Actual source code: mpiaij.c
1: #include <../src/mat/impls/aij/mpi/mpiaij.h>
2: #include <petsc/private/vecimpl.h>
3: #include <petsc/private/sfimpl.h>
4: #include <petsc/private/isimpl.h>
5: #include <petscblaslapack.h>
6: #include <petscsf.h>
7: #include <petsc/private/hashmapi.h>
9: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
10: {
11: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
13: PetscFunctionBegin;
14: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
15: PetscCall(MatStashDestroy_Private(&mat->stash));
16: PetscCall(VecDestroy(&aij->diag));
17: PetscCall(MatDestroy(&aij->A));
18: PetscCall(MatDestroy(&aij->B));
19: #if defined(PETSC_USE_CTABLE)
20: PetscCall(PetscHMapIDestroy(&aij->colmap));
21: #else
22: PetscCall(PetscFree(aij->colmap));
23: #endif
24: PetscCall(PetscFree(aij->garray));
25: PetscCall(VecDestroy(&aij->lvec));
26: PetscCall(VecScatterDestroy(&aij->Mvctx));
27: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
28: PetscCall(PetscFree(aij->ld));
30: PetscCall(PetscFree(mat->data));
32: /* may be created by MatCreateMPIAIJSumSeqAIJSymbolic */
33: PetscCall(PetscObjectCompose((PetscObject)mat, "MatMergeSeqsToMPI", NULL));
35: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
36: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatStoreValues_C", NULL));
37: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatRetrieveValues_C", NULL));
38: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatIsTranspose_C", NULL));
39: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocation_C", NULL));
40: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetPreallocation_C", NULL));
41: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocationCSR_C", NULL));
42: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
43: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpibaij_C", NULL));
44: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisbaij_C", NULL));
45: #if defined(PETSC_HAVE_CUDA)
46: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcusparse_C", NULL));
47: #endif
48: #if defined(PETSC_HAVE_HIP)
49: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijhipsparse_C", NULL));
50: #endif
51: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
52: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijkokkos_C", NULL));
53: #endif
54: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
55: #if defined(PETSC_HAVE_ELEMENTAL)
56: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_elemental_C", NULL));
57: #endif
58: #if defined(PETSC_HAVE_SCALAPACK)
59: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_scalapack_C", NULL));
60: #endif
61: #if defined(PETSC_HAVE_HYPRE)
62: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_hypre_C", NULL));
63: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", NULL));
64: #endif
65: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
66: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_is_mpiaij_C", NULL));
67: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpiaij_C", NULL));
68: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetUseScalableIncreaseOverlap_C", NULL));
69: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijperm_C", NULL));
70: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijsell_C", NULL));
71: #if defined(PETSC_HAVE_MKL_SPARSE)
72: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijmkl_C", NULL));
73: #endif
74: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcrl_C", NULL));
75: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
76: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisell_C", NULL));
77: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetPreallocationCOO_C", NULL));
78: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetValuesCOO_C", NULL));
79: PetscFunctionReturn(PETSC_SUCCESS);
80: }
82: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and MatAssemblyEnd_MPI_Hash() */
83: #define TYPE AIJ
84: #define TYPE_AIJ
85: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
86: #undef TYPE
87: #undef TYPE_AIJ
89: static PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
90: {
91: Mat B;
93: PetscFunctionBegin;
94: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
95: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
96: PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
97: PetscCall(MatDestroy(&B));
98: PetscFunctionReturn(PETSC_SUCCESS);
99: }
101: static PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
102: {
103: Mat B;
105: PetscFunctionBegin;
106: PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
107: PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
108: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
109: PetscFunctionReturn(PETSC_SUCCESS);
110: }
112: /*MC
113: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
115: This matrix type is identical to` MATSEQAIJ` when constructed with a single process communicator,
116: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
117: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
118: for communicators controlling multiple processes. It is recommended that you call both of
119: the above preallocation routines for simplicity.
121: Options Database Key:
122: . -mat_type aij - sets the matrix type to `MATAIJ` during a call to `MatSetFromOptions()`
124: Developer Note:
125: Level: beginner
127: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, `MATAIJKOKKOS`,and also automatically switches over to use inodes when
128: enough exist.
130: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`
131: M*/
133: /*MC
134: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
136: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
137: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
138: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
139: for communicators controlling multiple processes. It is recommended that you call both of
140: the above preallocation routines for simplicity.
142: Options Database Key:
143: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`
145: Level: beginner
147: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
148: M*/
150: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
151: {
152: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
154: PetscFunctionBegin;
155: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_VIENNACL)
156: A->boundtocpu = flg;
157: #endif
158: if (a->A) PetscCall(MatBindToCPU(a->A, flg));
159: if (a->B) PetscCall(MatBindToCPU(a->B, flg));
161: /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
162: * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
163: * to differ from the parent matrix. */
164: if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
165: if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));
166: PetscFunctionReturn(PETSC_SUCCESS);
167: }
169: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
170: {
171: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
173: PetscFunctionBegin;
174: if (mat->A) {
175: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
176: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
177: }
178: PetscFunctionReturn(PETSC_SUCCESS);
179: }
181: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
182: {
183: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
184: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
185: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
186: const PetscInt *ia, *ib;
187: const MatScalar *aa, *bb, *aav, *bav;
188: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
189: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
191: PetscFunctionBegin;
192: *keptrows = NULL;
194: ia = a->i;
195: ib = b->i;
196: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
197: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
198: for (i = 0; i < m; i++) {
199: na = ia[i + 1] - ia[i];
200: nb = ib[i + 1] - ib[i];
201: if (!na && !nb) {
202: cnt++;
203: goto ok1;
204: }
205: aa = aav + ia[i];
206: for (j = 0; j < na; j++) {
207: if (aa[j] != 0.0) goto ok1;
208: }
209: bb = PetscSafePointerPlusOffset(bav, ib[i]);
210: for (j = 0; j < nb; j++) {
211: if (bb[j] != 0.0) goto ok1;
212: }
213: cnt++;
214: ok1:;
215: }
216: PetscCall(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
217: if (!n0rows) {
218: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
219: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
220: PetscFunctionReturn(PETSC_SUCCESS);
221: }
222: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
223: cnt = 0;
224: for (i = 0; i < m; i++) {
225: na = ia[i + 1] - ia[i];
226: nb = ib[i + 1] - ib[i];
227: if (!na && !nb) continue;
228: aa = aav + ia[i];
229: for (j = 0; j < na; j++) {
230: if (aa[j] != 0.0) {
231: rows[cnt++] = rstart + i;
232: goto ok2;
233: }
234: }
235: bb = PetscSafePointerPlusOffset(bav, ib[i]);
236: for (j = 0; j < nb; j++) {
237: if (bb[j] != 0.0) {
238: rows[cnt++] = rstart + i;
239: goto ok2;
240: }
241: }
242: ok2:;
243: }
244: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
245: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
246: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
247: PetscFunctionReturn(PETSC_SUCCESS);
248: }
250: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
251: {
252: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
253: PetscBool cong;
255: PetscFunctionBegin;
256: PetscCall(MatHasCongruentLayouts(Y, &cong));
257: if (Y->assembled && cong) {
258: PetscCall(MatDiagonalSet(aij->A, D, is));
259: } else {
260: PetscCall(MatDiagonalSet_Default(Y, D, is));
261: }
262: PetscFunctionReturn(PETSC_SUCCESS);
263: }
265: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
266: {
267: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
268: PetscInt i, rstart, nrows, *rows;
270: PetscFunctionBegin;
271: *zrows = NULL;
272: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
273: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
274: for (i = 0; i < nrows; i++) rows[i] += rstart;
275: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
276: PetscFunctionReturn(PETSC_SUCCESS);
277: }
279: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
280: {
281: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
282: PetscInt i, m, n, *garray = aij->garray;
283: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
284: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
285: PetscReal *work;
286: const PetscScalar *dummy;
288: PetscFunctionBegin;
289: PetscCall(MatGetSize(A, &m, &n));
290: PetscCall(PetscCalloc1(n, &work));
291: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
292: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
293: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
294: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
295: if (type == NORM_2) {
296: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
297: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
298: } else if (type == NORM_1) {
299: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
300: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
301: } else if (type == NORM_INFINITY) {
302: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
303: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), work[garray[b_aij->j[i]]]);
304: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
305: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
306: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
307: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
308: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
309: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
310: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
311: if (type == NORM_INFINITY) {
312: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
313: } else {
314: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
315: }
316: PetscCall(PetscFree(work));
317: if (type == NORM_2) {
318: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
319: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
320: for (i = 0; i < n; i++) reductions[i] /= m;
321: }
322: PetscFunctionReturn(PETSC_SUCCESS);
323: }
325: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
326: {
327: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
328: IS sis, gis;
329: const PetscInt *isis, *igis;
330: PetscInt n, *iis, nsis, ngis, rstart, i;
332: PetscFunctionBegin;
333: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
334: PetscCall(MatFindNonzeroRows(a->B, &gis));
335: PetscCall(ISGetSize(gis, &ngis));
336: PetscCall(ISGetSize(sis, &nsis));
337: PetscCall(ISGetIndices(sis, &isis));
338: PetscCall(ISGetIndices(gis, &igis));
340: PetscCall(PetscMalloc1(ngis + nsis, &iis));
341: PetscCall(PetscArraycpy(iis, igis, ngis));
342: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
343: n = ngis + nsis;
344: PetscCall(PetscSortRemoveDupsInt(&n, iis));
345: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
346: for (i = 0; i < n; i++) iis[i] += rstart;
347: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
349: PetscCall(ISRestoreIndices(sis, &isis));
350: PetscCall(ISRestoreIndices(gis, &igis));
351: PetscCall(ISDestroy(&sis));
352: PetscCall(ISDestroy(&gis));
353: PetscFunctionReturn(PETSC_SUCCESS);
354: }
356: /*
357: Local utility routine that creates a mapping from the global column
358: number to the local number in the off-diagonal part of the local
359: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
360: a slightly higher hash table cost; without it it is not scalable (each processor
361: has an order N integer array but is fast to access.
362: */
363: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
364: {
365: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
366: PetscInt n = aij->B->cmap->n, i;
368: PetscFunctionBegin;
369: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
370: #if defined(PETSC_USE_CTABLE)
371: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
372: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
373: #else
374: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
375: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
376: #endif
377: PetscFunctionReturn(PETSC_SUCCESS);
378: }
380: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
381: do { \
382: if (col <= lastcol1) low1 = 0; \
383: else high1 = nrow1; \
384: lastcol1 = col; \
385: while (high1 - low1 > 5) { \
386: t = (low1 + high1) / 2; \
387: if (rp1[t] > col) high1 = t; \
388: else low1 = t; \
389: } \
390: for (_i = low1; _i < high1; _i++) { \
391: if (rp1[_i] > col) break; \
392: if (rp1[_i] == col) { \
393: if (addv == ADD_VALUES) { \
394: ap1[_i] += value; \
395: /* Not sure LogFlops will slow dow the code or not */ \
396: (void)PetscLogFlops(1.0); \
397: } else ap1[_i] = value; \
398: goto a_noinsert; \
399: } \
400: } \
401: if (value == 0.0 && ignorezeroentries && row != col) { \
402: low1 = 0; \
403: high1 = nrow1; \
404: goto a_noinsert; \
405: } \
406: if (nonew == 1) { \
407: low1 = 0; \
408: high1 = nrow1; \
409: goto a_noinsert; \
410: } \
411: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
412: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
413: N = nrow1++ - 1; \
414: a->nz++; \
415: high1++; \
416: /* shift up all the later entries in this row */ \
417: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
418: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
419: rp1[_i] = col; \
420: ap1[_i] = value; \
421: a_noinsert:; \
422: ailen[row] = nrow1; \
423: } while (0)
425: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
426: do { \
427: if (col <= lastcol2) low2 = 0; \
428: else high2 = nrow2; \
429: lastcol2 = col; \
430: while (high2 - low2 > 5) { \
431: t = (low2 + high2) / 2; \
432: if (rp2[t] > col) high2 = t; \
433: else low2 = t; \
434: } \
435: for (_i = low2; _i < high2; _i++) { \
436: if (rp2[_i] > col) break; \
437: if (rp2[_i] == col) { \
438: if (addv == ADD_VALUES) { \
439: ap2[_i] += value; \
440: (void)PetscLogFlops(1.0); \
441: } else ap2[_i] = value; \
442: goto b_noinsert; \
443: } \
444: } \
445: if (value == 0.0 && ignorezeroentries) { \
446: low2 = 0; \
447: high2 = nrow2; \
448: goto b_noinsert; \
449: } \
450: if (nonew == 1) { \
451: low2 = 0; \
452: high2 = nrow2; \
453: goto b_noinsert; \
454: } \
455: 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); \
456: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
457: N = nrow2++ - 1; \
458: b->nz++; \
459: high2++; \
460: /* shift up all the later entries in this row */ \
461: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
462: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
463: rp2[_i] = col; \
464: ap2[_i] = value; \
465: b_noinsert:; \
466: bilen[row] = nrow2; \
467: } while (0)
469: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
470: {
471: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
472: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
473: PetscInt l, *garray = mat->garray, diag;
474: PetscScalar *aa, *ba;
476: PetscFunctionBegin;
477: /* code only works for square matrices A */
479: /* find size of row to the left of the diagonal part */
480: PetscCall(MatGetOwnershipRange(A, &diag, NULL));
481: row = row - diag;
482: for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
483: if (garray[b->j[b->i[row] + l]] > diag) break;
484: }
485: if (l) {
486: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
487: PetscCall(PetscArraycpy(ba + b->i[row], v, l));
488: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
489: }
491: /* diagonal part */
492: if (a->i[row + 1] - a->i[row]) {
493: PetscCall(MatSeqAIJGetArray(mat->A, &aa));
494: PetscCall(PetscArraycpy(aa + a->i[row], v + l, (a->i[row + 1] - a->i[row])));
495: PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
496: }
498: /* right of diagonal part */
499: if (b->i[row + 1] - b->i[row] - l) {
500: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
501: PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
502: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
503: }
504: PetscFunctionReturn(PETSC_SUCCESS);
505: }
507: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
508: {
509: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
510: PetscScalar value = 0.0;
511: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
512: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
513: PetscBool roworiented = aij->roworiented;
515: /* Some Variables required in the macro */
516: Mat A = aij->A;
517: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
518: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
519: PetscBool ignorezeroentries = a->ignorezeroentries;
520: Mat B = aij->B;
521: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
522: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
523: MatScalar *aa, *ba;
524: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
525: PetscInt nonew;
526: MatScalar *ap1, *ap2;
528: PetscFunctionBegin;
529: PetscCall(MatSeqAIJGetArray(A, &aa));
530: PetscCall(MatSeqAIJGetArray(B, &ba));
531: for (i = 0; i < m; i++) {
532: if (im[i] < 0) continue;
533: 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);
534: if (im[i] >= rstart && im[i] < rend) {
535: row = im[i] - rstart;
536: lastcol1 = -1;
537: rp1 = PetscSafePointerPlusOffset(aj, ai[row]);
538: ap1 = PetscSafePointerPlusOffset(aa, ai[row]);
539: rmax1 = aimax[row];
540: nrow1 = ailen[row];
541: low1 = 0;
542: high1 = nrow1;
543: lastcol2 = -1;
544: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
545: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
546: rmax2 = bimax[row];
547: nrow2 = bilen[row];
548: low2 = 0;
549: high2 = nrow2;
551: for (j = 0; j < n; j++) {
552: if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
553: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
554: if (in[j] >= cstart && in[j] < cend) {
555: col = in[j] - cstart;
556: nonew = a->nonew;
557: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
558: } else if (in[j] < 0) {
559: continue;
560: } else {
561: 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);
562: if (mat->was_assembled) {
563: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
564: #if defined(PETSC_USE_CTABLE)
565: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
566: col--;
567: #else
568: col = aij->colmap[in[j]] - 1;
569: #endif
570: if (col < 0 && !((Mat_SeqAIJ *)aij->B->data)->nonew) { /* col < 0 means in[j] is a new col for B */
571: PetscCall(MatDisAssemble_MPIAIJ(mat)); /* Change aij->B from reduced/local format to expanded/global format */
572: col = in[j];
573: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
574: B = aij->B;
575: b = (Mat_SeqAIJ *)B->data;
576: bimax = b->imax;
577: bi = b->i;
578: bilen = b->ilen;
579: bj = b->j;
580: ba = b->a;
581: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
582: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
583: rmax2 = bimax[row];
584: nrow2 = bilen[row];
585: low2 = 0;
586: high2 = nrow2;
587: bm = aij->B->rmap->n;
588: ba = b->a;
589: } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
590: if (1 == ((Mat_SeqAIJ *)aij->B->data)->nonew) {
591: 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]));
592: } 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]);
593: }
594: } else col = in[j];
595: nonew = b->nonew;
596: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
597: }
598: }
599: } else {
600: 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]);
601: if (!aij->donotstash) {
602: mat->assembled = PETSC_FALSE;
603: if (roworiented) {
604: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
605: } else {
606: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
607: }
608: }
609: }
610: }
611: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
612: PetscCall(MatSeqAIJRestoreArray(B, &ba));
613: PetscFunctionReturn(PETSC_SUCCESS);
614: }
616: /*
617: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
618: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
619: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
620: */
621: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
622: {
623: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
624: Mat A = aij->A; /* diagonal part of the matrix */
625: Mat B = aij->B; /* off-diagonal part of the matrix */
626: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
627: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
628: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
629: PetscInt *ailen = a->ilen, *aj = a->j;
630: PetscInt *bilen = b->ilen, *bj = b->j;
631: PetscInt am = aij->A->rmap->n, j;
632: PetscInt diag_so_far = 0, dnz;
633: PetscInt offd_so_far = 0, onz;
635: PetscFunctionBegin;
636: /* Iterate over all rows of the matrix */
637: for (j = 0; j < am; j++) {
638: dnz = onz = 0;
639: /* Iterate over all non-zero columns of the current row */
640: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
641: /* If column is in the diagonal */
642: if (mat_j[col] >= cstart && mat_j[col] < cend) {
643: aj[diag_so_far++] = mat_j[col] - cstart;
644: dnz++;
645: } else { /* off-diagonal entries */
646: bj[offd_so_far++] = mat_j[col];
647: onz++;
648: }
649: }
650: ailen[j] = dnz;
651: bilen[j] = onz;
652: }
653: PetscFunctionReturn(PETSC_SUCCESS);
654: }
656: /*
657: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
658: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
659: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
660: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
661: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
662: */
663: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
664: {
665: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
666: Mat A = aij->A; /* diagonal part of the matrix */
667: Mat B = aij->B; /* off-diagonal part of the matrix */
668: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
669: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
670: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
671: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
672: PetscInt *ailen = a->ilen, *aj = a->j;
673: PetscInt *bilen = b->ilen, *bj = b->j;
674: PetscInt am = aij->A->rmap->n, j;
675: 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. */
676: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
677: PetscScalar *aa = a->a, *ba = b->a;
679: PetscFunctionBegin;
680: /* Iterate over all rows of the matrix */
681: for (j = 0; j < am; j++) {
682: dnz_row = onz_row = 0;
683: rowstart_offd = full_offd_i[j];
684: rowstart_diag = full_diag_i[j];
685: /* Iterate over all non-zero columns of the current row */
686: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
687: /* If column is in the diagonal */
688: if (mat_j[col] >= cstart && mat_j[col] < cend) {
689: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
690: aa[rowstart_diag + dnz_row] = mat_a[col];
691: dnz_row++;
692: } else { /* off-diagonal entries */
693: bj[rowstart_offd + onz_row] = mat_j[col];
694: ba[rowstart_offd + onz_row] = mat_a[col];
695: onz_row++;
696: }
697: }
698: ailen[j] = dnz_row;
699: bilen[j] = onz_row;
700: }
701: PetscFunctionReturn(PETSC_SUCCESS);
702: }
704: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
705: {
706: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
707: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
708: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
710: PetscFunctionBegin;
711: for (i = 0; i < m; i++) {
712: if (idxm[i] < 0) continue; /* negative row */
713: 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);
714: 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);
715: row = idxm[i] - rstart;
716: for (j = 0; j < n; j++) {
717: if (idxn[j] < 0) continue; /* negative column */
718: 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);
719: if (idxn[j] >= cstart && idxn[j] < cend) {
720: col = idxn[j] - cstart;
721: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
722: } else {
723: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
724: #if defined(PETSC_USE_CTABLE)
725: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
726: col--;
727: #else
728: col = aij->colmap[idxn[j]] - 1;
729: #endif
730: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
731: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
732: }
733: }
734: }
735: PetscFunctionReturn(PETSC_SUCCESS);
736: }
738: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
739: {
740: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
741: PetscInt nstash, reallocs;
743: PetscFunctionBegin;
744: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
746: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
747: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
748: PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
749: PetscFunctionReturn(PETSC_SUCCESS);
750: }
752: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
753: {
754: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
755: PetscMPIInt n;
756: PetscInt i, j, rstart, ncols, flg;
757: PetscInt *row, *col;
758: PetscBool other_disassembled;
759: PetscScalar *val;
761: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
763: PetscFunctionBegin;
764: if (!aij->donotstash && !mat->nooffprocentries) {
765: while (1) {
766: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
767: if (!flg) break;
769: for (i = 0; i < n;) {
770: /* Now identify the consecutive vals belonging to the same row */
771: for (j = i, rstart = row[j]; j < n; j++) {
772: if (row[j] != rstart) break;
773: }
774: if (j < n) ncols = j - i;
775: else ncols = n - i;
776: /* Now assemble all these values with a single function call */
777: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
778: i = j;
779: }
780: }
781: PetscCall(MatStashScatterEnd_Private(&mat->stash));
782: }
783: #if defined(PETSC_HAVE_DEVICE)
784: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
785: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
786: if (mat->boundtocpu) {
787: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
788: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
789: }
790: #endif
791: PetscCall(MatAssemblyBegin(aij->A, mode));
792: PetscCall(MatAssemblyEnd(aij->A, mode));
794: /* determine if any processor has disassembled, if so we must
795: also disassemble ourself, in order that we may reassemble. */
796: /*
797: if nonzero structure of submatrix B cannot change then we know that
798: no processor disassembled thus we can skip this stuff
799: */
800: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
801: PetscCall(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
802: if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
803: PetscCall(MatDisAssemble_MPIAIJ(mat));
804: }
805: }
806: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
807: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
808: #if defined(PETSC_HAVE_DEVICE)
809: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
810: #endif
811: PetscCall(MatAssemblyBegin(aij->B, mode));
812: PetscCall(MatAssemblyEnd(aij->B, mode));
814: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
816: aij->rowvalues = NULL;
818: PetscCall(VecDestroy(&aij->diag));
820: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
821: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
822: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
823: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
824: }
825: #if defined(PETSC_HAVE_DEVICE)
826: mat->offloadmask = PETSC_OFFLOAD_BOTH;
827: #endif
828: PetscFunctionReturn(PETSC_SUCCESS);
829: }
831: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
832: {
833: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
835: PetscFunctionBegin;
836: PetscCall(MatZeroEntries(l->A));
837: PetscCall(MatZeroEntries(l->B));
838: PetscFunctionReturn(PETSC_SUCCESS);
839: }
841: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
842: {
843: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
844: PetscInt *lrows;
845: PetscInt r, len;
846: PetscBool cong;
848: PetscFunctionBegin;
849: /* get locally owned rows */
850: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
851: PetscCall(MatHasCongruentLayouts(A, &cong));
852: /* fix right-hand side if needed */
853: if (x && b) {
854: const PetscScalar *xx;
855: PetscScalar *bb;
857: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
858: PetscCall(VecGetArrayRead(x, &xx));
859: PetscCall(VecGetArray(b, &bb));
860: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
861: PetscCall(VecRestoreArrayRead(x, &xx));
862: PetscCall(VecRestoreArray(b, &bb));
863: }
865: if (diag != 0.0 && cong) {
866: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
867: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
868: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
869: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
870: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
871: PetscInt nnwA, nnwB;
872: PetscBool nnzA, nnzB;
874: nnwA = aijA->nonew;
875: nnwB = aijB->nonew;
876: nnzA = aijA->keepnonzeropattern;
877: nnzB = aijB->keepnonzeropattern;
878: if (!nnzA) {
879: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
880: aijA->nonew = 0;
881: }
882: if (!nnzB) {
883: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
884: aijB->nonew = 0;
885: }
886: /* Must zero here before the next loop */
887: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
888: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
889: for (r = 0; r < len; ++r) {
890: const PetscInt row = lrows[r] + A->rmap->rstart;
891: if (row >= A->cmap->N) continue;
892: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
893: }
894: aijA->nonew = nnwA;
895: aijB->nonew = nnwB;
896: } else {
897: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
898: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
899: }
900: PetscCall(PetscFree(lrows));
901: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
902: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
904: /* only change matrix nonzero state if pattern was allowed to be changed */
905: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
906: PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
907: PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
908: }
909: PetscFunctionReturn(PETSC_SUCCESS);
910: }
912: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
913: {
914: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
915: PetscMPIInt n = A->rmap->n;
916: PetscInt i, j, r, m, len = 0;
917: PetscInt *lrows, *owners = A->rmap->range;
918: PetscMPIInt p = 0;
919: PetscSFNode *rrows;
920: PetscSF sf;
921: const PetscScalar *xx;
922: PetscScalar *bb, *mask, *aij_a;
923: Vec xmask, lmask;
924: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
925: const PetscInt *aj, *ii, *ridx;
926: PetscScalar *aa;
928: PetscFunctionBegin;
929: /* Create SF where leaves are input rows and roots are owned rows */
930: PetscCall(PetscMalloc1(n, &lrows));
931: for (r = 0; r < n; ++r) lrows[r] = -1;
932: PetscCall(PetscMalloc1(N, &rrows));
933: for (r = 0; r < N; ++r) {
934: const PetscInt idx = rows[r];
935: 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);
936: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
937: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
938: }
939: rrows[r].rank = p;
940: rrows[r].index = rows[r] - owners[p];
941: }
942: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
943: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
944: /* Collect flags for rows to be zeroed */
945: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
946: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
947: PetscCall(PetscSFDestroy(&sf));
948: /* Compress and put in row numbers */
949: for (r = 0; r < n; ++r)
950: if (lrows[r] >= 0) lrows[len++] = r;
951: /* zero diagonal part of matrix */
952: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
953: /* handle off-diagonal part of matrix */
954: PetscCall(MatCreateVecs(A, &xmask, NULL));
955: PetscCall(VecDuplicate(l->lvec, &lmask));
956: PetscCall(VecGetArray(xmask, &bb));
957: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
958: PetscCall(VecRestoreArray(xmask, &bb));
959: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
960: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
961: PetscCall(VecDestroy(&xmask));
962: if (x && b) { /* this code is buggy when the row and column layout don't match */
963: PetscBool cong;
965: PetscCall(MatHasCongruentLayouts(A, &cong));
966: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
967: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
968: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
969: PetscCall(VecGetArrayRead(l->lvec, &xx));
970: PetscCall(VecGetArray(b, &bb));
971: }
972: PetscCall(VecGetArray(lmask, &mask));
973: /* remove zeroed rows of off-diagonal matrix */
974: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
975: ii = aij->i;
976: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
977: /* loop over all elements of off process part of matrix zeroing removed columns*/
978: if (aij->compressedrow.use) {
979: m = aij->compressedrow.nrows;
980: ii = aij->compressedrow.i;
981: ridx = aij->compressedrow.rindex;
982: for (i = 0; i < m; i++) {
983: n = ii[i + 1] - ii[i];
984: aj = aij->j + ii[i];
985: aa = aij_a + ii[i];
987: for (j = 0; j < n; j++) {
988: if (PetscAbsScalar(mask[*aj])) {
989: if (b) bb[*ridx] -= *aa * xx[*aj];
990: *aa = 0.0;
991: }
992: aa++;
993: aj++;
994: }
995: ridx++;
996: }
997: } else { /* do not use compressed row format */
998: m = l->B->rmap->n;
999: for (i = 0; i < m; i++) {
1000: n = ii[i + 1] - ii[i];
1001: aj = aij->j + ii[i];
1002: aa = aij_a + ii[i];
1003: for (j = 0; j < n; j++) {
1004: if (PetscAbsScalar(mask[*aj])) {
1005: if (b) bb[i] -= *aa * xx[*aj];
1006: *aa = 0.0;
1007: }
1008: aa++;
1009: aj++;
1010: }
1011: }
1012: }
1013: if (x && b) {
1014: PetscCall(VecRestoreArray(b, &bb));
1015: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1016: }
1017: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1018: PetscCall(VecRestoreArray(lmask, &mask));
1019: PetscCall(VecDestroy(&lmask));
1020: PetscCall(PetscFree(lrows));
1022: /* only change matrix nonzero state if pattern was allowed to be changed */
1023: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1024: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1025: PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1026: }
1027: PetscFunctionReturn(PETSC_SUCCESS);
1028: }
1030: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1031: {
1032: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1033: PetscInt nt;
1034: VecScatter Mvctx = a->Mvctx;
1036: PetscFunctionBegin;
1037: PetscCall(VecGetLocalSize(xx, &nt));
1038: 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);
1039: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1040: PetscUseTypeMethod(a->A, mult, xx, yy);
1041: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1042: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1043: PetscFunctionReturn(PETSC_SUCCESS);
1044: }
1046: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1047: {
1048: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1050: PetscFunctionBegin;
1051: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1052: PetscFunctionReturn(PETSC_SUCCESS);
1053: }
1055: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1056: {
1057: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1058: VecScatter Mvctx = a->Mvctx;
1060: PetscFunctionBegin;
1061: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1062: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1063: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1064: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1065: PetscFunctionReturn(PETSC_SUCCESS);
1066: }
1068: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1069: {
1070: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1072: PetscFunctionBegin;
1073: /* do nondiagonal part */
1074: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1075: /* do local part */
1076: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1077: /* add partial results together */
1078: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1079: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1080: PetscFunctionReturn(PETSC_SUCCESS);
1081: }
1083: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1084: {
1085: MPI_Comm comm;
1086: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1087: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1088: IS Me, Notme;
1089: PetscInt M, N, first, last, *notme, i;
1090: PetscBool lf;
1091: PetscMPIInt size;
1093: PetscFunctionBegin;
1094: /* Easy test: symmetric diagonal block */
1095: PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1096: PetscCall(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1097: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1098: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1099: PetscCallMPI(MPI_Comm_size(comm, &size));
1100: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1102: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1103: PetscCall(MatGetSize(Amat, &M, &N));
1104: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1105: PetscCall(PetscMalloc1(N - last + first, ¬me));
1106: for (i = 0; i < first; i++) notme[i] = i;
1107: for (i = last; i < M; i++) notme[i - last + first] = i;
1108: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1109: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1110: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1111: Aoff = Aoffs[0];
1112: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1113: Boff = Boffs[0];
1114: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1115: PetscCall(MatDestroyMatrices(1, &Aoffs));
1116: PetscCall(MatDestroyMatrices(1, &Boffs));
1117: PetscCall(ISDestroy(&Me));
1118: PetscCall(ISDestroy(&Notme));
1119: PetscCall(PetscFree(notme));
1120: PetscFunctionReturn(PETSC_SUCCESS);
1121: }
1123: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1124: {
1125: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1127: PetscFunctionBegin;
1128: /* do nondiagonal part */
1129: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1130: /* do local part */
1131: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1132: /* add partial results together */
1133: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1134: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1135: PetscFunctionReturn(PETSC_SUCCESS);
1136: }
1138: /*
1139: This only works correctly for square matrices where the subblock A->A is the
1140: diagonal block
1141: */
1142: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1143: {
1144: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1146: PetscFunctionBegin;
1147: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1148: 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");
1149: PetscCall(MatGetDiagonal(a->A, v));
1150: PetscFunctionReturn(PETSC_SUCCESS);
1151: }
1153: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1154: {
1155: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1157: PetscFunctionBegin;
1158: PetscCall(MatScale(a->A, aa));
1159: PetscCall(MatScale(a->B, aa));
1160: PetscFunctionReturn(PETSC_SUCCESS);
1161: }
1163: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1164: {
1165: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1166: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1167: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1168: const PetscInt *garray = aij->garray;
1169: const PetscScalar *aa, *ba;
1170: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1171: PetscInt64 nz, hnz;
1172: PetscInt *rowlens;
1173: PetscInt *colidxs;
1174: PetscScalar *matvals;
1175: PetscMPIInt rank;
1177: PetscFunctionBegin;
1178: PetscCall(PetscViewerSetUp(viewer));
1180: M = mat->rmap->N;
1181: N = mat->cmap->N;
1182: m = mat->rmap->n;
1183: rs = mat->rmap->rstart;
1184: cs = mat->cmap->rstart;
1185: nz = A->nz + B->nz;
1187: /* write matrix header */
1188: header[0] = MAT_FILE_CLASSID;
1189: header[1] = M;
1190: header[2] = N;
1191: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1192: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1193: if (rank == 0) {
1194: if (hnz > PETSC_MAX_INT) header[3] = PETSC_MAX_INT;
1195: else header[3] = (PetscInt)hnz;
1196: }
1197: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1199: /* fill in and store row lengths */
1200: PetscCall(PetscMalloc1(m, &rowlens));
1201: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1202: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1203: PetscCall(PetscFree(rowlens));
1205: /* fill in and store column indices */
1206: PetscCall(PetscMalloc1(nz, &colidxs));
1207: for (cnt = 0, i = 0; i < m; i++) {
1208: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1209: if (garray[B->j[jb]] > cs) break;
1210: colidxs[cnt++] = garray[B->j[jb]];
1211: }
1212: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1213: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1214: }
1215: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1216: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1217: PetscCall(PetscFree(colidxs));
1219: /* fill in and store nonzero values */
1220: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1221: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1222: PetscCall(PetscMalloc1(nz, &matvals));
1223: for (cnt = 0, i = 0; i < m; i++) {
1224: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1225: if (garray[B->j[jb]] > cs) break;
1226: matvals[cnt++] = ba[jb];
1227: }
1228: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1229: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1230: }
1231: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1232: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1233: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1234: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1235: PetscCall(PetscFree(matvals));
1237: /* write block size option to the viewer's .info file */
1238: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1239: PetscFunctionReturn(PETSC_SUCCESS);
1240: }
1242: #include <petscdraw.h>
1243: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1244: {
1245: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1246: PetscMPIInt rank = aij->rank, size = aij->size;
1247: PetscBool isdraw, iascii, isbinary;
1248: PetscViewer sviewer;
1249: PetscViewerFormat format;
1251: PetscFunctionBegin;
1252: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1253: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1254: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1255: if (iascii) {
1256: PetscCall(PetscViewerGetFormat(viewer, &format));
1257: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1258: PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1259: PetscCall(PetscMalloc1(size, &nz));
1260: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1261: for (i = 0; i < (PetscInt)size; i++) {
1262: nmax = PetscMax(nmax, nz[i]);
1263: nmin = PetscMin(nmin, nz[i]);
1264: navg += nz[i];
1265: }
1266: PetscCall(PetscFree(nz));
1267: navg = navg / size;
1268: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1269: PetscFunctionReturn(PETSC_SUCCESS);
1270: }
1271: PetscCall(PetscViewerGetFormat(viewer, &format));
1272: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1273: MatInfo info;
1274: PetscInt *inodes = NULL;
1276: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1277: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1278: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1279: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1280: if (!inodes) {
1281: 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,
1282: (double)info.memory));
1283: } else {
1284: 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,
1285: (double)info.memory));
1286: }
1287: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1288: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1289: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1290: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1291: PetscCall(PetscViewerFlush(viewer));
1292: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1293: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1294: PetscCall(VecScatterView(aij->Mvctx, viewer));
1295: PetscFunctionReturn(PETSC_SUCCESS);
1296: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1297: PetscInt inodecount, inodelimit, *inodes;
1298: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1299: if (inodes) {
1300: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1301: } else {
1302: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1303: }
1304: PetscFunctionReturn(PETSC_SUCCESS);
1305: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1306: PetscFunctionReturn(PETSC_SUCCESS);
1307: }
1308: } else if (isbinary) {
1309: if (size == 1) {
1310: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1311: PetscCall(MatView(aij->A, viewer));
1312: } else {
1313: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1314: }
1315: PetscFunctionReturn(PETSC_SUCCESS);
1316: } else if (iascii && size == 1) {
1317: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1318: PetscCall(MatView(aij->A, viewer));
1319: PetscFunctionReturn(PETSC_SUCCESS);
1320: } else if (isdraw) {
1321: PetscDraw draw;
1322: PetscBool isnull;
1323: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1324: PetscCall(PetscDrawIsNull(draw, &isnull));
1325: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1326: }
1328: { /* assemble the entire matrix onto first processor */
1329: Mat A = NULL, Av;
1330: IS isrow, iscol;
1332: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1333: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1334: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1335: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1336: /* The commented code uses MatCreateSubMatrices instead */
1337: /*
1338: Mat *AA, A = NULL, Av;
1339: IS isrow,iscol;
1341: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1342: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1343: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1344: if (rank == 0) {
1345: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1346: A = AA[0];
1347: Av = AA[0];
1348: }
1349: PetscCall(MatDestroySubMatrices(1,&AA));
1350: */
1351: PetscCall(ISDestroy(&iscol));
1352: PetscCall(ISDestroy(&isrow));
1353: /*
1354: Everyone has to call to draw the matrix since the graphics waits are
1355: synchronized across all processors that share the PetscDraw object
1356: */
1357: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1358: if (rank == 0) {
1359: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1360: PetscCall(MatView_SeqAIJ(Av, sviewer));
1361: }
1362: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1363: PetscCall(MatDestroy(&A));
1364: }
1365: PetscFunctionReturn(PETSC_SUCCESS);
1366: }
1368: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1369: {
1370: PetscBool iascii, isdraw, issocket, isbinary;
1372: PetscFunctionBegin;
1373: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1374: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1375: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1376: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1377: if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1378: PetscFunctionReturn(PETSC_SUCCESS);
1379: }
1381: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1382: {
1383: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1384: Vec bb1 = NULL;
1385: PetscBool hasop;
1387: PetscFunctionBegin;
1388: if (flag == SOR_APPLY_UPPER) {
1389: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1390: PetscFunctionReturn(PETSC_SUCCESS);
1391: }
1393: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1395: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1396: if (flag & SOR_ZERO_INITIAL_GUESS) {
1397: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1398: its--;
1399: }
1401: while (its--) {
1402: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1403: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1405: /* update rhs: bb1 = bb - B*x */
1406: PetscCall(VecScale(mat->lvec, -1.0));
1407: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1409: /* local sweep */
1410: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1411: }
1412: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1413: if (flag & SOR_ZERO_INITIAL_GUESS) {
1414: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1415: its--;
1416: }
1417: while (its--) {
1418: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1419: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1421: /* update rhs: bb1 = bb - B*x */
1422: PetscCall(VecScale(mat->lvec, -1.0));
1423: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1425: /* local sweep */
1426: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1427: }
1428: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1429: if (flag & SOR_ZERO_INITIAL_GUESS) {
1430: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1431: its--;
1432: }
1433: while (its--) {
1434: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1435: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1437: /* update rhs: bb1 = bb - B*x */
1438: PetscCall(VecScale(mat->lvec, -1.0));
1439: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1441: /* local sweep */
1442: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1443: }
1444: } else if (flag & SOR_EISENSTAT) {
1445: Vec xx1;
1447: PetscCall(VecDuplicate(bb, &xx1));
1448: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1450: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1451: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1452: if (!mat->diag) {
1453: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1454: PetscCall(MatGetDiagonal(matin, mat->diag));
1455: }
1456: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1457: if (hasop) {
1458: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1459: } else {
1460: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1461: }
1462: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1464: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1466: /* local sweep */
1467: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1468: PetscCall(VecAXPY(xx, 1.0, xx1));
1469: PetscCall(VecDestroy(&xx1));
1470: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1472: PetscCall(VecDestroy(&bb1));
1474: matin->factorerrortype = mat->A->factorerrortype;
1475: PetscFunctionReturn(PETSC_SUCCESS);
1476: }
1478: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1479: {
1480: Mat aA, aB, Aperm;
1481: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1482: PetscScalar *aa, *ba;
1483: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1484: PetscSF rowsf, sf;
1485: IS parcolp = NULL;
1486: PetscBool done;
1488: PetscFunctionBegin;
1489: PetscCall(MatGetLocalSize(A, &m, &n));
1490: PetscCall(ISGetIndices(rowp, &rwant));
1491: PetscCall(ISGetIndices(colp, &cwant));
1492: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1494: /* Invert row permutation to find out where my rows should go */
1495: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1496: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1497: PetscCall(PetscSFSetFromOptions(rowsf));
1498: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1499: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1500: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1502: /* Invert column permutation to find out where my columns should go */
1503: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1504: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1505: PetscCall(PetscSFSetFromOptions(sf));
1506: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1507: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1508: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1509: PetscCall(PetscSFDestroy(&sf));
1511: PetscCall(ISRestoreIndices(rowp, &rwant));
1512: PetscCall(ISRestoreIndices(colp, &cwant));
1513: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1515: /* Find out where my gcols should go */
1516: PetscCall(MatGetSize(aB, NULL, &ng));
1517: PetscCall(PetscMalloc1(ng, &gcdest));
1518: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1519: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1520: PetscCall(PetscSFSetFromOptions(sf));
1521: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1522: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1523: PetscCall(PetscSFDestroy(&sf));
1525: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1526: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1527: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1528: for (i = 0; i < m; i++) {
1529: PetscInt row = rdest[i];
1530: PetscMPIInt rowner;
1531: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1532: for (j = ai[i]; j < ai[i + 1]; j++) {
1533: PetscInt col = cdest[aj[j]];
1534: PetscMPIInt cowner;
1535: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1536: if (rowner == cowner) dnnz[i]++;
1537: else onnz[i]++;
1538: }
1539: for (j = bi[i]; j < bi[i + 1]; j++) {
1540: PetscInt col = gcdest[bj[j]];
1541: PetscMPIInt cowner;
1542: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1543: if (rowner == cowner) dnnz[i]++;
1544: else onnz[i]++;
1545: }
1546: }
1547: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1548: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1549: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1550: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1551: PetscCall(PetscSFDestroy(&rowsf));
1553: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1554: PetscCall(MatSeqAIJGetArray(aA, &aa));
1555: PetscCall(MatSeqAIJGetArray(aB, &ba));
1556: for (i = 0; i < m; i++) {
1557: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1558: PetscInt j0, rowlen;
1559: rowlen = ai[i + 1] - ai[i];
1560: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1561: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1562: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1563: }
1564: rowlen = bi[i + 1] - bi[i];
1565: for (j0 = j = 0; j < rowlen; j0 = j) {
1566: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1567: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1568: }
1569: }
1570: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1571: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1572: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1573: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1574: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1575: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1576: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1577: PetscCall(PetscFree3(work, rdest, cdest));
1578: PetscCall(PetscFree(gcdest));
1579: if (parcolp) PetscCall(ISDestroy(&colp));
1580: *B = Aperm;
1581: PetscFunctionReturn(PETSC_SUCCESS);
1582: }
1584: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1585: {
1586: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1588: PetscFunctionBegin;
1589: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1590: if (ghosts) *ghosts = aij->garray;
1591: PetscFunctionReturn(PETSC_SUCCESS);
1592: }
1594: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1595: {
1596: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1597: Mat A = mat->A, B = mat->B;
1598: PetscLogDouble isend[5], irecv[5];
1600: PetscFunctionBegin;
1601: info->block_size = 1.0;
1602: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1604: isend[0] = info->nz_used;
1605: isend[1] = info->nz_allocated;
1606: isend[2] = info->nz_unneeded;
1607: isend[3] = info->memory;
1608: isend[4] = info->mallocs;
1610: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1612: isend[0] += info->nz_used;
1613: isend[1] += info->nz_allocated;
1614: isend[2] += info->nz_unneeded;
1615: isend[3] += info->memory;
1616: isend[4] += info->mallocs;
1617: if (flag == MAT_LOCAL) {
1618: info->nz_used = isend[0];
1619: info->nz_allocated = isend[1];
1620: info->nz_unneeded = isend[2];
1621: info->memory = isend[3];
1622: info->mallocs = isend[4];
1623: } else if (flag == MAT_GLOBAL_MAX) {
1624: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1626: info->nz_used = irecv[0];
1627: info->nz_allocated = irecv[1];
1628: info->nz_unneeded = irecv[2];
1629: info->memory = irecv[3];
1630: info->mallocs = irecv[4];
1631: } else if (flag == MAT_GLOBAL_SUM) {
1632: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1634: info->nz_used = irecv[0];
1635: info->nz_allocated = irecv[1];
1636: info->nz_unneeded = irecv[2];
1637: info->memory = irecv[3];
1638: info->mallocs = irecv[4];
1639: }
1640: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1641: info->fill_ratio_needed = 0;
1642: info->factor_mallocs = 0;
1643: PetscFunctionReturn(PETSC_SUCCESS);
1644: }
1646: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1647: {
1648: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1650: PetscFunctionBegin;
1651: switch (op) {
1652: case MAT_NEW_NONZERO_LOCATIONS:
1653: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1654: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1655: case MAT_KEEP_NONZERO_PATTERN:
1656: case MAT_NEW_NONZERO_LOCATION_ERR:
1657: case MAT_USE_INODES:
1658: case MAT_IGNORE_ZERO_ENTRIES:
1659: case MAT_FORM_EXPLICIT_TRANSPOSE:
1660: MatCheckPreallocated(A, 1);
1661: PetscCall(MatSetOption(a->A, op, flg));
1662: PetscCall(MatSetOption(a->B, op, flg));
1663: break;
1664: case MAT_ROW_ORIENTED:
1665: MatCheckPreallocated(A, 1);
1666: a->roworiented = flg;
1668: PetscCall(MatSetOption(a->A, op, flg));
1669: PetscCall(MatSetOption(a->B, op, flg));
1670: break;
1671: case MAT_FORCE_DIAGONAL_ENTRIES:
1672: case MAT_SORTED_FULL:
1673: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1674: break;
1675: case MAT_IGNORE_OFF_PROC_ENTRIES:
1676: a->donotstash = flg;
1677: break;
1678: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1679: case MAT_SPD:
1680: case MAT_SYMMETRIC:
1681: case MAT_STRUCTURALLY_SYMMETRIC:
1682: case MAT_HERMITIAN:
1683: case MAT_SYMMETRY_ETERNAL:
1684: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1685: case MAT_SPD_ETERNAL:
1686: /* if the diagonal matrix is square it inherits some of the properties above */
1687: break;
1688: case MAT_SUBMAT_SINGLEIS:
1689: A->submat_singleis = flg;
1690: break;
1691: case MAT_STRUCTURE_ONLY:
1692: /* The option is handled directly by MatSetOption() */
1693: break;
1694: default:
1695: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1696: }
1697: PetscFunctionReturn(PETSC_SUCCESS);
1698: }
1700: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1701: {
1702: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1703: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1704: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1705: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1706: PetscInt *cmap, *idx_p;
1708: PetscFunctionBegin;
1709: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1710: mat->getrowactive = PETSC_TRUE;
1712: if (!mat->rowvalues && (idx || v)) {
1713: /*
1714: allocate enough space to hold information from the longest row.
1715: */
1716: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1717: PetscInt max = 1, tmp;
1718: for (i = 0; i < matin->rmap->n; i++) {
1719: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1720: if (max < tmp) max = tmp;
1721: }
1722: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1723: }
1725: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1726: lrow = row - rstart;
1728: pvA = &vworkA;
1729: pcA = &cworkA;
1730: pvB = &vworkB;
1731: pcB = &cworkB;
1732: if (!v) {
1733: pvA = NULL;
1734: pvB = NULL;
1735: }
1736: if (!idx) {
1737: pcA = NULL;
1738: if (!v) pcB = NULL;
1739: }
1740: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1741: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1742: nztot = nzA + nzB;
1744: cmap = mat->garray;
1745: if (v || idx) {
1746: if (nztot) {
1747: /* Sort by increasing column numbers, assuming A and B already sorted */
1748: PetscInt imark = -1;
1749: if (v) {
1750: *v = v_p = mat->rowvalues;
1751: for (i = 0; i < nzB; i++) {
1752: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1753: else break;
1754: }
1755: imark = i;
1756: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1757: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1758: }
1759: if (idx) {
1760: *idx = idx_p = mat->rowindices;
1761: if (imark > -1) {
1762: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1763: } else {
1764: for (i = 0; i < nzB; i++) {
1765: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1766: else break;
1767: }
1768: imark = i;
1769: }
1770: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1771: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1772: }
1773: } else {
1774: if (idx) *idx = NULL;
1775: if (v) *v = NULL;
1776: }
1777: }
1778: *nz = nztot;
1779: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1780: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1781: PetscFunctionReturn(PETSC_SUCCESS);
1782: }
1784: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1785: {
1786: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1788: PetscFunctionBegin;
1789: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1790: aij->getrowactive = PETSC_FALSE;
1791: PetscFunctionReturn(PETSC_SUCCESS);
1792: }
1794: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1795: {
1796: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1797: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1798: PetscInt i, j, cstart = mat->cmap->rstart;
1799: PetscReal sum = 0.0;
1800: const MatScalar *v, *amata, *bmata;
1802: PetscFunctionBegin;
1803: if (aij->size == 1) {
1804: PetscCall(MatNorm(aij->A, type, norm));
1805: } else {
1806: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1807: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1808: if (type == NORM_FROBENIUS) {
1809: v = amata;
1810: for (i = 0; i < amat->nz; i++) {
1811: sum += PetscRealPart(PetscConj(*v) * (*v));
1812: v++;
1813: }
1814: v = bmata;
1815: for (i = 0; i < bmat->nz; i++) {
1816: sum += PetscRealPart(PetscConj(*v) * (*v));
1817: v++;
1818: }
1819: PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1820: *norm = PetscSqrtReal(*norm);
1821: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1822: } else if (type == NORM_1) { /* max column norm */
1823: PetscReal *tmp, *tmp2;
1824: PetscInt *jj, *garray = aij->garray;
1825: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1826: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1827: *norm = 0.0;
1828: v = amata;
1829: jj = amat->j;
1830: for (j = 0; j < amat->nz; j++) {
1831: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1832: v++;
1833: }
1834: v = bmata;
1835: jj = bmat->j;
1836: for (j = 0; j < bmat->nz; j++) {
1837: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1838: v++;
1839: }
1840: PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1841: for (j = 0; j < mat->cmap->N; j++) {
1842: if (tmp2[j] > *norm) *norm = tmp2[j];
1843: }
1844: PetscCall(PetscFree(tmp));
1845: PetscCall(PetscFree(tmp2));
1846: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1847: } else if (type == NORM_INFINITY) { /* max row norm */
1848: PetscReal ntemp = 0.0;
1849: for (j = 0; j < aij->A->rmap->n; j++) {
1850: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1851: sum = 0.0;
1852: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1853: sum += PetscAbsScalar(*v);
1854: v++;
1855: }
1856: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1857: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1858: sum += PetscAbsScalar(*v);
1859: v++;
1860: }
1861: if (sum > ntemp) ntemp = sum;
1862: }
1863: PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1864: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1865: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1866: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1867: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1868: }
1869: PetscFunctionReturn(PETSC_SUCCESS);
1870: }
1872: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1873: {
1874: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1875: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1876: 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;
1877: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1878: Mat B, A_diag, *B_diag;
1879: const MatScalar *pbv, *bv;
1881: PetscFunctionBegin;
1882: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1883: ma = A->rmap->n;
1884: na = A->cmap->n;
1885: mb = a->B->rmap->n;
1886: nb = a->B->cmap->n;
1887: ai = Aloc->i;
1888: aj = Aloc->j;
1889: bi = Bloc->i;
1890: bj = Bloc->j;
1891: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1892: PetscInt *d_nnz, *g_nnz, *o_nnz;
1893: PetscSFNode *oloc;
1894: PETSC_UNUSED PetscSF sf;
1896: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1897: /* compute d_nnz for preallocation */
1898: PetscCall(PetscArrayzero(d_nnz, na));
1899: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1900: /* compute local off-diagonal contributions */
1901: PetscCall(PetscArrayzero(g_nnz, nb));
1902: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1903: /* map those to global */
1904: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1905: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1906: PetscCall(PetscSFSetFromOptions(sf));
1907: PetscCall(PetscArrayzero(o_nnz, na));
1908: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1909: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1910: PetscCall(PetscSFDestroy(&sf));
1912: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1913: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1914: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1915: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1916: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1917: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1918: } else {
1919: B = *matout;
1920: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1921: }
1923: b = (Mat_MPIAIJ *)B->data;
1924: A_diag = a->A;
1925: B_diag = &b->A;
1926: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1927: A_diag_ncol = A_diag->cmap->N;
1928: B_diag_ilen = sub_B_diag->ilen;
1929: B_diag_i = sub_B_diag->i;
1931: /* Set ilen for diagonal of B */
1932: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1934: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1935: very quickly (=without using MatSetValues), because all writes are local. */
1936: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1937: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1939: /* copy over the B part */
1940: PetscCall(PetscMalloc1(bi[mb], &cols));
1941: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1942: pbv = bv;
1943: row = A->rmap->rstart;
1944: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1945: cols_tmp = cols;
1946: for (i = 0; i < mb; i++) {
1947: ncol = bi[i + 1] - bi[i];
1948: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1949: row++;
1950: if (pbv) pbv += ncol;
1951: if (cols_tmp) cols_tmp += ncol;
1952: }
1953: PetscCall(PetscFree(cols));
1954: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1956: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1957: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1958: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1959: *matout = B;
1960: } else {
1961: PetscCall(MatHeaderMerge(A, &B));
1962: }
1963: PetscFunctionReturn(PETSC_SUCCESS);
1964: }
1966: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1967: {
1968: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1969: Mat a = aij->A, b = aij->B;
1970: PetscInt s1, s2, s3;
1972: PetscFunctionBegin;
1973: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1974: if (rr) {
1975: PetscCall(VecGetLocalSize(rr, &s1));
1976: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1977: /* Overlap communication with computation. */
1978: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1979: }
1980: if (ll) {
1981: PetscCall(VecGetLocalSize(ll, &s1));
1982: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1983: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1984: }
1985: /* scale the diagonal block */
1986: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1988: if (rr) {
1989: /* Do a scatter end and then right scale the off-diagonal block */
1990: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1991: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1992: }
1993: PetscFunctionReturn(PETSC_SUCCESS);
1994: }
1996: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1997: {
1998: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2000: PetscFunctionBegin;
2001: PetscCall(MatSetUnfactored(a->A));
2002: PetscFunctionReturn(PETSC_SUCCESS);
2003: }
2005: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2006: {
2007: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2008: Mat a, b, c, d;
2009: PetscBool flg;
2011: PetscFunctionBegin;
2012: a = matA->A;
2013: b = matA->B;
2014: c = matB->A;
2015: d = matB->B;
2017: PetscCall(MatEqual(a, c, &flg));
2018: if (flg) PetscCall(MatEqual(b, d, &flg));
2019: PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2020: PetscFunctionReturn(PETSC_SUCCESS);
2021: }
2023: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2024: {
2025: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2026: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2028: PetscFunctionBegin;
2029: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2030: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2031: /* because of the column compression in the off-processor part of the matrix a->B,
2032: the number of columns in a->B and b->B may be different, hence we cannot call
2033: the MatCopy() directly on the two parts. If need be, we can provide a more
2034: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2035: then copying the submatrices */
2036: PetscCall(MatCopy_Basic(A, B, str));
2037: } else {
2038: PetscCall(MatCopy(a->A, b->A, str));
2039: PetscCall(MatCopy(a->B, b->B, str));
2040: }
2041: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2042: PetscFunctionReturn(PETSC_SUCCESS);
2043: }
2045: /*
2046: Computes the number of nonzeros per row needed for preallocation when X and Y
2047: have different nonzero structure.
2048: */
2049: 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)
2050: {
2051: PetscInt i, j, k, nzx, nzy;
2053: PetscFunctionBegin;
2054: /* Set the number of nonzeros in the new matrix */
2055: for (i = 0; i < m; i++) {
2056: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2057: nzx = xi[i + 1] - xi[i];
2058: nzy = yi[i + 1] - yi[i];
2059: nnz[i] = 0;
2060: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2061: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2062: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2063: nnz[i]++;
2064: }
2065: for (; k < nzy; k++) nnz[i]++;
2066: }
2067: PetscFunctionReturn(PETSC_SUCCESS);
2068: }
2070: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2071: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2072: {
2073: PetscInt m = Y->rmap->N;
2074: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2075: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2077: PetscFunctionBegin;
2078: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2079: PetscFunctionReturn(PETSC_SUCCESS);
2080: }
2082: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2083: {
2084: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2086: PetscFunctionBegin;
2087: if (str == SAME_NONZERO_PATTERN) {
2088: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2089: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2090: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2091: PetscCall(MatAXPY_Basic(Y, a, X, str));
2092: } else {
2093: Mat B;
2094: PetscInt *nnz_d, *nnz_o;
2096: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2097: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2098: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2099: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2100: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2101: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2102: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2103: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2104: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2105: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2106: PetscCall(MatHeaderMerge(Y, &B));
2107: PetscCall(PetscFree(nnz_d));
2108: PetscCall(PetscFree(nnz_o));
2109: }
2110: PetscFunctionReturn(PETSC_SUCCESS);
2111: }
2113: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2115: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2116: {
2117: PetscFunctionBegin;
2118: if (PetscDefined(USE_COMPLEX)) {
2119: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2121: PetscCall(MatConjugate_SeqAIJ(aij->A));
2122: PetscCall(MatConjugate_SeqAIJ(aij->B));
2123: }
2124: PetscFunctionReturn(PETSC_SUCCESS);
2125: }
2127: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2128: {
2129: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2131: PetscFunctionBegin;
2132: PetscCall(MatRealPart(a->A));
2133: PetscCall(MatRealPart(a->B));
2134: PetscFunctionReturn(PETSC_SUCCESS);
2135: }
2137: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2138: {
2139: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2141: PetscFunctionBegin;
2142: PetscCall(MatImaginaryPart(a->A));
2143: PetscCall(MatImaginaryPart(a->B));
2144: PetscFunctionReturn(PETSC_SUCCESS);
2145: }
2147: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2148: {
2149: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2150: PetscInt i, *idxb = NULL, m = A->rmap->n;
2151: PetscScalar *va, *vv;
2152: Vec vB, vA;
2153: const PetscScalar *vb;
2155: PetscFunctionBegin;
2156: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2157: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2159: PetscCall(VecGetArrayWrite(vA, &va));
2160: if (idx) {
2161: for (i = 0; i < m; i++) {
2162: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2163: }
2164: }
2166: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2167: PetscCall(PetscMalloc1(m, &idxb));
2168: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2170: PetscCall(VecGetArrayWrite(v, &vv));
2171: PetscCall(VecGetArrayRead(vB, &vb));
2172: for (i = 0; i < m; i++) {
2173: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2174: vv[i] = vb[i];
2175: if (idx) idx[i] = a->garray[idxb[i]];
2176: } else {
2177: vv[i] = va[i];
2178: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2179: }
2180: }
2181: PetscCall(VecRestoreArrayWrite(vA, &vv));
2182: PetscCall(VecRestoreArrayWrite(vA, &va));
2183: PetscCall(VecRestoreArrayRead(vB, &vb));
2184: PetscCall(PetscFree(idxb));
2185: PetscCall(VecDestroy(&vA));
2186: PetscCall(VecDestroy(&vB));
2187: PetscFunctionReturn(PETSC_SUCCESS);
2188: }
2190: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2191: {
2192: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2193: Vec vB, vA;
2195: PetscFunctionBegin;
2196: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2197: PetscCall(MatGetRowSumAbs(a->A, vA));
2198: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2199: PetscCall(MatGetRowSumAbs(a->B, vB));
2200: PetscCall(VecAXPY(vA, 1.0, vB));
2201: PetscCall(VecDestroy(&vB));
2202: PetscCall(VecCopy(vA, v));
2203: PetscCall(VecDestroy(&vA));
2204: PetscFunctionReturn(PETSC_SUCCESS);
2205: }
2207: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2208: {
2209: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2210: PetscInt m = A->rmap->n, n = A->cmap->n;
2211: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2212: PetscInt *cmap = mat->garray;
2213: PetscInt *diagIdx, *offdiagIdx;
2214: Vec diagV, offdiagV;
2215: PetscScalar *a, *diagA, *offdiagA;
2216: const PetscScalar *ba, *bav;
2217: PetscInt r, j, col, ncols, *bi, *bj;
2218: Mat B = mat->B;
2219: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2221: PetscFunctionBegin;
2222: /* When a process holds entire A and other processes have no entry */
2223: if (A->cmap->N == n) {
2224: PetscCall(VecGetArrayWrite(v, &diagA));
2225: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2226: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2227: PetscCall(VecDestroy(&diagV));
2228: PetscCall(VecRestoreArrayWrite(v, &diagA));
2229: PetscFunctionReturn(PETSC_SUCCESS);
2230: } else if (n == 0) {
2231: if (m) {
2232: PetscCall(VecGetArrayWrite(v, &a));
2233: for (r = 0; r < m; r++) {
2234: a[r] = 0.0;
2235: if (idx) idx[r] = -1;
2236: }
2237: PetscCall(VecRestoreArrayWrite(v, &a));
2238: }
2239: PetscFunctionReturn(PETSC_SUCCESS);
2240: }
2242: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2243: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2244: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2245: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2247: /* Get offdiagIdx[] for implicit 0.0 */
2248: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2249: ba = bav;
2250: bi = b->i;
2251: bj = b->j;
2252: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2253: for (r = 0; r < m; r++) {
2254: ncols = bi[r + 1] - bi[r];
2255: if (ncols == A->cmap->N - n) { /* Brow is dense */
2256: offdiagA[r] = *ba;
2257: offdiagIdx[r] = cmap[0];
2258: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2259: offdiagA[r] = 0.0;
2261: /* Find first hole in the cmap */
2262: for (j = 0; j < ncols; j++) {
2263: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2264: if (col > j && j < cstart) {
2265: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2266: break;
2267: } else if (col > j + n && j >= cstart) {
2268: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2269: break;
2270: }
2271: }
2272: if (j == ncols && ncols < A->cmap->N - n) {
2273: /* a hole is outside compressed Bcols */
2274: if (ncols == 0) {
2275: if (cstart) {
2276: offdiagIdx[r] = 0;
2277: } else offdiagIdx[r] = cend;
2278: } else { /* ncols > 0 */
2279: offdiagIdx[r] = cmap[ncols - 1] + 1;
2280: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2281: }
2282: }
2283: }
2285: for (j = 0; j < ncols; j++) {
2286: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2287: offdiagA[r] = *ba;
2288: offdiagIdx[r] = cmap[*bj];
2289: }
2290: ba++;
2291: bj++;
2292: }
2293: }
2295: PetscCall(VecGetArrayWrite(v, &a));
2296: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2297: for (r = 0; r < m; ++r) {
2298: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2299: a[r] = diagA[r];
2300: if (idx) idx[r] = cstart + diagIdx[r];
2301: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2302: a[r] = diagA[r];
2303: if (idx) {
2304: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2305: idx[r] = cstart + diagIdx[r];
2306: } else idx[r] = offdiagIdx[r];
2307: }
2308: } else {
2309: a[r] = offdiagA[r];
2310: if (idx) idx[r] = offdiagIdx[r];
2311: }
2312: }
2313: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2314: PetscCall(VecRestoreArrayWrite(v, &a));
2315: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2316: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2317: PetscCall(VecDestroy(&diagV));
2318: PetscCall(VecDestroy(&offdiagV));
2319: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2320: PetscFunctionReturn(PETSC_SUCCESS);
2321: }
2323: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2324: {
2325: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2326: PetscInt m = A->rmap->n, n = A->cmap->n;
2327: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2328: PetscInt *cmap = mat->garray;
2329: PetscInt *diagIdx, *offdiagIdx;
2330: Vec diagV, offdiagV;
2331: PetscScalar *a, *diagA, *offdiagA;
2332: const PetscScalar *ba, *bav;
2333: PetscInt r, j, col, ncols, *bi, *bj;
2334: Mat B = mat->B;
2335: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2337: PetscFunctionBegin;
2338: /* When a process holds entire A and other processes have no entry */
2339: if (A->cmap->N == n) {
2340: PetscCall(VecGetArrayWrite(v, &diagA));
2341: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2342: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2343: PetscCall(VecDestroy(&diagV));
2344: PetscCall(VecRestoreArrayWrite(v, &diagA));
2345: PetscFunctionReturn(PETSC_SUCCESS);
2346: } else if (n == 0) {
2347: if (m) {
2348: PetscCall(VecGetArrayWrite(v, &a));
2349: for (r = 0; r < m; r++) {
2350: a[r] = PETSC_MAX_REAL;
2351: if (idx) idx[r] = -1;
2352: }
2353: PetscCall(VecRestoreArrayWrite(v, &a));
2354: }
2355: PetscFunctionReturn(PETSC_SUCCESS);
2356: }
2358: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2359: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2360: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2361: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2363: /* Get offdiagIdx[] for implicit 0.0 */
2364: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2365: ba = bav;
2366: bi = b->i;
2367: bj = b->j;
2368: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2369: for (r = 0; r < m; r++) {
2370: ncols = bi[r + 1] - bi[r];
2371: if (ncols == A->cmap->N - n) { /* Brow is dense */
2372: offdiagA[r] = *ba;
2373: offdiagIdx[r] = cmap[0];
2374: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2375: offdiagA[r] = 0.0;
2377: /* Find first hole in the cmap */
2378: for (j = 0; j < ncols; j++) {
2379: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2380: if (col > j && j < cstart) {
2381: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2382: break;
2383: } else if (col > j + n && j >= cstart) {
2384: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2385: break;
2386: }
2387: }
2388: if (j == ncols && ncols < A->cmap->N - n) {
2389: /* a hole is outside compressed Bcols */
2390: if (ncols == 0) {
2391: if (cstart) {
2392: offdiagIdx[r] = 0;
2393: } else offdiagIdx[r] = cend;
2394: } else { /* ncols > 0 */
2395: offdiagIdx[r] = cmap[ncols - 1] + 1;
2396: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2397: }
2398: }
2399: }
2401: for (j = 0; j < ncols; j++) {
2402: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2403: offdiagA[r] = *ba;
2404: offdiagIdx[r] = cmap[*bj];
2405: }
2406: ba++;
2407: bj++;
2408: }
2409: }
2411: PetscCall(VecGetArrayWrite(v, &a));
2412: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2413: for (r = 0; r < m; ++r) {
2414: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2415: a[r] = diagA[r];
2416: if (idx) idx[r] = cstart + diagIdx[r];
2417: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2418: a[r] = diagA[r];
2419: if (idx) {
2420: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2421: idx[r] = cstart + diagIdx[r];
2422: } else idx[r] = offdiagIdx[r];
2423: }
2424: } else {
2425: a[r] = offdiagA[r];
2426: if (idx) idx[r] = offdiagIdx[r];
2427: }
2428: }
2429: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2430: PetscCall(VecRestoreArrayWrite(v, &a));
2431: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2432: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2433: PetscCall(VecDestroy(&diagV));
2434: PetscCall(VecDestroy(&offdiagV));
2435: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2436: PetscFunctionReturn(PETSC_SUCCESS);
2437: }
2439: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2440: {
2441: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2442: PetscInt m = A->rmap->n, n = A->cmap->n;
2443: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2444: PetscInt *cmap = mat->garray;
2445: PetscInt *diagIdx, *offdiagIdx;
2446: Vec diagV, offdiagV;
2447: PetscScalar *a, *diagA, *offdiagA;
2448: const PetscScalar *ba, *bav;
2449: PetscInt r, j, col, ncols, *bi, *bj;
2450: Mat B = mat->B;
2451: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2453: PetscFunctionBegin;
2454: /* When a process holds entire A and other processes have no entry */
2455: if (A->cmap->N == n) {
2456: PetscCall(VecGetArrayWrite(v, &diagA));
2457: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2458: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2459: PetscCall(VecDestroy(&diagV));
2460: PetscCall(VecRestoreArrayWrite(v, &diagA));
2461: PetscFunctionReturn(PETSC_SUCCESS);
2462: } else if (n == 0) {
2463: if (m) {
2464: PetscCall(VecGetArrayWrite(v, &a));
2465: for (r = 0; r < m; r++) {
2466: a[r] = PETSC_MIN_REAL;
2467: if (idx) idx[r] = -1;
2468: }
2469: PetscCall(VecRestoreArrayWrite(v, &a));
2470: }
2471: PetscFunctionReturn(PETSC_SUCCESS);
2472: }
2474: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2475: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2476: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2477: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2479: /* Get offdiagIdx[] for implicit 0.0 */
2480: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2481: ba = bav;
2482: bi = b->i;
2483: bj = b->j;
2484: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2485: for (r = 0; r < m; r++) {
2486: ncols = bi[r + 1] - bi[r];
2487: if (ncols == A->cmap->N - n) { /* Brow is dense */
2488: offdiagA[r] = *ba;
2489: offdiagIdx[r] = cmap[0];
2490: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2491: offdiagA[r] = 0.0;
2493: /* Find first hole in the cmap */
2494: for (j = 0; j < ncols; j++) {
2495: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2496: if (col > j && j < cstart) {
2497: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2498: break;
2499: } else if (col > j + n && j >= cstart) {
2500: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2501: break;
2502: }
2503: }
2504: if (j == ncols && ncols < A->cmap->N - n) {
2505: /* a hole is outside compressed Bcols */
2506: if (ncols == 0) {
2507: if (cstart) {
2508: offdiagIdx[r] = 0;
2509: } else offdiagIdx[r] = cend;
2510: } else { /* ncols > 0 */
2511: offdiagIdx[r] = cmap[ncols - 1] + 1;
2512: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2513: }
2514: }
2515: }
2517: for (j = 0; j < ncols; j++) {
2518: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2519: offdiagA[r] = *ba;
2520: offdiagIdx[r] = cmap[*bj];
2521: }
2522: ba++;
2523: bj++;
2524: }
2525: }
2527: PetscCall(VecGetArrayWrite(v, &a));
2528: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2529: for (r = 0; r < m; ++r) {
2530: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2531: a[r] = diagA[r];
2532: if (idx) idx[r] = cstart + diagIdx[r];
2533: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2534: a[r] = diagA[r];
2535: if (idx) {
2536: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2537: idx[r] = cstart + diagIdx[r];
2538: } else idx[r] = offdiagIdx[r];
2539: }
2540: } else {
2541: a[r] = offdiagA[r];
2542: if (idx) idx[r] = offdiagIdx[r];
2543: }
2544: }
2545: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2546: PetscCall(VecRestoreArrayWrite(v, &a));
2547: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2548: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2549: PetscCall(VecDestroy(&diagV));
2550: PetscCall(VecDestroy(&offdiagV));
2551: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2552: PetscFunctionReturn(PETSC_SUCCESS);
2553: }
2555: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2556: {
2557: Mat *dummy;
2559: PetscFunctionBegin;
2560: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2561: *newmat = *dummy;
2562: PetscCall(PetscFree(dummy));
2563: PetscFunctionReturn(PETSC_SUCCESS);
2564: }
2566: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2567: {
2568: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2570: PetscFunctionBegin;
2571: PetscCall(MatInvertBlockDiagonal(a->A, values));
2572: A->factorerrortype = a->A->factorerrortype;
2573: PetscFunctionReturn(PETSC_SUCCESS);
2574: }
2576: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2577: {
2578: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2580: PetscFunctionBegin;
2581: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2582: PetscCall(MatSetRandom(aij->A, rctx));
2583: if (x->assembled) {
2584: PetscCall(MatSetRandom(aij->B, rctx));
2585: } else {
2586: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2587: }
2588: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2589: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2590: PetscFunctionReturn(PETSC_SUCCESS);
2591: }
2593: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2594: {
2595: PetscFunctionBegin;
2596: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2597: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2598: PetscFunctionReturn(PETSC_SUCCESS);
2599: }
2601: /*@
2602: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2604: Not Collective
2606: Input Parameter:
2607: . A - the matrix
2609: Output Parameter:
2610: . nz - the number of nonzeros
2612: Level: advanced
2614: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2615: @*/
2616: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2617: {
2618: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2619: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2620: PetscBool isaij;
2622: PetscFunctionBegin;
2623: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2624: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2625: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2626: PetscFunctionReturn(PETSC_SUCCESS);
2627: }
2629: /*@
2630: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2632: Collective
2634: Input Parameters:
2635: + A - the matrix
2636: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2638: Level: advanced
2640: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2641: @*/
2642: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2643: {
2644: PetscFunctionBegin;
2645: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2646: PetscFunctionReturn(PETSC_SUCCESS);
2647: }
2649: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2650: {
2651: PetscBool sc = PETSC_FALSE, flg;
2653: PetscFunctionBegin;
2654: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2655: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2656: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2657: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2658: PetscOptionsHeadEnd();
2659: PetscFunctionReturn(PETSC_SUCCESS);
2660: }
2662: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2663: {
2664: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2665: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2667: PetscFunctionBegin;
2668: if (!Y->preallocated) {
2669: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2670: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2671: PetscInt nonew = aij->nonew;
2672: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2673: aij->nonew = nonew;
2674: }
2675: PetscCall(MatShift_Basic(Y, a));
2676: PetscFunctionReturn(PETSC_SUCCESS);
2677: }
2679: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2680: {
2681: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2683: PetscFunctionBegin;
2684: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2685: PetscCall(MatMissingDiagonal(a->A, missing, d));
2686: if (d) {
2687: PetscInt rstart;
2688: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2689: *d += rstart;
2690: }
2691: PetscFunctionReturn(PETSC_SUCCESS);
2692: }
2694: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2695: {
2696: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2698: PetscFunctionBegin;
2699: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2700: PetscFunctionReturn(PETSC_SUCCESS);
2701: }
2703: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2704: {
2705: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2707: PetscFunctionBegin;
2708: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2709: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2710: PetscFunctionReturn(PETSC_SUCCESS);
2711: }
2713: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2714: MatGetRow_MPIAIJ,
2715: MatRestoreRow_MPIAIJ,
2716: MatMult_MPIAIJ,
2717: /* 4*/ MatMultAdd_MPIAIJ,
2718: MatMultTranspose_MPIAIJ,
2719: MatMultTransposeAdd_MPIAIJ,
2720: NULL,
2721: NULL,
2722: NULL,
2723: /*10*/ NULL,
2724: NULL,
2725: NULL,
2726: MatSOR_MPIAIJ,
2727: MatTranspose_MPIAIJ,
2728: /*15*/ MatGetInfo_MPIAIJ,
2729: MatEqual_MPIAIJ,
2730: MatGetDiagonal_MPIAIJ,
2731: MatDiagonalScale_MPIAIJ,
2732: MatNorm_MPIAIJ,
2733: /*20*/ MatAssemblyBegin_MPIAIJ,
2734: MatAssemblyEnd_MPIAIJ,
2735: MatSetOption_MPIAIJ,
2736: MatZeroEntries_MPIAIJ,
2737: /*24*/ MatZeroRows_MPIAIJ,
2738: NULL,
2739: NULL,
2740: NULL,
2741: NULL,
2742: /*29*/ MatSetUp_MPI_Hash,
2743: NULL,
2744: NULL,
2745: MatGetDiagonalBlock_MPIAIJ,
2746: NULL,
2747: /*34*/ MatDuplicate_MPIAIJ,
2748: NULL,
2749: NULL,
2750: NULL,
2751: NULL,
2752: /*39*/ MatAXPY_MPIAIJ,
2753: MatCreateSubMatrices_MPIAIJ,
2754: MatIncreaseOverlap_MPIAIJ,
2755: MatGetValues_MPIAIJ,
2756: MatCopy_MPIAIJ,
2757: /*44*/ MatGetRowMax_MPIAIJ,
2758: MatScale_MPIAIJ,
2759: MatShift_MPIAIJ,
2760: MatDiagonalSet_MPIAIJ,
2761: MatZeroRowsColumns_MPIAIJ,
2762: /*49*/ MatSetRandom_MPIAIJ,
2763: MatGetRowIJ_MPIAIJ,
2764: MatRestoreRowIJ_MPIAIJ,
2765: NULL,
2766: NULL,
2767: /*54*/ MatFDColoringCreate_MPIXAIJ,
2768: NULL,
2769: MatSetUnfactored_MPIAIJ,
2770: MatPermute_MPIAIJ,
2771: NULL,
2772: /*59*/ MatCreateSubMatrix_MPIAIJ,
2773: MatDestroy_MPIAIJ,
2774: MatView_MPIAIJ,
2775: NULL,
2776: NULL,
2777: /*64*/ NULL,
2778: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2779: NULL,
2780: NULL,
2781: NULL,
2782: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2783: MatGetRowMinAbs_MPIAIJ,
2784: NULL,
2785: NULL,
2786: NULL,
2787: NULL,
2788: /*75*/ MatFDColoringApply_AIJ,
2789: MatSetFromOptions_MPIAIJ,
2790: NULL,
2791: NULL,
2792: MatFindZeroDiagonals_MPIAIJ,
2793: /*80*/ NULL,
2794: NULL,
2795: NULL,
2796: /*83*/ MatLoad_MPIAIJ,
2797: NULL,
2798: NULL,
2799: NULL,
2800: NULL,
2801: NULL,
2802: /*89*/ NULL,
2803: NULL,
2804: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2805: NULL,
2806: NULL,
2807: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2808: NULL,
2809: NULL,
2810: NULL,
2811: MatBindToCPU_MPIAIJ,
2812: /*99*/ MatProductSetFromOptions_MPIAIJ,
2813: NULL,
2814: NULL,
2815: MatConjugate_MPIAIJ,
2816: NULL,
2817: /*104*/ MatSetValuesRow_MPIAIJ,
2818: MatRealPart_MPIAIJ,
2819: MatImaginaryPart_MPIAIJ,
2820: NULL,
2821: NULL,
2822: /*109*/ NULL,
2823: NULL,
2824: MatGetRowMin_MPIAIJ,
2825: NULL,
2826: MatMissingDiagonal_MPIAIJ,
2827: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2828: NULL,
2829: MatGetGhosts_MPIAIJ,
2830: NULL,
2831: NULL,
2832: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2833: NULL,
2834: NULL,
2835: NULL,
2836: MatGetMultiProcBlock_MPIAIJ,
2837: /*124*/ MatFindNonzeroRows_MPIAIJ,
2838: MatGetColumnReductions_MPIAIJ,
2839: MatInvertBlockDiagonal_MPIAIJ,
2840: MatInvertVariableBlockDiagonal_MPIAIJ,
2841: MatCreateSubMatricesMPI_MPIAIJ,
2842: /*129*/ NULL,
2843: NULL,
2844: NULL,
2845: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2846: NULL,
2847: /*134*/ NULL,
2848: NULL,
2849: NULL,
2850: NULL,
2851: NULL,
2852: /*139*/ MatSetBlockSizes_MPIAIJ,
2853: NULL,
2854: NULL,
2855: MatFDColoringSetUp_MPIXAIJ,
2856: MatFindOffBlockDiagonalEntries_MPIAIJ,
2857: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2858: /*145*/ NULL,
2859: NULL,
2860: NULL,
2861: MatCreateGraph_Simple_AIJ,
2862: NULL,
2863: /*150*/ NULL,
2864: MatEliminateZeros_MPIAIJ,
2865: MatGetRowSumAbs_MPIAIJ,
2866: NULL};
2868: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2869: {
2870: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2872: PetscFunctionBegin;
2873: PetscCall(MatStoreValues(aij->A));
2874: PetscCall(MatStoreValues(aij->B));
2875: PetscFunctionReturn(PETSC_SUCCESS);
2876: }
2878: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2879: {
2880: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2882: PetscFunctionBegin;
2883: PetscCall(MatRetrieveValues(aij->A));
2884: PetscCall(MatRetrieveValues(aij->B));
2885: PetscFunctionReturn(PETSC_SUCCESS);
2886: }
2888: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2889: {
2890: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2891: PetscMPIInt size;
2893: PetscFunctionBegin;
2894: if (B->hash_active) {
2895: B->ops[0] = b->cops;
2896: B->hash_active = PETSC_FALSE;
2897: }
2898: PetscCall(PetscLayoutSetUp(B->rmap));
2899: PetscCall(PetscLayoutSetUp(B->cmap));
2901: #if defined(PETSC_USE_CTABLE)
2902: PetscCall(PetscHMapIDestroy(&b->colmap));
2903: #else
2904: PetscCall(PetscFree(b->colmap));
2905: #endif
2906: PetscCall(PetscFree(b->garray));
2907: PetscCall(VecDestroy(&b->lvec));
2908: PetscCall(VecScatterDestroy(&b->Mvctx));
2910: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2912: MatSeqXAIJGetOptions_Private(b->B);
2913: PetscCall(MatDestroy(&b->B));
2914: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2915: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2916: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2917: PetscCall(MatSetType(b->B, MATSEQAIJ));
2918: MatSeqXAIJRestoreOptions_Private(b->B);
2920: MatSeqXAIJGetOptions_Private(b->A);
2921: PetscCall(MatDestroy(&b->A));
2922: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2923: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2924: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2925: PetscCall(MatSetType(b->A, MATSEQAIJ));
2926: MatSeqXAIJRestoreOptions_Private(b->A);
2928: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2929: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2930: B->preallocated = PETSC_TRUE;
2931: B->was_assembled = PETSC_FALSE;
2932: B->assembled = PETSC_FALSE;
2933: PetscFunctionReturn(PETSC_SUCCESS);
2934: }
2936: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2937: {
2938: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2940: PetscFunctionBegin;
2942: PetscCall(PetscLayoutSetUp(B->rmap));
2943: PetscCall(PetscLayoutSetUp(B->cmap));
2945: #if defined(PETSC_USE_CTABLE)
2946: PetscCall(PetscHMapIDestroy(&b->colmap));
2947: #else
2948: PetscCall(PetscFree(b->colmap));
2949: #endif
2950: PetscCall(PetscFree(b->garray));
2951: PetscCall(VecDestroy(&b->lvec));
2952: PetscCall(VecScatterDestroy(&b->Mvctx));
2954: PetscCall(MatResetPreallocation(b->A));
2955: PetscCall(MatResetPreallocation(b->B));
2956: B->preallocated = PETSC_TRUE;
2957: B->was_assembled = PETSC_FALSE;
2958: B->assembled = PETSC_FALSE;
2959: PetscFunctionReturn(PETSC_SUCCESS);
2960: }
2962: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2963: {
2964: Mat mat;
2965: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2967: PetscFunctionBegin;
2968: *newmat = NULL;
2969: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2970: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2971: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2972: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2973: a = (Mat_MPIAIJ *)mat->data;
2975: mat->factortype = matin->factortype;
2976: mat->assembled = matin->assembled;
2977: mat->insertmode = NOT_SET_VALUES;
2979: a->size = oldmat->size;
2980: a->rank = oldmat->rank;
2981: a->donotstash = oldmat->donotstash;
2982: a->roworiented = oldmat->roworiented;
2983: a->rowindices = NULL;
2984: a->rowvalues = NULL;
2985: a->getrowactive = PETSC_FALSE;
2987: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2988: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2989: if (matin->hash_active) {
2990: PetscCall(MatSetUp(mat));
2991: } else {
2992: mat->preallocated = matin->preallocated;
2993: if (oldmat->colmap) {
2994: #if defined(PETSC_USE_CTABLE)
2995: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2996: #else
2997: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2998: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
2999: #endif
3000: } else a->colmap = NULL;
3001: if (oldmat->garray) {
3002: PetscInt len;
3003: len = oldmat->B->cmap->n;
3004: PetscCall(PetscMalloc1(len + 1, &a->garray));
3005: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3006: } else a->garray = NULL;
3008: /* It may happen MatDuplicate is called with a non-assembled matrix
3009: In fact, MatDuplicate only requires the matrix to be preallocated
3010: This may happen inside a DMCreateMatrix_Shell */
3011: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3012: if (oldmat->Mvctx) {
3013: a->Mvctx = oldmat->Mvctx;
3014: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3015: }
3016: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3017: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3018: }
3019: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3020: *newmat = mat;
3021: PetscFunctionReturn(PETSC_SUCCESS);
3022: }
3024: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3025: {
3026: PetscBool isbinary, ishdf5;
3028: PetscFunctionBegin;
3031: /* force binary viewer to load .info file if it has not yet done so */
3032: PetscCall(PetscViewerSetUp(viewer));
3033: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3034: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3035: if (isbinary) {
3036: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3037: } else if (ishdf5) {
3038: #if defined(PETSC_HAVE_HDF5)
3039: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3040: #else
3041: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3042: #endif
3043: } else {
3044: 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);
3045: }
3046: PetscFunctionReturn(PETSC_SUCCESS);
3047: }
3049: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3050: {
3051: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3052: PetscInt *rowidxs, *colidxs;
3053: PetscScalar *matvals;
3055: PetscFunctionBegin;
3056: PetscCall(PetscViewerSetUp(viewer));
3058: /* read in matrix header */
3059: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3060: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3061: M = header[1];
3062: N = header[2];
3063: nz = header[3];
3064: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3065: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3066: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3068: /* set block sizes from the viewer's .info file */
3069: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3070: /* set global sizes if not set already */
3071: if (mat->rmap->N < 0) mat->rmap->N = M;
3072: if (mat->cmap->N < 0) mat->cmap->N = N;
3073: PetscCall(PetscLayoutSetUp(mat->rmap));
3074: PetscCall(PetscLayoutSetUp(mat->cmap));
3076: /* check if the matrix sizes are correct */
3077: PetscCall(MatGetSize(mat, &rows, &cols));
3078: 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);
3080: /* read in row lengths and build row indices */
3081: PetscCall(MatGetLocalSize(mat, &m, NULL));
3082: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3083: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3084: rowidxs[0] = 0;
3085: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3086: if (nz != PETSC_MAX_INT) {
3087: PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3088: 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);
3089: }
3091: /* read in column indices and matrix values */
3092: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3093: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3094: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3095: /* store matrix indices and values */
3096: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3097: PetscCall(PetscFree(rowidxs));
3098: PetscCall(PetscFree2(colidxs, matvals));
3099: PetscFunctionReturn(PETSC_SUCCESS);
3100: }
3102: /* Not scalable because of ISAllGather() unless getting all columns. */
3103: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3104: {
3105: IS iscol_local;
3106: PetscBool isstride;
3107: PetscMPIInt lisstride = 0, gisstride;
3109: PetscFunctionBegin;
3110: /* check if we are grabbing all columns*/
3111: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3113: if (isstride) {
3114: PetscInt start, len, mstart, mlen;
3115: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3116: PetscCall(ISGetLocalSize(iscol, &len));
3117: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3118: if (mstart == start && mlen - mstart == len) lisstride = 1;
3119: }
3121: PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3122: if (gisstride) {
3123: PetscInt N;
3124: PetscCall(MatGetSize(mat, NULL, &N));
3125: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3126: PetscCall(ISSetIdentity(iscol_local));
3127: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3128: } else {
3129: PetscInt cbs;
3130: PetscCall(ISGetBlockSize(iscol, &cbs));
3131: PetscCall(ISAllGather(iscol, &iscol_local));
3132: PetscCall(ISSetBlockSize(iscol_local, cbs));
3133: }
3135: *isseq = iscol_local;
3136: PetscFunctionReturn(PETSC_SUCCESS);
3137: }
3139: /*
3140: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3141: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3143: Input Parameters:
3144: + mat - matrix
3145: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3146: i.e., mat->rstart <= isrow[i] < mat->rend
3147: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3148: i.e., mat->cstart <= iscol[i] < mat->cend
3150: Output Parameters:
3151: + isrow_d - sequential row index set for retrieving mat->A
3152: . iscol_d - sequential column index set for retrieving mat->A
3153: . iscol_o - sequential column index set for retrieving mat->B
3154: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3155: */
3156: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3157: {
3158: Vec x, cmap;
3159: const PetscInt *is_idx;
3160: PetscScalar *xarray, *cmaparray;
3161: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3162: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3163: Mat B = a->B;
3164: Vec lvec = a->lvec, lcmap;
3165: PetscInt i, cstart, cend, Bn = B->cmap->N;
3166: MPI_Comm comm;
3167: VecScatter Mvctx = a->Mvctx;
3169: PetscFunctionBegin;
3170: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3171: PetscCall(ISGetLocalSize(iscol, &ncols));
3173: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3174: PetscCall(MatCreateVecs(mat, &x, NULL));
3175: PetscCall(VecSet(x, -1.0));
3176: PetscCall(VecDuplicate(x, &cmap));
3177: PetscCall(VecSet(cmap, -1.0));
3179: /* Get start indices */
3180: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3181: isstart -= ncols;
3182: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3184: PetscCall(ISGetIndices(iscol, &is_idx));
3185: PetscCall(VecGetArray(x, &xarray));
3186: PetscCall(VecGetArray(cmap, &cmaparray));
3187: PetscCall(PetscMalloc1(ncols, &idx));
3188: for (i = 0; i < ncols; i++) {
3189: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3190: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3191: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3192: }
3193: PetscCall(VecRestoreArray(x, &xarray));
3194: PetscCall(VecRestoreArray(cmap, &cmaparray));
3195: PetscCall(ISRestoreIndices(iscol, &is_idx));
3197: /* Get iscol_d */
3198: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3199: PetscCall(ISGetBlockSize(iscol, &i));
3200: PetscCall(ISSetBlockSize(*iscol_d, i));
3202: /* Get isrow_d */
3203: PetscCall(ISGetLocalSize(isrow, &m));
3204: rstart = mat->rmap->rstart;
3205: PetscCall(PetscMalloc1(m, &idx));
3206: PetscCall(ISGetIndices(isrow, &is_idx));
3207: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3208: PetscCall(ISRestoreIndices(isrow, &is_idx));
3210: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3211: PetscCall(ISGetBlockSize(isrow, &i));
3212: PetscCall(ISSetBlockSize(*isrow_d, i));
3214: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3215: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3216: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3218: PetscCall(VecDuplicate(lvec, &lcmap));
3220: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3221: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3223: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3224: /* off-process column indices */
3225: count = 0;
3226: PetscCall(PetscMalloc1(Bn, &idx));
3227: PetscCall(PetscMalloc1(Bn, &cmap1));
3229: PetscCall(VecGetArray(lvec, &xarray));
3230: PetscCall(VecGetArray(lcmap, &cmaparray));
3231: for (i = 0; i < Bn; i++) {
3232: if (PetscRealPart(xarray[i]) > -1.0) {
3233: idx[count] = i; /* local column index in off-diagonal part B */
3234: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3235: count++;
3236: }
3237: }
3238: PetscCall(VecRestoreArray(lvec, &xarray));
3239: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3241: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3242: /* cannot ensure iscol_o has same blocksize as iscol! */
3244: PetscCall(PetscFree(idx));
3245: *garray = cmap1;
3247: PetscCall(VecDestroy(&x));
3248: PetscCall(VecDestroy(&cmap));
3249: PetscCall(VecDestroy(&lcmap));
3250: PetscFunctionReturn(PETSC_SUCCESS);
3251: }
3253: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3254: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3255: {
3256: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3257: Mat M = NULL;
3258: MPI_Comm comm;
3259: IS iscol_d, isrow_d, iscol_o;
3260: Mat Asub = NULL, Bsub = NULL;
3261: PetscInt n;
3263: PetscFunctionBegin;
3264: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3266: if (call == MAT_REUSE_MATRIX) {
3267: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3268: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3269: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3271: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3272: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3274: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3275: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3277: /* Update diagonal and off-diagonal portions of submat */
3278: asub = (Mat_MPIAIJ *)(*submat)->data;
3279: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3280: PetscCall(ISGetLocalSize(iscol_o, &n));
3281: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3282: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3283: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3285: } else { /* call == MAT_INITIAL_MATRIX) */
3286: const PetscInt *garray;
3287: PetscInt BsubN;
3289: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3290: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3292: /* Create local submatrices Asub and Bsub */
3293: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3294: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3296: /* Create submatrix M */
3297: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3299: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3300: asub = (Mat_MPIAIJ *)M->data;
3302: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3303: n = asub->B->cmap->N;
3304: if (BsubN > n) {
3305: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3306: const PetscInt *idx;
3307: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3308: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3310: PetscCall(PetscMalloc1(n, &idx_new));
3311: j = 0;
3312: PetscCall(ISGetIndices(iscol_o, &idx));
3313: for (i = 0; i < n; i++) {
3314: if (j >= BsubN) break;
3315: while (subgarray[i] > garray[j]) j++;
3317: if (subgarray[i] == garray[j]) {
3318: idx_new[i] = idx[j++];
3319: } 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]);
3320: }
3321: PetscCall(ISRestoreIndices(iscol_o, &idx));
3323: PetscCall(ISDestroy(&iscol_o));
3324: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3326: } else if (BsubN < n) {
3327: 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);
3328: }
3330: PetscCall(PetscFree(garray));
3331: *submat = M;
3333: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3334: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3335: PetscCall(ISDestroy(&isrow_d));
3337: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3338: PetscCall(ISDestroy(&iscol_d));
3340: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3341: PetscCall(ISDestroy(&iscol_o));
3342: }
3343: PetscFunctionReturn(PETSC_SUCCESS);
3344: }
3346: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3347: {
3348: IS iscol_local = NULL, isrow_d;
3349: PetscInt csize;
3350: PetscInt n, i, j, start, end;
3351: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3352: MPI_Comm comm;
3354: PetscFunctionBegin;
3355: /* If isrow has same processor distribution as mat,
3356: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3357: if (call == MAT_REUSE_MATRIX) {
3358: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3359: if (isrow_d) {
3360: sameRowDist = PETSC_TRUE;
3361: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3362: } else {
3363: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3364: if (iscol_local) {
3365: sameRowDist = PETSC_TRUE;
3366: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3367: }
3368: }
3369: } else {
3370: /* Check if isrow has same processor distribution as mat */
3371: sameDist[0] = PETSC_FALSE;
3372: PetscCall(ISGetLocalSize(isrow, &n));
3373: if (!n) {
3374: sameDist[0] = PETSC_TRUE;
3375: } else {
3376: PetscCall(ISGetMinMax(isrow, &i, &j));
3377: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3378: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3379: }
3381: /* Check if iscol has same processor distribution as mat */
3382: sameDist[1] = PETSC_FALSE;
3383: PetscCall(ISGetLocalSize(iscol, &n));
3384: if (!n) {
3385: sameDist[1] = PETSC_TRUE;
3386: } else {
3387: PetscCall(ISGetMinMax(iscol, &i, &j));
3388: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3389: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3390: }
3392: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3393: PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3394: sameRowDist = tsameDist[0];
3395: }
3397: if (sameRowDist) {
3398: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3399: /* isrow and iscol have same processor distribution as mat */
3400: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3401: PetscFunctionReturn(PETSC_SUCCESS);
3402: } else { /* sameRowDist */
3403: /* isrow has same processor distribution as mat */
3404: if (call == MAT_INITIAL_MATRIX) {
3405: PetscBool sorted;
3406: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3407: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3408: PetscCall(ISGetSize(iscol, &i));
3409: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3411: PetscCall(ISSorted(iscol_local, &sorted));
3412: if (sorted) {
3413: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3414: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3415: PetscFunctionReturn(PETSC_SUCCESS);
3416: }
3417: } else { /* call == MAT_REUSE_MATRIX */
3418: IS iscol_sub;
3419: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3420: if (iscol_sub) {
3421: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3422: PetscFunctionReturn(PETSC_SUCCESS);
3423: }
3424: }
3425: }
3426: }
3428: /* General case: iscol -> iscol_local which has global size of iscol */
3429: if (call == MAT_REUSE_MATRIX) {
3430: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3431: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3432: } else {
3433: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3434: }
3436: PetscCall(ISGetLocalSize(iscol, &csize));
3437: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3439: if (call == MAT_INITIAL_MATRIX) {
3440: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3441: PetscCall(ISDestroy(&iscol_local));
3442: }
3443: PetscFunctionReturn(PETSC_SUCCESS);
3444: }
3446: /*@C
3447: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3448: and "off-diagonal" part of the matrix in CSR format.
3450: Collective
3452: Input Parameters:
3453: + comm - MPI communicator
3454: . A - "diagonal" portion of matrix
3455: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3456: - garray - global index of `B` columns
3458: Output Parameter:
3459: . mat - the matrix, with input `A` as its local diagonal matrix
3461: Level: advanced
3463: Notes:
3464: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3466: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3468: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3469: @*/
3470: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3471: {
3472: Mat_MPIAIJ *maij;
3473: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3474: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3475: const PetscScalar *oa;
3476: Mat Bnew;
3477: PetscInt m, n, N;
3478: MatType mpi_mat_type;
3480: PetscFunctionBegin;
3481: PetscCall(MatCreate(comm, mat));
3482: PetscCall(MatGetSize(A, &m, &n));
3483: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3484: 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);
3485: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3486: /* 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); */
3488: /* Get global columns of mat */
3489: PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3491: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3492: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3493: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3494: PetscCall(MatSetType(*mat, mpi_mat_type));
3496: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3497: maij = (Mat_MPIAIJ *)(*mat)->data;
3499: (*mat)->preallocated = PETSC_TRUE;
3501: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3502: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3504: /* Set A as diagonal portion of *mat */
3505: maij->A = A;
3507: nz = oi[m];
3508: for (i = 0; i < nz; i++) {
3509: col = oj[i];
3510: oj[i] = garray[col];
3511: }
3513: /* Set Bnew as off-diagonal portion of *mat */
3514: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3515: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3516: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3517: bnew = (Mat_SeqAIJ *)Bnew->data;
3518: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3519: maij->B = Bnew;
3521: 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);
3523: b->free_a = PETSC_FALSE;
3524: b->free_ij = PETSC_FALSE;
3525: PetscCall(MatDestroy(&B));
3527: bnew->free_a = PETSC_TRUE;
3528: bnew->free_ij = PETSC_TRUE;
3530: /* condense columns of maij->B */
3531: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3532: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3533: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3534: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3535: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3536: PetscFunctionReturn(PETSC_SUCCESS);
3537: }
3539: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3541: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3542: {
3543: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3544: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3545: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3546: Mat M, Msub, B = a->B;
3547: MatScalar *aa;
3548: Mat_SeqAIJ *aij;
3549: PetscInt *garray = a->garray, *colsub, Ncols;
3550: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3551: IS iscol_sub, iscmap;
3552: const PetscInt *is_idx, *cmap;
3553: PetscBool allcolumns = PETSC_FALSE;
3554: MPI_Comm comm;
3556: PetscFunctionBegin;
3557: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3558: if (call == MAT_REUSE_MATRIX) {
3559: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3560: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3561: PetscCall(ISGetLocalSize(iscol_sub, &count));
3563: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3564: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3566: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3567: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3569: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3571: } else { /* call == MAT_INITIAL_MATRIX) */
3572: PetscBool flg;
3574: PetscCall(ISGetLocalSize(iscol, &n));
3575: PetscCall(ISGetSize(iscol, &Ncols));
3577: /* (1) iscol -> nonscalable iscol_local */
3578: /* Check for special case: each processor gets entire matrix columns */
3579: PetscCall(ISIdentity(iscol_local, &flg));
3580: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3581: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3582: if (allcolumns) {
3583: iscol_sub = iscol_local;
3584: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3585: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3587: } else {
3588: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3589: PetscInt *idx, *cmap1, k;
3590: PetscCall(PetscMalloc1(Ncols, &idx));
3591: PetscCall(PetscMalloc1(Ncols, &cmap1));
3592: PetscCall(ISGetIndices(iscol_local, &is_idx));
3593: count = 0;
3594: k = 0;
3595: for (i = 0; i < Ncols; i++) {
3596: j = is_idx[i];
3597: if (j >= cstart && j < cend) {
3598: /* diagonal part of mat */
3599: idx[count] = j;
3600: cmap1[count++] = i; /* column index in submat */
3601: } else if (Bn) {
3602: /* off-diagonal part of mat */
3603: if (j == garray[k]) {
3604: idx[count] = j;
3605: cmap1[count++] = i; /* column index in submat */
3606: } else if (j > garray[k]) {
3607: while (j > garray[k] && k < Bn - 1) k++;
3608: if (j == garray[k]) {
3609: idx[count] = j;
3610: cmap1[count++] = i; /* column index in submat */
3611: }
3612: }
3613: }
3614: }
3615: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3617: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3618: PetscCall(ISGetBlockSize(iscol, &cbs));
3619: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3621: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3622: }
3624: /* (3) Create sequential Msub */
3625: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3626: }
3628: PetscCall(ISGetLocalSize(iscol_sub, &count));
3629: aij = (Mat_SeqAIJ *)(Msub)->data;
3630: ii = aij->i;
3631: PetscCall(ISGetIndices(iscmap, &cmap));
3633: /*
3634: m - number of local rows
3635: Ncols - number of columns (same on all processors)
3636: rstart - first row in new global matrix generated
3637: */
3638: PetscCall(MatGetSize(Msub, &m, NULL));
3640: if (call == MAT_INITIAL_MATRIX) {
3641: /* (4) Create parallel newmat */
3642: PetscMPIInt rank, size;
3643: PetscInt csize;
3645: PetscCallMPI(MPI_Comm_size(comm, &size));
3646: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3648: /*
3649: Determine the number of non-zeros in the diagonal and off-diagonal
3650: portions of the matrix in order to do correct preallocation
3651: */
3653: /* first get start and end of "diagonal" columns */
3654: PetscCall(ISGetLocalSize(iscol, &csize));
3655: if (csize == PETSC_DECIDE) {
3656: PetscCall(ISGetSize(isrow, &mglobal));
3657: if (mglobal == Ncols) { /* square matrix */
3658: nlocal = m;
3659: } else {
3660: nlocal = Ncols / size + ((Ncols % size) > rank);
3661: }
3662: } else {
3663: nlocal = csize;
3664: }
3665: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3666: rstart = rend - nlocal;
3667: 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);
3669: /* next, compute all the lengths */
3670: jj = aij->j;
3671: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3672: olens = dlens + m;
3673: for (i = 0; i < m; i++) {
3674: jend = ii[i + 1] - ii[i];
3675: olen = 0;
3676: dlen = 0;
3677: for (j = 0; j < jend; j++) {
3678: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3679: else dlen++;
3680: jj++;
3681: }
3682: olens[i] = olen;
3683: dlens[i] = dlen;
3684: }
3686: PetscCall(ISGetBlockSize(isrow, &bs));
3687: PetscCall(ISGetBlockSize(iscol, &cbs));
3689: PetscCall(MatCreate(comm, &M));
3690: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3691: PetscCall(MatSetBlockSizes(M, bs, cbs));
3692: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3693: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3694: PetscCall(PetscFree(dlens));
3696: } else { /* call == MAT_REUSE_MATRIX */
3697: M = *newmat;
3698: PetscCall(MatGetLocalSize(M, &i, NULL));
3699: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3700: PetscCall(MatZeroEntries(M));
3701: /*
3702: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3703: rather than the slower MatSetValues().
3704: */
3705: M->was_assembled = PETSC_TRUE;
3706: M->assembled = PETSC_FALSE;
3707: }
3709: /* (5) Set values of Msub to *newmat */
3710: PetscCall(PetscMalloc1(count, &colsub));
3711: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3713: jj = aij->j;
3714: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3715: for (i = 0; i < m; i++) {
3716: row = rstart + i;
3717: nz = ii[i + 1] - ii[i];
3718: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3719: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3720: jj += nz;
3721: aa += nz;
3722: }
3723: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3724: PetscCall(ISRestoreIndices(iscmap, &cmap));
3726: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3727: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3729: PetscCall(PetscFree(colsub));
3731: /* save Msub, iscol_sub and iscmap used in processor for next request */
3732: if (call == MAT_INITIAL_MATRIX) {
3733: *newmat = M;
3734: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3735: PetscCall(MatDestroy(&Msub));
3737: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3738: PetscCall(ISDestroy(&iscol_sub));
3740: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3741: PetscCall(ISDestroy(&iscmap));
3743: if (iscol_local) {
3744: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3745: PetscCall(ISDestroy(&iscol_local));
3746: }
3747: }
3748: PetscFunctionReturn(PETSC_SUCCESS);
3749: }
3751: /*
3752: Not great since it makes two copies of the submatrix, first an SeqAIJ
3753: in local and then by concatenating the local matrices the end result.
3754: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3756: This requires a sequential iscol with all indices.
3757: */
3758: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3759: {
3760: PetscMPIInt rank, size;
3761: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3762: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3763: Mat M, Mreuse;
3764: MatScalar *aa, *vwork;
3765: MPI_Comm comm;
3766: Mat_SeqAIJ *aij;
3767: PetscBool colflag, allcolumns = PETSC_FALSE;
3769: PetscFunctionBegin;
3770: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3771: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3772: PetscCallMPI(MPI_Comm_size(comm, &size));
3774: /* Check for special case: each processor gets entire matrix columns */
3775: PetscCall(ISIdentity(iscol, &colflag));
3776: PetscCall(ISGetLocalSize(iscol, &n));
3777: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3778: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3780: if (call == MAT_REUSE_MATRIX) {
3781: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3782: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3783: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3784: } else {
3785: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3786: }
3788: /*
3789: m - number of local rows
3790: n - number of columns (same on all processors)
3791: rstart - first row in new global matrix generated
3792: */
3793: PetscCall(MatGetSize(Mreuse, &m, &n));
3794: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3795: if (call == MAT_INITIAL_MATRIX) {
3796: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3797: ii = aij->i;
3798: jj = aij->j;
3800: /*
3801: Determine the number of non-zeros in the diagonal and off-diagonal
3802: portions of the matrix in order to do correct preallocation
3803: */
3805: /* first get start and end of "diagonal" columns */
3806: if (csize == PETSC_DECIDE) {
3807: PetscCall(ISGetSize(isrow, &mglobal));
3808: if (mglobal == n) { /* square matrix */
3809: nlocal = m;
3810: } else {
3811: nlocal = n / size + ((n % size) > rank);
3812: }
3813: } else {
3814: nlocal = csize;
3815: }
3816: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3817: rstart = rend - nlocal;
3818: 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);
3820: /* next, compute all the lengths */
3821: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3822: olens = dlens + m;
3823: for (i = 0; i < m; i++) {
3824: jend = ii[i + 1] - ii[i];
3825: olen = 0;
3826: dlen = 0;
3827: for (j = 0; j < jend; j++) {
3828: if (*jj < rstart || *jj >= rend) olen++;
3829: else dlen++;
3830: jj++;
3831: }
3832: olens[i] = olen;
3833: dlens[i] = dlen;
3834: }
3835: PetscCall(MatCreate(comm, &M));
3836: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3837: PetscCall(MatSetBlockSizes(M, bs, cbs));
3838: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3839: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3840: PetscCall(PetscFree(dlens));
3841: } else {
3842: PetscInt ml, nl;
3844: M = *newmat;
3845: PetscCall(MatGetLocalSize(M, &ml, &nl));
3846: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3847: PetscCall(MatZeroEntries(M));
3848: /*
3849: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3850: rather than the slower MatSetValues().
3851: */
3852: M->was_assembled = PETSC_TRUE;
3853: M->assembled = PETSC_FALSE;
3854: }
3855: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3856: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3857: ii = aij->i;
3858: jj = aij->j;
3860: /* trigger copy to CPU if needed */
3861: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3862: for (i = 0; i < m; i++) {
3863: row = rstart + i;
3864: nz = ii[i + 1] - ii[i];
3865: cwork = jj;
3866: jj = PetscSafePointerPlusOffset(jj, nz);
3867: vwork = aa;
3868: aa = PetscSafePointerPlusOffset(aa, nz);
3869: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3870: }
3871: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3873: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3874: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3875: *newmat = M;
3877: /* save submatrix used in processor for next request */
3878: if (call == MAT_INITIAL_MATRIX) {
3879: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3880: PetscCall(MatDestroy(&Mreuse));
3881: }
3882: PetscFunctionReturn(PETSC_SUCCESS);
3883: }
3885: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3886: {
3887: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3888: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3889: const PetscInt *JJ;
3890: PetscBool nooffprocentries;
3891: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3893: PetscFunctionBegin;
3894: PetscCall(PetscLayoutSetUp(B->rmap));
3895: PetscCall(PetscLayoutSetUp(B->cmap));
3896: m = B->rmap->n;
3897: cstart = B->cmap->rstart;
3898: cend = B->cmap->rend;
3899: rstart = B->rmap->rstart;
3900: irstart = Ii[0];
3902: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3904: if (PetscDefined(USE_DEBUG)) {
3905: for (i = 0; i < m; i++) {
3906: nnz = Ii[i + 1] - Ii[i];
3907: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3908: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3909: 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]);
3910: 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);
3911: }
3912: }
3914: for (i = 0; i < m; i++) {
3915: nnz = Ii[i + 1] - Ii[i];
3916: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3917: nnz_max = PetscMax(nnz_max, nnz);
3918: d = 0;
3919: for (j = 0; j < nnz; j++) {
3920: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3921: }
3922: d_nnz[i] = d;
3923: o_nnz[i] = nnz - d;
3924: }
3925: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3926: PetscCall(PetscFree2(d_nnz, o_nnz));
3928: for (i = 0; i < m; i++) {
3929: ii = i + rstart;
3930: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3931: }
3932: nooffprocentries = B->nooffprocentries;
3933: B->nooffprocentries = PETSC_TRUE;
3934: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3935: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3936: B->nooffprocentries = nooffprocentries;
3938: /* count number of entries below block diagonal */
3939: PetscCall(PetscFree(Aij->ld));
3940: PetscCall(PetscCalloc1(m, &ld));
3941: Aij->ld = ld;
3942: for (i = 0; i < m; i++) {
3943: nnz = Ii[i + 1] - Ii[i];
3944: j = 0;
3945: while (j < nnz && J[j] < cstart) j++;
3946: ld[i] = j;
3947: if (J) J += nnz;
3948: }
3950: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3951: PetscFunctionReturn(PETSC_SUCCESS);
3952: }
3954: /*@
3955: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3956: (the default parallel PETSc format).
3958: Collective
3960: Input Parameters:
3961: + B - the matrix
3962: . i - the indices into `j` for the start of each local row (indices start with zero)
3963: . j - the column indices for each local row (indices start with zero)
3964: - v - optional values in the matrix
3966: Level: developer
3968: Notes:
3969: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3970: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3971: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3973: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3975: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3977: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3979: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3980: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3982: The format which is used for the sparse matrix input, is equivalent to a
3983: row-major ordering.. i.e for the following matrix, the input data expected is
3984: as shown
3985: .vb
3986: 1 0 0
3987: 2 0 3 P0
3988: -------
3989: 4 5 6 P1
3991: Process0 [P0] rows_owned=[0,1]
3992: i = {0,1,3} [size = nrow+1 = 2+1]
3993: j = {0,0,2} [size = 3]
3994: v = {1,2,3} [size = 3]
3996: Process1 [P1] rows_owned=[2]
3997: i = {0,3} [size = nrow+1 = 1+1]
3998: j = {0,1,2} [size = 3]
3999: v = {4,5,6} [size = 3]
4000: .ve
4002: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4003: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4004: @*/
4005: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4006: {
4007: PetscFunctionBegin;
4008: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4009: PetscFunctionReturn(PETSC_SUCCESS);
4010: }
4012: /*@
4013: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4014: (the default parallel PETSc format). For good matrix assembly performance
4015: the user should preallocate the matrix storage by setting the parameters
4016: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4018: Collective
4020: Input Parameters:
4021: + B - the matrix
4022: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4023: (same value is used for all local rows)
4024: . d_nnz - array containing the number of nonzeros in the various rows of the
4025: DIAGONAL portion of the local submatrix (possibly different for each row)
4026: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4027: The size of this array is equal to the number of local rows, i.e 'm'.
4028: For matrices that will be factored, you must leave room for (and set)
4029: the diagonal entry even if it is zero.
4030: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4031: submatrix (same value is used for all local rows).
4032: - o_nnz - array containing the number of nonzeros in the various rows of the
4033: OFF-DIAGONAL portion of the local submatrix (possibly different for
4034: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4035: structure. The size of this array is equal to the number
4036: of local rows, i.e 'm'.
4038: Example Usage:
4039: Consider the following 8x8 matrix with 34 non-zero values, that is
4040: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4041: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4042: as follows
4044: .vb
4045: 1 2 0 | 0 3 0 | 0 4
4046: Proc0 0 5 6 | 7 0 0 | 8 0
4047: 9 0 10 | 11 0 0 | 12 0
4048: -------------------------------------
4049: 13 0 14 | 15 16 17 | 0 0
4050: Proc1 0 18 0 | 19 20 21 | 0 0
4051: 0 0 0 | 22 23 0 | 24 0
4052: -------------------------------------
4053: Proc2 25 26 27 | 0 0 28 | 29 0
4054: 30 0 0 | 31 32 33 | 0 34
4055: .ve
4057: This can be represented as a collection of submatrices as
4058: .vb
4059: A B C
4060: D E F
4061: G H I
4062: .ve
4064: Where the submatrices A,B,C are owned by proc0, D,E,F are
4065: owned by proc1, G,H,I are owned by proc2.
4067: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4068: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4069: The 'M','N' parameters are 8,8, and have the same values on all procs.
4071: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4072: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4073: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4074: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4075: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4076: matrix, ans [DF] as another `MATSEQAIJ` matrix.
4078: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4079: allocated for every row of the local diagonal submatrix, and `o_nz`
4080: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4081: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4082: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4083: In this case, the values of `d_nz`, `o_nz` are
4084: .vb
4085: proc0 dnz = 2, o_nz = 2
4086: proc1 dnz = 3, o_nz = 2
4087: proc2 dnz = 1, o_nz = 4
4088: .ve
4089: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4090: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4091: for proc3. i.e we are using 12+15+10=37 storage locations to store
4092: 34 values.
4094: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4095: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4096: In the above case the values for `d_nnz`, `o_nnz` are
4097: .vb
4098: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4099: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4100: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4101: .ve
4102: Here the space allocated is sum of all the above values i.e 34, and
4103: hence pre-allocation is perfect.
4105: Level: intermediate
4107: Notes:
4108: If the *_nnz parameter is given then the *_nz parameter is ignored
4110: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4111: storage. The stored row and column indices begin with zero.
4112: See [Sparse Matrices](sec_matsparse) for details.
4114: The parallel matrix is partitioned such that the first m0 rows belong to
4115: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4116: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4118: The DIAGONAL portion of the local submatrix of a processor can be defined
4119: as the submatrix which is obtained by extraction the part corresponding to
4120: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4121: first row that belongs to the processor, r2 is the last row belonging to
4122: the this processor, and c1-c2 is range of indices of the local part of a
4123: vector suitable for applying the matrix to. This is an mxn matrix. In the
4124: common case of a square matrix, the row and column ranges are the same and
4125: the DIAGONAL part is also square. The remaining portion of the local
4126: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4128: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4130: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4131: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4132: You can also run with the option `-info` and look for messages with the string
4133: malloc in them to see if additional memory allocation was needed.
4135: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4136: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4137: @*/
4138: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4139: {
4140: PetscFunctionBegin;
4143: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4144: PetscFunctionReturn(PETSC_SUCCESS);
4145: }
4147: /*@
4148: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4149: CSR format for the local rows.
4151: Collective
4153: Input Parameters:
4154: + comm - MPI communicator
4155: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4156: . n - This value should be the same as the local size used in creating the
4157: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4158: calculated if `N` is given) For square matrices n is almost always `m`.
4159: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4160: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4161: . 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
4162: . j - global column indices
4163: - a - optional matrix values
4165: Output Parameter:
4166: . mat - the matrix
4168: Level: intermediate
4170: Notes:
4171: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4172: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4173: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4175: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4177: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4179: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4180: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4182: The format which is used for the sparse matrix input, is equivalent to a
4183: row-major ordering, i.e., for the following matrix, the input data expected is
4184: as shown
4185: .vb
4186: 1 0 0
4187: 2 0 3 P0
4188: -------
4189: 4 5 6 P1
4191: Process0 [P0] rows_owned=[0,1]
4192: i = {0,1,3} [size = nrow+1 = 2+1]
4193: j = {0,0,2} [size = 3]
4194: v = {1,2,3} [size = 3]
4196: Process1 [P1] rows_owned=[2]
4197: i = {0,3} [size = nrow+1 = 1+1]
4198: j = {0,1,2} [size = 3]
4199: v = {4,5,6} [size = 3]
4200: .ve
4202: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4203: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4204: @*/
4205: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4206: {
4207: PetscFunctionBegin;
4208: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4209: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4210: PetscCall(MatCreate(comm, mat));
4211: PetscCall(MatSetSizes(*mat, m, n, M, N));
4212: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4213: PetscCall(MatSetType(*mat, MATMPIAIJ));
4214: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4215: PetscFunctionReturn(PETSC_SUCCESS);
4216: }
4218: /*@
4219: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4220: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4221: from `MatCreateMPIAIJWithArrays()`
4223: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4225: Collective
4227: Input Parameters:
4228: + mat - the matrix
4229: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4230: . n - This value should be the same as the local size used in creating the
4231: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4232: calculated if N is given) For square matrices n is almost always m.
4233: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4234: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4235: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4236: . J - column indices
4237: - v - matrix values
4239: Level: deprecated
4241: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4242: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4243: @*/
4244: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4245: {
4246: PetscInt nnz, i;
4247: PetscBool nooffprocentries;
4248: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4249: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4250: PetscScalar *ad, *ao;
4251: PetscInt ldi, Iii, md;
4252: const PetscInt *Adi = Ad->i;
4253: PetscInt *ld = Aij->ld;
4255: PetscFunctionBegin;
4256: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4257: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4258: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4259: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4261: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4262: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4264: for (i = 0; i < m; i++) {
4265: if (PetscDefined(USE_DEBUG)) {
4266: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4267: 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);
4268: 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);
4269: }
4270: }
4271: nnz = Ii[i + 1] - Ii[i];
4272: Iii = Ii[i];
4273: ldi = ld[i];
4274: md = Adi[i + 1] - Adi[i];
4275: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4276: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4277: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4278: ad += md;
4279: ao += nnz - md;
4280: }
4281: nooffprocentries = mat->nooffprocentries;
4282: mat->nooffprocentries = PETSC_TRUE;
4283: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4284: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4285: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4286: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4287: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4288: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4289: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4290: mat->nooffprocentries = nooffprocentries;
4291: PetscFunctionReturn(PETSC_SUCCESS);
4292: }
4294: /*@
4295: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4297: Collective
4299: Input Parameters:
4300: + mat - the matrix
4301: - v - matrix values, stored by row
4303: Level: intermediate
4305: Notes:
4306: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4308: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4310: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4311: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4312: @*/
4313: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4314: {
4315: PetscInt nnz, i, m;
4316: PetscBool nooffprocentries;
4317: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4318: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4319: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4320: PetscScalar *ad, *ao;
4321: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4322: PetscInt ldi, Iii, md;
4323: PetscInt *ld = Aij->ld;
4325: PetscFunctionBegin;
4326: m = mat->rmap->n;
4328: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4329: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4330: Iii = 0;
4331: for (i = 0; i < m; i++) {
4332: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4333: ldi = ld[i];
4334: md = Adi[i + 1] - Adi[i];
4335: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4336: ad += md;
4337: if (ao) {
4338: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4339: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4340: ao += nnz - md;
4341: }
4342: Iii += nnz;
4343: }
4344: nooffprocentries = mat->nooffprocentries;
4345: mat->nooffprocentries = PETSC_TRUE;
4346: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4347: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4348: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4349: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4350: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4351: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4352: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4353: mat->nooffprocentries = nooffprocentries;
4354: PetscFunctionReturn(PETSC_SUCCESS);
4355: }
4357: /*@
4358: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4359: (the default parallel PETSc format). For good matrix assembly performance
4360: the user should preallocate the matrix storage by setting the parameters
4361: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4363: Collective
4365: Input Parameters:
4366: + comm - MPI communicator
4367: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4368: This value should be the same as the local size used in creating the
4369: y vector for the matrix-vector product y = Ax.
4370: . n - This value should be the same as the local size used in creating the
4371: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4372: calculated if N is given) For square matrices n is almost always m.
4373: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4374: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4375: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4376: (same value is used for all local rows)
4377: . d_nnz - array containing the number of nonzeros in the various rows of the
4378: DIAGONAL portion of the local submatrix (possibly different for each row)
4379: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4380: The size of this array is equal to the number of local rows, i.e 'm'.
4381: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4382: submatrix (same value is used for all local rows).
4383: - o_nnz - array containing the number of nonzeros in the various rows of the
4384: OFF-DIAGONAL portion of the local submatrix (possibly different for
4385: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4386: structure. The size of this array is equal to the number
4387: of local rows, i.e 'm'.
4389: Output Parameter:
4390: . A - the matrix
4392: Options Database Keys:
4393: + -mat_no_inode - Do not use inodes
4394: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4395: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4396: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4397: to be viewed as a matrix. Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.
4399: Level: intermediate
4401: Notes:
4402: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4403: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4404: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4406: If the *_nnz parameter is given then the *_nz parameter is ignored
4408: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4409: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4410: storage requirements for this matrix.
4412: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4413: processor than it must be used on all processors that share the object for
4414: that argument.
4416: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4417: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4419: The user MUST specify either the local or global matrix dimensions
4420: (possibly both).
4422: The parallel matrix is partitioned across processors such that the
4423: first `m0` rows belong to process 0, the next `m1` rows belong to
4424: process 1, the next `m2` rows belong to process 2, etc., where
4425: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4426: values corresponding to [m x N] submatrix.
4428: The columns are logically partitioned with the n0 columns belonging
4429: to 0th partition, the next n1 columns belonging to the next
4430: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4432: The DIAGONAL portion of the local submatrix on any given processor
4433: is the submatrix corresponding to the rows and columns m,n
4434: corresponding to the given processor. i.e diagonal matrix on
4435: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4436: etc. The remaining portion of the local submatrix [m x (N-n)]
4437: constitute the OFF-DIAGONAL portion. The example below better
4438: illustrates this concept.
4440: For a square global matrix we define each processor's diagonal portion
4441: to be its local rows and the corresponding columns (a square submatrix);
4442: each processor's off-diagonal portion encompasses the remainder of the
4443: local matrix (a rectangular submatrix).
4445: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4447: When calling this routine with a single process communicator, a matrix of
4448: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4449: type of communicator, use the construction mechanism
4450: .vb
4451: MatCreate(..., &A);
4452: MatSetType(A, MATMPIAIJ);
4453: MatSetSizes(A, m, n, M, N);
4454: MatMPIAIJSetPreallocation(A, ...);
4455: .ve
4457: By default, this format uses inodes (identical nodes) when possible.
4458: We search for consecutive rows with the same nonzero structure, thereby
4459: reusing matrix information to achieve increased efficiency.
4461: Example Usage:
4462: Consider the following 8x8 matrix with 34 non-zero values, that is
4463: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4464: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4465: as follows
4467: .vb
4468: 1 2 0 | 0 3 0 | 0 4
4469: Proc0 0 5 6 | 7 0 0 | 8 0
4470: 9 0 10 | 11 0 0 | 12 0
4471: -------------------------------------
4472: 13 0 14 | 15 16 17 | 0 0
4473: Proc1 0 18 0 | 19 20 21 | 0 0
4474: 0 0 0 | 22 23 0 | 24 0
4475: -------------------------------------
4476: Proc2 25 26 27 | 0 0 28 | 29 0
4477: 30 0 0 | 31 32 33 | 0 34
4478: .ve
4480: This can be represented as a collection of submatrices as
4482: .vb
4483: A B C
4484: D E F
4485: G H I
4486: .ve
4488: Where the submatrices A,B,C are owned by proc0, D,E,F are
4489: owned by proc1, G,H,I are owned by proc2.
4491: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4492: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4493: The 'M','N' parameters are 8,8, and have the same values on all procs.
4495: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4496: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4497: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4498: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4499: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4500: matrix, ans [DF] as another SeqAIJ matrix.
4502: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4503: allocated for every row of the local diagonal submatrix, and `o_nz`
4504: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4505: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4506: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4507: In this case, the values of `d_nz`,`o_nz` are
4508: .vb
4509: proc0 dnz = 2, o_nz = 2
4510: proc1 dnz = 3, o_nz = 2
4511: proc2 dnz = 1, o_nz = 4
4512: .ve
4513: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4514: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4515: for proc3. i.e we are using 12+15+10=37 storage locations to store
4516: 34 values.
4518: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4519: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4520: In the above case the values for d_nnz,o_nnz are
4521: .vb
4522: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4523: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4524: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4525: .ve
4526: Here the space allocated is sum of all the above values i.e 34, and
4527: hence pre-allocation is perfect.
4529: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4530: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4531: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4532: @*/
4533: 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)
4534: {
4535: PetscMPIInt size;
4537: PetscFunctionBegin;
4538: PetscCall(MatCreate(comm, A));
4539: PetscCall(MatSetSizes(*A, m, n, M, N));
4540: PetscCallMPI(MPI_Comm_size(comm, &size));
4541: if (size > 1) {
4542: PetscCall(MatSetType(*A, MATMPIAIJ));
4543: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4544: } else {
4545: PetscCall(MatSetType(*A, MATSEQAIJ));
4546: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4547: }
4548: PetscFunctionReturn(PETSC_SUCCESS);
4549: }
4551: /*MC
4552: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4554: Synopsis:
4555: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4557: Not Collective
4559: Input Parameter:
4560: . A - the `MATMPIAIJ` matrix
4562: Output Parameters:
4563: + Ad - the diagonal portion of the matrix
4564: . Ao - the off-diagonal portion of the matrix
4565: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4566: - ierr - error code
4568: Level: advanced
4570: Note:
4571: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4573: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4574: M*/
4576: /*MC
4577: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4579: Synopsis:
4580: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4582: Not Collective
4584: Input Parameters:
4585: + A - the `MATMPIAIJ` matrix
4586: . Ad - the diagonal portion of the matrix
4587: . Ao - the off-diagonal portion of the matrix
4588: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4589: - ierr - error code
4591: Level: advanced
4593: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4594: M*/
4596: /*@C
4597: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4599: Not Collective
4601: Input Parameter:
4602: . A - The `MATMPIAIJ` matrix
4604: Output Parameters:
4605: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4606: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4607: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4609: Level: intermediate
4611: Note:
4612: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4613: 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
4614: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4615: local column numbers to global column numbers in the original matrix.
4617: Fortran Notes:
4618: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4620: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4621: @*/
4622: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4623: {
4624: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4625: PetscBool flg;
4627: PetscFunctionBegin;
4628: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4629: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4630: if (Ad) *Ad = a->A;
4631: if (Ao) *Ao = a->B;
4632: if (colmap) *colmap = a->garray;
4633: PetscFunctionReturn(PETSC_SUCCESS);
4634: }
4636: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4637: {
4638: PetscInt m, N, i, rstart, nnz, Ii;
4639: PetscInt *indx;
4640: PetscScalar *values;
4641: MatType rootType;
4643: PetscFunctionBegin;
4644: PetscCall(MatGetSize(inmat, &m, &N));
4645: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4646: PetscInt *dnz, *onz, sum, bs, cbs;
4648: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4649: /* Check sum(n) = N */
4650: PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4651: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4653: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4654: rstart -= m;
4656: MatPreallocateBegin(comm, m, n, dnz, onz);
4657: for (i = 0; i < m; i++) {
4658: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4659: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4660: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4661: }
4663: PetscCall(MatCreate(comm, outmat));
4664: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4665: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4666: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4667: PetscCall(MatGetRootType_Private(inmat, &rootType));
4668: PetscCall(MatSetType(*outmat, rootType));
4669: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4670: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4671: MatPreallocateEnd(dnz, onz);
4672: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4673: }
4675: /* numeric phase */
4676: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4677: for (i = 0; i < m; i++) {
4678: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4679: Ii = i + rstart;
4680: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4681: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4682: }
4683: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4684: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4685: PetscFunctionReturn(PETSC_SUCCESS);
4686: }
4688: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4689: {
4690: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4692: PetscFunctionBegin;
4693: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4694: PetscCall(PetscFree(merge->id_r));
4695: PetscCall(PetscFree(merge->len_s));
4696: PetscCall(PetscFree(merge->len_r));
4697: PetscCall(PetscFree(merge->bi));
4698: PetscCall(PetscFree(merge->bj));
4699: PetscCall(PetscFree(merge->buf_ri[0]));
4700: PetscCall(PetscFree(merge->buf_ri));
4701: PetscCall(PetscFree(merge->buf_rj[0]));
4702: PetscCall(PetscFree(merge->buf_rj));
4703: PetscCall(PetscFree(merge->coi));
4704: PetscCall(PetscFree(merge->coj));
4705: PetscCall(PetscFree(merge->owners_co));
4706: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4707: PetscCall(PetscFree(merge));
4708: PetscFunctionReturn(PETSC_SUCCESS);
4709: }
4711: #include <../src/mat/utils/freespace.h>
4712: #include <petscbt.h>
4714: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4715: {
4716: MPI_Comm comm;
4717: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4718: PetscMPIInt size, rank, taga, *len_s;
4719: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4720: PetscInt proc, m;
4721: PetscInt **buf_ri, **buf_rj;
4722: PetscInt k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4723: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4724: MPI_Request *s_waits, *r_waits;
4725: MPI_Status *status;
4726: const MatScalar *aa, *a_a;
4727: MatScalar **abuf_r, *ba_i;
4728: Mat_Merge_SeqsToMPI *merge;
4729: PetscContainer container;
4731: PetscFunctionBegin;
4732: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4733: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4735: PetscCallMPI(MPI_Comm_size(comm, &size));
4736: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4738: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4739: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4740: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4741: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4742: aa = a_a;
4744: bi = merge->bi;
4745: bj = merge->bj;
4746: buf_ri = merge->buf_ri;
4747: buf_rj = merge->buf_rj;
4749: PetscCall(PetscMalloc1(size, &status));
4750: owners = merge->rowmap->range;
4751: len_s = merge->len_s;
4753: /* send and recv matrix values */
4754: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4755: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4757: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4758: for (proc = 0, k = 0; proc < size; proc++) {
4759: if (!len_s[proc]) continue;
4760: i = owners[proc];
4761: PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4762: k++;
4763: }
4765: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4766: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4767: PetscCall(PetscFree(status));
4769: PetscCall(PetscFree(s_waits));
4770: PetscCall(PetscFree(r_waits));
4772: /* insert mat values of mpimat */
4773: PetscCall(PetscMalloc1(N, &ba_i));
4774: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4776: for (k = 0; k < merge->nrecv; k++) {
4777: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4778: nrows = *buf_ri_k[k];
4779: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4780: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4781: }
4783: /* set values of ba */
4784: m = merge->rowmap->n;
4785: for (i = 0; i < m; i++) {
4786: arow = owners[rank] + i;
4787: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4788: bnzi = bi[i + 1] - bi[i];
4789: PetscCall(PetscArrayzero(ba_i, bnzi));
4791: /* add local non-zero vals of this proc's seqmat into ba */
4792: anzi = ai[arow + 1] - ai[arow];
4793: aj = a->j + ai[arow];
4794: aa = a_a + ai[arow];
4795: nextaj = 0;
4796: for (j = 0; nextaj < anzi; j++) {
4797: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4798: ba_i[j] += aa[nextaj++];
4799: }
4800: }
4802: /* add received vals into ba */
4803: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4804: /* i-th row */
4805: if (i == *nextrow[k]) {
4806: anzi = *(nextai[k] + 1) - *nextai[k];
4807: aj = buf_rj[k] + *nextai[k];
4808: aa = abuf_r[k] + *nextai[k];
4809: nextaj = 0;
4810: for (j = 0; nextaj < anzi; j++) {
4811: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4812: ba_i[j] += aa[nextaj++];
4813: }
4814: }
4815: nextrow[k]++;
4816: nextai[k]++;
4817: }
4818: }
4819: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4820: }
4821: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4822: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4823: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4825: PetscCall(PetscFree(abuf_r[0]));
4826: PetscCall(PetscFree(abuf_r));
4827: PetscCall(PetscFree(ba_i));
4828: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4829: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4830: PetscFunctionReturn(PETSC_SUCCESS);
4831: }
4833: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4834: {
4835: Mat B_mpi;
4836: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4837: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4838: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4839: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4840: PetscInt len, proc, *dnz, *onz, bs, cbs;
4841: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4842: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4843: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4844: MPI_Status *status;
4845: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4846: PetscBT lnkbt;
4847: Mat_Merge_SeqsToMPI *merge;
4848: PetscContainer container;
4850: PetscFunctionBegin;
4851: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4853: /* make sure it is a PETSc comm */
4854: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4855: PetscCallMPI(MPI_Comm_size(comm, &size));
4856: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4858: PetscCall(PetscNew(&merge));
4859: PetscCall(PetscMalloc1(size, &status));
4861: /* determine row ownership */
4862: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4863: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4864: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4865: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4866: PetscCall(PetscLayoutSetUp(merge->rowmap));
4867: PetscCall(PetscMalloc1(size, &len_si));
4868: PetscCall(PetscMalloc1(size, &merge->len_s));
4870: m = merge->rowmap->n;
4871: owners = merge->rowmap->range;
4873: /* determine the number of messages to send, their lengths */
4874: len_s = merge->len_s;
4876: len = 0; /* length of buf_si[] */
4877: merge->nsend = 0;
4878: for (proc = 0; proc < size; proc++) {
4879: len_si[proc] = 0;
4880: if (proc == rank) {
4881: len_s[proc] = 0;
4882: } else {
4883: len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4884: len_s[proc] = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4885: }
4886: if (len_s[proc]) {
4887: merge->nsend++;
4888: nrows = 0;
4889: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4890: if (ai[i + 1] > ai[i]) nrows++;
4891: }
4892: len_si[proc] = 2 * (nrows + 1);
4893: len += len_si[proc];
4894: }
4895: }
4897: /* determine the number and length of messages to receive for ij-structure */
4898: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4899: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4901: /* post the Irecv of j-structure */
4902: PetscCall(PetscCommGetNewTag(comm, &tagj));
4903: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4905: /* post the Isend of j-structure */
4906: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4908: for (proc = 0, k = 0; proc < size; proc++) {
4909: if (!len_s[proc]) continue;
4910: i = owners[proc];
4911: PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4912: k++;
4913: }
4915: /* receives and sends of j-structure are complete */
4916: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4917: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4919: /* send and recv i-structure */
4920: PetscCall(PetscCommGetNewTag(comm, &tagi));
4921: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4923: PetscCall(PetscMalloc1(len + 1, &buf_s));
4924: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4925: for (proc = 0, k = 0; proc < size; proc++) {
4926: if (!len_s[proc]) continue;
4927: /* form outgoing message for i-structure:
4928: buf_si[0]: nrows to be sent
4929: [1:nrows]: row index (global)
4930: [nrows+1:2*nrows+1]: i-structure index
4931: */
4932: nrows = len_si[proc] / 2 - 1;
4933: buf_si_i = buf_si + nrows + 1;
4934: buf_si[0] = nrows;
4935: buf_si_i[0] = 0;
4936: nrows = 0;
4937: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4938: anzi = ai[i + 1] - ai[i];
4939: if (anzi) {
4940: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4941: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4942: nrows++;
4943: }
4944: }
4945: PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4946: k++;
4947: buf_si += len_si[proc];
4948: }
4950: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4951: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4953: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4954: 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]));
4956: PetscCall(PetscFree(len_si));
4957: PetscCall(PetscFree(len_ri));
4958: PetscCall(PetscFree(rj_waits));
4959: PetscCall(PetscFree2(si_waits, sj_waits));
4960: PetscCall(PetscFree(ri_waits));
4961: PetscCall(PetscFree(buf_s));
4962: PetscCall(PetscFree(status));
4964: /* compute a local seq matrix in each processor */
4965: /* allocate bi array and free space for accumulating nonzero column info */
4966: PetscCall(PetscMalloc1(m + 1, &bi));
4967: bi[0] = 0;
4969: /* create and initialize a linked list */
4970: nlnk = N + 1;
4971: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4973: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4974: len = ai[owners[rank + 1]] - ai[owners[rank]];
4975: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4977: current_space = free_space;
4979: /* determine symbolic info for each local row */
4980: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4982: for (k = 0; k < merge->nrecv; k++) {
4983: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4984: nrows = *buf_ri_k[k];
4985: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4986: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4987: }
4989: MatPreallocateBegin(comm, m, n, dnz, onz);
4990: len = 0;
4991: for (i = 0; i < m; i++) {
4992: bnzi = 0;
4993: /* add local non-zero cols of this proc's seqmat into lnk */
4994: arow = owners[rank] + i;
4995: anzi = ai[arow + 1] - ai[arow];
4996: aj = a->j + ai[arow];
4997: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4998: bnzi += nlnk;
4999: /* add received col data into lnk */
5000: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5001: if (i == *nextrow[k]) { /* i-th row */
5002: anzi = *(nextai[k] + 1) - *nextai[k];
5003: aj = buf_rj[k] + *nextai[k];
5004: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5005: bnzi += nlnk;
5006: nextrow[k]++;
5007: nextai[k]++;
5008: }
5009: }
5010: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5012: /* if free space is not available, make more free space */
5013: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5014: /* copy data into free space, then initialize lnk */
5015: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5016: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5018: current_space->array += bnzi;
5019: current_space->local_used += bnzi;
5020: current_space->local_remaining -= bnzi;
5022: bi[i + 1] = bi[i] + bnzi;
5023: }
5025: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5027: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5028: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5029: PetscCall(PetscLLDestroy(lnk, lnkbt));
5031: /* create symbolic parallel matrix B_mpi */
5032: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5033: PetscCall(MatCreate(comm, &B_mpi));
5034: if (n == PETSC_DECIDE) {
5035: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5036: } else {
5037: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5038: }
5039: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5040: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5041: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5042: MatPreallocateEnd(dnz, onz);
5043: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5045: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5046: B_mpi->assembled = PETSC_FALSE;
5047: merge->bi = bi;
5048: merge->bj = bj;
5049: merge->buf_ri = buf_ri;
5050: merge->buf_rj = buf_rj;
5051: merge->coi = NULL;
5052: merge->coj = NULL;
5053: merge->owners_co = NULL;
5055: PetscCall(PetscCommDestroy(&comm));
5057: /* attach the supporting struct to B_mpi for reuse */
5058: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5059: PetscCall(PetscContainerSetPointer(container, merge));
5060: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5061: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5062: PetscCall(PetscContainerDestroy(&container));
5063: *mpimat = B_mpi;
5065: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5066: PetscFunctionReturn(PETSC_SUCCESS);
5067: }
5069: /*@
5070: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5071: matrices from each processor
5073: Collective
5075: Input Parameters:
5076: + comm - the communicators the parallel matrix will live on
5077: . seqmat - the input sequential matrices
5078: . m - number of local rows (or `PETSC_DECIDE`)
5079: . n - number of local columns (or `PETSC_DECIDE`)
5080: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5082: Output Parameter:
5083: . mpimat - the parallel matrix generated
5085: Level: advanced
5087: Note:
5088: The dimensions of the sequential matrix in each processor MUST be the same.
5089: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5090: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5092: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5093: @*/
5094: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5095: {
5096: PetscMPIInt size;
5098: PetscFunctionBegin;
5099: PetscCallMPI(MPI_Comm_size(comm, &size));
5100: if (size == 1) {
5101: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5102: if (scall == MAT_INITIAL_MATRIX) {
5103: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5104: } else {
5105: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5106: }
5107: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5108: PetscFunctionReturn(PETSC_SUCCESS);
5109: }
5110: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5111: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5112: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5113: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5114: PetscFunctionReturn(PETSC_SUCCESS);
5115: }
5117: /*@
5118: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5120: Not Collective
5122: Input Parameter:
5123: . A - the matrix
5125: Output Parameter:
5126: . A_loc - the local sequential matrix generated
5128: Level: developer
5130: Notes:
5131: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5132: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5133: `n` is the global column count obtained with `MatGetSize()`
5135: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5137: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5139: Destroy the matrix with `MatDestroy()`
5141: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5142: @*/
5143: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5144: {
5145: PetscBool mpi;
5147: PetscFunctionBegin;
5148: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5149: if (mpi) {
5150: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5151: } else {
5152: *A_loc = A;
5153: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5154: }
5155: PetscFunctionReturn(PETSC_SUCCESS);
5156: }
5158: /*@
5159: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5161: Not Collective
5163: Input Parameters:
5164: + A - the matrix
5165: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5167: Output Parameter:
5168: . A_loc - the local sequential matrix generated
5170: Level: developer
5172: Notes:
5173: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5174: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5175: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5177: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5179: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5180: 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
5181: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5182: 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.
5184: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5185: @*/
5186: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5187: {
5188: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5189: Mat_SeqAIJ *mat, *a, *b;
5190: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5191: const PetscScalar *aa, *ba, *aav, *bav;
5192: PetscScalar *ca, *cam;
5193: PetscMPIInt size;
5194: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5195: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5196: PetscBool match;
5198: PetscFunctionBegin;
5199: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5200: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5201: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5202: if (size == 1) {
5203: if (scall == MAT_INITIAL_MATRIX) {
5204: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5205: *A_loc = mpimat->A;
5206: } else if (scall == MAT_REUSE_MATRIX) {
5207: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5208: }
5209: PetscFunctionReturn(PETSC_SUCCESS);
5210: }
5212: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5213: a = (Mat_SeqAIJ *)mpimat->A->data;
5214: b = (Mat_SeqAIJ *)mpimat->B->data;
5215: ai = a->i;
5216: aj = a->j;
5217: bi = b->i;
5218: bj = b->j;
5219: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5220: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5221: aa = aav;
5222: ba = bav;
5223: if (scall == MAT_INITIAL_MATRIX) {
5224: PetscCall(PetscMalloc1(1 + am, &ci));
5225: ci[0] = 0;
5226: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5227: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5228: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5229: k = 0;
5230: for (i = 0; i < am; i++) {
5231: ncols_o = bi[i + 1] - bi[i];
5232: ncols_d = ai[i + 1] - ai[i];
5233: /* off-diagonal portion of A */
5234: for (jo = 0; jo < ncols_o; jo++) {
5235: col = cmap[*bj];
5236: if (col >= cstart) break;
5237: cj[k] = col;
5238: bj++;
5239: ca[k++] = *ba++;
5240: }
5241: /* diagonal portion of A */
5242: for (j = 0; j < ncols_d; j++) {
5243: cj[k] = cstart + *aj++;
5244: ca[k++] = *aa++;
5245: }
5246: /* off-diagonal portion of A */
5247: for (j = jo; j < ncols_o; j++) {
5248: cj[k] = cmap[*bj++];
5249: ca[k++] = *ba++;
5250: }
5251: }
5252: /* put together the new matrix */
5253: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5254: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5255: /* Since these are PETSc arrays, change flags to free them as necessary. */
5256: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5257: mat->free_a = PETSC_TRUE;
5258: mat->free_ij = PETSC_TRUE;
5259: mat->nonew = 0;
5260: } else if (scall == MAT_REUSE_MATRIX) {
5261: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5262: ci = mat->i;
5263: cj = mat->j;
5264: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5265: for (i = 0; i < am; i++) {
5266: /* off-diagonal portion of A */
5267: ncols_o = bi[i + 1] - bi[i];
5268: for (jo = 0; jo < ncols_o; jo++) {
5269: col = cmap[*bj];
5270: if (col >= cstart) break;
5271: *cam++ = *ba++;
5272: bj++;
5273: }
5274: /* diagonal portion of A */
5275: ncols_d = ai[i + 1] - ai[i];
5276: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5277: /* off-diagonal portion of A */
5278: for (j = jo; j < ncols_o; j++) {
5279: *cam++ = *ba++;
5280: bj++;
5281: }
5282: }
5283: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5284: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5285: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5286: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5287: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5288: PetscFunctionReturn(PETSC_SUCCESS);
5289: }
5291: /*@
5292: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5293: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5295: Not Collective
5297: Input Parameters:
5298: + A - the matrix
5299: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5301: Output Parameters:
5302: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5303: - A_loc - the local sequential matrix generated
5305: Level: developer
5307: Note:
5308: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5309: part, then those associated with the off-diagonal part (in its local ordering)
5311: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5312: @*/
5313: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5314: {
5315: Mat Ao, Ad;
5316: const PetscInt *cmap;
5317: PetscMPIInt size;
5318: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5320: PetscFunctionBegin;
5321: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5322: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5323: if (size == 1) {
5324: if (scall == MAT_INITIAL_MATRIX) {
5325: PetscCall(PetscObjectReference((PetscObject)Ad));
5326: *A_loc = Ad;
5327: } else if (scall == MAT_REUSE_MATRIX) {
5328: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5329: }
5330: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5331: PetscFunctionReturn(PETSC_SUCCESS);
5332: }
5333: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5334: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5335: if (f) {
5336: PetscCall((*f)(A, scall, glob, A_loc));
5337: } else {
5338: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5339: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5340: Mat_SeqAIJ *c;
5341: PetscInt *ai = a->i, *aj = a->j;
5342: PetscInt *bi = b->i, *bj = b->j;
5343: PetscInt *ci, *cj;
5344: const PetscScalar *aa, *ba;
5345: PetscScalar *ca;
5346: PetscInt i, j, am, dn, on;
5348: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5349: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5350: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5351: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5352: if (scall == MAT_INITIAL_MATRIX) {
5353: PetscInt k;
5354: PetscCall(PetscMalloc1(1 + am, &ci));
5355: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5356: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5357: ci[0] = 0;
5358: for (i = 0, k = 0; i < am; i++) {
5359: const PetscInt ncols_o = bi[i + 1] - bi[i];
5360: const PetscInt ncols_d = ai[i + 1] - ai[i];
5361: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5362: /* diagonal portion of A */
5363: for (j = 0; j < ncols_d; j++, k++) {
5364: cj[k] = *aj++;
5365: ca[k] = *aa++;
5366: }
5367: /* off-diagonal portion of A */
5368: for (j = 0; j < ncols_o; j++, k++) {
5369: cj[k] = dn + *bj++;
5370: ca[k] = *ba++;
5371: }
5372: }
5373: /* put together the new matrix */
5374: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5375: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5376: /* Since these are PETSc arrays, change flags to free them as necessary. */
5377: c = (Mat_SeqAIJ *)(*A_loc)->data;
5378: c->free_a = PETSC_TRUE;
5379: c->free_ij = PETSC_TRUE;
5380: c->nonew = 0;
5381: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5382: } else if (scall == MAT_REUSE_MATRIX) {
5383: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5384: for (i = 0; i < am; i++) {
5385: const PetscInt ncols_d = ai[i + 1] - ai[i];
5386: const PetscInt ncols_o = bi[i + 1] - bi[i];
5387: /* diagonal portion of A */
5388: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5389: /* off-diagonal portion of A */
5390: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5391: }
5392: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5393: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5394: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5395: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5396: if (glob) {
5397: PetscInt cst, *gidx;
5399: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5400: PetscCall(PetscMalloc1(dn + on, &gidx));
5401: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5402: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5403: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5404: }
5405: }
5406: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5407: PetscFunctionReturn(PETSC_SUCCESS);
5408: }
5410: /*@C
5411: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5413: Not Collective
5415: Input Parameters:
5416: + A - the matrix
5417: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5418: . row - index set of rows to extract (or `NULL`)
5419: - col - index set of columns to extract (or `NULL`)
5421: Output Parameter:
5422: . A_loc - the local sequential matrix generated
5424: Level: developer
5426: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5427: @*/
5428: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5429: {
5430: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5431: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5432: IS isrowa, iscola;
5433: Mat *aloc;
5434: PetscBool match;
5436: PetscFunctionBegin;
5437: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5438: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5439: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5440: if (!row) {
5441: start = A->rmap->rstart;
5442: end = A->rmap->rend;
5443: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5444: } else {
5445: isrowa = *row;
5446: }
5447: if (!col) {
5448: start = A->cmap->rstart;
5449: cmap = a->garray;
5450: nzA = a->A->cmap->n;
5451: nzB = a->B->cmap->n;
5452: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5453: ncols = 0;
5454: for (i = 0; i < nzB; i++) {
5455: if (cmap[i] < start) idx[ncols++] = cmap[i];
5456: else break;
5457: }
5458: imark = i;
5459: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5460: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5461: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5462: } else {
5463: iscola = *col;
5464: }
5465: if (scall != MAT_INITIAL_MATRIX) {
5466: PetscCall(PetscMalloc1(1, &aloc));
5467: aloc[0] = *A_loc;
5468: }
5469: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5470: if (!col) { /* attach global id of condensed columns */
5471: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5472: }
5473: *A_loc = aloc[0];
5474: PetscCall(PetscFree(aloc));
5475: if (!row) PetscCall(ISDestroy(&isrowa));
5476: if (!col) PetscCall(ISDestroy(&iscola));
5477: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5478: PetscFunctionReturn(PETSC_SUCCESS);
5479: }
5481: /*
5482: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5483: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5484: * on a global size.
5485: * */
5486: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5487: {
5488: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5489: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5490: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5491: PetscMPIInt owner;
5492: PetscSFNode *iremote, *oiremote;
5493: const PetscInt *lrowindices;
5494: PetscSF sf, osf;
5495: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5496: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5497: MPI_Comm comm;
5498: ISLocalToGlobalMapping mapping;
5499: const PetscScalar *pd_a, *po_a;
5501: PetscFunctionBegin;
5502: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5503: /* plocalsize is the number of roots
5504: * nrows is the number of leaves
5505: * */
5506: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5507: PetscCall(ISGetLocalSize(rows, &nrows));
5508: PetscCall(PetscCalloc1(nrows, &iremote));
5509: PetscCall(ISGetIndices(rows, &lrowindices));
5510: for (i = 0; i < nrows; i++) {
5511: /* Find a remote index and an owner for a row
5512: * The row could be local or remote
5513: * */
5514: owner = 0;
5515: lidx = 0;
5516: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5517: iremote[i].index = lidx;
5518: iremote[i].rank = owner;
5519: }
5520: /* Create SF to communicate how many nonzero columns for each row */
5521: PetscCall(PetscSFCreate(comm, &sf));
5522: /* SF will figure out the number of nonzero columns for each row, and their
5523: * offsets
5524: * */
5525: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5526: PetscCall(PetscSFSetFromOptions(sf));
5527: PetscCall(PetscSFSetUp(sf));
5529: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5530: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5531: PetscCall(PetscCalloc1(nrows, &pnnz));
5532: roffsets[0] = 0;
5533: roffsets[1] = 0;
5534: for (i = 0; i < plocalsize; i++) {
5535: /* diagonal */
5536: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5537: /* off-diagonal */
5538: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5539: /* compute offsets so that we relative location for each row */
5540: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5541: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5542: }
5543: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5544: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5545: /* 'r' means root, and 'l' means leaf */
5546: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5547: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5548: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5549: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5550: PetscCall(PetscSFDestroy(&sf));
5551: PetscCall(PetscFree(roffsets));
5552: PetscCall(PetscFree(nrcols));
5553: dntotalcols = 0;
5554: ontotalcols = 0;
5555: ncol = 0;
5556: for (i = 0; i < nrows; i++) {
5557: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5558: ncol = PetscMax(pnnz[i], ncol);
5559: /* diagonal */
5560: dntotalcols += nlcols[i * 2 + 0];
5561: /* off-diagonal */
5562: ontotalcols += nlcols[i * 2 + 1];
5563: }
5564: /* We do not need to figure the right number of columns
5565: * since all the calculations will be done by going through the raw data
5566: * */
5567: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5568: PetscCall(MatSetUp(*P_oth));
5569: PetscCall(PetscFree(pnnz));
5570: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5571: /* diagonal */
5572: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5573: /* off-diagonal */
5574: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5575: /* diagonal */
5576: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5577: /* off-diagonal */
5578: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5579: dntotalcols = 0;
5580: ontotalcols = 0;
5581: ntotalcols = 0;
5582: for (i = 0; i < nrows; i++) {
5583: owner = 0;
5584: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5585: /* Set iremote for diag matrix */
5586: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5587: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5588: iremote[dntotalcols].rank = owner;
5589: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5590: ilocal[dntotalcols++] = ntotalcols++;
5591: }
5592: /* off-diagonal */
5593: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5594: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5595: oiremote[ontotalcols].rank = owner;
5596: oilocal[ontotalcols++] = ntotalcols++;
5597: }
5598: }
5599: PetscCall(ISRestoreIndices(rows, &lrowindices));
5600: PetscCall(PetscFree(loffsets));
5601: PetscCall(PetscFree(nlcols));
5602: PetscCall(PetscSFCreate(comm, &sf));
5603: /* P serves as roots and P_oth is leaves
5604: * Diag matrix
5605: * */
5606: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5607: PetscCall(PetscSFSetFromOptions(sf));
5608: PetscCall(PetscSFSetUp(sf));
5610: PetscCall(PetscSFCreate(comm, &osf));
5611: /* off-diagonal */
5612: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5613: PetscCall(PetscSFSetFromOptions(osf));
5614: PetscCall(PetscSFSetUp(osf));
5615: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5616: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5617: /* operate on the matrix internal data to save memory */
5618: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5619: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5620: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5621: /* Convert to global indices for diag matrix */
5622: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5623: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5624: /* We want P_oth store global indices */
5625: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5626: /* Use memory scalable approach */
5627: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5628: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5629: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5630: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5631: /* Convert back to local indices */
5632: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5633: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5634: nout = 0;
5635: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5636: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5637: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5638: /* Exchange values */
5639: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5640: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5641: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5642: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5643: /* Stop PETSc from shrinking memory */
5644: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5645: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5646: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5647: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5648: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5649: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5650: PetscCall(PetscSFDestroy(&sf));
5651: PetscCall(PetscSFDestroy(&osf));
5652: PetscFunctionReturn(PETSC_SUCCESS);
5653: }
5655: /*
5656: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5657: * This supports MPIAIJ and MAIJ
5658: * */
5659: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5660: {
5661: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5662: Mat_SeqAIJ *p_oth;
5663: IS rows, map;
5664: PetscHMapI hamp;
5665: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5666: MPI_Comm comm;
5667: PetscSF sf, osf;
5668: PetscBool has;
5670: PetscFunctionBegin;
5671: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5672: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5673: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5674: * and then create a submatrix (that often is an overlapping matrix)
5675: * */
5676: if (reuse == MAT_INITIAL_MATRIX) {
5677: /* Use a hash table to figure out unique keys */
5678: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5679: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5680: count = 0;
5681: /* Assume that a->g is sorted, otherwise the following does not make sense */
5682: for (i = 0; i < a->B->cmap->n; i++) {
5683: key = a->garray[i] / dof;
5684: PetscCall(PetscHMapIHas(hamp, key, &has));
5685: if (!has) {
5686: mapping[i] = count;
5687: PetscCall(PetscHMapISet(hamp, key, count++));
5688: } else {
5689: /* Current 'i' has the same value the previous step */
5690: mapping[i] = count - 1;
5691: }
5692: }
5693: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5694: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5695: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5696: PetscCall(PetscCalloc1(htsize, &rowindices));
5697: off = 0;
5698: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5699: PetscCall(PetscHMapIDestroy(&hamp));
5700: PetscCall(PetscSortInt(htsize, rowindices));
5701: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5702: /* In case, the matrix was already created but users want to recreate the matrix */
5703: PetscCall(MatDestroy(P_oth));
5704: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5705: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5706: PetscCall(ISDestroy(&map));
5707: PetscCall(ISDestroy(&rows));
5708: } else if (reuse == MAT_REUSE_MATRIX) {
5709: /* If matrix was already created, we simply update values using SF objects
5710: * that as attached to the matrix earlier.
5711: */
5712: const PetscScalar *pd_a, *po_a;
5714: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5715: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5716: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5717: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5718: /* Update values in place */
5719: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5720: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5721: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5722: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5723: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5724: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5725: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5726: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5727: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5728: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5729: PetscFunctionReturn(PETSC_SUCCESS);
5730: }
5732: /*@C
5733: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5735: Collective
5737: Input Parameters:
5738: + A - the first matrix in `MATMPIAIJ` format
5739: . B - the second matrix in `MATMPIAIJ` format
5740: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5742: Output Parameters:
5743: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5744: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5745: - B_seq - the sequential matrix generated
5747: Level: developer
5749: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5750: @*/
5751: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5752: {
5753: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5754: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5755: IS isrowb, iscolb;
5756: Mat *bseq = NULL;
5758: PetscFunctionBegin;
5759: 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 ")",
5760: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5761: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5763: if (scall == MAT_INITIAL_MATRIX) {
5764: start = A->cmap->rstart;
5765: cmap = a->garray;
5766: nzA = a->A->cmap->n;
5767: nzB = a->B->cmap->n;
5768: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5769: ncols = 0;
5770: for (i = 0; i < nzB; i++) { /* row < local row index */
5771: if (cmap[i] < start) idx[ncols++] = cmap[i];
5772: else break;
5773: }
5774: imark = i;
5775: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5776: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5777: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5778: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5779: } else {
5780: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5781: isrowb = *rowb;
5782: iscolb = *colb;
5783: PetscCall(PetscMalloc1(1, &bseq));
5784: bseq[0] = *B_seq;
5785: }
5786: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5787: *B_seq = bseq[0];
5788: PetscCall(PetscFree(bseq));
5789: if (!rowb) {
5790: PetscCall(ISDestroy(&isrowb));
5791: } else {
5792: *rowb = isrowb;
5793: }
5794: if (!colb) {
5795: PetscCall(ISDestroy(&iscolb));
5796: } else {
5797: *colb = iscolb;
5798: }
5799: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5800: PetscFunctionReturn(PETSC_SUCCESS);
5801: }
5803: /*
5804: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5805: of the OFF-DIAGONAL portion of local A
5807: Collective
5809: Input Parameters:
5810: + A,B - the matrices in `MATMPIAIJ` format
5811: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5813: Output Parameter:
5814: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5815: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5816: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5817: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5819: Developer Note:
5820: This directly accesses information inside the VecScatter associated with the matrix-vector product
5821: for this matrix. This is not desirable..
5823: Level: developer
5825: */
5826: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5827: {
5828: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5829: Mat_SeqAIJ *b_oth;
5830: VecScatter ctx;
5831: MPI_Comm comm;
5832: const PetscMPIInt *rprocs, *sprocs;
5833: const PetscInt *srow, *rstarts, *sstarts;
5834: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5835: PetscInt i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5836: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5837: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5838: PetscMPIInt size, tag, rank, nreqs;
5840: PetscFunctionBegin;
5841: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5842: PetscCallMPI(MPI_Comm_size(comm, &size));
5844: 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 ")",
5845: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5846: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5847: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5849: if (size == 1) {
5850: startsj_s = NULL;
5851: bufa_ptr = NULL;
5852: *B_oth = NULL;
5853: PetscFunctionReturn(PETSC_SUCCESS);
5854: }
5856: ctx = a->Mvctx;
5857: tag = ((PetscObject)ctx)->tag;
5859: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5860: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5861: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5862: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5863: PetscCall(PetscMalloc1(nreqs, &reqs));
5864: rwaits = reqs;
5865: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5867: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5868: if (scall == MAT_INITIAL_MATRIX) {
5869: /* i-array */
5870: /* post receives */
5871: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5872: for (i = 0; i < nrecvs; i++) {
5873: rowlen = rvalues + rstarts[i] * rbs;
5874: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5875: PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5876: }
5878: /* pack the outgoing message */
5879: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5881: sstartsj[0] = 0;
5882: rstartsj[0] = 0;
5883: len = 0; /* total length of j or a array to be sent */
5884: if (nsends) {
5885: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5886: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5887: }
5888: for (i = 0; i < nsends; i++) {
5889: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5890: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5891: for (j = 0; j < nrows; j++) {
5892: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5893: for (l = 0; l < sbs; l++) {
5894: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5896: rowlen[j * sbs + l] = ncols;
5898: len += ncols;
5899: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5900: }
5901: k++;
5902: }
5903: PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5905: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5906: }
5907: /* recvs and sends of i-array are completed */
5908: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5909: PetscCall(PetscFree(svalues));
5911: /* allocate buffers for sending j and a arrays */
5912: PetscCall(PetscMalloc1(len + 1, &bufj));
5913: PetscCall(PetscMalloc1(len + 1, &bufa));
5915: /* create i-array of B_oth */
5916: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5918: b_othi[0] = 0;
5919: len = 0; /* total length of j or a array to be received */
5920: k = 0;
5921: for (i = 0; i < nrecvs; i++) {
5922: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5923: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5924: for (j = 0; j < nrows; j++) {
5925: b_othi[k + 1] = b_othi[k] + rowlen[j];
5926: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5927: k++;
5928: }
5929: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5930: }
5931: PetscCall(PetscFree(rvalues));
5933: /* allocate space for j and a arrays of B_oth */
5934: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5935: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5937: /* j-array */
5938: /* post receives of j-array */
5939: for (i = 0; i < nrecvs; i++) {
5940: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5941: PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5942: }
5944: /* pack the outgoing message j-array */
5945: if (nsends) k = sstarts[0];
5946: for (i = 0; i < nsends; i++) {
5947: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5948: bufJ = bufj + sstartsj[i];
5949: for (j = 0; j < nrows; j++) {
5950: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5951: for (ll = 0; ll < sbs; ll++) {
5952: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5953: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5954: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5955: }
5956: }
5957: PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5958: }
5960: /* recvs and sends of j-array are completed */
5961: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5962: } else if (scall == MAT_REUSE_MATRIX) {
5963: sstartsj = *startsj_s;
5964: rstartsj = *startsj_r;
5965: bufa = *bufa_ptr;
5966: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5967: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5968: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5970: /* a-array */
5971: /* post receives of a-array */
5972: for (i = 0; i < nrecvs; i++) {
5973: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5974: PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5975: }
5977: /* pack the outgoing message a-array */
5978: if (nsends) k = sstarts[0];
5979: for (i = 0; i < nsends; i++) {
5980: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5981: bufA = bufa + sstartsj[i];
5982: for (j = 0; j < nrows; j++) {
5983: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5984: for (ll = 0; ll < sbs; ll++) {
5985: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5986: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5987: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5988: }
5989: }
5990: PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5991: }
5992: /* recvs and sends of a-array are completed */
5993: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5994: PetscCall(PetscFree(reqs));
5996: if (scall == MAT_INITIAL_MATRIX) {
5997: /* put together the new matrix */
5998: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6000: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6001: /* Since these are PETSc arrays, change flags to free them as necessary. */
6002: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6003: b_oth->free_a = PETSC_TRUE;
6004: b_oth->free_ij = PETSC_TRUE;
6005: b_oth->nonew = 0;
6007: PetscCall(PetscFree(bufj));
6008: if (!startsj_s || !bufa_ptr) {
6009: PetscCall(PetscFree2(sstartsj, rstartsj));
6010: PetscCall(PetscFree(bufa_ptr));
6011: } else {
6012: *startsj_s = sstartsj;
6013: *startsj_r = rstartsj;
6014: *bufa_ptr = bufa;
6015: }
6016: } else if (scall == MAT_REUSE_MATRIX) {
6017: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6018: }
6020: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6021: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6022: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6023: PetscFunctionReturn(PETSC_SUCCESS);
6024: }
6026: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6027: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6029: #if defined(PETSC_HAVE_MKL_SPARSE)
6030: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6031: #endif
6032: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6033: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6034: #if defined(PETSC_HAVE_ELEMENTAL)
6035: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6036: #endif
6037: #if defined(PETSC_HAVE_SCALAPACK)
6038: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6039: #endif
6040: #if defined(PETSC_HAVE_HYPRE)
6041: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6042: #endif
6043: #if defined(PETSC_HAVE_CUDA)
6044: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6045: #endif
6046: #if defined(PETSC_HAVE_HIP)
6047: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6048: #endif
6049: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6050: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6051: #endif
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6053: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6054: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6056: /*
6057: Computes (B'*A')' since computing B*A directly is untenable
6059: n p p
6060: [ ] [ ] [ ]
6061: m [ A ] * n [ B ] = m [ C ]
6062: [ ] [ ] [ ]
6064: */
6065: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6066: {
6067: Mat At, Bt, Ct;
6069: PetscFunctionBegin;
6070: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6071: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6072: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6073: PetscCall(MatDestroy(&At));
6074: PetscCall(MatDestroy(&Bt));
6075: PetscCall(MatTransposeSetPrecursor(Ct, C));
6076: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6077: PetscCall(MatDestroy(&Ct));
6078: PetscFunctionReturn(PETSC_SUCCESS);
6079: }
6081: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6082: {
6083: PetscBool cisdense;
6085: PetscFunctionBegin;
6086: 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);
6087: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6088: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6089: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6090: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6091: PetscCall(MatSetUp(C));
6093: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6094: PetscFunctionReturn(PETSC_SUCCESS);
6095: }
6097: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6098: {
6099: Mat_Product *product = C->product;
6100: Mat A = product->A, B = product->B;
6102: PetscFunctionBegin;
6103: 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 ")",
6104: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6105: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6106: C->ops->productsymbolic = MatProductSymbolic_AB;
6107: PetscFunctionReturn(PETSC_SUCCESS);
6108: }
6110: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6111: {
6112: Mat_Product *product = C->product;
6114: PetscFunctionBegin;
6115: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6116: PetscFunctionReturn(PETSC_SUCCESS);
6117: }
6119: /*
6120: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6122: Input Parameters:
6124: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6125: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6127: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6129: For Set1, j1[] contains column indices of the nonzeros.
6130: 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
6131: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6132: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6134: Similar for Set2.
6136: This routine merges the two sets of nonzeros row by row and removes repeats.
6138: Output Parameters: (memory is allocated by the caller)
6140: i[],j[]: the CSR of the merged matrix, which has m rows.
6141: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6142: imap2[]: similar to imap1[], but for Set2.
6143: Note we order nonzeros row-by-row and from left to right.
6144: */
6145: 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[])
6146: {
6147: PetscInt r, m; /* Row index of mat */
6148: PetscCount t, t1, t2, b1, e1, b2, e2;
6150: PetscFunctionBegin;
6151: PetscCall(MatGetLocalSize(mat, &m, NULL));
6152: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6153: i[0] = 0;
6154: for (r = 0; r < m; r++) { /* Do row by row merging */
6155: b1 = rowBegin1[r];
6156: e1 = rowEnd1[r];
6157: b2 = rowBegin2[r];
6158: e2 = rowEnd2[r];
6159: while (b1 < e1 && b2 < e2) {
6160: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6161: j[t] = j1[b1];
6162: imap1[t1] = t;
6163: imap2[t2] = t;
6164: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6165: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6166: t1++;
6167: t2++;
6168: t++;
6169: } else if (j1[b1] < j2[b2]) {
6170: j[t] = j1[b1];
6171: imap1[t1] = t;
6172: b1 += jmap1[t1 + 1] - jmap1[t1];
6173: t1++;
6174: t++;
6175: } else {
6176: j[t] = j2[b2];
6177: imap2[t2] = t;
6178: b2 += jmap2[t2 + 1] - jmap2[t2];
6179: t2++;
6180: t++;
6181: }
6182: }
6183: /* Merge the remaining in either j1[] or j2[] */
6184: while (b1 < e1) {
6185: j[t] = j1[b1];
6186: imap1[t1] = t;
6187: b1 += jmap1[t1 + 1] - jmap1[t1];
6188: t1++;
6189: t++;
6190: }
6191: while (b2 < e2) {
6192: j[t] = j2[b2];
6193: imap2[t2] = t;
6194: b2 += jmap2[t2 + 1] - jmap2[t2];
6195: t2++;
6196: t++;
6197: }
6198: i[r + 1] = t;
6199: }
6200: PetscFunctionReturn(PETSC_SUCCESS);
6201: }
6203: /*
6204: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6206: Input Parameters:
6207: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6208: 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[]
6209: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6211: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6212: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6214: Output Parameters:
6215: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6216: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6217: 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,
6218: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6220: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6221: Atot: number of entries belonging to the diagonal block.
6222: Annz: number of unique nonzeros belonging to the diagonal block.
6223: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6224: repeats (i.e., same 'i,j' pair).
6225: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6226: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6228: Atot: number of entries belonging to the diagonal block
6229: Annz: number of unique nonzeros belonging to the diagonal block.
6231: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6233: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6234: */
6235: 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_)
6236: {
6237: PetscInt cstart, cend, rstart, rend, row, col;
6238: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6239: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6240: PetscCount k, m, p, q, r, s, mid;
6241: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6243: PetscFunctionBegin;
6244: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6245: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6246: m = rend - rstart;
6248: /* Skip negative rows */
6249: for (k = 0; k < n; k++)
6250: if (i[k] >= 0) break;
6252: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6253: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6254: */
6255: while (k < n) {
6256: row = i[k];
6257: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6258: for (s = k; s < n; s++)
6259: if (i[s] != row) break;
6261: /* Shift diag columns to range of [-PETSC_MAX_INT, -1] */
6262: for (p = k; p < s; p++) {
6263: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT;
6264: 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]);
6265: }
6266: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6267: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6268: rowBegin[row - rstart] = k;
6269: rowMid[row - rstart] = mid;
6270: rowEnd[row - rstart] = s;
6272: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6273: Atot += mid - k;
6274: Btot += s - mid;
6276: /* Count unique nonzeros of this diag row */
6277: for (p = k; p < mid;) {
6278: col = j[p];
6279: do {
6280: j[p] += PETSC_MAX_INT; /* Revert the modified diagonal indices */
6281: p++;
6282: } while (p < mid && j[p] == col);
6283: Annz++;
6284: }
6286: /* Count unique nonzeros of this offdiag row */
6287: for (p = mid; p < s;) {
6288: col = j[p];
6289: do {
6290: p++;
6291: } while (p < s && j[p] == col);
6292: Bnnz++;
6293: }
6294: k = s;
6295: }
6297: /* Allocation according to Atot, Btot, Annz, Bnnz */
6298: PetscCall(PetscMalloc1(Atot, &Aperm));
6299: PetscCall(PetscMalloc1(Btot, &Bperm));
6300: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6301: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6303: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6304: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6305: for (r = 0; r < m; r++) {
6306: k = rowBegin[r];
6307: mid = rowMid[r];
6308: s = rowEnd[r];
6309: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6310: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6311: Atot += mid - k;
6312: Btot += s - mid;
6314: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6315: for (p = k; p < mid;) {
6316: col = j[p];
6317: q = p;
6318: do {
6319: p++;
6320: } while (p < mid && j[p] == col);
6321: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6322: Annz++;
6323: }
6325: for (p = mid; p < s;) {
6326: col = j[p];
6327: q = p;
6328: do {
6329: p++;
6330: } while (p < s && j[p] == col);
6331: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6332: Bnnz++;
6333: }
6334: }
6335: /* Output */
6336: *Aperm_ = Aperm;
6337: *Annz_ = Annz;
6338: *Atot_ = Atot;
6339: *Ajmap_ = Ajmap;
6340: *Bperm_ = Bperm;
6341: *Bnnz_ = Bnnz;
6342: *Btot_ = Btot;
6343: *Bjmap_ = Bjmap;
6344: PetscFunctionReturn(PETSC_SUCCESS);
6345: }
6347: /*
6348: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6350: Input Parameters:
6351: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6352: nnz: number of unique nonzeros in the merged matrix
6353: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6354: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6356: Output Parameter: (memory is allocated by the caller)
6357: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6359: Example:
6360: nnz1 = 4
6361: nnz = 6
6362: imap = [1,3,4,5]
6363: jmap = [0,3,5,6,7]
6364: then,
6365: jmap_new = [0,0,3,3,5,6,7]
6366: */
6367: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6368: {
6369: PetscCount k, p;
6371: PetscFunctionBegin;
6372: jmap_new[0] = 0;
6373: p = nnz; /* p loops over jmap_new[] backwards */
6374: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6375: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6376: }
6377: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6378: PetscFunctionReturn(PETSC_SUCCESS);
6379: }
6381: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6382: {
6383: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6385: PetscFunctionBegin;
6386: PetscCall(PetscSFDestroy(&coo->sf));
6387: PetscCall(PetscFree(coo->Aperm1));
6388: PetscCall(PetscFree(coo->Bperm1));
6389: PetscCall(PetscFree(coo->Ajmap1));
6390: PetscCall(PetscFree(coo->Bjmap1));
6391: PetscCall(PetscFree(coo->Aimap2));
6392: PetscCall(PetscFree(coo->Bimap2));
6393: PetscCall(PetscFree(coo->Aperm2));
6394: PetscCall(PetscFree(coo->Bperm2));
6395: PetscCall(PetscFree(coo->Ajmap2));
6396: PetscCall(PetscFree(coo->Bjmap2));
6397: PetscCall(PetscFree(coo->Cperm1));
6398: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6399: PetscCall(PetscFree(coo));
6400: PetscFunctionReturn(PETSC_SUCCESS);
6401: }
6403: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6404: {
6405: MPI_Comm comm;
6406: PetscMPIInt rank, size;
6407: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6408: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6409: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6410: PetscContainer container;
6411: MatCOOStruct_MPIAIJ *coo;
6413: PetscFunctionBegin;
6414: PetscCall(PetscFree(mpiaij->garray));
6415: PetscCall(VecDestroy(&mpiaij->lvec));
6416: #if defined(PETSC_USE_CTABLE)
6417: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6418: #else
6419: PetscCall(PetscFree(mpiaij->colmap));
6420: #endif
6421: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6422: mat->assembled = PETSC_FALSE;
6423: mat->was_assembled = PETSC_FALSE;
6425: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6426: PetscCallMPI(MPI_Comm_size(comm, &size));
6427: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6428: PetscCall(PetscLayoutSetUp(mat->rmap));
6429: PetscCall(PetscLayoutSetUp(mat->cmap));
6430: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6431: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6432: PetscCall(MatGetLocalSize(mat, &m, &n));
6433: PetscCall(MatGetSize(mat, &M, &N));
6435: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6436: /* entries come first, then local rows, then remote rows. */
6437: PetscCount n1 = coo_n, *perm1;
6438: PetscInt *i1 = coo_i, *j1 = coo_j;
6440: PetscCall(PetscMalloc1(n1, &perm1));
6441: for (k = 0; k < n1; k++) perm1[k] = k;
6443: /* Manipulate indices so that entries with negative row or col indices will have smallest
6444: row indices, local entries will have greater but negative row indices, and remote entries
6445: will have positive row indices.
6446: */
6447: for (k = 0; k < n1; k++) {
6448: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT; /* e.g., -2^31, minimal to move them ahead */
6449: 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] */
6450: else {
6451: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6452: if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6453: }
6454: }
6456: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6457: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6459: /* Advance k to the first entry we need to take care of */
6460: for (k = 0; k < n1; k++)
6461: if (i1[k] > PETSC_MIN_INT) break;
6462: PetscInt i1start = k;
6464: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6465: for (; k < rem; k++) i1[k] += PETSC_MAX_INT; /* Revert row indices of local rows*/
6467: /* Send remote rows to their owner */
6468: /* Find which rows should be sent to which remote ranks*/
6469: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6470: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6471: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6472: const PetscInt *ranges;
6473: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6475: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6476: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6477: for (k = rem; k < n1;) {
6478: PetscMPIInt owner;
6479: PetscInt firstRow, lastRow;
6481: /* Locate a row range */
6482: firstRow = i1[k]; /* first row of this owner */
6483: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6484: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6486: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6487: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6489: /* All entries in [k,p) belong to this remote owner */
6490: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6491: PetscMPIInt *sendto2;
6492: PetscInt *nentries2;
6493: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6495: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6496: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6497: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6498: PetscCall(PetscFree2(sendto, nentries2));
6499: sendto = sendto2;
6500: nentries = nentries2;
6501: maxNsend = maxNsend2;
6502: }
6503: sendto[nsend] = owner;
6504: nentries[nsend] = p - k;
6505: PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6506: nsend++;
6507: k = p;
6508: }
6510: /* Build 1st SF to know offsets on remote to send data */
6511: PetscSF sf1;
6512: PetscInt nroots = 1, nroots2 = 0;
6513: PetscInt nleaves = nsend, nleaves2 = 0;
6514: PetscInt *offsets;
6515: PetscSFNode *iremote;
6517: PetscCall(PetscSFCreate(comm, &sf1));
6518: PetscCall(PetscMalloc1(nsend, &iremote));
6519: PetscCall(PetscMalloc1(nsend, &offsets));
6520: for (k = 0; k < nsend; k++) {
6521: iremote[k].rank = sendto[k];
6522: iremote[k].index = 0;
6523: nleaves2 += nentries[k];
6524: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6525: }
6526: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6527: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6528: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6529: PetscCall(PetscSFDestroy(&sf1));
6530: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6532: /* Build 2nd SF to send remote COOs to their owner */
6533: PetscSF sf2;
6534: nroots = nroots2;
6535: nleaves = nleaves2;
6536: PetscCall(PetscSFCreate(comm, &sf2));
6537: PetscCall(PetscSFSetFromOptions(sf2));
6538: PetscCall(PetscMalloc1(nleaves, &iremote));
6539: p = 0;
6540: for (k = 0; k < nsend; k++) {
6541: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6542: for (q = 0; q < nentries[k]; q++, p++) {
6543: iremote[p].rank = sendto[k];
6544: iremote[p].index = offsets[k] + q;
6545: }
6546: }
6547: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6549: /* Send the remote COOs to their owner */
6550: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6551: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6552: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6553: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6554: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6555: PetscInt *i1prem = i1 ? i1 + rem : NULL; /* silence ubsan warnings about pointer arithmetic on null pointer */
6556: PetscInt *j1prem = j1 ? j1 + rem : NULL;
6557: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6558: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6559: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6560: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6562: PetscCall(PetscFree(offsets));
6563: PetscCall(PetscFree2(sendto, nentries));
6565: /* Sort received COOs by row along with the permutation array */
6566: for (k = 0; k < n2; k++) perm2[k] = k;
6567: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6569: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6570: PetscCount *Cperm1;
6571: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6572: PetscCount *perm1prem = perm1 ? perm1 + rem : NULL;
6573: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6574: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6576: /* Support for HYPRE matrices, kind of a hack.
6577: Swap min column with diagonal so that diagonal values will go first */
6578: PetscBool hypre;
6579: const char *name;
6580: PetscCall(PetscObjectGetName((PetscObject)mat, &name));
6581: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
6582: if (hypre) {
6583: PetscInt *minj;
6584: PetscBT hasdiag;
6586: PetscCall(PetscBTCreate(m, &hasdiag));
6587: PetscCall(PetscMalloc1(m, &minj));
6588: for (k = 0; k < m; k++) minj[k] = PETSC_MAX_INT;
6589: for (k = i1start; k < rem; k++) {
6590: if (j1[k] < cstart || j1[k] >= cend) continue;
6591: const PetscInt rindex = i1[k] - rstart;
6592: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6593: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6594: }
6595: for (k = 0; k < n2; k++) {
6596: if (j2[k] < cstart || j2[k] >= cend) continue;
6597: const PetscInt rindex = i2[k] - rstart;
6598: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6599: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6600: }
6601: for (k = i1start; k < rem; k++) {
6602: const PetscInt rindex = i1[k] - rstart;
6603: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6604: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6605: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6606: }
6607: for (k = 0; k < n2; k++) {
6608: const PetscInt rindex = i2[k] - rstart;
6609: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6610: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6611: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6612: }
6613: PetscCall(PetscBTDestroy(&hasdiag));
6614: PetscCall(PetscFree(minj));
6615: }
6617: /* Split local COOs and received COOs into diag/offdiag portions */
6618: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6619: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6620: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6621: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6622: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6623: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6625: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6626: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6627: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6628: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6630: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6631: PetscInt *Ai, *Bi;
6632: PetscInt *Aj, *Bj;
6634: PetscCall(PetscMalloc1(m + 1, &Ai));
6635: PetscCall(PetscMalloc1(m + 1, &Bi));
6636: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6637: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6639: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6640: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6641: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6642: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6643: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6645: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6646: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6648: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6649: /* expect nonzeros in A/B most likely have local contributing entries */
6650: PetscInt Annz = Ai[m];
6651: PetscInt Bnnz = Bi[m];
6652: PetscCount *Ajmap1_new, *Bjmap1_new;
6654: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6655: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6657: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6658: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6660: PetscCall(PetscFree(Aimap1));
6661: PetscCall(PetscFree(Ajmap1));
6662: PetscCall(PetscFree(Bimap1));
6663: PetscCall(PetscFree(Bjmap1));
6664: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6665: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6666: PetscCall(PetscFree(perm1));
6667: PetscCall(PetscFree3(i2, j2, perm2));
6669: Ajmap1 = Ajmap1_new;
6670: Bjmap1 = Bjmap1_new;
6672: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6673: if (Annz < Annz1 + Annz2) {
6674: PetscInt *Aj_new;
6675: PetscCall(PetscMalloc1(Annz, &Aj_new));
6676: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6677: PetscCall(PetscFree(Aj));
6678: Aj = Aj_new;
6679: }
6681: if (Bnnz < Bnnz1 + Bnnz2) {
6682: PetscInt *Bj_new;
6683: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6684: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6685: PetscCall(PetscFree(Bj));
6686: Bj = Bj_new;
6687: }
6689: /* Create new submatrices for on-process and off-process coupling */
6690: PetscScalar *Aa, *Ba;
6691: MatType rtype;
6692: Mat_SeqAIJ *a, *b;
6693: PetscObjectState state;
6694: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6695: PetscCall(PetscCalloc1(Bnnz, &Ba));
6696: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6697: if (cstart) {
6698: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6699: }
6701: PetscCall(MatGetRootType_Private(mat, &rtype));
6703: MatSeqXAIJGetOptions_Private(mpiaij->A);
6704: PetscCall(MatDestroy(&mpiaij->A));
6705: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6706: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6707: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6709: MatSeqXAIJGetOptions_Private(mpiaij->B);
6710: PetscCall(MatDestroy(&mpiaij->B));
6711: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6712: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6713: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6715: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6716: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6717: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6718: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6720: a = (Mat_SeqAIJ *)mpiaij->A->data;
6721: b = (Mat_SeqAIJ *)mpiaij->B->data;
6722: a->free_a = PETSC_TRUE;
6723: a->free_ij = PETSC_TRUE;
6724: b->free_a = PETSC_TRUE;
6725: b->free_ij = PETSC_TRUE;
6727: /* conversion must happen AFTER multiply setup */
6728: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6729: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6730: PetscCall(VecDestroy(&mpiaij->lvec));
6731: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6733: // Put the COO struct in a container and then attach that to the matrix
6734: PetscCall(PetscMalloc1(1, &coo));
6735: coo->n = coo_n;
6736: coo->sf = sf2;
6737: coo->sendlen = nleaves;
6738: coo->recvlen = nroots;
6739: coo->Annz = Annz;
6740: coo->Bnnz = Bnnz;
6741: coo->Annz2 = Annz2;
6742: coo->Bnnz2 = Bnnz2;
6743: coo->Atot1 = Atot1;
6744: coo->Atot2 = Atot2;
6745: coo->Btot1 = Btot1;
6746: coo->Btot2 = Btot2;
6747: coo->Ajmap1 = Ajmap1;
6748: coo->Aperm1 = Aperm1;
6749: coo->Bjmap1 = Bjmap1;
6750: coo->Bperm1 = Bperm1;
6751: coo->Aimap2 = Aimap2;
6752: coo->Ajmap2 = Ajmap2;
6753: coo->Aperm2 = Aperm2;
6754: coo->Bimap2 = Bimap2;
6755: coo->Bjmap2 = Bjmap2;
6756: coo->Bperm2 = Bperm2;
6757: coo->Cperm1 = Cperm1;
6758: // Allocate in preallocation. If not used, it has zero cost on host
6759: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6760: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6761: PetscCall(PetscContainerSetPointer(container, coo));
6762: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6763: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6764: PetscCall(PetscContainerDestroy(&container));
6765: PetscFunctionReturn(PETSC_SUCCESS);
6766: }
6768: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6769: {
6770: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6771: Mat A = mpiaij->A, B = mpiaij->B;
6772: PetscScalar *Aa, *Ba;
6773: PetscScalar *sendbuf, *recvbuf;
6774: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6775: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6776: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6777: const PetscCount *Cperm1;
6778: PetscContainer container;
6779: MatCOOStruct_MPIAIJ *coo;
6781: PetscFunctionBegin;
6782: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6783: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6784: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6785: sendbuf = coo->sendbuf;
6786: recvbuf = coo->recvbuf;
6787: Ajmap1 = coo->Ajmap1;
6788: Ajmap2 = coo->Ajmap2;
6789: Aimap2 = coo->Aimap2;
6790: Bjmap1 = coo->Bjmap1;
6791: Bjmap2 = coo->Bjmap2;
6792: Bimap2 = coo->Bimap2;
6793: Aperm1 = coo->Aperm1;
6794: Aperm2 = coo->Aperm2;
6795: Bperm1 = coo->Bperm1;
6796: Bperm2 = coo->Bperm2;
6797: Cperm1 = coo->Cperm1;
6799: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6800: PetscCall(MatSeqAIJGetArray(B, &Ba));
6802: /* Pack entries to be sent to remote */
6803: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6805: /* Send remote entries to their owner and overlap the communication with local computation */
6806: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6807: /* Add local entries to A and B */
6808: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6809: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6810: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6811: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6812: }
6813: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6814: PetscScalar sum = 0.0;
6815: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6816: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6817: }
6818: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6820: /* Add received remote entries to A and B */
6821: for (PetscCount i = 0; i < coo->Annz2; i++) {
6822: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6823: }
6824: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6825: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6826: }
6827: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6828: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6829: PetscFunctionReturn(PETSC_SUCCESS);
6830: }
6832: /*MC
6833: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6835: Options Database Keys:
6836: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6838: Level: beginner
6840: Notes:
6841: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6842: in this case the values associated with the rows and columns one passes in are set to zero
6843: in the matrix
6845: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6846: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6848: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6849: M*/
6850: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6851: {
6852: Mat_MPIAIJ *b;
6853: PetscMPIInt size;
6855: PetscFunctionBegin;
6856: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6858: PetscCall(PetscNew(&b));
6859: B->data = (void *)b;
6860: B->ops[0] = MatOps_Values;
6861: B->assembled = PETSC_FALSE;
6862: B->insertmode = NOT_SET_VALUES;
6863: b->size = size;
6865: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6867: /* build cache for off array entries formed */
6868: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6870: b->donotstash = PETSC_FALSE;
6871: b->colmap = NULL;
6872: b->garray = NULL;
6873: b->roworiented = PETSC_TRUE;
6875: /* stuff used for matrix vector multiply */
6876: b->lvec = NULL;
6877: b->Mvctx = NULL;
6879: /* stuff for MatGetRow() */
6880: b->rowindices = NULL;
6881: b->rowvalues = NULL;
6882: b->getrowactive = PETSC_FALSE;
6884: /* flexible pointer used in CUSPARSE classes */
6885: b->spptr = NULL;
6887: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6888: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6890: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6891: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6897: #if defined(PETSC_HAVE_CUDA)
6898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6899: #endif
6900: #if defined(PETSC_HAVE_HIP)
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6902: #endif
6903: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6905: #endif
6906: #if defined(PETSC_HAVE_MKL_SPARSE)
6907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6908: #endif
6909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6910: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6912: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6913: #if defined(PETSC_HAVE_ELEMENTAL)
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6915: #endif
6916: #if defined(PETSC_HAVE_SCALAPACK)
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6918: #endif
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6921: #if defined(PETSC_HAVE_HYPRE)
6922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6924: #endif
6925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6927: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6928: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6929: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6930: PetscFunctionReturn(PETSC_SUCCESS);
6931: }
6933: /*@
6934: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6935: and "off-diagonal" part of the matrix in CSR format.
6937: Collective
6939: Input Parameters:
6940: + comm - MPI communicator
6941: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6942: . n - This value should be the same as the local size used in creating the
6943: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6944: calculated if `N` is given) For square matrices `n` is almost always `m`.
6945: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6946: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6947: . 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
6948: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6949: . a - matrix values
6950: . 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
6951: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6952: - oa - matrix values
6954: Output Parameter:
6955: . mat - the matrix
6957: Level: advanced
6959: Notes:
6960: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6961: must free the arrays once the matrix has been destroyed and not before.
6963: The `i` and `j` indices are 0 based
6965: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6967: This sets local rows and cannot be used to set off-processor values.
6969: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6970: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6971: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6972: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6973: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6974: communication if it is known that only local entries will be set.
6976: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6977: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6978: @*/
6979: 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)
6980: {
6981: Mat_MPIAIJ *maij;
6983: PetscFunctionBegin;
6984: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6985: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6986: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6987: PetscCall(MatCreate(comm, mat));
6988: PetscCall(MatSetSizes(*mat, m, n, M, N));
6989: PetscCall(MatSetType(*mat, MATMPIAIJ));
6990: maij = (Mat_MPIAIJ *)(*mat)->data;
6992: (*mat)->preallocated = PETSC_TRUE;
6994: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6995: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6997: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6998: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7000: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7001: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7002: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7003: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7004: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7005: PetscFunctionReturn(PETSC_SUCCESS);
7006: }
7008: typedef struct {
7009: Mat *mp; /* intermediate products */
7010: PetscBool *mptmp; /* is the intermediate product temporary ? */
7011: PetscInt cp; /* number of intermediate products */
7013: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7014: PetscInt *startsj_s, *startsj_r;
7015: PetscScalar *bufa;
7016: Mat P_oth;
7018: /* may take advantage of merging product->B */
7019: Mat Bloc; /* B-local by merging diag and off-diag */
7021: /* cusparse does not have support to split between symbolic and numeric phases.
7022: When api_user is true, we don't need to update the numerical values
7023: of the temporary storage */
7024: PetscBool reusesym;
7026: /* support for COO values insertion */
7027: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7028: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7029: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7030: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7031: PetscSF sf; /* used for non-local values insertion and memory malloc */
7032: PetscMemType mtype;
7034: /* customization */
7035: PetscBool abmerge;
7036: PetscBool P_oth_bind;
7037: } MatMatMPIAIJBACKEND;
7039: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7040: {
7041: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7042: PetscInt i;
7044: PetscFunctionBegin;
7045: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7046: PetscCall(PetscFree(mmdata->bufa));
7047: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7048: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7049: PetscCall(MatDestroy(&mmdata->P_oth));
7050: PetscCall(MatDestroy(&mmdata->Bloc));
7051: PetscCall(PetscSFDestroy(&mmdata->sf));
7052: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7053: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7054: PetscCall(PetscFree(mmdata->own[0]));
7055: PetscCall(PetscFree(mmdata->own));
7056: PetscCall(PetscFree(mmdata->off[0]));
7057: PetscCall(PetscFree(mmdata->off));
7058: PetscCall(PetscFree(mmdata));
7059: PetscFunctionReturn(PETSC_SUCCESS);
7060: }
7062: /* Copy selected n entries with indices in idx[] of A to v[].
7063: If idx is NULL, copy the whole data array of A to v[]
7064: */
7065: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7066: {
7067: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7069: PetscFunctionBegin;
7070: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7071: if (f) {
7072: PetscCall((*f)(A, n, idx, v));
7073: } else {
7074: const PetscScalar *vv;
7076: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7077: if (n && idx) {
7078: PetscScalar *w = v;
7079: const PetscInt *oi = idx;
7080: PetscInt j;
7082: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7083: } else {
7084: PetscCall(PetscArraycpy(v, vv, n));
7085: }
7086: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7087: }
7088: PetscFunctionReturn(PETSC_SUCCESS);
7089: }
7091: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7092: {
7093: MatMatMPIAIJBACKEND *mmdata;
7094: PetscInt i, n_d, n_o;
7096: PetscFunctionBegin;
7097: MatCheckProduct(C, 1);
7098: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7099: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7100: if (!mmdata->reusesym) { /* update temporary matrices */
7101: 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));
7102: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7103: }
7104: mmdata->reusesym = PETSC_FALSE;
7106: for (i = 0; i < mmdata->cp; i++) {
7107: 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]);
7108: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7109: }
7110: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7111: PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];
7113: if (mmdata->mptmp[i]) continue;
7114: if (noff) {
7115: PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];
7117: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7118: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7119: n_o += noff;
7120: n_d += nown;
7121: } else {
7122: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7124: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7125: n_d += mm->nz;
7126: }
7127: }
7128: if (mmdata->hasoffproc) { /* offprocess insertion */
7129: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7130: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7131: }
7132: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7133: PetscFunctionReturn(PETSC_SUCCESS);
7134: }
7136: /* Support for Pt * A, A * P, or Pt * A * P */
7137: #define MAX_NUMBER_INTERMEDIATE 4
7138: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7139: {
7140: Mat_Product *product = C->product;
7141: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7142: Mat_MPIAIJ *a, *p;
7143: MatMatMPIAIJBACKEND *mmdata;
7144: ISLocalToGlobalMapping P_oth_l2g = NULL;
7145: IS glob = NULL;
7146: const char *prefix;
7147: char pprefix[256];
7148: const PetscInt *globidx, *P_oth_idx;
7149: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7150: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7151: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7152: /* type-0: consecutive, start from 0; type-1: consecutive with */
7153: /* a base offset; type-2: sparse with a local to global map table */
7154: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7156: MatProductType ptype;
7157: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7158: PetscMPIInt size;
7160: PetscFunctionBegin;
7161: MatCheckProduct(C, 1);
7162: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7163: ptype = product->type;
7164: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7165: ptype = MATPRODUCT_AB;
7166: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7167: }
7168: switch (ptype) {
7169: case MATPRODUCT_AB:
7170: A = product->A;
7171: P = product->B;
7172: m = A->rmap->n;
7173: n = P->cmap->n;
7174: M = A->rmap->N;
7175: N = P->cmap->N;
7176: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7177: break;
7178: case MATPRODUCT_AtB:
7179: P = product->A;
7180: A = product->B;
7181: m = P->cmap->n;
7182: n = A->cmap->n;
7183: M = P->cmap->N;
7184: N = A->cmap->N;
7185: hasoffproc = PETSC_TRUE;
7186: break;
7187: case MATPRODUCT_PtAP:
7188: A = product->A;
7189: P = product->B;
7190: m = P->cmap->n;
7191: n = P->cmap->n;
7192: M = P->cmap->N;
7193: N = P->cmap->N;
7194: hasoffproc = PETSC_TRUE;
7195: break;
7196: default:
7197: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7198: }
7199: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7200: if (size == 1) hasoffproc = PETSC_FALSE;
7202: /* defaults */
7203: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7204: mp[i] = NULL;
7205: mptmp[i] = PETSC_FALSE;
7206: rmapt[i] = -1;
7207: cmapt[i] = -1;
7208: rmapa[i] = NULL;
7209: cmapa[i] = NULL;
7210: }
7212: /* customization */
7213: PetscCall(PetscNew(&mmdata));
7214: mmdata->reusesym = product->api_user;
7215: if (ptype == MATPRODUCT_AB) {
7216: if (product->api_user) {
7217: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7218: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7219: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7220: PetscOptionsEnd();
7221: } else {
7222: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7223: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7224: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7225: PetscOptionsEnd();
7226: }
7227: } else if (ptype == MATPRODUCT_PtAP) {
7228: if (product->api_user) {
7229: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7230: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7231: PetscOptionsEnd();
7232: } else {
7233: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7234: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7235: PetscOptionsEnd();
7236: }
7237: }
7238: a = (Mat_MPIAIJ *)A->data;
7239: p = (Mat_MPIAIJ *)P->data;
7240: PetscCall(MatSetSizes(C, m, n, M, N));
7241: PetscCall(PetscLayoutSetUp(C->rmap));
7242: PetscCall(PetscLayoutSetUp(C->cmap));
7243: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7244: PetscCall(MatGetOptionsPrefix(C, &prefix));
7246: cp = 0;
7247: switch (ptype) {
7248: case MATPRODUCT_AB: /* A * P */
7249: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7251: /* A_diag * P_local (merged or not) */
7252: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7253: /* P is product->B */
7254: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7255: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7256: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7257: PetscCall(MatProductSetFill(mp[cp], product->fill));
7258: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7259: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7260: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7261: mp[cp]->product->api_user = product->api_user;
7262: PetscCall(MatProductSetFromOptions(mp[cp]));
7263: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7264: PetscCall(ISGetIndices(glob, &globidx));
7265: rmapt[cp] = 1;
7266: cmapt[cp] = 2;
7267: cmapa[cp] = globidx;
7268: mptmp[cp] = PETSC_FALSE;
7269: cp++;
7270: } else { /* A_diag * P_diag and A_diag * P_off */
7271: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7272: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7273: PetscCall(MatProductSetFill(mp[cp], product->fill));
7274: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7275: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7276: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7277: mp[cp]->product->api_user = product->api_user;
7278: PetscCall(MatProductSetFromOptions(mp[cp]));
7279: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7280: rmapt[cp] = 1;
7281: cmapt[cp] = 1;
7282: mptmp[cp] = PETSC_FALSE;
7283: cp++;
7284: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7285: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7286: PetscCall(MatProductSetFill(mp[cp], product->fill));
7287: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7288: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7289: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7290: mp[cp]->product->api_user = product->api_user;
7291: PetscCall(MatProductSetFromOptions(mp[cp]));
7292: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7293: rmapt[cp] = 1;
7294: cmapt[cp] = 2;
7295: cmapa[cp] = p->garray;
7296: mptmp[cp] = PETSC_FALSE;
7297: cp++;
7298: }
7300: /* A_off * P_other */
7301: if (mmdata->P_oth) {
7302: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7303: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7304: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7305: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7306: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7307: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7308: PetscCall(MatProductSetFill(mp[cp], product->fill));
7309: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7310: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7311: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7312: mp[cp]->product->api_user = product->api_user;
7313: PetscCall(MatProductSetFromOptions(mp[cp]));
7314: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7315: rmapt[cp] = 1;
7316: cmapt[cp] = 2;
7317: cmapa[cp] = P_oth_idx;
7318: mptmp[cp] = PETSC_FALSE;
7319: cp++;
7320: }
7321: break;
7323: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7324: /* A is product->B */
7325: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7326: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7327: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7328: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7329: PetscCall(MatProductSetFill(mp[cp], product->fill));
7330: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7331: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7332: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7333: mp[cp]->product->api_user = product->api_user;
7334: PetscCall(MatProductSetFromOptions(mp[cp]));
7335: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7336: PetscCall(ISGetIndices(glob, &globidx));
7337: rmapt[cp] = 2;
7338: rmapa[cp] = globidx;
7339: cmapt[cp] = 2;
7340: cmapa[cp] = globidx;
7341: mptmp[cp] = PETSC_FALSE;
7342: cp++;
7343: } else {
7344: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7345: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7346: PetscCall(MatProductSetFill(mp[cp], product->fill));
7347: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7348: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7349: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7350: mp[cp]->product->api_user = product->api_user;
7351: PetscCall(MatProductSetFromOptions(mp[cp]));
7352: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7353: PetscCall(ISGetIndices(glob, &globidx));
7354: rmapt[cp] = 1;
7355: cmapt[cp] = 2;
7356: cmapa[cp] = globidx;
7357: mptmp[cp] = PETSC_FALSE;
7358: cp++;
7359: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7360: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7361: PetscCall(MatProductSetFill(mp[cp], product->fill));
7362: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7363: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7364: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7365: mp[cp]->product->api_user = product->api_user;
7366: PetscCall(MatProductSetFromOptions(mp[cp]));
7367: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7368: rmapt[cp] = 2;
7369: rmapa[cp] = p->garray;
7370: cmapt[cp] = 2;
7371: cmapa[cp] = globidx;
7372: mptmp[cp] = PETSC_FALSE;
7373: cp++;
7374: }
7375: break;
7376: case MATPRODUCT_PtAP:
7377: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7378: /* P is product->B */
7379: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7380: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7381: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7382: PetscCall(MatProductSetFill(mp[cp], product->fill));
7383: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7384: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7385: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7386: mp[cp]->product->api_user = product->api_user;
7387: PetscCall(MatProductSetFromOptions(mp[cp]));
7388: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7389: PetscCall(ISGetIndices(glob, &globidx));
7390: rmapt[cp] = 2;
7391: rmapa[cp] = globidx;
7392: cmapt[cp] = 2;
7393: cmapa[cp] = globidx;
7394: mptmp[cp] = PETSC_FALSE;
7395: cp++;
7396: if (mmdata->P_oth) {
7397: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7398: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7399: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7400: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7401: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7402: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7403: PetscCall(MatProductSetFill(mp[cp], product->fill));
7404: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7405: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7406: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7407: mp[cp]->product->api_user = product->api_user;
7408: PetscCall(MatProductSetFromOptions(mp[cp]));
7409: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7410: mptmp[cp] = PETSC_TRUE;
7411: cp++;
7412: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7413: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7414: PetscCall(MatProductSetFill(mp[cp], product->fill));
7415: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7416: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7417: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7418: mp[cp]->product->api_user = product->api_user;
7419: PetscCall(MatProductSetFromOptions(mp[cp]));
7420: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7421: rmapt[cp] = 2;
7422: rmapa[cp] = globidx;
7423: cmapt[cp] = 2;
7424: cmapa[cp] = P_oth_idx;
7425: mptmp[cp] = PETSC_FALSE;
7426: cp++;
7427: }
7428: break;
7429: default:
7430: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7431: }
7432: /* sanity check */
7433: if (size > 1)
7434: 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);
7436: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7437: for (i = 0; i < cp; i++) {
7438: mmdata->mp[i] = mp[i];
7439: mmdata->mptmp[i] = mptmp[i];
7440: }
7441: mmdata->cp = cp;
7442: C->product->data = mmdata;
7443: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7444: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7446: /* memory type */
7447: mmdata->mtype = PETSC_MEMTYPE_HOST;
7448: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7449: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7450: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7451: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7452: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7453: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7455: /* prepare coo coordinates for values insertion */
7457: /* count total nonzeros of those intermediate seqaij Mats
7458: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7459: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7460: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7461: */
7462: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7463: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7464: if (mptmp[cp]) continue;
7465: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7466: const PetscInt *rmap = rmapa[cp];
7467: const PetscInt mr = mp[cp]->rmap->n;
7468: const PetscInt rs = C->rmap->rstart;
7469: const PetscInt re = C->rmap->rend;
7470: const PetscInt *ii = mm->i;
7471: for (i = 0; i < mr; i++) {
7472: const PetscInt gr = rmap[i];
7473: const PetscInt nz = ii[i + 1] - ii[i];
7474: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7475: else ncoo_oown += nz; /* this row is local */
7476: }
7477: } else ncoo_d += mm->nz;
7478: }
7480: /*
7481: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7483: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7485: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7487: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7488: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7489: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7491: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7492: 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.
7493: */
7494: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7495: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7497: /* gather (i,j) of nonzeros inserted by remote procs */
7498: if (hasoffproc) {
7499: PetscSF msf;
7500: PetscInt ncoo2, *coo_i2, *coo_j2;
7502: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7503: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7504: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7506: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7507: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7508: PetscInt *idxoff = mmdata->off[cp];
7509: PetscInt *idxown = mmdata->own[cp];
7510: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7511: const PetscInt *rmap = rmapa[cp];
7512: const PetscInt *cmap = cmapa[cp];
7513: const PetscInt *ii = mm->i;
7514: PetscInt *coi = coo_i + ncoo_o;
7515: PetscInt *coj = coo_j + ncoo_o;
7516: const PetscInt mr = mp[cp]->rmap->n;
7517: const PetscInt rs = C->rmap->rstart;
7518: const PetscInt re = C->rmap->rend;
7519: const PetscInt cs = C->cmap->rstart;
7520: for (i = 0; i < mr; i++) {
7521: const PetscInt *jj = mm->j + ii[i];
7522: const PetscInt gr = rmap[i];
7523: const PetscInt nz = ii[i + 1] - ii[i];
7524: if (gr < rs || gr >= re) { /* this is an offproc row */
7525: for (j = ii[i]; j < ii[i + 1]; j++) {
7526: *coi++ = gr;
7527: *idxoff++ = j;
7528: }
7529: if (!cmapt[cp]) { /* already global */
7530: for (j = 0; j < nz; j++) *coj++ = jj[j];
7531: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7532: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7533: } else { /* offdiag */
7534: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7535: }
7536: ncoo_o += nz;
7537: } else { /* this is a local row */
7538: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7539: }
7540: }
7541: }
7542: mmdata->off[cp + 1] = idxoff;
7543: mmdata->own[cp + 1] = idxown;
7544: }
7546: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7547: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7548: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7549: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7550: ncoo = ncoo_d + ncoo_oown + ncoo2;
7551: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7552: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7553: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7554: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7555: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7556: PetscCall(PetscFree2(coo_i, coo_j));
7557: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7558: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7559: coo_i = coo_i2;
7560: coo_j = coo_j2;
7561: } else { /* no offproc values insertion */
7562: ncoo = ncoo_d;
7563: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7565: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7566: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7567: PetscCall(PetscSFSetUp(mmdata->sf));
7568: }
7569: mmdata->hasoffproc = hasoffproc;
7571: /* gather (i,j) of nonzeros inserted locally */
7572: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7573: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7574: PetscInt *coi = coo_i + ncoo_d;
7575: PetscInt *coj = coo_j + ncoo_d;
7576: const PetscInt *jj = mm->j;
7577: const PetscInt *ii = mm->i;
7578: const PetscInt *cmap = cmapa[cp];
7579: const PetscInt *rmap = rmapa[cp];
7580: const PetscInt mr = mp[cp]->rmap->n;
7581: const PetscInt rs = C->rmap->rstart;
7582: const PetscInt re = C->rmap->rend;
7583: const PetscInt cs = C->cmap->rstart;
7585: if (mptmp[cp]) continue;
7586: if (rmapt[cp] == 1) { /* consecutive rows */
7587: /* fill coo_i */
7588: for (i = 0; i < mr; i++) {
7589: const PetscInt gr = i + rs;
7590: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7591: }
7592: /* fill coo_j */
7593: if (!cmapt[cp]) { /* type-0, already global */
7594: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7595: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7596: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7597: } else { /* type-2, local to global for sparse columns */
7598: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7599: }
7600: ncoo_d += mm->nz;
7601: } else if (rmapt[cp] == 2) { /* sparse rows */
7602: for (i = 0; i < mr; i++) {
7603: const PetscInt *jj = mm->j + ii[i];
7604: const PetscInt gr = rmap[i];
7605: const PetscInt nz = ii[i + 1] - ii[i];
7606: if (gr >= rs && gr < re) { /* local rows */
7607: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7608: if (!cmapt[cp]) { /* type-0, already global */
7609: for (j = 0; j < nz; j++) *coj++ = jj[j];
7610: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7611: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7612: } else { /* type-2, local to global for sparse columns */
7613: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7614: }
7615: ncoo_d += nz;
7616: }
7617: }
7618: }
7619: }
7620: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7621: PetscCall(ISDestroy(&glob));
7622: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7623: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7624: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7625: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7627: /* preallocate with COO data */
7628: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7629: PetscCall(PetscFree2(coo_i, coo_j));
7630: PetscFunctionReturn(PETSC_SUCCESS);
7631: }
7633: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7634: {
7635: Mat_Product *product = mat->product;
7636: #if defined(PETSC_HAVE_DEVICE)
7637: PetscBool match = PETSC_FALSE;
7638: PetscBool usecpu = PETSC_FALSE;
7639: #else
7640: PetscBool match = PETSC_TRUE;
7641: #endif
7643: PetscFunctionBegin;
7644: MatCheckProduct(mat, 1);
7645: #if defined(PETSC_HAVE_DEVICE)
7646: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7647: if (match) { /* we can always fallback to the CPU if requested */
7648: switch (product->type) {
7649: case MATPRODUCT_AB:
7650: if (product->api_user) {
7651: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7652: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7653: PetscOptionsEnd();
7654: } else {
7655: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7656: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7657: PetscOptionsEnd();
7658: }
7659: break;
7660: case MATPRODUCT_AtB:
7661: if (product->api_user) {
7662: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7663: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7664: PetscOptionsEnd();
7665: } else {
7666: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7667: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7668: PetscOptionsEnd();
7669: }
7670: break;
7671: case MATPRODUCT_PtAP:
7672: if (product->api_user) {
7673: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7674: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7675: PetscOptionsEnd();
7676: } else {
7677: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7678: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7679: PetscOptionsEnd();
7680: }
7681: break;
7682: default:
7683: break;
7684: }
7685: match = (PetscBool)!usecpu;
7686: }
7687: #endif
7688: if (match) {
7689: switch (product->type) {
7690: case MATPRODUCT_AB:
7691: case MATPRODUCT_AtB:
7692: case MATPRODUCT_PtAP:
7693: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7694: break;
7695: default:
7696: break;
7697: }
7698: }
7699: /* fallback to MPIAIJ ops */
7700: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7701: PetscFunctionReturn(PETSC_SUCCESS);
7702: }
7704: /*
7705: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7707: n - the number of block indices in cc[]
7708: cc - the block indices (must be large enough to contain the indices)
7709: */
7710: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7711: {
7712: PetscInt cnt = -1, nidx, j;
7713: const PetscInt *idx;
7715: PetscFunctionBegin;
7716: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7717: if (nidx) {
7718: cnt = 0;
7719: cc[cnt] = idx[0] / bs;
7720: for (j = 1; j < nidx; j++) {
7721: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7722: }
7723: }
7724: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7725: *n = cnt + 1;
7726: PetscFunctionReturn(PETSC_SUCCESS);
7727: }
7729: /*
7730: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7732: ncollapsed - the number of block indices
7733: collapsed - the block indices (must be large enough to contain the indices)
7734: */
7735: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7736: {
7737: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7739: PetscFunctionBegin;
7740: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7741: for (i = start + 1; i < start + bs; i++) {
7742: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7743: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7744: cprevtmp = cprev;
7745: cprev = merged;
7746: merged = cprevtmp;
7747: }
7748: *ncollapsed = nprev;
7749: if (collapsed) *collapsed = cprev;
7750: PetscFunctionReturn(PETSC_SUCCESS);
7751: }
7753: /*
7754: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7756: Input Parameter:
7757: . Amat - matrix
7758: - symmetrize - make the result symmetric
7759: + scale - scale with diagonal
7761: Output Parameter:
7762: . a_Gmat - output scalar graph >= 0
7764: */
7765: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7766: {
7767: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7768: MPI_Comm comm;
7769: Mat Gmat;
7770: PetscBool ismpiaij, isseqaij;
7771: Mat a, b, c;
7772: MatType jtype;
7774: PetscFunctionBegin;
7775: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7776: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7777: PetscCall(MatGetSize(Amat, &MM, &NN));
7778: PetscCall(MatGetBlockSize(Amat, &bs));
7779: nloc = (Iend - Istart) / bs;
7781: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7782: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7783: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7785: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7786: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7787: implementation */
7788: if (bs > 1) {
7789: PetscCall(MatGetType(Amat, &jtype));
7790: PetscCall(MatCreate(comm, &Gmat));
7791: PetscCall(MatSetType(Gmat, jtype));
7792: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7793: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7794: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7795: PetscInt *d_nnz, *o_nnz;
7796: MatScalar *aa, val, *AA;
7797: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7798: if (isseqaij) {
7799: a = Amat;
7800: b = NULL;
7801: } else {
7802: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7803: a = d->A;
7804: b = d->B;
7805: }
7806: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7807: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7808: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7809: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7810: const PetscInt *cols1, *cols2;
7811: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7812: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7813: nnz[brow / bs] = nc2 / bs;
7814: if (nc2 % bs) ok = 0;
7815: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7816: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7817: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7818: if (nc1 != nc2) ok = 0;
7819: else {
7820: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7821: if (cols1[jj] != cols2[jj]) ok = 0;
7822: if (cols1[jj] % bs != jj % bs) ok = 0;
7823: }
7824: }
7825: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7826: }
7827: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7828: if (!ok) {
7829: PetscCall(PetscFree2(d_nnz, o_nnz));
7830: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7831: goto old_bs;
7832: }
7833: }
7834: }
7835: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7836: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7837: PetscCall(PetscFree2(d_nnz, o_nnz));
7838: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7839: // diag
7840: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7841: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7842: ai = aseq->i;
7843: n = ai[brow + 1] - ai[brow];
7844: aj = aseq->j + ai[brow];
7845: for (int k = 0; k < n; k += bs) { // block columns
7846: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7847: val = 0;
7848: if (index_size == 0) {
7849: for (int ii = 0; ii < bs; ii++) { // rows in block
7850: aa = aseq->a + ai[brow + ii] + k;
7851: for (int jj = 0; jj < bs; jj++) { // columns in block
7852: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7853: }
7854: }
7855: } else { // use (index,index) value if provided
7856: for (int iii = 0; iii < index_size; iii++) { // rows in block
7857: int ii = index[iii];
7858: aa = aseq->a + ai[brow + ii] + k;
7859: for (int jjj = 0; jjj < index_size; jjj++) { // columns in block
7860: int jj = index[jjj];
7861: val += PetscAbs(PetscRealPart(aa[jj]));
7862: }
7863: }
7864: }
7865: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7866: AA[k / bs] = val;
7867: }
7868: grow = Istart / bs + brow / bs;
7869: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7870: }
7871: // off-diag
7872: if (ismpiaij) {
7873: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7874: const PetscScalar *vals;
7875: const PetscInt *cols, *garray = aij->garray;
7876: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7877: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7878: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7879: for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7880: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7881: AA[k / bs] = 0;
7882: AJ[cidx] = garray[cols[k]] / bs;
7883: }
7884: nc = ncols / bs;
7885: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7886: if (index_size == 0) {
7887: for (int ii = 0; ii < bs; ii++) { // rows in block
7888: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7889: for (int k = 0; k < ncols; k += bs) {
7890: for (int jj = 0; jj < bs; jj++) { // cols in block
7891: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7892: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7893: }
7894: }
7895: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7896: }
7897: } else { // use (index,index) value if provided
7898: for (int iii = 0; iii < index_size; iii++) { // rows in block
7899: int ii = index[iii];
7900: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7901: for (int k = 0; k < ncols; k += bs) {
7902: for (int jjj = 0; jjj < index_size; jjj++) { // cols in block
7903: int jj = index[jjj];
7904: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7905: }
7906: }
7907: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7908: }
7909: }
7910: grow = Istart / bs + brow / bs;
7911: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7912: }
7913: }
7914: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7915: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7916: PetscCall(PetscFree2(AA, AJ));
7917: } else {
7918: const PetscScalar *vals;
7919: const PetscInt *idx;
7920: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7921: old_bs:
7922: /*
7923: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7924: */
7925: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7926: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7927: if (isseqaij) {
7928: PetscInt max_d_nnz;
7929: /*
7930: Determine exact preallocation count for (sequential) scalar matrix
7931: */
7932: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7933: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7934: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7935: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7936: PetscCall(PetscFree3(w0, w1, w2));
7937: } else if (ismpiaij) {
7938: Mat Daij, Oaij;
7939: const PetscInt *garray;
7940: PetscInt max_d_nnz;
7941: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7942: /*
7943: Determine exact preallocation count for diagonal block portion of scalar matrix
7944: */
7945: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7946: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7947: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7948: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7949: PetscCall(PetscFree3(w0, w1, w2));
7950: /*
7951: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7952: */
7953: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7954: o_nnz[jj] = 0;
7955: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7956: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7957: o_nnz[jj] += ncols;
7958: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7959: }
7960: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7961: }
7962: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7963: /* get scalar copy (norms) of matrix */
7964: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7965: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7966: PetscCall(PetscFree2(d_nnz, o_nnz));
7967: for (Ii = Istart; Ii < Iend; Ii++) {
7968: PetscInt dest_row = Ii / bs;
7969: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7970: for (jj = 0; jj < ncols; jj++) {
7971: PetscInt dest_col = idx[jj] / bs;
7972: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7973: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7974: }
7975: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7976: }
7977: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7978: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7979: }
7980: } else {
7981: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7982: else {
7983: Gmat = Amat;
7984: PetscCall(PetscObjectReference((PetscObject)Gmat));
7985: }
7986: if (isseqaij) {
7987: a = Gmat;
7988: b = NULL;
7989: } else {
7990: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7991: a = d->A;
7992: b = d->B;
7993: }
7994: if (filter >= 0 || scale) {
7995: /* take absolute value of each entry */
7996: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7997: MatInfo info;
7998: PetscScalar *avals;
7999: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8000: PetscCall(MatSeqAIJGetArray(c, &avals));
8001: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8002: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8003: }
8004: }
8005: }
8006: if (symmetrize) {
8007: PetscBool isset, issym;
8008: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8009: if (!isset || !issym) {
8010: Mat matTrans;
8011: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8012: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8013: PetscCall(MatDestroy(&matTrans));
8014: }
8015: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8016: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8017: if (scale) {
8018: /* scale c for all diagonal values = 1 or -1 */
8019: Vec diag;
8020: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8021: PetscCall(MatGetDiagonal(Gmat, diag));
8022: PetscCall(VecReciprocal(diag));
8023: PetscCall(VecSqrtAbs(diag));
8024: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8025: PetscCall(VecDestroy(&diag));
8026: }
8027: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8029: if (filter >= 0) {
8030: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8031: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8032: }
8033: *a_Gmat = Gmat;
8034: PetscFunctionReturn(PETSC_SUCCESS);
8035: }
8037: /*
8038: Special version for direct calls from Fortran
8039: */
8041: /* Change these macros so can be used in void function */
8042: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8043: #undef PetscCall
8044: #define PetscCall(...) \
8045: do { \
8046: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8047: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8048: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8049: return; \
8050: } \
8051: } while (0)
8053: #undef SETERRQ
8054: #define SETERRQ(comm, ierr, ...) \
8055: do { \
8056: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8057: return; \
8058: } while (0)
8060: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8061: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8062: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8063: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8064: #else
8065: #endif
8066: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8067: {
8068: Mat mat = *mmat;
8069: PetscInt m = *mm, n = *mn;
8070: InsertMode addv = *maddv;
8071: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8072: PetscScalar value;
8074: MatCheckPreallocated(mat, 1);
8075: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8076: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8077: {
8078: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8079: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8080: PetscBool roworiented = aij->roworiented;
8082: /* Some Variables required in the macro */
8083: Mat A = aij->A;
8084: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8085: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8086: MatScalar *aa;
8087: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8088: Mat B = aij->B;
8089: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8090: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8091: MatScalar *ba;
8092: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8093: * cannot use "#if defined" inside a macro. */
8094: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8096: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8097: PetscInt nonew = a->nonew;
8098: MatScalar *ap1, *ap2;
8100: PetscFunctionBegin;
8101: PetscCall(MatSeqAIJGetArray(A, &aa));
8102: PetscCall(MatSeqAIJGetArray(B, &ba));
8103: for (i = 0; i < m; i++) {
8104: if (im[i] < 0) continue;
8105: 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);
8106: if (im[i] >= rstart && im[i] < rend) {
8107: row = im[i] - rstart;
8108: lastcol1 = -1;
8109: rp1 = aj + ai[row];
8110: ap1 = aa + ai[row];
8111: rmax1 = aimax[row];
8112: nrow1 = ailen[row];
8113: low1 = 0;
8114: high1 = nrow1;
8115: lastcol2 = -1;
8116: rp2 = bj + bi[row];
8117: ap2 = ba + bi[row];
8118: rmax2 = bimax[row];
8119: nrow2 = bilen[row];
8120: low2 = 0;
8121: high2 = nrow2;
8123: for (j = 0; j < n; j++) {
8124: if (roworiented) value = v[i * n + j];
8125: else value = v[i + j * m];
8126: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8127: if (in[j] >= cstart && in[j] < cend) {
8128: col = in[j] - cstart;
8129: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8130: } else if (in[j] < 0) continue;
8131: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8132: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8133: } else {
8134: if (mat->was_assembled) {
8135: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8136: #if defined(PETSC_USE_CTABLE)
8137: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8138: col--;
8139: #else
8140: col = aij->colmap[in[j]] - 1;
8141: #endif
8142: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8143: PetscCall(MatDisAssemble_MPIAIJ(mat));
8144: col = in[j];
8145: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8146: B = aij->B;
8147: b = (Mat_SeqAIJ *)B->data;
8148: bimax = b->imax;
8149: bi = b->i;
8150: bilen = b->ilen;
8151: bj = b->j;
8152: rp2 = bj + bi[row];
8153: ap2 = ba + bi[row];
8154: rmax2 = bimax[row];
8155: nrow2 = bilen[row];
8156: low2 = 0;
8157: high2 = nrow2;
8158: bm = aij->B->rmap->n;
8159: ba = b->a;
8160: inserted = PETSC_FALSE;
8161: }
8162: } else col = in[j];
8163: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8164: }
8165: }
8166: } else if (!aij->donotstash) {
8167: if (roworiented) {
8168: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8169: } else {
8170: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8171: }
8172: }
8173: }
8174: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8175: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8176: }
8177: PetscFunctionReturnVoid();
8178: }
8180: /* Undefining these here since they were redefined from their original definition above! No
8181: * other PETSc functions should be defined past this point, as it is impossible to recover the
8182: * original definitions */
8183: #undef PetscCall
8184: #undef SETERRQ