Actual source code: mpiaij.c
petsc-3.12.5 2020-03-29
1: #include <../src/mat/impls/aij/mpi/mpiaij.h>
2: #include <petsc/private/vecimpl.h>
3: #include <petsc/private/vecscatterimpl.h>
4: #include <petsc/private/isimpl.h>
5: #include <petscblaslapack.h>
6: #include <petscsf.h>
7: #include <petsc/private/hashmapi.h>
9: /*MC
10: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
12: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
13: and MATMPIAIJ otherwise. As a result, for single process communicators,
14: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
15: for communicators controlling multiple processes. It is recommended that you call both of
16: the above preallocation routines for simplicity.
18: Options Database Keys:
19: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
21: Developer Notes:
22: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
23: enough exist.
25: Level: beginner
27: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ
28: M*/
30: /*MC
31: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
33: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
34: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
35: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
36: for communicators controlling multiple processes. It is recommended that you call both of
37: the above preallocation routines for simplicity.
39: Options Database Keys:
40: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
42: Level: beginner
44: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
45: M*/
47: static PetscErrorCode MatPinToCPU_MPIAIJ(Mat A,PetscBool flg)
48: {
49: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
53: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL)
54: A->pinnedtocpu = flg;
55: #endif
56: if (a->A) {
57: MatPinToCPU(a->A,flg);
58: }
59: if (a->B) {
60: MatPinToCPU(a->B,flg);
61: }
62: return(0);
63: }
66: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
67: {
69: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
72: if (mat->A) {
73: MatSetBlockSizes(mat->A,rbs,cbs);
74: MatSetBlockSizes(mat->B,rbs,1);
75: }
76: return(0);
77: }
79: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
80: {
81: PetscErrorCode ierr;
82: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
83: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data;
84: Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data;
85: const PetscInt *ia,*ib;
86: const MatScalar *aa,*bb;
87: PetscInt na,nb,i,j,*rows,cnt=0,n0rows;
88: PetscInt m = M->rmap->n,rstart = M->rmap->rstart;
91: *keptrows = 0;
92: ia = a->i;
93: ib = b->i;
94: for (i=0; i<m; i++) {
95: na = ia[i+1] - ia[i];
96: nb = ib[i+1] - ib[i];
97: if (!na && !nb) {
98: cnt++;
99: goto ok1;
100: }
101: aa = a->a + ia[i];
102: for (j=0; j<na; j++) {
103: if (aa[j] != 0.0) goto ok1;
104: }
105: bb = b->a + ib[i];
106: for (j=0; j <nb; j++) {
107: if (bb[j] != 0.0) goto ok1;
108: }
109: cnt++;
110: ok1:;
111: }
112: MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
113: if (!n0rows) return(0);
114: PetscMalloc1(M->rmap->n-cnt,&rows);
115: cnt = 0;
116: for (i=0; i<m; i++) {
117: na = ia[i+1] - ia[i];
118: nb = ib[i+1] - ib[i];
119: if (!na && !nb) continue;
120: aa = a->a + ia[i];
121: for (j=0; j<na;j++) {
122: if (aa[j] != 0.0) {
123: rows[cnt++] = rstart + i;
124: goto ok2;
125: }
126: }
127: bb = b->a + ib[i];
128: for (j=0; j<nb; j++) {
129: if (bb[j] != 0.0) {
130: rows[cnt++] = rstart + i;
131: goto ok2;
132: }
133: }
134: ok2:;
135: }
136: ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
137: return(0);
138: }
140: PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
141: {
142: PetscErrorCode ierr;
143: Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data;
144: PetscBool cong;
147: MatHasCongruentLayouts(Y,&cong);
148: if (Y->assembled && cong) {
149: MatDiagonalSet(aij->A,D,is);
150: } else {
151: MatDiagonalSet_Default(Y,D,is);
152: }
153: return(0);
154: }
156: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
157: {
158: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data;
160: PetscInt i,rstart,nrows,*rows;
163: *zrows = NULL;
164: MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
165: MatGetOwnershipRange(M,&rstart,NULL);
166: for (i=0; i<nrows; i++) rows[i] += rstart;
167: ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
168: return(0);
169: }
171: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
172: {
174: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
175: PetscInt i,n,*garray = aij->garray;
176: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data;
177: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data;
178: PetscReal *work;
181: MatGetSize(A,NULL,&n);
182: PetscCalloc1(n,&work);
183: if (type == NORM_2) {
184: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
185: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
186: }
187: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
188: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
189: }
190: } else if (type == NORM_1) {
191: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
192: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
193: }
194: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
195: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
196: }
197: } else if (type == NORM_INFINITY) {
198: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
199: work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
200: }
201: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
202: work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
203: }
205: } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
206: if (type == NORM_INFINITY) {
207: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
208: } else {
209: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
210: }
211: PetscFree(work);
212: if (type == NORM_2) {
213: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
214: }
215: return(0);
216: }
218: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
219: {
220: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
221: IS sis,gis;
222: PetscErrorCode ierr;
223: const PetscInt *isis,*igis;
224: PetscInt n,*iis,nsis,ngis,rstart,i;
227: MatFindOffBlockDiagonalEntries(a->A,&sis);
228: MatFindNonzeroRows(a->B,&gis);
229: ISGetSize(gis,&ngis);
230: ISGetSize(sis,&nsis);
231: ISGetIndices(sis,&isis);
232: ISGetIndices(gis,&igis);
234: PetscMalloc1(ngis+nsis,&iis);
235: PetscArraycpy(iis,igis,ngis);
236: PetscArraycpy(iis+ngis,isis,nsis);
237: n = ngis + nsis;
238: PetscSortRemoveDupsInt(&n,iis);
239: MatGetOwnershipRange(A,&rstart,NULL);
240: for (i=0; i<n; i++) iis[i] += rstart;
241: ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);
243: ISRestoreIndices(sis,&isis);
244: ISRestoreIndices(gis,&igis);
245: ISDestroy(&sis);
246: ISDestroy(&gis);
247: return(0);
248: }
250: /*
251: Distributes a SeqAIJ matrix across a set of processes. Code stolen from
252: MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.
254: Only for square matrices
256: Used by a preconditioner, hence PETSC_EXTERN
257: */
258: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
259: {
260: PetscMPIInt rank,size;
261: PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
263: Mat mat;
264: Mat_SeqAIJ *gmata;
265: PetscMPIInt tag;
266: MPI_Status status;
267: PetscBool aij;
268: MatScalar *gmataa,*ao,*ad,*gmataarestore=0;
271: MPI_Comm_rank(comm,&rank);
272: MPI_Comm_size(comm,&size);
273: if (!rank) {
274: PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
275: if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
276: }
277: if (reuse == MAT_INITIAL_MATRIX) {
278: MatCreate(comm,&mat);
279: MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
280: MatGetBlockSizes(gmat,&bses[0],&bses[1]);
281: MPI_Bcast(bses,2,MPIU_INT,0,comm);
282: MatSetBlockSizes(mat,bses[0],bses[1]);
283: MatSetType(mat,MATAIJ);
284: PetscMalloc1(size+1,&rowners);
285: PetscMalloc2(m,&dlens,m,&olens);
286: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
288: rowners[0] = 0;
289: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
290: rstart = rowners[rank];
291: rend = rowners[rank+1];
292: PetscObjectGetNewTag((PetscObject)mat,&tag);
293: if (!rank) {
294: gmata = (Mat_SeqAIJ*) gmat->data;
295: /* send row lengths to all processors */
296: for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
297: for (i=1; i<size; i++) {
298: MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
299: }
300: /* determine number diagonal and off-diagonal counts */
301: PetscArrayzero(olens,m);
302: PetscCalloc1(m,&ld);
303: jj = 0;
304: for (i=0; i<m; i++) {
305: for (j=0; j<dlens[i]; j++) {
306: if (gmata->j[jj] < rstart) ld[i]++;
307: if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
308: jj++;
309: }
310: }
311: /* send column indices to other processes */
312: for (i=1; i<size; i++) {
313: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
314: MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
315: MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
316: }
318: /* send numerical values to other processes */
319: for (i=1; i<size; i++) {
320: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
321: MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
322: }
323: gmataa = gmata->a;
324: gmataj = gmata->j;
326: } else {
327: /* receive row lengths */
328: MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
329: /* receive column indices */
330: MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
331: PetscMalloc2(nz,&gmataa,nz,&gmataj);
332: MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
333: /* determine number diagonal and off-diagonal counts */
334: PetscArrayzero(olens,m);
335: PetscCalloc1(m,&ld);
336: jj = 0;
337: for (i=0; i<m; i++) {
338: for (j=0; j<dlens[i]; j++) {
339: if (gmataj[jj] < rstart) ld[i]++;
340: if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
341: jj++;
342: }
343: }
344: /* receive numerical values */
345: PetscArrayzero(gmataa,nz);
346: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
347: }
348: /* set preallocation */
349: for (i=0; i<m; i++) {
350: dlens[i] -= olens[i];
351: }
352: MatSeqAIJSetPreallocation(mat,0,dlens);
353: MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
355: for (i=0; i<m; i++) {
356: dlens[i] += olens[i];
357: }
358: cnt = 0;
359: for (i=0; i<m; i++) {
360: row = rstart + i;
361: MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
362: cnt += dlens[i];
363: }
364: if (rank) {
365: PetscFree2(gmataa,gmataj);
366: }
367: PetscFree2(dlens,olens);
368: PetscFree(rowners);
370: ((Mat_MPIAIJ*)(mat->data))->ld = ld;
372: *inmat = mat;
373: } else { /* column indices are already set; only need to move over numerical values from process 0 */
374: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
375: Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
376: mat = *inmat;
377: PetscObjectGetNewTag((PetscObject)mat,&tag);
378: if (!rank) {
379: /* send numerical values to other processes */
380: gmata = (Mat_SeqAIJ*) gmat->data;
381: MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
382: gmataa = gmata->a;
383: for (i=1; i<size; i++) {
384: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
385: MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
386: }
387: nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
388: } else {
389: /* receive numerical values from process 0*/
390: nz = Ad->nz + Ao->nz;
391: PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
392: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
393: }
394: /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
395: ld = ((Mat_MPIAIJ*)(mat->data))->ld;
396: ad = Ad->a;
397: ao = Ao->a;
398: if (mat->rmap->n) {
399: i = 0;
400: nz = ld[i]; PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
401: nz = Ad->i[i+1] - Ad->i[i]; PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
402: }
403: for (i=1; i<mat->rmap->n; i++) {
404: nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
405: nz = Ad->i[i+1] - Ad->i[i]; PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
406: }
407: i--;
408: if (mat->rmap->n) {
409: nz = Ao->i[i+1] - Ao->i[i] - ld[i]; PetscArraycpy(ao,gmataa,nz);
410: }
411: if (rank) {
412: PetscFree(gmataarestore);
413: }
414: }
415: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
416: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
417: return(0);
418: }
420: /*
421: Local utility routine that creates a mapping from the global column
422: number to the local number in the off-diagonal part of the local
423: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
424: a slightly higher hash table cost; without it it is not scalable (each processor
425: has an order N integer array but is fast to acess.
426: */
427: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
428: {
429: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
431: PetscInt n = aij->B->cmap->n,i;
434: if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
435: #if defined(PETSC_USE_CTABLE)
436: PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
437: for (i=0; i<n; i++) {
438: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
439: }
440: #else
441: PetscCalloc1(mat->cmap->N+1,&aij->colmap);
442: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
443: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
444: #endif
445: return(0);
446: }
448: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \
449: { \
450: if (col <= lastcol1) low1 = 0; \
451: else high1 = nrow1; \
452: lastcol1 = col;\
453: while (high1-low1 > 5) { \
454: t = (low1+high1)/2; \
455: if (rp1[t] > col) high1 = t; \
456: else low1 = t; \
457: } \
458: for (_i=low1; _i<high1; _i++) { \
459: if (rp1[_i] > col) break; \
460: if (rp1[_i] == col) { \
461: if (addv == ADD_VALUES) { \
462: ap1[_i] += value; \
463: /* Not sure LogFlops will slow dow the code or not */ \
464: (void)PetscLogFlops(1.0); \
465: } \
466: else ap1[_i] = value; \
467: inserted = PETSC_TRUE; \
468: goto a_noinsert; \
469: } \
470: } \
471: if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
472: if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \
473: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
474: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
475: N = nrow1++ - 1; a->nz++; high1++; \
476: /* shift up all the later entries in this row */ \
477: PetscArraymove(rp1+_i+1,rp1+_i,N-_i+1);\
478: PetscArraymove(ap1+_i+1,ap1+_i,N-_i+1);\
479: rp1[_i] = col; \
480: ap1[_i] = value; \
481: A->nonzerostate++;\
482: a_noinsert: ; \
483: ailen[row] = nrow1; \
484: }
486: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
487: { \
488: if (col <= lastcol2) low2 = 0; \
489: else high2 = nrow2; \
490: lastcol2 = col; \
491: while (high2-low2 > 5) { \
492: t = (low2+high2)/2; \
493: if (rp2[t] > col) high2 = t; \
494: else low2 = t; \
495: } \
496: for (_i=low2; _i<high2; _i++) { \
497: if (rp2[_i] > col) break; \
498: if (rp2[_i] == col) { \
499: if (addv == ADD_VALUES) { \
500: ap2[_i] += value; \
501: (void)PetscLogFlops(1.0); \
502: } \
503: else ap2[_i] = value; \
504: inserted = PETSC_TRUE; \
505: goto b_noinsert; \
506: } \
507: } \
508: if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
509: if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
510: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
511: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
512: N = nrow2++ - 1; b->nz++; high2++; \
513: /* shift up all the later entries in this row */ \
514: PetscArraymove(rp2+_i+1,rp2+_i,N-_i+1);\
515: PetscArraymove(ap2+_i+1,ap2+_i,N-_i+1);\
516: rp2[_i] = col; \
517: ap2[_i] = value; \
518: B->nonzerostate++; \
519: b_noinsert: ; \
520: bilen[row] = nrow2; \
521: }
523: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
524: {
525: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
526: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
528: PetscInt l,*garray = mat->garray,diag;
531: /* code only works for square matrices A */
533: /* find size of row to the left of the diagonal part */
534: MatGetOwnershipRange(A,&diag,0);
535: row = row - diag;
536: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
537: if (garray[b->j[b->i[row]+l]] > diag) break;
538: }
539: PetscArraycpy(b->a+b->i[row],v,l);
541: /* diagonal part */
542: PetscArraycpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row]));
544: /* right of diagonal part */
545: PetscArraycpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],b->i[row+1]-b->i[row]-l);
546: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
547: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && (l || (a->i[row+1]-a->i[row]) || (b->i[row+1]-b->i[row]-l))) A->offloadmask = PETSC_OFFLOAD_CPU;
548: #endif
549: return(0);
550: }
552: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
553: {
554: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
555: PetscScalar value = 0.0;
557: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
558: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
559: PetscBool roworiented = aij->roworiented;
561: /* Some Variables required in the macro */
562: Mat A = aij->A;
563: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
564: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
565: MatScalar *aa = a->a;
566: PetscBool ignorezeroentries = a->ignorezeroentries;
567: Mat B = aij->B;
568: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
569: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
570: MatScalar *ba = b->a;
571: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
572: * cannot use "#if defined" inside a macro. */
573: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
575: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
576: PetscInt nonew;
577: MatScalar *ap1,*ap2;
580: for (i=0; i<m; i++) {
581: if (im[i] < 0) continue;
582: #if defined(PETSC_USE_DEBUG)
583: 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);
584: #endif
585: if (im[i] >= rstart && im[i] < rend) {
586: row = im[i] - rstart;
587: lastcol1 = -1;
588: rp1 = aj + ai[row];
589: ap1 = aa + ai[row];
590: rmax1 = aimax[row];
591: nrow1 = ailen[row];
592: low1 = 0;
593: high1 = nrow1;
594: lastcol2 = -1;
595: rp2 = bj + bi[row];
596: ap2 = ba + bi[row];
597: rmax2 = bimax[row];
598: nrow2 = bilen[row];
599: low2 = 0;
600: high2 = nrow2;
602: for (j=0; j<n; j++) {
603: if (v) value = roworiented ? v[i*n+j] : v[i+j*m];
604: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
605: if (in[j] >= cstart && in[j] < cend) {
606: col = in[j] - cstart;
607: nonew = a->nonew;
608: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
609: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
610: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
611: #endif
612: } else if (in[j] < 0) continue;
613: #if defined(PETSC_USE_DEBUG)
614: 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);
615: #endif
616: else {
617: if (mat->was_assembled) {
618: if (!aij->colmap) {
619: MatCreateColmap_MPIAIJ_Private(mat);
620: }
621: #if defined(PETSC_USE_CTABLE)
622: PetscTableFind(aij->colmap,in[j]+1,&col);
623: col--;
624: #else
625: col = aij->colmap[in[j]] - 1;
626: #endif
627: if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
628: MatDisAssemble_MPIAIJ(mat);
629: col = in[j];
630: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
631: B = aij->B;
632: b = (Mat_SeqAIJ*)B->data;
633: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
634: rp2 = bj + bi[row];
635: ap2 = ba + bi[row];
636: rmax2 = bimax[row];
637: nrow2 = bilen[row];
638: low2 = 0;
639: high2 = nrow2;
640: bm = aij->B->rmap->n;
641: ba = b->a;
642: inserted = PETSC_FALSE;
643: } else if (col < 0) {
644: if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) {
645: PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);
646: } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]);
647: }
648: } else col = in[j];
649: nonew = b->nonew;
650: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
651: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
652: if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
653: #endif
654: }
655: }
656: } else {
657: 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]);
658: if (!aij->donotstash) {
659: mat->assembled = PETSC_FALSE;
660: if (roworiented) {
661: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
662: } else {
663: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
664: }
665: }
666: }
667: }
668: return(0);
669: }
671: /*
672: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
673: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
674: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
675: */
676: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[])
677: {
678: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
679: Mat A = aij->A; /* diagonal part of the matrix */
680: Mat B = aij->B; /* offdiagonal part of the matrix */
681: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
682: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
683: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col;
684: PetscInt *ailen = a->ilen,*aj = a->j;
685: PetscInt *bilen = b->ilen,*bj = b->j;
686: PetscInt am = aij->A->rmap->n,j;
687: PetscInt diag_so_far = 0,dnz;
688: PetscInt offd_so_far = 0,onz;
691: /* Iterate over all rows of the matrix */
692: for (j=0; j<am; j++) {
693: dnz = onz = 0;
694: /* Iterate over all non-zero columns of the current row */
695: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
696: /* If column is in the diagonal */
697: if (mat_j[col] >= cstart && mat_j[col] < cend) {
698: aj[diag_so_far++] = mat_j[col] - cstart;
699: dnz++;
700: } else { /* off-diagonal entries */
701: bj[offd_so_far++] = mat_j[col];
702: onz++;
703: }
704: }
705: ailen[j] = dnz;
706: bilen[j] = onz;
707: }
708: return(0);
709: }
711: /*
712: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
713: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
714: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
715: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
716: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
717: */
718: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[])
719: {
720: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
721: Mat A = aij->A; /* diagonal part of the matrix */
722: Mat B = aij->B; /* offdiagonal part of the matrix */
723: Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data;
724: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
725: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
726: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend;
727: PetscInt *ailen = a->ilen,*aj = a->j;
728: PetscInt *bilen = b->ilen,*bj = b->j;
729: PetscInt am = aij->A->rmap->n,j;
730: PetscInt *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */
731: PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd;
732: PetscScalar *aa = a->a,*ba = b->a;
735: /* Iterate over all rows of the matrix */
736: for (j=0; j<am; j++) {
737: dnz_row = onz_row = 0;
738: rowstart_offd = full_offd_i[j];
739: rowstart_diag = full_diag_i[j];
740: /* Iterate over all non-zero columns of the current row */
741: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
742: /* If column is in the diagonal */
743: if (mat_j[col] >= cstart && mat_j[col] < cend) {
744: aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
745: aa[rowstart_diag+dnz_row] = mat_a[col];
746: dnz_row++;
747: } else { /* off-diagonal entries */
748: bj[rowstart_offd+onz_row] = mat_j[col];
749: ba[rowstart_offd+onz_row] = mat_a[col];
750: onz_row++;
751: }
752: }
753: ailen[j] = dnz_row;
754: bilen[j] = onz_row;
755: }
756: return(0);
757: }
759: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
760: {
761: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
763: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
764: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
767: for (i=0; i<m; i++) {
768: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
769: 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);
770: if (idxm[i] >= rstart && idxm[i] < rend) {
771: row = idxm[i] - rstart;
772: for (j=0; j<n; j++) {
773: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
774: 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);
775: if (idxn[j] >= cstart && idxn[j] < cend) {
776: col = idxn[j] - cstart;
777: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
778: } else {
779: if (!aij->colmap) {
780: MatCreateColmap_MPIAIJ_Private(mat);
781: }
782: #if defined(PETSC_USE_CTABLE)
783: PetscTableFind(aij->colmap,idxn[j]+1,&col);
784: col--;
785: #else
786: col = aij->colmap[idxn[j]] - 1;
787: #endif
788: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
789: else {
790: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
791: }
792: }
793: }
794: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
795: }
796: return(0);
797: }
799: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);
801: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
802: {
803: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
805: PetscInt nstash,reallocs;
808: if (aij->donotstash || mat->nooffprocentries) return(0);
810: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
811: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
812: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
813: return(0);
814: }
816: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
817: {
818: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
819: Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data;
821: PetscMPIInt n;
822: PetscInt i,j,rstart,ncols,flg;
823: PetscInt *row,*col;
824: PetscBool other_disassembled;
825: PetscScalar *val;
827: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
830: if (!aij->donotstash && !mat->nooffprocentries) {
831: while (1) {
832: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
833: if (!flg) break;
835: for (i=0; i<n; ) {
836: /* Now identify the consecutive vals belonging to the same row */
837: for (j=i,rstart=row[j]; j<n; j++) {
838: if (row[j] != rstart) break;
839: }
840: if (j < n) ncols = j-i;
841: else ncols = n-i;
842: /* Now assemble all these values with a single function call */
843: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);
845: i = j;
846: }
847: }
848: MatStashScatterEnd_Private(&mat->stash);
849: }
850: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
851: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
852: #endif
853: MatAssemblyBegin(aij->A,mode);
854: MatAssemblyEnd(aij->A,mode);
856: /* determine if any processor has disassembled, if so we must
857: also disassemble ourself, in order that we may reassemble. */
858: /*
859: if nonzero structure of submatrix B cannot change then we know that
860: no processor disassembled thus we can skip this stuff
861: */
862: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
863: MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
864: if (mat->was_assembled && !other_disassembled) {
865: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
866: aij->B->offloadmask = PETSC_OFFLOAD_BOTH; /* do not copy on the GPU when assembling inside MatDisAssemble_MPIAIJ */
867: #endif
868: MatDisAssemble_MPIAIJ(mat);
869: }
870: }
871: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
872: MatSetUpMultiply_MPIAIJ(mat);
873: }
874: MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
875: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
876: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
877: #endif
878: MatAssemblyBegin(aij->B,mode);
879: MatAssemblyEnd(aij->B,mode);
881: PetscFree2(aij->rowvalues,aij->rowindices);
883: aij->rowvalues = 0;
885: VecDestroy(&aij->diag);
886: if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
888: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
889: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
890: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
891: MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
892: }
893: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
894: mat->offloadmask = PETSC_OFFLOAD_BOTH;
895: #endif
896: return(0);
897: }
899: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
900: {
901: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
905: MatZeroEntries(l->A);
906: MatZeroEntries(l->B);
907: return(0);
908: }
910: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
911: {
912: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
913: PetscObjectState sA, sB;
914: PetscInt *lrows;
915: PetscInt r, len;
916: PetscBool cong, lch, gch;
917: PetscErrorCode ierr;
920: /* get locally owned rows */
921: MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);
922: MatHasCongruentLayouts(A,&cong);
923: /* fix right hand side if needed */
924: if (x && b) {
925: const PetscScalar *xx;
926: PetscScalar *bb;
928: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
929: VecGetArrayRead(x, &xx);
930: VecGetArray(b, &bb);
931: for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
932: VecRestoreArrayRead(x, &xx);
933: VecRestoreArray(b, &bb);
934: }
936: sA = mat->A->nonzerostate;
937: sB = mat->B->nonzerostate;
939: if (diag != 0.0 && cong) {
940: MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
941: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
942: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
943: Mat_SeqAIJ *aijA = (Mat_SeqAIJ*)mat->A->data;
944: Mat_SeqAIJ *aijB = (Mat_SeqAIJ*)mat->B->data;
945: PetscInt nnwA, nnwB;
946: PetscBool nnzA, nnzB;
948: nnwA = aijA->nonew;
949: nnwB = aijB->nonew;
950: nnzA = aijA->keepnonzeropattern;
951: nnzB = aijB->keepnonzeropattern;
952: if (!nnzA) {
953: PetscInfo(mat->A,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n");
954: aijA->nonew = 0;
955: }
956: if (!nnzB) {
957: PetscInfo(mat->B,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n");
958: aijB->nonew = 0;
959: }
960: /* Must zero here before the next loop */
961: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
962: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
963: for (r = 0; r < len; ++r) {
964: const PetscInt row = lrows[r] + A->rmap->rstart;
965: if (row >= A->cmap->N) continue;
966: MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
967: }
968: aijA->nonew = nnwA;
969: aijB->nonew = nnwB;
970: } else {
971: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
972: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
973: }
974: PetscFree(lrows);
975: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
976: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
978: /* reduce nonzerostate */
979: lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
980: MPIU_Allreduce(&lch,&gch,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)A));
981: if (gch) A->nonzerostate++;
982: return(0);
983: }
985: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
986: {
987: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
988: PetscErrorCode ierr;
989: PetscMPIInt n = A->rmap->n;
990: PetscInt i,j,r,m,p = 0,len = 0;
991: PetscInt *lrows,*owners = A->rmap->range;
992: PetscSFNode *rrows;
993: PetscSF sf;
994: const PetscScalar *xx;
995: PetscScalar *bb,*mask;
996: Vec xmask,lmask;
997: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data;
998: const PetscInt *aj, *ii,*ridx;
999: PetscScalar *aa;
1002: /* Create SF where leaves are input rows and roots are owned rows */
1003: PetscMalloc1(n, &lrows);
1004: for (r = 0; r < n; ++r) lrows[r] = -1;
1005: PetscMalloc1(N, &rrows);
1006: for (r = 0; r < N; ++r) {
1007: const PetscInt idx = rows[r];
1008: if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
1009: if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
1010: PetscLayoutFindOwner(A->rmap,idx,&p);
1011: }
1012: rrows[r].rank = p;
1013: rrows[r].index = rows[r] - owners[p];
1014: }
1015: PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1016: PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1017: /* Collect flags for rows to be zeroed */
1018: PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1019: PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1020: PetscSFDestroy(&sf);
1021: /* Compress and put in row numbers */
1022: for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1023: /* zero diagonal part of matrix */
1024: MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
1025: /* handle off diagonal part of matrix */
1026: MatCreateVecs(A,&xmask,NULL);
1027: VecDuplicate(l->lvec,&lmask);
1028: VecGetArray(xmask,&bb);
1029: for (i=0; i<len; i++) bb[lrows[i]] = 1;
1030: VecRestoreArray(xmask,&bb);
1031: VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1032: VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1033: VecDestroy(&xmask);
1034: if (x && b) { /* this code is buggy when the row and column layout don't match */
1035: PetscBool cong;
1037: MatHasCongruentLayouts(A,&cong);
1038: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
1039: VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1040: VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1041: VecGetArrayRead(l->lvec,&xx);
1042: VecGetArray(b,&bb);
1043: }
1044: VecGetArray(lmask,&mask);
1045: /* remove zeroed rows of off diagonal matrix */
1046: ii = aij->i;
1047: for (i=0; i<len; i++) {
1048: PetscArrayzero(aij->a + ii[lrows[i]],ii[lrows[i]+1] - ii[lrows[i]]);
1049: }
1050: /* loop over all elements of off process part of matrix zeroing removed columns*/
1051: if (aij->compressedrow.use) {
1052: m = aij->compressedrow.nrows;
1053: ii = aij->compressedrow.i;
1054: ridx = aij->compressedrow.rindex;
1055: for (i=0; i<m; i++) {
1056: n = ii[i+1] - ii[i];
1057: aj = aij->j + ii[i];
1058: aa = aij->a + ii[i];
1060: for (j=0; j<n; j++) {
1061: if (PetscAbsScalar(mask[*aj])) {
1062: if (b) bb[*ridx] -= *aa*xx[*aj];
1063: *aa = 0.0;
1064: }
1065: aa++;
1066: aj++;
1067: }
1068: ridx++;
1069: }
1070: } else { /* do not use compressed row format */
1071: m = l->B->rmap->n;
1072: for (i=0; i<m; i++) {
1073: n = ii[i+1] - ii[i];
1074: aj = aij->j + ii[i];
1075: aa = aij->a + ii[i];
1076: for (j=0; j<n; j++) {
1077: if (PetscAbsScalar(mask[*aj])) {
1078: if (b) bb[i] -= *aa*xx[*aj];
1079: *aa = 0.0;
1080: }
1081: aa++;
1082: aj++;
1083: }
1084: }
1085: }
1086: if (x && b) {
1087: VecRestoreArray(b,&bb);
1088: VecRestoreArrayRead(l->lvec,&xx);
1089: }
1090: VecRestoreArray(lmask,&mask);
1091: VecDestroy(&lmask);
1092: PetscFree(lrows);
1094: /* only change matrix nonzero state if pattern was allowed to be changed */
1095: if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
1096: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1097: MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1098: }
1099: return(0);
1100: }
1102: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1103: {
1104: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1106: PetscInt nt;
1107: VecScatter Mvctx = a->Mvctx;
1110: VecGetLocalSize(xx,&nt);
1111: if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
1113: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1114: (*a->A->ops->mult)(a->A,xx,yy);
1115: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1116: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1117: return(0);
1118: }
1120: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1121: {
1122: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1126: MatMultDiagonalBlock(a->A,bb,xx);
1127: return(0);
1128: }
1130: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1131: {
1132: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1134: VecScatter Mvctx = a->Mvctx;
1137: if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1;
1138: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1139: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1140: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1141: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1142: return(0);
1143: }
1145: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1146: {
1147: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1151: /* do nondiagonal part */
1152: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1153: /* do local part */
1154: (*a->A->ops->multtranspose)(a->A,xx,yy);
1155: /* add partial results together */
1156: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1157: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1158: return(0);
1159: }
1161: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f)
1162: {
1163: MPI_Comm comm;
1164: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1165: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1166: IS Me,Notme;
1168: PetscInt M,N,first,last,*notme,i;
1169: PetscBool lf;
1170: PetscMPIInt size;
1173: /* Easy test: symmetric diagonal block */
1174: Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1175: MatIsTranspose(Adia,Bdia,tol,&lf);
1176: MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));
1177: if (!*f) return(0);
1178: PetscObjectGetComm((PetscObject)Amat,&comm);
1179: MPI_Comm_size(comm,&size);
1180: if (size == 1) return(0);
1182: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1183: MatGetSize(Amat,&M,&N);
1184: MatGetOwnershipRange(Amat,&first,&last);
1185: PetscMalloc1(N-last+first,¬me);
1186: for (i=0; i<first; i++) notme[i] = i;
1187: for (i=last; i<M; i++) notme[i-last+first] = i;
1188: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1189: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1190: MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1191: Aoff = Aoffs[0];
1192: MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1193: Boff = Boffs[0];
1194: MatIsTranspose(Aoff,Boff,tol,f);
1195: MatDestroyMatrices(1,&Aoffs);
1196: MatDestroyMatrices(1,&Boffs);
1197: ISDestroy(&Me);
1198: ISDestroy(&Notme);
1199: PetscFree(notme);
1200: return(0);
1201: }
1203: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f)
1204: {
1208: MatIsTranspose_MPIAIJ(A,A,tol,f);
1209: return(0);
1210: }
1212: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1213: {
1214: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1218: /* do nondiagonal part */
1219: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1220: /* do local part */
1221: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1222: /* add partial results together */
1223: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1224: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1225: return(0);
1226: }
1228: /*
1229: This only works correctly for square matrices where the subblock A->A is the
1230: diagonal block
1231: */
1232: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1233: {
1235: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1238: if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1239: if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1240: MatGetDiagonal(a->A,v);
1241: return(0);
1242: }
1244: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1245: {
1246: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1250: MatScale(a->A,aa);
1251: MatScale(a->B,aa);
1252: return(0);
1253: }
1255: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1256: {
1257: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1261: #if defined(PETSC_USE_LOG)
1262: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1263: #endif
1264: MatStashDestroy_Private(&mat->stash);
1265: VecDestroy(&aij->diag);
1266: MatDestroy(&aij->A);
1267: MatDestroy(&aij->B);
1268: #if defined(PETSC_USE_CTABLE)
1269: PetscTableDestroy(&aij->colmap);
1270: #else
1271: PetscFree(aij->colmap);
1272: #endif
1273: PetscFree(aij->garray);
1274: VecDestroy(&aij->lvec);
1275: VecScatterDestroy(&aij->Mvctx);
1276: if (aij->Mvctx_mpi1) {VecScatterDestroy(&aij->Mvctx_mpi1);}
1277: PetscFree2(aij->rowvalues,aij->rowindices);
1278: PetscFree(aij->ld);
1279: PetscFree(mat->data);
1281: PetscObjectChangeTypeName((PetscObject)mat,0);
1282: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1283: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1284: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1285: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1286: PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);
1287: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1288: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1289: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1290: #if defined(PETSC_HAVE_ELEMENTAL)
1291: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1292: #endif
1293: #if defined(PETSC_HAVE_HYPRE)
1294: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);
1295: PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);
1296: #endif
1297: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);
1298: PetscObjectComposeFunction((PetscObject)mat,"MatPtAP_is_mpiaij_C",NULL);
1299: return(0);
1300: }
1302: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1303: {
1304: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1305: Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data;
1306: Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data;
1308: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
1309: int fd;
1310: PetscInt nz,header[4],*row_lengths,*range=0,rlen,i;
1311: PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0;
1312: PetscScalar *column_values;
1313: PetscInt message_count,flowcontrolcount;
1314: FILE *file;
1317: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1318: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1319: nz = A->nz + B->nz;
1320: PetscViewerBinaryGetDescriptor(viewer,&fd);
1321: if (!rank) {
1322: header[0] = MAT_FILE_CLASSID;
1323: header[1] = mat->rmap->N;
1324: header[2] = mat->cmap->N;
1326: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1327: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1328: /* get largest number of rows any processor has */
1329: rlen = mat->rmap->n;
1330: range = mat->rmap->range;
1331: for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]);
1332: } else {
1333: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1334: rlen = mat->rmap->n;
1335: }
1337: /* load up the local row counts */
1338: PetscMalloc1(rlen+1,&row_lengths);
1339: for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1341: /* store the row lengths to the file */
1342: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1343: if (!rank) {
1344: PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1345: for (i=1; i<size; i++) {
1346: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1347: rlen = range[i+1] - range[i];
1348: MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1349: PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1350: }
1351: PetscViewerFlowControlEndMaster(viewer,&message_count);
1352: } else {
1353: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1354: MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1355: PetscViewerFlowControlEndWorker(viewer,&message_count);
1356: }
1357: PetscFree(row_lengths);
1359: /* load up the local column indices */
1360: nzmax = nz; /* th processor needs space a largest processor needs */
1361: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1362: PetscMalloc1(nzmax+1,&column_indices);
1363: cnt = 0;
1364: for (i=0; i<mat->rmap->n; i++) {
1365: for (j=B->i[i]; j<B->i[i+1]; j++) {
1366: if ((col = garray[B->j[j]]) > cstart) break;
1367: column_indices[cnt++] = col;
1368: }
1369: for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart;
1370: for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]];
1371: }
1372: 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);
1374: /* store the column indices to the file */
1375: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1376: if (!rank) {
1377: MPI_Status status;
1378: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1379: for (i=1; i<size; i++) {
1380: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1381: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1382: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1383: MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1384: PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1385: }
1386: PetscViewerFlowControlEndMaster(viewer,&message_count);
1387: } else {
1388: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1389: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1390: MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1391: PetscViewerFlowControlEndWorker(viewer,&message_count);
1392: }
1393: PetscFree(column_indices);
1395: /* load up the local column values */
1396: PetscMalloc1(nzmax+1,&column_values);
1397: cnt = 0;
1398: for (i=0; i<mat->rmap->n; i++) {
1399: for (j=B->i[i]; j<B->i[i+1]; j++) {
1400: if (garray[B->j[j]] > cstart) break;
1401: column_values[cnt++] = B->a[j];
1402: }
1403: for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k];
1404: for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j];
1405: }
1406: 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);
1408: /* store the column values to the file */
1409: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1410: if (!rank) {
1411: MPI_Status status;
1412: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1413: for (i=1; i<size; i++) {
1414: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1415: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1416: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1417: MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));
1418: PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1419: }
1420: PetscViewerFlowControlEndMaster(viewer,&message_count);
1421: } else {
1422: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1423: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1424: MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1425: PetscViewerFlowControlEndWorker(viewer,&message_count);
1426: }
1427: PetscFree(column_values);
1429: PetscViewerBinaryGetInfoPointer(viewer,&file);
1430: if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs));
1431: return(0);
1432: }
1434: #include <petscdraw.h>
1435: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1436: {
1437: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1438: PetscErrorCode ierr;
1439: PetscMPIInt rank = aij->rank,size = aij->size;
1440: PetscBool isdraw,iascii,isbinary;
1441: PetscViewer sviewer;
1442: PetscViewerFormat format;
1445: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1446: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1447: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1448: if (iascii) {
1449: PetscViewerGetFormat(viewer,&format);
1450: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1451: PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz;
1452: PetscMalloc1(size,&nz);
1453: MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
1454: for (i=0; i<(PetscInt)size; i++) {
1455: nmax = PetscMax(nmax,nz[i]);
1456: nmin = PetscMin(nmin,nz[i]);
1457: navg += nz[i];
1458: }
1459: PetscFree(nz);
1460: navg = navg/size;
1461: PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);
1462: return(0);
1463: }
1464: PetscViewerGetFormat(viewer,&format);
1465: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1466: MatInfo info;
1467: PetscBool inodes;
1469: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1470: MatGetInfo(mat,MAT_LOCAL,&info);
1471: MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1472: PetscViewerASCIIPushSynchronized(viewer);
1473: if (!inodes) {
1474: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n",
1475: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1476: } else {
1477: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n",
1478: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1479: }
1480: MatGetInfo(aij->A,MAT_LOCAL,&info);
1481: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1482: MatGetInfo(aij->B,MAT_LOCAL,&info);
1483: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1484: PetscViewerFlush(viewer);
1485: PetscViewerASCIIPopSynchronized(viewer);
1486: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1487: VecScatterView(aij->Mvctx,viewer);
1488: return(0);
1489: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1490: PetscInt inodecount,inodelimit,*inodes;
1491: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1492: if (inodes) {
1493: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1494: } else {
1495: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1496: }
1497: return(0);
1498: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1499: return(0);
1500: }
1501: } else if (isbinary) {
1502: if (size == 1) {
1503: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1504: MatView(aij->A,viewer);
1505: } else {
1506: MatView_MPIAIJ_Binary(mat,viewer);
1507: }
1508: return(0);
1509: } else if (iascii && size == 1) {
1510: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1511: MatView(aij->A,viewer);
1512: return(0);
1513: } else if (isdraw) {
1514: PetscDraw draw;
1515: PetscBool isnull;
1516: PetscViewerDrawGetDraw(viewer,0,&draw);
1517: PetscDrawIsNull(draw,&isnull);
1518: if (isnull) return(0);
1519: }
1521: { /* assemble the entire matrix onto first processor */
1522: Mat A = NULL, Av;
1523: IS isrow,iscol;
1525: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1526: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1527: MatCreateSubMatrix(mat,isrow,iscol,MAT_INITIAL_MATRIX,&A);
1528: MatMPIAIJGetSeqAIJ(A,&Av,NULL,NULL);
1529: /* The commented code uses MatCreateSubMatrices instead */
1530: /*
1531: Mat *AA, A = NULL, Av;
1532: IS isrow,iscol;
1534: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1535: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1536: MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA);
1537: if (!rank) {
1538: PetscObjectReference((PetscObject)AA[0]);
1539: A = AA[0];
1540: Av = AA[0];
1541: }
1542: MatDestroySubMatrices(1,&AA);
1543: */
1544: ISDestroy(&iscol);
1545: ISDestroy(&isrow);
1546: /*
1547: Everyone has to call to draw the matrix since the graphics waits are
1548: synchronized across all processors that share the PetscDraw object
1549: */
1550: PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1551: if (!rank) {
1552: if (((PetscObject)mat)->name) {
1553: PetscObjectSetName((PetscObject)Av,((PetscObject)mat)->name);
1554: }
1555: MatView_SeqAIJ(Av,sviewer);
1556: }
1557: PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1558: PetscViewerFlush(viewer);
1559: MatDestroy(&A);
1560: }
1561: return(0);
1562: }
1564: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1565: {
1567: PetscBool iascii,isdraw,issocket,isbinary;
1570: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1571: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1572: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1573: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1574: if (iascii || isdraw || isbinary || issocket) {
1575: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1576: }
1577: return(0);
1578: }
1580: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1581: {
1582: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1584: Vec bb1 = 0;
1585: PetscBool hasop;
1588: if (flag == SOR_APPLY_UPPER) {
1589: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1590: return(0);
1591: }
1593: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1594: VecDuplicate(bb,&bb1);
1595: }
1597: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1598: if (flag & SOR_ZERO_INITIAL_GUESS) {
1599: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1600: its--;
1601: }
1603: while (its--) {
1604: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1605: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1607: /* update rhs: bb1 = bb - B*x */
1608: VecScale(mat->lvec,-1.0);
1609: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1611: /* local sweep */
1612: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1613: }
1614: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1615: if (flag & SOR_ZERO_INITIAL_GUESS) {
1616: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1617: its--;
1618: }
1619: while (its--) {
1620: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1621: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1623: /* update rhs: bb1 = bb - B*x */
1624: VecScale(mat->lvec,-1.0);
1625: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1627: /* local sweep */
1628: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1629: }
1630: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1631: if (flag & SOR_ZERO_INITIAL_GUESS) {
1632: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1633: its--;
1634: }
1635: while (its--) {
1636: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1637: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1639: /* update rhs: bb1 = bb - B*x */
1640: VecScale(mat->lvec,-1.0);
1641: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1643: /* local sweep */
1644: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1645: }
1646: } else if (flag & SOR_EISENSTAT) {
1647: Vec xx1;
1649: VecDuplicate(bb,&xx1);
1650: (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);
1652: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1653: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1654: if (!mat->diag) {
1655: MatCreateVecs(matin,&mat->diag,NULL);
1656: MatGetDiagonal(matin,mat->diag);
1657: }
1658: MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1659: if (hasop) {
1660: MatMultDiagonalBlock(matin,xx,bb1);
1661: } else {
1662: VecPointwiseMult(bb1,mat->diag,xx);
1663: }
1664: VecAYPX(bb1,(omega-2.0)/omega,bb);
1666: MatMultAdd(mat->B,mat->lvec,bb1,bb1);
1668: /* local sweep */
1669: (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1670: VecAXPY(xx,1.0,xx1);
1671: VecDestroy(&xx1);
1672: } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");
1674: VecDestroy(&bb1);
1676: matin->factorerrortype = mat->A->factorerrortype;
1677: return(0);
1678: }
1680: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1681: {
1682: Mat aA,aB,Aperm;
1683: const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1684: PetscScalar *aa,*ba;
1685: PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1686: PetscSF rowsf,sf;
1687: IS parcolp = NULL;
1688: PetscBool done;
1692: MatGetLocalSize(A,&m,&n);
1693: ISGetIndices(rowp,&rwant);
1694: ISGetIndices(colp,&cwant);
1695: PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);
1697: /* Invert row permutation to find out where my rows should go */
1698: PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1699: PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1700: PetscSFSetFromOptions(rowsf);
1701: for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1702: PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1703: PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1705: /* Invert column permutation to find out where my columns should go */
1706: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1707: PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1708: PetscSFSetFromOptions(sf);
1709: for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1710: PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1711: PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1712: PetscSFDestroy(&sf);
1714: ISRestoreIndices(rowp,&rwant);
1715: ISRestoreIndices(colp,&cwant);
1716: MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);
1718: /* Find out where my gcols should go */
1719: MatGetSize(aB,NULL,&ng);
1720: PetscMalloc1(ng,&gcdest);
1721: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1722: PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1723: PetscSFSetFromOptions(sf);
1724: PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1725: PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1726: PetscSFDestroy(&sf);
1728: PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1729: MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1730: MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1731: for (i=0; i<m; i++) {
1732: PetscInt row = rdest[i],rowner;
1733: PetscLayoutFindOwner(A->rmap,row,&rowner);
1734: for (j=ai[i]; j<ai[i+1]; j++) {
1735: PetscInt cowner,col = cdest[aj[j]];
1736: PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1737: if (rowner == cowner) dnnz[i]++;
1738: else onnz[i]++;
1739: }
1740: for (j=bi[i]; j<bi[i+1]; j++) {
1741: PetscInt cowner,col = gcdest[bj[j]];
1742: PetscLayoutFindOwner(A->cmap,col,&cowner);
1743: if (rowner == cowner) dnnz[i]++;
1744: else onnz[i]++;
1745: }
1746: }
1747: PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);
1748: PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);
1749: PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);
1750: PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);
1751: PetscSFDestroy(&rowsf);
1753: MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);
1754: MatSeqAIJGetArray(aA,&aa);
1755: MatSeqAIJGetArray(aB,&ba);
1756: for (i=0; i<m; i++) {
1757: PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */
1758: PetscInt j0,rowlen;
1759: rowlen = ai[i+1] - ai[i];
1760: for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */
1761: for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]];
1762: MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);
1763: }
1764: rowlen = bi[i+1] - bi[i];
1765: for (j0=j=0; j<rowlen; j0=j) {
1766: for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]];
1767: MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);
1768: }
1769: }
1770: MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);
1771: MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);
1772: MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1773: MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1774: MatSeqAIJRestoreArray(aA,&aa);
1775: MatSeqAIJRestoreArray(aB,&ba);
1776: PetscFree4(dnnz,onnz,tdnnz,tonnz);
1777: PetscFree3(work,rdest,cdest);
1778: PetscFree(gcdest);
1779: if (parcolp) {ISDestroy(&colp);}
1780: *B = Aperm;
1781: return(0);
1782: }
1784: PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
1785: {
1786: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1790: MatGetSize(aij->B,NULL,nghosts);
1791: if (ghosts) *ghosts = aij->garray;
1792: return(0);
1793: }
1795: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1796: {
1797: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1798: Mat A = mat->A,B = mat->B;
1800: PetscLogDouble isend[5],irecv[5];
1803: info->block_size = 1.0;
1804: MatGetInfo(A,MAT_LOCAL,info);
1806: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1807: isend[3] = info->memory; isend[4] = info->mallocs;
1809: MatGetInfo(B,MAT_LOCAL,info);
1811: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1812: isend[3] += info->memory; isend[4] += info->mallocs;
1813: if (flag == MAT_LOCAL) {
1814: info->nz_used = isend[0];
1815: info->nz_allocated = isend[1];
1816: info->nz_unneeded = isend[2];
1817: info->memory = isend[3];
1818: info->mallocs = isend[4];
1819: } else if (flag == MAT_GLOBAL_MAX) {
1820: MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_MAX,PetscObjectComm((PetscObject)matin));
1822: info->nz_used = irecv[0];
1823: info->nz_allocated = irecv[1];
1824: info->nz_unneeded = irecv[2];
1825: info->memory = irecv[3];
1826: info->mallocs = irecv[4];
1827: } else if (flag == MAT_GLOBAL_SUM) {
1828: MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_SUM,PetscObjectComm((PetscObject)matin));
1830: info->nz_used = irecv[0];
1831: info->nz_allocated = irecv[1];
1832: info->nz_unneeded = irecv[2];
1833: info->memory = irecv[3];
1834: info->mallocs = irecv[4];
1835: }
1836: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1837: info->fill_ratio_needed = 0;
1838: info->factor_mallocs = 0;
1839: return(0);
1840: }
1842: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1843: {
1844: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1848: switch (op) {
1849: case MAT_NEW_NONZERO_LOCATIONS:
1850: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1851: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1852: case MAT_KEEP_NONZERO_PATTERN:
1853: case MAT_NEW_NONZERO_LOCATION_ERR:
1854: case MAT_USE_INODES:
1855: case MAT_IGNORE_ZERO_ENTRIES:
1856: MatCheckPreallocated(A,1);
1857: MatSetOption(a->A,op,flg);
1858: MatSetOption(a->B,op,flg);
1859: break;
1860: case MAT_ROW_ORIENTED:
1861: MatCheckPreallocated(A,1);
1862: a->roworiented = flg;
1864: MatSetOption(a->A,op,flg);
1865: MatSetOption(a->B,op,flg);
1866: break;
1867: case MAT_NEW_DIAGONALS:
1868: case MAT_SORTED_FULL:
1869: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1870: break;
1871: case MAT_IGNORE_OFF_PROC_ENTRIES:
1872: a->donotstash = flg;
1873: break;
1874: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1875: case MAT_SPD:
1876: case MAT_SYMMETRIC:
1877: case MAT_STRUCTURALLY_SYMMETRIC:
1878: case MAT_HERMITIAN:
1879: case MAT_SYMMETRY_ETERNAL:
1880: break;
1881: case MAT_SUBMAT_SINGLEIS:
1882: A->submat_singleis = flg;
1883: break;
1884: case MAT_STRUCTURE_ONLY:
1885: /* The option is handled directly by MatSetOption() */
1886: break;
1887: default:
1888: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1889: }
1890: return(0);
1891: }
1893: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1894: {
1895: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1896: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1898: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1899: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1900: PetscInt *cmap,*idx_p;
1903: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1904: mat->getrowactive = PETSC_TRUE;
1906: if (!mat->rowvalues && (idx || v)) {
1907: /*
1908: allocate enough space to hold information from the longest row.
1909: */
1910: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1911: PetscInt max = 1,tmp;
1912: for (i=0; i<matin->rmap->n; i++) {
1913: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1914: if (max < tmp) max = tmp;
1915: }
1916: PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1917: }
1919: if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1920: lrow = row - rstart;
1922: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1923: if (!v) {pvA = 0; pvB = 0;}
1924: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1925: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1926: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1927: nztot = nzA + nzB;
1929: cmap = mat->garray;
1930: if (v || idx) {
1931: if (nztot) {
1932: /* Sort by increasing column numbers, assuming A and B already sorted */
1933: PetscInt imark = -1;
1934: if (v) {
1935: *v = v_p = mat->rowvalues;
1936: for (i=0; i<nzB; i++) {
1937: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1938: else break;
1939: }
1940: imark = i;
1941: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1942: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1943: }
1944: if (idx) {
1945: *idx = idx_p = mat->rowindices;
1946: if (imark > -1) {
1947: for (i=0; i<imark; i++) {
1948: idx_p[i] = cmap[cworkB[i]];
1949: }
1950: } else {
1951: for (i=0; i<nzB; i++) {
1952: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1953: else break;
1954: }
1955: imark = i;
1956: }
1957: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1958: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1959: }
1960: } else {
1961: if (idx) *idx = 0;
1962: if (v) *v = 0;
1963: }
1964: }
1965: *nz = nztot;
1966: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1967: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1968: return(0);
1969: }
1971: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1972: {
1973: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1976: if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1977: aij->getrowactive = PETSC_FALSE;
1978: return(0);
1979: }
1981: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1982: {
1983: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1984: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1986: PetscInt i,j,cstart = mat->cmap->rstart;
1987: PetscReal sum = 0.0;
1988: MatScalar *v;
1991: if (aij->size == 1) {
1992: MatNorm(aij->A,type,norm);
1993: } else {
1994: if (type == NORM_FROBENIUS) {
1995: v = amat->a;
1996: for (i=0; i<amat->nz; i++) {
1997: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1998: }
1999: v = bmat->a;
2000: for (i=0; i<bmat->nz; i++) {
2001: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
2002: }
2003: MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
2004: *norm = PetscSqrtReal(*norm);
2005: PetscLogFlops(2*amat->nz+2*bmat->nz);
2006: } else if (type == NORM_1) { /* max column norm */
2007: PetscReal *tmp,*tmp2;
2008: PetscInt *jj,*garray = aij->garray;
2009: PetscCalloc1(mat->cmap->N+1,&tmp);
2010: PetscMalloc1(mat->cmap->N+1,&tmp2);
2011: *norm = 0.0;
2012: v = amat->a; jj = amat->j;
2013: for (j=0; j<amat->nz; j++) {
2014: tmp[cstart + *jj++] += PetscAbsScalar(*v); v++;
2015: }
2016: v = bmat->a; jj = bmat->j;
2017: for (j=0; j<bmat->nz; j++) {
2018: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
2019: }
2020: MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
2021: for (j=0; j<mat->cmap->N; j++) {
2022: if (tmp2[j] > *norm) *norm = tmp2[j];
2023: }
2024: PetscFree(tmp);
2025: PetscFree(tmp2);
2026: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
2027: } else if (type == NORM_INFINITY) { /* max row norm */
2028: PetscReal ntemp = 0.0;
2029: for (j=0; j<aij->A->rmap->n; j++) {
2030: v = amat->a + amat->i[j];
2031: sum = 0.0;
2032: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
2033: sum += PetscAbsScalar(*v); v++;
2034: }
2035: v = bmat->a + bmat->i[j];
2036: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
2037: sum += PetscAbsScalar(*v); v++;
2038: }
2039: if (sum > ntemp) ntemp = sum;
2040: }
2041: MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
2042: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
2043: } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
2044: }
2045: return(0);
2046: }
2048: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
2049: {
2050: Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b;
2051: Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag;
2052: PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,row,*cols,*cols_tmp,*B_diag_ilen,i,ncol,A_diag_ncol;
2053: const PetscInt *ai,*aj,*bi,*bj,*B_diag_i;
2054: PetscErrorCode ierr;
2055: Mat B,A_diag,*B_diag;
2056: const MatScalar *array;
2059: ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
2060: ai = Aloc->i; aj = Aloc->j;
2061: bi = Bloc->i; bj = Bloc->j;
2062: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2063: PetscInt *d_nnz,*g_nnz,*o_nnz;
2064: PetscSFNode *oloc;
2065: PETSC_UNUSED PetscSF sf;
2067: PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
2068: /* compute d_nnz for preallocation */
2069: PetscArrayzero(d_nnz,na);
2070: for (i=0; i<ai[ma]; i++) {
2071: d_nnz[aj[i]]++;
2072: }
2073: /* compute local off-diagonal contributions */
2074: PetscArrayzero(g_nnz,nb);
2075: for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
2076: /* map those to global */
2077: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2078: PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2079: PetscSFSetFromOptions(sf);
2080: PetscArrayzero(o_nnz,na);
2081: PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2082: PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2083: PetscSFDestroy(&sf);
2085: MatCreate(PetscObjectComm((PetscObject)A),&B);
2086: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2087: MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2088: MatSetType(B,((PetscObject)A)->type_name);
2089: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2090: PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2091: } else {
2092: B = *matout;
2093: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2094: }
2096: b = (Mat_MPIAIJ*)B->data;
2097: A_diag = a->A;
2098: B_diag = &b->A;
2099: sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data;
2100: A_diag_ncol = A_diag->cmap->N;
2101: B_diag_ilen = sub_B_diag->ilen;
2102: B_diag_i = sub_B_diag->i;
2104: /* Set ilen for diagonal of B */
2105: for (i=0; i<A_diag_ncol; i++) {
2106: B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i];
2107: }
2109: /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
2110: very quickly (=without using MatSetValues), because all writes are local. */
2111: MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);
2113: /* copy over the B part */
2114: PetscMalloc1(bi[mb],&cols);
2115: array = Bloc->a;
2116: row = A->rmap->rstart;
2117: for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2118: cols_tmp = cols;
2119: for (i=0; i<mb; i++) {
2120: ncol = bi[i+1]-bi[i];
2121: MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2122: row++;
2123: array += ncol; cols_tmp += ncol;
2124: }
2125: PetscFree(cols);
2127: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2128: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2129: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2130: *matout = B;
2131: } else {
2132: MatHeaderMerge(A,&B);
2133: }
2134: return(0);
2135: }
2137: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2138: {
2139: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2140: Mat a = aij->A,b = aij->B;
2142: PetscInt s1,s2,s3;
2145: MatGetLocalSize(mat,&s2,&s3);
2146: if (rr) {
2147: VecGetLocalSize(rr,&s1);
2148: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2149: /* Overlap communication with computation. */
2150: VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2151: }
2152: if (ll) {
2153: VecGetLocalSize(ll,&s1);
2154: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2155: (*b->ops->diagonalscale)(b,ll,0);
2156: }
2157: /* scale the diagonal block */
2158: (*a->ops->diagonalscale)(a,ll,rr);
2160: if (rr) {
2161: /* Do a scatter end and then right scale the off-diagonal block */
2162: VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2163: (*b->ops->diagonalscale)(b,0,aij->lvec);
2164: }
2165: return(0);
2166: }
2168: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2169: {
2170: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2174: MatSetUnfactored(a->A);
2175: return(0);
2176: }
2178: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag)
2179: {
2180: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2181: Mat a,b,c,d;
2182: PetscBool flg;
2186: a = matA->A; b = matA->B;
2187: c = matB->A; d = matB->B;
2189: MatEqual(a,c,&flg);
2190: if (flg) {
2191: MatEqual(b,d,&flg);
2192: }
2193: MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2194: return(0);
2195: }
2197: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2198: {
2200: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2201: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
2204: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2205: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2206: /* because of the column compression in the off-processor part of the matrix a->B,
2207: the number of columns in a->B and b->B may be different, hence we cannot call
2208: the MatCopy() directly on the two parts. If need be, we can provide a more
2209: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2210: then copying the submatrices */
2211: MatCopy_Basic(A,B,str);
2212: } else {
2213: MatCopy(a->A,b->A,str);
2214: MatCopy(a->B,b->B,str);
2215: }
2216: PetscObjectStateIncrease((PetscObject)B);
2217: return(0);
2218: }
2220: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2221: {
2225: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2226: return(0);
2227: }
2229: /*
2230: Computes the number of nonzeros per row needed for preallocation when X and Y
2231: have different nonzero structure.
2232: */
2233: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz)
2234: {
2235: PetscInt i,j,k,nzx,nzy;
2238: /* Set the number of nonzeros in the new matrix */
2239: for (i=0; i<m; i++) {
2240: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2241: nzx = xi[i+1] - xi[i];
2242: nzy = yi[i+1] - yi[i];
2243: nnz[i] = 0;
2244: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2245: for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2246: if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */
2247: nnz[i]++;
2248: }
2249: for (; k<nzy; k++) nnz[i]++;
2250: }
2251: return(0);
2252: }
2254: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2255: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2256: {
2258: PetscInt m = Y->rmap->N;
2259: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2260: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2263: MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2264: return(0);
2265: }
2267: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2268: {
2270: Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2271: PetscBLASInt bnz,one=1;
2272: Mat_SeqAIJ *x,*y;
2275: if (str == SAME_NONZERO_PATTERN) {
2276: PetscScalar alpha = a;
2277: x = (Mat_SeqAIJ*)xx->A->data;
2278: PetscBLASIntCast(x->nz,&bnz);
2279: y = (Mat_SeqAIJ*)yy->A->data;
2280: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2281: x = (Mat_SeqAIJ*)xx->B->data;
2282: y = (Mat_SeqAIJ*)yy->B->data;
2283: PetscBLASIntCast(x->nz,&bnz);
2284: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2285: PetscObjectStateIncrease((PetscObject)Y);
2286: /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU
2287: will be updated */
2288: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
2289: if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) {
2290: Y->offloadmask = PETSC_OFFLOAD_CPU;
2291: }
2292: #endif
2293: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2294: MatAXPY_Basic(Y,a,X,str);
2295: } else {
2296: Mat B;
2297: PetscInt *nnz_d,*nnz_o;
2298: PetscMalloc1(yy->A->rmap->N,&nnz_d);
2299: PetscMalloc1(yy->B->rmap->N,&nnz_o);
2300: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2301: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2302: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2303: MatSetBlockSizesFromMats(B,Y,Y);
2304: MatSetType(B,MATMPIAIJ);
2305: MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2306: MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2307: MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2308: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2309: MatHeaderReplace(Y,&B);
2310: PetscFree(nnz_d);
2311: PetscFree(nnz_o);
2312: }
2313: return(0);
2314: }
2316: extern PetscErrorCode MatConjugate_SeqAIJ(Mat);
2318: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2319: {
2320: #if defined(PETSC_USE_COMPLEX)
2322: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2325: MatConjugate_SeqAIJ(aij->A);
2326: MatConjugate_SeqAIJ(aij->B);
2327: #else
2329: #endif
2330: return(0);
2331: }
2333: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2334: {
2335: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2339: MatRealPart(a->A);
2340: MatRealPart(a->B);
2341: return(0);
2342: }
2344: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2345: {
2346: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2350: MatImaginaryPart(a->A);
2351: MatImaginaryPart(a->B);
2352: return(0);
2353: }
2355: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2356: {
2357: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2359: PetscInt i,*idxb = 0;
2360: PetscScalar *va,*vb;
2361: Vec vtmp;
2364: MatGetRowMaxAbs(a->A,v,idx);
2365: VecGetArray(v,&va);
2366: if (idx) {
2367: for (i=0; i<A->rmap->n; i++) {
2368: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2369: }
2370: }
2372: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2373: if (idx) {
2374: PetscMalloc1(A->rmap->n,&idxb);
2375: }
2376: MatGetRowMaxAbs(a->B,vtmp,idxb);
2377: VecGetArray(vtmp,&vb);
2379: for (i=0; i<A->rmap->n; i++) {
2380: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2381: va[i] = vb[i];
2382: if (idx) idx[i] = a->garray[idxb[i]];
2383: }
2384: }
2386: VecRestoreArray(v,&va);
2387: VecRestoreArray(vtmp,&vb);
2388: PetscFree(idxb);
2389: VecDestroy(&vtmp);
2390: return(0);
2391: }
2393: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2394: {
2395: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2397: PetscInt i,*idxb = 0;
2398: PetscScalar *va,*vb;
2399: Vec vtmp;
2402: MatGetRowMinAbs(a->A,v,idx);
2403: VecGetArray(v,&va);
2404: if (idx) {
2405: for (i=0; i<A->cmap->n; i++) {
2406: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2407: }
2408: }
2410: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2411: if (idx) {
2412: PetscMalloc1(A->rmap->n,&idxb);
2413: }
2414: MatGetRowMinAbs(a->B,vtmp,idxb);
2415: VecGetArray(vtmp,&vb);
2417: for (i=0; i<A->rmap->n; i++) {
2418: if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2419: va[i] = vb[i];
2420: if (idx) idx[i] = a->garray[idxb[i]];
2421: }
2422: }
2424: VecRestoreArray(v,&va);
2425: VecRestoreArray(vtmp,&vb);
2426: PetscFree(idxb);
2427: VecDestroy(&vtmp);
2428: return(0);
2429: }
2431: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2432: {
2433: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2434: PetscInt n = A->rmap->n;
2435: PetscInt cstart = A->cmap->rstart;
2436: PetscInt *cmap = mat->garray;
2437: PetscInt *diagIdx, *offdiagIdx;
2438: Vec diagV, offdiagV;
2439: PetscScalar *a, *diagA, *offdiagA;
2440: PetscInt r;
2444: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2445: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);
2446: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);
2447: MatGetRowMin(mat->A, diagV, diagIdx);
2448: MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2449: VecGetArray(v, &a);
2450: VecGetArray(diagV, &diagA);
2451: VecGetArray(offdiagV, &offdiagA);
2452: for (r = 0; r < n; ++r) {
2453: if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2454: a[r] = diagA[r];
2455: idx[r] = cstart + diagIdx[r];
2456: } else {
2457: a[r] = offdiagA[r];
2458: idx[r] = cmap[offdiagIdx[r]];
2459: }
2460: }
2461: VecRestoreArray(v, &a);
2462: VecRestoreArray(diagV, &diagA);
2463: VecRestoreArray(offdiagV, &offdiagA);
2464: VecDestroy(&diagV);
2465: VecDestroy(&offdiagV);
2466: PetscFree2(diagIdx, offdiagIdx);
2467: return(0);
2468: }
2470: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2471: {
2472: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2473: PetscInt n = A->rmap->n;
2474: PetscInt cstart = A->cmap->rstart;
2475: PetscInt *cmap = mat->garray;
2476: PetscInt *diagIdx, *offdiagIdx;
2477: Vec diagV, offdiagV;
2478: PetscScalar *a, *diagA, *offdiagA;
2479: PetscInt r;
2483: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2484: VecCreateSeq(PETSC_COMM_SELF, n, &diagV);
2485: VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);
2486: MatGetRowMax(mat->A, diagV, diagIdx);
2487: MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2488: VecGetArray(v, &a);
2489: VecGetArray(diagV, &diagA);
2490: VecGetArray(offdiagV, &offdiagA);
2491: for (r = 0; r < n; ++r) {
2492: if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2493: a[r] = diagA[r];
2494: idx[r] = cstart + diagIdx[r];
2495: } else {
2496: a[r] = offdiagA[r];
2497: idx[r] = cmap[offdiagIdx[r]];
2498: }
2499: }
2500: VecRestoreArray(v, &a);
2501: VecRestoreArray(diagV, &diagA);
2502: VecRestoreArray(offdiagV, &offdiagA);
2503: VecDestroy(&diagV);
2504: VecDestroy(&offdiagV);
2505: PetscFree2(diagIdx, offdiagIdx);
2506: return(0);
2507: }
2509: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2510: {
2512: Mat *dummy;
2515: MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2516: *newmat = *dummy;
2517: PetscFree(dummy);
2518: return(0);
2519: }
2521: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2522: {
2523: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data;
2527: MatInvertBlockDiagonal(a->A,values);
2528: A->factorerrortype = a->A->factorerrortype;
2529: return(0);
2530: }
2532: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2533: {
2535: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data;
2538: if (!x->assembled && !x->preallocated) SETERRQ(PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2539: MatSetRandom(aij->A,rctx);
2540: if (x->assembled) {
2541: MatSetRandom(aij->B,rctx);
2542: } else {
2543: MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B,x->cmap->rstart,x->cmap->rend,rctx);
2544: }
2545: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2546: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2547: return(0);
2548: }
2550: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2551: {
2553: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2554: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2555: return(0);
2556: }
2558: /*@
2559: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2561: Collective on Mat
2563: Input Parameters:
2564: + A - the matrix
2565: - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)
2567: Level: advanced
2569: @*/
2570: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2571: {
2572: PetscErrorCode ierr;
2575: PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2576: return(0);
2577: }
2579: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2580: {
2581: PetscErrorCode ierr;
2582: PetscBool sc = PETSC_FALSE,flg;
2585: PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2586: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2587: PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2588: if (flg) {
2589: MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2590: }
2591: PetscOptionsTail();
2592: return(0);
2593: }
2595: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2596: {
2598: Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data;
2599: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data;
2602: if (!Y->preallocated) {
2603: MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2604: } else if (!aij->nz) {
2605: PetscInt nonew = aij->nonew;
2606: MatSeqAIJSetPreallocation(maij->A,1,NULL);
2607: aij->nonew = nonew;
2608: }
2609: MatShift_Basic(Y,a);
2610: return(0);
2611: }
2613: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d)
2614: {
2615: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2619: if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2620: MatMissingDiagonal(a->A,missing,d);
2621: if (d) {
2622: PetscInt rstart;
2623: MatGetOwnershipRange(A,&rstart,NULL);
2624: *d += rstart;
2626: }
2627: return(0);
2628: }
2630: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
2631: {
2632: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2636: MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);
2637: return(0);
2638: }
2640: /* -------------------------------------------------------------------*/
2641: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2642: MatGetRow_MPIAIJ,
2643: MatRestoreRow_MPIAIJ,
2644: MatMult_MPIAIJ,
2645: /* 4*/ MatMultAdd_MPIAIJ,
2646: MatMultTranspose_MPIAIJ,
2647: MatMultTransposeAdd_MPIAIJ,
2648: 0,
2649: 0,
2650: 0,
2651: /*10*/ 0,
2652: 0,
2653: 0,
2654: MatSOR_MPIAIJ,
2655: MatTranspose_MPIAIJ,
2656: /*15*/ MatGetInfo_MPIAIJ,
2657: MatEqual_MPIAIJ,
2658: MatGetDiagonal_MPIAIJ,
2659: MatDiagonalScale_MPIAIJ,
2660: MatNorm_MPIAIJ,
2661: /*20*/ MatAssemblyBegin_MPIAIJ,
2662: MatAssemblyEnd_MPIAIJ,
2663: MatSetOption_MPIAIJ,
2664: MatZeroEntries_MPIAIJ,
2665: /*24*/ MatZeroRows_MPIAIJ,
2666: 0,
2667: 0,
2668: 0,
2669: 0,
2670: /*29*/ MatSetUp_MPIAIJ,
2671: 0,
2672: 0,
2673: MatGetDiagonalBlock_MPIAIJ,
2674: 0,
2675: /*34*/ MatDuplicate_MPIAIJ,
2676: 0,
2677: 0,
2678: 0,
2679: 0,
2680: /*39*/ MatAXPY_MPIAIJ,
2681: MatCreateSubMatrices_MPIAIJ,
2682: MatIncreaseOverlap_MPIAIJ,
2683: MatGetValues_MPIAIJ,
2684: MatCopy_MPIAIJ,
2685: /*44*/ MatGetRowMax_MPIAIJ,
2686: MatScale_MPIAIJ,
2687: MatShift_MPIAIJ,
2688: MatDiagonalSet_MPIAIJ,
2689: MatZeroRowsColumns_MPIAIJ,
2690: /*49*/ MatSetRandom_MPIAIJ,
2691: 0,
2692: 0,
2693: 0,
2694: 0,
2695: /*54*/ MatFDColoringCreate_MPIXAIJ,
2696: 0,
2697: MatSetUnfactored_MPIAIJ,
2698: MatPermute_MPIAIJ,
2699: 0,
2700: /*59*/ MatCreateSubMatrix_MPIAIJ,
2701: MatDestroy_MPIAIJ,
2702: MatView_MPIAIJ,
2703: 0,
2704: MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ,
2705: /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ,
2706: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2707: 0,
2708: 0,
2709: 0,
2710: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2711: MatGetRowMinAbs_MPIAIJ,
2712: 0,
2713: 0,
2714: 0,
2715: 0,
2716: /*75*/ MatFDColoringApply_AIJ,
2717: MatSetFromOptions_MPIAIJ,
2718: 0,
2719: 0,
2720: MatFindZeroDiagonals_MPIAIJ,
2721: /*80*/ 0,
2722: 0,
2723: 0,
2724: /*83*/ MatLoad_MPIAIJ,
2725: MatIsSymmetric_MPIAIJ,
2726: 0,
2727: 0,
2728: 0,
2729: 0,
2730: /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
2731: MatMatMultSymbolic_MPIAIJ_MPIAIJ,
2732: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2733: MatPtAP_MPIAIJ_MPIAIJ,
2734: MatPtAPSymbolic_MPIAIJ_MPIAIJ,
2735: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2736: 0,
2737: 0,
2738: 0,
2739: MatPinToCPU_MPIAIJ,
2740: /*99*/ 0,
2741: 0,
2742: 0,
2743: MatConjugate_MPIAIJ,
2744: 0,
2745: /*104*/MatSetValuesRow_MPIAIJ,
2746: MatRealPart_MPIAIJ,
2747: MatImaginaryPart_MPIAIJ,
2748: 0,
2749: 0,
2750: /*109*/0,
2751: 0,
2752: MatGetRowMin_MPIAIJ,
2753: 0,
2754: MatMissingDiagonal_MPIAIJ,
2755: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2756: 0,
2757: MatGetGhosts_MPIAIJ,
2758: 0,
2759: 0,
2760: /*119*/0,
2761: 0,
2762: 0,
2763: 0,
2764: MatGetMultiProcBlock_MPIAIJ,
2765: /*124*/MatFindNonzeroRows_MPIAIJ,
2766: MatGetColumnNorms_MPIAIJ,
2767: MatInvertBlockDiagonal_MPIAIJ,
2768: MatInvertVariableBlockDiagonal_MPIAIJ,
2769: MatCreateSubMatricesMPI_MPIAIJ,
2770: /*129*/0,
2771: MatTransposeMatMult_MPIAIJ_MPIAIJ,
2772: MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
2773: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2774: 0,
2775: /*134*/0,
2776: 0,
2777: MatRARt_MPIAIJ_MPIAIJ,
2778: 0,
2779: 0,
2780: /*139*/MatSetBlockSizes_MPIAIJ,
2781: 0,
2782: 0,
2783: MatFDColoringSetUp_MPIXAIJ,
2784: MatFindOffBlockDiagonalEntries_MPIAIJ,
2785: /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ
2786: };
2788: /* ----------------------------------------------------------------------------------------*/
2790: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2791: {
2792: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2796: MatStoreValues(aij->A);
2797: MatStoreValues(aij->B);
2798: return(0);
2799: }
2801: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2802: {
2803: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2807: MatRetrieveValues(aij->A);
2808: MatRetrieveValues(aij->B);
2809: return(0);
2810: }
2812: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2813: {
2814: Mat_MPIAIJ *b;
2816: PetscMPIInt size;
2819: PetscLayoutSetUp(B->rmap);
2820: PetscLayoutSetUp(B->cmap);
2821: b = (Mat_MPIAIJ*)B->data;
2823: #if defined(PETSC_USE_CTABLE)
2824: PetscTableDestroy(&b->colmap);
2825: #else
2826: PetscFree(b->colmap);
2827: #endif
2828: PetscFree(b->garray);
2829: VecDestroy(&b->lvec);
2830: VecScatterDestroy(&b->Mvctx);
2832: /* Because the B will have been resized we simply destroy it and create a new one each time */
2833: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
2834: MatDestroy(&b->B);
2835: MatCreate(PETSC_COMM_SELF,&b->B);
2836: MatSetSizes(b->B,B->rmap->n,size > 1 ? B->cmap->N : 0,B->rmap->n,size > 1 ? B->cmap->N : 0);
2837: MatSetBlockSizesFromMats(b->B,B,B);
2838: MatSetType(b->B,MATSEQAIJ);
2839: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2841: if (!B->preallocated) {
2842: MatCreate(PETSC_COMM_SELF,&b->A);
2843: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2844: MatSetBlockSizesFromMats(b->A,B,B);
2845: MatSetType(b->A,MATSEQAIJ);
2846: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2847: }
2849: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2850: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
2851: B->preallocated = PETSC_TRUE;
2852: B->was_assembled = PETSC_FALSE;
2853: B->assembled = PETSC_FALSE;
2854: return(0);
2855: }
2857: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2858: {
2859: Mat_MPIAIJ *b;
2864: PetscLayoutSetUp(B->rmap);
2865: PetscLayoutSetUp(B->cmap);
2866: b = (Mat_MPIAIJ*)B->data;
2868: #if defined(PETSC_USE_CTABLE)
2869: PetscTableDestroy(&b->colmap);
2870: #else
2871: PetscFree(b->colmap);
2872: #endif
2873: PetscFree(b->garray);
2874: VecDestroy(&b->lvec);
2875: VecScatterDestroy(&b->Mvctx);
2877: MatResetPreallocation(b->A);
2878: MatResetPreallocation(b->B);
2879: B->preallocated = PETSC_TRUE;
2880: B->was_assembled = PETSC_FALSE;
2881: B->assembled = PETSC_FALSE;
2882: return(0);
2883: }
2885: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2886: {
2887: Mat mat;
2888: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
2892: *newmat = 0;
2893: MatCreate(PetscObjectComm((PetscObject)matin),&mat);
2894: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
2895: MatSetBlockSizesFromMats(mat,matin,matin);
2896: MatSetType(mat,((PetscObject)matin)->type_name);
2897: a = (Mat_MPIAIJ*)mat->data;
2899: mat->factortype = matin->factortype;
2900: mat->assembled = PETSC_TRUE;
2901: mat->insertmode = NOT_SET_VALUES;
2902: mat->preallocated = PETSC_TRUE;
2904: a->size = oldmat->size;
2905: a->rank = oldmat->rank;
2906: a->donotstash = oldmat->donotstash;
2907: a->roworiented = oldmat->roworiented;
2908: a->rowindices = 0;
2909: a->rowvalues = 0;
2910: a->getrowactive = PETSC_FALSE;
2912: PetscLayoutReference(matin->rmap,&mat->rmap);
2913: PetscLayoutReference(matin->cmap,&mat->cmap);
2915: if (oldmat->colmap) {
2916: #if defined(PETSC_USE_CTABLE)
2917: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2918: #else
2919: PetscMalloc1(mat->cmap->N,&a->colmap);
2920: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
2921: PetscArraycpy(a->colmap,oldmat->colmap,mat->cmap->N);
2922: #endif
2923: } else a->colmap = 0;
2924: if (oldmat->garray) {
2925: PetscInt len;
2926: len = oldmat->B->cmap->n;
2927: PetscMalloc1(len+1,&a->garray);
2928: PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
2929: if (len) { PetscArraycpy(a->garray,oldmat->garray,len); }
2930: } else a->garray = 0;
2932: /* It may happen MatDuplicate is called with a non-assembled matrix
2933: In fact, MatDuplicate only requires the matrix to be preallocated
2934: This may happen inside a DMCreateMatrix_Shell */
2935: if (oldmat->lvec) {
2936: VecDuplicate(oldmat->lvec,&a->lvec);
2937: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
2938: }
2939: if (oldmat->Mvctx) {
2940: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2941: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
2942: }
2943: if (oldmat->Mvctx_mpi1) {
2944: VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);
2945: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);
2946: }
2948: MatDuplicate(oldmat->A,cpvalues,&a->A);
2949: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
2950: MatDuplicate(oldmat->B,cpvalues,&a->B);
2951: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
2952: PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2953: *newmat = mat;
2954: return(0);
2955: }
2957: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
2958: {
2959: PetscBool isbinary, ishdf5;
2965: /* force binary viewer to load .info file if it has not yet done so */
2966: PetscViewerSetUp(viewer);
2967: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
2968: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);
2969: if (isbinary) {
2970: MatLoad_MPIAIJ_Binary(newMat,viewer);
2971: } else if (ishdf5) {
2972: #if defined(PETSC_HAVE_HDF5)
2973: MatLoad_AIJ_HDF5(newMat,viewer);
2974: #else
2975: SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
2976: #endif
2977: } else {
2978: SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name);
2979: }
2980: return(0);
2981: }
2983: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat newMat, PetscViewer viewer)
2984: {
2985: PetscScalar *vals,*svals;
2986: MPI_Comm comm;
2988: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
2989: PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0;
2990: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
2991: PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols;
2992: PetscInt cend,cstart,n,*rowners;
2993: int fd;
2994: PetscInt bs = newMat->rmap->bs;
2997: PetscObjectGetComm((PetscObject)viewer,&comm);
2998: MPI_Comm_size(comm,&size);
2999: MPI_Comm_rank(comm,&rank);
3000: PetscViewerBinaryGetDescriptor(viewer,&fd);
3001: if (!rank) {
3002: PetscBinaryRead(fd,(char*)header,4,NULL,PETSC_INT);
3003: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3004: if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ");
3005: }
3007: PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");
3008: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
3009: PetscOptionsEnd();
3010: if (bs < 0) bs = 1;
3012: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3013: M = header[1]; N = header[2];
3015: /* If global sizes are set, check if they are consistent with that given in the file */
3016: if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M);
3017: if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N);
3019: /* determine ownership of all (block) rows */
3020: if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs);
3021: if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */
3022: else m = newMat->rmap->n; /* Set by user */
3024: PetscMalloc1(size+1,&rowners);
3025: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
3027: /* First process needs enough room for process with most rows */
3028: if (!rank) {
3029: mmax = rowners[1];
3030: for (i=2; i<=size; i++) {
3031: mmax = PetscMax(mmax, rowners[i]);
3032: }
3033: } else mmax = -1; /* unused, but compilers complain */
3035: rowners[0] = 0;
3036: for (i=2; i<=size; i++) {
3037: rowners[i] += rowners[i-1];
3038: }
3039: rstart = rowners[rank];
3040: rend = rowners[rank+1];
3042: /* distribute row lengths to all processors */
3043: PetscMalloc2(m,&ourlens,m,&offlens);
3044: if (!rank) {
3045: PetscBinaryRead(fd,ourlens,m,NULL,PETSC_INT);
3046: PetscMalloc1(mmax,&rowlengths);
3047: PetscCalloc1(size,&procsnz);
3048: for (j=0; j<m; j++) {
3049: procsnz[0] += ourlens[j];
3050: }
3051: for (i=1; i<size; i++) {
3052: PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],NULL,PETSC_INT);
3053: /* calculate the number of nonzeros on each processor */
3054: for (j=0; j<rowners[i+1]-rowners[i]; j++) {
3055: procsnz[i] += rowlengths[j];
3056: }
3057: MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
3058: }
3059: PetscFree(rowlengths);
3060: } else {
3061: MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
3062: }
3064: if (!rank) {
3065: /* determine max buffer needed and allocate it */
3066: maxnz = 0;
3067: for (i=0; i<size; i++) {
3068: maxnz = PetscMax(maxnz,procsnz[i]);
3069: }
3070: PetscMalloc1(maxnz,&cols);
3072: /* read in my part of the matrix column indices */
3073: nz = procsnz[0];
3074: PetscMalloc1(nz,&mycols);
3075: PetscBinaryRead(fd,mycols,nz,NULL,PETSC_INT);
3077: /* read in every one elses and ship off */
3078: for (i=1; i<size; i++) {
3079: nz = procsnz[i];
3080: PetscBinaryRead(fd,cols,nz,NULL,PETSC_INT);
3081: MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
3082: }
3083: PetscFree(cols);
3084: } else {
3085: /* determine buffer space needed for message */
3086: nz = 0;
3087: for (i=0; i<m; i++) {
3088: nz += ourlens[i];
3089: }
3090: PetscMalloc1(nz,&mycols);
3092: /* receive message of column indices*/
3093: MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
3094: }
3096: /* determine column ownership if matrix is not square */
3097: if (N != M) {
3098: if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank);
3099: else n = newMat->cmap->n;
3100: MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
3101: cstart = cend - n;
3102: } else {
3103: cstart = rstart;
3104: cend = rend;
3105: n = cend - cstart;
3106: }
3108: /* loop over local rows, determining number of off diagonal entries */
3109: PetscArrayzero(offlens,m);
3110: jj = 0;
3111: for (i=0; i<m; i++) {
3112: for (j=0; j<ourlens[i]; j++) {
3113: if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
3114: jj++;
3115: }
3116: }
3118: for (i=0; i<m; i++) {
3119: ourlens[i] -= offlens[i];
3120: }
3121: MatSetSizes(newMat,m,n,M,N);
3123: if (bs > 1) {MatSetBlockSize(newMat,bs);}
3125: MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);
3127: for (i=0; i<m; i++) {
3128: ourlens[i] += offlens[i];
3129: }
3131: if (!rank) {
3132: PetscMalloc1(maxnz+1,&vals);
3134: /* read in my part of the matrix numerical values */
3135: nz = procsnz[0];
3136: PetscBinaryRead(fd,vals,nz,NULL,PETSC_SCALAR);
3138: /* insert into matrix */
3139: jj = rstart;
3140: smycols = mycols;
3141: svals = vals;
3142: for (i=0; i<m; i++) {
3143: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3144: smycols += ourlens[i];
3145: svals += ourlens[i];
3146: jj++;
3147: }
3149: /* read in other processors and ship out */
3150: for (i=1; i<size; i++) {
3151: nz = procsnz[i];
3152: PetscBinaryRead(fd,vals,nz,NULL,PETSC_SCALAR);
3153: MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3154: }
3155: PetscFree(procsnz);
3156: } else {
3157: /* receive numeric values */
3158: PetscMalloc1(nz+1,&vals);
3160: /* receive message of values*/
3161: MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);
3163: /* insert into matrix */
3164: jj = rstart;
3165: smycols = mycols;
3166: svals = vals;
3167: for (i=0; i<m; i++) {
3168: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3169: smycols += ourlens[i];
3170: svals += ourlens[i];
3171: jj++;
3172: }
3173: }
3174: PetscFree2(ourlens,offlens);
3175: PetscFree(vals);
3176: PetscFree(mycols);
3177: PetscFree(rowners);
3178: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3179: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3180: return(0);
3181: }
3183: /* Not scalable because of ISAllGather() unless getting all columns. */
3184: PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq)
3185: {
3187: IS iscol_local;
3188: PetscBool isstride;
3189: PetscMPIInt lisstride=0,gisstride;
3192: /* check if we are grabbing all columns*/
3193: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);
3195: if (isstride) {
3196: PetscInt start,len,mstart,mlen;
3197: ISStrideGetInfo(iscol,&start,NULL);
3198: ISGetLocalSize(iscol,&len);
3199: MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
3200: if (mstart == start && mlen-mstart == len) lisstride = 1;
3201: }
3203: MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3204: if (gisstride) {
3205: PetscInt N;
3206: MatGetSize(mat,NULL,&N);
3207: ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);
3208: ISSetIdentity(iscol_local);
3209: PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3210: } else {
3211: PetscInt cbs;
3212: ISGetBlockSize(iscol,&cbs);
3213: ISAllGather(iscol,&iscol_local);
3214: ISSetBlockSize(iscol_local,cbs);
3215: }
3217: *isseq = iscol_local;
3218: return(0);
3219: }
3221: /*
3222: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3223: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3225: Input Parameters:
3226: mat - matrix
3227: isrow - parallel row index set; its local indices are a subset of local columns of mat,
3228: i.e., mat->rstart <= isrow[i] < mat->rend
3229: iscol - parallel column index set; its local indices are a subset of local columns of mat,
3230: i.e., mat->cstart <= iscol[i] < mat->cend
3231: Output Parameter:
3232: isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A
3233: iscol_o - sequential column index set for retrieving mat->B
3234: garray - column map; garray[i] indicates global location of iscol_o[i] in iscol
3235: */
3236: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[])
3237: {
3239: Vec x,cmap;
3240: const PetscInt *is_idx;
3241: PetscScalar *xarray,*cmaparray;
3242: PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count;
3243: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3244: Mat B=a->B;
3245: Vec lvec=a->lvec,lcmap;
3246: PetscInt i,cstart,cend,Bn=B->cmap->N;
3247: MPI_Comm comm;
3248: VecScatter Mvctx=a->Mvctx;
3251: PetscObjectGetComm((PetscObject)mat,&comm);
3252: ISGetLocalSize(iscol,&ncols);
3254: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3255: MatCreateVecs(mat,&x,NULL);
3256: VecSet(x,-1.0);
3257: VecDuplicate(x,&cmap);
3258: VecSet(cmap,-1.0);
3260: /* Get start indices */
3261: MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);
3262: isstart -= ncols;
3263: MatGetOwnershipRangeColumn(mat,&cstart,&cend);
3265: ISGetIndices(iscol,&is_idx);
3266: VecGetArray(x,&xarray);
3267: VecGetArray(cmap,&cmaparray);
3268: PetscMalloc1(ncols,&idx);
3269: for (i=0; i<ncols; i++) {
3270: xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i];
3271: cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */
3272: idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */
3273: }
3274: VecRestoreArray(x,&xarray);
3275: VecRestoreArray(cmap,&cmaparray);
3276: ISRestoreIndices(iscol,&is_idx);
3278: /* Get iscol_d */
3279: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);
3280: ISGetBlockSize(iscol,&i);
3281: ISSetBlockSize(*iscol_d,i);
3283: /* Get isrow_d */
3284: ISGetLocalSize(isrow,&m);
3285: rstart = mat->rmap->rstart;
3286: PetscMalloc1(m,&idx);
3287: ISGetIndices(isrow,&is_idx);
3288: for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart;
3289: ISRestoreIndices(isrow,&is_idx);
3291: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);
3292: ISGetBlockSize(isrow,&i);
3293: ISSetBlockSize(*isrow_d,i);
3295: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3296: VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3297: VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3299: VecDuplicate(lvec,&lcmap);
3301: VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3302: VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3304: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3305: /* off-process column indices */
3306: count = 0;
3307: PetscMalloc1(Bn,&idx);
3308: PetscMalloc1(Bn,&cmap1);
3310: VecGetArray(lvec,&xarray);
3311: VecGetArray(lcmap,&cmaparray);
3312: for (i=0; i<Bn; i++) {
3313: if (PetscRealPart(xarray[i]) > -1.0) {
3314: idx[count] = i; /* local column index in off-diagonal part B */
3315: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3316: count++;
3317: }
3318: }
3319: VecRestoreArray(lvec,&xarray);
3320: VecRestoreArray(lcmap,&cmaparray);
3322: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);
3323: /* cannot ensure iscol_o has same blocksize as iscol! */
3325: PetscFree(idx);
3326: *garray = cmap1;
3328: VecDestroy(&x);
3329: VecDestroy(&cmap);
3330: VecDestroy(&lcmap);
3331: return(0);
3332: }
3334: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3335: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat)
3336: {
3338: Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub;
3339: Mat M = NULL;
3340: MPI_Comm comm;
3341: IS iscol_d,isrow_d,iscol_o;
3342: Mat Asub = NULL,Bsub = NULL;
3343: PetscInt n;
3346: PetscObjectGetComm((PetscObject)mat,&comm);
3348: if (call == MAT_REUSE_MATRIX) {
3349: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3350: PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);
3351: if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse");
3353: PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);
3354: if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse");
3356: PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);
3357: if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse");
3359: /* Update diagonal and off-diagonal portions of submat */
3360: asub = (Mat_MPIAIJ*)(*submat)->data;
3361: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);
3362: ISGetLocalSize(iscol_o,&n);
3363: if (n) {
3364: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);
3365: }
3366: MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3367: MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);
3369: } else { /* call == MAT_INITIAL_MATRIX) */
3370: const PetscInt *garray;
3371: PetscInt BsubN;
3373: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3374: ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);
3376: /* Create local submatrices Asub and Bsub */
3377: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);
3378: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);
3380: /* Create submatrix M */
3381: MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);
3383: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3384: asub = (Mat_MPIAIJ*)M->data;
3386: ISGetLocalSize(iscol_o,&BsubN);
3387: n = asub->B->cmap->N;
3388: if (BsubN > n) {
3389: /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */
3390: const PetscInt *idx;
3391: PetscInt i,j,*idx_new,*subgarray = asub->garray;
3392: PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);
3394: PetscMalloc1(n,&idx_new);
3395: j = 0;
3396: ISGetIndices(iscol_o,&idx);
3397: for (i=0; i<n; i++) {
3398: if (j >= BsubN) break;
3399: while (subgarray[i] > garray[j]) j++;
3401: if (subgarray[i] == garray[j]) {
3402: idx_new[i] = idx[j++];
3403: } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]);
3404: }
3405: ISRestoreIndices(iscol_o,&idx);
3407: ISDestroy(&iscol_o);
3408: ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);
3410: } else if (BsubN < n) {
3411: SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N);
3412: }
3414: PetscFree(garray);
3415: *submat = M;
3417: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3418: PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);
3419: ISDestroy(&isrow_d);
3421: PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);
3422: ISDestroy(&iscol_d);
3424: PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);
3425: ISDestroy(&iscol_o);
3426: }
3427: return(0);
3428: }
3430: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3431: {
3433: IS iscol_local=NULL,isrow_d;
3434: PetscInt csize;
3435: PetscInt n,i,j,start,end;
3436: PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2];
3437: MPI_Comm comm;
3440: /* If isrow has same processor distribution as mat,
3441: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3442: if (call == MAT_REUSE_MATRIX) {
3443: PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);
3444: if (isrow_d) {
3445: sameRowDist = PETSC_TRUE;
3446: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3447: } else {
3448: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);
3449: if (iscol_local) {
3450: sameRowDist = PETSC_TRUE;
3451: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3452: }
3453: }
3454: } else {
3455: /* Check if isrow has same processor distribution as mat */
3456: sameDist[0] = PETSC_FALSE;
3457: ISGetLocalSize(isrow,&n);
3458: if (!n) {
3459: sameDist[0] = PETSC_TRUE;
3460: } else {
3461: ISGetMinMax(isrow,&i,&j);
3462: MatGetOwnershipRange(mat,&start,&end);
3463: if (i >= start && j < end) {
3464: sameDist[0] = PETSC_TRUE;
3465: }
3466: }
3468: /* Check if iscol has same processor distribution as mat */
3469: sameDist[1] = PETSC_FALSE;
3470: ISGetLocalSize(iscol,&n);
3471: if (!n) {
3472: sameDist[1] = PETSC_TRUE;
3473: } else {
3474: ISGetMinMax(iscol,&i,&j);
3475: MatGetOwnershipRangeColumn(mat,&start,&end);
3476: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3477: }
3479: PetscObjectGetComm((PetscObject)mat,&comm);
3480: MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);
3481: sameRowDist = tsameDist[0];
3482: }
3484: if (sameRowDist) {
3485: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3486: /* isrow and iscol have same processor distribution as mat */
3487: MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);
3488: return(0);
3489: } else { /* sameRowDist */
3490: /* isrow has same processor distribution as mat */
3491: if (call == MAT_INITIAL_MATRIX) {
3492: PetscBool sorted;
3493: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3494: ISGetLocalSize(iscol_local,&n); /* local size of iscol_local = global columns of newmat */
3495: ISGetSize(iscol,&i);
3496: if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i);
3498: ISSorted(iscol_local,&sorted);
3499: if (sorted) {
3500: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3501: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);
3502: return(0);
3503: }
3504: } else { /* call == MAT_REUSE_MATRIX */
3505: IS iscol_sub;
3506: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3507: if (iscol_sub) {
3508: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);
3509: return(0);
3510: }
3511: }
3512: }
3513: }
3515: /* General case: iscol -> iscol_local which has global size of iscol */
3516: if (call == MAT_REUSE_MATRIX) {
3517: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3518: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3519: } else {
3520: if (!iscol_local) {
3521: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3522: }
3523: }
3525: ISGetLocalSize(iscol,&csize);
3526: MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);
3528: if (call == MAT_INITIAL_MATRIX) {
3529: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3530: ISDestroy(&iscol_local);
3531: }
3532: return(0);
3533: }
3535: /*@C
3536: MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal"
3537: and "off-diagonal" part of the matrix in CSR format.
3539: Collective
3541: Input Parameters:
3542: + comm - MPI communicator
3543: . A - "diagonal" portion of matrix
3544: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3545: - garray - global index of B columns
3547: Output Parameter:
3548: . mat - the matrix, with input A as its local diagonal matrix
3549: Level: advanced
3551: Notes:
3552: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3553: A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore.
3555: .seealso: MatCreateMPIAIJWithSplitArrays()
3556: @*/
3557: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat)
3558: {
3560: Mat_MPIAIJ *maij;
3561: Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew;
3562: PetscInt *oi=b->i,*oj=b->j,i,nz,col;
3563: PetscScalar *oa=b->a;
3564: Mat Bnew;
3565: PetscInt m,n,N;
3568: MatCreate(comm,mat);
3569: MatGetSize(A,&m,&n);
3570: if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N);
3571: if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs);
3572: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3573: /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */
3575: /* Get global columns of mat */
3576: MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);
3578: MatSetSizes(*mat,m,n,PETSC_DECIDE,N);
3579: MatSetType(*mat,MATMPIAIJ);
3580: MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);
3581: maij = (Mat_MPIAIJ*)(*mat)->data;
3583: (*mat)->preallocated = PETSC_TRUE;
3585: PetscLayoutSetUp((*mat)->rmap);
3586: PetscLayoutSetUp((*mat)->cmap);
3588: /* Set A as diagonal portion of *mat */
3589: maij->A = A;
3591: nz = oi[m];
3592: for (i=0; i<nz; i++) {
3593: col = oj[i];
3594: oj[i] = garray[col];
3595: }
3597: /* Set Bnew as off-diagonal portion of *mat */
3598: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);
3599: bnew = (Mat_SeqAIJ*)Bnew->data;
3600: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3601: maij->B = Bnew;
3603: if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N);
3605: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3606: b->free_a = PETSC_FALSE;
3607: b->free_ij = PETSC_FALSE;
3608: MatDestroy(&B);
3610: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3611: bnew->free_a = PETSC_TRUE;
3612: bnew->free_ij = PETSC_TRUE;
3614: /* condense columns of maij->B */
3615: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
3616: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3617: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3618: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
3619: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3620: return(0);
3621: }
3623: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*);
3625: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat)
3626: {
3628: PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs;
3629: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3630: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3631: Mat M,Msub,B=a->B;
3632: MatScalar *aa;
3633: Mat_SeqAIJ *aij;
3634: PetscInt *garray = a->garray,*colsub,Ncols;
3635: PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend;
3636: IS iscol_sub,iscmap;
3637: const PetscInt *is_idx,*cmap;
3638: PetscBool allcolumns=PETSC_FALSE;
3639: MPI_Comm comm;
3642: PetscObjectGetComm((PetscObject)mat,&comm);
3644: if (call == MAT_REUSE_MATRIX) {
3645: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3646: if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse");
3647: ISGetLocalSize(iscol_sub,&count);
3649: PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);
3650: if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse");
3652: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);
3653: if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3655: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);
3657: } else { /* call == MAT_INITIAL_MATRIX) */
3658: PetscBool flg;
3660: ISGetLocalSize(iscol,&n);
3661: ISGetSize(iscol,&Ncols);
3663: /* (1) iscol -> nonscalable iscol_local */
3664: /* Check for special case: each processor gets entire matrix columns */
3665: ISIdentity(iscol_local,&flg);
3666: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3667: if (allcolumns) {
3668: iscol_sub = iscol_local;
3669: PetscObjectReference((PetscObject)iscol_local);
3670: ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);
3672: } else {
3673: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3674: PetscInt *idx,*cmap1,k;
3675: PetscMalloc1(Ncols,&idx);
3676: PetscMalloc1(Ncols,&cmap1);
3677: ISGetIndices(iscol_local,&is_idx);
3678: count = 0;
3679: k = 0;
3680: for (i=0; i<Ncols; i++) {
3681: j = is_idx[i];
3682: if (j >= cstart && j < cend) {
3683: /* diagonal part of mat */
3684: idx[count] = j;
3685: cmap1[count++] = i; /* column index in submat */
3686: } else if (Bn) {
3687: /* off-diagonal part of mat */
3688: if (j == garray[k]) {
3689: idx[count] = j;
3690: cmap1[count++] = i; /* column index in submat */
3691: } else if (j > garray[k]) {
3692: while (j > garray[k] && k < Bn-1) k++;
3693: if (j == garray[k]) {
3694: idx[count] = j;
3695: cmap1[count++] = i; /* column index in submat */
3696: }
3697: }
3698: }
3699: }
3700: ISRestoreIndices(iscol_local,&is_idx);
3702: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);
3703: ISGetBlockSize(iscol,&cbs);
3704: ISSetBlockSize(iscol_sub,cbs);
3706: ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);
3707: }
3709: /* (3) Create sequential Msub */
3710: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);
3711: }
3713: ISGetLocalSize(iscol_sub,&count);
3714: aij = (Mat_SeqAIJ*)(Msub)->data;
3715: ii = aij->i;
3716: ISGetIndices(iscmap,&cmap);
3718: /*
3719: m - number of local rows
3720: Ncols - number of columns (same on all processors)
3721: rstart - first row in new global matrix generated
3722: */
3723: MatGetSize(Msub,&m,NULL);
3725: if (call == MAT_INITIAL_MATRIX) {
3726: /* (4) Create parallel newmat */
3727: PetscMPIInt rank,size;
3728: PetscInt csize;
3730: MPI_Comm_size(comm,&size);
3731: MPI_Comm_rank(comm,&rank);
3733: /*
3734: Determine the number of non-zeros in the diagonal and off-diagonal
3735: portions of the matrix in order to do correct preallocation
3736: */
3738: /* first get start and end of "diagonal" columns */
3739: ISGetLocalSize(iscol,&csize);
3740: if (csize == PETSC_DECIDE) {
3741: ISGetSize(isrow,&mglobal);
3742: if (mglobal == Ncols) { /* square matrix */
3743: nlocal = m;
3744: } else {
3745: nlocal = Ncols/size + ((Ncols % size) > rank);
3746: }
3747: } else {
3748: nlocal = csize;
3749: }
3750: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3751: rstart = rend - nlocal;
3752: if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols);
3754: /* next, compute all the lengths */
3755: jj = aij->j;
3756: PetscMalloc1(2*m+1,&dlens);
3757: olens = dlens + m;
3758: for (i=0; i<m; i++) {
3759: jend = ii[i+1] - ii[i];
3760: olen = 0;
3761: dlen = 0;
3762: for (j=0; j<jend; j++) {
3763: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3764: else dlen++;
3765: jj++;
3766: }
3767: olens[i] = olen;
3768: dlens[i] = dlen;
3769: }
3771: ISGetBlockSize(isrow,&bs);
3772: ISGetBlockSize(iscol,&cbs);
3774: MatCreate(comm,&M);
3775: MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);
3776: MatSetBlockSizes(M,bs,cbs);
3777: MatSetType(M,((PetscObject)mat)->type_name);
3778: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3779: PetscFree(dlens);
3781: } else { /* call == MAT_REUSE_MATRIX */
3782: M = *newmat;
3783: MatGetLocalSize(M,&i,NULL);
3784: if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3785: MatZeroEntries(M);
3786: /*
3787: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3788: rather than the slower MatSetValues().
3789: */
3790: M->was_assembled = PETSC_TRUE;
3791: M->assembled = PETSC_FALSE;
3792: }
3794: /* (5) Set values of Msub to *newmat */
3795: PetscMalloc1(count,&colsub);
3796: MatGetOwnershipRange(M,&rstart,NULL);
3798: jj = aij->j;
3799: aa = aij->a;
3800: for (i=0; i<m; i++) {
3801: row = rstart + i;
3802: nz = ii[i+1] - ii[i];
3803: for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]];
3804: MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);
3805: jj += nz; aa += nz;
3806: }
3807: ISRestoreIndices(iscmap,&cmap);
3809: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3810: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3812: PetscFree(colsub);
3814: /* save Msub, iscol_sub and iscmap used in processor for next request */
3815: if (call == MAT_INITIAL_MATRIX) {
3816: *newmat = M;
3817: PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);
3818: MatDestroy(&Msub);
3820: PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);
3821: ISDestroy(&iscol_sub);
3823: PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);
3824: ISDestroy(&iscmap);
3826: if (iscol_local) {
3827: PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);
3828: ISDestroy(&iscol_local);
3829: }
3830: }
3831: return(0);
3832: }
3834: /*
3835: Not great since it makes two copies of the submatrix, first an SeqAIJ
3836: in local and then by concatenating the local matrices the end result.
3837: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3839: Note: This requires a sequential iscol with all indices.
3840: */
3841: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3842: {
3844: PetscMPIInt rank,size;
3845: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3846: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3847: Mat M,Mreuse;
3848: MatScalar *aa,*vwork;
3849: MPI_Comm comm;
3850: Mat_SeqAIJ *aij;
3851: PetscBool colflag,allcolumns=PETSC_FALSE;
3854: PetscObjectGetComm((PetscObject)mat,&comm);
3855: MPI_Comm_rank(comm,&rank);
3856: MPI_Comm_size(comm,&size);
3858: /* Check for special case: each processor gets entire matrix columns */
3859: ISIdentity(iscol,&colflag);
3860: ISGetLocalSize(iscol,&n);
3861: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3863: if (call == MAT_REUSE_MATRIX) {
3864: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3865: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3866: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);
3867: } else {
3868: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);
3869: }
3871: /*
3872: m - number of local rows
3873: n - number of columns (same on all processors)
3874: rstart - first row in new global matrix generated
3875: */
3876: MatGetSize(Mreuse,&m,&n);
3877: MatGetBlockSizes(Mreuse,&bs,&cbs);
3878: if (call == MAT_INITIAL_MATRIX) {
3879: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3880: ii = aij->i;
3881: jj = aij->j;
3883: /*
3884: Determine the number of non-zeros in the diagonal and off-diagonal
3885: portions of the matrix in order to do correct preallocation
3886: */
3888: /* first get start and end of "diagonal" columns */
3889: if (csize == PETSC_DECIDE) {
3890: ISGetSize(isrow,&mglobal);
3891: if (mglobal == n) { /* square matrix */
3892: nlocal = m;
3893: } else {
3894: nlocal = n/size + ((n % size) > rank);
3895: }
3896: } else {
3897: nlocal = csize;
3898: }
3899: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3900: rstart = rend - nlocal;
3901: 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);
3903: /* next, compute all the lengths */
3904: PetscMalloc1(2*m+1,&dlens);
3905: olens = dlens + m;
3906: for (i=0; i<m; i++) {
3907: jend = ii[i+1] - ii[i];
3908: olen = 0;
3909: dlen = 0;
3910: for (j=0; j<jend; j++) {
3911: if (*jj < rstart || *jj >= rend) olen++;
3912: else dlen++;
3913: jj++;
3914: }
3915: olens[i] = olen;
3916: dlens[i] = dlen;
3917: }
3918: MatCreate(comm,&M);
3919: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3920: MatSetBlockSizes(M,bs,cbs);
3921: MatSetType(M,((PetscObject)mat)->type_name);
3922: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3923: PetscFree(dlens);
3924: } else {
3925: PetscInt ml,nl;
3927: M = *newmat;
3928: MatGetLocalSize(M,&ml,&nl);
3929: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3930: MatZeroEntries(M);
3931: /*
3932: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3933: rather than the slower MatSetValues().
3934: */
3935: M->was_assembled = PETSC_TRUE;
3936: M->assembled = PETSC_FALSE;
3937: }
3938: MatGetOwnershipRange(M,&rstart,&rend);
3939: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3940: ii = aij->i;
3941: jj = aij->j;
3942: aa = aij->a;
3943: for (i=0; i<m; i++) {
3944: row = rstart + i;
3945: nz = ii[i+1] - ii[i];
3946: cwork = jj; jj += nz;
3947: vwork = aa; aa += nz;
3948: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3949: }
3951: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3952: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3953: *newmat = M;
3955: /* save submatrix used in processor for next request */
3956: if (call == MAT_INITIAL_MATRIX) {
3957: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3958: MatDestroy(&Mreuse);
3959: }
3960: return(0);
3961: }
3963: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3964: {
3965: PetscInt m,cstart, cend,j,nnz,i,d;
3966: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3967: const PetscInt *JJ;
3969: PetscBool nooffprocentries;
3972: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
3974: PetscLayoutSetUp(B->rmap);
3975: PetscLayoutSetUp(B->cmap);
3976: m = B->rmap->n;
3977: cstart = B->cmap->rstart;
3978: cend = B->cmap->rend;
3979: rstart = B->rmap->rstart;
3981: PetscCalloc2(m,&d_nnz,m,&o_nnz);
3983: #if defined(PETSC_USE_DEBUG)
3984: for (i=0; i<m; i++) {
3985: nnz = Ii[i+1]- Ii[i];
3986: JJ = J + Ii[i];
3987: if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3988: if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]);
3989: if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
3990: }
3991: #endif
3993: for (i=0; i<m; i++) {
3994: nnz = Ii[i+1]- Ii[i];
3995: JJ = J + Ii[i];
3996: nnz_max = PetscMax(nnz_max,nnz);
3997: d = 0;
3998: for (j=0; j<nnz; j++) {
3999: if (cstart <= JJ[j] && JJ[j] < cend) d++;
4000: }
4001: d_nnz[i] = d;
4002: o_nnz[i] = nnz - d;
4003: }
4004: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
4005: PetscFree2(d_nnz,o_nnz);
4007: for (i=0; i<m; i++) {
4008: ii = i + rstart;
4009: MatSetValues_MPIAIJ(B,1,&ii,Ii[i+1] - Ii[i],J+Ii[i], v ? v + Ii[i] : NULL,INSERT_VALUES);
4010: }
4011: nooffprocentries = B->nooffprocentries;
4012: B->nooffprocentries = PETSC_TRUE;
4013: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4014: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4015: B->nooffprocentries = nooffprocentries;
4017: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
4018: return(0);
4019: }
4021: /*@
4022: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
4023: (the default parallel PETSc format).
4025: Collective
4027: Input Parameters:
4028: + B - the matrix
4029: . i - the indices into j for the start of each local row (starts with zero)
4030: . j - the column indices for each local row (starts with zero)
4031: - v - optional values in the matrix
4033: Level: developer
4035: Notes:
4036: The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc;
4037: thus you CANNOT change the matrix entries by changing the values of v[] after you have
4038: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4040: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4042: The format which is used for the sparse matrix input, is equivalent to a
4043: row-major ordering.. i.e for the following matrix, the input data expected is
4044: as shown
4046: $ 1 0 0
4047: $ 2 0 3 P0
4048: $ -------
4049: $ 4 5 6 P1
4050: $
4051: $ Process0 [P0]: rows_owned=[0,1]
4052: $ i = {0,1,3} [size = nrow+1 = 2+1]
4053: $ j = {0,0,2} [size = 3]
4054: $ v = {1,2,3} [size = 3]
4055: $
4056: $ Process1 [P1]: rows_owned=[2]
4057: $ i = {0,3} [size = nrow+1 = 1+1]
4058: $ j = {0,1,2} [size = 3]
4059: $ v = {4,5,6} [size = 3]
4061: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ,
4062: MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
4063: @*/
4064: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
4065: {
4069: PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4070: return(0);
4071: }
4073: /*@C
4074: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
4075: (the default parallel PETSc format). For good matrix assembly performance
4076: the user should preallocate the matrix storage by setting the parameters
4077: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4078: performance can be increased by more than a factor of 50.
4080: Collective
4082: Input Parameters:
4083: + B - the matrix
4084: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4085: (same value is used for all local rows)
4086: . d_nnz - array containing the number of nonzeros in the various rows of the
4087: DIAGONAL portion of the local submatrix (possibly different for each row)
4088: or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
4089: The size of this array is equal to the number of local rows, i.e 'm'.
4090: For matrices that will be factored, you must leave room for (and set)
4091: the diagonal entry even if it is zero.
4092: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4093: submatrix (same value is used for all local rows).
4094: - o_nnz - array containing the number of nonzeros in the various rows of the
4095: OFF-DIAGONAL portion of the local submatrix (possibly different for
4096: each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
4097: structure. The size of this array is equal to the number
4098: of local rows, i.e 'm'.
4100: If the *_nnz parameter is given then the *_nz parameter is ignored
4102: The AIJ format (also called the Yale sparse matrix format or
4103: compressed row storage (CSR)), is fully compatible with standard Fortran 77
4104: storage. The stored row and column indices begin with zero.
4105: See Users-Manual: ch_mat for details.
4107: The parallel matrix is partitioned such that the first m0 rows belong to
4108: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4109: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4111: The DIAGONAL portion of the local submatrix of a processor can be defined
4112: as the submatrix which is obtained by extraction the part corresponding to
4113: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4114: first row that belongs to the processor, r2 is the last row belonging to
4115: the this processor, and c1-c2 is range of indices of the local part of a
4116: vector suitable for applying the matrix to. This is an mxn matrix. In the
4117: common case of a square matrix, the row and column ranges are the same and
4118: the DIAGONAL part is also square. The remaining portion of the local
4119: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4121: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4123: You can call MatGetInfo() to get information on how effective the preallocation was;
4124: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4125: You can also run with the option -info and look for messages with the string
4126: malloc in them to see if additional memory allocation was needed.
4128: Example usage:
4130: Consider the following 8x8 matrix with 34 non-zero values, that is
4131: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4132: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4133: as follows:
4135: .vb
4136: 1 2 0 | 0 3 0 | 0 4
4137: Proc0 0 5 6 | 7 0 0 | 8 0
4138: 9 0 10 | 11 0 0 | 12 0
4139: -------------------------------------
4140: 13 0 14 | 15 16 17 | 0 0
4141: Proc1 0 18 0 | 19 20 21 | 0 0
4142: 0 0 0 | 22 23 0 | 24 0
4143: -------------------------------------
4144: Proc2 25 26 27 | 0 0 28 | 29 0
4145: 30 0 0 | 31 32 33 | 0 34
4146: .ve
4148: This can be represented as a collection of submatrices as:
4150: .vb
4151: A B C
4152: D E F
4153: G H I
4154: .ve
4156: Where the submatrices A,B,C are owned by proc0, D,E,F are
4157: owned by proc1, G,H,I are owned by proc2.
4159: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4160: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4161: The 'M','N' parameters are 8,8, and have the same values on all procs.
4163: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4164: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4165: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4166: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4167: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4168: matrix, ans [DF] as another SeqAIJ matrix.
4170: When d_nz, o_nz parameters are specified, d_nz storage elements are
4171: allocated for every row of the local diagonal submatrix, and o_nz
4172: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4173: One way to choose d_nz and o_nz is to use the max nonzerors per local
4174: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4175: In this case, the values of d_nz,o_nz are:
4176: .vb
4177: proc0 : dnz = 2, o_nz = 2
4178: proc1 : dnz = 3, o_nz = 2
4179: proc2 : dnz = 1, o_nz = 4
4180: .ve
4181: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4182: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4183: for proc3. i.e we are using 12+15+10=37 storage locations to store
4184: 34 values.
4186: When d_nnz, o_nnz parameters are specified, the storage is specified
4187: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4188: In the above case the values for d_nnz,o_nnz are:
4189: .vb
4190: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4191: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4192: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4193: .ve
4194: Here the space allocated is sum of all the above values i.e 34, and
4195: hence pre-allocation is perfect.
4197: Level: intermediate
4199: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
4200: MATMPIAIJ, MatGetInfo(), PetscSplitOwnership()
4201: @*/
4202: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
4203: {
4209: PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
4210: return(0);
4211: }
4213: /*@
4214: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
4215: CSR format for the local rows.
4217: Collective
4219: Input Parameters:
4220: + comm - MPI communicator
4221: . m - number of local rows (Cannot be PETSC_DECIDE)
4222: . n - This value should be the same as the local size used in creating the
4223: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4224: calculated if N is given) For square matrices n is almost always m.
4225: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4226: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4227: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4228: . j - column indices
4229: - a - matrix values
4231: Output Parameter:
4232: . mat - the matrix
4234: Level: intermediate
4236: Notes:
4237: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4238: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4239: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4241: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4243: The format which is used for the sparse matrix input, is equivalent to a
4244: row-major ordering.. i.e for the following matrix, the input data expected is
4245: as shown
4247: Once you have created the matrix you can update it with new numerical values using MatUpdateMPIAIJWithArrays
4249: $ 1 0 0
4250: $ 2 0 3 P0
4251: $ -------
4252: $ 4 5 6 P1
4253: $
4254: $ Process0 [P0]: rows_owned=[0,1]
4255: $ i = {0,1,3} [size = nrow+1 = 2+1]
4256: $ j = {0,0,2} [size = 3]
4257: $ v = {1,2,3} [size = 3]
4258: $
4259: $ Process1 [P1]: rows_owned=[2]
4260: $ i = {0,3} [size = nrow+1 = 1+1]
4261: $ j = {0,1,2} [size = 3]
4262: $ v = {4,5,6} [size = 3]
4264: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4265: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4266: @*/
4267: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4268: {
4272: if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4273: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4274: MatCreate(comm,mat);
4275: MatSetSizes(*mat,m,n,M,N);
4276: /* MatSetBlockSizes(M,bs,cbs); */
4277: MatSetType(*mat,MATMPIAIJ);
4278: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4279: return(0);
4280: }
4282: /*@
4283: MatUpdateMPIAIJWithArrays - updates a MPI AIJ matrix using arrays that contain in standard
4284: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical
4286: Collective
4288: Input Parameters:
4289: + mat - the matrix
4290: . m - number of local rows (Cannot be PETSC_DECIDE)
4291: . n - This value should be the same as the local size used in creating the
4292: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4293: calculated if N is given) For square matrices n is almost always m.
4294: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4295: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4296: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4297: . J - column indices
4298: - v - matrix values
4300: Level: intermediate
4302: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4303: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4304: @*/
4305: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
4306: {
4308: PetscInt cstart,nnz,i,j;
4309: PetscInt *ld;
4310: PetscBool nooffprocentries;
4311: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*)mat->data;
4312: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)Aij->A->data, *Ao = (Mat_SeqAIJ*)Aij->B->data;
4313: PetscScalar *ad = Ad->a, *ao = Ao->a;
4314: const PetscInt *Adi = Ad->i;
4315: PetscInt ldi,Iii,md;
4318: if (Ii[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4319: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4320: if (m != mat->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4321: if (n != mat->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4323: cstart = mat->cmap->rstart;
4324: if (!Aij->ld) {
4325: /* count number of entries below block diagonal */
4326: PetscCalloc1(m,&ld);
4327: Aij->ld = ld;
4328: for (i=0; i<m; i++) {
4329: nnz = Ii[i+1]- Ii[i];
4330: j = 0;
4331: while (J[j] < cstart && j < nnz) {j++;}
4332: J += nnz;
4333: ld[i] = j;
4334: }
4335: } else {
4336: ld = Aij->ld;
4337: }
4339: for (i=0; i<m; i++) {
4340: nnz = Ii[i+1]- Ii[i];
4341: Iii = Ii[i];
4342: ldi = ld[i];
4343: md = Adi[i+1]-Adi[i];
4344: PetscArraycpy(ao,v + Iii,ldi);
4345: PetscArraycpy(ad,v + Iii + ldi,md);
4346: PetscArraycpy(ao + ldi,v + Iii + ldi + md,nnz - ldi - md);
4347: ad += md;
4348: ao += nnz - md;
4349: }
4350: nooffprocentries = mat->nooffprocentries;
4351: mat->nooffprocentries = PETSC_TRUE;
4352: PetscObjectStateIncrease((PetscObject)Aij->A);
4353: PetscObjectStateIncrease((PetscObject)Aij->B);
4354: PetscObjectStateIncrease((PetscObject)mat);
4355: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
4356: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
4357: mat->nooffprocentries = nooffprocentries;
4358: return(0);
4359: }
4361: /*@C
4362: MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4363: (the default parallel PETSc format). For good matrix assembly performance
4364: the user should preallocate the matrix storage by setting the parameters
4365: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4366: performance can be increased by more than a factor of 50.
4368: Collective
4370: Input Parameters:
4371: + comm - MPI communicator
4372: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4373: This value should be the same as the local size used in creating the
4374: y vector for the matrix-vector product y = Ax.
4375: . n - This value should be the same as the local size used in creating the
4376: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4377: calculated if N is given) For square matrices n is almost always m.
4378: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4379: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4380: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4381: (same value is used for all local rows)
4382: . d_nnz - array containing the number of nonzeros in the various rows of the
4383: DIAGONAL portion of the local submatrix (possibly different for each row)
4384: or NULL, if d_nz is used to specify the nonzero structure.
4385: The size of this array is equal to the number of local rows, i.e 'm'.
4386: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4387: submatrix (same value is used for all local rows).
4388: - o_nnz - array containing the number of nonzeros in the various rows of the
4389: OFF-DIAGONAL portion of the local submatrix (possibly different for
4390: each row) or NULL, if o_nz is used to specify the nonzero
4391: structure. The size of this array is equal to the number
4392: of local rows, i.e 'm'.
4394: Output Parameter:
4395: . A - the matrix
4397: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4398: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4399: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
4401: Notes:
4402: If the *_nnz parameter is given then the *_nz parameter is ignored
4404: m,n,M,N parameters specify the size of the matrix, and its partitioning across
4405: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4406: storage requirements for this matrix.
4408: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
4409: processor than it must be used on all processors that share the object for
4410: that argument.
4412: The user MUST specify either the local or global matrix dimensions
4413: (possibly both).
4415: The parallel matrix is partitioned across processors such that the
4416: first m0 rows belong to process 0, the next m1 rows belong to
4417: process 1, the next m2 rows belong to process 2 etc.. where
4418: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4419: values corresponding to [m x N] submatrix.
4421: The columns are logically partitioned with the n0 columns belonging
4422: to 0th partition, the next n1 columns belonging to the next
4423: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4425: The DIAGONAL portion of the local submatrix on any given processor
4426: is the submatrix corresponding to the rows and columns m,n
4427: corresponding to the given processor. i.e diagonal matrix on
4428: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4429: etc. The remaining portion of the local submatrix [m x (N-n)]
4430: constitute the OFF-DIAGONAL portion. The example below better
4431: illustrates this concept.
4433: For a square global matrix we define each processor's diagonal portion
4434: to be its local rows and the corresponding columns (a square submatrix);
4435: each processor's off-diagonal portion encompasses the remainder of the
4436: local matrix (a rectangular submatrix).
4438: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4440: When calling this routine with a single process communicator, a matrix of
4441: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
4442: type of communicator, use the construction mechanism
4443: .vb
4444: MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4445: .ve
4447: $ MatCreate(...,&A);
4448: $ MatSetType(A,MATMPIAIJ);
4449: $ MatSetSizes(A, m,n,M,N);
4450: $ MatMPIAIJSetPreallocation(A,...);
4452: By default, this format uses inodes (identical nodes) when possible.
4453: We search for consecutive rows with the same nonzero structure, thereby
4454: reusing matrix information to achieve increased efficiency.
4456: Options Database Keys:
4457: + -mat_no_inode - Do not use inodes
4458: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4462: Example usage:
4464: Consider the following 8x8 matrix with 34 non-zero values, that is
4465: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4466: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4467: as follows
4469: .vb
4470: 1 2 0 | 0 3 0 | 0 4
4471: Proc0 0 5 6 | 7 0 0 | 8 0
4472: 9 0 10 | 11 0 0 | 12 0
4473: -------------------------------------
4474: 13 0 14 | 15 16 17 | 0 0
4475: Proc1 0 18 0 | 19 20 21 | 0 0
4476: 0 0 0 | 22 23 0 | 24 0
4477: -------------------------------------
4478: Proc2 25 26 27 | 0 0 28 | 29 0
4479: 30 0 0 | 31 32 33 | 0 34
4480: .ve
4482: This can be represented as a collection of submatrices as
4484: .vb
4485: A B C
4486: D E F
4487: G H I
4488: .ve
4490: Where the submatrices A,B,C are owned by proc0, D,E,F are
4491: owned by proc1, G,H,I are owned by proc2.
4493: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4494: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4495: The 'M','N' parameters are 8,8, and have the same values on all procs.
4497: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4498: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4499: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4500: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4501: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4502: matrix, ans [DF] as another SeqAIJ matrix.
4504: When d_nz, o_nz parameters are specified, d_nz storage elements are
4505: allocated for every row of the local diagonal submatrix, and o_nz
4506: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4507: One way to choose d_nz and o_nz is to use the max nonzerors per local
4508: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4509: In this case, the values of d_nz,o_nz are
4510: .vb
4511: proc0 : dnz = 2, o_nz = 2
4512: proc1 : dnz = 3, o_nz = 2
4513: proc2 : dnz = 1, o_nz = 4
4514: .ve
4515: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4516: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4517: for proc3. i.e we are using 12+15+10=37 storage locations to store
4518: 34 values.
4520: When d_nnz, o_nnz parameters are specified, the storage is specified
4521: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4522: In the above case the values for d_nnz,o_nnz are
4523: .vb
4524: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4525: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4526: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4527: .ve
4528: Here the space allocated is sum of all the above values i.e 34, and
4529: hence pre-allocation is perfect.
4531: Level: intermediate
4533: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4534: MATMPIAIJ, MatCreateMPIAIJWithArrays()
4535: @*/
4536: 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)
4537: {
4539: PetscMPIInt size;
4542: MatCreate(comm,A);
4543: MatSetSizes(*A,m,n,M,N);
4544: MPI_Comm_size(comm,&size);
4545: if (size > 1) {
4546: MatSetType(*A,MATMPIAIJ);
4547: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4548: } else {
4549: MatSetType(*A,MATSEQAIJ);
4550: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4551: }
4552: return(0);
4553: }
4555: /*@C
4556: MatMPIAIJGetSeqAIJ - Returns the local piece of this distributed matrix
4557:
4558: Not collective
4559:
4560: Input Parameter:
4561: . A - The MPIAIJ matrix
4563: Output Parameters:
4564: + Ad - The local diagonal block as a SeqAIJ matrix
4565: . Ao - The local off-diagonal block as a SeqAIJ matrix
4566: - colmap - An array mapping local column numbers of Ao to global column numbers of the parallel matrix
4568: Note: The rows in Ad and Ao are in [0, Nr), where Nr is the number of local rows on this process. The columns
4569: in Ad are in [0, Nc) where Nc is the number of local columns. The columns are Ao are in [0, Nco), where Nco is
4570: the number of nonzero columns in the local off-diagonal piece of the matrix A. The array colmap maps these
4571: local column numbers to global column numbers in the original matrix.
4573: Level: intermediate
4575: .seealso: MatMPIAIJGetLocalMat(), MatMPIAIJGetLocalMatCondensed(), MatCreateAIJ(), MATMPIAJ, MATSEQAIJ
4576: @*/
4577: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
4578: {
4579: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4580: PetscBool flg;
4584: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);
4585: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input");
4586: if (Ad) *Ad = a->A;
4587: if (Ao) *Ao = a->B;
4588: if (colmap) *colmap = a->garray;
4589: return(0);
4590: }
4592: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4593: {
4595: PetscInt m,N,i,rstart,nnz,Ii;
4596: PetscInt *indx;
4597: PetscScalar *values;
4600: MatGetSize(inmat,&m,&N);
4601: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4602: PetscInt *dnz,*onz,sum,bs,cbs;
4604: if (n == PETSC_DECIDE) {
4605: PetscSplitOwnership(comm,&n,&N);
4606: }
4607: /* Check sum(n) = N */
4608: MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4609: if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N);
4611: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4612: rstart -= m;
4614: MatPreallocateInitialize(comm,m,n,dnz,onz);
4615: for (i=0; i<m; i++) {
4616: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4617: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4618: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4619: }
4621: MatCreate(comm,outmat);
4622: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4623: MatGetBlockSizes(inmat,&bs,&cbs);
4624: MatSetBlockSizes(*outmat,bs,cbs);
4625: MatSetType(*outmat,MATAIJ);
4626: MatSeqAIJSetPreallocation(*outmat,0,dnz);
4627: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4628: MatPreallocateFinalize(dnz,onz);
4629: }
4631: /* numeric phase */
4632: MatGetOwnershipRange(*outmat,&rstart,NULL);
4633: for (i=0; i<m; i++) {
4634: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4635: Ii = i + rstart;
4636: MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4637: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4638: }
4639: MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
4640: MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
4641: return(0);
4642: }
4644: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4645: {
4646: PetscErrorCode ierr;
4647: PetscMPIInt rank;
4648: PetscInt m,N,i,rstart,nnz;
4649: size_t len;
4650: const PetscInt *indx;
4651: PetscViewer out;
4652: char *name;
4653: Mat B;
4654: const PetscScalar *values;
4657: MatGetLocalSize(A,&m,0);
4658: MatGetSize(A,0,&N);
4659: /* Should this be the type of the diagonal block of A? */
4660: MatCreate(PETSC_COMM_SELF,&B);
4661: MatSetSizes(B,m,N,m,N);
4662: MatSetBlockSizesFromMats(B,A,A);
4663: MatSetType(B,MATSEQAIJ);
4664: MatSeqAIJSetPreallocation(B,0,NULL);
4665: MatGetOwnershipRange(A,&rstart,0);
4666: for (i=0; i<m; i++) {
4667: MatGetRow(A,i+rstart,&nnz,&indx,&values);
4668: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4669: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4670: }
4671: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4672: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4674: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
4675: PetscStrlen(outfile,&len);
4676: PetscMalloc1(len+5,&name);
4677: sprintf(name,"%s.%d",outfile,rank);
4678: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4679: PetscFree(name);
4680: MatView(B,out);
4681: PetscViewerDestroy(&out);
4682: MatDestroy(&B);
4683: return(0);
4684: }
4686: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4687: {
4688: PetscErrorCode ierr;
4689: Mat_Merge_SeqsToMPI *merge;
4690: PetscContainer container;
4693: PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);
4694: if (container) {
4695: PetscContainerGetPointer(container,(void**)&merge);
4696: PetscFree(merge->id_r);
4697: PetscFree(merge->len_s);
4698: PetscFree(merge->len_r);
4699: PetscFree(merge->bi);
4700: PetscFree(merge->bj);
4701: PetscFree(merge->buf_ri[0]);
4702: PetscFree(merge->buf_ri);
4703: PetscFree(merge->buf_rj[0]);
4704: PetscFree(merge->buf_rj);
4705: PetscFree(merge->coi);
4706: PetscFree(merge->coj);
4707: PetscFree(merge->owners_co);
4708: PetscLayoutDestroy(&merge->rowmap);
4709: PetscFree(merge);
4710: PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4711: }
4712: MatDestroy_MPIAIJ(A);
4713: return(0);
4714: }
4716: #include <../src/mat/utils/freespace.h>
4717: #include <petscbt.h>
4719: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4720: {
4721: PetscErrorCode ierr;
4722: MPI_Comm comm;
4723: Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data;
4724: PetscMPIInt size,rank,taga,*len_s;
4725: PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4726: PetscInt proc,m;
4727: PetscInt **buf_ri,**buf_rj;
4728: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4729: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
4730: MPI_Request *s_waits,*r_waits;
4731: MPI_Status *status;
4732: MatScalar *aa=a->a;
4733: MatScalar **abuf_r,*ba_i;
4734: Mat_Merge_SeqsToMPI *merge;
4735: PetscContainer container;
4738: PetscObjectGetComm((PetscObject)mpimat,&comm);
4739: PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);
4741: MPI_Comm_size(comm,&size);
4742: MPI_Comm_rank(comm,&rank);
4744: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4745: PetscContainerGetPointer(container,(void**)&merge);
4747: bi = merge->bi;
4748: bj = merge->bj;
4749: buf_ri = merge->buf_ri;
4750: buf_rj = merge->buf_rj;
4752: PetscMalloc1(size,&status);
4753: owners = merge->rowmap->range;
4754: len_s = merge->len_s;
4756: /* send and recv matrix values */
4757: /*-----------------------------*/
4758: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4759: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
4761: PetscMalloc1(merge->nsend+1,&s_waits);
4762: for (proc=0,k=0; proc<size; proc++) {
4763: if (!len_s[proc]) continue;
4764: i = owners[proc];
4765: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4766: k++;
4767: }
4769: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4770: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4771: PetscFree(status);
4773: PetscFree(s_waits);
4774: PetscFree(r_waits);
4776: /* insert mat values of mpimat */
4777: /*----------------------------*/
4778: PetscMalloc1(N,&ba_i);
4779: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4781: for (k=0; k<merge->nrecv; k++) {
4782: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4783: nrows = *(buf_ri_k[k]);
4784: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
4785: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4786: }
4788: /* set values of ba */
4789: m = merge->rowmap->n;
4790: for (i=0; i<m; i++) {
4791: arow = owners[rank] + i;
4792: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
4793: bnzi = bi[i+1] - bi[i];
4794: PetscArrayzero(ba_i,bnzi);
4796: /* add local non-zero vals of this proc's seqmat into ba */
4797: anzi = ai[arow+1] - ai[arow];
4798: aj = a->j + ai[arow];
4799: aa = a->a + ai[arow];
4800: nextaj = 0;
4801: for (j=0; nextaj<anzi; j++) {
4802: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4803: ba_i[j] += aa[nextaj++];
4804: }
4805: }
4807: /* add received vals into ba */
4808: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4809: /* i-th row */
4810: if (i == *nextrow[k]) {
4811: anzi = *(nextai[k]+1) - *nextai[k];
4812: aj = buf_rj[k] + *(nextai[k]);
4813: aa = abuf_r[k] + *(nextai[k]);
4814: nextaj = 0;
4815: for (j=0; nextaj<anzi; j++) {
4816: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4817: ba_i[j] += aa[nextaj++];
4818: }
4819: }
4820: nextrow[k]++; nextai[k]++;
4821: }
4822: }
4823: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4824: }
4825: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4826: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
4828: PetscFree(abuf_r[0]);
4829: PetscFree(abuf_r);
4830: PetscFree(ba_i);
4831: PetscFree3(buf_ri_k,nextrow,nextai);
4832: PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4833: return(0);
4834: }
4836: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4837: {
4838: PetscErrorCode ierr;
4839: Mat B_mpi;
4840: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4841: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4842: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
4843: PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4844: PetscInt len,proc,*dnz,*onz,bs,cbs;
4845: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4846: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4847: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
4848: MPI_Status *status;
4849: PetscFreeSpaceList free_space=NULL,current_space=NULL;
4850: PetscBT lnkbt;
4851: Mat_Merge_SeqsToMPI *merge;
4852: PetscContainer container;
4855: PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);
4857: /* make sure it is a PETSc comm */
4858: PetscCommDuplicate(comm,&comm,NULL);
4859: MPI_Comm_size(comm,&size);
4860: MPI_Comm_rank(comm,&rank);
4862: PetscNew(&merge);
4863: PetscMalloc1(size,&status);
4865: /* determine row ownership */
4866: /*---------------------------------------------------------*/
4867: PetscLayoutCreate(comm,&merge->rowmap);
4868: PetscLayoutSetLocalSize(merge->rowmap,m);
4869: PetscLayoutSetSize(merge->rowmap,M);
4870: PetscLayoutSetBlockSize(merge->rowmap,1);
4871: PetscLayoutSetUp(merge->rowmap);
4872: PetscMalloc1(size,&len_si);
4873: PetscMalloc1(size,&merge->len_s);
4875: m = merge->rowmap->n;
4876: owners = merge->rowmap->range;
4878: /* determine the number of messages to send, their lengths */
4879: /*---------------------------------------------------------*/
4880: len_s = merge->len_s;
4882: len = 0; /* length of buf_si[] */
4883: merge->nsend = 0;
4884: for (proc=0; proc<size; proc++) {
4885: len_si[proc] = 0;
4886: if (proc == rank) {
4887: len_s[proc] = 0;
4888: } else {
4889: len_si[proc] = owners[proc+1] - owners[proc] + 1;
4890: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4891: }
4892: if (len_s[proc]) {
4893: merge->nsend++;
4894: nrows = 0;
4895: for (i=owners[proc]; i<owners[proc+1]; i++) {
4896: if (ai[i+1] > ai[i]) nrows++;
4897: }
4898: len_si[proc] = 2*(nrows+1);
4899: len += len_si[proc];
4900: }
4901: }
4903: /* determine the number and length of messages to receive for ij-structure */
4904: /*-------------------------------------------------------------------------*/
4905: PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4906: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
4908: /* post the Irecv of j-structure */
4909: /*-------------------------------*/
4910: PetscCommGetNewTag(comm,&tagj);
4911: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
4913: /* post the Isend of j-structure */
4914: /*--------------------------------*/
4915: PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);
4917: for (proc=0, k=0; proc<size; proc++) {
4918: if (!len_s[proc]) continue;
4919: i = owners[proc];
4920: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4921: k++;
4922: }
4924: /* receives and sends of j-structure are complete */
4925: /*------------------------------------------------*/
4926: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4927: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
4929: /* send and recv i-structure */
4930: /*---------------------------*/
4931: PetscCommGetNewTag(comm,&tagi);
4932: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
4934: PetscMalloc1(len+1,&buf_s);
4935: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4936: for (proc=0,k=0; proc<size; proc++) {
4937: if (!len_s[proc]) continue;
4938: /* form outgoing message for i-structure:
4939: buf_si[0]: nrows to be sent
4940: [1:nrows]: row index (global)
4941: [nrows+1:2*nrows+1]: i-structure index
4942: */
4943: /*-------------------------------------------*/
4944: nrows = len_si[proc]/2 - 1;
4945: buf_si_i = buf_si + nrows+1;
4946: buf_si[0] = nrows;
4947: buf_si_i[0] = 0;
4948: nrows = 0;
4949: for (i=owners[proc]; i<owners[proc+1]; i++) {
4950: anzi = ai[i+1] - ai[i];
4951: if (anzi) {
4952: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4953: buf_si[nrows+1] = i-owners[proc]; /* local row index */
4954: nrows++;
4955: }
4956: }
4957: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4958: k++;
4959: buf_si += len_si[proc];
4960: }
4962: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4963: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
4965: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4966: for (i=0; i<merge->nrecv; i++) {
4967: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4968: }
4970: PetscFree(len_si);
4971: PetscFree(len_ri);
4972: PetscFree(rj_waits);
4973: PetscFree2(si_waits,sj_waits);
4974: PetscFree(ri_waits);
4975: PetscFree(buf_s);
4976: PetscFree(status);
4978: /* compute a local seq matrix in each processor */
4979: /*----------------------------------------------*/
4980: /* allocate bi array and free space for accumulating nonzero column info */
4981: PetscMalloc1(m+1,&bi);
4982: bi[0] = 0;
4984: /* create and initialize a linked list */
4985: nlnk = N+1;
4986: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
4988: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4989: len = ai[owners[rank+1]] - ai[owners[rank]];
4990: PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);
4992: current_space = free_space;
4994: /* determine symbolic info for each local row */
4995: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4997: for (k=0; k<merge->nrecv; k++) {
4998: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4999: nrows = *buf_ri_k[k];
5000: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
5001: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
5002: }
5004: MatPreallocateInitialize(comm,m,n,dnz,onz);
5005: len = 0;
5006: for (i=0; i<m; i++) {
5007: bnzi = 0;
5008: /* add local non-zero cols of this proc's seqmat into lnk */
5009: arow = owners[rank] + i;
5010: anzi = ai[arow+1] - ai[arow];
5011: aj = a->j + ai[arow];
5012: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
5013: bnzi += nlnk;
5014: /* add received col data into lnk */
5015: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
5016: if (i == *nextrow[k]) { /* i-th row */
5017: anzi = *(nextai[k]+1) - *nextai[k];
5018: aj = buf_rj[k] + *nextai[k];
5019: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
5020: bnzi += nlnk;
5021: nextrow[k]++; nextai[k]++;
5022: }
5023: }
5024: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5026: /* if free space is not available, make more free space */
5027: if (current_space->local_remaining<bnzi) {
5028: PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);
5029: nspacedouble++;
5030: }
5031: /* copy data into free space, then initialize lnk */
5032: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
5033: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
5035: current_space->array += bnzi;
5036: current_space->local_used += bnzi;
5037: current_space->local_remaining -= bnzi;
5039: bi[i+1] = bi[i] + bnzi;
5040: }
5042: PetscFree3(buf_ri_k,nextrow,nextai);
5044: PetscMalloc1(bi[m]+1,&bj);
5045: PetscFreeSpaceContiguous(&free_space,bj);
5046: PetscLLDestroy(lnk,lnkbt);
5048: /* create symbolic parallel matrix B_mpi */
5049: /*---------------------------------------*/
5050: MatGetBlockSizes(seqmat,&bs,&cbs);
5051: MatCreate(comm,&B_mpi);
5052: if (n==PETSC_DECIDE) {
5053: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
5054: } else {
5055: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5056: }
5057: MatSetBlockSizes(B_mpi,bs,cbs);
5058: MatSetType(B_mpi,MATMPIAIJ);
5059: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
5060: MatPreallocateFinalize(dnz,onz);
5061: MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
5063: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5064: B_mpi->assembled = PETSC_FALSE;
5065: B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
5066: merge->bi = bi;
5067: merge->bj = bj;
5068: merge->buf_ri = buf_ri;
5069: merge->buf_rj = buf_rj;
5070: merge->coi = NULL;
5071: merge->coj = NULL;
5072: merge->owners_co = NULL;
5074: PetscCommDestroy(&comm);
5076: /* attach the supporting struct to B_mpi for reuse */
5077: PetscContainerCreate(PETSC_COMM_SELF,&container);
5078: PetscContainerSetPointer(container,merge);
5079: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
5080: PetscContainerDestroy(&container);
5081: *mpimat = B_mpi;
5083: PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
5084: return(0);
5085: }
5087: /*@C
5088: MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential
5089: matrices from each processor
5091: Collective
5093: Input Parameters:
5094: + comm - the communicators the parallel matrix will live on
5095: . seqmat - the input sequential matrices
5096: . m - number of local rows (or PETSC_DECIDE)
5097: . n - number of local columns (or PETSC_DECIDE)
5098: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5100: Output Parameter:
5101: . mpimat - the parallel matrix generated
5103: Level: advanced
5105: Notes:
5106: The dimensions of the sequential matrix in each processor MUST be the same.
5107: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5108: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
5109: @*/
5110: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
5111: {
5113: PetscMPIInt size;
5116: MPI_Comm_size(comm,&size);
5117: if (size == 1) {
5118: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
5119: if (scall == MAT_INITIAL_MATRIX) {
5120: MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
5121: } else {
5122: MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
5123: }
5124: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
5125: return(0);
5126: }
5127: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
5128: if (scall == MAT_INITIAL_MATRIX) {
5129: MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
5130: }
5131: MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
5132: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
5133: return(0);
5134: }
5136: /*@
5137: MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with
5138: mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
5139: with MatGetSize()
5141: Not Collective
5143: Input Parameters:
5144: + A - the matrix
5145: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5147: Output Parameter:
5148: . A_loc - the local sequential matrix generated
5150: Level: developer
5152: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed()
5154: @*/
5155: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
5156: {
5158: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
5159: Mat_SeqAIJ *mat,*a,*b;
5160: PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
5161: MatScalar *aa,*ba,*cam;
5162: PetscScalar *ca;
5163: PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
5164: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
5165: PetscBool match;
5166: MPI_Comm comm;
5167: PetscMPIInt size;
5170: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);
5171: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5172: PetscObjectGetComm((PetscObject)A,&comm);
5173: MPI_Comm_size(comm,&size);
5174: if (size == 1 && scall == MAT_REUSE_MATRIX) return(0);
5176: PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
5177: a = (Mat_SeqAIJ*)(mpimat->A)->data;
5178: b = (Mat_SeqAIJ*)(mpimat->B)->data;
5179: ai = a->i; aj = a->j; bi = b->i; bj = b->j;
5180: aa = a->a; ba = b->a;
5181: if (scall == MAT_INITIAL_MATRIX) {
5182: if (size == 1) {
5183: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);
5184: return(0);
5185: }
5187: PetscMalloc1(1+am,&ci);
5188: ci[0] = 0;
5189: for (i=0; i<am; i++) {
5190: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
5191: }
5192: PetscMalloc1(1+ci[am],&cj);
5193: PetscMalloc1(1+ci[am],&ca);
5194: k = 0;
5195: for (i=0; i<am; i++) {
5196: ncols_o = bi[i+1] - bi[i];
5197: ncols_d = ai[i+1] - ai[i];
5198: /* off-diagonal portion of A */
5199: for (jo=0; jo<ncols_o; jo++) {
5200: col = cmap[*bj];
5201: if (col >= cstart) break;
5202: cj[k] = col; bj++;
5203: ca[k++] = *ba++;
5204: }
5205: /* diagonal portion of A */
5206: for (j=0; j<ncols_d; j++) {
5207: cj[k] = cstart + *aj++;
5208: ca[k++] = *aa++;
5209: }
5210: /* off-diagonal portion of A */
5211: for (j=jo; j<ncols_o; j++) {
5212: cj[k] = cmap[*bj++];
5213: ca[k++] = *ba++;
5214: }
5215: }
5216: /* put together the new matrix */
5217: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
5218: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5219: /* Since these are PETSc arrays, change flags to free them as necessary. */
5220: mat = (Mat_SeqAIJ*)(*A_loc)->data;
5221: mat->free_a = PETSC_TRUE;
5222: mat->free_ij = PETSC_TRUE;
5223: mat->nonew = 0;
5224: } else if (scall == MAT_REUSE_MATRIX) {
5225: mat=(Mat_SeqAIJ*)(*A_loc)->data;
5226: ci = mat->i; cj = mat->j; cam = mat->a;
5227: for (i=0; i<am; i++) {
5228: /* off-diagonal portion of A */
5229: ncols_o = bi[i+1] - bi[i];
5230: for (jo=0; jo<ncols_o; jo++) {
5231: col = cmap[*bj];
5232: if (col >= cstart) break;
5233: *cam++ = *ba++; bj++;
5234: }
5235: /* diagonal portion of A */
5236: ncols_d = ai[i+1] - ai[i];
5237: for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5238: /* off-diagonal portion of A */
5239: for (j=jo; j<ncols_o; j++) {
5240: *cam++ = *ba++; bj++;
5241: }
5242: }
5243: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5244: PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5245: return(0);
5246: }
5248: /*@C
5249: MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns
5251: Not Collective
5253: Input Parameters:
5254: + A - the matrix
5255: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5256: - row, col - index sets of rows and columns to extract (or NULL)
5258: Output Parameter:
5259: . A_loc - the local sequential matrix generated
5261: Level: developer
5263: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()
5265: @*/
5266: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5267: {
5268: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5270: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5271: IS isrowa,iscola;
5272: Mat *aloc;
5273: PetscBool match;
5276: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5277: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5278: PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5279: if (!row) {
5280: start = A->rmap->rstart; end = A->rmap->rend;
5281: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5282: } else {
5283: isrowa = *row;
5284: }
5285: if (!col) {
5286: start = A->cmap->rstart;
5287: cmap = a->garray;
5288: nzA = a->A->cmap->n;
5289: nzB = a->B->cmap->n;
5290: PetscMalloc1(nzA+nzB, &idx);
5291: ncols = 0;
5292: for (i=0; i<nzB; i++) {
5293: if (cmap[i] < start) idx[ncols++] = cmap[i];
5294: else break;
5295: }
5296: imark = i;
5297: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5298: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5299: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5300: } else {
5301: iscola = *col;
5302: }
5303: if (scall != MAT_INITIAL_MATRIX) {
5304: PetscMalloc1(1,&aloc);
5305: aloc[0] = *A_loc;
5306: }
5307: MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5308: if (!col) { /* attach global id of condensed columns */
5309: PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);
5310: }
5311: *A_loc = aloc[0];
5312: PetscFree(aloc);
5313: if (!row) {
5314: ISDestroy(&isrowa);
5315: }
5316: if (!col) {
5317: ISDestroy(&iscola);
5318: }
5319: PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5320: return(0);
5321: }
5323: /*
5324: * Destroy a mat that may be composed with PetscSF communication objects.
5325: * The SF objects were created in MatCreateSeqSubMatrixWithRows_Private.
5326: * */
5327: PetscErrorCode MatDestroy_SeqAIJ_PetscSF(Mat mat)
5328: {
5329: PetscSF sf,osf;
5330: IS map;
5331: PetscErrorCode ierr;
5334: PetscObjectQuery((PetscObject)mat,"diagsf",(PetscObject*)&sf);
5335: PetscObjectQuery((PetscObject)mat,"offdiagsf",(PetscObject*)&osf);
5336: PetscSFDestroy(&sf);
5337: PetscSFDestroy(&osf);
5338: PetscObjectQuery((PetscObject)mat,"aoffdiagtopothmapping",(PetscObject*)&map);
5339: ISDestroy(&map);
5340: MatDestroy_SeqAIJ(mat);
5341: return(0);
5342: }
5344: /*
5345: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5346: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5347: * on a global size.
5348: * */
5349: PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P,IS rows,Mat *P_oth)
5350: {
5351: Mat_MPIAIJ *p=(Mat_MPIAIJ*)P->data;
5352: Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data,*p_oth;
5353: PetscInt plocalsize,nrows,*ilocal,*oilocal,i,owner,lidx,*nrcols,*nlcols,ncol;
5354: PetscSFNode *iremote,*oiremote;
5355: const PetscInt *lrowindices;
5356: PetscErrorCode ierr;
5357: PetscSF sf,osf;
5358: PetscInt pcstart,*roffsets,*loffsets,*pnnz,j;
5359: PetscInt ontotalcols,dntotalcols,ntotalcols,nout;
5360: MPI_Comm comm;
5361: ISLocalToGlobalMapping mapping;
5364: PetscObjectGetComm((PetscObject)P,&comm);
5365: /* plocalsize is the number of roots
5366: * nrows is the number of leaves
5367: * */
5368: MatGetLocalSize(P,&plocalsize,NULL);
5369: ISGetLocalSize(rows,&nrows);
5370: PetscCalloc1(nrows,&iremote);
5371: ISGetIndices(rows,&lrowindices);
5372: for (i=0;i<nrows;i++) {
5373: /* Find a remote index and an owner for a row
5374: * The row could be local or remote
5375: * */
5376: owner = 0;
5377: lidx = 0;
5378: PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,&lidx);
5379: iremote[i].index = lidx;
5380: iremote[i].rank = owner;
5381: }
5382: /* Create SF to communicate how many nonzero columns for each row */
5383: PetscSFCreate(comm,&sf);
5384: /* SF will figure out the number of nonzero colunms for each row, and their
5385: * offsets
5386: * */
5387: PetscSFSetGraph(sf,plocalsize,nrows,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5388: PetscSFSetFromOptions(sf);
5389: PetscSFSetUp(sf);
5391: PetscCalloc1(2*(plocalsize+1),&roffsets);
5392: PetscCalloc1(2*plocalsize,&nrcols);
5393: PetscCalloc1(nrows,&pnnz);
5394: roffsets[0] = 0;
5395: roffsets[1] = 0;
5396: for (i=0;i<plocalsize;i++) {
5397: /* diag */
5398: nrcols[i*2+0] = pd->i[i+1] - pd->i[i];
5399: /* off diag */
5400: nrcols[i*2+1] = po->i[i+1] - po->i[i];
5401: /* compute offsets so that we relative location for each row */
5402: roffsets[(i+1)*2+0] = roffsets[i*2+0] + nrcols[i*2+0];
5403: roffsets[(i+1)*2+1] = roffsets[i*2+1] + nrcols[i*2+1];
5404: }
5405: PetscCalloc1(2*nrows,&nlcols);
5406: PetscCalloc1(2*nrows,&loffsets);
5407: /* 'r' means root, and 'l' means leaf */
5408: PetscSFBcastBegin(sf,MPIU_2INT,nrcols,nlcols);
5409: PetscSFBcastBegin(sf,MPIU_2INT,roffsets,loffsets);
5410: PetscSFBcastEnd(sf,MPIU_2INT,nrcols,nlcols);
5411: PetscSFBcastEnd(sf,MPIU_2INT,roffsets,loffsets);
5412: PetscSFDestroy(&sf);
5413: PetscFree(roffsets);
5414: PetscFree(nrcols);
5415: dntotalcols = 0;
5416: ontotalcols = 0;
5417: ncol = 0;
5418: for (i=0;i<nrows;i++) {
5419: pnnz[i] = nlcols[i*2+0] + nlcols[i*2+1];
5420: ncol = PetscMax(pnnz[i],ncol);
5421: /* diag */
5422: dntotalcols += nlcols[i*2+0];
5423: /* off diag */
5424: ontotalcols += nlcols[i*2+1];
5425: }
5426: /* We do not need to figure the right number of columns
5427: * since all the calculations will be done by going through the raw data
5428: * */
5429: MatCreateSeqAIJ(PETSC_COMM_SELF,nrows,ncol,0,pnnz,P_oth);
5430: MatSetUp(*P_oth);
5431: PetscFree(pnnz);
5432: p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5433: /* diag */
5434: PetscCalloc1(dntotalcols,&iremote);
5435: /* off diag */
5436: PetscCalloc1(ontotalcols,&oiremote);
5437: /* diag */
5438: PetscCalloc1(dntotalcols,&ilocal);
5439: /* off diag */
5440: PetscCalloc1(ontotalcols,&oilocal);
5441: dntotalcols = 0;
5442: ontotalcols = 0;
5443: ntotalcols = 0;
5444: for (i=0;i<nrows;i++) {
5445: owner = 0;
5446: PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,NULL);
5447: /* Set iremote for diag matrix */
5448: for (j=0;j<nlcols[i*2+0];j++) {
5449: iremote[dntotalcols].index = loffsets[i*2+0] + j;
5450: iremote[dntotalcols].rank = owner;
5451: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5452: ilocal[dntotalcols++] = ntotalcols++;
5453: }
5454: /* off diag */
5455: for (j=0;j<nlcols[i*2+1];j++) {
5456: oiremote[ontotalcols].index = loffsets[i*2+1] + j;
5457: oiremote[ontotalcols].rank = owner;
5458: oilocal[ontotalcols++] = ntotalcols++;
5459: }
5460: }
5461: ISRestoreIndices(rows,&lrowindices);
5462: PetscFree(loffsets);
5463: PetscFree(nlcols);
5464: PetscSFCreate(comm,&sf);
5465: /* P serves as roots and P_oth is leaves
5466: * Diag matrix
5467: * */
5468: PetscSFSetGraph(sf,pd->i[plocalsize],dntotalcols,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5469: PetscSFSetFromOptions(sf);
5470: PetscSFSetUp(sf);
5472: PetscSFCreate(comm,&osf);
5473: /* Off diag */
5474: PetscSFSetGraph(osf,po->i[plocalsize],ontotalcols,oilocal,PETSC_OWN_POINTER,oiremote,PETSC_OWN_POINTER);
5475: PetscSFSetFromOptions(osf);
5476: PetscSFSetUp(osf);
5477: /* We operate on the matrix internal data for saving memory */
5478: PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5479: PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5480: MatGetOwnershipRangeColumn(P,&pcstart,NULL);
5481: /* Convert to global indices for diag matrix */
5482: for (i=0;i<pd->i[plocalsize];i++) pd->j[i] += pcstart;
5483: PetscSFBcastBegin(sf,MPIU_INT,pd->j,p_oth->j);
5484: /* We want P_oth store global indices */
5485: ISLocalToGlobalMappingCreate(comm,1,p->B->cmap->n,p->garray,PETSC_COPY_VALUES,&mapping);
5486: /* Use memory scalable approach */
5487: ISLocalToGlobalMappingSetType(mapping,ISLOCALTOGLOBALMAPPINGHASH);
5488: ISLocalToGlobalMappingApply(mapping,po->i[plocalsize],po->j,po->j);
5489: PetscSFBcastBegin(osf,MPIU_INT,po->j,p_oth->j);
5490: PetscSFBcastEnd(sf,MPIU_INT,pd->j,p_oth->j);
5491: /* Convert back to local indices */
5492: for (i=0;i<pd->i[plocalsize];i++) pd->j[i] -= pcstart;
5493: PetscSFBcastEnd(osf,MPIU_INT,po->j,p_oth->j);
5494: nout = 0;
5495: ISGlobalToLocalMappingApply(mapping,IS_GTOLM_DROP,po->i[plocalsize],po->j,&nout,po->j);
5496: if (nout != po->i[plocalsize]) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP,"n %D does not equal to nout %D \n",po->i[plocalsize],nout);
5497: ISLocalToGlobalMappingDestroy(&mapping);
5498: /* Exchange values */
5499: PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5500: PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5501: /* Stop PETSc from shrinking memory */
5502: for (i=0;i<nrows;i++) p_oth->ilen[i] = p_oth->imax[i];
5503: MatAssemblyBegin(*P_oth,MAT_FINAL_ASSEMBLY);
5504: MatAssemblyEnd(*P_oth,MAT_FINAL_ASSEMBLY);
5505: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5506: PetscObjectCompose((PetscObject)*P_oth,"diagsf",(PetscObject)sf);
5507: PetscObjectCompose((PetscObject)*P_oth,"offdiagsf",(PetscObject)osf);
5508: /* ``New MatDestroy" takes care of PetscSF objects as well */
5509: (*P_oth)->ops->destroy = MatDestroy_SeqAIJ_PetscSF;
5510: return(0);
5511: }
5513: /*
5514: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5515: * This supports MPIAIJ and MAIJ
5516: * */
5517: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A,Mat P,PetscInt dof,MatReuse reuse,Mat *P_oth)
5518: {
5519: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data;
5520: Mat_SeqAIJ *p_oth;
5521: Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data;
5522: IS rows,map;
5523: PetscHMapI hamp;
5524: PetscInt i,htsize,*rowindices,off,*mapping,key,count;
5525: MPI_Comm comm;
5526: PetscSF sf,osf;
5527: PetscBool has;
5528: PetscErrorCode ierr;
5531: PetscObjectGetComm((PetscObject)A,&comm);
5532: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,P,0,0);
5533: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5534: * and then create a submatrix (that often is an overlapping matrix)
5535: * */
5536: if (reuse==MAT_INITIAL_MATRIX) {
5537: /* Use a hash table to figure out unique keys */
5538: PetscHMapICreate(&hamp);
5539: PetscHMapIResize(hamp,a->B->cmap->n);
5540: PetscCalloc1(a->B->cmap->n,&mapping);
5541: count = 0;
5542: /* Assume that a->g is sorted, otherwise the following does not make sense */
5543: for (i=0;i<a->B->cmap->n;i++) {
5544: key = a->garray[i]/dof;
5545: PetscHMapIHas(hamp,key,&has);
5546: if (!has) {
5547: mapping[i] = count;
5548: PetscHMapISet(hamp,key,count++);
5549: } else {
5550: /* Current 'i' has the same value the previous step */
5551: mapping[i] = count-1;
5552: }
5553: }
5554: ISCreateGeneral(comm,a->B->cmap->n,mapping,PETSC_OWN_POINTER,&map);
5555: PetscHMapIGetSize(hamp,&htsize);
5556: if (htsize!=count) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP," Size of hash map %D is inconsistent with count %D \n",htsize,count);
5557: PetscCalloc1(htsize,&rowindices);
5558: off = 0;
5559: PetscHMapIGetKeys(hamp,&off,rowindices);
5560: PetscHMapIDestroy(&hamp);
5561: PetscSortInt(htsize,rowindices);
5562: ISCreateGeneral(comm,htsize,rowindices,PETSC_OWN_POINTER,&rows);
5563: /* In case, the matrix was already created but users want to recreate the matrix */
5564: MatDestroy(P_oth);
5565: MatCreateSeqSubMatrixWithRows_Private(P,rows,P_oth);
5566: PetscObjectCompose((PetscObject)*P_oth,"aoffdiagtopothmapping",(PetscObject)map);
5567: ISDestroy(&rows);
5568: } else if (reuse==MAT_REUSE_MATRIX) {
5569: /* If matrix was already created, we simply update values using SF objects
5570: * that as attached to the matrix ealier.
5571: * */
5572: PetscObjectQuery((PetscObject)*P_oth,"diagsf",(PetscObject*)&sf);
5573: PetscObjectQuery((PetscObject)*P_oth,"offdiagsf",(PetscObject*)&osf);
5574: if (!sf || !osf) {
5575: SETERRQ(comm,PETSC_ERR_ARG_NULL,"Matrix is not initialized yet \n");
5576: }
5577: p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5578: /* Update values in place */
5579: PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5580: PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5581: PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5582: PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5583: } else {
5584: SETERRQ(comm,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown reuse type \n");
5585: }
5586: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,P,0,0);
5587: return(0);
5588: }
5590: /*@C
5591: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5593: Collective on Mat
5595: Input Parameters:
5596: + A,B - the matrices in mpiaij format
5597: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5598: - rowb, colb - index sets of rows and columns of B to extract (or NULL)
5600: Output Parameter:
5601: + rowb, colb - index sets of rows and columns of B to extract
5602: - B_seq - the sequential matrix generated
5604: Level: developer
5606: @*/
5607: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5608: {
5609: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5611: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5612: IS isrowb,iscolb;
5613: Mat *bseq=NULL;
5616: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5617: 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);
5618: }
5619: PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5621: if (scall == MAT_INITIAL_MATRIX) {
5622: start = A->cmap->rstart;
5623: cmap = a->garray;
5624: nzA = a->A->cmap->n;
5625: nzB = a->B->cmap->n;
5626: PetscMalloc1(nzA+nzB, &idx);
5627: ncols = 0;
5628: for (i=0; i<nzB; i++) { /* row < local row index */
5629: if (cmap[i] < start) idx[ncols++] = cmap[i];
5630: else break;
5631: }
5632: imark = i;
5633: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
5634: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5635: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5636: ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5637: } else {
5638: if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5639: isrowb = *rowb; iscolb = *colb;
5640: PetscMalloc1(1,&bseq);
5641: bseq[0] = *B_seq;
5642: }
5643: MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5644: *B_seq = bseq[0];
5645: PetscFree(bseq);
5646: if (!rowb) {
5647: ISDestroy(&isrowb);
5648: } else {
5649: *rowb = isrowb;
5650: }
5651: if (!colb) {
5652: ISDestroy(&iscolb);
5653: } else {
5654: *colb = iscolb;
5655: }
5656: PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5657: return(0);
5658: }
5660: /*
5661: MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5662: of the OFF-DIAGONAL portion of local A
5664: Collective on Mat
5666: Input Parameters:
5667: + A,B - the matrices in mpiaij format
5668: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5670: Output Parameter:
5671: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5672: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5673: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5674: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5676: Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product
5677: for this matrix. This is not desirable..
5679: Level: developer
5681: */
5682: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5683: {
5684: PetscErrorCode ierr;
5685: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5686: Mat_SeqAIJ *b_oth;
5687: VecScatter ctx;
5688: MPI_Comm comm;
5689: const PetscMPIInt *rprocs,*sprocs;
5690: const PetscInt *srow,*rstarts,*sstarts;
5691: PetscInt *rowlen,*bufj,*bufJ,ncols = 0,aBn=a->B->cmap->n,row,*b_othi,*b_othj,*rvalues=NULL,*svalues=NULL,*cols,sbs,rbs;
5692: PetscInt i,j,k=0,l,ll,nrecvs,nsends,nrows,*rstartsj = 0,*sstartsj,len;
5693: PetscScalar *b_otha,*bufa,*bufA,*vals = NULL;
5694: MPI_Request *rwaits = NULL,*swaits = NULL;
5695: MPI_Status rstatus;
5696: PetscMPIInt jj,size,tag,rank,nsends_mpi,nrecvs_mpi;
5699: PetscObjectGetComm((PetscObject)A,&comm);
5700: MPI_Comm_size(comm,&size);
5702: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5703: 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);
5704: }
5705: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5706: MPI_Comm_rank(comm,&rank);
5708: if (size == 1) {
5709: startsj_s = NULL;
5710: bufa_ptr = NULL;
5711: *B_oth = NULL;
5712: return(0);
5713: }
5715: ctx = a->Mvctx;
5716: tag = ((PetscObject)ctx)->tag;
5718: if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use");
5719: VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&srow,&sprocs,&sbs);
5720: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5721: VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL/*indices not needed*/,&rprocs,&rbs);
5722: PetscMPIIntCast(nsends,&nsends_mpi);
5723: PetscMPIIntCast(nrecvs,&nrecvs_mpi);
5724: PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);
5726: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5727: if (scall == MAT_INITIAL_MATRIX) {
5728: /* i-array */
5729: /*---------*/
5730: /* post receives */
5731: if (nrecvs) {PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);} /* rstarts can be NULL when nrecvs=0 */
5732: for (i=0; i<nrecvs; i++) {
5733: rowlen = rvalues + rstarts[i]*rbs;
5734: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5735: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5736: }
5738: /* pack the outgoing message */
5739: PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);
5741: sstartsj[0] = 0;
5742: rstartsj[0] = 0;
5743: len = 0; /* total length of j or a array to be sent */
5744: if (nsends) {
5745: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5746: PetscMalloc1(sbs*(sstarts[nsends]-sstarts[0]),&svalues);
5747: }
5748: for (i=0; i<nsends; i++) {
5749: rowlen = svalues + (sstarts[i]-sstarts[0])*sbs;
5750: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5751: for (j=0; j<nrows; j++) {
5752: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5753: for (l=0; l<sbs; l++) {
5754: MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */
5756: rowlen[j*sbs+l] = ncols;
5758: len += ncols;
5759: MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
5760: }
5761: k++;
5762: }
5763: MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5765: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5766: }
5767: /* recvs and sends of i-array are completed */
5768: i = nrecvs;
5769: while (i--) {
5770: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5771: }
5772: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5773: PetscFree(svalues);
5775: /* allocate buffers for sending j and a arrays */
5776: PetscMalloc1(len+1,&bufj);
5777: PetscMalloc1(len+1,&bufa);
5779: /* create i-array of B_oth */
5780: PetscMalloc1(aBn+2,&b_othi);
5782: b_othi[0] = 0;
5783: len = 0; /* total length of j or a array to be received */
5784: k = 0;
5785: for (i=0; i<nrecvs; i++) {
5786: rowlen = rvalues + (rstarts[i]-rstarts[0])*rbs;
5787: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of rows to be received */
5788: for (j=0; j<nrows; j++) {
5789: b_othi[k+1] = b_othi[k] + rowlen[j];
5790: PetscIntSumError(rowlen[j],len,&len);
5791: k++;
5792: }
5793: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5794: }
5795: PetscFree(rvalues);
5797: /* allocate space for j and a arrrays of B_oth */
5798: PetscMalloc1(b_othi[aBn]+1,&b_othj);
5799: PetscMalloc1(b_othi[aBn]+1,&b_otha);
5801: /* j-array */
5802: /*---------*/
5803: /* post receives of j-array */
5804: for (i=0; i<nrecvs; i++) {
5805: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5806: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5807: }
5809: /* pack the outgoing message j-array */
5810: if (nsends) k = sstarts[0];
5811: for (i=0; i<nsends; i++) {
5812: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5813: bufJ = bufj+sstartsj[i];
5814: for (j=0; j<nrows; j++) {
5815: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5816: for (ll=0; ll<sbs; ll++) {
5817: MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5818: for (l=0; l<ncols; l++) {
5819: *bufJ++ = cols[l];
5820: }
5821: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5822: }
5823: }
5824: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5825: }
5827: /* recvs and sends of j-array are completed */
5828: i = nrecvs;
5829: while (i--) {
5830: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5831: }
5832: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5833: } else if (scall == MAT_REUSE_MATRIX) {
5834: sstartsj = *startsj_s;
5835: rstartsj = *startsj_r;
5836: bufa = *bufa_ptr;
5837: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5838: b_otha = b_oth->a;
5839: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5841: /* a-array */
5842: /*---------*/
5843: /* post receives of a-array */
5844: for (i=0; i<nrecvs; i++) {
5845: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5846: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5847: }
5849: /* pack the outgoing message a-array */
5850: if (nsends) k = sstarts[0];
5851: for (i=0; i<nsends; i++) {
5852: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5853: bufA = bufa+sstartsj[i];
5854: for (j=0; j<nrows; j++) {
5855: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5856: for (ll=0; ll<sbs; ll++) {
5857: MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5858: for (l=0; l<ncols; l++) {
5859: *bufA++ = vals[l];
5860: }
5861: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5862: }
5863: }
5864: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5865: }
5866: /* recvs and sends of a-array are completed */
5867: i = nrecvs;
5868: while (i--) {
5869: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5870: }
5871: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5872: PetscFree2(rwaits,swaits);
5874: if (scall == MAT_INITIAL_MATRIX) {
5875: /* put together the new matrix */
5876: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);
5878: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5879: /* Since these are PETSc arrays, change flags to free them as necessary. */
5880: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5881: b_oth->free_a = PETSC_TRUE;
5882: b_oth->free_ij = PETSC_TRUE;
5883: b_oth->nonew = 0;
5885: PetscFree(bufj);
5886: if (!startsj_s || !bufa_ptr) {
5887: PetscFree2(sstartsj,rstartsj);
5888: PetscFree(bufa_ptr);
5889: } else {
5890: *startsj_s = sstartsj;
5891: *startsj_r = rstartsj;
5892: *bufa_ptr = bufa;
5893: }
5894: }
5896: VecScatterRestoreRemote_Private(ctx,PETSC_TRUE,&nsends,&sstarts,&srow,&sprocs,&sbs);
5897: VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE,&nrecvs,&rstarts,NULL,&rprocs,&rbs);
5898: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5899: return(0);
5900: }
5902: /*@C
5903: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
5905: Not Collective
5907: Input Parameters:
5908: . A - The matrix in mpiaij format
5910: Output Parameter:
5911: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5912: . colmap - A map from global column index to local index into lvec
5913: - multScatter - A scatter from the argument of a matrix-vector product to lvec
5915: Level: developer
5917: @*/
5918: #if defined(PETSC_USE_CTABLE)
5919: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5920: #else
5921: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5922: #endif
5923: {
5924: Mat_MPIAIJ *a;
5931: a = (Mat_MPIAIJ*) A->data;
5932: if (lvec) *lvec = a->lvec;
5933: if (colmap) *colmap = a->colmap;
5934: if (multScatter) *multScatter = a->Mvctx;
5935: return(0);
5936: }
5938: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
5939: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
5940: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*);
5941: #if defined(PETSC_HAVE_MKL_SPARSE)
5942: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*);
5943: #endif
5944: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
5945: #if defined(PETSC_HAVE_ELEMENTAL)
5946: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
5947: #endif
5948: #if defined(PETSC_HAVE_HYPRE)
5949: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
5950: PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
5951: #endif
5952: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
5953: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*);
5954: PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
5956: /*
5957: Computes (B'*A')' since computing B*A directly is untenable
5959: n p p
5960: ( ) ( ) ( )
5961: m ( A ) * n ( B ) = m ( C )
5962: ( ) ( ) ( )
5964: */
5965: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5966: {
5968: Mat At,Bt,Ct;
5971: MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5972: MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5973: MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5974: MatDestroy(&At);
5975: MatDestroy(&Bt);
5976: MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5977: MatDestroy(&Ct);
5978: return(0);
5979: }
5981: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
5982: {
5984: PetscInt m=A->rmap->n,n=B->cmap->n;
5985: Mat Cmat;
5988: 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);
5989: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
5990: MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5991: MatSetBlockSizesFromMats(Cmat,A,B);
5992: MatSetType(Cmat,MATMPIDENSE);
5993: MatMPIDenseSetPreallocation(Cmat,NULL);
5994: MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5995: MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
5997: Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
5999: *C = Cmat;
6000: return(0);
6001: }
6003: /* ----------------------------------------------------------------*/
6004: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
6005: {
6009: if (scall == MAT_INITIAL_MATRIX) {
6010: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
6011: MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
6012: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
6013: }
6014: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
6015: MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
6016: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
6017: return(0);
6018: }
6020: /*MC
6021: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6023: Options Database Keys:
6024: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
6026: Level: beginner
6028: Notes:
6029: MatSetValues() may be called for this matrix type with a NULL argument for the numerical values,
6030: in this case the values associated with the rows and columns one passes in are set to zero
6031: in the matrix
6033: MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
6034: space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored
6036: .seealso: MatCreateAIJ()
6037: M*/
6039: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6040: {
6041: Mat_MPIAIJ *b;
6043: PetscMPIInt size;
6046: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
6048: PetscNewLog(B,&b);
6049: B->data = (void*)b;
6050: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
6051: B->assembled = PETSC_FALSE;
6052: B->insertmode = NOT_SET_VALUES;
6053: b->size = size;
6055: MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
6057: /* build cache for off array entries formed */
6058: MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);
6060: b->donotstash = PETSC_FALSE;
6061: b->colmap = 0;
6062: b->garray = 0;
6063: b->roworiented = PETSC_TRUE;
6065: /* stuff used for matrix vector multiply */
6066: b->lvec = NULL;
6067: b->Mvctx = NULL;
6069: /* stuff for MatGetRow() */
6070: b->rowindices = 0;
6071: b->rowvalues = 0;
6072: b->getrowactive = PETSC_FALSE;
6074: /* flexible pointer used in CUSP/CUSPARSE classes */
6075: b->spptr = NULL;
6077: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
6078: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
6079: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
6080: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
6081: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
6082: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);
6083: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
6084: PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
6085: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
6086: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);
6087: #if defined(PETSC_HAVE_MKL_SPARSE)
6088: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);
6089: #endif
6090: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
6091: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
6092: #if defined(PETSC_HAVE_ELEMENTAL)
6093: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
6094: #endif
6095: #if defined(PETSC_HAVE_HYPRE)
6096: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);
6097: #endif
6098: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);
6099: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);
6100: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);
6101: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);
6102: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);
6103: #if defined(PETSC_HAVE_HYPRE)
6104: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
6105: #endif
6106: PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_mpiaij_C",MatPtAP_IS_XAIJ);
6107: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
6108: return(0);
6109: }
6111: /*@C
6112: MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
6113: and "off-diagonal" part of the matrix in CSR format.
6115: Collective
6117: Input Parameters:
6118: + comm - MPI communicator
6119: . m - number of local rows (Cannot be PETSC_DECIDE)
6120: . n - This value should be the same as the local size used in creating the
6121: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
6122: calculated if N is given) For square matrices n is almost always m.
6123: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
6124: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
6125: . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
6126: . j - column indices
6127: . a - matrix values
6128: . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix
6129: . oj - column indices
6130: - oa - matrix values
6132: Output Parameter:
6133: . mat - the matrix
6135: Level: advanced
6137: Notes:
6138: The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6139: must free the arrays once the matrix has been destroyed and not before.
6141: The i and j indices are 0 based
6143: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix
6145: This sets local rows and cannot be used to set off-processor values.
6147: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6148: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6149: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6150: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6151: keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
6152: communication if it is known that only local entries will be set.
6154: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
6155: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
6156: @*/
6157: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
6158: {
6160: Mat_MPIAIJ *maij;
6163: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
6164: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
6165: if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
6166: MatCreate(comm,mat);
6167: MatSetSizes(*mat,m,n,M,N);
6168: MatSetType(*mat,MATMPIAIJ);
6169: maij = (Mat_MPIAIJ*) (*mat)->data;
6171: (*mat)->preallocated = PETSC_TRUE;
6173: PetscLayoutSetUp((*mat)->rmap);
6174: PetscLayoutSetUp((*mat)->cmap);
6176: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
6177: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);
6179: MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
6180: MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
6181: MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
6182: MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);
6184: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
6185: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
6186: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
6187: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
6188: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
6189: return(0);
6190: }
6192: /*
6193: Special version for direct calls from Fortran
6194: */
6195: #include <petsc/private/fortranimpl.h>
6197: /* Change these macros so can be used in void function */
6198: #undef CHKERRQ
6199: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
6200: #undef SETERRQ2
6201: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
6202: #undef SETERRQ3
6203: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
6204: #undef SETERRQ
6205: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)
6207: #if defined(PETSC_HAVE_FORTRAN_CAPS)
6208: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
6209: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
6210: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
6211: #else
6212: #endif
6213: PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
6214: {
6215: Mat mat = *mmat;
6216: PetscInt m = *mm, n = *mn;
6217: InsertMode addv = *maddv;
6218: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
6219: PetscScalar value;
6222: MatCheckPreallocated(mat,1);
6223: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
6225: #if defined(PETSC_USE_DEBUG)
6226: else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
6227: #endif
6228: {
6229: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
6230: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
6231: PetscBool roworiented = aij->roworiented;
6233: /* Some Variables required in the macro */
6234: Mat A = aij->A;
6235: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
6236: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
6237: MatScalar *aa = a->a;
6238: PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
6239: Mat B = aij->B;
6240: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
6241: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
6242: MatScalar *ba = b->a;
6243: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
6244: * cannot use "#if defined" inside a macro. */
6245: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
6247: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
6248: PetscInt nonew = a->nonew;
6249: MatScalar *ap1,*ap2;
6252: for (i=0; i<m; i++) {
6253: if (im[i] < 0) continue;
6254: #if defined(PETSC_USE_DEBUG)
6255: 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);
6256: #endif
6257: if (im[i] >= rstart && im[i] < rend) {
6258: row = im[i] - rstart;
6259: lastcol1 = -1;
6260: rp1 = aj + ai[row];
6261: ap1 = aa + ai[row];
6262: rmax1 = aimax[row];
6263: nrow1 = ailen[row];
6264: low1 = 0;
6265: high1 = nrow1;
6266: lastcol2 = -1;
6267: rp2 = bj + bi[row];
6268: ap2 = ba + bi[row];
6269: rmax2 = bimax[row];
6270: nrow2 = bilen[row];
6271: low2 = 0;
6272: high2 = nrow2;
6274: for (j=0; j<n; j++) {
6275: if (roworiented) value = v[i*n+j];
6276: else value = v[i+j*m];
6277: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
6278: if (in[j] >= cstart && in[j] < cend) {
6279: col = in[j] - cstart;
6280: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
6281: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
6282: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
6283: #endif
6284: } else if (in[j] < 0) continue;
6285: #if defined(PETSC_USE_DEBUG)
6286: /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */
6287: 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);}
6288: #endif
6289: else {
6290: if (mat->was_assembled) {
6291: if (!aij->colmap) {
6292: MatCreateColmap_MPIAIJ_Private(mat);
6293: }
6294: #if defined(PETSC_USE_CTABLE)
6295: PetscTableFind(aij->colmap,in[j]+1,&col);
6296: col--;
6297: #else
6298: col = aij->colmap[in[j]] - 1;
6299: #endif
6300: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
6301: MatDisAssemble_MPIAIJ(mat);
6302: col = in[j];
6303: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
6304: B = aij->B;
6305: b = (Mat_SeqAIJ*)B->data;
6306: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
6307: rp2 = bj + bi[row];
6308: ap2 = ba + bi[row];
6309: rmax2 = bimax[row];
6310: nrow2 = bilen[row];
6311: low2 = 0;
6312: high2 = nrow2;
6313: bm = aij->B->rmap->n;
6314: ba = b->a;
6315: inserted = PETSC_FALSE;
6316: }
6317: } else col = in[j];
6318: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
6319: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
6320: if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
6321: #endif
6322: }
6323: }
6324: } else if (!aij->donotstash) {
6325: if (roworiented) {
6326: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6327: } else {
6328: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6329: }
6330: }
6331: }
6332: }
6333: PetscFunctionReturnVoid();
6334: }