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

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

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

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

 82: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and  MatAssemblyEnd_MPI_Hash() */
 83: #define TYPE AIJ
 84: #define TYPE_AIJ
 85: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
 86: #undef TYPE
 87: #undef TYPE_AIJ

 89: static PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
 90: {
 91:   Mat B;

 93:   PetscFunctionBegin;
 94:   PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
 95:   PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
 96:   PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
 97:   PetscCall(MatDestroy(&B));
 98:   PetscFunctionReturn(PETSC_SUCCESS);
 99: }

101: static PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
102: {
103:   Mat B;

105:   PetscFunctionBegin;
106:   PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
107:   PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
108:   PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
109:   PetscFunctionReturn(PETSC_SUCCESS);
110: }

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

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

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

124:   Developer Note:
125:   Level: beginner

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

130: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`
131: M*/

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

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

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

145:   Level: beginner

147: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
148: M*/

150: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
151: {
152:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

154:   PetscFunctionBegin;
155: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_VIENNACL)
156:   A->boundtocpu = flg;
157: #endif
158:   if (a->A) PetscCall(MatBindToCPU(a->A, flg));
159:   if (a->B) PetscCall(MatBindToCPU(a->B, flg));

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

169: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
170: {
171:   Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;

173:   PetscFunctionBegin;
174:   if (mat->A) {
175:     PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
176:     PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
177:   }
178:   PetscFunctionReturn(PETSC_SUCCESS);
179: }

181: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
182: {
183:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ *)M->data;
184:   Mat_SeqAIJ      *a   = (Mat_SeqAIJ *)mat->A->data;
185:   Mat_SeqAIJ      *b   = (Mat_SeqAIJ *)mat->B->data;
186:   const PetscInt  *ia, *ib;
187:   const MatScalar *aa, *bb, *aav, *bav;
188:   PetscInt         na, nb, i, j, *rows, cnt = 0, n0rows;
189:   PetscInt         m = M->rmap->n, rstart = M->rmap->rstart;

191:   PetscFunctionBegin;
192:   *keptrows = NULL;

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

250: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
251: {
252:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
253:   PetscBool   cong;

255:   PetscFunctionBegin;
256:   PetscCall(MatHasCongruentLayouts(Y, &cong));
257:   if (Y->assembled && cong) {
258:     PetscCall(MatDiagonalSet(aij->A, D, is));
259:   } else {
260:     PetscCall(MatDiagonalSet_Default(Y, D, is));
261:   }
262:   PetscFunctionReturn(PETSC_SUCCESS);
263: }

265: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
266: {
267:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
268:   PetscInt    i, rstart, nrows, *rows;

270:   PetscFunctionBegin;
271:   *zrows = NULL;
272:   PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
273:   PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
274:   for (i = 0; i < nrows; i++) rows[i] += rstart;
275:   PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
276:   PetscFunctionReturn(PETSC_SUCCESS);
277: }

279: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
280: {
281:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ *)A->data;
282:   PetscInt           i, m, n, *garray = aij->garray;
283:   Mat_SeqAIJ        *a_aij = (Mat_SeqAIJ *)aij->A->data;
284:   Mat_SeqAIJ        *b_aij = (Mat_SeqAIJ *)aij->B->data;
285:   PetscReal         *work;
286:   const PetscScalar *dummy;

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

325: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
326: {
327:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)A->data;
328:   IS              sis, gis;
329:   const PetscInt *isis, *igis;
330:   PetscInt        n, *iis, nsis, ngis, rstart, i;

332:   PetscFunctionBegin;
333:   PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
334:   PetscCall(MatFindNonzeroRows(a->B, &gis));
335:   PetscCall(ISGetSize(gis, &ngis));
336:   PetscCall(ISGetSize(sis, &nsis));
337:   PetscCall(ISGetIndices(sis, &isis));
338:   PetscCall(ISGetIndices(gis, &igis));

340:   PetscCall(PetscMalloc1(ngis + nsis, &iis));
341:   PetscCall(PetscArraycpy(iis, igis, ngis));
342:   PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
343:   n = ngis + nsis;
344:   PetscCall(PetscSortRemoveDupsInt(&n, iis));
345:   PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
346:   for (i = 0; i < n; i++) iis[i] += rstart;
347:   PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));

349:   PetscCall(ISRestoreIndices(sis, &isis));
350:   PetscCall(ISRestoreIndices(gis, &igis));
351:   PetscCall(ISDestroy(&sis));
352:   PetscCall(ISDestroy(&gis));
353:   PetscFunctionReturn(PETSC_SUCCESS);
354: }

356: /*
357:   Local utility routine that creates a mapping from the global column
358: number to the local number in the off-diagonal part of the local
359: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at
360: a slightly higher hash table cost; without it it is not scalable (each processor
361: has an order N integer array but is fast to access.
362: */
363: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
364: {
365:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
366:   PetscInt    n   = aij->B->cmap->n, i;

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

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

426: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
427:   do { \
428:     if (col <= lastcol2) low2 = 0; \
429:     else high2 = nrow2; \
430:     lastcol2 = col; \
431:     while (high2 - low2 > 5) { \
432:       t = (low2 + high2) / 2; \
433:       if (rp2[t] > col) high2 = t; \
434:       else low2 = t; \
435:     } \
436:     for (_i = low2; _i < high2; _i++) { \
437:       if (rp2[_i] > col) break; \
438:       if (rp2[_i] == col) { \
439:         if (addv == ADD_VALUES) { \
440:           ap2[_i] += value; \
441:           (void)PetscLogFlops(1.0); \
442:         } else ap2[_i] = value; \
443:         goto b_noinsert; \
444:       } \
445:     } \
446:     if (value == 0.0 && ignorezeroentries) { \
447:       low2  = 0; \
448:       high2 = nrow2; \
449:       goto b_noinsert; \
450:     } \
451:     if (nonew == 1) { \
452:       low2  = 0; \
453:       high2 = nrow2; \
454:       goto b_noinsert; \
455:     } \
456:     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); \
457:     MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
458:     N = nrow2++ - 1; \
459:     b->nz++; \
460:     high2++; \
461:     /* shift up all the later entries in this row */ \
462:     PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
463:     PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
464:     rp2[_i] = col; \
465:     ap2[_i] = value; \
466:     B->nonzerostate++; \
467:   b_noinsert:; \
468:     bilen[row] = nrow2; \
469:   } while (0)

471: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
472: {
473:   Mat_MPIAIJ  *mat = (Mat_MPIAIJ *)A->data;
474:   Mat_SeqAIJ  *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
475:   PetscInt     l, *garray                         = mat->garray, diag;
476:   PetscScalar *aa, *ba;

478:   PetscFunctionBegin;
479:   /* code only works for square matrices A */

481:   /* find size of row to the left of the diagonal part */
482:   PetscCall(MatGetOwnershipRange(A, &diag, NULL));
483:   row = row - diag;
484:   for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
485:     if (garray[b->j[b->i[row] + l]] > diag) break;
486:   }
487:   if (l) {
488:     PetscCall(MatSeqAIJGetArray(mat->B, &ba));
489:     PetscCall(PetscArraycpy(ba + b->i[row], v, l));
490:     PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
491:   }

493:   /* diagonal part */
494:   if (a->i[row + 1] - a->i[row]) {
495:     PetscCall(MatSeqAIJGetArray(mat->A, &aa));
496:     PetscCall(PetscArraycpy(aa + a->i[row], v + l, (a->i[row + 1] - a->i[row])));
497:     PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
498:   }

500:   /* right of diagonal part */
501:   if (b->i[row + 1] - b->i[row] - l) {
502:     PetscCall(MatSeqAIJGetArray(mat->B, &ba));
503:     PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
504:     PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
505:   }
506:   PetscFunctionReturn(PETSC_SUCCESS);
507: }

509: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
510: {
511:   Mat_MPIAIJ *aij   = (Mat_MPIAIJ *)mat->data;
512:   PetscScalar value = 0.0;
513:   PetscInt    i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
514:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
515:   PetscBool   roworiented = aij->roworiented;

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

530:   PetscFunctionBegin;
531:   PetscCall(MatSeqAIJGetArray(A, &aa));
532:   PetscCall(MatSeqAIJGetArray(B, &ba));
533:   for (i = 0; i < m; i++) {
534:     if (im[i] < 0) continue;
535:     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);
536:     if (im[i] >= rstart && im[i] < rend) {
537:       row      = im[i] - rstart;
538:       lastcol1 = -1;
539:       rp1      = PetscSafePointerPlusOffset(aj, ai[row]);
540:       ap1      = PetscSafePointerPlusOffset(aa, ai[row]);
541:       rmax1    = aimax[row];
542:       nrow1    = ailen[row];
543:       low1     = 0;
544:       high1    = nrow1;
545:       lastcol2 = -1;
546:       rp2      = PetscSafePointerPlusOffset(bj, bi[row]);
547:       ap2      = PetscSafePointerPlusOffset(ba, bi[row]);
548:       rmax2    = bimax[row];
549:       nrow2    = bilen[row];
550:       low2     = 0;
551:       high2    = nrow2;

553:       for (j = 0; j < n; j++) {
554:         if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
555:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
556:         if (in[j] >= cstart && in[j] < cend) {
557:           col   = in[j] - cstart;
558:           nonew = a->nonew;
559:           MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
560:         } else if (in[j] < 0) {
561:           continue;
562:         } else {
563:           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);
564:           if (mat->was_assembled) {
565:             if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
566: #if defined(PETSC_USE_CTABLE)
567:             PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
568:             col--;
569: #else
570:             col = aij->colmap[in[j]] - 1;
571: #endif
572:             if (col < 0 && !((Mat_SeqAIJ *)aij->B->data)->nonew) { /* col < 0 means in[j] is a new col for B */
573:               PetscCall(MatDisAssemble_MPIAIJ(mat));               /* Change aij->B from reduced/local format to expanded/global format */
574:               col = in[j];
575:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
576:               B     = aij->B;
577:               b     = (Mat_SeqAIJ *)B->data;
578:               bimax = b->imax;
579:               bi    = b->i;
580:               bilen = b->ilen;
581:               bj    = b->j;
582:               ba    = b->a;
583:               rp2   = bj + bi[row];
584:               ap2   = ba + bi[row];
585:               rmax2 = bimax[row];
586:               nrow2 = bilen[row];
587:               low2  = 0;
588:               high2 = nrow2;
589:               bm    = aij->B->rmap->n;
590:               ba    = b->a;
591:             } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
592:               if (1 == ((Mat_SeqAIJ *)aij->B->data)->nonew) {
593:                 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]));
594:               } 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]);
595:             }
596:           } else col = in[j];
597:           nonew = b->nonew;
598:           MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
599:         }
600:       }
601:     } else {
602:       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]);
603:       if (!aij->donotstash) {
604:         mat->assembled = PETSC_FALSE;
605:         if (roworiented) {
606:           PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
607:         } else {
608:           PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
609:         }
610:       }
611:     }
612:   }
613:   PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
614:   PetscCall(MatSeqAIJRestoreArray(B, &ba));
615:   PetscFunctionReturn(PETSC_SUCCESS);
616: }

618: /*
619:     This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
620:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
621:     No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
622: */
623: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
624: {
625:   Mat_MPIAIJ *aij    = (Mat_MPIAIJ *)mat->data;
626:   Mat         A      = aij->A; /* diagonal part of the matrix */
627:   Mat         B      = aij->B; /* off-diagonal part of the matrix */
628:   Mat_SeqAIJ *a      = (Mat_SeqAIJ *)A->data;
629:   Mat_SeqAIJ *b      = (Mat_SeqAIJ *)B->data;
630:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
631:   PetscInt   *ailen = a->ilen, *aj = a->j;
632:   PetscInt   *bilen = b->ilen, *bj = b->j;
633:   PetscInt    am          = aij->A->rmap->n, j;
634:   PetscInt    diag_so_far = 0, dnz;
635:   PetscInt    offd_so_far = 0, onz;

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

658: /*
659:     This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
660:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
661:     No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
662:     Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
663:     would not be true and the more complex MatSetValues_MPIAIJ has to be used.
664: */
665: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
666: {
667:   Mat_MPIAIJ  *aij  = (Mat_MPIAIJ *)mat->data;
668:   Mat          A    = aij->A; /* diagonal part of the matrix */
669:   Mat          B    = aij->B; /* off-diagonal part of the matrix */
670:   Mat_SeqAIJ  *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
671:   Mat_SeqAIJ  *a      = (Mat_SeqAIJ *)A->data;
672:   Mat_SeqAIJ  *b      = (Mat_SeqAIJ *)B->data;
673:   PetscInt     cstart = mat->cmap->rstart, cend = mat->cmap->rend;
674:   PetscInt    *ailen = a->ilen, *aj = a->j;
675:   PetscInt    *bilen = b->ilen, *bj = b->j;
676:   PetscInt     am          = aij->A->rmap->n, j;
677:   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. */
678:   PetscInt     col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
679:   PetscScalar *aa = a->a, *ba = b->a;

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

706: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
707: {
708:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
709:   PetscInt    i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
710:   PetscInt    cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;

712:   PetscFunctionBegin;
713:   for (i = 0; i < m; i++) {
714:     if (idxm[i] < 0) continue; /* negative row */
715:     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);
716:     PetscCheck(idxm[i] >= rstart && idxm[i] < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported, row requested %" PetscInt_FMT " range [%" PetscInt_FMT " %" PetscInt_FMT ")", idxm[i], rstart, rend);
717:     row = idxm[i] - rstart;
718:     for (j = 0; j < n; j++) {
719:       if (idxn[j] < 0) continue; /* negative column */
720:       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);
721:       if (idxn[j] >= cstart && idxn[j] < cend) {
722:         col = idxn[j] - cstart;
723:         PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
724:       } else {
725:         if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
726: #if defined(PETSC_USE_CTABLE)
727:         PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
728:         col--;
729: #else
730:         col = aij->colmap[idxn[j]] - 1;
731: #endif
732:         if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
733:         else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
734:       }
735:     }
736:   }
737:   PetscFunctionReturn(PETSC_SUCCESS);
738: }

740: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
741: {
742:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
743:   PetscInt    nstash, reallocs;

745:   PetscFunctionBegin;
746:   if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);

748:   PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
749:   PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
750:   PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
751:   PetscFunctionReturn(PETSC_SUCCESS);
752: }

754: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
755: {
756:   Mat_MPIAIJ  *aij = (Mat_MPIAIJ *)mat->data;
757:   PetscMPIInt  n;
758:   PetscInt     i, j, rstart, ncols, flg;
759:   PetscInt    *row, *col;
760:   PetscBool    other_disassembled;
761:   PetscScalar *val;

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

765:   PetscFunctionBegin;
766:   if (!aij->donotstash && !mat->nooffprocentries) {
767:     while (1) {
768:       PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
769:       if (!flg) break;

771:       for (i = 0; i < n;) {
772:         /* Now identify the consecutive vals belonging to the same row */
773:         for (j = i, rstart = row[j]; j < n; j++) {
774:           if (row[j] != rstart) break;
775:         }
776:         if (j < n) ncols = j - i;
777:         else ncols = n - i;
778:         /* Now assemble all these values with a single function call */
779:         PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
780:         i = j;
781:       }
782:     }
783:     PetscCall(MatStashScatterEnd_Private(&mat->stash));
784:   }
785: #if defined(PETSC_HAVE_DEVICE)
786:   if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
787:   /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
788:   if (mat->boundtocpu) {
789:     PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
790:     PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
791:   }
792: #endif
793:   PetscCall(MatAssemblyBegin(aij->A, mode));
794:   PetscCall(MatAssemblyEnd(aij->A, mode));

796:   /* determine if any processor has disassembled, if so we must
797:      also disassemble ourself, in order that we may reassemble. */
798:   /*
799:      if nonzero structure of submatrix B cannot change then we know that
800:      no processor disassembled thus we can skip this stuff
801:   */
802:   if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
803:     PetscCall(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
804:     if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
805:       PetscCall(MatDisAssemble_MPIAIJ(mat));
806:     }
807:   }
808:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
809:   PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
810: #if defined(PETSC_HAVE_DEVICE)
811:   if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
812: #endif
813:   PetscCall(MatAssemblyBegin(aij->B, mode));
814:   PetscCall(MatAssemblyEnd(aij->B, mode));

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

818:   aij->rowvalues = NULL;

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

822:   /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
823:   if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
824:     PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
825:     PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
826:   }
827: #if defined(PETSC_HAVE_DEVICE)
828:   mat->offloadmask = PETSC_OFFLOAD_BOTH;
829: #endif
830:   PetscFunctionReturn(PETSC_SUCCESS);
831: }

833: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
834: {
835:   Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;

837:   PetscFunctionBegin;
838:   PetscCall(MatZeroEntries(l->A));
839:   PetscCall(MatZeroEntries(l->B));
840:   PetscFunctionReturn(PETSC_SUCCESS);
841: }

843: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
844: {
845:   Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
846:   PetscInt   *lrows;
847:   PetscInt    r, len;
848:   PetscBool   cong;

850:   PetscFunctionBegin;
851:   /* get locally owned rows */
852:   PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
853:   PetscCall(MatHasCongruentLayouts(A, &cong));
854:   /* fix right-hand side if needed */
855:   if (x && b) {
856:     const PetscScalar *xx;
857:     PetscScalar       *bb;

859:     PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
860:     PetscCall(VecGetArrayRead(x, &xx));
861:     PetscCall(VecGetArray(b, &bb));
862:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
863:     PetscCall(VecRestoreArrayRead(x, &xx));
864:     PetscCall(VecRestoreArray(b, &bb));
865:   }

867:   if (diag != 0.0 && cong) {
868:     PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
869:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
870:   } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
871:     Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
872:     Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
873:     PetscInt    nnwA, nnwB;
874:     PetscBool   nnzA, nnzB;

876:     nnwA = aijA->nonew;
877:     nnwB = aijB->nonew;
878:     nnzA = aijA->keepnonzeropattern;
879:     nnzB = aijB->keepnonzeropattern;
880:     if (!nnzA) {
881:       PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
882:       aijA->nonew = 0;
883:     }
884:     if (!nnzB) {
885:       PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
886:       aijB->nonew = 0;
887:     }
888:     /* Must zero here before the next loop */
889:     PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
890:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
891:     for (r = 0; r < len; ++r) {
892:       const PetscInt row = lrows[r] + A->rmap->rstart;
893:       if (row >= A->cmap->N) continue;
894:       PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
895:     }
896:     aijA->nonew = nnwA;
897:     aijB->nonew = nnwB;
898:   } else {
899:     PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
900:     PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
901:   }
902:   PetscCall(PetscFree(lrows));
903:   PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
904:   PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));

906:   /* only change matrix nonzero state if pattern was allowed to be changed */
907:   if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
908:     PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
909:     PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
910:   }
911:   PetscFunctionReturn(PETSC_SUCCESS);
912: }

914: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
915: {
916:   Mat_MPIAIJ        *l = (Mat_MPIAIJ *)A->data;
917:   PetscMPIInt        n = A->rmap->n;
918:   PetscInt           i, j, r, m, len = 0;
919:   PetscInt          *lrows, *owners = A->rmap->range;
920:   PetscMPIInt        p = 0;
921:   PetscSFNode       *rrows;
922:   PetscSF            sf;
923:   const PetscScalar *xx;
924:   PetscScalar       *bb, *mask, *aij_a;
925:   Vec                xmask, lmask;
926:   Mat_SeqAIJ        *aij = (Mat_SeqAIJ *)l->B->data;
927:   const PetscInt    *aj, *ii, *ridx;
928:   PetscScalar       *aa;

930:   PetscFunctionBegin;
931:   /* Create SF where leaves are input rows and roots are owned rows */
932:   PetscCall(PetscMalloc1(n, &lrows));
933:   for (r = 0; r < n; ++r) lrows[r] = -1;
934:   PetscCall(PetscMalloc1(N, &rrows));
935:   for (r = 0; r < N; ++r) {
936:     const PetscInt idx = rows[r];
937:     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);
938:     if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
939:       PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
940:     }
941:     rrows[r].rank  = p;
942:     rrows[r].index = rows[r] - owners[p];
943:   }
944:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
945:   PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
946:   /* Collect flags for rows to be zeroed */
947:   PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
948:   PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
949:   PetscCall(PetscSFDestroy(&sf));
950:   /* Compress and put in row numbers */
951:   for (r = 0; r < n; ++r)
952:     if (lrows[r] >= 0) lrows[len++] = r;
953:   /* zero diagonal part of matrix */
954:   PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
955:   /* handle off-diagonal part of matrix */
956:   PetscCall(MatCreateVecs(A, &xmask, NULL));
957:   PetscCall(VecDuplicate(l->lvec, &lmask));
958:   PetscCall(VecGetArray(xmask, &bb));
959:   for (i = 0; i < len; i++) bb[lrows[i]] = 1;
960:   PetscCall(VecRestoreArray(xmask, &bb));
961:   PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
962:   PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
963:   PetscCall(VecDestroy(&xmask));
964:   if (x && b) { /* this code is buggy when the row and column layout don't match */
965:     PetscBool cong;

967:     PetscCall(MatHasCongruentLayouts(A, &cong));
968:     PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
969:     PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
970:     PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
971:     PetscCall(VecGetArrayRead(l->lvec, &xx));
972:     PetscCall(VecGetArray(b, &bb));
973:   }
974:   PetscCall(VecGetArray(lmask, &mask));
975:   /* remove zeroed rows of off-diagonal matrix */
976:   PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
977:   ii = aij->i;
978:   for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
979:   /* loop over all elements of off process part of matrix zeroing removed columns*/
980:   if (aij->compressedrow.use) {
981:     m    = aij->compressedrow.nrows;
982:     ii   = aij->compressedrow.i;
983:     ridx = aij->compressedrow.rindex;
984:     for (i = 0; i < m; i++) {
985:       n  = ii[i + 1] - ii[i];
986:       aj = aij->j + ii[i];
987:       aa = aij_a + ii[i];

989:       for (j = 0; j < n; j++) {
990:         if (PetscAbsScalar(mask[*aj])) {
991:           if (b) bb[*ridx] -= *aa * xx[*aj];
992:           *aa = 0.0;
993:         }
994:         aa++;
995:         aj++;
996:       }
997:       ridx++;
998:     }
999:   } else { /* do not use compressed row format */
1000:     m = l->B->rmap->n;
1001:     for (i = 0; i < m; i++) {
1002:       n  = ii[i + 1] - ii[i];
1003:       aj = aij->j + ii[i];
1004:       aa = aij_a + ii[i];
1005:       for (j = 0; j < n; j++) {
1006:         if (PetscAbsScalar(mask[*aj])) {
1007:           if (b) bb[i] -= *aa * xx[*aj];
1008:           *aa = 0.0;
1009:         }
1010:         aa++;
1011:         aj++;
1012:       }
1013:     }
1014:   }
1015:   if (x && b) {
1016:     PetscCall(VecRestoreArray(b, &bb));
1017:     PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1018:   }
1019:   PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1020:   PetscCall(VecRestoreArray(lmask, &mask));
1021:   PetscCall(VecDestroy(&lmask));
1022:   PetscCall(PetscFree(lrows));

1024:   /* only change matrix nonzero state if pattern was allowed to be changed */
1025:   if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1026:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1027:     PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1028:   }
1029:   PetscFunctionReturn(PETSC_SUCCESS);
1030: }

1032: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1033: {
1034:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1035:   PetscInt    nt;
1036:   VecScatter  Mvctx = a->Mvctx;

1038:   PetscFunctionBegin;
1039:   PetscCall(VecGetLocalSize(xx, &nt));
1040:   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);
1041:   PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1042:   PetscUseTypeMethod(a->A, mult, xx, yy);
1043:   PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1044:   PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1045:   PetscFunctionReturn(PETSC_SUCCESS);
1046: }

1048: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1049: {
1050:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1052:   PetscFunctionBegin;
1053:   PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1054:   PetscFunctionReturn(PETSC_SUCCESS);
1055: }

1057: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1058: {
1059:   Mat_MPIAIJ *a     = (Mat_MPIAIJ *)A->data;
1060:   VecScatter  Mvctx = a->Mvctx;

1062:   PetscFunctionBegin;
1063:   PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1064:   PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1065:   PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1066:   PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1067:   PetscFunctionReturn(PETSC_SUCCESS);
1068: }

1070: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1071: {
1072:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1074:   PetscFunctionBegin;
1075:   /* do nondiagonal part */
1076:   PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1077:   /* do local part */
1078:   PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1079:   /* add partial results together */
1080:   PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1081:   PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1082:   PetscFunctionReturn(PETSC_SUCCESS);
1083: }

1085: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1086: {
1087:   MPI_Comm    comm;
1088:   Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1089:   Mat         Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1090:   IS          Me, Notme;
1091:   PetscInt    M, N, first, last, *notme, i;
1092:   PetscBool   lf;
1093:   PetscMPIInt size;

1095:   PetscFunctionBegin;
1096:   /* Easy test: symmetric diagonal block */
1097:   PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1098:   PetscCall(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1099:   if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1100:   PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1101:   PetscCallMPI(MPI_Comm_size(comm, &size));
1102:   if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);

1104:   /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1105:   PetscCall(MatGetSize(Amat, &M, &N));
1106:   PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1107:   PetscCall(PetscMalloc1(N - last + first, &notme));
1108:   for (i = 0; i < first; i++) notme[i] = i;
1109:   for (i = last; i < M; i++) notme[i - last + first] = i;
1110:   PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1111:   PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1112:   PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1113:   Aoff = Aoffs[0];
1114:   PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1115:   Boff = Boffs[0];
1116:   PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1117:   PetscCall(MatDestroyMatrices(1, &Aoffs));
1118:   PetscCall(MatDestroyMatrices(1, &Boffs));
1119:   PetscCall(ISDestroy(&Me));
1120:   PetscCall(ISDestroy(&Notme));
1121:   PetscCall(PetscFree(notme));
1122:   PetscFunctionReturn(PETSC_SUCCESS);
1123: }

1125: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1126: {
1127:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1129:   PetscFunctionBegin;
1130:   /* do nondiagonal part */
1131:   PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1132:   /* do local part */
1133:   PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1134:   /* add partial results together */
1135:   PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1136:   PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1137:   PetscFunctionReturn(PETSC_SUCCESS);
1138: }

1140: /*
1141:   This only works correctly for square matrices where the subblock A->A is the
1142:    diagonal block
1143: */
1144: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1145: {
1146:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1148:   PetscFunctionBegin;
1149:   PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1150:   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");
1151:   PetscCall(MatGetDiagonal(a->A, v));
1152:   PetscFunctionReturn(PETSC_SUCCESS);
1153: }

1155: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1156: {
1157:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1159:   PetscFunctionBegin;
1160:   PetscCall(MatScale(a->A, aa));
1161:   PetscCall(MatScale(a->B, aa));
1162:   PetscFunctionReturn(PETSC_SUCCESS);
1163: }

1165: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1166: {
1167:   Mat_MPIAIJ        *aij    = (Mat_MPIAIJ *)mat->data;
1168:   Mat_SeqAIJ        *A      = (Mat_SeqAIJ *)aij->A->data;
1169:   Mat_SeqAIJ        *B      = (Mat_SeqAIJ *)aij->B->data;
1170:   const PetscInt    *garray = aij->garray;
1171:   const PetscScalar *aa, *ba;
1172:   PetscInt           header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1173:   PetscInt64         nz, hnz;
1174:   PetscInt          *rowlens;
1175:   PetscInt          *colidxs;
1176:   PetscScalar       *matvals;
1177:   PetscMPIInt        rank;

1179:   PetscFunctionBegin;
1180:   PetscCall(PetscViewerSetUp(viewer));

1182:   M  = mat->rmap->N;
1183:   N  = mat->cmap->N;
1184:   m  = mat->rmap->n;
1185:   rs = mat->rmap->rstart;
1186:   cs = mat->cmap->rstart;
1187:   nz = A->nz + B->nz;

1189:   /* write matrix header */
1190:   header[0] = MAT_FILE_CLASSID;
1191:   header[1] = M;
1192:   header[2] = N;
1193:   PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1194:   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1195:   if (rank == 0) {
1196:     if (hnz > PETSC_MAX_INT) header[3] = PETSC_MAX_INT;
1197:     else header[3] = (PetscInt)hnz;
1198:   }
1199:   PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));

1201:   /* fill in and store row lengths  */
1202:   PetscCall(PetscMalloc1(m, &rowlens));
1203:   for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1204:   PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1205:   PetscCall(PetscFree(rowlens));

1207:   /* fill in and store column indices */
1208:   PetscCall(PetscMalloc1(nz, &colidxs));
1209:   for (cnt = 0, i = 0; i < m; i++) {
1210:     for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1211:       if (garray[B->j[jb]] > cs) break;
1212:       colidxs[cnt++] = garray[B->j[jb]];
1213:     }
1214:     for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1215:     for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1216:   }
1217:   PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1218:   PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1219:   PetscCall(PetscFree(colidxs));

1221:   /* fill in and store nonzero values */
1222:   PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1223:   PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1224:   PetscCall(PetscMalloc1(nz, &matvals));
1225:   for (cnt = 0, i = 0; i < m; i++) {
1226:     for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1227:       if (garray[B->j[jb]] > cs) break;
1228:       matvals[cnt++] = ba[jb];
1229:     }
1230:     for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1231:     for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1232:   }
1233:   PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1234:   PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1235:   PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1236:   PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1237:   PetscCall(PetscFree(matvals));

1239:   /* write block size option to the viewer's .info file */
1240:   PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1241:   PetscFunctionReturn(PETSC_SUCCESS);
1242: }

1244: #include <petscdraw.h>
1245: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1246: {
1247:   Mat_MPIAIJ       *aij  = (Mat_MPIAIJ *)mat->data;
1248:   PetscMPIInt       rank = aij->rank, size = aij->size;
1249:   PetscBool         isdraw, iascii, isbinary;
1250:   PetscViewer       sviewer;
1251:   PetscViewerFormat format;

1253:   PetscFunctionBegin;
1254:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1255:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1256:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1257:   if (iascii) {
1258:     PetscCall(PetscViewerGetFormat(viewer, &format));
1259:     if (format == PETSC_VIEWER_LOAD_BALANCE) {
1260:       PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1261:       PetscCall(PetscMalloc1(size, &nz));
1262:       PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1263:       for (i = 0; i < (PetscInt)size; i++) {
1264:         nmax = PetscMax(nmax, nz[i]);
1265:         nmin = PetscMin(nmin, nz[i]);
1266:         navg += nz[i];
1267:       }
1268:       PetscCall(PetscFree(nz));
1269:       navg = navg / size;
1270:       PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT "  avg %" PetscInt_FMT "  max %" PetscInt_FMT "\n", nmin, navg, nmax));
1271:       PetscFunctionReturn(PETSC_SUCCESS);
1272:     }
1273:     PetscCall(PetscViewerGetFormat(viewer, &format));
1274:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1275:       MatInfo   info;
1276:       PetscInt *inodes = NULL;

1278:       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1279:       PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1280:       PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1281:       PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1282:       if (!inodes) {
1283:         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,
1284:                                                      (double)info.memory));
1285:       } else {
1286:         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,
1287:                                                      (double)info.memory));
1288:       }
1289:       PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1290:       PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1291:       PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1292:       PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1293:       PetscCall(PetscViewerFlush(viewer));
1294:       PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1295:       PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1296:       PetscCall(VecScatterView(aij->Mvctx, viewer));
1297:       PetscFunctionReturn(PETSC_SUCCESS);
1298:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
1299:       PetscInt inodecount, inodelimit, *inodes;
1300:       PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1301:       if (inodes) {
1302:         PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1303:       } else {
1304:         PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1305:       }
1306:       PetscFunctionReturn(PETSC_SUCCESS);
1307:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1308:       PetscFunctionReturn(PETSC_SUCCESS);
1309:     }
1310:   } else if (isbinary) {
1311:     if (size == 1) {
1312:       PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1313:       PetscCall(MatView(aij->A, viewer));
1314:     } else {
1315:       PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1316:     }
1317:     PetscFunctionReturn(PETSC_SUCCESS);
1318:   } else if (iascii && size == 1) {
1319:     PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1320:     PetscCall(MatView(aij->A, viewer));
1321:     PetscFunctionReturn(PETSC_SUCCESS);
1322:   } else if (isdraw) {
1323:     PetscDraw draw;
1324:     PetscBool isnull;
1325:     PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1326:     PetscCall(PetscDrawIsNull(draw, &isnull));
1327:     if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1328:   }

1330:   { /* assemble the entire matrix onto first processor */
1331:     Mat A = NULL, Av;
1332:     IS  isrow, iscol;

1334:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1335:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1336:     PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1337:     PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1338:     /*  The commented code uses MatCreateSubMatrices instead */
1339:     /*
1340:     Mat *AA, A = NULL, Av;
1341:     IS  isrow,iscol;

1343:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1344:     PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1345:     PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1346:     if (rank == 0) {
1347:        PetscCall(PetscObjectReference((PetscObject)AA[0]));
1348:        A    = AA[0];
1349:        Av   = AA[0];
1350:     }
1351:     PetscCall(MatDestroySubMatrices(1,&AA));
1352: */
1353:     PetscCall(ISDestroy(&iscol));
1354:     PetscCall(ISDestroy(&isrow));
1355:     /*
1356:        Everyone has to call to draw the matrix since the graphics waits are
1357:        synchronized across all processors that share the PetscDraw object
1358:     */
1359:     PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1360:     if (rank == 0) {
1361:       if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1362:       PetscCall(MatView_SeqAIJ(Av, sviewer));
1363:     }
1364:     PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1365:     PetscCall(MatDestroy(&A));
1366:   }
1367:   PetscFunctionReturn(PETSC_SUCCESS);
1368: }

1370: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1371: {
1372:   PetscBool iascii, isdraw, issocket, isbinary;

1374:   PetscFunctionBegin;
1375:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1376:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1377:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1378:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1379:   if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1380:   PetscFunctionReturn(PETSC_SUCCESS);
1381: }

1383: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1384: {
1385:   Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1386:   Vec         bb1 = NULL;
1387:   PetscBool   hasop;

1389:   PetscFunctionBegin;
1390:   if (flag == SOR_APPLY_UPPER) {
1391:     PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1392:     PetscFunctionReturn(PETSC_SUCCESS);
1393:   }

1395:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));

1397:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1398:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1399:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1400:       its--;
1401:     }

1403:     while (its--) {
1404:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1405:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1407:       /* update rhs: bb1 = bb - B*x */
1408:       PetscCall(VecScale(mat->lvec, -1.0));
1409:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1411:       /* local sweep */
1412:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1413:     }
1414:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1415:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1416:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1417:       its--;
1418:     }
1419:     while (its--) {
1420:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1421:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1423:       /* update rhs: bb1 = bb - B*x */
1424:       PetscCall(VecScale(mat->lvec, -1.0));
1425:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1427:       /* local sweep */
1428:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1429:     }
1430:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1431:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1432:       PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1433:       its--;
1434:     }
1435:     while (its--) {
1436:       PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1437:       PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));

1439:       /* update rhs: bb1 = bb - B*x */
1440:       PetscCall(VecScale(mat->lvec, -1.0));
1441:       PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));

1443:       /* local sweep */
1444:       PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1445:     }
1446:   } else if (flag & SOR_EISENSTAT) {
1447:     Vec xx1;

1449:     PetscCall(VecDuplicate(bb, &xx1));
1450:     PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));

1452:     PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1453:     PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1454:     if (!mat->diag) {
1455:       PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1456:       PetscCall(MatGetDiagonal(matin, mat->diag));
1457:     }
1458:     PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1459:     if (hasop) {
1460:       PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1461:     } else {
1462:       PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1463:     }
1464:     PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));

1466:     PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));

1468:     /* local sweep */
1469:     PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1470:     PetscCall(VecAXPY(xx, 1.0, xx1));
1471:     PetscCall(VecDestroy(&xx1));
1472:   } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");

1474:   PetscCall(VecDestroy(&bb1));

1476:   matin->factorerrortype = mat->A->factorerrortype;
1477:   PetscFunctionReturn(PETSC_SUCCESS);
1478: }

1480: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1481: {
1482:   Mat             aA, aB, Aperm;
1483:   const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1484:   PetscScalar    *aa, *ba;
1485:   PetscInt        i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1486:   PetscSF         rowsf, sf;
1487:   IS              parcolp = NULL;
1488:   PetscBool       done;

1490:   PetscFunctionBegin;
1491:   PetscCall(MatGetLocalSize(A, &m, &n));
1492:   PetscCall(ISGetIndices(rowp, &rwant));
1493:   PetscCall(ISGetIndices(colp, &cwant));
1494:   PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));

1496:   /* Invert row permutation to find out where my rows should go */
1497:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1498:   PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1499:   PetscCall(PetscSFSetFromOptions(rowsf));
1500:   for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1501:   PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1502:   PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));

1504:   /* Invert column permutation to find out where my columns should go */
1505:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1506:   PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1507:   PetscCall(PetscSFSetFromOptions(sf));
1508:   for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1509:   PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1510:   PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1511:   PetscCall(PetscSFDestroy(&sf));

1513:   PetscCall(ISRestoreIndices(rowp, &rwant));
1514:   PetscCall(ISRestoreIndices(colp, &cwant));
1515:   PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));

1517:   /* Find out where my gcols should go */
1518:   PetscCall(MatGetSize(aB, NULL, &ng));
1519:   PetscCall(PetscMalloc1(ng, &gcdest));
1520:   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1521:   PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1522:   PetscCall(PetscSFSetFromOptions(sf));
1523:   PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1524:   PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1525:   PetscCall(PetscSFDestroy(&sf));

1527:   PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1528:   PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1529:   PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1530:   for (i = 0; i < m; i++) {
1531:     PetscInt    row = rdest[i];
1532:     PetscMPIInt rowner;
1533:     PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1534:     for (j = ai[i]; j < ai[i + 1]; j++) {
1535:       PetscInt    col = cdest[aj[j]];
1536:       PetscMPIInt cowner;
1537:       PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1538:       if (rowner == cowner) dnnz[i]++;
1539:       else onnz[i]++;
1540:     }
1541:     for (j = bi[i]; j < bi[i + 1]; j++) {
1542:       PetscInt    col = gcdest[bj[j]];
1543:       PetscMPIInt cowner;
1544:       PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1545:       if (rowner == cowner) dnnz[i]++;
1546:       else onnz[i]++;
1547:     }
1548:   }
1549:   PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1550:   PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1551:   PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1552:   PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1553:   PetscCall(PetscSFDestroy(&rowsf));

1555:   PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1556:   PetscCall(MatSeqAIJGetArray(aA, &aa));
1557:   PetscCall(MatSeqAIJGetArray(aB, &ba));
1558:   for (i = 0; i < m; i++) {
1559:     PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1560:     PetscInt  j0, rowlen;
1561:     rowlen = ai[i + 1] - ai[i];
1562:     for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1563:       for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1564:       PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1565:     }
1566:     rowlen = bi[i + 1] - bi[i];
1567:     for (j0 = j = 0; j < rowlen; j0 = j) {
1568:       for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1569:       PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1570:     }
1571:   }
1572:   PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1573:   PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1574:   PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1575:   PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1576:   PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1577:   PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1578:   PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1579:   PetscCall(PetscFree3(work, rdest, cdest));
1580:   PetscCall(PetscFree(gcdest));
1581:   if (parcolp) PetscCall(ISDestroy(&colp));
1582:   *B = Aperm;
1583:   PetscFunctionReturn(PETSC_SUCCESS);
1584: }

1586: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1587: {
1588:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

1590:   PetscFunctionBegin;
1591:   PetscCall(MatGetSize(aij->B, NULL, nghosts));
1592:   if (ghosts) *ghosts = aij->garray;
1593:   PetscFunctionReturn(PETSC_SUCCESS);
1594: }

1596: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1597: {
1598:   Mat_MPIAIJ    *mat = (Mat_MPIAIJ *)matin->data;
1599:   Mat            A = mat->A, B = mat->B;
1600:   PetscLogDouble isend[5], irecv[5];

1602:   PetscFunctionBegin;
1603:   info->block_size = 1.0;
1604:   PetscCall(MatGetInfo(A, MAT_LOCAL, info));

1606:   isend[0] = info->nz_used;
1607:   isend[1] = info->nz_allocated;
1608:   isend[2] = info->nz_unneeded;
1609:   isend[3] = info->memory;
1610:   isend[4] = info->mallocs;

1612:   PetscCall(MatGetInfo(B, MAT_LOCAL, info));

1614:   isend[0] += info->nz_used;
1615:   isend[1] += info->nz_allocated;
1616:   isend[2] += info->nz_unneeded;
1617:   isend[3] += info->memory;
1618:   isend[4] += info->mallocs;
1619:   if (flag == MAT_LOCAL) {
1620:     info->nz_used      = isend[0];
1621:     info->nz_allocated = isend[1];
1622:     info->nz_unneeded  = isend[2];
1623:     info->memory       = isend[3];
1624:     info->mallocs      = isend[4];
1625:   } else if (flag == MAT_GLOBAL_MAX) {
1626:     PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));

1628:     info->nz_used      = irecv[0];
1629:     info->nz_allocated = irecv[1];
1630:     info->nz_unneeded  = irecv[2];
1631:     info->memory       = irecv[3];
1632:     info->mallocs      = irecv[4];
1633:   } else if (flag == MAT_GLOBAL_SUM) {
1634:     PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));

1636:     info->nz_used      = irecv[0];
1637:     info->nz_allocated = irecv[1];
1638:     info->nz_unneeded  = irecv[2];
1639:     info->memory       = irecv[3];
1640:     info->mallocs      = irecv[4];
1641:   }
1642:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1643:   info->fill_ratio_needed = 0;
1644:   info->factor_mallocs    = 0;
1645:   PetscFunctionReturn(PETSC_SUCCESS);
1646: }

1648: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1649: {
1650:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

1652:   PetscFunctionBegin;
1653:   switch (op) {
1654:   case MAT_NEW_NONZERO_LOCATIONS:
1655:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1656:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1657:   case MAT_KEEP_NONZERO_PATTERN:
1658:   case MAT_NEW_NONZERO_LOCATION_ERR:
1659:   case MAT_USE_INODES:
1660:   case MAT_IGNORE_ZERO_ENTRIES:
1661:   case MAT_FORM_EXPLICIT_TRANSPOSE:
1662:     MatCheckPreallocated(A, 1);
1663:     PetscCall(MatSetOption(a->A, op, flg));
1664:     PetscCall(MatSetOption(a->B, op, flg));
1665:     break;
1666:   case MAT_ROW_ORIENTED:
1667:     MatCheckPreallocated(A, 1);
1668:     a->roworiented = flg;

1670:     PetscCall(MatSetOption(a->A, op, flg));
1671:     PetscCall(MatSetOption(a->B, op, flg));
1672:     break;
1673:   case MAT_FORCE_DIAGONAL_ENTRIES:
1674:   case MAT_SORTED_FULL:
1675:     PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1676:     break;
1677:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1678:     a->donotstash = flg;
1679:     break;
1680:   /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1681:   case MAT_SPD:
1682:   case MAT_SYMMETRIC:
1683:   case MAT_STRUCTURALLY_SYMMETRIC:
1684:   case MAT_HERMITIAN:
1685:   case MAT_SYMMETRY_ETERNAL:
1686:   case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1687:   case MAT_SPD_ETERNAL:
1688:     /* if the diagonal matrix is square it inherits some of the properties above */
1689:     break;
1690:   case MAT_SUBMAT_SINGLEIS:
1691:     A->submat_singleis = flg;
1692:     break;
1693:   case MAT_STRUCTURE_ONLY:
1694:     /* The option is handled directly by MatSetOption() */
1695:     break;
1696:   default:
1697:     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1698:   }
1699:   PetscFunctionReturn(PETSC_SUCCESS);
1700: }

1702: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1703: {
1704:   Mat_MPIAIJ  *mat = (Mat_MPIAIJ *)matin->data;
1705:   PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1706:   PetscInt     i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1707:   PetscInt     nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1708:   PetscInt    *cmap, *idx_p;

1710:   PetscFunctionBegin;
1711:   PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1712:   mat->getrowactive = PETSC_TRUE;

1714:   if (!mat->rowvalues && (idx || v)) {
1715:     /*
1716:         allocate enough space to hold information from the longest row.
1717:     */
1718:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1719:     PetscInt    max = 1, tmp;
1720:     for (i = 0; i < matin->rmap->n; i++) {
1721:       tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1722:       if (max < tmp) max = tmp;
1723:     }
1724:     PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1725:   }

1727:   PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1728:   lrow = row - rstart;

1730:   pvA = &vworkA;
1731:   pcA = &cworkA;
1732:   pvB = &vworkB;
1733:   pcB = &cworkB;
1734:   if (!v) {
1735:     pvA = NULL;
1736:     pvB = NULL;
1737:   }
1738:   if (!idx) {
1739:     pcA = NULL;
1740:     if (!v) pcB = NULL;
1741:   }
1742:   PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1743:   PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1744:   nztot = nzA + nzB;

1746:   cmap = mat->garray;
1747:   if (v || idx) {
1748:     if (nztot) {
1749:       /* Sort by increasing column numbers, assuming A and B already sorted */
1750:       PetscInt imark = -1;
1751:       if (v) {
1752:         *v = v_p = mat->rowvalues;
1753:         for (i = 0; i < nzB; i++) {
1754:           if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1755:           else break;
1756:         }
1757:         imark = i;
1758:         for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1759:         for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1760:       }
1761:       if (idx) {
1762:         *idx = idx_p = mat->rowindices;
1763:         if (imark > -1) {
1764:           for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1765:         } else {
1766:           for (i = 0; i < nzB; i++) {
1767:             if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1768:             else break;
1769:           }
1770:           imark = i;
1771:         }
1772:         for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1773:         for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1774:       }
1775:     } else {
1776:       if (idx) *idx = NULL;
1777:       if (v) *v = NULL;
1778:     }
1779:   }
1780:   *nz = nztot;
1781:   PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1782:   PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1783:   PetscFunctionReturn(PETSC_SUCCESS);
1784: }

1786: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1787: {
1788:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

1790:   PetscFunctionBegin;
1791:   PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1792:   aij->getrowactive = PETSC_FALSE;
1793:   PetscFunctionReturn(PETSC_SUCCESS);
1794: }

1796: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1797: {
1798:   Mat_MPIAIJ      *aij  = (Mat_MPIAIJ *)mat->data;
1799:   Mat_SeqAIJ      *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1800:   PetscInt         i, j, cstart = mat->cmap->rstart;
1801:   PetscReal        sum = 0.0;
1802:   const MatScalar *v, *amata, *bmata;

1804:   PetscFunctionBegin;
1805:   if (aij->size == 1) {
1806:     PetscCall(MatNorm(aij->A, type, norm));
1807:   } else {
1808:     PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1809:     PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1810:     if (type == NORM_FROBENIUS) {
1811:       v = amata;
1812:       for (i = 0; i < amat->nz; i++) {
1813:         sum += PetscRealPart(PetscConj(*v) * (*v));
1814:         v++;
1815:       }
1816:       v = bmata;
1817:       for (i = 0; i < bmat->nz; i++) {
1818:         sum += PetscRealPart(PetscConj(*v) * (*v));
1819:         v++;
1820:       }
1821:       PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1822:       *norm = PetscSqrtReal(*norm);
1823:       PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1824:     } else if (type == NORM_1) { /* max column norm */
1825:       PetscReal *tmp, *tmp2;
1826:       PetscInt  *jj, *garray = aij->garray;
1827:       PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1828:       PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1829:       *norm = 0.0;
1830:       v     = amata;
1831:       jj    = amat->j;
1832:       for (j = 0; j < amat->nz; j++) {
1833:         tmp[cstart + *jj++] += PetscAbsScalar(*v);
1834:         v++;
1835:       }
1836:       v  = bmata;
1837:       jj = bmat->j;
1838:       for (j = 0; j < bmat->nz; j++) {
1839:         tmp[garray[*jj++]] += PetscAbsScalar(*v);
1840:         v++;
1841:       }
1842:       PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1843:       for (j = 0; j < mat->cmap->N; j++) {
1844:         if (tmp2[j] > *norm) *norm = tmp2[j];
1845:       }
1846:       PetscCall(PetscFree(tmp));
1847:       PetscCall(PetscFree(tmp2));
1848:       PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1849:     } else if (type == NORM_INFINITY) { /* max row norm */
1850:       PetscReal ntemp = 0.0;
1851:       for (j = 0; j < aij->A->rmap->n; j++) {
1852:         v   = PetscSafePointerPlusOffset(amata, amat->i[j]);
1853:         sum = 0.0;
1854:         for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1855:           sum += PetscAbsScalar(*v);
1856:           v++;
1857:         }
1858:         v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1859:         for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1860:           sum += PetscAbsScalar(*v);
1861:           v++;
1862:         }
1863:         if (sum > ntemp) ntemp = sum;
1864:       }
1865:       PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1866:       PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1867:     } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1868:     PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1869:     PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1870:   }
1871:   PetscFunctionReturn(PETSC_SUCCESS);
1872: }

1874: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1875: {
1876:   Mat_MPIAIJ      *a    = (Mat_MPIAIJ *)A->data, *b;
1877:   Mat_SeqAIJ      *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1878:   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;
1879:   const PetscInt  *ai, *aj, *bi, *bj, *B_diag_i;
1880:   Mat              B, A_diag, *B_diag;
1881:   const MatScalar *pbv, *bv;

1883:   PetscFunctionBegin;
1884:   if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1885:   ma = A->rmap->n;
1886:   na = A->cmap->n;
1887:   mb = a->B->rmap->n;
1888:   nb = a->B->cmap->n;
1889:   ai = Aloc->i;
1890:   aj = Aloc->j;
1891:   bi = Bloc->i;
1892:   bj = Bloc->j;
1893:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1894:     PetscInt            *d_nnz, *g_nnz, *o_nnz;
1895:     PetscSFNode         *oloc;
1896:     PETSC_UNUSED PetscSF sf;

1898:     PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1899:     /* compute d_nnz for preallocation */
1900:     PetscCall(PetscArrayzero(d_nnz, na));
1901:     for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1902:     /* compute local off-diagonal contributions */
1903:     PetscCall(PetscArrayzero(g_nnz, nb));
1904:     for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1905:     /* map those to global */
1906:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1907:     PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1908:     PetscCall(PetscSFSetFromOptions(sf));
1909:     PetscCall(PetscArrayzero(o_nnz, na));
1910:     PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1911:     PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1912:     PetscCall(PetscSFDestroy(&sf));

1914:     PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1915:     PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1916:     PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1917:     PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1918:     PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1919:     PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1920:   } else {
1921:     B = *matout;
1922:     PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1923:   }

1925:   b           = (Mat_MPIAIJ *)B->data;
1926:   A_diag      = a->A;
1927:   B_diag      = &b->A;
1928:   sub_B_diag  = (Mat_SeqAIJ *)(*B_diag)->data;
1929:   A_diag_ncol = A_diag->cmap->N;
1930:   B_diag_ilen = sub_B_diag->ilen;
1931:   B_diag_i    = sub_B_diag->i;

1933:   /* Set ilen for diagonal of B */
1934:   for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];

1936:   /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1937:   very quickly (=without using MatSetValues), because all writes are local. */
1938:   PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1939:   PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));

1941:   /* copy over the B part */
1942:   PetscCall(PetscMalloc1(bi[mb], &cols));
1943:   PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1944:   pbv = bv;
1945:   row = A->rmap->rstart;
1946:   for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1947:   cols_tmp = cols;
1948:   for (i = 0; i < mb; i++) {
1949:     ncol = bi[i + 1] - bi[i];
1950:     PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1951:     row++;
1952:     if (pbv) pbv += ncol;
1953:     if (cols_tmp) cols_tmp += ncol;
1954:   }
1955:   PetscCall(PetscFree(cols));
1956:   PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));

1958:   PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1959:   PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1960:   if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1961:     *matout = B;
1962:   } else {
1963:     PetscCall(MatHeaderMerge(A, &B));
1964:   }
1965:   PetscFunctionReturn(PETSC_SUCCESS);
1966: }

1968: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1969: {
1970:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1971:   Mat         a = aij->A, b = aij->B;
1972:   PetscInt    s1, s2, s3;

1974:   PetscFunctionBegin;
1975:   PetscCall(MatGetLocalSize(mat, &s2, &s3));
1976:   if (rr) {
1977:     PetscCall(VecGetLocalSize(rr, &s1));
1978:     PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1979:     /* Overlap communication with computation. */
1980:     PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1981:   }
1982:   if (ll) {
1983:     PetscCall(VecGetLocalSize(ll, &s1));
1984:     PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1985:     PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1986:   }
1987:   /* scale  the diagonal block */
1988:   PetscUseTypeMethod(a, diagonalscale, ll, rr);

1990:   if (rr) {
1991:     /* Do a scatter end and then right scale the off-diagonal block */
1992:     PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1993:     PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1994:   }
1995:   PetscFunctionReturn(PETSC_SUCCESS);
1996: }

1998: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1999: {
2000:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2002:   PetscFunctionBegin;
2003:   PetscCall(MatSetUnfactored(a->A));
2004:   PetscFunctionReturn(PETSC_SUCCESS);
2005: }

2007: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2008: {
2009:   Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2010:   Mat         a, b, c, d;
2011:   PetscBool   flg;

2013:   PetscFunctionBegin;
2014:   a = matA->A;
2015:   b = matA->B;
2016:   c = matB->A;
2017:   d = matB->B;

2019:   PetscCall(MatEqual(a, c, &flg));
2020:   if (flg) PetscCall(MatEqual(b, d, &flg));
2021:   PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2022:   PetscFunctionReturn(PETSC_SUCCESS);
2023: }

2025: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2026: {
2027:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2028:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;

2030:   PetscFunctionBegin;
2031:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2032:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2033:     /* because of the column compression in the off-processor part of the matrix a->B,
2034:        the number of columns in a->B and b->B may be different, hence we cannot call
2035:        the MatCopy() directly on the two parts. If need be, we can provide a more
2036:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2037:        then copying the submatrices */
2038:     PetscCall(MatCopy_Basic(A, B, str));
2039:   } else {
2040:     PetscCall(MatCopy(a->A, b->A, str));
2041:     PetscCall(MatCopy(a->B, b->B, str));
2042:   }
2043:   PetscCall(PetscObjectStateIncrease((PetscObject)B));
2044:   PetscFunctionReturn(PETSC_SUCCESS);
2045: }

2047: /*
2048:    Computes the number of nonzeros per row needed for preallocation when X and Y
2049:    have different nonzero structure.
2050: */
2051: 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)
2052: {
2053:   PetscInt i, j, k, nzx, nzy;

2055:   PetscFunctionBegin;
2056:   /* Set the number of nonzeros in the new matrix */
2057:   for (i = 0; i < m; i++) {
2058:     const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2059:     nzx    = xi[i + 1] - xi[i];
2060:     nzy    = yi[i + 1] - yi[i];
2061:     nnz[i] = 0;
2062:     for (j = 0, k = 0; j < nzx; j++) {                                /* Point in X */
2063:       for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2064:       if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++;             /* Skip duplicate */
2065:       nnz[i]++;
2066:     }
2067:     for (; k < nzy; k++) nnz[i]++;
2068:   }
2069:   PetscFunctionReturn(PETSC_SUCCESS);
2070: }

2072: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2073: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2074: {
2075:   PetscInt    m = Y->rmap->N;
2076:   Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2077:   Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;

2079:   PetscFunctionBegin;
2080:   PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2081:   PetscFunctionReturn(PETSC_SUCCESS);
2082: }

2084: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2085: {
2086:   Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;

2088:   PetscFunctionBegin;
2089:   if (str == SAME_NONZERO_PATTERN) {
2090:     PetscCall(MatAXPY(yy->A, a, xx->A, str));
2091:     PetscCall(MatAXPY(yy->B, a, xx->B, str));
2092:   } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2093:     PetscCall(MatAXPY_Basic(Y, a, X, str));
2094:   } else {
2095:     Mat       B;
2096:     PetscInt *nnz_d, *nnz_o;

2098:     PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2099:     PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2100:     PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2101:     PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2102:     PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2103:     PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2104:     PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2105:     PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2106:     PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2107:     PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2108:     PetscCall(MatHeaderMerge(Y, &B));
2109:     PetscCall(PetscFree(nnz_d));
2110:     PetscCall(PetscFree(nnz_o));
2111:   }
2112:   PetscFunctionReturn(PETSC_SUCCESS);
2113: }

2115: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);

2117: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2118: {
2119:   PetscFunctionBegin;
2120:   if (PetscDefined(USE_COMPLEX)) {
2121:     Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2123:     PetscCall(MatConjugate_SeqAIJ(aij->A));
2124:     PetscCall(MatConjugate_SeqAIJ(aij->B));
2125:   }
2126:   PetscFunctionReturn(PETSC_SUCCESS);
2127: }

2129: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2130: {
2131:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2133:   PetscFunctionBegin;
2134:   PetscCall(MatRealPart(a->A));
2135:   PetscCall(MatRealPart(a->B));
2136:   PetscFunctionReturn(PETSC_SUCCESS);
2137: }

2139: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2140: {
2141:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2143:   PetscFunctionBegin;
2144:   PetscCall(MatImaginaryPart(a->A));
2145:   PetscCall(MatImaginaryPart(a->B));
2146:   PetscFunctionReturn(PETSC_SUCCESS);
2147: }

2149: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2150: {
2151:   Mat_MPIAIJ        *a = (Mat_MPIAIJ *)A->data;
2152:   PetscInt           i, *idxb = NULL, m = A->rmap->n;
2153:   PetscScalar       *va, *vv;
2154:   Vec                vB, vA;
2155:   const PetscScalar *vb;

2157:   PetscFunctionBegin;
2158:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2159:   PetscCall(MatGetRowMaxAbs(a->A, vA, idx));

2161:   PetscCall(VecGetArrayWrite(vA, &va));
2162:   if (idx) {
2163:     for (i = 0; i < m; i++) {
2164:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2165:     }
2166:   }

2168:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2169:   PetscCall(PetscMalloc1(m, &idxb));
2170:   PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));

2172:   PetscCall(VecGetArrayWrite(v, &vv));
2173:   PetscCall(VecGetArrayRead(vB, &vb));
2174:   for (i = 0; i < m; i++) {
2175:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2176:       vv[i] = vb[i];
2177:       if (idx) idx[i] = a->garray[idxb[i]];
2178:     } else {
2179:       vv[i] = va[i];
2180:       if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2181:     }
2182:   }
2183:   PetscCall(VecRestoreArrayWrite(vA, &vv));
2184:   PetscCall(VecRestoreArrayWrite(vA, &va));
2185:   PetscCall(VecRestoreArrayRead(vB, &vb));
2186:   PetscCall(PetscFree(idxb));
2187:   PetscCall(VecDestroy(&vA));
2188:   PetscCall(VecDestroy(&vB));
2189:   PetscFunctionReturn(PETSC_SUCCESS);
2190: }

2192: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2193: {
2194:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2195:   PetscInt    m = A->rmap->n;
2196:   Vec         vB, vA;

2198:   PetscFunctionBegin;
2199:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2200:   PetscCall(MatGetRowSumAbs(a->A, vA));
2201:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2202:   PetscCall(MatGetRowSumAbs(a->B, vB));
2203:   PetscCall(VecAXPY(vA, 1.0, vB));
2204:   PetscCall(VecDestroy(&vB));
2205:   PetscCall(VecCopy(vA, v));
2206:   PetscCall(VecDestroy(&vA));
2207:   PetscFunctionReturn(PETSC_SUCCESS);
2208: }

2210: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2211: {
2212:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2213:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2214:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2215:   PetscInt          *cmap = mat->garray;
2216:   PetscInt          *diagIdx, *offdiagIdx;
2217:   Vec                diagV, offdiagV;
2218:   PetscScalar       *a, *diagA, *offdiagA;
2219:   const PetscScalar *ba, *bav;
2220:   PetscInt           r, j, col, ncols, *bi, *bj;
2221:   Mat                B = mat->B;
2222:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2224:   PetscFunctionBegin;
2225:   /* When a process holds entire A and other processes have no entry */
2226:   if (A->cmap->N == n) {
2227:     PetscCall(VecGetArrayWrite(v, &diagA));
2228:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2229:     PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2230:     PetscCall(VecDestroy(&diagV));
2231:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2232:     PetscFunctionReturn(PETSC_SUCCESS);
2233:   } else if (n == 0) {
2234:     if (m) {
2235:       PetscCall(VecGetArrayWrite(v, &a));
2236:       for (r = 0; r < m; r++) {
2237:         a[r] = 0.0;
2238:         if (idx) idx[r] = -1;
2239:       }
2240:       PetscCall(VecRestoreArrayWrite(v, &a));
2241:     }
2242:     PetscFunctionReturn(PETSC_SUCCESS);
2243:   }

2245:   PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2246:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2247:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2248:   PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));

2250:   /* Get offdiagIdx[] for implicit 0.0 */
2251:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2252:   ba = bav;
2253:   bi = b->i;
2254:   bj = b->j;
2255:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2256:   for (r = 0; r < m; r++) {
2257:     ncols = bi[r + 1] - bi[r];
2258:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2259:       offdiagA[r]   = *ba;
2260:       offdiagIdx[r] = cmap[0];
2261:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2262:       offdiagA[r] = 0.0;

2264:       /* Find first hole in the cmap */
2265:       for (j = 0; j < ncols; j++) {
2266:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2267:         if (col > j && j < cstart) {
2268:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2269:           break;
2270:         } else if (col > j + n && j >= cstart) {
2271:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2272:           break;
2273:         }
2274:       }
2275:       if (j == ncols && ncols < A->cmap->N - n) {
2276:         /* a hole is outside compressed Bcols */
2277:         if (ncols == 0) {
2278:           if (cstart) {
2279:             offdiagIdx[r] = 0;
2280:           } else offdiagIdx[r] = cend;
2281:         } else { /* ncols > 0 */
2282:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2283:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2284:         }
2285:       }
2286:     }

2288:     for (j = 0; j < ncols; j++) {
2289:       if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2290:         offdiagA[r]   = *ba;
2291:         offdiagIdx[r] = cmap[*bj];
2292:       }
2293:       ba++;
2294:       bj++;
2295:     }
2296:   }

2298:   PetscCall(VecGetArrayWrite(v, &a));
2299:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2300:   for (r = 0; r < m; ++r) {
2301:     if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2302:       a[r] = diagA[r];
2303:       if (idx) idx[r] = cstart + diagIdx[r];
2304:     } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2305:       a[r] = diagA[r];
2306:       if (idx) {
2307:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2308:           idx[r] = cstart + diagIdx[r];
2309:         } else idx[r] = offdiagIdx[r];
2310:       }
2311:     } else {
2312:       a[r] = offdiagA[r];
2313:       if (idx) idx[r] = offdiagIdx[r];
2314:     }
2315:   }
2316:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2317:   PetscCall(VecRestoreArrayWrite(v, &a));
2318:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2319:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2320:   PetscCall(VecDestroy(&diagV));
2321:   PetscCall(VecDestroy(&offdiagV));
2322:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2323:   PetscFunctionReturn(PETSC_SUCCESS);
2324: }

2326: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2327: {
2328:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2329:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2330:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2331:   PetscInt          *cmap = mat->garray;
2332:   PetscInt          *diagIdx, *offdiagIdx;
2333:   Vec                diagV, offdiagV;
2334:   PetscScalar       *a, *diagA, *offdiagA;
2335:   const PetscScalar *ba, *bav;
2336:   PetscInt           r, j, col, ncols, *bi, *bj;
2337:   Mat                B = mat->B;
2338:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2340:   PetscFunctionBegin;
2341:   /* When a process holds entire A and other processes have no entry */
2342:   if (A->cmap->N == n) {
2343:     PetscCall(VecGetArrayWrite(v, &diagA));
2344:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2345:     PetscCall(MatGetRowMin(mat->A, diagV, idx));
2346:     PetscCall(VecDestroy(&diagV));
2347:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2348:     PetscFunctionReturn(PETSC_SUCCESS);
2349:   } else if (n == 0) {
2350:     if (m) {
2351:       PetscCall(VecGetArrayWrite(v, &a));
2352:       for (r = 0; r < m; r++) {
2353:         a[r] = PETSC_MAX_REAL;
2354:         if (idx) idx[r] = -1;
2355:       }
2356:       PetscCall(VecRestoreArrayWrite(v, &a));
2357:     }
2358:     PetscFunctionReturn(PETSC_SUCCESS);
2359:   }

2361:   PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2362:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2363:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2364:   PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));

2366:   /* Get offdiagIdx[] for implicit 0.0 */
2367:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2368:   ba = bav;
2369:   bi = b->i;
2370:   bj = b->j;
2371:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2372:   for (r = 0; r < m; r++) {
2373:     ncols = bi[r + 1] - bi[r];
2374:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2375:       offdiagA[r]   = *ba;
2376:       offdiagIdx[r] = cmap[0];
2377:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2378:       offdiagA[r] = 0.0;

2380:       /* Find first hole in the cmap */
2381:       for (j = 0; j < ncols; j++) {
2382:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2383:         if (col > j && j < cstart) {
2384:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2385:           break;
2386:         } else if (col > j + n && j >= cstart) {
2387:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2388:           break;
2389:         }
2390:       }
2391:       if (j == ncols && ncols < A->cmap->N - n) {
2392:         /* a hole is outside compressed Bcols */
2393:         if (ncols == 0) {
2394:           if (cstart) {
2395:             offdiagIdx[r] = 0;
2396:           } else offdiagIdx[r] = cend;
2397:         } else { /* ncols > 0 */
2398:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2399:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2400:         }
2401:       }
2402:     }

2404:     for (j = 0; j < ncols; j++) {
2405:       if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2406:         offdiagA[r]   = *ba;
2407:         offdiagIdx[r] = cmap[*bj];
2408:       }
2409:       ba++;
2410:       bj++;
2411:     }
2412:   }

2414:   PetscCall(VecGetArrayWrite(v, &a));
2415:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2416:   for (r = 0; r < m; ++r) {
2417:     if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2418:       a[r] = diagA[r];
2419:       if (idx) idx[r] = cstart + diagIdx[r];
2420:     } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2421:       a[r] = diagA[r];
2422:       if (idx) {
2423:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2424:           idx[r] = cstart + diagIdx[r];
2425:         } else idx[r] = offdiagIdx[r];
2426:       }
2427:     } else {
2428:       a[r] = offdiagA[r];
2429:       if (idx) idx[r] = offdiagIdx[r];
2430:     }
2431:   }
2432:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2433:   PetscCall(VecRestoreArrayWrite(v, &a));
2434:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2435:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2436:   PetscCall(VecDestroy(&diagV));
2437:   PetscCall(VecDestroy(&offdiagV));
2438:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2439:   PetscFunctionReturn(PETSC_SUCCESS);
2440: }

2442: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2443: {
2444:   Mat_MPIAIJ        *mat = (Mat_MPIAIJ *)A->data;
2445:   PetscInt           m = A->rmap->n, n = A->cmap->n;
2446:   PetscInt           cstart = A->cmap->rstart, cend = A->cmap->rend;
2447:   PetscInt          *cmap = mat->garray;
2448:   PetscInt          *diagIdx, *offdiagIdx;
2449:   Vec                diagV, offdiagV;
2450:   PetscScalar       *a, *diagA, *offdiagA;
2451:   const PetscScalar *ba, *bav;
2452:   PetscInt           r, j, col, ncols, *bi, *bj;
2453:   Mat                B = mat->B;
2454:   Mat_SeqAIJ        *b = (Mat_SeqAIJ *)B->data;

2456:   PetscFunctionBegin;
2457:   /* When a process holds entire A and other processes have no entry */
2458:   if (A->cmap->N == n) {
2459:     PetscCall(VecGetArrayWrite(v, &diagA));
2460:     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2461:     PetscCall(MatGetRowMax(mat->A, diagV, idx));
2462:     PetscCall(VecDestroy(&diagV));
2463:     PetscCall(VecRestoreArrayWrite(v, &diagA));
2464:     PetscFunctionReturn(PETSC_SUCCESS);
2465:   } else if (n == 0) {
2466:     if (m) {
2467:       PetscCall(VecGetArrayWrite(v, &a));
2468:       for (r = 0; r < m; r++) {
2469:         a[r] = PETSC_MIN_REAL;
2470:         if (idx) idx[r] = -1;
2471:       }
2472:       PetscCall(VecRestoreArrayWrite(v, &a));
2473:     }
2474:     PetscFunctionReturn(PETSC_SUCCESS);
2475:   }

2477:   PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2478:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2479:   PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2480:   PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));

2482:   /* Get offdiagIdx[] for implicit 0.0 */
2483:   PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2484:   ba = bav;
2485:   bi = b->i;
2486:   bj = b->j;
2487:   PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2488:   for (r = 0; r < m; r++) {
2489:     ncols = bi[r + 1] - bi[r];
2490:     if (ncols == A->cmap->N - n) { /* Brow is dense */
2491:       offdiagA[r]   = *ba;
2492:       offdiagIdx[r] = cmap[0];
2493:     } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2494:       offdiagA[r] = 0.0;

2496:       /* Find first hole in the cmap */
2497:       for (j = 0; j < ncols; j++) {
2498:         col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2499:         if (col > j && j < cstart) {
2500:           offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2501:           break;
2502:         } else if (col > j + n && j >= cstart) {
2503:           offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2504:           break;
2505:         }
2506:       }
2507:       if (j == ncols && ncols < A->cmap->N - n) {
2508:         /* a hole is outside compressed Bcols */
2509:         if (ncols == 0) {
2510:           if (cstart) {
2511:             offdiagIdx[r] = 0;
2512:           } else offdiagIdx[r] = cend;
2513:         } else { /* ncols > 0 */
2514:           offdiagIdx[r] = cmap[ncols - 1] + 1;
2515:           if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2516:         }
2517:       }
2518:     }

2520:     for (j = 0; j < ncols; j++) {
2521:       if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2522:         offdiagA[r]   = *ba;
2523:         offdiagIdx[r] = cmap[*bj];
2524:       }
2525:       ba++;
2526:       bj++;
2527:     }
2528:   }

2530:   PetscCall(VecGetArrayWrite(v, &a));
2531:   PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2532:   for (r = 0; r < m; ++r) {
2533:     if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2534:       a[r] = diagA[r];
2535:       if (idx) idx[r] = cstart + diagIdx[r];
2536:     } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2537:       a[r] = diagA[r];
2538:       if (idx) {
2539:         if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2540:           idx[r] = cstart + diagIdx[r];
2541:         } else idx[r] = offdiagIdx[r];
2542:       }
2543:     } else {
2544:       a[r] = offdiagA[r];
2545:       if (idx) idx[r] = offdiagIdx[r];
2546:     }
2547:   }
2548:   PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2549:   PetscCall(VecRestoreArrayWrite(v, &a));
2550:   PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2551:   PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2552:   PetscCall(VecDestroy(&diagV));
2553:   PetscCall(VecDestroy(&offdiagV));
2554:   PetscCall(PetscFree2(diagIdx, offdiagIdx));
2555:   PetscFunctionReturn(PETSC_SUCCESS);
2556: }

2558: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2559: {
2560:   Mat *dummy;

2562:   PetscFunctionBegin;
2563:   PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2564:   *newmat = *dummy;
2565:   PetscCall(PetscFree(dummy));
2566:   PetscFunctionReturn(PETSC_SUCCESS);
2567: }

2569: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2570: {
2571:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2573:   PetscFunctionBegin;
2574:   PetscCall(MatInvertBlockDiagonal(a->A, values));
2575:   A->factorerrortype = a->A->factorerrortype;
2576:   PetscFunctionReturn(PETSC_SUCCESS);
2577: }

2579: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2580: {
2581:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;

2583:   PetscFunctionBegin;
2584:   PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2585:   PetscCall(MatSetRandom(aij->A, rctx));
2586:   if (x->assembled) {
2587:     PetscCall(MatSetRandom(aij->B, rctx));
2588:   } else {
2589:     PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2590:   }
2591:   PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2592:   PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2593:   PetscFunctionReturn(PETSC_SUCCESS);
2594: }

2596: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2597: {
2598:   PetscFunctionBegin;
2599:   if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2600:   else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2601:   PetscFunctionReturn(PETSC_SUCCESS);
2602: }

2604: /*@
2605:   MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank

2607:   Not Collective

2609:   Input Parameter:
2610: . A - the matrix

2612:   Output Parameter:
2613: . nz - the number of nonzeros

2615:   Level: advanced

2617: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2618: @*/
2619: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2620: {
2621:   Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2622:   Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2623:   PetscBool   isaij;

2625:   PetscFunctionBegin;
2626:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2627:   PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2628:   *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2629:   PetscFunctionReturn(PETSC_SUCCESS);
2630: }

2632: /*@
2633:   MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap

2635:   Collective

2637:   Input Parameters:
2638: + A  - the matrix
2639: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)

2641:   Level: advanced

2643: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2644: @*/
2645: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2646: {
2647:   PetscFunctionBegin;
2648:   PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2649:   PetscFunctionReturn(PETSC_SUCCESS);
2650: }

2652: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2653: {
2654:   PetscBool sc = PETSC_FALSE, flg;

2656:   PetscFunctionBegin;
2657:   PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2658:   if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2659:   PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2660:   if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2661:   PetscOptionsHeadEnd();
2662:   PetscFunctionReturn(PETSC_SUCCESS);
2663: }

2665: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2666: {
2667:   Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2668:   Mat_SeqAIJ *aij  = (Mat_SeqAIJ *)maij->A->data;

2670:   PetscFunctionBegin;
2671:   if (!Y->preallocated) {
2672:     PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2673:   } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2674:     PetscInt nonew = aij->nonew;
2675:     PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2676:     aij->nonew = nonew;
2677:   }
2678:   PetscCall(MatShift_Basic(Y, a));
2679:   PetscFunctionReturn(PETSC_SUCCESS);
2680: }

2682: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2683: {
2684:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2686:   PetscFunctionBegin;
2687:   PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2688:   PetscCall(MatMissingDiagonal(a->A, missing, d));
2689:   if (d) {
2690:     PetscInt rstart;
2691:     PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2692:     *d += rstart;
2693:   }
2694:   PetscFunctionReturn(PETSC_SUCCESS);
2695: }

2697: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2698: {
2699:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2701:   PetscFunctionBegin;
2702:   PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2703:   PetscFunctionReturn(PETSC_SUCCESS);
2704: }

2706: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2707: {
2708:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2710:   PetscFunctionBegin;
2711:   PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep));        // possibly keep zero diagonal coefficients
2712:   PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2713:   PetscFunctionReturn(PETSC_SUCCESS);
2714: }

2716: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2717:                                        MatGetRow_MPIAIJ,
2718:                                        MatRestoreRow_MPIAIJ,
2719:                                        MatMult_MPIAIJ,
2720:                                        /* 4*/ MatMultAdd_MPIAIJ,
2721:                                        MatMultTranspose_MPIAIJ,
2722:                                        MatMultTransposeAdd_MPIAIJ,
2723:                                        NULL,
2724:                                        NULL,
2725:                                        NULL,
2726:                                        /*10*/ NULL,
2727:                                        NULL,
2728:                                        NULL,
2729:                                        MatSOR_MPIAIJ,
2730:                                        MatTranspose_MPIAIJ,
2731:                                        /*15*/ MatGetInfo_MPIAIJ,
2732:                                        MatEqual_MPIAIJ,
2733:                                        MatGetDiagonal_MPIAIJ,
2734:                                        MatDiagonalScale_MPIAIJ,
2735:                                        MatNorm_MPIAIJ,
2736:                                        /*20*/ MatAssemblyBegin_MPIAIJ,
2737:                                        MatAssemblyEnd_MPIAIJ,
2738:                                        MatSetOption_MPIAIJ,
2739:                                        MatZeroEntries_MPIAIJ,
2740:                                        /*24*/ MatZeroRows_MPIAIJ,
2741:                                        NULL,
2742:                                        NULL,
2743:                                        NULL,
2744:                                        NULL,
2745:                                        /*29*/ MatSetUp_MPI_Hash,
2746:                                        NULL,
2747:                                        NULL,
2748:                                        MatGetDiagonalBlock_MPIAIJ,
2749:                                        NULL,
2750:                                        /*34*/ MatDuplicate_MPIAIJ,
2751:                                        NULL,
2752:                                        NULL,
2753:                                        NULL,
2754:                                        NULL,
2755:                                        /*39*/ MatAXPY_MPIAIJ,
2756:                                        MatCreateSubMatrices_MPIAIJ,
2757:                                        MatIncreaseOverlap_MPIAIJ,
2758:                                        MatGetValues_MPIAIJ,
2759:                                        MatCopy_MPIAIJ,
2760:                                        /*44*/ MatGetRowMax_MPIAIJ,
2761:                                        MatScale_MPIAIJ,
2762:                                        MatShift_MPIAIJ,
2763:                                        MatDiagonalSet_MPIAIJ,
2764:                                        MatZeroRowsColumns_MPIAIJ,
2765:                                        /*49*/ MatSetRandom_MPIAIJ,
2766:                                        MatGetRowIJ_MPIAIJ,
2767:                                        MatRestoreRowIJ_MPIAIJ,
2768:                                        NULL,
2769:                                        NULL,
2770:                                        /*54*/ MatFDColoringCreate_MPIXAIJ,
2771:                                        NULL,
2772:                                        MatSetUnfactored_MPIAIJ,
2773:                                        MatPermute_MPIAIJ,
2774:                                        NULL,
2775:                                        /*59*/ MatCreateSubMatrix_MPIAIJ,
2776:                                        MatDestroy_MPIAIJ,
2777:                                        MatView_MPIAIJ,
2778:                                        NULL,
2779:                                        NULL,
2780:                                        /*64*/ NULL,
2781:                                        MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2782:                                        NULL,
2783:                                        NULL,
2784:                                        NULL,
2785:                                        /*69*/ MatGetRowMaxAbs_MPIAIJ,
2786:                                        MatGetRowMinAbs_MPIAIJ,
2787:                                        NULL,
2788:                                        NULL,
2789:                                        NULL,
2790:                                        NULL,
2791:                                        /*75*/ MatFDColoringApply_AIJ,
2792:                                        MatSetFromOptions_MPIAIJ,
2793:                                        NULL,
2794:                                        NULL,
2795:                                        MatFindZeroDiagonals_MPIAIJ,
2796:                                        /*80*/ NULL,
2797:                                        NULL,
2798:                                        NULL,
2799:                                        /*83*/ MatLoad_MPIAIJ,
2800:                                        NULL,
2801:                                        NULL,
2802:                                        NULL,
2803:                                        NULL,
2804:                                        NULL,
2805:                                        /*89*/ NULL,
2806:                                        NULL,
2807:                                        MatMatMultNumeric_MPIAIJ_MPIAIJ,
2808:                                        NULL,
2809:                                        NULL,
2810:                                        /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2811:                                        NULL,
2812:                                        NULL,
2813:                                        NULL,
2814:                                        MatBindToCPU_MPIAIJ,
2815:                                        /*99*/ MatProductSetFromOptions_MPIAIJ,
2816:                                        NULL,
2817:                                        NULL,
2818:                                        MatConjugate_MPIAIJ,
2819:                                        NULL,
2820:                                        /*104*/ MatSetValuesRow_MPIAIJ,
2821:                                        MatRealPart_MPIAIJ,
2822:                                        MatImaginaryPart_MPIAIJ,
2823:                                        NULL,
2824:                                        NULL,
2825:                                        /*109*/ NULL,
2826:                                        NULL,
2827:                                        MatGetRowMin_MPIAIJ,
2828:                                        NULL,
2829:                                        MatMissingDiagonal_MPIAIJ,
2830:                                        /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2831:                                        NULL,
2832:                                        MatGetGhosts_MPIAIJ,
2833:                                        NULL,
2834:                                        NULL,
2835:                                        /*119*/ MatMultDiagonalBlock_MPIAIJ,
2836:                                        NULL,
2837:                                        NULL,
2838:                                        NULL,
2839:                                        MatGetMultiProcBlock_MPIAIJ,
2840:                                        /*124*/ MatFindNonzeroRows_MPIAIJ,
2841:                                        MatGetColumnReductions_MPIAIJ,
2842:                                        MatInvertBlockDiagonal_MPIAIJ,
2843:                                        MatInvertVariableBlockDiagonal_MPIAIJ,
2844:                                        MatCreateSubMatricesMPI_MPIAIJ,
2845:                                        /*129*/ NULL,
2846:                                        NULL,
2847:                                        NULL,
2848:                                        MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2849:                                        NULL,
2850:                                        /*134*/ NULL,
2851:                                        NULL,
2852:                                        NULL,
2853:                                        NULL,
2854:                                        NULL,
2855:                                        /*139*/ MatSetBlockSizes_MPIAIJ,
2856:                                        NULL,
2857:                                        NULL,
2858:                                        MatFDColoringSetUp_MPIXAIJ,
2859:                                        MatFindOffBlockDiagonalEntries_MPIAIJ,
2860:                                        MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2861:                                        /*145*/ NULL,
2862:                                        NULL,
2863:                                        NULL,
2864:                                        MatCreateGraph_Simple_AIJ,
2865:                                        NULL,
2866:                                        /*150*/ NULL,
2867:                                        MatEliminateZeros_MPIAIJ,
2868:                                        MatGetRowSumAbs_MPIAIJ};

2870: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2871: {
2872:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2874:   PetscFunctionBegin;
2875:   PetscCall(MatStoreValues(aij->A));
2876:   PetscCall(MatStoreValues(aij->B));
2877:   PetscFunctionReturn(PETSC_SUCCESS);
2878: }

2880: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2881: {
2882:   Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;

2884:   PetscFunctionBegin;
2885:   PetscCall(MatRetrieveValues(aij->A));
2886:   PetscCall(MatRetrieveValues(aij->B));
2887:   PetscFunctionReturn(PETSC_SUCCESS);
2888: }

2890: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2891: {
2892:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2893:   PetscMPIInt size;

2895:   PetscFunctionBegin;
2896:   if (B->hash_active) {
2897:     B->ops[0]      = b->cops;
2898:     B->hash_active = PETSC_FALSE;
2899:   }
2900:   PetscCall(PetscLayoutSetUp(B->rmap));
2901:   PetscCall(PetscLayoutSetUp(B->cmap));

2903: #if defined(PETSC_USE_CTABLE)
2904:   PetscCall(PetscHMapIDestroy(&b->colmap));
2905: #else
2906:   PetscCall(PetscFree(b->colmap));
2907: #endif
2908:   PetscCall(PetscFree(b->garray));
2909:   PetscCall(VecDestroy(&b->lvec));
2910:   PetscCall(VecScatterDestroy(&b->Mvctx));

2912:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));

2914:   MatSeqXAIJGetOptions_Private(b->B);
2915:   PetscCall(MatDestroy(&b->B));
2916:   PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2917:   PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2918:   PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2919:   PetscCall(MatSetType(b->B, MATSEQAIJ));
2920:   MatSeqXAIJRestoreOptions_Private(b->B);

2922:   MatSeqXAIJGetOptions_Private(b->A);
2923:   PetscCall(MatDestroy(&b->A));
2924:   PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2925:   PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2926:   PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2927:   PetscCall(MatSetType(b->A, MATSEQAIJ));
2928:   MatSeqXAIJRestoreOptions_Private(b->A);

2930:   PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2931:   PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2932:   B->preallocated  = PETSC_TRUE;
2933:   B->was_assembled = PETSC_FALSE;
2934:   B->assembled     = PETSC_FALSE;
2935:   PetscFunctionReturn(PETSC_SUCCESS);
2936: }

2938: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2939: {
2940:   Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;

2942:   PetscFunctionBegin;
2944:   PetscCall(PetscLayoutSetUp(B->rmap));
2945:   PetscCall(PetscLayoutSetUp(B->cmap));

2947: #if defined(PETSC_USE_CTABLE)
2948:   PetscCall(PetscHMapIDestroy(&b->colmap));
2949: #else
2950:   PetscCall(PetscFree(b->colmap));
2951: #endif
2952:   PetscCall(PetscFree(b->garray));
2953:   PetscCall(VecDestroy(&b->lvec));
2954:   PetscCall(VecScatterDestroy(&b->Mvctx));

2956:   PetscCall(MatResetPreallocation(b->A));
2957:   PetscCall(MatResetPreallocation(b->B));
2958:   B->preallocated  = PETSC_TRUE;
2959:   B->was_assembled = PETSC_FALSE;
2960:   B->assembled     = PETSC_FALSE;
2961:   PetscFunctionReturn(PETSC_SUCCESS);
2962: }

2964: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2965: {
2966:   Mat         mat;
2967:   Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;

2969:   PetscFunctionBegin;
2970:   *newmat = NULL;
2971:   PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2972:   PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2973:   PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2974:   PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2975:   a = (Mat_MPIAIJ *)mat->data;

2977:   mat->factortype = matin->factortype;
2978:   mat->assembled  = matin->assembled;
2979:   mat->insertmode = NOT_SET_VALUES;

2981:   a->size         = oldmat->size;
2982:   a->rank         = oldmat->rank;
2983:   a->donotstash   = oldmat->donotstash;
2984:   a->roworiented  = oldmat->roworiented;
2985:   a->rowindices   = NULL;
2986:   a->rowvalues    = NULL;
2987:   a->getrowactive = PETSC_FALSE;

2989:   PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2990:   PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2991:   if (matin->hash_active) {
2992:     PetscCall(MatSetUp(mat));
2993:   } else {
2994:     mat->preallocated = matin->preallocated;
2995:     if (oldmat->colmap) {
2996: #if defined(PETSC_USE_CTABLE)
2997:       PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2998: #else
2999:       PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3000:       PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3001: #endif
3002:     } else a->colmap = NULL;
3003:     if (oldmat->garray) {
3004:       PetscInt len;
3005:       len = oldmat->B->cmap->n;
3006:       PetscCall(PetscMalloc1(len + 1, &a->garray));
3007:       if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3008:     } else a->garray = NULL;

3010:     /* It may happen MatDuplicate is called with a non-assembled matrix
3011:       In fact, MatDuplicate only requires the matrix to be preallocated
3012:       This may happen inside a DMCreateMatrix_Shell */
3013:     if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3014:     if (oldmat->Mvctx) PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3015:     PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3016:     PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3017:   }
3018:   PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3019:   *newmat = mat;
3020:   PetscFunctionReturn(PETSC_SUCCESS);
3021: }

3023: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3024: {
3025:   PetscBool isbinary, ishdf5;

3027:   PetscFunctionBegin;
3030:   /* force binary viewer to load .info file if it has not yet done so */
3031:   PetscCall(PetscViewerSetUp(viewer));
3032:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3033:   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3034:   if (isbinary) {
3035:     PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3036:   } else if (ishdf5) {
3037: #if defined(PETSC_HAVE_HDF5)
3038:     PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3039: #else
3040:     SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3041: #endif
3042:   } else {
3043:     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);
3044:   }
3045:   PetscFunctionReturn(PETSC_SUCCESS);
3046: }

3048: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3049: {
3050:   PetscInt     header[4], M, N, m, nz, rows, cols, sum, i;
3051:   PetscInt    *rowidxs, *colidxs;
3052:   PetscScalar *matvals;

3054:   PetscFunctionBegin;
3055:   PetscCall(PetscViewerSetUp(viewer));

3057:   /* read in matrix header */
3058:   PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3059:   PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3060:   M  = header[1];
3061:   N  = header[2];
3062:   nz = header[3];
3063:   PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3064:   PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3065:   PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");

3067:   /* set block sizes from the viewer's .info file */
3068:   PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3069:   /* set global sizes if not set already */
3070:   if (mat->rmap->N < 0) mat->rmap->N = M;
3071:   if (mat->cmap->N < 0) mat->cmap->N = N;
3072:   PetscCall(PetscLayoutSetUp(mat->rmap));
3073:   PetscCall(PetscLayoutSetUp(mat->cmap));

3075:   /* check if the matrix sizes are correct */
3076:   PetscCall(MatGetSize(mat, &rows, &cols));
3077:   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);

3079:   /* read in row lengths and build row indices */
3080:   PetscCall(MatGetLocalSize(mat, &m, NULL));
3081:   PetscCall(PetscMalloc1(m + 1, &rowidxs));
3082:   PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3083:   rowidxs[0] = 0;
3084:   for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3085:   if (nz != PETSC_MAX_INT) {
3086:     PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3087:     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);
3088:   }

3090:   /* read in column indices and matrix values */
3091:   PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3092:   PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3093:   PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3094:   /* store matrix indices and values */
3095:   PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3096:   PetscCall(PetscFree(rowidxs));
3097:   PetscCall(PetscFree2(colidxs, matvals));
3098:   PetscFunctionReturn(PETSC_SUCCESS);
3099: }

3101: /* Not scalable because of ISAllGather() unless getting all columns. */
3102: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3103: {
3104:   IS          iscol_local;
3105:   PetscBool   isstride;
3106:   PetscMPIInt lisstride = 0, gisstride;

3108:   PetscFunctionBegin;
3109:   /* check if we are grabbing all columns*/
3110:   PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));

3112:   if (isstride) {
3113:     PetscInt start, len, mstart, mlen;
3114:     PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3115:     PetscCall(ISGetLocalSize(iscol, &len));
3116:     PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3117:     if (mstart == start && mlen - mstart == len) lisstride = 1;
3118:   }

3120:   PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3121:   if (gisstride) {
3122:     PetscInt N;
3123:     PetscCall(MatGetSize(mat, NULL, &N));
3124:     PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3125:     PetscCall(ISSetIdentity(iscol_local));
3126:     PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3127:   } else {
3128:     PetscInt cbs;
3129:     PetscCall(ISGetBlockSize(iscol, &cbs));
3130:     PetscCall(ISAllGather(iscol, &iscol_local));
3131:     PetscCall(ISSetBlockSize(iscol_local, cbs));
3132:   }

3134:   *isseq = iscol_local;
3135:   PetscFunctionReturn(PETSC_SUCCESS);
3136: }

3138: /*
3139:  Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3140:  (see MatCreateSubMatrix_MPIAIJ_nonscalable)

3142:  Input Parameters:
3143: +   mat - matrix
3144: .   isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3145:            i.e., mat->rstart <= isrow[i] < mat->rend
3146: -   iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3147:            i.e., mat->cstart <= iscol[i] < mat->cend

3149:  Output Parameters:
3150: +   isrow_d - sequential row index set for retrieving mat->A
3151: .   iscol_d - sequential  column index set for retrieving mat->A
3152: .   iscol_o - sequential column index set for retrieving mat->B
3153: -   garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3154:  */
3155: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3156: {
3157:   Vec             x, cmap;
3158:   const PetscInt *is_idx;
3159:   PetscScalar    *xarray, *cmaparray;
3160:   PetscInt        ncols, isstart, *idx, m, rstart, *cmap1, count;
3161:   Mat_MPIAIJ     *a    = (Mat_MPIAIJ *)mat->data;
3162:   Mat             B    = a->B;
3163:   Vec             lvec = a->lvec, lcmap;
3164:   PetscInt        i, cstart, cend, Bn = B->cmap->N;
3165:   MPI_Comm        comm;
3166:   VecScatter      Mvctx = a->Mvctx;

3168:   PetscFunctionBegin;
3169:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3170:   PetscCall(ISGetLocalSize(iscol, &ncols));

3172:   /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3173:   PetscCall(MatCreateVecs(mat, &x, NULL));
3174:   PetscCall(VecSet(x, -1.0));
3175:   PetscCall(VecDuplicate(x, &cmap));
3176:   PetscCall(VecSet(cmap, -1.0));

3178:   /* Get start indices */
3179:   PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3180:   isstart -= ncols;
3181:   PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));

3183:   PetscCall(ISGetIndices(iscol, &is_idx));
3184:   PetscCall(VecGetArray(x, &xarray));
3185:   PetscCall(VecGetArray(cmap, &cmaparray));
3186:   PetscCall(PetscMalloc1(ncols, &idx));
3187:   for (i = 0; i < ncols; i++) {
3188:     xarray[is_idx[i] - cstart]    = (PetscScalar)is_idx[i];
3189:     cmaparray[is_idx[i] - cstart] = i + isstart;        /* global index of iscol[i] */
3190:     idx[i]                        = is_idx[i] - cstart; /* local index of iscol[i]  */
3191:   }
3192:   PetscCall(VecRestoreArray(x, &xarray));
3193:   PetscCall(VecRestoreArray(cmap, &cmaparray));
3194:   PetscCall(ISRestoreIndices(iscol, &is_idx));

3196:   /* Get iscol_d */
3197:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3198:   PetscCall(ISGetBlockSize(iscol, &i));
3199:   PetscCall(ISSetBlockSize(*iscol_d, i));

3201:   /* Get isrow_d */
3202:   PetscCall(ISGetLocalSize(isrow, &m));
3203:   rstart = mat->rmap->rstart;
3204:   PetscCall(PetscMalloc1(m, &idx));
3205:   PetscCall(ISGetIndices(isrow, &is_idx));
3206:   for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3207:   PetscCall(ISRestoreIndices(isrow, &is_idx));

3209:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3210:   PetscCall(ISGetBlockSize(isrow, &i));
3211:   PetscCall(ISSetBlockSize(*isrow_d, i));

3213:   /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3214:   PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3215:   PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));

3217:   PetscCall(VecDuplicate(lvec, &lcmap));

3219:   PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3220:   PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));

3222:   /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3223:   /* off-process column indices */
3224:   count = 0;
3225:   PetscCall(PetscMalloc1(Bn, &idx));
3226:   PetscCall(PetscMalloc1(Bn, &cmap1));

3228:   PetscCall(VecGetArray(lvec, &xarray));
3229:   PetscCall(VecGetArray(lcmap, &cmaparray));
3230:   for (i = 0; i < Bn; i++) {
3231:     if (PetscRealPart(xarray[i]) > -1.0) {
3232:       idx[count]   = i;                                     /* local column index in off-diagonal part B */
3233:       cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3234:       count++;
3235:     }
3236:   }
3237:   PetscCall(VecRestoreArray(lvec, &xarray));
3238:   PetscCall(VecRestoreArray(lcmap, &cmaparray));

3240:   PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3241:   /* cannot ensure iscol_o has same blocksize as iscol! */

3243:   PetscCall(PetscFree(idx));
3244:   *garray = cmap1;

3246:   PetscCall(VecDestroy(&x));
3247:   PetscCall(VecDestroy(&cmap));
3248:   PetscCall(VecDestroy(&lcmap));
3249:   PetscFunctionReturn(PETSC_SUCCESS);
3250: }

3252: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3253: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3254: {
3255:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3256:   Mat         M = NULL;
3257:   MPI_Comm    comm;
3258:   IS          iscol_d, isrow_d, iscol_o;
3259:   Mat         Asub = NULL, Bsub = NULL;
3260:   PetscInt    n;

3262:   PetscFunctionBegin;
3263:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));

3265:   if (call == MAT_REUSE_MATRIX) {
3266:     /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3267:     PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3268:     PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");

3270:     PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3271:     PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");

3273:     PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3274:     PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");

3276:     /* Update diagonal and off-diagonal portions of submat */
3277:     asub = (Mat_MPIAIJ *)(*submat)->data;
3278:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3279:     PetscCall(ISGetLocalSize(iscol_o, &n));
3280:     if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3281:     PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3282:     PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));

3284:   } else { /* call == MAT_INITIAL_MATRIX) */
3285:     const PetscInt *garray;
3286:     PetscInt        BsubN;

3288:     /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3289:     PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));

3291:     /* Create local submatrices Asub and Bsub */
3292:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3293:     PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));

3295:     /* Create submatrix M */
3296:     PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));

3298:     /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3299:     asub = (Mat_MPIAIJ *)M->data;

3301:     PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3302:     n = asub->B->cmap->N;
3303:     if (BsubN > n) {
3304:       /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3305:       const PetscInt *idx;
3306:       PetscInt        i, j, *idx_new, *subgarray = asub->garray;
3307:       PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));

3309:       PetscCall(PetscMalloc1(n, &idx_new));
3310:       j = 0;
3311:       PetscCall(ISGetIndices(iscol_o, &idx));
3312:       for (i = 0; i < n; i++) {
3313:         if (j >= BsubN) break;
3314:         while (subgarray[i] > garray[j]) j++;

3316:         if (subgarray[i] == garray[j]) {
3317:           idx_new[i] = idx[j++];
3318:         } 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]);
3319:       }
3320:       PetscCall(ISRestoreIndices(iscol_o, &idx));

3322:       PetscCall(ISDestroy(&iscol_o));
3323:       PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));

3325:     } else if (BsubN < n) {
3326:       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);
3327:     }

3329:     PetscCall(PetscFree(garray));
3330:     *submat = M;

3332:     /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3333:     PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3334:     PetscCall(ISDestroy(&isrow_d));

3336:     PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3337:     PetscCall(ISDestroy(&iscol_d));

3339:     PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3340:     PetscCall(ISDestroy(&iscol_o));
3341:   }
3342:   PetscFunctionReturn(PETSC_SUCCESS);
3343: }

3345: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3346: {
3347:   IS        iscol_local = NULL, isrow_d;
3348:   PetscInt  csize;
3349:   PetscInt  n, i, j, start, end;
3350:   PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3351:   MPI_Comm  comm;

3353:   PetscFunctionBegin;
3354:   /* If isrow has same processor distribution as mat,
3355:      call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3356:   if (call == MAT_REUSE_MATRIX) {
3357:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3358:     if (isrow_d) {
3359:       sameRowDist  = PETSC_TRUE;
3360:       tsameDist[1] = PETSC_TRUE; /* sameColDist */
3361:     } else {
3362:       PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3363:       if (iscol_local) {
3364:         sameRowDist  = PETSC_TRUE;
3365:         tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3366:       }
3367:     }
3368:   } else {
3369:     /* Check if isrow has same processor distribution as mat */
3370:     sameDist[0] = PETSC_FALSE;
3371:     PetscCall(ISGetLocalSize(isrow, &n));
3372:     if (!n) {
3373:       sameDist[0] = PETSC_TRUE;
3374:     } else {
3375:       PetscCall(ISGetMinMax(isrow, &i, &j));
3376:       PetscCall(MatGetOwnershipRange(mat, &start, &end));
3377:       if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3378:     }

3380:     /* Check if iscol has same processor distribution as mat */
3381:     sameDist[1] = PETSC_FALSE;
3382:     PetscCall(ISGetLocalSize(iscol, &n));
3383:     if (!n) {
3384:       sameDist[1] = PETSC_TRUE;
3385:     } else {
3386:       PetscCall(ISGetMinMax(iscol, &i, &j));
3387:       PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3388:       if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3389:     }

3391:     PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3392:     PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3393:     sameRowDist = tsameDist[0];
3394:   }

3396:   if (sameRowDist) {
3397:     if (tsameDist[1]) { /* sameRowDist & sameColDist */
3398:       /* isrow and iscol have same processor distribution as mat */
3399:       PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3400:       PetscFunctionReturn(PETSC_SUCCESS);
3401:     } else { /* sameRowDist */
3402:       /* isrow has same processor distribution as mat */
3403:       if (call == MAT_INITIAL_MATRIX) {
3404:         PetscBool sorted;
3405:         PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3406:         PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3407:         PetscCall(ISGetSize(iscol, &i));
3408:         PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);

3410:         PetscCall(ISSorted(iscol_local, &sorted));
3411:         if (sorted) {
3412:           /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3413:           PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3414:           PetscFunctionReturn(PETSC_SUCCESS);
3415:         }
3416:       } else { /* call == MAT_REUSE_MATRIX */
3417:         IS iscol_sub;
3418:         PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3419:         if (iscol_sub) {
3420:           PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3421:           PetscFunctionReturn(PETSC_SUCCESS);
3422:         }
3423:       }
3424:     }
3425:   }

3427:   /* General case: iscol -> iscol_local which has global size of iscol */
3428:   if (call == MAT_REUSE_MATRIX) {
3429:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3430:     PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3431:   } else {
3432:     if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3433:   }

3435:   PetscCall(ISGetLocalSize(iscol, &csize));
3436:   PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));

3438:   if (call == MAT_INITIAL_MATRIX) {
3439:     PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3440:     PetscCall(ISDestroy(&iscol_local));
3441:   }
3442:   PetscFunctionReturn(PETSC_SUCCESS);
3443: }

3445: /*@C
3446:   MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3447:   and "off-diagonal" part of the matrix in CSR format.

3449:   Collective

3451:   Input Parameters:
3452: + comm   - MPI communicator
3453: . A      - "diagonal" portion of matrix
3454: . B      - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3455: - garray - global index of `B` columns

3457:   Output Parameter:
3458: . mat - the matrix, with input `A` as its local diagonal matrix

3460:   Level: advanced

3462:   Notes:
3463:   See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.

3465:   `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.

3467: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3468: @*/
3469: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3470: {
3471:   Mat_MPIAIJ        *maij;
3472:   Mat_SeqAIJ        *b  = (Mat_SeqAIJ *)B->data, *bnew;
3473:   PetscInt          *oi = b->i, *oj = b->j, i, nz, col;
3474:   const PetscScalar *oa;
3475:   Mat                Bnew;
3476:   PetscInt           m, n, N;
3477:   MatType            mpi_mat_type;

3479:   PetscFunctionBegin;
3480:   PetscCall(MatCreate(comm, mat));
3481:   PetscCall(MatGetSize(A, &m, &n));
3482:   PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3483:   PetscCheck(PetscAbs(A->rmap->bs) == PetscAbs(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);
3484:   /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3485:   /* 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); */

3487:   /* Get global columns of mat */
3488:   PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));

3490:   PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3491:   /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3492:   PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3493:   PetscCall(MatSetType(*mat, mpi_mat_type));

3495:   if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3496:   maij = (Mat_MPIAIJ *)(*mat)->data;

3498:   (*mat)->preallocated = PETSC_TRUE;

3500:   PetscCall(PetscLayoutSetUp((*mat)->rmap));
3501:   PetscCall(PetscLayoutSetUp((*mat)->cmap));

3503:   /* Set A as diagonal portion of *mat */
3504:   maij->A = A;

3506:   nz = oi[m];
3507:   for (i = 0; i < nz; i++) {
3508:     col   = oj[i];
3509:     oj[i] = garray[col];
3510:   }

3512:   /* Set Bnew as off-diagonal portion of *mat */
3513:   PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3514:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3515:   PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3516:   bnew        = (Mat_SeqAIJ *)Bnew->data;
3517:   bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3518:   maij->B     = Bnew;

3520:   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);

3522:   b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3523:   b->free_a       = PETSC_FALSE;
3524:   b->free_ij      = PETSC_FALSE;
3525:   PetscCall(MatDestroy(&B));

3527:   bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3528:   bnew->free_a       = PETSC_TRUE;
3529:   bnew->free_ij      = PETSC_TRUE;

3531:   /* condense columns of maij->B */
3532:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3533:   PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3534:   PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3535:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3536:   PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3537:   PetscFunctionReturn(PETSC_SUCCESS);
3538: }

3540: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);

3542: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3543: {
3544:   PetscInt        i, m, n, rstart, row, rend, nz, j, bs, cbs;
3545:   PetscInt       *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3546:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)mat->data;
3547:   Mat             M, Msub, B = a->B;
3548:   MatScalar      *aa;
3549:   Mat_SeqAIJ     *aij;
3550:   PetscInt       *garray = a->garray, *colsub, Ncols;
3551:   PetscInt        count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3552:   IS              iscol_sub, iscmap;
3553:   const PetscInt *is_idx, *cmap;
3554:   PetscBool       allcolumns = PETSC_FALSE;
3555:   MPI_Comm        comm;

3557:   PetscFunctionBegin;
3558:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3559:   if (call == MAT_REUSE_MATRIX) {
3560:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3561:     PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3562:     PetscCall(ISGetLocalSize(iscol_sub, &count));

3564:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3565:     PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");

3567:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3568:     PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");

3570:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));

3572:   } else { /* call == MAT_INITIAL_MATRIX) */
3573:     PetscBool flg;

3575:     PetscCall(ISGetLocalSize(iscol, &n));
3576:     PetscCall(ISGetSize(iscol, &Ncols));

3578:     /* (1) iscol -> nonscalable iscol_local */
3579:     /* Check for special case: each processor gets entire matrix columns */
3580:     PetscCall(ISIdentity(iscol_local, &flg));
3581:     if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3582:     PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3583:     if (allcolumns) {
3584:       iscol_sub = iscol_local;
3585:       PetscCall(PetscObjectReference((PetscObject)iscol_local));
3586:       PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));

3588:     } else {
3589:       /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3590:       PetscInt *idx, *cmap1, k;
3591:       PetscCall(PetscMalloc1(Ncols, &idx));
3592:       PetscCall(PetscMalloc1(Ncols, &cmap1));
3593:       PetscCall(ISGetIndices(iscol_local, &is_idx));
3594:       count = 0;
3595:       k     = 0;
3596:       for (i = 0; i < Ncols; i++) {
3597:         j = is_idx[i];
3598:         if (j >= cstart && j < cend) {
3599:           /* diagonal part of mat */
3600:           idx[count]     = j;
3601:           cmap1[count++] = i; /* column index in submat */
3602:         } else if (Bn) {
3603:           /* off-diagonal part of mat */
3604:           if (j == garray[k]) {
3605:             idx[count]     = j;
3606:             cmap1[count++] = i; /* column index in submat */
3607:           } else if (j > garray[k]) {
3608:             while (j > garray[k] && k < Bn - 1) k++;
3609:             if (j == garray[k]) {
3610:               idx[count]     = j;
3611:               cmap1[count++] = i; /* column index in submat */
3612:             }
3613:           }
3614:         }
3615:       }
3616:       PetscCall(ISRestoreIndices(iscol_local, &is_idx));

3618:       PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3619:       PetscCall(ISGetBlockSize(iscol, &cbs));
3620:       PetscCall(ISSetBlockSize(iscol_sub, cbs));

3622:       PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3623:     }

3625:     /* (3) Create sequential Msub */
3626:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3627:   }

3629:   PetscCall(ISGetLocalSize(iscol_sub, &count));
3630:   aij = (Mat_SeqAIJ *)(Msub)->data;
3631:   ii  = aij->i;
3632:   PetscCall(ISGetIndices(iscmap, &cmap));

3634:   /*
3635:       m - number of local rows
3636:       Ncols - number of columns (same on all processors)
3637:       rstart - first row in new global matrix generated
3638:   */
3639:   PetscCall(MatGetSize(Msub, &m, NULL));

3641:   if (call == MAT_INITIAL_MATRIX) {
3642:     /* (4) Create parallel newmat */
3643:     PetscMPIInt rank, size;
3644:     PetscInt    csize;

3646:     PetscCallMPI(MPI_Comm_size(comm, &size));
3647:     PetscCallMPI(MPI_Comm_rank(comm, &rank));

3649:     /*
3650:         Determine the number of non-zeros in the diagonal and off-diagonal
3651:         portions of the matrix in order to do correct preallocation
3652:     */

3654:     /* first get start and end of "diagonal" columns */
3655:     PetscCall(ISGetLocalSize(iscol, &csize));
3656:     if (csize == PETSC_DECIDE) {
3657:       PetscCall(ISGetSize(isrow, &mglobal));
3658:       if (mglobal == Ncols) { /* square matrix */
3659:         nlocal = m;
3660:       } else {
3661:         nlocal = Ncols / size + ((Ncols % size) > rank);
3662:       }
3663:     } else {
3664:       nlocal = csize;
3665:     }
3666:     PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3667:     rstart = rend - nlocal;
3668:     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);

3670:     /* next, compute all the lengths */
3671:     jj = aij->j;
3672:     PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3673:     olens = dlens + m;
3674:     for (i = 0; i < m; i++) {
3675:       jend = ii[i + 1] - ii[i];
3676:       olen = 0;
3677:       dlen = 0;
3678:       for (j = 0; j < jend; j++) {
3679:         if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3680:         else dlen++;
3681:         jj++;
3682:       }
3683:       olens[i] = olen;
3684:       dlens[i] = dlen;
3685:     }

3687:     PetscCall(ISGetBlockSize(isrow, &bs));
3688:     PetscCall(ISGetBlockSize(iscol, &cbs));

3690:     PetscCall(MatCreate(comm, &M));
3691:     PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3692:     PetscCall(MatSetBlockSizes(M, bs, cbs));
3693:     PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3694:     PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3695:     PetscCall(PetscFree(dlens));

3697:   } else { /* call == MAT_REUSE_MATRIX */
3698:     M = *newmat;
3699:     PetscCall(MatGetLocalSize(M, &i, NULL));
3700:     PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3701:     PetscCall(MatZeroEntries(M));
3702:     /*
3703:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3704:        rather than the slower MatSetValues().
3705:     */
3706:     M->was_assembled = PETSC_TRUE;
3707:     M->assembled     = PETSC_FALSE;
3708:   }

3710:   /* (5) Set values of Msub to *newmat */
3711:   PetscCall(PetscMalloc1(count, &colsub));
3712:   PetscCall(MatGetOwnershipRange(M, &rstart, NULL));

3714:   jj = aij->j;
3715:   PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3716:   for (i = 0; i < m; i++) {
3717:     row = rstart + i;
3718:     nz  = ii[i + 1] - ii[i];
3719:     for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3720:     PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3721:     jj += nz;
3722:     aa += nz;
3723:   }
3724:   PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3725:   PetscCall(ISRestoreIndices(iscmap, &cmap));

3727:   PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3728:   PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));

3730:   PetscCall(PetscFree(colsub));

3732:   /* save Msub, iscol_sub and iscmap used in processor for next request */
3733:   if (call == MAT_INITIAL_MATRIX) {
3734:     *newmat = M;
3735:     PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3736:     PetscCall(MatDestroy(&Msub));

3738:     PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3739:     PetscCall(ISDestroy(&iscol_sub));

3741:     PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3742:     PetscCall(ISDestroy(&iscmap));

3744:     if (iscol_local) {
3745:       PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3746:       PetscCall(ISDestroy(&iscol_local));
3747:     }
3748:   }
3749:   PetscFunctionReturn(PETSC_SUCCESS);
3750: }

3752: /*
3753:     Not great since it makes two copies of the submatrix, first an SeqAIJ
3754:   in local and then by concatenating the local matrices the end result.
3755:   Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()

3757:   This requires a sequential iscol with all indices.
3758: */
3759: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3760: {
3761:   PetscMPIInt rank, size;
3762:   PetscInt    i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3763:   PetscInt   *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3764:   Mat         M, Mreuse;
3765:   MatScalar  *aa, *vwork;
3766:   MPI_Comm    comm;
3767:   Mat_SeqAIJ *aij;
3768:   PetscBool   colflag, allcolumns = PETSC_FALSE;

3770:   PetscFunctionBegin;
3771:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3772:   PetscCallMPI(MPI_Comm_rank(comm, &rank));
3773:   PetscCallMPI(MPI_Comm_size(comm, &size));

3775:   /* Check for special case: each processor gets entire matrix columns */
3776:   PetscCall(ISIdentity(iscol, &colflag));
3777:   PetscCall(ISGetLocalSize(iscol, &n));
3778:   if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3779:   PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));

3781:   if (call == MAT_REUSE_MATRIX) {
3782:     PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3783:     PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3784:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3785:   } else {
3786:     PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3787:   }

3789:   /*
3790:       m - number of local rows
3791:       n - number of columns (same on all processors)
3792:       rstart - first row in new global matrix generated
3793:   */
3794:   PetscCall(MatGetSize(Mreuse, &m, &n));
3795:   PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3796:   if (call == MAT_INITIAL_MATRIX) {
3797:     aij = (Mat_SeqAIJ *)(Mreuse)->data;
3798:     ii  = aij->i;
3799:     jj  = aij->j;

3801:     /*
3802:         Determine the number of non-zeros in the diagonal and off-diagonal
3803:         portions of the matrix in order to do correct preallocation
3804:     */

3806:     /* first get start and end of "diagonal" columns */
3807:     if (csize == PETSC_DECIDE) {
3808:       PetscCall(ISGetSize(isrow, &mglobal));
3809:       if (mglobal == n) { /* square matrix */
3810:         nlocal = m;
3811:       } else {
3812:         nlocal = n / size + ((n % size) > rank);
3813:       }
3814:     } else {
3815:       nlocal = csize;
3816:     }
3817:     PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3818:     rstart = rend - nlocal;
3819:     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);

3821:     /* next, compute all the lengths */
3822:     PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3823:     olens = dlens + m;
3824:     for (i = 0; i < m; i++) {
3825:       jend = ii[i + 1] - ii[i];
3826:       olen = 0;
3827:       dlen = 0;
3828:       for (j = 0; j < jend; j++) {
3829:         if (*jj < rstart || *jj >= rend) olen++;
3830:         else dlen++;
3831:         jj++;
3832:       }
3833:       olens[i] = olen;
3834:       dlens[i] = dlen;
3835:     }
3836:     PetscCall(MatCreate(comm, &M));
3837:     PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3838:     PetscCall(MatSetBlockSizes(M, bs, cbs));
3839:     PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3840:     PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3841:     PetscCall(PetscFree(dlens));
3842:   } else {
3843:     PetscInt ml, nl;

3845:     M = *newmat;
3846:     PetscCall(MatGetLocalSize(M, &ml, &nl));
3847:     PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3848:     PetscCall(MatZeroEntries(M));
3849:     /*
3850:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3851:        rather than the slower MatSetValues().
3852:     */
3853:     M->was_assembled = PETSC_TRUE;
3854:     M->assembled     = PETSC_FALSE;
3855:   }
3856:   PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3857:   aij = (Mat_SeqAIJ *)(Mreuse)->data;
3858:   ii  = aij->i;
3859:   jj  = aij->j;

3861:   /* trigger copy to CPU if needed */
3862:   PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3863:   for (i = 0; i < m; i++) {
3864:     row   = rstart + i;
3865:     nz    = ii[i + 1] - ii[i];
3866:     cwork = jj;
3867:     jj    = PetscSafePointerPlusOffset(jj, nz);
3868:     vwork = aa;
3869:     aa    = PetscSafePointerPlusOffset(aa, nz);
3870:     PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3871:   }
3872:   PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));

3874:   PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3875:   PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3876:   *newmat = M;

3878:   /* save submatrix used in processor for next request */
3879:   if (call == MAT_INITIAL_MATRIX) {
3880:     PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3881:     PetscCall(MatDestroy(&Mreuse));
3882:   }
3883:   PetscFunctionReturn(PETSC_SUCCESS);
3884: }

3886: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3887: {
3888:   PetscInt        m, cstart, cend, j, nnz, i, d, *ld;
3889:   PetscInt       *d_nnz, *o_nnz, nnz_max = 0, rstart, ii;
3890:   const PetscInt *JJ;
3891:   PetscBool       nooffprocentries;
3892:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)B->data;

3894:   PetscFunctionBegin;
3895:   PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);

3897:   PetscCall(PetscLayoutSetUp(B->rmap));
3898:   PetscCall(PetscLayoutSetUp(B->cmap));
3899:   m      = B->rmap->n;
3900:   cstart = B->cmap->rstart;
3901:   cend   = B->cmap->rend;
3902:   rstart = B->rmap->rstart;

3904:   PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));

3906:   if (PetscDefined(USE_DEBUG)) {
3907:     for (i = 0; i < m; i++) {
3908:       nnz = Ii[i + 1] - Ii[i];
3909:       JJ  = PetscSafePointerPlusOffset(J, Ii[i]);
3910:       PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3911:       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]);
3912:       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);
3913:     }
3914:   }

3916:   for (i = 0; i < m; i++) {
3917:     nnz     = Ii[i + 1] - Ii[i];
3918:     JJ      = PetscSafePointerPlusOffset(J, Ii[i]);
3919:     nnz_max = PetscMax(nnz_max, nnz);
3920:     d       = 0;
3921:     for (j = 0; j < nnz; j++) {
3922:       if (cstart <= JJ[j] && JJ[j] < cend) d++;
3923:     }
3924:     d_nnz[i] = d;
3925:     o_nnz[i] = nnz - d;
3926:   }
3927:   PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3928:   PetscCall(PetscFree2(d_nnz, o_nnz));

3930:   for (i = 0; i < m; i++) {
3931:     ii = i + rstart;
3932:     PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i]), PetscSafePointerPlusOffset(v, Ii[i]), INSERT_VALUES));
3933:   }
3934:   nooffprocentries    = B->nooffprocentries;
3935:   B->nooffprocentries = PETSC_TRUE;
3936:   PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3937:   PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3938:   B->nooffprocentries = nooffprocentries;

3940:   /* count number of entries below block diagonal */
3941:   PetscCall(PetscFree(Aij->ld));
3942:   PetscCall(PetscCalloc1(m, &ld));
3943:   Aij->ld = ld;
3944:   for (i = 0; i < m; i++) {
3945:     nnz = Ii[i + 1] - Ii[i];
3946:     j   = 0;
3947:     while (j < nnz && J[j] < cstart) j++;
3948:     ld[i] = j;
3949:     if (J) J += nnz;
3950:   }

3952:   PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3953:   PetscFunctionReturn(PETSC_SUCCESS);
3954: }

3956: /*@
3957:   MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3958:   (the default parallel PETSc format).

3960:   Collective

3962:   Input Parameters:
3963: + B - the matrix
3964: . i - the indices into `j` for the start of each local row (indices start with zero)
3965: . j - the column indices for each local row (indices start with zero)
3966: - v - optional values in the matrix

3968:   Level: developer

3970:   Notes:
3971:   The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3972:   thus you CANNOT change the matrix entries by changing the values of `v` after you have
3973:   called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.

3975:   The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.

3977:   A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.

3979:   You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.

3981:   If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3982:   `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.

3984:   The format which is used for the sparse matrix input, is equivalent to a
3985:   row-major ordering.. i.e for the following matrix, the input data expected is
3986:   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: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4005:           `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
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:   Example 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: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4138:           `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
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 (of length m+1); that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4164: . j    - global 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:   Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`

4181:   If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4182:   `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.

4184:   The format which is used for the sparse matrix input, is equivalent to a
4185:   row-major ordering, i.e., for the following matrix, the input data expected is
4186:   as shown
4187: .vb
4188:         1 0 0
4189:         2 0 3     P0
4190:        -------
4191:         4 5 6     P1

4193:      Process0 [P0] rows_owned=[0,1]
4194:         i =  {0,1,3}  [size = nrow+1  = 2+1]
4195:         j =  {0,0,2}  [size = 3]
4196:         v =  {1,2,3}  [size = 3]

4198:      Process1 [P1] rows_owned=[2]
4199:         i =  {0,3}    [size = nrow+1  = 1+1]
4200:         j =  {0,1,2}  [size = 3]
4201:         v =  {4,5,6}  [size = 3]
4202: .ve

4204: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4205:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4206: @*/
4207: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4208: {
4209:   PetscFunctionBegin;
4210:   PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4211:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4212:   PetscCall(MatCreate(comm, mat));
4213:   PetscCall(MatSetSizes(*mat, m, n, M, N));
4214:   /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4215:   PetscCall(MatSetType(*mat, MATMPIAIJ));
4216:   PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4217:   PetscFunctionReturn(PETSC_SUCCESS);
4218: }

4220: /*@
4221:   MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4222:   CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4223:   from `MatCreateMPIAIJWithArrays()`

4225:   Deprecated: Use `MatUpdateMPIAIJWithArray()`

4227:   Collective

4229:   Input Parameters:
4230: + mat - the matrix
4231: . m   - number of local rows (Cannot be `PETSC_DECIDE`)
4232: . n   - This value should be the same as the local size used in creating the
4233:        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4234:        calculated if N is given) For square matrices n is almost always m.
4235: . M   - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4236: . N   - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4237: . Ii  - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4238: . J   - column indices
4239: - v   - matrix values

4241:   Level: deprecated

4243: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4244:           `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4245: @*/
4246: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4247: {
4248:   PetscInt        nnz, i;
4249:   PetscBool       nooffprocentries;
4250:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)mat->data;
4251:   Mat_SeqAIJ     *Ad  = (Mat_SeqAIJ *)Aij->A->data;
4252:   PetscScalar    *ad, *ao;
4253:   PetscInt        ldi, Iii, md;
4254:   const PetscInt *Adi = Ad->i;
4255:   PetscInt       *ld  = Aij->ld;

4257:   PetscFunctionBegin;
4258:   PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4259:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4260:   PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4261:   PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");

4263:   PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4264:   PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));

4266:   for (i = 0; i < m; i++) {
4267:     if (PetscDefined(USE_DEBUG)) {
4268:       for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4269:         PetscCheck(J[j] >= J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is not sorted", j - Ii[i], J[j], i);
4270:         PetscCheck(J[j] != J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is identical to previous entry", j - Ii[i], J[j], i);
4271:       }
4272:     }
4273:     nnz = Ii[i + 1] - Ii[i];
4274:     Iii = Ii[i];
4275:     ldi = ld[i];
4276:     md  = Adi[i + 1] - Adi[i];
4277:     PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4278:     PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4279:     PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4280:     ad += md;
4281:     ao += nnz - md;
4282:   }
4283:   nooffprocentries      = mat->nooffprocentries;
4284:   mat->nooffprocentries = PETSC_TRUE;
4285:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4286:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4287:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4288:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4289:   PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4290:   PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4291:   PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4292:   mat->nooffprocentries = nooffprocentries;
4293:   PetscFunctionReturn(PETSC_SUCCESS);
4294: }

4296: /*@
4297:   MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values

4299:   Collective

4301:   Input Parameters:
4302: + mat - the matrix
4303: - v   - matrix values, stored by row

4305:   Level: intermediate

4307:   Notes:
4308:   The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`

4310:   The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly

4312: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4313:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4314: @*/
4315: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4316: {
4317:   PetscInt        nnz, i, m;
4318:   PetscBool       nooffprocentries;
4319:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *)mat->data;
4320:   Mat_SeqAIJ     *Ad  = (Mat_SeqAIJ *)Aij->A->data;
4321:   Mat_SeqAIJ     *Ao  = (Mat_SeqAIJ *)Aij->B->data;
4322:   PetscScalar    *ad, *ao;
4323:   const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4324:   PetscInt        ldi, Iii, md;
4325:   PetscInt       *ld = Aij->ld;

4327:   PetscFunctionBegin;
4328:   m = mat->rmap->n;

4330:   PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4331:   PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4332:   Iii = 0;
4333:   for (i = 0; i < m; i++) {
4334:     nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4335:     ldi = ld[i];
4336:     md  = Adi[i + 1] - Adi[i];
4337:     PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4338:     ad += md;
4339:     if (ao) {
4340:       PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4341:       PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4342:       ao += nnz - md;
4343:     }
4344:     Iii += nnz;
4345:   }
4346:   nooffprocentries      = mat->nooffprocentries;
4347:   mat->nooffprocentries = PETSC_TRUE;
4348:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4349:   PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4350:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4351:   PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4352:   PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4353:   PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4354:   PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4355:   mat->nooffprocentries = nooffprocentries;
4356:   PetscFunctionReturn(PETSC_SUCCESS);
4357: }

4359: /*@C
4360:   MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4361:   (the default parallel PETSc format).  For good matrix assembly performance
4362:   the user should preallocate the matrix storage by setting the parameters
4363:   `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).

4365:   Collective

4367:   Input Parameters:
4368: + comm  - MPI communicator
4369: . m     - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4370:            This value should be the same as the local size used in creating the
4371:            y vector for the matrix-vector product y = Ax.
4372: . n     - This value should be the same as the local size used in creating the
4373:        x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4374:        calculated if N is given) For square matrices n is almost always m.
4375: . M     - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4376: . N     - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4377: . d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
4378:            (same value is used for all local rows)
4379: . d_nnz - array containing the number of nonzeros in the various rows of the
4380:            DIAGONAL portion of the local submatrix (possibly different for each row)
4381:            or `NULL`, if `d_nz` is used to specify the nonzero structure.
4382:            The size of this array is equal to the number of local rows, i.e 'm'.
4383: . o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
4384:            submatrix (same value is used for all local rows).
4385: - o_nnz - array containing the number of nonzeros in the various rows of the
4386:            OFF-DIAGONAL portion of the local submatrix (possibly different for
4387:            each row) or `NULL`, if `o_nz` is used to specify the nonzero
4388:            structure. The size of this array is equal to the number
4389:            of local rows, i.e 'm'.

4391:   Output Parameter:
4392: . A - the matrix

4394:   Options Database Keys:
4395: + -mat_no_inode                     - Do not use inodes
4396: . -mat_inode_limit <limit>          - Sets inode limit (max limit=5)
4397: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4398:         See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the vecscatter be viewed as a matrix.
4399:         Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.

4401:   Level: intermediate

4403:   Notes:
4404:   It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4405:   MatXXXXSetPreallocation() paradigm instead of this routine directly.
4406:   [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]

4408:   If the *_nnz parameter is given then the *_nz parameter is ignored

4410:   The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4411:   processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4412:   storage requirements for this matrix.

4414:   If `PETSC_DECIDE` or  `PETSC_DETERMINE` is used for a particular argument on one
4415:   processor than it must be used on all processors that share the object for
4416:   that argument.

4418:   The user MUST specify either the local or global matrix dimensions
4419:   (possibly both).

4421:   The parallel matrix is partitioned across processors such that the
4422:   first m0 rows belong to process 0, the next m1 rows belong to
4423:   process 1, the next m2 rows belong to process 2 etc.. where
4424:   m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4425:   values corresponding to [m x N] submatrix.

4427:   The columns are logically partitioned with the n0 columns belonging
4428:   to 0th partition, the next n1 columns belonging to the next
4429:   partition etc.. where n0,n1,n2... are the input parameter 'n'.

4431:   The DIAGONAL portion of the local submatrix on any given processor
4432:   is the submatrix corresponding to the rows and columns m,n
4433:   corresponding to the given processor. i.e diagonal matrix on
4434:   process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4435:   etc. The remaining portion of the local submatrix [m x (N-n)]
4436:   constitute the OFF-DIAGONAL portion. The example below better
4437:   illustrates this concept.

4439:   For a square global matrix we define each processor's diagonal portion
4440:   to be its local rows and the corresponding columns (a square submatrix);
4441:   each processor's off-diagonal portion encompasses the remainder of the
4442:   local matrix (a rectangular submatrix).

4444:   If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.

4446:   When calling this routine with a single process communicator, a matrix of
4447:   type `MATSEQAIJ` is returned.  If a matrix of type `MATMPIAIJ` is desired for this
4448:   type of communicator, use the construction mechanism
4449: .vb
4450:   MatCreate(..., &A);
4451:   MatSetType(A, MATMPIAIJ);
4452:   MatSetSizes(A, m, n, M, N);
4453:   MatMPIAIJSetPreallocation(A, ...);
4454: .ve

4456:   By default, this format uses inodes (identical nodes) when possible.
4457:   We search for consecutive rows with the same nonzero structure, thereby
4458:   reusing matrix information to achieve increased efficiency.

4460:   Example Usage:
4461:   Consider the following 8x8 matrix with 34 non-zero values, that is
4462:   assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4463:   proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4464:   as follows

4466: .vb
4467:             1  2  0  |  0  3  0  |  0  4
4468:     Proc0   0  5  6  |  7  0  0  |  8  0
4469:             9  0 10  | 11  0  0  | 12  0
4470:     -------------------------------------
4471:            13  0 14  | 15 16 17  |  0  0
4472:     Proc1   0 18  0  | 19 20 21  |  0  0
4473:             0  0  0  | 22 23  0  | 24  0
4474:     -------------------------------------
4475:     Proc2  25 26 27  |  0  0 28  | 29  0
4476:            30  0  0  | 31 32 33  |  0 34
4477: .ve

4479:   This can be represented as a collection of submatrices as

4481: .vb
4482:       A B C
4483:       D E F
4484:       G H I
4485: .ve

4487:   Where the submatrices A,B,C are owned by proc0, D,E,F are
4488:   owned by proc1, G,H,I are owned by proc2.

4490:   The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4491:   The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4492:   The 'M','N' parameters are 8,8, and have the same values on all procs.

4494:   The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4495:   submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4496:   corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4497:   Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4498:   part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4499:   matrix, ans [DF] as another SeqAIJ matrix.

4501:   When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4502:   allocated for every row of the local diagonal submatrix, and `o_nz`
4503:   storage locations are allocated for every row of the OFF-DIAGONAL submat.
4504:   One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4505:   rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4506:   In this case, the values of `d_nz`,`o_nz` are
4507: .vb
4508:      proc0  dnz = 2, o_nz = 2
4509:      proc1  dnz = 3, o_nz = 2
4510:      proc2  dnz = 1, o_nz = 4
4511: .ve
4512:   We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4513:   translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4514:   for proc3. i.e we are using 12+15+10=37 storage locations to store
4515:   34 values.

4517:   When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4518:   for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4519:   In the above case the values for d_nnz,o_nnz are
4520: .vb
4521:      proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4522:      proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4523:      proc2 d_nnz = [1,1]   and o_nnz = [4,4]
4524: .ve
4525:   Here the space allocated is sum of all the above values i.e 34, and
4526:   hence pre-allocation is perfect.

4528: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4529:           `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`
4530: @*/
4531: 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)
4532: {
4533:   PetscMPIInt size;

4535:   PetscFunctionBegin;
4536:   PetscCall(MatCreate(comm, A));
4537:   PetscCall(MatSetSizes(*A, m, n, M, N));
4538:   PetscCallMPI(MPI_Comm_size(comm, &size));
4539:   if (size > 1) {
4540:     PetscCall(MatSetType(*A, MATMPIAIJ));
4541:     PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4542:   } else {
4543:     PetscCall(MatSetType(*A, MATSEQAIJ));
4544:     PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4545:   }
4546:   PetscFunctionReturn(PETSC_SUCCESS);
4547: }

4549: /*MC
4550:     MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix

4552:     Synopsis:
4553:     MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)

4555:     Not Collective

4557:     Input Parameter:
4558: .   A - the `MATMPIAIJ` matrix

4560:     Output Parameters:
4561: +   Ad - the diagonal portion of the matrix
4562: .   Ao - the off-diagonal portion of the matrix
4563: .   colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4564: -   ierr - error code

4566:      Level: advanced

4568:     Note:
4569:     Use  `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`

4571: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4572: M*/

4574: /*MC
4575:     MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`

4577:     Synopsis:
4578:     MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)

4580:     Not Collective

4582:     Input Parameters:
4583: +   A - the `MATMPIAIJ` matrix
4584: .   Ad - the diagonal portion of the matrix
4585: .   Ao - the off-diagonal portion of the matrix
4586: .   colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4587: -   ierr - error code

4589:      Level: advanced

4591: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4592: M*/

4594: /*@C
4595:   MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix

4597:   Not Collective

4599:   Input Parameter:
4600: . A - The `MATMPIAIJ` matrix

4602:   Output Parameters:
4603: + Ad     - The local diagonal block as a `MATSEQAIJ` matrix
4604: . Ao     - The local off-diagonal block as a `MATSEQAIJ` matrix
4605: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix

4607:   Level: intermediate

4609:   Note:
4610:   The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4611:   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
4612:   the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4613:   local column numbers to global column numbers in the original matrix.

4615:   Fortran Notes:
4616:   `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`

4618: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4619: @*/
4620: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4621: {
4622:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4623:   PetscBool   flg;

4625:   PetscFunctionBegin;
4626:   PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4627:   PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4628:   if (Ad) *Ad = a->A;
4629:   if (Ao) *Ao = a->B;
4630:   if (colmap) *colmap = a->garray;
4631:   PetscFunctionReturn(PETSC_SUCCESS);
4632: }

4634: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4635: {
4636:   PetscInt     m, N, i, rstart, nnz, Ii;
4637:   PetscInt    *indx;
4638:   PetscScalar *values;
4639:   MatType      rootType;

4641:   PetscFunctionBegin;
4642:   PetscCall(MatGetSize(inmat, &m, &N));
4643:   if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4644:     PetscInt *dnz, *onz, sum, bs, cbs;

4646:     if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4647:     /* Check sum(n) = N */
4648:     PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4649:     PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);

4651:     PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4652:     rstart -= m;

4654:     MatPreallocateBegin(comm, m, n, dnz, onz);
4655:     for (i = 0; i < m; i++) {
4656:       PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4657:       PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4658:       PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4659:     }

4661:     PetscCall(MatCreate(comm, outmat));
4662:     PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4663:     PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4664:     PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4665:     PetscCall(MatGetRootType_Private(inmat, &rootType));
4666:     PetscCall(MatSetType(*outmat, rootType));
4667:     PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4668:     PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4669:     MatPreallocateEnd(dnz, onz);
4670:     PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4671:   }

4673:   /* numeric phase */
4674:   PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4675:   for (i = 0; i < m; i++) {
4676:     PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4677:     Ii = i + rstart;
4678:     PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4679:     PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4680:   }
4681:   PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4682:   PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4683:   PetscFunctionReturn(PETSC_SUCCESS);
4684: }

4686: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4687: {
4688:   Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;

4690:   PetscFunctionBegin;
4691:   if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4692:   PetscCall(PetscFree(merge->id_r));
4693:   PetscCall(PetscFree(merge->len_s));
4694:   PetscCall(PetscFree(merge->len_r));
4695:   PetscCall(PetscFree(merge->bi));
4696:   PetscCall(PetscFree(merge->bj));
4697:   PetscCall(PetscFree(merge->buf_ri[0]));
4698:   PetscCall(PetscFree(merge->buf_ri));
4699:   PetscCall(PetscFree(merge->buf_rj[0]));
4700:   PetscCall(PetscFree(merge->buf_rj));
4701:   PetscCall(PetscFree(merge->coi));
4702:   PetscCall(PetscFree(merge->coj));
4703:   PetscCall(PetscFree(merge->owners_co));
4704:   PetscCall(PetscLayoutDestroy(&merge->rowmap));
4705:   PetscCall(PetscFree(merge));
4706:   PetscFunctionReturn(PETSC_SUCCESS);
4707: }

4709: #include <../src/mat/utils/freespace.h>
4710: #include <petscbt.h>

4712: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4713: {
4714:   MPI_Comm             comm;
4715:   Mat_SeqAIJ          *a = (Mat_SeqAIJ *)seqmat->data;
4716:   PetscMPIInt          size, rank, taga, *len_s;
4717:   PetscInt             N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4718:   PetscInt             proc, m;
4719:   PetscInt           **buf_ri, **buf_rj;
4720:   PetscInt             k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4721:   PetscInt             nrows, **buf_ri_k, **nextrow, **nextai;
4722:   MPI_Request         *s_waits, *r_waits;
4723:   MPI_Status          *status;
4724:   const MatScalar     *aa, *a_a;
4725:   MatScalar          **abuf_r, *ba_i;
4726:   Mat_Merge_SeqsToMPI *merge;
4727:   PetscContainer       container;

4729:   PetscFunctionBegin;
4730:   PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4731:   PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));

4733:   PetscCallMPI(MPI_Comm_size(comm, &size));
4734:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

4736:   PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4737:   PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4738:   PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4739:   PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4740:   aa = a_a;

4742:   bi     = merge->bi;
4743:   bj     = merge->bj;
4744:   buf_ri = merge->buf_ri;
4745:   buf_rj = merge->buf_rj;

4747:   PetscCall(PetscMalloc1(size, &status));
4748:   owners = merge->rowmap->range;
4749:   len_s  = merge->len_s;

4751:   /* send and recv matrix values */
4752:   PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4753:   PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));

4755:   PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4756:   for (proc = 0, k = 0; proc < size; proc++) {
4757:     if (!len_s[proc]) continue;
4758:     i = owners[proc];
4759:     PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4760:     k++;
4761:   }

4763:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4764:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4765:   PetscCall(PetscFree(status));

4767:   PetscCall(PetscFree(s_waits));
4768:   PetscCall(PetscFree(r_waits));

4770:   /* insert mat values of mpimat */
4771:   PetscCall(PetscMalloc1(N, &ba_i));
4772:   PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));

4774:   for (k = 0; k < merge->nrecv; k++) {
4775:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4776:     nrows       = *buf_ri_k[k];
4777:     nextrow[k]  = buf_ri_k[k] + 1;           /* next row number of k-th recved i-structure */
4778:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure  */
4779:   }

4781:   /* set values of ba */
4782:   m = merge->rowmap->n;
4783:   for (i = 0; i < m; i++) {
4784:     arow = owners[rank] + i;
4785:     bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4786:     bnzi = bi[i + 1] - bi[i];
4787:     PetscCall(PetscArrayzero(ba_i, bnzi));

4789:     /* add local non-zero vals of this proc's seqmat into ba */
4790:     anzi   = ai[arow + 1] - ai[arow];
4791:     aj     = a->j + ai[arow];
4792:     aa     = a_a + ai[arow];
4793:     nextaj = 0;
4794:     for (j = 0; nextaj < anzi; j++) {
4795:       if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4796:         ba_i[j] += aa[nextaj++];
4797:       }
4798:     }

4800:     /* add received vals into ba */
4801:     for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4802:       /* i-th row */
4803:       if (i == *nextrow[k]) {
4804:         anzi   = *(nextai[k] + 1) - *nextai[k];
4805:         aj     = buf_rj[k] + *nextai[k];
4806:         aa     = abuf_r[k] + *nextai[k];
4807:         nextaj = 0;
4808:         for (j = 0; nextaj < anzi; j++) {
4809:           if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4810:             ba_i[j] += aa[nextaj++];
4811:           }
4812:         }
4813:         nextrow[k]++;
4814:         nextai[k]++;
4815:       }
4816:     }
4817:     PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4818:   }
4819:   PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4820:   PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4821:   PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));

4823:   PetscCall(PetscFree(abuf_r[0]));
4824:   PetscCall(PetscFree(abuf_r));
4825:   PetscCall(PetscFree(ba_i));
4826:   PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4827:   PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4828:   PetscFunctionReturn(PETSC_SUCCESS);
4829: }

4831: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4832: {
4833:   Mat                  B_mpi;
4834:   Mat_SeqAIJ          *a = (Mat_SeqAIJ *)seqmat->data;
4835:   PetscMPIInt          size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4836:   PetscInt           **buf_rj, **buf_ri, **buf_ri_k;
4837:   PetscInt             M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4838:   PetscInt             len, proc, *dnz, *onz, bs, cbs;
4839:   PetscInt             k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4840:   PetscInt             nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4841:   MPI_Request         *si_waits, *sj_waits, *ri_waits, *rj_waits;
4842:   MPI_Status          *status;
4843:   PetscFreeSpaceList   free_space = NULL, current_space = NULL;
4844:   PetscBT              lnkbt;
4845:   Mat_Merge_SeqsToMPI *merge;
4846:   PetscContainer       container;

4848:   PetscFunctionBegin;
4849:   PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));

4851:   /* make sure it is a PETSc comm */
4852:   PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4853:   PetscCallMPI(MPI_Comm_size(comm, &size));
4854:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

4856:   PetscCall(PetscNew(&merge));
4857:   PetscCall(PetscMalloc1(size, &status));

4859:   /* determine row ownership */
4860:   PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4861:   PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4862:   PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4863:   PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4864:   PetscCall(PetscLayoutSetUp(merge->rowmap));
4865:   PetscCall(PetscMalloc1(size, &len_si));
4866:   PetscCall(PetscMalloc1(size, &merge->len_s));

4868:   m      = merge->rowmap->n;
4869:   owners = merge->rowmap->range;

4871:   /* determine the number of messages to send, their lengths */
4872:   len_s = merge->len_s;

4874:   len          = 0; /* length of buf_si[] */
4875:   merge->nsend = 0;
4876:   for (proc = 0; proc < size; proc++) {
4877:     len_si[proc] = 0;
4878:     if (proc == rank) {
4879:       len_s[proc] = 0;
4880:     } else {
4881:       len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4882:       len_s[proc]  = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4883:     }
4884:     if (len_s[proc]) {
4885:       merge->nsend++;
4886:       nrows = 0;
4887:       for (i = owners[proc]; i < owners[proc + 1]; i++) {
4888:         if (ai[i + 1] > ai[i]) nrows++;
4889:       }
4890:       len_si[proc] = 2 * (nrows + 1);
4891:       len += len_si[proc];
4892:     }
4893:   }

4895:   /* determine the number and length of messages to receive for ij-structure */
4896:   PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4897:   PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));

4899:   /* post the Irecv of j-structure */
4900:   PetscCall(PetscCommGetNewTag(comm, &tagj));
4901:   PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));

4903:   /* post the Isend of j-structure */
4904:   PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));

4906:   for (proc = 0, k = 0; proc < size; proc++) {
4907:     if (!len_s[proc]) continue;
4908:     i = owners[proc];
4909:     PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4910:     k++;
4911:   }

4913:   /* receives and sends of j-structure are complete */
4914:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4915:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));

4917:   /* send and recv i-structure */
4918:   PetscCall(PetscCommGetNewTag(comm, &tagi));
4919:   PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));

4921:   PetscCall(PetscMalloc1(len + 1, &buf_s));
4922:   buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4923:   for (proc = 0, k = 0; proc < size; proc++) {
4924:     if (!len_s[proc]) continue;
4925:     /* form outgoing message for i-structure:
4926:          buf_si[0]:                 nrows to be sent
4927:                [1:nrows]:           row index (global)
4928:                [nrows+1:2*nrows+1]: i-structure index
4929:     */
4930:     nrows       = len_si[proc] / 2 - 1;
4931:     buf_si_i    = buf_si + nrows + 1;
4932:     buf_si[0]   = nrows;
4933:     buf_si_i[0] = 0;
4934:     nrows       = 0;
4935:     for (i = owners[proc]; i < owners[proc + 1]; i++) {
4936:       anzi = ai[i + 1] - ai[i];
4937:       if (anzi) {
4938:         buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4939:         buf_si[nrows + 1]   = i - owners[proc];       /* local row index */
4940:         nrows++;
4941:       }
4942:     }
4943:     PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4944:     k++;
4945:     buf_si += len_si[proc];
4946:   }

4948:   if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4949:   if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));

4951:   PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4952:   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]));

4954:   PetscCall(PetscFree(len_si));
4955:   PetscCall(PetscFree(len_ri));
4956:   PetscCall(PetscFree(rj_waits));
4957:   PetscCall(PetscFree2(si_waits, sj_waits));
4958:   PetscCall(PetscFree(ri_waits));
4959:   PetscCall(PetscFree(buf_s));
4960:   PetscCall(PetscFree(status));

4962:   /* compute a local seq matrix in each processor */
4963:   /* allocate bi array and free space for accumulating nonzero column info */
4964:   PetscCall(PetscMalloc1(m + 1, &bi));
4965:   bi[0] = 0;

4967:   /* create and initialize a linked list */
4968:   nlnk = N + 1;
4969:   PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));

4971:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4972:   len = ai[owners[rank + 1]] - ai[owners[rank]];
4973:   PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));

4975:   current_space = free_space;

4977:   /* determine symbolic info for each local row */
4978:   PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));

4980:   for (k = 0; k < merge->nrecv; k++) {
4981:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4982:     nrows       = *buf_ri_k[k];
4983:     nextrow[k]  = buf_ri_k[k] + 1;           /* next row number of k-th recved i-structure */
4984:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure  */
4985:   }

4987:   MatPreallocateBegin(comm, m, n, dnz, onz);
4988:   len = 0;
4989:   for (i = 0; i < m; i++) {
4990:     bnzi = 0;
4991:     /* add local non-zero cols of this proc's seqmat into lnk */
4992:     arow = owners[rank] + i;
4993:     anzi = ai[arow + 1] - ai[arow];
4994:     aj   = a->j + ai[arow];
4995:     PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4996:     bnzi += nlnk;
4997:     /* add received col data into lnk */
4998:     for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4999:       if (i == *nextrow[k]) {            /* i-th row */
5000:         anzi = *(nextai[k] + 1) - *nextai[k];
5001:         aj   = buf_rj[k] + *nextai[k];
5002:         PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5003:         bnzi += nlnk;
5004:         nextrow[k]++;
5005:         nextai[k]++;
5006:       }
5007:     }
5008:     if (len < bnzi) len = bnzi; /* =max(bnzi) */

5010:     /* if free space is not available, make more free space */
5011:     if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), &current_space));
5012:     /* copy data into free space, then initialize lnk */
5013:     PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5014:     PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));

5016:     current_space->array += bnzi;
5017:     current_space->local_used += bnzi;
5018:     current_space->local_remaining -= bnzi;

5020:     bi[i + 1] = bi[i] + bnzi;
5021:   }

5023:   PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));

5025:   PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5026:   PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5027:   PetscCall(PetscLLDestroy(lnk, lnkbt));

5029:   /* create symbolic parallel matrix B_mpi */
5030:   PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5031:   PetscCall(MatCreate(comm, &B_mpi));
5032:   if (n == PETSC_DECIDE) {
5033:     PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5034:   } else {
5035:     PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5036:   }
5037:   PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5038:   PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5039:   PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5040:   MatPreallocateEnd(dnz, onz);
5041:   PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));

5043:   /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5044:   B_mpi->assembled = PETSC_FALSE;
5045:   merge->bi        = bi;
5046:   merge->bj        = bj;
5047:   merge->buf_ri    = buf_ri;
5048:   merge->buf_rj    = buf_rj;
5049:   merge->coi       = NULL;
5050:   merge->coj       = NULL;
5051:   merge->owners_co = NULL;

5053:   PetscCall(PetscCommDestroy(&comm));

5055:   /* attach the supporting struct to B_mpi for reuse */
5056:   PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5057:   PetscCall(PetscContainerSetPointer(container, merge));
5058:   PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5059:   PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5060:   PetscCall(PetscContainerDestroy(&container));
5061:   *mpimat = B_mpi;

5063:   PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5064:   PetscFunctionReturn(PETSC_SUCCESS);
5065: }

5067: /*@C
5068:   MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5069:   matrices from each processor

5071:   Collective

5073:   Input Parameters:
5074: + comm   - the communicators the parallel matrix will live on
5075: . seqmat - the input sequential matrices
5076: . m      - number of local rows (or `PETSC_DECIDE`)
5077: . n      - number of local columns (or `PETSC_DECIDE`)
5078: - scall  - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5080:   Output Parameter:
5081: . mpimat - the parallel matrix generated

5083:   Level: advanced

5085:   Note:
5086:   The dimensions of the sequential matrix in each processor MUST be the same.
5087:   The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5088:   destroyed when mpimat is destroyed. Call `PetscObjectQuery()` to access seqmat.

5090: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5091: @*/
5092: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5093: {
5094:   PetscMPIInt size;

5096:   PetscFunctionBegin;
5097:   PetscCallMPI(MPI_Comm_size(comm, &size));
5098:   if (size == 1) {
5099:     PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5100:     if (scall == MAT_INITIAL_MATRIX) {
5101:       PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5102:     } else {
5103:       PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5104:     }
5105:     PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5106:     PetscFunctionReturn(PETSC_SUCCESS);
5107:   }
5108:   PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5109:   if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5110:   PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5111:   PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5112:   PetscFunctionReturn(PETSC_SUCCESS);
5113: }

5115: /*@
5116:   MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.

5118:   Not Collective

5120:   Input Parameter:
5121: . A - the matrix

5123:   Output Parameter:
5124: . A_loc - the local sequential matrix generated

5126:   Level: developer

5128:   Notes:
5129:   The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5130:   with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5131:   `n` is the global column count obtained with `MatGetSize()`

5133:   In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.

5135:   For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.

5137:   Destroy the matrix with `MatDestroy()`

5139: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5140: @*/
5141: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5142: {
5143:   PetscBool mpi;

5145:   PetscFunctionBegin;
5146:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5147:   if (mpi) {
5148:     PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5149:   } else {
5150:     *A_loc = A;
5151:     PetscCall(PetscObjectReference((PetscObject)*A_loc));
5152:   }
5153:   PetscFunctionReturn(PETSC_SUCCESS);
5154: }

5156: /*@
5157:   MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.

5159:   Not Collective

5161:   Input Parameters:
5162: + A     - the matrix
5163: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5165:   Output Parameter:
5166: . A_loc - the local sequential matrix generated

5168:   Level: developer

5170:   Notes:
5171:   The matrix is created by taking all `A`'s local rows and putting them into a sequential
5172:   matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5173:   `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.

5175:   In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.

5177:   When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5178:   with its reference count increased by one. Hence changing values of `A_loc` changes `A`. If `MAT_REUSE_MATRIX` is requested on a sequential matrix
5179:   then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5180:   and then call this routine with `MAT_REUSE_MATRIX`. In this case, one can modify the values of `A_loc` without affecting the original sequential matrix.

5182: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5183: @*/
5184: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5185: {
5186:   Mat_MPIAIJ        *mpimat = (Mat_MPIAIJ *)A->data;
5187:   Mat_SeqAIJ        *mat, *a, *b;
5188:   PetscInt          *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5189:   const PetscScalar *aa, *ba, *aav, *bav;
5190:   PetscScalar       *ca, *cam;
5191:   PetscMPIInt        size;
5192:   PetscInt           am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5193:   PetscInt          *ci, *cj, col, ncols_d, ncols_o, jo;
5194:   PetscBool          match;

5196:   PetscFunctionBegin;
5197:   PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5198:   PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5199:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5200:   if (size == 1) {
5201:     if (scall == MAT_INITIAL_MATRIX) {
5202:       PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5203:       *A_loc = mpimat->A;
5204:     } else if (scall == MAT_REUSE_MATRIX) {
5205:       PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5206:     }
5207:     PetscFunctionReturn(PETSC_SUCCESS);
5208:   }

5210:   PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5211:   a  = (Mat_SeqAIJ *)mpimat->A->data;
5212:   b  = (Mat_SeqAIJ *)mpimat->B->data;
5213:   ai = a->i;
5214:   aj = a->j;
5215:   bi = b->i;
5216:   bj = b->j;
5217:   PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5218:   PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5219:   aa = aav;
5220:   ba = bav;
5221:   if (scall == MAT_INITIAL_MATRIX) {
5222:     PetscCall(PetscMalloc1(1 + am, &ci));
5223:     ci[0] = 0;
5224:     for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5225:     PetscCall(PetscMalloc1(1 + ci[am], &cj));
5226:     PetscCall(PetscMalloc1(1 + ci[am], &ca));
5227:     k = 0;
5228:     for (i = 0; i < am; i++) {
5229:       ncols_o = bi[i + 1] - bi[i];
5230:       ncols_d = ai[i + 1] - ai[i];
5231:       /* off-diagonal portion of A */
5232:       for (jo = 0; jo < ncols_o; jo++) {
5233:         col = cmap[*bj];
5234:         if (col >= cstart) break;
5235:         cj[k] = col;
5236:         bj++;
5237:         ca[k++] = *ba++;
5238:       }
5239:       /* diagonal portion of A */
5240:       for (j = 0; j < ncols_d; j++) {
5241:         cj[k]   = cstart + *aj++;
5242:         ca[k++] = *aa++;
5243:       }
5244:       /* off-diagonal portion of A */
5245:       for (j = jo; j < ncols_o; j++) {
5246:         cj[k]   = cmap[*bj++];
5247:         ca[k++] = *ba++;
5248:       }
5249:     }
5250:     /* put together the new matrix */
5251:     PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5252:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5253:     /* Since these are PETSc arrays, change flags to free them as necessary. */
5254:     mat          = (Mat_SeqAIJ *)(*A_loc)->data;
5255:     mat->free_a  = PETSC_TRUE;
5256:     mat->free_ij = PETSC_TRUE;
5257:     mat->nonew   = 0;
5258:   } else if (scall == MAT_REUSE_MATRIX) {
5259:     mat = (Mat_SeqAIJ *)(*A_loc)->data;
5260:     ci  = mat->i;
5261:     cj  = mat->j;
5262:     PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5263:     for (i = 0; i < am; i++) {
5264:       /* off-diagonal portion of A */
5265:       ncols_o = bi[i + 1] - bi[i];
5266:       for (jo = 0; jo < ncols_o; jo++) {
5267:         col = cmap[*bj];
5268:         if (col >= cstart) break;
5269:         *cam++ = *ba++;
5270:         bj++;
5271:       }
5272:       /* diagonal portion of A */
5273:       ncols_d = ai[i + 1] - ai[i];
5274:       for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5275:       /* off-diagonal portion of A */
5276:       for (j = jo; j < ncols_o; j++) {
5277:         *cam++ = *ba++;
5278:         bj++;
5279:       }
5280:     }
5281:     PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5282:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5283:   PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5284:   PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5285:   PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5286:   PetscFunctionReturn(PETSC_SUCCESS);
5287: }

5289: /*@
5290:   MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5291:   mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part

5293:   Not Collective

5295:   Input Parameters:
5296: + A     - the matrix
5297: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5299:   Output Parameters:
5300: + glob  - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5301: - A_loc - the local sequential matrix generated

5303:   Level: developer

5305:   Note:
5306:   This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5307:   part, then those associated with the off-diagonal part (in its local ordering)

5309: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5310: @*/
5311: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5312: {
5313:   Mat             Ao, Ad;
5314:   const PetscInt *cmap;
5315:   PetscMPIInt     size;
5316:   PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);

5318:   PetscFunctionBegin;
5319:   PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5320:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5321:   if (size == 1) {
5322:     if (scall == MAT_INITIAL_MATRIX) {
5323:       PetscCall(PetscObjectReference((PetscObject)Ad));
5324:       *A_loc = Ad;
5325:     } else if (scall == MAT_REUSE_MATRIX) {
5326:       PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5327:     }
5328:     if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5329:     PetscFunctionReturn(PETSC_SUCCESS);
5330:   }
5331:   PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5332:   PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5333:   if (f) {
5334:     PetscCall((*f)(A, scall, glob, A_loc));
5335:   } else {
5336:     Mat_SeqAIJ        *a = (Mat_SeqAIJ *)Ad->data;
5337:     Mat_SeqAIJ        *b = (Mat_SeqAIJ *)Ao->data;
5338:     Mat_SeqAIJ        *c;
5339:     PetscInt          *ai = a->i, *aj = a->j;
5340:     PetscInt          *bi = b->i, *bj = b->j;
5341:     PetscInt          *ci, *cj;
5342:     const PetscScalar *aa, *ba;
5343:     PetscScalar       *ca;
5344:     PetscInt           i, j, am, dn, on;

5346:     PetscCall(MatGetLocalSize(Ad, &am, &dn));
5347:     PetscCall(MatGetLocalSize(Ao, NULL, &on));
5348:     PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5349:     PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5350:     if (scall == MAT_INITIAL_MATRIX) {
5351:       PetscInt k;
5352:       PetscCall(PetscMalloc1(1 + am, &ci));
5353:       PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5354:       PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5355:       ci[0] = 0;
5356:       for (i = 0, k = 0; i < am; i++) {
5357:         const PetscInt ncols_o = bi[i + 1] - bi[i];
5358:         const PetscInt ncols_d = ai[i + 1] - ai[i];
5359:         ci[i + 1]              = ci[i] + ncols_o + ncols_d;
5360:         /* diagonal portion of A */
5361:         for (j = 0; j < ncols_d; j++, k++) {
5362:           cj[k] = *aj++;
5363:           ca[k] = *aa++;
5364:         }
5365:         /* off-diagonal portion of A */
5366:         for (j = 0; j < ncols_o; j++, k++) {
5367:           cj[k] = dn + *bj++;
5368:           ca[k] = *ba++;
5369:         }
5370:       }
5371:       /* put together the new matrix */
5372:       PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5373:       /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5374:       /* Since these are PETSc arrays, change flags to free them as necessary. */
5375:       c          = (Mat_SeqAIJ *)(*A_loc)->data;
5376:       c->free_a  = PETSC_TRUE;
5377:       c->free_ij = PETSC_TRUE;
5378:       c->nonew   = 0;
5379:       PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5380:     } else if (scall == MAT_REUSE_MATRIX) {
5381:       PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5382:       for (i = 0; i < am; i++) {
5383:         const PetscInt ncols_d = ai[i + 1] - ai[i];
5384:         const PetscInt ncols_o = bi[i + 1] - bi[i];
5385:         /* diagonal portion of A */
5386:         for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5387:         /* off-diagonal portion of A */
5388:         for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5389:       }
5390:       PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5391:     } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5392:     PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5393:     PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5394:     if (glob) {
5395:       PetscInt cst, *gidx;

5397:       PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5398:       PetscCall(PetscMalloc1(dn + on, &gidx));
5399:       for (i = 0; i < dn; i++) gidx[i] = cst + i;
5400:       for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5401:       PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5402:     }
5403:   }
5404:   PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5405:   PetscFunctionReturn(PETSC_SUCCESS);
5406: }

5408: /*@C
5409:   MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns

5411:   Not Collective

5413:   Input Parameters:
5414: + A     - the matrix
5415: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5416: . row   - index set of rows to extract (or `NULL`)
5417: - col   - index set of columns to extract (or `NULL`)

5419:   Output Parameter:
5420: . A_loc - the local sequential matrix generated

5422:   Level: developer

5424: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5425: @*/
5426: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5427: {
5428:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5429:   PetscInt    i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5430:   IS          isrowa, iscola;
5431:   Mat        *aloc;
5432:   PetscBool   match;

5434:   PetscFunctionBegin;
5435:   PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5436:   PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5437:   PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5438:   if (!row) {
5439:     start = A->rmap->rstart;
5440:     end   = A->rmap->rend;
5441:     PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5442:   } else {
5443:     isrowa = *row;
5444:   }
5445:   if (!col) {
5446:     start = A->cmap->rstart;
5447:     cmap  = a->garray;
5448:     nzA   = a->A->cmap->n;
5449:     nzB   = a->B->cmap->n;
5450:     PetscCall(PetscMalloc1(nzA + nzB, &idx));
5451:     ncols = 0;
5452:     for (i = 0; i < nzB; i++) {
5453:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5454:       else break;
5455:     }
5456:     imark = i;
5457:     for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5458:     for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5459:     PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5460:   } else {
5461:     iscola = *col;
5462:   }
5463:   if (scall != MAT_INITIAL_MATRIX) {
5464:     PetscCall(PetscMalloc1(1, &aloc));
5465:     aloc[0] = *A_loc;
5466:   }
5467:   PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5468:   if (!col) { /* attach global id of condensed columns */
5469:     PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5470:   }
5471:   *A_loc = aloc[0];
5472:   PetscCall(PetscFree(aloc));
5473:   if (!row) PetscCall(ISDestroy(&isrowa));
5474:   if (!col) PetscCall(ISDestroy(&iscola));
5475:   PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5476:   PetscFunctionReturn(PETSC_SUCCESS);
5477: }

5479: /*
5480:  * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5481:  * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5482:  * on a global size.
5483:  * */
5484: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5485: {
5486:   Mat_MPIAIJ            *p  = (Mat_MPIAIJ *)P->data;
5487:   Mat_SeqAIJ            *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5488:   PetscInt               plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5489:   PetscMPIInt            owner;
5490:   PetscSFNode           *iremote, *oiremote;
5491:   const PetscInt        *lrowindices;
5492:   PetscSF                sf, osf;
5493:   PetscInt               pcstart, *roffsets, *loffsets, *pnnz, j;
5494:   PetscInt               ontotalcols, dntotalcols, ntotalcols, nout;
5495:   MPI_Comm               comm;
5496:   ISLocalToGlobalMapping mapping;
5497:   const PetscScalar     *pd_a, *po_a;

5499:   PetscFunctionBegin;
5500:   PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5501:   /* plocalsize is the number of roots
5502:    * nrows is the number of leaves
5503:    * */
5504:   PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5505:   PetscCall(ISGetLocalSize(rows, &nrows));
5506:   PetscCall(PetscCalloc1(nrows, &iremote));
5507:   PetscCall(ISGetIndices(rows, &lrowindices));
5508:   for (i = 0; i < nrows; i++) {
5509:     /* Find a remote index and an owner for a row
5510:      * The row could be local or remote
5511:      * */
5512:     owner = 0;
5513:     lidx  = 0;
5514:     PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5515:     iremote[i].index = lidx;
5516:     iremote[i].rank  = owner;
5517:   }
5518:   /* Create SF to communicate how many nonzero columns for each row */
5519:   PetscCall(PetscSFCreate(comm, &sf));
5520:   /* SF will figure out the number of nonzero columns for each row, and their
5521:    * offsets
5522:    * */
5523:   PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5524:   PetscCall(PetscSFSetFromOptions(sf));
5525:   PetscCall(PetscSFSetUp(sf));

5527:   PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5528:   PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5529:   PetscCall(PetscCalloc1(nrows, &pnnz));
5530:   roffsets[0] = 0;
5531:   roffsets[1] = 0;
5532:   for (i = 0; i < plocalsize; i++) {
5533:     /* diagonal */
5534:     nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5535:     /* off-diagonal */
5536:     nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5537:     /* compute offsets so that we relative location for each row */
5538:     roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5539:     roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5540:   }
5541:   PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5542:   PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5543:   /* 'r' means root, and 'l' means leaf */
5544:   PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5545:   PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5546:   PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5547:   PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5548:   PetscCall(PetscSFDestroy(&sf));
5549:   PetscCall(PetscFree(roffsets));
5550:   PetscCall(PetscFree(nrcols));
5551:   dntotalcols = 0;
5552:   ontotalcols = 0;
5553:   ncol        = 0;
5554:   for (i = 0; i < nrows; i++) {
5555:     pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5556:     ncol    = PetscMax(pnnz[i], ncol);
5557:     /* diagonal */
5558:     dntotalcols += nlcols[i * 2 + 0];
5559:     /* off-diagonal */
5560:     ontotalcols += nlcols[i * 2 + 1];
5561:   }
5562:   /* We do not need to figure the right number of columns
5563:    * since all the calculations will be done by going through the raw data
5564:    * */
5565:   PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5566:   PetscCall(MatSetUp(*P_oth));
5567:   PetscCall(PetscFree(pnnz));
5568:   p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5569:   /* diagonal */
5570:   PetscCall(PetscCalloc1(dntotalcols, &iremote));
5571:   /* off-diagonal */
5572:   PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5573:   /* diagonal */
5574:   PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5575:   /* off-diagonal */
5576:   PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5577:   dntotalcols = 0;
5578:   ontotalcols = 0;
5579:   ntotalcols  = 0;
5580:   for (i = 0; i < nrows; i++) {
5581:     owner = 0;
5582:     PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5583:     /* Set iremote for diag matrix */
5584:     for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5585:       iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5586:       iremote[dntotalcols].rank  = owner;
5587:       /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5588:       ilocal[dntotalcols++] = ntotalcols++;
5589:     }
5590:     /* off-diagonal */
5591:     for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5592:       oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5593:       oiremote[ontotalcols].rank  = owner;
5594:       oilocal[ontotalcols++]      = ntotalcols++;
5595:     }
5596:   }
5597:   PetscCall(ISRestoreIndices(rows, &lrowindices));
5598:   PetscCall(PetscFree(loffsets));
5599:   PetscCall(PetscFree(nlcols));
5600:   PetscCall(PetscSFCreate(comm, &sf));
5601:   /* P serves as roots and P_oth is leaves
5602:    * Diag matrix
5603:    * */
5604:   PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5605:   PetscCall(PetscSFSetFromOptions(sf));
5606:   PetscCall(PetscSFSetUp(sf));

5608:   PetscCall(PetscSFCreate(comm, &osf));
5609:   /* off-diagonal */
5610:   PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5611:   PetscCall(PetscSFSetFromOptions(osf));
5612:   PetscCall(PetscSFSetUp(osf));
5613:   PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5614:   PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5615:   /* operate on the matrix internal data to save memory */
5616:   PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5617:   PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5618:   PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5619:   /* Convert to global indices for diag matrix */
5620:   for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5621:   PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5622:   /* We want P_oth store global indices */
5623:   PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5624:   /* Use memory scalable approach */
5625:   PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5626:   PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5627:   PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5628:   PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5629:   /* Convert back to local indices */
5630:   for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5631:   PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5632:   nout = 0;
5633:   PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5634:   PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5635:   PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5636:   /* Exchange values */
5637:   PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5638:   PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5639:   PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5640:   PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5641:   /* Stop PETSc from shrinking memory */
5642:   for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5643:   PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5644:   PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5645:   /* Attach PetscSF objects to P_oth so that we can reuse it later */
5646:   PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5647:   PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5648:   PetscCall(PetscSFDestroy(&sf));
5649:   PetscCall(PetscSFDestroy(&osf));
5650:   PetscFunctionReturn(PETSC_SUCCESS);
5651: }

5653: /*
5654:  * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5655:  * This supports MPIAIJ and MAIJ
5656:  * */
5657: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5658: {
5659:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5660:   Mat_SeqAIJ *p_oth;
5661:   IS          rows, map;
5662:   PetscHMapI  hamp;
5663:   PetscInt    i, htsize, *rowindices, off, *mapping, key, count;
5664:   MPI_Comm    comm;
5665:   PetscSF     sf, osf;
5666:   PetscBool   has;

5668:   PetscFunctionBegin;
5669:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5670:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5671:   /* If it is the first time, create an index set of off-diag nonzero columns of A,
5672:    *  and then create a submatrix (that often is an overlapping matrix)
5673:    * */
5674:   if (reuse == MAT_INITIAL_MATRIX) {
5675:     /* Use a hash table to figure out unique keys */
5676:     PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5677:     PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5678:     count = 0;
5679:     /* Assume that  a->g is sorted, otherwise the following does not make sense */
5680:     for (i = 0; i < a->B->cmap->n; i++) {
5681:       key = a->garray[i] / dof;
5682:       PetscCall(PetscHMapIHas(hamp, key, &has));
5683:       if (!has) {
5684:         mapping[i] = count;
5685:         PetscCall(PetscHMapISet(hamp, key, count++));
5686:       } else {
5687:         /* Current 'i' has the same value the previous step */
5688:         mapping[i] = count - 1;
5689:       }
5690:     }
5691:     PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5692:     PetscCall(PetscHMapIGetSize(hamp, &htsize));
5693:     PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5694:     PetscCall(PetscCalloc1(htsize, &rowindices));
5695:     off = 0;
5696:     PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5697:     PetscCall(PetscHMapIDestroy(&hamp));
5698:     PetscCall(PetscSortInt(htsize, rowindices));
5699:     PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5700:     /* In case, the matrix was already created but users want to recreate the matrix */
5701:     PetscCall(MatDestroy(P_oth));
5702:     PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5703:     PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5704:     PetscCall(ISDestroy(&map));
5705:     PetscCall(ISDestroy(&rows));
5706:   } else if (reuse == MAT_REUSE_MATRIX) {
5707:     /* If matrix was already created, we simply update values using SF objects
5708:      * that as attached to the matrix earlier.
5709:      */
5710:     const PetscScalar *pd_a, *po_a;

5712:     PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5713:     PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5714:     PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5715:     p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5716:     /* Update values in place */
5717:     PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5718:     PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5719:     PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5720:     PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5721:     PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5722:     PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5723:     PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5724:     PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5725:   } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5726:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5727:   PetscFunctionReturn(PETSC_SUCCESS);
5728: }

5730: /*@C
5731:   MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`

5733:   Collective

5735:   Input Parameters:
5736: + A     - the first matrix in `MATMPIAIJ` format
5737: . B     - the second matrix in `MATMPIAIJ` format
5738: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5740:   Output Parameters:
5741: + rowb  - On input index sets of rows of B to extract (or `NULL`), modified on output
5742: . colb  - On input index sets of columns of B to extract (or `NULL`), modified on output
5743: - B_seq - the sequential matrix generated

5745:   Level: developer

5747: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5748: @*/
5749: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5750: {
5751:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5752:   PetscInt   *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5753:   IS          isrowb, iscolb;
5754:   Mat        *bseq = NULL;

5756:   PetscFunctionBegin;
5757:   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 ")",
5758:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5759:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));

5761:   if (scall == MAT_INITIAL_MATRIX) {
5762:     start = A->cmap->rstart;
5763:     cmap  = a->garray;
5764:     nzA   = a->A->cmap->n;
5765:     nzB   = a->B->cmap->n;
5766:     PetscCall(PetscMalloc1(nzA + nzB, &idx));
5767:     ncols = 0;
5768:     for (i = 0; i < nzB; i++) { /* row < local row index */
5769:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5770:       else break;
5771:     }
5772:     imark = i;
5773:     for (i = 0; i < nzA; i++) idx[ncols++] = start + i;   /* local rows */
5774:     for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5775:     PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5776:     PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5777:   } else {
5778:     PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5779:     isrowb = *rowb;
5780:     iscolb = *colb;
5781:     PetscCall(PetscMalloc1(1, &bseq));
5782:     bseq[0] = *B_seq;
5783:   }
5784:   PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5785:   *B_seq = bseq[0];
5786:   PetscCall(PetscFree(bseq));
5787:   if (!rowb) {
5788:     PetscCall(ISDestroy(&isrowb));
5789:   } else {
5790:     *rowb = isrowb;
5791:   }
5792:   if (!colb) {
5793:     PetscCall(ISDestroy(&iscolb));
5794:   } else {
5795:     *colb = iscolb;
5796:   }
5797:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5798:   PetscFunctionReturn(PETSC_SUCCESS);
5799: }

5801: /*
5802:     MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5803:     of the OFF-DIAGONAL portion of local A

5805:     Collective

5807:    Input Parameters:
5808: +    A,B - the matrices in `MATMPIAIJ` format
5809: -    scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`

5811:    Output Parameter:
5812: +    startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5813: .    startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5814: .    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5815: -    B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N

5817:     Developer Note:
5818:     This directly accesses information inside the VecScatter associated with the matrix-vector product
5819:      for this matrix. This is not desirable..

5821:     Level: developer

5823: */
5824: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5825: {
5826:   Mat_MPIAIJ        *a = (Mat_MPIAIJ *)A->data;
5827:   Mat_SeqAIJ        *b_oth;
5828:   VecScatter         ctx;
5829:   MPI_Comm           comm;
5830:   const PetscMPIInt *rprocs, *sprocs;
5831:   const PetscInt    *srow, *rstarts, *sstarts;
5832:   PetscInt          *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5833:   PetscInt           i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5834:   PetscScalar       *b_otha, *bufa, *bufA, *vals = NULL;
5835:   MPI_Request       *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5836:   PetscMPIInt        size, tag, rank, nreqs;

5838:   PetscFunctionBegin;
5839:   PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5840:   PetscCallMPI(MPI_Comm_size(comm, &size));

5842:   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 ")",
5843:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5844:   PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5845:   PetscCallMPI(MPI_Comm_rank(comm, &rank));

5847:   if (size == 1) {
5848:     startsj_s = NULL;
5849:     bufa_ptr  = NULL;
5850:     *B_oth    = NULL;
5851:     PetscFunctionReturn(PETSC_SUCCESS);
5852:   }

5854:   ctx = a->Mvctx;
5855:   tag = ((PetscObject)ctx)->tag;

5857:   PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5858:   /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5859:   PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5860:   PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5861:   PetscCall(PetscMalloc1(nreqs, &reqs));
5862:   rwaits = reqs;
5863:   swaits = PetscSafePointerPlusOffset(reqs, nrecvs);

5865:   if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5866:   if (scall == MAT_INITIAL_MATRIX) {
5867:     /* i-array */
5868:     /*  post receives */
5869:     if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5870:     for (i = 0; i < nrecvs; i++) {
5871:       rowlen = rvalues + rstarts[i] * rbs;
5872:       nrows  = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5873:       PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5874:     }

5876:     /* pack the outgoing message */
5877:     PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));

5879:     sstartsj[0] = 0;
5880:     rstartsj[0] = 0;
5881:     len         = 0; /* total length of j or a array to be sent */
5882:     if (nsends) {
5883:       k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5884:       PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5885:     }
5886:     for (i = 0; i < nsends; i++) {
5887:       rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5888:       nrows  = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5889:       for (j = 0; j < nrows; j++) {
5890:         row = srow[k] + B->rmap->range[rank]; /* global row idx */
5891:         for (l = 0; l < sbs; l++) {
5892:           PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */

5894:           rowlen[j * sbs + l] = ncols;

5896:           len += ncols;
5897:           PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5898:         }
5899:         k++;
5900:       }
5901:       PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));

5903:       sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5904:     }
5905:     /* recvs and sends of i-array are completed */
5906:     if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5907:     PetscCall(PetscFree(svalues));

5909:     /* allocate buffers for sending j and a arrays */
5910:     PetscCall(PetscMalloc1(len + 1, &bufj));
5911:     PetscCall(PetscMalloc1(len + 1, &bufa));

5913:     /* create i-array of B_oth */
5914:     PetscCall(PetscMalloc1(aBn + 2, &b_othi));

5916:     b_othi[0] = 0;
5917:     len       = 0; /* total length of j or a array to be received */
5918:     k         = 0;
5919:     for (i = 0; i < nrecvs; i++) {
5920:       rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5921:       nrows  = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5922:       for (j = 0; j < nrows; j++) {
5923:         b_othi[k + 1] = b_othi[k] + rowlen[j];
5924:         PetscCall(PetscIntSumError(rowlen[j], len, &len));
5925:         k++;
5926:       }
5927:       rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5928:     }
5929:     PetscCall(PetscFree(rvalues));

5931:     /* allocate space for j and a arrays of B_oth */
5932:     PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5933:     PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));

5935:     /* j-array */
5936:     /*  post receives of j-array */
5937:     for (i = 0; i < nrecvs; i++) {
5938:       nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5939:       PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5940:     }

5942:     /* pack the outgoing message j-array */
5943:     if (nsends) k = sstarts[0];
5944:     for (i = 0; i < nsends; i++) {
5945:       nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5946:       bufJ  = bufj + sstartsj[i];
5947:       for (j = 0; j < nrows; j++) {
5948:         row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5949:         for (ll = 0; ll < sbs; ll++) {
5950:           PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5951:           for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5952:           PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5953:         }
5954:       }
5955:       PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5956:     }

5958:     /* recvs and sends of j-array are completed */
5959:     if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5960:   } else if (scall == MAT_REUSE_MATRIX) {
5961:     sstartsj = *startsj_s;
5962:     rstartsj = *startsj_r;
5963:     bufa     = *bufa_ptr;
5964:     b_oth    = (Mat_SeqAIJ *)(*B_oth)->data;
5965:     PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5966:   } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");

5968:   /* a-array */
5969:   /*  post receives of a-array */
5970:   for (i = 0; i < nrecvs; i++) {
5971:     nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5972:     PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5973:   }

5975:   /* pack the outgoing message a-array */
5976:   if (nsends) k = sstarts[0];
5977:   for (i = 0; i < nsends; i++) {
5978:     nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5979:     bufA  = bufa + sstartsj[i];
5980:     for (j = 0; j < nrows; j++) {
5981:       row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5982:       for (ll = 0; ll < sbs; ll++) {
5983:         PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5984:         for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5985:         PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5986:       }
5987:     }
5988:     PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5989:   }
5990:   /* recvs and sends of a-array are completed */
5991:   if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5992:   PetscCall(PetscFree(reqs));

5994:   if (scall == MAT_INITIAL_MATRIX) {
5995:     /* put together the new matrix */
5996:     PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));

5998:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5999:     /* Since these are PETSc arrays, change flags to free them as necessary. */
6000:     b_oth          = (Mat_SeqAIJ *)(*B_oth)->data;
6001:     b_oth->free_a  = PETSC_TRUE;
6002:     b_oth->free_ij = PETSC_TRUE;
6003:     b_oth->nonew   = 0;

6005:     PetscCall(PetscFree(bufj));
6006:     if (!startsj_s || !bufa_ptr) {
6007:       PetscCall(PetscFree2(sstartsj, rstartsj));
6008:       PetscCall(PetscFree(bufa_ptr));
6009:     } else {
6010:       *startsj_s = sstartsj;
6011:       *startsj_r = rstartsj;
6012:       *bufa_ptr  = bufa;
6013:     }
6014:   } else if (scall == MAT_REUSE_MATRIX) {
6015:     PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6016:   }

6018:   PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6019:   PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6020:   PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6021:   PetscFunctionReturn(PETSC_SUCCESS);
6022: }

6024: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6025: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6026: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6027: #if defined(PETSC_HAVE_MKL_SPARSE)
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6029: #endif
6030: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6031: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6032: #if defined(PETSC_HAVE_ELEMENTAL)
6033: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6034: #endif
6035: #if defined(PETSC_HAVE_SCALAPACK)
6036: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6037: #endif
6038: #if defined(PETSC_HAVE_HYPRE)
6039: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6040: #endif
6041: #if defined(PETSC_HAVE_CUDA)
6042: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6043: #endif
6044: #if defined(PETSC_HAVE_HIP)
6045: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6046: #endif
6047: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6048: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6049: #endif
6050: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6051: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6052: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);

6054: /*
6055:     Computes (B'*A')' since computing B*A directly is untenable

6057:                n                       p                          p
6058:         [             ]       [             ]         [                 ]
6059:       m [      A      ]  *  n [       B     ]   =   m [         C       ]
6060:         [             ]       [             ]         [                 ]

6062: */
6063: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6064: {
6065:   Mat At, Bt, Ct;

6067:   PetscFunctionBegin;
6068:   PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6069:   PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6070:   PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6071:   PetscCall(MatDestroy(&At));
6072:   PetscCall(MatDestroy(&Bt));
6073:   PetscCall(MatTransposeSetPrecursor(Ct, C));
6074:   PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6075:   PetscCall(MatDestroy(&Ct));
6076:   PetscFunctionReturn(PETSC_SUCCESS);
6077: }

6079: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6080: {
6081:   PetscBool cisdense;

6083:   PetscFunctionBegin;
6084:   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);
6085:   PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6086:   PetscCall(MatSetBlockSizesFromMats(C, A, B));
6087:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6088:   if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6089:   PetscCall(MatSetUp(C));

6091:   C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6092:   PetscFunctionReturn(PETSC_SUCCESS);
6093: }

6095: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6096: {
6097:   Mat_Product *product = C->product;
6098:   Mat          A = product->A, B = product->B;

6100:   PetscFunctionBegin;
6101:   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 ")",
6102:              A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6103:   C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6104:   C->ops->productsymbolic = MatProductSymbolic_AB;
6105:   PetscFunctionReturn(PETSC_SUCCESS);
6106: }

6108: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6109: {
6110:   Mat_Product *product = C->product;

6112:   PetscFunctionBegin;
6113:   if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6114:   PetscFunctionReturn(PETSC_SUCCESS);
6115: }

6117: /*
6118:    Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix

6120:   Input Parameters:

6122:     j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6123:     j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)

6125:     mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat

6127:     For Set1, j1[] contains column indices of the nonzeros.
6128:     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
6129:     respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6130:     but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.

6132:     Similar for Set2.

6134:     This routine merges the two sets of nonzeros row by row and removes repeats.

6136:   Output Parameters: (memory is allocated by the caller)

6138:     i[],j[]: the CSR of the merged matrix, which has m rows.
6139:     imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6140:     imap2[]: similar to imap1[], but for Set2.
6141:     Note we order nonzeros row-by-row and from left to right.
6142: */
6143: 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[])
6144: {
6145:   PetscInt   r, m; /* Row index of mat */
6146:   PetscCount t, t1, t2, b1, e1, b2, e2;

6148:   PetscFunctionBegin;
6149:   PetscCall(MatGetLocalSize(mat, &m, NULL));
6150:   t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6151:   i[0]        = 0;
6152:   for (r = 0; r < m; r++) { /* Do row by row merging */
6153:     b1 = rowBegin1[r];
6154:     e1 = rowEnd1[r];
6155:     b2 = rowBegin2[r];
6156:     e2 = rowEnd2[r];
6157:     while (b1 < e1 && b2 < e2) {
6158:       if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6159:         j[t]      = j1[b1];
6160:         imap1[t1] = t;
6161:         imap2[t2] = t;
6162:         b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6163:         b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6164:         t1++;
6165:         t2++;
6166:         t++;
6167:       } else if (j1[b1] < j2[b2]) {
6168:         j[t]      = j1[b1];
6169:         imap1[t1] = t;
6170:         b1 += jmap1[t1 + 1] - jmap1[t1];
6171:         t1++;
6172:         t++;
6173:       } else {
6174:         j[t]      = j2[b2];
6175:         imap2[t2] = t;
6176:         b2 += jmap2[t2 + 1] - jmap2[t2];
6177:         t2++;
6178:         t++;
6179:       }
6180:     }
6181:     /* Merge the remaining in either j1[] or j2[] */
6182:     while (b1 < e1) {
6183:       j[t]      = j1[b1];
6184:       imap1[t1] = t;
6185:       b1 += jmap1[t1 + 1] - jmap1[t1];
6186:       t1++;
6187:       t++;
6188:     }
6189:     while (b2 < e2) {
6190:       j[t]      = j2[b2];
6191:       imap2[t2] = t;
6192:       b2 += jmap2[t2 + 1] - jmap2[t2];
6193:       t2++;
6194:       t++;
6195:     }
6196:     i[r + 1] = t;
6197:   }
6198:   PetscFunctionReturn(PETSC_SUCCESS);
6199: }

6201: /*
6202:   Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block

6204:   Input Parameters:
6205:     mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6206:     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[]
6207:       respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.

6209:       i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6210:       i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.

6212:   Output Parameters:
6213:     j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6214:     rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6215:       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,
6216:       and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.

6218:     Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6219:       Atot: number of entries belonging to the diagonal block.
6220:       Annz: number of unique nonzeros belonging to the diagonal block.
6221:       Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6222:         repeats (i.e., same 'i,j' pair).
6223:       Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6224:         is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.

6226:       Atot: number of entries belonging to the diagonal block
6227:       Annz: number of unique nonzeros belonging to the diagonal block.

6229:     Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.

6231:     Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6232: */
6233: 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_)
6234: {
6235:   PetscInt    cstart, cend, rstart, rend, row, col;
6236:   PetscCount  Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6237:   PetscCount  Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6238:   PetscCount  k, m, p, q, r, s, mid;
6239:   PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;

6241:   PetscFunctionBegin;
6242:   PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6243:   PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6244:   m = rend - rstart;

6246:   /* Skip negative rows */
6247:   for (k = 0; k < n; k++)
6248:     if (i[k] >= 0) break;

6250:   /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6251:      fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6252:   */
6253:   while (k < n) {
6254:     row = i[k];
6255:     /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6256:     for (s = k; s < n; s++)
6257:       if (i[s] != row) break;

6259:     /* Shift diag columns to range of [-PETSC_MAX_INT, -1] */
6260:     for (p = k; p < s; p++) {
6261:       if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT;
6262:       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]);
6263:     }
6264:     PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6265:     PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6266:     rowBegin[row - rstart] = k;
6267:     rowMid[row - rstart]   = mid;
6268:     rowEnd[row - rstart]   = s;

6270:     /* Count nonzeros of this diag/offdiag row, which might have repeats */
6271:     Atot += mid - k;
6272:     Btot += s - mid;

6274:     /* Count unique nonzeros of this diag row */
6275:     for (p = k; p < mid;) {
6276:       col = j[p];
6277:       do {
6278:         j[p] += PETSC_MAX_INT; /* Revert the modified diagonal indices */
6279:         p++;
6280:       } while (p < mid && j[p] == col);
6281:       Annz++;
6282:     }

6284:     /* Count unique nonzeros of this offdiag row */
6285:     for (p = mid; p < s;) {
6286:       col = j[p];
6287:       do {
6288:         p++;
6289:       } while (p < s && j[p] == col);
6290:       Bnnz++;
6291:     }
6292:     k = s;
6293:   }

6295:   /* Allocation according to Atot, Btot, Annz, Bnnz */
6296:   PetscCall(PetscMalloc1(Atot, &Aperm));
6297:   PetscCall(PetscMalloc1(Btot, &Bperm));
6298:   PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6299:   PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));

6301:   /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6302:   Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6303:   for (r = 0; r < m; r++) {
6304:     k   = rowBegin[r];
6305:     mid = rowMid[r];
6306:     s   = rowEnd[r];
6307:     PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6308:     PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6309:     Atot += mid - k;
6310:     Btot += s - mid;

6312:     /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6313:     for (p = k; p < mid;) {
6314:       col = j[p];
6315:       q   = p;
6316:       do {
6317:         p++;
6318:       } while (p < mid && j[p] == col);
6319:       Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6320:       Annz++;
6321:     }

6323:     for (p = mid; p < s;) {
6324:       col = j[p];
6325:       q   = p;
6326:       do {
6327:         p++;
6328:       } while (p < s && j[p] == col);
6329:       Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6330:       Bnnz++;
6331:     }
6332:   }
6333:   /* Output */
6334:   *Aperm_ = Aperm;
6335:   *Annz_  = Annz;
6336:   *Atot_  = Atot;
6337:   *Ajmap_ = Ajmap;
6338:   *Bperm_ = Bperm;
6339:   *Bnnz_  = Bnnz;
6340:   *Btot_  = Btot;
6341:   *Bjmap_ = Bjmap;
6342:   PetscFunctionReturn(PETSC_SUCCESS);
6343: }

6345: /*
6346:   Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix

6348:   Input Parameters:
6349:     nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6350:     nnz:  number of unique nonzeros in the merged matrix
6351:     imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6352:     jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set

6354:   Output Parameter: (memory is allocated by the caller)
6355:     jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set

6357:   Example:
6358:     nnz1 = 4
6359:     nnz  = 6
6360:     imap = [1,3,4,5]
6361:     jmap = [0,3,5,6,7]
6362:    then,
6363:     jmap_new = [0,0,3,3,5,6,7]
6364: */
6365: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6366: {
6367:   PetscCount k, p;

6369:   PetscFunctionBegin;
6370:   jmap_new[0] = 0;
6371:   p           = nnz;                /* p loops over jmap_new[] backwards */
6372:   for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6373:     for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6374:   }
6375:   for (; p >= 0; p--) jmap_new[p] = jmap[0];
6376:   PetscFunctionReturn(PETSC_SUCCESS);
6377: }

6379: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6380: {
6381:   MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;

6383:   PetscFunctionBegin;
6384:   PetscCall(PetscSFDestroy(&coo->sf));
6385:   PetscCall(PetscFree(coo->Aperm1));
6386:   PetscCall(PetscFree(coo->Bperm1));
6387:   PetscCall(PetscFree(coo->Ajmap1));
6388:   PetscCall(PetscFree(coo->Bjmap1));
6389:   PetscCall(PetscFree(coo->Aimap2));
6390:   PetscCall(PetscFree(coo->Bimap2));
6391:   PetscCall(PetscFree(coo->Aperm2));
6392:   PetscCall(PetscFree(coo->Bperm2));
6393:   PetscCall(PetscFree(coo->Ajmap2));
6394:   PetscCall(PetscFree(coo->Bjmap2));
6395:   PetscCall(PetscFree(coo->Cperm1));
6396:   PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6397:   PetscCall(PetscFree(coo));
6398:   PetscFunctionReturn(PETSC_SUCCESS);
6399: }

6401: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6402: {
6403:   MPI_Comm             comm;
6404:   PetscMPIInt          rank, size;
6405:   PetscInt             m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6406:   PetscCount           k, p, q, rem;                           /* Loop variables over coo arrays */
6407:   Mat_MPIAIJ          *mpiaij = (Mat_MPIAIJ *)mat->data;
6408:   PetscContainer       container;
6409:   MatCOOStruct_MPIAIJ *coo;

6411:   PetscFunctionBegin;
6412:   PetscCall(PetscFree(mpiaij->garray));
6413:   PetscCall(VecDestroy(&mpiaij->lvec));
6414: #if defined(PETSC_USE_CTABLE)
6415:   PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6416: #else
6417:   PetscCall(PetscFree(mpiaij->colmap));
6418: #endif
6419:   PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6420:   mat->assembled     = PETSC_FALSE;
6421:   mat->was_assembled = PETSC_FALSE;

6423:   PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6424:   PetscCallMPI(MPI_Comm_size(comm, &size));
6425:   PetscCallMPI(MPI_Comm_rank(comm, &rank));
6426:   PetscCall(PetscLayoutSetUp(mat->rmap));
6427:   PetscCall(PetscLayoutSetUp(mat->cmap));
6428:   PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6429:   PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6430:   PetscCall(MatGetLocalSize(mat, &m, &n));
6431:   PetscCall(MatGetSize(mat, &M, &N));

6433:   /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6434:   /* entries come first, then local rows, then remote rows.                     */
6435:   PetscCount n1 = coo_n, *perm1;
6436:   PetscInt  *i1 = coo_i, *j1 = coo_j;

6438:   PetscCall(PetscMalloc1(n1, &perm1));
6439:   for (k = 0; k < n1; k++) perm1[k] = k;

6441:   /* Manipulate indices so that entries with negative row or col indices will have smallest
6442:      row indices, local entries will have greater but negative row indices, and remote entries
6443:      will have positive row indices.
6444:   */
6445:   for (k = 0; k < n1; k++) {
6446:     if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT;                /* e.g., -2^31, minimal to move them ahead */
6447:     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] */
6448:     else {
6449:       PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6450:       if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6451:     }
6452:   }

6454:   /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6455:   PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));

6457:   /* Advance k to the first entry we need to take care of */
6458:   for (k = 0; k < n1; k++)
6459:     if (i1[k] > PETSC_MIN_INT) break;
6460:   PetscInt i1start = k;

6462:   PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6463:   for (; k < rem; k++) i1[k] += PETSC_MAX_INT;                                    /* Revert row indices of local rows*/

6465:   /*           Send remote rows to their owner                                  */
6466:   /* Find which rows should be sent to which remote ranks*/
6467:   PetscInt        nsend = 0; /* Number of MPI ranks to send data to */
6468:   PetscMPIInt    *sendto;    /* [nsend], storing remote ranks */
6469:   PetscInt       *nentries;  /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6470:   const PetscInt *ranges;
6471:   PetscInt        maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */

6473:   PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6474:   PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6475:   for (k = rem; k < n1;) {
6476:     PetscMPIInt owner;
6477:     PetscInt    firstRow, lastRow;

6479:     /* Locate a row range */
6480:     firstRow = i1[k]; /* first row of this owner */
6481:     PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6482:     lastRow = ranges[owner + 1] - 1; /* last row of this owner */

6484:     /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6485:     PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));

6487:     /* All entries in [k,p) belong to this remote owner */
6488:     if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6489:       PetscMPIInt *sendto2;
6490:       PetscInt    *nentries2;
6491:       PetscInt     maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;

6493:       PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6494:       PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6495:       PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6496:       PetscCall(PetscFree2(sendto, nentries2));
6497:       sendto   = sendto2;
6498:       nentries = nentries2;
6499:       maxNsend = maxNsend2;
6500:     }
6501:     sendto[nsend]   = owner;
6502:     nentries[nsend] = p - k;
6503:     PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6504:     nsend++;
6505:     k = p;
6506:   }

6508:   /* Build 1st SF to know offsets on remote to send data */
6509:   PetscSF      sf1;
6510:   PetscInt     nroots = 1, nroots2 = 0;
6511:   PetscInt     nleaves = nsend, nleaves2 = 0;
6512:   PetscInt    *offsets;
6513:   PetscSFNode *iremote;

6515:   PetscCall(PetscSFCreate(comm, &sf1));
6516:   PetscCall(PetscMalloc1(nsend, &iremote));
6517:   PetscCall(PetscMalloc1(nsend, &offsets));
6518:   for (k = 0; k < nsend; k++) {
6519:     iremote[k].rank  = sendto[k];
6520:     iremote[k].index = 0;
6521:     nleaves2 += nentries[k];
6522:     PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6523:   }
6524:   PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6525:   PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6526:   PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6527:   PetscCall(PetscSFDestroy(&sf1));
6528:   PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);

6530:   /* Build 2nd SF to send remote COOs to their owner */
6531:   PetscSF sf2;
6532:   nroots  = nroots2;
6533:   nleaves = nleaves2;
6534:   PetscCall(PetscSFCreate(comm, &sf2));
6535:   PetscCall(PetscSFSetFromOptions(sf2));
6536:   PetscCall(PetscMalloc1(nleaves, &iremote));
6537:   p = 0;
6538:   for (k = 0; k < nsend; k++) {
6539:     PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6540:     for (q = 0; q < nentries[k]; q++, p++) {
6541:       iremote[p].rank  = sendto[k];
6542:       iremote[p].index = offsets[k] + q;
6543:     }
6544:   }
6545:   PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));

6547:   /* Send the remote COOs to their owner */
6548:   PetscInt    n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6549:   PetscCount *perm2;                 /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6550:   PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6551:   PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6552:   PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6553:   PetscInt *i1prem = i1 ? i1 + rem : NULL; /* silence ubsan warnings about pointer arithmetic on null pointer */
6554:   PetscInt *j1prem = j1 ? j1 + rem : NULL;
6555:   PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6556:   PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6557:   PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6558:   PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));

6560:   PetscCall(PetscFree(offsets));
6561:   PetscCall(PetscFree2(sendto, nentries));

6563:   /* Sort received COOs by row along with the permutation array     */
6564:   for (k = 0; k < n2; k++) perm2[k] = k;
6565:   PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));

6567:   /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6568:   PetscCount *Cperm1;
6569:   PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6570:   PetscCount *perm1prem = perm1 ? perm1 + rem : NULL;
6571:   PetscCall(PetscMalloc1(nleaves, &Cperm1));
6572:   PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));

6574:   /* Support for HYPRE matrices, kind of a hack.
6575:      Swap min column with diagonal so that diagonal values will go first */
6576:   PetscBool   hypre;
6577:   const char *name;
6578:   PetscCall(PetscObjectGetName((PetscObject)mat, &name));
6579:   PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
6580:   if (hypre) {
6581:     PetscInt *minj;
6582:     PetscBT   hasdiag;

6584:     PetscCall(PetscBTCreate(m, &hasdiag));
6585:     PetscCall(PetscMalloc1(m, &minj));
6586:     for (k = 0; k < m; k++) minj[k] = PETSC_MAX_INT;
6587:     for (k = i1start; k < rem; k++) {
6588:       if (j1[k] < cstart || j1[k] >= cend) continue;
6589:       const PetscInt rindex = i1[k] - rstart;
6590:       if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6591:       minj[rindex] = PetscMin(minj[rindex], j1[k]);
6592:     }
6593:     for (k = 0; k < n2; k++) {
6594:       if (j2[k] < cstart || j2[k] >= cend) continue;
6595:       const PetscInt rindex = i2[k] - rstart;
6596:       if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6597:       minj[rindex] = PetscMin(minj[rindex], j2[k]);
6598:     }
6599:     for (k = i1start; k < rem; k++) {
6600:       const PetscInt rindex = i1[k] - rstart;
6601:       if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6602:       if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6603:       else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6604:     }
6605:     for (k = 0; k < n2; k++) {
6606:       const PetscInt rindex = i2[k] - rstart;
6607:       if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6608:       if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6609:       else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6610:     }
6611:     PetscCall(PetscBTDestroy(&hasdiag));
6612:     PetscCall(PetscFree(minj));
6613:   }

6615:   /* Split local COOs and received COOs into diag/offdiag portions */
6616:   PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6617:   PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6618:   PetscCount  Annz1, Bnnz1, Atot1, Btot1;
6619:   PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6620:   PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6621:   PetscCount  Annz2, Bnnz2, Atot2, Btot2;

6623:   PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6624:   PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6625:   PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6626:   PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));

6628:   /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6629:   PetscInt *Ai, *Bi;
6630:   PetscInt *Aj, *Bj;

6632:   PetscCall(PetscMalloc1(m + 1, &Ai));
6633:   PetscCall(PetscMalloc1(m + 1, &Bi));
6634:   PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6635:   PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));

6637:   PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6638:   PetscCall(PetscMalloc1(Annz1, &Aimap1));
6639:   PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6640:   PetscCall(PetscMalloc1(Annz2, &Aimap2));
6641:   PetscCall(PetscMalloc1(Bnnz2, &Bimap2));

6643:   PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6644:   PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));

6646:   /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we     */
6647:   /* expect nonzeros in A/B most likely have local contributing entries        */
6648:   PetscInt    Annz = Ai[m];
6649:   PetscInt    Bnnz = Bi[m];
6650:   PetscCount *Ajmap1_new, *Bjmap1_new;

6652:   PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6653:   PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));

6655:   PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6656:   PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));

6658:   PetscCall(PetscFree(Aimap1));
6659:   PetscCall(PetscFree(Ajmap1));
6660:   PetscCall(PetscFree(Bimap1));
6661:   PetscCall(PetscFree(Bjmap1));
6662:   PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6663:   PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6664:   PetscCall(PetscFree(perm1));
6665:   PetscCall(PetscFree3(i2, j2, perm2));

6667:   Ajmap1 = Ajmap1_new;
6668:   Bjmap1 = Bjmap1_new;

6670:   /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6671:   if (Annz < Annz1 + Annz2) {
6672:     PetscInt *Aj_new;
6673:     PetscCall(PetscMalloc1(Annz, &Aj_new));
6674:     PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6675:     PetscCall(PetscFree(Aj));
6676:     Aj = Aj_new;
6677:   }

6679:   if (Bnnz < Bnnz1 + Bnnz2) {
6680:     PetscInt *Bj_new;
6681:     PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6682:     PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6683:     PetscCall(PetscFree(Bj));
6684:     Bj = Bj_new;
6685:   }

6687:   /* Create new submatrices for on-process and off-process coupling                  */
6688:   PetscScalar     *Aa, *Ba;
6689:   MatType          rtype;
6690:   Mat_SeqAIJ      *a, *b;
6691:   PetscObjectState state;
6692:   PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6693:   PetscCall(PetscCalloc1(Bnnz, &Ba));
6694:   /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6695:   if (cstart) {
6696:     for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6697:   }

6699:   PetscCall(MatGetRootType_Private(mat, &rtype));

6701:   MatSeqXAIJGetOptions_Private(mpiaij->A);
6702:   PetscCall(MatDestroy(&mpiaij->A));
6703:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6704:   PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6705:   MatSeqXAIJRestoreOptions_Private(mpiaij->A);

6707:   MatSeqXAIJGetOptions_Private(mpiaij->B);
6708:   PetscCall(MatDestroy(&mpiaij->B));
6709:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6710:   PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6711:   MatSeqXAIJRestoreOptions_Private(mpiaij->B);

6713:   PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6714:   mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6715:   state              = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6716:   PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));

6718:   a               = (Mat_SeqAIJ *)mpiaij->A->data;
6719:   b               = (Mat_SeqAIJ *)mpiaij->B->data;
6720:   a->singlemalloc = b->singlemalloc = PETSC_FALSE; /* Let newmat own Ai,Aj,Aa,Bi,Bj,Ba */
6721:   a->free_a = b->free_a = PETSC_TRUE;
6722:   a->free_ij = b->free_ij = PETSC_TRUE;

6724:   /* conversion must happen AFTER multiply setup */
6725:   PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6726:   PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6727:   PetscCall(VecDestroy(&mpiaij->lvec));
6728:   PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));

6730:   // Put the COO struct in a container and then attach that to the matrix
6731:   PetscCall(PetscMalloc1(1, &coo));
6732:   coo->n       = coo_n;
6733:   coo->sf      = sf2;
6734:   coo->sendlen = nleaves;
6735:   coo->recvlen = nroots;
6736:   coo->Annz    = Annz;
6737:   coo->Bnnz    = Bnnz;
6738:   coo->Annz2   = Annz2;
6739:   coo->Bnnz2   = Bnnz2;
6740:   coo->Atot1   = Atot1;
6741:   coo->Atot2   = Atot2;
6742:   coo->Btot1   = Btot1;
6743:   coo->Btot2   = Btot2;
6744:   coo->Ajmap1  = Ajmap1;
6745:   coo->Aperm1  = Aperm1;
6746:   coo->Bjmap1  = Bjmap1;
6747:   coo->Bperm1  = Bperm1;
6748:   coo->Aimap2  = Aimap2;
6749:   coo->Ajmap2  = Ajmap2;
6750:   coo->Aperm2  = Aperm2;
6751:   coo->Bimap2  = Bimap2;
6752:   coo->Bjmap2  = Bjmap2;
6753:   coo->Bperm2  = Bperm2;
6754:   coo->Cperm1  = Cperm1;
6755:   // Allocate in preallocation. If not used, it has zero cost on host
6756:   PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6757:   PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6758:   PetscCall(PetscContainerSetPointer(container, coo));
6759:   PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6760:   PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6761:   PetscCall(PetscContainerDestroy(&container));
6762:   PetscFunctionReturn(PETSC_SUCCESS);
6763: }

6765: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6766: {
6767:   Mat_MPIAIJ          *mpiaij = (Mat_MPIAIJ *)mat->data;
6768:   Mat                  A = mpiaij->A, B = mpiaij->B;
6769:   PetscScalar         *Aa, *Ba;
6770:   PetscScalar         *sendbuf, *recvbuf;
6771:   const PetscCount    *Ajmap1, *Ajmap2, *Aimap2;
6772:   const PetscCount    *Bjmap1, *Bjmap2, *Bimap2;
6773:   const PetscCount    *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6774:   const PetscCount    *Cperm1;
6775:   PetscContainer       container;
6776:   MatCOOStruct_MPIAIJ *coo;

6778:   PetscFunctionBegin;
6779:   PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6780:   PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6781:   PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6782:   sendbuf = coo->sendbuf;
6783:   recvbuf = coo->recvbuf;
6784:   Ajmap1  = coo->Ajmap1;
6785:   Ajmap2  = coo->Ajmap2;
6786:   Aimap2  = coo->Aimap2;
6787:   Bjmap1  = coo->Bjmap1;
6788:   Bjmap2  = coo->Bjmap2;
6789:   Bimap2  = coo->Bimap2;
6790:   Aperm1  = coo->Aperm1;
6791:   Aperm2  = coo->Aperm2;
6792:   Bperm1  = coo->Bperm1;
6793:   Bperm2  = coo->Bperm2;
6794:   Cperm1  = coo->Cperm1;

6796:   PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6797:   PetscCall(MatSeqAIJGetArray(B, &Ba));

6799:   /* Pack entries to be sent to remote */
6800:   for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];

6802:   /* Send remote entries to their owner and overlap the communication with local computation */
6803:   PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6804:   /* Add local entries to A and B */
6805:   for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6806:     PetscScalar sum = 0.0;                     /* Do partial summation first to improve numerical stability */
6807:     for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6808:     Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6809:   }
6810:   for (PetscCount i = 0; i < coo->Bnnz; i++) {
6811:     PetscScalar sum = 0.0;
6812:     for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6813:     Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6814:   }
6815:   PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));

6817:   /* Add received remote entries to A and B */
6818:   for (PetscCount i = 0; i < coo->Annz2; i++) {
6819:     for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6820:   }
6821:   for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6822:     for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6823:   }
6824:   PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6825:   PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6826:   PetscFunctionReturn(PETSC_SUCCESS);
6827: }

6829: /*MC
6830:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

6832:    Options Database Keys:
6833: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`

6835:    Level: beginner

6837:    Notes:
6838:    `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6839:     in this case the values associated with the rows and columns one passes in are set to zero
6840:     in the matrix

6842:     `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6843:     space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored

6845: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6846: M*/
6847: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6848: {
6849:   Mat_MPIAIJ *b;
6850:   PetscMPIInt size;

6852:   PetscFunctionBegin;
6853:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));

6855:   PetscCall(PetscNew(&b));
6856:   B->data       = (void *)b;
6857:   B->ops[0]     = MatOps_Values;
6858:   B->assembled  = PETSC_FALSE;
6859:   B->insertmode = NOT_SET_VALUES;
6860:   b->size       = size;

6862:   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));

6864:   /* build cache for off array entries formed */
6865:   PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));

6867:   b->donotstash  = PETSC_FALSE;
6868:   b->colmap      = NULL;
6869:   b->garray      = NULL;
6870:   b->roworiented = PETSC_TRUE;

6872:   /* stuff used for matrix vector multiply */
6873:   b->lvec  = NULL;
6874:   b->Mvctx = NULL;

6876:   /* stuff for MatGetRow() */
6877:   b->rowindices   = NULL;
6878:   b->rowvalues    = NULL;
6879:   b->getrowactive = PETSC_FALSE;

6881:   /* flexible pointer used in CUSPARSE classes */
6882:   b->spptr = NULL;

6884:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6885:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6886:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6887:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6888:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6889:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6890:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6891:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6892:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6893:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6894: #if defined(PETSC_HAVE_CUDA)
6895:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6896: #endif
6897: #if defined(PETSC_HAVE_HIP)
6898:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6899: #endif
6900: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6901:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6902: #endif
6903: #if defined(PETSC_HAVE_MKL_SPARSE)
6904:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6905: #endif
6906:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6907:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6908:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6909:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6910: #if defined(PETSC_HAVE_ELEMENTAL)
6911:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6912: #endif
6913: #if defined(PETSC_HAVE_SCALAPACK)
6914:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6915: #endif
6916:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6917:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6918: #if defined(PETSC_HAVE_HYPRE)
6919:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6920:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6921: #endif
6922:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6923:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6924:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6925:   PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6926:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6927:   PetscFunctionReturn(PETSC_SUCCESS);
6928: }

6930: /*@C
6931:   MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6932:   and "off-diagonal" part of the matrix in CSR format.

6934:   Collective

6936:   Input Parameters:
6937: + comm - MPI communicator
6938: . m    - number of local rows (Cannot be `PETSC_DECIDE`)
6939: . n    - This value should be the same as the local size used in creating the
6940:          x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6941:          calculated if `N` is given) For square matrices `n` is almost always `m`.
6942: . M    - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6943: . N    - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6944: . 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
6945: . j    - column indices, which must be local, i.e., based off the start column of the diagonal portion
6946: . a    - matrix values
6947: . 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
6948: . oj   - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6949: - oa   - matrix values

6951:   Output Parameter:
6952: . mat - the matrix

6954:   Level: advanced

6956:   Notes:
6957:   The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6958:   must free the arrays once the matrix has been destroyed and not before.

6960:   The `i` and `j` indices are 0 based

6962:   See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix

6964:   This sets local rows and cannot be used to set off-processor values.

6966:   Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6967:   legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6968:   not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6969:   the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6970:   keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6971:   communication if it is known that only local entries will be set.

6973: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6974:           `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6975: @*/
6976: 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)
6977: {
6978:   Mat_MPIAIJ *maij;

6980:   PetscFunctionBegin;
6981:   PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6982:   PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6983:   PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6984:   PetscCall(MatCreate(comm, mat));
6985:   PetscCall(MatSetSizes(*mat, m, n, M, N));
6986:   PetscCall(MatSetType(*mat, MATMPIAIJ));
6987:   maij = (Mat_MPIAIJ *)(*mat)->data;

6989:   (*mat)->preallocated = PETSC_TRUE;

6991:   PetscCall(PetscLayoutSetUp((*mat)->rmap));
6992:   PetscCall(PetscLayoutSetUp((*mat)->cmap));

6994:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6995:   PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));

6997:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6998:   PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
6999:   PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7000:   PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7001:   PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7002:   PetscFunctionReturn(PETSC_SUCCESS);
7003: }

7005: typedef struct {
7006:   Mat       *mp;    /* intermediate products */
7007:   PetscBool *mptmp; /* is the intermediate product temporary ? */
7008:   PetscInt   cp;    /* number of intermediate products */

7010:   /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7011:   PetscInt    *startsj_s, *startsj_r;
7012:   PetscScalar *bufa;
7013:   Mat          P_oth;

7015:   /* may take advantage of merging product->B */
7016:   Mat Bloc; /* B-local by merging diag and off-diag */

7018:   /* cusparse does not have support to split between symbolic and numeric phases.
7019:      When api_user is true, we don't need to update the numerical values
7020:      of the temporary storage */
7021:   PetscBool reusesym;

7023:   /* support for COO values insertion */
7024:   PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7025:   PetscInt   **own;           /* own[i] points to address of on-process COO indices for Mat mp[i] */
7026:   PetscInt   **off;           /* off[i] points to address of off-process COO indices for Mat mp[i] */
7027:   PetscBool    hasoffproc;    /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7028:   PetscSF      sf;            /* used for non-local values insertion and memory malloc */
7029:   PetscMemType mtype;

7031:   /* customization */
7032:   PetscBool abmerge;
7033:   PetscBool P_oth_bind;
7034: } MatMatMPIAIJBACKEND;

7036: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7037: {
7038:   MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7039:   PetscInt             i;

7041:   PetscFunctionBegin;
7042:   PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7043:   PetscCall(PetscFree(mmdata->bufa));
7044:   PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7045:   PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7046:   PetscCall(MatDestroy(&mmdata->P_oth));
7047:   PetscCall(MatDestroy(&mmdata->Bloc));
7048:   PetscCall(PetscSFDestroy(&mmdata->sf));
7049:   for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7050:   PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7051:   PetscCall(PetscFree(mmdata->own[0]));
7052:   PetscCall(PetscFree(mmdata->own));
7053:   PetscCall(PetscFree(mmdata->off[0]));
7054:   PetscCall(PetscFree(mmdata->off));
7055:   PetscCall(PetscFree(mmdata));
7056:   PetscFunctionReturn(PETSC_SUCCESS);
7057: }

7059: /* Copy selected n entries with indices in idx[] of A to v[].
7060:    If idx is NULL, copy the whole data array of A to v[]
7061:  */
7062: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7063: {
7064:   PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);

7066:   PetscFunctionBegin;
7067:   PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7068:   if (f) {
7069:     PetscCall((*f)(A, n, idx, v));
7070:   } else {
7071:     const PetscScalar *vv;

7073:     PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7074:     if (n && idx) {
7075:       PetscScalar    *w  = v;
7076:       const PetscInt *oi = idx;
7077:       PetscInt        j;

7079:       for (j = 0; j < n; j++) *w++ = vv[*oi++];
7080:     } else {
7081:       PetscCall(PetscArraycpy(v, vv, n));
7082:     }
7083:     PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7084:   }
7085:   PetscFunctionReturn(PETSC_SUCCESS);
7086: }

7088: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7089: {
7090:   MatMatMPIAIJBACKEND *mmdata;
7091:   PetscInt             i, n_d, n_o;

7093:   PetscFunctionBegin;
7094:   MatCheckProduct(C, 1);
7095:   PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7096:   mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7097:   if (!mmdata->reusesym) { /* update temporary matrices */
7098:     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));
7099:     if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7100:   }
7101:   mmdata->reusesym = PETSC_FALSE;

7103:   for (i = 0; i < mmdata->cp; i++) {
7104:     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]);
7105:     PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7106:   }
7107:   for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7108:     PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];

7110:     if (mmdata->mptmp[i]) continue;
7111:     if (noff) {
7112:       PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];

7114:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7115:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7116:       n_o += noff;
7117:       n_d += nown;
7118:     } else {
7119:       Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;

7121:       PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7122:       n_d += mm->nz;
7123:     }
7124:   }
7125:   if (mmdata->hasoffproc) { /* offprocess insertion */
7126:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7127:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7128:   }
7129:   PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7130:   PetscFunctionReturn(PETSC_SUCCESS);
7131: }

7133: /* Support for Pt * A, A * P, or Pt * A * P */
7134: #define MAX_NUMBER_INTERMEDIATE 4
7135: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7136: {
7137:   Mat_Product           *product = C->product;
7138:   Mat                    A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7139:   Mat_MPIAIJ            *a, *p;
7140:   MatMatMPIAIJBACKEND   *mmdata;
7141:   ISLocalToGlobalMapping P_oth_l2g = NULL;
7142:   IS                     glob      = NULL;
7143:   const char            *prefix;
7144:   char                   pprefix[256];
7145:   const PetscInt        *globidx, *P_oth_idx;
7146:   PetscInt               i, j, cp, m, n, M, N, *coo_i, *coo_j;
7147:   PetscCount             ncoo, ncoo_d, ncoo_o, ncoo_oown;
7148:   PetscInt               cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7149:                                                                                          /* type-0: consecutive, start from 0; type-1: consecutive with */
7150:                                                                                          /* a base offset; type-2: sparse with a local to global map table */
7151:   const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE];       /* col/row local to global map array (table) for type-2 map type */

7153:   MatProductType ptype;
7154:   PetscBool      mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7155:   PetscMPIInt    size;

7157:   PetscFunctionBegin;
7158:   MatCheckProduct(C, 1);
7159:   PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7160:   ptype = product->type;
7161:   if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7162:     ptype                                          = MATPRODUCT_AB;
7163:     product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7164:   }
7165:   switch (ptype) {
7166:   case MATPRODUCT_AB:
7167:     A          = product->A;
7168:     P          = product->B;
7169:     m          = A->rmap->n;
7170:     n          = P->cmap->n;
7171:     M          = A->rmap->N;
7172:     N          = P->cmap->N;
7173:     hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7174:     break;
7175:   case MATPRODUCT_AtB:
7176:     P          = product->A;
7177:     A          = product->B;
7178:     m          = P->cmap->n;
7179:     n          = A->cmap->n;
7180:     M          = P->cmap->N;
7181:     N          = A->cmap->N;
7182:     hasoffproc = PETSC_TRUE;
7183:     break;
7184:   case MATPRODUCT_PtAP:
7185:     A          = product->A;
7186:     P          = product->B;
7187:     m          = P->cmap->n;
7188:     n          = P->cmap->n;
7189:     M          = P->cmap->N;
7190:     N          = P->cmap->N;
7191:     hasoffproc = PETSC_TRUE;
7192:     break;
7193:   default:
7194:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7195:   }
7196:   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7197:   if (size == 1) hasoffproc = PETSC_FALSE;

7199:   /* defaults */
7200:   for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7201:     mp[i]    = NULL;
7202:     mptmp[i] = PETSC_FALSE;
7203:     rmapt[i] = -1;
7204:     cmapt[i] = -1;
7205:     rmapa[i] = NULL;
7206:     cmapa[i] = NULL;
7207:   }

7209:   /* customization */
7210:   PetscCall(PetscNew(&mmdata));
7211:   mmdata->reusesym = product->api_user;
7212:   if (ptype == MATPRODUCT_AB) {
7213:     if (product->api_user) {
7214:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7215:       PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7216:       PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7217:       PetscOptionsEnd();
7218:     } else {
7219:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7220:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7221:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7222:       PetscOptionsEnd();
7223:     }
7224:   } else if (ptype == MATPRODUCT_PtAP) {
7225:     if (product->api_user) {
7226:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7227:       PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7228:       PetscOptionsEnd();
7229:     } else {
7230:       PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7231:       PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7232:       PetscOptionsEnd();
7233:     }
7234:   }
7235:   a = (Mat_MPIAIJ *)A->data;
7236:   p = (Mat_MPIAIJ *)P->data;
7237:   PetscCall(MatSetSizes(C, m, n, M, N));
7238:   PetscCall(PetscLayoutSetUp(C->rmap));
7239:   PetscCall(PetscLayoutSetUp(C->cmap));
7240:   PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7241:   PetscCall(MatGetOptionsPrefix(C, &prefix));

7243:   cp = 0;
7244:   switch (ptype) {
7245:   case MATPRODUCT_AB: /* A * P */
7246:     PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));

7248:     /* A_diag * P_local (merged or not) */
7249:     if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7250:       /* P is product->B */
7251:       PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7252:       PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7253:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7254:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7255:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7256:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7257:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7258:       mp[cp]->product->api_user = product->api_user;
7259:       PetscCall(MatProductSetFromOptions(mp[cp]));
7260:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7261:       PetscCall(ISGetIndices(glob, &globidx));
7262:       rmapt[cp] = 1;
7263:       cmapt[cp] = 2;
7264:       cmapa[cp] = globidx;
7265:       mptmp[cp] = PETSC_FALSE;
7266:       cp++;
7267:     } else { /* A_diag * P_diag and A_diag * P_off */
7268:       PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7269:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7270:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7271:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7272:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7273:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7274:       mp[cp]->product->api_user = product->api_user;
7275:       PetscCall(MatProductSetFromOptions(mp[cp]));
7276:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7277:       rmapt[cp] = 1;
7278:       cmapt[cp] = 1;
7279:       mptmp[cp] = PETSC_FALSE;
7280:       cp++;
7281:       PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7282:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7283:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7284:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7285:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7286:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7287:       mp[cp]->product->api_user = product->api_user;
7288:       PetscCall(MatProductSetFromOptions(mp[cp]));
7289:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7290:       rmapt[cp] = 1;
7291:       cmapt[cp] = 2;
7292:       cmapa[cp] = p->garray;
7293:       mptmp[cp] = PETSC_FALSE;
7294:       cp++;
7295:     }

7297:     /* A_off * P_other */
7298:     if (mmdata->P_oth) {
7299:       PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7300:       PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7301:       PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7302:       PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7303:       PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7304:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7305:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7306:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7307:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7308:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7309:       mp[cp]->product->api_user = product->api_user;
7310:       PetscCall(MatProductSetFromOptions(mp[cp]));
7311:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7312:       rmapt[cp] = 1;
7313:       cmapt[cp] = 2;
7314:       cmapa[cp] = P_oth_idx;
7315:       mptmp[cp] = PETSC_FALSE;
7316:       cp++;
7317:     }
7318:     break;

7320:   case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7321:     /* A is product->B */
7322:     PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7323:     if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7324:       PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7325:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7326:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7327:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7328:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7329:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7330:       mp[cp]->product->api_user = product->api_user;
7331:       PetscCall(MatProductSetFromOptions(mp[cp]));
7332:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7333:       PetscCall(ISGetIndices(glob, &globidx));
7334:       rmapt[cp] = 2;
7335:       rmapa[cp] = globidx;
7336:       cmapt[cp] = 2;
7337:       cmapa[cp] = globidx;
7338:       mptmp[cp] = PETSC_FALSE;
7339:       cp++;
7340:     } else {
7341:       PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7342:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7343:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7344:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7345:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7346:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7347:       mp[cp]->product->api_user = product->api_user;
7348:       PetscCall(MatProductSetFromOptions(mp[cp]));
7349:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7350:       PetscCall(ISGetIndices(glob, &globidx));
7351:       rmapt[cp] = 1;
7352:       cmapt[cp] = 2;
7353:       cmapa[cp] = globidx;
7354:       mptmp[cp] = PETSC_FALSE;
7355:       cp++;
7356:       PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7357:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7358:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7359:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7360:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7361:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7362:       mp[cp]->product->api_user = product->api_user;
7363:       PetscCall(MatProductSetFromOptions(mp[cp]));
7364:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7365:       rmapt[cp] = 2;
7366:       rmapa[cp] = p->garray;
7367:       cmapt[cp] = 2;
7368:       cmapa[cp] = globidx;
7369:       mptmp[cp] = PETSC_FALSE;
7370:       cp++;
7371:     }
7372:     break;
7373:   case MATPRODUCT_PtAP:
7374:     PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7375:     /* P is product->B */
7376:     PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7377:     PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7378:     PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7379:     PetscCall(MatProductSetFill(mp[cp], product->fill));
7380:     PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7381:     PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7382:     PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7383:     mp[cp]->product->api_user = product->api_user;
7384:     PetscCall(MatProductSetFromOptions(mp[cp]));
7385:     PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7386:     PetscCall(ISGetIndices(glob, &globidx));
7387:     rmapt[cp] = 2;
7388:     rmapa[cp] = globidx;
7389:     cmapt[cp] = 2;
7390:     cmapa[cp] = globidx;
7391:     mptmp[cp] = PETSC_FALSE;
7392:     cp++;
7393:     if (mmdata->P_oth) {
7394:       PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7395:       PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7396:       PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7397:       PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7398:       PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7399:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7400:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7401:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7402:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7403:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7404:       mp[cp]->product->api_user = product->api_user;
7405:       PetscCall(MatProductSetFromOptions(mp[cp]));
7406:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7407:       mptmp[cp] = PETSC_TRUE;
7408:       cp++;
7409:       PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7410:       PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7411:       PetscCall(MatProductSetFill(mp[cp], product->fill));
7412:       PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7413:       PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7414:       PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7415:       mp[cp]->product->api_user = product->api_user;
7416:       PetscCall(MatProductSetFromOptions(mp[cp]));
7417:       PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7418:       rmapt[cp] = 2;
7419:       rmapa[cp] = globidx;
7420:       cmapt[cp] = 2;
7421:       cmapa[cp] = P_oth_idx;
7422:       mptmp[cp] = PETSC_FALSE;
7423:       cp++;
7424:     }
7425:     break;
7426:   default:
7427:     SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7428:   }
7429:   /* sanity check */
7430:   if (size > 1)
7431:     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);

7433:   PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7434:   for (i = 0; i < cp; i++) {
7435:     mmdata->mp[i]    = mp[i];
7436:     mmdata->mptmp[i] = mptmp[i];
7437:   }
7438:   mmdata->cp             = cp;
7439:   C->product->data       = mmdata;
7440:   C->product->destroy    = MatDestroy_MatMatMPIAIJBACKEND;
7441:   C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;

7443:   /* memory type */
7444:   mmdata->mtype = PETSC_MEMTYPE_HOST;
7445:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7446:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7447:   PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7448:   if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7449:   else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7450:   else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;

7452:   /* prepare coo coordinates for values insertion */

7454:   /* count total nonzeros of those intermediate seqaij Mats
7455:     ncoo_d:    # of nonzeros of matrices that do not have offproc entries
7456:     ncoo_o:    # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7457:     ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7458:   */
7459:   for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7460:     Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7461:     if (mptmp[cp]) continue;
7462:     if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7463:       const PetscInt *rmap = rmapa[cp];
7464:       const PetscInt  mr   = mp[cp]->rmap->n;
7465:       const PetscInt  rs   = C->rmap->rstart;
7466:       const PetscInt  re   = C->rmap->rend;
7467:       const PetscInt *ii   = mm->i;
7468:       for (i = 0; i < mr; i++) {
7469:         const PetscInt gr = rmap[i];
7470:         const PetscInt nz = ii[i + 1] - ii[i];
7471:         if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7472:         else ncoo_oown += nz;                  /* this row is local */
7473:       }
7474:     } else ncoo_d += mm->nz;
7475:   }

7477:   /*
7478:     ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc

7480:     ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.

7482:     off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].

7484:     off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7485:     own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7486:     so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.

7488:     coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7489:     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.
7490:   */
7491:   PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7492:   PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));

7494:   /* gather (i,j) of nonzeros inserted by remote procs */
7495:   if (hasoffproc) {
7496:     PetscSF  msf;
7497:     PetscInt ncoo2, *coo_i2, *coo_j2;

7499:     PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7500:     PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7501:     PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */

7503:     for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7504:       Mat_SeqAIJ *mm     = (Mat_SeqAIJ *)mp[cp]->data;
7505:       PetscInt   *idxoff = mmdata->off[cp];
7506:       PetscInt   *idxown = mmdata->own[cp];
7507:       if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7508:         const PetscInt *rmap = rmapa[cp];
7509:         const PetscInt *cmap = cmapa[cp];
7510:         const PetscInt *ii   = mm->i;
7511:         PetscInt       *coi  = coo_i + ncoo_o;
7512:         PetscInt       *coj  = coo_j + ncoo_o;
7513:         const PetscInt  mr   = mp[cp]->rmap->n;
7514:         const PetscInt  rs   = C->rmap->rstart;
7515:         const PetscInt  re   = C->rmap->rend;
7516:         const PetscInt  cs   = C->cmap->rstart;
7517:         for (i = 0; i < mr; i++) {
7518:           const PetscInt *jj = mm->j + ii[i];
7519:           const PetscInt  gr = rmap[i];
7520:           const PetscInt  nz = ii[i + 1] - ii[i];
7521:           if (gr < rs || gr >= re) { /* this is an offproc row */
7522:             for (j = ii[i]; j < ii[i + 1]; j++) {
7523:               *coi++    = gr;
7524:               *idxoff++ = j;
7525:             }
7526:             if (!cmapt[cp]) { /* already global */
7527:               for (j = 0; j < nz; j++) *coj++ = jj[j];
7528:             } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7529:               for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7530:             } else { /* offdiag */
7531:               for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7532:             }
7533:             ncoo_o += nz;
7534:           } else { /* this is a local row */
7535:             for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7536:           }
7537:         }
7538:       }
7539:       mmdata->off[cp + 1] = idxoff;
7540:       mmdata->own[cp + 1] = idxown;
7541:     }

7543:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7544:     PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7545:     PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7546:     PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7547:     ncoo = ncoo_d + ncoo_oown + ncoo2;
7548:     PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7549:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7550:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7551:     PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7552:     PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7553:     PetscCall(PetscFree2(coo_i, coo_j));
7554:     /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7555:     PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7556:     coo_i = coo_i2;
7557:     coo_j = coo_j2;
7558:   } else { /* no offproc values insertion */
7559:     ncoo = ncoo_d;
7560:     PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));

7562:     PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7563:     PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7564:     PetscCall(PetscSFSetUp(mmdata->sf));
7565:   }
7566:   mmdata->hasoffproc = hasoffproc;

7568:   /* gather (i,j) of nonzeros inserted locally */
7569:   for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7570:     Mat_SeqAIJ     *mm   = (Mat_SeqAIJ *)mp[cp]->data;
7571:     PetscInt       *coi  = coo_i + ncoo_d;
7572:     PetscInt       *coj  = coo_j + ncoo_d;
7573:     const PetscInt *jj   = mm->j;
7574:     const PetscInt *ii   = mm->i;
7575:     const PetscInt *cmap = cmapa[cp];
7576:     const PetscInt *rmap = rmapa[cp];
7577:     const PetscInt  mr   = mp[cp]->rmap->n;
7578:     const PetscInt  rs   = C->rmap->rstart;
7579:     const PetscInt  re   = C->rmap->rend;
7580:     const PetscInt  cs   = C->cmap->rstart;

7582:     if (mptmp[cp]) continue;
7583:     if (rmapt[cp] == 1) { /* consecutive rows */
7584:       /* fill coo_i */
7585:       for (i = 0; i < mr; i++) {
7586:         const PetscInt gr = i + rs;
7587:         for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7588:       }
7589:       /* fill coo_j */
7590:       if (!cmapt[cp]) { /* type-0, already global */
7591:         PetscCall(PetscArraycpy(coj, jj, mm->nz));
7592:       } else if (cmapt[cp] == 1) {                        /* type-1, local to global for consecutive columns of C */
7593:         for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7594:       } else {                                            /* type-2, local to global for sparse columns */
7595:         for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7596:       }
7597:       ncoo_d += mm->nz;
7598:     } else if (rmapt[cp] == 2) { /* sparse rows */
7599:       for (i = 0; i < mr; i++) {
7600:         const PetscInt *jj = mm->j + ii[i];
7601:         const PetscInt  gr = rmap[i];
7602:         const PetscInt  nz = ii[i + 1] - ii[i];
7603:         if (gr >= rs && gr < re) { /* local rows */
7604:           for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7605:           if (!cmapt[cp]) { /* type-0, already global */
7606:             for (j = 0; j < nz; j++) *coj++ = jj[j];
7607:           } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7608:             for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7609:           } else { /* type-2, local to global for sparse columns */
7610:             for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7611:           }
7612:           ncoo_d += nz;
7613:         }
7614:       }
7615:     }
7616:   }
7617:   if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7618:   PetscCall(ISDestroy(&glob));
7619:   if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7620:   PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7621:   /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7622:   PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));

7624:   /* preallocate with COO data */
7625:   PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7626:   PetscCall(PetscFree2(coo_i, coo_j));
7627:   PetscFunctionReturn(PETSC_SUCCESS);
7628: }

7630: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7631: {
7632:   Mat_Product *product = mat->product;
7633: #if defined(PETSC_HAVE_DEVICE)
7634:   PetscBool match  = PETSC_FALSE;
7635:   PetscBool usecpu = PETSC_FALSE;
7636: #else
7637:   PetscBool match = PETSC_TRUE;
7638: #endif

7640:   PetscFunctionBegin;
7641:   MatCheckProduct(mat, 1);
7642: #if defined(PETSC_HAVE_DEVICE)
7643:   if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7644:   if (match) { /* we can always fallback to the CPU if requested */
7645:     switch (product->type) {
7646:     case MATPRODUCT_AB:
7647:       if (product->api_user) {
7648:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7649:         PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7650:         PetscOptionsEnd();
7651:       } else {
7652:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7653:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7654:         PetscOptionsEnd();
7655:       }
7656:       break;
7657:     case MATPRODUCT_AtB:
7658:       if (product->api_user) {
7659:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7660:         PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7661:         PetscOptionsEnd();
7662:       } else {
7663:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7664:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7665:         PetscOptionsEnd();
7666:       }
7667:       break;
7668:     case MATPRODUCT_PtAP:
7669:       if (product->api_user) {
7670:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7671:         PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7672:         PetscOptionsEnd();
7673:       } else {
7674:         PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7675:         PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7676:         PetscOptionsEnd();
7677:       }
7678:       break;
7679:     default:
7680:       break;
7681:     }
7682:     match = (PetscBool)!usecpu;
7683:   }
7684: #endif
7685:   if (match) {
7686:     switch (product->type) {
7687:     case MATPRODUCT_AB:
7688:     case MATPRODUCT_AtB:
7689:     case MATPRODUCT_PtAP:
7690:       mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7691:       break;
7692:     default:
7693:       break;
7694:     }
7695:   }
7696:   /* fallback to MPIAIJ ops */
7697:   if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7698:   PetscFunctionReturn(PETSC_SUCCESS);
7699: }

7701: /*
7702:    Produces a set of block column indices of the matrix row, one for each block represented in the original row

7704:    n - the number of block indices in cc[]
7705:    cc - the block indices (must be large enough to contain the indices)
7706: */
7707: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7708: {
7709:   PetscInt        cnt = -1, nidx, j;
7710:   const PetscInt *idx;

7712:   PetscFunctionBegin;
7713:   PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7714:   if (nidx) {
7715:     cnt     = 0;
7716:     cc[cnt] = idx[0] / bs;
7717:     for (j = 1; j < nidx; j++) {
7718:       if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7719:     }
7720:   }
7721:   PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7722:   *n = cnt + 1;
7723:   PetscFunctionReturn(PETSC_SUCCESS);
7724: }

7726: /*
7727:     Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows

7729:     ncollapsed - the number of block indices
7730:     collapsed - the block indices (must be large enough to contain the indices)
7731: */
7732: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7733: {
7734:   PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;

7736:   PetscFunctionBegin;
7737:   PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7738:   for (i = start + 1; i < start + bs; i++) {
7739:     PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7740:     PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7741:     cprevtmp = cprev;
7742:     cprev    = merged;
7743:     merged   = cprevtmp;
7744:   }
7745:   *ncollapsed = nprev;
7746:   if (collapsed) *collapsed = cprev;
7747:   PetscFunctionReturn(PETSC_SUCCESS);
7748: }

7750: /*
7751:  MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix

7753:  Input Parameter:
7754:  . Amat - matrix
7755:  - symmetrize - make the result symmetric
7756:  + scale - scale with diagonal

7758:  Output Parameter:
7759:  . a_Gmat - output scalar graph >= 0

7761: */
7762: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7763: {
7764:   PetscInt  Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7765:   MPI_Comm  comm;
7766:   Mat       Gmat;
7767:   PetscBool ismpiaij, isseqaij;
7768:   Mat       a, b, c;
7769:   MatType   jtype;

7771:   PetscFunctionBegin;
7772:   PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7773:   PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7774:   PetscCall(MatGetSize(Amat, &MM, &NN));
7775:   PetscCall(MatGetBlockSize(Amat, &bs));
7776:   nloc = (Iend - Istart) / bs;

7778:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7779:   PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7780:   PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");

7782:   /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7783:   /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7784:      implementation */
7785:   if (bs > 1) {
7786:     PetscCall(MatGetType(Amat, &jtype));
7787:     PetscCall(MatCreate(comm, &Gmat));
7788:     PetscCall(MatSetType(Gmat, jtype));
7789:     PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7790:     PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7791:     if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7792:       PetscInt  *d_nnz, *o_nnz;
7793:       MatScalar *aa, val, *AA;
7794:       PetscInt  *aj, *ai, *AJ, nc, nmax = 0;
7795:       if (isseqaij) {
7796:         a = Amat;
7797:         b = NULL;
7798:       } else {
7799:         Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7800:         a             = d->A;
7801:         b             = d->B;
7802:       }
7803:       PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7804:       PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7805:       for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7806:         PetscInt       *nnz = (c == a) ? d_nnz : o_nnz;
7807:         const PetscInt *cols1, *cols2;
7808:         for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7809:           PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7810:           nnz[brow / bs] = nc2 / bs;
7811:           if (nc2 % bs) ok = 0;
7812:           if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7813:           for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7814:             PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7815:             if (nc1 != nc2) ok = 0;
7816:             else {
7817:               for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7818:                 if (cols1[jj] != cols2[jj]) ok = 0;
7819:                 if (cols1[jj] % bs != jj % bs) ok = 0;
7820:               }
7821:             }
7822:             PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7823:           }
7824:           PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7825:           if (!ok) {
7826:             PetscCall(PetscFree2(d_nnz, o_nnz));
7827:             PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7828:             goto old_bs;
7829:           }
7830:         }
7831:       }
7832:       PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7833:       PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7834:       PetscCall(PetscFree2(d_nnz, o_nnz));
7835:       PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7836:       // diag
7837:       for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7838:         Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7839:         ai               = aseq->i;
7840:         n                = ai[brow + 1] - ai[brow];
7841:         aj               = aseq->j + ai[brow];
7842:         for (int k = 0; k < n; k += bs) {        // block columns
7843:           AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7844:           val        = 0;
7845:           if (index_size == 0) {
7846:             for (int ii = 0; ii < bs; ii++) { // rows in block
7847:               aa = aseq->a + ai[brow + ii] + k;
7848:               for (int jj = 0; jj < bs; jj++) {         // columns in block
7849:                 val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7850:               }
7851:             }
7852:           } else {                                       // use (index,index) value if provided
7853:             for (int iii = 0; iii < index_size; iii++) { // rows in block
7854:               int ii = index[iii];
7855:               aa     = aseq->a + ai[brow + ii] + k;
7856:               for (int jjj = 0; jjj < index_size; jjj++) { // columns in block
7857:                 int jj = index[jjj];
7858:                 val += PetscAbs(PetscRealPart(aa[jj]));
7859:               }
7860:             }
7861:           }
7862:           PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7863:           AA[k / bs] = val;
7864:         }
7865:         grow = Istart / bs + brow / bs;
7866:         PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, INSERT_VALUES));
7867:       }
7868:       // off-diag
7869:       if (ismpiaij) {
7870:         Mat_MPIAIJ        *aij = (Mat_MPIAIJ *)Amat->data;
7871:         const PetscScalar *vals;
7872:         const PetscInt    *cols, *garray = aij->garray;
7873:         PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7874:         for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7875:           PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7876:           for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7877:             PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7878:             AA[k / bs] = 0;
7879:             AJ[cidx]   = garray[cols[k]] / bs;
7880:           }
7881:           nc = ncols / bs;
7882:           PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7883:           if (index_size == 0) {
7884:             for (int ii = 0; ii < bs; ii++) { // rows in block
7885:               PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7886:               for (int k = 0; k < ncols; k += bs) {
7887:                 for (int jj = 0; jj < bs; jj++) { // cols in block
7888:                   PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7889:                   AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7890:                 }
7891:               }
7892:               PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7893:             }
7894:           } else {                                       // use (index,index) value if provided
7895:             for (int iii = 0; iii < index_size; iii++) { // rows in block
7896:               int ii = index[iii];
7897:               PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7898:               for (int k = 0; k < ncols; k += bs) {
7899:                 for (int jjj = 0; jjj < index_size; jjj++) { // cols in block
7900:                   int jj = index[jjj];
7901:                   AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7902:                 }
7903:               }
7904:               PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7905:             }
7906:           }
7907:           grow = Istart / bs + brow / bs;
7908:           PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, INSERT_VALUES));
7909:         }
7910:       }
7911:       PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7912:       PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7913:       PetscCall(PetscFree2(AA, AJ));
7914:     } else {
7915:       const PetscScalar *vals;
7916:       const PetscInt    *idx;
7917:       PetscInt          *d_nnz, *o_nnz, *w0, *w1, *w2;
7918:     old_bs:
7919:       /*
7920:        Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7921:        */
7922:       PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7923:       PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7924:       if (isseqaij) {
7925:         PetscInt max_d_nnz;
7926:         /*
7927:          Determine exact preallocation count for (sequential) scalar matrix
7928:          */
7929:         PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7930:         max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7931:         PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7932:         for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7933:         PetscCall(PetscFree3(w0, w1, w2));
7934:       } else if (ismpiaij) {
7935:         Mat             Daij, Oaij;
7936:         const PetscInt *garray;
7937:         PetscInt        max_d_nnz;
7938:         PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7939:         /*
7940:          Determine exact preallocation count for diagonal block portion of scalar matrix
7941:          */
7942:         PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7943:         max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7944:         PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7945:         for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7946:         PetscCall(PetscFree3(w0, w1, w2));
7947:         /*
7948:          Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7949:          */
7950:         for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7951:           o_nnz[jj] = 0;
7952:           for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7953:             PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7954:             o_nnz[jj] += ncols;
7955:             PetscCall(