmap->N*sizeof(PetscInt),&nnz_o);
2268: MatCreate(((PetscObject)Y)->comm,&B);
2269: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2270: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2271: MatSetBlockSizes(B,Y->rmap->bs,Y->cmap->bs);
2272: MatSetType(B,MATMPIAIJ);
2273: MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2274: MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2275: MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2276: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2277: MatHeaderReplace(Y,B);
2278: PetscFree(nnz_d);
2279: PetscFree(nnz_o);
2280: }
2281: return(0);
2282: }
2284: extern PetscErrorCode MatConjugate_SeqAIJ(Mat);
2288: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2289: {
2290: #if defined(PETSC_USE_COMPLEX)
2292: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2295: MatConjugate_SeqAIJ(aij->A);
2296: MatConjugate_SeqAIJ(aij->B);
2297: #else
2299: #endif
2300: return(0);
2301: }
2305: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2306: {
2307: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2311: MatRealPart(a->A);
2312: MatRealPart(a->B);
2313: return(0);
2314: }
2318: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2319: {
2320: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2324: MatImaginaryPart(a->A);
2325: MatImaginaryPart(a->B);
2326: return(0);
2327: }
2329: #ifdef PETSC_HAVE_PBGL
2331: #include <boost/parallel/mpi/bsp_process_group.hpp>
2332: #include <boost/graph/distributed/ilu_default_graph.hpp>
2333: #include <boost/graph/distributed/ilu_0_block.hpp>
2334: #include <boost/graph/distributed/ilu_preconditioner.hpp>
2335: #include <boost/graph/distributed/petsc/interface.hpp>
2336: #include <boost/multi_array.hpp>
2337: #include <boost/parallel/distributed_property_map->hpp>
2341: /*
2342: This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
2343: */
2344: PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2345: {
2346: namespace petsc = boost::distributed::petsc;
2347:
2348: namespace graph_dist = boost::graph::distributed;
2349: using boost::graph::distributed::ilu_default::process_group_type;
2350: using boost::graph::ilu_permuted;
2352: PetscBool row_identity, col_identity;
2353: PetscContainer c;
2354: PetscInt m, n, M, N;
2355: PetscErrorCode ierr;
2358: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu");
2359: ISIdentity(isrow, &row_identity);
2360: ISIdentity(iscol, &col_identity);
2361: if (!row_identity || !col_identity) {
2362: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU");
2363: }
2365: process_group_type pg;
2366: typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type;
2367: lgraph_type* lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg));
2368: lgraph_type& level_graph = *lgraph_p;
2369: graph_dist::ilu_default::graph_type& graph(level_graph.graph);
2371: petsc::read_matrix(A, graph, get(boost::edge_weight, graph));
2372: ilu_permuted(level_graph);
2374: /* put together the new matrix */
2375: MatCreate(((PetscObject)A)->comm, fact);
2376: MatGetLocalSize(A, &m, &n);
2377: MatGetSize(A, &M, &N);
2378: MatSetSizes(fact, m, n, M, N);
2379: MatSetBlockSizes(fact,A->rmap->bs,A->cmap->bs);
2380: MatSetType(fact, ((PetscObject)A)->type_name);
2381: MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);
2382: MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);
2384: PetscContainerCreate(((PetscObject)A)->comm, &c);
2385: PetscContainerSetPointer(c, lgraph_p);
2386: PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c);
2387: PetscContainerDestroy(&c);
2388: return(0);
2389: }
2393: PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info)
2394: {
2396: return(0);
2397: }
2401: /*
2402: This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
2403: */
2404: PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x)
2405: {
2406: namespace graph_dist = boost::graph::distributed;
2408: typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type;
2409: lgraph_type* lgraph_p;
2410: PetscContainer c;
2414: PetscObjectQuery((PetscObject) A, "graph", (PetscObject *) &c);
2415: PetscContainerGetPointer(c, (void **) &lgraph_p);
2416: VecCopy(b, x);
2418: PetscScalar* array_x;
2419: VecGetArray(x, &array_x);
2420: PetscInt sx;
2421: VecGetSize(x, &sx);
2422:
2423: PetscScalar* array_b;
2424: VecGetArray(b, &array_b);
2425: PetscInt sb;
2426: VecGetSize(b, &sb);
2428: lgraph_type& level_graph = *lgraph_p;
2429: graph_dist::ilu_default::graph_type& graph(level_graph.graph);
2431: typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type;
2432: array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]),
2433: ref_x(array_x, boost::extents[num_vertices(graph)]);
2435: typedef boost::iterator_property_map<array_ref_type::iterator,
2436: boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type;
2437: gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)),
2438: vector_x(ref_x.begin(), get(boost::vertex_index, graph));
2439:
2440: ilu_set_solve(*lgraph_p, vector_b, vector_x);
2442: return(0);
2443: }
2444: #endif
2446: typedef struct { /* used by MatGetRedundantMatrix() for reusing matredundant */
2447: PetscInt nzlocal,nsends,nrecvs;
2448: PetscMPIInt *send_rank,*recv_rank;
2449: PetscInt *sbuf_nz,*rbuf_nz,*sbuf_j,**rbuf_j;
2450: PetscScalar *sbuf_a,**rbuf_a;
2451: PetscErrorCode (*Destroy)(Mat);
2452: } Mat_Redundant;
2456: PetscErrorCode PetscContainerDestroy_MatRedundant(void *ptr)
2457: {
2458: PetscErrorCode ierr;
2459: Mat_Redundant *redund=(Mat_Redundant*)ptr;
2460: PetscInt i;
2463: PetscFree2(redund->send_rank,redund->recv_rank);
2464: PetscFree(redund->sbuf_j);
2465: PetscFree(redund->sbuf_a);
2466: for (i=0; i<redund->nrecvs; i++){
2467: PetscFree(redund->rbuf_j[i]);
2468: PetscFree(redund->rbuf_a[i]);
2469: }
2470: PetscFree4(redund->sbuf_nz,redund->rbuf_nz,redund->rbuf_j,redund->rbuf_a);
2471: PetscFree(redund);
2472: return(0);
2473: }
2477: PetscErrorCode MatDestroy_MatRedundant(Mat A)
2478: {
2479: PetscErrorCode ierr;
2480: PetscContainer container;
2481: Mat_Redundant *redund=PETSC_NULL;
2484: PetscObjectQuery((PetscObject)A,"Mat_Redundant",(PetscObject *)&container);
2485: if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit");
2486: PetscContainerGetPointer(container,(void **)&redund);
2487: A->ops->destroy = redund->Destroy;
2488: PetscObjectCompose((PetscObject)A,"Mat_Redundant",0);
2489: if (A->ops->destroy) {
2490: (*A->ops->destroy)(A);
2491: }
2492: return(0);
2493: }
2497: PetscErrorCode MatGetRedundantMatrix_MPIAIJ(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,PetscInt mlocal_sub,MatReuse reuse,Mat *matredundant)
2498: {
2499: PetscMPIInt rank,size;
2500: MPI_Comm comm=((PetscObject)mat)->comm;
2502: PetscInt nsends=0,nrecvs=0,i,rownz_max=0;
2503: PetscMPIInt *send_rank=PETSC_NULL,*recv_rank=PETSC_NULL;
2504: PetscInt *rowrange=mat->rmap->range;
2505: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2506: Mat A=aij->A,B=aij->B,C=*matredundant;
2507: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data;
2508: PetscScalar *sbuf_a;
2509: PetscInt nzlocal=a->nz+b->nz;
2510: PetscInt j,cstart=mat->cmap->rstart,cend=mat->cmap->rend,row,nzA,nzB,ncols,*cworkA,*cworkB;
2511: PetscInt rstart=mat->rmap->rstart,rend=mat->rmap->rend,*bmap=aij->garray,M,N;
2512: PetscInt *cols,ctmp,lwrite,*rptr,l,*sbuf_j;
2513: MatScalar *aworkA,*aworkB;
2514: PetscScalar *vals;
2515: PetscMPIInt tag1,tag2,tag3,imdex;
2516: MPI_Request *s_waits1=PETSC_NULL,*s_waits2=PETSC_NULL,*s_waits3=PETSC_NULL,
2517: *r_waits1=PETSC_NULL,*r_waits2=PETSC_NULL,*r_waits3=PETSC_NULL;
2518: MPI_Status recv_status,*send_status;
2519: PetscInt *sbuf_nz=PETSC_NULL,*rbuf_nz=PETSC_NULL,count;
2520: PetscInt **rbuf_j=PETSC_NULL;
2521: PetscScalar **rbuf_a=PETSC_NULL;
2522: Mat_Redundant *redund=PETSC_NULL;
2523: PetscContainer container;
2526: MPI_Comm_rank(comm,&rank);
2527: MPI_Comm_size(comm,&size);
2529: if (reuse == MAT_REUSE_MATRIX) {
2530: MatGetSize(C,&M,&N);
2531: if (M != N || M != mat->rmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong global size");
2532: MatGetLocalSize(C,&M,&N);
2533: if (M != N || M != mlocal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong local size");
2534: PetscObjectQuery((PetscObject)C,"Mat_Redundant",(PetscObject *)&container);
2535: if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit");
2536: PetscContainerGetPointer(container,(void **)&redund);
2537: if (nzlocal != redund->nzlocal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong nzlocal");
2539: nsends = redund->nsends;
2540: nrecvs = redund->nrecvs;
2541: send_rank = redund->send_rank;
2542: recv_rank = redund->recv_rank;
2543: sbuf_nz = redund->sbuf_nz;
2544: rbuf_nz = redund->rbuf_nz;
2545: sbuf_j = redund->sbuf_j;
2546: sbuf_a = redund->sbuf_a;
2547: rbuf_j = redund->rbuf_j;
2548: rbuf_a = redund->rbuf_a;
2549: }
2551: if (reuse == MAT_INITIAL_MATRIX){
2552: PetscMPIInt subrank,subsize;
2553: PetscInt nleftover,np_subcomm;
2554: /* get the destination processors' id send_rank, nsends and nrecvs */
2555: MPI_Comm_rank(subcomm,&subrank);
2556: MPI_Comm_size(subcomm,&subsize);
2557: PetscMalloc2(size,PetscMPIInt,&send_rank,size,PetscMPIInt,&recv_rank);
2558: np_subcomm = size/nsubcomm;
2559: nleftover = size - nsubcomm*np_subcomm;
2560: nsends = 0; nrecvs = 0;
2561: for (i=0; i<size; i++){ /* i=rank*/
2562: if (subrank == i/nsubcomm && rank != i){ /* my_subrank == other's subrank */
2563: send_rank[nsends] = i; nsends++;
2564: recv_rank[nrecvs++] = i;
2565: }
2566: }
2567: if (rank >= size - nleftover){/* this proc is a leftover processor */
2568: i = size-nleftover-1;
2569: j = 0;
2570: while (j < nsubcomm - nleftover){
2571: send_rank[nsends++] = i;
2572: i--; j++;
2573: }
2574: }
2576: if (nleftover && subsize == size/nsubcomm && subrank==subsize-1){ /* this proc recvs from leftover processors */
2577: for (i=0; i<nleftover; i++){
2578: recv_rank[nrecvs++] = size-nleftover+i;
2579: }
2580: }
2582: /* allocate sbuf_j, sbuf_a */
2583: i = nzlocal + rowrange[rank+1] - rowrange[rank] + 2;
2584: PetscMalloc(i*sizeof(PetscInt),&sbuf_j);
2585: PetscMalloc((nzlocal+1)*sizeof(PetscScalar),&sbuf_a);
2586: } /* endof if (reuse == MAT_INITIAL_MATRIX) */
2588: /* copy mat's local entries into the buffers */
2589: if (reuse == MAT_INITIAL_MATRIX){
2590: rownz_max = 0;
2591: rptr = sbuf_j;
2592: cols = sbuf_j + rend-rstart + 1;
2593: vals = sbuf_a;
2594: rptr[0] = 0;
2595: for (i=0; i<rend-rstart; i++){
2596: row = i + rstart;
2597: nzA = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
2598: ncols = nzA + nzB;
2599: cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
2600: aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
2601: /* load the column indices for this row into cols */
2602: lwrite = 0;
2603: for (l=0; l<nzB; l++) {
2604: if ((ctmp = bmap[cworkB[l]]) < cstart){
2605: vals[lwrite] = aworkB[l];
2606: cols[lwrite++] = ctmp;
2607: }
2608: }
2609: for (l=0; l<nzA; l++){
2610: vals[lwrite] = aworkA[l];
2611: cols[lwrite++] = cstart + cworkA[l];
2612: }
2613: for (l=0; l<nzB; l++) {
2614: if ((ctmp = bmap[cworkB[l]]) >= cend){
2615: vals[lwrite] = aworkB[l];
2616: cols[lwrite++] = ctmp;
2617: }
2618: }
2619: vals += ncols;
2620: cols += ncols;
2621: rptr[i+1] = rptr[i] + ncols;
2622: if (rownz_max < ncols) rownz_max = ncols;
2623: }
2624: if (rptr[rend-rstart] != a->nz + b->nz) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB, "rptr[%d] %d != %d + %d",rend-rstart,rptr[rend-rstart+1],a->nz,b->nz);
2625: } else { /* only copy matrix values into sbuf_a */
2626: rptr = sbuf_j;
2627: vals = sbuf_a;
2628: rptr[0] = 0;
2629: for (i=0; i<rend-rstart; i++){
2630: row = i + rstart;
2631: nzA = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
2632: ncols = nzA + nzB;
2633: cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
2634: aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
2635: lwrite = 0;
2636: for (l=0; l<nzB; l++) {
2637: if ((ctmp = bmap[cworkB[l]]) < cstart) vals[lwrite++] = aworkB[l];
2638: }
2639: for (l=0; l<nzA; l++) vals[lwrite++] = aworkA[l];
2640: for (l=0; l<nzB; l++) {
2641: if ((ctmp = bmap[cworkB[l]]) >= cend) vals[lwrite++] = aworkB[l];
2642: }
2643: vals += ncols;
2644: rptr[i+1] = rptr[i] + ncols;
2645: }
2646: } /* endof if (reuse == MAT_INITIAL_MATRIX) */
2648: /* send nzlocal to others, and recv other's nzlocal */
2649: /*--------------------------------------------------*/
2650: if (reuse == MAT_INITIAL_MATRIX){
2651: PetscMalloc2(3*(nsends + nrecvs)+1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
2652: s_waits2 = s_waits3 + nsends;
2653: s_waits1 = s_waits2 + nsends;
2654: r_waits1 = s_waits1 + nsends;
2655: r_waits2 = r_waits1 + nrecvs;
2656: r_waits3 = r_waits2 + nrecvs;
2657: } else {
2658: PetscMalloc2(nsends + nrecvs +1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
2659: r_waits3 = s_waits3 + nsends;
2660: }
2662: PetscObjectGetNewTag((PetscObject)mat,&tag3);
2663: if (reuse == MAT_INITIAL_MATRIX){
2664: /* get new tags to keep the communication clean */
2665: PetscObjectGetNewTag((PetscObject)mat,&tag1);
2666: PetscObjectGetNewTag((PetscObject)mat,&tag2);
2667: PetscMalloc4(nsends,PetscInt,&sbuf_nz,nrecvs,PetscInt,&rbuf_nz,nrecvs,PetscInt*,&rbuf_j,nrecvs,PetscScalar*,&rbuf_a);
2669: /* post receives of other's nzlocal */
2670: for (i=0; i<nrecvs; i++){
2671: MPI_Irecv(rbuf_nz+i,1,MPIU_INT,MPI_ANY_SOURCE,tag1,comm,r_waits1+i);
2672: }
2673: /* send nzlocal to others */
2674: for (i=0; i<nsends; i++){
2675: sbuf_nz[i] = nzlocal;
2676: MPI_Isend(sbuf_nz+i,1,MPIU_INT,send_rank[i],tag1,comm,s_waits1+i);
2677: }
2678: /* wait on receives of nzlocal; allocate space for rbuf_j, rbuf_a */
2679: count = nrecvs;
2680: while (count) {
2681: MPI_Waitany(nrecvs,r_waits1,&imdex,&recv_status);
2682: recv_rank[imdex] = recv_status.MPI_SOURCE;
2683: /* allocate rbuf_a and rbuf_j; then post receives of rbuf_j */
2684: PetscMalloc((rbuf_nz[imdex]+1)*sizeof(PetscScalar),&rbuf_a[imdex]);
2686: i = rowrange[recv_status.MPI_SOURCE+1] - rowrange[recv_status.MPI_SOURCE]; /* number of expected mat->i */
2687: rbuf_nz[imdex] += i + 2;
2688: PetscMalloc(rbuf_nz[imdex]*sizeof(PetscInt),&rbuf_j[imdex]);
2689: MPI_Irecv(rbuf_j[imdex],rbuf_nz[imdex],MPIU_INT,recv_status.MPI_SOURCE,tag2,comm,r_waits2+imdex);
2690: count--;
2691: }
2692: /* wait on sends of nzlocal */
2693: if (nsends) {MPI_Waitall(nsends,s_waits1,send_status);}
2694: /* send mat->i,j to others, and recv from other's */
2695: /*------------------------------------------------*/
2696: for (i=0; i<nsends; i++){
2697: j = nzlocal + rowrange[rank+1] - rowrange[rank] + 1;
2698: MPI_Isend(sbuf_j,j,MPIU_INT,send_rank[i],tag2,comm,s_waits2+i);
2699: }
2700: /* wait on receives of mat->i,j */
2701: /*------------------------------*/
2702: count = nrecvs;
2703: while (count) {
2704: MPI_Waitany(nrecvs,r_waits2,&imdex,&recv_status);
2705: if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2706: count--;
2707: }
2708: /* wait on sends of mat->i,j */
2709: /*---------------------------*/
2710: if (nsends) {
2711: MPI_Waitall(nsends,s_waits2,send_status);
2712: }
2713: } /* endof if (reuse == MAT_INITIAL_MATRIX) */
2715: /* post receives, send and receive mat->a */
2716: /*----------------------------------------*/
2717: for (imdex=0; imdex<nrecvs; imdex++) {
2718: MPI_Irecv(rbuf_a[imdex],rbuf_nz[imdex],MPIU_SCALAR,recv_rank[imdex],tag3,comm,r_waits3+imdex);
2719: }
2720: for (i=0; i<nsends; i++){
2721: MPI_Isend(sbuf_a,nzlocal,MPIU_SCALAR,send_rank[i],tag3,comm,s_waits3+i);
2722: }
2723: count = nrecvs;
2724: while (count) {
2725: MPI_Waitany(nrecvs,r_waits3,&imdex,&recv_status);
2726: if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2727: count--;
2728: }
2729: if (nsends) {
2730: MPI_Waitall(nsends,s_waits3,send_status);
2731: }
2733: PetscFree2(s_waits3,send_status);
2735: /* create redundant matrix */
2736: /*-------------------------*/
2737: if (reuse == MAT_INITIAL_MATRIX){
2738: /* compute rownz_max for preallocation */
2739: for (imdex=0; imdex<nrecvs; imdex++){
2740: j = rowrange[recv_rank[imdex]+1] - rowrange[recv_rank[imdex]];
2741: rptr = rbuf_j[imdex];
2742: for (i=0; i<j; i++){
2743: ncols = rptr[i+1] - rptr[i];
2744: if (rownz_max < ncols) rownz_max = ncols;
2745: }
2746: }
2748: MatCreate(subcomm,&C);
2749: MatSetSizes(C,mlocal_sub,mlocal_sub,PETSC_DECIDE,PETSC_DECIDE);
2750: MatSetBlockSizes(C,mat->rmap->bs,mat->cmap->bs);
2751: MatSetFromOptions(C);
2752: MatSeqAIJSetPreallocation(C,rownz_max,PETSC_NULL);
2753: MatMPIAIJSetPreallocation(C,rownz_max,PETSC_NULL,rownz_max,PETSC_NULL);
2754: } else {
2755: C = *matredundant;
2756: }
2758: /* insert local matrix entries */
2759: rptr = sbuf_j;
2760: cols = sbuf_j + rend-rstart + 1;
2761: vals = sbuf_a;
2762: for (i=0; i<rend-rstart; i++){
2763: row = i + rstart;
2764: ncols = rptr[i+1] - rptr[i];
2765: MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2766: vals += ncols;
2767: cols += ncols;
2768: }
2769: /* insert received matrix entries */
2770: for (imdex=0; imdex<nrecvs; imdex++){
2771: rstart = rowrange[recv_rank[imdex]];
2772: rend = rowrange[recv_rank[imdex]+1];
2773: rptr = rbuf_j[imdex];
2774: cols = rbuf_j[imdex] + rend-rstart + 1;
2775: vals = rbuf_a[imdex];
2776: for (i=0; i<rend-rstart; i++){
2777: row = i + rstart;
2778: ncols = rptr[i+1] - rptr[i];
2779: MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2780: vals += ncols;
2781: cols += ncols;
2782: }
2783: }
2784: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2785: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2786: MatGetSize(C,&M,&N);
2787: if (M != mat->rmap->N || N != mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"redundant mat size %d != input mat size %d",M,mat->rmap->N);
2788: if (reuse == MAT_INITIAL_MATRIX) {
2789: PetscContainer container;
2790: *matredundant = C;
2791: /* create a supporting struct and attach it to C for reuse */
2792: PetscNewLog(C,Mat_Redundant,&redund);
2793: PetscContainerCreate(PETSC_COMM_SELF,&container);
2794: PetscContainerSetPointer(container,redund);
2795: PetscContainerSetUserDestroy(container,PetscContainerDestroy_MatRedundant);
2796: PetscObjectCompose((PetscObject)C,"Mat_Redundant",(PetscObject)container);
2797: PetscContainerDestroy(&container);
2799: redund->nzlocal = nzlocal;
2800: redund->nsends = nsends;
2801: redund->nrecvs = nrecvs;
2802: redund->send_rank = send_rank;
2803: redund->recv_rank = recv_rank;
2804: redund->sbuf_nz = sbuf_nz;
2805: redund->rbuf_nz = rbuf_nz;
2806: redund->sbuf_j = sbuf_j;
2807: redund->sbuf_a = sbuf_a;
2808: redund->rbuf_j = rbuf_j;
2809: redund->rbuf_a = rbuf_a;
2811: redund->Destroy = C->ops->destroy;
2812: C->ops->destroy = MatDestroy_MatRedundant;
2813: }
2814: return(0);
2815: }
2819: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2820: {
2821: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2823: PetscInt i,*idxb = 0;
2824: PetscScalar *va,*vb;
2825: Vec vtmp;
2828: MatGetRowMaxAbs(a->A,v,idx);
2829: VecGetArray(v,&va);
2830: if (idx) {
2831: for (i=0; i<A->rmap->n; i++) {
2832: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2833: }
2834: }
2836: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2837: if (idx) {
2838: PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);
2839: }
2840: MatGetRowMaxAbs(a->B,vtmp,idxb);
2841: VecGetArray(vtmp,&vb);
2843: for (i=0; i<A->rmap->n; i++){
2844: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2845: va[i] = vb[i];
2846: if (idx) idx[i] = a->garray[idxb[i]];
2847: }
2848: }
2850: VecRestoreArray(v,&va);
2851: VecRestoreArray(vtmp,&vb);
2852: PetscFree(idxb);
2853: VecDestroy(&vtmp);
2854: return(0);
2855: }
2859: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2860: {
2861: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2863: PetscInt i,*idxb = 0;
2864: PetscScalar *va,*vb;
2865: Vec vtmp;
2868: MatGetRowMinAbs(a->A,v,idx);
2869: VecGetArray(v,&va);
2870: if (idx) {
2871: for (i=0; i<A->cmap->n; i++) {
2872: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2873: }
2874: }
2876: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2877: if (idx) {
2878: PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);
2879: }
2880: MatGetRowMinAbs(a->B,vtmp,idxb);
2881: VecGetArray(vtmp,&vb);
2883: for (i=0; i<A->rmap->n; i++){
2884: if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2885: va[i] = vb[i];
2886: if (idx) idx[i] = a->garray[idxb[i]];
2887: }
2888: }
2890: VecRestoreArray(v,&va);
2891: VecRestoreArray(vtmp,&vb);
2892: PetscFree(idxb);
2893: VecDestroy(&vtmp);
2894: return(0);
2895: }
2899: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2900: {
2901: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
2902: PetscInt n = A->rmap->n;
2903: PetscInt cstart = A->cmap->rstart;
2904: PetscInt *cmap = mat->garray;
2905: PetscInt *diagIdx, *offdiagIdx;
2906: Vec diagV, offdiagV;
2907: PetscScalar *a, *diagA, *offdiagA;
2908: PetscInt r;
2912: PetscMalloc2(n,PetscInt,&diagIdx,n,PetscInt,&offdiagIdx);
2913: VecCreateSeq(((PetscObject)A)->comm, n, &diagV);
2914: VecCreateSeq(((PetscObject)A)->comm, n, &offdiagV);
2915: MatGetRowMin(mat->A, diagV, diagIdx);
2916: MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2917: VecGetArray(v, &a);
2918: VecGetArray(diagV, &diagA);
2919: VecGetArray(offdiagV, &offdiagA);
2920: for(r = 0; r < n; ++r) {
2921: if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2922: a[r] = diagA[r];
2923: idx[r] = cstart + diagIdx[r];
2924: } else {
2925: a[r] = offdiagA[r];
2926: idx[r] = cmap[offdiagIdx[r]];
2927: }
2928: }
2929: VecRestoreArray(v, &a);
2930: VecRestoreArray(diagV, &diagA);
2931: VecRestoreArray(offdiagV, &offdiagA);
2932: VecDestroy(&diagV);
2933: VecDestroy(&offdiagV);
2934: PetscFree2(diagIdx, offdiagIdx);
2935: return(0);
2936: }
2940: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2941: {
2942: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
2943: PetscInt n = A->rmap->n;
2944: PetscInt cstart = A->cmap->rstart;
2945: PetscInt *cmap = mat->garray;
2946: PetscInt *diagIdx, *offdiagIdx;
2947: Vec diagV, offdiagV;
2948: PetscScalar *a, *diagA, *offdiagA;
2949: PetscInt r;
2953: PetscMalloc2(n,PetscInt,&diagIdx,n,PetscInt,&offdiagIdx);
2954: VecCreateSeq(((PetscObject)A)->comm, n, &diagV);
2955: VecCreateSeq(((PetscObject)A)->comm, n, &offdiagV);
2956: MatGetRowMax(mat->A, diagV, diagIdx);
2957: MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2958: VecGetArray(v, &a);
2959: VecGetArray(diagV, &diagA);
2960: VecGetArray(offdiagV, &offdiagA);
2961: for(r = 0; r < n; ++r) {
2962: if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2963: a[r] = diagA[r];
2964: idx[r] = cstart + diagIdx[r];
2965: } else {
2966: a[r] = offdiagA[r];
2967: idx[r] = cmap[offdiagIdx[r]];
2968: }
2969: }
2970: VecRestoreArray(v, &a);
2971: VecRestoreArray(diagV, &diagA);
2972: VecRestoreArray(offdiagV, &offdiagA);
2973: VecDestroy(&diagV);
2974: VecDestroy(&offdiagV);
2975: PetscFree2(diagIdx, offdiagIdx);
2976: return(0);
2977: }
2981: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2982: {
2984: Mat *dummy;
2987: MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2988: *newmat = *dummy;
2989: PetscFree(dummy);
2990: return(0);
2991: }
2993: extern PetscErrorCode MatFDColoringApply_AIJ(Mat,MatFDColoring,Vec,MatStructure*,void*);
2997: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2998: {
2999: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data;
3003: MatInvertBlockDiagonal(a->A,values);
3004: return(0);
3005: }
3008: /* -------------------------------------------------------------------*/
3009: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
3010: MatGetRow_MPIAIJ,
3011: MatRestoreRow_MPIAIJ,
3012: MatMult_MPIAIJ,
3013: /* 4*/ MatMultAdd_MPIAIJ,
3014: MatMultTranspose_MPIAIJ,
3015: MatMultTransposeAdd_MPIAIJ,
3016: #ifdef PETSC_HAVE_PBGL
3017: MatSolve_MPIAIJ,
3018: #else
3019: 0,
3020: #endif
3021: 0,
3022: 0,
3023: /*10*/ 0,
3024: 0,
3025: 0,
3026: MatSOR_MPIAIJ,
3027: MatTranspose_MPIAIJ,
3028: /*15*/ MatGetInfo_MPIAIJ,
3029: MatEqual_MPIAIJ,
3030: MatGetDiagonal_MPIAIJ,
3031: MatDiagonalScale_MPIAIJ,
3032: MatNorm_MPIAIJ,
3033: /*20*/ MatAssemblyBegin_MPIAIJ,
3034: MatAssemblyEnd_MPIAIJ,
3035: MatSetOption_MPIAIJ,
3036: MatZeroEntries_MPIAIJ,
3037: /*24*/ MatZeroRows_MPIAIJ,
3038: 0,
3039: #ifdef PETSC_HAVE_PBGL
3040: 0,
3041: #else
3042: 0,
3043: #endif
3044: 0,
3045: 0,
3046: /*29*/ MatSetUp_MPIAIJ,
3047: #ifdef PETSC_HAVE_PBGL
3048: 0,
3049: #else
3050: 0,
3051: #endif
3052: 0,
3053: 0,
3054: 0,
3055: /*34*/ MatDuplicate_MPIAIJ,
3056: 0,
3057: 0,
3058: 0,
3059: 0,
3060: /*39*/ MatAXPY_MPIAIJ,
3061: MatGetSubMatrices_MPIAIJ,
3062: MatIncreaseOverlap_MPIAIJ,
3063: MatGetValues_MPIAIJ,
3064: MatCopy_MPIAIJ,
3065: /*44*/ MatGetRowMax_MPIAIJ,
3066: MatScale_MPIAIJ,
3067: 0,
3068: 0,
3069: MatZeroRowsColumns_MPIAIJ,
3070: /*49*/ 0,
3071: 0,
3072: 0,
3073: 0,
3074: 0,
3075: /*54*/ MatFDColoringCreate_MPIAIJ,
3076: 0,
3077: MatSetUnfactored_MPIAIJ,
3078: 0, /* MatPermute_MPIAIJ, impl currently broken */
3079: 0,
3080: /*59*/ MatGetSubMatrix_MPIAIJ,
3081: MatDestroy_MPIAIJ,
3082: MatView_MPIAIJ,
3083: 0,
3084: 0,
3085: /*64*/ 0,
3086: 0,
3087: 0,
3088: 0,
3089: 0,
3090: /*69*/ MatGetRowMaxAbs_MPIAIJ,
3091: MatGetRowMinAbs_MPIAIJ,
3092: 0,
3093: MatSetColoring_MPIAIJ,
3094: #if defined(PETSC_HAVE_ADIC)
3095: MatSetValuesAdic_MPIAIJ,
3096: #else
3097: 0,
3098: #endif
3099: MatSetValuesAdifor_MPIAIJ,
3100: /*75*/ MatFDColoringApply_AIJ,
3101: 0,
3102: 0,
3103: 0,
3104: MatFindZeroDiagonals_MPIAIJ,
3105: /*80*/ 0,
3106: 0,
3107: 0,
3108: /*83*/ MatLoad_MPIAIJ,
3109: 0,
3110: 0,
3111: 0,
3112: 0,
3113: 0,
3114: /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
3115: MatMatMultSymbolic_MPIAIJ_MPIAIJ,
3116: MatMatMultNumeric_MPIAIJ_MPIAIJ,
3117: MatPtAP_Basic,
3118: MatPtAPSymbolic_MPIAIJ,
3119: /*94*/ MatPtAPNumeric_MPIAIJ,
3120: 0,
3121: 0,
3122: 0,
3123: 0,
3124: /*99*/ 0,
3125: MatPtAPSymbolic_MPIAIJ_MPIAIJ,
3126: MatPtAPNumeric_MPIAIJ_MPIAIJ,
3127: MatConjugate_MPIAIJ,
3128: 0,
3129: /*104*/MatSetValuesRow_MPIAIJ,
3130: MatRealPart_MPIAIJ,
3131: MatImaginaryPart_MPIAIJ,
3132: 0,
3133: 0,
3134: /*109*/0,
3135: MatGetRedundantMatrix_MPIAIJ,
3136: MatGetRowMin_MPIAIJ,
3137: 0,
3138: 0,
3139: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
3140: 0,
3141: 0,
3142: 0,
3143: 0,
3144: /*119*/0,
3145: 0,
3146: 0,
3147: 0,
3148: MatGetMultiProcBlock_MPIAIJ,
3149: /*124*/MatFindNonzeroRows_MPIAIJ,
3150: MatGetColumnNorms_MPIAIJ,
3151: MatInvertBlockDiagonal_MPIAIJ,
3152: 0,
3153: MatGetSubMatricesParallel_MPIAIJ,
3154: /*129*/0,
3155: MatTransposeMatMult_MPIAIJ_MPIAIJ,
3156: MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
3157: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
3158: 0,
3159: /*134*/0,
3160: 0,
3161: 0,
3162: 0,
3163: 0
3164: };
3166: /* ----------------------------------------------------------------------------------------*/
3168: EXTERN_C_BEGIN
3171: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
3172: {
3173: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
3177: MatStoreValues(aij->A);
3178: MatStoreValues(aij->B);
3179: return(0);
3180: }
3181: EXTERN_C_END
3183: EXTERN_C_BEGIN
3186: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
3187: {
3188: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
3192: MatRetrieveValues(aij->A);
3193: MatRetrieveValues(aij->B);
3194: return(0);
3195: }
3196: EXTERN_C_END
3198: EXTERN_C_BEGIN
3201: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3202: {
3203: Mat_MPIAIJ *b;
3205: PetscInt i;
3206: PetscBool d_realalloc = PETSC_FALSE,o_realalloc = PETSC_FALSE;
3209: if (d_nz >= 0 || d_nnz) d_realalloc = PETSC_TRUE;
3210: if (o_nz >= 0 || o_nnz) o_realalloc = PETSC_TRUE;
3211: if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
3212: if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
3213: if (d_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
3214: if (o_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);
3216: PetscLayoutSetUp(B->rmap);
3217: PetscLayoutSetUp(B->cmap);
3218: if (d_nnz) {
3219: for (i=0; i<B->rmap->n; i++) {
3220: if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than 0: local row %D value %D",i,d_nnz[i]);
3221: }
3222: }
3223: if (o_nnz) {
3224: for (i=0; i<B->rmap->n; i++) {
3225: if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than 0: local row %D value %D",i,o_nnz[i]);
3226: }
3227: }
3228: b = (Mat_MPIAIJ*)B->data;
3230: if (!B->preallocated) {
3231: /* Explicitly create 2 MATSEQAIJ matrices. */
3232: MatCreate(PETSC_COMM_SELF,&b->A);
3233: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
3234: MatSetBlockSizes(b->A,B->rmap->bs,B->cmap->bs);
3235: MatSetType(b->A,MATSEQAIJ);
3236: PetscLogObjectParent(B,b->A);
3237: MatCreate(PETSC_COMM_SELF,&b->B);
3238: MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
3239: MatSetBlockSizes(b->B,B->rmap->bs,B->cmap->bs);
3240: MatSetType(b->B,MATSEQAIJ);
3241: PetscLogObjectParent(B,b->B);
3242: }
3244: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
3245: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
3246: /* Do not error if the user did not give real preallocation information. Ugly because this would overwrite a previous user call to MatSetOption(). */
3247: if (!d_realalloc) {MatSetOption(b->A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);}
3248: if (!o_realalloc) {MatSetOption(b->B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);}
3249: B->preallocated = PETSC_TRUE;
3250: return(0);
3251: }
3252: EXTERN_C_END
3256: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3257: {
3258: Mat mat;
3259: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
3263: *newmat = 0;
3264: MatCreate(((PetscObject)matin)->comm,&mat);
3265: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3266: MatSetBlockSizes(mat,matin->rmap->bs,matin->cmap->bs);
3267: MatSetType(mat,((PetscObject)matin)->type_name);
3268: PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
3269: a = (Mat_MPIAIJ*)mat->data;
3270:
3271: mat->factortype = matin->factortype;
3272: mat->rmap->bs = matin->rmap->bs;
3273: mat->cmap->bs = matin->cmap->bs;
3274: mat->assembled = PETSC_TRUE;
3275: mat->insertmode = NOT_SET_VALUES;
3276: mat->preallocated = PETSC_TRUE;
3278: a->size = oldmat->size;
3279: a->rank = oldmat->rank;
3280: a->donotstash = oldmat->donotstash;
3281: a->roworiented = oldmat->roworiented;
3282: a->rowindices = 0;
3283: a->rowvalues = 0;
3284: a->getrowactive = PETSC_FALSE;
3286: PetscLayoutReference(matin->rmap,&mat->rmap);
3287: PetscLayoutReference(matin->cmap,&mat->cmap);
3289: if (oldmat->colmap) {
3290: #if defined (PETSC_USE_CTABLE)
3291: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3292: #else
3293: PetscMalloc((mat->cmap->N)*sizeof(PetscInt),&a->colmap);
3294: PetscLogObjectMemory(mat,(mat->cmap->N)*sizeof(PetscInt));
3295: PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));
3296: #endif
3297: } else a->colmap = 0;
3298: if (oldmat->garray) {
3299: PetscInt len;
3300: len = oldmat->B->cmap->n;
3301: PetscMalloc((len+1)*sizeof(PetscInt),&a->garray);
3302: PetscLogObjectMemory(mat,len*sizeof(PetscInt));
3303: if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
3304: } else a->garray = 0;
3305:
3306: VecDuplicate(oldmat->lvec,&a->lvec);
3307: PetscLogObjectParent(mat,a->lvec);
3308: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3309: PetscLogObjectParent(mat,a->Mvctx);
3310: MatDuplicate(oldmat->A,cpvalues,&a->A);
3311: PetscLogObjectParent(mat,a->A);
3312: MatDuplicate(oldmat->B,cpvalues,&a->B);
3313: PetscLogObjectParent(mat,a->B);
3314: PetscFListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3315: *newmat = mat;
3316: return(0);
3317: }
3323: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3324: {
3325: PetscScalar *vals,*svals;
3326: MPI_Comm comm = ((PetscObject)viewer)->comm;
3328: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
3329: PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols;
3330: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
3331: PetscInt *ourlens = PETSC_NULL,*procsnz = PETSC_NULL,*offlens = PETSC_NULL,jj,*mycols,*smycols;
3332: PetscInt cend,cstart,n,*rowners,sizesset=1;
3333: int fd;
3336: MPI_Comm_size(comm,&size);
3337: MPI_Comm_rank(comm,&rank);
3338: if (!rank) {
3339: PetscViewerBinaryGetDescriptor(viewer,&fd);
3340: PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
3341: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3342: }
3344: if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0;
3346: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3347: M = header[1]; N = header[2];
3348: /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */
3349: if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M;
3350: if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N;
3351:
3352: /* If global sizes are set, check if they are consistent with that given in the file */
3353: if (sizesset) {
3354: MatGetSize(newMat,&grows,&gcols);
3355: }
3356: if (sizesset && newMat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows);
3357: if (sizesset && newMat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols);
3359: /* determine ownership of all rows */
3360: if (newMat->rmap->n < 0 ) m = M/size + ((M % size) > rank); /* PETSC_DECIDE */
3361: else m = newMat->rmap->n; /* Set by user */
3362:
3363: PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
3364: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
3366: /* First process needs enough room for process with most rows */
3367: if (!rank) {
3368: mmax = rowners[1];
3369: for(i=2; i<=size; i++) {
3370: mmax = PetscMax(mmax, rowners[i]);
3371: }
3372: } else mmax = m;
3374: rowners[0] = 0;
3375: for (i=2; i<=size; i++) {
3376: rowners[i] += rowners[i-1];
3377: }
3378: rstart = rowners[rank];
3379: rend = rowners[rank+1];
3381: /* distribute row lengths to all processors */
3382: PetscMalloc2(mmax,PetscInt,&ourlens,mmax,PetscInt,&offlens);
3383: if (!rank) {
3384: PetscBinaryRead(fd,ourlens,m,PETSC_INT);
3385: PetscMalloc(mmax*sizeof(PetscInt),&rowlengths);
3386: PetscMalloc(size*sizeof(PetscInt),&procsnz);
3387: PetscMemzero(procsnz,size*sizeof(PetscInt));
3388: for (j=0; j<m; j++) {
3389: procsnz[0] += ourlens[j];
3390: }
3391: for (i=1; i<size; i++) {
3392: PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
3393: /* calculate the number of nonzeros on each processor */
3394: for (j=0; j<rowners[i+1]-rowners[i]; j++) {
3395: procsnz[i] += rowlengths[j];
3396: }
3397: MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
3398: }
3399: PetscFree(rowlengths);
3400: } else {
3401: MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
3402: }
3404: if (!rank) {
3405: /* determine max buffer needed and allocate it */
3406: maxnz = 0;
3407: for (i=0; i<size; i++) {
3408: maxnz = PetscMax(maxnz,procsnz[i]);
3409: }
3410: PetscMalloc(maxnz*sizeof(PetscInt),&cols);
3412: /* read in my part of the matrix column indices */
3413: nz = procsnz[0];
3414: PetscMalloc(nz*sizeof(PetscInt),&mycols);
3415: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3417: /* read in every one elses and ship off */
3418: for (i=1; i<size; i++) {
3419: nz = procsnz[i];
3420: PetscBinaryRead(fd,cols,nz,PETSC_INT);
3421: MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
3422: }
3423: PetscFree(cols);
3424: } else {
3425: /* determine buffer space needed for message */
3426: nz = 0;
3427: for (i=0; i<m; i++) {
3428: nz += ourlens[i];
3429: }
3430: PetscMalloc(nz*sizeof(PetscInt),&mycols);
3432: /* receive message of column indices*/
3433: MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
3434: }
3436: /* determine column ownership if matrix is not square */
3437: if (N != M) {
3438: if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank);
3439: else n = newMat->cmap->n;
3440: MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
3441: cstart = cend - n;
3442: } else {
3443: cstart = rstart;
3444: cend = rend;
3445: n = cend - cstart;
3446: }
3448: /* loop over local rows, determining number of off diagonal entries */
3449: PetscMemzero(offlens,m*sizeof(PetscInt));
3450: jj = 0;
3451: for (i=0; i<m; i++) {
3452: for (j=0; j<ourlens[i]; j++) {
3453: if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
3454: jj++;
3455: }
3456: }
3458: for (i=0; i<m; i++) {
3459: ourlens[i] -= offlens[i];
3460: }
3461: if (!sizesset) {
3462: MatSetSizes(newMat,m,n,M,N);
3463: }
3464: MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);
3466: for (i=0; i<m; i++) {
3467: ourlens[i] += offlens[i];
3468: }
3470: if (!rank) {
3471: PetscMalloc((maxnz+1)*sizeof(PetscScalar),&vals);
3473: /* read in my part of the matrix numerical values */
3474: nz = procsnz[0];
3475: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3476:
3477: /* insert into matrix */
3478: jj = rstart;
3479: smycols = mycols;
3480: svals = vals;
3481: for (i=0; i<m; i++) {
3482: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3483: smycols += ourlens[i];
3484: svals += ourlens[i];
3485: jj++;
3486: }
3488: /* read in other processors and ship out */
3489: for (i=1; i<size; i++) {
3490: nz = procsnz[i];
3491: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3492: MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3493: }
3494: PetscFree(procsnz);
3495: } else {
3496: /* receive numeric values */
3497: PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);
3499: /* receive message of values*/
3500: MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);
3502: /* insert into matrix */
3503: jj = rstart;
3504: smycols = mycols;
3505: svals = vals;
3506: for (i=0; i<m; i++) {
3507: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3508: smycols += ourlens[i];
3509: svals += ourlens[i];
3510: jj++;
3511: }
3512: }
3513: PetscFree2(ourlens,offlens);
3514: PetscFree(vals);
3515: PetscFree(mycols);
3516: PetscFree(rowners);
3518: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3519: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3520: return(0);
3521: }
3525: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3526: {
3528: IS iscol_local;
3529: PetscInt csize;
3532: ISGetLocalSize(iscol,&csize);
3533: if (call == MAT_REUSE_MATRIX) {
3534: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3535: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3536: } else {
3537: PetscInt cbs;
3538: ISGetBlockSize(iscol,&cbs);
3539: ISAllGather(iscol,&iscol_local);
3540: ISSetBlockSize(iscol_local,cbs);
3541: }
3542: MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
3543: if (call == MAT_INITIAL_MATRIX) {
3544: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3545: ISDestroy(&iscol_local);
3546: }
3547: return(0);
3548: }
3550: extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*);
3553: /*
3554: Not great since it makes two copies of the submatrix, first an SeqAIJ
3555: in local and then by concatenating the local matrices the end result.
3556: Writing it directly would be much like MatGetSubMatrices_MPIAIJ()
3558: Note: This requires a sequential iscol with all indices.
3559: */
3560: PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3561: {
3563: PetscMPIInt rank,size;
3564: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3565: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol;
3566: PetscBool allcolumns, colflag;
3567: Mat M,Mreuse;
3568: MatScalar *vwork,*aa;
3569: MPI_Comm comm = ((PetscObject)mat)->comm;
3570: Mat_SeqAIJ *aij;
3574: MPI_Comm_rank(comm,&rank);
3575: MPI_Comm_size(comm,&size);
3577: ISIdentity(iscol,&colflag);
3578: ISGetLocalSize(iscol,&ncol);
3579: if (colflag && ncol == mat->cmap->N){
3580: allcolumns = PETSC_TRUE;
3581: } else {
3582: allcolumns = PETSC_FALSE;
3583: }
3584: if (call == MAT_REUSE_MATRIX) {
3585: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);
3586: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3587: MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);
3588: } else {
3589: MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);
3590: }
3592: /*
3593: m - number of local rows
3594: n - number of columns (same on all processors)
3595: rstart - first row in new global matrix generated
3596: */
3597: MatGetSize(Mreuse,&m,&n);
3598: MatGetBlockSizes(Mreuse,&bs,&cbs);
3599: if (call == MAT_INITIAL_MATRIX) {
3600: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3601: ii = aij->i;
3602: jj = aij->j;
3604: /*
3605: Determine the number of non-zeros in the diagonal and off-diagonal
3606: portions of the matrix in order to do correct preallocation
3607: */
3609: /* first get start and end of "diagonal" columns */
3610: if (csize == PETSC_DECIDE) {
3611: ISGetSize(isrow,&mglobal);
3612: if (mglobal == n) { /* square matrix */
3613: nlocal = m;
3614: } else {
3615: nlocal = n/size + ((n % size) > rank);
3616: }
3617: } else {
3618: nlocal = csize;
3619: }
3620: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3621: rstart = rend - nlocal;
3622: if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
3624: /* next, compute all the lengths */
3625: PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
3626: olens = dlens + m;
3627: for (i=0; i<m; i++) {
3628: jend = ii[i+1] - ii[i];
3629: olen = 0;
3630: dlen = 0;
3631: for (j=0; j<jend; j++) {
3632: if (*jj < rstart || *jj >= rend) olen++;
3633: else dlen++;
3634: jj++;
3635: }
3636: olens[i] = olen;
3637: dlens[i] = dlen;
3638: }
3639: MatCreate(comm,&M);
3640: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3641: MatSetBlockSizes(M,bs,cbs);
3642: MatSetType(M,((PetscObject)mat)->type_name);
3643: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3644: PetscFree(dlens);
3645: } else {
3646: PetscInt ml,nl;
3648: M = *newmat;
3649: MatGetLocalSize(M,&ml,&nl);
3650: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3651: MatZeroEntries(M);
3652: /*
3653: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3654: rather than the slower MatSetValues().
3655: */
3656: M->was_assembled = PETSC_TRUE;
3657: M->assembled = PETSC_FALSE;
3658: }
3659: MatGetOwnershipRange(M,&rstart,&rend);
3660: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3661: ii = aij->i;
3662: jj = aij->j;
3663: aa = aij->a;
3664: for (i=0; i<m; i++) {
3665: row = rstart + i;
3666: nz = ii[i+1] - ii[i];
3667: cwork = jj; jj += nz;
3668: vwork = aa; aa += nz;
3669: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3670: }
3672: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3673: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3674: *newmat = M;
3676: /* save submatrix used in processor for next request */
3677: if (call == MAT_INITIAL_MATRIX) {
3678: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3679: MatDestroy(&Mreuse);
3680: }
3682: return(0);
3683: }
3685: EXTERN_C_BEGIN
3688: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3689: {
3690: PetscInt m,cstart, cend,j,nnz,i,d;
3691: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3692: const PetscInt *JJ;
3693: PetscScalar *values;
3697: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
3699: PetscLayoutSetUp(B->rmap);
3700: PetscLayoutSetUp(B->cmap);
3701: m = B->rmap->n;
3702: cstart = B->cmap->rstart;
3703: cend = B->cmap->rend;
3704: rstart = B->rmap->rstart;
3706: PetscMalloc2(m,PetscInt,&d_nnz,m,PetscInt,&o_nnz);
3708: #if defined(PETSC_USE_DEBUGGING)
3709: for (i=0; i<m; i++) {
3710: nnz = Ii[i+1]- Ii[i];
3711: JJ = J + Ii[i];
3712: if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3713: if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j);
3714: if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
3715: }
3716: #endif
3718: for (i=0; i<m; i++) {
3719: nnz = Ii[i+1]- Ii[i];
3720: JJ = J + Ii[i];
3721: nnz_max = PetscMax(nnz_max,nnz);
3722: d = 0;
3723: for (j=0; j<nnz; j++) {
3724: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3725: }
3726: d_nnz[i] = d;
3727: o_nnz[i] = nnz - d;
3728: }
3729: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3730: PetscFree2(d_nnz,o_nnz);
3732: if (v) values = (PetscScalar*)v;
3733: else {
3734: PetscMalloc((nnz_max+1)*sizeof(PetscScalar),&values);
3735: PetscMemzero(values,nnz_max*sizeof(PetscScalar));
3736: }
3738: for (i=0; i<m; i++) {
3739: ii = i + rstart;
3740: nnz = Ii[i+1]- Ii[i];
3741: MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
3742: }
3743: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3744: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3746: if (!v) {
3747: PetscFree(values);
3748: }
3749: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3750: return(0);
3751: }
3752: EXTERN_C_END
3756: /*@
3757: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3758: (the default parallel PETSc format).
3760: Collective on MPI_Comm
3762: Input Parameters:
3763: + B - the matrix
3764: . i - the indices into j for the start of each local row (starts with zero)
3765: . j - the column indices for each local row (starts with zero)
3766: - v - optional values in the matrix
3768: Level: developer
3770: Notes:
3771: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3772: thus you CANNOT change the matrix entries by changing the values of a[] after you have
3773: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
3775: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
3777: The format which is used for the sparse matrix input, is equivalent to a
3778: row-major ordering.. i.e for the following matrix, the input data expected is
3779: as shown:
3781: 1 0 0
3782: 2 0 3 P0
3783: -------
3784: 4 5 6 P1
3786: Process0 [P0]: rows_owned=[0,1]
3787: i = {0,1,3} [size = nrow+1 = 2+1]
3788: j = {0,0,2} [size = nz = 6]
3789: v = {1,2,3} [size = nz = 6]
3791: Process1 [P1]: rows_owned=[2]
3792: i = {0,3} [size = nrow+1 = 1+1]
3793: j = {0,1,2} [size = nz = 6]
3794: v = {4,5,6} [size = nz = 6]
3796: .keywords: matrix, aij, compressed row, sparse, parallel
3798: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ,
3799: MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
3800: @*/
3801: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
3802: {
3806: PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3807: return(0);
3808: }
3812: /*@C
3813: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
3814: (the default parallel PETSc format). For good matrix assembly performance
3815: the user should preallocate the matrix storage by setting the parameters
3816: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
3817: performance can be increased by more than a factor of 50.
3819: Collective on MPI_Comm
3821: Input Parameters:
3822: + A - the matrix
3823: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
3824: (same value is used for all local rows)
3825: . d_nnz - array containing the number of nonzeros in the various rows of the
3826: DIAGONAL portion of the local submatrix (possibly different for each row)
3827: or PETSC_NULL, if d_nz is used to specify the nonzero structure.
3828: The size of this array is equal to the number of local rows, i.e 'm'.
3829: For matrices that will be factored, you must leave room for (and set)
3830: the diagonal entry even if it is zero.
3831: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
3832: submatrix (same value is used for all local rows).
3833: - o_nnz - array containing the number of nonzeros in the various rows of the
3834: OFF-DIAGONAL portion of the local submatrix (possibly different for
3835: each row) or PETSC_NULL, if o_nz is used to specify the nonzero
3836: structure. The size of this array is equal to the number
3837: of local rows, i.e 'm'.
3839: If the *_nnz parameter is given then the *_nz parameter is ignored
3841: The AIJ format (also called the Yale sparse matrix format or
3842: compressed row storage (CSR)), is fully compatible with standard Fortran 77
3843: storage. The stored row and column indices begin with zero.
3844: See the <A href="../../docs/manual.pdf#nameddest=ch_mat">Mat chapter of the users manual</A> for details.
3846: The parallel matrix is partitioned such that the first m0 rows belong to
3847: process 0, the next m1 rows belong to process 1, the next m2 rows belong
3848: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
3850: The DIAGONAL portion of the local submatrix of a processor can be defined
3851: as the submatrix which is obtained by extraction the part corresponding to
3852: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
3853: first row that belongs to the processor, r2 is the last row belonging to
3854: the this processor, and c1-c2 is range of indices of the local part of a
3855: vector suitable for applying the matrix to. This is an mxn matrix. In the
3856: common case of a square matrix, the row and column ranges are the same and
3857: the DIAGONAL part is also square. The remaining portion of the local
3858: submatrix (mxN) constitute the OFF-DIAGONAL portion.
3860: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
3862: You can call MatGetInfo() to get information on how effective the preallocation was;
3863: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3864: You can also run with the option -info and look for messages with the string
3865: malloc in them to see if additional memory allocation was needed.
3867: Example usage:
3868:
3869: Consider the following 8x8 matrix with 34 non-zero values, that is
3870: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3871: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
3872: as follows:
3874: .vb
3875: 1 2 0 | 0 3 0 | 0 4
3876: Proc0 0 5 6 | 7 0 0 | 8 0
3877: 9 0 10 | 11 0 0 | 12 0
3878: -------------------------------------
3879: 13 0 14 | 15 16 17 | 0 0
3880: Proc1 0 18 0 | 19 20 21 | 0 0
3881: 0 0 0 | 22 23 0 | 24 0
3882: -------------------------------------
3883: Proc2 25 26 27 | 0 0 28 | 29 0
3884: 30 0 0 | 31 32 33 | 0 34
3885: .ve
3887: This can be represented as a collection of submatrices as:
3889: .vb
3890: A B C
3891: D E F
3892: G H I
3893: .ve
3895: Where the submatrices A,B,C are owned by proc0, D,E,F are
3896: owned by proc1, G,H,I are owned by proc2.
3898: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3899: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3900: The 'M','N' parameters are 8,8, and have the same values on all procs.
3902: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3903: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3904: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3905: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3906: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3907: matrix, ans [DF] as another SeqAIJ matrix.
3909: When d_nz, o_nz parameters are specified, d_nz storage elements are
3910: allocated for every row of the local diagonal submatrix, and o_nz
3911: storage locations are allocated for every row of the OFF-DIAGONAL submat.
3912: One way to choose d_nz and o_nz is to use the max nonzerors per local
3913: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
3914: In this case, the values of d_nz,o_nz are:
3915: .vb
3916: proc0 : dnz = 2, o_nz = 2
3917: proc1 : dnz = 3, o_nz = 2
3918: proc2 : dnz = 1, o_nz = 4
3919: .ve
3920: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3921: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3922: for proc3. i.e we are using 12+15+10=37 storage locations to store
3923: 34 values.
3925: When d_nnz, o_nnz parameters are specified, the storage is specified
3926: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3927: In the above case the values for d_nnz,o_nnz are:
3928: .vb
3929: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3930: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3931: proc2: d_nnz = [1,1] and o_nnz = [4,4]
3932: .ve
3933: Here the space allocated is sum of all the above values i.e 34, and
3934: hence pre-allocation is perfect.
3936: Level: intermediate
3938: .keywords: matrix, aij, compressed row, sparse, parallel
3940: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
3941: MPIAIJ, MatGetInfo()
3942: @*/
3943: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3944: {
3950: PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
3951: return(0);
3952: }
3956: /*@
3957: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
3958: CSR format the local rows.
3960: Collective on MPI_Comm
3962: Input Parameters:
3963: + comm - MPI communicator
3964: . m - number of local rows (Cannot be PETSC_DECIDE)
3965: . n - This value should be the same as the local size used in creating the
3966: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3967: calculated if N is given) For square matrices n is almost always m.
3968: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3969: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3970: . i - row indices
3971: . j - column indices
3972: - a - matrix values
3974: Output Parameter:
3975: . mat - the matrix
3977: Level: intermediate
3979: Notes:
3980: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3981: thus you CANNOT change the matrix entries by changing the values of a[] after you have
3982: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
3984: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
3986: The format which is used for the sparse matrix input, is equivalent to a
3987: row-major ordering.. i.e for the following matrix, the input data expected is
3988: as shown:
3990: 1 0 0
3991: 2 0 3 P0
3992: -------
3993: 4 5 6 P1
3995: Process0 [P0]: rows_owned=[0,1]
3996: i = {0,1,3} [size = nrow+1 = 2+1]
3997: j = {0,0,2} [size = nz = 6]
3998: v = {1,2,3} [size = nz = 6]
4000: Process1 [P1]: rows_owned=[2]
4001: i = {0,3} [size = nrow+1 = 1+1]
4002: j = {0,1,2} [size = nz = 6]
4003: v = {4,5,6} [size = nz = 6]
4005: .keywords: matrix, aij, compressed row, sparse, parallel
4007: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4008: MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
4009: @*/
4010: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4011: {
4015: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4016: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4017: MatCreate(comm,mat);
4018: MatSetSizes(*mat,m,n,M,N);
4019: /* MatSetBlockSizes(M,bs,cbs); */
4020: MatSetType(*mat,MATMPIAIJ);
4021: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4022: return(0);
4023: }
4027: /*@C
4028: MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4029: (the default parallel PETSc format). For good matrix assembly performance
4030: the user should preallocate the matrix storage by setting the parameters
4031: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4032: performance can be increased by more than a factor of 50.
4034: Collective on MPI_Comm
4036: Input Parameters:
4037: + comm - MPI communicator
4038: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4039: This value should be the same as the local size used in creating the
4040: y vector for the matrix-vector product y = Ax.
4041: . n - This value should be the same as the local size used in creating the
4042: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4043: calculated if N is given) For square matrices n is almost always m.
4044: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4045: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4046: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4047: (same value is used for all local rows)
4048: . d_nnz - array containing the number of nonzeros in the various rows of the
4049: DIAGONAL portion of the local submatrix (possibly different for each row)
4050: or PETSC_NULL, if d_nz is used to specify the nonzero structure.
4051: The size of this array is equal to the number of local rows, i.e 'm'.
4052: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4053: submatrix (same value is used for all local rows).
4054: - o_nnz - array containing the number of nonzeros in the various rows of the
4055: OFF-DIAGONAL portion of the local submatrix (possibly different for
4056: each row) or PETSC_NULL, if o_nz is used to specify the nonzero
4057: structure. The size of this array is equal to the number
4058: of local rows, i.e 'm'.
4060: Output Parameter:
4061: . A - the matrix
4063: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4064: MatXXXXSetPreallocation() paradgm instead of this routine directly.
4065: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
4067: Notes:
4068: If the *_nnz parameter is given then the *_nz parameter is ignored
4070: m,n,M,N parameters specify the size of the matrix, and its partitioning across
4071: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4072: storage requirements for this matrix.
4074: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
4075: processor than it must be used on all processors that share the object for
4076: that argument.
4078: The user MUST specify either the local or global matrix dimensions
4079: (possibly both).
4081: The parallel matrix is partitioned across processors such that the
4082: first m0 rows belong to process 0, the next m1 rows belong to
4083: process 1, the next m2 rows belong to process 2 etc.. where
4084: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4085: values corresponding to [m x N] submatrix.
4087: The columns are logically partitioned with the n0 columns belonging
4088: to 0th partition, the next n1 columns belonging to the next
4089: partition etc.. where n0,n1,n2... are the the input parameter 'n'.
4091: The DIAGONAL portion of the local submatrix on any given processor
4092: is the submatrix corresponding to the rows and columns m,n
4093: corresponding to the given processor. i.e diagonal matrix on
4094: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4095: etc. The remaining portion of the local submatrix [m x (N-n)]
4096: constitute the OFF-DIAGONAL portion. The example below better
4097: illustrates this concept.
4099: For a square global matrix we define each processor's diagonal portion
4100: to be its local rows and the corresponding columns (a square submatrix);
4101: each processor's off-diagonal portion encompasses the remainder of the
4102: local matrix (a rectangular submatrix).
4104: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4106: When calling this routine with a single process communicator, a matrix of
4107: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
4108: type of communicator, use the construction mechanism:
4109: MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4110:
4111: By default, this format uses inodes (identical nodes) when possible.
4112: We search for consecutive rows with the same nonzero structure, thereby
4113: reusing matrix information to achieve increased efficiency.
4115: Options Database Keys:
4116: + -mat_no_inode - Do not use inodes
4117: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4118: - -mat_aij_oneindex - Internally use indexing starting at 1
4119: rather than 0. Note that when calling MatSetValues(),
4120: the user still MUST index entries starting at 0!
4123: Example usage:
4124:
4125: Consider the following 8x8 matrix with 34 non-zero values, that is
4126: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4127: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4128: as follows:
4130: .vb
4131: 1 2 0 | 0 3 0 | 0 4
4132: Proc0 0 5 6 | 7 0 0 | 8 0
4133: 9 0 10 | 11 0 0 | 12 0
4134: -------------------------------------
4135: 13 0 14 | 15 16 17 | 0 0
4136: Proc1 0 18 0 | 19 20 21 | 0 0
4137: 0 0 0 | 22 23 0 | 24 0
4138: -------------------------------------
4139: Proc2 25 26 27 | 0 0 28 | 29 0
4140: 30 0 0 | 31 32 33 | 0 34
4141: .ve
4143: This can be represented as a collection of submatrices as:
4145: .vb
4146: A B C
4147: D E F
4148: G H I
4149: .ve
4151: Where the submatrices A,B,C are owned by proc0, D,E,F are
4152: owned by proc1, G,H,I are owned by proc2.
4154: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4155: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4156: The 'M','N' parameters are 8,8, and have the same values on all procs.
4158: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4159: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4160: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4161: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4162: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4163: matrix, ans [DF] as another SeqAIJ matrix.
4165: When d_nz, o_nz parameters are specified, d_nz storage elements are
4166: allocated for every row of the local diagonal submatrix, and o_nz
4167: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4168: One way to choose d_nz and o_nz is to use the max nonzerors per local
4169: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4170: In this case, the values of d_nz,o_nz are:
4171: .vb
4172: proc0 : dnz = 2, o_nz = 2
4173: proc1 : dnz = 3, o_nz = 2
4174: proc2 : dnz = 1, o_nz = 4
4175: .ve
4176: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4177: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4178: for proc3. i.e we are using 12+15+10=37 storage locations to store
4179: 34 values.
4181: When d_nnz, o_nnz parameters are specified, the storage is specified
4182: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4183: In the above case the values for d_nnz,o_nnz are:
4184: .vb
4185: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4186: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4187: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4188: .ve
4189: Here the space allocated is sum of all the above values i.e 34, and
4190: hence pre-allocation is perfect.
4192: Level: intermediate
4194: .keywords: matrix, aij, compressed row, sparse, parallel
4196: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4197: MPIAIJ, MatCreateMPIAIJWithArrays()
4198: @*/
4199: PetscErrorCode MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
4200: {
4202: PetscMPIInt size;
4205: MatCreate(comm,A);
4206: MatSetSizes(*A,m,n,M,N);
4207: MPI_Comm_size(comm,&size);
4208: if (size > 1) {
4209: MatSetType(*A,MATMPIAIJ);
4210: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4211: } else {
4212: MatSetType(*A,MATSEQAIJ);
4213: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4214: }
4215: return(0);
4216: }
4220: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[])
4221: {
4222: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4225: *Ad = a->A;
4226: *Ao = a->B;
4227: *colmap = a->garray;
4228: return(0);
4229: }
4233: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
4234: {
4236: PetscInt i;
4237: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4240: if (coloring->ctype == IS_COLORING_GLOBAL) {
4241: ISColoringValue *allcolors,*colors;
4242: ISColoring ocoloring;
4244: /* set coloring for diagonal portion */
4245: MatSetColoring_SeqAIJ(a->A,coloring);
4247: /* set coloring for off-diagonal portion */
4248: ISAllGatherColors(((PetscObject)A)->comm,coloring->n,coloring->colors,PETSC_NULL,&allcolors);
4249: PetscMalloc((a->B->cmap->n+1)*sizeof(ISColoringValue),&colors);
4250: for (i=0; i<a->B->cmap->n; i++) {
4251: colors[i] = allcolors[a->garray[i]];
4252: }
4253: PetscFree(allcolors);
4254: ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);
4255: MatSetColoring_SeqAIJ(a->B,ocoloring);
4256: ISColoringDestroy(&ocoloring);
4257: } else if (coloring->ctype == IS_COLORING_GHOSTED) {
4258: ISColoringValue *colors;
4259: PetscInt *larray;
4260: ISColoring ocoloring;
4262: /* set coloring for diagonal portion */
4263: PetscMalloc((a->A->cmap->n+1)*sizeof(PetscInt),&larray);
4264: for (i=0; i<a->A->cmap->n; i++) {
4265: larray[i] = i + A->cmap->rstart;
4266: }
4267: ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,PETSC_NULL,larray);
4268: PetscMalloc((a->A->cmap->n+1)*sizeof(ISColoringValue),&colors);
4269: for (i=0; i<a->A->cmap->n; i++) {
4270: colors[i] = coloring->colors[larray[i]];
4271: }
4272: PetscFree(larray);
4273: ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,&ocoloring);
4274: MatSetColoring_SeqAIJ(a->A,ocoloring);
4275: ISColoringDestroy(&ocoloring);
4277: /* set coloring for off-diagonal portion */
4278: PetscMalloc((a->B->cmap->n+1)*sizeof(PetscInt),&larray);
4279: ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,PETSC_NULL,larray);
4280: PetscMalloc((a->B->cmap->n+1)*sizeof(ISColoringValue),&colors);
4281: for (i=0; i<a->B->cmap->n; i++) {
4282: colors[i] = coloring->colors[larray[i]];
4283: }
4284: PetscFree(larray);
4285: ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);
4286: MatSetColoring_SeqAIJ(a->B,ocoloring);
4287: ISColoringDestroy(&ocoloring);
4288: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);
4290: return(0);
4291: }
4293: #if defined(PETSC_HAVE_ADIC)
4296: PetscErrorCode MatSetValuesAdic_MPIAIJ(Mat A,void *advalues)
4297: {
4298: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4302: MatSetValuesAdic_SeqAIJ(a->A,advalues);
4303: MatSetValuesAdic_SeqAIJ(a->B,advalues);
4304: return(0);
4305: }
4306: #endif
4310: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
4311: {
4312: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4316: MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
4317: MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
4318: return(0);
4319: }
4323: PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJSymbolic(MPI_Comm comm,Mat inmat,PetscInt n,Mat *outmat)
4324: {
4326: PetscInt m,N,i,rstart,nnz,*dnz,*onz,sum,bs,cbs;
4327: PetscInt *indx;
4330: /* This routine will ONLY return MPIAIJ type matrix */
4331: MatGetSize(inmat,&m,&N);
4332: MatGetBlockSizes(inmat,&bs,&cbs);
4333: if (n == PETSC_DECIDE){
4334: PetscSplitOwnership(comm,&n,&N);
4335: }
4336: /* Check sum(n) = N */
4337: MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4338: if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N);
4339:
4340: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4341: rstart -= m;
4343: MatPreallocateInitialize(comm,m,n,dnz,onz);
4344: for (i=0;i<m;i++) {
4345: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
4346: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4347: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
4348: }
4349:
4350: MatCreate(comm,outmat);
4351: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4352: MatSetBlockSizes(*outmat,bs,cbs);
4353: MatSetType(*outmat,MATMPIAIJ);
4354: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4355: MatPreallocateFinalize(dnz,onz);
4356: return(0);
4357: }
4361: PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJNumeric(MPI_Comm comm,Mat inmat,PetscInt n,Mat outmat)
4362: {
4364: PetscInt m,N,i,rstart,nnz,Ii;
4365: PetscInt *indx;
4366: PetscScalar *values;
4369: MatGetSize(inmat,&m,&N);
4370: MatGetOwnershipRange(outmat,&rstart,PETSC_NULL);
4371: for (i=0;i<m;i++) {
4372: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4373: Ii = i + rstart;
4374: MatSetValues_MPIAIJ(outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4375: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4376: }
4377: MatAssemblyBegin(outmat,MAT_FINAL_ASSEMBLY);
4378: MatAssemblyEnd(outmat,MAT_FINAL_ASSEMBLY);
4379: return(0);
4380: }
4384: /*@
4385: MatCreateMPIAIJConcatenateSeqAIJ - Creates a single large PETSc matrix by concatenating sequential
4386: matrices from each processor
4388: Collective on MPI_Comm
4390: Input Parameters:
4391: + comm - the communicators the parallel matrix will live on
4392: . inmat - the input sequential matrices
4393: . n - number of local columns (or PETSC_DECIDE)
4394: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4396: Output Parameter:
4397: . outmat - the parallel matrix generated
4399: Level: advanced
4401: Notes: The number of columns of the matrix in EACH processor MUST be the same.
4403: @*/
4404: PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4405: {
4409: PetscLogEventBegin(MAT_Merge,inmat,0,0,0);
4410: if (scall == MAT_INITIAL_MATRIX){
4411: MatCreateMPIAIJConcatenateSeqAIJSymbolic(comm,inmat,n,outmat);
4412: }
4413: MatCreateMPIAIJConcatenateSeqAIJNumeric(comm,inmat,n,*outmat);
4414: PetscLogEventEnd(MAT_Merge,inmat,0,0,0);
4415: return(0);
4416: }
4420: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4421: {
4422: PetscErrorCode ierr;
4423: PetscMPIInt rank;
4424: PetscInt m,N,i,rstart,nnz;
4425: size_t len;
4426: const PetscInt *indx;
4427: PetscViewer out;
4428: char *name;
4429: Mat B;
4430: const PetscScalar *values;
4433: MatGetLocalSize(A,&m,0);
4434: MatGetSize(A,0,&N);
4435: /* Should this be the type of the diagonal block of A? */
4436: MatCreate(PETSC_COMM_SELF,&B);
4437: MatSetSizes(B,m,N,m,N);
4438: MatSetBlockSizes(B,A->rmap->bs,A->cmap->bs);
4439: MatSetType(B,MATSEQAIJ);
4440: MatSeqAIJSetPreallocation(B,0,PETSC_NULL);
4441: MatGetOwnershipRange(A,&rstart,0);
4442: for (i=0;i<m;i++) {
4443: MatGetRow(A,i+rstart,&nnz,&indx,&values);
4444: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4445: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4446: }
4447: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4448: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4450: MPI_Comm_rank(((PetscObject)A)->comm,&rank);
4451: PetscStrlen(outfile,&len);
4452: PetscMalloc((len+5)*sizeof(char),&name);
4453: sprintf(name,"%s.%d",outfile,rank);
4454: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4455: PetscFree(name);
4456: MatView(B,out);
4457: PetscViewerDestroy(&out);
4458: MatDestroy(&B);
4459: return(0);
4460: }
4462: extern PetscErrorCode MatDestroy_MPIAIJ(Mat);
4465: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4466: {
4467: PetscErrorCode ierr;
4468: Mat_Merge_SeqsToMPI *merge;
4469: PetscContainer container;
4472: PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);
4473: if (container) {
4474: PetscContainerGetPointer(container,(void **)&merge);
4475: PetscFree(merge->id_r);
4476: PetscFree(merge->len_s);
4477: PetscFree(merge->len_r);
4478: PetscFree(merge->bi);
4479: PetscFree(merge->bj);
4480: PetscFree(merge->buf_ri[0]);
4481: PetscFree(merge->buf_ri);
4482: PetscFree(merge->buf_rj[0]);
4483: PetscFree(merge->buf_rj);
4484: PetscFree(merge->coi);
4485: PetscFree(merge->coj);
4486: PetscFree(merge->owners_co);
4487: PetscLayoutDestroy(&merge->rowmap);
4488: PetscFree(merge);
4489: PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4490: }
4491: MatDestroy_MPIAIJ(A);
4492: return(0);
4493: }
4495: #include <../src/mat/utils/freespace.h>
4496: #include <petscbt.h>
4500: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4501: {
4502: PetscErrorCode ierr;
4503: MPI_Comm comm=((PetscObject)mpimat)->comm;
4504: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4505: PetscMPIInt size,rank,taga,*len_s;
4506: PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj=a->j;
4507: PetscInt proc,m;
4508: PetscInt **buf_ri,**buf_rj;
4509: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4510: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
4511: MPI_Request *s_waits,*r_waits;
4512: MPI_Status *status;
4513: MatScalar *aa=a->a;
4514: MatScalar **abuf_r,*ba_i;
4515: Mat_Merge_SeqsToMPI *merge;
4516: PetscContainer container;
4519: PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);
4521: MPI_Comm_size(comm,&size);
4522: MPI_Comm_rank(comm,&rank);
4524: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject *)&container);
4525: PetscContainerGetPointer(container,(void **)&merge);
4527: bi = merge->bi;
4528: bj = merge->bj;
4529: buf_ri = merge->buf_ri;
4530: buf_rj = merge->buf_rj;
4532: PetscMalloc(size*sizeof(MPI_Status),&status);
4533: owners = merge->rowmap->range;
4534: len_s = merge->len_s;
4536: /* send and recv matrix values */
4537: /*-----------------------------*/
4538: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4539: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
4541: PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);
4542: for (proc=0,k=0; proc<size; proc++){
4543: if (!len_s[proc]) continue;
4544: i = owners[proc];
4545: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4546: k++;
4547: }
4549: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4550: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4551: PetscFree(status);
4553: PetscFree(s_waits);
4554: PetscFree(r_waits);
4556: /* insert mat values of mpimat */
4557: /*----------------------------*/
4558: PetscMalloc(N*sizeof(PetscScalar),&ba_i);
4559: PetscMalloc3(merge->nrecv,PetscInt*,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextai);
4561: for (k=0; k<merge->nrecv; k++){
4562: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4563: nrows = *(buf_ri_k[k]);
4564: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
4565: nextai[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */
4566: }
4568: /* set values of ba */
4569: m = merge->rowmap->n;
4570: for (i=0; i<m; i++) {
4571: arow = owners[rank] + i;
4572: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
4573: bnzi = bi[i+1] - bi[i];
4574: PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));
4576: /* add local non-zero vals of this proc's seqmat into ba */
4577: anzi = ai[arow+1] - ai[arow];
4578: aj = a->j + ai[arow];
4579: aa = a->a + ai[arow];
4580: nextaj = 0;
4581: for (j=0; nextaj<anzi; j++){
4582: if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
4583: ba_i[j] += aa[nextaj++];
4584: }
4585: }
4587: /* add received vals into ba */
4588: for (k=0; k<merge->nrecv; k++){ /* k-th received message */
4589: /* i-th row */
4590: if (i == *nextrow[k]) {
4591: anzi = *(nextai[k]+1) - *nextai[k];
4592: aj = buf_rj[k] + *(nextai[k]);
4593: aa = abuf_r[k] + *(nextai[k]);
4594: nextaj = 0;
4595: for (j=0; nextaj<anzi; j++){
4596: if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
4597: ba_i[j] += aa[nextaj++];
4598: }
4599: }
4600: nextrow[k]++; nextai[k]++;
4601: }
4602: }
4603: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4604: }
4605: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4606: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
4608: PetscFree(abuf_r[0]);
4609: PetscFree(abuf_r);
4610: PetscFree(ba_i);
4611: PetscFree3(buf_ri_k,nextrow,nextai);
4612: PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4613: return(0);
4614: }
4616: extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat);
4620: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4621: {
4622: PetscErrorCode ierr;
4623: Mat B_mpi;
4624: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4625: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4626: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
4627: PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4628: PetscInt len,proc,*dnz,*onz,bs,cbs;
4629: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4630: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4631: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
4632: MPI_Status *status;
4633: PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
4634: PetscBT lnkbt;
4635: Mat_Merge_SeqsToMPI *merge;
4636: PetscContainer container;
4639: PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);
4641: /* make sure it is a PETSc comm */
4642: PetscCommDuplicate(comm,&comm,PETSC_NULL);
4643: MPI_Comm_size(comm,&size);
4644: MPI_Comm_rank(comm,&rank);
4646: PetscNew(Mat_Merge_SeqsToMPI,&merge);
4647: PetscMalloc(size*sizeof(MPI_Status),&status);
4649: /* determine row ownership */
4650: /*---------------------------------------------------------*/
4651: PetscLayoutCreate(comm,&merge->rowmap);
4652: PetscLayoutSetLocalSize(merge->rowmap,m);
4653: PetscLayoutSetSize(merge->rowmap,M);
4654: PetscLayoutSetBlockSize(merge->rowmap,1);
4655: PetscLayoutSetUp(merge->rowmap);
4656: PetscMalloc(size*sizeof(PetscMPIInt),&len_si);
4657: PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);
4659: m = merge->rowmap->n;
4660: M = merge->rowmap->N;
4661: owners = merge->rowmap->range;
4663: /* determine the number of messages to send, their lengths */
4664: /*---------------------------------------------------------*/
4665: len_s = merge->len_s;
4667: len = 0; /* length of buf_si[] */
4668: merge->nsend = 0;
4669: for (proc=0; proc<size; proc++){
4670: len_si[proc] = 0;
4671: if (proc == rank){
4672: len_s[proc] = 0;
4673: } else {
4674: len_si[proc] = owners[proc+1] - owners[proc] + 1;
4675: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4676: }
4677: if (len_s[proc]) {
4678: merge->nsend++;
4679: nrows = 0;
4680: for (i=owners[proc]; i<owners[proc+1]; i++){
4681: if (ai[i+1] > ai[i]) nrows++;
4682: }
4683: len_si[proc] = 2*(nrows+1);
4684: len += len_si[proc];
4685: }
4686: }
4688: /* determine the number and length of messages to receive for ij-structure */
4689: /*-------------------------------------------------------------------------*/
4690: PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);
4691: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
4693: /* post the Irecv of j-structure */
4694: /*-------------------------------*/
4695: PetscCommGetNewTag(comm,&tagj);
4696: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
4698: /* post the Isend of j-structure */
4699: /*--------------------------------*/
4700: PetscMalloc2(merge->nsend,MPI_Request,&si_waits,merge->nsend,MPI_Request,&sj_waits);
4702: for (proc=0, k=0; proc<size; proc++){
4703: if (!len_s[proc]) continue;
4704: i = owners[proc];
4705: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4706: k++;
4707: }
4709: /* receives and sends of j-structure are complete */
4710: /*------------------------------------------------*/
4711: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4712: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
4714: /* send and recv i-structure */
4715: /*---------------------------*/
4716: PetscCommGetNewTag(comm,&tagi);
4717: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
4719: PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);
4720: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4721: for (proc=0,k=0; proc<size; proc++){
4722: if (!len_s[proc]) continue;
4723: /* form outgoing message for i-structure:
4724: buf_si[0]: nrows to be sent
4725: [1:nrows]: row index (global)
4726: [nrows+1:2*nrows+1]: i-structure index
4727: */
4728: /*-------------------------------------------*/
4729: nrows = len_si[proc]/2 - 1;
4730: buf_si_i = buf_si + nrows+1;
4731: buf_si[0] = nrows;
4732: buf_si_i[0] = 0;
4733: nrows = 0;
4734: for (i=owners[proc]; i<owners[proc+1]; i++){
4735: anzi = ai[i+1] - ai[i];
4736: if (anzi) {
4737: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4738: buf_si[nrows+1] = i-owners[proc]; /* local row index */
4739: nrows++;
4740: }
4741: }
4742: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4743: k++;
4744: buf_si += len_si[proc];
4745: }
4747: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4748: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
4750: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4751: for (i=0; i<merge->nrecv; i++){
4752: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4753: }
4755: PetscFree(len_si);
4756: PetscFree(len_ri);
4757: PetscFree(rj_waits);
4758: PetscFree2(si_waits,sj_waits);
4759: PetscFree(ri_waits);
4760: PetscFree(buf_s);
4761: PetscFree(status);
4763: /* compute a local seq matrix in each processor */
4764: /*----------------------------------------------*/
4765: /* allocate bi array and free space for accumulating nonzero column info */
4766: PetscMalloc((m+1)*sizeof(PetscInt),&bi);
4767: bi[0] = 0;
4769: /* create and initialize a linked list */
4770: nlnk = N+1;
4771: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
4773: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4774: len = 0;
4775: len = ai[owners[rank+1]] - ai[owners[rank]];
4776: PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);
4777: current_space = free_space;
4779: /* determine symbolic info for each local row */
4780: PetscMalloc3(merge->nrecv,PetscInt*,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextai);
4782: for (k=0; k<merge->nrecv; k++){
4783: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4784: nrows = *buf_ri_k[k];
4785: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4786: nextai[k] = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure */
4787: }
4789: MatPreallocateInitialize(comm,m,n,dnz,onz);
4790: len = 0;
4791: for (i=0;i<m;i++) {
4792: bnzi = 0;
4793: /* add local non-zero cols of this proc's seqmat into lnk */
4794: arow = owners[rank] + i;
4795: anzi = ai[arow+1] - ai[arow];
4796: aj = a->j + ai[arow];
4797: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4798: bnzi += nlnk;
4799: /* add received col data into lnk */
4800: for (k=0; k<merge->nrecv; k++){ /* k-th received message */
4801: if (i == *nextrow[k]) { /* i-th row */
4802: anzi = *(nextai[k]+1) - *nextai[k];
4803: aj = buf_rj[k] + *nextai[k];
4804: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4805: bnzi += nlnk;
4806: nextrow[k]++; nextai[k]++;
4807: }
4808: }
4809: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4811: /* if free space is not available, make more free space */
4812: if (current_space->local_remaining<bnzi) {
4813: PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);
4814: nspacedouble++;
4815: }
4816: /* copy data into free space, then initialize lnk */
4817: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4818: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
4820: current_space->array += bnzi;
4821: current_space->local_used += bnzi;
4822: current_space->local_remaining -= bnzi;
4824: bi[i+1] = bi[i] + bnzi;
4825: }
4827: PetscFree3(buf_ri_k,nextrow,nextai);
4829: PetscMalloc((bi[m]+1)*sizeof(PetscInt),&bj);
4830: PetscFreeSpaceContiguous(&free_space,bj);
4831: PetscLLDestroy(lnk,lnkbt);
4833: /* create symbolic parallel matrix B_mpi */
4834: /*---------------------------------------*/
4835: MatGetBlockSizes(seqmat,&bs,&cbs);
4836: MatCreate(comm,&B_mpi);
4837: if (n==PETSC_DECIDE) {
4838: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4839: } else {
4840: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4841: }
4842: MatSetBlockSizes(B_mpi,bs,cbs);
4843: MatSetType(B_mpi,MATMPIAIJ);
4844: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4845: MatPreallocateFinalize(dnz,onz);
4846: MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
4848: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4849: B_mpi->assembled = PETSC_FALSE;
4850: B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
4851: merge->bi = bi;
4852: merge->bj = bj;
4853: merge->buf_ri = buf_ri;
4854: merge->buf_rj = buf_rj;
4855: merge->coi = PETSC_NULL;
4856: merge->coj = PETSC_NULL;
4857: merge->owners_co = PETSC_NULL;
4859: PetscCommDestroy(&comm);
4861: /* attach the supporting struct to B_mpi for reuse */
4862: PetscContainerCreate(PETSC_COMM_SELF,&container);
4863: PetscContainerSetPointer(container,merge);
4864: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4865: PetscContainerDestroy(&container);
4866: *mpimat = B_mpi;
4868: PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4869: return(0);
4870: }
4874: /*@C
4875: MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential
4876: matrices from each processor
4878: Collective on MPI_Comm
4880: Input Parameters:
4881: + comm - the communicators the parallel matrix will live on
4882: . seqmat - the input sequential matrices
4883: . m - number of local rows (or PETSC_DECIDE)
4884: . n - number of local columns (or PETSC_DECIDE)
4885: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4887: Output Parameter:
4888: . mpimat - the parallel matrix generated
4890: Level: advanced
4892: Notes:
4893: The dimensions of the sequential matrix in each processor MUST be the same.
4894: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4895: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4896: @*/
4897: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4898: {
4899: PetscErrorCode ierr;
4900: PetscMPIInt size;
4903: MPI_Comm_size(comm,&size);
4904: if (size == 1){
4905: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4906: if (scall == MAT_INITIAL_MATRIX){
4907: MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4908: } else {
4909: MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4910: }
4911: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4912: return(0);
4913: }
4914: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4915: if (scall == MAT_INITIAL_MATRIX){
4916: MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4917: }
4918: MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4919: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4920: return(0);
4921: }
4925: /*@
4926: MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with
4927: mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4928: with MatGetSize()
4930: Not Collective
4932: Input Parameters:
4933: + A - the matrix
4934: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4936: Output Parameter:
4937: . A_loc - the local sequential matrix generated
4939: Level: developer
4941: .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed()
4943: @*/
4944: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4945: {
4946: PetscErrorCode ierr;
4947: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
4948: Mat_SeqAIJ *mat,*a=(Mat_SeqAIJ*)(mpimat->A)->data,*b=(Mat_SeqAIJ*)(mpimat->B)->data;
4949: PetscInt *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*cmap=mpimat->garray;
4950: MatScalar *aa=a->a,*ba=b->a,*cam;
4951: PetscScalar *ca;
4952: PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
4953: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
4954: PetscBool match;
4957: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
4958: if (!match) SETERRQ(((PetscObject)A)->comm, PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
4959: PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
4960: if (scall == MAT_INITIAL_MATRIX){
4961: PetscMalloc((1+am)*sizeof(PetscInt),&ci);
4962: ci[0] = 0;
4963: for (i=0; i<am; i++){
4964: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
4965: }
4966: PetscMalloc((1+ci[am])*sizeof(PetscInt),&cj);
4967: PetscMalloc((1+ci[am])*sizeof(PetscScalar),&ca);
4968: k = 0;
4969: for (i=0; i<am; i++) {
4970: ncols_o = bi[i+1] - bi[i];
4971: ncols_d = ai[i+1] - ai[i];
4972: /* off-diagonal portion of A */
4973: for (jo=0; jo<ncols_o; jo++) {
4974: col = cmap[*bj];
4975: if (col >= cstart) break;
4976: cj[k] = col; bj++;
4977: ca[k++] = *ba++;
4978: }
4979: /* diagonal portion of A */
4980: for (j=0; j<ncols_d; j++) {
4981: cj[k] = cstart + *aj++;
4982: ca[k++] = *aa++;
4983: }
4984: /* off-diagonal portion of A */
4985: for (j=jo; j<ncols_o; j++) {
4986: cj[k] = cmap[*bj++];
4987: ca[k++] = *ba++;
4988: }
4989: }
4990: /* put together the new matrix */
4991: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
4992: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4993: /* Since these are PETSc arrays, change flags to free them as necessary. */
4994: mat = (Mat_SeqAIJ*)(*A_loc)->data;
4995: mat->free_a = PETSC_TRUE;
4996: mat->free_ij = PETSC_TRUE;
4997: mat->nonew = 0;
4998: } else if (scall == MAT_REUSE_MATRIX){
4999: mat=(Mat_SeqAIJ*)(*A_loc)->data;
5000: ci = mat->i; cj = mat->j; cam = mat->a;
5001: for (i=0; i<am; i++) {
5002: /* off-diagonal portion of A */
5003: ncols_o = bi[i+1] - bi[i];
5004: for (jo=0; jo<ncols_o; jo++) {
5005: col = cmap[*bj];
5006: if (col >= cstart) break;
5007: *cam++ = *ba++; bj++;
5008: }
5009: /* diagonal portion of A */
5010: ncols_d = ai[i+1] - ai[i];
5011: for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5012: /* off-diagonal portion of A */
5013: for (j=jo; j<ncols_o; j++) {
5014: *cam++ = *ba++; bj++;
5015: }
5016: }
5017: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5018: PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5019: return(0);
5020: }
5024: /*@C
5025: MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns
5027: Not Collective
5029: Input Parameters:
5030: + A - the matrix
5031: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5032: - row, col - index sets of rows and columns to extract (or PETSC_NULL)
5034: Output Parameter:
5035: . A_loc - the local sequential matrix generated
5037: Level: developer
5039: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()
5041: @*/
5042: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5043: {
5044: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5045: PetscErrorCode ierr;
5046: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5047: IS isrowa,iscola;
5048: Mat *aloc;
5049: PetscBool match;
5052: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5053: if (!match) SETERRQ(((PetscObject)A)->comm, PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
5054: PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5055: if (!row){
5056: start = A->rmap->rstart; end = A->rmap->rend;
5057: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5058: } else {
5059: isrowa = *row;
5060: }
5061: if (!col){
5062: start = A->cmap->rstart;
5063: cmap = a->garray;
5064: nzA = a->A->cmap->n;
5065: nzB = a->B->cmap->n;
5066: PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
5067: ncols = 0;
5068: for (i=0; i<nzB; i++) {
5069: if (cmap[i] < start) idx[ncols++] = cmap[i];
5070: else break;
5071: }
5072: imark = i;
5073: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5074: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5075: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5076: } else {
5077: iscola = *col;
5078: }
5079: if (scall != MAT_INITIAL_MATRIX){
5080: PetscMalloc(sizeof(Mat),&aloc);
5081: aloc[0] = *A_loc;
5082: }
5083: MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5084: *A_loc = aloc[0];
5085: PetscFree(aloc);
5086: if (!row){
5087: ISDestroy(&isrowa);
5088: }
5089: if (!col){
5090: ISDestroy(&iscola);
5091: }
5092: PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5093: return(0);
5094: }
5098: /*@C
5099: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5101: Collective on Mat
5103: Input Parameters:
5104: + A,B - the matrices in mpiaij format
5105: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5106: - rowb, colb - index sets of rows and columns of B to extract (or PETSC_NULL)
5108: Output Parameter:
5109: + rowb, colb - index sets of rows and columns of B to extract
5110: - B_seq - the sequential matrix generated
5112: Level: developer
5114: @*/
5115: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5116: {
5117: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5118: PetscErrorCode ierr;
5119: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5120: IS isrowb,iscolb;
5121: Mat *bseq=PETSC_NULL;
5122:
5124: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){
5125: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5126: }
5127: PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5128:
5129: if (scall == MAT_INITIAL_MATRIX){
5130: start = A->cmap->rstart;
5131: cmap = a->garray;
5132: nzA = a->A->cmap->n;
5133: nzB = a->B->cmap->n;
5134: PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
5135: ncols = 0;
5136: for (i=0; i<nzB; i++) { /* row < local row index */
5137: if (cmap[i] < start) idx[ncols++] = cmap[i];
5138: else break;
5139: }
5140: imark = i;
5141: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
5142: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5143: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5144: ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5145: } else {
5146: if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5147: isrowb = *rowb; iscolb = *colb;
5148: PetscMalloc(sizeof(Mat),&bseq);
5149: bseq[0] = *B_seq;
5150: }
5151: MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5152: *B_seq = bseq[0];
5153: PetscFree(bseq);
5154: if (!rowb){
5155: ISDestroy(&isrowb);
5156: } else {
5157: *rowb = isrowb;
5158: }
5159: if (!colb){
5160: ISDestroy(&iscolb);
5161: } else {
5162: *colb = iscolb;
5163: }
5164: PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5165: return(0);
5166: }
5170: /*
5171: MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5172: of the OFF-DIAGONAL portion of local A
5174: Collective on Mat
5176: Input Parameters:
5177: + A,B - the matrices in mpiaij format
5178: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5180: Output Parameter:
5181: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or PETSC_NULL)
5182: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or PETSC_NULL)
5183: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or PETSC_NULL)
5184: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5186: Level: developer
5188: */
5189: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5190: {
5191: VecScatter_MPI_General *gen_to,*gen_from;
5192: PetscErrorCode ierr;
5193: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5194: Mat_SeqAIJ *b_oth;
5195: VecScatter ctx=a->Mvctx;
5196: MPI_Comm comm=((PetscObject)ctx)->comm;
5197: PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank;
5198: PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj;
5199: PetscScalar *rvalues,*svalues;
5200: MatScalar *b_otha,*bufa,*bufA;
5201: PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
5202: MPI_Request *rwaits = PETSC_NULL,*swaits = PETSC_NULL;
5203: MPI_Status *sstatus,rstatus;
5204: PetscMPIInt jj;
5205: PetscInt *cols,sbs,rbs;
5206: PetscScalar *vals;
5209: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){
5210: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5211: }
5212: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5213: MPI_Comm_rank(comm,&rank);
5215: gen_to = (VecScatter_MPI_General*)ctx->todata;
5216: gen_from = (VecScatter_MPI_General*)ctx->fromdata;
5217: rvalues = gen_from->values; /* holds the length of receiving row */
5218: svalues = gen_to->values; /* holds the length of sending row */
5219: nrecvs = gen_from->n;
5220: nsends = gen_to->n;
5222: PetscMalloc2(nrecvs,MPI_Request,&rwaits,nsends,MPI_Request,&swaits);
5223: srow = gen_to->indices; /* local row index to be sent */
5224: sstarts = gen_to->starts;
5225: sprocs = gen_to->procs;
5226: sstatus = gen_to->sstatus;
5227: sbs = gen_to->bs;
5228: rstarts = gen_from->starts;
5229: rprocs = gen_from->procs;
5230: rbs = gen_from->bs;
5232: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5233: if (scall == MAT_INITIAL_MATRIX){
5234: /* i-array */
5235: /*---------*/
5236: /* post receives */
5237: for (i=0; i<nrecvs; i++){
5238: rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
5239: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5240: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5241: }
5243: /* pack the outgoing message */
5244: PetscMalloc2(nsends+1,PetscInt,&sstartsj,nrecvs+1,PetscInt,&rstartsj);
5245: sstartsj[0] = 0; rstartsj[0] = 0;
5246: len = 0; /* total length of j or a array to be sent */
5247: k = 0;
5248: for (i=0; i<nsends; i++){
5249: rowlen = (PetscInt*)svalues + sstarts[i]*sbs;
5250: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5251: for (j=0; j<nrows; j++) {
5252: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5253: for (l=0; l<sbs; l++){
5254: MatGetRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL); /* rowlength */
5255: rowlen[j*sbs+l] = ncols;
5256: len += ncols;
5257: MatRestoreRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL);
5258: }
5259: k++;
5260: }
5261: MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5262: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5263: }
5264: /* recvs and sends of i-array are completed */
5265: i = nrecvs;
5266: while (i--) {
5267: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5268: }
5269: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5271: /* allocate buffers for sending j and a arrays */
5272: PetscMalloc((len+1)*sizeof(PetscInt),&bufj);
5273: PetscMalloc((len+1)*sizeof(PetscScalar),&bufa);
5275: /* create i-array of B_oth */
5276: PetscMalloc((aBn+2)*sizeof(PetscInt),&b_othi);
5277: b_othi[0] = 0;
5278: len = 0; /* total length of j or a array to be received */
5279: k = 0;
5280: for (i=0; i<nrecvs; i++){
5281: rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
5282: nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */
5283: for (j=0; j<nrows; j++) {
5284: b_othi[k+1] = b_othi[k] + rowlen[j];
5285: len += rowlen[j]; k++;
5286: }
5287: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5288: }
5290: /* allocate space for j and a arrrays of B_oth */
5291: PetscMalloc((b_othi[aBn]+1)*sizeof(PetscInt),&b_othj);
5292: PetscMalloc((b_othi[aBn]+1)*sizeof(MatScalar),&b_otha);
5294: /* j-array */
5295: /*---------*/
5296: /* post receives of j-array */
5297: for (i=0; i<nrecvs; i++){
5298: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5299: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5300: }
5302: /* pack the outgoing message j-array */
5303: k = 0;
5304: for (i=0; i<nsends; i++){
5305: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5306: bufJ = bufj+sstartsj[i];
5307: for (j=0; j<nrows; j++) {
5308: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5309: for (ll=0; ll<sbs; ll++){
5310: MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
5311: for (l=0; l<ncols; l++){
5312: *bufJ++ = cols[l];
5313: }
5314: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
5315: }
5316: }
5317: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5318: }
5320: /* recvs and sends of j-array are completed */
5321: i = nrecvs;
5322: while (i--) {
5323: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5324: }
5325: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5326: } else if (scall == MAT_REUSE_MATRIX){
5327: sstartsj = *startsj_s;
5328: rstartsj = *startsj_r;
5329: bufa = *bufa_ptr;
5330: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5331: b_otha = b_oth->a;
5332: } else {
5333: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5334: }
5336: /* a-array */
5337: /*---------*/
5338: /* post receives of a-array */
5339: for (i=0; i<nrecvs; i++){
5340: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5341: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5342: }
5344: /* pack the outgoing message a-array */
5345: k = 0;
5346: for (i=0; i<nsends; i++){
5347: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5348: bufA = bufa+sstartsj[i];
5349: for (j=0; j<nrows; j++) {
5350: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5351: for (ll=0; ll<sbs; ll++){
5352: MatGetRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
5353: for (l=0; l<ncols; l++){
5354: *bufA++ = vals[l];
5355: }
5356: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
5357: }
5358: }
5359: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5360: }
5361: /* recvs and sends of a-array are completed */
5362: i = nrecvs;
5363: while (i--) {
5364: MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5365: }
5366: if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5367: PetscFree2(rwaits,swaits);
5369: if (scall == MAT_INITIAL_MATRIX){
5370: /* put together the new matrix */
5371: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);
5373: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5374: /* Since these are PETSc arrays, change flags to free them as necessary. */
5375: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5376: b_oth->free_a = PETSC_TRUE;
5377: b_oth->free_ij = PETSC_TRUE;
5378: b_oth->nonew = 0;
5380: PetscFree(bufj);
5381: if (!startsj_s || !bufa_ptr){
5382: PetscFree2(sstartsj,rstartsj);
5383: PetscFree(bufa_ptr);
5384: } else {
5385: *startsj_s = sstartsj;
5386: *startsj_r = rstartsj;
5387: *bufa_ptr = bufa;
5388: }
5389: }
5390: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5391: return(0);
5392: }
5396: /*@C
5397: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
5399: Not Collective
5401: Input Parameters:
5402: . A - The matrix in mpiaij format
5404: Output Parameter:
5405: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5406: . colmap - A map from global column index to local index into lvec
5407: - multScatter - A scatter from the argument of a matrix-vector product to lvec
5409: Level: developer
5411: @*/
5412: #if defined (PETSC_USE_CTABLE)
5413: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5414: #else
5415: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5416: #endif
5417: {
5418: Mat_MPIAIJ *a;
5425: a = (Mat_MPIAIJ *) A->data;
5426: if (lvec) *lvec = a->lvec;
5427: if (colmap) *colmap = a->colmap;
5428: if (multScatter) *multScatter = a->Mvctx;
5429: return(0);
5430: }
5432: EXTERN_C_BEGIN
5433: extern PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,const MatType,MatReuse,Mat*);
5434: extern PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,const MatType,MatReuse,Mat*);
5435: extern PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,const MatType,MatReuse,Mat*);
5436: EXTERN_C_END
5440: /*
5441: Computes (B'*A')' since computing B*A directly is untenable
5443: n p p
5444: ( ) ( ) ( )
5445: m ( A ) * n ( B ) = m ( C )
5446: ( ) ( ) ( )
5448: */
5449: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5450: {
5451: PetscErrorCode ierr;
5452: Mat At,Bt,Ct;
5455: MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5456: MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5457: MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5458: MatDestroy(&At);
5459: MatDestroy(&Bt);
5460: MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5461: MatDestroy(&Ct);
5462: return(0);
5463: }
5467: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
5468: {
5470: PetscInt m=A->rmap->n,n=B->cmap->n;
5471: Mat Cmat;
5474: if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n);
5475: MatCreate(((PetscObject)A)->comm,&Cmat);
5476: MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5477: MatSetBlockSizes(Cmat,A->rmap->bs,B->cmap->bs);
5478: MatSetType(Cmat,MATMPIDENSE);
5479: MatMPIDenseSetPreallocation(Cmat,PETSC_NULL);
5480: MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5481: MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
5482: *C = Cmat;
5483: (*C)->ops->matmult = MatMatMult_MPIDense_MPIAIJ;
5484: return(0);
5485: }
5487: /* ----------------------------------------------------------------*/
5490: PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
5491: {
5495: if (scall == MAT_INITIAL_MATRIX){
5496: MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
5497: }
5498: MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
5499: return(0);
5500: }
5502: EXTERN_C_BEGIN
5503: #if defined(PETSC_HAVE_MUMPS)
5504: extern PetscErrorCode MatGetFactor_aij_mumps(Mat,MatFactorType,Mat*);
5505: #endif
5506: #if defined(PETSC_HAVE_PASTIX)
5507: extern PetscErrorCode MatGetFactor_mpiaij_pastix(Mat,MatFactorType,Mat*);
5508: #endif
5509: #if defined(PETSC_HAVE_SUPERLU_DIST)
5510: extern PetscErrorCode MatGetFactor_mpiaij_superlu_dist(Mat,MatFactorType,Mat*);
5511: #endif
5512: #if defined(PETSC_HAVE_SPOOLES)
5513: extern PetscErrorCode MatGetFactor_mpiaij_spooles(Mat,MatFactorType,Mat*);
5514: #endif
5515: EXTERN_C_END
5517: /*MC
5518: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
5520: Options Database Keys:
5521: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
5523: Level: beginner
5525: .seealso: MatCreateAIJ()
5526: M*/
5528: EXTERN_C_BEGIN
5531: PetscErrorCode MatCreate_MPIAIJ(Mat B)
5532: {
5533: Mat_MPIAIJ *b;
5535: PetscMPIInt size;
5538: MPI_Comm_size(((PetscObject)B)->comm,&size);
5540: PetscNewLog(B,Mat_MPIAIJ,&b);
5541: B->data = (void*)b;
5542: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5543: B->assembled = PETSC_FALSE;
5545: B->insertmode = NOT_SET_VALUES;
5546: b->size = size;
5547: MPI_Comm_rank(((PetscObject)B)->comm,&b->rank);
5549: /* build cache for off array entries formed */
5550: MatStashCreate_Private(((PetscObject)B)->comm,1,&B->stash);
5551: b->donotstash = PETSC_FALSE;
5552: b->colmap = 0;
5553: b->garray = 0;
5554: b->roworiented = PETSC_TRUE;
5556: /* stuff used for matrix vector multiply */
5557: b->lvec = PETSC_NULL;
5558: b->Mvctx = PETSC_NULL;
5560: /* stuff for MatGetRow() */
5561: b->rowindices = 0;
5562: b->rowvalues = 0;
5563: b->getrowactive = PETSC_FALSE;
5565: #if defined(PETSC_HAVE_SPOOLES)
5566: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_spooles_C",
5567: "MatGetFactor_mpiaij_spooles",
5568: MatGetFactor_mpiaij_spooles);
5569: #endif
5570: #if defined(PETSC_HAVE_MUMPS)
5571: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_mumps_C",
5572: "MatGetFactor_aij_mumps",
5573: MatGetFactor_aij_mumps);
5574: #endif
5575: #if defined(PETSC_HAVE_PASTIX)
5576: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_pastix_C",
5577: "MatGetFactor_mpiaij_pastix",
5578: MatGetFactor_mpiaij_pastix);
5579: #endif
5580: #if defined(PETSC_HAVE_SUPERLU_DIST)
5581: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_superlu_dist_C",
5582: "MatGetFactor_mpiaij_superlu_dist",
5583: MatGetFactor_mpiaij_superlu_dist);
5584: #endif
5585: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
5586: "MatStoreValues_MPIAIJ",
5587: MatStoreValues_MPIAIJ);
5588: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
5589: "MatRetrieveValues_MPIAIJ",
5590: MatRetrieveValues_MPIAIJ);
5591: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
5592: "MatGetDiagonalBlock_MPIAIJ",
5593: MatGetDiagonalBlock_MPIAIJ);
5594: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatIsTranspose_C",
5595: "MatIsTranspose_MPIAIJ",
5596: MatIsTranspose_MPIAIJ);
5597: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocation_C",
5598: "MatMPIAIJSetPreallocation_MPIAIJ",
5599: MatMPIAIJSetPreallocation_MPIAIJ);
5600: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",
5601: "MatMPIAIJSetPreallocationCSR_MPIAIJ",
5602: MatMPIAIJSetPreallocationCSR_MPIAIJ);
5603: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
5604: "MatDiagonalScaleLocal_MPIAIJ",
5605: MatDiagonalScaleLocal_MPIAIJ);
5606: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",
5607: "MatConvert_MPIAIJ_MPIAIJPERM",
5608: MatConvert_MPIAIJ_MPIAIJPERM);
5609: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",
5610: "MatConvert_MPIAIJ_MPIAIJCRL",
5611: MatConvert_MPIAIJ_MPIAIJCRL);
5612: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",
5613: "MatConvert_MPIAIJ_MPISBAIJ",
5614: MatConvert_MPIAIJ_MPISBAIJ);
5615: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",
5616: "MatMatMult_MPIDense_MPIAIJ",
5617: MatMatMult_MPIDense_MPIAIJ);
5618: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",
5619: "MatMatMultSymbolic_MPIDense_MPIAIJ",
5620: MatMatMultSymbolic_MPIDense_MPIAIJ);
5621: PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",
5622: "MatMatMultNumeric_MPIDense_MPIAIJ",
5623: MatMatMultNumeric_MPIDense_MPIAIJ);
5624: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5625: return(0);
5626: }
5627: EXTERN_C_END
5631: /*@
5632: MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5633: and "off-diagonal" part of the matrix in CSR format.
5635: Collective on MPI_Comm
5637: Input Parameters:
5638: + comm - MPI communicator
5639: . m - number of local rows (Cannot be PETSC_DECIDE)
5640: . n - This value should be the same as the local size used in creating the
5641: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5642: calculated if N is given) For square matrices n is almost always m.
5643: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5644: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5645: . i - row indices for "diagonal" portion of matrix
5646: . j - column indices
5647: . a - matrix values
5648: . oi - row indices for "off-diagonal" portion of matrix
5649: . oj - column indices
5650: - oa - matrix values
5652: Output Parameter:
5653: . mat - the matrix
5655: Level: advanced
5657: Notes:
5658: The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5659: must free the arrays once the matrix has been destroyed and not before.
5661: The i and j indices are 0 based
5662:
5663: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix
5665: This sets local rows and cannot be used to set off-processor values.
5667: You cannot later use MatSetValues() to change values in this matrix.
5669: .keywords: matrix, aij, compressed row, sparse, parallel
5671: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5672: MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5673: @*/
5674: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],
5675: PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
5676: {
5678: Mat_MPIAIJ *maij;
5681: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5682: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5683: if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5684: MatCreate(comm,mat);
5685: MatSetSizes(*mat,m,n,M,N);
5686: MatSetType(*mat,MATMPIAIJ);
5687: maij = (Mat_MPIAIJ*) (*mat)->data;
5688: maij->donotstash = PETSC_TRUE;
5689: (*mat)->preallocated = PETSC_TRUE;
5691: PetscLayoutSetUp((*mat)->rmap);
5692: PetscLayoutSetUp((*mat)->cmap);
5694: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5695: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);
5697: MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5698: MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5699: MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5700: MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);
5702: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5703: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5704: return(0);
5705: }
5707: /*
5708: Special version for direct calls from Fortran
5709: */
5710: #include <petsc-private/fortranimpl.h>
5712: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5713: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
5714: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5715: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
5716: #endif
5718: /* Change these macros so can be used in void function */
5719: #undef CHKERRQ
5720: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
5721: #undef SETERRQ2
5722: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5723: #undef SETERRQ3
5724: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5725: #undef SETERRQ
5726: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)
5728: EXTERN_C_BEGIN
5731: void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
5732: {
5733: Mat mat = *mmat;
5734: PetscInt m = *mm, n = *mn;
5735: InsertMode addv = *maddv;
5736: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
5737: PetscScalar value;
5738: PetscErrorCode ierr;
5740: MatCheckPreallocated(mat,1);
5741: if (mat->insertmode == NOT_SET_VALUES) {
5742: mat->insertmode = addv;
5743: }
5744: #if defined(PETSC_USE_DEBUG)
5745: else if (mat->insertmode != addv) {
5746: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
5747: }
5748: #endif
5749: {
5750: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
5751: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
5752: PetscBool roworiented = aij->roworiented;
5754: /* Some Variables required in the macro */
5755: Mat A = aij->A;
5756: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
5757: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
5758: MatScalar *aa = a->a;
5759: PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
5760: Mat B = aij->B;
5761: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
5762: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
5763: MatScalar *ba = b->a;
5765: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
5766: PetscInt nonew = a->nonew;
5767: MatScalar *ap1,*ap2;
5770: for (i=0; i<m; i++) {
5771: if (im[i] < 0) continue;
5772: #if defined(PETSC_USE_DEBUG)
5773: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
5774: #endif
5775: if (im[i] >= rstart && im[i] < rend) {
5776: row = im[i] - rstart;
5777: lastcol1 = -1;
5778: rp1 = aj + ai[row];
5779: ap1 = aa + ai[row];
5780: rmax1 = aimax[row];
5781: nrow1 = ailen[row];
5782: low1 = 0;
5783: high1 = nrow1;
5784: lastcol2 = -1;
5785: rp2 = bj + bi[row];
5786: ap2 = ba + bi[row];
5787: rmax2 = bimax[row];
5788: nrow2 = bilen[row];
5789: low2 = 0;
5790: high2 = nrow2;
5792: for (j=0; j<n; j++) {
5793: if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
5794: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
5795: if (in[j] >= cstart && in[j] < cend){
5796: col = in[j] - cstart;
5797: MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
5798: } else if (in[j] < 0) continue;
5799: #if defined(PETSC_USE_DEBUG)
5800: else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
5801: #endif
5802: else {
5803: if (mat->was_assembled) {
5804: if (!aij->colmap) {
5805: MatCreateColmap_MPIAIJ_Private(mat);
5806: }
5807: #if defined (PETSC_USE_CTABLE)
5808: PetscTableFind(aij->colmap,in[j]+1,&col);
5809: col--;
5810: #else
5811: col = aij->colmap[in[j]] - 1;
5812: #endif
5813: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
5814: MatDisAssemble_MPIAIJ(mat);
5815: col = in[j];
5816: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
5817: B = aij->B;
5818: b = (Mat_SeqAIJ*)B->data;
5819: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
5820: rp2 = bj + bi[row];
5821: ap2 = ba + bi[row];
5822: rmax2 = bimax[row];
5823: nrow2 = bilen[row];
5824: low2 = 0;
5825: high2 = nrow2;
5826: bm = aij->B->rmap->n;
5827: ba = b->a;
5828: }
5829: } else col = in[j];
5830: MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
5831: }
5832: }
5833: } else {
5834: if (!aij->donotstash) {
5835: if (roworiented) {
5836: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5837: } else {
5838: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5839: }
5840: }
5841: }
5842: }}
5843: PetscFunctionReturnVoid();
5844: }
5845: EXTERN_C_END