Actual source code: matnest.c
petsc-3.13.0 2020-03-29
1: #include <../src/mat/impls/nest/matnestimpl.h>
2: #include <../src/mat/impls/aij/seq/aij.h>
3: #include <petscsf.h>
5: static PetscErrorCode MatSetUp_NestIS_Private(Mat,PetscInt,const IS[],PetscInt,const IS[]);
6: static PetscErrorCode MatCreateVecs_Nest(Mat,Vec*,Vec*);
7: static PetscErrorCode MatReset_Nest(Mat);
9: PETSC_INTERN PetscErrorCode MatConvert_Nest_IS(Mat,MatType,MatReuse,Mat*);
11: /* private functions */
12: static PetscErrorCode MatNestGetSizes_Private(Mat A,PetscInt *m,PetscInt *n,PetscInt *M,PetscInt *N)
13: {
14: Mat_Nest *bA = (Mat_Nest*)A->data;
15: PetscInt i,j;
19: *m = *n = *M = *N = 0;
20: for (i=0; i<bA->nr; i++) { /* rows */
21: PetscInt sm,sM;
22: ISGetLocalSize(bA->isglobal.row[i],&sm);
23: ISGetSize(bA->isglobal.row[i],&sM);
24: *m += sm;
25: *M += sM;
26: }
27: for (j=0; j<bA->nc; j++) { /* cols */
28: PetscInt sn,sN;
29: ISGetLocalSize(bA->isglobal.col[j],&sn);
30: ISGetSize(bA->isglobal.col[j],&sN);
31: *n += sn;
32: *N += sN;
33: }
34: return(0);
35: }
37: /* operations */
38: static PetscErrorCode MatMult_Nest(Mat A,Vec x,Vec y)
39: {
40: Mat_Nest *bA = (Mat_Nest*)A->data;
41: Vec *bx = bA->right,*by = bA->left;
42: PetscInt i,j,nr = bA->nr,nc = bA->nc;
46: for (i=0; i<nr; i++) {VecGetSubVector(y,bA->isglobal.row[i],&by[i]);}
47: for (i=0; i<nc; i++) {VecGetSubVector(x,bA->isglobal.col[i],&bx[i]);}
48: for (i=0; i<nr; i++) {
49: VecZeroEntries(by[i]);
50: for (j=0; j<nc; j++) {
51: if (!bA->m[i][j]) continue;
52: /* y[i] <- y[i] + A[i][j] * x[j] */
53: MatMultAdd(bA->m[i][j],bx[j],by[i],by[i]);
54: }
55: }
56: for (i=0; i<nr; i++) {VecRestoreSubVector(y,bA->isglobal.row[i],&by[i]);}
57: for (i=0; i<nc; i++) {VecRestoreSubVector(x,bA->isglobal.col[i],&bx[i]);}
58: return(0);
59: }
61: static PetscErrorCode MatMultAdd_Nest(Mat A,Vec x,Vec y,Vec z)
62: {
63: Mat_Nest *bA = (Mat_Nest*)A->data;
64: Vec *bx = bA->right,*bz = bA->left;
65: PetscInt i,j,nr = bA->nr,nc = bA->nc;
69: for (i=0; i<nr; i++) {VecGetSubVector(z,bA->isglobal.row[i],&bz[i]);}
70: for (i=0; i<nc; i++) {VecGetSubVector(x,bA->isglobal.col[i],&bx[i]);}
71: for (i=0; i<nr; i++) {
72: if (y != z) {
73: Vec by;
74: VecGetSubVector(y,bA->isglobal.row[i],&by);
75: VecCopy(by,bz[i]);
76: VecRestoreSubVector(y,bA->isglobal.row[i],&by);
77: }
78: for (j=0; j<nc; j++) {
79: if (!bA->m[i][j]) continue;
80: /* y[i] <- y[i] + A[i][j] * x[j] */
81: MatMultAdd(bA->m[i][j],bx[j],bz[i],bz[i]);
82: }
83: }
84: for (i=0; i<nr; i++) {VecRestoreSubVector(z,bA->isglobal.row[i],&bz[i]);}
85: for (i=0; i<nc; i++) {VecRestoreSubVector(x,bA->isglobal.col[i],&bx[i]);}
86: return(0);
87: }
89: typedef struct {
90: Mat *workC; /* array of Mat with specific containers depending on the underlying MatMatMult implementation */
91: PetscScalar *tarray; /* buffer for storing all temporary products A[i][j] B[j] */
92: PetscInt *dm,*dn,k; /* displacements and number of submatrices */
93: } Nest_Dense;
95: PETSC_INTERN PetscErrorCode MatMatMultNumeric_Nest_Dense(Mat A,Mat B,Mat C)
96: {
97: Mat_Nest *bA = (Mat_Nest*)A->data;
98: PetscContainer container;
99: Nest_Dense *contents;
100: Mat viewB,viewC,seq,productB,workC;
101: const PetscScalar *barray;
102: PetscScalar *carray;
103: PetscInt i,j,M,N,nr = bA->nr,nc = bA->nc,ldb,ldc;
104: PetscErrorCode ierr;
107: PetscObjectQuery((PetscObject)C,"workC",(PetscObject*)&container);
108: if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exist");
109: PetscContainerGetPointer(container,(void**)&contents);
110: MatDenseGetLDA(B,&ldb);
111: MatDenseGetLDA(C,&ldc);
112: MatGetSize(B,NULL,&N);
113: MatZeroEntries(C);
114: MatDenseGetArrayRead(B,&barray);
115: MatDenseGetArray(C,&carray);
116: for (i=0; i<nr; i++) {
117: ISGetSize(bA->isglobal.row[i],&M);
118: MatCreateDense(PetscObjectComm((PetscObject)A),contents->dm[i+1]-contents->dm[i],PETSC_DECIDE,M,N,carray+contents->dm[i],&viewC);
119: MatDenseGetLocalMatrix(viewC,&seq);
120: MatSeqDenseSetLDA(seq,ldc);
121: for (j=0; j<nc; j++) {
122: if (!bA->m[i][j]) continue;
123: ISGetSize(bA->isglobal.col[j],&M);
124: MatCreateDense(PetscObjectComm((PetscObject)A),contents->dn[j+1]-contents->dn[j],PETSC_DECIDE,M,N,(PetscScalar*)(barray+contents->dn[j]),&viewB);
125: MatDenseGetLocalMatrix(viewB,&seq);
126: MatSeqDenseSetLDA(seq,ldb);
128: /* MatMatMultNumeric(bA->m[i][j],viewB,contents->workC[i*nc + j]); */
129: workC = contents->workC[i*nc + j];
130: productB = workC->product->B;
131: workC->product->B = viewB; /* use newly created dense matrix viewB */
132: (workC->ops->productnumeric)(workC);
133: MatDestroy(&viewB);
134: workC->product->B = productB; /* resume original B */
136: /* C[i] <- workC + C[i] */
137: MatAXPY(viewC,1.0,contents->workC[i*nc + j],SAME_NONZERO_PATTERN);
138: }
139: MatDestroy(&viewC);
140: }
141: MatDenseRestoreArray(C,&carray);
142: MatDenseRestoreArrayRead(B,&barray);
144: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
145: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
146: return(0);
147: }
149: PetscErrorCode MatNest_DenseDestroy(void *ctx)
150: {
151: Nest_Dense *contents = (Nest_Dense*)ctx;
152: PetscInt i;
156: PetscFree(contents->tarray);
157: for (i=0; i<contents->k; i++) {
158: MatDestroy(contents->workC + i);
159: }
160: PetscFree3(contents->dm,contents->dn,contents->workC);
161: PetscFree(contents);
162: return(0);
163: }
165: PETSC_INTERN PetscErrorCode MatMatMultSymbolic_Nest_Dense(Mat A,Mat B,PetscReal fill,Mat C)
166: {
167: Mat_Nest *bA = (Mat_Nest*)A->data;
168: Mat viewB,viewSeq,workC;
169: const PetscScalar *barray;
170: PetscInt i,j,M,N,m,nr = bA->nr,nc = bA->nc,maxm = 0,ldb;
171: PetscContainer container;
172: Nest_Dense *contents=NULL;
173: PetscErrorCode ierr;
176: if (!C->assembled) {
177: MatGetSize(B,NULL,&N);
178: MatGetLocalSize(A,&m,NULL);
179: MatGetSize(A,&M,NULL);
181: MatSetSizes(C,m,PETSC_DECIDE,M,N);
182: MatSetType(C,MATDENSE);
183: MatSeqDenseSetPreallocation(C,NULL);
184: MatMPIDenseSetPreallocation(C,NULL);
185: }
187: PetscNew(&contents);
188: PetscContainerCreate(PetscObjectComm((PetscObject)A),&container);
189: PetscContainerSetPointer(container,contents);
190: PetscContainerSetUserDestroy(container,MatNest_DenseDestroy);
191: PetscObjectCompose((PetscObject)C,"workC",(PetscObject)container);
192: PetscContainerDestroy(&container);
193: PetscCalloc3(nr+1,&contents->dm,nc+1,&contents->dn,nr*nc,&contents->workC);
194: contents->k = nr*nc;
195: for (i=0; i<nr; i++) {
196: ISGetLocalSize(bA->isglobal.row[i],contents->dm + i+1);
197: maxm = PetscMax(maxm,contents->dm[i+1]);
198: contents->dm[i+1] += contents->dm[i];
199: }
200: for (i=0; i<nc; i++) {
201: ISGetLocalSize(bA->isglobal.col[i],contents->dn + i+1);
202: contents->dn[i+1] += contents->dn[i];
203: }
204: PetscMalloc1(maxm*N,&contents->tarray);
205: MatDenseGetLDA(B,&ldb);
206: MatGetSize(B,NULL,&N);
207: MatDenseGetArrayRead(B,&barray);
208: /* loops are permuted compared to MatMatMultNumeric so that viewB is created only once per column of A */
209: for (j=0; j<nc; j++) {
210: ISGetSize(bA->isglobal.col[j],&M);
211: MatCreateDense(PetscObjectComm((PetscObject)A),contents->dn[j+1]-contents->dn[j],PETSC_DECIDE,M,N,(PetscScalar*)(barray+contents->dn[j]),&viewB);
212: MatDenseGetLocalMatrix(viewB,&viewSeq);
213: MatSeqDenseSetLDA(viewSeq,ldb);
214: for (i=0; i<nr; i++) {
215: if (!bA->m[i][j]) continue;
216: /* MatMatMultSymbolic may attach a specific container (depending on MatType of bA->m[i][j]) to workC[i][j] */
218: MatProductCreate(bA->m[i][j],viewB,NULL,&contents->workC[i*nc + j]);
219: workC = contents->workC[i*nc + j];
220: MatProductSetType(workC,MATPRODUCT_AB);
221: MatProductSetAlgorithm(workC,"default");
222: MatProductSetFill(workC,fill);
223: MatProductSetFromOptions(workC);
224: MatProductSymbolic(workC);
226: MatDenseGetLocalMatrix(workC,&viewSeq);
227: /* free the memory allocated in MatMatMultSymbolic, since tarray will be shared by all Mat */
228: MatSeqDenseSetPreallocation(viewSeq,contents->tarray);
229: }
230: MatDestroy(&viewB);
231: }
232: MatDenseRestoreArrayRead(B,&barray);
234: C->ops->matmultnumeric = MatMatMultNumeric_Nest_Dense;
235: return(0);
236: }
238: /* --------------------------------------------------------- */
239: static PetscErrorCode MatProductSetFromOptions_Nest_Dense_AB(Mat C)
240: {
242: C->ops->matmultsymbolic = MatMatMultSymbolic_Nest_Dense;
243: C->ops->productsymbolic = MatProductSymbolic_AB;
244: return(0);
245: }
247: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_Nest_Dense(Mat C)
248: {
250: Mat_Product *product = C->product;
253: if (product->type == MATPRODUCT_AB) {
254: MatProductSetFromOptions_Nest_Dense_AB(C);
255: } else SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_SUP,"MatProduct type is not supported");
256: return(0);
257: }
258: /* --------------------------------------------------------- */
260: static PetscErrorCode MatMultTranspose_Nest(Mat A,Vec x,Vec y)
261: {
262: Mat_Nest *bA = (Mat_Nest*)A->data;
263: Vec *bx = bA->left,*by = bA->right;
264: PetscInt i,j,nr = bA->nr,nc = bA->nc;
268: for (i=0; i<nr; i++) {VecGetSubVector(x,bA->isglobal.row[i],&bx[i]);}
269: for (i=0; i<nc; i++) {VecGetSubVector(y,bA->isglobal.col[i],&by[i]);}
270: for (j=0; j<nc; j++) {
271: VecZeroEntries(by[j]);
272: for (i=0; i<nr; i++) {
273: if (!bA->m[i][j]) continue;
274: /* y[j] <- y[j] + (A[i][j])^T * x[i] */
275: MatMultTransposeAdd(bA->m[i][j],bx[i],by[j],by[j]);
276: }
277: }
278: for (i=0; i<nr; i++) {VecRestoreSubVector(x,bA->isglobal.row[i],&bx[i]);}
279: for (i=0; i<nc; i++) {VecRestoreSubVector(y,bA->isglobal.col[i],&by[i]);}
280: return(0);
281: }
283: static PetscErrorCode MatMultTransposeAdd_Nest(Mat A,Vec x,Vec y,Vec z)
284: {
285: Mat_Nest *bA = (Mat_Nest*)A->data;
286: Vec *bx = bA->left,*bz = bA->right;
287: PetscInt i,j,nr = bA->nr,nc = bA->nc;
291: for (i=0; i<nr; i++) {VecGetSubVector(x,bA->isglobal.row[i],&bx[i]);}
292: for (i=0; i<nc; i++) {VecGetSubVector(z,bA->isglobal.col[i],&bz[i]);}
293: for (j=0; j<nc; j++) {
294: if (y != z) {
295: Vec by;
296: VecGetSubVector(y,bA->isglobal.col[j],&by);
297: VecCopy(by,bz[j]);
298: VecRestoreSubVector(y,bA->isglobal.col[j],&by);
299: }
300: for (i=0; i<nr; i++) {
301: if (!bA->m[i][j]) continue;
302: /* z[j] <- y[j] + (A[i][j])^T * x[i] */
303: MatMultTransposeAdd(bA->m[i][j],bx[i],bz[j],bz[j]);
304: }
305: }
306: for (i=0; i<nr; i++) {VecRestoreSubVector(x,bA->isglobal.row[i],&bx[i]);}
307: for (i=0; i<nc; i++) {VecRestoreSubVector(z,bA->isglobal.col[i],&bz[i]);}
308: return(0);
309: }
311: static PetscErrorCode MatTranspose_Nest(Mat A,MatReuse reuse,Mat *B)
312: {
313: Mat_Nest *bA = (Mat_Nest*)A->data, *bC;
314: Mat C;
315: PetscInt i,j,nr = bA->nr,nc = bA->nc;
319: if (reuse == MAT_INPLACE_MATRIX && nr != nc) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square nested matrix only for in-place");
321: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
322: Mat *subs;
323: IS *is_row,*is_col;
325: PetscCalloc1(nr * nc,&subs);
326: PetscMalloc2(nr,&is_row,nc,&is_col);
327: MatNestGetISs(A,is_row,is_col);
328: if (reuse == MAT_INPLACE_MATRIX) {
329: for (i=0; i<nr; i++) {
330: for (j=0; j<nc; j++) {
331: subs[i + nr * j] = bA->m[i][j];
332: }
333: }
334: }
336: MatCreateNest(PetscObjectComm((PetscObject)A),nc,is_col,nr,is_row,subs,&C);
337: PetscFree(subs);
338: PetscFree2(is_row,is_col);
339: } else {
340: C = *B;
341: }
343: bC = (Mat_Nest*)C->data;
344: for (i=0; i<nr; i++) {
345: for (j=0; j<nc; j++) {
346: if (bA->m[i][j]) {
347: MatTranspose(bA->m[i][j], reuse, &(bC->m[j][i]));
348: } else {
349: bC->m[j][i] = NULL;
350: }
351: }
352: }
354: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
355: *B = C;
356: } else {
357: MatHeaderMerge(A, &C);
358: }
359: return(0);
360: }
362: static PetscErrorCode MatNestDestroyISList(PetscInt n,IS **list)
363: {
365: IS *lst = *list;
366: PetscInt i;
369: if (!lst) return(0);
370: for (i=0; i<n; i++) if (lst[i]) {ISDestroy(&lst[i]);}
371: PetscFree(lst);
372: *list = NULL;
373: return(0);
374: }
376: static PetscErrorCode MatReset_Nest(Mat A)
377: {
378: Mat_Nest *vs = (Mat_Nest*)A->data;
379: PetscInt i,j;
383: /* release the matrices and the place holders */
384: MatNestDestroyISList(vs->nr,&vs->isglobal.row);
385: MatNestDestroyISList(vs->nc,&vs->isglobal.col);
386: MatNestDestroyISList(vs->nr,&vs->islocal.row);
387: MatNestDestroyISList(vs->nc,&vs->islocal.col);
389: PetscFree(vs->row_len);
390: PetscFree(vs->col_len);
391: PetscFree(vs->nnzstate);
393: PetscFree2(vs->left,vs->right);
395: /* release the matrices and the place holders */
396: if (vs->m) {
397: for (i=0; i<vs->nr; i++) {
398: for (j=0; j<vs->nc; j++) {
399: MatDestroy(&vs->m[i][j]);
400: }
401: PetscFree(vs->m[i]);
402: }
403: PetscFree(vs->m);
404: }
406: /* restore defaults */
407: vs->nr = 0;
408: vs->nc = 0;
409: vs->splitassembly = PETSC_FALSE;
410: return(0);
411: }
413: static PetscErrorCode MatDestroy_Nest(Mat A)
414: {
417: MatReset_Nest(A);
418: PetscFree(A->data);
420: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSubMat_C",0);
421: PetscObjectComposeFunction((PetscObject)A,"MatNestSetSubMat_C",0);
422: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSubMats_C",0);
423: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSize_C",0);
424: PetscObjectComposeFunction((PetscObject)A,"MatNestGetISs_C",0);
425: PetscObjectComposeFunction((PetscObject)A,"MatNestGetLocalISs_C",0);
426: PetscObjectComposeFunction((PetscObject)A,"MatNestSetVecType_C",0);
427: PetscObjectComposeFunction((PetscObject)A,"MatNestSetSubMats_C",0);
428: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_mpiaij_C",0);
429: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_seqaij_C",0);
430: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_aij_C",0);
431: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_is_C",0);
432: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_seqdense_C",NULL);
433: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_mpidense_C",NULL);
434: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_dense_C",NULL);
435: return(0);
436: }
438: static PetscErrorCode MatMissingDiagonal_Nest(Mat mat,PetscBool *missing,PetscInt *dd)
439: {
440: Mat_Nest *vs = (Mat_Nest*)mat->data;
441: PetscInt i;
445: if (dd) *dd = 0;
446: if (!vs->nr) {
447: *missing = PETSC_TRUE;
448: return(0);
449: }
450: *missing = PETSC_FALSE;
451: for (i = 0; i < vs->nr && !(*missing); i++) {
452: *missing = PETSC_TRUE;
453: if (vs->m[i][i]) {
454: MatMissingDiagonal(vs->m[i][i],missing,NULL);
455: if (*missing && dd) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"First missing entry not yet implemented");
456: }
457: }
458: return(0);
459: }
461: static PetscErrorCode MatAssemblyBegin_Nest(Mat A,MatAssemblyType type)
462: {
463: Mat_Nest *vs = (Mat_Nest*)A->data;
464: PetscInt i,j;
466: PetscBool nnzstate = PETSC_FALSE;
469: for (i=0; i<vs->nr; i++) {
470: for (j=0; j<vs->nc; j++) {
471: PetscObjectState subnnzstate = 0;
472: if (vs->m[i][j]) {
473: MatAssemblyBegin(vs->m[i][j],type);
474: if (!vs->splitassembly) {
475: /* Note: split assembly will fail if the same block appears more than once (even indirectly through a nested
476: * sub-block). This could be fixed by adding a flag to Mat so that there was a way to check if a Mat was
477: * already performing an assembly, but the result would by more complicated and appears to offer less
478: * potential for diagnostics and correctness checking. Split assembly should be fixed once there is an
479: * interface for libraries to make asynchronous progress in "user-defined non-blocking collectives".
480: */
481: MatAssemblyEnd(vs->m[i][j],type);
482: MatGetNonzeroState(vs->m[i][j],&subnnzstate);
483: }
484: }
485: nnzstate = (PetscBool)(nnzstate || vs->nnzstate[i*vs->nc+j] != subnnzstate);
486: vs->nnzstate[i*vs->nc+j] = subnnzstate;
487: }
488: }
489: if (nnzstate) A->nonzerostate++;
490: return(0);
491: }
493: static PetscErrorCode MatAssemblyEnd_Nest(Mat A, MatAssemblyType type)
494: {
495: Mat_Nest *vs = (Mat_Nest*)A->data;
496: PetscInt i,j;
500: for (i=0; i<vs->nr; i++) {
501: for (j=0; j<vs->nc; j++) {
502: if (vs->m[i][j]) {
503: if (vs->splitassembly) {
504: MatAssemblyEnd(vs->m[i][j],type);
505: }
506: }
507: }
508: }
509: return(0);
510: }
512: static PetscErrorCode MatNestFindNonzeroSubMatRow(Mat A,PetscInt row,Mat *B)
513: {
515: Mat_Nest *vs = (Mat_Nest*)A->data;
516: PetscInt j;
517: Mat sub;
520: sub = (row < vs->nc) ? vs->m[row][row] : (Mat)NULL; /* Prefer to find on the diagonal */
521: for (j=0; !sub && j<vs->nc; j++) sub = vs->m[row][j];
522: if (sub) {MatSetUp(sub);} /* Ensure that the sizes are available */
523: *B = sub;
524: return(0);
525: }
527: static PetscErrorCode MatNestFindNonzeroSubMatCol(Mat A,PetscInt col,Mat *B)
528: {
530: Mat_Nest *vs = (Mat_Nest*)A->data;
531: PetscInt i;
532: Mat sub;
535: sub = (col < vs->nr) ? vs->m[col][col] : (Mat)NULL; /* Prefer to find on the diagonal */
536: for (i=0; !sub && i<vs->nr; i++) sub = vs->m[i][col];
537: if (sub) {MatSetUp(sub);} /* Ensure that the sizes are available */
538: *B = sub;
539: return(0);
540: }
542: static PetscErrorCode MatNestFindIS(Mat A,PetscInt n,const IS list[],IS is,PetscInt *found)
543: {
545: PetscInt i;
546: PetscBool flg;
552: *found = -1;
553: for (i=0; i<n; i++) {
554: if (!list[i]) continue;
555: ISEqualUnsorted(list[i],is,&flg);
556: if (flg) {
557: *found = i;
558: return(0);
559: }
560: }
561: SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Could not find index set");
562: return(0);
563: }
565: /* Get a block row as a new MatNest */
566: static PetscErrorCode MatNestGetRow(Mat A,PetscInt row,Mat *B)
567: {
568: Mat_Nest *vs = (Mat_Nest*)A->data;
569: char keyname[256];
573: *B = NULL;
574: PetscSNPrintf(keyname,sizeof(keyname),"NestRow_%D",row);
575: PetscObjectQuery((PetscObject)A,keyname,(PetscObject*)B);
576: if (*B) return(0);
578: MatCreateNest(PetscObjectComm((PetscObject)A),1,NULL,vs->nc,vs->isglobal.col,vs->m[row],B);
580: (*B)->assembled = A->assembled;
582: PetscObjectCompose((PetscObject)A,keyname,(PetscObject)*B);
583: PetscObjectDereference((PetscObject)*B); /* Leave the only remaining reference in the composition */
584: return(0);
585: }
587: static PetscErrorCode MatNestFindSubMat(Mat A,struct MatNestISPair *is,IS isrow,IS iscol,Mat *B)
588: {
589: Mat_Nest *vs = (Mat_Nest*)A->data;
591: PetscInt row,col;
592: PetscBool same,isFullCol,isFullColGlobal;
595: /* Check if full column space. This is a hack */
596: isFullCol = PETSC_FALSE;
597: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&same);
598: if (same) {
599: PetscInt n,first,step,i,an,am,afirst,astep;
600: ISStrideGetInfo(iscol,&first,&step);
601: ISGetLocalSize(iscol,&n);
602: isFullCol = PETSC_TRUE;
603: for (i=0,an=A->cmap->rstart; i<vs->nc; i++) {
604: PetscObjectTypeCompare((PetscObject)is->col[i],ISSTRIDE,&same);
605: ISGetLocalSize(is->col[i],&am);
606: if (same) {
607: ISStrideGetInfo(is->col[i],&afirst,&astep);
608: if (afirst != an || astep != step) isFullCol = PETSC_FALSE;
609: } else isFullCol = PETSC_FALSE;
610: an += am;
611: }
612: if (an != A->cmap->rstart+n) isFullCol = PETSC_FALSE;
613: }
614: MPIU_Allreduce(&isFullCol,&isFullColGlobal,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)iscol));
616: if (isFullColGlobal && vs->nc > 1) {
617: PetscInt row;
618: MatNestFindIS(A,vs->nr,is->row,isrow,&row);
619: MatNestGetRow(A,row,B);
620: } else {
621: MatNestFindIS(A,vs->nr,is->row,isrow,&row);
622: MatNestFindIS(A,vs->nc,is->col,iscol,&col);
623: if (!vs->m[row][col]) {
624: PetscInt lr,lc;
626: MatCreate(PetscObjectComm((PetscObject)A),&vs->m[row][col]);
627: ISGetLocalSize(vs->isglobal.row[row],&lr);
628: ISGetLocalSize(vs->isglobal.col[col],&lc);
629: MatSetSizes(vs->m[row][col],lr,lc,PETSC_DECIDE,PETSC_DECIDE);
630: MatSetType(vs->m[row][col],MATAIJ);
631: MatSeqAIJSetPreallocation(vs->m[row][col],0,NULL);
632: MatMPIAIJSetPreallocation(vs->m[row][col],0,NULL,0,NULL);
633: MatSetUp(vs->m[row][col]);
634: MatAssemblyBegin(vs->m[row][col],MAT_FINAL_ASSEMBLY);
635: MatAssemblyEnd(vs->m[row][col],MAT_FINAL_ASSEMBLY);
636: }
637: *B = vs->m[row][col];
638: }
639: return(0);
640: }
642: /*
643: TODO: This does not actually returns a submatrix we can modify
644: */
645: static PetscErrorCode MatCreateSubMatrix_Nest(Mat A,IS isrow,IS iscol,MatReuse reuse,Mat *B)
646: {
648: Mat_Nest *vs = (Mat_Nest*)A->data;
649: Mat sub;
652: MatNestFindSubMat(A,&vs->isglobal,isrow,iscol,&sub);
653: switch (reuse) {
654: case MAT_INITIAL_MATRIX:
655: if (sub) { PetscObjectReference((PetscObject)sub); }
656: *B = sub;
657: break;
658: case MAT_REUSE_MATRIX:
659: if (sub != *B) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Submatrix was not used before in this call");
660: break;
661: case MAT_IGNORE_MATRIX: /* Nothing to do */
662: break;
663: case MAT_INPLACE_MATRIX: /* Nothing to do */
664: SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MAT_INPLACE_MATRIX is not supported yet");
665: break;
666: }
667: return(0);
668: }
670: PetscErrorCode MatGetLocalSubMatrix_Nest(Mat A,IS isrow,IS iscol,Mat *B)
671: {
673: Mat_Nest *vs = (Mat_Nest*)A->data;
674: Mat sub;
677: MatNestFindSubMat(A,&vs->islocal,isrow,iscol,&sub);
678: /* We allow the submatrix to be NULL, perhaps it would be better for the user to return an empty matrix instead */
679: if (sub) {PetscObjectReference((PetscObject)sub);}
680: *B = sub;
681: return(0);
682: }
684: static PetscErrorCode MatRestoreLocalSubMatrix_Nest(Mat A,IS isrow,IS iscol,Mat *B)
685: {
687: Mat_Nest *vs = (Mat_Nest*)A->data;
688: Mat sub;
691: MatNestFindSubMat(A,&vs->islocal,isrow,iscol,&sub);
692: if (*B != sub) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Local submatrix has not been gotten");
693: if (sub) {
694: if (((PetscObject)sub)->refct <= 1) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Local submatrix has had reference count decremented too many times");
695: MatDestroy(B);
696: }
697: return(0);
698: }
700: static PetscErrorCode MatGetDiagonal_Nest(Mat A,Vec v)
701: {
702: Mat_Nest *bA = (Mat_Nest*)A->data;
703: PetscInt i;
707: for (i=0; i<bA->nr; i++) {
708: Vec bv;
709: VecGetSubVector(v,bA->isglobal.row[i],&bv);
710: if (bA->m[i][i]) {
711: MatGetDiagonal(bA->m[i][i],bv);
712: } else {
713: VecSet(bv,0.0);
714: }
715: VecRestoreSubVector(v,bA->isglobal.row[i],&bv);
716: }
717: return(0);
718: }
720: static PetscErrorCode MatDiagonalScale_Nest(Mat A,Vec l,Vec r)
721: {
722: Mat_Nest *bA = (Mat_Nest*)A->data;
723: Vec bl,*br;
724: PetscInt i,j;
728: PetscCalloc1(bA->nc,&br);
729: if (r) {
730: for (j=0; j<bA->nc; j++) {VecGetSubVector(r,bA->isglobal.col[j],&br[j]);}
731: }
732: bl = NULL;
733: for (i=0; i<bA->nr; i++) {
734: if (l) {
735: VecGetSubVector(l,bA->isglobal.row[i],&bl);
736: }
737: for (j=0; j<bA->nc; j++) {
738: if (bA->m[i][j]) {
739: MatDiagonalScale(bA->m[i][j],bl,br[j]);
740: }
741: }
742: if (l) {
743: VecRestoreSubVector(l,bA->isglobal.row[i],&bl);
744: }
745: }
746: if (r) {
747: for (j=0; j<bA->nc; j++) {VecRestoreSubVector(r,bA->isglobal.col[j],&br[j]);}
748: }
749: PetscFree(br);
750: return(0);
751: }
753: static PetscErrorCode MatScale_Nest(Mat A,PetscScalar a)
754: {
755: Mat_Nest *bA = (Mat_Nest*)A->data;
756: PetscInt i,j;
760: for (i=0; i<bA->nr; i++) {
761: for (j=0; j<bA->nc; j++) {
762: if (bA->m[i][j]) {
763: MatScale(bA->m[i][j],a);
764: }
765: }
766: }
767: return(0);
768: }
770: static PetscErrorCode MatShift_Nest(Mat A,PetscScalar a)
771: {
772: Mat_Nest *bA = (Mat_Nest*)A->data;
773: PetscInt i;
775: PetscBool nnzstate = PETSC_FALSE;
778: for (i=0; i<bA->nr; i++) {
779: PetscObjectState subnnzstate = 0;
780: if (!bA->m[i][i]) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"No support for shifting an empty diagonal block, insert a matrix in block (%D,%D)",i,i);
781: MatShift(bA->m[i][i],a);
782: MatGetNonzeroState(bA->m[i][i],&subnnzstate);
783: nnzstate = (PetscBool)(nnzstate || bA->nnzstate[i*bA->nc+i] != subnnzstate);
784: bA->nnzstate[i*bA->nc+i] = subnnzstate;
785: }
786: if (nnzstate) A->nonzerostate++;
787: return(0);
788: }
790: static PetscErrorCode MatDiagonalSet_Nest(Mat A,Vec D,InsertMode is)
791: {
792: Mat_Nest *bA = (Mat_Nest*)A->data;
793: PetscInt i;
795: PetscBool nnzstate = PETSC_FALSE;
798: for (i=0; i<bA->nr; i++) {
799: PetscObjectState subnnzstate = 0;
800: Vec bv;
801: VecGetSubVector(D,bA->isglobal.row[i],&bv);
802: if (bA->m[i][i]) {
803: MatDiagonalSet(bA->m[i][i],bv,is);
804: MatGetNonzeroState(bA->m[i][i],&subnnzstate);
805: }
806: VecRestoreSubVector(D,bA->isglobal.row[i],&bv);
807: nnzstate = (PetscBool)(nnzstate || bA->nnzstate[i*bA->nc+i] != subnnzstate);
808: bA->nnzstate[i*bA->nc+i] = subnnzstate;
809: }
810: if (nnzstate) A->nonzerostate++;
811: return(0);
812: }
814: static PetscErrorCode MatSetRandom_Nest(Mat A,PetscRandom rctx)
815: {
816: Mat_Nest *bA = (Mat_Nest*)A->data;
817: PetscInt i,j;
821: for (i=0; i<bA->nr; i++) {
822: for (j=0; j<bA->nc; j++) {
823: if (bA->m[i][j]) {
824: MatSetRandom(bA->m[i][j],rctx);
825: }
826: }
827: }
828: return(0);
829: }
831: static PetscErrorCode MatCreateVecs_Nest(Mat A,Vec *right,Vec *left)
832: {
833: Mat_Nest *bA = (Mat_Nest*)A->data;
834: Vec *L,*R;
835: MPI_Comm comm;
836: PetscInt i,j;
840: PetscObjectGetComm((PetscObject)A,&comm);
841: if (right) {
842: /* allocate R */
843: PetscMalloc1(bA->nc, &R);
844: /* Create the right vectors */
845: for (j=0; j<bA->nc; j++) {
846: for (i=0; i<bA->nr; i++) {
847: if (bA->m[i][j]) {
848: MatCreateVecs(bA->m[i][j],&R[j],NULL);
849: break;
850: }
851: }
852: if (i==bA->nr) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Mat(Nest) contains a null column.");
853: }
854: VecCreateNest(comm,bA->nc,bA->isglobal.col,R,right);
855: /* hand back control to the nest vector */
856: for (j=0; j<bA->nc; j++) {
857: VecDestroy(&R[j]);
858: }
859: PetscFree(R);
860: }
862: if (left) {
863: /* allocate L */
864: PetscMalloc1(bA->nr, &L);
865: /* Create the left vectors */
866: for (i=0; i<bA->nr; i++) {
867: for (j=0; j<bA->nc; j++) {
868: if (bA->m[i][j]) {
869: MatCreateVecs(bA->m[i][j],NULL,&L[i]);
870: break;
871: }
872: }
873: if (j==bA->nc) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Mat(Nest) contains a null row.");
874: }
876: VecCreateNest(comm,bA->nr,bA->isglobal.row,L,left);
877: for (i=0; i<bA->nr; i++) {
878: VecDestroy(&L[i]);
879: }
881: PetscFree(L);
882: }
883: return(0);
884: }
886: static PetscErrorCode MatView_Nest(Mat A,PetscViewer viewer)
887: {
888: Mat_Nest *bA = (Mat_Nest*)A->data;
889: PetscBool isascii,viewSub = PETSC_FALSE;
890: PetscInt i,j;
894: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
895: if (isascii) {
897: PetscOptionsGetBool(((PetscObject)A)->options,((PetscObject)A)->prefix,"-mat_view_nest_sub",&viewSub,NULL);
898: PetscViewerASCIIPrintf(viewer,"Matrix object: \n");
899: PetscViewerASCIIPushTab(viewer);
900: PetscViewerASCIIPrintf(viewer, "type=nest, rows=%D, cols=%D \n",bA->nr,bA->nc);
902: PetscViewerASCIIPrintf(viewer,"MatNest structure: \n");
903: for (i=0; i<bA->nr; i++) {
904: for (j=0; j<bA->nc; j++) {
905: MatType type;
906: char name[256] = "",prefix[256] = "";
907: PetscInt NR,NC;
908: PetscBool isNest = PETSC_FALSE;
910: if (!bA->m[i][j]) {
911: PetscViewerASCIIPrintf(viewer, "(%D,%D) : NULL \n",i,j);
912: continue;
913: }
914: MatGetSize(bA->m[i][j],&NR,&NC);
915: MatGetType(bA->m[i][j], &type);
916: if (((PetscObject)bA->m[i][j])->name) {PetscSNPrintf(name,sizeof(name),"name=\"%s\", ",((PetscObject)bA->m[i][j])->name);}
917: if (((PetscObject)bA->m[i][j])->prefix) {PetscSNPrintf(prefix,sizeof(prefix),"prefix=\"%s\", ",((PetscObject)bA->m[i][j])->prefix);}
918: PetscObjectTypeCompare((PetscObject)bA->m[i][j],MATNEST,&isNest);
920: PetscViewerASCIIPrintf(viewer,"(%D,%D) : %s%stype=%s, rows=%D, cols=%D \n",i,j,name,prefix,type,NR,NC);
922: if (isNest || viewSub) {
923: PetscViewerASCIIPushTab(viewer); /* push1 */
924: MatView(bA->m[i][j],viewer);
925: PetscViewerASCIIPopTab(viewer); /* pop1 */
926: }
927: }
928: }
929: PetscViewerASCIIPopTab(viewer); /* pop0 */
930: }
931: return(0);
932: }
934: static PetscErrorCode MatZeroEntries_Nest(Mat A)
935: {
936: Mat_Nest *bA = (Mat_Nest*)A->data;
937: PetscInt i,j;
941: for (i=0; i<bA->nr; i++) {
942: for (j=0; j<bA->nc; j++) {
943: if (!bA->m[i][j]) continue;
944: MatZeroEntries(bA->m[i][j]);
945: }
946: }
947: return(0);
948: }
950: static PetscErrorCode MatCopy_Nest(Mat A,Mat B,MatStructure str)
951: {
952: Mat_Nest *bA = (Mat_Nest*)A->data,*bB = (Mat_Nest*)B->data;
953: PetscInt i,j,nr = bA->nr,nc = bA->nc;
955: PetscBool nnzstate = PETSC_FALSE;
958: if (nr != bB->nr || nc != bB->nc) SETERRQ4(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Cannot copy a Mat_Nest of block size (%D,%D) to a Mat_Nest of block size (%D,%D)",bB->nr,bB->nc,nr,nc);
959: for (i=0; i<nr; i++) {
960: for (j=0; j<nc; j++) {
961: PetscObjectState subnnzstate = 0;
962: if (bA->m[i][j] && bB->m[i][j]) {
963: MatCopy(bA->m[i][j],bB->m[i][j],str);
964: } else if (bA->m[i][j] || bB->m[i][j]) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"Matrix block does not exist at %D,%D",i,j);
965: MatGetNonzeroState(bB->m[i][j],&subnnzstate);
966: nnzstate = (PetscBool)(nnzstate || bB->nnzstate[i*nc+j] != subnnzstate);
967: bB->nnzstate[i*nc+j] = subnnzstate;
968: }
969: }
970: if (nnzstate) B->nonzerostate++;
971: return(0);
972: }
974: static PetscErrorCode MatAXPY_Nest(Mat Y,PetscScalar a,Mat X,MatStructure str)
975: {
976: Mat_Nest *bY = (Mat_Nest*)Y->data,*bX = (Mat_Nest*)X->data;
977: PetscInt i,j,nr = bY->nr,nc = bY->nc;
979: PetscBool nnzstate = PETSC_FALSE;
982: if (nr != bX->nr || nc != bX->nc) SETERRQ4(PetscObjectComm((PetscObject)Y),PETSC_ERR_ARG_INCOMP,"Cannot AXPY a MatNest of block size (%D,%D) with a MatNest of block size (%D,%D)",bX->nr,bX->nc,nr,nc);
983: for (i=0; i<nr; i++) {
984: for (j=0; j<nc; j++) {
985: PetscObjectState subnnzstate = 0;
986: if (bY->m[i][j] && bX->m[i][j]) {
987: MatAXPY(bY->m[i][j],a,bX->m[i][j],str);
988: } else if (bX->m[i][j]) {
989: Mat M;
991: if (str != DIFFERENT_NONZERO_PATTERN) SETERRQ2(PetscObjectComm((PetscObject)Y),PETSC_ERR_ARG_INCOMP,"Matrix block does not exist at %D,%D. Use DIFFERENT_NONZERO_PATTERN",i,j);
992: MatDuplicate(bX->m[i][j],MAT_COPY_VALUES,&M);
993: MatNestSetSubMat(Y,i,j,M);
994: MatDestroy(&M);
995: }
996: if (bY->m[i][j]) { MatGetNonzeroState(bY->m[i][j],&subnnzstate); }
997: nnzstate = (PetscBool)(nnzstate || bY->nnzstate[i*nc+j] != subnnzstate);
998: bY->nnzstate[i*nc+j] = subnnzstate;
999: }
1000: }
1001: if (nnzstate) Y->nonzerostate++;
1002: return(0);
1003: }
1005: static PetscErrorCode MatDuplicate_Nest(Mat A,MatDuplicateOption op,Mat *B)
1006: {
1007: Mat_Nest *bA = (Mat_Nest*)A->data;
1008: Mat *b;
1009: PetscInt i,j,nr = bA->nr,nc = bA->nc;
1013: PetscMalloc1(nr*nc,&b);
1014: for (i=0; i<nr; i++) {
1015: for (j=0; j<nc; j++) {
1016: if (bA->m[i][j]) {
1017: MatDuplicate(bA->m[i][j],op,&b[i*nc+j]);
1018: } else {
1019: b[i*nc+j] = NULL;
1020: }
1021: }
1022: }
1023: MatCreateNest(PetscObjectComm((PetscObject)A),nr,bA->isglobal.row,nc,bA->isglobal.col,b,B);
1024: /* Give the new MatNest exclusive ownership */
1025: for (i=0; i<nr*nc; i++) {
1026: MatDestroy(&b[i]);
1027: }
1028: PetscFree(b);
1030: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
1031: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
1032: return(0);
1033: }
1035: /* nest api */
1036: PetscErrorCode MatNestGetSubMat_Nest(Mat A,PetscInt idxm,PetscInt jdxm,Mat *mat)
1037: {
1038: Mat_Nest *bA = (Mat_Nest*)A->data;
1041: if (idxm >= bA->nr) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm,bA->nr-1);
1042: if (jdxm >= bA->nc) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Col too large: row %D max %D",jdxm,bA->nc-1);
1043: *mat = bA->m[idxm][jdxm];
1044: return(0);
1045: }
1047: /*@
1048: MatNestGetSubMat - Returns a single, sub-matrix from a nest matrix.
1050: Not collective
1052: Input Parameters:
1053: + A - nest matrix
1054: . idxm - index of the matrix within the nest matrix
1055: - jdxm - index of the matrix within the nest matrix
1057: Output Parameter:
1058: . sub - matrix at index idxm,jdxm within the nest matrix
1060: Level: developer
1062: .seealso: MatNestGetSize(), MatNestGetSubMats(), MatNestCreate(), MATNEST, MatNestSetSubMat(),
1063: MatNestGetLocalISs(), MatNestGetISs()
1064: @*/
1065: PetscErrorCode MatNestGetSubMat(Mat A,PetscInt idxm,PetscInt jdxm,Mat *sub)
1066: {
1070: PetscUseMethod(A,"MatNestGetSubMat_C",(Mat,PetscInt,PetscInt,Mat*),(A,idxm,jdxm,sub));
1071: return(0);
1072: }
1074: PetscErrorCode MatNestSetSubMat_Nest(Mat A,PetscInt idxm,PetscInt jdxm,Mat mat)
1075: {
1076: Mat_Nest *bA = (Mat_Nest*)A->data;
1077: PetscInt m,n,M,N,mi,ni,Mi,Ni;
1081: if (idxm >= bA->nr) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm,bA->nr-1);
1082: if (jdxm >= bA->nc) SETERRQ2(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Col too large: row %D max %D",jdxm,bA->nc-1);
1083: MatGetLocalSize(mat,&m,&n);
1084: MatGetSize(mat,&M,&N);
1085: ISGetLocalSize(bA->isglobal.row[idxm],&mi);
1086: ISGetSize(bA->isglobal.row[idxm],&Mi);
1087: ISGetLocalSize(bA->isglobal.col[jdxm],&ni);
1088: ISGetSize(bA->isglobal.col[jdxm],&Ni);
1089: if (M != Mi || N != Ni) SETERRQ4(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"Submatrix dimension (%D,%D) incompatible with nest block (%D,%D)",M,N,Mi,Ni);
1090: if (m != mi || n != ni) SETERRQ4(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"Submatrix local dimension (%D,%D) incompatible with nest block (%D,%D)",m,n,mi,ni);
1092: /* do not increase object state */
1093: if (mat == bA->m[idxm][jdxm]) return(0);
1095: PetscObjectReference((PetscObject)mat);
1096: MatDestroy(&bA->m[idxm][jdxm]);
1097: bA->m[idxm][jdxm] = mat;
1098: PetscObjectStateIncrease((PetscObject)A);
1099: MatGetNonzeroState(mat,&bA->nnzstate[idxm*bA->nc+jdxm]);
1100: A->nonzerostate++;
1101: return(0);
1102: }
1104: /*@
1105: MatNestSetSubMat - Set a single submatrix in the nest matrix.
1107: Logically collective on the submatrix communicator
1109: Input Parameters:
1110: + A - nest matrix
1111: . idxm - index of the matrix within the nest matrix
1112: . jdxm - index of the matrix within the nest matrix
1113: - sub - matrix at index idxm,jdxm within the nest matrix
1115: Notes:
1116: The new submatrix must have the same size and communicator as that block of the nest.
1118: This increments the reference count of the submatrix.
1120: Level: developer
1122: .seealso: MatNestSetSubMats(), MatNestGetSubMats(), MatNestGetLocalISs(), MATNEST, MatNestCreate(),
1123: MatNestGetSubMat(), MatNestGetISs(), MatNestGetSize()
1124: @*/
1125: PetscErrorCode MatNestSetSubMat(Mat A,PetscInt idxm,PetscInt jdxm,Mat sub)
1126: {
1130: PetscUseMethod(A,"MatNestSetSubMat_C",(Mat,PetscInt,PetscInt,Mat),(A,idxm,jdxm,sub));
1131: return(0);
1132: }
1134: PetscErrorCode MatNestGetSubMats_Nest(Mat A,PetscInt *M,PetscInt *N,Mat ***mat)
1135: {
1136: Mat_Nest *bA = (Mat_Nest*)A->data;
1139: if (M) *M = bA->nr;
1140: if (N) *N = bA->nc;
1141: if (mat) *mat = bA->m;
1142: return(0);
1143: }
1145: /*@C
1146: MatNestGetSubMats - Returns the entire two dimensional array of matrices defining a nest matrix.
1148: Not collective
1150: Input Parameters:
1151: . A - nest matrix
1153: Output Parameter:
1154: + M - number of rows in the nest matrix
1155: . N - number of cols in the nest matrix
1156: - mat - 2d array of matrices
1158: Notes:
1160: The user should not free the array mat.
1162: In Fortran, this routine has a calling sequence
1163: $ call MatNestGetSubMats(A, M, N, mat, ierr)
1164: where the space allocated for the optional argument mat is assumed large enough (if provided).
1166: Level: developer
1168: .seealso: MatNestGetSize(), MatNestGetSubMat(), MatNestGetLocalISs(), MATNEST, MatNestCreate(),
1169: MatNestSetSubMats(), MatNestGetISs(), MatNestSetSubMat()
1170: @*/
1171: PetscErrorCode MatNestGetSubMats(Mat A,PetscInt *M,PetscInt *N,Mat ***mat)
1172: {
1176: PetscUseMethod(A,"MatNestGetSubMats_C",(Mat,PetscInt*,PetscInt*,Mat***),(A,M,N,mat));
1177: return(0);
1178: }
1180: PetscErrorCode MatNestGetSize_Nest(Mat A,PetscInt *M,PetscInt *N)
1181: {
1182: Mat_Nest *bA = (Mat_Nest*)A->data;
1185: if (M) *M = bA->nr;
1186: if (N) *N = bA->nc;
1187: return(0);
1188: }
1190: /*@
1191: MatNestGetSize - Returns the size of the nest matrix.
1193: Not collective
1195: Input Parameters:
1196: . A - nest matrix
1198: Output Parameter:
1199: + M - number of rows in the nested mat
1200: - N - number of cols in the nested mat
1202: Notes:
1204: Level: developer
1206: .seealso: MatNestGetSubMat(), MatNestGetSubMats(), MATNEST, MatNestCreate(), MatNestGetLocalISs(),
1207: MatNestGetISs()
1208: @*/
1209: PetscErrorCode MatNestGetSize(Mat A,PetscInt *M,PetscInt *N)
1210: {
1214: PetscUseMethod(A,"MatNestGetSize_C",(Mat,PetscInt*,PetscInt*),(A,M,N));
1215: return(0);
1216: }
1218: static PetscErrorCode MatNestGetISs_Nest(Mat A,IS rows[],IS cols[])
1219: {
1220: Mat_Nest *vs = (Mat_Nest*)A->data;
1221: PetscInt i;
1224: if (rows) for (i=0; i<vs->nr; i++) rows[i] = vs->isglobal.row[i];
1225: if (cols) for (i=0; i<vs->nc; i++) cols[i] = vs->isglobal.col[i];
1226: return(0);
1227: }
1229: /*@C
1230: MatNestGetISs - Returns the index sets partitioning the row and column spaces
1232: Not collective
1234: Input Parameters:
1235: . A - nest matrix
1237: Output Parameter:
1238: + rows - array of row index sets
1239: - cols - array of column index sets
1241: Level: advanced
1243: Notes:
1244: The user must have allocated arrays of the correct size. The reference count is not increased on the returned ISs.
1246: .seealso: MatNestGetSubMat(), MatNestGetSubMats(), MatNestGetSize(), MatNestGetLocalISs(), MATNEST,
1247: MatNestCreate(), MatNestGetSubMats(), MatNestSetSubMats()
1248: @*/
1249: PetscErrorCode MatNestGetISs(Mat A,IS rows[],IS cols[])
1250: {
1255: PetscUseMethod(A,"MatNestGetISs_C",(Mat,IS[],IS[]),(A,rows,cols));
1256: return(0);
1257: }
1259: static PetscErrorCode MatNestGetLocalISs_Nest(Mat A,IS rows[],IS cols[])
1260: {
1261: Mat_Nest *vs = (Mat_Nest*)A->data;
1262: PetscInt i;
1265: if (rows) for (i=0; i<vs->nr; i++) rows[i] = vs->islocal.row[i];
1266: if (cols) for (i=0; i<vs->nc; i++) cols[i] = vs->islocal.col[i];
1267: return(0);
1268: }
1270: /*@C
1271: MatNestGetLocalISs - Returns the index sets partitioning the row and column spaces
1273: Not collective
1275: Input Parameters:
1276: . A - nest matrix
1278: Output Parameter:
1279: + rows - array of row index sets (or NULL to ignore)
1280: - cols - array of column index sets (or NULL to ignore)
1282: Level: advanced
1284: Notes:
1285: The user must have allocated arrays of the correct size. The reference count is not increased on the returned ISs.
1287: .seealso: MatNestGetSubMat(), MatNestGetSubMats(), MatNestGetSize(), MatNestGetISs(), MatNestCreate(),
1288: MATNEST, MatNestSetSubMats(), MatNestSetSubMat()
1289: @*/
1290: PetscErrorCode MatNestGetLocalISs(Mat A,IS rows[],IS cols[])
1291: {
1296: PetscUseMethod(A,"MatNestGetLocalISs_C",(Mat,IS[],IS[]),(A,rows,cols));
1297: return(0);
1298: }
1300: PetscErrorCode MatNestSetVecType_Nest(Mat A,VecType vtype)
1301: {
1303: PetscBool flg;
1306: PetscStrcmp(vtype,VECNEST,&flg);
1307: /* In reality, this only distinguishes VECNEST and "other" */
1308: if (flg) A->ops->getvecs = MatCreateVecs_Nest;
1309: else A->ops->getvecs = (PetscErrorCode (*)(Mat,Vec*,Vec*)) 0;
1310: return(0);
1311: }
1313: /*@C
1314: MatNestSetVecType - Sets the type of Vec returned by MatCreateVecs()
1316: Not collective
1318: Input Parameters:
1319: + A - nest matrix
1320: - vtype - type to use for creating vectors
1322: Notes:
1324: Level: developer
1326: .seealso: MatCreateVecs(), MATNEST, MatNestCreate()
1327: @*/
1328: PetscErrorCode MatNestSetVecType(Mat A,VecType vtype)
1329: {
1333: PetscTryMethod(A,"MatNestSetVecType_C",(Mat,VecType),(A,vtype));
1334: return(0);
1335: }
1337: PetscErrorCode MatNestSetSubMats_Nest(Mat A,PetscInt nr,const IS is_row[],PetscInt nc,const IS is_col[],const Mat a[])
1338: {
1339: Mat_Nest *s = (Mat_Nest*)A->data;
1340: PetscInt i,j,m,n,M,N;
1342: PetscBool cong;
1345: MatReset_Nest(A);
1347: s->nr = nr;
1348: s->nc = nc;
1350: /* Create space for submatrices */
1351: PetscMalloc1(nr,&s->m);
1352: for (i=0; i<nr; i++) {
1353: PetscMalloc1(nc,&s->m[i]);
1354: }
1355: for (i=0; i<nr; i++) {
1356: for (j=0; j<nc; j++) {
1357: s->m[i][j] = a[i*nc+j];
1358: if (a[i*nc+j]) {
1359: PetscObjectReference((PetscObject)a[i*nc+j]);
1360: }
1361: }
1362: }
1364: MatSetUp_NestIS_Private(A,nr,is_row,nc,is_col);
1366: PetscMalloc1(nr,&s->row_len);
1367: PetscMalloc1(nc,&s->col_len);
1368: for (i=0; i<nr; i++) s->row_len[i]=-1;
1369: for (j=0; j<nc; j++) s->col_len[j]=-1;
1371: PetscCalloc1(nr*nc,&s->nnzstate);
1372: for (i=0; i<nr; i++) {
1373: for (j=0; j<nc; j++) {
1374: if (s->m[i][j]) {
1375: MatGetNonzeroState(s->m[i][j],&s->nnzstate[i*nc+j]);
1376: }
1377: }
1378: }
1380: MatNestGetSizes_Private(A,&m,&n,&M,&N);
1382: PetscLayoutSetSize(A->rmap,M);
1383: PetscLayoutSetLocalSize(A->rmap,m);
1384: PetscLayoutSetSize(A->cmap,N);
1385: PetscLayoutSetLocalSize(A->cmap,n);
1387: PetscLayoutSetUp(A->rmap);
1388: PetscLayoutSetUp(A->cmap);
1390: /* disable operations that are not supported for non-square matrices,
1391: or matrices for which is_row != is_col */
1392: MatHasCongruentLayouts(A,&cong);
1393: if (cong && nr != nc) cong = PETSC_FALSE;
1394: if (cong) {
1395: for (i = 0; cong && i < nr; i++) {
1396: ISEqualUnsorted(s->isglobal.row[i],s->isglobal.col[i],&cong);
1397: }
1398: }
1399: if (!cong) {
1400: A->ops->missingdiagonal = NULL;
1401: A->ops->getdiagonal = NULL;
1402: A->ops->shift = NULL;
1403: A->ops->diagonalset = NULL;
1404: }
1406: PetscCalloc2(nr,&s->left,nc,&s->right);
1407: PetscObjectStateIncrease((PetscObject)A);
1408: A->nonzerostate++;
1409: return(0);
1410: }
1412: /*@
1413: MatNestSetSubMats - Sets the nested submatrices
1415: Collective on Mat
1417: Input Parameter:
1418: + A - nested matrix
1419: . nr - number of nested row blocks
1420: . is_row - index sets for each nested row block, or NULL to make contiguous
1421: . nc - number of nested column blocks
1422: . is_col - index sets for each nested column block, or NULL to make contiguous
1423: - a - row-aligned array of nr*nc submatrices, empty submatrices can be passed using NULL
1425: Notes: this always resets any submatrix information previously set
1427: Level: advanced
1429: .seealso: MatCreateNest(), MATNEST, MatNestSetSubMat(), MatNestGetSubMat(), MatNestGetSubMats()
1430: @*/
1431: PetscErrorCode MatNestSetSubMats(Mat A,PetscInt nr,const IS is_row[],PetscInt nc,const IS is_col[],const Mat a[])
1432: {
1434: PetscInt i;
1438: if (nr < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Number of rows cannot be negative");
1439: if (nr && is_row) {
1442: }
1443: if (nc < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Number of columns cannot be negative");
1444: if (nc && is_col) {
1447: }
1449: PetscUseMethod(A,"MatNestSetSubMats_C",(Mat,PetscInt,const IS[],PetscInt,const IS[],const Mat[]),(A,nr,is_row,nc,is_col,a));
1450: return(0);
1451: }
1453: static PetscErrorCode MatNestCreateAggregateL2G_Private(Mat A,PetscInt n,const IS islocal[],const IS isglobal[],PetscBool colflg,ISLocalToGlobalMapping *ltog)
1454: {
1456: PetscBool flg;
1457: PetscInt i,j,m,mi,*ix;
1460: *ltog = NULL;
1461: for (i=0,m=0,flg=PETSC_FALSE; i<n; i++) {
1462: if (islocal[i]) {
1463: ISGetLocalSize(islocal[i],&mi);
1464: flg = PETSC_TRUE; /* We found a non-trivial entry */
1465: } else {
1466: ISGetLocalSize(isglobal[i],&mi);
1467: }
1468: m += mi;
1469: }
1470: if (!flg) return(0);
1472: PetscMalloc1(m,&ix);
1473: for (i=0,m=0; i<n; i++) {
1474: ISLocalToGlobalMapping smap = NULL;
1475: Mat sub = NULL;
1476: PetscSF sf;
1477: PetscLayout map;
1478: const PetscInt *ix2;
1480: if (!colflg) {
1481: MatNestFindNonzeroSubMatRow(A,i,&sub);
1482: } else {
1483: MatNestFindNonzeroSubMatCol(A,i,&sub);
1484: }
1485: if (sub) {
1486: if (!colflg) {
1487: MatGetLocalToGlobalMapping(sub,&smap,NULL);
1488: } else {
1489: MatGetLocalToGlobalMapping(sub,NULL,&smap);
1490: }
1491: }
1492: /*
1493: Now we need to extract the monolithic global indices that correspond to the given split global indices.
1494: In many/most cases, we only want MatGetLocalSubMatrix() to work, in which case we only need to know the size of the local spaces.
1495: */
1496: ISGetIndices(isglobal[i],&ix2);
1497: if (islocal[i]) {
1498: PetscInt *ilocal,*iremote;
1499: PetscInt mil,nleaves;
1501: ISGetLocalSize(islocal[i],&mi);
1502: if (!smap) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"Missing local to global map");
1503: for (j=0; j<mi; j++) ix[m+j] = j;
1504: ISLocalToGlobalMappingApply(smap,mi,ix+m,ix+m);
1506: /* PetscSFSetGraphLayout does not like negative indices */
1507: PetscMalloc2(mi,&ilocal,mi,&iremote);
1508: for (j=0, nleaves = 0; j<mi; j++) {
1509: if (ix[m+j] < 0) continue;
1510: ilocal[nleaves] = j;
1511: iremote[nleaves] = ix[m+j];
1512: nleaves++;
1513: }
1514: ISGetLocalSize(isglobal[i],&mil);
1515: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1516: PetscLayoutCreate(PetscObjectComm((PetscObject)A),&map);
1517: PetscLayoutSetLocalSize(map,mil);
1518: PetscLayoutSetUp(map);
1519: PetscSFSetGraphLayout(sf,map,nleaves,ilocal,PETSC_USE_POINTER,iremote);
1520: PetscLayoutDestroy(&map);
1521: PetscSFBcastBegin(sf,MPIU_INT,ix2,ix + m);
1522: PetscSFBcastEnd(sf,MPIU_INT,ix2,ix + m);
1523: PetscSFDestroy(&sf);
1524: PetscFree2(ilocal,iremote);
1525: } else {
1526: ISGetLocalSize(isglobal[i],&mi);
1527: for (j=0; j<mi; j++) ix[m+j] = ix2[i];
1528: }
1529: ISRestoreIndices(isglobal[i],&ix2);
1530: m += mi;
1531: }
1532: ISLocalToGlobalMappingCreate(PetscObjectComm((PetscObject)A),1,m,ix,PETSC_OWN_POINTER,ltog);
1533: return(0);
1534: }
1537: /* If an IS was provided, there is nothing Nest needs to do, otherwise Nest will build a strided IS */
1538: /*
1539: nprocessors = NP
1540: Nest x^T = ((g_0,g_1,...g_nprocs-1), (h_0,h_1,...h_NP-1))
1541: proc 0: => (g_0,h_0,)
1542: proc 1: => (g_1,h_1,)
1543: ...
1544: proc nprocs-1: => (g_NP-1,h_NP-1,)
1546: proc 0: proc 1: proc nprocs-1:
1547: is[0] = (0,1,2,...,nlocal(g_0)-1) (0,1,...,nlocal(g_1)-1) (0,1,...,nlocal(g_NP-1))
1549: proc 0:
1550: is[1] = (nlocal(g_0),nlocal(g_0)+1,...,nlocal(g_0)+nlocal(h_0)-1)
1551: proc 1:
1552: is[1] = (nlocal(g_1),nlocal(g_1)+1,...,nlocal(g_1)+nlocal(h_1)-1)
1554: proc NP-1:
1555: is[1] = (nlocal(g_NP-1),nlocal(g_NP-1)+1,...,nlocal(g_NP-1)+nlocal(h_NP-1)-1)
1556: */
1557: static PetscErrorCode MatSetUp_NestIS_Private(Mat A,PetscInt nr,const IS is_row[],PetscInt nc,const IS is_col[])
1558: {
1559: Mat_Nest *vs = (Mat_Nest*)A->data;
1560: PetscInt i,j,offset,n,nsum,bs;
1562: Mat sub = NULL;
1565: PetscMalloc1(nr,&vs->isglobal.row);
1566: PetscMalloc1(nc,&vs->isglobal.col);
1567: if (is_row) { /* valid IS is passed in */
1568: /* refs on is[] are incremeneted */
1569: for (i=0; i<vs->nr; i++) {
1570: PetscObjectReference((PetscObject)is_row[i]);
1572: vs->isglobal.row[i] = is_row[i];
1573: }
1574: } else { /* Create the ISs by inspecting sizes of a submatrix in each row */
1575: nsum = 0;
1576: for (i=0; i<vs->nr; i++) { /* Add up the local sizes to compute the aggregate offset */
1577: MatNestFindNonzeroSubMatRow(A,i,&sub);
1578: if (!sub) SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"No nonzero submatrix in row %D",i);
1579: MatGetLocalSize(sub,&n,NULL);
1580: if (n < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Sizes have not yet been set for submatrix");
1581: nsum += n;
1582: }
1583: MPI_Scan(&nsum,&offset,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)A));
1584: offset -= nsum;
1585: for (i=0; i<vs->nr; i++) {
1586: MatNestFindNonzeroSubMatRow(A,i,&sub);
1587: MatGetLocalSize(sub,&n,NULL);
1588: MatGetBlockSize(sub,&bs);
1589: ISCreateStride(PetscObjectComm((PetscObject)sub),n,offset,1,&vs->isglobal.row[i]);
1590: ISSetBlockSize(vs->isglobal.row[i],bs);
1591: offset += n;
1592: }
1593: }
1595: if (is_col) { /* valid IS is passed in */
1596: /* refs on is[] are incremeneted */
1597: for (j=0; j<vs->nc; j++) {
1598: PetscObjectReference((PetscObject)is_col[j]);
1600: vs->isglobal.col[j] = is_col[j];
1601: }
1602: } else { /* Create the ISs by inspecting sizes of a submatrix in each column */
1603: offset = A->cmap->rstart;
1604: nsum = 0;
1605: for (j=0; j<vs->nc; j++) {
1606: MatNestFindNonzeroSubMatCol(A,j,&sub);
1607: if (!sub) SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"No nonzero submatrix in column %D",i);
1608: MatGetLocalSize(sub,NULL,&n);
1609: if (n < 0) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONGSTATE,"Sizes have not yet been set for submatrix");
1610: nsum += n;
1611: }
1612: MPI_Scan(&nsum,&offset,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)A));
1613: offset -= nsum;
1614: for (j=0; j<vs->nc; j++) {
1615: MatNestFindNonzeroSubMatCol(A,j,&sub);
1616: MatGetLocalSize(sub,NULL,&n);
1617: MatGetBlockSize(sub,&bs);
1618: ISCreateStride(PetscObjectComm((PetscObject)sub),n,offset,1,&vs->isglobal.col[j]);
1619: ISSetBlockSize(vs->isglobal.col[j],bs);
1620: offset += n;
1621: }
1622: }
1624: /* Set up the local ISs */
1625: PetscMalloc1(vs->nr,&vs->islocal.row);
1626: PetscMalloc1(vs->nc,&vs->islocal.col);
1627: for (i=0,offset=0; i<vs->nr; i++) {
1628: IS isloc;
1629: ISLocalToGlobalMapping rmap = NULL;
1630: PetscInt nlocal,bs;
1631: MatNestFindNonzeroSubMatRow(A,i,&sub);
1632: if (sub) {MatGetLocalToGlobalMapping(sub,&rmap,NULL);}
1633: if (rmap) {
1634: MatGetBlockSize(sub,&bs);
1635: ISLocalToGlobalMappingGetSize(rmap,&nlocal);
1636: ISCreateStride(PETSC_COMM_SELF,nlocal,offset,1,&isloc);
1637: ISSetBlockSize(isloc,bs);
1638: } else {
1639: nlocal = 0;
1640: isloc = NULL;
1641: }
1642: vs->islocal.row[i] = isloc;
1643: offset += nlocal;
1644: }
1645: for (i=0,offset=0; i<vs->nc; i++) {
1646: IS isloc;
1647: ISLocalToGlobalMapping cmap = NULL;
1648: PetscInt nlocal,bs;
1649: MatNestFindNonzeroSubMatCol(A,i,&sub);
1650: if (sub) {MatGetLocalToGlobalMapping(sub,NULL,&cmap);}
1651: if (cmap) {
1652: MatGetBlockSize(sub,&bs);
1653: ISLocalToGlobalMappingGetSize(cmap,&nlocal);
1654: ISCreateStride(PETSC_COMM_SELF,nlocal,offset,1,&isloc);
1655: ISSetBlockSize(isloc,bs);
1656: } else {
1657: nlocal = 0;
1658: isloc = NULL;
1659: }
1660: vs->islocal.col[i] = isloc;
1661: offset += nlocal;
1662: }
1664: /* Set up the aggregate ISLocalToGlobalMapping */
1665: {
1666: ISLocalToGlobalMapping rmap,cmap;
1667: MatNestCreateAggregateL2G_Private(A,vs->nr,vs->islocal.row,vs->isglobal.row,PETSC_FALSE,&rmap);
1668: MatNestCreateAggregateL2G_Private(A,vs->nc,vs->islocal.col,vs->isglobal.col,PETSC_TRUE,&cmap);
1669: if (rmap && cmap) {MatSetLocalToGlobalMapping(A,rmap,cmap);}
1670: ISLocalToGlobalMappingDestroy(&rmap);
1671: ISLocalToGlobalMappingDestroy(&cmap);
1672: }
1674: #if defined(PETSC_USE_DEBUG)
1675: for (i=0; i<vs->nr; i++) {
1676: for (j=0; j<vs->nc; j++) {
1677: PetscInt m,n,M,N,mi,ni,Mi,Ni;
1678: Mat B = vs->m[i][j];
1679: if (!B) continue;
1680: MatGetSize(B,&M,&N);
1681: MatGetLocalSize(B,&m,&n);
1682: ISGetSize(vs->isglobal.row[i],&Mi);
1683: ISGetSize(vs->isglobal.col[j],&Ni);
1684: ISGetLocalSize(vs->isglobal.row[i],&mi);
1685: ISGetLocalSize(vs->isglobal.col[j],&ni);
1686: if (M != Mi || N != Ni) SETERRQ6(PetscObjectComm((PetscObject)sub),PETSC_ERR_ARG_INCOMP,"Global sizes (%D,%D) of nested submatrix (%D,%D) do not agree with space defined by index sets (%D,%D)",M,N,i,j,Mi,Ni);
1687: if (m != mi || n != ni) SETERRQ6(PetscObjectComm((PetscObject)sub),PETSC_ERR_ARG_INCOMP,"Local sizes (%D,%D) of nested submatrix (%D,%D) do not agree with space defined by index sets (%D,%D)",m,n,i,j,mi,ni);
1688: }
1689: }
1690: #endif
1692: /* Set A->assembled if all non-null blocks are currently assembled */
1693: for (i=0; i<vs->nr; i++) {
1694: for (j=0; j<vs->nc; j++) {
1695: if (vs->m[i][j] && !vs->m[i][j]->assembled) return(0);
1696: }
1697: }
1698: A->assembled = PETSC_TRUE;
1699: return(0);
1700: }
1702: /*@C
1703: MatCreateNest - Creates a new matrix containing several nested submatrices, each stored separately
1705: Collective on Mat
1707: Input Parameter:
1708: + comm - Communicator for the new Mat
1709: . nr - number of nested row blocks
1710: . is_row - index sets for each nested row block, or NULL to make contiguous
1711: . nc - number of nested column blocks
1712: . is_col - index sets for each nested column block, or NULL to make contiguous
1713: - a - row-aligned array of nr*nc submatrices, empty submatrices can be passed using NULL
1715: Output Parameter:
1716: . B - new matrix
1718: Level: advanced
1720: .seealso: MatCreate(), VecCreateNest(), DMCreateMatrix(), MATNEST, MatNestSetSubMat(),
1721: MatNestGetSubMat(), MatNestGetLocalISs(), MatNestGetSize(),
1722: MatNestGetISs(), MatNestSetSubMats(), MatNestGetSubMats()
1723: @*/
1724: PetscErrorCode MatCreateNest(MPI_Comm comm,PetscInt nr,const IS is_row[],PetscInt nc,const IS is_col[],const Mat a[],Mat *B)
1725: {
1726: Mat A;
1730: *B = 0;
1731: MatCreate(comm,&A);
1732: MatSetType(A,MATNEST);
1733: A->preallocated = PETSC_TRUE;
1734: MatNestSetSubMats(A,nr,is_row,nc,is_col,a);
1735: *B = A;
1736: return(0);
1737: }
1739: static PetscErrorCode MatConvert_Nest_SeqAIJ_fast(Mat A,MatType newtype,MatReuse reuse,Mat *newmat)
1740: {
1741: Mat_Nest *nest = (Mat_Nest*)A->data;
1742: Mat *trans;
1743: PetscScalar **avv;
1744: PetscScalar *vv;
1745: PetscInt **aii,**ajj;
1746: PetscInt *ii,*jj,*ci;
1747: PetscInt nr,nc,nnz,i,j;
1748: PetscBool done;
1752: MatGetSize(A,&nr,&nc);
1753: if (reuse == MAT_REUSE_MATRIX) {
1754: PetscInt rnr;
1756: MatGetRowIJ(*newmat,0,PETSC_FALSE,PETSC_FALSE,&rnr,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);
1757: if (!done) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_PLIB,"MatGetRowIJ");
1758: if (rnr != nr) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_USER,"Cannot reuse matrix, wrong number of rows");
1759: MatSeqAIJGetArray(*newmat,&vv);
1760: }
1761: /* extract CSR for nested SeqAIJ matrices */
1762: nnz = 0;
1763: PetscCalloc4(nest->nr*nest->nc,&aii,nest->nr*nest->nc,&ajj,nest->nr*nest->nc,&avv,nest->nr*nest->nc,&trans);
1764: for (i=0; i<nest->nr; ++i) {
1765: for (j=0; j<nest->nc; ++j) {
1766: Mat B = nest->m[i][j];
1767: if (B) {
1768: PetscScalar *naa;
1769: PetscInt *nii,*njj,nnr;
1770: PetscBool istrans;
1772: PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&istrans);
1773: if (istrans) {
1774: Mat Bt;
1776: MatTransposeGetMat(B,&Bt);
1777: MatTranspose(Bt,MAT_INITIAL_MATRIX,&trans[i*nest->nc+j]);
1778: B = trans[i*nest->nc+j];
1779: }
1780: MatGetRowIJ(B,0,PETSC_FALSE,PETSC_FALSE,&nnr,(const PetscInt**)&nii,(const PetscInt**)&njj,&done);
1781: if (!done) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"MatGetRowIJ");
1782: MatSeqAIJGetArray(B,&naa);
1783: nnz += nii[nnr];
1785: aii[i*nest->nc+j] = nii;
1786: ajj[i*nest->nc+j] = njj;
1787: avv[i*nest->nc+j] = naa;
1788: }
1789: }
1790: }
1791: if (reuse != MAT_REUSE_MATRIX) {
1792: PetscMalloc1(nr+1,&ii);
1793: PetscMalloc1(nnz,&jj);
1794: PetscMalloc1(nnz,&vv);
1795: } else {
1796: if (nnz != ii[nr]) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_USER,"Cannot reuse matrix, wrong number of nonzeros");
1797: }
1799: /* new row pointer */
1800: PetscArrayzero(ii,nr+1);
1801: for (i=0; i<nest->nr; ++i) {
1802: PetscInt ncr,rst;
1804: ISStrideGetInfo(nest->isglobal.row[i],&rst,NULL);
1805: ISGetLocalSize(nest->isglobal.row[i],&ncr);
1806: for (j=0; j<nest->nc; ++j) {
1807: if (aii[i*nest->nc+j]) {
1808: PetscInt *nii = aii[i*nest->nc+j];
1809: PetscInt ir;
1811: for (ir=rst; ir<ncr+rst; ++ir) {
1812: ii[ir+1] += nii[1]-nii[0];
1813: nii++;
1814: }
1815: }
1816: }
1817: }
1818: for (i=0; i<nr; i++) ii[i+1] += ii[i];
1820: /* construct CSR for the new matrix */
1821: PetscCalloc1(nr,&ci);
1822: for (i=0; i<nest->nr; ++i) {
1823: PetscInt ncr,rst;
1825: ISStrideGetInfo(nest->isglobal.row[i],&rst,NULL);
1826: ISGetLocalSize(nest->isglobal.row[i],&ncr);
1827: for (j=0; j<nest->nc; ++j) {
1828: if (aii[i*nest->nc+j]) {
1829: PetscScalar *nvv = avv[i*nest->nc+j];
1830: PetscInt *nii = aii[i*nest->nc+j];
1831: PetscInt *njj = ajj[i*nest->nc+j];
1832: PetscInt ir,cst;
1834: ISStrideGetInfo(nest->isglobal.col[j],&cst,NULL);
1835: for (ir=rst; ir<ncr+rst; ++ir) {
1836: PetscInt ij,rsize = nii[1]-nii[0],ist = ii[ir]+ci[ir];
1838: for (ij=0;ij<rsize;ij++) {
1839: jj[ist+ij] = *njj+cst;
1840: vv[ist+ij] = *nvv;
1841: njj++;
1842: nvv++;
1843: }
1844: ci[ir] += rsize;
1845: nii++;
1846: }
1847: }
1848: }
1849: }
1850: PetscFree(ci);
1852: /* restore info */
1853: for (i=0; i<nest->nr; ++i) {
1854: for (j=0; j<nest->nc; ++j) {
1855: Mat B = nest->m[i][j];
1856: if (B) {
1857: PetscInt nnr = 0, k = i*nest->nc+j;
1859: B = (trans[k] ? trans[k] : B);
1860: MatRestoreRowIJ(B,0,PETSC_FALSE,PETSC_FALSE,&nnr,(const PetscInt**)&aii[k],(const PetscInt**)&ajj[k],&done);
1861: if (!done) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"MatRestoreRowIJ");
1862: MatSeqAIJRestoreArray(B,&avv[k]);
1863: MatDestroy(&trans[k]);
1864: }
1865: }
1866: }
1867: PetscFree4(aii,ajj,avv,trans);
1869: /* finalize newmat */
1870: if (reuse == MAT_INITIAL_MATRIX) {
1871: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),nr,nc,ii,jj,vv,newmat);
1872: } else if (reuse == MAT_INPLACE_MATRIX) {
1873: Mat B;
1875: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),nr,nc,ii,jj,vv,&B);
1876: MatHeaderReplace(A,&B);
1877: }
1878: MatAssemblyBegin(*newmat,MAT_FINAL_ASSEMBLY);
1879: MatAssemblyEnd(*newmat,MAT_FINAL_ASSEMBLY);
1880: {
1881: Mat_SeqAIJ *a = (Mat_SeqAIJ*)((*newmat)->data);
1882: a->free_a = PETSC_TRUE;
1883: a->free_ij = PETSC_TRUE;
1884: }
1885: return(0);
1886: }
1888: PETSC_INTERN PetscErrorCode MatConvert_Nest_AIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat)
1889: {
1891: Mat_Nest *nest = (Mat_Nest*)A->data;
1892: PetscInt m,n,M,N,i,j,k,*dnnz,*onnz,rstart;
1893: PetscInt cstart,cend;
1894: PetscMPIInt size;
1895: Mat C;
1898: MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
1899: if (size == 1) { /* look for a special case with SeqAIJ matrices and strided-1, contiguous, blocks */
1900: PetscInt nf;
1901: PetscBool fast;
1903: PetscStrcmp(newtype,MATAIJ,&fast);
1904: if (!fast) {
1905: PetscStrcmp(newtype,MATSEQAIJ,&fast);
1906: }
1907: for (i=0; i<nest->nr && fast; ++i) {
1908: for (j=0; j<nest->nc && fast; ++j) {
1909: Mat B = nest->m[i][j];
1910: if (B) {
1911: PetscObjectTypeCompare((PetscObject)B,MATSEQAIJ,&fast);
1912: if (!fast) {
1913: PetscBool istrans;
1915: PetscObjectTypeCompare((PetscObject)B,MATTRANSPOSEMAT,&istrans);
1916: if (istrans) {
1917: Mat Bt;
1919: MatTransposeGetMat(B,&Bt);
1920: PetscObjectTypeCompare((PetscObject)Bt,MATSEQAIJ,&fast);
1921: }
1922: }
1923: }
1924: }
1925: }
1926: for (i=0, nf=0; i<nest->nr && fast; ++i) {
1927: PetscObjectTypeCompare((PetscObject)nest->isglobal.row[i],ISSTRIDE,&fast);
1928: if (fast) {
1929: PetscInt f,s;
1931: ISStrideGetInfo(nest->isglobal.row[i],&f,&s);
1932: if (f != nf || s != 1) { fast = PETSC_FALSE; }
1933: else {
1934: ISGetSize(nest->isglobal.row[i],&f);
1935: nf += f;
1936: }
1937: }
1938: }
1939: for (i=0, nf=0; i<nest->nc && fast; ++i) {
1940: PetscObjectTypeCompare((PetscObject)nest->isglobal.col[i],ISSTRIDE,&fast);
1941: if (fast) {
1942: PetscInt f,s;
1944: ISStrideGetInfo(nest->isglobal.col[i],&f,&s);
1945: if (f != nf || s != 1) { fast = PETSC_FALSE; }
1946: else {
1947: ISGetSize(nest->isglobal.col[i],&f);
1948: nf += f;
1949: }
1950: }
1951: }
1952: if (fast) {
1953: MatConvert_Nest_SeqAIJ_fast(A,newtype,reuse,newmat);
1954: return(0);
1955: }
1956: }
1957: MatGetSize(A,&M,&N);
1958: MatGetLocalSize(A,&m,&n);
1959: MatGetOwnershipRangeColumn(A,&cstart,&cend);
1960: switch (reuse) {
1961: case MAT_INITIAL_MATRIX:
1962: MatCreate(PetscObjectComm((PetscObject)A),&C);
1963: MatSetType(C,newtype);
1964: MatSetSizes(C,m,n,M,N);
1965: *newmat = C;
1966: break;
1967: case MAT_REUSE_MATRIX:
1968: C = *newmat;
1969: break;
1970: default: SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"MatReuse");
1971: }
1972: PetscMalloc1(2*m,&dnnz);
1973: onnz = dnnz + m;
1974: for (k=0; k<m; k++) {
1975: dnnz[k] = 0;
1976: onnz[k] = 0;
1977: }
1978: for (j=0; j<nest->nc; ++j) {
1979: IS bNis;
1980: PetscInt bN;
1981: const PetscInt *bNindices;
1982: /* Using global column indices and ISAllGather() is not scalable. */
1983: ISAllGather(nest->isglobal.col[j], &bNis);
1984: ISGetSize(bNis, &bN);
1985: ISGetIndices(bNis,&bNindices);
1986: for (i=0; i<nest->nr; ++i) {
1987: PetscSF bmsf;
1988: PetscSFNode *iremote;
1989: Mat B;
1990: PetscInt bm, *sub_dnnz,*sub_onnz, br;
1991: const PetscInt *bmindices;
1992: B = nest->m[i][j];
1993: if (!B) continue;
1994: ISGetLocalSize(nest->isglobal.row[i],&bm);
1995: ISGetIndices(nest->isglobal.row[i],&bmindices);
1996: PetscSFCreate(PetscObjectComm((PetscObject)A), &bmsf);
1997: PetscMalloc1(bm,&iremote);
1998: PetscMalloc1(bm,&sub_dnnz);
1999: PetscMalloc1(bm,&sub_onnz);
2000: for (k = 0; k < bm; ++k){
2001: sub_dnnz[k] = 0;
2002: sub_onnz[k] = 0;
2003: }
2004: /*
2005: Locate the owners for all of the locally-owned global row indices for this row block.
2006: These determine the roots of PetscSF used to communicate preallocation data to row owners.
2007: The roots correspond to the dnnz and onnz entries; thus, there are two roots per row.
2008: */
2009: MatGetOwnershipRange(B,&rstart,NULL);
2010: for (br = 0; br < bm; ++br) {
2011: PetscInt row = bmindices[br], brncols, col;
2012: const PetscInt *brcols;
2013: PetscInt rowrel = 0; /* row's relative index on its owner rank */
2014: PetscMPIInt rowowner = 0;
2015: PetscLayoutFindOwnerIndex(A->rmap,row,&rowowner,&rowrel);
2016: /* how many roots */
2017: iremote[br].rank = rowowner; iremote[br].index = rowrel; /* edge from bmdnnz to dnnz */
2018: /* get nonzero pattern */
2019: MatGetRow(B,br+rstart,&brncols,&brcols,NULL);
2020: for (k=0; k<brncols; k++) {
2021: col = bNindices[brcols[k]];
2022: if (col>=A->cmap->range[rowowner] && col<A->cmap->range[rowowner+1]) {
2023: sub_dnnz[br]++;
2024: } else {
2025: sub_onnz[br]++;
2026: }
2027: }
2028: MatRestoreRow(B,br+rstart,&brncols,&brcols,NULL);
2029: }
2030: ISRestoreIndices(nest->isglobal.row[i],&bmindices);
2031: /* bsf will have to take care of disposing of bedges. */
2032: PetscSFSetGraph(bmsf,m,bm,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
2033: PetscSFReduceBegin(bmsf,MPIU_INT,sub_dnnz,dnnz,MPI_SUM);
2034: PetscSFReduceEnd(bmsf,MPIU_INT,sub_dnnz,dnnz,MPI_SUM);
2035: PetscSFReduceBegin(bmsf,MPIU_INT,sub_onnz,onnz,MPI_SUM);
2036: PetscSFReduceEnd(bmsf,MPIU_INT,sub_onnz,onnz,MPI_SUM);
2037: PetscFree(sub_dnnz);
2038: PetscFree(sub_onnz);
2039: PetscSFDestroy(&bmsf);
2040: }
2041: ISRestoreIndices(bNis,&bNindices);
2042: ISDestroy(&bNis);
2043: }
2044: /* Resize preallocation if overestimated */
2045: for (i=0;i<m;i++) {
2046: dnnz[i] = PetscMin(dnnz[i],A->cmap->n);
2047: onnz[i] = PetscMin(onnz[i],A->cmap->N - A->cmap->n);
2048: }
2049: MatSeqAIJSetPreallocation(C,0,dnnz);
2050: MatMPIAIJSetPreallocation(C,0,dnnz,0,onnz);
2051: PetscFree(dnnz);
2053: /* Fill by row */
2054: for (j=0; j<nest->nc; ++j) {
2055: /* Using global column indices and ISAllGather() is not scalable. */
2056: IS bNis;
2057: PetscInt bN;
2058: const PetscInt *bNindices;
2059: ISAllGather(nest->isglobal.col[j], &bNis);
2060: ISGetSize(bNis,&bN);
2061: ISGetIndices(bNis,&bNindices);
2062: for (i=0; i<nest->nr; ++i) {
2063: Mat B;
2064: PetscInt bm, br;
2065: const PetscInt *bmindices;
2066: B = nest->m[i][j];
2067: if (!B) continue;
2068: ISGetLocalSize(nest->isglobal.row[i],&bm);
2069: ISGetIndices(nest->isglobal.row[i],&bmindices);
2070: MatGetOwnershipRange(B,&rstart,NULL);
2071: for (br = 0; br < bm; ++br) {
2072: PetscInt row = bmindices[br], brncols, *cols;
2073: const PetscInt *brcols;
2074: const PetscScalar *brcoldata;
2075: MatGetRow(B,br+rstart,&brncols,&brcols,&brcoldata);
2076: PetscMalloc1(brncols,&cols);
2077: for (k=0; k<brncols; k++) cols[k] = bNindices[brcols[k]];
2078: /*
2079: Nest blocks are required to be nonoverlapping -- otherwise nest and monolithic index layouts wouldn't match.
2080: Thus, we could use INSERT_VALUES, but I prefer ADD_VALUES.
2081: */
2082: MatSetValues(C,1,&row,brncols,cols,brcoldata,ADD_VALUES);
2083: MatRestoreRow(B,br+rstart,&brncols,&brcols,&brcoldata);
2084: PetscFree(cols);
2085: }
2086: ISRestoreIndices(nest->isglobal.row[i],&bmindices);
2087: }
2088: ISRestoreIndices(bNis,&bNindices);
2089: ISDestroy(&bNis);
2090: }
2091: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2092: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2093: return(0);
2094: }
2096: PetscErrorCode MatHasOperation_Nest(Mat mat,MatOperation op,PetscBool *has)
2097: {
2098: Mat_Nest *bA = (Mat_Nest*)mat->data;
2099: PetscInt i,j,nr = bA->nr,nc = bA->nc;
2100: PetscBool flg;
2104: *has = PETSC_FALSE;
2105: if (op == MATOP_MULT_TRANSPOSE || op == MATOP_MAT_MULT) {
2106: for (j=0; j<nc; j++) {
2107: for (i=0; i<nr; i++) {
2108: if (!bA->m[i][j]) continue;
2109: MatHasOperation(bA->m[i][j],op,&flg);
2110: if (!flg) return(0);
2111: }
2112: }
2113: }
2114: if (((void**)mat->ops)[op] || (op == MATOP_MAT_MULT && flg)) *has = PETSC_TRUE;
2115: return(0);
2116: }
2118: /*MC
2119: MATNEST - MATNEST = "nest" - Matrix type consisting of nested submatrices, each stored separately.
2121: Level: intermediate
2123: Notes:
2124: This matrix type permits scalable use of PCFieldSplit and avoids the large memory costs of extracting submatrices.
2125: It allows the use of symmetric and block formats for parts of multi-physics simulations.
2126: It is usually used with DMComposite and DMCreateMatrix()
2128: Each of the submatrices lives on the same MPI communicator as the original nest matrix (though they can have zero
2129: rows/columns on some processes.) Thus this is not meant for cases where the submatrices live on far fewer processes
2130: than the nest matrix.
2132: .seealso: MatCreate(), MatType, MatCreateNest(), MatNestSetSubMat(), MatNestGetSubMat(),
2133: VecCreateNest(), DMCreateMatrix(), DMCOMPOSITE, MatNestSetVecType(), MatNestGetLocalISs(),
2134: MatNestGetISs(), MatNestSetSubMats(), MatNestGetSubMats()
2135: M*/
2136: PETSC_EXTERN PetscErrorCode MatCreate_Nest(Mat A)
2137: {
2138: Mat_Nest *s;
2142: PetscNewLog(A,&s);
2143: A->data = (void*)s;
2145: s->nr = -1;
2146: s->nc = -1;
2147: s->m = NULL;
2148: s->splitassembly = PETSC_FALSE;
2150: PetscMemzero(A->ops,sizeof(*A->ops));
2152: A->ops->mult = MatMult_Nest;
2153: A->ops->multadd = MatMultAdd_Nest;
2154: A->ops->multtranspose = MatMultTranspose_Nest;
2155: A->ops->multtransposeadd = MatMultTransposeAdd_Nest;
2156: A->ops->transpose = MatTranspose_Nest;
2157: A->ops->assemblybegin = MatAssemblyBegin_Nest;
2158: A->ops->assemblyend = MatAssemblyEnd_Nest;
2159: A->ops->zeroentries = MatZeroEntries_Nest;
2160: A->ops->copy = MatCopy_Nest;
2161: A->ops->axpy = MatAXPY_Nest;
2162: A->ops->duplicate = MatDuplicate_Nest;
2163: A->ops->createsubmatrix = MatCreateSubMatrix_Nest;
2164: A->ops->destroy = MatDestroy_Nest;
2165: A->ops->view = MatView_Nest;
2166: A->ops->getvecs = 0; /* Use VECNEST by calling MatNestSetVecType(A,VECNEST) */
2167: A->ops->getlocalsubmatrix = MatGetLocalSubMatrix_Nest;
2168: A->ops->restorelocalsubmatrix = MatRestoreLocalSubMatrix_Nest;
2169: A->ops->getdiagonal = MatGetDiagonal_Nest;
2170: A->ops->diagonalscale = MatDiagonalScale_Nest;
2171: A->ops->scale = MatScale_Nest;
2172: A->ops->shift = MatShift_Nest;
2173: A->ops->diagonalset = MatDiagonalSet_Nest;
2174: A->ops->setrandom = MatSetRandom_Nest;
2175: A->ops->hasoperation = MatHasOperation_Nest;
2176: A->ops->missingdiagonal = MatMissingDiagonal_Nest;
2178: A->spptr = 0;
2179: A->assembled = PETSC_FALSE;
2181: /* expose Nest api's */
2182: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSubMat_C", MatNestGetSubMat_Nest);
2183: PetscObjectComposeFunction((PetscObject)A,"MatNestSetSubMat_C", MatNestSetSubMat_Nest);
2184: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSubMats_C", MatNestGetSubMats_Nest);
2185: PetscObjectComposeFunction((PetscObject)A,"MatNestGetSize_C", MatNestGetSize_Nest);
2186: PetscObjectComposeFunction((PetscObject)A,"MatNestGetISs_C", MatNestGetISs_Nest);
2187: PetscObjectComposeFunction((PetscObject)A,"MatNestGetLocalISs_C", MatNestGetLocalISs_Nest);
2188: PetscObjectComposeFunction((PetscObject)A,"MatNestSetVecType_C", MatNestSetVecType_Nest);
2189: PetscObjectComposeFunction((PetscObject)A,"MatNestSetSubMats_C", MatNestSetSubMats_Nest);
2190: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_mpiaij_C", MatConvert_Nest_AIJ);
2191: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_seqaij_C", MatConvert_Nest_AIJ);
2192: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_aij_C", MatConvert_Nest_AIJ);
2193: PetscObjectComposeFunction((PetscObject)A,"MatConvert_nest_is_C", MatConvert_Nest_IS);
2194: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_seqdense_C",MatProductSetFromOptions_Nest_Dense);
2195: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_mpidense_C",MatProductSetFromOptions_Nest_Dense);
2196: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_nest_dense_C",MatProductSetFromOptions_Nest_Dense);
2198: PetscObjectChangeTypeName((PetscObject)A,MATNEST);
2199: return(0);
2200: }