Actual source code: mpiaij.c

  1: #include <../src/mat/impls/aij/mpi/mpiaij.h>
  2: #include <petsc/private/vecimpl.h>
  3: #include <petsc/private/sfimpl.h>
  4: #include <petsc/private/isimpl.h>
  5: #include <petscblaslapack.h>
  6: #include <petscsf.h>
  7: #include <petsc/private/hashmapi.h>

  9: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
 10: {
 11:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

 13:   PetscFunctionBegin;
 14: #if defined(PETSC_USE_LOG)
 15:   PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
 16: #endif
 17:   PetscCall(MatStashDestroy_Private(&mat->stash));
 18:   PetscCall(VecDestroy(&aij->diag));
 19:   PetscCall(MatDestroy(&aij->A));
 20:   PetscCall(MatDestroy(&aij->B));
 21: #if defined(PETSC_USE_CTABLE)
 22:   PetscCall(PetscHMapIDestroy(&aij->colmap));
 23: #else
 24:   PetscCall(PetscFree(aij->colmap));
 25: #endif
 26:   PetscCall(PetscFree(aij->garray));
 27:   PetscCall(VecDestroy(&aij->lvec));
 28:   PetscCall(VecScatterDestroy(&aij->Mvctx));
 29:   PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
 30:   PetscCall(PetscFree(aij->ld));

 32:   /* Free COO */
 33:   PetscCall(MatResetPreallocationCOO_MPIAIJ(mat));

 35:   PetscCall(PetscFree(mat->data));

 37:   /* may be created by MatCreateMPIAIJSumSeqAIJSymbolic */
 38:   PetscCall(PetscObjectCompose((PetscObject)mat, "MatMergeSeqsToMPI", NULL));

 40:   PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
 41:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatStoreValues_C", NULL));
 42:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatRetrieveValues_C", NULL));
 43:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatIsTranspose_C", NULL));
 44:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocation_C", NULL));
 45:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetPreallocation_C", NULL));
 46:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocationCSR_C", NULL));
 47:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
 48:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpibaij_C", NULL));
 49:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisbaij_C", NULL));
 50: #if defined(PETSC_HAVE_CUDA)
 51:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcusparse_C", NULL));
 52: #endif
 53: #if defined(PETSC_HAVE_HIP)
 54:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijhipsparse_C", NULL));
 55: #endif
 56: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
 57:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijkokkos_C", NULL));
 58: #endif
 59:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
 60: #if defined(PETSC_HAVE_ELEMENTAL)
 61:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_elemental_C", NULL));
 62: #endif
 63: #if defined(PETSC_HAVE_SCALAPACK)
 64:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_scalapack_C", NULL));
 65: #endif
 66: #if defined(PETSC_HAVE_HYPRE)
 67:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_hypre_C", NULL));
 68:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", NULL));
 69: #endif
 70:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
 71:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_is_mpiaij_C", NULL));
 72:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpiaij_C", NULL));
 73:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetUseScalableIncreaseOverlap_C", NULL));
 74:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijperm_C", NULL));
 75:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijsell_C", NULL));
 76: #if defined(PETSC_HAVE_MKL_SPARSE)
 77:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijmkl_C", NULL));
 78: #endif
 79:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcrl_C", NULL));
 80:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
 81:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisell_C", NULL));
 82:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetPreallocationCOO_C", NULL));
 83:   PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetValuesCOO_C", NULL));
 84:   PetscFunctionReturn(PETSC_SUCCESS);
 85: }

 87: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and  MatAssemblyEnd_MPI_Hash() */
 88: #define TYPE AIJ
 89: #define TYPE_AIJ
 90: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
 91: #undef TYPE
 92: #undef TYPE_AIJ

 94: PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
 95: {
 96:   Mat B;

 98:   PetscFunctionBegin;
 99:   PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
100:   PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
101:   PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
102:   PetscCall(MatDestroy(&B));
103:   PetscFunctionReturn(PETSC_SUCCESS);
104: }

106: PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
107: {
108:   Mat B;

110:   PetscFunctionBegin;
111:   PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
112:   PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
113:   PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
114:   PetscFunctionReturn(PETSC_SUCCESS);
115: }

117: /*MC
118:    MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.

120:    This matrix type is identical to` MATSEQAIJ` when constructed with a single process communicator,
121:    and `MATMPIAIJ` otherwise.  As a result, for single process communicators,
122:   `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
123:   for communicators controlling multiple processes.  It is recommended that you call both of
124:   the above preallocation routines for simplicity.

126:    Options Database Key:
127: . -mat_type aij - sets the matrix type to `MATAIJ` during a call to `MatSetFromOptions()`

129:   Developer Note:
130:   Level: beginner

132:     Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, `MATAIJKOKKOS`,and also automatically switches over to use inodes when
133:    enough exist.

135: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`
136: M*/

138: /*MC
139:    MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.

141:    This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
142:    and `MATMPIAIJCRL` otherwise.  As a result, for single process communicators,
143:    `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
144:   for communicators controlling multiple processes.  It is recommended that you call both of
145:   the above preallocation routines for simplicity.

147:    Options Database Key:
148: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`

150:   Level: beginner

152: .seealso: [](chapter_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
153: M*/

155: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
156: {
157:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

159:   PetscFunctionBegin;
160: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_VIENNACL)
161:   A->boundtocpu = flg;
162: #endif
163:   if (a->A) PetscCall(MatBindToCPU(a->A, flg));
164:   if (a->B) PetscCall(MatBindToCPU(a->B, flg));

166:   /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
167:    * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
168:    * to differ from the parent matrix. */
169:   if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
170:   if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));

172:   PetscFunctionReturn(PETSC_SUCCESS);
173: }

175: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
176: {
177:   Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;

179:   PetscFunctionBegin;
180:   if (mat->A) {
181:     PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
182:     PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
183:   }
184:   PetscFunctionReturn(PETSC_SUCCESS);
185: }

187: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
188: {
189:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ *)M->data;
190:   Mat_SeqAIJ      *a   = (Mat_SeqAIJ *)mat->A->data;
191:   Mat_SeqAIJ      *b   = (Mat_SeqAIJ *)mat->B->data;
192:   const PetscInt  *ia, *ib;
193:   const MatScalar *aa, *bb, *aav, *bav;
194:   PetscInt         na, nb, i, j, *rows, cnt = 0, n0rows;
195:   PetscInt         m = M->rmap->n, rstart = M->rmap->rstart;

197:   PetscFunctionBegin;
198:   *keptrows = NULL;

200:   ia = a->i;
201:   ib = b->i;
202:   PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
203:   PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
204:   for (i = 0; i < m; i++) {
205:     na = ia[i + 1] - ia[i];
206:     nb = ib[i + 1] - ib[i];
207:     if (!na && !nb) {
208:       cnt++;
209:       goto ok1;
210:     }
211:     aa = aav + ia[i];
212:     for (j = 0; j < na; j++) {
213:       if (aa[j] != 0.0) goto ok1;
214:     }
215:     bb = bav + ib[i];
216:     for (j = 0; j < nb; j++) {
217:       if (bb[j] != 0.0) goto ok1;
218:     }
219:     cnt++;
220:   ok1:;
221:   }
222:   PetscCall(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
223:   if (!n0rows) {
224:     PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
225:     PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
226:     PetscFunctionReturn(PETSC_SUCCESS);
227:   }
228:   PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
229:   cnt = 0;
230:   for (i = 0; i < m; i++) {
231:     na = ia[i + 1] - ia[i];
232:     nb = ib[i + 1] - ib[i];
233:     if (!na && !nb) continue;
234:     aa = aav + ia[i];
235:     for (j = 0; j < na; j++) {
236:       if (aa[j] != 0.0) {
237:         rows[cnt++] = rstart + i;
238:         goto ok2;
239:       }
240:     }
241:     bb = bav + ib[i];
242:     for (j = 0; j < nb; j++) {
243:       if (bb[j] != 0.0) {
244:         rows[cnt++] = rstart + i;
245:         goto ok2;
246:       }
247:     }
248:   ok2:;
249:   }
250:   PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
251:   PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
252:   PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
253:   PetscFunctionReturn(PETSC_SUCCESS);
254: }

256: PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
257: {
258:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
259:   PetscBool   cong;

261:   PetscFunctionBegin;
262:   PetscCall(MatHasCongruentLayouts(Y, &cong));
263:   if (Y->assembled && cong) {
264:     PetscCall(MatDiagonalSet(aij->A, D, is));
265:   } else {
266:     PetscCall(MatDiagonalSet_Default(Y, D, is));
267:   }
268:   PetscFunctionReturn(PETSC_SUCCESS);
269: }

271: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
272: {
273:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
274:   PetscInt    i, rstart, nrows, *rows;

276:   PetscFunctionBegin;
277:   *zrows = NULL;
278:   PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
279:   PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
280:   for (i = 0; i < nrows; i++) rows[i] += rstart;
281:   PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
282:   PetscFunctionReturn(PETSC_SUCCESS);
283: }

285: PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
286: {
287:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ *)A->data;
288:   PetscInt           i, m, n, *garray = aij->garray;
289:   Mat_SeqAIJ        *a_aij = (Mat_SeqAIJ *)aij->A->data;
290:   Mat_SeqAIJ        *b_aij = (Mat_SeqAIJ *)aij->B->data;
291:   PetscReal         *work;
292:   const PetscScalar *dummy;

294:   PetscFunctionBegin;
295:   PetscCall(MatGetSize(A, &m, &n));
296:   PetscCall(PetscCalloc1(n, &work));
297:   PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
298:   PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
299:   PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
300:   PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
301:   if (type == NORM_2) {
302:     for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
303:     for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
304:   } else if (type == NORM_1) {
305:     for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
306:     for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
307:   } else if (type == NORM_INFINITY) {
308:     for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
309:     for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), work[garray[b_aij->j[i]]]);
310:   } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
311:     for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
312:     for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
313:   } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
314:     for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
315:     for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
316:   } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
317:   if (type == NORM_INFINITY) {
318:     PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
319:   } else {
320:     PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
321:   }
322:   PetscCall(PetscFree(work));
323:   if (type == NORM_2) {
324:     for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
325:   } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
326:     for (i = 0; i < n; i++) reductions[i] /= m;
327:   }
328:   PetscFunctionReturn(PETSC_SUCCESS);
329: }

331: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
332: {
333:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)A->data;
334:   IS              sis, gis;
335:   const PetscInt *isis, *igis;
336:   PetscInt        n, *iis, nsis, ngis, rstart, i;

338:   PetscFunctionBegin;
339:   PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
340:   PetscCall(MatFindNonzeroRows(a->B, &gis));
341:   PetscCall(ISGetSize(gis, &ngis));
342:   PetscCall(ISGetSize(sis, &nsis));
343:   PetscCall(ISGetIndices(sis, &isis));
344:   PetscCall(ISGetIndices(gis, &igis));

346:   PetscCall(PetscMalloc1(ngis + nsis, &iis));
347:   PetscCall(PetscArraycpy(iis, igis, ngis));
348:   PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
349:   n = ngis + nsis;
350:   PetscCall(PetscSortRemoveDupsInt(&n, iis));
351:   PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
352:   for (i = 0; i < n; i++) iis[i] += rstart;
353:   PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));

355:   PetscCall(ISRestoreIndices(sis, &isis));
356:   PetscCall(ISRestoreIndices(gis, &igis));
357:   PetscCall(ISDestroy(&sis));
358:   PetscCall(ISDestroy(&gis));
359:   PetscFunctionReturn(PETSC_SUCCESS);
360: }

362: /*
363:   Local utility routine that creates a mapping from the global column
364: number to the local number in the off-diagonal part of the local
365: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at
366: a slightly higher hash table cost; without it it is not scalable (each processor
367: has an order N integer array but is fast to access.
368: */
369: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
370: {
371:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
372:   PetscInt    n   = aij->B->cmap->n, i;

374:   PetscFunctionBegin;
375:   PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
376: #if defined(PETSC_USE_CTABLE)
377:   PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
378:   for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
379: #else
380:   PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
381:   for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
382: #endif
383:   PetscFunctionReturn(PETSC_SUCCESS);
384: }

386: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
387:   { \
388:     if (col <= lastcol1) low1 = 0; \
389:     else high1 = nrow1; \
390:     lastcol1 = col; \
391:     while (high1 - low1 > 5) { \
392:       t = (low1 + high1) / 2; \
393:       if (rp1[t] > col) high1 = t; \
394:       else low1 = t; \
395:     } \
396:     for (_i = low1; _i < high1; _i++) { \
397:       if (rp1[_i] > col) break; \
398:       if (rp1[_i] == col) { \
399:         if (addv == ADD_VALUES) { \
400:           ap1[_i] += value; \
401:           /* Not sure LogFlops will slow dow the code or not */ \
402:           (void)PetscLogFlops(1.0); \
403:         } else ap1[_i] = value; \
404:         goto a_noinsert; \
405:       } \
406:     } \
407:     if (value == 0.0 && ignorezeroentries && row != col) { \
408:       low1  = 0; \
409:       high1 = nrow1; \
410:       goto a_noinsert; \
411:     } \
412:     if (nonew == 1) { \
413:       low1  = 0; \
414:       high1 = nrow1; \
415:       goto a_noinsert; \
416:     } \
417:     PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
418:     MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
419:     N = nrow1++ - 1; \
420:     a->nz++; \
421:     high1++; \
422:     /* shift up all the later entries in this row */ \
423:     PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
424:     PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
425:     rp1[_i] = col; \
426:     ap1[_i] = value; \
427:     A->nonzerostate++; \
428:   a_noinsert:; \
429:     ailen[row] = nrow1; \
430:   }

432: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
433:   { \
434:     if (col <= lastcol2) low2 = 0; \
435:     else high2 = nrow2; \
436:     lastcol2 = col; \
437:     while (high2 - low2 > 5) { \
438:       t = (low2 + high2) / 2; \
439:       if (rp2[t] > col) high2 = t; \
440:       else low2 = t; \
441:     } \
442:     for (_i = low2; _i < high2; _i++) { \
443:       if (rp2[_i] > col) break; \
444:       if (rp2[_i] == col) { \
445:         if (addv == ADD_VALUES) { \
446:           ap2[_i] += value; \
447:           (void)PetscLogFlops(1.0); \
448:         } else ap2[_i] = value; \
449:         goto b_noinsert; \
450:       } \
451:     } \
452:     if (value == 0.0 && ignorezeroentries) { \
453:       low2  = 0; \
454:       high2 = nrow2; \
455:       goto b_noinsert; \
456:     } \
457:     if (nonew == 1) { \
458:       low2  = 0; \
459:       high2 = nrow2; \
460:       goto b_noinsert; \
461:     } \
462:     PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
463:     MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
464:     N = nrow2++ - 1; \
465:     b->nz++; \
466:     high2++; \
467:     /* shift up all the later entries in this row */ \
468:     PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
469:     PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
470:     rp2[_i] = col; \
471:     ap2[_i] = value; \
472:     B->nonzerostate++; \
473:   b_noinsert:; \
474:     bilen[row] = nrow2; \
475:   }

477: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
478: {
479:   Mat_MPIAIJ  *mat = (Mat_MPIAIJ *)A->data;
480:   Mat_SeqAIJ  *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
481:   PetscInt     l, *garray                         = mat->garray, diag;
482:   PetscScalar *aa, *ba;

484:   PetscFunctionBegin;
485:   /* code only works for square matrices A */

487:   /* find size of row to the left of the diagonal part */
488:   PetscCall(MatGetOwnershipRange(A, &diag, NULL));
489:   row = row - diag;
490:   for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
491:     if (garray[b->j[b->i[row] + l]] > diag) break;
492:   }
493:   if (l) {
494:     PetscCall(MatSeqAIJGetArray(mat->B, &ba));
495:     PetscCall(PetscArraycpy(ba + b->i[row], v, l));
496:     PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
497:   }

499:   /* diagonal part */
500:   if (a->i[row + 1] - a->i[row]) {
501:     PetscCall(MatSeqAIJGetArray(mat->A, &aa));
502:     PetscCall(PetscArraycpy(aa + a->i[row], v + l, (a->i[row + 1] - a->i[row])));
503:     PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
504:   }

506:   /* right of diagonal part */
507:   if (b->i[row + 1] - b->i[row] - l) {
508:     PetscCall(MatSeqAIJGetArray(mat->B, &ba));
509:     PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
510:     PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
511:   }
512:   PetscFunctionReturn(PETSC_SUCCESS);
513: }

515: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
516: {
517:   Mat_MPIAIJ *aij   = (Mat_MPIAIJ *)mat->data;
518:   PetscScalar value = 0.0;
519:   PetscInt    i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
520:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
521:   PetscBool   roworiented = aij->roworiented;

523:   /* Some Variables required in the macro */
524:   Mat         A     = aij->A;
525:   Mat_SeqAIJ *a     = (Mat_SeqAIJ *)A->data;
526:   PetscInt   *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
527:   PetscBool   ignorezeroentries = a->ignorezeroentries;
528:   Mat         B                 = aij->B;
529:   Mat_SeqAIJ *b                 = (Mat_SeqAIJ *)B->data;
530:   PetscInt   *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
531:   MatScalar  *aa, *ba;
532:   PetscInt   *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
533:   PetscInt    nonew;
534:   MatScalar  *ap1, *ap2;

536:   PetscFunctionBegin;
537:   PetscCall(MatSeqAIJGetArray(A, &aa));
538:   PetscCall(MatSeqAIJGetArray(B, &ba));
539:   for (i = 0; i < m; i++) {
540:     if (im[i] < 0) continue;
541:     PetscCheck(im[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], mat->rmap->N - 1);
542:     if (im[i] >= rstart && im[i] < rend) {
543:       row      = im[i] - rstart;
544:       lastcol1 = -1;
545:       rp1      = aj + ai[row];
546:       ap1      = aa + ai[row];
547:       rmax1    = aimax[row];
548:       nrow1    = ailen[row];
549:       low1     = 0;
550:       high1    = nrow1;
551:       lastcol2 = -1;
552:       rp2      = bj + bi[row];
553:       ap2      = ba + bi[row];
554:       rmax2    = bimax[row];
555:       nrow2    = bilen[row];
556:       low2     = 0;
557:       high2    = nrow2;

559:       for (j = 0; j < n; j++) {
560:         if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
561:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
562:         if (in[j] >= cstart && in[j] < cend) {
563:           col   = in[j] - cstart;
564:           nonew = a->nonew;
565:           MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
566:         } else if (in[j] < 0) {
567:           continue;
568:         } else {
569:           PetscCheck(in[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
570:           if (mat->was_assembled) {
571:             if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
572: #if defined(PETSC_USE_CTABLE)
573:             PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
574:             col--;
575: #else
576:             col = aij->colmap[in[j]] - 1;
577: #endif
578:             if (col < 0 && !((Mat_SeqAIJ *)(aij->B->data))->nonew) { /* col < 0 means in[j] is a new col for B */
579:               PetscCall(MatDisAssemble_MPIAIJ(mat));                 /* Change aij->B from reduced/local format to expanded/global format */
580:               col = in[j];
581:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
582:               B     = aij->B;
583:               b     = (Mat_SeqAIJ *)B->data;
584:               bimax = b->imax;
585:               bi    = b->i;
586:               bilen = b->ilen;
587:               bj    = b->j;
588:               ba    = b->a;
589:               rp2   = bj + bi[row];
590:               ap2   = ba + bi[row];
591:               rmax2 = bimax[row];
592:               nrow2 = bilen[row];
593:               low2  = 0;
594:               high2 = nrow2;
595:               bm    = aij->B->rmap->n;
596:               ba    = b->a;
597:             } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
598:               if (1 == ((Mat_SeqAIJ *)(aij->B->data))->nonew) {
599:                 PetscCall(PetscInfo(mat, "Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%" PetscInt_FMT ",%" PetscInt_FMT ")\n", (double)PetscRealPart(value), im[i], in[j]));
600:               } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", im[i], in[j]);
601:             }
602:           } else col = in[j];
603:           nonew = b->nonew;
604:           MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
605:         }
606:       }
607:     } else {
608:       PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Setting off process row %" PetscInt_FMT " even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set", im[i]);
609:       if (!aij->donotstash) {
610:         mat->assembled = PETSC_FALSE;
611:         if (roworiented) {
612:           PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
613:         } else {
614:           PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
615:         }
616:       }
617:     }
618:   }
619:   PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
620:   PetscCall(MatSeqAIJRestoreArray(B, &ba));
621:   PetscFunctionReturn(PETSC_SUCCESS);
622: }

624: /*
625:     This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
626:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
627:     No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
628: */
629: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
630: {
631:   Mat_MPIAIJ *aij    = (Mat_MPIAIJ *)mat->data;
632:   Mat         A      = aij->A; /* diagonal part of the matrix */
633:   Mat         B      = aij->B; /* offdiagonal part of the matrix */
634:   Mat_SeqAIJ *a      = (Mat_SeqAIJ *)A->data;
635:   Mat_SeqAIJ *b      = (Mat_SeqAIJ *)B->data;
636:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
637:   PetscInt   *ailen = a->ilen, *aj = a->j;
638:   PetscInt   *bilen = b->ilen, *bj = b->j;
639:   PetscInt    am          = aij->A->rmap->n, j;
640:   PetscInt    diag_so_far = 0, dnz;
641:   PetscInt    offd_so_far = 0, onz;

643:   PetscFunctionBegin;
644:   /* Iterate over all rows of the matrix */
645:   for (j = 0; j < am; j++) {
646:     dnz = onz = 0;
647:     /*  Iterate over all non-zero columns of the current row */
648:     for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
649:       /* If column is in the diagonal */
650:       if (mat_j[col] >= cstart && mat_j[col] < cend) {
651:         aj[diag_so_far++] = mat_j[col] - cstart;
652:         dnz++;
653:       } else { /* off-diagonal entries */
654:         bj[offd_so_far++] = mat_j[col];
655:         onz++;
656:       }
657:     }
658:     ailen[j] = dnz;
659:     bilen[j] = onz;
660:   }
661:   PetscFunctionReturn(PETSC_SUCCESS);
662: }

664: /*
665:     This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
666:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
667:     No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
668:     Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
669:     would not be true and the more complex MatSetValues_MPIAIJ has to be used.
670: */
671: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
672: {
673:   Mat_MPIAIJ  *aij  = (Mat_MPIAIJ *)mat->data;
674:   Mat          A    = aij->A; /* diagonal part of the matrix */
675:   Mat          B    = aij->B; /* offdiagonal part of the matrix */
676:   Mat_SeqAIJ  *aijd = (Mat_SeqAIJ *)(aij->A)->data, *aijo = (Mat_SeqAIJ *)(aij->B)->data;
677:   Mat_SeqAIJ  *a      = (Mat_SeqAIJ *)A->data;
678:   Mat_SeqAIJ  *b      = (Mat_SeqAIJ *)B->data;
679:   PetscInt     cstart = mat->cmap->rstart, cend = mat->cmap->rend;
680:   PetscInt    *ailen = a->ilen, *aj = a->j;
681:   PetscInt    *bilen = b->ilen, *bj = b->j;
682:   PetscInt     am          = aij->A->rmap->n, j;
683:   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. */
684:   PetscInt     col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
685:   PetscScalar *aa = a->a, *ba = b->a;

687:   PetscFunctionBegin;
688:   /* Iterate over all rows of the matrix */
689:   for (j = 0; j < am; j++) {
690:     dnz_row = onz_row = 0;
691:     rowstart_offd     = full_offd_i[j];
692:     rowstart_diag     = full_diag_i[j];
693:     /*  Iterate over all non-zero columns of the current row */
694:     for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
695:       /* If column is in the diagonal */
696:       if (mat_j[col] >= cstart && mat_j[col] < cend) {
697:         aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
698:         aa[rowstart_diag + dnz_row] = mat_a[col];
699:         dnz_row++;
700:       } else { /* off-diagonal entries */
701:         bj[rowstart_offd + onz_row] = mat_j[col];
702:         ba[rowstart_offd + onz_row] = mat_a[col];
703:         onz_row++;
704:       }
705:     }
706:     ailen[j] = dnz_row;
707:     bilen[j] = onz_row;
708:   }
709:   PetscFunctionReturn(PETSC_SUCCESS);
710: }

712: PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
713: {
714:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
715:   PetscInt    i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
716:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;

718:   PetscFunctionBegin;
719:   for (i = 0; i < m; i++) {
720:     if (idxm[i] < 0) continue; /* negative row */
721:     PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, idxm[i], mat->rmap->N - 1);
722:     if (idxm[i] >= rstart && idxm[i] < rend) {
723:       row = idxm[i] - rstart;
724:       for (j = 0; j < n; j++) {
725:         if (idxn[j] < 0) continue; /* negative column */
726:         PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, idxn[j], mat->cmap->N - 1);
727:         if (idxn[j] >= cstart && idxn[j] < cend) {
728:           col = idxn[j] - cstart;
729:           PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
730:         } else {
731:           if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
732: #if defined(PETSC_USE_CTABLE)
733:           PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
734:           col--;
735: #else
736:           col = aij->colmap[idxn[j]] - 1;
737: #endif
738:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
739:           else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
740:         }
741:       }
742:     } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported");
743:   }
744:   PetscFunctionReturn(PETSC_SUCCESS);
745: }

747: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
748: {
749:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
750:   PetscInt    nstash, reallocs;

752:   PetscFunctionBegin;
753:   if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);

755:   PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
756:   PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
757:   PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
758:   PetscFunctionReturn(PETSC_SUCCESS);
759: }

761: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
762: {
763:   Mat_MPIAIJ  *aij = (Mat_MPIAIJ *)mat->data;
764:   PetscMPIInt  n;
765:   PetscInt     i, j, rstart, ncols, flg;
766:   PetscInt    *row, *col;
767:   PetscBool    other_disassembled;
768:   PetscScalar *val;

770:   /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */

772:   PetscFunctionBegin;
773:   if (!aij->donotstash && !mat->nooffprocentries) {
774:     while (1) {
775:       PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
776:       if (!flg) break;

778:       for (i = 0; i < n;) {
779:         /* Now identify the consecutive vals belonging to the same row */
780:         for (j = i, rstart = row[j]; j < n; j++) {
781:           if (row[j] != rstart) break;
782:         }
783:         if (j < n) ncols = j - i;
784:         else ncols = n - i;
785:         /* Now assemble all these values with a single function call */
786:         PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
787:         i = j;
788:       }
789:     }
790:     PetscCall(MatStashScatterEnd_Private(&mat->stash));
791:   }
792: #if defined(PETSC_HAVE_DEVICE)
793:   if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
794:   /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
795:   if (mat->boundtocpu) {
796:     PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
797:     PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
798:   }
799: #endif
800:   PetscCall(MatAssemblyBegin(aij->A, mode));
801:   PetscCall(MatAssemblyEnd(aij->A, mode));

803:   /* determine if any processor has disassembled, if so we must
804:      also disassemble ourself, in order that we may reassemble. */
805:   /*
806:      if nonzero structure of submatrix B cannot change then we know that
807:      no processor disassembled thus we can skip this stuff
808:   */
809:   if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
810:     PetscCall(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
811:     if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
812:       PetscCall(MatDisAssemble_MPIAIJ(mat));
813:     }
814:   }
815:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
816:   PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
817: #if defined(PETSC_HAVE_DEVICE)
818:   if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
819: #endif
820:   PetscCall(MatAssemblyBegin(aij->B, mode));
821:   PetscCall(MatAssemblyEnd(aij->B, mode));

823:   PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));

825:   aij->rowvalues = NULL;

827:   PetscCall(VecDestroy(&aij->diag));

829:   /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
830:   if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)(aij->A->data))->nonew) {
831:     PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
832:     PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
833:   }
834: #if defined(PETSC_HAVE_DEVICE)
835:   mat->offloadmask = PETSC_OFFLOAD_BOTH;
836: #endif
837:   PetscFunctionReturn(PETSC_SUCCESS);
838: }

840: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
841: {
842:   Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;

844:   PetscFunctionBegin;
845:   PetscCall(MatZeroEntries(l->A));
846:   PetscCall(MatZeroEntries(l->B));
847:   PetscFunctionReturn(PETSC_SUCCESS);
848: }

850: PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
851: {
852:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ *)A->data;
853:   PetscObjectState sA, sB;
854:   PetscInt        *lrows;
855:   PetscInt         r, len;
856:   PetscBool        cong, lch, gch;

858:   PetscFunctionBegin;
859:   /* get locally owned rows */
860:   PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
861:   PetscCall(MatHasCongruentLayouts(A, &cong));
862:   /* fix right hand side if needed */
863:   if (x && b) {
864:     const PetscScalar *xx;
865:     PetscScalar       *bb;

867:     PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
868:     PetscCall(VecGetArrayRead(x, &xx));
869:     PetscCall(VecGetArray(b, &bb));
870:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
871:     PetscCall(VecRestoreArrayRead(x, &xx));
872:     PetscCall(VecRestoreArray(b, &bb));
873:   }

875:   sA = mat->A->nonzerostate;
876:   sB = mat->B->nonzerostate;

878:   if (diag != 0.0 && cong) {
879:     PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
880:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
881:   } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
882:     Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
883:     Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
884:     PetscInt    nnwA, nnwB;
885:     PetscBool   nnzA, nnzB;

887:     nnwA = aijA->nonew;
888:     nnwB = aijB->nonew;
889:     nnzA = aijA->keepnonzeropattern;
890:     nnzB = aijB->keepnonzeropattern;
891:     if (!nnzA) {
892:       PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
893:       aijA->nonew = 0;
894:     }
895:     if (!nnzB) {
896:       PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
897:       aijB->nonew = 0;
898:     }
899:     /* Must zero here before the next loop */
900:     PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
901:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
902:     for (r = 0; r < len; ++r) {
903:       const PetscInt row = lrows[r] + A->rmap->rstart;
904:       if (row >= A->cmap->N) continue;
905:       PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
906:     }
907:     aijA->nonew = nnwA;
908:     aijB->nonew = nnwB;
909:   } else {
910:     PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
911:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
912:   }
913:   PetscCall(PetscFree(lrows));
914:   PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
915:   PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));

917:   /* reduce nonzerostate */
918:   lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
919:   PetscCall(MPIU_Allreduce(&lch, &gch, 1, MPIU_BOOL, MPI_LOR, PetscObjectComm((PetscObject)A)));
920:   if (gch) A->nonzerostate++;
921:   PetscFunctionReturn(PETSC_SUCCESS);
922: }

924: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
925: {
926:   Mat_MPIAIJ        *l = (Mat_MPIAIJ *)A->data;
927:   PetscMPIInt        n = A->rmap->n;
928:   PetscInt           i, j, r, m, len = 0;
929:   PetscInt          *lrows, *owners = A->rmap->range;
930:   PetscMPIInt        p = 0;
931:   PetscSFNode       *rrows;
932:   PetscSF            sf;
933:   const PetscScalar *xx;
934:   PetscScalar       *bb, *mask, *aij_a;
935:   Vec                xmask, lmask;
936:   Mat_SeqAIJ        *aij = (Mat_SeqAIJ *)l->B->data;
937:   const PetscInt    *aj, *ii, *ridx;
938:   PetscScalar       *aa;

940:   PetscFunctionBegin;
941:   /* Create SF where leaves are input rows and roots are owned rows */
942:   PetscCall(PetscMalloc1(n, &lrows));
943:   for (r = 0; r < n; ++r) lrows[r] = -1;
944:   PetscCall(PetscMalloc1(N, &rrows));
945:   for (r = 0; r < N; ++r) {
946:     const PetscInt idx = rows[r];
947:     PetscCheck(idx >= 0 && A->rmap->N > idx, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " out of range [0,%" PetscInt_FMT ")", idx, A->rmap->N);
948:     if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
949:       PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
950:     }
951:     rrows[r].rank  = p;
952:     rrows[r].index = rows[r] - owners[p];
953:   }
954:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
955:   PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
956:   /* Collect flags for rows to be zeroed */
957:   PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
958:   PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
959:   PetscCall(PetscSFDestroy(&sf));
960:   /* Compress and put in row numbers */
961:   for (r = 0; r < n; ++r)
962:     if (lrows[r] >= 0) lrows[len++] = r;
963:   /* zero diagonal part of matrix */
964:   PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
965:   /* handle off diagonal part of matrix */
966:   PetscCall(MatCreateVecs(A, &xmask, NULL));
967:   PetscCall(VecDuplicate(l->lvec, &lmask));
968:   PetscCall(VecGetArray(xmask, &bb));
969:   for (i = 0; i < len; i++) bb[lrows[i]] = 1;
970:   PetscCall(VecRestoreArray(xmask, &bb));
971:   PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
972:   PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
973:   PetscCall(VecDestroy(&xmask));
974:   if (x && b) { /* this code is buggy when the row and column layout don't match */
975:     PetscBool cong;

977:     PetscCall(MatHasCongruentLayouts(A, &cong));
978:     PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
979:     PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
980:     PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
981:     PetscCall(VecGetArrayRead(l->lvec, &xx));
982:     PetscCall(VecGetArray(b, &bb));
983:   }
984:   PetscCall(VecGetArray(lmask, &mask));
985:   /* remove zeroed rows of off diagonal matrix */
986:   PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
987:   ii = aij->i;
988:   for (i = 0; i < len; i++) PetscCall(PetscArrayzero(aij_a + ii[lrows[i]], ii[lrows[i] + 1] - ii[lrows[i]]));
989:   /* loop over all elements of off process part of matrix zeroing removed columns*/
990:   if (aij->compressedrow.use) {
991:     m    = aij->compressedrow.nrows;
992:     ii   = aij->compressedrow.i;
993:     ridx = aij->compressedrow.rindex;
994:     for (i = 0; i < m; i++) {
995:       n  = ii[i + 1] - ii[i];
996:       aj = aij->j + ii[i];
997:       aa = aij_a + ii[i];

999:       for (j = 0; j < n; j++) {
1000:         if (PetscAbsScalar(mask[*aj])) {
1001:           if (b) bb[*ridx] -= *aa * xx[*aj];
1002:           *aa = 0.0;
1003:         }
1004:         aa++;
1005:         aj++;
1006:       }
1007:       ridx++;
1008:     }
1009:   } else { /* do not use compressed row format */
1010:     m = l->B->rmap->n;
1011:     for (i = 0; i < m; i++) {
1012:       n  = ii[i + 1] - ii[i];
1013:       aj = aij->j + ii[i];
1014:       aa = aij_a + ii[i];
1015:       for (j = 0; j < n; j++) {
1016:         if (PetscAbsScalar(mask[*aj])) {
1017:           if (b) bb[i] -= *aa * xx[*aj];
1018:           *aa = 0.0;
1019:         }
1020:         aa++;
1021:         aj++;
1022:       }
1023:     }
1024:   }
1025:   if (x && b) {
1026:     PetscCall(VecRestoreArray(b, &bb));
1027:     PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1028:   }
1029:   PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1030:   PetscCall(VecRestoreArray(lmask, &mask));
1031:   PetscCall(VecDestroy(&lmask));
1032:   PetscCall(PetscFree(lrows));

1034:   /* only change matrix nonzero state if pattern was allowed to be changed */
1035:   if (!((Mat_SeqAIJ *)(l->A->data))->keepnonzeropattern) {
1036:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1037:     PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1038:   }
1039:   PetscFunctionReturn(PETSC_SUCCESS);
1040: }

1042: PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1043: {
1044:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1045:   PetscInt    nt;
1046:   VecScatter  Mvctx = a->Mvctx;

1048:   PetscFunctionBegin;
1049:   PetscCall(VecGetLocalSize(xx, &nt));
1050:   PetscCheck(nt == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A (%" PetscInt_FMT ") and xx (%" PetscInt_FMT ")", A->cmap->n, nt);
1051:   PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1052:   PetscUseTypeMethod(a->A, mult, xx, yy);
1053:   PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1054:   PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1055:   PetscFunctionReturn(PETSC_SUCCESS);
1056: }

1058: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1059: {
1060:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1062:   PetscFunctionBegin;
1063:   PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1064:   PetscFunctionReturn(PETSC_SUCCESS);
1065: }

1067: PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1068: {
1069:   Mat_MPIAIJ *a     = (Mat_MPIAIJ *)A->data;
1070:   VecScatter  Mvctx = a->Mvctx;

1072:   PetscFunctionBegin;
1073:   PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1074:   PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1075:   PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1076:   PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1077:   PetscFunctionReturn(PETSC_SUCCESS);
1078: }

1080: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1081: {
1082:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1084:   PetscFunctionBegin;
1085:   /* do nondiagonal part */
1086:   PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1087:   /* do local part */
1088:   PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1089:   /* add partial results together */
1090:   PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1091:   PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1092:   PetscFunctionReturn(PETSC_SUCCESS);
1093: }

1095: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1096: {
1097:   MPI_Comm    comm;
1098:   Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1099:   Mat         Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1100:   IS          Me, Notme;
1101:   PetscInt    M, N, first, last, *notme, i;
1102:   PetscBool   lf;
1103:   PetscMPIInt size;

1105:   PetscFunctionBegin;
1106:   /* Easy test: symmetric diagonal block */
1107:   PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1108:   PetscCall(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1109:   if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1110:   PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1111:   PetscCallMPI(MPI_Comm_size(comm, &size));
1112:   if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);

1114:   /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1115:   PetscCall(MatGetSize(Amat, &M, &N));
1116:   PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1117:   PetscCall(PetscMalloc1(N - last + first, &notme));
1118:   for (i = 0; i < first; i++) notme[i] = i;
1119:   for (i = last; i < M; i++) notme[i - last + first] = i;
1120:   PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1121:   PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1122:   PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1123:   Aoff = Aoffs[0];
1124:   PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1125:   Boff = Boffs[0];
1126:   PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1127:   PetscCall(MatDestroyMatrices(1, &Aoffs));
1128:   PetscCall(MatDestroyMatrices(1, &Boffs));
1129:   PetscCall(ISDestroy(&Me));
1130:   PetscCall(ISDestroy(&Notme));
1131:   PetscCall(PetscFree(notme));
1132:   PetscFunctionReturn(PETSC_SUCCESS);
1133: }

1135: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A, PetscReal tol, PetscBool *f)
1136: {
1137:   PetscFunctionBegin;
1138:   PetscCall(MatIsTranspose_MPIAIJ(A, A, tol, f));
1139:   PetscFunctionReturn(PETSC_SUCCESS);
1140: }

1142: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1143: {
1144:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1146:   PetscFunctionBegin;
1147:   /* do nondiagonal part */
1148:   PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1149:   /* do local part */
1150:   PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1151:   /* add partial results together */
1152:   PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1153:   PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1154:   PetscFunctionReturn(PETSC_SUCCESS);
1155: }

1157: /*
1158:   This only works correctly for square matrices where the subblock A->A is the
1159:    diagonal block
1160: */
1161: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1162: {
1163:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1165:   PetscFunctionBegin;
1166:   PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1167:   PetscCheck(A->rmap->rstart == A->cmap->rstart && A->rmap->rend == A->cmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "row partition must equal col partition");
1168:   PetscCall(MatGetDiagonal(a->A, v));
1169:   PetscFunctionReturn(PETSC_SUCCESS);
1170: }

1172: PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1173: {
1174:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1176:   PetscFunctionBegin;
1177:   PetscCall(MatScale(a->A, aa));
1178:   PetscCall(MatScale(a->B, aa));
1179:   PetscFunctionReturn(PETSC_SUCCESS);
1180: }

1182: /* Free COO stuff; must match allocation methods in MatSetPreallocationCOO_MPIAIJ() */
1183: PETSC_INTERN PetscErrorCode MatResetPreallocationCOO_MPIAIJ(Mat mat)
1184: {
1185:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

1187:   PetscFunctionBegin;
1188:   PetscCall(PetscSFDestroy(&aij->coo_sf));
1189:   PetscCall(PetscFree(aij->Aperm1));
1190:   PetscCall(PetscFree(aij->Bperm1));
1191:   PetscCall(PetscFree(aij->Ajmap1));
1192:   PetscCall(PetscFree(aij->Bjmap1));

1194:   PetscCall(PetscFree(aij->Aimap2));
1195:   PetscCall(PetscFree(aij->Bimap2));
1196:   PetscCall(PetscFree(aij->Aperm2));
1197:   PetscCall(PetscFree(aij->Bperm2));
1198:   PetscCall(PetscFree(aij->Ajmap2));
1199:   PetscCall(PetscFree(aij->Bjmap2));

1201:   PetscCall(PetscFree2(aij->sendbuf, aij->recvbuf));
1202:   PetscCall(PetscFree(aij->Cperm1));
1203:   PetscFunctionReturn(PETSC_SUCCESS);
1204: }

1206: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1207: {
1208:   Mat_MPIAIJ        *aij    = (Mat_MPIAIJ *)mat->data;
1209:   Mat_SeqAIJ        *A      = (Mat_SeqAIJ *)aij->A->data;
1210:   Mat_SeqAIJ        *B      = (Mat_SeqAIJ *)aij->B->data;
1211:   const PetscInt    *garray = aij->garray;
1212:   const PetscScalar *aa, *ba;
1213:   PetscInt           header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1214:   PetscInt64         nz, hnz;
1215:   PetscInt          *rowlens;
1216:   PetscInt          *colidxs;
1217:   PetscScalar       *matvals;
1218:   PetscMPIInt        rank;

1220:   PetscFunctionBegin;
1221:   PetscCall(PetscViewerSetUp(viewer));

1223:   M  = mat->rmap->N;
1224:   N  = mat->cmap->N;
1225:   m  = mat->rmap->n;
1226:   rs = mat->rmap->rstart;
1227:   cs = mat->cmap->rstart;
1228:   nz = A->nz + B->nz;

1230:   /* write matrix header */
1231:   header[0] = MAT_FILE_CLASSID;
1232:   header[1] = M;
1233:   header[2] = N;
1234:   PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1235:   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1236:   if (rank == 0) PetscCall(PetscIntCast(hnz, &header[3]));
1237:   PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));

1239:   /* fill in and store row lengths  */
1240:   PetscCall(PetscMalloc1(m, &rowlens));
1241:   for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1242:   PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1243:   PetscCall(PetscFree(rowlens));

1245:   /* fill in and store column indices */
1246:   PetscCall(PetscMalloc1(nz, &colidxs));
1247:   for (cnt = 0, i = 0; i < m; i++) {
1248:     for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1249:       if (garray[B->j[jb]] > cs) break;
1250:       colidxs[cnt++] = garray[B->j[jb]];
1251:     }
1252:     for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1253:     for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1254:   }
1255:   PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1256:   PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1257:   PetscCall(PetscFree(colidxs));

1259:   /* fill in and store nonzero values */
1260:   PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1261:   PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1262:   PetscCall(PetscMalloc1(nz, &matvals));
1263:   for (cnt = 0, i = 0; i < m; i++) {
1264:     for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1265:       if (garray[B->j[jb]] > cs) break;
1266:       matvals[cnt++] = ba[jb];
1267:     }
1268:     for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1269:     for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1270:   }
1271:   PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1272:   PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1273:   PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1274:   PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1275:   PetscCall(PetscFree(matvals));

1277:   /* write block size option to the viewer's .info file */
1278:   PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1279:   PetscFunctionReturn(PETSC_SUCCESS);
1280: }

1282: #include <petscdraw.h>
1283: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1284: {
1285:   Mat_MPIAIJ       *aij  = (Mat_MPIAIJ *)mat->data;
1286:   PetscMPIInt       rank = aij->rank, size = aij->size;
1287:   PetscBool         isdraw, iascii, isbinary;
1288:   PetscViewer       sviewer;
1289:   PetscViewerFormat format;

1291:   PetscFunctionBegin;
1292:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1293:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1294:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1295:   if (iascii) {
1296:     PetscCall(PetscViewerGetFormat(viewer, &format));
1297:     if (format == PETSC_VIEWER_LOAD_BALANCE) {
1298:       PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)(aij->A->data))->nz + ((Mat_SeqAIJ *)(aij->B->data))->nz;
1299:       PetscCall(PetscMalloc1(size, &nz));
1300:       PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1301:       for (i = 0; i < (PetscInt)size; i++) {
1302:         nmax = PetscMax(nmax, nz[i]);
1303:         nmin = PetscMin(nmin, nz[i]);
1304:         navg += nz[i];
1305:       }
1306:       PetscCall(PetscFree(nz));
1307:       navg = navg / size;
1308:       PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT "  avg %" PetscInt_FMT "  max %" PetscInt_FMT "\n", nmin, navg, nmax));
1309:       PetscFunctionReturn(PETSC_SUCCESS);
1310:     }
1311:     PetscCall(PetscViewerGetFormat(viewer, &format));
1312:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1313:       MatInfo   info;
1314:       PetscInt *inodes = NULL;

1316:       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1317:       PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1318:       PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1319:       PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1320:       if (!inodes) {
1321:         PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " mem %g, not using I-node routines\n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
1322:                                                      (double)info.memory));
1323:       } else {
1324:         PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " mem %g, using I-node routines\n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
1325:                                                      (double)info.memory));
1326:       }
1327:       PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1328:       PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1329:       PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1330:       PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1331:       PetscCall(PetscViewerFlush(viewer));
1332:       PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1333:       PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1334:       PetscCall(VecScatterView(aij->Mvctx, viewer));
1335:       PetscFunctionReturn(PETSC_SUCCESS);
1336:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
1337:       PetscInt inodecount, inodelimit, *inodes;
1338:       PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1339:       if (inodes) {
1340:         PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1341:       } else {
1342:         PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1343:       }
1344:       PetscFunctionReturn(PETSC_SUCCESS);
1345:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1346:       PetscFunctionReturn(PETSC_SUCCESS);
1347:     }
1348:   } else if (isbinary) {
1349:     if (size == 1) {
1350:       PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1351:       PetscCall(MatView(aij->A, viewer));
1352:     } else {
1353:       PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1354:     }
1355:     PetscFunctionReturn(PETSC_SUCCESS);
1356:   } else if (iascii && size == 1) {
1357:     PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1358:     PetscCall(MatView(aij->A, viewer));
1359:     PetscFunctionReturn(PETSC_SUCCESS);
1360:   } else if (isdraw) {
1361:     PetscDraw draw;
1362:     PetscBool isnull;
1363:     PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1364:     PetscCall(PetscDrawIsNull(draw, &isnull));
1365:     if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1366:   }

1368:   { /* assemble the entire matrix onto first processor */
1369:     Mat A = NULL, Av;
1370:     IS  isrow, iscol;

1372:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1373:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1374:     PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1375:     PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1376:     /*  The commented code uses MatCreateSubMatrices instead */
1377:     /*
1378:     Mat *AA, A = NULL, Av;
1379:     IS  isrow,iscol;

1381:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1382:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1383:     PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1384:     if (rank == 0) {
1385:        PetscCall(PetscObjectReference((PetscObject)AA[0]));
1386:        A    = AA[0];
1387:        Av   = AA[0];
1388:     }
1389:     PetscCall(MatDestroySubMatrices(1,&AA));
1390: */
1391:     PetscCall(ISDestroy(&iscol));
1392:     PetscCall(ISDestroy(&isrow));
1393:     /*
1394:        Everyone has to call to draw the matrix since the graphics waits are
1395:        synchronized across all processors that share the PetscDraw object
1396:     */
1397:     PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1398:     if (rank == 0) {
1399:       if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1400:       PetscCall(MatView_SeqAIJ(Av, sviewer));
1401:     }
1402:     PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1403:     PetscCall(PetscViewerFlush(viewer));
1404:     PetscCall(MatDestroy(&A));
1405:   }
1406:   PetscFunctionReturn(PETSC_SUCCESS);
1407: }

1409: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1410: {
1411:   PetscBool iascii, isdraw, issocket, isbinary;

1413:   PetscFunctionBegin;
1414:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1415:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1416:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1417:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1418:   if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1419:   PetscFunctionReturn(PETSC_SUCCESS);
1420: }

1422: PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1423: {
1424:   Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1425:   Vec         bb1 = NULL;
1426:   PetscBool   hasop;

1428:   PetscFunctionBegin;
1429:   if (flag == SOR_APPLY_UPPER) {
1430:     PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1431:     PetscFunctionReturn(PETSC_SUCCESS);
1432:   }

1434:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));

1436:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1437:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1438:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1439:       its--;
1440:     }

1442:     while (its--) {
1443:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1444:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1446:       /* update rhs: bb1 = bb - B*x */
1447:       PetscCall(VecScale(mat->lvec, -1.0));
1448:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1450:       /* local sweep */
1451:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1452:     }
1453:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1454:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1455:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1456:       its--;
1457:     }
1458:     while (its--) {
1459:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1460:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1462:       /* update rhs: bb1 = bb - B*x */
1463:       PetscCall(VecScale(mat->lvec, -1.0));
1464:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1466:       /* local sweep */
1467:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1468:     }
1469:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1470:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1471:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1472:       its--;
1473:     }
1474:     while (its--) {
1475:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1476:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1478:       /* update rhs: bb1 = bb - B*x */
1479:       PetscCall(VecScale(mat->lvec, -1.0));
1480:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1482:       /* local sweep */
1483:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1484:     }
1485:   } else if (flag & SOR_EISENSTAT) {
1486:     Vec xx1;

1488:     PetscCall(VecDuplicate(bb, &xx1));
1489:     PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));

1491:     PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1492:     PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1493:     if (!mat->diag) {
1494:       PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1495:       PetscCall(MatGetDiagonal(matin, mat->diag));
1496:     }
1497:     PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1498:     if (hasop) {
1499:       PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1500:     } else {
1501:       PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1502:     }
1503:     PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));

1505:     PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));

1507:     /* local sweep */
1508:     PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1509:     PetscCall(VecAXPY(xx, 1.0, xx1));
1510:     PetscCall(VecDestroy(&xx1));
1511:   } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");

1513:   PetscCall(VecDestroy(&bb1));

1515:   matin->factorerrortype = mat->A->factorerrortype;
1516:   PetscFunctionReturn(PETSC_SUCCESS);
1517: }

1519: PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1520: {
1521:   Mat             aA, aB, Aperm;
1522:   const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1523:   PetscScalar    *aa, *ba;
1524:   PetscInt        i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1525:   PetscSF         rowsf, sf;
1526:   IS              parcolp = NULL;
1527:   PetscBool       done;

1529:   PetscFunctionBegin;
1530:   PetscCall(MatGetLocalSize(A, &m, &n));
1531:   PetscCall(ISGetIndices(rowp, &rwant));
1532:   PetscCall(ISGetIndices(colp, &cwant));
1533:   PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));

1535:   /* Invert row permutation to find out where my rows should go */
1536:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1537:   PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1538:   PetscCall(PetscSFSetFromOptions(rowsf));
1539:   for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1540:   PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1541:   PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));

1543:   /* Invert column permutation to find out where my columns should go */
1544:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1545:   PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1546:   PetscCall(PetscSFSetFromOptions(sf));
1547:   for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1548:   PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1549:   PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1550:   PetscCall(PetscSFDestroy(&sf));

1552:   PetscCall(ISRestoreIndices(rowp, &rwant));
1553:   PetscCall(ISRestoreIndices(colp, &cwant));
1554:   PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));

1556:   /* Find out where my gcols should go */
1557:   PetscCall(MatGetSize(aB, NULL, &ng));
1558:   PetscCall(PetscMalloc1(ng, &gcdest));
1559:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1560:   PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1561:   PetscCall(PetscSFSetFromOptions(sf));
1562:   PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1563:   PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1564:   PetscCall(PetscSFDestroy(&sf));

1566:   PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1567:   PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1568:   PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1569:   for (i = 0; i < m; i++) {
1570:     PetscInt    row = rdest[i];
1571:     PetscMPIInt rowner;
1572:     PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1573:     for (j = ai[i]; j < ai[i + 1]; j++) {
1574:       PetscInt    col = cdest[aj[j]];
1575:       PetscMPIInt cowner;
1576:       PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1577:       if (rowner == cowner) dnnz[i]++;
1578:       else onnz[i]++;
1579:     }
1580:     for (j = bi[i]; j < bi[i + 1]; j++) {
1581:       PetscInt    col = gcdest[bj[j]];
1582:       PetscMPIInt cowner;
1583:       PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1584:       if (rowner == cowner) dnnz[i]++;
1585:       else onnz[i]++;
1586:     }
1587:   }
1588:   PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1589:   PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1590:   PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1591:   PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1592:   PetscCall(PetscSFDestroy(&rowsf));

1594:   PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1595:   PetscCall(MatSeqAIJGetArray(aA, &aa));
1596:   PetscCall(MatSeqAIJGetArray(aB, &ba));
1597:   for (i = 0; i < m; i++) {
1598:     PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1599:     PetscInt  j0, rowlen;
1600:     rowlen = ai[i + 1] - ai[i];
1601:     for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1602:       for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1603:       PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1604:     }
1605:     rowlen = bi[i + 1] - bi[i];
1606:     for (j0 = j = 0; j < rowlen; j0 = j) {
1607:       for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1608:       PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1609:     }
1610:   }
1611:   PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1612:   PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1613:   PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1614:   PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1615:   PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1616:   PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1617:   PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1618:   PetscCall(PetscFree3(work, rdest, cdest));
1619:   PetscCall(PetscFree(gcdest));
1620:   if (parcolp) PetscCall(ISDestroy(&colp));
1621:   *B = Aperm;
1622:   PetscFunctionReturn(PETSC_SUCCESS);
1623: }

1625: PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1626: {
1627:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

1629:   PetscFunctionBegin;
1630:   PetscCall(MatGetSize(aij->B, NULL, nghosts));
1631:   if (ghosts) *ghosts = aij->garray;
1632:   PetscFunctionReturn(PETSC_SUCCESS);
1633: }

1635: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1636: {
1637:   Mat_MPIAIJ    *mat = (Mat_MPIAIJ *)matin->data;
1638:   Mat            A = mat->A, B = mat->B;
1639:   PetscLogDouble isend[5], irecv[5];

1641:   PetscFunctionBegin;
1642:   info->block_size = 1.0;
1643:   PetscCall(MatGetInfo(A, MAT_LOCAL, info));

1645:   isend[0] = info->nz_used;
1646:   isend[1] = info->nz_allocated;
1647:   isend[2] = info->nz_unneeded;
1648:   isend[3] = info->memory;
1649:   isend[4] = info->mallocs;

1651:   PetscCall(MatGetInfo(B, MAT_LOCAL, info));

1653:   isend[0] += info->nz_used;
1654:   isend[1] += info->nz_allocated;
1655:   isend[2] += info->nz_unneeded;
1656:   isend[3] += info->memory;
1657:   isend[4] += info->mallocs;
1658:   if (flag == MAT_LOCAL) {
1659:     info->nz_used      = isend[0];
1660:     info->nz_allocated = isend[1];
1661:     info->nz_unneeded  = isend[2];
1662:     info->memory       = isend[3];
1663:     info->mallocs      = isend[4];
1664:   } else if (flag == MAT_GLOBAL_MAX) {
1665:     PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));

1667:     info->nz_used      = irecv[0];
1668:     info->nz_allocated = irecv[1];
1669:     info->nz_unneeded  = irecv[2];
1670:     info->memory       = irecv[3];
1671:     info->mallocs      = irecv[4];
1672:   } else if (flag == MAT_GLOBAL_SUM) {
1673:     PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));

1675:     info->nz_used      = irecv[0];
1676:     info->nz_allocated = irecv[1];
1677:     info->nz_unneeded  = irecv[2];
1678:     info->memory       = irecv[3];
1679:     info->mallocs      = irecv[4];
1680:   }
1681:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1682:   info->fill_ratio_needed = 0;
1683:   info->factor_mallocs    = 0;
1684:   PetscFunctionReturn(PETSC_SUCCESS);
1685: }

1687: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1688: {
1689:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1691:   PetscFunctionBegin;
1692:   switch (op) {
1693:   case MAT_NEW_NONZERO_LOCATIONS:
1694:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1695:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1696:   case MAT_KEEP_NONZERO_PATTERN:
1697:   case MAT_NEW_NONZERO_LOCATION_ERR:
1698:   case MAT_USE_INODES:
1699:   case MAT_IGNORE_ZERO_ENTRIES:
1700:   case MAT_FORM_EXPLICIT_TRANSPOSE:
1701:     MatCheckPreallocated(A, 1);
1702:     PetscCall(MatSetOption(a->A, op, flg));
1703:     PetscCall(MatSetOption(a->B, op, flg));
1704:     break;
1705:   case MAT_ROW_ORIENTED:
1706:     MatCheckPreallocated(A, 1);
1707:     a->roworiented = flg;

1709:     PetscCall(MatSetOption(a->A, op, flg));
1710:     PetscCall(MatSetOption(a->B, op, flg));
1711:     break;
1712:   case MAT_FORCE_DIAGONAL_ENTRIES:
1713:   case MAT_SORTED_FULL:
1714:     PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1715:     break;
1716:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1717:     a->donotstash = flg;
1718:     break;
1719:   /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1720:   case MAT_SPD:
1721:   case MAT_SYMMETRIC:
1722:   case MAT_STRUCTURALLY_SYMMETRIC:
1723:   case MAT_HERMITIAN:
1724:   case MAT_SYMMETRY_ETERNAL:
1725:   case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1726:   case MAT_SPD_ETERNAL:
1727:     /* if the diagonal matrix is square it inherits some of the properties above */
1728:     break;
1729:   case MAT_SUBMAT_SINGLEIS:
1730:     A->submat_singleis = flg;
1731:     break;
1732:   case MAT_STRUCTURE_ONLY:
1733:     /* The option is handled directly by MatSetOption() */
1734:     break;
1735:   default:
1736:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1737:   }
1738:   PetscFunctionReturn(PETSC_SUCCESS);
1739: }

1741: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1742: {
1743:   Mat_MPIAIJ  *mat = (Mat_MPIAIJ *)matin->data;
1744:   PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1745:   PetscInt     i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1746:   PetscInt     nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1747:   PetscInt    *cmap, *idx_p;

1749:   PetscFunctionBegin;
1750:   PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1751:   mat->getrowactive = PETSC_TRUE;

1753:   if (!mat->rowvalues && (idx || v)) {
1754:     /*
1755:         allocate enough space to hold information from the longest row.
1756:     */
1757:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1758:     PetscInt    max = 1, tmp;
1759:     for (i = 0; i < matin->rmap->n; i++) {
1760:       tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1761:       if (max < tmp) max = tmp;
1762:     }
1763:     PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1764:   }

1766:   PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1767:   lrow = row - rstart;

1769:   pvA = &vworkA;
1770:   pcA = &cworkA;
1771:   pvB = &vworkB;
1772:   pcB = &cworkB;
1773:   if (!v) {
1774:     pvA = NULL;
1775:     pvB = NULL;
1776:   }
1777:   if (!idx) {
1778:     pcA = NULL;
1779:     if (!v) pcB = NULL;
1780:   }
1781:   PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1782:   PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1783:   nztot = nzA + nzB;

1785:   cmap = mat->garray;
1786:   if (v || idx) {
1787:     if (nztot) {
1788:       /* Sort by increasing column numbers, assuming A and B already sorted */
1789:       PetscInt imark = -1;
1790:       if (v) {
1791:         *v = v_p = mat->rowvalues;
1792:         for (i = 0; i < nzB; i++) {
1793:           if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1794:           else break;
1795:         }
1796:         imark = i;
1797:         for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1798:         for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1799:       }
1800:       if (idx) {
1801:         *idx = idx_p = mat->rowindices;
1802:         if (imark > -1) {
1803:           for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1804:         } else {
1805:           for (i = 0; i < nzB; i++) {
1806:             if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1807:             else break;
1808:           }
1809:           imark = i;
1810:         }
1811:         for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1812:         for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1813:       }
1814:     } else {
1815:       if (idx) *idx = NULL;
1816:       if (v) *v = NULL;
1817:     }
1818:   }
1819:   *nz = nztot;
1820:   PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1821:   PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1822:   PetscFunctionReturn(PETSC_SUCCESS);
1823: }

1825: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1826: {
1827:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

1829:   PetscFunctionBegin;
1830:   PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1831:   aij->getrowactive = PETSC_FALSE;
1832:   PetscFunctionReturn(PETSC_SUCCESS);
1833: }

1835: PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1836: {
1837:   Mat_MPIAIJ      *aij  = (Mat_MPIAIJ *)mat->data;
1838:   Mat_SeqAIJ      *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1839:   PetscInt         i, j, cstart = mat->cmap->rstart;
1840:   PetscReal        sum = 0.0;
1841:   const MatScalar *v, *amata, *bmata;

1843:   PetscFunctionBegin;
1844:   if (aij->size == 1) {
1845:     PetscCall(MatNorm(aij->A, type, norm));
1846:   } else {
1847:     PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1848:     PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1849:     if (type == NORM_FROBENIUS) {
1850:       v = amata;
1851:       for (i = 0; i < amat->nz; i++) {
1852:         sum += PetscRealPart(PetscConj(*v) * (*v));
1853:         v++;
1854:       }
1855:       v = bmata;
1856:       for (i = 0; i < bmat->nz; i++) {
1857:         sum += PetscRealPart(PetscConj(*v) * (*v));
1858:         v++;
1859:       }
1860:       PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1861:       *norm = PetscSqrtReal(*norm);
1862:       PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1863:     } else if (type == NORM_1) { /* max column norm */
1864:       PetscReal *tmp, *tmp2;
1865:       PetscInt  *jj, *garray = aij->garray;
1866:       PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1867:       PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1868:       *norm = 0.0;
1869:       v     = amata;
1870:       jj    = amat->j;
1871:       for (j = 0; j < amat->nz; j++) {
1872:         tmp[cstart + *jj++] += PetscAbsScalar(*v);
1873:         v++;
1874:       }
1875:       v  = bmata;
1876:       jj = bmat->j;
1877:       for (j = 0; j < bmat->nz; j++) {
1878:         tmp[garray[*jj++]] += PetscAbsScalar(*v);
1879:         v++;
1880:       }
1881:       PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1882:       for (j = 0; j < mat->cmap->N; j++) {
1883:         if (tmp2[j] > *norm) *norm = tmp2[j];
1884:       }
1885:       PetscCall(PetscFree(tmp));
1886:       PetscCall(PetscFree(tmp2));
1887:       PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1888:     } else if (type == NORM_INFINITY) { /* max row norm */
1889:       PetscReal ntemp = 0.0;
1890:       for (j = 0; j < aij->A->rmap->n; j++) {
1891:         v   = amata + amat->i[j];
1892:         sum = 0.0;
1893:         for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1894:           sum += PetscAbsScalar(*v);
1895:           v++;
1896:         }
1897:         v = bmata + bmat->i[j];
1898:         for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1899:           sum += PetscAbsScalar(*v);
1900:           v++;
1901:         }
1902:         if (sum > ntemp) ntemp = sum;
1903:       }
1904:       PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1905:       PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1906:     } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1907:     PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1908:     PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1909:   }
1910:   PetscFunctionReturn(PETSC_SUCCESS);
1911: }

1913: PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1914: {
1915:   Mat_MPIAIJ      *a    = (Mat_MPIAIJ *)A->data, *b;
1916:   Mat_SeqAIJ      *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1917:   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;
1918:   const PetscInt  *ai, *aj, *bi, *bj, *B_diag_i;
1919:   Mat              B, A_diag, *B_diag;
1920:   const MatScalar *pbv, *bv;

1922:   PetscFunctionBegin;
1923:   if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1924:   ma = A->rmap->n;
1925:   na = A->cmap->n;
1926:   mb = a->B->rmap->n;
1927:   nb = a->B->cmap->n;
1928:   ai = Aloc->i;
1929:   aj = Aloc->j;
1930:   bi = Bloc->i;
1931:   bj = Bloc->j;
1932:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1933:     PetscInt            *d_nnz, *g_nnz, *o_nnz;
1934:     PetscSFNode         *oloc;
1935:     PETSC_UNUSED PetscSF sf;

1937:     PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1938:     /* compute d_nnz for preallocation */
1939:     PetscCall(PetscArrayzero(d_nnz, na));
1940:     for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1941:     /* compute local off-diagonal contributions */
1942:     PetscCall(PetscArrayzero(g_nnz, nb));
1943:     for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1944:     /* map those to global */
1945:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1946:     PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1947:     PetscCall(PetscSFSetFromOptions(sf));
1948:     PetscCall(PetscArrayzero(o_nnz, na));
1949:     PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1950:     PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1951:     PetscCall(PetscSFDestroy(&sf));

1953:     PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1954:     PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1955:     PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1956:     PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1957:     PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1958:     PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1959:   } else {
1960:     B = *matout;
1961:     PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1962:   }

1964:   b           = (Mat_MPIAIJ *)B->data;
1965:   A_diag      = a->A;
1966:   B_diag      = &b->A;
1967:   sub_B_diag  = (Mat_SeqAIJ *)(*B_diag)->data;
1968:   A_diag_ncol = A_diag->cmap->N;
1969:   B_diag_ilen = sub_B_diag->ilen;
1970:   B_diag_i    = sub_B_diag->i;

1972:   /* Set ilen for diagonal of B */
1973:   for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];

1975:   /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
1976:   very quickly (=without using MatSetValues), because all writes are local. */
1977:   PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1978:   PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));

1980:   /* copy over the B part */
1981:   PetscCall(PetscMalloc1(bi[mb], &cols));
1982:   PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1983:   pbv = bv;
1984:   row = A->rmap->rstart;
1985:   for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1986:   cols_tmp = cols;
1987:   for (i = 0; i < mb; i++) {
1988:     ncol = bi[i + 1] - bi[i];
1989:     PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1990:     row++;
1991:     pbv += ncol;
1992:     cols_tmp += ncol;
1993:   }
1994:   PetscCall(PetscFree(cols));
1995:   PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));

1997:   PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1998:   PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1999:   if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2000:     *matout = B;
2001:   } else {
2002:     PetscCall(MatHeaderMerge(A, &B));
2003:   }
2004:   PetscFunctionReturn(PETSC_SUCCESS);
2005: }

2007: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
2008: {
2009:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2010:   Mat         a = aij->A, b = aij->B;
2011:   PetscInt    s1, s2, s3;

2013:   PetscFunctionBegin;
2014:   PetscCall(MatGetLocalSize(mat, &s2, &s3));
2015:   if (rr) {
2016:     PetscCall(VecGetLocalSize(rr, &s1));
2017:     PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
2018:     /* Overlap communication with computation. */
2019:     PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2020:   }
2021:   if (ll) {
2022:     PetscCall(VecGetLocalSize(ll, &s1));
2023:     PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
2024:     PetscUseTypeMethod(b, diagonalscale, ll, NULL);
2025:   }
2026:   /* scale  the diagonal block */
2027:   PetscUseTypeMethod(a, diagonalscale, ll, rr);

2029:   if (rr) {
2030:     /* Do a scatter end and then right scale the off-diagonal block */
2031:     PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2032:     PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
2033:   }
2034:   PetscFunctionReturn(PETSC_SUCCESS);
2035: }

2037: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2038: {
2039:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2041:   PetscFunctionBegin;
2042:   PetscCall(MatSetUnfactored(a->A));
2043:   PetscFunctionReturn(PETSC_SUCCESS);
2044: }

2046: PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2047: {
2048:   Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2049:   Mat         a, b, c, d;
2050:   PetscBool   flg;

2052:   PetscFunctionBegin;
2053:   a = matA->A;
2054:   b = matA->B;
2055:   c = matB->A;
2056:   d = matB->B;

2058:   PetscCall(MatEqual(a, c, &flg));
2059:   if (flg) PetscCall(MatEqual(b, d, &flg));
2060:   PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2061:   PetscFunctionReturn(PETSC_SUCCESS);
2062: }

2064: PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2065: {
2066:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2067:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;

2069:   PetscFunctionBegin;
2070:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2071:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2072:     /* because of the column compression in the off-processor part of the matrix a->B,
2073:        the number of columns in a->B and b->B may be different, hence we cannot call
2074:        the MatCopy() directly on the two parts. If need be, we can provide a more
2075:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2076:        then copying the submatrices */
2077:     PetscCall(MatCopy_Basic(A, B, str));
2078:   } else {
2079:     PetscCall(MatCopy(a->A, b->A, str));
2080:     PetscCall(MatCopy(a->B, b->B, str));
2081:   }
2082:   PetscCall(PetscObjectStateIncrease((PetscObject)B));
2083:   PetscFunctionReturn(PETSC_SUCCESS);
2084: }

2086: /*
2087:    Computes the number of nonzeros per row needed for preallocation when X and Y
2088:    have different nonzero structure.
2089: */
2090: 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)
2091: {
2092:   PetscInt i, j, k, nzx, nzy;

2094:   PetscFunctionBegin;
2095:   /* Set the number of nonzeros in the new matrix */
2096:   for (i = 0; i < m; i++) {
2097:     const PetscInt *xjj = xj + xi[i], *yjj = yj + yi[i];
2098:     nzx    = xi[i + 1] - xi[i];
2099:     nzy    = yi[i + 1] - yi[i];
2100:     nnz[i] = 0;
2101:     for (j = 0, k = 0; j < nzx; j++) {                                /* Point in X */
2102:       for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2103:       if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++;             /* Skip duplicate */
2104:       nnz[i]++;
2105:     }
2106:     for (; k < nzy; k++) nnz[i]++;
2107:   }
2108:   PetscFunctionReturn(PETSC_SUCCESS);
2109: }

2111: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2112: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2113: {
2114:   PetscInt    m = Y->rmap->N;
2115:   Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2116:   Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;

2118:   PetscFunctionBegin;
2119:   PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2120:   PetscFunctionReturn(PETSC_SUCCESS);
2121: }

2123: PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2124: {
2125:   Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;

2127:   PetscFunctionBegin;
2128:   if (str == SAME_NONZERO_PATTERN) {
2129:     PetscCall(MatAXPY(yy->A, a, xx->A, str));
2130:     PetscCall(MatAXPY(yy->B, a, xx->B, str));
2131:   } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2132:     PetscCall(MatAXPY_Basic(Y, a, X, str));
2133:   } else {
2134:     Mat       B;
2135:     PetscInt *nnz_d, *nnz_o;

2137:     PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2138:     PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2139:     PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2140:     PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2141:     PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2142:     PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2143:     PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2144:     PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2145:     PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2146:     PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2147:     PetscCall(MatHeaderMerge(Y, &B));
2148:     PetscCall(PetscFree(nnz_d));
2149:     PetscCall(PetscFree(nnz_o));
2150:   }
2151:   PetscFunctionReturn(PETSC_SUCCESS);
2152: }

2154: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);

2156: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2157: {
2158:   PetscFunctionBegin;
2159:   if (PetscDefined(USE_COMPLEX)) {
2160:     Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2162:     PetscCall(MatConjugate_SeqAIJ(aij->A));
2163:     PetscCall(MatConjugate_SeqAIJ(aij->B));
2164:   }
2165:   PetscFunctionReturn(PETSC_SUCCESS);
2166: }

2168: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2169: {
2170:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2172:   PetscFunctionBegin;
2173:   PetscCall(MatRealPart(a->A));
2174:   PetscCall(MatRealPart(a->B));
2175:   PetscFunctionReturn(PETSC_SUCCESS);
2176: }

2178: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2179: {
2180:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2182:   PetscFunctionBegin;
2183:   PetscCall(MatImaginaryPart(a->A));
2184:   PetscCall(MatImaginaryPart(a->B));
2185:   PetscFunctionReturn(PETSC_SUCCESS);
2186: }

2188: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2189: {
2190:   Mat_MPIAIJ        *a = (Mat_MPIAIJ *)A->data;
2191:   PetscInt           i, *idxb = NULL, m = A->rmap->n;
2192:   PetscScalar       *va, *vv;
2193:   Vec                vB, vA;
2194:   const PetscScalar *vb;

2196:   PetscFunctionBegin;
2197:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2198:   PetscCall(MatGetRowMaxAbs(a->A, vA, idx));

2200:   PetscCall(VecGetArrayWrite(vA, &va));
2201:   if (idx) {
2202:     for (i = 0; i < m; i++) {
2203:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2204:     }
2205:   }

2207:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2208:   PetscCall(PetscMalloc1(m, &idxb));
2209:   PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));

2211:   PetscCall(VecGetArrayWrite(v, &vv));
2212:   PetscCall(VecGetArrayRead(vB, &vb));
2213:   for (i = 0; i < m; i++) {
2214:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2215:       vv[i] = vb[i];
2216:       if (idx) idx[i] = a->garray[idxb[i]];
2217:     } else {
2218:       vv[i] = va[i];
2219:       if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2220:     }
2221:   }
2222:   PetscCall(VecRestoreArrayWrite(vA, &vv));
2223:   PetscCall(VecRestoreArrayWrite(vA, &va));
2224:   PetscCall(VecRestoreArrayRead(vB, &vb));
2225:   PetscCall(PetscFree(idxb));
2226:   PetscCall(VecDestroy(&vA));
2227:   PetscCall(VecDestroy(&vB));
2228:   PetscFunctionReturn(PETSC_SUCCESS);
2229: }

2231: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2232: {
2233:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2234:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2235:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2236:   PetscInt          *cmap = mat->garray;
2237:   PetscInt          *diagIdx, *offdiagIdx;
2238:   Vec                diagV, offdiagV;
2239:   PetscScalar       *a, *diagA, *offdiagA;
2240:   const PetscScalar *ba, *bav;
2241:   PetscInt           r, j, col, ncols, *bi, *bj;
2242:   Mat                B = mat->B;
2243:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2245:   PetscFunctionBegin;
2246:   /* When a process holds entire A and other processes have no entry */
2247:   if (A->cmap->N == n) {
2248:     PetscCall(VecGetArrayWrite(v, &diagA));
2249:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2250:     PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2251:     PetscCall(VecDestroy(&diagV));
2252:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2253:     PetscFunctionReturn(PETSC_SUCCESS);
2254:   } else if (n == 0) {
2255:     if (m) {
2256:       PetscCall(VecGetArrayWrite(v, &a));
2257:       for (r = 0; r < m; r++) {
2258:         a[r] = 0.0;
2259:         if (idx) idx[r] = -1;
2260:       }
2261:       PetscCall(VecRestoreArrayWrite(v, &a));
2262:     }
2263:     PetscFunctionReturn(PETSC_SUCCESS);
2264:   }

2266:   PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2267:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2268:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2269:   PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));

2271:   /* Get offdiagIdx[] for implicit 0.0 */
2272:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2273:   ba = bav;
2274:   bi = b->i;
2275:   bj = b->j;
2276:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2277:   for (r = 0; r < m; r++) {
2278:     ncols = bi[r + 1] - bi[r];
2279:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2280:       offdiagA[r]   = *ba;
2281:       offdiagIdx[r] = cmap[0];
2282:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2283:       offdiagA[r] = 0.0;

2285:       /* Find first hole in the cmap */
2286:       for (j = 0; j < ncols; j++) {
2287:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2288:         if (col > j && j < cstart) {
2289:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2290:           break;
2291:         } else if (col > j + n && j >= cstart) {
2292:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2293:           break;
2294:         }
2295:       }
2296:       if (j == ncols && ncols < A->cmap->N - n) {
2297:         /* a hole is outside compressed Bcols */
2298:         if (ncols == 0) {
2299:           if (cstart) {
2300:             offdiagIdx[r] = 0;
2301:           } else offdiagIdx[r] = cend;
2302:         } else { /* ncols > 0 */
2303:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2304:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2305:         }
2306:       }
2307:     }

2309:     for (j = 0; j < ncols; j++) {
2310:       if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2311:         offdiagA[r]   = *ba;
2312:         offdiagIdx[r] = cmap[*bj];
2313:       }
2314:       ba++;
2315:       bj++;
2316:     }
2317:   }

2319:   PetscCall(VecGetArrayWrite(v, &a));
2320:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2321:   for (r = 0; r < m; ++r) {
2322:     if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2323:       a[r] = diagA[r];
2324:       if (idx) idx[r] = cstart + diagIdx[r];
2325:     } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2326:       a[r] = diagA[r];
2327:       if (idx) {
2328:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2329:           idx[r] = cstart + diagIdx[r];
2330:         } else idx[r] = offdiagIdx[r];
2331:       }
2332:     } else {
2333:       a[r] = offdiagA[r];
2334:       if (idx) idx[r] = offdiagIdx[r];
2335:     }
2336:   }
2337:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2338:   PetscCall(VecRestoreArrayWrite(v, &a));
2339:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2340:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2341:   PetscCall(VecDestroy(&diagV));
2342:   PetscCall(VecDestroy(&offdiagV));
2343:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2344:   PetscFunctionReturn(PETSC_SUCCESS);
2345: }

2347: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2348: {
2349:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2350:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2351:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2352:   PetscInt          *cmap = mat->garray;
2353:   PetscInt          *diagIdx, *offdiagIdx;
2354:   Vec                diagV, offdiagV;
2355:   PetscScalar       *a, *diagA, *offdiagA;
2356:   const PetscScalar *ba, *bav;
2357:   PetscInt           r, j, col, ncols, *bi, *bj;
2358:   Mat                B = mat->B;
2359:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2361:   PetscFunctionBegin;
2362:   /* When a process holds entire A and other processes have no entry */
2363:   if (A->cmap->N == n) {
2364:     PetscCall(VecGetArrayWrite(v, &diagA));
2365:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2366:     PetscCall(MatGetRowMin(mat->A, diagV, idx));
2367:     PetscCall(VecDestroy(&diagV));
2368:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2369:     PetscFunctionReturn(PETSC_SUCCESS);
2370:   } else if (n == 0) {
2371:     if (m) {
2372:       PetscCall(VecGetArrayWrite(v, &a));
2373:       for (r = 0; r < m; r++) {
2374:         a[r] = PETSC_MAX_REAL;
2375:         if (idx) idx[r] = -1;
2376:       }
2377:       PetscCall(VecRestoreArrayWrite(v, &a));
2378:     }
2379:     PetscFunctionReturn(PETSC_SUCCESS);
2380:   }

2382:   PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2383:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2384:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2385:   PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));

2387:   /* Get offdiagIdx[] for implicit 0.0 */
2388:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2389:   ba = bav;
2390:   bi = b->i;
2391:   bj = b->j;
2392:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2393:   for (r = 0; r < m; r++) {
2394:     ncols = bi[r + 1] - bi[r];
2395:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2396:       offdiagA[r]   = *ba;
2397:       offdiagIdx[r] = cmap[0];
2398:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2399:       offdiagA[r] = 0.0;

2401:       /* Find first hole in the cmap */
2402:       for (j = 0; j < ncols; j++) {
2403:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2404:         if (col > j && j < cstart) {
2405:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2406:           break;
2407:         } else if (col > j + n && j >= cstart) {
2408:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2409:           break;
2410:         }
2411:       }
2412:       if (j == ncols && ncols < A->cmap->N - n) {
2413:         /* a hole is outside compressed Bcols */
2414:         if (ncols == 0) {
2415:           if (cstart) {
2416:             offdiagIdx[r] = 0;
2417:           } else offdiagIdx[r] = cend;
2418:         } else { /* ncols > 0 */
2419:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2420:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2421:         }
2422:       }
2423:     }

2425:     for (j = 0; j < ncols; j++) {
2426:       if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2427:         offdiagA[r]   = *ba;
2428:         offdiagIdx[r] = cmap[*bj];
2429:       }
2430:       ba++;
2431:       bj++;
2432:     }
2433:   }

2435:   PetscCall(VecGetArrayWrite(v, &a));
2436:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2437:   for (r = 0; r < m; ++r) {
2438:     if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2439:       a[r] = diagA[r];
2440:       if (idx) idx[r] = cstart + diagIdx[r];
2441:     } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2442:       a[r] = diagA[r];
2443:       if (idx) {
2444:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2445:           idx[r] = cstart + diagIdx[r];
2446:         } else idx[r] = offdiagIdx[r];
2447:       }
2448:     } else {
2449:       a[r] = offdiagA[r];
2450:       if (idx) idx[r] = offdiagIdx[r];
2451:     }
2452:   }
2453:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2454:   PetscCall(VecRestoreArrayWrite(v, &a));
2455:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2456:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2457:   PetscCall(VecDestroy(&diagV));
2458:   PetscCall(VecDestroy(&offdiagV));
2459:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2460:   PetscFunctionReturn(PETSC_SUCCESS);
2461: }

2463: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2464: {
2465:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2466:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2467:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2468:   PetscInt          *cmap = mat->garray;
2469:   PetscInt          *diagIdx, *offdiagIdx;
2470:   Vec                diagV, offdiagV;
2471:   PetscScalar       *a, *diagA, *offdiagA;
2472:   const PetscScalar *ba, *bav;
2473:   PetscInt           r, j, col, ncols, *bi, *bj;
2474:   Mat                B = mat->B;
2475:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2477:   PetscFunctionBegin;
2478:   /* When a process holds entire A and other processes have no entry */
2479:   if (A->cmap->N == n) {
2480:     PetscCall(VecGetArrayWrite(v, &diagA));
2481:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2482:     PetscCall(MatGetRowMax(mat->A, diagV, idx));
2483:     PetscCall(VecDestroy(&diagV));
2484:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2485:     PetscFunctionReturn(PETSC_SUCCESS);
2486:   } else if (n == 0) {
2487:     if (m) {
2488:       PetscCall(VecGetArrayWrite(v, &a));
2489:       for (r = 0; r < m; r++) {
2490:         a[r] = PETSC_MIN_REAL;
2491:         if (idx) idx[r] = -1;
2492:       }
2493:       PetscCall(VecRestoreArrayWrite(v, &a));
2494:     }
2495:     PetscFunctionReturn(PETSC_SUCCESS);
2496:   }

2498:   PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2499:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2500:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2501:   PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));

2503:   /* Get offdiagIdx[] for implicit 0.0 */
2504:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2505:   ba = bav;
2506:   bi = b->i;
2507:   bj = b->j;
2508:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2509:   for (r = 0; r < m; r++) {
2510:     ncols = bi[r + 1] - bi[r];
2511:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2512:       offdiagA[r]   = *ba;
2513:       offdiagIdx[r] = cmap[0];
2514:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2515:       offdiagA[r] = 0.0;

2517:       /* Find first hole in the cmap */
2518:       for (j = 0; j < ncols; j++) {
2519:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2520:         if (col > j && j < cstart) {
2521:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2522:           break;
2523:         } else if (col > j + n && j >= cstart) {
2524:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2525:           break;
2526:         }
2527:       }
2528:       if (j == ncols && ncols < A->cmap->N - n) {
2529:         /* a hole is outside compressed Bcols */
2530:         if (ncols == 0) {
2531:           if (cstart) {
2532:             offdiagIdx[r] = 0;
2533:           } else offdiagIdx[r] = cend;
2534:         } else { /* ncols > 0 */
2535:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2536:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2537:         }
2538:       }
2539:     }

2541:     for (j = 0; j < ncols; j++) {
2542:       if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2543:         offdiagA[r]   = *ba;
2544:         offdiagIdx[r] = cmap[*bj];
2545:       }
2546:       ba++;
2547:       bj++;
2548:     }
2549:   }

2551:   PetscCall(VecGetArrayWrite(v, &a));
2552:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2553:   for (r = 0; r < m; ++r) {
2554:     if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2555:       a[r] = diagA[r];
2556:       if (idx) idx[r] = cstart + diagIdx[r];
2557:     } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2558:       a[r] = diagA[r];
2559:       if (idx) {
2560:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2561:           idx[r] = cstart + diagIdx[r];
2562:         } else idx[r] = offdiagIdx[r];
2563:       }
2564:     } else {
2565:       a[r] = offdiagA[r];
2566:       if (idx) idx[r] = offdiagIdx[r];
2567:     }
2568:   }
2569:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2570:   PetscCall(VecRestoreArrayWrite(v, &a));
2571:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2572:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2573:   PetscCall(VecDestroy(&diagV));
2574:   PetscCall(VecDestroy(&offdiagV));
2575:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2576:   PetscFunctionReturn(PETSC_SUCCESS);
2577: }

2579: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2580: {
2581:   Mat *dummy;

2583:   PetscFunctionBegin;
2584:   PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2585:   *newmat = *dummy;
2586:   PetscCall(PetscFree(dummy));
2587:   PetscFunctionReturn(PETSC_SUCCESS);
2588: }

2590: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2591: {
2592:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2594:   PetscFunctionBegin;
2595:   PetscCall(MatInvertBlockDiagonal(a->A, values));
2596:   A->factorerrortype = a->A->factorerrortype;
2597:   PetscFunctionReturn(PETSC_SUCCESS);
2598: }

2600: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2601: {
2602:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;

2604:   PetscFunctionBegin;
2605:   PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2606:   PetscCall(MatSetRandom(aij->A, rctx));
2607:   if (x->assembled) {
2608:     PetscCall(MatSetRandom(aij->B, rctx));
2609:   } else {
2610:     PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2611:   }
2612:   PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2613:   PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2614:   PetscFunctionReturn(PETSC_SUCCESS);
2615: }

2617: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2618: {
2619:   PetscFunctionBegin;
2620:   if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2621:   else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2622:   PetscFunctionReturn(PETSC_SUCCESS);
2623: }

2625: /*@
2626:    MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank

2628:    Not Collective

2630:    Input Parameter:
2631: .    A - the matrix

2633:    Output Parameter:
2634: .    nz - the number of nonzeros

2636:  Level: advanced

2638: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `Mat`
2639: @*/
2640: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2641: {
2642:   Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2643:   Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;

2645:   PetscFunctionBegin;
2646:   *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2647:   PetscFunctionReturn(PETSC_SUCCESS);
2648: }

2650: /*@
2651:    MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap

2653:    Collective

2655:    Input Parameters:
2656: +    A - the matrix
2657: -    sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)

2659:  Level: advanced

2661: .seealso: [](chapter_matrices), `Mat`, `Mat`, `MATMPIAIJ`
2662: @*/
2663: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2664: {
2665:   PetscFunctionBegin;
2666:   PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2667:   PetscFunctionReturn(PETSC_SUCCESS);
2668: }

2670: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2671: {
2672:   PetscBool sc = PETSC_FALSE, flg;

2674:   PetscFunctionBegin;
2675:   PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2676:   if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2677:   PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2678:   if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2679:   PetscOptionsHeadEnd();
2680:   PetscFunctionReturn(PETSC_SUCCESS);
2681: }

2683: PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2684: {
2685:   Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2686:   Mat_SeqAIJ *aij  = (Mat_SeqAIJ *)maij->A->data;

2688:   PetscFunctionBegin;
2689:   if (!Y->preallocated) {
2690:     PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2691:   } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2692:     PetscInt nonew = aij->nonew;
2693:     PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2694:     aij->nonew = nonew;
2695:   }
2696:   PetscCall(MatShift_Basic(Y, a));
2697:   PetscFunctionReturn(PETSC_SUCCESS);
2698: }

2700: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2701: {
2702:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2704:   PetscFunctionBegin;
2705:   PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2706:   PetscCall(MatMissingDiagonal(a->A, missing, d));
2707:   if (d) {
2708:     PetscInt rstart;
2709:     PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2710:     *d += rstart;
2711:   }
2712:   PetscFunctionReturn(PETSC_SUCCESS);
2713: }

2715: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2716: {
2717:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2719:   PetscFunctionBegin;
2720:   PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2721:   PetscFunctionReturn(PETSC_SUCCESS);
2722: }

2724: PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A)
2725: {
2726:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2728:   PetscFunctionBegin;
2729:   PetscCall(MatEliminateZeros(a->A));
2730:   PetscCall(MatEliminateZeros(a->B));
2731:   PetscFunctionReturn(PETSC_SUCCESS);
2732: }

2734: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2735:                                        MatGetRow_MPIAIJ,
2736:                                        MatRestoreRow_MPIAIJ,
2737:                                        MatMult_MPIAIJ,
2738:                                        /* 4*/ MatMultAdd_MPIAIJ,
2739:                                        MatMultTranspose_MPIAIJ,
2740:                                        MatMultTransposeAdd_MPIAIJ,
2741:                                        NULL,
2742:                                        NULL,
2743:                                        NULL,
2744:                                        /*10*/ NULL,
2745:                                        NULL,
2746:                                        NULL,
2747:                                        MatSOR_MPIAIJ,
2748:                                        MatTranspose_MPIAIJ,
2749:                                        /*15*/ MatGetInfo_MPIAIJ,
2750:                                        MatEqual_MPIAIJ,
2751:                                        MatGetDiagonal_MPIAIJ,
2752:                                        MatDiagonalScale_MPIAIJ,
2753:                                        MatNorm_MPIAIJ,
2754:                                        /*20*/ MatAssemblyBegin_MPIAIJ,
2755:                                        MatAssemblyEnd_MPIAIJ,
2756:                                        MatSetOption_MPIAIJ,
2757:                                        MatZeroEntries_MPIAIJ,
2758:                                        /*24*/ MatZeroRows_MPIAIJ,
2759:                                        NULL,
2760:                                        NULL,
2761:                                        NULL,
2762:                                        NULL,
2763:                                        /*29*/ MatSetUp_MPI_Hash,
2764:                                        NULL,
2765:                                        NULL,
2766:                                        MatGetDiagonalBlock_MPIAIJ,
2767:                                        NULL,
2768:                                        /*34*/ MatDuplicate_MPIAIJ,
2769:                                        NULL,
2770:                                        NULL,
2771:                                        NULL,
2772:                                        NULL,
2773:                                        /*39*/ MatAXPY_MPIAIJ,
2774:                                        MatCreateSubMatrices_MPIAIJ,
2775:                                        MatIncreaseOverlap_MPIAIJ,
2776:                                        MatGetValues_MPIAIJ,
2777:                                        MatCopy_MPIAIJ,
2778:                                        /*44*/ MatGetRowMax_MPIAIJ,
2779:                                        MatScale_MPIAIJ,
2780:                                        MatShift_MPIAIJ,
2781:                                        MatDiagonalSet_MPIAIJ,
2782:                                        MatZeroRowsColumns_MPIAIJ,
2783:                                        /*49*/ MatSetRandom_MPIAIJ,
2784:                                        MatGetRowIJ_MPIAIJ,
2785:                                        MatRestoreRowIJ_MPIAIJ,
2786:                                        NULL,
2787:                                        NULL,
2788:                                        /*54*/ MatFDColoringCreate_MPIXAIJ,
2789:                                        NULL,
2790:                                        MatSetUnfactored_MPIAIJ,
2791:                                        MatPermute_MPIAIJ,
2792:                                        NULL,
2793:                                        /*59*/ MatCreateSubMatrix_MPIAIJ,
2794:                                        MatDestroy_MPIAIJ,
2795:                                        MatView_MPIAIJ,
2796:                                        NULL,
2797:                                        NULL,
2798:                                        /*64*/ NULL,
2799:                                        MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2800:                                        NULL,
2801:                                        NULL,
2802:                                        NULL,
2803:                                        /*69*/ MatGetRowMaxAbs_MPIAIJ,
2804:                                        MatGetRowMinAbs_MPIAIJ,
2805:                                        NULL,
2806:                                        NULL,
2807:                                        NULL,
2808:                                        NULL,
2809:                                        /*75*/ MatFDColoringApply_AIJ,
2810:                                        MatSetFromOptions_MPIAIJ,
2811:                                        NULL,
2812:                                        NULL,
2813:                                        MatFindZeroDiagonals_MPIAIJ,
2814:                                        /*80*/ NULL,
2815:                                        NULL,
2816:                                        NULL,
2817:                                        /*83*/ MatLoad_MPIAIJ,
2818:                                        MatIsSymmetric_MPIAIJ,
2819:                                        NULL,
2820:                                        NULL,
2821:                                        NULL,
2822:                                        NULL,
2823:                                        /*89*/ NULL,
2824:                                        NULL,
2825:                                        MatMatMultNumeric_MPIAIJ_MPIAIJ,
2826:                                        NULL,
2827:                                        NULL,
2828:                                        /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2829:                                        NULL,
2830:                                        NULL,
2831:                                        NULL,
2832:                                        MatBindToCPU_MPIAIJ,
2833:                                        /*99*/ MatProductSetFromOptions_MPIAIJ,
2834:                                        NULL,
2835:                                        NULL,
2836:                                        MatConjugate_MPIAIJ,
2837:                                        NULL,
2838:                                        /*104*/ MatSetValuesRow_MPIAIJ,
2839:                                        MatRealPart_MPIAIJ,
2840:                                        MatImaginaryPart_MPIAIJ,
2841:                                        NULL,
2842:                                        NULL,
2843:                                        /*109*/ NULL,
2844:                                        NULL,
2845:                                        MatGetRowMin_MPIAIJ,
2846:                                        NULL,
2847:                                        MatMissingDiagonal_MPIAIJ,
2848:                                        /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2849:                                        NULL,
2850:                                        MatGetGhosts_MPIAIJ,
2851:                                        NULL,
2852:                                        NULL,
2853:                                        /*119*/ MatMultDiagonalBlock_MPIAIJ,
2854:                                        NULL,
2855:                                        NULL,
2856:                                        NULL,
2857:                                        MatGetMultiProcBlock_MPIAIJ,
2858:                                        /*124*/ MatFindNonzeroRows_MPIAIJ,
2859:                                        MatGetColumnReductions_MPIAIJ,
2860:                                        MatInvertBlockDiagonal_MPIAIJ,
2861:                                        MatInvertVariableBlockDiagonal_MPIAIJ,
2862:                                        MatCreateSubMatricesMPI_MPIAIJ,
2863:                                        /*129*/ NULL,
2864:                                        NULL,
2865:                                        NULL,
2866:                                        MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2867:                                        NULL,
2868:                                        /*134*/ NULL,
2869:                                        NULL,
2870:                                        NULL,
2871:                                        NULL,
2872:                                        NULL,
2873:                                        /*139*/ MatSetBlockSizes_MPIAIJ,
2874:                                        NULL,
2875:                                        NULL,
2876:                                        MatFDColoringSetUp_MPIXAIJ,
2877:                                        MatFindOffBlockDiagonalEntries_MPIAIJ,
2878:                                        MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2879:                                        /*145*/ NULL,
2880:                                        NULL,
2881:                                        NULL,
2882:                                        MatCreateGraph_Simple_AIJ,
2883:                                        NULL,
2884:                                        /*150*/ NULL,
2885:                                        MatEliminateZeros_MPIAIJ};

2887: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2888: {
2889:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2891:   PetscFunctionBegin;
2892:   PetscCall(MatStoreValues(aij->A));
2893:   PetscCall(MatStoreValues(aij->B));
2894:   PetscFunctionReturn(PETSC_SUCCESS);
2895: }

2897: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2898: {
2899:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2901:   PetscFunctionBegin;
2902:   PetscCall(MatRetrieveValues(aij->A));
2903:   PetscCall(MatRetrieveValues(aij->B));
2904:   PetscFunctionReturn(PETSC_SUCCESS);
2905: }

2907: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2908: {
2909:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2910:   PetscMPIInt size;

2912:   PetscFunctionBegin;
2913:   if (B->hash_active) {
2914:     PetscCall(PetscMemcpy(&B->ops, &b->cops, sizeof(*(B->ops))));
2915:     B->hash_active = PETSC_FALSE;
2916:   }
2917:   PetscCall(PetscLayoutSetUp(B->rmap));
2918:   PetscCall(PetscLayoutSetUp(B->cmap));

2920: #if defined(PETSC_USE_CTABLE)
2921:   PetscCall(PetscHMapIDestroy(&b->colmap));
2922: #else
2923:   PetscCall(PetscFree(b->colmap));
2924: #endif
2925:   PetscCall(PetscFree(b->garray));
2926:   PetscCall(VecDestroy(&b->lvec));
2927:   PetscCall(VecScatterDestroy(&b->Mvctx));

2929:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2930:   PetscCall(MatDestroy(&b->B));
2931:   PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2932:   PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2933:   PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2934:   PetscCall(MatSetType(b->B, MATSEQAIJ));

2936:   PetscCall(MatDestroy(&b->A));
2937:   PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2938:   PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2939:   PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2940:   PetscCall(MatSetType(b->A, MATSEQAIJ));

2942:   PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2943:   PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2944:   B->preallocated  = PETSC_TRUE;
2945:   B->was_assembled = PETSC_FALSE;
2946:   B->assembled     = PETSC_FALSE;
2947:   PetscFunctionReturn(PETSC_SUCCESS);
2948: }

2950: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2951: {
2952:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;

2954:   PetscFunctionBegin;
2956:   PetscCall(PetscLayoutSetUp(B->rmap));
2957:   PetscCall(PetscLayoutSetUp(B->cmap));

2959: #if defined(PETSC_USE_CTABLE)
2960:   PetscCall(PetscHMapIDestroy(&b->colmap));
2961: #else
2962:   PetscCall(PetscFree(b->colmap));
2963: #endif
2964:   PetscCall(PetscFree(b->garray));
2965:   PetscCall(VecDestroy(&b->lvec));
2966:   PetscCall(VecScatterDestroy(&b->Mvctx));

2968:   PetscCall(MatResetPreallocation(b->A));
2969:   PetscCall(MatResetPreallocation(b->B));
2970:   B->preallocated  = PETSC_TRUE;
2971:   B->was_assembled = PETSC_FALSE;
2972:   B->assembled     = PETSC_FALSE;
2973:   PetscFunctionReturn(PETSC_SUCCESS);
2974: }

2976: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2977: {
2978:   Mat         mat;
2979:   Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;

2981:   PetscFunctionBegin;
2982:   *newmat = NULL;
2983:   PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2984:   PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2985:   PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2986:   PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2987:   a = (Mat_MPIAIJ *)mat->data;

2989:   mat->factortype   = matin->factortype;
2990:   mat->assembled    = matin->assembled;
2991:   mat->insertmode   = NOT_SET_VALUES;
2992:   mat->preallocated = matin->preallocated;

2994:   a->size         = oldmat->size;
2995:   a->rank         = oldmat->rank;
2996:   a->donotstash   = oldmat->donotstash;
2997:   a->roworiented  = oldmat->roworiented;
2998:   a->rowindices   = NULL;
2999:   a->rowvalues    = NULL;
3000:   a->getrowactive = PETSC_FALSE;

3002:   PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
3003:   PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));

3005:   if (oldmat->colmap) {
3006: #if defined(PETSC_USE_CTABLE)
3007:     PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3008: #else
3009:     PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3010:     PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3011: #endif
3012:   } else a->colmap = NULL;
3013:   if (oldmat->garray) {
3014:     PetscInt len;
3015:     len = oldmat->B->cmap->n;
3016:     PetscCall(PetscMalloc1(len + 1, &a->garray));
3017:     if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3018:   } else a->garray = NULL;

3020:   /* It may happen MatDuplicate is called with a non-assembled matrix
3021:      In fact, MatDuplicate only requires the matrix to be preallocated
3022:      This may happen inside a DMCreateMatrix_Shell */
3023:   if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3024:   if (oldmat->Mvctx) PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3025:   PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3026:   PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3027:   PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3028:   *newmat = mat;
3029:   PetscFunctionReturn(PETSC_SUCCESS);
3030: }

3032: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3033: {
3034:   PetscBool isbinary, ishdf5;

3036:   PetscFunctionBegin;
3039:   /* force binary viewer to load .info file if it has not yet done so */
3040:   PetscCall(PetscViewerSetUp(viewer));
3041:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3042:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3043:   if (isbinary) {
3044:     PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3045:   } else if (ishdf5) {
3046: #if defined(PETSC_HAVE_HDF5)
3047:     PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3048: #else
3049:     SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3050: #endif
3051:   } else {
3052:     SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "Viewer type %s not yet supported for reading %s matrices", ((PetscObject)viewer)->type_name, ((PetscObject)newMat)->type_name);
3053:   }
3054:   PetscFunctionReturn(PETSC_SUCCESS);
3055: }

3057: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3058: {
3059:   PetscInt     header[4], M, N, m, nz, rows, cols, sum, i;
3060:   PetscInt    *rowidxs, *colidxs;
3061:   PetscScalar *matvals;

3063:   PetscFunctionBegin;
3064:   PetscCall(PetscViewerSetUp(viewer));

3066:   /* read in matrix header */
3067:   PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3068:   PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3069:   M  = header[1];
3070:   N  = header[2];
3071:   nz = header[3];
3072:   PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3073:   PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3074:   PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");

3076:   /* set block sizes from the viewer's .info file */
3077:   PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3078:   /* set global sizes if not set already */
3079:   if (mat->rmap->N < 0) mat->rmap->N = M;
3080:   if (mat->cmap->N < 0) mat->cmap->N = N;
3081:   PetscCall(PetscLayoutSetUp(mat->rmap));
3082:   PetscCall(PetscLayoutSetUp(mat->cmap));

3084:   /* check if the matrix sizes are correct */
3085:   PetscCall(MatGetSize(mat, &rows, &cols));
3086:   PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols);

3088:   /* read in row lengths and build row indices */
3089:   PetscCall(MatGetLocalSize(mat, &m, NULL));
3090:   PetscCall(PetscMalloc1(m + 1, &rowidxs));
3091:   PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3092:   rowidxs[0] = 0;
3093:   for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3094:   PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3095:   PetscCheck(sum == nz, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum);
3096:   /* read in column indices and matrix values */
3097:   PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3098:   PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3099:   PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3100:   /* store matrix indices and values */
3101:   PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3102:   PetscCall(PetscFree(rowidxs));
3103:   PetscCall(PetscFree2(colidxs, matvals));
3104:   PetscFunctionReturn(PETSC_SUCCESS);
3105: }

3107: /* Not scalable because of ISAllGather() unless getting all columns. */
3108: PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3109: {
3110:   IS          iscol_local;
3111:   PetscBool   isstride;
3112:   PetscMPIInt lisstride = 0, gisstride;

3114:   PetscFunctionBegin;
3115:   /* check if we are grabbing all columns*/
3116:   PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));

3118:   if (isstride) {
3119:     PetscInt start, len, mstart, mlen;
3120:     PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3121:     PetscCall(ISGetLocalSize(iscol, &len));
3122:     PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3123:     if (mstart == start && mlen - mstart == len) lisstride = 1;
3124:   }

3126:   PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3127:   if (gisstride) {
3128:     PetscInt N;
3129:     PetscCall(MatGetSize(mat, NULL, &N));
3130:     PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3131:     PetscCall(ISSetIdentity(iscol_local));
3132:     PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3133:   } else {
3134:     PetscInt cbs;
3135:     PetscCall(ISGetBlockSize(iscol, &cbs));
3136:     PetscCall(ISAllGather(iscol, &iscol_local));
3137:     PetscCall(ISSetBlockSize(iscol_local, cbs));
3138:   }

3140:   *isseq = iscol_local;
3141:   PetscFunctionReturn(PETSC_SUCCESS);
3142: }

3144: /*
3145:  Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3146:  (see MatCreateSubMatrix_MPIAIJ_nonscalable)

3148:  Input Parameters:
3149: +   mat - matrix
3150: .   isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3151:            i.e., mat->rstart <= isrow[i] < mat->rend
3152: -   iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3153:            i.e., mat->cstart <= iscol[i] < mat->cend

3155:  Output Parameters:
3156: +   isrow_d - sequential row index set for retrieving mat->A
3157: .   iscol_d - sequential  column index set for retrieving mat->A
3158: .   iscol_o - sequential column index set for retrieving mat->B
3159: -   garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3160:  */
3161: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3162: {
3163:   Vec             x, cmap;
3164:   const PetscInt *is_idx;
3165:   PetscScalar    *xarray, *cmaparray;
3166:   PetscInt        ncols, isstart, *idx, m, rstart, *cmap1, count;
3167:   Mat_MPIAIJ     *a    = (Mat_MPIAIJ *)mat->data;
3168:   Mat             B    = a->B;
3169:   Vec             lvec = a->lvec, lcmap;
3170:   PetscInt        i, cstart, cend, Bn = B->cmap->N;
3171:   MPI_Comm        comm;
3172:   VecScatter      Mvctx = a->Mvctx;

3174:   PetscFunctionBegin;
3175:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3176:   PetscCall(ISGetLocalSize(iscol, &ncols));

3178:   /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3179:   PetscCall(MatCreateVecs(mat, &x, NULL));
3180:   PetscCall(VecSet(x, -1.0));
3181:   PetscCall(VecDuplicate(x, &cmap));
3182:   PetscCall(VecSet(cmap, -1.0));

3184:   /* Get start indices */
3185:   PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3186:   isstart -= ncols;
3187:   PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));

3189:   PetscCall(ISGetIndices(iscol, &is_idx));
3190:   PetscCall(VecGetArray(x, &xarray));
3191:   PetscCall(VecGetArray(cmap, &cmaparray));
3192:   PetscCall(PetscMalloc1(ncols, &idx));
3193:   for (i = 0; i < ncols; i++) {
3194:     xarray[is_idx[i] - cstart]    = (PetscScalar)is_idx[i];
3195:     cmaparray[is_idx[i] - cstart] = i + isstart;        /* global index of iscol[i] */
3196:     idx[i]                        = is_idx[i] - cstart; /* local index of iscol[i]  */
3197:   }
3198:   PetscCall(VecRestoreArray(x, &xarray));
3199:   PetscCall(VecRestoreArray(cmap, &cmaparray));
3200:   PetscCall(ISRestoreIndices(iscol, &is_idx));

3202:   /* Get iscol_d */
3203:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3204:   PetscCall(ISGetBlockSize(iscol, &i));
3205:   PetscCall(ISSetBlockSize(*iscol_d, i));

3207:   /* Get isrow_d */
3208:   PetscCall(ISGetLocalSize(isrow, &m));
3209:   rstart = mat->rmap->rstart;
3210:   PetscCall(PetscMalloc1(m, &idx));
3211:   PetscCall(ISGetIndices(isrow, &is_idx));
3212:   for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3213:   PetscCall(ISRestoreIndices(isrow, &is_idx));

3215:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3216:   PetscCall(ISGetBlockSize(isrow, &i));
3217:   PetscCall(ISSetBlockSize(*isrow_d, i));

3219:   /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3220:   PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3221:   PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));

3223:   PetscCall(VecDuplicate(lvec, &lcmap));

3225:   PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3226:   PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));

3228:   /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3229:   /* off-process column indices */
3230:   count = 0;
3231:   PetscCall(PetscMalloc1(Bn, &idx));
3232:   PetscCall(PetscMalloc1(Bn, &cmap1));

3234:   PetscCall(VecGetArray(lvec, &xarray));
3235:   PetscCall(VecGetArray(lcmap, &cmaparray));
3236:   for (i = 0; i < Bn; i++) {
3237:     if (PetscRealPart(xarray[i]) > -1.0) {
3238:       idx[count]   = i;                                     /* local column index in off-diagonal part B */
3239:       cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3240:       count++;
3241:     }
3242:   }
3243:   PetscCall(VecRestoreArray(lvec, &xarray));
3244:   PetscCall(VecRestoreArray(lcmap, &cmaparray));

3246:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3247:   /* cannot ensure iscol_o has same blocksize as iscol! */

3249:   PetscCall(PetscFree(idx));
3250:   *garray = cmap1;

3252:   PetscCall(VecDestroy(&x));
3253:   PetscCall(VecDestroy(&cmap));
3254:   PetscCall(VecDestroy(&lcmap));
3255:   PetscFunctionReturn(PETSC_SUCCESS);
3256: }

3258: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3259: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3260: {
3261:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3262:   Mat         M = NULL;
3263:   MPI_Comm    comm;
3264:   IS          iscol_d, isrow_d, iscol_o;
3265:   Mat         Asub = NULL, Bsub = NULL;
3266:   PetscInt    n;

3268:   PetscFunctionBegin;
3269:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));

3271:   if (call == MAT_REUSE_MATRIX) {
3272:     /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3273:     PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3274:     PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");

3276:     PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3277:     PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");

3279:     PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3280:     PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");

3282:     /* Update diagonal and off-diagonal portions of submat */
3283:     asub = (Mat_MPIAIJ *)(*submat)->data;
3284:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3285:     PetscCall(ISGetLocalSize(iscol_o, &n));
3286:     if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3287:     PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3288:     PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));

3290:   } else { /* call == MAT_INITIAL_MATRIX) */
3291:     const PetscInt *garray;
3292:     PetscInt        BsubN;

3294:     /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3295:     PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));

3297:     /* Create local submatrices Asub and Bsub */
3298:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3299:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));

3301:     /* Create submatrix M */
3302:     PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));

3304:     /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3305:     asub = (Mat_MPIAIJ *)M->data;

3307:     PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3308:     n = asub->B->cmap->N;
3309:     if (BsubN > n) {
3310:       /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3311:       const PetscInt *idx;
3312:       PetscInt        i, j, *idx_new, *subgarray = asub->garray;
3313:       PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));

3315:       PetscCall(PetscMalloc1(n, &idx_new));
3316:       j = 0;
3317:       PetscCall(ISGetIndices(iscol_o, &idx));
3318:       for (i = 0; i < n; i++) {
3319:         if (j >= BsubN) break;
3320:         while (subgarray[i] > garray[j]) j++;

3322:         if (subgarray[i] == garray[j]) {
3323:           idx_new[i] = idx[j++];
3324:         } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "subgarray[%" PetscInt_FMT "]=%" PetscInt_FMT " cannot < garray[%" PetscInt_FMT "]=%" PetscInt_FMT, i, subgarray[i], j, garray[j]);
3325:       }
3326:       PetscCall(ISRestoreIndices(iscol_o, &idx));

3328:       PetscCall(ISDestroy(&iscol_o));
3329:       PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));

3331:     } else if (BsubN < n) {
3332:       SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Columns of Bsub (%" PetscInt_FMT ") cannot be smaller than B's (%" PetscInt_FMT ")", BsubN, asub->B->cmap->N);
3333:     }

3335:     PetscCall(PetscFree(garray));
3336:     *submat = M;

3338:     /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3339:     PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3340:     PetscCall(ISDestroy(&isrow_d));

3342:     PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3343:     PetscCall(ISDestroy(&iscol_d));

3345:     PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3346:     PetscCall(ISDestroy(&iscol_o));
3347:   }
3348:   PetscFunctionReturn(PETSC_SUCCESS);
3349: }

3351: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3352: {
3353:   IS        iscol_local = NULL, isrow_d;
3354:   PetscInt  csize;
3355:   PetscInt  n, i, j, start, end;
3356:   PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3357:   MPI_Comm  comm;

3359:   PetscFunctionBegin;
3360:   /* If isrow has same processor distribution as mat,
3361:      call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3362:   if (call == MAT_REUSE_MATRIX) {
3363:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3364:     if (isrow_d) {
3365:       sameRowDist  = PETSC_TRUE;
3366:       tsameDist[1] = PETSC_TRUE; /* sameColDist */
3367:     } else {
3368:       PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3369:       if (iscol_local) {
3370:         sameRowDist  = PETSC_TRUE;
3371:         tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3372:       }
3373:     }
3374:   } else {
3375:     /* Check if isrow has same processor distribution as mat */
3376:     sameDist[0] = PETSC_FALSE;
3377:     PetscCall(ISGetLocalSize(isrow, &n));
3378:     if (!n) {
3379:       sameDist[0] = PETSC_TRUE;
3380:     } else {
3381:       PetscCall(ISGetMinMax(isrow, &i, &j));
3382:       PetscCall(MatGetOwnershipRange(mat, &start, &end));
3383:       if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3384:     }

3386:     /* Check if iscol has same processor distribution as mat */
3387:     sameDist[1] = PETSC_FALSE;
3388:     PetscCall(ISGetLocalSize(iscol, &n));
3389:     if (!n) {
3390:       sameDist[1] = PETSC_TRUE;
3391:     } else {
3392:       PetscCall(ISGetMinMax(iscol, &i, &j));
3393:       PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3394:       if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3395:     }

3397:     PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3398:     PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3399:     sameRowDist = tsameDist[0];
3400:   }

3402:   if (sameRowDist) {
3403:     if (tsameDist[1]) { /* sameRowDist & sameColDist */
3404:       /* isrow and iscol have same processor distribution as mat */
3405:       PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3406:       PetscFunctionReturn(PETSC_SUCCESS);
3407:     } else { /* sameRowDist */
3408:       /* isrow has same processor distribution as mat */
3409:       if (call == MAT_INITIAL_MATRIX) {
3410:         PetscBool sorted;
3411:         PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3412:         PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3413:         PetscCall(ISGetSize(iscol, &i));
3414:         PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);

3416:         PetscCall(ISSorted(iscol_local, &sorted));
3417:         if (sorted) {
3418:           /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3419:           PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3420:           PetscFunctionReturn(PETSC_SUCCESS);
3421:         }
3422:       } else { /* call == MAT_REUSE_MATRIX */
3423:         IS iscol_sub;
3424:         PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3425:         if (iscol_sub) {
3426:           PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3427:           PetscFunctionReturn(PETSC_SUCCESS);
3428:         }
3429:       }
3430:     }
3431:   }

3433:   /* General case: iscol -> iscol_local which has global size of iscol */
3434:   if (call == MAT_REUSE_MATRIX) {
3435:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3436:     PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3437:   } else {
3438:     if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3439:   }

3441:   PetscCall(ISGetLocalSize(iscol, &csize));
3442:   PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));

3444:   if (call == MAT_INITIAL_MATRIX) {
3445:     PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3446:     PetscCall(ISDestroy(&iscol_local));
3447:   }
3448:   PetscFunctionReturn(PETSC_SUCCESS);
3449: }

3451: /*@C
3452:      MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3453:          and "off-diagonal" part of the matrix in CSR format.

3455:    Collective

3457:    Input Parameters:
3458: +  comm - MPI communicator
3459: .  A - "diagonal" portion of matrix
3460: .  B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3461: -  garray - global index of `B` columns

3463:    Output Parameter:
3464: .   mat - the matrix, with input `A` as its local diagonal matrix

3466:   Level: advanced

3468:    Notes:
3469:    See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.

3471:    `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.

3473: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3474: @*/
3475: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3476: {
3477:   Mat_MPIAIJ        *maij;
3478:   Mat_SeqAIJ        *b  = (Mat_SeqAIJ *)B->data, *bnew;
3479:   PetscInt          *oi = b->i, *oj = b->j, i, nz, col;
3480:   const PetscScalar *oa;
3481:   Mat                Bnew;
3482:   PetscInt           m, n, N;
3483:   MatType            mpi_mat_type;

3485:   PetscFunctionBegin;
3486:   PetscCall(MatCreate(comm, mat));
3487:   PetscCall(MatGetSize(A, &m, &n));
3488:   PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3489:   PetscCheck(A->rmap->bs == B->rmap->bs, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "A row bs %" PetscInt_FMT " != B row bs %" PetscInt_FMT, A->rmap->bs, B->rmap->bs);
3490:   /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3491:   /* PetscCheck(A->cmap->bs == B->cmap->bs,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %" PetscInt_FMT " != B column bs %" PetscInt_FMT,A->cmap->bs,B->cmap->bs); */

3493:   /* Get global columns of mat */
3494:   PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));

3496:   PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3497:   /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3498:   PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3499:   PetscCall(MatSetType(*mat, mpi_mat_type));

3501:   PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3502:   maij = (Mat_MPIAIJ *)(*mat)->data;

3504:   (*mat)->preallocated = PETSC_TRUE;

3506:   PetscCall(PetscLayoutSetUp((*mat)->rmap));
3507:   PetscCall(PetscLayoutSetUp((*mat)->cmap));

3509:   /* Set A as diagonal portion of *mat */
3510:   maij->A = A;

3512:   nz = oi[m];
3513:   for (i = 0; i < nz; i++) {
3514:     col   = oj[i];
3515:     oj[i] = garray[col];
3516:   }

3518:   /* Set Bnew as off-diagonal portion of *mat */
3519:   PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3520:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3521:   PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3522:   bnew        = (Mat_SeqAIJ *)Bnew->data;
3523:   bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3524:   maij->B     = Bnew;

3526:   PetscCheck(B->rmap->N == Bnew->rmap->N, PETSC_COMM_SELF, PETSC_ERR_PLIB, "BN %" PetscInt_FMT " != BnewN %" PetscInt_FMT, B->rmap->N, Bnew->rmap->N);

3528:   b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3529:   b->free_a       = PETSC_FALSE;
3530:   b->free_ij      = PETSC_FALSE;
3531:   PetscCall(MatDestroy(&B));

3533:   bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3534:   bnew->free_a       = PETSC_TRUE;
3535:   bnew->free_ij      = PETSC_TRUE;

3537:   /* condense columns of maij->B */
3538:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3539:   PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3540:   PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3541:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3542:   PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3543:   PetscFunctionReturn(PETSC_SUCCESS);
3544: }

3546: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);

3548: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3549: {
3550:   PetscInt        i, m, n, rstart, row, rend, nz, j, bs, cbs;
3551:   PetscInt       *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3552:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)mat->data;
3553:   Mat             M, Msub, B = a->B;
3554:   MatScalar      *aa;
3555:   Mat_SeqAIJ     *aij;
3556:   PetscInt       *garray = a->garray, *colsub, Ncols;
3557:   PetscInt        count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3558:   IS              iscol_sub, iscmap;
3559:   const PetscInt *is_idx, *cmap;
3560:   PetscBool       allcolumns = PETSC_FALSE;
3561:   MPI_Comm        comm;

3563:   PetscFunctionBegin;
3564:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3565:   if (call == MAT_REUSE_MATRIX) {
3566:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3567:     PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3568:     PetscCall(ISGetLocalSize(iscol_sub, &count));

3570:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3571:     PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");

3573:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3574:     PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");

3576:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));

3578:   } else { /* call == MAT_INITIAL_MATRIX) */
3579:     PetscBool flg;

3581:     PetscCall(ISGetLocalSize(iscol, &n));
3582:     PetscCall(ISGetSize(iscol, &Ncols));

3584:     /* (1) iscol -> nonscalable iscol_local */
3585:     /* Check for special case: each processor gets entire matrix columns */
3586:     PetscCall(ISIdentity(iscol_local, &flg));
3587:     if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3588:     PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3589:     if (allcolumns) {
3590:       iscol_sub = iscol_local;
3591:       PetscCall(PetscObjectReference((PetscObject)iscol_local));
3592:       PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));

3594:     } else {
3595:       /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3596:       PetscInt *idx, *cmap1, k;
3597:       PetscCall(PetscMalloc1(Ncols, &idx));
3598:       PetscCall(PetscMalloc1(Ncols, &cmap1));
3599:       PetscCall(ISGetIndices(iscol_local, &is_idx));
3600:       count = 0;
3601:       k     = 0;
3602:       for (i = 0; i < Ncols; i++) {
3603:         j = is_idx[i];
3604:         if (j >= cstart && j < cend) {
3605:           /* diagonal part of mat */
3606:           idx[count]     = j;
3607:           cmap1[count++] = i; /* column index in submat */
3608:         } else if (Bn) {
3609:           /* off-diagonal part of mat */
3610:           if (j == garray[k]) {
3611:             idx[count]     = j;
3612:             cmap1[count++] = i; /* column index in submat */
3613:           } else if (j > garray[k]) {
3614:             while (j > garray[k] && k < Bn - 1) k++;
3615:             if (j == garray[k]) {
3616:               idx[count]     = j;
3617:               cmap1[count++] = i; /* column index in submat */
3618:             }
3619:           }
3620:         }
3621:       }
3622:       PetscCall(ISRestoreIndices(iscol_local, &is_idx));

3624:       PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3625:       PetscCall(ISGetBlockSize(iscol, &cbs));
3626:       PetscCall(ISSetBlockSize(iscol_sub, cbs));

3628:       PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3629:     }

3631:     /* (3) Create sequential Msub */
3632:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3633:   }

3635:   PetscCall(ISGetLocalSize(iscol_sub, &count));
3636:   aij = (Mat_SeqAIJ *)(Msub)->data;
3637:   ii  = aij->i;
3638:   PetscCall(ISGetIndices(iscmap, &cmap));

3640:   /*
3641:       m - number of local rows
3642:       Ncols - number of columns (same on all processors)
3643:       rstart - first row in new global matrix generated
3644:   */
3645:   PetscCall(MatGetSize(Msub, &m, NULL));

3647:   if (call == MAT_INITIAL_MATRIX) {
3648:     /* (4) Create parallel newmat */
3649:     PetscMPIInt rank, size;
3650:     PetscInt    csize;

3652:     PetscCallMPI(MPI_Comm_size(comm, &size));
3653:     PetscCallMPI(MPI_Comm_rank(comm, &rank));

3655:     /*
3656:         Determine the number of non-zeros in the diagonal and off-diagonal
3657:         portions of the matrix in order to do correct preallocation
3658:     */

3660:     /* first get start and end of "diagonal" columns */
3661:     PetscCall(ISGetLocalSize(iscol, &csize));
3662:     if (csize == PETSC_DECIDE) {
3663:       PetscCall(ISGetSize(isrow, &mglobal));
3664:       if (mglobal == Ncols) { /* square matrix */
3665:         nlocal = m;
3666:       } else {
3667:         nlocal = Ncols / size + ((Ncols % size) > rank);
3668:       }
3669:     } else {
3670:       nlocal = csize;
3671:     }
3672:     PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3673:     rstart = rend - nlocal;
3674:     PetscCheck(rank != size - 1 || rend == Ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, Ncols);

3676:     /* next, compute all the lengths */
3677:     jj = aij->j;
3678:     PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3679:     olens = dlens + m;
3680:     for (i = 0; i < m; i++) {
3681:       jend = ii[i + 1] - ii[i];
3682:       olen = 0;
3683:       dlen = 0;
3684:       for (j = 0; j < jend; j++) {
3685:         if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3686:         else dlen++;
3687:         jj++;
3688:       }
3689:       olens[i] = olen;
3690:       dlens[i] = dlen;
3691:     }

3693:     PetscCall(ISGetBlockSize(isrow, &bs));
3694:     PetscCall(ISGetBlockSize(iscol, &cbs));

3696:     PetscCall(MatCreate(comm, &M));
3697:     PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3698:     PetscCall(MatSetBlockSizes(M, bs, cbs));
3699:     PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3700:     PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3701:     PetscCall(PetscFree(dlens));

3703:   } else { /* call == MAT_REUSE_MATRIX */
3704:     M = *newmat;
3705:     PetscCall(MatGetLocalSize(M, &i, NULL));
3706:     PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3707:     PetscCall(MatZeroEntries(M));
3708:     /*
3709:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3710:        rather than the slower MatSetValues().
3711:     */
3712:     M->was_assembled = PETSC_TRUE;
3713:     M->assembled     = PETSC_FALSE;
3714:   }

3716:   /* (5) Set values of Msub to *newmat */
3717:   PetscCall(PetscMalloc1(count, &colsub));
3718:   PetscCall(MatGetOwnershipRange(M, &rstart, NULL));

3720:   jj = aij->j;
3721:   PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3722:   for (i = 0; i < m; i++) {
3723:     row = rstart + i;
3724:     nz  = ii[i + 1] - ii[i];
3725:     for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3726:     PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3727:     jj += nz;
3728:     aa += nz;
3729:   }
3730:   PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3731:   PetscCall(ISRestoreIndices(iscmap, &cmap));

3733:   PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3734:   PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));

3736:   PetscCall(PetscFree(colsub));

3738:   /* save Msub, iscol_sub and iscmap used in processor for next request */
3739:   if (call == MAT_INITIAL_MATRIX) {
3740:     *newmat = M;
3741:     PetscCall(PetscObjectCompose((PetscObject)(*newmat), "SubMatrix", (PetscObject)Msub));
3742:     PetscCall(MatDestroy(&Msub));

3744:     PetscCall(PetscObjectCompose((PetscObject)(*newmat), "SubIScol", (PetscObject)iscol_sub));
3745:     PetscCall(ISDestroy(&iscol_sub));

3747:     PetscCall(PetscObjectCompose((PetscObject)(*newmat), "Subcmap", (PetscObject)iscmap));
3748:     PetscCall(ISDestroy(&iscmap));

3750:     if (iscol_local) {
3751:       PetscCall(PetscObjectCompose((PetscObject)(*newmat), "ISAllGather", (PetscObject)iscol_local));
3752:       PetscCall(ISDestroy(&iscol_local));
3753:     }
3754:   }
3755:   PetscFunctionReturn(PETSC_SUCCESS);
3756: }

3758: /*
3759:     Not great since it makes two copies of the submatrix, first an SeqAIJ
3760:   in local and then by concatenating the local matrices the end result.
3761:   Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()

3763:   This requires a sequential iscol with all indices.
3764: */
3765: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3766: {
3767:   PetscMPIInt rank, size;
3768:   PetscInt    i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3769:   PetscInt   *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3770:   Mat         M, Mreuse;
3771:   MatScalar  *aa, *vwork;
3772:   MPI_Comm    comm;
3773:   Mat_SeqAIJ *aij;
3774:   PetscBool   colflag, allcolumns = PETSC_FALSE;

3776:   PetscFunctionBegin;
3777:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3778:   PetscCallMPI(MPI_Comm_rank(comm, &rank));
3779:   PetscCallMPI(MPI_Comm_size(comm, &size));

3781:   /* Check for special case: each processor gets entire matrix columns */
3782:   PetscCall(ISIdentity(iscol, &colflag));
3783:   PetscCall(ISGetLocalSize(iscol, &n));
3784:   if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3785:   PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));

3787:   if (call == MAT_REUSE_MATRIX) {
3788:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3789:     PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3790:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3791:   } else {
3792:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3793:   }

3795:   /*
3796:       m - number of local rows
3797:       n - number of columns (same on all processors)
3798:       rstart - first row in new global matrix generated
3799:   */
3800:   PetscCall(MatGetSize(Mreuse, &m, &n));
3801:   PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3802:   if (call == MAT_INITIAL_MATRIX) {
3803:     aij = (Mat_SeqAIJ *)(Mreuse)->data;
3804:     ii  = aij->i;
3805:     jj  = aij->j;

3807:     /*
3808:         Determine the number of non-zeros in the diagonal and off-diagonal
3809:         portions of the matrix in order to do correct preallocation
3810:     */

3812:     /* first get start and end of "diagonal" columns */
3813:     if (csize == PETSC_DECIDE) {
3814:       PetscCall(ISGetSize(isrow, &mglobal));
3815:       if (mglobal == n) { /* square matrix */
3816:         nlocal = m;
3817:       } else {
3818:         nlocal = n / size + ((n % size) > rank);
3819:       }
3820:     } else {
3821:       nlocal = csize;
3822:     }
3823:     PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3824:     rstart = rend - nlocal;
3825:     PetscCheck(rank != size - 1 || rend == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, n);

3827:     /* next, compute all the lengths */
3828:     PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3829:     olens = dlens + m;
3830:     for (i = 0; i < m; i++) {
3831:       jend = ii[i + 1] - ii[i];
3832:       olen = 0;
3833:       dlen = 0;
3834:       for (j = 0; j < jend; j++) {
3835:         if (*jj < rstart || *jj >= rend) olen++;
3836:         else dlen++;
3837:         jj++;
3838:       }
3839:       olens[i] = olen;
3840:       dlens[i] = dlen;
3841:     }
3842:     PetscCall(MatCreate(comm, &M));
3843:     PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3844:     PetscCall(MatSetBlockSizes(M, bs, cbs));
3845:     PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3846:     PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3847:     PetscCall(PetscFree(dlens));
3848:   } else {
3849:     PetscInt ml, nl;

3851:     M = *newmat;
3852:     PetscCall(MatGetLocalSize(M, &ml, &nl));
3853:     PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3854:     PetscCall(MatZeroEntries(M));
3855:     /*
3856:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3857:        rather than the slower MatSetValues().
3858:     */
3859:     M->was_assembled = PETSC_TRUE;
3860:     M->assembled     = PETSC_FALSE;
3861:   }
3862:   PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3863:   aij = (Mat_SeqAIJ *)(Mreuse)->data;
3864:   ii  = aij->i;
3865:   jj  = aij->j;

3867:   /* trigger copy to CPU if needed */
3868:   PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3869:   for (i = 0; i < m; i++) {
3870:     row   = rstart + i;
3871:     nz    = ii[i + 1] - ii[i];
3872:     cwork = jj;
3873:     jj += nz;
3874:     vwork = aa;
3875:     aa += nz;
3876:     PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3877:   }
3878:   PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));

3880:   PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3881:   PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3882:   *newmat = M;

3884:   /* save submatrix used in processor for next request */
3885:   if (call == MAT_INITIAL_MATRIX) {
3886:     PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3887:     PetscCall(MatDestroy(&Mreuse));
3888:   }
3889:   PetscFunctionReturn(PETSC_SUCCESS);
3890: }

3892: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3893: {
3894:   PetscInt        m, cstart, cend, j, nnz, i, d, *ld;
3895:   PetscInt       *d_nnz, *o_nnz, nnz_max = 0, rstart, ii;
3896:   const PetscInt *JJ;
3897:   PetscBool       nooffprocentries;
3898:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)B->data;

3900:   PetscFunctionBegin;
3901:   PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);

3903:   PetscCall(PetscLayoutSetUp(B->rmap));
3904:   PetscCall(PetscLayoutSetUp(B->cmap));
3905:   m      = B->rmap->n;
3906:   cstart = B->cmap->rstart;
3907:   cend   = B->cmap->rend;
3908:   rstart = B->rmap->rstart;

3910:   PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));

3912:   if (PetscDefined(USE_DEBUG)) {
3913:     for (i = 0; i < m; i++) {
3914:       nnz = Ii[i + 1] - Ii[i];
3915:       JJ  = J + Ii[i];
3916:       PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3917:       PetscCheck(!nnz || !(JJ[0] < 0), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Row %" PetscInt_FMT " starts with negative column index %" PetscInt_FMT, i, JJ[0]);
3918:       PetscCheck(!nnz || !(JJ[nnz - 1] >= B->cmap->N), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Row %" PetscInt_FMT " ends with too large a column index %" PetscInt_FMT " (max allowed %" PetscInt_FMT ")", i, JJ[nnz - 1], B->cmap->N);
3919:     }
3920:   }

3922:   for (i = 0; i < m; i++) {
3923:     nnz     = Ii[i + 1] - Ii[i];
3924:     JJ      = J + Ii[i];
3925:     nnz_max = PetscMax(nnz_max, nnz);
3926:     d       = 0;
3927:     for (j = 0; j < nnz; j++) {
3928:       if (cstart <= JJ[j] && JJ[j] < cend) d++;
3929:     }
3930:     d_nnz[i] = d;
3931:     o_nnz[i] = nnz - d;
3932:   }
3933:   PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3934:   PetscCall(PetscFree2(d_nnz, o_nnz));

3936:   for (i = 0; i < m; i++) {
3937:     ii = i + rstart;
3938:     PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], J + Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES));
3939:   }
3940:   nooffprocentries    = B->nooffprocentries;
3941:   B->nooffprocentries = PETSC_TRUE;
3942:   PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3943:   PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3944:   B->nooffprocentries = nooffprocentries;

3946:   /* count number of entries below block diagonal */
3947:   PetscCall(PetscFree(Aij->ld));
3948:   PetscCall(PetscCalloc1(m, &ld));
3949:   Aij->ld = ld;
3950:   for (i = 0; i < m; i++) {
3951:     nnz = Ii[i + 1] - Ii[i];
3952:     j   = 0;
3953:     while (j < nnz && J[j] < cstart) j++;
3954:     ld[i] = j;
3955:     J += nnz;
3956:   }

3958:   PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3959:   PetscFunctionReturn(PETSC_SUCCESS);
3960: }

3962: /*@
3963:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3964:    (the default parallel PETSc format).

3966:    Collective

3968:    Input Parameters:
3969: +  B - the matrix
3970: .  i - the indices into j for the start of each local row (starts with zero)
3971: .  j - the column indices for each local row (starts with zero)
3972: -  v - optional values in the matrix

3974:    Level: developer

3976:    Notes:
3977:        The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3978:      thus you CANNOT change the matrix entries by changing the values of `v` after you have
3979:      called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.

3981:        The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.

3983:        The format which is used for the sparse matrix input, is equivalent to a
3984:     row-major ordering.. i.e for the following matrix, the input data expected is
3985:     as shown

3987: .vb
3988:         1 0 0
3989:         2 0 3     P0
3990:        -------
3991:         4 5 6     P1

3993:      Process0 [P0] rows_owned=[0,1]
3994:         i =  {0,1,3}  [size = nrow+1  = 2+1]
3995:         j =  {0,0,2}  [size = 3]
3996:         v =  {1,2,3}  [size = 3]

3998:      Process1 [P1] rows_owned=[2]
3999:         i =  {0,3}    [size = nrow+1  = 1+1]
4000:         j =  {0,1,2}  [size = 3]
4001:         v =  {4,5,6}  [size = 3]
4002: .ve

4004: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`, `MATMPIAIJ`,
4005:           `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`
4006: @*/
4007: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4008: {
4009:   PetscFunctionBegin;
4010:   PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4011:   PetscFunctionReturn(PETSC_SUCCESS);
4012: }

4014: /*@C
4015:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4016:    (the default parallel PETSc format).  For good matrix assembly performance
4017:    the user should preallocate the matrix storage by setting the parameters
4018:    `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).

4020:    Collective

4022:    Input Parameters:
4023: +  B - the matrix
4024: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
4025:            (same value is used for all local rows)
4026: .  d_nnz - array containing the number of nonzeros in the various rows of the
4027:            DIAGONAL portion of the local submatrix (possibly different for each row)
4028:            or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4029:            The size of this array is equal to the number of local rows, i.e 'm'.
4030:            For matrices that will be factored, you must leave room for (and set)
4031:            the diagonal entry even if it is zero.
4032: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
4033:            submatrix (same value is used for all local rows).
4034: -  o_nnz - array containing the number of nonzeros in the various rows of the
4035:            OFF-DIAGONAL portion of the local submatrix (possibly different for
4036:            each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4037:            structure. The size of this array is equal to the number
4038:            of local rows, i.e 'm'.

4040:    Usage:
4041:    Consider the following 8x8 matrix with 34 non-zero values, that is
4042:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4043:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4044:    as follows

4046: .vb
4047:             1  2  0  |  0  3  0  |  0  4
4048:     Proc0   0  5  6  |  7  0  0  |  8  0
4049:             9  0 10  | 11  0  0  | 12  0
4050:     -------------------------------------
4051:            13  0 14  | 15 16 17  |  0  0
4052:     Proc1   0 18  0  | 19 20 21  |  0  0
4053:             0  0  0  | 22 23  0  | 24  0
4054:     -------------------------------------
4055:     Proc2  25 26 27  |  0  0 28  | 29  0
4056:            30  0  0  | 31 32 33  |  0 34
4057: .ve

4059:    This can be represented as a collection of submatrices as
4060: .vb
4061:       A B C
4062:       D E F
4063:       G H I
4064: .ve

4066:    Where the submatrices A,B,C are owned by proc0, D,E,F are
4067:    owned by proc1, G,H,I are owned by proc2.

4069:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4070:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4071:    The 'M','N' parameters are 8,8, and have the same values on all procs.

4073:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4074:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4075:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4076:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4077:    part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4078:    matrix, ans [DF] as another `MATSEQAIJ` matrix.

4080:    When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4081:    allocated for every row of the local diagonal submatrix, and `o_nz`
4082:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
4083:    One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4084:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4085:    In this case, the values of `d_nz`, `o_nz` are
4086: .vb
4087:      proc0  dnz = 2, o_nz = 2
4088:      proc1  dnz = 3, o_nz = 2
4089:      proc2  dnz = 1, o_nz = 4
4090: .ve
4091:    We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4092:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4093:    for proc3. i.e we are using 12+15+10=37 storage locations to store
4094:    34 values.

4096:    When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4097:    for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4098:    In the above case the values for `d_nnz`, `o_nnz` are
4099: .vb
4100:      proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4101:      proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4102:      proc2 d_nnz = [1,1]   and o_nnz = [4,4]
4103: .ve
4104:    Here the space allocated is sum of all the above values i.e 34, and
4105:    hence pre-allocation is perfect.

4107:    Level: intermediate

4109:    Notes:
4110:    If the *_nnz parameter is given then the *_nz parameter is ignored

4112:    The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4113:    storage.  The stored row and column indices begin with zero.
4114:    See [Sparse Matrices](sec_matsparse) for details.

4116:    The parallel matrix is partitioned such that the first m0 rows belong to
4117:    process 0, the next m1 rows belong to process 1, the next m2 rows belong
4118:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

4120:    The DIAGONAL portion of the local submatrix of a processor can be defined
4121:    as the submatrix which is obtained by extraction the part corresponding to
4122:    the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4123:    first row that belongs to the processor, r2 is the last row belonging to
4124:    the this processor, and c1-c2 is range of indices of the local part of a
4125:    vector suitable for applying the matrix to.  This is an mxn matrix.  In the
4126:    common case of a square matrix, the row and column ranges are the same and
4127:    the DIAGONAL part is also square. The remaining portion of the local
4128:    submatrix (mxN) constitute the OFF-DIAGONAL portion.

4130:    If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.

4132:    You can call `MatGetInfo()` to get information on how effective the preallocation was;
4133:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4134:    You can also run with the option `-info` and look for messages with the string
4135:    malloc in them to see if additional memory allocation was needed.

4137: .seealso: [](chapter_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4138:           `MATMPIAIJ`, `MatGetInfo()`, `PetscSplitOwnership()`
4139: @*/
4140: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4141: {
4142:   PetscFunctionBegin;
4145:   PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4146:   PetscFunctionReturn(PETSC_SUCCESS);
4147: }

4149: /*@
4150:      MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4151:          CSR format for the local rows.

4153:    Collective

4155:    Input Parameters:
4156: +  comm - MPI communicator
4157: .  m - number of local rows (Cannot be `PETSC_DECIDE`)
4158: .  n - This value should be the same as the local size used in creating the
4159:        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4160:        calculated if N is given) For square matrices n is almost always m.
4161: .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4162: .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4163: .   i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4164: .   j - column indices
4165: -   a - optional matrix values

4167:    Output Parameter:
4168: .   mat - the matrix

4170:    Level: intermediate

4172:    Notes:
4173:        The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4174:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
4175:      called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.

4177:        The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.

4179:        The format which is used for the sparse matrix input, is equivalent to a
4180:     row-major ordering.. i.e for the following matrix, the input data expected is
4181:     as shown

4183:        Once you have created the matrix you can update it with new numerical values using MatUpdateMPIAIJWithArrays
4184: .vb
4185:         1 0 0
4186:         2 0 3     P0
4187:        -------
4188:         4 5 6     P1

4190:      Process0 [P0] rows_owned=[0,1]
4191:         i =  {0,1,3}  [size = nrow+1  = 2+1]
4192:         j =  {0,0,2}  [size = 3]
4193:         v =  {1,2,3}  [size = 3]

4195:      Process1 [P1] rows_owned=[2]
4196:         i =  {0,3}    [size = nrow+1  = 1+1]
4197:         j =  {0,1,2}  [size = 3]
4198:         v =  {4,5,6}  [size = 3]
4199: .ve

4201: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIK`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4202:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`
4203: @*/
4204: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4205: {
4206:   PetscFunctionBegin;
4207:   PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4208:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4209:   PetscCall(MatCreate(comm, mat));
4210:   PetscCall(MatSetSizes(*mat, m, n, M, N));
4211:   /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4212:   PetscCall(MatSetType(*mat, MATMPIAIJ));
4213:   PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4214:   PetscFunctionReturn(PETSC_SUCCESS);
4215: }

4217: /*@
4218:      MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4219:      CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4220:      from `MatCreateMPIAIJWithArrays()`

4222:      Deprecated: Use `MatUpdateMPIAIJWithArray()`

4224:    Collective

4226:    Input Parameters:
4227: +  mat - the matrix
4228: .  m - number of local rows (Cannot be `PETSC_DECIDE`)
4229: .  n - This value should be the same as the local size used in creating the
4230:        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4231:        calculated if N is given) For square matrices n is almost always m.
4232: .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4233: .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4234: .  Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4235: .  J - column indices
4236: -  v - matrix values

4238:    Level: deprecated

4240: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4241:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatUpdateMPIAIJWithArray()`
4242: @*/
4243: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4244: {
4245:   PetscInt        nnz, i;
4246:   PetscBool       nooffprocentries;
4247:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)mat->data;
4248:   Mat_SeqAIJ     *Ad  = (Mat_SeqAIJ *)Aij->A->data;
4249:   PetscScalar    *ad, *ao;
4250:   PetscInt        ldi, Iii, md;
4251:   const PetscInt *Adi = Ad->i;
4252:   PetscInt       *ld  = Aij->ld;

4254:   PetscFunctionBegin;
4255:   PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4256:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4257:   PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4258:   PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");

4260:   PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4261:   PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));

4263:   for (i = 0; i < m; i++) {
4264:     nnz = Ii[i + 1] - Ii[i];
4265:     Iii = Ii[i];
4266:     ldi = ld[i];
4267:     md  = Adi[i + 1] - Adi[i];
4268:     PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4269:     PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4270:     PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4271:     ad += md;
4272:     ao += nnz - md;
4273:   }
4274:   nooffprocentries      = mat->nooffprocentries;
4275:   mat->nooffprocentries = PETSC_TRUE;
4276:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4277:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4278:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4279:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4280:   PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4281:   PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4282:   PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4283:   mat->nooffprocentries = nooffprocentries;
4284:   PetscFunctionReturn(PETSC_SUCCESS);
4285: }

4287: /*@
4288:      MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values

4290:    Collective

4292:    Input Parameters:
4293: +  mat - the matrix
4294: -  v - matrix values, stored by row

4296:    Level: intermediate

4298:    Note:
4299:    The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`

4301: .seealso: [](chapter_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4302:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatUpdateMPIAIJWithArrays()`
4303: @*/
4304: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4305: {
4306:   PetscInt        nnz, i, m;
4307:   PetscBool       nooffprocentries;
4308:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)mat->data;
4309:   Mat_SeqAIJ     *Ad  = (Mat_SeqAIJ *)Aij->A->data;
4310:   Mat_SeqAIJ     *Ao  = (Mat_SeqAIJ *)Aij->B->data;
4311:   PetscScalar    *ad, *ao;
4312:   const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4313:   PetscInt        ldi, Iii, md;
4314:   PetscInt       *ld = Aij->ld;

4316:   PetscFunctionBegin;
4317:   m = mat->rmap->n;

4319:   PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4320:   PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4321:   Iii = 0;
4322:   for (i = 0; i < m; i++) {
4323:     nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4324:     ldi = ld[i];
4325:     md  = Adi[i + 1] - Adi[i];
4326:     PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4327:     PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4328:     PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4329:     ad += md;
4330:     ao += nnz - md;
4331:     Iii += nnz;
4332:   }
4333:   nooffprocentries      = mat->nooffprocentries;
4334:   mat->nooffprocentries = PETSC_TRUE;
4335:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4336:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4337:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4338:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4339:   PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4340:   PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4341:   PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4342:   mat->nooffprocentries = nooffprocentries;
4343:   PetscFunctionReturn(PETSC_SUCCESS);
4344: }

4346: /*@C
4347:    MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4348:    (the default parallel PETSc format).  For good matrix assembly performance
4349:    the user should preallocate the matrix storage by setting the parameters
4350:    `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).

4352:    Collective

4354:    Input Parameters:
4355: +  comm - MPI communicator
4356: .  m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4357:            This value should be the same as the local size used in creating the
4358:            y vector for the matrix-vector product y = Ax.
4359: .  n - This value should be the same as the local size used in creating the
4360:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4361:        calculated if N is given) For square matrices n is almost always m.
4362: .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4363: .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4364: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
4365:            (same value is used for all local rows)
4366: .  d_nnz - array containing the number of nonzeros in the various rows of the
4367:            DIAGONAL portion of the local submatrix (possibly different for each row)
4368:            or `NULL`, if `d_nz` is used to specify the nonzero structure.
4369:            The size of this array is equal to the number of local rows, i.e 'm'.
4370: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
4371:            submatrix (same value is used for all local rows).
4372: -  o_nnz - array containing the number of nonzeros in the various rows of the
4373:            OFF-DIAGONAL portion of the local submatrix (possibly different for
4374:            each row) or `NULL`, if `o_nz` is used to specify the nonzero
4375:            structure. The size of this array is equal to the number
4376:            of local rows, i.e 'm'.

4378:    Output Parameter:
4379: .  A - the matrix

4381:    Options Database Keys:
4382: +  -mat_no_inode  - Do not use inodes
4383: .  -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4384: -  -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4385:         See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the vecscatter be viewed as a matrix.
4386:         Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.

4388:    Level: intermediate

4390:    Notes:
4391:    It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4392:    MatXXXXSetPreallocation() paradigm instead of this routine directly.
4393:    [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]

4395:    If the *_nnz parameter is given then the *_nz parameter is ignored

4397:    The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4398:    processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4399:    storage requirements for this matrix.

4401:    If `PETSC_DECIDE` or  `PETSC_DETERMINE` is used for a particular argument on one
4402:    processor than it must be used on all processors that share the object for
4403:    that argument.

4405:    The user MUST specify either the local or global matrix dimensions
4406:    (possibly both).

4408:    The parallel matrix is partitioned across processors such that the
4409:    first m0 rows belong to process 0, the next m1 rows belong to
4410:    process 1, the next m2 rows belong to process 2 etc.. where
4411:    m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4412:    values corresponding to [m x N] submatrix.

4414:    The columns are logically partitioned with the n0 columns belonging
4415:    to 0th partition, the next n1 columns belonging to the next
4416:    partition etc.. where n0,n1,n2... are the input parameter 'n'.

4418:    The DIAGONAL portion of the local submatrix on any given processor
4419:    is the submatrix corresponding to the rows and columns m,n
4420:    corresponding to the given processor. i.e diagonal matrix on
4421:    process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4422:    etc. The remaining portion of the local submatrix [m x (N-n)]
4423:    constitute the OFF-DIAGONAL portion. The example below better
4424:    illustrates this concept.

4426:    For a square global matrix we define each processor's diagonal portion
4427:    to be its local rows and the corresponding columns (a square submatrix);
4428:    each processor's off-diagonal portion encompasses the remainder of the
4429:    local matrix (a rectangular submatrix).

4431:    If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.

4433:    When calling this routine with a single process communicator, a matrix of
4434:    type `MATSEQAIJ` is returned.  If a matrix of type `MATMPIAIJ` is desired for this
4435:    type of communicator, use the construction mechanism
4436: .vb
4437:      MatCreate(...,&A);
4438:      MatSetType(A,MATMPIAIJ);
4439:      MatSetSizes(A, m,n,M,N);
4440:      MatMPIAIJSetPreallocation(A,...);
4441: .ve

4443:    By default, this format uses inodes (identical nodes) when possible.
4444:    We search for consecutive rows with the same nonzero structure, thereby
4445:    reusing matrix information to achieve increased efficiency.

4447:    Usage:
4448:    Consider the following 8x8 matrix with 34 non-zero values, that is
4449:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4450:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4451:    as follows

4453: .vb
4454:             1  2  0  |  0  3  0  |  0  4
4455:     Proc0   0  5  6  |  7  0  0  |  8  0
4456:             9  0 10  | 11  0  0  | 12  0
4457:     -------------------------------------
4458:            13  0 14  | 15 16 17  |  0  0
4459:     Proc1   0 18  0  | 19 20 21  |  0  0
4460:             0  0  0  | 22 23  0  | 24  0
4461:     -------------------------------------
4462:     Proc2  25 26 27  |  0  0 28  | 29  0
4463:            30  0  0  | 31 32 33  |  0 34
4464: .ve

4466:    This can be represented as a collection of submatrices as

4468: .vb
4469:       A B C
4470:       D E F
4471:       G H I
4472: .ve

4474:    Where the submatrices A,B,C are owned by proc0, D,E,F are
4475:    owned by proc1, G,H,I are owned by proc2.

4477:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4478:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4479:    The 'M','N' parameters are 8,8, and have the same values on all procs.

4481:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4482:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4483:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4484:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4485:    part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4486:    matrix, ans [DF] as another SeqAIJ matrix.

4488:    When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4489:    allocated for every row of the local diagonal submatrix, and `o_nz`
4490:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
4491:    One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4492:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4493:    In this case, the values of `d_nz`,`o_nz` are
4494: .vb
4495:      proc0  dnz = 2, o_nz = 2
4496:      proc1  dnz = 3, o_nz = 2
4497:      proc2  dnz = 1, o_nz = 4
4498: .ve
4499:    We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4500:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4501:    for proc3. i.e we are using 12+15+10=37 storage locations to store
4502:    34 values.

4504:    When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4505:    for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4506:    In the above case the values for d_nnz,o_nnz are
4507: .vb
4508:      proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4509:      proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4510:      proc2 d_nnz = [1,1]   and o_nnz = [4,4]
4511: .ve
4512:    Here the space allocated is sum of all the above values i.e 34, and
4513:    hence pre-allocation is perfect.

4515: .seealso: [](chapter_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4516:           `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`
4517: @*/
4518: 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)
4519: {
4520:   PetscMPIInt size;

4522:   PetscFunctionBegin;
4523:   PetscCall(MatCreate(comm, A));
4524:   PetscCall(MatSetSizes(*A, m, n, M, N));
4525:   PetscCallMPI(MPI_Comm_size(comm, &size));
4526:   if (size > 1) {
4527:     PetscCall(MatSetType(*A, MATMPIAIJ));
4528:     PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4529:   } else {
4530:     PetscCall(MatSetType(*A, MATSEQAIJ));
4531:     PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4532:   }
4533:   PetscFunctionReturn(PETSC_SUCCESS);
4534: }

4536: /*MC
4537:     MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix

4539:     Synopsis:
4540:     MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)

4542:     Not Collective

4544:     Input Parameter:
4545: .   A - the `MATMPIAIJ` matrix

4547:     Output Parameters:
4548: +   Ad - the diagonal portion of the matrix
4549: .   Ao - the off diagonal portion of the matrix
4550: .   colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4551: -   ierr - error code

4553:      Level: advanced

4555:     Note:
4556:     Use  `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`

4558: .seealso: [](chapter_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4559: M*/

4561: /*MC
4562:     MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`

4564:     Synopsis:
4565:     MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)

4567:     Not Collective

4569:     Input Parameters:
4570: +   A - the `MATMPIAIJ` matrix
4571: .   Ad - the diagonal portion of the matrix
4572: .   Ao - the off diagonal portion of the matrix
4573: .   colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4574: -   ierr - error code

4576:      Level: advanced

4578: .seealso: [](chapter_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4579: M*/

4581: /*@C
4582:   MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix

4584:   Not Collective

4586:   Input Parameter:
4587: . A - The `MATMPIAIJ` matrix

4589:   Output Parameters:
4590: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4591: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4592: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix

4594:   Level: intermediate

4596:   Note:
4597:   The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4598:   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
4599:   the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4600:   local column numbers to global column numbers in the original matrix.

4602:   Fortran Note:
4603:   `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`

4605: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATMPIAIJ`, `MATSEQAIJ`
4606: @*/
4607: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4608: {
4609:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4610:   PetscBool   flg;

4612:   PetscFunctionBegin;
4613:   PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4614:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4615:   if (Ad) *Ad = a->A;
4616:   if (Ao) *Ao = a->B;
4617:   if (colmap) *colmap = a->garray;
4618:   PetscFunctionReturn(PETSC_SUCCESS);
4619: }

4621: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4622: {
4623:   PetscInt     m, N, i, rstart, nnz, Ii;
4624:   PetscInt    *indx;
4625:   PetscScalar *values;
4626:   MatType      rootType;

4628:   PetscFunctionBegin;
4629:   PetscCall(MatGetSize(inmat, &m, &N));
4630:   if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4631:     PetscInt *dnz, *onz, sum, bs, cbs;

4633:     if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4634:     /* Check sum(n) = N */
4635:     PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4636:     PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);

4638:     PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4639:     rstart -= m;

4641:     MatPreallocateBegin(comm, m, n, dnz, onz);
4642:     for (i = 0; i < m; i++) {
4643:       PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4644:       PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4645:       PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4646:     }

4648:     PetscCall(MatCreate(comm, outmat));
4649:     PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4650:     PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4651:     PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4652:     PetscCall(MatGetRootType_Private(inmat, &rootType));
4653:     PetscCall(MatSetType(*outmat, rootType));
4654:     PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4655:     PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4656:     MatPreallocateEnd(dnz, onz);
4657:     PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4658:   }

4660:   /* numeric phase */
4661:   PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4662:   for (i = 0; i < m; i++) {
4663:     PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4664:     Ii = i + rstart;
4665:     PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4666:     PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4667:   }
4668:   PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4669:   PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4670:   PetscFunctionReturn(PETSC_SUCCESS);
4671: }

4673: PetscErrorCode MatFileSplit(Mat A, char *outfile)
4674: {
4675:   PetscMPIInt        rank;
4676:   PetscInt           m, N, i, rstart, nnz;
4677:   size_t             len;
4678:   const PetscInt    *indx;
4679:   PetscViewer        out;
4680:   char              *name;
4681:   Mat                B;
4682:   const PetscScalar *values;

4684:   PetscFunctionBegin;
4685:   PetscCall(MatGetLocalSize(A, &m, NULL));
4686:   PetscCall(MatGetSize(A, NULL, &N));
4687:   /* Should this be the type of the diagonal block of A? */
4688:   PetscCall(MatCreate(PETSC_COMM_SELF, &B));
4689:   PetscCall(MatSetSizes(B, m, N, m, N));
4690:   PetscCall(MatSetBlockSizesFromMats(B, A, A));
4691:   PetscCall(MatSetType(B, MATSEQAIJ));
4692:   PetscCall(MatSeqAIJSetPreallocation(B, 0, NULL));
4693:   PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
4694:   for (i = 0; i < m; i++) {
4695:     PetscCall(MatGetRow(A, i + rstart, &nnz, &indx, &values));
4696:     PetscCall(MatSetValues(B, 1, &i, nnz, indx, values, INSERT_VALUES));
4697:     PetscCall(MatRestoreRow(A, i + rstart, &nnz, &indx, &values));
4698:   }
4699:   PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
4700:   PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));

4702:   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)A), &rank));
4703:   PetscCall(PetscStrlen(outfile, &len));
4704:   PetscCall(PetscMalloc1(len + 6, &name));
4705:   PetscCall(PetscSNPrintf(name, len + 6, "%s.%d", outfile, rank));
4706:   PetscCall(PetscViewerBinaryOpen(PETSC_COMM_SELF, name, FILE_MODE_APPEND, &out));
4707:   PetscCall(PetscFree(name));
4708:   PetscCall(MatView(B, out));
4709:   PetscCall(PetscViewerDestroy(&out));
4710:   PetscCall(MatDestroy(&B));
4711:   PetscFunctionReturn(PETSC_SUCCESS);
4712: }

4714: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4715: {
4716:   Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;

4718:   PetscFunctionBegin;
4719:   if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4720:   PetscCall(PetscFree(merge->id_r));
4721:   PetscCall(PetscFree(merge->len_s));
4722:   PetscCall(PetscFree(merge->len_r));
4723:   PetscCall(PetscFree(merge->bi));
4724:   PetscCall(PetscFree(merge->bj));
4725:   PetscCall(PetscFree(merge->buf_ri[0]));
4726:   PetscCall(PetscFree(merge->buf_ri));
4727:   PetscCall(PetscFree(merge->buf_rj[0]));
4728:   PetscCall(PetscFree(merge->buf_rj));
4729:   PetscCall(PetscFree(merge->coi));
4730:   PetscCall(PetscFree(merge->coj));
4731:   PetscCall(PetscFree(merge->owners_co));
4732:   PetscCall(PetscLayoutDestroy(&merge->rowmap));
4733:   PetscCall(PetscFree(merge));
4734:   PetscFunctionReturn(PETSC_SUCCESS);
4735: }

4737: #include <../src/mat/utils/freespace.h>
4738: #include <petscbt.h>

4740: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4741: {
4742:   MPI_Comm             comm;
4743:   Mat_SeqAIJ          *a = (Mat_SeqAIJ *)seqmat->data;
4744:   PetscMPIInt          size, rank, taga, *len_s;
4745:   PetscInt             N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4746:   PetscInt             proc, m;
4747:   PetscInt           **buf_ri, **buf_rj;
4748:   PetscInt             k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4749:   PetscInt             nrows, **buf_ri_k, **nextrow, **nextai;
4750:   MPI_Request         *s_waits, *r_waits;
4751:   MPI_Status          *status;
4752:   const MatScalar     *aa, *a_a;
4753:   MatScalar          **abuf_r, *ba_i;
4754:   Mat_Merge_SeqsToMPI *merge;
4755:   PetscContainer       container;

4757:   PetscFunctionBegin;
4758:   PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4759:   PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));

4761:   PetscCallMPI(MPI_Comm_size(comm, &size));
4762:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

4764:   PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4765:   PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4766:   PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4767:   PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4768:   aa = a_a;

4770:   bi     = merge->bi;
4771:   bj     = merge->bj;
4772:   buf_ri = merge->buf_ri;
4773:   buf_rj = merge->buf_rj;

4775:   PetscCall(PetscMalloc1(size, &status));
4776:   owners = merge->rowmap->range;
4777:   len_s  = merge->len_s;

4779:   /* send and recv matrix values */
4780:   PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4781:   PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));

4783:   PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4784:   for (proc = 0, k = 0; proc < size; proc++) {
4785:     if (!len_s[proc]) continue;
4786:     i = owners[proc];
4787:     PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4788:     k++;
4789:   }

4791:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4792:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4793:   PetscCall(PetscFree(status));

4795:   PetscCall(PetscFree(s_waits));
4796:   PetscCall(PetscFree(r_waits));

4798:   /* insert mat values of mpimat */
4799:   PetscCall(PetscMalloc1(N, &ba_i));
4800:   PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));

4802:   for (k = 0; k < merge->nrecv; k++) {
4803:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4804:     nrows       = *(buf_ri_k[k]);
4805:     nextrow[k]  = buf_ri_k[k] + 1;           /* next row number of k-th recved i-structure */
4806:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure  */
4807:   }

4809:   /* set values of ba */
4810:   m = merge->rowmap->n;
4811:   for (i = 0; i < m; i++) {
4812:     arow = owners[rank] + i;
4813:     bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4814:     bnzi = bi[i + 1] - bi[i];
4815:     PetscCall(PetscArrayzero(ba_i, bnzi));

4817:     /* add local non-zero vals of this proc's seqmat into ba */
4818:     anzi   = ai[arow + 1] - ai[arow];
4819:     aj     = a->j + ai[arow];
4820:     aa     = a_a + ai[arow];
4821:     nextaj = 0;
4822:     for (j = 0; nextaj < anzi; j++) {
4823:       if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4824:         ba_i[j] += aa[nextaj++];
4825:       }
4826:     }

4828:     /* add received vals into ba */
4829:     for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4830:       /* i-th row */
4831:       if (i == *nextrow[k]) {
4832:         anzi   = *(nextai[k] + 1) - *nextai[k];
4833:         aj     = buf_rj[k] + *(nextai[k]);
4834:         aa     = abuf_r[k] + *(nextai[k]);
4835:         nextaj = 0;
4836:         for (j = 0; nextaj < anzi; j++) {
4837:           if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4838:             ba_i[j] += aa[nextaj++];
4839:           }
4840:         }
4841:         nextrow[k]++;
4842:         nextai[k]++;
4843:       }
4844:     }
4845:     PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4846:   }
4847:   PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4848:   PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4849:   PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));

4851:   PetscCall(PetscFree(abuf_r[0]));
4852:   PetscCall(PetscFree(abuf_r));
4853:   PetscCall(PetscFree(ba_i));
4854:   PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4855:   PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4856:   PetscFunctionReturn(PETSC_SUCCESS);
4857: }

4859: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4860: {
4861:   Mat                  B_mpi;
4862:   Mat_SeqAIJ          *a = (Mat_SeqAIJ *)seqmat->data;
4863:   PetscMPIInt          size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4864:   PetscInt           **buf_rj, **buf_ri, **buf_ri_k;
4865:   PetscInt             M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4866:   PetscInt             len, proc, *dnz, *onz, bs, cbs;
4867:   PetscInt             k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4868:   PetscInt             nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4869:   MPI_Request         *si_waits, *sj_waits, *ri_waits, *rj_waits;
4870:   MPI_Status          *status;
4871:   PetscFreeSpaceList   free_space = NULL, current_space = NULL;
4872:   PetscBT              lnkbt;
4873:   Mat_Merge_SeqsToMPI *merge;
4874:   PetscContainer       container;

4876:   PetscFunctionBegin;
4877:   PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));

4879:   /* make sure it is a PETSc comm */
4880:   PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4881:   PetscCallMPI(MPI_Comm_size(comm, &size));
4882:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

4884:   PetscCall(PetscNew(&merge));
4885:   PetscCall(PetscMalloc1(size, &status));

4887:   /* determine row ownership */
4888:   PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4889:   PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4890:   PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4891:   PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4892:   PetscCall(PetscLayoutSetUp(merge->rowmap));
4893:   PetscCall(PetscMalloc1(size, &len_si));
4894:   PetscCall(PetscMalloc1(size, &merge->len_s));

4896:   m      = merge->rowmap->n;
4897:   owners = merge->rowmap->range;

4899:   /* determine the number of messages to send, their lengths */
4900:   len_s = merge->len_s;

4902:   len          = 0; /* length of buf_si[] */
4903:   merge->nsend = 0;
4904:   for (proc = 0; proc < size; proc++) {
4905:     len_si[proc] = 0;
4906:     if (proc == rank) {
4907:       len_s[proc] = 0;
4908:     } else {
4909:       len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4910:       len_s[proc]  = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4911:     }
4912:     if (len_s[proc]) {
4913:       merge->nsend++;
4914:       nrows = 0;
4915:       for (i = owners[proc]; i < owners[proc + 1]; i++) {
4916:         if (ai[i + 1] > ai[i]) nrows++;
4917:       }
4918:       len_si[proc] = 2 * (nrows + 1);
4919:       len += len_si[proc];
4920:     }
4921:   }

4923:   /* determine the number and length of messages to receive for ij-structure */
4924:   PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4925:   PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));

4927:   /* post the Irecv of j-structure */
4928:   PetscCall(PetscCommGetNewTag(comm, &tagj));
4929:   PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));

4931:   /* post the Isend of j-structure */
4932:   PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));

4934:   for (proc = 0, k = 0; proc < size; proc++) {
4935:     if (!len_s[proc]) continue;
4936:     i = owners[proc];
4937:     PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4938:     k++;
4939:   }

4941:   /* receives and sends of j-structure are complete */
4942:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4943:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));

4945:   /* send and recv i-structure */
4946:   PetscCall(PetscCommGetNewTag(comm, &tagi));
4947:   PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));

4949:   PetscCall(PetscMalloc1(len + 1, &buf_s));
4950:   buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4951:   for (proc = 0, k = 0; proc < size; proc++) {
4952:     if (!len_s[proc]) continue;
4953:     /* form outgoing message for i-structure:
4954:          buf_si[0]:                 nrows to be sent
4955:                [1:nrows]:           row index (global)
4956:                [nrows+1:2*nrows+1]: i-structure index
4957:     */
4958:     nrows       = len_si[proc] / 2 - 1;
4959:     buf_si_i    = buf_si + nrows + 1;
4960:     buf_si[0]   = nrows;
4961:     buf_si_i[0] = 0;
4962:     nrows       = 0;
4963:     for (i = owners[proc]; i < owners[proc + 1]; i++) {
4964:       anzi = ai[i + 1] - ai[i];
4965:       if (anzi) {
4966:         buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4967:         buf_si[nrows + 1]   = i - owners[proc];       /* local row index */
4968:         nrows++;
4969:       }
4970:     }
4971:     PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4972:     k++;
4973:     buf_si += len_si[proc];
4974:   }

4976:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4977:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));

4979:   PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4980:   for (i = 0; i < merge->nrecv; i++) PetscCall(PetscInfo(seqmat, "recv len_ri=%d, len_rj=%d from [%d]\n", len_ri[i], merge->len_r[i], merge->id_r[i]));

4982:   PetscCall(PetscFree(len_si));
4983:   PetscCall(PetscFree(len_ri));
4984:   PetscCall(PetscFree(rj_waits));
4985:   PetscCall(PetscFree2(si_waits, sj_waits));
4986:   PetscCall(PetscFree(ri_waits));
4987:   PetscCall(PetscFree(buf_s));
4988:   PetscCall(PetscFree(status));

4990:   /* compute a local seq matrix in each processor */
4991:   /* allocate bi array and free space for accumulating nonzero column info */
4992:   PetscCall(PetscMalloc1(m + 1, &bi));
4993:   bi[0] = 0;

4995:   /* create and initialize a linked list */
4996:   nlnk = N + 1;
4997:   PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));

4999:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
5000:   len = ai[owners[rank + 1]] - ai[owners[rank]];
5001:   PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));

5003:   current_space = free_space;

5005:   /* determine symbolic info for each local row */
5006:   PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));

5008:   for (k = 0; k < merge->nrecv; k++) {
5009:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
5010:     nrows       = *buf_ri_k[k];
5011:     nextrow[k]  = buf_ri_k[k] + 1;           /* next row number of k-th recved i-structure */
5012:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure  */
5013:   }

5015:   MatPreallocateBegin(comm, m, n, dnz, onz);
5016:   len = 0;
5017:   for (i = 0; i < m; i++) {
5018:     bnzi = 0;
5019:     /* add local non-zero cols of this proc's seqmat into lnk */
5020:     arow = owners[rank] + i;
5021:     anzi = ai[arow + 1] - ai[arow];
5022:     aj   = a->j + ai[arow];
5023:     PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5024:     bnzi += nlnk;
5025:     /* add received col data into lnk */
5026:     for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5027:       if (i == *nextrow[k]) {            /* i-th row */
5028:         anzi = *(nextai[k] + 1) - *nextai[k];
5029:         aj   = buf_rj[k] + *nextai[k];
5030:         PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5031:         bnzi += nlnk;
5032:         nextrow[k]++;
5033:         nextai[k]++;
5034:       }
5035:     }
5036:     if (len < bnzi) len = bnzi; /* =max(bnzi) */

5038:     /* if free space is not available, make more free space */
5039:     if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), &current_space));
5040:     /* copy data into free space, then initialize lnk */
5041:     PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5042:     PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));

5044:     current_space->array += bnzi;
5045:     current_space->local_used += bnzi;
5046:     current_space->local_remaining -= bnzi;

5048:     bi[i + 1] = bi[i] + bnzi;
5049:   }

5051:   PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));

5053:   PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5054:   PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5055:   PetscCall(PetscLLDestroy(lnk, lnkbt));

5057:   /* create symbolic parallel matrix B_mpi */
5058:   PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5059:   PetscCall(MatCreate(comm, &B_mpi));
5060:   if (n == PETSC_DECIDE) {
5061:     PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5062:   } else {
5063:     PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5064:   }
5065:   PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5066:   PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5067:   PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5068:   MatPreallocateEnd(dnz, onz);
5069:   PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));

5071:   /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5072:   B_mpi->assembled = PETSC_FALSE;
5073:   merge->bi        = bi;
5074:   merge->bj        = bj;
5075:   merge->buf_ri    = buf_ri;
5076:   merge->buf_rj    = buf_rj;
5077:   merge->coi       = NULL;
5078:   merge->coj       = NULL;
5079:   merge->owners_co = NULL;

5081:   PetscCall(PetscCommDestroy(&comm));

5083:   /* attach the supporting struct to B_mpi for reuse */
5084:   PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5085:   PetscCall(PetscContainerSetPointer(container, merge));
5086:   PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5087:   PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5088:   PetscCall(PetscContainerDestroy(&container));
5089:   *mpimat = B_mpi;

5091:   PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5092:   PetscFunctionReturn(PETSC_SUCCESS);
5093: }

5095: /*@C
5096:       MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5097:                  matrices from each processor

5099:     Collective

5101:    Input Parameters:
5102: +    comm - the communicators the parallel matrix will live on
5103: .    seqmat - the input sequential matrices
5104: .    m - number of local rows (or `PETSC_DECIDE`)
5105: .    n - number of local columns (or `PETSC_DECIDE`)
5106: -    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5108:    Output Parameter:
5109: .    mpimat - the parallel matrix generated

5111:     Level: advanced

5113:    Note:
5114:      The dimensions of the sequential matrix in each processor MUST be the same.
5115:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5116:      destroyed when mpimat is destroyed. Call `PetscObjectQuery()` to access seqmat.

5118: seealso: [](chapter_matrices), `Mat`, `MatCreateAIJ()`
5119: @*/
5120: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5121: {
5122:   PetscMPIInt size;

5124:   PetscFunctionBegin;
5125:   PetscCallMPI(MPI_Comm_size(comm, &size));
5126:   if (size == 1) {
5127:     PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5128:     if (scall == MAT_INITIAL_MATRIX) {
5129:       PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5130:     } else {
5131:       PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5132:     }
5133:     PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5134:     PetscFunctionReturn(PETSC_SUCCESS);
5135:   }
5136:   PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5137:   if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5138:   PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5139:   PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5140:   PetscFunctionReturn(PETSC_SUCCESS);
5141: }

5143: /*@
5144:      MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix by taking its local rows and putting them into a sequential matrix with
5145:           mlocal rows and n columns. Where mlocal is obtained with `MatGetLocalSize()` and n is the global column count obtained
5146:           with `MatGetSize()`

5148:     Not Collective

5150:    Input Parameter:
5151: .    A - the matrix

5153:    Output Parameter:
5154: .    A_loc - the local sequential matrix generated

5156:     Level: developer

5158:    Notes:
5159:      In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.

5161:      Destroy the matrix with `MatDestroy()`

5163: .seealso: [](chapter_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5164: @*/
5165: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5166: {
5167:   PetscBool mpi;

5169:   PetscFunctionBegin;
5170:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5171:   if (mpi) {
5172:     PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5173:   } else {
5174:     *A_loc = A;
5175:     PetscCall(PetscObjectReference((PetscObject)*A_loc));
5176:   }
5177:   PetscFunctionReturn(PETSC_SUCCESS);
5178: }

5180: /*@
5181:      MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5182:           mlocal rows and n columns. Where mlocal is the row count obtained with `MatGetLocalSize()` and n is the global column count obtained
5183:           with `MatGetSize()`

5185:     Not Collective

5187:    Input Parameters:
5188: +    A - the matrix
5189: -    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5191:    Output Parameter:
5192: .    A_loc - the local sequential matrix generated

5194:     Level: developer

5196:    Notes:
5197:      In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.

5199:      When the communicator associated with `A` has size 1 and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A`.
5200:      If `MAT_REUSE_MATRIX` is requested with comm size 1, `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called.
5201:      This means that one can preallocate the proper sequential matrix first and then call this routine with `MAT_REUSE_MATRIX` to safely
5202:      modify the values of the returned `A_loc`.

5204: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5205: @*/
5206: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5207: {
5208:   Mat_MPIAIJ        *mpimat = (Mat_MPIAIJ *)A->data;
5209:   Mat_SeqAIJ        *mat, *a, *b;
5210:   PetscInt          *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5211:   const PetscScalar *aa, *ba, *aav, *bav;
5212:   PetscScalar       *ca, *cam;
5213:   PetscMPIInt        size;
5214:   PetscInt           am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5215:   PetscInt          *ci, *cj, col, ncols_d, ncols_o, jo;
5216:   PetscBool          match;

5218:   PetscFunctionBegin;
5219:   PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5220:   PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5221:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5222:   if (size == 1) {
5223:     if (scall == MAT_INITIAL_MATRIX) {
5224:       PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5225:       *A_loc = mpimat->A;
5226:     } else if (scall == MAT_REUSE_MATRIX) {
5227:       PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5228:     }
5229:     PetscFunctionReturn(PETSC_SUCCESS);
5230:   }

5232:   PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5233:   a  = (Mat_SeqAIJ *)(mpimat->A)->data;
5234:   b  = (Mat_SeqAIJ *)(mpimat->B)->data;
5235:   ai = a->i;
5236:   aj = a->j;
5237:   bi = b->i;
5238:   bj = b->j;
5239:   PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5240:   PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5241:   aa = aav;
5242:   ba = bav;
5243:   if (scall == MAT_INITIAL_MATRIX) {
5244:     PetscCall(PetscMalloc1(1 + am, &ci));
5245:     ci[0] = 0;
5246:     for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5247:     PetscCall(PetscMalloc1(1 + ci[am], &cj));
5248:     PetscCall(PetscMalloc1(1 + ci[am], &ca));
5249:     k = 0;
5250:     for (i = 0; i < am; i++) {
5251:       ncols_o = bi[i + 1] - bi[i];
5252:       ncols_d = ai[i + 1] - ai[i];
5253:       /* off-diagonal portion of A */
5254:       for (jo = 0; jo < ncols_o; jo++) {
5255:         col = cmap[*bj];
5256:         if (col >= cstart) break;
5257:         cj[k] = col;
5258:         bj++;
5259:         ca[k++] = *ba++;
5260:       }
5261:       /* diagonal portion of A */
5262:       for (j = 0; j < ncols_d; j++) {
5263:         cj[k]   = cstart + *aj++;
5264:         ca[k++] = *aa++;
5265:       }
5266:       /* off-diagonal portion of A */
5267:       for (j = jo; j < ncols_o; j++) {
5268:         cj[k]   = cmap[*bj++];
5269:         ca[k++] = *ba++;
5270:       }
5271:     }
5272:     /* put together the new matrix */
5273:     PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5274:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5275:     /* Since these are PETSc arrays, change flags to free them as necessary. */
5276:     mat          = (Mat_SeqAIJ *)(*A_loc)->data;
5277:     mat->free_a  = PETSC_TRUE;
5278:     mat->free_ij = PETSC_TRUE;
5279:     mat->nonew   = 0;
5280:   } else if (scall == MAT_REUSE_MATRIX) {
5281:     mat = (Mat_SeqAIJ *)(*A_loc)->data;
5282:     ci  = mat->i;
5283:     cj  = mat->j;
5284:     PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5285:     for (i = 0; i < am; i++) {
5286:       /* off-diagonal portion of A */
5287:       ncols_o = bi[i + 1] - bi[i];
5288:       for (jo = 0; jo < ncols_o; jo++) {
5289:         col = cmap[*bj];
5290:         if (col >= cstart) break;
5291:         *cam++ = *ba++;
5292:         bj++;
5293:       }
5294:       /* diagonal portion of A */
5295:       ncols_d = ai[i + 1] - ai[i];
5296:       for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5297:       /* off-diagonal portion of A */
5298:       for (j = jo; j < ncols_o; j++) {
5299:         *cam++ = *ba++;
5300:         bj++;
5301:       }
5302:     }
5303:     PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5304:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5305:   PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5306:   PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5307:   PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5308:   PetscFunctionReturn(PETSC_SUCCESS);
5309: }

5311: /*@
5312:      MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5313:           mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and offdiagonal part

5315:     Not Collective

5317:    Input Parameters:
5318: +    A - the matrix
5319: -    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5321:    Output Parameters:
5322: +    glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5323: -    A_loc - the local sequential matrix generated

5325:     Level: developer

5327:    Note:
5328:      This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5329:      part, then those associated with the off diagonal part (in its local ordering)

5331: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5332: @*/
5333: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5334: {
5335:   Mat             Ao, Ad;
5336:   const PetscInt *cmap;
5337:   PetscMPIInt     size;
5338:   PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);

5340:   PetscFunctionBegin;
5341:   PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5342:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5343:   if (size == 1) {
5344:     if (scall == MAT_INITIAL_MATRIX) {
5345:       PetscCall(PetscObjectReference((PetscObject)Ad));
5346:       *A_loc = Ad;
5347:     } else if (scall == MAT_REUSE_MATRIX) {
5348:       PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5349:     }
5350:     if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5351:     PetscFunctionReturn(PETSC_SUCCESS);
5352:   }
5353:   PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5354:   PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5355:   if (f) {
5356:     PetscCall((*f)(A, scall, glob, A_loc));
5357:   } else {
5358:     Mat_SeqAIJ        *a = (Mat_SeqAIJ *)Ad->data;
5359:     Mat_SeqAIJ        *b = (Mat_SeqAIJ *)Ao->data;
5360:     Mat_SeqAIJ        *c;
5361:     PetscInt          *ai = a->i, *aj = a->j;
5362:     PetscInt          *bi = b->i, *bj = b->j;
5363:     PetscInt          *ci, *cj;
5364:     const PetscScalar *aa, *ba;
5365:     PetscScalar       *ca;
5366:     PetscInt           i, j, am, dn, on;

5368:     PetscCall(MatGetLocalSize(Ad, &am, &dn));
5369:     PetscCall(MatGetLocalSize(Ao, NULL, &on));
5370:     PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5371:     PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5372:     if (scall == MAT_INITIAL_MATRIX) {
5373:       PetscInt k;
5374:       PetscCall(PetscMalloc1(1 + am, &ci));
5375:       PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5376:       PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5377:       ci[0] = 0;
5378:       for (i = 0, k = 0; i < am; i++) {
5379:         const PetscInt ncols_o = bi[i + 1] - bi[i];
5380:         const PetscInt ncols_d = ai[i + 1] - ai[i];
5381:         ci[i + 1]              = ci[i] + ncols_o + ncols_d;
5382:         /* diagonal portion of A */
5383:         for (j = 0; j < ncols_d; j++, k++) {
5384:           cj[k] = *aj++;
5385:           ca[k] = *aa++;
5386:         }
5387:         /* off-diagonal portion of A */
5388:         for (j = 0; j < ncols_o; j++, k++) {
5389:           cj[k] = dn + *bj++;
5390:           ca[k] = *ba++;
5391:         }
5392:       }
5393:       /* put together the new matrix */
5394:       PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5395:       /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5396:       /* Since these are PETSc arrays, change flags to free them as necessary. */
5397:       c          = (Mat_SeqAIJ *)(*A_loc)->data;
5398:       c->free_a  = PETSC_TRUE;
5399:       c->free_ij = PETSC_TRUE;
5400:       c->nonew   = 0;
5401:       PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5402:     } else if (scall == MAT_REUSE_MATRIX) {
5403:       PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5404:       for (i = 0; i < am; i++) {
5405:         const PetscInt ncols_d = ai[i + 1] - ai[i];
5406:         const PetscInt ncols_o = bi[i + 1] - bi[i];
5407:         /* diagonal portion of A */
5408:         for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5409:         /* off-diagonal portion of A */
5410:         for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5411:       }
5412:       PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5413:     } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5414:     PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5415:     PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5416:     if (glob) {
5417:       PetscInt cst, *gidx;

5419:       PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5420:       PetscCall(PetscMalloc1(dn + on, &gidx));
5421:       for (i = 0; i < dn; i++) gidx[i] = cst + i;
5422:       for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5423:       PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5424:     }
5425:   }
5426:   PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5427:   PetscFunctionReturn(PETSC_SUCCESS);
5428: }

5430: /*@C
5431:      MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns

5433:     Not Collective

5435:    Input Parameters:
5436: +    A - the matrix
5437: .    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5438: .    row - index set of rows to extract (or `NULL`)
5439: -    col - index set of columns to extract (or `NULL`)

5441:    Output Parameter:
5442: .    A_loc - the local sequential matrix generated

5444:     Level: developer

5446: .seealso: [](chapter_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5447: @*/
5448: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5449: {
5450:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5451:   PetscInt    i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5452:   IS          isrowa, iscola;
5453:   Mat        *aloc;
5454:   PetscBool   match;

5456:   PetscFunctionBegin;
5457:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5458:   PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5459:   PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5460:   if (!row) {
5461:     start = A->rmap->rstart;
5462:     end   = A->rmap->rend;
5463:     PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5464:   } else {
5465:     isrowa = *row;
5466:   }
5467:   if (!col) {
5468:     start = A->cmap->rstart;
5469:     cmap  = a->garray;
5470:     nzA   = a->A->cmap->n;
5471:     nzB   = a->B->cmap->n;
5472:     PetscCall(PetscMalloc1(nzA + nzB, &idx));
5473:     ncols = 0;
5474:     for (i = 0; i < nzB; i++) {
5475:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5476:       else break;
5477:     }
5478:     imark = i;
5479:     for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5480:     for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5481:     PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5482:   } else {
5483:     iscola = *col;
5484:   }
5485:   if (scall != MAT_INITIAL_MATRIX) {
5486:     PetscCall(PetscMalloc1(1, &aloc));
5487:     aloc[0] = *A_loc;
5488:   }
5489:   PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5490:   if (!col) { /* attach global id of condensed columns */
5491:     PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5492:   }
5493:   *A_loc = aloc[0];
5494:   PetscCall(PetscFree(aloc));
5495:   if (!row) PetscCall(ISDestroy(&isrowa));
5496:   if (!col) PetscCall(ISDestroy(&iscola));
5497:   PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5498:   PetscFunctionReturn(PETSC_SUCCESS);
5499: }

5501: /*
5502:  * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5503:  * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5504:  * on a global size.
5505:  * */
5506: PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5507: {
5508:   Mat_MPIAIJ            *p  = (Mat_MPIAIJ *)P->data;
5509:   Mat_SeqAIJ            *pd = (Mat_SeqAIJ *)(p->A)->data, *po = (Mat_SeqAIJ *)(p->B)->data, *p_oth;
5510:   PetscInt               plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5511:   PetscMPIInt            owner;
5512:   PetscSFNode           *iremote, *oiremote;
5513:   const PetscInt        *lrowindices;
5514:   PetscSF                sf, osf;
5515:   PetscInt               pcstart, *roffsets, *loffsets, *pnnz, j;
5516:   PetscInt               ontotalcols, dntotalcols, ntotalcols, nout;
5517:   MPI_Comm               comm;
5518:   ISLocalToGlobalMapping mapping;
5519:   const PetscScalar     *pd_a, *po_a;

5521:   PetscFunctionBegin;
5522:   PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5523:   /* plocalsize is the number of roots
5524:    * nrows is the number of leaves
5525:    * */
5526:   PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5527:   PetscCall(ISGetLocalSize(rows, &nrows));
5528:   PetscCall(PetscCalloc1(nrows, &iremote));
5529:   PetscCall(ISGetIndices(rows, &lrowindices));
5530:   for (i = 0; i < nrows; i++) {
5531:     /* Find a remote index and an owner for a row
5532:      * The row could be local or remote
5533:      * */
5534:     owner = 0;
5535:     lidx  = 0;
5536:     PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5537:     iremote[i].index = lidx;
5538:     iremote[i].rank  = owner;
5539:   }
5540:   /* Create SF to communicate how many nonzero columns for each row */
5541:   PetscCall(PetscSFCreate(comm, &sf));
5542:   /* SF will figure out the number of nonzero colunms for each row, and their
5543:    * offsets
5544:    * */
5545:   PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5546:   PetscCall(PetscSFSetFromOptions(sf));
5547:   PetscCall(PetscSFSetUp(sf));

5549:   PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5550:   PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5551:   PetscCall(PetscCalloc1(nrows, &pnnz));
5552:   roffsets[0] = 0;
5553:   roffsets[1] = 0;
5554:   for (i = 0; i < plocalsize; i++) {
5555:     /* diag */
5556:     nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5557:     /* off diag */
5558:     nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5559:     /* compute offsets so that we relative location for each row */
5560:     roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5561:     roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5562:   }
5563:   PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5564:   PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5565:   /* 'r' means root, and 'l' means leaf */
5566:   PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5567:   PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5568:   PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5569:   PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5570:   PetscCall(PetscSFDestroy(&sf));
5571:   PetscCall(PetscFree(roffsets));
5572:   PetscCall(PetscFree(nrcols));
5573:   dntotalcols = 0;
5574:   ontotalcols = 0;
5575:   ncol        = 0;
5576:   for (i = 0; i < nrows; i++) {
5577:     pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5578:     ncol    = PetscMax(pnnz[i], ncol);
5579:     /* diag */
5580:     dntotalcols += nlcols[i * 2 + 0];
5581:     /* off diag */
5582:     ontotalcols += nlcols[i * 2 + 1];
5583:   }
5584:   /* We do not need to figure the right number of columns
5585:    * since all the calculations will be done by going through the raw data
5586:    * */
5587:   PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5588:   PetscCall(MatSetUp(*P_oth));
5589:   PetscCall(PetscFree(pnnz));
5590:   p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5591:   /* diag */
5592:   PetscCall(PetscCalloc1(dntotalcols, &iremote));
5593:   /* off diag */
5594:   PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5595:   /* diag */
5596:   PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5597:   /* off diag */
5598:   PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5599:   dntotalcols = 0;
5600:   ontotalcols = 0;
5601:   ntotalcols  = 0;
5602:   for (i = 0; i < nrows; i++) {
5603:     owner = 0;
5604:     PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5605:     /* Set iremote for diag matrix */
5606:     for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5607:       iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5608:       iremote[dntotalcols].rank  = owner;
5609:       /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5610:       ilocal[dntotalcols++] = ntotalcols++;
5611:     }
5612:     /* off diag */
5613:     for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5614:       oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5615:       oiremote[ontotalcols].rank  = owner;
5616:       oilocal[ontotalcols++]      = ntotalcols++;
5617:     }
5618:   }
5619:   PetscCall(ISRestoreIndices(rows, &lrowindices));
5620:   PetscCall(PetscFree(loffsets));
5621:   PetscCall(PetscFree(nlcols));
5622:   PetscCall(PetscSFCreate(comm, &sf));
5623:   /* P serves as roots and P_oth is leaves
5624:    * Diag matrix
5625:    * */
5626:   PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5627:   PetscCall(PetscSFSetFromOptions(sf));
5628:   PetscCall(PetscSFSetUp(sf));

5630:   PetscCall(PetscSFCreate(comm, &osf));
5631:   /* Off diag */
5632:   PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5633:   PetscCall(PetscSFSetFromOptions(osf));
5634:   PetscCall(PetscSFSetUp(osf));
5635:   PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5636:   PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5637:   /* We operate on the matrix internal data for saving memory */
5638:   PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5639:   PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5640:   PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5641:   /* Convert to global indices for diag matrix */
5642:   for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5643:   PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5644:   /* We want P_oth store global indices */
5645:   PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5646:   /* Use memory scalable approach */
5647:   PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5648:   PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5649:   PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5650:   PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5651:   /* Convert back to local indices */
5652:   for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5653:   PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5654:   nout = 0;
5655:   PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5656:   PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5657:   PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5658:   /* Exchange values */
5659:   PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5660:   PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5661:   PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5662:   PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5663:   /* Stop PETSc from shrinking memory */
5664:   for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5665:   PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5666:   PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5667:   /* Attach PetscSF objects to P_oth so that we can reuse it later */
5668:   PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5669:   PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5670:   PetscCall(PetscSFDestroy(&sf));
5671:   PetscCall(PetscSFDestroy(&osf));
5672:   PetscFunctionReturn(PETSC_SUCCESS);
5673: }

5675: /*
5676:  * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5677:  * This supports MPIAIJ and MAIJ
5678:  * */
5679: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5680: {
5681:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5682:   Mat_SeqAIJ *p_oth;
5683:   IS          rows, map;
5684:   PetscHMapI  hamp;
5685:   PetscInt    i, htsize, *rowindices, off, *mapping, key, count;
5686:   MPI_Comm    comm;
5687:   PetscSF     sf, osf;
5688:   PetscBool   has;

5690:   PetscFunctionBegin;
5691:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5692:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5693:   /* If it is the first time, create an index set of off-diag nonzero columns of A,
5694:    *  and then create a submatrix (that often is an overlapping matrix)
5695:    * */
5696:   if (reuse == MAT_INITIAL_MATRIX) {
5697:     /* Use a hash table to figure out unique keys */
5698:     PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5699:     PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5700:     count = 0;
5701:     /* Assume that  a->g is sorted, otherwise the following does not make sense */
5702:     for (i = 0; i < a->B->cmap->n; i++) {
5703:       key = a->garray[i] / dof;
5704:       PetscCall(PetscHMapIHas(hamp, key, &has));
5705:       if (!has) {
5706:         mapping[i] = count;
5707:         PetscCall(PetscHMapISet(hamp, key, count++));
5708:       } else {
5709:         /* Current 'i' has the same value the previous step */
5710:         mapping[i] = count - 1;
5711:       }
5712:     }
5713:     PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5714:     PetscCall(PetscHMapIGetSize(hamp, &htsize));
5715:     PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5716:     PetscCall(PetscCalloc1(htsize, &rowindices));
5717:     off = 0;
5718:     PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5719:     PetscCall(PetscHMapIDestroy(&hamp));
5720:     PetscCall(PetscSortInt(htsize, rowindices));
5721:     PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5722:     /* In case, the matrix was already created but users want to recreate the matrix */
5723:     PetscCall(MatDestroy(P_oth));
5724:     PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5725:     PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5726:     PetscCall(ISDestroy(&map));
5727:     PetscCall(ISDestroy(&rows));
5728:   } else if (reuse == MAT_REUSE_MATRIX) {
5729:     /* If matrix was already created, we simply update values using SF objects
5730:      * that as attached to the matrix earlier.
5731:      */
5732:     const PetscScalar *pd_a, *po_a;

5734:     PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5735:     PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5736:     PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5737:     p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5738:     /* Update values in place */
5739:     PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5740:     PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5741:     PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5742:     PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5743:     PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5744:     PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5745:     PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5746:     PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5747:   } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5748:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5749:   PetscFunctionReturn(PETSC_SUCCESS);
5750: }

5752: /*@C
5753:   MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`

5755:   Collective

5757:   Input Parameters:
5758: + A - the first matrix in `MATMPIAIJ` format
5759: . B - the second matrix in `MATMPIAIJ` format
5760: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5762:   Output Parameters:
5763: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5764: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5765: - B_seq - the sequential matrix generated

5767:   Level: developer

5769: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5770: @*/
5771: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5772: {
5773:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5774:   PetscInt   *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5775:   IS          isrowb, iscolb;
5776:   Mat        *bseq = NULL;

5778:   PetscFunctionBegin;
5779:   PetscCheck(A->cmap->rstart == B->rmap->rstart && A->cmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",
5780:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5781:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));

5783:   if (scall == MAT_INITIAL_MATRIX) {
5784:     start = A->cmap->rstart;
5785:     cmap  = a->garray;
5786:     nzA   = a->A->cmap->n;
5787:     nzB   = a->B->cmap->n;
5788:     PetscCall(PetscMalloc1(nzA + nzB, &idx));
5789:     ncols = 0;
5790:     for (i = 0; i < nzB; i++) { /* row < local row index */
5791:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5792:       else break;
5793:     }
5794:     imark = i;
5795:     for (i = 0; i < nzA; i++) idx[ncols++] = start + i;   /* local rows */
5796:     for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5797:     PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5798:     PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5799:   } else {
5800:     PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5801:     isrowb = *rowb;
5802:     iscolb = *colb;
5803:     PetscCall(PetscMalloc1(1, &bseq));
5804:     bseq[0] = *B_seq;
5805:   }
5806:   PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5807:   *B_seq = bseq[0];
5808:   PetscCall(PetscFree(bseq));
5809:   if (!rowb) {
5810:     PetscCall(ISDestroy(&isrowb));
5811:   } else {
5812:     *rowb = isrowb;
5813:   }
5814:   if (!colb) {
5815:     PetscCall(ISDestroy(&iscolb));
5816:   } else {
5817:     *colb = iscolb;
5818:   }
5819:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5820:   PetscFunctionReturn(PETSC_SUCCESS);
5821: }

5823: /*
5824:     MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5825:     of the OFF-DIAGONAL portion of local A

5827:     Collective

5829:    Input Parameters:
5830: +    A,B - the matrices in `MATMPIAIJ` format
5831: -    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5833:    Output Parameter:
5834: +    startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5835: .    startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5836: .    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5837: -    B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N

5839:     Developer Note:
5840:     This directly accesses information inside the VecScatter associated with the matrix-vector product
5841:      for this matrix. This is not desirable..

5843:     Level: developer

5845: */
5846: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5847: {
5848:   Mat_MPIAIJ        *a = (Mat_MPIAIJ *)A->data;
5849:   Mat_SeqAIJ        *b_oth;
5850:   VecScatter         ctx;
5851:   MPI_Comm           comm;
5852:   const PetscMPIInt *rprocs, *sprocs;
5853:   const PetscInt    *srow, *rstarts, *sstarts;
5854:   PetscInt          *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5855:   PetscInt           i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5856:   PetscScalar       *b_otha, *bufa, *bufA, *vals = NULL;
5857:   MPI_Request       *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5858:   PetscMPIInt        size, tag, rank, nreqs;

5860:   PetscFunctionBegin;
5861:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5862:   PetscCallMPI(MPI_Comm_size(comm, &size));

5864:   PetscCheck(A->cmap->rstart == B->rmap->rstart && A->cmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",
5865:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5866:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5867:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

5869:   if (size == 1) {
5870:     startsj_s = NULL;
5871:     bufa_ptr  = NULL;
5872:     *B_oth    = NULL;
5873:     PetscFunctionReturn(PETSC_SUCCESS);
5874:   }

5876:   ctx = a->Mvctx;
5877:   tag = ((PetscObject)ctx)->tag;

5879:   PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5880:   /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5881:   PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5882:   PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5883:   PetscCall(PetscMalloc1(nreqs, &reqs));
5884:   rwaits = reqs;
5885:   swaits = reqs + nrecvs;

5887:   if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5888:   if (scall == MAT_INITIAL_MATRIX) {
5889:     /* i-array */
5890:     /*  post receives */
5891:     if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5892:     for (i = 0; i < nrecvs; i++) {
5893:       rowlen = rvalues + rstarts[i] * rbs;
5894:       nrows  = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5895:       PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5896:     }

5898:     /* pack the outgoing message */
5899:     PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));

5901:     sstartsj[0] = 0;
5902:     rstartsj[0] = 0;
5903:     len         = 0; /* total length of j or a array to be sent */
5904:     if (nsends) {
5905:       k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5906:       PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5907:     }
5908:     for (i = 0; i < nsends; i++) {
5909:       rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5910:       nrows  = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5911:       for (j = 0; j < nrows; j++) {
5912:         row = srow[k] + B->rmap->range[rank]; /* global row idx */
5913:         for (l = 0; l < sbs; l++) {
5914:           PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */

5916:           rowlen[j * sbs + l] = ncols;

5918:           len += ncols;
5919:           PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5920:         }
5921:         k++;
5922:       }
5923:       PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));

5925:       sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5926:     }
5927:     /* recvs and sends of i-array are completed */
5928:     if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5929:     PetscCall(PetscFree(svalues));

5931:     /* allocate buffers for sending j and a arrays */
5932:     PetscCall(PetscMalloc1(len + 1, &bufj));
5933:     PetscCall(PetscMalloc1(len + 1, &bufa));

5935:     /* create i-array of B_oth */
5936:     PetscCall(PetscMalloc1(aBn + 2, &b_othi));

5938:     b_othi[0] = 0;
5939:     len       = 0; /* total length of j or a array to be received */
5940:     k         = 0;
5941:     for (i = 0; i < nrecvs; i++) {
5942:       rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5943:       nrows  = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5944:       for (j = 0; j < nrows; j++) {
5945:         b_othi[k + 1] = b_othi[k] + rowlen[j];
5946:         PetscCall(PetscIntSumError(rowlen[j], len, &len));
5947:         k++;
5948:       }
5949:       rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5950:     }
5951:     PetscCall(PetscFree(rvalues));

5953:     /* allocate space for j and a arrays of B_oth */
5954:     PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5955:     PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));

5957:     /* j-array */
5958:     /*  post receives of j-array */
5959:     for (i = 0; i < nrecvs; i++) {
5960:       nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5961:       PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5962:     }

5964:     /* pack the outgoing message j-array */
5965:     if (nsends) k = sstarts[0];
5966:     for (i = 0; i < nsends; i++) {
5967:       nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5968:       bufJ  = bufj + sstartsj[i];
5969:       for (j = 0; j < nrows; j++) {
5970:         row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5971:         for (ll = 0; ll < sbs; ll++) {
5972:           PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5973:           for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5974:           PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5975:         }
5976:       }
5977:       PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5978:     }

5980:     /* recvs and sends of j-array are completed */
5981:     if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5982:   } else if (scall == MAT_REUSE_MATRIX) {
5983:     sstartsj = *startsj_s;
5984:     rstartsj = *startsj_r;
5985:     bufa     = *bufa_ptr;
5986:     b_oth    = (Mat_SeqAIJ *)(*B_oth)->data;
5987:     PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5988:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");

5990:   /* a-array */
5991:   /*  post receives of a-array */
5992:   for (i = 0; i < nrecvs; i++) {
5993:     nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5994:     PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5995:   }

5997:   /* pack the outgoing message a-array */
5998:   if (nsends) k = sstarts[0];
5999:   for (i = 0; i < nsends; i++) {
6000:     nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
6001:     bufA  = bufa + sstartsj[i];
6002:     for (j = 0; j < nrows; j++) {
6003:       row = srow[k++] + B->rmap->range[rank]; /* global row idx */
6004:       for (ll = 0; ll < sbs; ll++) {
6005:         PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
6006:         for (l = 0; l < ncols; l++) *bufA++ = vals[l];
6007:         PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
6008:       }
6009:     }
6010:     PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
6011:   }
6012:   /* recvs and sends of a-array are completed */
6013:   if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
6014:   PetscCall(PetscFree(reqs));

6016:   if (scall == MAT_INITIAL_MATRIX) {
6017:     /* put together the new matrix */
6018:     PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));

6020:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6021:     /* Since these are PETSc arrays, change flags to free them as necessary. */
6022:     b_oth          = (Mat_SeqAIJ *)(*B_oth)->data;
6023:     b_oth->free_a  = PETSC_TRUE;
6024:     b_oth->free_ij = PETSC_TRUE;
6025:     b_oth->nonew   = 0;

6027:     PetscCall(PetscFree(bufj));
6028:     if (!startsj_s || !bufa_ptr) {
6029:       PetscCall(PetscFree2(sstartsj, rstartsj));
6030:       PetscCall(PetscFree(bufa_ptr));
6031:     } else {
6032:       *startsj_s = sstartsj;
6033:       *startsj_r = rstartsj;
6034:       *bufa_ptr  = bufa;
6035:     }
6036:   } else if (scall == MAT_REUSE_MATRIX) {
6037:     PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6038:   }

6040:   PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6041:   PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6042:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6043:   PetscFunctionReturn(PETSC_SUCCESS);
6044: }

6046: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6047: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6048: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6049: #if defined(PETSC_HAVE_MKL_SPARSE)
6050: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6051: #endif
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6053: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6054: #if defined(PETSC_HAVE_ELEMENTAL)
6055: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: #if defined(PETSC_HAVE_SCALAPACK)
6058: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6059: #endif
6060: #if defined(PETSC_HAVE_HYPRE)
6061: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6062: #endif
6063: #if defined(PETSC_HAVE_CUDA)
6064: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6065: #endif
6066: #if defined(PETSC_HAVE_HIP)
6067: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6068: #endif
6069: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6070: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6071: #endif
6072: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6073: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6074: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);

6076: /*
6077:     Computes (B'*A')' since computing B*A directly is untenable

6079:                n                       p                          p
6080:         [             ]       [             ]         [                 ]
6081:       m [      A      ]  *  n [       B     ]   =   m [         C       ]
6082:         [             ]       [             ]         [                 ]

6084: */
6085: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6086: {
6087:   Mat At, Bt, Ct;

6089:   PetscFunctionBegin;
6090:   PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6091:   PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6092:   PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6093:   PetscCall(MatDestroy(&At));
6094:   PetscCall(MatDestroy(&Bt));
6095:   PetscCall(MatTransposeSetPrecursor(Ct, C));
6096:   PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6097:   PetscCall(MatDestroy(&Ct));
6098:   PetscFunctionReturn(PETSC_SUCCESS);
6099: }

6101: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6102: {
6103:   PetscBool cisdense;

6105:   PetscFunctionBegin;
6106:   PetscCheck(A->cmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "A->cmap->n %" PetscInt_FMT " != B->rmap->n %" PetscInt_FMT, A->cmap->n, B->rmap->n);
6107:   PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6108:   PetscCall(MatSetBlockSizesFromMats(C, A, B));
6109:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6110:   if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6111:   PetscCall(MatSetUp(C));

6113:   C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6114:   PetscFunctionReturn(PETSC_SUCCESS);
6115: }

6117: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6118: {
6119:   Mat_Product *product = C->product;
6120:   Mat          A = product->A, B = product->B;

6122:   PetscFunctionBegin;
6123:   PetscCheck(A->cmap->rstart == B->rmap->rstart && A->cmap->rend == B->rmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Matrix local dimensions are incompatible, (%" PetscInt_FMT ", %" PetscInt_FMT ") != (%" PetscInt_FMT ",%" PetscInt_FMT ")",
6124:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6125:   C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6126:   C->ops->productsymbolic = MatProductSymbolic_AB;
6127:   PetscFunctionReturn(PETSC_SUCCESS);
6128: }

6130: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6131: {
6132:   Mat_Product *product = C->product;

6134:   PetscFunctionBegin;
6135:   if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6136:   PetscFunctionReturn(PETSC_SUCCESS);
6137: }

6139: /*
6140:    Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix

6142:   Input Parameters:

6144:     j1,rowBegin1,rowEnd1,perm1,jmap1: describe the first set of nonzeros (Set1)
6145:     j2,rowBegin2,rowEnd2,perm2,jmap2: describe the second set of nonzeros (Set2)

6147:     mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat

6149:     For Set1, j1[] contains column indices of the nonzeros.
6150:     For the k-th row (0<=k<m), [rowBegin1[k],rowEnd1[k]) index into j1[] and point to the begin/end nonzero in row k
6151:     respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6152:     but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.

6154:     Similar for Set2.

6156:     This routine merges the two sets of nonzeros row by row and removes repeats.

6158:   Output Parameters: (memory is allocated by the caller)

6160:     i[],j[]: the CSR of the merged matrix, which has m rows.
6161:     imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6162:     imap2[]: similar to imap1[], but for Set2.
6163:     Note we order nonzeros row-by-row and from left to right.
6164: */
6165: static PetscErrorCode MatMergeEntries_Internal(Mat mat, const PetscInt j1[], const PetscInt j2[], const PetscCount rowBegin1[], const PetscCount rowEnd1[], const PetscCount rowBegin2[], const PetscCount rowEnd2[], const PetscCount jmap1[], const PetscCount jmap2[], PetscCount imap1[], PetscCount imap2[], PetscInt i[], PetscInt j[])
6166: {
6167:   PetscInt   r, m; /* Row index of mat */
6168:   PetscCount t, t1, t2, b1, e1, b2, e2;

6170:   PetscFunctionBegin;
6171:   PetscCall(MatGetLocalSize(mat, &m, NULL));
6172:   t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6173:   i[0]        = 0;
6174:   for (r = 0; r < m; r++) { /* Do row by row merging */
6175:     b1 = rowBegin1[r];
6176:     e1 = rowEnd1[r];
6177:     b2 = rowBegin2[r];
6178:     e2 = rowEnd2[r];
6179:     while (b1 < e1 && b2 < e2) {
6180:       if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6181:         j[t]      = j1[b1];
6182:         imap1[t1] = t;
6183:         imap2[t2] = t;
6184:         b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6185:         b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6186:         t1++;
6187:         t2++;
6188:         t++;
6189:       } else if (j1[b1] < j2[b2]) {
6190:         j[t]      = j1[b1];
6191:         imap1[t1] = t;
6192:         b1 += jmap1[t1 + 1] - jmap1[t1];
6193:         t1++;
6194:         t++;
6195:       } else {
6196:         j[t]      = j2[b2];
6197:         imap2[t2] = t;
6198:         b2 += jmap2[t2 + 1] - jmap2[t2];
6199:         t2++;
6200:         t++;
6201:       }
6202:     }
6203:     /* Merge the remaining in either j1[] or j2[] */
6204:     while (b1 < e1) {
6205:       j[t]      = j1[b1];
6206:       imap1[t1] = t;
6207:       b1 += jmap1[t1 + 1] - jmap1[t1];
6208:       t1++;
6209:       t++;
6210:     }
6211:     while (b2 < e2) {
6212:       j[t]      = j2[b2];
6213:       imap2[t2] = t;
6214:       b2 += jmap2[t2 + 1] - jmap2[t2];
6215:       t2++;
6216:       t++;
6217:     }
6218:     i[r + 1] = t;
6219:   }
6220:   PetscFunctionReturn(PETSC_SUCCESS);
6221: }

6223: /*
6224:   Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block

6226:   Input Parameters:
6227:     mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6228:     n,i[],j[],perm[]: there are n input entries, belonging to m rows. Row/col indices of the entries are stored in i[] and j[]
6229:       respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.

6231:       i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6232:       i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.

6234:   Output Parameters:
6235:     j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6236:     rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6237:       They contain indices pointing to j[]. For 0<=r<m, [rowBegin[r],rowMid[r]) point to begin/end entries of row r of the diagonal block,
6238:       and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.

6240:     Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6241:       Atot: number of entries belonging to the diagonal block.
6242:       Annz: number of unique nonzeros belonging to the diagonal block.
6243:       Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6244:         repeats (i.e., same 'i,j' pair).
6245:       Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6246:         is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.

6248:       Atot: number of entries belonging to the diagonal block
6249:       Annz: number of unique nonzeros belonging to the diagonal block.

6251:     Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.

6253:     Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6254: */
6255: static PetscErrorCode MatSplitEntries_Internal(Mat mat, PetscCount n, const PetscInt i[], PetscInt j[], PetscCount perm[], PetscCount rowBegin[], PetscCount rowMid[], PetscCount rowEnd[], PetscCount *Atot_, PetscCount **Aperm_, PetscCount *Annz_, PetscCount **Ajmap_, PetscCount *Btot_, PetscCount **Bperm_, PetscCount *Bnnz_, PetscCount **Bjmap_)
6256: {
6257:   PetscInt    cstart, cend, rstart, rend, row, col;
6258:   PetscCount  Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6259:   PetscCount  Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6260:   PetscCount  k, m, p, q, r, s, mid;
6261:   PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;

6263:   PetscFunctionBegin;
6264:   PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6265:   PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6266:   m = rend - rstart;

6268:   for (k = 0; k < n; k++) {
6269:     if (i[k] >= 0) break;
6270:   } /* Skip negative rows */

6272:   /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6273:      fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6274:   */
6275:   while (k < n) {
6276:     row = i[k];
6277:     /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6278:     for (s = k; s < n; s++)
6279:       if (i[s] != row) break;
6280:     for (p = k; p < s; p++) {
6281:       if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT; /* Shift diag columns to range of [-PETSC_MAX_INT, -1]  */
6282:       else PetscAssert((j[p] >= 0) && (j[p] <= mat->cmap->N), PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index %" PetscInt_FMT " is out of range", j[p]);
6283:     }
6284:     PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6285:     PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6286:     rowBegin[row - rstart] = k;
6287:     rowMid[row - rstart]   = mid;
6288:     rowEnd[row - rstart]   = s;

6290:     /* Count nonzeros of this diag/offdiag row, which might have repeats */
6291:     Atot += mid - k;
6292:     Btot += s - mid;

6294:     /* Count unique nonzeros of this diag/offdiag row */
6295:     for (p = k; p < mid;) {
6296:       col = j[p];
6297:       do {
6298:         j[p] += PETSC_MAX_INT;
6299:         p++;
6300:       } while (p < mid && j[p] == col); /* Revert the modified diagonal indices */
6301:       Annz++;
6302:     }

6304:     for (p = mid; p < s;) {
6305:       col = j[p];
6306:       do {
6307:         p++;
6308:       } while (p < s && j[p] == col);
6309:       Bnnz++;
6310:     }
6311:     k = s;
6312:   }

6314:   /* Allocation according to Atot, Btot, Annz, Bnnz */
6315:   PetscCall(PetscMalloc1(Atot, &Aperm));
6316:   PetscCall(PetscMalloc1(Btot, &Bperm));
6317:   PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6318:   PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));

6320:   /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6321:   Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6322:   for (r = 0; r < m; r++) {
6323:     k   = rowBegin[r];
6324:     mid = rowMid[r];
6325:     s   = rowEnd[r];
6326:     PetscCall(PetscArraycpy(Aperm + Atot, perm + k, mid - k));
6327:     PetscCall(PetscArraycpy(Bperm + Btot, perm + mid, s - mid));
6328:     Atot += mid - k;
6329:     Btot += s - mid;

6331:     /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6332:     for (p = k; p < mid;) {
6333:       col = j[p];
6334:       q   = p;
6335:       do {
6336:         p++;
6337:       } while (p < mid && j[p] == col);
6338:       Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6339:       Annz++;
6340:     }

6342:     for (p = mid; p < s;) {
6343:       col = j[p];
6344:       q   = p;
6345:       do {
6346:         p++;
6347:       } while (p < s && j[p] == col);
6348:       Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6349:       Bnnz++;
6350:     }
6351:   }
6352:   /* Output */
6353:   *Aperm_ = Aperm;
6354:   *Annz_  = Annz;
6355:   *Atot_  = Atot;
6356:   *Ajmap_ = Ajmap;
6357:   *Bperm_ = Bperm;
6358:   *Bnnz_  = Bnnz;
6359:   *Btot_  = Btot;
6360:   *Bjmap_ = Bjmap;
6361:   PetscFunctionReturn(PETSC_SUCCESS);
6362: }

6364: /*
6365:   Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix

6367:   Input Parameters:
6368:     nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6369:     nnz:  number of unique nonzeros in the merged matrix
6370:     imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6371:     jmap[nnz1+1]: i-th nonzeron in the set has jmap[i+1] - jmap[i] repeats in the set

6373:   Output Parameter: (memory is allocated by the caller)
6374:     jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set

6376:   Example:
6377:     nnz1 = 4
6378:     nnz  = 6
6379:     imap = [1,3,4,5]
6380:     jmap = [0,3,5,6,7]
6381:    then,
6382:     jmap_new = [0,0,3,3,5,6,7]
6383: */
6384: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6385: {
6386:   PetscCount k, p;

6388:   PetscFunctionBegin;
6389:   jmap_new[0] = 0;
6390:   p           = nnz;                /* p loops over jmap_new[] backwards */
6391:   for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6392:     for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6393:   }
6394:   for (; p >= 0; p--) jmap_new[p] = jmap[0];
6395:   PetscFunctionReturn(PETSC_SUCCESS);
6396: }

6398: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6399: {
6400:   MPI_Comm    comm;
6401:   PetscMPIInt rank, size;
6402:   PetscInt    m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6403:   PetscCount  k, p, q, rem;                           /* Loop variables over coo arrays */
6404:   Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;

6406:   PetscFunctionBegin;
6407:   PetscCall(PetscFree(mpiaij->garray));
6408:   PetscCall(VecDestroy(&mpiaij->lvec));
6409: #if defined(PETSC_USE_CTABLE)
6410:   PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6411: #else
6412:   PetscCall(PetscFree(mpiaij->colmap));
6413: #endif
6414:   PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6415:   mat->assembled     = PETSC_FALSE;
6416:   mat->was_assembled = PETSC_FALSE;
6417:   PetscCall(MatResetPreallocationCOO_MPIAIJ(mat));

6419:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6420:   PetscCallMPI(MPI_Comm_size(comm, &size));
6421:   PetscCallMPI(MPI_Comm_rank(comm, &rank));
6422:   PetscCall(PetscLayoutSetUp(mat->rmap));
6423:   PetscCall(PetscLayoutSetUp(mat->cmap));
6424:   PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6425:   PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6426:   PetscCall(MatGetLocalSize(mat, &m, &n));
6427:   PetscCall(MatGetSize(mat, &M, &N));

6429:   /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6430:   /* entries come first, then local rows, then remote rows.                     */
6431:   PetscCount n1 = coo_n, *perm1;
6432:   PetscInt  *i1 = coo_i, *j1 = coo_j;

6434:   PetscCall(PetscMalloc1(n1, &perm1));
6435:   for (k = 0; k < n1; k++) perm1[k] = k;

6437:   /* Manipulate indices so that entries with negative row or col indices will have smallest
6438:      row indices, local entries will have greater but negative row indices, and remote entries
6439:      will have positive row indices.
6440:   */
6441:   for (k = 0; k < n1; k++) {
6442:     if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT;                /* e.g., -2^31, minimal to move them ahead */
6443:     else if (i1[k] >= rstart && i1[k] < rend) i1[k] -= PETSC_MAX_INT; /* e.g., minus 2^31-1 to shift local rows to range of [-PETSC_MAX_INT, -1] */
6444:     else {
6445:       PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6446:       if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6447:     }
6448:   }

6450:   /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6451:   PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6452:   for (k = 0; k < n1; k++) {
6453:     if (i1[k] > PETSC_MIN_INT) break;
6454:   }                                                                               /* Advance k to the first entry we need to take care of */
6455:   PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6456:   for (; k < rem; k++) i1[k] += PETSC_MAX_INT;                                    /* Revert row indices of local rows*/

6458:   /*           Split local rows into diag/offdiag portions                      */
6459:   PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6460:   PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1, *Cperm1;
6461:   PetscCount  Annz1, Bnnz1, Atot1, Btot1;

6463:   PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6464:   PetscCall(PetscMalloc1(n1 - rem, &Cperm1));
6465:   PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));

6467:   /*           Send remote rows to their owner                                  */
6468:   /* Find which rows should be sent to which remote ranks*/
6469:   PetscInt        nsend = 0; /* Number of MPI ranks to send data to */
6470:   PetscMPIInt    *sendto;    /* [nsend], storing remote ranks */
6471:   PetscInt       *nentries;  /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6472:   const PetscInt *ranges;
6473:   PetscInt        maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */

6475:   PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6476:   PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6477:   for (k = rem; k < n1;) {
6478:     PetscMPIInt owner;
6479:     PetscInt    firstRow, lastRow;

6481:     /* Locate a row range */
6482:     firstRow = i1[k]; /* first row of this owner */
6483:     PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6484:     lastRow = ranges[owner + 1] - 1; /* last row of this owner */

6486:     /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6487:     PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));

6489:     /* All entries in [k,p) belong to this remote owner */
6490:     if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6491:       PetscMPIInt *sendto2;
6492:       PetscInt    *nentries2;
6493:       PetscInt     maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;

6495:       PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6496:       PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6497:       PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6498:       PetscCall(PetscFree2(sendto, nentries2));
6499:       sendto   = sendto2;
6500:       nentries = nentries2;
6501:       maxNsend = maxNsend2;
6502:     }
6503:     sendto[nsend]   = owner;
6504:     nentries[nsend] = p - k;
6505:     PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6506:     nsend++;
6507:     k = p;
6508:   }

6510:   /* Build 1st SF to know offsets on remote to send data */
6511:   PetscSF      sf1;
6512:   PetscInt     nroots = 1, nroots2 = 0;
6513:   PetscInt     nleaves = nsend, nleaves2 = 0;
6514:   PetscInt    *offsets;
6515:   PetscSFNode *iremote;

6517:   PetscCall(PetscSFCreate(comm, &sf1));
6518:   PetscCall(PetscMalloc1(nsend, &iremote));
6519:   PetscCall(PetscMalloc1(nsend, &offsets));
6520:   for (k = 0; k < nsend; k++) {
6521:     iremote[k].rank  = sendto[k];
6522:     iremote[k].index = 0;
6523:     nleaves2 += nentries[k];
6524:     PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6525:   }
6526:   PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6527:   PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6528:   PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6529:   PetscCall(PetscSFDestroy(&sf1));
6530:   PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT "", nleaves2, n1 - rem);

6532:   /* Build 2nd SF to send remote COOs to their owner */
6533:   PetscSF sf2;
6534:   nroots  = nroots2;
6535:   nleaves = nleaves2;
6536:   PetscCall(PetscSFCreate(comm, &sf2));
6537:   PetscCall(PetscSFSetFromOptions(sf2));
6538:   PetscCall(PetscMalloc1(nleaves, &iremote));
6539:   p = 0;
6540:   for (k = 0; k < nsend; k++) {
6541:     PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6542:     for (q = 0; q < nentries[k]; q++, p++) {
6543:       iremote[p].rank  = sendto[k];
6544:       iremote[p].index = offsets[k] + q;
6545:     }
6546:   }
6547:   PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));

6549:   /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6550:   PetscCall(PetscArraycpy(Cperm1, perm1 + rem, n1 - rem));

6552:   /* Send the remote COOs to their owner */
6553:   PetscInt    n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6554:   PetscCount *perm2;                 /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6555:   PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6556:   PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1 + rem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6557:   PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1 + rem, i2, MPI_REPLACE));
6558:   PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1 + rem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6559:   PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1 + rem, j2, MPI_REPLACE));

6561:   PetscCall(PetscFree(offsets));
6562:   PetscCall(PetscFree2(sendto, nentries));

6564:   /* Sort received COOs by row along with the permutation array     */
6565:   for (k = 0; k < n2; k++) perm2[k] = k;
6566:   PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));

6568:   /* Split received COOs into diag/offdiag portions                 */
6569:   PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6570:   PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6571:   PetscCount  Annz2, Bnnz2, Atot2, Btot2;

6573:   PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6574:   PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));

6576:   /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6577:   PetscInt *Ai, *Bi;
6578:   PetscInt *Aj, *Bj;

6580:   PetscCall(PetscMalloc1(m + 1, &Ai));
6581:   PetscCall(PetscMalloc1(m + 1, &Bi));
6582:   PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6583:   PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));

6585:   PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6586:   PetscCall(PetscMalloc1(Annz1, &Aimap1));
6587:   PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6588:   PetscCall(PetscMalloc1(Annz2, &Aimap2));
6589:   PetscCall(PetscMalloc1(Bnnz2, &Bimap2));

6591:   PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6592:   PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));

6594:   /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we     */
6595:   /* expect nonzeros in A/B most likely have local contributing entries        */
6596:   PetscInt    Annz = Ai[m];
6597:   PetscInt    Bnnz = Bi[m];
6598:   PetscCount *Ajmap1_new, *Bjmap1_new;

6600:   PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6601:   PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));

6603:   PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6604:   PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));

6606:   PetscCall(PetscFree(Aimap1));
6607:   PetscCall(PetscFree(Ajmap1));
6608:   PetscCall(PetscFree(Bimap1));
6609:   PetscCall(PetscFree(Bjmap1));
6610:   PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6611:   PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6612:   PetscCall(PetscFree(perm1));
6613:   PetscCall(PetscFree3(i2, j2, perm2));

6615:   Ajmap1 = Ajmap1_new;
6616:   Bjmap1 = Bjmap1_new;

6618:   /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6619:   if (Annz < Annz1 + Annz2) {
6620:     PetscInt *Aj_new;
6621:     PetscCall(PetscMalloc1(Annz, &Aj_new));
6622:     PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6623:     PetscCall(PetscFree(Aj));
6624:     Aj = Aj_new;
6625:   }

6627:   if (Bnnz < Bnnz1 + Bnnz2) {
6628:     PetscInt *Bj_new;
6629:     PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6630:     PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6631:     PetscCall(PetscFree(Bj));
6632:     Bj = Bj_new;
6633:   }

6635:   /* Create new submatrices for on-process and off-process coupling                  */
6636:   PetscScalar *Aa, *Ba;
6637:   MatType      rtype;
6638:   Mat_SeqAIJ  *a, *b;
6639:   PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6640:   PetscCall(PetscCalloc1(Bnnz, &Ba));
6641:   /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6642:   if (cstart) {
6643:     for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6644:   }
6645:   PetscCall(MatDestroy(&mpiaij->A));
6646:   PetscCall(MatDestroy(&mpiaij->B));
6647:   PetscCall(MatGetRootType_Private(mat, &rtype));
6648:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6649:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6650:   PetscCall(MatSetUpMultiply_MPIAIJ(mat));

6652:   a               = (Mat_SeqAIJ *)mpiaij->A->data;
6653:   b               = (Mat_SeqAIJ *)mpiaij->B->data;
6654:   a->singlemalloc = b->singlemalloc = PETSC_FALSE; /* Let newmat own Ai,Aj,Aa,Bi,Bj,Ba */
6655:   a->free_a = b->free_a = PETSC_TRUE;
6656:   a->free_ij = b->free_ij = PETSC_TRUE;

6658:   /* conversion must happen AFTER multiply setup */
6659:   PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6660:   PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6661:   PetscCall(VecDestroy(&mpiaij->lvec));
6662:   PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));

6664:   mpiaij->coo_n   = coo_n;
6665:   mpiaij->coo_sf  = sf2;
6666:   mpiaij->sendlen = nleaves;
6667:   mpiaij->recvlen = nroots;

6669:   mpiaij->Annz = Annz;
6670:   mpiaij->Bnnz = Bnnz;

6672:   mpiaij->Annz2 = Annz2;
6673:   mpiaij->Bnnz2 = Bnnz2;

6675:   mpiaij->Atot1 = Atot1;
6676:   mpiaij->Atot2 = Atot2;
6677:   mpiaij->Btot1 = Btot1;
6678:   mpiaij->Btot2 = Btot2;

6680:   mpiaij->Ajmap1 = Ajmap1;
6681:   mpiaij->Aperm1 = Aperm1;

6683:   mpiaij->Bjmap1 = Bjmap1;
6684:   mpiaij->Bperm1 = Bperm1;

6686:   mpiaij->Aimap2 = Aimap2;
6687:   mpiaij->Ajmap2 = Ajmap2;
6688:   mpiaij->Aperm2 = Aperm2;

6690:   mpiaij->Bimap2 = Bimap2;
6691:   mpiaij->Bjmap2 = Bjmap2;
6692:   mpiaij->Bperm2 = Bperm2;

6694:   mpiaij->Cperm1 = Cperm1;

6696:   /* Allocate in preallocation. If not used, it has zero cost on host */
6697:   PetscCall(PetscMalloc2(mpiaij->sendlen, &mpiaij->sendbuf, mpiaij->recvlen, &mpiaij->recvbuf));
6698:   PetscFunctionReturn(PETSC_SUCCESS);
6699: }

6701: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6702: {
6703:   Mat_MPIAIJ       *mpiaij = (Mat_MPIAIJ *)mat->data;
6704:   Mat               A = mpiaij->A, B = mpiaij->B;
6705:   PetscCount        Annz = mpiaij->Annz, Annz2 = mpiaij->Annz2, Bnnz = mpiaij->Bnnz, Bnnz2 = mpiaij->Bnnz2;
6706:   PetscScalar      *Aa, *Ba;
6707:   PetscScalar      *sendbuf = mpiaij->sendbuf;
6708:   PetscScalar      *recvbuf = mpiaij->recvbuf;
6709:   const PetscCount *Ajmap1 = mpiaij->Ajmap1, *Ajmap2 = mpiaij->Ajmap2, *Aimap2 = mpiaij->Aimap2;
6710:   const PetscCount *Bjmap1 = mpiaij->Bjmap1, *Bjmap2 = mpiaij->Bjmap2, *Bimap2 = mpiaij->Bimap2;
6711:   const PetscCount *Aperm1 = mpiaij->Aperm1, *Aperm2 = mpiaij->Aperm2, *Bperm1 = mpiaij->Bperm1, *Bperm2 = mpiaij->Bperm2;
6712:   const PetscCount *Cperm1 = mpiaij->Cperm1;

6714:   PetscFunctionBegin;
6715:   PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6716:   PetscCall(MatSeqAIJGetArray(B, &Ba));

6718:   /* Pack entries to be sent to remote */
6719:   for (PetscCount i = 0; i < mpiaij->sendlen; i++) sendbuf[i] = v[Cperm1[i]];

6721:   /* Send remote entries to their owner and overlap the communication with local computation */
6722:   PetscCall(PetscSFReduceWithMemTypeBegin(mpiaij->coo_sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6723:   /* Add local entries to A and B */
6724:   for (PetscCount i = 0; i < Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6725:     PetscScalar sum = 0.0;                /* Do partial summation first to improve numerical stability */
6726:     for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6727:     Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6728:   }
6729:   for (PetscCount i = 0; i < Bnnz; i++) {
6730:     PetscScalar sum = 0.0;
6731:     for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6732:     Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6733:   }
6734:   PetscCall(PetscSFReduceEnd(mpiaij->coo_sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));

6736:   /* Add received remote entries to A and B */
6737:   for (PetscCount i = 0; i < Annz2; i++) {
6738:     for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6739:   }
6740:   for (PetscCount i = 0; i < Bnnz2; i++) {
6741:     for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6742:   }
6743:   PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6744:   PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6745:   PetscFunctionReturn(PETSC_SUCCESS);
6746: }

6748: /*MC
6749:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

6751:    Options Database Keys:
6752: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`

6754:    Level: beginner

6756:    Notes:
6757:    `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6758:     in this case the values associated with the rows and columns one passes in are set to zero
6759:     in the matrix

6761:     `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6762:     space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored

6764: .seealso: [](chapter_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6765: M*/
6766: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6767: {
6768:   Mat_MPIAIJ *b;
6769:   PetscMPIInt size;

6771:   PetscFunctionBegin;
6772:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));

6774:   PetscCall(PetscNew(&b));
6775:   B->data = (void *)b;
6776:   PetscCall(PetscMemcpy(B->ops, &MatOps_Values, sizeof(struct _MatOps)));
6777:   B->assembled  = PETSC_FALSE;
6778:   B->insertmode = NOT_SET_VALUES;
6779:   b->size       = size;

6781:   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));

6783:   /* build cache for off array entries formed */
6784:   PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));

6786:   b->donotstash  = PETSC_FALSE;
6787:   b->colmap      = NULL;
6788:   b->garray      = NULL;
6789:   b->roworiented = PETSC_TRUE;

6791:   /* stuff used for matrix vector multiply */
6792:   b->lvec  = NULL;
6793:   b->Mvctx = NULL;

6795:   /* stuff for MatGetRow() */
6796:   b->rowindices   = NULL;
6797:   b->rowvalues    = NULL;
6798:   b->getrowactive = PETSC_FALSE;

6800:   /* flexible pointer used in CUSPARSE classes */
6801:   b->spptr = NULL;

6803:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6804:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6805:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6806:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6807:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6808:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6809:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6810:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6811:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6812:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6813: #if defined(PETSC_HAVE_CUDA)
6814:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6815: #endif
6816: #if defined(PETSC_HAVE_HIP)
6817:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6818: #endif
6819: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6820:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6821: #endif
6822: #if defined(PETSC_HAVE_MKL_SPARSE)
6823:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6824: #endif
6825:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6826:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6827:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6828:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6829: #if defined(PETSC_HAVE_ELEMENTAL)
6830:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6831: #endif
6832: #if defined(PETSC_HAVE_SCALAPACK)
6833:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6834: #endif
6835:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6836:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6837: #if defined(PETSC_HAVE_HYPRE)
6838:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6839:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6840: #endif
6841:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6842:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6843:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6844:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6845:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6846:   PetscFunctionReturn(PETSC_SUCCESS);
6847: }

6849: /*@C
6850:      MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6851:          and "off-diagonal" part of the matrix in CSR format.

6853:    Collective

6855:    Input Parameters:
6856: +  comm - MPI communicator
6857: .  m - number of local rows (Cannot be `PETSC_DECIDE`)
6858: .  n - This value should be the same as the local size used in creating the
6859:        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
6860:        calculated if `N` is given) For square matrices `n` is almost always `m`.
6861: .  M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6862: .  N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6863: .   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
6864: .   j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6865: .   a - matrix values
6866: .   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
6867: .   oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6868: -   oa - matrix values

6870:    Output Parameter:
6871: .   mat - the matrix

6873:    Level: advanced

6875:    Notes:
6876:        The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6877:        must free the arrays once the matrix has been destroyed and not before.

6879:        The `i` and `j` indices are 0 based

6881:        See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix

6883:        This sets local rows and cannot be used to set off-processor values.

6885:        Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6886:        legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6887:        not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6888:        the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6889:        keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6890:        communication if it is known that only local entries will be set.

6892: .seealso: [](chapter_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6893:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6894: @*/
6895: 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)
6896: {
6897:   Mat_MPIAIJ *maij;

6899:   PetscFunctionBegin;
6900:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6901:   PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6902:   PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6903:   PetscCall(MatCreate(comm, mat));
6904:   PetscCall(MatSetSizes(*mat, m, n, M, N));
6905:   PetscCall(MatSetType(*mat, MATMPIAIJ));
6906:   maij = (Mat_MPIAIJ *)(*mat)->data;

6908:   (*mat)->preallocated = PETSC_TRUE;

6910:   PetscCall(PetscLayoutSetUp((*mat)->rmap));
6911:   PetscCall(PetscLayoutSetUp((*mat)->cmap));

6913:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6914:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));

6916:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6917:   PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
6918:   PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
6919:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
6920:   PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
6921:   PetscFunctionReturn(PETSC_SUCCESS);
6922: }

6924: typedef struct {
6925:   Mat       *mp;    /* intermediate products */
6926:   PetscBool *mptmp; /* is the intermediate product temporary ? */
6927:   PetscInt   cp;    /* number of intermediate products */

6929:   /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
6930:   PetscInt    *startsj_s, *startsj_r;
6931:   PetscScalar *bufa;
6932:   Mat          P_oth;

6934:   /* may take advantage of merging product->B */
6935:   Mat Bloc; /* B-local by merging diag and off-diag */

6937:   /* cusparse does not have support to split between symbolic and numeric phases.
6938:      When api_user is true, we don't need to update the numerical values
6939:      of the temporary storage */
6940:   PetscBool reusesym;

6942:   /* support for COO values insertion */
6943:   PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
6944:   PetscInt   **own;           /* own[i] points to address of on-process COO indices for Mat mp[i] */
6945:   PetscInt   **off;           /* off[i] points to address of off-process COO indices for Mat mp[i] */
6946:   PetscBool    hasoffproc;    /* if true, have off-process values insertion (i.e. AtB or PtAP) */
6947:   PetscSF      sf;            /* used for non-local values insertion and memory malloc */
6948:   PetscMemType mtype;

6950:   /* customization */
6951:   PetscBool abmerge;
6952:   PetscBool P_oth_bind;
6953: } MatMatMPIAIJBACKEND;

6955: PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
6956: {
6957:   MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
6958:   PetscInt             i;

6960:   PetscFunctionBegin;
6961:   PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
6962:   PetscCall(PetscFree(mmdata->bufa));
6963:   PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
6964:   PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
6965:   PetscCall(MatDestroy(&mmdata->P_oth));
6966:   PetscCall(MatDestroy(&mmdata->Bloc));
6967:   PetscCall(PetscSFDestroy(&mmdata->sf));
6968:   for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
6969:   PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
6970:   PetscCall(PetscFree(mmdata->own[0]));
6971:   PetscCall(PetscFree(mmdata->own));
6972:   PetscCall(PetscFree(mmdata->off[0]));
6973:   PetscCall(PetscFree(mmdata->off));
6974:   PetscCall(PetscFree(mmdata));
6975:   PetscFunctionReturn(PETSC_SUCCESS);
6976: }

6978: /* Copy selected n entries with indices in idx[] of A to v[].
6979:    If idx is NULL, copy the whole data array of A to v[]
6980:  */
6981: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
6982: {
6983:   PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);

6985:   PetscFunctionBegin;
6986:   PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
6987:   if (f) {
6988:     PetscCall((*f)(A, n, idx, v));
6989:   } else {
6990:     const PetscScalar *vv;

6992:     PetscCall(MatSeqAIJGetArrayRead(A, &vv));
6993:     if (n && idx) {
6994:       PetscScalar    *w  = v;
6995:       const PetscInt *oi = idx;
6996:       PetscInt        j;

6998:       for (j = 0; j < n; j++) *w++ = vv[*oi++];
6999:     } else {
7000:       PetscCall(PetscArraycpy(v, vv, n));
7001:     }
7002:     PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7003:   }
7004:   PetscFunctionReturn(PETSC_SUCCESS);
7005: }

7007: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7008: {
7009:   MatMatMPIAIJBACKEND *mmdata;
7010:   PetscInt             i, n_d, n_o;

7012:   PetscFunctionBegin;
7013:   MatCheckProduct(C, 1);
7014:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7015:   mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7016:   if (!mmdata->reusesym) { /* update temporary matrices */
7017:     if (mmdata->P_oth) PetscCall(MatGetBrowsOfAoCols_MPIAIJ(C->product->A, C->product->B, MAT_REUSE_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7018:     if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7019:   }
7020:   mmdata->reusesym = PETSC_FALSE;

7022:   for (i = 0; i < mmdata->cp; i++) {
7023:     PetscCheck(mmdata->mp[i]->ops->productnumeric, PetscObjectComm((PetscObject)mmdata->mp[i]), PETSC_ERR_PLIB, "Missing numeric op for %s", MatProductTypes[mmdata->mp[i]->product->type]);
7024:     PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7025:   }
7026:   for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7027:     PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];

7029:     if (mmdata->mptmp[i]) continue;
7030:     if (noff) {
7031:       PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];

7033:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7034:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7035:       n_o += noff;
7036:       n_d += nown;
7037:     } else {
7038:       Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;

7040:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7041:       n_d += mm->nz;
7042:     }
7043:   }
7044:   if (mmdata->hasoffproc) { /* offprocess insertion */
7045:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7046:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7047:   }
7048:   PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7049:   PetscFunctionReturn(PETSC_SUCCESS);
7050: }

7052: /* Support for Pt * A, A * P, or Pt * A * P */
7053: #define MAX_NUMBER_INTERMEDIATE 4
7054: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7055: {
7056:   Mat_Product           *product = C->product;
7057:   Mat                    A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7058:   Mat_MPIAIJ            *a, *p;
7059:   MatMatMPIAIJBACKEND   *mmdata;
7060:   ISLocalToGlobalMapping P_oth_l2g = NULL;
7061:   IS                     glob      = NULL;
7062:   const char            *prefix;
7063:   char                   pprefix[256];
7064:   const PetscInt        *globidx, *P_oth_idx;
7065:   PetscInt               i, j, cp, m, n, M, N, *coo_i, *coo_j;
7066:   PetscCount             ncoo, ncoo_d, ncoo_o, ncoo_oown;
7067:   PetscInt               cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7068:                                                                                          /* type-0: consecutive, start from 0; type-1: consecutive with */
7069:                                                                                          /* a base offset; type-2: sparse with a local to global map table */
7070:   const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE];       /* col/row local to global map array (table) for type-2 map type */

7072:   MatProductType ptype;
7073:   PetscBool      mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7074:   PetscMPIInt    size;

7076:   PetscFunctionBegin;
7077:   MatCheckProduct(C, 1);
7078:   PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7079:   ptype = product->type;
7080:   if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7081:     ptype                                          = MATPRODUCT_AB;
7082:     product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7083:   }
7084:   switch (ptype) {
7085:   case MATPRODUCT_AB:
7086:     A          = product->A;
7087:     P          = product->B;
7088:     m          = A->rmap->n;
7089:     n          = P->cmap->n;
7090:     M          = A->rmap->N;
7091:     N          = P->cmap->N;
7092:     hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7093:     break;
7094:   case MATPRODUCT_AtB:
7095:     P          = product->A;
7096:     A          = product->B;
7097:     m          = P->cmap->n;
7098:     n          = A->cmap->n;
7099:     M          = P->cmap->N;
7100:     N          = A->cmap->N;
7101:     hasoffproc = PETSC_TRUE;
7102:     break;
7103:   case MATPRODUCT_PtAP:
7104:     A          = product->A;
7105:     P          = product->B;
7106:     m          = P->cmap->n;
7107:     n          = P->cmap->n;
7108:     M          = P->cmap->N;
7109:     N          = P->cmap->N;
7110:     hasoffproc = PETSC_TRUE;
7111:     break;
7112:   default:
7113:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7114:   }
7115:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7116:   if (size == 1) hasoffproc = PETSC_FALSE;

7118:   /* defaults */
7119:   for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7120:     mp[i]    = NULL;
7121:     mptmp[i] = PETSC_FALSE;
7122:     rmapt[i] = -1;
7123:     cmapt[i] = -1;
7124:     rmapa[i] = NULL;
7125:     cmapa[i] = NULL;
7126:   }

7128:   /* customization */
7129:   PetscCall(PetscNew(&mmdata));
7130:   mmdata->reusesym = product->api_user;
7131:   if (ptype == MATPRODUCT_AB) {
7132:     if (product->api_user) {
7133:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7134:       PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7135:       PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7136:       PetscOptionsEnd();
7137:     } else {
7138:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7139:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7140:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7141:       PetscOptionsEnd();
7142:     }
7143:   } else if (ptype == MATPRODUCT_PtAP) {
7144:     if (product->api_user) {
7145:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7146:       PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7147:       PetscOptionsEnd();
7148:     } else {
7149:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7150:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7151:       PetscOptionsEnd();
7152:     }
7153:   }
7154:   a = (Mat_MPIAIJ *)A->data;
7155:   p = (Mat_MPIAIJ *)P->data;
7156:   PetscCall(MatSetSizes(C, m, n, M, N));
7157:   PetscCall(PetscLayoutSetUp(C->rmap));
7158:   PetscCall(PetscLayoutSetUp(C->cmap));
7159:   PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7160:   PetscCall(MatGetOptionsPrefix(C, &prefix));

7162:   cp = 0;
7163:   switch (ptype) {
7164:   case MATPRODUCT_AB: /* A * P */
7165:     PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));

7167:     /* A_diag * P_local (merged or not) */
7168:     if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7169:       /* P is product->B */
7170:       PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7171:       PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7172:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7173:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7174:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7175:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7176:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7177:       mp[cp]->product->api_user = product->api_user;
7178:       PetscCall(MatProductSetFromOptions(mp[cp]));
7179:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7180:       PetscCall(ISGetIndices(glob, &globidx));
7181:       rmapt[cp] = 1;
7182:       cmapt[cp] = 2;
7183:       cmapa[cp] = globidx;
7184:       mptmp[cp] = PETSC_FALSE;
7185:       cp++;
7186:     } else { /* A_diag * P_diag and A_diag * P_off */
7187:       PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7188:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7189:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7190:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7191:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7192:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7193:       mp[cp]->product->api_user = product->api_user;
7194:       PetscCall(MatProductSetFromOptions(mp[cp]));
7195:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7196:       rmapt[cp] = 1;
7197:       cmapt[cp] = 1;
7198:       mptmp[cp] = PETSC_FALSE;
7199:       cp++;
7200:       PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7201:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7202:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7203:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7204:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7205:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7206:       mp[cp]->product->api_user = product->api_user;
7207:       PetscCall(MatProductSetFromOptions(mp[cp]));
7208:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7209:       rmapt[cp] = 1;
7210:       cmapt[cp] = 2;
7211:       cmapa[cp] = p->garray;
7212:       mptmp[cp] = PETSC_FALSE;
7213:       cp++;
7214:     }

7216:     /* A_off * P_other */
7217:     if (mmdata->P_oth) {
7218:       PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7219:       PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7220:       PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)(a->B))->type_name));
7221:       PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7222:       PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7223:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7224:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7225:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7226:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7227:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7228:       mp[cp]->product->api_user = product->api_user;
7229:       PetscCall(MatProductSetFromOptions(mp[cp]));
7230:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7231:       rmapt[cp] = 1;
7232:       cmapt[cp] = 2;
7233:       cmapa[cp] = P_oth_idx;
7234:       mptmp[cp] = PETSC_FALSE;
7235:       cp++;
7236:     }
7237:     break;

7239:   case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7240:     /* A is product->B */
7241:     PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7242:     if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7243:       PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7244:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7245:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7246:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7247:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7248:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7249:       mp[cp]->product->api_user = product->api_user;
7250:       PetscCall(MatProductSetFromOptions(mp[cp]));
7251:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7252:       PetscCall(ISGetIndices(glob, &globidx));
7253:       rmapt[cp] = 2;
7254:       rmapa[cp] = globidx;
7255:       cmapt[cp] = 2;
7256:       cmapa[cp] = globidx;
7257:       mptmp[cp] = PETSC_FALSE;
7258:       cp++;
7259:     } else {
7260:       PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7261:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7262:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7263:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7264:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7265:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7266:       mp[cp]->product->api_user = product->api_user;
7267:       PetscCall(MatProductSetFromOptions(mp[cp]));
7268:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7269:       PetscCall(ISGetIndices(glob, &globidx));
7270:       rmapt[cp] = 1;
7271:       cmapt[cp] = 2;
7272:       cmapa[cp] = globidx;
7273:       mptmp[cp] = PETSC_FALSE;
7274:       cp++;
7275:       PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7276:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7277:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7278:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7279:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7280:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7281:       mp[cp]->product->api_user = product->api_user;
7282:       PetscCall(MatProductSetFromOptions(mp[cp]));
7283:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7284:       rmapt[cp] = 2;
7285:       rmapa[cp] = p->garray;
7286:       cmapt[cp] = 2;
7287:       cmapa[cp] = globidx;
7288:       mptmp[cp] = PETSC_FALSE;
7289:       cp++;
7290:     }
7291:     break;
7292:   case MATPRODUCT_PtAP:
7293:     PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7294:     /* P is product->B */
7295:     PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7296:     PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7297:     PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7298:     PetscCall(MatProductSetFill(mp[cp], product->fill));
7299:     PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7300:     PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7301:     PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7302:     mp[cp]->product->api_user = product->api_user;
7303:     PetscCall(MatProductSetFromOptions(mp[cp]));
7304:     PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7305:     PetscCall(ISGetIndices(glob, &globidx));
7306:     rmapt[cp] = 2;
7307:     rmapa[cp] = globidx;
7308:     cmapt[cp] = 2;
7309:     cmapa[cp] = globidx;
7310:     mptmp[cp] = PETSC_FALSE;
7311:     cp++;
7312:     if (mmdata->P_oth) {
7313:       PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7314:       PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7315:       PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)(a->B))->type_name));
7316:       PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7317:       PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7318:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7319:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7320:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7321:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7322:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7323:       mp[cp]->product->api_user = product->api_user;
7324:       PetscCall(MatProductSetFromOptions(mp[cp]));
7325:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7326:       mptmp[cp] = PETSC_TRUE;
7327:       cp++;
7328:       PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7329:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7330:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7331:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7332:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7333:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7334:       mp[cp]->product->api_user = product->api_user;
7335:       PetscCall(MatProductSetFromOptions(mp[cp]));
7336:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7337:       rmapt[cp] = 2;
7338:       rmapa[cp] = globidx;
7339:       cmapt[cp] = 2;
7340:       cmapa[cp] = P_oth_idx;
7341:       mptmp[cp] = PETSC_FALSE;
7342:       cp++;
7343:     }
7344:     break;
7345:   default:
7346:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7347:   }
7348:   /* sanity check */
7349:   if (size > 1)
7350:     for (i = 0; i < cp; i++) PetscCheck(rmapt[i] != 2 || hasoffproc, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected offproc map type for product %" PetscInt_FMT, i);

7352:   PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7353:   for (i = 0; i < cp; i++) {
7354:     mmdata->mp[i]    = mp[i];
7355:     mmdata->mptmp[i] = mptmp[i];
7356:   }
7357:   mmdata->cp             = cp;
7358:   C->product->data       = mmdata;
7359:   C->product->destroy    = MatDestroy_MatMatMPIAIJBACKEND;
7360:   C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;

7362:   /* memory type */
7363:   mmdata->mtype = PETSC_MEMTYPE_HOST;
7364:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7365:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7366:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7367:   if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7368:   else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7369:   else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;

7371:   /* prepare coo coordinates for values insertion */

7373:   /* count total nonzeros of those intermediate seqaij Mats
7374:     ncoo_d:    # of nonzeros of matrices that do not have offproc entries
7375:     ncoo_o:    # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7376:     ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7377:   */
7378:   for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7379:     Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7380:     if (mptmp[cp]) continue;
7381:     if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7382:       const PetscInt *rmap = rmapa[cp];
7383:       const PetscInt  mr   = mp[cp]->rmap->n;
7384:       const PetscInt  rs   = C->rmap->rstart;
7385:       const PetscInt  re   = C->rmap->rend;
7386:       const PetscInt *ii   = mm->i;
7387:       for (i = 0; i < mr; i++) {
7388:         const PetscInt gr = rmap[i];
7389:         const PetscInt nz = ii[i + 1] - ii[i];
7390:         if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7391:         else ncoo_oown += nz;                  /* this row is local */
7392:       }
7393:     } else ncoo_d += mm->nz;
7394:   }

7396:   /*
7397:     ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc

7399:     ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.

7401:     off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].

7403:     off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7404:     own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7405:     so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.

7407:     coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7408:     Ex. coo_i[]: the beginning part (of size ncoo_d + ncoo_oown) stores i of local nonzeros, and the remaining part stores i of nonzeros I will receive.
7409:   */
7410:   PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7411:   PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));

7413:   /* gather (i,j) of nonzeros inserted by remote procs */
7414:   if (hasoffproc) {
7415:     PetscSF  msf;
7416:     PetscInt ncoo2, *coo_i2, *coo_j2;

7418:     PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7419:     PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7420:     PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */

7422:     for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7423:       Mat_SeqAIJ *mm     = (Mat_SeqAIJ *)mp[cp]->data;
7424:       PetscInt   *idxoff = mmdata->off[cp];
7425:       PetscInt   *idxown = mmdata->own[cp];
7426:       if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7427:         const PetscInt *rmap = rmapa[cp];
7428:         const PetscInt *cmap = cmapa[cp];
7429:         const PetscInt *ii   = mm->i;
7430:         PetscInt       *coi  = coo_i + ncoo_o;
7431:         PetscInt       *coj  = coo_j + ncoo_o;
7432:         const PetscInt  mr   = mp[cp]->rmap->n;
7433:         const PetscInt  rs   = C->rmap->rstart;
7434:         const PetscInt  re   = C->rmap->rend;
7435:         const PetscInt  cs   = C->cmap->rstart;
7436:         for (i = 0; i < mr; i++) {
7437:           const PetscInt *jj = mm->j + ii[i];
7438:           const PetscInt  gr = rmap[i];
7439:           const PetscInt  nz = ii[i + 1] - ii[i];
7440:           if (gr < rs || gr >= re) { /* this is an offproc row */
7441:             for (j = ii[i]; j < ii[i + 1]; j++) {
7442:               *coi++    = gr;
7443:               *idxoff++ = j;
7444:             }
7445:             if (!cmapt[cp]) { /* already global */
7446:               for (j = 0; j < nz; j++) *coj++ = jj[j];
7447:             } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7448:               for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7449:             } else { /* offdiag */
7450:               for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7451:             }
7452:             ncoo_o += nz;
7453:           } else { /* this is a local row */
7454:             for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7455:           }
7456:         }
7457:       }
7458:       mmdata->off[cp + 1] = idxoff;
7459:       mmdata->own[cp + 1] = idxown;
7460:     }

7462:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7463:     PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7464:     PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7465:     PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7466:     ncoo = ncoo_d + ncoo_oown + ncoo2;
7467:     PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7468:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7469:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7470:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7471:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7472:     PetscCall(PetscFree2(coo_i, coo_j));
7473:     /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7474:     PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7475:     coo_i = coo_i2;
7476:     coo_j = coo_j2;
7477:   } else { /* no offproc values insertion */
7478:     ncoo = ncoo_d;
7479:     PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));

7481:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7482:     PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7483:     PetscCall(PetscSFSetUp(mmdata->sf));
7484:   }
7485:   mmdata->hasoffproc = hasoffproc;

7487:   /* gather (i,j) of nonzeros inserted locally */
7488:   for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7489:     Mat_SeqAIJ     *mm   = (Mat_SeqAIJ *)mp[cp]->data;
7490:     PetscInt       *coi  = coo_i + ncoo_d;
7491:     PetscInt       *coj  = coo_j + ncoo_d;
7492:     const PetscInt *jj   = mm->j;
7493:     const PetscInt *ii   = mm->i;
7494:     const PetscInt *cmap = cmapa[cp];
7495:     const PetscInt *rmap = rmapa[cp];
7496:     const PetscInt  mr   = mp[cp]->rmap->n;
7497:     const PetscInt  rs   = C->rmap->rstart;
7498:     const PetscInt  re   = C->rmap->rend;
7499:     const PetscInt  cs   = C->cmap->rstart;

7501:     if (mptmp[cp]) continue;
7502:     if (rmapt[cp] == 1) { /* consecutive rows */
7503:       /* fill coo_i */
7504:       for (i = 0; i < mr; i++) {
7505:         const PetscInt gr = i + rs;
7506:         for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7507:       }
7508:       /* fill coo_j */
7509:       if (!cmapt[cp]) { /* type-0, already global */
7510:         PetscCall(PetscArraycpy(coj, jj, mm->nz));
7511:       } else if (cmapt[cp] == 1) {                        /* type-1, local to global for consecutive columns of C */
7512:         for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7513:       } else {                                            /* type-2, local to global for sparse columns */
7514:         for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7515:       }
7516:       ncoo_d += mm->nz;
7517:     } else if (rmapt[cp] == 2) { /* sparse rows */
7518:       for (i = 0; i < mr; i++) {
7519:         const PetscInt *jj = mm->j + ii[i];
7520:         const PetscInt  gr = rmap[i];
7521:         const PetscInt  nz = ii[i + 1] - ii[i];
7522:         if (gr >= rs && gr < re) { /* local rows */
7523:           for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7524:           if (!cmapt[cp]) { /* type-0, already global */
7525:             for (j = 0; j < nz; j++) *coj++ = jj[j];
7526:           } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7527:             for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7528:           } else { /* type-2, local to global for sparse columns */
7529:             for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7530:           }
7531:           ncoo_d += nz;
7532:         }
7533:       }
7534:     }
7535:   }
7536:   if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7537:   PetscCall(ISDestroy(&glob));
7538:   if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7539:   PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7540:   /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7541:   PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));

7543:   /* preallocate with COO data */
7544:   PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7545:   PetscCall(PetscFree2(coo_i, coo_j));
7546:   PetscFunctionReturn(PETSC_SUCCESS);
7547: }

7549: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7550: {
7551:   Mat_Product *product = mat->product;
7552: #if defined(PETSC_HAVE_DEVICE)
7553:   PetscBool match  = PETSC_FALSE;
7554:   PetscBool usecpu = PETSC_FALSE;
7555: #else
7556:   PetscBool match = PETSC_TRUE;
7557: #endif

7559:   PetscFunctionBegin;
7560:   MatCheckProduct(mat, 1);
7561: #if defined(PETSC_HAVE_DEVICE)
7562:   if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7563:   if (match) { /* we can always fallback to the CPU if requested */
7564:     switch (product->type) {
7565:     case MATPRODUCT_AB:
7566:       if (product->api_user) {
7567:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7568:         PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7569:         PetscOptionsEnd();
7570:       } else {
7571:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7572:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7573:         PetscOptionsEnd();
7574:       }
7575:       break;
7576:     case MATPRODUCT_AtB:
7577:       if (product->api_user) {
7578:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7579:         PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7580:         PetscOptionsEnd();
7581:       } else {
7582:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7583:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7584:         PetscOptionsEnd();
7585:       }
7586:       break;
7587:     case MATPRODUCT_PtAP:
7588:       if (product->api_user) {
7589:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7590:         PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7591:         PetscOptionsEnd();
7592:       } else {
7593:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7594:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7595:         PetscOptionsEnd();
7596:       }
7597:       break;
7598:     default:
7599:       break;
7600:     }
7601:     match = (PetscBool)!usecpu;
7602:   }
7603: #endif
7604:   if (match) {
7605:     switch (product->type) {
7606:     case MATPRODUCT_AB:
7607:     case MATPRODUCT_AtB:
7608:     case MATPRODUCT_PtAP:
7609:       mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7610:       break;
7611:     default:
7612:       break;
7613:     }
7614:   }
7615:   /* fallback to MPIAIJ ops */
7616:   if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7617:   PetscFunctionReturn(PETSC_SUCCESS);
7618: }

7620: /*
7621:    Produces a set of block column indices of the matrix row, one for each block represented in the original row

7623:    n - the number of block indices in cc[]
7624:    cc - the block indices (must be large enough to contain the indices)
7625: */
7626: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7627: {
7628:   PetscInt        cnt = -1, nidx, j;
7629:   const PetscInt *idx;

7631:   PetscFunctionBegin;
7632:   PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7633:   if (nidx) {
7634:     cnt     = 0;
7635:     cc[cnt] = idx[0] / bs;
7636:     for (j = 1; j < nidx; j++) {
7637:       if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7638:     }
7639:   }
7640:   PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7641:   *n = cnt + 1;
7642:   PetscFunctionReturn(PETSC_SUCCESS);
7643: }

7645: /*
7646:     Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows

7648:     ncollapsed - the number of block indices
7649:     collapsed - the block indices (must be large enough to contain the indices)
7650: */
7651: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7652: {
7653:   PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;

7655:   PetscFunctionBegin;
7656:   PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7657:   for (i = start + 1; i < start + bs; i++) {
7658:     PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7659:     PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7660:     cprevtmp = cprev;
7661:     cprev    = merged;
7662:     merged   = cprevtmp;
7663:   }
7664:   *ncollapsed = nprev;
7665:   if (collapsed) *collapsed = cprev;
7666:   PetscFunctionReturn(PETSC_SUCCESS);
7667: }

7669: /*
7670:    This will eventually be folded into MatCreateGraph_AIJ() for optimal performance
7671: */
7672: static PetscErrorCode MatFilter_AIJ(Mat Gmat, PetscReal vfilter, Mat *filteredG)
7673: {
7674:   PetscInt           Istart, Iend, ncols, nnz0, nnz1, NN, MM, nloc;
7675:   Mat                tGmat;
7676:   MPI_Comm           comm;
7677:   const PetscScalar *vals;
7678:   const PetscInt    *idx;
7679:   PetscInt          *d_nnz, *o_nnz, kk, *garray = NULL, *AJ, maxcols = 0;
7680:   MatScalar         *AA; // this is checked in graph
7681:   PetscBool          isseqaij;
7682:   Mat                a, b, c;
7683:   MatType            jtype;

7685:   PetscFunctionBegin;
7686:   PetscCall(PetscObjectGetComm((PetscObject)Gmat, &comm));
7687:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)Gmat, MATSEQAIJ, &isseqaij));
7688:   PetscCall(MatGetType(Gmat, &jtype));
7689:   PetscCall(MatCreate(comm, &tGmat));
7690:   PetscCall(MatSetType(tGmat, jtype));

7692:   /* TODO GPU: this can be called when filter = 0 -> Probably provide MatAIJThresholdCompress that compresses the entries below a threshold?
7693:                Also, if the matrix is symmetric, can we skip this
7694:                operation? It can be very expensive on large matrices. */

7696:   // global sizes
7697:   PetscCall(MatGetSize(Gmat, &MM, &NN));
7698:   PetscCall(MatGetOwnershipRange(Gmat, &Istart, &Iend));
7699:   nloc = Iend - Istart;
7700:   PetscCall(PetscMalloc2(nloc, &d_nnz, nloc, &o_nnz));
7701:   if (isseqaij) {
7702:     a = Gmat;
7703:     b = NULL;
7704:   } else {
7705:     Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7706:     a             = d->A;
7707:     b             = d->B;
7708:     garray        = d->garray;
7709:   }
7710:   /* Determine upper bound on non-zeros needed in new filtered matrix */
7711:   for (PetscInt row = 0; row < nloc; row++) {
7712:     PetscCall(MatGetRow(a, row, &ncols, NULL, NULL));
7713:     d_nnz[row] = ncols;
7714:     if (ncols > maxcols) maxcols = ncols;
7715:     PetscCall(MatRestoreRow(a, row, &ncols, NULL, NULL));
7716:   }
7717:   if (b) {
7718:     for (PetscInt row = 0; row < nloc; row++) {
7719:       PetscCall(MatGetRow(b, row, &ncols, NULL, NULL));
7720:       o_nnz[row] = ncols;
7721:       if (ncols > maxcols) maxcols = ncols;
7722:       PetscCall(MatRestoreRow(b, row, &ncols, NULL, NULL));
7723:     }
7724:   }
7725:   PetscCall(MatSetSizes(tGmat, nloc, nloc, MM, MM));
7726:   PetscCall(MatSetBlockSizes(tGmat, 1, 1));
7727:   PetscCall(MatSeqAIJSetPreallocation(tGmat, 0, d_nnz));
7728:   PetscCall(MatMPIAIJSetPreallocation(tGmat, 0, d_nnz, 0, o_nnz));
7729:   PetscCall(MatSetOption(tGmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7730:   PetscCall(PetscFree2(d_nnz, o_nnz));
7731:   //
7732:   PetscCall(PetscMalloc2(maxcols, &AA, maxcols, &AJ));
7733:   nnz0 = nnz1 = 0;
7734:   for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7735:     for (PetscInt row = 0, grow = Istart, ncol_row, jj; row < nloc; row++, grow++) {
7736:       PetscCall(MatGetRow(c, row, &ncols, &idx, &vals));
7737:       for (ncol_row = jj = 0; jj < ncols; jj++, nnz0++) {
7738:         PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7739:         if (PetscRealPart(sv) > vfilter) {
7740:           nnz1++;
7741:           PetscInt cid = idx[jj] + Istart; //diag
7742:           if (c != a) cid = garray[idx[jj]];
7743:           AA[ncol_row] = vals[jj];
7744:           AJ[ncol_row] = cid;
7745:           ncol_row++;
7746:         }
7747:       }
7748:       PetscCall(MatRestoreRow(c, row, &ncols, &idx, &vals));
7749:       PetscCall(MatSetValues(tGmat, 1, &grow, ncol_row, AJ, AA, INSERT_VALUES));
7750:     }
7751:   }
7752:   PetscCall(PetscFree2(AA, AJ));
7753:   PetscCall(MatAssemblyBegin(tGmat, MAT_FINAL_ASSEMBLY));
7754:   PetscCall(MatAssemblyEnd(tGmat, MAT_FINAL_ASSEMBLY));
7755:   PetscCall(MatPropagateSymmetryOptions(Gmat, tGmat)); /* Normal Mat options are not relevant ? */

7757:   PetscCall(PetscInfo(tGmat, "\t %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%" PetscInt_FMT ", max row size %d)\n", (!nnz0) ? 1. : 100. * (double)nnz1 / (double)nnz0, (double)vfilter, (!nloc) ? 1. : (double)nnz0 / (double)nloc, MM, (int)maxcols));

7759:   *filteredG = tGmat;
7760:   PetscCall(MatViewFromOptions(tGmat, NULL, "-mat_filter_graph_view"));
7761:   PetscFunctionReturn(PETSC_SUCCESS);
7762: }

7764: /*
7765:  MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix

7767:  Input Parameter:
7768:  . Amat - matrix
7769:  - symmetrize - make the result symmetric
7770:  + scale - scale with diagonal

7772:  Output Parameter:
7773:  . a_Gmat - output scalar graph >= 0

7775: */
7776: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, Mat *a_Gmat)
7777: {
7778:   PetscInt  Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7779:   MPI_Comm  comm;
7780:   Mat       Gmat;
7781:   PetscBool ismpiaij, isseqaij;
7782:   Mat       a, b, c;
7783:   MatType   jtype;

7785:   PetscFunctionBegin;
7786:   PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7787:   PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7788:   PetscCall(MatGetSize(Amat, &MM, &NN));
7789:   PetscCall(MatGetBlockSize(Amat, &bs));
7790:   nloc = (Iend - Istart) / bs;

7792:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7793:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7794:   PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");

7796:   /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7797:   /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7798:      implementation */
7799:   if (bs > 1) {
7800:     PetscCall(MatGetType(Amat, &jtype));
7801:     PetscCall(MatCreate(comm, &Gmat));
7802:     PetscCall(MatSetType(Gmat, jtype));
7803:     PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7804:     PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7805:     if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7806:       PetscInt  *d_nnz, *o_nnz;
7807:       MatScalar *aa, val, *AA;
7808:       PetscInt  *aj, *ai, *AJ, nc, nmax = 0;
7809:       if (isseqaij) {
7810:         a = Amat;
7811:         b = NULL;
7812:       } else {
7813:         Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7814:         a             = d->A;
7815:         b             = d->B;
7816:       }
7817:       PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7818:       PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7819:       for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7820:         PetscInt       *nnz = (c == a) ? d_nnz : o_nnz;
7821:         const PetscInt *cols1, *cols2;
7822:         for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7823:           PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7824:           nnz[brow / bs] = nc2 / bs;
7825:           if (nc2 % bs) ok = 0;
7826:           if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7827:           for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7828:             PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7829:             if (nc1 != nc2) ok = 0;
7830:             else {
7831:               for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7832:                 if (cols1[jj] != cols2[jj]) ok = 0;
7833:                 if (cols1[jj] % bs != jj % bs) ok = 0;
7834:               }
7835:             }
7836:             PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7837:           }
7838:           PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7839:           if (!ok) {
7840:             PetscCall(PetscFree2(d_nnz, o_nnz));
7841:             PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7842:             goto old_bs;
7843:           }
7844:         }
7845:       }
7846:       PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7847:       PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7848:       PetscCall(PetscFree2(d_nnz, o_nnz));
7849:       PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7850:       // diag
7851:       for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7852:         Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7853:         ai               = aseq->i;
7854:         n                = ai[brow + 1] - ai[brow];
7855:         aj               = aseq->j + ai[brow];
7856:         for (int k = 0; k < n; k += bs) {        // block columns
7857:           AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7858:           val        = 0;
7859:           for (int ii = 0; ii < bs; ii++) { // rows in block
7860:             aa = aseq->a + ai[brow + ii] + k;
7861:             for (int jj = 0; jj < bs; jj++) {         // columns in block
7862:               val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7863:             }
7864:           }
7865:           PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7866:           AA[k / bs] = val;
7867:         }
7868:         grow = Istart / bs + brow / bs;
7869:         PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, INSERT_VALUES));
7870:       }
7871:       // off-diag
7872:       if (ismpiaij) {
7873:         Mat_MPIAIJ        *aij = (Mat_MPIAIJ *)Amat->data;
7874:         const PetscScalar *vals;
7875:         const PetscInt    *cols, *garray = aij->garray;
7876:         PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7877:         for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7878:           PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7879:           for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7880:             PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7881:             AA[k / bs] = 0;
7882:             AJ[cidx]   = garray[cols[k]] / bs;
7883:           }
7884:           nc = ncols / bs;
7885:           PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7886:           for (int ii = 0; ii < bs; ii++) { // rows in block
7887:             PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7888:             for (int k = 0; k < ncols; k += bs) {
7889:               for (int jj = 0; jj < bs; jj++) { // cols in block
7890:                 PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7891:                 AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7892:               }
7893:             }
7894:             PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7895:           }
7896:           grow = Istart / bs + brow / bs;
7897:           PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, INSERT_VALUES));
7898:         }
7899:       }
7900:       PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7901:       PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7902:       PetscCall(PetscFree2(AA, AJ));
7903:     } else {
7904:       const PetscScalar *vals;
7905:       const PetscInt    *idx;
7906:       PetscInt          *d_nnz, *o_nnz, *w0, *w1, *w2;
7907:     old_bs:
7908:       /*
7909:        Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7910:        */
7911:       PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7912:       PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7913:       if (isseqaij) {
7914:         PetscInt max_d_nnz;
7915:         /*
7916:          Determine exact preallocation count for (sequential) scalar matrix
7917:          */
7918:         PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7919:         max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7920:         PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7921:         for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7922:         PetscCall(PetscFree3(w0, w1, w2));
7923:       } else if (ismpiaij) {
7924:         Mat             Daij, Oaij;
7925:         const PetscInt *garray;
7926:         PetscInt        max_d_nnz;
7927:         PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7928:         /*
7929:          Determine exact preallocation count for diagonal block portion of scalar matrix
7930:          */
7931:         PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7932:         max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7933:         PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7934:         for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7935:         PetscCall(PetscFree3(w0, w1, w2));
7936:         /*
7937:          Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7938:          */
7939:         for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7940:           o_nnz[jj] = 0;
7941:           for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7942:             PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7943:             o_nnz[jj] += ncols;
7944:             PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7945:           }
7946:           if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7947:         }
7948:       } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7949:       /* get scalar copy (norms) of matrix */
7950:       PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7951:       PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7952:       PetscCall(PetscFree2(d_nnz, o_nnz));
7953:       for (Ii = Istart; Ii < Iend; Ii++) {
7954:         PetscInt dest_row = Ii / bs;
7955:         PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7956:         for (jj = 0; jj < ncols; jj++) {
7957:           PetscInt    dest_col = idx[jj] / bs;
7958:           PetscScalar sv       = PetscAbs(PetscRealPart(vals[jj]));
7959:           PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7960:         }
7961:         PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7962:       }
7963:       PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7964:       PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7965:     }
7966:   } else {
7967:     if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7968:     else {
7969:       Gmat = Amat;
7970:       PetscCall(PetscObjectReference((PetscObject)Gmat));
7971:     }
7972:     if (isseqaij) {
7973:       a = Gmat;
7974:       b = NULL;
7975:     } else {
7976:       Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7977:       a             = d->A;
7978:       b             = d->B;
7979:     }
7980:     if (filter >= 0 || scale) {
7981:       /* take absolute value of each entry */
7982:       for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7983:         MatInfo      info;
7984:         PetscScalar *avals;
7985:         PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
7986:         PetscCall(MatSeqAIJGetArray(c, &avals));
7987:         for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
7988:         PetscCall(MatSeqAIJRestoreArray(c, &avals));
7989:       }
7990:     }
7991:   }
7992:   if (symmetrize) {
7993:     PetscBool isset, issym;
7994:     PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
7995:     if (!isset || !issym) {
7996:       Mat matTrans;
7997:       PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
7998:       PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
7999:       PetscCall(MatDestroy(&matTrans));
8000:     }
8001:     PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8002:   } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8003:   if (scale) {
8004:     /* scale c for all diagonal values = 1 or -1 */
8005:     Vec diag;
8006:     PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8007:     PetscCall(MatGetDiagonal(Gmat, diag));
8008:     PetscCall(VecReciprocal(diag));
8009:     PetscCall(VecSqrtAbs(diag));
8010:     PetscCall(MatDiagonalScale(Gmat, diag, diag));
8011:     PetscCall(VecDestroy(&diag));
8012:   }
8013:   PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));

8015:   if (filter >= 0) {
8016:     Mat Fmat = NULL; /* some silly compiler needs this */

8018:     PetscCall(MatFilter_AIJ(Gmat, filter, &Fmat));
8019:     PetscCall(MatDestroy(&Gmat));
8020:     Gmat = Fmat;
8021:   }
8022:   *a_Gmat = Gmat;
8023:   PetscFunctionReturn(PETSC_SUCCESS);
8024: }

8026: /*
8027:     Special version for direct calls from Fortran
8028: */
8029: #include <petsc/private/fortranimpl.h>

8031: /* Change these macros so can be used in void function */
8032: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8033: #undef PetscCall
8034: #define PetscCall(...) \
8035:   do { \
8036:     PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8037:     if (PetscUnlikely(ierr_msv_mpiaij)) { \
8038:       *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8039:       return; \
8040:     } \
8041:   } while (0)

8043: #undef SETERRQ
8044: #define SETERRQ(comm, ierr, ...) \
8045:   do { \
8046:     *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8047:     return; \
8048:   } while (0)

8050: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8051:   #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8052: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8053:   #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8054: #else
8055: #endif
8056: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8057: {
8058:   Mat         mat = *mmat;
8059:   PetscInt    m = *mm, n = *mn;
8060:   InsertMode  addv = *maddv;
8061:   Mat_MPIAIJ *aij  = (Mat_MPIAIJ *)mat->data;
8062:   PetscScalar value;

8064:   MatCheckPreallocated(mat, 1);
8065:   if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8066:   else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8067:   {
8068:     PetscInt  i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8069:     PetscInt  cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8070:     PetscBool roworiented = aij->roworiented;

8072:     /* Some Variables required in the macro */
8073:     Mat         A     = aij->A;
8074:     Mat_SeqAIJ *a     = (Mat_SeqAIJ *)A->data;
8075:     PetscInt   *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8076:     MatScalar  *aa;
8077:     PetscBool   ignorezeroentries = (((a->ignorezeroentries) && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8078:     Mat         B                 = aij->B;
8079:     Mat_SeqAIJ *b                 = (Mat_SeqAIJ *)B->data;
8080:     PetscInt   *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8081:     MatScalar  *ba;
8082:     /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8083:      * cannot use "#if defined" inside a macro. */
8084:     PETSC_UNUSED PetscBool inserted = PETSC_FALSE;

8086:     PetscInt  *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8087:     PetscInt   nonew = a->nonew;
8088:     MatScalar *ap1, *ap2;

8090:     PetscFunctionBegin;
8091:     PetscCall(MatSeqAIJGetArray(A, &aa));
8092:     PetscCall(MatSeqAIJGetArray(B, &ba));
8093:     for (i = 0; i < m; i++) {
8094:       if (im[i] < 0) continue;
8095:       PetscCheck(im[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], mat->rmap->N - 1);
8096:       if (im[i] >= rstart && im[i] < rend) {
8097:         row      = im[i] - rstart;
8098:         lastcol1 = -1;
8099:         rp1      = aj + ai[row];
8100:         ap1      = aa + ai[row];
8101:         rmax1    = aimax[row];
8102:         nrow1    = ailen[row];
8103:         low1     = 0;
8104:         high1    = nrow1;
8105:         lastcol2 = -1;
8106:         rp2      = bj + bi[row];
8107:         ap2      = ba + bi[row];
8108:         rmax2    = bimax[row];
8109:         nrow2    = bilen[row];
8110:         low2     = 0;
8111:         high2    = nrow2;

8113:         for (j = 0; j < n; j++) {
8114:           if (roworiented) value = v[i * n + j];
8115:           else value = v[i + j * m];
8116:           if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8117:           if (in[j] >= cstart && in[j] < cend) {
8118:             col = in[j] - cstart;
8119:             MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8120:           } else if (in[j] < 0) continue;
8121:           else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8122:             SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8123:           } else {
8124:             if (mat->was_assembled) {
8125:               if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8126: #if defined(PETSC_USE_CTABLE)
8127:               PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8128:               col--;
8129: #else
8130:               col = aij->colmap[in[j]] - 1;
8131: #endif
8132:               if (col < 0 && !((Mat_SeqAIJ *)(aij->A->data))->nonew) {
8133:                 PetscCall(MatDisAssemble_MPIAIJ(mat));
8134:                 col = in[j];
8135:                 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8136:                 B        = aij->B;
8137:                 b        = (Mat_SeqAIJ *)B->data;
8138:                 bimax    = b->imax;
8139:                 bi       = b->i;
8140:                 bilen    = b->ilen;
8141:                 bj       = b->j;
8142:                 rp2      = bj + bi[row];
8143:                 ap2      = ba + bi[row];
8144:                 rmax2    = bimax[row];
8145:                 nrow2    = bilen[row];
8146:                 low2     = 0;
8147:                 high2    = nrow2;
8148:                 bm       = aij->B->rmap->n;
8149:                 ba       = b->a;
8150:                 inserted = PETSC_FALSE;
8151:               }
8152:             } else col = in[j];
8153:             MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8154:           }
8155:         }
8156:       } else if (!aij->donotstash) {
8157:         if (roworiented) {
8158:           PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8159:         } else {
8160:           PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8161:         }
8162:       }
8163:     }
8164:     PetscCall(MatSeqAIJRestoreArray(A, &aa));
8165:     PetscCall(MatSeqAIJRestoreArray(B, &ba));
8166:   }
8167:   PetscFunctionReturnVoid();
8168: }

8170: /* Undefining these here since they were redefined from their original definition above! No
8171:  * other PETSc functions should be defined past this point, as it is impossible to recover the
8172:  * original definitions */
8173: #undef PetscCall
8174: #undef SETERRQ