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));
167: PetscFunctionReturn(PETSC_SUCCESS);
168: }
170: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
171: {
172: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
174: PetscFunctionBegin;
175: if (mat->A) {
176: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
177: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
178: }
179: PetscFunctionReturn(PETSC_SUCCESS);
180: }
182: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
183: {
184: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
185: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
186: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
187: const PetscInt *ia, *ib;
188: const MatScalar *aa, *bb, *aav, *bav;
189: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
190: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
192: PetscFunctionBegin;
193: *keptrows = NULL;
195: ia = a->i;
196: ib = b->i;
197: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
198: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
199: for (i = 0; i < m; i++) {
200: na = ia[i + 1] - ia[i];
201: nb = ib[i + 1] - ib[i];
202: if (!na && !nb) {
203: cnt++;
204: goto ok1;
205: }
206: aa = aav + ia[i];
207: for (j = 0; j < na; j++) {
208: if (aa[j] != 0.0) goto ok1;
209: }
210: bb = bav ? bav + ib[i] : NULL;
211: for (j = 0; j < nb; j++) {
212: if (bb[j] != 0.0) goto ok1;
213: }
214: cnt++;
215: ok1:;
216: }
217: PetscCall(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
218: if (!n0rows) {
219: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
220: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
221: PetscFunctionReturn(PETSC_SUCCESS);
222: }
223: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
224: cnt = 0;
225: for (i = 0; i < m; i++) {
226: na = ia[i + 1] - ia[i];
227: nb = ib[i + 1] - ib[i];
228: if (!na && !nb) continue;
229: aa = aav + ia[i];
230: for (j = 0; j < na; j++) {
231: if (aa[j] != 0.0) {
232: rows[cnt++] = rstart + i;
233: goto ok2;
234: }
235: }
236: bb = bav ? bav + ib[i] : NULL;
237: for (j = 0; j < nb; j++) {
238: if (bb[j] != 0.0) {
239: rows[cnt++] = rstart + i;
240: goto ok2;
241: }
242: }
243: ok2:;
244: }
245: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
246: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
247: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
248: PetscFunctionReturn(PETSC_SUCCESS);
249: }
251: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
252: {
253: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
254: PetscBool cong;
256: PetscFunctionBegin;
257: PetscCall(MatHasCongruentLayouts(Y, &cong));
258: if (Y->assembled && cong) {
259: PetscCall(MatDiagonalSet(aij->A, D, is));
260: } else {
261: PetscCall(MatDiagonalSet_Default(Y, D, is));
262: }
263: PetscFunctionReturn(PETSC_SUCCESS);
264: }
266: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
267: {
268: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
269: PetscInt i, rstart, nrows, *rows;
271: PetscFunctionBegin;
272: *zrows = NULL;
273: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
274: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
275: for (i = 0; i < nrows; i++) rows[i] += rstart;
276: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
277: PetscFunctionReturn(PETSC_SUCCESS);
278: }
280: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
281: {
282: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
283: PetscInt i, m, n, *garray = aij->garray;
284: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
285: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
286: PetscReal *work;
287: const PetscScalar *dummy;
289: PetscFunctionBegin;
290: PetscCall(MatGetSize(A, &m, &n));
291: PetscCall(PetscCalloc1(n, &work));
292: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
293: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
294: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
295: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
296: if (type == NORM_2) {
297: 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]);
298: 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]);
299: } else if (type == NORM_1) {
300: 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]);
301: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
302: } else if (type == NORM_INFINITY) {
303: 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]]);
304: 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]]]);
305: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
306: 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]);
307: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
308: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
309: 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]);
310: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
311: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
312: if (type == NORM_INFINITY) {
313: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
314: } else {
315: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
316: }
317: PetscCall(PetscFree(work));
318: if (type == NORM_2) {
319: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
320: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
321: for (i = 0; i < n; i++) reductions[i] /= m;
322: }
323: PetscFunctionReturn(PETSC_SUCCESS);
324: }
326: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
327: {
328: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
329: IS sis, gis;
330: const PetscInt *isis, *igis;
331: PetscInt n, *iis, nsis, ngis, rstart, i;
333: PetscFunctionBegin;
334: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
335: PetscCall(MatFindNonzeroRows(a->B, &gis));
336: PetscCall(ISGetSize(gis, &ngis));
337: PetscCall(ISGetSize(sis, &nsis));
338: PetscCall(ISGetIndices(sis, &isis));
339: PetscCall(ISGetIndices(gis, &igis));
341: PetscCall(PetscMalloc1(ngis + nsis, &iis));
342: PetscCall(PetscArraycpy(iis, igis, ngis));
343: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
344: n = ngis + nsis;
345: PetscCall(PetscSortRemoveDupsInt(&n, iis));
346: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
347: for (i = 0; i < n; i++) iis[i] += rstart;
348: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
350: PetscCall(ISRestoreIndices(sis, &isis));
351: PetscCall(ISRestoreIndices(gis, &igis));
352: PetscCall(ISDestroy(&sis));
353: PetscCall(ISDestroy(&gis));
354: PetscFunctionReturn(PETSC_SUCCESS);
355: }
357: /*
358: Local utility routine that creates a mapping from the global column
359: number to the local number in the off-diagonal part of the local
360: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
361: a slightly higher hash table cost; without it it is not scalable (each processor
362: has an order N integer array but is fast to access.
363: */
364: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
365: {
366: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
367: PetscInt n = aij->B->cmap->n, i;
369: PetscFunctionBegin;
370: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
371: #if defined(PETSC_USE_CTABLE)
372: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
373: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
374: #else
375: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
376: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
377: #endif
378: PetscFunctionReturn(PETSC_SUCCESS);
379: }
381: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
382: do { \
383: if (col <= lastcol1) low1 = 0; \
384: else high1 = nrow1; \
385: lastcol1 = col; \
386: while (high1 - low1 > 5) { \
387: t = (low1 + high1) / 2; \
388: if (rp1[t] > col) high1 = t; \
389: else low1 = t; \
390: } \
391: for (_i = low1; _i < high1; _i++) { \
392: if (rp1[_i] > col) break; \
393: if (rp1[_i] == col) { \
394: if (addv == ADD_VALUES) { \
395: ap1[_i] += value; \
396: /* Not sure LogFlops will slow dow the code or not */ \
397: (void)PetscLogFlops(1.0); \
398: } else ap1[_i] = value; \
399: goto a_noinsert; \
400: } \
401: } \
402: if (value == 0.0 && ignorezeroentries && row != col) { \
403: low1 = 0; \
404: high1 = nrow1; \
405: goto a_noinsert; \
406: } \
407: if (nonew == 1) { \
408: low1 = 0; \
409: high1 = nrow1; \
410: goto a_noinsert; \
411: } \
412: 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); \
413: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
414: N = nrow1++ - 1; \
415: a->nz++; \
416: high1++; \
417: /* shift up all the later entries in this row */ \
418: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
419: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
420: rp1[_i] = col; \
421: ap1[_i] = value; \
422: A->nonzerostate++; \
423: a_noinsert:; \
424: ailen[row] = nrow1; \
425: } while (0)
427: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
428: do { \
429: if (col <= lastcol2) low2 = 0; \
430: else high2 = nrow2; \
431: lastcol2 = col; \
432: while (high2 - low2 > 5) { \
433: t = (low2 + high2) / 2; \
434: if (rp2[t] > col) high2 = t; \
435: else low2 = t; \
436: } \
437: for (_i = low2; _i < high2; _i++) { \
438: if (rp2[_i] > col) break; \
439: if (rp2[_i] == col) { \
440: if (addv == ADD_VALUES) { \
441: ap2[_i] += value; \
442: (void)PetscLogFlops(1.0); \
443: } else ap2[_i] = value; \
444: goto b_noinsert; \
445: } \
446: } \
447: if (value == 0.0 && ignorezeroentries) { \
448: low2 = 0; \
449: high2 = nrow2; \
450: goto b_noinsert; \
451: } \
452: if (nonew == 1) { \
453: low2 = 0; \
454: high2 = nrow2; \
455: goto b_noinsert; \
456: } \
457: 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); \
458: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
459: N = nrow2++ - 1; \
460: b->nz++; \
461: high2++; \
462: /* shift up all the later entries in this row */ \
463: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
464: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
465: rp2[_i] = col; \
466: ap2[_i] = value; \
467: B->nonzerostate++; \
468: b_noinsert:; \
469: bilen[row] = nrow2; \
470: } while (0)
472: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
473: {
474: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
475: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
476: PetscInt l, *garray = mat->garray, diag;
477: PetscScalar *aa, *ba;
479: PetscFunctionBegin;
480: /* code only works for square matrices A */
482: /* find size of row to the left of the diagonal part */
483: PetscCall(MatGetOwnershipRange(A, &diag, NULL));
484: row = row - diag;
485: for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
486: if (garray[b->j[b->i[row] + l]] > diag) break;
487: }
488: if (l) {
489: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
490: PetscCall(PetscArraycpy(ba + b->i[row], v, l));
491: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
492: }
494: /* diagonal part */
495: if (a->i[row + 1] - a->i[row]) {
496: PetscCall(MatSeqAIJGetArray(mat->A, &aa));
497: PetscCall(PetscArraycpy(aa + a->i[row], v + l, (a->i[row + 1] - a->i[row])));
498: PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
499: }
501: /* right of diagonal part */
502: if (b->i[row + 1] - b->i[row] - l) {
503: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
504: PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
505: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
506: }
507: PetscFunctionReturn(PETSC_SUCCESS);
508: }
510: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
511: {
512: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
513: PetscScalar value = 0.0;
514: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
515: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
516: PetscBool roworiented = aij->roworiented;
518: /* Some Variables required in the macro */
519: Mat A = aij->A;
520: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
521: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
522: PetscBool ignorezeroentries = a->ignorezeroentries;
523: Mat B = aij->B;
524: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
525: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
526: MatScalar *aa, *ba;
527: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
528: PetscInt nonew;
529: MatScalar *ap1, *ap2;
531: PetscFunctionBegin;
532: PetscCall(MatSeqAIJGetArray(A, &aa));
533: PetscCall(MatSeqAIJGetArray(B, &ba));
534: for (i = 0; i < m; i++) {
535: if (im[i] < 0) continue;
536: 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);
537: if (im[i] >= rstart && im[i] < rend) {
538: row = im[i] - rstart;
539: lastcol1 = -1;
540: rp1 = aj ? aj + ai[row] : NULL;
541: ap1 = aa ? aa + ai[row] : NULL;
542: rmax1 = aimax[row];
543: nrow1 = ailen[row];
544: low1 = 0;
545: high1 = nrow1;
546: lastcol2 = -1;
547: rp2 = bj ? bj + bi[row] : NULL;
548: ap2 = ba ? ba + bi[row] : NULL;
549: rmax2 = bimax[row];
550: nrow2 = bilen[row];
551: low2 = 0;
552: high2 = nrow2;
554: for (j = 0; j < n; j++) {
555: if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
556: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
557: if (in[j] >= cstart && in[j] < cend) {
558: col = in[j] - cstart;
559: nonew = a->nonew;
560: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
561: } else if (in[j] < 0) {
562: continue;
563: } else {
564: 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);
565: if (mat->was_assembled) {
566: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
567: #if defined(PETSC_USE_CTABLE)
568: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
569: col--;
570: #else
571: col = aij->colmap[in[j]] - 1;
572: #endif
573: if (col < 0 && !((Mat_SeqAIJ *)(aij->B->data))->nonew) { /* col < 0 means in[j] is a new col for B */
574: PetscCall(MatDisAssemble_MPIAIJ(mat)); /* Change aij->B from reduced/local format to expanded/global format */
575: col = in[j];
576: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
577: B = aij->B;
578: b = (Mat_SeqAIJ *)B->data;
579: bimax = b->imax;
580: bi = b->i;
581: bilen = b->ilen;
582: bj = b->j;
583: ba = b->a;
584: rp2 = bj + bi[row];
585: ap2 = ba + bi[row];
586: rmax2 = bimax[row];
587: nrow2 = bilen[row];
588: low2 = 0;
589: high2 = nrow2;
590: bm = aij->B->rmap->n;
591: ba = b->a;
592: } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
593: if (1 == ((Mat_SeqAIJ *)(aij->B->data))->nonew) {
594: 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]));
595: } 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]);
596: }
597: } else col = in[j];
598: nonew = b->nonew;
599: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
600: }
601: }
602: } else {
603: 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]);
604: if (!aij->donotstash) {
605: mat->assembled = PETSC_FALSE;
606: if (roworiented) {
607: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v ? v + i * n : NULL, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
608: } else {
609: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v ? v + i : NULL, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
610: }
611: }
612: }
613: }
614: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
615: PetscCall(MatSeqAIJRestoreArray(B, &ba));
616: PetscFunctionReturn(PETSC_SUCCESS);
617: }
619: /*
620: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
621: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
622: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
623: */
624: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
625: {
626: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
627: Mat A = aij->A; /* diagonal part of the matrix */
628: Mat B = aij->B; /* off-diagonal part of the matrix */
629: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
630: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
631: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
632: PetscInt *ailen = a->ilen, *aj = a->j;
633: PetscInt *bilen = b->ilen, *bj = b->j;
634: PetscInt am = aij->A->rmap->n, j;
635: PetscInt diag_so_far = 0, dnz;
636: PetscInt offd_so_far = 0, onz;
638: PetscFunctionBegin;
639: /* Iterate over all rows of the matrix */
640: for (j = 0; j < am; j++) {
641: dnz = onz = 0;
642: /* Iterate over all non-zero columns of the current row */
643: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
644: /* If column is in the diagonal */
645: if (mat_j[col] >= cstart && mat_j[col] < cend) {
646: aj[diag_so_far++] = mat_j[col] - cstart;
647: dnz++;
648: } else { /* off-diagonal entries */
649: bj[offd_so_far++] = mat_j[col];
650: onz++;
651: }
652: }
653: ailen[j] = dnz;
654: bilen[j] = onz;
655: }
656: PetscFunctionReturn(PETSC_SUCCESS);
657: }
659: /*
660: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
661: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
662: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
663: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
664: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
665: */
666: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
667: {
668: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
669: Mat A = aij->A; /* diagonal part of the matrix */
670: Mat B = aij->B; /* off-diagonal part of the matrix */
671: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)(aij->A)->data, *aijo = (Mat_SeqAIJ *)(aij->B)->data;
672: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
673: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
674: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
675: PetscInt *ailen = a->ilen, *aj = a->j;
676: PetscInt *bilen = b->ilen, *bj = b->j;
677: PetscInt am = aij->A->rmap->n, j;
678: 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. */
679: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
680: PetscScalar *aa = a->a, *ba = b->a;
682: PetscFunctionBegin;
683: /* Iterate over all rows of the matrix */
684: for (j = 0; j < am; j++) {
685: dnz_row = onz_row = 0;
686: rowstart_offd = full_offd_i[j];
687: rowstart_diag = full_diag_i[j];
688: /* Iterate over all non-zero columns of the current row */
689: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
690: /* If column is in the diagonal */
691: if (mat_j[col] >= cstart && mat_j[col] < cend) {
692: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
693: aa[rowstart_diag + dnz_row] = mat_a[col];
694: dnz_row++;
695: } else { /* off-diagonal entries */
696: bj[rowstart_offd + onz_row] = mat_j[col];
697: ba[rowstart_offd + onz_row] = mat_a[col];
698: onz_row++;
699: }
700: }
701: ailen[j] = dnz_row;
702: bilen[j] = onz_row;
703: }
704: PetscFunctionReturn(PETSC_SUCCESS);
705: }
707: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
708: {
709: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
710: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
711: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
713: PetscFunctionBegin;
714: for (i = 0; i < m; i++) {
715: if (idxm[i] < 0) continue; /* negative row */
716: 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);
717: 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);
718: row = idxm[i] - rstart;
719: for (j = 0; j < n; j++) {
720: if (idxn[j] < 0) continue; /* negative column */
721: 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);
722: if (idxn[j] >= cstart && idxn[j] < cend) {
723: col = idxn[j] - cstart;
724: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
725: } else {
726: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
727: #if defined(PETSC_USE_CTABLE)
728: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
729: col--;
730: #else
731: col = aij->colmap[idxn[j]] - 1;
732: #endif
733: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
734: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
735: }
736: }
737: }
738: PetscFunctionReturn(PETSC_SUCCESS);
739: }
741: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
742: {
743: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
744: PetscInt nstash, reallocs;
746: PetscFunctionBegin;
747: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
749: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
750: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
751: PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
752: PetscFunctionReturn(PETSC_SUCCESS);
753: }
755: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
756: {
757: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
758: PetscMPIInt n;
759: PetscInt i, j, rstart, ncols, flg;
760: PetscInt *row, *col;
761: PetscBool other_disassembled;
762: PetscScalar *val;
764: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
766: PetscFunctionBegin;
767: if (!aij->donotstash && !mat->nooffprocentries) {
768: while (1) {
769: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
770: if (!flg) break;
772: for (i = 0; i < n;) {
773: /* Now identify the consecutive vals belonging to the same row */
774: for (j = i, rstart = row[j]; j < n; j++) {
775: if (row[j] != rstart) break;
776: }
777: if (j < n) ncols = j - i;
778: else ncols = n - i;
779: /* Now assemble all these values with a single function call */
780: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
781: i = j;
782: }
783: }
784: PetscCall(MatStashScatterEnd_Private(&mat->stash));
785: }
786: #if defined(PETSC_HAVE_DEVICE)
787: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
788: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
789: if (mat->boundtocpu) {
790: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
791: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
792: }
793: #endif
794: PetscCall(MatAssemblyBegin(aij->A, mode));
795: PetscCall(MatAssemblyEnd(aij->A, mode));
797: /* determine if any processor has disassembled, if so we must
798: also disassemble ourself, in order that we may reassemble. */
799: /*
800: if nonzero structure of submatrix B cannot change then we know that
801: no processor disassembled thus we can skip this stuff
802: */
803: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
804: PetscCall(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
805: if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
806: PetscCall(MatDisAssemble_MPIAIJ(mat));
807: }
808: }
809: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
810: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
811: #if defined(PETSC_HAVE_DEVICE)
812: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
813: #endif
814: PetscCall(MatAssemblyBegin(aij->B, mode));
815: PetscCall(MatAssemblyEnd(aij->B, mode));
817: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
819: aij->rowvalues = NULL;
821: PetscCall(VecDestroy(&aij->diag));
823: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
824: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)(aij->A->data))->nonew) {
825: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
826: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
827: }
828: #if defined(PETSC_HAVE_DEVICE)
829: mat->offloadmask = PETSC_OFFLOAD_BOTH;
830: #endif
831: PetscFunctionReturn(PETSC_SUCCESS);
832: }
834: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
835: {
836: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
838: PetscFunctionBegin;
839: PetscCall(MatZeroEntries(l->A));
840: PetscCall(MatZeroEntries(l->B));
841: PetscFunctionReturn(PETSC_SUCCESS);
842: }
844: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
845: {
846: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
847: PetscObjectState sA, sB;
848: PetscInt *lrows;
849: PetscInt r, len;
850: PetscBool cong, lch, gch;
852: PetscFunctionBegin;
853: /* get locally owned rows */
854: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
855: PetscCall(MatHasCongruentLayouts(A, &cong));
856: /* fix right hand side if needed */
857: if (x && b) {
858: const PetscScalar *xx;
859: PetscScalar *bb;
861: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
862: PetscCall(VecGetArrayRead(x, &xx));
863: PetscCall(VecGetArray(b, &bb));
864: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
865: PetscCall(VecRestoreArrayRead(x, &xx));
866: PetscCall(VecRestoreArray(b, &bb));
867: }
869: sA = mat->A->nonzerostate;
870: sB = mat->B->nonzerostate;
872: if (diag != 0.0 && cong) {
873: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
874: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
875: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
876: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
877: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
878: PetscInt nnwA, nnwB;
879: PetscBool nnzA, nnzB;
881: nnwA = aijA->nonew;
882: nnwB = aijB->nonew;
883: nnzA = aijA->keepnonzeropattern;
884: nnzB = aijB->keepnonzeropattern;
885: if (!nnzA) {
886: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
887: aijA->nonew = 0;
888: }
889: if (!nnzB) {
890: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
891: aijB->nonew = 0;
892: }
893: /* Must zero here before the next loop */
894: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
895: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
896: for (r = 0; r < len; ++r) {
897: const PetscInt row = lrows[r] + A->rmap->rstart;
898: if (row >= A->cmap->N) continue;
899: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
900: }
901: aijA->nonew = nnwA;
902: aijB->nonew = nnwB;
903: } else {
904: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
905: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
906: }
907: PetscCall(PetscFree(lrows));
908: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
909: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
911: /* reduce nonzerostate */
912: lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
913: PetscCall(MPIU_Allreduce(&lch, &gch, 1, MPIU_BOOL, MPI_LOR, PetscObjectComm((PetscObject)A)));
914: if (gch) A->nonzerostate++;
915: PetscFunctionReturn(PETSC_SUCCESS);
916: }
918: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
919: {
920: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
921: PetscMPIInt n = A->rmap->n;
922: PetscInt i, j, r, m, len = 0;
923: PetscInt *lrows, *owners = A->rmap->range;
924: PetscMPIInt p = 0;
925: PetscSFNode *rrows;
926: PetscSF sf;
927: const PetscScalar *xx;
928: PetscScalar *bb, *mask, *aij_a;
929: Vec xmask, lmask;
930: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
931: const PetscInt *aj, *ii, *ridx;
932: PetscScalar *aa;
934: PetscFunctionBegin;
935: /* Create SF where leaves are input rows and roots are owned rows */
936: PetscCall(PetscMalloc1(n, &lrows));
937: for (r = 0; r < n; ++r) lrows[r] = -1;
938: PetscCall(PetscMalloc1(N, &rrows));
939: for (r = 0; r < N; ++r) {
940: const PetscInt idx = rows[r];
941: 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);
942: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
943: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
944: }
945: rrows[r].rank = p;
946: rrows[r].index = rows[r] - owners[p];
947: }
948: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
949: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
950: /* Collect flags for rows to be zeroed */
951: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
952: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
953: PetscCall(PetscSFDestroy(&sf));
954: /* Compress and put in row numbers */
955: for (r = 0; r < n; ++r)
956: if (lrows[r] >= 0) lrows[len++] = r;
957: /* zero diagonal part of matrix */
958: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
959: /* handle off-diagonal part of matrix */
960: PetscCall(MatCreateVecs(A, &xmask, NULL));
961: PetscCall(VecDuplicate(l->lvec, &lmask));
962: PetscCall(VecGetArray(xmask, &bb));
963: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
964: PetscCall(VecRestoreArray(xmask, &bb));
965: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
966: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
967: PetscCall(VecDestroy(&xmask));
968: if (x && b) { /* this code is buggy when the row and column layout don't match */
969: PetscBool cong;
971: PetscCall(MatHasCongruentLayouts(A, &cong));
972: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
973: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
974: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
975: PetscCall(VecGetArrayRead(l->lvec, &xx));
976: PetscCall(VecGetArray(b, &bb));
977: }
978: PetscCall(VecGetArray(lmask, &mask));
979: /* remove zeroed rows of off-diagonal matrix */
980: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
981: ii = aij->i;
982: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(aij_a + ii[lrows[i]], ii[lrows[i] + 1] - ii[lrows[i]]));
983: /* loop over all elements of off process part of matrix zeroing removed columns*/
984: if (aij->compressedrow.use) {
985: m = aij->compressedrow.nrows;
986: ii = aij->compressedrow.i;
987: ridx = aij->compressedrow.rindex;
988: for (i = 0; i < m; i++) {
989: n = ii[i + 1] - ii[i];
990: aj = aij->j + ii[i];
991: aa = aij_a + ii[i];
993: for (j = 0; j < n; j++) {
994: if (PetscAbsScalar(mask[*aj])) {
995: if (b) bb[*ridx] -= *aa * xx[*aj];
996: *aa = 0.0;
997: }
998: aa++;
999: aj++;
1000: }
1001: ridx++;
1002: }
1003: } else { /* do not use compressed row format */
1004: m = l->B->rmap->n;
1005: for (i = 0; i < m; i++) {
1006: n = ii[i + 1] - ii[i];
1007: aj = aij->j + ii[i];
1008: aa = aij_a + ii[i];
1009: for (j = 0; j < n; j++) {
1010: if (PetscAbsScalar(mask[*aj])) {
1011: if (b) bb[i] -= *aa * xx[*aj];
1012: *aa = 0.0;
1013: }
1014: aa++;
1015: aj++;
1016: }
1017: }
1018: }
1019: if (x && b) {
1020: PetscCall(VecRestoreArray(b, &bb));
1021: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1022: }
1023: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1024: PetscCall(VecRestoreArray(lmask, &mask));
1025: PetscCall(VecDestroy(&lmask));
1026: PetscCall(PetscFree(lrows));
1028: /* only change matrix nonzero state if pattern was allowed to be changed */
1029: if (!((Mat_SeqAIJ *)(l->A->data))->keepnonzeropattern) {
1030: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1031: PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1032: }
1033: PetscFunctionReturn(PETSC_SUCCESS);
1034: }
1036: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1037: {
1038: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1039: PetscInt nt;
1040: VecScatter Mvctx = a->Mvctx;
1042: PetscFunctionBegin;
1043: PetscCall(VecGetLocalSize(xx, &nt));
1044: 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);
1045: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1046: PetscUseTypeMethod(a->A, mult, xx, yy);
1047: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1048: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1049: PetscFunctionReturn(PETSC_SUCCESS);
1050: }
1052: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1053: {
1054: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1056: PetscFunctionBegin;
1057: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1058: PetscFunctionReturn(PETSC_SUCCESS);
1059: }
1061: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1062: {
1063: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1064: VecScatter Mvctx = a->Mvctx;
1066: PetscFunctionBegin;
1067: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1068: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1069: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1070: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1071: PetscFunctionReturn(PETSC_SUCCESS);
1072: }
1074: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1075: {
1076: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1078: PetscFunctionBegin;
1079: /* do nondiagonal part */
1080: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1081: /* do local part */
1082: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1083: /* add partial results together */
1084: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1085: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1086: PetscFunctionReturn(PETSC_SUCCESS);
1087: }
1089: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1090: {
1091: MPI_Comm comm;
1092: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1093: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1094: IS Me, Notme;
1095: PetscInt M, N, first, last, *notme, i;
1096: PetscBool lf;
1097: PetscMPIInt size;
1099: PetscFunctionBegin;
1100: /* Easy test: symmetric diagonal block */
1101: PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1102: PetscCall(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1103: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1104: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1105: PetscCallMPI(MPI_Comm_size(comm, &size));
1106: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1108: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1109: PetscCall(MatGetSize(Amat, &M, &N));
1110: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1111: PetscCall(PetscMalloc1(N - last + first, ¬me));
1112: for (i = 0; i < first; i++) notme[i] = i;
1113: for (i = last; i < M; i++) notme[i - last + first] = i;
1114: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1115: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1116: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1117: Aoff = Aoffs[0];
1118: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1119: Boff = Boffs[0];
1120: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1121: PetscCall(MatDestroyMatrices(1, &Aoffs));
1122: PetscCall(MatDestroyMatrices(1, &Boffs));
1123: PetscCall(ISDestroy(&Me));
1124: PetscCall(ISDestroy(&Notme));
1125: PetscCall(PetscFree(notme));
1126: PetscFunctionReturn(PETSC_SUCCESS);
1127: }
1129: static PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A, PetscReal tol, PetscBool *f)
1130: {
1131: PetscFunctionBegin;
1132: PetscCall(MatIsTranspose_MPIAIJ(A, A, tol, f));
1133: PetscFunctionReturn(PETSC_SUCCESS);
1134: }
1136: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1137: {
1138: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1140: PetscFunctionBegin;
1141: /* do nondiagonal part */
1142: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1143: /* do local part */
1144: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1145: /* add partial results together */
1146: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1147: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1148: PetscFunctionReturn(PETSC_SUCCESS);
1149: }
1151: /*
1152: This only works correctly for square matrices where the subblock A->A is the
1153: diagonal block
1154: */
1155: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1156: {
1157: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1159: PetscFunctionBegin;
1160: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1161: 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");
1162: PetscCall(MatGetDiagonal(a->A, v));
1163: PetscFunctionReturn(PETSC_SUCCESS);
1164: }
1166: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1167: {
1168: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1170: PetscFunctionBegin;
1171: PetscCall(MatScale(a->A, aa));
1172: PetscCall(MatScale(a->B, aa));
1173: PetscFunctionReturn(PETSC_SUCCESS);
1174: }
1176: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1177: {
1178: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1179: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1180: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1181: const PetscInt *garray = aij->garray;
1182: const PetscScalar *aa, *ba;
1183: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1184: PetscInt64 nz, hnz;
1185: PetscInt *rowlens;
1186: PetscInt *colidxs;
1187: PetscScalar *matvals;
1188: PetscMPIInt rank;
1190: PetscFunctionBegin;
1191: PetscCall(PetscViewerSetUp(viewer));
1193: M = mat->rmap->N;
1194: N = mat->cmap->N;
1195: m = mat->rmap->n;
1196: rs = mat->rmap->rstart;
1197: cs = mat->cmap->rstart;
1198: nz = A->nz + B->nz;
1200: /* write matrix header */
1201: header[0] = MAT_FILE_CLASSID;
1202: header[1] = M;
1203: header[2] = N;
1204: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1205: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1206: if (rank == 0) {
1207: if (hnz > PETSC_MAX_INT) header[3] = PETSC_MAX_INT;
1208: else header[3] = (PetscInt)hnz;
1209: }
1210: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1212: /* fill in and store row lengths */
1213: PetscCall(PetscMalloc1(m, &rowlens));
1214: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1215: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1216: PetscCall(PetscFree(rowlens));
1218: /* fill in and store column indices */
1219: PetscCall(PetscMalloc1(nz, &colidxs));
1220: for (cnt = 0, i = 0; i < m; i++) {
1221: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1222: if (garray[B->j[jb]] > cs) break;
1223: colidxs[cnt++] = garray[B->j[jb]];
1224: }
1225: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1226: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1227: }
1228: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1229: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1230: PetscCall(PetscFree(colidxs));
1232: /* fill in and store nonzero values */
1233: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1234: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1235: PetscCall(PetscMalloc1(nz, &matvals));
1236: for (cnt = 0, i = 0; i < m; i++) {
1237: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1238: if (garray[B->j[jb]] > cs) break;
1239: matvals[cnt++] = ba[jb];
1240: }
1241: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1242: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1243: }
1244: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1245: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1246: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1247: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1248: PetscCall(PetscFree(matvals));
1250: /* write block size option to the viewer's .info file */
1251: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1252: PetscFunctionReturn(PETSC_SUCCESS);
1253: }
1255: #include <petscdraw.h>
1256: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1257: {
1258: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1259: PetscMPIInt rank = aij->rank, size = aij->size;
1260: PetscBool isdraw, iascii, isbinary;
1261: PetscViewer sviewer;
1262: PetscViewerFormat format;
1264: PetscFunctionBegin;
1265: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1266: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1267: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1268: if (iascii) {
1269: PetscCall(PetscViewerGetFormat(viewer, &format));
1270: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1271: PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)(aij->A->data))->nz + ((Mat_SeqAIJ *)(aij->B->data))->nz;
1272: PetscCall(PetscMalloc1(size, &nz));
1273: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1274: for (i = 0; i < (PetscInt)size; i++) {
1275: nmax = PetscMax(nmax, nz[i]);
1276: nmin = PetscMin(nmin, nz[i]);
1277: navg += nz[i];
1278: }
1279: PetscCall(PetscFree(nz));
1280: navg = navg / size;
1281: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1282: PetscFunctionReturn(PETSC_SUCCESS);
1283: }
1284: PetscCall(PetscViewerGetFormat(viewer, &format));
1285: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1286: MatInfo info;
1287: PetscInt *inodes = NULL;
1289: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1290: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1291: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1292: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1293: if (!inodes) {
1294: 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,
1295: (double)info.memory));
1296: } else {
1297: 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,
1298: (double)info.memory));
1299: }
1300: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1301: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1302: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1303: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1304: PetscCall(PetscViewerFlush(viewer));
1305: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1306: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1307: PetscCall(VecScatterView(aij->Mvctx, viewer));
1308: PetscFunctionReturn(PETSC_SUCCESS);
1309: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1310: PetscInt inodecount, inodelimit, *inodes;
1311: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1312: if (inodes) {
1313: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1314: } else {
1315: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1316: }
1317: PetscFunctionReturn(PETSC_SUCCESS);
1318: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1319: PetscFunctionReturn(PETSC_SUCCESS);
1320: }
1321: } else if (isbinary) {
1322: if (size == 1) {
1323: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1324: PetscCall(MatView(aij->A, viewer));
1325: } else {
1326: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1327: }
1328: PetscFunctionReturn(PETSC_SUCCESS);
1329: } else if (iascii && size == 1) {
1330: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1331: PetscCall(MatView(aij->A, viewer));
1332: PetscFunctionReturn(PETSC_SUCCESS);
1333: } else if (isdraw) {
1334: PetscDraw draw;
1335: PetscBool isnull;
1336: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1337: PetscCall(PetscDrawIsNull(draw, &isnull));
1338: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1339: }
1341: { /* assemble the entire matrix onto first processor */
1342: Mat A = NULL, Av;
1343: IS isrow, iscol;
1345: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1346: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1347: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1348: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1349: /* The commented code uses MatCreateSubMatrices instead */
1350: /*
1351: Mat *AA, A = NULL, Av;
1352: IS isrow,iscol;
1354: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1355: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1356: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1357: if (rank == 0) {
1358: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1359: A = AA[0];
1360: Av = AA[0];
1361: }
1362: PetscCall(MatDestroySubMatrices(1,&AA));
1363: */
1364: PetscCall(ISDestroy(&iscol));
1365: PetscCall(ISDestroy(&isrow));
1366: /*
1367: Everyone has to call to draw the matrix since the graphics waits are
1368: synchronized across all processors that share the PetscDraw object
1369: */
1370: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1371: if (rank == 0) {
1372: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1373: PetscCall(MatView_SeqAIJ(Av, sviewer));
1374: }
1375: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1376: PetscCall(PetscViewerFlush(viewer));
1377: PetscCall(MatDestroy(&A));
1378: }
1379: PetscFunctionReturn(PETSC_SUCCESS);
1380: }
1382: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1383: {
1384: PetscBool iascii, isdraw, issocket, isbinary;
1386: PetscFunctionBegin;
1387: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1388: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1389: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1390: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1391: if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1392: PetscFunctionReturn(PETSC_SUCCESS);
1393: }
1395: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1396: {
1397: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1398: Vec bb1 = NULL;
1399: PetscBool hasop;
1401: PetscFunctionBegin;
1402: if (flag == SOR_APPLY_UPPER) {
1403: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1404: PetscFunctionReturn(PETSC_SUCCESS);
1405: }
1407: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1409: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1410: if (flag & SOR_ZERO_INITIAL_GUESS) {
1411: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1412: its--;
1413: }
1415: while (its--) {
1416: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1417: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1419: /* update rhs: bb1 = bb - B*x */
1420: PetscCall(VecScale(mat->lvec, -1.0));
1421: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1423: /* local sweep */
1424: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1425: }
1426: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1427: if (flag & SOR_ZERO_INITIAL_GUESS) {
1428: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1429: its--;
1430: }
1431: while (its--) {
1432: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1433: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1435: /* update rhs: bb1 = bb - B*x */
1436: PetscCall(VecScale(mat->lvec, -1.0));
1437: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1439: /* local sweep */
1440: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1441: }
1442: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1443: if (flag & SOR_ZERO_INITIAL_GUESS) {
1444: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1445: its--;
1446: }
1447: while (its--) {
1448: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1449: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1451: /* update rhs: bb1 = bb - B*x */
1452: PetscCall(VecScale(mat->lvec, -1.0));
1453: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1455: /* local sweep */
1456: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1457: }
1458: } else if (flag & SOR_EISENSTAT) {
1459: Vec xx1;
1461: PetscCall(VecDuplicate(bb, &xx1));
1462: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1464: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1465: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1466: if (!mat->diag) {
1467: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1468: PetscCall(MatGetDiagonal(matin, mat->diag));
1469: }
1470: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1471: if (hasop) {
1472: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1473: } else {
1474: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1475: }
1476: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1478: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1480: /* local sweep */
1481: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1482: PetscCall(VecAXPY(xx, 1.0, xx1));
1483: PetscCall(VecDestroy(&xx1));
1484: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1486: PetscCall(VecDestroy(&bb1));
1488: matin->factorerrortype = mat->A->factorerrortype;
1489: PetscFunctionReturn(PETSC_SUCCESS);
1490: }
1492: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1493: {
1494: Mat aA, aB, Aperm;
1495: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1496: PetscScalar *aa, *ba;
1497: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1498: PetscSF rowsf, sf;
1499: IS parcolp = NULL;
1500: PetscBool done;
1502: PetscFunctionBegin;
1503: PetscCall(MatGetLocalSize(A, &m, &n));
1504: PetscCall(ISGetIndices(rowp, &rwant));
1505: PetscCall(ISGetIndices(colp, &cwant));
1506: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1508: /* Invert row permutation to find out where my rows should go */
1509: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1510: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1511: PetscCall(PetscSFSetFromOptions(rowsf));
1512: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1513: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1514: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1516: /* Invert column permutation to find out where my columns should go */
1517: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1518: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1519: PetscCall(PetscSFSetFromOptions(sf));
1520: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1521: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1522: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1523: PetscCall(PetscSFDestroy(&sf));
1525: PetscCall(ISRestoreIndices(rowp, &rwant));
1526: PetscCall(ISRestoreIndices(colp, &cwant));
1527: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1529: /* Find out where my gcols should go */
1530: PetscCall(MatGetSize(aB, NULL, &ng));
1531: PetscCall(PetscMalloc1(ng, &gcdest));
1532: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1533: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1534: PetscCall(PetscSFSetFromOptions(sf));
1535: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1536: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1537: PetscCall(PetscSFDestroy(&sf));
1539: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1540: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1541: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1542: for (i = 0; i < m; i++) {
1543: PetscInt row = rdest[i];
1544: PetscMPIInt rowner;
1545: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1546: for (j = ai[i]; j < ai[i + 1]; j++) {
1547: PetscInt col = cdest[aj[j]];
1548: PetscMPIInt cowner;
1549: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1550: if (rowner == cowner) dnnz[i]++;
1551: else onnz[i]++;
1552: }
1553: for (j = bi[i]; j < bi[i + 1]; j++) {
1554: PetscInt col = gcdest[bj[j]];
1555: PetscMPIInt cowner;
1556: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1557: if (rowner == cowner) dnnz[i]++;
1558: else onnz[i]++;
1559: }
1560: }
1561: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1562: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1563: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1564: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1565: PetscCall(PetscSFDestroy(&rowsf));
1567: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1568: PetscCall(MatSeqAIJGetArray(aA, &aa));
1569: PetscCall(MatSeqAIJGetArray(aB, &ba));
1570: for (i = 0; i < m; i++) {
1571: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1572: PetscInt j0, rowlen;
1573: rowlen = ai[i + 1] - ai[i];
1574: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1575: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1576: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1577: }
1578: rowlen = bi[i + 1] - bi[i];
1579: for (j0 = j = 0; j < rowlen; j0 = j) {
1580: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1581: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1582: }
1583: }
1584: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1585: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1586: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1587: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1588: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1589: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1590: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1591: PetscCall(PetscFree3(work, rdest, cdest));
1592: PetscCall(PetscFree(gcdest));
1593: if (parcolp) PetscCall(ISDestroy(&colp));
1594: *B = Aperm;
1595: PetscFunctionReturn(PETSC_SUCCESS);
1596: }
1598: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1599: {
1600: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1602: PetscFunctionBegin;
1603: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1604: if (ghosts) *ghosts = aij->garray;
1605: PetscFunctionReturn(PETSC_SUCCESS);
1606: }
1608: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1609: {
1610: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1611: Mat A = mat->A, B = mat->B;
1612: PetscLogDouble isend[5], irecv[5];
1614: PetscFunctionBegin;
1615: info->block_size = 1.0;
1616: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1618: isend[0] = info->nz_used;
1619: isend[1] = info->nz_allocated;
1620: isend[2] = info->nz_unneeded;
1621: isend[3] = info->memory;
1622: isend[4] = info->mallocs;
1624: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1626: isend[0] += info->nz_used;
1627: isend[1] += info->nz_allocated;
1628: isend[2] += info->nz_unneeded;
1629: isend[3] += info->memory;
1630: isend[4] += info->mallocs;
1631: if (flag == MAT_LOCAL) {
1632: info->nz_used = isend[0];
1633: info->nz_allocated = isend[1];
1634: info->nz_unneeded = isend[2];
1635: info->memory = isend[3];
1636: info->mallocs = isend[4];
1637: } else if (flag == MAT_GLOBAL_MAX) {
1638: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1640: info->nz_used = irecv[0];
1641: info->nz_allocated = irecv[1];
1642: info->nz_unneeded = irecv[2];
1643: info->memory = irecv[3];
1644: info->mallocs = irecv[4];
1645: } else if (flag == MAT_GLOBAL_SUM) {
1646: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1648: info->nz_used = irecv[0];
1649: info->nz_allocated = irecv[1];
1650: info->nz_unneeded = irecv[2];
1651: info->memory = irecv[3];
1652: info->mallocs = irecv[4];
1653: }
1654: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1655: info->fill_ratio_needed = 0;
1656: info->factor_mallocs = 0;
1657: PetscFunctionReturn(PETSC_SUCCESS);
1658: }
1660: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1661: {
1662: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1664: PetscFunctionBegin;
1665: switch (op) {
1666: case MAT_NEW_NONZERO_LOCATIONS:
1667: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1668: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1669: case MAT_KEEP_NONZERO_PATTERN:
1670: case MAT_NEW_NONZERO_LOCATION_ERR:
1671: case MAT_USE_INODES:
1672: case MAT_IGNORE_ZERO_ENTRIES:
1673: case MAT_FORM_EXPLICIT_TRANSPOSE:
1674: MatCheckPreallocated(A, 1);
1675: PetscCall(MatSetOption(a->A, op, flg));
1676: PetscCall(MatSetOption(a->B, op, flg));
1677: break;
1678: case MAT_ROW_ORIENTED:
1679: MatCheckPreallocated(A, 1);
1680: a->roworiented = flg;
1682: PetscCall(MatSetOption(a->A, op, flg));
1683: PetscCall(MatSetOption(a->B, op, flg));
1684: break;
1685: case MAT_FORCE_DIAGONAL_ENTRIES:
1686: case MAT_SORTED_FULL:
1687: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1688: break;
1689: case MAT_IGNORE_OFF_PROC_ENTRIES:
1690: a->donotstash = flg;
1691: break;
1692: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1693: case MAT_SPD:
1694: case MAT_SYMMETRIC:
1695: case MAT_STRUCTURALLY_SYMMETRIC:
1696: case MAT_HERMITIAN:
1697: case MAT_SYMMETRY_ETERNAL:
1698: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1699: case MAT_SPD_ETERNAL:
1700: /* if the diagonal matrix is square it inherits some of the properties above */
1701: break;
1702: case MAT_SUBMAT_SINGLEIS:
1703: A->submat_singleis = flg;
1704: break;
1705: case MAT_STRUCTURE_ONLY:
1706: /* The option is handled directly by MatSetOption() */
1707: break;
1708: default:
1709: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1710: }
1711: PetscFunctionReturn(PETSC_SUCCESS);
1712: }
1714: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1715: {
1716: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1717: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1718: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1719: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1720: PetscInt *cmap, *idx_p;
1722: PetscFunctionBegin;
1723: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1724: mat->getrowactive = PETSC_TRUE;
1726: if (!mat->rowvalues && (idx || v)) {
1727: /*
1728: allocate enough space to hold information from the longest row.
1729: */
1730: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1731: PetscInt max = 1, tmp;
1732: for (i = 0; i < matin->rmap->n; i++) {
1733: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1734: if (max < tmp) max = tmp;
1735: }
1736: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1737: }
1739: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1740: lrow = row - rstart;
1742: pvA = &vworkA;
1743: pcA = &cworkA;
1744: pvB = &vworkB;
1745: pcB = &cworkB;
1746: if (!v) {
1747: pvA = NULL;
1748: pvB = NULL;
1749: }
1750: if (!idx) {
1751: pcA = NULL;
1752: if (!v) pcB = NULL;
1753: }
1754: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1755: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1756: nztot = nzA + nzB;
1758: cmap = mat->garray;
1759: if (v || idx) {
1760: if (nztot) {
1761: /* Sort by increasing column numbers, assuming A and B already sorted */
1762: PetscInt imark = -1;
1763: if (v) {
1764: *v = v_p = mat->rowvalues;
1765: for (i = 0; i < nzB; i++) {
1766: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1767: else break;
1768: }
1769: imark = i;
1770: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1771: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1772: }
1773: if (idx) {
1774: *idx = idx_p = mat->rowindices;
1775: if (imark > -1) {
1776: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1777: } else {
1778: for (i = 0; i < nzB; i++) {
1779: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1780: else break;
1781: }
1782: imark = i;
1783: }
1784: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1785: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1786: }
1787: } else {
1788: if (idx) *idx = NULL;
1789: if (v) *v = NULL;
1790: }
1791: }
1792: *nz = nztot;
1793: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1794: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1795: PetscFunctionReturn(PETSC_SUCCESS);
1796: }
1798: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1799: {
1800: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1802: PetscFunctionBegin;
1803: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1804: aij->getrowactive = PETSC_FALSE;
1805: PetscFunctionReturn(PETSC_SUCCESS);
1806: }
1808: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1809: {
1810: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1811: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1812: PetscInt i, j, cstart = mat->cmap->rstart;
1813: PetscReal sum = 0.0;
1814: const MatScalar *v, *amata, *bmata;
1816: PetscFunctionBegin;
1817: if (aij->size == 1) {
1818: PetscCall(MatNorm(aij->A, type, norm));
1819: } else {
1820: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1821: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1822: if (type == NORM_FROBENIUS) {
1823: v = amata;
1824: for (i = 0; i < amat->nz; i++) {
1825: sum += PetscRealPart(PetscConj(*v) * (*v));
1826: v++;
1827: }
1828: v = bmata;
1829: for (i = 0; i < bmat->nz; i++) {
1830: sum += PetscRealPart(PetscConj(*v) * (*v));
1831: v++;
1832: }
1833: PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1834: *norm = PetscSqrtReal(*norm);
1835: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1836: } else if (type == NORM_1) { /* max column norm */
1837: PetscReal *tmp, *tmp2;
1838: PetscInt *jj, *garray = aij->garray;
1839: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1840: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1841: *norm = 0.0;
1842: v = amata;
1843: jj = amat->j;
1844: for (j = 0; j < amat->nz; j++) {
1845: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1846: v++;
1847: }
1848: v = bmata;
1849: jj = bmat->j;
1850: for (j = 0; j < bmat->nz; j++) {
1851: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1852: v++;
1853: }
1854: PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1855: for (j = 0; j < mat->cmap->N; j++) {
1856: if (tmp2[j] > *norm) *norm = tmp2[j];
1857: }
1858: PetscCall(PetscFree(tmp));
1859: PetscCall(PetscFree(tmp2));
1860: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1861: } else if (type == NORM_INFINITY) { /* max row norm */
1862: PetscReal ntemp = 0.0;
1863: for (j = 0; j < aij->A->rmap->n; j++) {
1864: v = amata + amat->i[j];
1865: sum = 0.0;
1866: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1867: sum += PetscAbsScalar(*v);
1868: v++;
1869: }
1870: v = bmata + bmat->i[j];
1871: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1872: sum += PetscAbsScalar(*v);
1873: v++;
1874: }
1875: if (sum > ntemp) ntemp = sum;
1876: }
1877: PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1878: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1879: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1880: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1881: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1882: }
1883: PetscFunctionReturn(PETSC_SUCCESS);
1884: }
1886: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1887: {
1888: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1889: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1890: 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;
1891: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1892: Mat B, A_diag, *B_diag;
1893: const MatScalar *pbv, *bv;
1895: PetscFunctionBegin;
1896: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1897: ma = A->rmap->n;
1898: na = A->cmap->n;
1899: mb = a->B->rmap->n;
1900: nb = a->B->cmap->n;
1901: ai = Aloc->i;
1902: aj = Aloc->j;
1903: bi = Bloc->i;
1904: bj = Bloc->j;
1905: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1906: PetscInt *d_nnz, *g_nnz, *o_nnz;
1907: PetscSFNode *oloc;
1908: PETSC_UNUSED PetscSF sf;
1910: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1911: /* compute d_nnz for preallocation */
1912: PetscCall(PetscArrayzero(d_nnz, na));
1913: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1914: /* compute local off-diagonal contributions */
1915: PetscCall(PetscArrayzero(g_nnz, nb));
1916: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1917: /* map those to global */
1918: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1919: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1920: PetscCall(PetscSFSetFromOptions(sf));
1921: PetscCall(PetscArrayzero(o_nnz, na));
1922: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1923: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1924: PetscCall(PetscSFDestroy(&sf));
1926: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1927: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1928: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1929: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1930: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1931: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1932: } else {
1933: B = *matout;
1934: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1935: }
1937: b = (Mat_MPIAIJ *)B->data;
1938: A_diag = a->A;
1939: B_diag = &b->A;
1940: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1941: A_diag_ncol = A_diag->cmap->N;
1942: B_diag_ilen = sub_B_diag->ilen;
1943: B_diag_i = sub_B_diag->i;
1945: /* Set ilen for diagonal of B */
1946: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1948: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1949: very quickly (=without using MatSetValues), because all writes are local. */
1950: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1951: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1953: /* copy over the B part */
1954: PetscCall(PetscMalloc1(bi[mb], &cols));
1955: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1956: pbv = bv;
1957: row = A->rmap->rstart;
1958: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1959: cols_tmp = cols;
1960: for (i = 0; i < mb; i++) {
1961: ncol = bi[i + 1] - bi[i];
1962: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1963: row++;
1964: if (pbv) pbv += ncol;
1965: if (cols_tmp) cols_tmp += ncol;
1966: }
1967: PetscCall(PetscFree(cols));
1968: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1970: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1971: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1972: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1973: *matout = B;
1974: } else {
1975: PetscCall(MatHeaderMerge(A, &B));
1976: }
1977: PetscFunctionReturn(PETSC_SUCCESS);
1978: }
1980: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1981: {
1982: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1983: Mat a = aij->A, b = aij->B;
1984: PetscInt s1, s2, s3;
1986: PetscFunctionBegin;
1987: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1988: if (rr) {
1989: PetscCall(VecGetLocalSize(rr, &s1));
1990: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1991: /* Overlap communication with computation. */
1992: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1993: }
1994: if (ll) {
1995: PetscCall(VecGetLocalSize(ll, &s1));
1996: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1997: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1998: }
1999: /* scale the diagonal block */
2000: PetscUseTypeMethod(a, diagonalscale, ll, rr);
2002: if (rr) {
2003: /* Do a scatter end and then right scale the off-diagonal block */
2004: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2005: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
2006: }
2007: PetscFunctionReturn(PETSC_SUCCESS);
2008: }
2010: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2011: {
2012: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2014: PetscFunctionBegin;
2015: PetscCall(MatSetUnfactored(a->A));
2016: PetscFunctionReturn(PETSC_SUCCESS);
2017: }
2019: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2020: {
2021: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2022: Mat a, b, c, d;
2023: PetscBool flg;
2025: PetscFunctionBegin;
2026: a = matA->A;
2027: b = matA->B;
2028: c = matB->A;
2029: d = matB->B;
2031: PetscCall(MatEqual(a, c, &flg));
2032: if (flg) PetscCall(MatEqual(b, d, &flg));
2033: PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2034: PetscFunctionReturn(PETSC_SUCCESS);
2035: }
2037: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2038: {
2039: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2040: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2042: PetscFunctionBegin;
2043: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2044: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2045: /* because of the column compression in the off-processor part of the matrix a->B,
2046: the number of columns in a->B and b->B may be different, hence we cannot call
2047: the MatCopy() directly on the two parts. If need be, we can provide a more
2048: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2049: then copying the submatrices */
2050: PetscCall(MatCopy_Basic(A, B, str));
2051: } else {
2052: PetscCall(MatCopy(a->A, b->A, str));
2053: PetscCall(MatCopy(a->B, b->B, str));
2054: }
2055: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2056: PetscFunctionReturn(PETSC_SUCCESS);
2057: }
2059: /*
2060: Computes the number of nonzeros per row needed for preallocation when X and Y
2061: have different nonzero structure.
2062: */
2063: 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)
2064: {
2065: PetscInt i, j, k, nzx, nzy;
2067: PetscFunctionBegin;
2068: /* Set the number of nonzeros in the new matrix */
2069: for (i = 0; i < m; i++) {
2070: const PetscInt *xjj = xj + xi[i], *yjj = yj + yi[i];
2071: nzx = xi[i + 1] - xi[i];
2072: nzy = yi[i + 1] - yi[i];
2073: nnz[i] = 0;
2074: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2075: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2076: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2077: nnz[i]++;
2078: }
2079: for (; k < nzy; k++) nnz[i]++;
2080: }
2081: PetscFunctionReturn(PETSC_SUCCESS);
2082: }
2084: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2085: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2086: {
2087: PetscInt m = Y->rmap->N;
2088: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2089: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2091: PetscFunctionBegin;
2092: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2093: PetscFunctionReturn(PETSC_SUCCESS);
2094: }
2096: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2097: {
2098: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2100: PetscFunctionBegin;
2101: if (str == SAME_NONZERO_PATTERN) {
2102: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2103: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2104: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2105: PetscCall(MatAXPY_Basic(Y, a, X, str));
2106: } else {
2107: Mat B;
2108: PetscInt *nnz_d, *nnz_o;
2110: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2111: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2112: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2113: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2114: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2115: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2116: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2117: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2118: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2119: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2120: PetscCall(MatHeaderMerge(Y, &B));
2121: PetscCall(PetscFree(nnz_d));
2122: PetscCall(PetscFree(nnz_o));
2123: }
2124: PetscFunctionReturn(PETSC_SUCCESS);
2125: }
2127: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2129: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2130: {
2131: PetscFunctionBegin;
2132: if (PetscDefined(USE_COMPLEX)) {
2133: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2135: PetscCall(MatConjugate_SeqAIJ(aij->A));
2136: PetscCall(MatConjugate_SeqAIJ(aij->B));
2137: }
2138: PetscFunctionReturn(PETSC_SUCCESS);
2139: }
2141: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2142: {
2143: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2145: PetscFunctionBegin;
2146: PetscCall(MatRealPart(a->A));
2147: PetscCall(MatRealPart(a->B));
2148: PetscFunctionReturn(PETSC_SUCCESS);
2149: }
2151: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2152: {
2153: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2155: PetscFunctionBegin;
2156: PetscCall(MatImaginaryPart(a->A));
2157: PetscCall(MatImaginaryPart(a->B));
2158: PetscFunctionReturn(PETSC_SUCCESS);
2159: }
2161: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2162: {
2163: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2164: PetscInt i, *idxb = NULL, m = A->rmap->n;
2165: PetscScalar *va, *vv;
2166: Vec vB, vA;
2167: const PetscScalar *vb;
2169: PetscFunctionBegin;
2170: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2171: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2173: PetscCall(VecGetArrayWrite(vA, &va));
2174: if (idx) {
2175: for (i = 0; i < m; i++) {
2176: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2177: }
2178: }
2180: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2181: PetscCall(PetscMalloc1(m, &idxb));
2182: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2184: PetscCall(VecGetArrayWrite(v, &vv));
2185: PetscCall(VecGetArrayRead(vB, &vb));
2186: for (i = 0; i < m; i++) {
2187: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2188: vv[i] = vb[i];
2189: if (idx) idx[i] = a->garray[idxb[i]];
2190: } else {
2191: vv[i] = va[i];
2192: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2193: }
2194: }
2195: PetscCall(VecRestoreArrayWrite(vA, &vv));
2196: PetscCall(VecRestoreArrayWrite(vA, &va));
2197: PetscCall(VecRestoreArrayRead(vB, &vb));
2198: PetscCall(PetscFree(idxb));
2199: PetscCall(VecDestroy(&vA));
2200: PetscCall(VecDestroy(&vB));
2201: PetscFunctionReturn(PETSC_SUCCESS);
2202: }
2204: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2205: {
2206: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2207: PetscInt m = A->rmap->n, n = A->cmap->n;
2208: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2209: PetscInt *cmap = mat->garray;
2210: PetscInt *diagIdx, *offdiagIdx;
2211: Vec diagV, offdiagV;
2212: PetscScalar *a, *diagA, *offdiagA;
2213: const PetscScalar *ba, *bav;
2214: PetscInt r, j, col, ncols, *bi, *bj;
2215: Mat B = mat->B;
2216: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2218: PetscFunctionBegin;
2219: /* When a process holds entire A and other processes have no entry */
2220: if (A->cmap->N == n) {
2221: PetscCall(VecGetArrayWrite(v, &diagA));
2222: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2223: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2224: PetscCall(VecDestroy(&diagV));
2225: PetscCall(VecRestoreArrayWrite(v, &diagA));
2226: PetscFunctionReturn(PETSC_SUCCESS);
2227: } else if (n == 0) {
2228: if (m) {
2229: PetscCall(VecGetArrayWrite(v, &a));
2230: for (r = 0; r < m; r++) {
2231: a[r] = 0.0;
2232: if (idx) idx[r] = -1;
2233: }
2234: PetscCall(VecRestoreArrayWrite(v, &a));
2235: }
2236: PetscFunctionReturn(PETSC_SUCCESS);
2237: }
2239: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2240: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2241: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2242: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2244: /* Get offdiagIdx[] for implicit 0.0 */
2245: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2246: ba = bav;
2247: bi = b->i;
2248: bj = b->j;
2249: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2250: for (r = 0; r < m; r++) {
2251: ncols = bi[r + 1] - bi[r];
2252: if (ncols == A->cmap->N - n) { /* Brow is dense */
2253: offdiagA[r] = *ba;
2254: offdiagIdx[r] = cmap[0];
2255: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2256: offdiagA[r] = 0.0;
2258: /* Find first hole in the cmap */
2259: for (j = 0; j < ncols; j++) {
2260: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2261: if (col > j && j < cstart) {
2262: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2263: break;
2264: } else if (col > j + n && j >= cstart) {
2265: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2266: break;
2267: }
2268: }
2269: if (j == ncols && ncols < A->cmap->N - n) {
2270: /* a hole is outside compressed Bcols */
2271: if (ncols == 0) {
2272: if (cstart) {
2273: offdiagIdx[r] = 0;
2274: } else offdiagIdx[r] = cend;
2275: } else { /* ncols > 0 */
2276: offdiagIdx[r] = cmap[ncols - 1] + 1;
2277: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2278: }
2279: }
2280: }
2282: for (j = 0; j < ncols; j++) {
2283: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2284: offdiagA[r] = *ba;
2285: offdiagIdx[r] = cmap[*bj];
2286: }
2287: ba++;
2288: bj++;
2289: }
2290: }
2292: PetscCall(VecGetArrayWrite(v, &a));
2293: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2294: for (r = 0; r < m; ++r) {
2295: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2296: a[r] = diagA[r];
2297: if (idx) idx[r] = cstart + diagIdx[r];
2298: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2299: a[r] = diagA[r];
2300: if (idx) {
2301: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2302: idx[r] = cstart + diagIdx[r];
2303: } else idx[r] = offdiagIdx[r];
2304: }
2305: } else {
2306: a[r] = offdiagA[r];
2307: if (idx) idx[r] = offdiagIdx[r];
2308: }
2309: }
2310: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2311: PetscCall(VecRestoreArrayWrite(v, &a));
2312: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2313: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2314: PetscCall(VecDestroy(&diagV));
2315: PetscCall(VecDestroy(&offdiagV));
2316: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2317: PetscFunctionReturn(PETSC_SUCCESS);
2318: }
2320: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2321: {
2322: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2323: PetscInt m = A->rmap->n, n = A->cmap->n;
2324: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2325: PetscInt *cmap = mat->garray;
2326: PetscInt *diagIdx, *offdiagIdx;
2327: Vec diagV, offdiagV;
2328: PetscScalar *a, *diagA, *offdiagA;
2329: const PetscScalar *ba, *bav;
2330: PetscInt r, j, col, ncols, *bi, *bj;
2331: Mat B = mat->B;
2332: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2334: PetscFunctionBegin;
2335: /* When a process holds entire A and other processes have no entry */
2336: if (A->cmap->N == n) {
2337: PetscCall(VecGetArrayWrite(v, &diagA));
2338: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2339: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2340: PetscCall(VecDestroy(&diagV));
2341: PetscCall(VecRestoreArrayWrite(v, &diagA));
2342: PetscFunctionReturn(PETSC_SUCCESS);
2343: } else if (n == 0) {
2344: if (m) {
2345: PetscCall(VecGetArrayWrite(v, &a));
2346: for (r = 0; r < m; r++) {
2347: a[r] = PETSC_MAX_REAL;
2348: if (idx) idx[r] = -1;
2349: }
2350: PetscCall(VecRestoreArrayWrite(v, &a));
2351: }
2352: PetscFunctionReturn(PETSC_SUCCESS);
2353: }
2355: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2356: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2357: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2358: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2360: /* Get offdiagIdx[] for implicit 0.0 */
2361: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2362: ba = bav;
2363: bi = b->i;
2364: bj = b->j;
2365: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2366: for (r = 0; r < m; r++) {
2367: ncols = bi[r + 1] - bi[r];
2368: if (ncols == A->cmap->N - n) { /* Brow is dense */
2369: offdiagA[r] = *ba;
2370: offdiagIdx[r] = cmap[0];
2371: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2372: offdiagA[r] = 0.0;
2374: /* Find first hole in the cmap */
2375: for (j = 0; j < ncols; j++) {
2376: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2377: if (col > j && j < cstart) {
2378: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2379: break;
2380: } else if (col > j + n && j >= cstart) {
2381: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2382: break;
2383: }
2384: }
2385: if (j == ncols && ncols < A->cmap->N - n) {
2386: /* a hole is outside compressed Bcols */
2387: if (ncols == 0) {
2388: if (cstart) {
2389: offdiagIdx[r] = 0;
2390: } else offdiagIdx[r] = cend;
2391: } else { /* ncols > 0 */
2392: offdiagIdx[r] = cmap[ncols - 1] + 1;
2393: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2394: }
2395: }
2396: }
2398: for (j = 0; j < ncols; j++) {
2399: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2400: offdiagA[r] = *ba;
2401: offdiagIdx[r] = cmap[*bj];
2402: }
2403: ba++;
2404: bj++;
2405: }
2406: }
2408: PetscCall(VecGetArrayWrite(v, &a));
2409: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2410: for (r = 0; r < m; ++r) {
2411: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2412: a[r] = diagA[r];
2413: if (idx) idx[r] = cstart + diagIdx[r];
2414: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2415: a[r] = diagA[r];
2416: if (idx) {
2417: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2418: idx[r] = cstart + diagIdx[r];
2419: } else idx[r] = offdiagIdx[r];
2420: }
2421: } else {
2422: a[r] = offdiagA[r];
2423: if (idx) idx[r] = offdiagIdx[r];
2424: }
2425: }
2426: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2427: PetscCall(VecRestoreArrayWrite(v, &a));
2428: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2429: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2430: PetscCall(VecDestroy(&diagV));
2431: PetscCall(VecDestroy(&offdiagV));
2432: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2433: PetscFunctionReturn(PETSC_SUCCESS);
2434: }
2436: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2437: {
2438: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2439: PetscInt m = A->rmap->n, n = A->cmap->n;
2440: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2441: PetscInt *cmap = mat->garray;
2442: PetscInt *diagIdx, *offdiagIdx;
2443: Vec diagV, offdiagV;
2444: PetscScalar *a, *diagA, *offdiagA;
2445: const PetscScalar *ba, *bav;
2446: PetscInt r, j, col, ncols, *bi, *bj;
2447: Mat B = mat->B;
2448: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2450: PetscFunctionBegin;
2451: /* When a process holds entire A and other processes have no entry */
2452: if (A->cmap->N == n) {
2453: PetscCall(VecGetArrayWrite(v, &diagA));
2454: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2455: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2456: PetscCall(VecDestroy(&diagV));
2457: PetscCall(VecRestoreArrayWrite(v, &diagA));
2458: PetscFunctionReturn(PETSC_SUCCESS);
2459: } else if (n == 0) {
2460: if (m) {
2461: PetscCall(VecGetArrayWrite(v, &a));
2462: for (r = 0; r < m; r++) {
2463: a[r] = PETSC_MIN_REAL;
2464: if (idx) idx[r] = -1;
2465: }
2466: PetscCall(VecRestoreArrayWrite(v, &a));
2467: }
2468: PetscFunctionReturn(PETSC_SUCCESS);
2469: }
2471: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2472: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2473: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2474: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2476: /* Get offdiagIdx[] for implicit 0.0 */
2477: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2478: ba = bav;
2479: bi = b->i;
2480: bj = b->j;
2481: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2482: for (r = 0; r < m; r++) {
2483: ncols = bi[r + 1] - bi[r];
2484: if (ncols == A->cmap->N - n) { /* Brow is dense */
2485: offdiagA[r] = *ba;
2486: offdiagIdx[r] = cmap[0];
2487: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2488: offdiagA[r] = 0.0;
2490: /* Find first hole in the cmap */
2491: for (j = 0; j < ncols; j++) {
2492: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2493: if (col > j && j < cstart) {
2494: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2495: break;
2496: } else if (col > j + n && j >= cstart) {
2497: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2498: break;
2499: }
2500: }
2501: if (j == ncols && ncols < A->cmap->N - n) {
2502: /* a hole is outside compressed Bcols */
2503: if (ncols == 0) {
2504: if (cstart) {
2505: offdiagIdx[r] = 0;
2506: } else offdiagIdx[r] = cend;
2507: } else { /* ncols > 0 */
2508: offdiagIdx[r] = cmap[ncols - 1] + 1;
2509: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2510: }
2511: }
2512: }
2514: for (j = 0; j < ncols; j++) {
2515: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2516: offdiagA[r] = *ba;
2517: offdiagIdx[r] = cmap[*bj];
2518: }
2519: ba++;
2520: bj++;
2521: }
2522: }
2524: PetscCall(VecGetArrayWrite(v, &a));
2525: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2526: for (r = 0; r < m; ++r) {
2527: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2528: a[r] = diagA[r];
2529: if (idx) idx[r] = cstart + diagIdx[r];
2530: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2531: a[r] = diagA[r];
2532: if (idx) {
2533: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2534: idx[r] = cstart + diagIdx[r];
2535: } else idx[r] = offdiagIdx[r];
2536: }
2537: } else {
2538: a[r] = offdiagA[r];
2539: if (idx) idx[r] = offdiagIdx[r];
2540: }
2541: }
2542: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2543: PetscCall(VecRestoreArrayWrite(v, &a));
2544: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2545: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2546: PetscCall(VecDestroy(&diagV));
2547: PetscCall(VecDestroy(&offdiagV));
2548: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2549: PetscFunctionReturn(PETSC_SUCCESS);
2550: }
2552: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2553: {
2554: Mat *dummy;
2556: PetscFunctionBegin;
2557: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2558: *newmat = *dummy;
2559: PetscCall(PetscFree(dummy));
2560: PetscFunctionReturn(PETSC_SUCCESS);
2561: }
2563: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2564: {
2565: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2567: PetscFunctionBegin;
2568: PetscCall(MatInvertBlockDiagonal(a->A, values));
2569: A->factorerrortype = a->A->factorerrortype;
2570: PetscFunctionReturn(PETSC_SUCCESS);
2571: }
2573: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2574: {
2575: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2577: PetscFunctionBegin;
2578: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2579: PetscCall(MatSetRandom(aij->A, rctx));
2580: if (x->assembled) {
2581: PetscCall(MatSetRandom(aij->B, rctx));
2582: } else {
2583: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2584: }
2585: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2586: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2587: PetscFunctionReturn(PETSC_SUCCESS);
2588: }
2590: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2591: {
2592: PetscFunctionBegin;
2593: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2594: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2595: PetscFunctionReturn(PETSC_SUCCESS);
2596: }
2598: /*@
2599: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2601: Not Collective
2603: Input Parameter:
2604: . A - the matrix
2606: Output Parameter:
2607: . nz - the number of nonzeros
2609: Level: advanced
2611: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2612: @*/
2613: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2614: {
2615: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2616: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2617: PetscBool isaij;
2619: PetscFunctionBegin;
2620: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2621: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2622: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2623: PetscFunctionReturn(PETSC_SUCCESS);
2624: }
2626: /*@
2627: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2629: Collective
2631: Input Parameters:
2632: + A - the matrix
2633: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2635: Level: advanced
2637: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2638: @*/
2639: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2640: {
2641: PetscFunctionBegin;
2642: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2643: PetscFunctionReturn(PETSC_SUCCESS);
2644: }
2646: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2647: {
2648: PetscBool sc = PETSC_FALSE, flg;
2650: PetscFunctionBegin;
2651: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2652: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2653: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2654: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2655: PetscOptionsHeadEnd();
2656: PetscFunctionReturn(PETSC_SUCCESS);
2657: }
2659: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2660: {
2661: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2662: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2664: PetscFunctionBegin;
2665: if (!Y->preallocated) {
2666: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2667: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2668: PetscInt nonew = aij->nonew;
2669: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2670: aij->nonew = nonew;
2671: }
2672: PetscCall(MatShift_Basic(Y, a));
2673: PetscFunctionReturn(PETSC_SUCCESS);
2674: }
2676: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2677: {
2678: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2680: PetscFunctionBegin;
2681: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2682: PetscCall(MatMissingDiagonal(a->A, missing, d));
2683: if (d) {
2684: PetscInt rstart;
2685: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2686: *d += rstart;
2687: }
2688: PetscFunctionReturn(PETSC_SUCCESS);
2689: }
2691: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2692: {
2693: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2695: PetscFunctionBegin;
2696: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2697: PetscFunctionReturn(PETSC_SUCCESS);
2698: }
2700: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2701: {
2702: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2704: PetscFunctionBegin;
2705: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2706: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2707: PetscFunctionReturn(PETSC_SUCCESS);
2708: }
2710: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2711: MatGetRow_MPIAIJ,
2712: MatRestoreRow_MPIAIJ,
2713: MatMult_MPIAIJ,
2714: /* 4*/ MatMultAdd_MPIAIJ,
2715: MatMultTranspose_MPIAIJ,
2716: MatMultTransposeAdd_MPIAIJ,
2717: NULL,
2718: NULL,
2719: NULL,
2720: /*10*/ NULL,
2721: NULL,
2722: NULL,
2723: MatSOR_MPIAIJ,
2724: MatTranspose_MPIAIJ,
2725: /*15*/ MatGetInfo_MPIAIJ,
2726: MatEqual_MPIAIJ,
2727: MatGetDiagonal_MPIAIJ,
2728: MatDiagonalScale_MPIAIJ,
2729: MatNorm_MPIAIJ,
2730: /*20*/ MatAssemblyBegin_MPIAIJ,
2731: MatAssemblyEnd_MPIAIJ,
2732: MatSetOption_MPIAIJ,
2733: MatZeroEntries_MPIAIJ,
2734: /*24*/ MatZeroRows_MPIAIJ,
2735: NULL,
2736: NULL,
2737: NULL,
2738: NULL,
2739: /*29*/ MatSetUp_MPI_Hash,
2740: NULL,
2741: NULL,
2742: MatGetDiagonalBlock_MPIAIJ,
2743: NULL,
2744: /*34*/ MatDuplicate_MPIAIJ,
2745: NULL,
2746: NULL,
2747: NULL,
2748: NULL,
2749: /*39*/ MatAXPY_MPIAIJ,
2750: MatCreateSubMatrices_MPIAIJ,
2751: MatIncreaseOverlap_MPIAIJ,
2752: MatGetValues_MPIAIJ,
2753: MatCopy_MPIAIJ,
2754: /*44*/ MatGetRowMax_MPIAIJ,
2755: MatScale_MPIAIJ,
2756: MatShift_MPIAIJ,
2757: MatDiagonalSet_MPIAIJ,
2758: MatZeroRowsColumns_MPIAIJ,
2759: /*49*/ MatSetRandom_MPIAIJ,
2760: MatGetRowIJ_MPIAIJ,
2761: MatRestoreRowIJ_MPIAIJ,
2762: NULL,
2763: NULL,
2764: /*54*/ MatFDColoringCreate_MPIXAIJ,
2765: NULL,
2766: MatSetUnfactored_MPIAIJ,
2767: MatPermute_MPIAIJ,
2768: NULL,
2769: /*59*/ MatCreateSubMatrix_MPIAIJ,
2770: MatDestroy_MPIAIJ,
2771: MatView_MPIAIJ,
2772: NULL,
2773: NULL,
2774: /*64*/ NULL,
2775: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2776: NULL,
2777: NULL,
2778: NULL,
2779: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2780: MatGetRowMinAbs_MPIAIJ,
2781: NULL,
2782: NULL,
2783: NULL,
2784: NULL,
2785: /*75*/ MatFDColoringApply_AIJ,
2786: MatSetFromOptions_MPIAIJ,
2787: NULL,
2788: NULL,
2789: MatFindZeroDiagonals_MPIAIJ,
2790: /*80*/ NULL,
2791: NULL,
2792: NULL,
2793: /*83*/ MatLoad_MPIAIJ,
2794: MatIsSymmetric_MPIAIJ,
2795: NULL,
2796: NULL,
2797: NULL,
2798: NULL,
2799: /*89*/ NULL,
2800: NULL,
2801: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2802: NULL,
2803: NULL,
2804: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2805: NULL,
2806: NULL,
2807: NULL,
2808: MatBindToCPU_MPIAIJ,
2809: /*99*/ MatProductSetFromOptions_MPIAIJ,
2810: NULL,
2811: NULL,
2812: MatConjugate_MPIAIJ,
2813: NULL,
2814: /*104*/ MatSetValuesRow_MPIAIJ,
2815: MatRealPart_MPIAIJ,
2816: MatImaginaryPart_MPIAIJ,
2817: NULL,
2818: NULL,
2819: /*109*/ NULL,
2820: NULL,
2821: MatGetRowMin_MPIAIJ,
2822: NULL,
2823: MatMissingDiagonal_MPIAIJ,
2824: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2825: NULL,
2826: MatGetGhosts_MPIAIJ,
2827: NULL,
2828: NULL,
2829: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2830: NULL,
2831: NULL,
2832: NULL,
2833: MatGetMultiProcBlock_MPIAIJ,
2834: /*124*/ MatFindNonzeroRows_MPIAIJ,
2835: MatGetColumnReductions_MPIAIJ,
2836: MatInvertBlockDiagonal_MPIAIJ,
2837: MatInvertVariableBlockDiagonal_MPIAIJ,
2838: MatCreateSubMatricesMPI_MPIAIJ,
2839: /*129*/ NULL,
2840: NULL,
2841: NULL,
2842: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2843: NULL,
2844: /*134*/ NULL,
2845: NULL,
2846: NULL,
2847: NULL,
2848: NULL,
2849: /*139*/ MatSetBlockSizes_MPIAIJ,
2850: NULL,
2851: NULL,
2852: MatFDColoringSetUp_MPIXAIJ,
2853: MatFindOffBlockDiagonalEntries_MPIAIJ,
2854: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2855: /*145*/ NULL,
2856: NULL,
2857: NULL,
2858: MatCreateGraph_Simple_AIJ,
2859: NULL,
2860: /*150*/ NULL,
2861: MatEliminateZeros_MPIAIJ};
2863: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2864: {
2865: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2867: PetscFunctionBegin;
2868: PetscCall(MatStoreValues(aij->A));
2869: PetscCall(MatStoreValues(aij->B));
2870: PetscFunctionReturn(PETSC_SUCCESS);
2871: }
2873: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2874: {
2875: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2877: PetscFunctionBegin;
2878: PetscCall(MatRetrieveValues(aij->A));
2879: PetscCall(MatRetrieveValues(aij->B));
2880: PetscFunctionReturn(PETSC_SUCCESS);
2881: }
2883: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2884: {
2885: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2886: PetscMPIInt size;
2888: PetscFunctionBegin;
2889: if (B->hash_active) {
2890: B->ops[0] = b->cops;
2891: B->hash_active = PETSC_FALSE;
2892: }
2893: PetscCall(PetscLayoutSetUp(B->rmap));
2894: PetscCall(PetscLayoutSetUp(B->cmap));
2896: #if defined(PETSC_USE_CTABLE)
2897: PetscCall(PetscHMapIDestroy(&b->colmap));
2898: #else
2899: PetscCall(PetscFree(b->colmap));
2900: #endif
2901: PetscCall(PetscFree(b->garray));
2902: PetscCall(VecDestroy(&b->lvec));
2903: PetscCall(VecScatterDestroy(&b->Mvctx));
2905: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2906: PetscCall(MatDestroy(&b->B));
2907: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2908: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2909: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2910: PetscCall(MatSetType(b->B, MATSEQAIJ));
2912: PetscCall(MatDestroy(&b->A));
2913: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2914: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2915: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2916: PetscCall(MatSetType(b->A, MATSEQAIJ));
2918: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2919: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2920: B->preallocated = PETSC_TRUE;
2921: B->was_assembled = PETSC_FALSE;
2922: B->assembled = PETSC_FALSE;
2923: PetscFunctionReturn(PETSC_SUCCESS);
2924: }
2926: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2927: {
2928: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2930: PetscFunctionBegin;
2932: PetscCall(PetscLayoutSetUp(B->rmap));
2933: PetscCall(PetscLayoutSetUp(B->cmap));
2935: #if defined(PETSC_USE_CTABLE)
2936: PetscCall(PetscHMapIDestroy(&b->colmap));
2937: #else
2938: PetscCall(PetscFree(b->colmap));
2939: #endif
2940: PetscCall(PetscFree(b->garray));
2941: PetscCall(VecDestroy(&b->lvec));
2942: PetscCall(VecScatterDestroy(&b->Mvctx));
2944: PetscCall(MatResetPreallocation(b->A));
2945: PetscCall(MatResetPreallocation(b->B));
2946: B->preallocated = PETSC_TRUE;
2947: B->was_assembled = PETSC_FALSE;
2948: B->assembled = PETSC_FALSE;
2949: PetscFunctionReturn(PETSC_SUCCESS);
2950: }
2952: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2953: {
2954: Mat mat;
2955: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2957: PetscFunctionBegin;
2958: *newmat = NULL;
2959: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2960: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2961: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2962: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2963: a = (Mat_MPIAIJ *)mat->data;
2965: mat->factortype = matin->factortype;
2966: mat->assembled = matin->assembled;
2967: mat->insertmode = NOT_SET_VALUES;
2969: a->size = oldmat->size;
2970: a->rank = oldmat->rank;
2971: a->donotstash = oldmat->donotstash;
2972: a->roworiented = oldmat->roworiented;
2973: a->rowindices = NULL;
2974: a->rowvalues = NULL;
2975: a->getrowactive = PETSC_FALSE;
2977: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2978: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2979: if (matin->hash_active) {
2980: PetscCall(MatSetUp(mat));
2981: } else {
2982: mat->preallocated = matin->preallocated;
2983: if (oldmat->colmap) {
2984: #if defined(PETSC_USE_CTABLE)
2985: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2986: #else
2987: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2988: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
2989: #endif
2990: } else a->colmap = NULL;
2991: if (oldmat->garray) {
2992: PetscInt len;
2993: len = oldmat->B->cmap->n;
2994: PetscCall(PetscMalloc1(len + 1, &a->garray));
2995: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
2996: } else a->garray = NULL;
2998: /* It may happen MatDuplicate is called with a non-assembled matrix
2999: In fact, MatDuplicate only requires the matrix to be preallocated
3000: This may happen inside a DMCreateMatrix_Shell */
3001: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3002: if (oldmat->Mvctx) PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3003: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3004: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3005: }
3006: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3007: *newmat = mat;
3008: PetscFunctionReturn(PETSC_SUCCESS);
3009: }
3011: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3012: {
3013: PetscBool isbinary, ishdf5;
3015: PetscFunctionBegin;
3018: /* force binary viewer to load .info file if it has not yet done so */
3019: PetscCall(PetscViewerSetUp(viewer));
3020: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3021: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3022: if (isbinary) {
3023: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3024: } else if (ishdf5) {
3025: #if defined(PETSC_HAVE_HDF5)
3026: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3027: #else
3028: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3029: #endif
3030: } else {
3031: 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);
3032: }
3033: PetscFunctionReturn(PETSC_SUCCESS);
3034: }
3036: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3037: {
3038: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3039: PetscInt *rowidxs, *colidxs;
3040: PetscScalar *matvals;
3042: PetscFunctionBegin;
3043: PetscCall(PetscViewerSetUp(viewer));
3045: /* read in matrix header */
3046: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3047: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3048: M = header[1];
3049: N = header[2];
3050: nz = header[3];
3051: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3052: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3053: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3055: /* set block sizes from the viewer's .info file */
3056: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3057: /* set global sizes if not set already */
3058: if (mat->rmap->N < 0) mat->rmap->N = M;
3059: if (mat->cmap->N < 0) mat->cmap->N = N;
3060: PetscCall(PetscLayoutSetUp(mat->rmap));
3061: PetscCall(PetscLayoutSetUp(mat->cmap));
3063: /* check if the matrix sizes are correct */
3064: PetscCall(MatGetSize(mat, &rows, &cols));
3065: 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);
3067: /* read in row lengths and build row indices */
3068: PetscCall(MatGetLocalSize(mat, &m, NULL));
3069: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3070: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3071: rowidxs[0] = 0;
3072: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3073: if (nz != PETSC_MAX_INT) {
3074: PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3075: 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);
3076: }
3078: /* read in column indices and matrix values */
3079: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3080: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3081: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3082: /* store matrix indices and values */
3083: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3084: PetscCall(PetscFree(rowidxs));
3085: PetscCall(PetscFree2(colidxs, matvals));
3086: PetscFunctionReturn(PETSC_SUCCESS);
3087: }
3089: /* Not scalable because of ISAllGather() unless getting all columns. */
3090: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3091: {
3092: IS iscol_local;
3093: PetscBool isstride;
3094: PetscMPIInt lisstride = 0, gisstride;
3096: PetscFunctionBegin;
3097: /* check if we are grabbing all columns*/
3098: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3100: if (isstride) {
3101: PetscInt start, len, mstart, mlen;
3102: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3103: PetscCall(ISGetLocalSize(iscol, &len));
3104: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3105: if (mstart == start && mlen - mstart == len) lisstride = 1;
3106: }
3108: PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3109: if (gisstride) {
3110: PetscInt N;
3111: PetscCall(MatGetSize(mat, NULL, &N));
3112: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3113: PetscCall(ISSetIdentity(iscol_local));
3114: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3115: } else {
3116: PetscInt cbs;
3117: PetscCall(ISGetBlockSize(iscol, &cbs));
3118: PetscCall(ISAllGather(iscol, &iscol_local));
3119: PetscCall(ISSetBlockSize(iscol_local, cbs));
3120: }
3122: *isseq = iscol_local;
3123: PetscFunctionReturn(PETSC_SUCCESS);
3124: }
3126: /*
3127: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3128: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3130: Input Parameters:
3131: + mat - matrix
3132: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3133: i.e., mat->rstart <= isrow[i] < mat->rend
3134: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3135: i.e., mat->cstart <= iscol[i] < mat->cend
3137: Output Parameters:
3138: + isrow_d - sequential row index set for retrieving mat->A
3139: . iscol_d - sequential column index set for retrieving mat->A
3140: . iscol_o - sequential column index set for retrieving mat->B
3141: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3142: */
3143: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3144: {
3145: Vec x, cmap;
3146: const PetscInt *is_idx;
3147: PetscScalar *xarray, *cmaparray;
3148: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3149: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3150: Mat B = a->B;
3151: Vec lvec = a->lvec, lcmap;
3152: PetscInt i, cstart, cend, Bn = B->cmap->N;
3153: MPI_Comm comm;
3154: VecScatter Mvctx = a->Mvctx;
3156: PetscFunctionBegin;
3157: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3158: PetscCall(ISGetLocalSize(iscol, &ncols));
3160: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3161: PetscCall(MatCreateVecs(mat, &x, NULL));
3162: PetscCall(VecSet(x, -1.0));
3163: PetscCall(VecDuplicate(x, &cmap));
3164: PetscCall(VecSet(cmap, -1.0));
3166: /* Get start indices */
3167: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3168: isstart -= ncols;
3169: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3171: PetscCall(ISGetIndices(iscol, &is_idx));
3172: PetscCall(VecGetArray(x, &xarray));
3173: PetscCall(VecGetArray(cmap, &cmaparray));
3174: PetscCall(PetscMalloc1(ncols, &idx));
3175: for (i = 0; i < ncols; i++) {
3176: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3177: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3178: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3179: }
3180: PetscCall(VecRestoreArray(x, &xarray));
3181: PetscCall(VecRestoreArray(cmap, &cmaparray));
3182: PetscCall(ISRestoreIndices(iscol, &is_idx));
3184: /* Get iscol_d */
3185: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3186: PetscCall(ISGetBlockSize(iscol, &i));
3187: PetscCall(ISSetBlockSize(*iscol_d, i));
3189: /* Get isrow_d */
3190: PetscCall(ISGetLocalSize(isrow, &m));
3191: rstart = mat->rmap->rstart;
3192: PetscCall(PetscMalloc1(m, &idx));
3193: PetscCall(ISGetIndices(isrow, &is_idx));
3194: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3195: PetscCall(ISRestoreIndices(isrow, &is_idx));
3197: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3198: PetscCall(ISGetBlockSize(isrow, &i));
3199: PetscCall(ISSetBlockSize(*isrow_d, i));
3201: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3202: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3203: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3205: PetscCall(VecDuplicate(lvec, &lcmap));
3207: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3208: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3210: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3211: /* off-process column indices */
3212: count = 0;
3213: PetscCall(PetscMalloc1(Bn, &idx));
3214: PetscCall(PetscMalloc1(Bn, &cmap1));
3216: PetscCall(VecGetArray(lvec, &xarray));
3217: PetscCall(VecGetArray(lcmap, &cmaparray));
3218: for (i = 0; i < Bn; i++) {
3219: if (PetscRealPart(xarray[i]) > -1.0) {
3220: idx[count] = i; /* local column index in off-diagonal part B */
3221: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3222: count++;
3223: }
3224: }
3225: PetscCall(VecRestoreArray(lvec, &xarray));
3226: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3228: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3229: /* cannot ensure iscol_o has same blocksize as iscol! */
3231: PetscCall(PetscFree(idx));
3232: *garray = cmap1;
3234: PetscCall(VecDestroy(&x));
3235: PetscCall(VecDestroy(&cmap));
3236: PetscCall(VecDestroy(&lcmap));
3237: PetscFunctionReturn(PETSC_SUCCESS);
3238: }
3240: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3241: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3242: {
3243: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3244: Mat M = NULL;
3245: MPI_Comm comm;
3246: IS iscol_d, isrow_d, iscol_o;
3247: Mat Asub = NULL, Bsub = NULL;
3248: PetscInt n;
3250: PetscFunctionBegin;
3251: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3253: if (call == MAT_REUSE_MATRIX) {
3254: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3255: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3256: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3258: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3259: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3261: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3262: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3264: /* Update diagonal and off-diagonal portions of submat */
3265: asub = (Mat_MPIAIJ *)(*submat)->data;
3266: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3267: PetscCall(ISGetLocalSize(iscol_o, &n));
3268: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3269: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3270: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3272: } else { /* call == MAT_INITIAL_MATRIX) */
3273: const PetscInt *garray;
3274: PetscInt BsubN;
3276: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3277: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3279: /* Create local submatrices Asub and Bsub */
3280: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3281: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3283: /* Create submatrix M */
3284: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3286: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3287: asub = (Mat_MPIAIJ *)M->data;
3289: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3290: n = asub->B->cmap->N;
3291: if (BsubN > n) {
3292: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3293: const PetscInt *idx;
3294: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3295: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3297: PetscCall(PetscMalloc1(n, &idx_new));
3298: j = 0;
3299: PetscCall(ISGetIndices(iscol_o, &idx));
3300: for (i = 0; i < n; i++) {
3301: if (j >= BsubN) break;
3302: while (subgarray[i] > garray[j]) j++;
3304: if (subgarray[i] == garray[j]) {
3305: idx_new[i] = idx[j++];
3306: } 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]);
3307: }
3308: PetscCall(ISRestoreIndices(iscol_o, &idx));
3310: PetscCall(ISDestroy(&iscol_o));
3311: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3313: } else if (BsubN < n) {
3314: 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);
3315: }
3317: PetscCall(PetscFree(garray));
3318: *submat = M;
3320: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3321: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3322: PetscCall(ISDestroy(&isrow_d));
3324: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3325: PetscCall(ISDestroy(&iscol_d));
3327: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3328: PetscCall(ISDestroy(&iscol_o));
3329: }
3330: PetscFunctionReturn(PETSC_SUCCESS);
3331: }
3333: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3334: {
3335: IS iscol_local = NULL, isrow_d;
3336: PetscInt csize;
3337: PetscInt n, i, j, start, end;
3338: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3339: MPI_Comm comm;
3341: PetscFunctionBegin;
3342: /* If isrow has same processor distribution as mat,
3343: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3344: if (call == MAT_REUSE_MATRIX) {
3345: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3346: if (isrow_d) {
3347: sameRowDist = PETSC_TRUE;
3348: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3349: } else {
3350: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3351: if (iscol_local) {
3352: sameRowDist = PETSC_TRUE;
3353: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3354: }
3355: }
3356: } else {
3357: /* Check if isrow has same processor distribution as mat */
3358: sameDist[0] = PETSC_FALSE;
3359: PetscCall(ISGetLocalSize(isrow, &n));
3360: if (!n) {
3361: sameDist[0] = PETSC_TRUE;
3362: } else {
3363: PetscCall(ISGetMinMax(isrow, &i, &j));
3364: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3365: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3366: }
3368: /* Check if iscol has same processor distribution as mat */
3369: sameDist[1] = PETSC_FALSE;
3370: PetscCall(ISGetLocalSize(iscol, &n));
3371: if (!n) {
3372: sameDist[1] = PETSC_TRUE;
3373: } else {
3374: PetscCall(ISGetMinMax(iscol, &i, &j));
3375: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3376: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3377: }
3379: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3380: PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3381: sameRowDist = tsameDist[0];
3382: }
3384: if (sameRowDist) {
3385: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3386: /* isrow and iscol have same processor distribution as mat */
3387: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3388: PetscFunctionReturn(PETSC_SUCCESS);
3389: } else { /* sameRowDist */
3390: /* isrow has same processor distribution as mat */
3391: if (call == MAT_INITIAL_MATRIX) {
3392: PetscBool sorted;
3393: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3394: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3395: PetscCall(ISGetSize(iscol, &i));
3396: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3398: PetscCall(ISSorted(iscol_local, &sorted));
3399: if (sorted) {
3400: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3401: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3402: PetscFunctionReturn(PETSC_SUCCESS);
3403: }
3404: } else { /* call == MAT_REUSE_MATRIX */
3405: IS iscol_sub;
3406: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3407: if (iscol_sub) {
3408: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3409: PetscFunctionReturn(PETSC_SUCCESS);
3410: }
3411: }
3412: }
3413: }
3415: /* General case: iscol -> iscol_local which has global size of iscol */
3416: if (call == MAT_REUSE_MATRIX) {
3417: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3418: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3419: } else {
3420: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3421: }
3423: PetscCall(ISGetLocalSize(iscol, &csize));
3424: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3426: if (call == MAT_INITIAL_MATRIX) {
3427: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3428: PetscCall(ISDestroy(&iscol_local));
3429: }
3430: PetscFunctionReturn(PETSC_SUCCESS);
3431: }
3433: /*@C
3434: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3435: and "off-diagonal" part of the matrix in CSR format.
3437: Collective
3439: Input Parameters:
3440: + comm - MPI communicator
3441: . A - "diagonal" portion of matrix
3442: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3443: - garray - global index of `B` columns
3445: Output Parameter:
3446: . mat - the matrix, with input `A` as its local diagonal matrix
3448: Level: advanced
3450: Notes:
3451: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3453: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3455: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3456: @*/
3457: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3458: {
3459: Mat_MPIAIJ *maij;
3460: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3461: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3462: const PetscScalar *oa;
3463: Mat Bnew;
3464: PetscInt m, n, N;
3465: MatType mpi_mat_type;
3467: PetscFunctionBegin;
3468: PetscCall(MatCreate(comm, mat));
3469: PetscCall(MatGetSize(A, &m, &n));
3470: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3471: 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);
3472: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3473: /* 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); */
3475: /* Get global columns of mat */
3476: PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3478: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3479: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3480: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3481: PetscCall(MatSetType(*mat, mpi_mat_type));
3483: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3484: maij = (Mat_MPIAIJ *)(*mat)->data;
3486: (*mat)->preallocated = PETSC_TRUE;
3488: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3489: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3491: /* Set A as diagonal portion of *mat */
3492: maij->A = A;
3494: nz = oi[m];
3495: for (i = 0; i < nz; i++) {
3496: col = oj[i];
3497: oj[i] = garray[col];
3498: }
3500: /* Set Bnew as off-diagonal portion of *mat */
3501: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3502: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3503: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3504: bnew = (Mat_SeqAIJ *)Bnew->data;
3505: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3506: maij->B = Bnew;
3508: 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);
3510: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3511: b->free_a = PETSC_FALSE;
3512: b->free_ij = PETSC_FALSE;
3513: PetscCall(MatDestroy(&B));
3515: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3516: bnew->free_a = PETSC_TRUE;
3517: bnew->free_ij = PETSC_TRUE;
3519: /* condense columns of maij->B */
3520: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3521: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3522: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3523: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3524: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3525: PetscFunctionReturn(PETSC_SUCCESS);
3526: }
3528: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3530: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3531: {
3532: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3533: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3534: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3535: Mat M, Msub, B = a->B;
3536: MatScalar *aa;
3537: Mat_SeqAIJ *aij;
3538: PetscInt *garray = a->garray, *colsub, Ncols;
3539: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3540: IS iscol_sub, iscmap;
3541: const PetscInt *is_idx, *cmap;
3542: PetscBool allcolumns = PETSC_FALSE;
3543: MPI_Comm comm;
3545: PetscFunctionBegin;
3546: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3547: if (call == MAT_REUSE_MATRIX) {
3548: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3549: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3550: PetscCall(ISGetLocalSize(iscol_sub, &count));
3552: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3553: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3555: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3556: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3558: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3560: } else { /* call == MAT_INITIAL_MATRIX) */
3561: PetscBool flg;
3563: PetscCall(ISGetLocalSize(iscol, &n));
3564: PetscCall(ISGetSize(iscol, &Ncols));
3566: /* (1) iscol -> nonscalable iscol_local */
3567: /* Check for special case: each processor gets entire matrix columns */
3568: PetscCall(ISIdentity(iscol_local, &flg));
3569: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3570: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3571: if (allcolumns) {
3572: iscol_sub = iscol_local;
3573: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3574: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3576: } else {
3577: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3578: PetscInt *idx, *cmap1, k;
3579: PetscCall(PetscMalloc1(Ncols, &idx));
3580: PetscCall(PetscMalloc1(Ncols, &cmap1));
3581: PetscCall(ISGetIndices(iscol_local, &is_idx));
3582: count = 0;
3583: k = 0;
3584: for (i = 0; i < Ncols; i++) {
3585: j = is_idx[i];
3586: if (j >= cstart && j < cend) {
3587: /* diagonal part of mat */
3588: idx[count] = j;
3589: cmap1[count++] = i; /* column index in submat */
3590: } else if (Bn) {
3591: /* off-diagonal part of mat */
3592: if (j == garray[k]) {
3593: idx[count] = j;
3594: cmap1[count++] = i; /* column index in submat */
3595: } else if (j > garray[k]) {
3596: while (j > garray[k] && k < Bn - 1) k++;
3597: if (j == garray[k]) {
3598: idx[count] = j;
3599: cmap1[count++] = i; /* column index in submat */
3600: }
3601: }
3602: }
3603: }
3604: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3606: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3607: PetscCall(ISGetBlockSize(iscol, &cbs));
3608: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3610: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3611: }
3613: /* (3) Create sequential Msub */
3614: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3615: }
3617: PetscCall(ISGetLocalSize(iscol_sub, &count));
3618: aij = (Mat_SeqAIJ *)(Msub)->data;
3619: ii = aij->i;
3620: PetscCall(ISGetIndices(iscmap, &cmap));
3622: /*
3623: m - number of local rows
3624: Ncols - number of columns (same on all processors)
3625: rstart - first row in new global matrix generated
3626: */
3627: PetscCall(MatGetSize(Msub, &m, NULL));
3629: if (call == MAT_INITIAL_MATRIX) {
3630: /* (4) Create parallel newmat */
3631: PetscMPIInt rank, size;
3632: PetscInt csize;
3634: PetscCallMPI(MPI_Comm_size(comm, &size));
3635: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3637: /*
3638: Determine the number of non-zeros in the diagonal and off-diagonal
3639: portions of the matrix in order to do correct preallocation
3640: */
3642: /* first get start and end of "diagonal" columns */
3643: PetscCall(ISGetLocalSize(iscol, &csize));
3644: if (csize == PETSC_DECIDE) {
3645: PetscCall(ISGetSize(isrow, &mglobal));
3646: if (mglobal == Ncols) { /* square matrix */
3647: nlocal = m;
3648: } else {
3649: nlocal = Ncols / size + ((Ncols % size) > rank);
3650: }
3651: } else {
3652: nlocal = csize;
3653: }
3654: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3655: rstart = rend - nlocal;
3656: 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);
3658: /* next, compute all the lengths */
3659: jj = aij->j;
3660: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3661: olens = dlens + m;
3662: for (i = 0; i < m; i++) {
3663: jend = ii[i + 1] - ii[i];
3664: olen = 0;
3665: dlen = 0;
3666: for (j = 0; j < jend; j++) {
3667: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3668: else dlen++;
3669: jj++;
3670: }
3671: olens[i] = olen;
3672: dlens[i] = dlen;
3673: }
3675: PetscCall(ISGetBlockSize(isrow, &bs));
3676: PetscCall(ISGetBlockSize(iscol, &cbs));
3678: PetscCall(MatCreate(comm, &M));
3679: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3680: PetscCall(MatSetBlockSizes(M, bs, cbs));
3681: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3682: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3683: PetscCall(PetscFree(dlens));
3685: } else { /* call == MAT_REUSE_MATRIX */
3686: M = *newmat;
3687: PetscCall(MatGetLocalSize(M, &i, NULL));
3688: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3689: PetscCall(MatZeroEntries(M));
3690: /*
3691: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3692: rather than the slower MatSetValues().
3693: */
3694: M->was_assembled = PETSC_TRUE;
3695: M->assembled = PETSC_FALSE;
3696: }
3698: /* (5) Set values of Msub to *newmat */
3699: PetscCall(PetscMalloc1(count, &colsub));
3700: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3702: jj = aij->j;
3703: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3704: for (i = 0; i < m; i++) {
3705: row = rstart + i;
3706: nz = ii[i + 1] - ii[i];
3707: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3708: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3709: jj += nz;
3710: aa += nz;
3711: }
3712: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3713: PetscCall(ISRestoreIndices(iscmap, &cmap));
3715: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3716: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3718: PetscCall(PetscFree(colsub));
3720: /* save Msub, iscol_sub and iscmap used in processor for next request */
3721: if (call == MAT_INITIAL_MATRIX) {
3722: *newmat = M;
3723: PetscCall(PetscObjectCompose((PetscObject)(*newmat), "SubMatrix", (PetscObject)Msub));
3724: PetscCall(MatDestroy(&Msub));
3726: PetscCall(PetscObjectCompose((PetscObject)(*newmat), "SubIScol", (PetscObject)iscol_sub));
3727: PetscCall(ISDestroy(&iscol_sub));
3729: PetscCall(PetscObjectCompose((PetscObject)(*newmat), "Subcmap", (PetscObject)iscmap));
3730: PetscCall(ISDestroy(&iscmap));
3732: if (iscol_local) {
3733: PetscCall(PetscObjectCompose((PetscObject)(*newmat), "ISAllGather", (PetscObject)iscol_local));
3734: PetscCall(ISDestroy(&iscol_local));
3735: }
3736: }
3737: PetscFunctionReturn(PETSC_SUCCESS);
3738: }
3740: /*
3741: Not great since it makes two copies of the submatrix, first an SeqAIJ
3742: in local and then by concatenating the local matrices the end result.
3743: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3745: This requires a sequential iscol with all indices.
3746: */
3747: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3748: {
3749: PetscMPIInt rank, size;
3750: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3751: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3752: Mat M, Mreuse;
3753: MatScalar *aa, *vwork;
3754: MPI_Comm comm;
3755: Mat_SeqAIJ *aij;
3756: PetscBool colflag, allcolumns = PETSC_FALSE;
3758: PetscFunctionBegin;
3759: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3760: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3761: PetscCallMPI(MPI_Comm_size(comm, &size));
3763: /* Check for special case: each processor gets entire matrix columns */
3764: PetscCall(ISIdentity(iscol, &colflag));
3765: PetscCall(ISGetLocalSize(iscol, &n));
3766: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3767: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3769: if (call == MAT_REUSE_MATRIX) {
3770: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3771: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3772: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3773: } else {
3774: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3775: }
3777: /*
3778: m - number of local rows
3779: n - number of columns (same on all processors)
3780: rstart - first row in new global matrix generated
3781: */
3782: PetscCall(MatGetSize(Mreuse, &m, &n));
3783: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3784: if (call == MAT_INITIAL_MATRIX) {
3785: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3786: ii = aij->i;
3787: jj = aij->j;
3789: /*
3790: Determine the number of non-zeros in the diagonal and off-diagonal
3791: portions of the matrix in order to do correct preallocation
3792: */
3794: /* first get start and end of "diagonal" columns */
3795: if (csize == PETSC_DECIDE) {
3796: PetscCall(ISGetSize(isrow, &mglobal));
3797: if (mglobal == n) { /* square matrix */
3798: nlocal = m;
3799: } else {
3800: nlocal = n / size + ((n % size) > rank);
3801: }
3802: } else {
3803: nlocal = csize;
3804: }
3805: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3806: rstart = rend - nlocal;
3807: 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);
3809: /* next, compute all the lengths */
3810: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3811: olens = dlens + m;
3812: for (i = 0; i < m; i++) {
3813: jend = ii[i + 1] - ii[i];
3814: olen = 0;
3815: dlen = 0;
3816: for (j = 0; j < jend; j++) {
3817: if (*jj < rstart || *jj >= rend) olen++;
3818: else dlen++;
3819: jj++;
3820: }
3821: olens[i] = olen;
3822: dlens[i] = dlen;
3823: }
3824: PetscCall(MatCreate(comm, &M));
3825: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3826: PetscCall(MatSetBlockSizes(M, bs, cbs));
3827: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3828: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3829: PetscCall(PetscFree(dlens));
3830: } else {
3831: PetscInt ml, nl;
3833: M = *newmat;
3834: PetscCall(MatGetLocalSize(M, &ml, &nl));
3835: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3836: PetscCall(MatZeroEntries(M));
3837: /*
3838: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3839: rather than the slower MatSetValues().
3840: */
3841: M->was_assembled = PETSC_TRUE;
3842: M->assembled = PETSC_FALSE;
3843: }
3844: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3845: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3846: ii = aij->i;
3847: jj = aij->j;
3849: /* trigger copy to CPU if needed */
3850: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3851: for (i = 0; i < m; i++) {
3852: row = rstart + i;
3853: nz = ii[i + 1] - ii[i];
3854: cwork = jj;
3855: jj += nz;
3856: vwork = aa;
3857: aa += nz;
3858: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3859: }
3860: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3862: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3863: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3864: *newmat = M;
3866: /* save submatrix used in processor for next request */
3867: if (call == MAT_INITIAL_MATRIX) {
3868: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3869: PetscCall(MatDestroy(&Mreuse));
3870: }
3871: PetscFunctionReturn(PETSC_SUCCESS);
3872: }
3874: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3875: {
3876: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3877: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii;
3878: const PetscInt *JJ;
3879: PetscBool nooffprocentries;
3880: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3882: PetscFunctionBegin;
3883: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
3885: PetscCall(PetscLayoutSetUp(B->rmap));
3886: PetscCall(PetscLayoutSetUp(B->cmap));
3887: m = B->rmap->n;
3888: cstart = B->cmap->rstart;
3889: cend = B->cmap->rend;
3890: rstart = B->rmap->rstart;
3892: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3894: if (PetscDefined(USE_DEBUG)) {
3895: for (i = 0; i < m; i++) {
3896: nnz = Ii[i + 1] - Ii[i];
3897: JJ = J ? J + Ii[i] : NULL;
3898: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3899: 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]);
3900: 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);
3901: }
3902: }
3904: for (i = 0; i < m; i++) {
3905: nnz = Ii[i + 1] - Ii[i];
3906: JJ = J ? J + Ii[i] : NULL;
3907: nnz_max = PetscMax(nnz_max, nnz);
3908: d = 0;
3909: for (j = 0; j < nnz; j++) {
3910: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3911: }
3912: d_nnz[i] = d;
3913: o_nnz[i] = nnz - d;
3914: }
3915: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3916: PetscCall(PetscFree2(d_nnz, o_nnz));
3918: for (i = 0; i < m; i++) {
3919: ii = i + rstart;
3920: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], J ? J + Ii[i] : NULL, v ? v + Ii[i] : NULL, INSERT_VALUES));
3921: }
3922: nooffprocentries = B->nooffprocentries;
3923: B->nooffprocentries = PETSC_TRUE;
3924: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3925: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3926: B->nooffprocentries = nooffprocentries;
3928: /* count number of entries below block diagonal */
3929: PetscCall(PetscFree(Aij->ld));
3930: PetscCall(PetscCalloc1(m, &ld));
3931: Aij->ld = ld;
3932: for (i = 0; i < m; i++) {
3933: nnz = Ii[i + 1] - Ii[i];
3934: j = 0;
3935: while (j < nnz && J[j] < cstart) j++;
3936: ld[i] = j;
3937: if (J) J += nnz;
3938: }
3940: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3941: PetscFunctionReturn(PETSC_SUCCESS);
3942: }
3944: /*@
3945: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3946: (the default parallel PETSc format).
3948: Collective
3950: Input Parameters:
3951: + B - the matrix
3952: . i - the indices into j for the start of each local row (starts with zero)
3953: . j - the column indices for each local row (starts with zero)
3954: - v - optional values in the matrix
3956: Level: developer
3958: Notes:
3959: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3960: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3961: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3963: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3965: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3967: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3969: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3970: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3972: The format which is used for the sparse matrix input, is equivalent to a
3973: row-major ordering.. i.e for the following matrix, the input data expected is
3974: as shown
3975: .vb
3976: 1 0 0
3977: 2 0 3 P0
3978: -------
3979: 4 5 6 P1
3981: Process0 [P0] rows_owned=[0,1]
3982: i = {0,1,3} [size = nrow+1 = 2+1]
3983: j = {0,0,2} [size = 3]
3984: v = {1,2,3} [size = 3]
3986: Process1 [P1] rows_owned=[2]
3987: i = {0,3} [size = nrow+1 = 1+1]
3988: j = {0,1,2} [size = 3]
3989: v = {4,5,6} [size = 3]
3990: .ve
3992: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
3993: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
3994: @*/
3995: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
3996: {
3997: PetscFunctionBegin;
3998: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
3999: PetscFunctionReturn(PETSC_SUCCESS);
4000: }
4002: /*@C
4003: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4004: (the default parallel PETSc format). For good matrix assembly performance
4005: the user should preallocate the matrix storage by setting the parameters
4006: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4008: Collective
4010: Input Parameters:
4011: + B - the matrix
4012: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4013: (same value is used for all local rows)
4014: . d_nnz - array containing the number of nonzeros in the various rows of the
4015: DIAGONAL portion of the local submatrix (possibly different for each row)
4016: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4017: The size of this array is equal to the number of local rows, i.e 'm'.
4018: For matrices that will be factored, you must leave room for (and set)
4019: the diagonal entry even if it is zero.
4020: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4021: submatrix (same value is used for all local rows).
4022: - o_nnz - array containing the number of nonzeros in the various rows of the
4023: OFF-DIAGONAL portion of the local submatrix (possibly different for
4024: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4025: structure. The size of this array is equal to the number
4026: of local rows, i.e 'm'.
4028: Example Usage:
4029: Consider the following 8x8 matrix with 34 non-zero values, that is
4030: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4031: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4032: as follows
4034: .vb
4035: 1 2 0 | 0 3 0 | 0 4
4036: Proc0 0 5 6 | 7 0 0 | 8 0
4037: 9 0 10 | 11 0 0 | 12 0
4038: -------------------------------------
4039: 13 0 14 | 15 16 17 | 0 0
4040: Proc1 0 18 0 | 19 20 21 | 0 0
4041: 0 0 0 | 22 23 0 | 24 0
4042: -------------------------------------
4043: Proc2 25 26 27 | 0 0 28 | 29 0
4044: 30 0 0 | 31 32 33 | 0 34
4045: .ve
4047: This can be represented as a collection of submatrices as
4048: .vb
4049: A B C
4050: D E F
4051: G H I
4052: .ve
4054: Where the submatrices A,B,C are owned by proc0, D,E,F are
4055: owned by proc1, G,H,I are owned by proc2.
4057: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4058: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4059: The 'M','N' parameters are 8,8, and have the same values on all procs.
4061: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4062: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4063: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4064: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4065: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4066: matrix, ans [DF] as another `MATSEQAIJ` matrix.
4068: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4069: allocated for every row of the local diagonal submatrix, and `o_nz`
4070: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4071: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4072: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4073: In this case, the values of `d_nz`, `o_nz` are
4074: .vb
4075: proc0 dnz = 2, o_nz = 2
4076: proc1 dnz = 3, o_nz = 2
4077: proc2 dnz = 1, o_nz = 4
4078: .ve
4079: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4080: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4081: for proc3. i.e we are using 12+15+10=37 storage locations to store
4082: 34 values.
4084: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4085: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4086: In the above case the values for `d_nnz`, `o_nnz` are
4087: .vb
4088: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4089: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4090: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4091: .ve
4092: Here the space allocated is sum of all the above values i.e 34, and
4093: hence pre-allocation is perfect.
4095: Level: intermediate
4097: Notes:
4098: If the *_nnz parameter is given then the *_nz parameter is ignored
4100: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4101: storage. The stored row and column indices begin with zero.
4102: See [Sparse Matrices](sec_matsparse) for details.
4104: The parallel matrix is partitioned such that the first m0 rows belong to
4105: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4106: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4108: The DIAGONAL portion of the local submatrix of a processor can be defined
4109: as the submatrix which is obtained by extraction the part corresponding to
4110: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4111: first row that belongs to the processor, r2 is the last row belonging to
4112: the this processor, and c1-c2 is range of indices of the local part of a
4113: vector suitable for applying the matrix to. This is an mxn matrix. In the
4114: common case of a square matrix, the row and column ranges are the same and
4115: the DIAGONAL part is also square. The remaining portion of the local
4116: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4118: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4120: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4121: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4122: You can also run with the option `-info` and look for messages with the string
4123: malloc in them to see if additional memory allocation was needed.
4125: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4126: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4127: @*/
4128: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4129: {
4130: PetscFunctionBegin;
4133: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4134: PetscFunctionReturn(PETSC_SUCCESS);
4135: }
4137: /*@
4138: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4139: CSR format for the local rows.
4141: Collective
4143: Input Parameters:
4144: + comm - MPI communicator
4145: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4146: . n - This value should be the same as the local size used in creating the
4147: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4148: calculated if N is given) For square matrices n is almost always m.
4149: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4150: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4151: . 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
4152: . j - global column indices
4153: - a - optional matrix values
4155: Output Parameter:
4156: . mat - the matrix
4158: Level: intermediate
4160: Notes:
4161: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4162: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4163: called this routine. Use `MatCreateMPIAIJWithSplitArray()` to avoid needing to copy the arrays.
4165: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4167: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4169: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4170: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4172: The format which is used for the sparse matrix input, is equivalent to a
4173: row-major ordering.. i.e for the following matrix, the input data expected is
4174: as shown
4175: .vb
4176: 1 0 0
4177: 2 0 3 P0
4178: -------
4179: 4 5 6 P1
4181: Process0 [P0] rows_owned=[0,1]
4182: i = {0,1,3} [size = nrow+1 = 2+1]
4183: j = {0,0,2} [size = 3]
4184: v = {1,2,3} [size = 3]
4186: Process1 [P1] rows_owned=[2]
4187: i = {0,3} [size = nrow+1 = 1+1]
4188: j = {0,1,2} [size = 3]
4189: v = {4,5,6} [size = 3]
4190: .ve
4192: .seealso: [](ch_matrices), `Mat`, `MATMPIAIK`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4193: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4194: @*/
4195: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4196: {
4197: PetscFunctionBegin;
4198: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4199: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4200: PetscCall(MatCreate(comm, mat));
4201: PetscCall(MatSetSizes(*mat, m, n, M, N));
4202: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4203: PetscCall(MatSetType(*mat, MATMPIAIJ));
4204: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4205: PetscFunctionReturn(PETSC_SUCCESS);
4206: }
4208: /*@
4209: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4210: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4211: from `MatCreateMPIAIJWithArrays()`
4213: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4215: Collective
4217: Input Parameters:
4218: + mat - the matrix
4219: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4220: . n - This value should be the same as the local size used in creating the
4221: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4222: calculated if N is given) For square matrices n is almost always m.
4223: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4224: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4225: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4226: . J - column indices
4227: - v - matrix values
4229: Level: deprecated
4231: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4232: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4233: @*/
4234: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4235: {
4236: PetscInt nnz, i;
4237: PetscBool nooffprocentries;
4238: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4239: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4240: PetscScalar *ad, *ao;
4241: PetscInt ldi, Iii, md;
4242: const PetscInt *Adi = Ad->i;
4243: PetscInt *ld = Aij->ld;
4245: PetscFunctionBegin;
4246: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4247: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4248: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4249: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4251: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4252: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4254: for (i = 0; i < m; i++) {
4255: if (PetscDefined(USE_DEBUG)) {
4256: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4257: 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);
4258: 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);
4259: }
4260: }
4261: nnz = Ii[i + 1] - Ii[i];
4262: Iii = Ii[i];
4263: ldi = ld[i];
4264: md = Adi[i + 1] - Adi[i];
4265: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4266: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4267: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4268: ad += md;
4269: ao += nnz - md;
4270: }
4271: nooffprocentries = mat->nooffprocentries;
4272: mat->nooffprocentries = PETSC_TRUE;
4273: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4274: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4275: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4276: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4277: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4278: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4279: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4280: mat->nooffprocentries = nooffprocentries;
4281: PetscFunctionReturn(PETSC_SUCCESS);
4282: }
4284: /*@
4285: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4287: Collective
4289: Input Parameters:
4290: + mat - the matrix
4291: - v - matrix values, stored by row
4293: Level: intermediate
4295: Notes:
4296: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4298: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4300: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4301: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4302: @*/
4303: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4304: {
4305: PetscInt nnz, i, m;
4306: PetscBool nooffprocentries;
4307: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4308: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4309: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4310: PetscScalar *ad, *ao;
4311: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4312: PetscInt ldi, Iii, md;
4313: PetscInt *ld = Aij->ld;
4315: PetscFunctionBegin;
4316: m = mat->rmap->n;
4318: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4319: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4320: Iii = 0;
4321: for (i = 0; i < m; i++) {
4322: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4323: ldi = ld[i];
4324: md = Adi[i + 1] - Adi[i];
4325: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4326: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4327: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4328: ad += md;
4329: ao += nnz - md;
4330: Iii += nnz;
4331: }
4332: nooffprocentries = mat->nooffprocentries;
4333: mat->nooffprocentries = PETSC_TRUE;
4334: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4335: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4336: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4337: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4338: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4339: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4340: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4341: mat->nooffprocentries = nooffprocentries;
4342: PetscFunctionReturn(PETSC_SUCCESS);
4343: }
4345: /*@C
4346: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4347: (the default parallel PETSc format). For good matrix assembly performance
4348: the user should preallocate the matrix storage by setting the parameters
4349: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4351: Collective
4353: Input Parameters:
4354: + comm - MPI communicator
4355: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4356: This value should be the same as the local size used in creating the
4357: y vector for the matrix-vector product y = Ax.
4358: . n - This value should be the same as the local size used in creating the
4359: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4360: calculated if N is given) For square matrices n is almost always m.
4361: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4362: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4363: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4364: (same value is used for all local rows)
4365: . d_nnz - array containing the number of nonzeros in the various rows of the
4366: DIAGONAL portion of the local submatrix (possibly different for each row)
4367: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4368: The size of this array is equal to the number of local rows, i.e 'm'.
4369: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4370: submatrix (same value is used for all local rows).
4371: - o_nnz - array containing the number of nonzeros in the various rows of the
4372: OFF-DIAGONAL portion of the local submatrix (possibly different for
4373: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4374: structure. The size of this array is equal to the number
4375: of local rows, i.e 'm'.
4377: Output Parameter:
4378: . A - the matrix
4380: Options Database Keys:
4381: + -mat_no_inode - Do not use inodes
4382: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4383: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4384: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the vecscatter be viewed as a matrix.
4385: Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.
4387: Level: intermediate
4389: Notes:
4390: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4391: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4392: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4394: If the *_nnz parameter is given then the *_nz parameter is ignored
4396: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4397: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4398: storage requirements for this matrix.
4400: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4401: processor than it must be used on all processors that share the object for
4402: that argument.
4404: The user MUST specify either the local or global matrix dimensions
4405: (possibly both).
4407: The parallel matrix is partitioned across processors such that the
4408: first m0 rows belong to process 0, the next m1 rows belong to
4409: process 1, the next m2 rows belong to process 2 etc.. where
4410: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4411: values corresponding to [m x N] submatrix.
4413: The columns are logically partitioned with the n0 columns belonging
4414: to 0th partition, the next n1 columns belonging to the next
4415: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4417: The DIAGONAL portion of the local submatrix on any given processor
4418: is the submatrix corresponding to the rows and columns m,n
4419: corresponding to the given processor. i.e diagonal matrix on
4420: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4421: etc. The remaining portion of the local submatrix [m x (N-n)]
4422: constitute the OFF-DIAGONAL portion. The example below better
4423: illustrates this concept.
4425: For a square global matrix we define each processor's diagonal portion
4426: to be its local rows and the corresponding columns (a square submatrix);
4427: each processor's off-diagonal portion encompasses the remainder of the
4428: local matrix (a rectangular submatrix).
4430: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4432: When calling this routine with a single process communicator, a matrix of
4433: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4434: type of communicator, use the construction mechanism
4435: .vb
4436: MatCreate(..., &A);
4437: MatSetType(A, MATMPIAIJ);
4438: MatSetSizes(A, m, n, M, N);
4439: MatMPIAIJSetPreallocation(A, ...);
4440: .ve
4442: By default, this format uses inodes (identical nodes) when possible.
4443: We search for consecutive rows with the same nonzero structure, thereby
4444: reusing matrix information to achieve increased efficiency.
4446: Example Usage:
4447: Consider the following 8x8 matrix with 34 non-zero values, that is
4448: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4449: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4450: as follows
4452: .vb
4453: 1 2 0 | 0 3 0 | 0 4
4454: Proc0 0 5 6 | 7 0 0 | 8 0
4455: 9 0 10 | 11 0 0 | 12 0
4456: -------------------------------------
4457: 13 0 14 | 15 16 17 | 0 0
4458: Proc1 0 18 0 | 19 20 21 | 0 0
4459: 0 0 0 | 22 23 0 | 24 0
4460: -------------------------------------
4461: Proc2 25 26 27 | 0 0 28 | 29 0
4462: 30 0 0 | 31 32 33 | 0 34
4463: .ve
4465: This can be represented as a collection of submatrices as
4467: .vb
4468: A B C
4469: D E F
4470: G H I
4471: .ve
4473: Where the submatrices A,B,C are owned by proc0, D,E,F are
4474: owned by proc1, G,H,I are owned by proc2.
4476: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4477: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4478: The 'M','N' parameters are 8,8, and have the same values on all procs.
4480: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4481: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4482: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4483: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4484: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4485: matrix, ans [DF] as another SeqAIJ matrix.
4487: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4488: allocated for every row of the local diagonal submatrix, and `o_nz`
4489: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4490: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4491: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4492: In this case, the values of `d_nz`,`o_nz` are
4493: .vb
4494: proc0 dnz = 2, o_nz = 2
4495: proc1 dnz = 3, o_nz = 2
4496: proc2 dnz = 1, o_nz = 4
4497: .ve
4498: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4499: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4500: for proc3. i.e we are using 12+15+10=37 storage locations to store
4501: 34 values.
4503: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4504: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4505: In the above case the values for d_nnz,o_nnz are
4506: .vb
4507: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4508: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4509: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4510: .ve
4511: Here the space allocated is sum of all the above values i.e 34, and
4512: hence pre-allocation is perfect.
4514: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4515: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`
4516: @*/
4517: 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)
4518: {
4519: PetscMPIInt size;
4521: PetscFunctionBegin;
4522: PetscCall(MatCreate(comm, A));
4523: PetscCall(MatSetSizes(*A, m, n, M, N));
4524: PetscCallMPI(MPI_Comm_size(comm, &size));
4525: if (size > 1) {
4526: PetscCall(MatSetType(*A, MATMPIAIJ));
4527: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4528: } else {
4529: PetscCall(MatSetType(*A, MATSEQAIJ));
4530: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4531: }
4532: PetscFunctionReturn(PETSC_SUCCESS);
4533: }
4535: /*MC
4536: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4538: Synopsis:
4539: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4541: Not Collective
4543: Input Parameter:
4544: . A - the `MATMPIAIJ` matrix
4546: Output Parameters:
4547: + Ad - the diagonal portion of the matrix
4548: . Ao - the off-diagonal portion of the matrix
4549: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4550: - ierr - error code
4552: Level: advanced
4554: Note:
4555: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4557: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4558: M*/
4560: /*MC
4561: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4563: Synopsis:
4564: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4566: Not Collective
4568: Input Parameters:
4569: + A - the `MATMPIAIJ` matrix
4570: . Ad - the diagonal portion of the matrix
4571: . Ao - the off-diagonal portion of the matrix
4572: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4573: - ierr - error code
4575: Level: advanced
4577: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4578: M*/
4580: /*@C
4581: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4583: Not Collective
4585: Input Parameter:
4586: . A - The `MATMPIAIJ` matrix
4588: Output Parameters:
4589: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4590: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4591: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4593: Level: intermediate
4595: Note:
4596: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4597: 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
4598: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4599: local column numbers to global column numbers in the original matrix.
4601: Fortran Notes:
4602: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4604: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4605: @*/
4606: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4607: {
4608: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4609: PetscBool flg;
4611: PetscFunctionBegin;
4612: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4613: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4614: if (Ad) *Ad = a->A;
4615: if (Ao) *Ao = a->B;
4616: if (colmap) *colmap = a->garray;
4617: PetscFunctionReturn(PETSC_SUCCESS);
4618: }
4620: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4621: {
4622: PetscInt m, N, i, rstart, nnz, Ii;
4623: PetscInt *indx;
4624: PetscScalar *values;
4625: MatType rootType;
4627: PetscFunctionBegin;
4628: PetscCall(MatGetSize(inmat, &m, &N));
4629: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4630: PetscInt *dnz, *onz, sum, bs, cbs;
4632: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4633: /* Check sum(n) = N */
4634: PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4635: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4637: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4638: rstart -= m;
4640: MatPreallocateBegin(comm, m, n, dnz, onz);
4641: for (i = 0; i < m; i++) {
4642: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4643: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4644: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4645: }
4647: PetscCall(MatCreate(comm, outmat));
4648: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4649: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4650: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4651: PetscCall(MatGetRootType_Private(inmat, &rootType));
4652: PetscCall(MatSetType(*outmat, rootType));
4653: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4654: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4655: MatPreallocateEnd(dnz, onz);
4656: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4657: }
4659: /* numeric phase */
4660: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4661: for (i = 0; i < m; i++) {
4662: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4663: Ii = i + rstart;
4664: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4665: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4666: }
4667: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4668: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4669: PetscFunctionReturn(PETSC_SUCCESS);
4670: }
4672: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4673: {
4674: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4676: PetscFunctionBegin;
4677: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4678: PetscCall(PetscFree(merge->id_r));
4679: PetscCall(PetscFree(merge->len_s));
4680: PetscCall(PetscFree(merge->len_r));
4681: PetscCall(PetscFree(merge->bi));
4682: PetscCall(PetscFree(merge->bj));
4683: PetscCall(PetscFree(merge->buf_ri[0]));
4684: PetscCall(PetscFree(merge->buf_ri));
4685: PetscCall(PetscFree(merge->buf_rj[0]));
4686: PetscCall(PetscFree(merge->buf_rj));
4687: PetscCall(PetscFree(merge->coi));
4688: PetscCall(PetscFree(merge->coj));
4689: PetscCall(PetscFree(merge->owners_co));
4690: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4691: PetscCall(PetscFree(merge));
4692: PetscFunctionReturn(PETSC_SUCCESS);
4693: }
4695: #include <../src/mat/utils/freespace.h>
4696: #include <petscbt.h>
4698: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4699: {
4700: MPI_Comm comm;
4701: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4702: PetscMPIInt size, rank, taga, *len_s;
4703: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4704: PetscInt proc, m;
4705: PetscInt **buf_ri, **buf_rj;
4706: PetscInt k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4707: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4708: MPI_Request *s_waits, *r_waits;
4709: MPI_Status *status;
4710: const MatScalar *aa, *a_a;
4711: MatScalar **abuf_r, *ba_i;
4712: Mat_Merge_SeqsToMPI *merge;
4713: PetscContainer container;
4715: PetscFunctionBegin;
4716: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4717: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4719: PetscCallMPI(MPI_Comm_size(comm, &size));
4720: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4722: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4723: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4724: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4725: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4726: aa = a_a;
4728: bi = merge->bi;
4729: bj = merge->bj;
4730: buf_ri = merge->buf_ri;
4731: buf_rj = merge->buf_rj;
4733: PetscCall(PetscMalloc1(size, &status));
4734: owners = merge->rowmap->range;
4735: len_s = merge->len_s;
4737: /* send and recv matrix values */
4738: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4739: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4741: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4742: for (proc = 0, k = 0; proc < size; proc++) {
4743: if (!len_s[proc]) continue;
4744: i = owners[proc];
4745: PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4746: k++;
4747: }
4749: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4750: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4751: PetscCall(PetscFree(status));
4753: PetscCall(PetscFree(s_waits));
4754: PetscCall(PetscFree(r_waits));
4756: /* insert mat values of mpimat */
4757: PetscCall(PetscMalloc1(N, &ba_i));
4758: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4760: for (k = 0; k < merge->nrecv; k++) {
4761: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4762: nrows = *(buf_ri_k[k]);
4763: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4764: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4765: }
4767: /* set values of ba */
4768: m = merge->rowmap->n;
4769: for (i = 0; i < m; i++) {
4770: arow = owners[rank] + i;
4771: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4772: bnzi = bi[i + 1] - bi[i];
4773: PetscCall(PetscArrayzero(ba_i, bnzi));
4775: /* add local non-zero vals of this proc's seqmat into ba */
4776: anzi = ai[arow + 1] - ai[arow];
4777: aj = a->j + ai[arow];
4778: aa = a_a + ai[arow];
4779: nextaj = 0;
4780: for (j = 0; nextaj < anzi; j++) {
4781: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4782: ba_i[j] += aa[nextaj++];
4783: }
4784: }
4786: /* add received vals into ba */
4787: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4788: /* i-th row */
4789: if (i == *nextrow[k]) {
4790: anzi = *(nextai[k] + 1) - *nextai[k];
4791: aj = buf_rj[k] + *(nextai[k]);
4792: aa = abuf_r[k] + *(nextai[k]);
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: }
4799: nextrow[k]++;
4800: nextai[k]++;
4801: }
4802: }
4803: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4804: }
4805: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4806: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4807: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4809: PetscCall(PetscFree(abuf_r[0]));
4810: PetscCall(PetscFree(abuf_r));
4811: PetscCall(PetscFree(ba_i));
4812: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4813: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4814: PetscFunctionReturn(PETSC_SUCCESS);
4815: }
4817: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4818: {
4819: Mat B_mpi;
4820: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4821: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4822: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4823: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4824: PetscInt len, proc, *dnz, *onz, bs, cbs;
4825: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4826: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4827: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4828: MPI_Status *status;
4829: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4830: PetscBT lnkbt;
4831: Mat_Merge_SeqsToMPI *merge;
4832: PetscContainer container;
4834: PetscFunctionBegin;
4835: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4837: /* make sure it is a PETSc comm */
4838: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4839: PetscCallMPI(MPI_Comm_size(comm, &size));
4840: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4842: PetscCall(PetscNew(&merge));
4843: PetscCall(PetscMalloc1(size, &status));
4845: /* determine row ownership */
4846: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4847: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4848: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4849: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4850: PetscCall(PetscLayoutSetUp(merge->rowmap));
4851: PetscCall(PetscMalloc1(size, &len_si));
4852: PetscCall(PetscMalloc1(size, &merge->len_s));
4854: m = merge->rowmap->n;
4855: owners = merge->rowmap->range;
4857: /* determine the number of messages to send, their lengths */
4858: len_s = merge->len_s;
4860: len = 0; /* length of buf_si[] */
4861: merge->nsend = 0;
4862: for (proc = 0; proc < size; proc++) {
4863: len_si[proc] = 0;
4864: if (proc == rank) {
4865: len_s[proc] = 0;
4866: } else {
4867: len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4868: len_s[proc] = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4869: }
4870: if (len_s[proc]) {
4871: merge->nsend++;
4872: nrows = 0;
4873: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4874: if (ai[i + 1] > ai[i]) nrows++;
4875: }
4876: len_si[proc] = 2 * (nrows + 1);
4877: len += len_si[proc];
4878: }
4879: }
4881: /* determine the number and length of messages to receive for ij-structure */
4882: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4883: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4885: /* post the Irecv of j-structure */
4886: PetscCall(PetscCommGetNewTag(comm, &tagj));
4887: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4889: /* post the Isend of j-structure */
4890: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4892: for (proc = 0, k = 0; proc < size; proc++) {
4893: if (!len_s[proc]) continue;
4894: i = owners[proc];
4895: PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4896: k++;
4897: }
4899: /* receives and sends of j-structure are complete */
4900: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4901: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4903: /* send and recv i-structure */
4904: PetscCall(PetscCommGetNewTag(comm, &tagi));
4905: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4907: PetscCall(PetscMalloc1(len + 1, &buf_s));
4908: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4909: for (proc = 0, k = 0; proc < size; proc++) {
4910: if (!len_s[proc]) continue;
4911: /* form outgoing message for i-structure:
4912: buf_si[0]: nrows to be sent
4913: [1:nrows]: row index (global)
4914: [nrows+1:2*nrows+1]: i-structure index
4915: */
4916: nrows = len_si[proc] / 2 - 1;
4917: buf_si_i = buf_si + nrows + 1;
4918: buf_si[0] = nrows;
4919: buf_si_i[0] = 0;
4920: nrows = 0;
4921: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4922: anzi = ai[i + 1] - ai[i];
4923: if (anzi) {
4924: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4925: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4926: nrows++;
4927: }
4928: }
4929: PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4930: k++;
4931: buf_si += len_si[proc];
4932: }
4934: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4935: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4937: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4938: 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]));
4940: PetscCall(PetscFree(len_si));
4941: PetscCall(PetscFree(len_ri));
4942: PetscCall(PetscFree(rj_waits));
4943: PetscCall(PetscFree2(si_waits, sj_waits));
4944: PetscCall(PetscFree(ri_waits));
4945: PetscCall(PetscFree(buf_s));
4946: PetscCall(PetscFree(status));
4948: /* compute a local seq matrix in each processor */
4949: /* allocate bi array and free space for accumulating nonzero column info */
4950: PetscCall(PetscMalloc1(m + 1, &bi));
4951: bi[0] = 0;
4953: /* create and initialize a linked list */
4954: nlnk = N + 1;
4955: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4957: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4958: len = ai[owners[rank + 1]] - ai[owners[rank]];
4959: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4961: current_space = free_space;
4963: /* determine symbolic info for each local row */
4964: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4966: for (k = 0; k < merge->nrecv; k++) {
4967: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4968: nrows = *buf_ri_k[k];
4969: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4970: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4971: }
4973: MatPreallocateBegin(comm, m, n, dnz, onz);
4974: len = 0;
4975: for (i = 0; i < m; i++) {
4976: bnzi = 0;
4977: /* add local non-zero cols of this proc's seqmat into lnk */
4978: arow = owners[rank] + i;
4979: anzi = ai[arow + 1] - ai[arow];
4980: aj = a->j + ai[arow];
4981: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4982: bnzi += nlnk;
4983: /* add received col data into lnk */
4984: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4985: if (i == *nextrow[k]) { /* i-th row */
4986: anzi = *(nextai[k] + 1) - *nextai[k];
4987: aj = buf_rj[k] + *nextai[k];
4988: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4989: bnzi += nlnk;
4990: nextrow[k]++;
4991: nextai[k]++;
4992: }
4993: }
4994: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4996: /* if free space is not available, make more free space */
4997: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
4998: /* copy data into free space, then initialize lnk */
4999: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5000: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5002: current_space->array += bnzi;
5003: current_space->local_used += bnzi;
5004: current_space->local_remaining -= bnzi;
5006: bi[i + 1] = bi[i] + bnzi;
5007: }
5009: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5011: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5012: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5013: PetscCall(PetscLLDestroy(lnk, lnkbt));
5015: /* create symbolic parallel matrix B_mpi */
5016: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5017: PetscCall(MatCreate(comm, &B_mpi));
5018: if (n == PETSC_DECIDE) {
5019: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5020: } else {
5021: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5022: }
5023: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5024: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5025: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5026: MatPreallocateEnd(dnz, onz);
5027: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5029: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5030: B_mpi->assembled = PETSC_FALSE;
5031: merge->bi = bi;
5032: merge->bj = bj;
5033: merge->buf_ri = buf_ri;
5034: merge->buf_rj = buf_rj;
5035: merge->coi = NULL;
5036: merge->coj = NULL;
5037: merge->owners_co = NULL;
5039: PetscCall(PetscCommDestroy(&comm));
5041: /* attach the supporting struct to B_mpi for reuse */
5042: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5043: PetscCall(PetscContainerSetPointer(container, merge));
5044: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5045: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5046: PetscCall(PetscContainerDestroy(&container));
5047: *mpimat = B_mpi;
5049: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5050: PetscFunctionReturn(PETSC_SUCCESS);
5051: }
5053: /*@C
5054: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5055: matrices from each processor
5057: Collective
5059: Input Parameters:
5060: + comm - the communicators the parallel matrix will live on
5061: . seqmat - the input sequential matrices
5062: . m - number of local rows (or `PETSC_DECIDE`)
5063: . n - number of local columns (or `PETSC_DECIDE`)
5064: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5066: Output Parameter:
5067: . mpimat - the parallel matrix generated
5069: Level: advanced
5071: Note:
5072: The dimensions of the sequential matrix in each processor MUST be the same.
5073: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5074: destroyed when mpimat is destroyed. Call `PetscObjectQuery()` to access seqmat.
5076: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5077: @*/
5078: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5079: {
5080: PetscMPIInt size;
5082: PetscFunctionBegin;
5083: PetscCallMPI(MPI_Comm_size(comm, &size));
5084: if (size == 1) {
5085: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5086: if (scall == MAT_INITIAL_MATRIX) {
5087: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5088: } else {
5089: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5090: }
5091: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5092: PetscFunctionReturn(PETSC_SUCCESS);
5093: }
5094: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5095: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5096: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5097: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5098: PetscFunctionReturn(PETSC_SUCCESS);
5099: }
5101: /*@
5102: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5104: Not Collective
5106: Input Parameter:
5107: . A - the matrix
5109: Output Parameter:
5110: . A_loc - the local sequential matrix generated
5112: Level: developer
5114: Notes:
5115: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5116: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5117: `n` is the global column count obtained with `MatGetSize()`
5119: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5121: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5123: Destroy the matrix with `MatDestroy()`
5125: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5126: @*/
5127: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5128: {
5129: PetscBool mpi;
5131: PetscFunctionBegin;
5132: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5133: if (mpi) {
5134: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5135: } else {
5136: *A_loc = A;
5137: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5138: }
5139: PetscFunctionReturn(PETSC_SUCCESS);
5140: }
5142: /*@
5143: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5145: Not Collective
5147: Input Parameters:
5148: + A - the matrix
5149: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5151: Output Parameter:
5152: . A_loc - the local sequential matrix generated
5154: Level: developer
5156: Notes:
5157: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5158: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5159: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5161: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5163: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5164: 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
5165: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5166: 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.
5168: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5169: @*/
5170: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5171: {
5172: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5173: Mat_SeqAIJ *mat, *a, *b;
5174: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5175: const PetscScalar *aa, *ba, *aav, *bav;
5176: PetscScalar *ca, *cam;
5177: PetscMPIInt size;
5178: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5179: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5180: PetscBool match;
5182: PetscFunctionBegin;
5183: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5184: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5185: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5186: if (size == 1) {
5187: if (scall == MAT_INITIAL_MATRIX) {
5188: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5189: *A_loc = mpimat->A;
5190: } else if (scall == MAT_REUSE_MATRIX) {
5191: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5192: }
5193: PetscFunctionReturn(PETSC_SUCCESS);
5194: }
5196: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5197: a = (Mat_SeqAIJ *)(mpimat->A)->data;
5198: b = (Mat_SeqAIJ *)(mpimat->B)->data;
5199: ai = a->i;
5200: aj = a->j;
5201: bi = b->i;
5202: bj = b->j;
5203: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5204: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5205: aa = aav;
5206: ba = bav;
5207: if (scall == MAT_INITIAL_MATRIX) {
5208: PetscCall(PetscMalloc1(1 + am, &ci));
5209: ci[0] = 0;
5210: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5211: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5212: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5213: k = 0;
5214: for (i = 0; i < am; i++) {
5215: ncols_o = bi[i + 1] - bi[i];
5216: ncols_d = ai[i + 1] - ai[i];
5217: /* off-diagonal portion of A */
5218: for (jo = 0; jo < ncols_o; jo++) {
5219: col = cmap[*bj];
5220: if (col >= cstart) break;
5221: cj[k] = col;
5222: bj++;
5223: ca[k++] = *ba++;
5224: }
5225: /* diagonal portion of A */
5226: for (j = 0; j < ncols_d; j++) {
5227: cj[k] = cstart + *aj++;
5228: ca[k++] = *aa++;
5229: }
5230: /* off-diagonal portion of A */
5231: for (j = jo; j < ncols_o; j++) {
5232: cj[k] = cmap[*bj++];
5233: ca[k++] = *ba++;
5234: }
5235: }
5236: /* put together the new matrix */
5237: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5238: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5239: /* Since these are PETSc arrays, change flags to free them as necessary. */
5240: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5241: mat->free_a = PETSC_TRUE;
5242: mat->free_ij = PETSC_TRUE;
5243: mat->nonew = 0;
5244: } else if (scall == MAT_REUSE_MATRIX) {
5245: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5246: ci = mat->i;
5247: cj = mat->j;
5248: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5249: for (i = 0; i < am; i++) {
5250: /* off-diagonal portion of A */
5251: ncols_o = bi[i + 1] - bi[i];
5252: for (jo = 0; jo < ncols_o; jo++) {
5253: col = cmap[*bj];
5254: if (col >= cstart) break;
5255: *cam++ = *ba++;
5256: bj++;
5257: }
5258: /* diagonal portion of A */
5259: ncols_d = ai[i + 1] - ai[i];
5260: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5261: /* off-diagonal portion of A */
5262: for (j = jo; j < ncols_o; j++) {
5263: *cam++ = *ba++;
5264: bj++;
5265: }
5266: }
5267: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5268: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5269: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5270: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5271: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5272: PetscFunctionReturn(PETSC_SUCCESS);
5273: }
5275: /*@
5276: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5277: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5279: Not Collective
5281: Input Parameters:
5282: + A - the matrix
5283: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5285: Output Parameters:
5286: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5287: - A_loc - the local sequential matrix generated
5289: Level: developer
5291: Note:
5292: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5293: part, then those associated with the off-diagonal part (in its local ordering)
5295: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5296: @*/
5297: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5298: {
5299: Mat Ao, Ad;
5300: const PetscInt *cmap;
5301: PetscMPIInt size;
5302: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5304: PetscFunctionBegin;
5305: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5306: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5307: if (size == 1) {
5308: if (scall == MAT_INITIAL_MATRIX) {
5309: PetscCall(PetscObjectReference((PetscObject)Ad));
5310: *A_loc = Ad;
5311: } else if (scall == MAT_REUSE_MATRIX) {
5312: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5313: }
5314: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5315: PetscFunctionReturn(PETSC_SUCCESS);
5316: }
5317: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5318: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5319: if (f) {
5320: PetscCall((*f)(A, scall, glob, A_loc));
5321: } else {
5322: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5323: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5324: Mat_SeqAIJ *c;
5325: PetscInt *ai = a->i, *aj = a->j;
5326: PetscInt *bi = b->i, *bj = b->j;
5327: PetscInt *ci, *cj;
5328: const PetscScalar *aa, *ba;
5329: PetscScalar *ca;
5330: PetscInt i, j, am, dn, on;
5332: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5333: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5334: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5335: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5336: if (scall == MAT_INITIAL_MATRIX) {
5337: PetscInt k;
5338: PetscCall(PetscMalloc1(1 + am, &ci));
5339: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5340: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5341: ci[0] = 0;
5342: for (i = 0, k = 0; i < am; i++) {
5343: const PetscInt ncols_o = bi[i + 1] - bi[i];
5344: const PetscInt ncols_d = ai[i + 1] - ai[i];
5345: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5346: /* diagonal portion of A */
5347: for (j = 0; j < ncols_d; j++, k++) {
5348: cj[k] = *aj++;
5349: ca[k] = *aa++;
5350: }
5351: /* off-diagonal portion of A */
5352: for (j = 0; j < ncols_o; j++, k++) {
5353: cj[k] = dn + *bj++;
5354: ca[k] = *ba++;
5355: }
5356: }
5357: /* put together the new matrix */
5358: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5359: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5360: /* Since these are PETSc arrays, change flags to free them as necessary. */
5361: c = (Mat_SeqAIJ *)(*A_loc)->data;
5362: c->free_a = PETSC_TRUE;
5363: c->free_ij = PETSC_TRUE;
5364: c->nonew = 0;
5365: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5366: } else if (scall == MAT_REUSE_MATRIX) {
5367: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5368: for (i = 0; i < am; i++) {
5369: const PetscInt ncols_d = ai[i + 1] - ai[i];
5370: const PetscInt ncols_o = bi[i + 1] - bi[i];
5371: /* diagonal portion of A */
5372: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5373: /* off-diagonal portion of A */
5374: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5375: }
5376: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5377: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5378: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5379: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5380: if (glob) {
5381: PetscInt cst, *gidx;
5383: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5384: PetscCall(PetscMalloc1(dn + on, &gidx));
5385: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5386: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5387: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5388: }
5389: }
5390: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5391: PetscFunctionReturn(PETSC_SUCCESS);
5392: }
5394: /*@C
5395: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5397: Not Collective
5399: Input Parameters:
5400: + A - the matrix
5401: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5402: . row - index set of rows to extract (or `NULL`)
5403: - col - index set of columns to extract (or `NULL`)
5405: Output Parameter:
5406: . A_loc - the local sequential matrix generated
5408: Level: developer
5410: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5411: @*/
5412: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5413: {
5414: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5415: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5416: IS isrowa, iscola;
5417: Mat *aloc;
5418: PetscBool match;
5420: PetscFunctionBegin;
5421: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5422: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5423: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5424: if (!row) {
5425: start = A->rmap->rstart;
5426: end = A->rmap->rend;
5427: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5428: } else {
5429: isrowa = *row;
5430: }
5431: if (!col) {
5432: start = A->cmap->rstart;
5433: cmap = a->garray;
5434: nzA = a->A->cmap->n;
5435: nzB = a->B->cmap->n;
5436: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5437: ncols = 0;
5438: for (i = 0; i < nzB; i++) {
5439: if (cmap[i] < start) idx[ncols++] = cmap[i];
5440: else break;
5441: }
5442: imark = i;
5443: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5444: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5445: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5446: } else {
5447: iscola = *col;
5448: }
5449: if (scall != MAT_INITIAL_MATRIX) {
5450: PetscCall(PetscMalloc1(1, &aloc));
5451: aloc[0] = *A_loc;
5452: }
5453: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5454: if (!col) { /* attach global id of condensed columns */
5455: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5456: }
5457: *A_loc = aloc[0];
5458: PetscCall(PetscFree(aloc));
5459: if (!row) PetscCall(ISDestroy(&isrowa));
5460: if (!col) PetscCall(ISDestroy(&iscola));
5461: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5462: PetscFunctionReturn(PETSC_SUCCESS);
5463: }
5465: /*
5466: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5467: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5468: * on a global size.
5469: * */
5470: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5471: {
5472: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5473: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)(p->A)->data, *po = (Mat_SeqAIJ *)(p->B)->data, *p_oth;
5474: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5475: PetscMPIInt owner;
5476: PetscSFNode *iremote, *oiremote;
5477: const PetscInt *lrowindices;
5478: PetscSF sf, osf;
5479: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5480: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5481: MPI_Comm comm;
5482: ISLocalToGlobalMapping mapping;
5483: const PetscScalar *pd_a, *po_a;
5485: PetscFunctionBegin;
5486: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5487: /* plocalsize is the number of roots
5488: * nrows is the number of leaves
5489: * */
5490: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5491: PetscCall(ISGetLocalSize(rows, &nrows));
5492: PetscCall(PetscCalloc1(nrows, &iremote));
5493: PetscCall(ISGetIndices(rows, &lrowindices));
5494: for (i = 0; i < nrows; i++) {
5495: /* Find a remote index and an owner for a row
5496: * The row could be local or remote
5497: * */
5498: owner = 0;
5499: lidx = 0;
5500: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5501: iremote[i].index = lidx;
5502: iremote[i].rank = owner;
5503: }
5504: /* Create SF to communicate how many nonzero columns for each row */
5505: PetscCall(PetscSFCreate(comm, &sf));
5506: /* SF will figure out the number of nonzero columns for each row, and their
5507: * offsets
5508: * */
5509: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5510: PetscCall(PetscSFSetFromOptions(sf));
5511: PetscCall(PetscSFSetUp(sf));
5513: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5514: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5515: PetscCall(PetscCalloc1(nrows, &pnnz));
5516: roffsets[0] = 0;
5517: roffsets[1] = 0;
5518: for (i = 0; i < plocalsize; i++) {
5519: /* diagonal */
5520: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5521: /* off-diagonal */
5522: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5523: /* compute offsets so that we relative location for each row */
5524: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5525: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5526: }
5527: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5528: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5529: /* 'r' means root, and 'l' means leaf */
5530: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5531: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5532: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5533: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5534: PetscCall(PetscSFDestroy(&sf));
5535: PetscCall(PetscFree(roffsets));
5536: PetscCall(PetscFree(nrcols));
5537: dntotalcols = 0;
5538: ontotalcols = 0;
5539: ncol = 0;
5540: for (i = 0; i < nrows; i++) {
5541: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5542: ncol = PetscMax(pnnz[i], ncol);
5543: /* diagonal */
5544: dntotalcols += nlcols[i * 2 + 0];
5545: /* off-diagonal */
5546: ontotalcols += nlcols[i * 2 + 1];
5547: }
5548: /* We do not need to figure the right number of columns
5549: * since all the calculations will be done by going through the raw data
5550: * */
5551: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5552: PetscCall(MatSetUp(*P_oth));
5553: PetscCall(PetscFree(pnnz));
5554: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5555: /* diagonal */
5556: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5557: /* off-diagonal */
5558: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5559: /* diagonal */
5560: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5561: /* off-diagonal */
5562: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5563: dntotalcols = 0;
5564: ontotalcols = 0;
5565: ntotalcols = 0;
5566: for (i = 0; i < nrows; i++) {
5567: owner = 0;
5568: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5569: /* Set iremote for diag matrix */
5570: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5571: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5572: iremote[dntotalcols].rank = owner;
5573: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5574: ilocal[dntotalcols++] = ntotalcols++;
5575: }
5576: /* off-diagonal */
5577: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5578: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5579: oiremote[ontotalcols].rank = owner;
5580: oilocal[ontotalcols++] = ntotalcols++;
5581: }
5582: }
5583: PetscCall(ISRestoreIndices(rows, &lrowindices));
5584: PetscCall(PetscFree(loffsets));
5585: PetscCall(PetscFree(nlcols));
5586: PetscCall(PetscSFCreate(comm, &sf));
5587: /* P serves as roots and P_oth is leaves
5588: * Diag matrix
5589: * */
5590: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5591: PetscCall(PetscSFSetFromOptions(sf));
5592: PetscCall(PetscSFSetUp(sf));
5594: PetscCall(PetscSFCreate(comm, &osf));
5595: /* off-diagonal */
5596: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5597: PetscCall(PetscSFSetFromOptions(osf));
5598: PetscCall(PetscSFSetUp(osf));
5599: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5600: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5601: /* operate on the matrix internal data to save memory */
5602: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5603: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5604: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5605: /* Convert to global indices for diag matrix */
5606: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5607: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5608: /* We want P_oth store global indices */
5609: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5610: /* Use memory scalable approach */
5611: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5612: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5613: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5614: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5615: /* Convert back to local indices */
5616: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5617: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5618: nout = 0;
5619: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5620: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5621: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5622: /* Exchange values */
5623: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5624: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5625: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5626: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5627: /* Stop PETSc from shrinking memory */
5628: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5629: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5630: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5631: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5632: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5633: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5634: PetscCall(PetscSFDestroy(&sf));
5635: PetscCall(PetscSFDestroy(&osf));
5636: PetscFunctionReturn(PETSC_SUCCESS);
5637: }
5639: /*
5640: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5641: * This supports MPIAIJ and MAIJ
5642: * */
5643: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5644: {
5645: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5646: Mat_SeqAIJ *p_oth;
5647: IS rows, map;
5648: PetscHMapI hamp;
5649: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5650: MPI_Comm comm;
5651: PetscSF sf, osf;
5652: PetscBool has;
5654: PetscFunctionBegin;
5655: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5656: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5657: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5658: * and then create a submatrix (that often is an overlapping matrix)
5659: * */
5660: if (reuse == MAT_INITIAL_MATRIX) {
5661: /* Use a hash table to figure out unique keys */
5662: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5663: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5664: count = 0;
5665: /* Assume that a->g is sorted, otherwise the following does not make sense */
5666: for (i = 0; i < a->B->cmap->n; i++) {
5667: key = a->garray[i] / dof;
5668: PetscCall(PetscHMapIHas(hamp, key, &has));
5669: if (!has) {
5670: mapping[i] = count;
5671: PetscCall(PetscHMapISet(hamp, key, count++));
5672: } else {
5673: /* Current 'i' has the same value the previous step */
5674: mapping[i] = count - 1;
5675: }
5676: }
5677: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5678: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5679: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5680: PetscCall(PetscCalloc1(htsize, &rowindices));
5681: off = 0;
5682: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5683: PetscCall(PetscHMapIDestroy(&hamp));
5684: PetscCall(PetscSortInt(htsize, rowindices));
5685: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5686: /* In case, the matrix was already created but users want to recreate the matrix */
5687: PetscCall(MatDestroy(P_oth));
5688: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5689: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5690: PetscCall(ISDestroy(&map));
5691: PetscCall(ISDestroy(&rows));
5692: } else if (reuse == MAT_REUSE_MATRIX) {
5693: /* If matrix was already created, we simply update values using SF objects
5694: * that as attached to the matrix earlier.
5695: */
5696: const PetscScalar *pd_a, *po_a;
5698: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5699: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5700: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5701: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5702: /* Update values in place */
5703: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5704: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5705: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5706: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5707: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5708: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5709: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5710: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5711: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5712: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5713: PetscFunctionReturn(PETSC_SUCCESS);
5714: }
5716: /*@C
5717: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5719: Collective
5721: Input Parameters:
5722: + A - the first matrix in `MATMPIAIJ` format
5723: . B - the second matrix in `MATMPIAIJ` format
5724: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5726: Output Parameters:
5727: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5728: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5729: - B_seq - the sequential matrix generated
5731: Level: developer
5733: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5734: @*/
5735: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5736: {
5737: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5738: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5739: IS isrowb, iscolb;
5740: Mat *bseq = NULL;
5742: PetscFunctionBegin;
5743: 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 ")",
5744: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5745: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5747: if (scall == MAT_INITIAL_MATRIX) {
5748: start = A->cmap->rstart;
5749: cmap = a->garray;
5750: nzA = a->A->cmap->n;
5751: nzB = a->B->cmap->n;
5752: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5753: ncols = 0;
5754: for (i = 0; i < nzB; i++) { /* row < local row index */
5755: if (cmap[i] < start) idx[ncols++] = cmap[i];
5756: else break;
5757: }
5758: imark = i;
5759: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5760: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5761: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5762: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5763: } else {
5764: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5765: isrowb = *rowb;
5766: iscolb = *colb;
5767: PetscCall(PetscMalloc1(1, &bseq));
5768: bseq[0] = *B_seq;
5769: }
5770: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5771: *B_seq = bseq[0];
5772: PetscCall(PetscFree(bseq));
5773: if (!rowb) {
5774: PetscCall(ISDestroy(&isrowb));
5775: } else {
5776: *rowb = isrowb;
5777: }
5778: if (!colb) {
5779: PetscCall(ISDestroy(&iscolb));
5780: } else {
5781: *colb = iscolb;
5782: }
5783: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5784: PetscFunctionReturn(PETSC_SUCCESS);
5785: }
5787: /*
5788: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5789: of the OFF-DIAGONAL portion of local A
5791: Collective
5793: Input Parameters:
5794: + A,B - the matrices in `MATMPIAIJ` format
5795: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5797: Output Parameter:
5798: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5799: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5800: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5801: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5803: Developer Note:
5804: This directly accesses information inside the VecScatter associated with the matrix-vector product
5805: for this matrix. This is not desirable..
5807: Level: developer
5809: */
5810: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5811: {
5812: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5813: Mat_SeqAIJ *b_oth;
5814: VecScatter ctx;
5815: MPI_Comm comm;
5816: const PetscMPIInt *rprocs, *sprocs;
5817: const PetscInt *srow, *rstarts, *sstarts;
5818: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5819: PetscInt i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5820: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5821: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5822: PetscMPIInt size, tag, rank, nreqs;
5824: PetscFunctionBegin;
5825: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5826: PetscCallMPI(MPI_Comm_size(comm, &size));
5828: 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 ")",
5829: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5830: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5831: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5833: if (size == 1) {
5834: startsj_s = NULL;
5835: bufa_ptr = NULL;
5836: *B_oth = NULL;
5837: PetscFunctionReturn(PETSC_SUCCESS);
5838: }
5840: ctx = a->Mvctx;
5841: tag = ((PetscObject)ctx)->tag;
5843: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5844: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5845: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5846: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5847: PetscCall(PetscMalloc1(nreqs, &reqs));
5848: rwaits = reqs;
5849: swaits = reqs + nrecvs;
5851: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5852: if (scall == MAT_INITIAL_MATRIX) {
5853: /* i-array */
5854: /* post receives */
5855: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5856: for (i = 0; i < nrecvs; i++) {
5857: rowlen = rvalues + rstarts[i] * rbs;
5858: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5859: PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5860: }
5862: /* pack the outgoing message */
5863: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5865: sstartsj[0] = 0;
5866: rstartsj[0] = 0;
5867: len = 0; /* total length of j or a array to be sent */
5868: if (nsends) {
5869: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5870: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5871: }
5872: for (i = 0; i < nsends; i++) {
5873: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5874: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5875: for (j = 0; j < nrows; j++) {
5876: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5877: for (l = 0; l < sbs; l++) {
5878: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5880: rowlen[j * sbs + l] = ncols;
5882: len += ncols;
5883: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5884: }
5885: k++;
5886: }
5887: PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5889: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5890: }
5891: /* recvs and sends of i-array are completed */
5892: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5893: PetscCall(PetscFree(svalues));
5895: /* allocate buffers for sending j and a arrays */
5896: PetscCall(PetscMalloc1(len + 1, &bufj));
5897: PetscCall(PetscMalloc1(len + 1, &bufa));
5899: /* create i-array of B_oth */
5900: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5902: b_othi[0] = 0;
5903: len = 0; /* total length of j or a array to be received */
5904: k = 0;
5905: for (i = 0; i < nrecvs; i++) {
5906: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5907: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5908: for (j = 0; j < nrows; j++) {
5909: b_othi[k + 1] = b_othi[k] + rowlen[j];
5910: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5911: k++;
5912: }
5913: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5914: }
5915: PetscCall(PetscFree(rvalues));
5917: /* allocate space for j and a arrays of B_oth */
5918: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5919: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5921: /* j-array */
5922: /* post receives of j-array */
5923: for (i = 0; i < nrecvs; i++) {
5924: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5925: PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5926: }
5928: /* pack the outgoing message j-array */
5929: if (nsends) k = sstarts[0];
5930: for (i = 0; i < nsends; i++) {
5931: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5932: bufJ = bufj + sstartsj[i];
5933: for (j = 0; j < nrows; j++) {
5934: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5935: for (ll = 0; ll < sbs; ll++) {
5936: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5937: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5938: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5939: }
5940: }
5941: PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5942: }
5944: /* recvs and sends of j-array are completed */
5945: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5946: } else if (scall == MAT_REUSE_MATRIX) {
5947: sstartsj = *startsj_s;
5948: rstartsj = *startsj_r;
5949: bufa = *bufa_ptr;
5950: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5951: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5952: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5954: /* a-array */
5955: /* post receives of a-array */
5956: for (i = 0; i < nrecvs; i++) {
5957: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5958: PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5959: }
5961: /* pack the outgoing message a-array */
5962: if (nsends) k = sstarts[0];
5963: for (i = 0; i < nsends; i++) {
5964: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5965: bufA = bufa + sstartsj[i];
5966: for (j = 0; j < nrows; j++) {
5967: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5968: for (ll = 0; ll < sbs; ll++) {
5969: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5970: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5971: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5972: }
5973: }
5974: PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5975: }
5976: /* recvs and sends of a-array are completed */
5977: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5978: PetscCall(PetscFree(reqs));
5980: if (scall == MAT_INITIAL_MATRIX) {
5981: /* put together the new matrix */
5982: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
5984: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5985: /* Since these are PETSc arrays, change flags to free them as necessary. */
5986: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5987: b_oth->free_a = PETSC_TRUE;
5988: b_oth->free_ij = PETSC_TRUE;
5989: b_oth->nonew = 0;
5991: PetscCall(PetscFree(bufj));
5992: if (!startsj_s || !bufa_ptr) {
5993: PetscCall(PetscFree2(sstartsj, rstartsj));
5994: PetscCall(PetscFree(bufa_ptr));
5995: } else {
5996: *startsj_s = sstartsj;
5997: *startsj_r = rstartsj;
5998: *bufa_ptr = bufa;
5999: }
6000: } else if (scall == MAT_REUSE_MATRIX) {
6001: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6002: }
6004: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6005: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6006: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6007: PetscFunctionReturn(PETSC_SUCCESS);
6008: }
6010: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6011: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6012: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6013: #if defined(PETSC_HAVE_MKL_SPARSE)
6014: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6015: #endif
6016: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6017: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6018: #if defined(PETSC_HAVE_ELEMENTAL)
6019: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6020: #endif
6021: #if defined(PETSC_HAVE_SCALAPACK)
6022: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6023: #endif
6024: #if defined(PETSC_HAVE_HYPRE)
6025: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6026: #endif
6027: #if defined(PETSC_HAVE_CUDA)
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6029: #endif
6030: #if defined(PETSC_HAVE_HIP)
6031: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6032: #endif
6033: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6034: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6035: #endif
6036: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6037: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6038: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6040: /*
6041: Computes (B'*A')' since computing B*A directly is untenable
6043: n p p
6044: [ ] [ ] [ ]
6045: m [ A ] * n [ B ] = m [ C ]
6046: [ ] [ ] [ ]
6048: */
6049: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6050: {
6051: Mat At, Bt, Ct;
6053: PetscFunctionBegin;
6054: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6055: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6056: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6057: PetscCall(MatDestroy(&At));
6058: PetscCall(MatDestroy(&Bt));
6059: PetscCall(MatTransposeSetPrecursor(Ct, C));
6060: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6061: PetscCall(MatDestroy(&Ct));
6062: PetscFunctionReturn(PETSC_SUCCESS);
6063: }
6065: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6066: {
6067: PetscBool cisdense;
6069: PetscFunctionBegin;
6070: 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);
6071: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6072: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6073: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6074: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6075: PetscCall(MatSetUp(C));
6077: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6078: PetscFunctionReturn(PETSC_SUCCESS);
6079: }
6081: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6082: {
6083: Mat_Product *product = C->product;
6084: Mat A = product->A, B = product->B;
6086: PetscFunctionBegin;
6087: 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 ")",
6088: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6089: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6090: C->ops->productsymbolic = MatProductSymbolic_AB;
6091: PetscFunctionReturn(PETSC_SUCCESS);
6092: }
6094: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6095: {
6096: Mat_Product *product = C->product;
6098: PetscFunctionBegin;
6099: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6100: PetscFunctionReturn(PETSC_SUCCESS);
6101: }
6103: /*
6104: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6106: Input Parameters:
6108: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6109: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6111: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6113: For Set1, j1[] contains column indices of the nonzeros.
6114: 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
6115: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6116: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6118: Similar for Set2.
6120: This routine merges the two sets of nonzeros row by row and removes repeats.
6122: Output Parameters: (memory is allocated by the caller)
6124: i[],j[]: the CSR of the merged matrix, which has m rows.
6125: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6126: imap2[]: similar to imap1[], but for Set2.
6127: Note we order nonzeros row-by-row and from left to right.
6128: */
6129: 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[])
6130: {
6131: PetscInt r, m; /* Row index of mat */
6132: PetscCount t, t1, t2, b1, e1, b2, e2;
6134: PetscFunctionBegin;
6135: PetscCall(MatGetLocalSize(mat, &m, NULL));
6136: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6137: i[0] = 0;
6138: for (r = 0; r < m; r++) { /* Do row by row merging */
6139: b1 = rowBegin1[r];
6140: e1 = rowEnd1[r];
6141: b2 = rowBegin2[r];
6142: e2 = rowEnd2[r];
6143: while (b1 < e1 && b2 < e2) {
6144: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6145: j[t] = j1[b1];
6146: imap1[t1] = t;
6147: imap2[t2] = t;
6148: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6149: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6150: t1++;
6151: t2++;
6152: t++;
6153: } else if (j1[b1] < j2[b2]) {
6154: j[t] = j1[b1];
6155: imap1[t1] = t;
6156: b1 += jmap1[t1 + 1] - jmap1[t1];
6157: t1++;
6158: t++;
6159: } else {
6160: j[t] = j2[b2];
6161: imap2[t2] = t;
6162: b2 += jmap2[t2 + 1] - jmap2[t2];
6163: t2++;
6164: t++;
6165: }
6166: }
6167: /* Merge the remaining in either j1[] or j2[] */
6168: while (b1 < e1) {
6169: j[t] = j1[b1];
6170: imap1[t1] = t;
6171: b1 += jmap1[t1 + 1] - jmap1[t1];
6172: t1++;
6173: t++;
6174: }
6175: while (b2 < e2) {
6176: j[t] = j2[b2];
6177: imap2[t2] = t;
6178: b2 += jmap2[t2 + 1] - jmap2[t2];
6179: t2++;
6180: t++;
6181: }
6182: i[r + 1] = t;
6183: }
6184: PetscFunctionReturn(PETSC_SUCCESS);
6185: }
6187: /*
6188: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6190: Input Parameters:
6191: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6192: 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[]
6193: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6195: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6196: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6198: Output Parameters:
6199: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6200: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6201: 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,
6202: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6204: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6205: Atot: number of entries belonging to the diagonal block.
6206: Annz: number of unique nonzeros belonging to the diagonal block.
6207: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6208: repeats (i.e., same 'i,j' pair).
6209: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6210: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6212: Atot: number of entries belonging to the diagonal block
6213: Annz: number of unique nonzeros belonging to the diagonal block.
6215: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6217: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6218: */
6219: 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_)
6220: {
6221: PetscInt cstart, cend, rstart, rend, row, col;
6222: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6223: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6224: PetscCount k, m, p, q, r, s, mid;
6225: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6227: PetscFunctionBegin;
6228: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6229: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6230: m = rend - rstart;
6232: /* Skip negative rows */
6233: for (k = 0; k < n; k++)
6234: if (i[k] >= 0) break;
6236: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6237: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6238: */
6239: while (k < n) {
6240: row = i[k];
6241: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6242: for (s = k; s < n; s++)
6243: if (i[s] != row) break;
6245: /* Shift diag columns to range of [-PETSC_MAX_INT, -1] */
6246: for (p = k; p < s; p++) {
6247: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT;
6248: 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]);
6249: }
6250: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6251: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6252: rowBegin[row - rstart] = k;
6253: rowMid[row - rstart] = mid;
6254: rowEnd[row - rstart] = s;
6256: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6257: Atot += mid - k;
6258: Btot += s - mid;
6260: /* Count unique nonzeros of this diag row */
6261: for (p = k; p < mid;) {
6262: col = j[p];
6263: do {
6264: j[p] += PETSC_MAX_INT; /* Revert the modified diagonal indices */
6265: p++;
6266: } while (p < mid && j[p] == col);
6267: Annz++;
6268: }
6270: /* Count unique nonzeros of this offdiag row */
6271: for (p = mid; p < s;) {
6272: col = j[p];
6273: do {
6274: p++;
6275: } while (p < s && j[p] == col);
6276: Bnnz++;
6277: }
6278: k = s;
6279: }
6281: /* Allocation according to Atot, Btot, Annz, Bnnz */
6282: PetscCall(PetscMalloc1(Atot, &Aperm));
6283: PetscCall(PetscMalloc1(Btot, &Bperm));
6284: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6285: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6287: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6288: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6289: for (r = 0; r < m; r++) {
6290: k = rowBegin[r];
6291: mid = rowMid[r];
6292: s = rowEnd[r];
6293: PetscCall(PetscArraycpy(Aperm + Atot, perm + k, mid - k));
6294: PetscCall(PetscArraycpy(Bperm + Btot, perm + mid, s - mid));
6295: Atot += mid - k;
6296: Btot += s - mid;
6298: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6299: for (p = k; p < mid;) {
6300: col = j[p];
6301: q = p;
6302: do {
6303: p++;
6304: } while (p < mid && j[p] == col);
6305: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6306: Annz++;
6307: }
6309: for (p = mid; p < s;) {
6310: col = j[p];
6311: q = p;
6312: do {
6313: p++;
6314: } while (p < s && j[p] == col);
6315: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6316: Bnnz++;
6317: }
6318: }
6319: /* Output */
6320: *Aperm_ = Aperm;
6321: *Annz_ = Annz;
6322: *Atot_ = Atot;
6323: *Ajmap_ = Ajmap;
6324: *Bperm_ = Bperm;
6325: *Bnnz_ = Bnnz;
6326: *Btot_ = Btot;
6327: *Bjmap_ = Bjmap;
6328: PetscFunctionReturn(PETSC_SUCCESS);
6329: }
6331: /*
6332: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6334: Input Parameters:
6335: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6336: nnz: number of unique nonzeros in the merged matrix
6337: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6338: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6340: Output Parameter: (memory is allocated by the caller)
6341: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6343: Example:
6344: nnz1 = 4
6345: nnz = 6
6346: imap = [1,3,4,5]
6347: jmap = [0,3,5,6,7]
6348: then,
6349: jmap_new = [0,0,3,3,5,6,7]
6350: */
6351: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6352: {
6353: PetscCount k, p;
6355: PetscFunctionBegin;
6356: jmap_new[0] = 0;
6357: p = nnz; /* p loops over jmap_new[] backwards */
6358: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6359: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6360: }
6361: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6362: PetscFunctionReturn(PETSC_SUCCESS);
6363: }
6365: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6366: {
6367: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6369: PetscFunctionBegin;
6370: PetscCall(PetscSFDestroy(&coo->sf));
6371: PetscCall(PetscFree(coo->Aperm1));
6372: PetscCall(PetscFree(coo->Bperm1));
6373: PetscCall(PetscFree(coo->Ajmap1));
6374: PetscCall(PetscFree(coo->Bjmap1));
6375: PetscCall(PetscFree(coo->Aimap2));
6376: PetscCall(PetscFree(coo->Bimap2));
6377: PetscCall(PetscFree(coo->Aperm2));
6378: PetscCall(PetscFree(coo->Bperm2));
6379: PetscCall(PetscFree(coo->Ajmap2));
6380: PetscCall(PetscFree(coo->Bjmap2));
6381: PetscCall(PetscFree(coo->Cperm1));
6382: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6383: PetscCall(PetscFree(coo));
6384: PetscFunctionReturn(PETSC_SUCCESS);
6385: }
6387: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6388: {
6389: MPI_Comm comm;
6390: PetscMPIInt rank, size;
6391: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6392: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6393: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6394: PetscContainer container;
6395: MatCOOStruct_MPIAIJ *coo;
6397: PetscFunctionBegin;
6398: PetscCall(PetscFree(mpiaij->garray));
6399: PetscCall(VecDestroy(&mpiaij->lvec));
6400: #if defined(PETSC_USE_CTABLE)
6401: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6402: #else
6403: PetscCall(PetscFree(mpiaij->colmap));
6404: #endif
6405: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6406: mat->assembled = PETSC_FALSE;
6407: mat->was_assembled = PETSC_FALSE;
6409: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6410: PetscCallMPI(MPI_Comm_size(comm, &size));
6411: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6412: PetscCall(PetscLayoutSetUp(mat->rmap));
6413: PetscCall(PetscLayoutSetUp(mat->cmap));
6414: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6415: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6416: PetscCall(MatGetLocalSize(mat, &m, &n));
6417: PetscCall(MatGetSize(mat, &M, &N));
6419: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6420: /* entries come first, then local rows, then remote rows. */
6421: PetscCount n1 = coo_n, *perm1;
6422: PetscInt *i1 = coo_i, *j1 = coo_j;
6424: PetscCall(PetscMalloc1(n1, &perm1));
6425: for (k = 0; k < n1; k++) perm1[k] = k;
6427: /* Manipulate indices so that entries with negative row or col indices will have smallest
6428: row indices, local entries will have greater but negative row indices, and remote entries
6429: will have positive row indices.
6430: */
6431: for (k = 0; k < n1; k++) {
6432: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT; /* e.g., -2^31, minimal to move them ahead */
6433: 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] */
6434: else {
6435: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6436: if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6437: }
6438: }
6440: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6441: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6443: /* Advance k to the first entry we need to take care of */
6444: for (k = 0; k < n1; k++)
6445: if (i1[k] > PETSC_MIN_INT) break;
6446: PetscInt i1start = k;
6448: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6449: for (; k < rem; k++) i1[k] += PETSC_MAX_INT; /* Revert row indices of local rows*/
6451: /* Send remote rows to their owner */
6452: /* Find which rows should be sent to which remote ranks*/
6453: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6454: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6455: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6456: const PetscInt *ranges;
6457: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6459: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6460: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6461: for (k = rem; k < n1;) {
6462: PetscMPIInt owner;
6463: PetscInt firstRow, lastRow;
6465: /* Locate a row range */
6466: firstRow = i1[k]; /* first row of this owner */
6467: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6468: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6470: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6471: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6473: /* All entries in [k,p) belong to this remote owner */
6474: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6475: PetscMPIInt *sendto2;
6476: PetscInt *nentries2;
6477: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6479: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6480: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6481: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6482: PetscCall(PetscFree2(sendto, nentries2));
6483: sendto = sendto2;
6484: nentries = nentries2;
6485: maxNsend = maxNsend2;
6486: }
6487: sendto[nsend] = owner;
6488: nentries[nsend] = p - k;
6489: PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6490: nsend++;
6491: k = p;
6492: }
6494: /* Build 1st SF to know offsets on remote to send data */
6495: PetscSF sf1;
6496: PetscInt nroots = 1, nroots2 = 0;
6497: PetscInt nleaves = nsend, nleaves2 = 0;
6498: PetscInt *offsets;
6499: PetscSFNode *iremote;
6501: PetscCall(PetscSFCreate(comm, &sf1));
6502: PetscCall(PetscMalloc1(nsend, &iremote));
6503: PetscCall(PetscMalloc1(nsend, &offsets));
6504: for (k = 0; k < nsend; k++) {
6505: iremote[k].rank = sendto[k];
6506: iremote[k].index = 0;
6507: nleaves2 += nentries[k];
6508: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6509: }
6510: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6511: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6512: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6513: PetscCall(PetscSFDestroy(&sf1));
6514: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT "", nleaves2, n1 - rem);
6516: /* Build 2nd SF to send remote COOs to their owner */
6517: PetscSF sf2;
6518: nroots = nroots2;
6519: nleaves = nleaves2;
6520: PetscCall(PetscSFCreate(comm, &sf2));
6521: PetscCall(PetscSFSetFromOptions(sf2));
6522: PetscCall(PetscMalloc1(nleaves, &iremote));
6523: p = 0;
6524: for (k = 0; k < nsend; k++) {
6525: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6526: for (q = 0; q < nentries[k]; q++, p++) {
6527: iremote[p].rank = sendto[k];
6528: iremote[p].index = offsets[k] + q;
6529: }
6530: }
6531: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6533: /* Send the remote COOs to their owner */
6534: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6535: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6536: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6537: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1 + rem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6538: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1 + rem, i2, MPI_REPLACE));
6539: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1 + rem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6540: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1 + rem, j2, MPI_REPLACE));
6542: PetscCall(PetscFree(offsets));
6543: PetscCall(PetscFree2(sendto, nentries));
6545: /* Sort received COOs by row along with the permutation array */
6546: for (k = 0; k < n2; k++) perm2[k] = k;
6547: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6549: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6550: PetscCount *Cperm1;
6551: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6552: PetscCall(PetscArraycpy(Cperm1, perm1 + rem, nleaves));
6554: /* Support for HYPRE matrices, kind of a hack.
6555: Swap min column with diagonal so that diagonal values will go first */
6556: PetscBool hypre;
6557: const char *name;
6558: PetscCall(PetscObjectGetName((PetscObject)mat, &name));
6559: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
6560: if (hypre) {
6561: PetscInt *minj;
6562: PetscBT hasdiag;
6564: PetscCall(PetscBTCreate(m, &hasdiag));
6565: PetscCall(PetscMalloc1(m, &minj));
6566: for (k = 0; k < m; k++) minj[k] = PETSC_MAX_INT;
6567: for (k = i1start; k < rem; k++) {
6568: if (j1[k] < cstart || j1[k] >= cend) continue;
6569: const PetscInt rindex = i1[k] - rstart;
6570: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6571: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6572: }
6573: for (k = 0; k < n2; k++) {
6574: if (j2[k] < cstart || j2[k] >= cend) continue;
6575: const PetscInt rindex = i2[k] - rstart;
6576: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6577: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6578: }
6579: for (k = i1start; k < rem; k++) {
6580: const PetscInt rindex = i1[k] - rstart;
6581: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6582: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6583: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6584: }
6585: for (k = 0; k < n2; k++) {
6586: const PetscInt rindex = i2[k] - rstart;
6587: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6588: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6589: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6590: }
6591: PetscCall(PetscBTDestroy(&hasdiag));
6592: PetscCall(PetscFree(minj));
6593: }
6595: /* Split local COOs and received COOs into diag/offdiag portions */
6596: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6597: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6598: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6599: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6600: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6601: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6603: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6604: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6605: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6606: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6608: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6609: PetscInt *Ai, *Bi;
6610: PetscInt *Aj, *Bj;
6612: PetscCall(PetscMalloc1(m + 1, &Ai));
6613: PetscCall(PetscMalloc1(m + 1, &Bi));
6614: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6615: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6617: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6618: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6619: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6620: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6621: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6623: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6624: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6626: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6627: /* expect nonzeros in A/B most likely have local contributing entries */
6628: PetscInt Annz = Ai[m];
6629: PetscInt Bnnz = Bi[m];
6630: PetscCount *Ajmap1_new, *Bjmap1_new;
6632: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6633: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6635: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6636: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6638: PetscCall(PetscFree(Aimap1));
6639: PetscCall(PetscFree(Ajmap1));
6640: PetscCall(PetscFree(Bimap1));
6641: PetscCall(PetscFree(Bjmap1));
6642: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6643: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6644: PetscCall(PetscFree(perm1));
6645: PetscCall(PetscFree3(i2, j2, perm2));
6647: Ajmap1 = Ajmap1_new;
6648: Bjmap1 = Bjmap1_new;
6650: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6651: if (Annz < Annz1 + Annz2) {
6652: PetscInt *Aj_new;
6653: PetscCall(PetscMalloc1(Annz, &Aj_new));
6654: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6655: PetscCall(PetscFree(Aj));
6656: Aj = Aj_new;
6657: }
6659: if (Bnnz < Bnnz1 + Bnnz2) {
6660: PetscInt *Bj_new;
6661: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6662: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6663: PetscCall(PetscFree(Bj));
6664: Bj = Bj_new;
6665: }
6667: /* Create new submatrices for on-process and off-process coupling */
6668: PetscScalar *Aa, *Ba;
6669: MatType rtype;
6670: Mat_SeqAIJ *a, *b;
6671: PetscObjectState state;
6672: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6673: PetscCall(PetscCalloc1(Bnnz, &Ba));
6674: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6675: if (cstart) {
6676: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6677: }
6678: PetscCall(MatDestroy(&mpiaij->A));
6679: PetscCall(MatDestroy(&mpiaij->B));
6680: PetscCall(MatGetRootType_Private(mat, &rtype));
6681: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6682: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6683: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6684: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6685: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6686: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6688: a = (Mat_SeqAIJ *)mpiaij->A->data;
6689: b = (Mat_SeqAIJ *)mpiaij->B->data;
6690: a->singlemalloc = b->singlemalloc = PETSC_FALSE; /* Let newmat own Ai,Aj,Aa,Bi,Bj,Ba */
6691: a->free_a = b->free_a = PETSC_TRUE;
6692: a->free_ij = b->free_ij = PETSC_TRUE;
6694: /* conversion must happen AFTER multiply setup */
6695: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6696: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6697: PetscCall(VecDestroy(&mpiaij->lvec));
6698: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6700: // Put the COO struct in a container and then attach that to the matrix
6701: PetscCall(PetscMalloc1(1, &coo));
6702: coo->n = coo_n;
6703: coo->sf = sf2;
6704: coo->sendlen = nleaves;
6705: coo->recvlen = nroots;
6706: coo->Annz = Annz;
6707: coo->Bnnz = Bnnz;
6708: coo->Annz2 = Annz2;
6709: coo->Bnnz2 = Bnnz2;
6710: coo->Atot1 = Atot1;
6711: coo->Atot2 = Atot2;
6712: coo->Btot1 = Btot1;
6713: coo->Btot2 = Btot2;
6714: coo->Ajmap1 = Ajmap1;
6715: coo->Aperm1 = Aperm1;
6716: coo->Bjmap1 = Bjmap1;
6717: coo->Bperm1 = Bperm1;
6718: coo->Aimap2 = Aimap2;
6719: coo->Ajmap2 = Ajmap2;
6720: coo->Aperm2 = Aperm2;
6721: coo->Bimap2 = Bimap2;
6722: coo->Bjmap2 = Bjmap2;
6723: coo->Bperm2 = Bperm2;
6724: coo->Cperm1 = Cperm1;
6725: // Allocate in preallocation. If not used, it has zero cost on host
6726: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6727: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6728: PetscCall(PetscContainerSetPointer(container, coo));
6729: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6730: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6731: PetscCall(PetscContainerDestroy(&container));
6732: PetscFunctionReturn(PETSC_SUCCESS);
6733: }
6735: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6736: {
6737: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6738: Mat A = mpiaij->A, B = mpiaij->B;
6739: PetscScalar *Aa, *Ba;
6740: PetscScalar *sendbuf, *recvbuf;
6741: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6742: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6743: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6744: const PetscCount *Cperm1;
6745: PetscContainer container;
6746: MatCOOStruct_MPIAIJ *coo;
6748: PetscFunctionBegin;
6749: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6750: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6751: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6752: sendbuf = coo->sendbuf;
6753: recvbuf = coo->recvbuf;
6754: Ajmap1 = coo->Ajmap1;
6755: Ajmap2 = coo->Ajmap2;
6756: Aimap2 = coo->Aimap2;
6757: Bjmap1 = coo->Bjmap1;
6758: Bjmap2 = coo->Bjmap2;
6759: Bimap2 = coo->Bimap2;
6760: Aperm1 = coo->Aperm1;
6761: Aperm2 = coo->Aperm2;
6762: Bperm1 = coo->Bperm1;
6763: Bperm2 = coo->Bperm2;
6764: Cperm1 = coo->Cperm1;
6766: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6767: PetscCall(MatSeqAIJGetArray(B, &Ba));
6769: /* Pack entries to be sent to remote */
6770: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6772: /* Send remote entries to their owner and overlap the communication with local computation */
6773: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6774: /* Add local entries to A and B */
6775: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6776: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6777: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6778: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6779: }
6780: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6781: PetscScalar sum = 0.0;
6782: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6783: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6784: }
6785: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6787: /* Add received remote entries to A and B */
6788: for (PetscCount i = 0; i < coo->Annz2; i++) {
6789: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6790: }
6791: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6792: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6793: }
6794: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6795: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6796: PetscFunctionReturn(PETSC_SUCCESS);
6797: }
6799: /*MC
6800: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6802: Options Database Keys:
6803: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6805: Level: beginner
6807: Notes:
6808: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6809: in this case the values associated with the rows and columns one passes in are set to zero
6810: in the matrix
6812: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6813: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6815: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6816: M*/
6817: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6818: {
6819: Mat_MPIAIJ *b;
6820: PetscMPIInt size;
6822: PetscFunctionBegin;
6823: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6825: PetscCall(PetscNew(&b));
6826: B->data = (void *)b;
6827: B->ops[0] = MatOps_Values;
6828: B->assembled = PETSC_FALSE;
6829: B->insertmode = NOT_SET_VALUES;
6830: b->size = size;
6832: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6834: /* build cache for off array entries formed */
6835: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6837: b->donotstash = PETSC_FALSE;
6838: b->colmap = NULL;
6839: b->garray = NULL;
6840: b->roworiented = PETSC_TRUE;
6842: /* stuff used for matrix vector multiply */
6843: b->lvec = NULL;
6844: b->Mvctx = NULL;
6846: /* stuff for MatGetRow() */
6847: b->rowindices = NULL;
6848: b->rowvalues = NULL;
6849: b->getrowactive = PETSC_FALSE;
6851: /* flexible pointer used in CUSPARSE classes */
6852: b->spptr = NULL;
6854: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6855: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6856: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6857: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6858: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6859: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6860: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6861: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6862: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6863: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6864: #if defined(PETSC_HAVE_CUDA)
6865: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6866: #endif
6867: #if defined(PETSC_HAVE_HIP)
6868: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6869: #endif
6870: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6871: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6872: #endif
6873: #if defined(PETSC_HAVE_MKL_SPARSE)
6874: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6875: #endif
6876: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6877: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6878: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6879: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6880: #if defined(PETSC_HAVE_ELEMENTAL)
6881: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6882: #endif
6883: #if defined(PETSC_HAVE_SCALAPACK)
6884: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6885: #endif
6886: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6887: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6888: #if defined(PETSC_HAVE_HYPRE)
6889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6890: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6891: #endif
6892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6896: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6897: PetscFunctionReturn(PETSC_SUCCESS);
6898: }
6900: /*@C
6901: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6902: and "off-diagonal" part of the matrix in CSR format.
6904: Collective
6906: Input Parameters:
6907: + comm - MPI communicator
6908: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6909: . n - This value should be the same as the local size used in creating the
6910: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
6911: calculated if `N` is given) For square matrices `n` is almost always `m`.
6912: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6913: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6914: . 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
6915: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6916: . a - matrix values
6917: . 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
6918: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6919: - oa - matrix values
6921: Output Parameter:
6922: . mat - the matrix
6924: Level: advanced
6926: Notes:
6927: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6928: must free the arrays once the matrix has been destroyed and not before.
6930: The `i` and `j` indices are 0 based
6932: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6934: This sets local rows and cannot be used to set off-processor values.
6936: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6937: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6938: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6939: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6940: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6941: communication if it is known that only local entries will be set.
6943: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6944: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6945: @*/
6946: 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)
6947: {
6948: Mat_MPIAIJ *maij;
6950: PetscFunctionBegin;
6951: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6952: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6953: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6954: PetscCall(MatCreate(comm, mat));
6955: PetscCall(MatSetSizes(*mat, m, n, M, N));
6956: PetscCall(MatSetType(*mat, MATMPIAIJ));
6957: maij = (Mat_MPIAIJ *)(*mat)->data;
6959: (*mat)->preallocated = PETSC_TRUE;
6961: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6962: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6964: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6965: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
6967: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6968: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
6969: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
6970: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
6971: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
6972: PetscFunctionReturn(PETSC_SUCCESS);
6973: }
6975: typedef struct {
6976: Mat *mp; /* intermediate products */
6977: PetscBool *mptmp; /* is the intermediate product temporary ? */
6978: PetscInt cp; /* number of intermediate products */
6980: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
6981: PetscInt *startsj_s, *startsj_r;
6982: PetscScalar *bufa;
6983: Mat P_oth;
6985: /* may take advantage of merging product->B */
6986: Mat Bloc; /* B-local by merging diag and off-diag */
6988: /* cusparse does not have support to split between symbolic and numeric phases.
6989: When api_user is true, we don't need to update the numerical values
6990: of the temporary storage */
6991: PetscBool reusesym;
6993: /* support for COO values insertion */
6994: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
6995: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
6996: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
6997: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
6998: PetscSF sf; /* used for non-local values insertion and memory malloc */
6999: PetscMemType mtype;
7001: /* customization */
7002: PetscBool abmerge;
7003: PetscBool P_oth_bind;
7004: } MatMatMPIAIJBACKEND;
7006: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7007: {
7008: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7009: PetscInt i;
7011: PetscFunctionBegin;
7012: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7013: PetscCall(PetscFree(mmdata->bufa));
7014: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7015: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7016: PetscCall(MatDestroy(&mmdata->P_oth));
7017: PetscCall(MatDestroy(&mmdata->Bloc));
7018: PetscCall(PetscSFDestroy(&mmdata->sf));
7019: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7020: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7021: PetscCall(PetscFree(mmdata->own[0]));
7022: PetscCall(PetscFree(mmdata->own));
7023: PetscCall(PetscFree(mmdata->off[0]));
7024: PetscCall(PetscFree(mmdata->off));
7025: PetscCall(PetscFree(mmdata));
7026: PetscFunctionReturn(PETSC_SUCCESS);
7027: }
7029: /* Copy selected n entries with indices in idx[] of A to v[].
7030: If idx is NULL, copy the whole data array of A to v[]
7031: */
7032: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7033: {
7034: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7036: PetscFunctionBegin;
7037: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7038: if (f) {
7039: PetscCall((*f)(A, n, idx, v));
7040: } else {
7041: const PetscScalar *vv;
7043: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7044: if (n && idx) {
7045: PetscScalar *w = v;
7046: const PetscInt *oi = idx;
7047: PetscInt j;
7049: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7050: } else {
7051: PetscCall(PetscArraycpy(v, vv, n));
7052: }
7053: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7054: }
7055: PetscFunctionReturn(PETSC_SUCCESS);
7056: }
7058: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7059: {
7060: MatMatMPIAIJBACKEND *mmdata;
7061: PetscInt i, n_d, n_o;
7063: PetscFunctionBegin;
7064: MatCheckProduct(C, 1);
7065: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7066: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7067: if (!mmdata->reusesym) { /* update temporary matrices */
7068: 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));
7069: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7070: }
7071: mmdata->reusesym = PETSC_FALSE;
7073: for (i = 0; i < mmdata->cp; i++) {
7074: 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]);
7075: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7076: }
7077: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7078: PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];
7080: if (mmdata->mptmp[i]) continue;
7081: if (noff) {
7082: PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];
7084: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7085: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7086: n_o += noff;
7087: n_d += nown;
7088: } else {
7089: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7091: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7092: n_d += mm->nz;
7093: }
7094: }
7095: if (mmdata->hasoffproc) { /* offprocess insertion */
7096: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7097: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7098: }
7099: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7100: PetscFunctionReturn(PETSC_SUCCESS);
7101: }
7103: /* Support for Pt * A, A * P, or Pt * A * P */
7104: #define MAX_NUMBER_INTERMEDIATE 4
7105: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7106: {
7107: Mat_Product *product = C->product;
7108: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7109: Mat_MPIAIJ *a, *p;
7110: MatMatMPIAIJBACKEND *mmdata;
7111: ISLocalToGlobalMapping P_oth_l2g = NULL;
7112: IS glob = NULL;
7113: const char *prefix;
7114: char pprefix[256];
7115: const PetscInt *globidx, *P_oth_idx;
7116: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7117: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7118: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7119: /* type-0: consecutive, start from 0; type-1: consecutive with */
7120: /* a base offset; type-2: sparse with a local to global map table */
7121: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7123: MatProductType ptype;
7124: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7125: PetscMPIInt size;
7127: PetscFunctionBegin;
7128: MatCheckProduct(C, 1);
7129: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7130: ptype = product->type;
7131: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7132: ptype = MATPRODUCT_AB;
7133: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7134: }
7135: switch (ptype) {
7136: case MATPRODUCT_AB:
7137: A = product->A;
7138: P = product->B;
7139: m = A->rmap->n;
7140: n = P->cmap->n;
7141: M = A->rmap->N;
7142: N = P->cmap->N;
7143: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7144: break;
7145: case MATPRODUCT_AtB:
7146: P = product->A;
7147: A = product->B;
7148: m = P->cmap->n;
7149: n = A->cmap->n;
7150: M = P->cmap->N;
7151: N = A->cmap->N;
7152: hasoffproc = PETSC_TRUE;
7153: break;
7154: case MATPRODUCT_PtAP:
7155: A = product->A;
7156: P = product->B;
7157: m = P->cmap->n;
7158: n = P->cmap->n;
7159: M = P->cmap->N;
7160: N = P->cmap->N;
7161: hasoffproc = PETSC_TRUE;
7162: break;
7163: default:
7164: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7165: }
7166: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7167: if (size == 1) hasoffproc = PETSC_FALSE;
7169: /* defaults */
7170: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7171: mp[i] = NULL;
7172: mptmp[i] = PETSC_FALSE;
7173: rmapt[i] = -1;
7174: cmapt[i] = -1;
7175: rmapa[i] = NULL;
7176: cmapa[i] = NULL;
7177: }
7179: /* customization */
7180: PetscCall(PetscNew(&mmdata));
7181: mmdata->reusesym = product->api_user;
7182: if (ptype == MATPRODUCT_AB) {
7183: if (product->api_user) {
7184: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7185: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7186: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7187: PetscOptionsEnd();
7188: } else {
7189: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7190: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7191: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7192: PetscOptionsEnd();
7193: }
7194: } else if (ptype == MATPRODUCT_PtAP) {
7195: if (product->api_user) {
7196: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7197: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7198: PetscOptionsEnd();
7199: } else {
7200: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7201: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7202: PetscOptionsEnd();
7203: }
7204: }
7205: a = (Mat_MPIAIJ *)A->data;
7206: p = (Mat_MPIAIJ *)P->data;
7207: PetscCall(MatSetSizes(C, m, n, M, N));
7208: PetscCall(PetscLayoutSetUp(C->rmap));
7209: PetscCall(PetscLayoutSetUp(C->cmap));
7210: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7211: PetscCall(MatGetOptionsPrefix(C, &prefix));
7213: cp = 0;
7214: switch (ptype) {
7215: case MATPRODUCT_AB: /* A * P */
7216: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7218: /* A_diag * P_local (merged or not) */
7219: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7220: /* P is product->B */
7221: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7222: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7223: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7224: PetscCall(MatProductSetFill(mp[cp], product->fill));
7225: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7226: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7227: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7228: mp[cp]->product->api_user = product->api_user;
7229: PetscCall(MatProductSetFromOptions(mp[cp]));
7230: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7231: PetscCall(ISGetIndices(glob, &globidx));
7232: rmapt[cp] = 1;
7233: cmapt[cp] = 2;
7234: cmapa[cp] = globidx;
7235: mptmp[cp] = PETSC_FALSE;
7236: cp++;
7237: } else { /* A_diag * P_diag and A_diag * P_off */
7238: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7239: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7240: PetscCall(MatProductSetFill(mp[cp], product->fill));
7241: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7242: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7243: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7244: mp[cp]->product->api_user = product->api_user;
7245: PetscCall(MatProductSetFromOptions(mp[cp]));
7246: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7247: rmapt[cp] = 1;
7248: cmapt[cp] = 1;
7249: mptmp[cp] = PETSC_FALSE;
7250: cp++;
7251: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7252: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7253: PetscCall(MatProductSetFill(mp[cp], product->fill));
7254: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7255: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7256: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7257: mp[cp]->product->api_user = product->api_user;
7258: PetscCall(MatProductSetFromOptions(mp[cp]));
7259: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7260: rmapt[cp] = 1;
7261: cmapt[cp] = 2;
7262: cmapa[cp] = p->garray;
7263: mptmp[cp] = PETSC_FALSE;
7264: cp++;
7265: }
7267: /* A_off * P_other */
7268: if (mmdata->P_oth) {
7269: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7270: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7271: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)(a->B))->type_name));
7272: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7273: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7274: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7275: PetscCall(MatProductSetFill(mp[cp], product->fill));
7276: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7277: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7278: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7279: mp[cp]->product->api_user = product->api_user;
7280: PetscCall(MatProductSetFromOptions(mp[cp]));
7281: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7282: rmapt[cp] = 1;
7283: cmapt[cp] = 2;
7284: cmapa[cp] = P_oth_idx;
7285: mptmp[cp] = PETSC_FALSE;
7286: cp++;
7287: }
7288: break;
7290: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7291: /* A is product->B */
7292: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7293: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7294: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7295: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7296: PetscCall(MatProductSetFill(mp[cp], product->fill));
7297: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7298: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7299: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7300: mp[cp]->product->api_user = product->api_user;
7301: PetscCall(MatProductSetFromOptions(mp[cp]));
7302: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7303: PetscCall(ISGetIndices(glob, &globidx));
7304: rmapt[cp] = 2;
7305: rmapa[cp] = globidx;
7306: cmapt[cp] = 2;
7307: cmapa[cp] = globidx;
7308: mptmp[cp] = PETSC_FALSE;
7309: cp++;
7310: } else {
7311: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7312: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7313: PetscCall(MatProductSetFill(mp[cp], product->fill));
7314: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7315: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7316: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7317: mp[cp]->product->api_user = product->api_user;
7318: PetscCall(MatProductSetFromOptions(mp[cp]));
7319: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7320: PetscCall(ISGetIndices(glob, &globidx));
7321: rmapt[cp] = 1;
7322: cmapt[cp] = 2;
7323: cmapa[cp] = globidx;
7324: mptmp[cp] = PETSC_FALSE;
7325: cp++;
7326: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7327: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7328: PetscCall(MatProductSetFill(mp[cp], product->fill));
7329: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7330: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7331: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7332: mp[cp]->product->api_user = product->api_user;
7333: PetscCall(MatProductSetFromOptions(mp[cp]));
7334: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7335: rmapt[cp] = 2;
7336: rmapa[cp] = p->garray;
7337: cmapt[cp] = 2;
7338: cmapa[cp] = globidx;
7339: mptmp[cp] = PETSC_FALSE;
7340: cp++;
7341: }
7342: break;
7343: case MATPRODUCT_PtAP:
7344: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7345: /* P is product->B */
7346: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7347: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7348: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7349: PetscCall(MatProductSetFill(mp[cp], product->fill));
7350: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7351: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7352: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7353: mp[cp]->product->api_user = product->api_user;
7354: PetscCall(MatProductSetFromOptions(mp[cp]));
7355: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7356: PetscCall(ISGetIndices(glob, &globidx));
7357: rmapt[cp] = 2;
7358: rmapa[cp] = globidx;
7359: cmapt[cp] = 2;
7360: cmapa[cp] = globidx;
7361: mptmp[cp] = PETSC_FALSE;
7362: cp++;
7363: if (mmdata->P_oth) {
7364: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7365: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7366: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)(a->B))->type_name));
7367: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7368: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7369: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7370: PetscCall(MatProductSetFill(mp[cp], product->fill));
7371: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7372: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7373: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7374: mp[cp]->product->api_user = product->api_user;
7375: PetscCall(MatProductSetFromOptions(mp[cp]));
7376: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7377: mptmp[cp] = PETSC_TRUE;
7378: cp++;
7379: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7380: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7381: PetscCall(MatProductSetFill(mp[cp], product->fill));
7382: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7383: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7384: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7385: mp[cp]->product->api_user = product->api_user;
7386: PetscCall(MatProductSetFromOptions(mp[cp]));
7387: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7388: rmapt[cp] = 2;
7389: rmapa[cp] = globidx;
7390: cmapt[cp] = 2;
7391: cmapa[cp] = P_oth_idx;
7392: mptmp[cp] = PETSC_FALSE;
7393: cp++;
7394: }
7395: break;
7396: default:
7397: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7398: }
7399: /* sanity check */
7400: if (size > 1)
7401: 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);
7403: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7404: for (i = 0; i < cp; i++) {
7405: mmdata->mp[i] = mp[i];
7406: mmdata->mptmp[i] = mptmp[i];
7407: }
7408: mmdata->cp = cp;
7409: C->product->data = mmdata;
7410: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7411: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7413: /* memory type */
7414: mmdata->mtype = PETSC_MEMTYPE_HOST;
7415: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7416: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7417: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7418: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7419: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7420: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7422: /* prepare coo coordinates for values insertion */
7424: /* count total nonzeros of those intermediate seqaij Mats
7425: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7426: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7427: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7428: */
7429: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7430: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7431: if (mptmp[cp]) continue;
7432: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7433: const PetscInt *rmap = rmapa[cp];
7434: const PetscInt mr = mp[cp]->rmap->n;
7435: const PetscInt rs = C->rmap->rstart;
7436: const PetscInt re = C->rmap->rend;
7437: const PetscInt *ii = mm->i;
7438: for (i = 0; i < mr; i++) {
7439: const PetscInt gr = rmap[i];
7440: const PetscInt nz = ii[i + 1] - ii[i];
7441: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7442: else ncoo_oown += nz; /* this row is local */
7443: }
7444: } else ncoo_d += mm->nz;
7445: }
7447: /*
7448: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7450: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7452: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7454: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7455: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7456: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7458: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7459: 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.
7460: */
7461: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7462: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7464: /* gather (i,j) of nonzeros inserted by remote procs */
7465: if (hasoffproc) {
7466: PetscSF msf;
7467: PetscInt ncoo2, *coo_i2, *coo_j2;
7469: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7470: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7471: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7473: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7474: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7475: PetscInt *idxoff = mmdata->off[cp];
7476: PetscInt *idxown = mmdata->own[cp];
7477: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7478: const PetscInt *rmap = rmapa[cp];
7479: const PetscInt *cmap = cmapa[cp];
7480: const PetscInt *ii = mm->i;
7481: PetscInt *coi = coo_i + ncoo_o;
7482: PetscInt *coj = coo_j + ncoo_o;
7483: const PetscInt mr = mp[cp]->rmap->n;
7484: const PetscInt rs = C->rmap->rstart;
7485: const PetscInt re = C->rmap->rend;
7486: const PetscInt cs = C->cmap->rstart;
7487: for (i = 0; i < mr; i++) {
7488: const PetscInt *jj = mm->j + ii[i];
7489: const PetscInt gr = rmap[i];
7490: const PetscInt nz = ii[i + 1] - ii[i];
7491: if (gr < rs || gr >= re) { /* this is an offproc row */
7492: for (j = ii[i]; j < ii[i + 1]; j++) {
7493: *coi++ = gr;
7494: *idxoff++ = j;
7495: }
7496: if (!cmapt[cp]) { /* already global */
7497: for (j = 0; j < nz; j++) *coj++ = jj[j];
7498: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7499: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7500: } else { /* offdiag */
7501: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7502: }
7503: ncoo_o += nz;
7504: } else { /* this is a local row */
7505: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7506: }
7507: }
7508: }
7509: mmdata->off[cp + 1] = idxoff;
7510: mmdata->own[cp + 1] = idxown;
7511: }
7513: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7514: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7515: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7516: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7517: ncoo = ncoo_d + ncoo_oown + ncoo2;
7518: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7519: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7520: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7521: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7522: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7523: PetscCall(PetscFree2(coo_i, coo_j));
7524: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7525: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7526: coo_i = coo_i2;
7527: coo_j = coo_j2;
7528: } else { /* no offproc values insertion */
7529: ncoo = ncoo_d;
7530: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7532: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7533: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7534: PetscCall(PetscSFSetUp(mmdata->sf));
7535: }
7536: mmdata->hasoffproc = hasoffproc;
7538: /* gather (i,j) of nonzeros inserted locally */
7539: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7540: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7541: PetscInt *coi = coo_i + ncoo_d;
7542: PetscInt *coj = coo_j + ncoo_d;
7543: const PetscInt *jj = mm->j;
7544: const PetscInt *ii = mm->i;
7545: const PetscInt *cmap = cmapa[cp];
7546: const PetscInt *rmap = rmapa[cp];
7547: const PetscInt mr = mp[cp]->rmap->n;
7548: const PetscInt rs = C->rmap->rstart;
7549: const PetscInt re = C->rmap->rend;
7550: const PetscInt cs = C->cmap->rstart;
7552: if (mptmp[cp]) continue;
7553: if (rmapt[cp] == 1) { /* consecutive rows */
7554: /* fill coo_i */
7555: for (i = 0; i < mr; i++) {
7556: const PetscInt gr = i + rs;
7557: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7558: }
7559: /* fill coo_j */
7560: if (!cmapt[cp]) { /* type-0, already global */
7561: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7562: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7563: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7564: } else { /* type-2, local to global for sparse columns */
7565: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7566: }
7567: ncoo_d += mm->nz;
7568: } else if (rmapt[cp] == 2) { /* sparse rows */
7569: for (i = 0; i < mr; i++) {
7570: const PetscInt *jj = mm->j + ii[i];
7571: const PetscInt gr = rmap[i];
7572: const PetscInt nz = ii[i + 1] - ii[i];
7573: if (gr >= rs && gr < re) { /* local rows */
7574: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7575: if (!cmapt[cp]) { /* type-0, already global */
7576: for (j = 0; j < nz; j++) *coj++ = jj[j];
7577: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7578: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7579: } else { /* type-2, local to global for sparse columns */
7580: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7581: }
7582: ncoo_d += nz;
7583: }
7584: }
7585: }
7586: }
7587: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7588: PetscCall(ISDestroy(&glob));
7589: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7590: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7591: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7592: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7594: /* preallocate with COO data */
7595: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7596: PetscCall(PetscFree2(coo_i, coo_j));
7597: PetscFunctionReturn(PETSC_SUCCESS);
7598: }
7600: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7601: {
7602: Mat_Product *product = mat->product;
7603: #if defined(PETSC_HAVE_DEVICE)
7604: PetscBool match = PETSC_FALSE;
7605: PetscBool usecpu = PETSC_FALSE;
7606: #else
7607: PetscBool match = PETSC_TRUE;
7608: #endif
7610: PetscFunctionBegin;
7611: MatCheckProduct(mat, 1);
7612: #if defined(PETSC_HAVE_DEVICE)
7613: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7614: if (match) { /* we can always fallback to the CPU if requested */
7615: switch (product->type) {
7616: case MATPRODUCT_AB:
7617: if (product->api_user) {
7618: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7619: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7620: PetscOptionsEnd();
7621: } else {
7622: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7623: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7624: PetscOptionsEnd();
7625: }
7626: break;
7627: case MATPRODUCT_AtB:
7628: if (product->api_user) {
7629: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7630: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7631: PetscOptionsEnd();
7632: } else {
7633: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7634: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7635: PetscOptionsEnd();
7636: }
7637: break;
7638: case MATPRODUCT_PtAP:
7639: if (product->api_user) {
7640: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7641: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7642: PetscOptionsEnd();
7643: } else {
7644: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7645: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7646: PetscOptionsEnd();
7647: }
7648: break;
7649: default:
7650: break;
7651: }
7652: match = (PetscBool)!usecpu;
7653: }
7654: #endif
7655: if (match) {
7656: switch (product->type) {
7657: case MATPRODUCT_AB:
7658: case MATPRODUCT_AtB:
7659: case MATPRODUCT_PtAP:
7660: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7661: break;
7662: default:
7663: break;
7664: }
7665: }
7666: /* fallback to MPIAIJ ops */
7667: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7668: PetscFunctionReturn(PETSC_SUCCESS);
7669: }
7671: /*
7672: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7674: n - the number of block indices in cc[]
7675: cc - the block indices (must be large enough to contain the indices)
7676: */
7677: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7678: {
7679: PetscInt cnt = -1, nidx, j;
7680: const PetscInt *idx;
7682: PetscFunctionBegin;
7683: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7684: if (nidx) {
7685: cnt = 0;
7686: cc[cnt] = idx[0] / bs;
7687: for (j = 1; j < nidx; j++) {
7688: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7689: }
7690: }
7691: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7692: *n = cnt + 1;
7693: PetscFunctionReturn(PETSC_SUCCESS);
7694: }
7696: /*
7697: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7699: ncollapsed - the number of block indices
7700: collapsed - the block indices (must be large enough to contain the indices)
7701: */
7702: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7703: {
7704: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7706: PetscFunctionBegin;
7707: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7708: for (i = start + 1; i < start + bs; i++) {
7709: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7710: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7711: cprevtmp = cprev;
7712: cprev = merged;
7713: merged = cprevtmp;
7714: }
7715: *ncollapsed = nprev;
7716: if (collapsed) *collapsed = cprev;
7717: PetscFunctionReturn(PETSC_SUCCESS);
7718: }
7720: /*
7721: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7723: Input Parameter:
7724: . Amat - matrix
7725: - symmetrize - make the result symmetric
7726: + scale - scale with diagonal
7728: Output Parameter:
7729: . a_Gmat - output scalar graph >= 0
7731: */
7732: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, Mat *a_Gmat)
7733: {
7734: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7735: MPI_Comm comm;
7736: Mat Gmat;
7737: PetscBool ismpiaij, isseqaij;
7738: Mat a, b, c;
7739: MatType jtype;
7741: PetscFunctionBegin;
7742: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7743: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7744: PetscCall(MatGetSize(Amat, &MM, &NN));
7745: PetscCall(MatGetBlockSize(Amat, &bs));
7746: nloc = (Iend - Istart) / bs;
7748: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7749: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7750: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7752: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7753: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7754: implementation */
7755: if (bs > 1) {
7756: PetscCall(MatGetType(Amat, &jtype));
7757: PetscCall(MatCreate(comm, &Gmat));
7758: PetscCall(MatSetType(Gmat, jtype));
7759: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7760: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7761: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7762: PetscInt *d_nnz, *o_nnz;
7763: MatScalar *aa, val, *AA;
7764: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7765: if (isseqaij) {
7766: a = Amat;
7767: b = NULL;
7768: } else {
7769: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7770: a = d->A;
7771: b = d->B;
7772: }
7773: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7774: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7775: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7776: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7777: const PetscInt *cols1, *cols2;
7778: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7779: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7780: nnz[brow / bs] = nc2 / bs;
7781: if (nc2 % bs) ok = 0;
7782: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7783: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7784: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7785: if (nc1 != nc2) ok = 0;
7786: else {
7787: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7788: if (cols1[jj] != cols2[jj]) ok = 0;
7789: if (cols1[jj] % bs != jj % bs) ok = 0;
7790: }
7791: }
7792: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7793: }
7794: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7795: if (!ok) {
7796: PetscCall(PetscFree2(d_nnz, o_nnz));
7797: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7798: goto old_bs;
7799: }
7800: }
7801: }
7802: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7803: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7804: PetscCall(PetscFree2(d_nnz, o_nnz));
7805: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7806: // diag
7807: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7808: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7809: ai = aseq->i;
7810: n = ai[brow + 1] - ai[brow];
7811: aj = aseq->j + ai[brow];
7812: for (int k = 0; k < n; k += bs) { // block columns
7813: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7814: val = 0;
7815: for (int ii = 0; ii < bs; ii++) { // rows in block
7816: aa = aseq->a + ai[brow + ii] + k;
7817: for (int jj = 0; jj < bs; jj++) { // columns in block
7818: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7819: }
7820: }
7821: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7822: AA[k / bs] = val;
7823: }
7824: grow = Istart / bs + brow / bs;
7825: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, INSERT_VALUES));
7826: }
7827: // off-diag
7828: if (ismpiaij) {
7829: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7830: const PetscScalar *vals;
7831: const PetscInt *cols, *garray = aij->garray;
7832: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7833: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7834: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7835: for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7836: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7837: AA[k / bs] = 0;
7838: AJ[cidx] = garray[cols[k]] / bs;
7839: }
7840: nc = ncols / bs;
7841: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7842: for (int ii = 0; ii < bs; ii++) { // rows in block
7843: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7844: for (int k = 0; k < ncols; k += bs) {
7845: for (int jj = 0; jj < bs; jj++) { // cols in block
7846: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7847: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7848: }
7849: }
7850: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7851: }
7852: grow = Istart / bs + brow / bs;
7853: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, INSERT_VALUES));
7854: }
7855: }
7856: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7857: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7858: PetscCall(PetscFree2(AA, AJ));
7859: } else {
7860: const PetscScalar *vals;
7861: const PetscInt *idx;
7862: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7863: old_bs:
7864: /*
7865: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7866: */
7867: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7868: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7869: if (isseqaij) {
7870: PetscInt max_d_nnz;
7871: /*
7872: Determine exact preallocation count for (sequential) scalar matrix
7873: */
7874: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7875: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7876: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7877: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7878: PetscCall(PetscFree3(w0, w1, w2));
7879: } else if (ismpiaij) {
7880: Mat Daij, Oaij;
7881: const PetscInt *garray;
7882: PetscInt max_d_nnz;
7883: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7884: /*
7885: Determine exact preallocation count for diagonal block portion of scalar matrix
7886: */
7887: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7888: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7889: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7890: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7891: PetscCall(PetscFree3(w0, w1, w2));
7892: /*
7893: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7894: */
7895: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7896: o_nnz[jj] = 0;
7897: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7898: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7899: o_nnz[jj] += ncols;
7900: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7901: }
7902: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7903: }
7904: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7905: /* get scalar copy (norms) of matrix */
7906: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7907: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7908: PetscCall(PetscFree2(d_nnz, o_nnz));
7909: for (Ii = Istart; Ii < Iend; Ii++) {
7910: PetscInt dest_row = Ii / bs;
7911: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7912: for (jj = 0; jj < ncols; jj++) {
7913: PetscInt dest_col = idx[jj] / bs;
7914: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7915: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7916: }
7917: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7918: }
7919: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7920: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7921: }
7922: } else {
7923: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7924: else {
7925: Gmat = Amat;
7926: PetscCall(PetscObjectReference((PetscObject)Gmat));
7927: }
7928: if (isseqaij) {
7929: a = Gmat;
7930: b = NULL;
7931: } else {
7932: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7933: a = d->A;
7934: b = d->B;
7935: }
7936: if (filter >= 0 || scale) {
7937: /* take absolute value of each entry */
7938: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7939: MatInfo info;
7940: PetscScalar *avals;
7941: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
7942: PetscCall(MatSeqAIJGetArray(c, &avals));
7943: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
7944: PetscCall(MatSeqAIJRestoreArray(c, &avals));
7945: }
7946: }
7947: }
7948: if (symmetrize) {
7949: PetscBool isset, issym;
7950: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
7951: if (!isset || !issym) {
7952: Mat matTrans;
7953: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
7954: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
7955: PetscCall(MatDestroy(&matTrans));
7956: }
7957: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
7958: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
7959: if (scale) {
7960: /* scale c for all diagonal values = 1 or -1 */
7961: Vec diag;
7962: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
7963: PetscCall(MatGetDiagonal(Gmat, diag));
7964: PetscCall(VecReciprocal(diag));
7965: PetscCall(VecSqrtAbs(diag));
7966: PetscCall(MatDiagonalScale(Gmat, diag, diag));
7967: PetscCall(VecDestroy(&diag));
7968: }
7969: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
7971: if (filter >= 0) {
7972: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
7973: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
7974: }
7975: *a_Gmat = Gmat;
7976: PetscFunctionReturn(PETSC_SUCCESS);
7977: }
7979: /*
7980: Special version for direct calls from Fortran
7981: */
7982: #include <petsc/private/fortranimpl.h>
7984: /* Change these macros so can be used in void function */
7985: /* Identical to PetscCallVoid, except it assigns to *_ierr */
7986: #undef PetscCall
7987: #define PetscCall(...) \
7988: do { \
7989: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
7990: if (PetscUnlikely(ierr_msv_mpiaij)) { \
7991: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
7992: return; \
7993: } \
7994: } while (0)
7996: #undef SETERRQ
7997: #define SETERRQ(comm, ierr, ...) \
7998: do { \
7999: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8000: return; \
8001: } while (0)
8003: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8004: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8005: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8006: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8007: #else
8008: #endif
8009: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8010: {
8011: Mat mat = *mmat;
8012: PetscInt m = *mm, n = *mn;
8013: InsertMode addv = *maddv;
8014: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8015: PetscScalar value;
8017: MatCheckPreallocated(mat, 1);
8018: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8019: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8020: {
8021: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8022: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8023: PetscBool roworiented = aij->roworiented;
8025: /* Some Variables required in the macro */
8026: Mat A = aij->A;
8027: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8028: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8029: MatScalar *aa;
8030: PetscBool ignorezeroentries = (((a->ignorezeroentries) && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8031: Mat B = aij->B;
8032: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8033: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8034: MatScalar *ba;
8035: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8036: * cannot use "#if defined" inside a macro. */
8037: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8039: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8040: PetscInt nonew = a->nonew;
8041: MatScalar *ap1, *ap2;
8043: PetscFunctionBegin;
8044: PetscCall(MatSeqAIJGetArray(A, &aa));
8045: PetscCall(MatSeqAIJGetArray(B, &ba));
8046: for (i = 0; i < m; i++) {
8047: if (im[i] < 0) continue;
8048: 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);
8049: if (im[i] >= rstart && im[i] < rend) {
8050: row = im[i] - rstart;
8051: lastcol1 = -1;
8052: rp1 = aj + ai[row];
8053: ap1 = aa + ai[row];
8054: rmax1 = aimax[row];
8055: nrow1 = ailen[row];
8056: low1 = 0;
8057: high1 = nrow1;
8058: lastcol2 = -1;
8059: rp2 = bj + bi[row];
8060: ap2 = ba + bi[row];
8061: rmax2 = bimax[row];
8062: nrow2 = bilen[row];
8063: low2 = 0;
8064: high2 = nrow2;
8066: for (j = 0; j < n; j++) {
8067: if (roworiented) value = v[i * n + j];
8068: else value = v[i + j * m];
8069: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8070: if (in[j] >= cstart && in[j] < cend) {
8071: col = in[j] - cstart;
8072: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8073: } else if (in[j] < 0) continue;
8074: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8075: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8076: } else {
8077: if (mat->was_assembled) {
8078: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8079: #if defined(PETSC_USE_CTABLE)
8080: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8081: col--;
8082: #else
8083: col = aij->colmap[in[j]] - 1;
8084: #endif
8085: if (col < 0 && !((Mat_SeqAIJ *)(aij->A->data))->nonew) {
8086: PetscCall(MatDisAssemble_MPIAIJ(mat));
8087: col = in[j];
8088: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8089: B = aij->B;
8090: b = (Mat_SeqAIJ *)B->data;
8091: bimax = b->imax;
8092: bi = b->i;
8093: bilen = b->ilen;
8094: bj = b->j;
8095: rp2 = bj + bi[row];
8096: ap2 = ba + bi[row];
8097: rmax2 = bimax[row];
8098: nrow2 = bilen[row];
8099: low2 = 0;
8100: high2 = nrow2;
8101: bm = aij->B->rmap->n;
8102: ba = b->a;
8103: inserted = PETSC_FALSE;
8104: }
8105: } else col = in[j];
8106: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8107: }
8108: }
8109: } else if (!aij->donotstash) {
8110: if (roworiented) {
8111: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8112: } else {
8113: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8114: }
8115: }
8116: }
8117: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8118: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8119: }
8120: PetscFunctionReturnVoid();
8121: }
8123: /* Undefining these here since they were redefined from their original definition above! No
8124: * other PETSc functions should be defined past this point, as it is impossible to recover the
8125: * original definitions */
8126: #undef PetscCall
8127: #undef SETERRQ