Actual source code: mpiaij.c
1: #include <../src/mat/impls/aij/mpi/mpiaij.h>
2: #include <petsc/private/vecimpl.h>
3: #include <petsc/private/sfimpl.h>
4: #include <petsc/private/isimpl.h>
5: #include <petscblaslapack.h>
6: #include <petscsf.h>
7: #include <petsc/private/hashmapi.h>
9: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
10: {
11: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
13: PetscFunctionBegin;
14: PetscCall(PetscLogObjectState((PetscObject)mat, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT, mat->rmap->N, mat->cmap->N));
15: PetscCall(MatStashDestroy_Private(&mat->stash));
16: PetscCall(VecDestroy(&aij->diag));
17: PetscCall(MatDestroy(&aij->A));
18: PetscCall(MatDestroy(&aij->B));
19: #if defined(PETSC_USE_CTABLE)
20: PetscCall(PetscHMapIDestroy(&aij->colmap));
21: #else
22: PetscCall(PetscFree(aij->colmap));
23: #endif
24: PetscCall(PetscFree(aij->garray));
25: PetscCall(VecDestroy(&aij->lvec));
26: PetscCall(VecScatterDestroy(&aij->Mvctx));
27: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
28: PetscCall(PetscFree(aij->ld));
30: PetscCall(PetscFree(mat->data));
32: /* may be created by MatCreateMPIAIJSumSeqAIJSymbolic */
33: PetscCall(PetscObjectCompose((PetscObject)mat, "MatMergeSeqsToMPI", NULL));
35: PetscCall(PetscObjectChangeTypeName((PetscObject)mat, NULL));
36: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatStoreValues_C", NULL));
37: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatRetrieveValues_C", NULL));
38: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatIsTranspose_C", NULL));
39: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocation_C", NULL));
40: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatResetPreallocation_C", NULL));
41: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetPreallocationCSR_C", NULL));
42: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatDiagonalScaleLocal_C", NULL));
43: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpibaij_C", NULL));
44: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisbaij_C", NULL));
45: #if defined(PETSC_HAVE_CUDA)
46: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcusparse_C", NULL));
47: #endif
48: #if defined(PETSC_HAVE_HIP)
49: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijhipsparse_C", NULL));
50: #endif
51: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
52: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijkokkos_C", NULL));
53: #endif
54: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpidense_C", NULL));
55: #if defined(PETSC_HAVE_ELEMENTAL)
56: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_elemental_C", NULL));
57: #endif
58: #if defined(PETSC_HAVE_SCALAPACK)
59: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_scalapack_C", NULL));
60: #endif
61: #if defined(PETSC_HAVE_HYPRE)
62: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_hypre_C", NULL));
63: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", NULL));
64: #endif
65: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
66: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_is_mpiaij_C", NULL));
67: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatProductSetFromOptions_mpiaij_mpiaij_C", NULL));
68: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatMPIAIJSetUseScalableIncreaseOverlap_C", NULL));
69: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijperm_C", NULL));
70: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijsell_C", NULL));
71: #if defined(PETSC_HAVE_MKL_SPARSE)
72: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijmkl_C", NULL));
73: #endif
74: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpiaijcrl_C", NULL));
75: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_is_C", NULL));
76: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatConvert_mpiaij_mpisell_C", NULL));
77: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetPreallocationCOO_C", NULL));
78: PetscCall(PetscObjectComposeFunction((PetscObject)mat, "MatSetValuesCOO_C", NULL));
79: PetscFunctionReturn(PETSC_SUCCESS);
80: }
82: /* defines MatSetValues_MPI_Hash(), MatAssemblyBegin_MPI_Hash(), and MatAssemblyEnd_MPI_Hash() */
83: #define TYPE AIJ
84: #define TYPE_AIJ
85: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
86: #undef TYPE
87: #undef TYPE_AIJ
89: static PetscErrorCode MatGetRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
90: {
91: Mat B;
93: PetscFunctionBegin;
94: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, &B));
95: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject)B));
96: PetscCall(MatGetRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
97: PetscCall(MatDestroy(&B));
98: PetscFunctionReturn(PETSC_SUCCESS);
99: }
101: static PetscErrorCode MatRestoreRowIJ_MPIAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
102: {
103: Mat B;
105: PetscFunctionBegin;
106: PetscCall(PetscObjectQuery((PetscObject)A, "MatGetRowIJ_MPIAIJ", (PetscObject *)&B));
107: PetscCall(MatRestoreRowIJ(B, oshift, symmetric, inodecompressed, m, ia, ja, done));
108: PetscCall(PetscObjectCompose((PetscObject)A, "MatGetRowIJ_MPIAIJ", NULL));
109: PetscFunctionReturn(PETSC_SUCCESS);
110: }
112: /*MC
113: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
115: This matrix type is identical to` MATSEQAIJ` when constructed with a single process communicator,
116: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
117: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
118: for communicators controlling multiple processes. It is recommended that you call both of
119: the above preallocation routines for simplicity.
121: Options Database Key:
122: . -mat_type aij - sets the matrix type to `MATAIJ` during a call to `MatSetFromOptions()`
124: Developer Note:
125: Level: beginner
127: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, `MATAIJKOKKOS`,and also automatically switches over to use inodes when
128: enough exist.
130: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`
131: M*/
133: /*MC
134: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
136: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
137: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
138: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
139: for communicators controlling multiple processes. It is recommended that you call both of
140: the above preallocation routines for simplicity.
142: Options Database Key:
143: . -mat_type aijcrl - sets the matrix type to `MATMPIAIJCRL` during a call to `MatSetFromOptions()`
145: Level: beginner
147: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
148: M*/
150: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A, PetscBool flg)
151: {
152: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
154: PetscFunctionBegin;
155: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_VIENNACL)
156: A->boundtocpu = flg;
157: #endif
158: if (a->A) PetscCall(MatBindToCPU(a->A, flg));
159: if (a->B) PetscCall(MatBindToCPU(a->B, flg));
161: /* In addition to binding the diagonal and off-diagonal matrices, bind the local vectors used for matrix-vector products.
162: * This maybe seems a little odd for a MatBindToCPU() call to do, but it makes no sense for the binding of these vectors
163: * to differ from the parent matrix. */
164: if (a->lvec) PetscCall(VecBindToCPU(a->lvec, flg));
165: if (a->diag) PetscCall(VecBindToCPU(a->diag, flg));
166: PetscFunctionReturn(PETSC_SUCCESS);
167: }
169: static PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
170: {
171: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
173: PetscFunctionBegin;
174: if (mat->A) {
175: PetscCall(MatSetBlockSizes(mat->A, rbs, cbs));
176: PetscCall(MatSetBlockSizes(mat->B, rbs, 1));
177: }
178: PetscFunctionReturn(PETSC_SUCCESS);
179: }
181: static PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M, IS *keptrows)
182: {
183: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)M->data;
184: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data;
185: Mat_SeqAIJ *b = (Mat_SeqAIJ *)mat->B->data;
186: const PetscInt *ia, *ib;
187: const MatScalar *aa, *bb, *aav, *bav;
188: PetscInt na, nb, i, j, *rows, cnt = 0, n0rows;
189: PetscInt m = M->rmap->n, rstart = M->rmap->rstart;
191: PetscFunctionBegin;
192: *keptrows = NULL;
194: ia = a->i;
195: ib = b->i;
196: PetscCall(MatSeqAIJGetArrayRead(mat->A, &aav));
197: PetscCall(MatSeqAIJGetArrayRead(mat->B, &bav));
198: for (i = 0; i < m; i++) {
199: na = ia[i + 1] - ia[i];
200: nb = ib[i + 1] - ib[i];
201: if (!na && !nb) {
202: cnt++;
203: goto ok1;
204: }
205: aa = aav + ia[i];
206: for (j = 0; j < na; j++) {
207: if (aa[j] != 0.0) goto ok1;
208: }
209: bb = PetscSafePointerPlusOffset(bav, ib[i]);
210: for (j = 0; j < nb; j++) {
211: if (bb[j] != 0.0) goto ok1;
212: }
213: cnt++;
214: ok1:;
215: }
216: PetscCall(MPIU_Allreduce(&cnt, &n0rows, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)M)));
217: if (!n0rows) {
218: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
219: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
220: PetscFunctionReturn(PETSC_SUCCESS);
221: }
222: PetscCall(PetscMalloc1(M->rmap->n - cnt, &rows));
223: cnt = 0;
224: for (i = 0; i < m; i++) {
225: na = ia[i + 1] - ia[i];
226: nb = ib[i + 1] - ib[i];
227: if (!na && !nb) continue;
228: aa = aav + ia[i];
229: for (j = 0; j < na; j++) {
230: if (aa[j] != 0.0) {
231: rows[cnt++] = rstart + i;
232: goto ok2;
233: }
234: }
235: bb = PetscSafePointerPlusOffset(bav, ib[i]);
236: for (j = 0; j < nb; j++) {
237: if (bb[j] != 0.0) {
238: rows[cnt++] = rstart + i;
239: goto ok2;
240: }
241: }
242: ok2:;
243: }
244: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), cnt, rows, PETSC_OWN_POINTER, keptrows));
245: PetscCall(MatSeqAIJRestoreArrayRead(mat->A, &aav));
246: PetscCall(MatSeqAIJRestoreArrayRead(mat->B, &bav));
247: PetscFunctionReturn(PETSC_SUCCESS);
248: }
250: static PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y, Vec D, InsertMode is)
251: {
252: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Y->data;
253: PetscBool cong;
255: PetscFunctionBegin;
256: PetscCall(MatHasCongruentLayouts(Y, &cong));
257: if (Y->assembled && cong) {
258: PetscCall(MatDiagonalSet(aij->A, D, is));
259: } else {
260: PetscCall(MatDiagonalSet_Default(Y, D, is));
261: }
262: PetscFunctionReturn(PETSC_SUCCESS);
263: }
265: static PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M, IS *zrows)
266: {
267: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)M->data;
268: PetscInt i, rstart, nrows, *rows;
270: PetscFunctionBegin;
271: *zrows = NULL;
272: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(aij->A, &nrows, &rows));
273: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
274: for (i = 0; i < nrows; i++) rows[i] += rstart;
275: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)M), nrows, rows, PETSC_OWN_POINTER, zrows));
276: PetscFunctionReturn(PETSC_SUCCESS);
277: }
279: static PetscErrorCode MatGetColumnReductions_MPIAIJ(Mat A, PetscInt type, PetscReal *reductions)
280: {
281: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
282: PetscInt i, m, n, *garray = aij->garray;
283: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ *)aij->A->data;
284: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ *)aij->B->data;
285: PetscReal *work;
286: const PetscScalar *dummy;
288: PetscFunctionBegin;
289: PetscCall(MatGetSize(A, &m, &n));
290: PetscCall(PetscCalloc1(n, &work));
291: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
292: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
293: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
294: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
295: if (type == NORM_2) {
296: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
297: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
298: } else if (type == NORM_1) {
299: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
300: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
301: } else if (type == NORM_INFINITY) {
302: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
303: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), work[garray[b_aij->j[i]]]);
304: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
305: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
306: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
307: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
308: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
309: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
310: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
311: if (type == NORM_INFINITY) {
312: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
313: } else {
314: PetscCall(MPIU_Allreduce(work, reductions, n, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
315: }
316: PetscCall(PetscFree(work));
317: if (type == NORM_2) {
318: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
319: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
320: for (i = 0; i < n; i++) reductions[i] /= m;
321: }
322: PetscFunctionReturn(PETSC_SUCCESS);
323: }
325: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
326: {
327: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
328: IS sis, gis;
329: const PetscInt *isis, *igis;
330: PetscInt n, *iis, nsis, ngis, rstart, i;
332: PetscFunctionBegin;
333: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
334: PetscCall(MatFindNonzeroRows(a->B, &gis));
335: PetscCall(ISGetSize(gis, &ngis));
336: PetscCall(ISGetSize(sis, &nsis));
337: PetscCall(ISGetIndices(sis, &isis));
338: PetscCall(ISGetIndices(gis, &igis));
340: PetscCall(PetscMalloc1(ngis + nsis, &iis));
341: PetscCall(PetscArraycpy(iis, igis, ngis));
342: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
343: n = ngis + nsis;
344: PetscCall(PetscSortRemoveDupsInt(&n, iis));
345: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
346: for (i = 0; i < n; i++) iis[i] += rstart;
347: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
349: PetscCall(ISRestoreIndices(sis, &isis));
350: PetscCall(ISRestoreIndices(gis, &igis));
351: PetscCall(ISDestroy(&sis));
352: PetscCall(ISDestroy(&gis));
353: PetscFunctionReturn(PETSC_SUCCESS);
354: }
356: /*
357: Local utility routine that creates a mapping from the global column
358: number to the local number in the off-diagonal part of the local
359: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
360: a slightly higher hash table cost; without it it is not scalable (each processor
361: has an order N integer array but is fast to access.
362: */
363: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
364: {
365: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
366: PetscInt n = aij->B->cmap->n, i;
368: PetscFunctionBegin;
369: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
370: #if defined(PETSC_USE_CTABLE)
371: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
372: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
373: #else
374: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
375: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
376: #endif
377: PetscFunctionReturn(PETSC_SUCCESS);
378: }
380: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
381: do { \
382: if (col <= lastcol1) low1 = 0; \
383: else high1 = nrow1; \
384: lastcol1 = col; \
385: while (high1 - low1 > 5) { \
386: t = (low1 + high1) / 2; \
387: if (rp1[t] > col) high1 = t; \
388: else low1 = t; \
389: } \
390: for (_i = low1; _i < high1; _i++) { \
391: if (rp1[_i] > col) break; \
392: if (rp1[_i] == col) { \
393: if (addv == ADD_VALUES) { \
394: ap1[_i] += value; \
395: /* Not sure LogFlops will slow dow the code or not */ \
396: (void)PetscLogFlops(1.0); \
397: } else ap1[_i] = value; \
398: goto a_noinsert; \
399: } \
400: } \
401: if (value == 0.0 && ignorezeroentries && row != col) { \
402: low1 = 0; \
403: high1 = nrow1; \
404: goto a_noinsert; \
405: } \
406: if (nonew == 1) { \
407: low1 = 0; \
408: high1 = nrow1; \
409: goto a_noinsert; \
410: } \
411: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
412: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
413: N = nrow1++ - 1; \
414: a->nz++; \
415: high1++; \
416: /* shift up all the later entries in this row */ \
417: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
418: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
419: rp1[_i] = col; \
420: ap1[_i] = value; \
421: A->nonzerostate++; \
422: a_noinsert:; \
423: ailen[row] = nrow1; \
424: } while (0)
426: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
427: do { \
428: if (col <= lastcol2) low2 = 0; \
429: else high2 = nrow2; \
430: lastcol2 = col; \
431: while (high2 - low2 > 5) { \
432: t = (low2 + high2) / 2; \
433: if (rp2[t] > col) high2 = t; \
434: else low2 = t; \
435: } \
436: for (_i = low2; _i < high2; _i++) { \
437: if (rp2[_i] > col) break; \
438: if (rp2[_i] == col) { \
439: if (addv == ADD_VALUES) { \
440: ap2[_i] += value; \
441: (void)PetscLogFlops(1.0); \
442: } else ap2[_i] = value; \
443: goto b_noinsert; \
444: } \
445: } \
446: if (value == 0.0 && ignorezeroentries) { \
447: low2 = 0; \
448: high2 = nrow2; \
449: goto b_noinsert; \
450: } \
451: if (nonew == 1) { \
452: low2 = 0; \
453: high2 = nrow2; \
454: goto b_noinsert; \
455: } \
456: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
457: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
458: N = nrow2++ - 1; \
459: b->nz++; \
460: high2++; \
461: /* shift up all the later entries in this row */ \
462: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
463: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
464: rp2[_i] = col; \
465: ap2[_i] = value; \
466: B->nonzerostate++; \
467: b_noinsert:; \
468: bilen[row] = nrow2; \
469: } while (0)
471: static PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A, PetscInt row, const PetscScalar v[])
472: {
473: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
474: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->A->data, *b = (Mat_SeqAIJ *)mat->B->data;
475: PetscInt l, *garray = mat->garray, diag;
476: PetscScalar *aa, *ba;
478: PetscFunctionBegin;
479: /* code only works for square matrices A */
481: /* find size of row to the left of the diagonal part */
482: PetscCall(MatGetOwnershipRange(A, &diag, NULL));
483: row = row - diag;
484: for (l = 0; l < b->i[row + 1] - b->i[row]; l++) {
485: if (garray[b->j[b->i[row] + l]] > diag) break;
486: }
487: if (l) {
488: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
489: PetscCall(PetscArraycpy(ba + b->i[row], v, l));
490: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
491: }
493: /* diagonal part */
494: if (a->i[row + 1] - a->i[row]) {
495: PetscCall(MatSeqAIJGetArray(mat->A, &aa));
496: PetscCall(PetscArraycpy(aa + a->i[row], v + l, (a->i[row + 1] - a->i[row])));
497: PetscCall(MatSeqAIJRestoreArray(mat->A, &aa));
498: }
500: /* right of diagonal part */
501: if (b->i[row + 1] - b->i[row] - l) {
502: PetscCall(MatSeqAIJGetArray(mat->B, &ba));
503: PetscCall(PetscArraycpy(ba + b->i[row] + l, v + l + a->i[row + 1] - a->i[row], b->i[row + 1] - b->i[row] - l));
504: PetscCall(MatSeqAIJRestoreArray(mat->B, &ba));
505: }
506: PetscFunctionReturn(PETSC_SUCCESS);
507: }
509: PetscErrorCode MatSetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode addv)
510: {
511: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
512: PetscScalar value = 0.0;
513: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
514: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
515: PetscBool roworiented = aij->roworiented;
517: /* Some Variables required in the macro */
518: Mat A = aij->A;
519: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
520: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
521: PetscBool ignorezeroentries = a->ignorezeroentries;
522: Mat B = aij->B;
523: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
524: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
525: MatScalar *aa, *ba;
526: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
527: PetscInt nonew;
528: MatScalar *ap1, *ap2;
530: PetscFunctionBegin;
531: PetscCall(MatSeqAIJGetArray(A, &aa));
532: PetscCall(MatSeqAIJGetArray(B, &ba));
533: for (i = 0; i < m; i++) {
534: if (im[i] < 0) continue;
535: PetscCheck(im[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, im[i], mat->rmap->N - 1);
536: if (im[i] >= rstart && im[i] < rend) {
537: row = im[i] - rstart;
538: lastcol1 = -1;
539: rp1 = PetscSafePointerPlusOffset(aj, ai[row]);
540: ap1 = PetscSafePointerPlusOffset(aa, ai[row]);
541: rmax1 = aimax[row];
542: nrow1 = ailen[row];
543: low1 = 0;
544: high1 = nrow1;
545: lastcol2 = -1;
546: rp2 = PetscSafePointerPlusOffset(bj, bi[row]);
547: ap2 = PetscSafePointerPlusOffset(ba, bi[row]);
548: rmax2 = bimax[row];
549: nrow2 = bilen[row];
550: low2 = 0;
551: high2 = nrow2;
553: for (j = 0; j < n; j++) {
554: if (v) value = roworiented ? v[i * n + j] : v[i + j * m];
555: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
556: if (in[j] >= cstart && in[j] < cend) {
557: col = in[j] - cstart;
558: nonew = a->nonew;
559: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
560: } else if (in[j] < 0) {
561: continue;
562: } else {
563: PetscCheck(in[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
564: if (mat->was_assembled) {
565: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
566: #if defined(PETSC_USE_CTABLE)
567: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col)); /* map global col ids to local ones */
568: col--;
569: #else
570: col = aij->colmap[in[j]] - 1;
571: #endif
572: if (col < 0 && !((Mat_SeqAIJ *)aij->B->data)->nonew) { /* col < 0 means in[j] is a new col for B */
573: PetscCall(MatDisAssemble_MPIAIJ(mat)); /* Change aij->B from reduced/local format to expanded/global format */
574: col = in[j];
575: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
576: B = aij->B;
577: b = (Mat_SeqAIJ *)B->data;
578: bimax = b->imax;
579: bi = b->i;
580: bilen = b->ilen;
581: bj = b->j;
582: ba = b->a;
583: rp2 = bj + bi[row];
584: ap2 = ba + bi[row];
585: rmax2 = bimax[row];
586: nrow2 = bilen[row];
587: low2 = 0;
588: high2 = nrow2;
589: bm = aij->B->rmap->n;
590: ba = b->a;
591: } else if (col < 0 && !(ignorezeroentries && value == 0.0)) {
592: if (1 == ((Mat_SeqAIJ *)aij->B->data)->nonew) {
593: PetscCall(PetscInfo(mat, "Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%" PetscInt_FMT ",%" PetscInt_FMT ")\n", (double)PetscRealPart(value), im[i], in[j]));
594: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", im[i], in[j]);
595: }
596: } else col = in[j];
597: nonew = b->nonew;
598: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
599: }
600: }
601: } else {
602: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Setting off process row %" PetscInt_FMT " even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set", im[i]);
603: if (!aij->donotstash) {
604: mat->assembled = PETSC_FALSE;
605: if (roworiented) {
606: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i * n), (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
607: } else {
608: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, PetscSafePointerPlusOffset(v, i), m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
609: }
610: }
611: }
612: }
613: PetscCall(MatSeqAIJRestoreArray(A, &aa)); /* aa, bb might have been free'd due to reallocation above. But we don't access them here */
614: PetscCall(MatSeqAIJRestoreArray(B, &ba));
615: PetscFunctionReturn(PETSC_SUCCESS);
616: }
618: /*
619: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
620: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
621: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
622: */
623: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[])
624: {
625: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
626: Mat A = aij->A; /* diagonal part of the matrix */
627: Mat B = aij->B; /* off-diagonal part of the matrix */
628: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
629: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
630: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, col;
631: PetscInt *ailen = a->ilen, *aj = a->j;
632: PetscInt *bilen = b->ilen, *bj = b->j;
633: PetscInt am = aij->A->rmap->n, j;
634: PetscInt diag_so_far = 0, dnz;
635: PetscInt offd_so_far = 0, onz;
637: PetscFunctionBegin;
638: /* Iterate over all rows of the matrix */
639: for (j = 0; j < am; j++) {
640: dnz = onz = 0;
641: /* Iterate over all non-zero columns of the current row */
642: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
643: /* If column is in the diagonal */
644: if (mat_j[col] >= cstart && mat_j[col] < cend) {
645: aj[diag_so_far++] = mat_j[col] - cstart;
646: dnz++;
647: } else { /* off-diagonal entries */
648: bj[offd_so_far++] = mat_j[col];
649: onz++;
650: }
651: }
652: ailen[j] = dnz;
653: bilen[j] = onz;
654: }
655: PetscFunctionReturn(PETSC_SUCCESS);
656: }
658: /*
659: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
660: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
661: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
662: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
663: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
664: */
665: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat, const PetscInt mat_j[], const PetscInt mat_i[], const PetscScalar mat_a[])
666: {
667: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
668: Mat A = aij->A; /* diagonal part of the matrix */
669: Mat B = aij->B; /* off-diagonal part of the matrix */
670: Mat_SeqAIJ *aijd = (Mat_SeqAIJ *)aij->A->data, *aijo = (Mat_SeqAIJ *)aij->B->data;
671: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
672: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
673: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend;
674: PetscInt *ailen = a->ilen, *aj = a->j;
675: PetscInt *bilen = b->ilen, *bj = b->j;
676: PetscInt am = aij->A->rmap->n, j;
677: PetscInt *full_diag_i = aijd->i, *full_offd_i = aijo->i; /* These variables can also include non-local elements, which are set at a later point. */
678: PetscInt col, dnz_row, onz_row, rowstart_diag, rowstart_offd;
679: PetscScalar *aa = a->a, *ba = b->a;
681: PetscFunctionBegin;
682: /* Iterate over all rows of the matrix */
683: for (j = 0; j < am; j++) {
684: dnz_row = onz_row = 0;
685: rowstart_offd = full_offd_i[j];
686: rowstart_diag = full_diag_i[j];
687: /* Iterate over all non-zero columns of the current row */
688: for (col = mat_i[j]; col < mat_i[j + 1]; col++) {
689: /* If column is in the diagonal */
690: if (mat_j[col] >= cstart && mat_j[col] < cend) {
691: aj[rowstart_diag + dnz_row] = mat_j[col] - cstart;
692: aa[rowstart_diag + dnz_row] = mat_a[col];
693: dnz_row++;
694: } else { /* off-diagonal entries */
695: bj[rowstart_offd + onz_row] = mat_j[col];
696: ba[rowstart_offd + onz_row] = mat_a[col];
697: onz_row++;
698: }
699: }
700: ailen[j] = dnz_row;
701: bilen[j] = onz_row;
702: }
703: PetscFunctionReturn(PETSC_SUCCESS);
704: }
706: static PetscErrorCode MatGetValues_MPIAIJ(Mat mat, PetscInt m, const PetscInt idxm[], PetscInt n, const PetscInt idxn[], PetscScalar v[])
707: {
708: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
709: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
710: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
712: PetscFunctionBegin;
713: for (i = 0; i < m; i++) {
714: if (idxm[i] < 0) continue; /* negative row */
715: PetscCheck(idxm[i] < mat->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, idxm[i], mat->rmap->N - 1);
716: PetscCheck(idxm[i] >= rstart && idxm[i] < rend, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only local values currently supported, row requested %" PetscInt_FMT " range [%" PetscInt_FMT " %" PetscInt_FMT ")", idxm[i], rstart, rend);
717: row = idxm[i] - rstart;
718: for (j = 0; j < n; j++) {
719: if (idxn[j] < 0) continue; /* negative column */
720: PetscCheck(idxn[j] < mat->cmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, idxn[j], mat->cmap->N - 1);
721: if (idxn[j] >= cstart && idxn[j] < cend) {
722: col = idxn[j] - cstart;
723: PetscCall(MatGetValues(aij->A, 1, &row, 1, &col, v + i * n + j));
724: } else {
725: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
726: #if defined(PETSC_USE_CTABLE)
727: PetscCall(PetscHMapIGetWithDefault(aij->colmap, idxn[j] + 1, 0, &col));
728: col--;
729: #else
730: col = aij->colmap[idxn[j]] - 1;
731: #endif
732: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v + i * n + j) = 0.0;
733: else PetscCall(MatGetValues(aij->B, 1, &row, 1, &col, v + i * n + j));
734: }
735: }
736: }
737: PetscFunctionReturn(PETSC_SUCCESS);
738: }
740: static PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat, MatAssemblyType mode)
741: {
742: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
743: PetscInt nstash, reallocs;
745: PetscFunctionBegin;
746: if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(PETSC_SUCCESS);
748: PetscCall(MatStashScatterBegin_Private(mat, &mat->stash, mat->rmap->range));
749: PetscCall(MatStashGetInfo_Private(&mat->stash, &nstash, &reallocs));
750: PetscCall(PetscInfo(aij->A, "Stash has %" PetscInt_FMT " entries, uses %" PetscInt_FMT " mallocs.\n", nstash, reallocs));
751: PetscFunctionReturn(PETSC_SUCCESS);
752: }
754: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat, MatAssemblyType mode)
755: {
756: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
757: PetscMPIInt n;
758: PetscInt i, j, rstart, ncols, flg;
759: PetscInt *row, *col;
760: PetscBool other_disassembled;
761: PetscScalar *val;
763: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
765: PetscFunctionBegin;
766: if (!aij->donotstash && !mat->nooffprocentries) {
767: while (1) {
768: PetscCall(MatStashScatterGetMesg_Private(&mat->stash, &n, &row, &col, &val, &flg));
769: if (!flg) break;
771: for (i = 0; i < n;) {
772: /* Now identify the consecutive vals belonging to the same row */
773: for (j = i, rstart = row[j]; j < n; j++) {
774: if (row[j] != rstart) break;
775: }
776: if (j < n) ncols = j - i;
777: else ncols = n - i;
778: /* Now assemble all these values with a single function call */
779: PetscCall(MatSetValues_MPIAIJ(mat, 1, row + i, ncols, col + i, val + i, mat->insertmode));
780: i = j;
781: }
782: }
783: PetscCall(MatStashScatterEnd_Private(&mat->stash));
784: }
785: #if defined(PETSC_HAVE_DEVICE)
786: if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
787: /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
788: if (mat->boundtocpu) {
789: PetscCall(MatBindToCPU(aij->A, PETSC_TRUE));
790: PetscCall(MatBindToCPU(aij->B, PETSC_TRUE));
791: }
792: #endif
793: PetscCall(MatAssemblyBegin(aij->A, mode));
794: PetscCall(MatAssemblyEnd(aij->A, mode));
796: /* determine if any processor has disassembled, if so we must
797: also disassemble ourself, in order that we may reassemble. */
798: /*
799: if nonzero structure of submatrix B cannot change then we know that
800: no processor disassembled thus we can skip this stuff
801: */
802: if (!((Mat_SeqAIJ *)aij->B->data)->nonew) {
803: PetscCall(MPIU_Allreduce(&mat->was_assembled, &other_disassembled, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
804: if (mat->was_assembled && !other_disassembled) { /* mat on this rank has reduced off-diag B with local col ids, but globally it does not */
805: PetscCall(MatDisAssemble_MPIAIJ(mat));
806: }
807: }
808: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) PetscCall(MatSetUpMultiply_MPIAIJ(mat));
809: PetscCall(MatSetOption(aij->B, MAT_USE_INODES, PETSC_FALSE));
810: #if defined(PETSC_HAVE_DEVICE)
811: if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
812: #endif
813: PetscCall(MatAssemblyBegin(aij->B, mode));
814: PetscCall(MatAssemblyEnd(aij->B, mode));
816: PetscCall(PetscFree2(aij->rowvalues, aij->rowindices));
818: aij->rowvalues = NULL;
820: PetscCall(VecDestroy(&aij->diag));
822: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
823: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ *)aij->A->data)->nonew) {
824: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
825: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
826: }
827: #if defined(PETSC_HAVE_DEVICE)
828: mat->offloadmask = PETSC_OFFLOAD_BOTH;
829: #endif
830: PetscFunctionReturn(PETSC_SUCCESS);
831: }
833: static PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
834: {
835: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
837: PetscFunctionBegin;
838: PetscCall(MatZeroEntries(l->A));
839: PetscCall(MatZeroEntries(l->B));
840: PetscFunctionReturn(PETSC_SUCCESS);
841: }
843: static PetscErrorCode MatZeroRows_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
844: {
845: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
846: PetscInt *lrows;
847: PetscInt r, len;
848: PetscBool cong;
850: PetscFunctionBegin;
851: /* get locally owned rows */
852: PetscCall(MatZeroRowsMapLocal_Private(A, N, rows, &len, &lrows));
853: PetscCall(MatHasCongruentLayouts(A, &cong));
854: /* fix right hand side if needed */
855: if (x && b) {
856: const PetscScalar *xx;
857: PetscScalar *bb;
859: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
860: PetscCall(VecGetArrayRead(x, &xx));
861: PetscCall(VecGetArray(b, &bb));
862: for (r = 0; r < len; ++r) bb[lrows[r]] = diag * xx[lrows[r]];
863: PetscCall(VecRestoreArrayRead(x, &xx));
864: PetscCall(VecRestoreArray(b, &bb));
865: }
867: if (diag != 0.0 && cong) {
868: PetscCall(MatZeroRows(mat->A, len, lrows, diag, NULL, NULL));
869: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
870: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
871: Mat_SeqAIJ *aijA = (Mat_SeqAIJ *)mat->A->data;
872: Mat_SeqAIJ *aijB = (Mat_SeqAIJ *)mat->B->data;
873: PetscInt nnwA, nnwB;
874: PetscBool nnzA, nnzB;
876: nnwA = aijA->nonew;
877: nnwB = aijB->nonew;
878: nnzA = aijA->keepnonzeropattern;
879: nnzB = aijB->keepnonzeropattern;
880: if (!nnzA) {
881: PetscCall(PetscInfo(mat->A, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n"));
882: aijA->nonew = 0;
883: }
884: if (!nnzB) {
885: PetscCall(PetscInfo(mat->B, "Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n"));
886: aijB->nonew = 0;
887: }
888: /* Must zero here before the next loop */
889: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
890: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
891: for (r = 0; r < len; ++r) {
892: const PetscInt row = lrows[r] + A->rmap->rstart;
893: if (row >= A->cmap->N) continue;
894: PetscCall(MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES));
895: }
896: aijA->nonew = nnwA;
897: aijB->nonew = nnwB;
898: } else {
899: PetscCall(MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL));
900: PetscCall(MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL));
901: }
902: PetscCall(PetscFree(lrows));
903: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
904: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
906: /* only change matrix nonzero state if pattern was allowed to be changed */
907: if (!((Mat_SeqAIJ *)mat->A->data)->keepnonzeropattern || !((Mat_SeqAIJ *)mat->A->data)->nonew) {
908: PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
909: PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
910: }
911: PetscFunctionReturn(PETSC_SUCCESS);
912: }
914: static PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
915: {
916: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
917: PetscMPIInt n = A->rmap->n;
918: PetscInt i, j, r, m, len = 0;
919: PetscInt *lrows, *owners = A->rmap->range;
920: PetscMPIInt p = 0;
921: PetscSFNode *rrows;
922: PetscSF sf;
923: const PetscScalar *xx;
924: PetscScalar *bb, *mask, *aij_a;
925: Vec xmask, lmask;
926: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)l->B->data;
927: const PetscInt *aj, *ii, *ridx;
928: PetscScalar *aa;
930: PetscFunctionBegin;
931: /* Create SF where leaves are input rows and roots are owned rows */
932: PetscCall(PetscMalloc1(n, &lrows));
933: for (r = 0; r < n; ++r) lrows[r] = -1;
934: PetscCall(PetscMalloc1(N, &rrows));
935: for (r = 0; r < N; ++r) {
936: const PetscInt idx = rows[r];
937: PetscCheck(idx >= 0 && A->rmap->N > idx, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %" PetscInt_FMT " out of range [0,%" PetscInt_FMT ")", idx, A->rmap->N);
938: if (idx < owners[p] || owners[p + 1] <= idx) { /* short-circuit the search if the last p owns this row too */
939: PetscCall(PetscLayoutFindOwner(A->rmap, idx, &p));
940: }
941: rrows[r].rank = p;
942: rrows[r].index = rows[r] - owners[p];
943: }
944: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
945: PetscCall(PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER));
946: /* Collect flags for rows to be zeroed */
947: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
948: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *)rows, lrows, MPI_LOR));
949: PetscCall(PetscSFDestroy(&sf));
950: /* Compress and put in row numbers */
951: for (r = 0; r < n; ++r)
952: if (lrows[r] >= 0) lrows[len++] = r;
953: /* zero diagonal part of matrix */
954: PetscCall(MatZeroRowsColumns(l->A, len, lrows, diag, x, b));
955: /* handle off-diagonal part of matrix */
956: PetscCall(MatCreateVecs(A, &xmask, NULL));
957: PetscCall(VecDuplicate(l->lvec, &lmask));
958: PetscCall(VecGetArray(xmask, &bb));
959: for (i = 0; i < len; i++) bb[lrows[i]] = 1;
960: PetscCall(VecRestoreArray(xmask, &bb));
961: PetscCall(VecScatterBegin(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
962: PetscCall(VecScatterEnd(l->Mvctx, xmask, lmask, ADD_VALUES, SCATTER_FORWARD));
963: PetscCall(VecDestroy(&xmask));
964: if (x && b) { /* this code is buggy when the row and column layout don't match */
965: PetscBool cong;
967: PetscCall(MatHasCongruentLayouts(A, &cong));
968: PetscCheck(cong, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need matching row/col layout");
969: PetscCall(VecScatterBegin(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
970: PetscCall(VecScatterEnd(l->Mvctx, x, l->lvec, INSERT_VALUES, SCATTER_FORWARD));
971: PetscCall(VecGetArrayRead(l->lvec, &xx));
972: PetscCall(VecGetArray(b, &bb));
973: }
974: PetscCall(VecGetArray(lmask, &mask));
975: /* remove zeroed rows of off-diagonal matrix */
976: PetscCall(MatSeqAIJGetArray(l->B, &aij_a));
977: ii = aij->i;
978: for (i = 0; i < len; i++) PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aij_a, ii[lrows[i]]), ii[lrows[i] + 1] - ii[lrows[i]]));
979: /* loop over all elements of off process part of matrix zeroing removed columns*/
980: if (aij->compressedrow.use) {
981: m = aij->compressedrow.nrows;
982: ii = aij->compressedrow.i;
983: ridx = aij->compressedrow.rindex;
984: for (i = 0; i < m; i++) {
985: n = ii[i + 1] - ii[i];
986: aj = aij->j + ii[i];
987: aa = aij_a + ii[i];
989: for (j = 0; j < n; j++) {
990: if (PetscAbsScalar(mask[*aj])) {
991: if (b) bb[*ridx] -= *aa * xx[*aj];
992: *aa = 0.0;
993: }
994: aa++;
995: aj++;
996: }
997: ridx++;
998: }
999: } else { /* do not use compressed row format */
1000: m = l->B->rmap->n;
1001: for (i = 0; i < m; i++) {
1002: n = ii[i + 1] - ii[i];
1003: aj = aij->j + ii[i];
1004: aa = aij_a + ii[i];
1005: for (j = 0; j < n; j++) {
1006: if (PetscAbsScalar(mask[*aj])) {
1007: if (b) bb[i] -= *aa * xx[*aj];
1008: *aa = 0.0;
1009: }
1010: aa++;
1011: aj++;
1012: }
1013: }
1014: }
1015: if (x && b) {
1016: PetscCall(VecRestoreArray(b, &bb));
1017: PetscCall(VecRestoreArrayRead(l->lvec, &xx));
1018: }
1019: PetscCall(MatSeqAIJRestoreArray(l->B, &aij_a));
1020: PetscCall(VecRestoreArray(lmask, &mask));
1021: PetscCall(VecDestroy(&lmask));
1022: PetscCall(PetscFree(lrows));
1024: /* only change matrix nonzero state if pattern was allowed to be changed */
1025: if (!((Mat_SeqAIJ *)l->A->data)->nonew) {
1026: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1027: PetscCall(MPIU_Allreduce(&state, &A->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)A)));
1028: }
1029: PetscFunctionReturn(PETSC_SUCCESS);
1030: }
1032: static PetscErrorCode MatMult_MPIAIJ(Mat A, Vec xx, Vec yy)
1033: {
1034: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1035: PetscInt nt;
1036: VecScatter Mvctx = a->Mvctx;
1038: PetscFunctionBegin;
1039: PetscCall(VecGetLocalSize(xx, &nt));
1040: PetscCheck(nt == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Incompatible partition of A (%" PetscInt_FMT ") and xx (%" PetscInt_FMT ")", A->cmap->n, nt);
1041: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1042: PetscUseTypeMethod(a->A, mult, xx, yy);
1043: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1044: PetscUseTypeMethod(a->B, multadd, a->lvec, yy, yy);
1045: PetscFunctionReturn(PETSC_SUCCESS);
1046: }
1048: static PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A, Vec bb, Vec xx)
1049: {
1050: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1052: PetscFunctionBegin;
1053: PetscCall(MatMultDiagonalBlock(a->A, bb, xx));
1054: PetscFunctionReturn(PETSC_SUCCESS);
1055: }
1057: static PetscErrorCode MatMultAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1058: {
1059: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1060: VecScatter Mvctx = a->Mvctx;
1062: PetscFunctionBegin;
1063: PetscCall(VecScatterBegin(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1064: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
1065: PetscCall(VecScatterEnd(Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
1066: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
1067: PetscFunctionReturn(PETSC_SUCCESS);
1068: }
1070: static PetscErrorCode MatMultTranspose_MPIAIJ(Mat A, Vec xx, Vec yy)
1071: {
1072: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1074: PetscFunctionBegin;
1075: /* do nondiagonal part */
1076: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1077: /* do local part */
1078: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
1079: /* add partial results together */
1080: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1081: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
1082: PetscFunctionReturn(PETSC_SUCCESS);
1083: }
1085: static PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat, Mat Bmat, PetscReal tol, PetscBool *f)
1086: {
1087: MPI_Comm comm;
1088: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)Amat->data, *Bij = (Mat_MPIAIJ *)Bmat->data;
1089: Mat Adia = Aij->A, Bdia = Bij->A, Aoff, Boff, *Aoffs, *Boffs;
1090: IS Me, Notme;
1091: PetscInt M, N, first, last, *notme, i;
1092: PetscBool lf;
1093: PetscMPIInt size;
1095: PetscFunctionBegin;
1096: /* Easy test: symmetric diagonal block */
1097: PetscCall(MatIsTranspose(Adia, Bdia, tol, &lf));
1098: PetscCall(MPIU_Allreduce(&lf, f, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)Amat)));
1099: if (!*f) PetscFunctionReturn(PETSC_SUCCESS);
1100: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
1101: PetscCallMPI(MPI_Comm_size(comm, &size));
1102: if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
1104: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1105: PetscCall(MatGetSize(Amat, &M, &N));
1106: PetscCall(MatGetOwnershipRange(Amat, &first, &last));
1107: PetscCall(PetscMalloc1(N - last + first, ¬me));
1108: for (i = 0; i < first; i++) notme[i] = i;
1109: for (i = last; i < M; i++) notme[i - last + first] = i;
1110: PetscCall(ISCreateGeneral(MPI_COMM_SELF, N - last + first, notme, PETSC_COPY_VALUES, &Notme));
1111: PetscCall(ISCreateStride(MPI_COMM_SELF, last - first, first, 1, &Me));
1112: PetscCall(MatCreateSubMatrices(Amat, 1, &Me, &Notme, MAT_INITIAL_MATRIX, &Aoffs));
1113: Aoff = Aoffs[0];
1114: PetscCall(MatCreateSubMatrices(Bmat, 1, &Notme, &Me, MAT_INITIAL_MATRIX, &Boffs));
1115: Boff = Boffs[0];
1116: PetscCall(MatIsTranspose(Aoff, Boff, tol, f));
1117: PetscCall(MatDestroyMatrices(1, &Aoffs));
1118: PetscCall(MatDestroyMatrices(1, &Boffs));
1119: PetscCall(ISDestroy(&Me));
1120: PetscCall(ISDestroy(&Notme));
1121: PetscCall(PetscFree(notme));
1122: PetscFunctionReturn(PETSC_SUCCESS);
1123: }
1125: static PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A, PetscReal tol, PetscBool *f)
1126: {
1127: PetscFunctionBegin;
1128: PetscCall(MatIsTranspose_MPIAIJ(A, A, tol, f));
1129: PetscFunctionReturn(PETSC_SUCCESS);
1130: }
1132: static PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1133: {
1134: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1136: PetscFunctionBegin;
1137: /* do nondiagonal part */
1138: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1139: /* do local part */
1140: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1141: /* add partial results together */
1142: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1143: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1144: PetscFunctionReturn(PETSC_SUCCESS);
1145: }
1147: /*
1148: This only works correctly for square matrices where the subblock A->A is the
1149: diagonal block
1150: */
1151: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1152: {
1153: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1155: PetscFunctionBegin;
1156: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1157: 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");
1158: PetscCall(MatGetDiagonal(a->A, v));
1159: PetscFunctionReturn(PETSC_SUCCESS);
1160: }
1162: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1163: {
1164: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1166: PetscFunctionBegin;
1167: PetscCall(MatScale(a->A, aa));
1168: PetscCall(MatScale(a->B, aa));
1169: PetscFunctionReturn(PETSC_SUCCESS);
1170: }
1172: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1173: {
1174: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1175: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1176: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1177: const PetscInt *garray = aij->garray;
1178: const PetscScalar *aa, *ba;
1179: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1180: PetscInt64 nz, hnz;
1181: PetscInt *rowlens;
1182: PetscInt *colidxs;
1183: PetscScalar *matvals;
1184: PetscMPIInt rank;
1186: PetscFunctionBegin;
1187: PetscCall(PetscViewerSetUp(viewer));
1189: M = mat->rmap->N;
1190: N = mat->cmap->N;
1191: m = mat->rmap->n;
1192: rs = mat->rmap->rstart;
1193: cs = mat->cmap->rstart;
1194: nz = A->nz + B->nz;
1196: /* write matrix header */
1197: header[0] = MAT_FILE_CLASSID;
1198: header[1] = M;
1199: header[2] = N;
1200: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1201: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1202: if (rank == 0) {
1203: if (hnz > PETSC_MAX_INT) header[3] = PETSC_MAX_INT;
1204: else header[3] = (PetscInt)hnz;
1205: }
1206: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1208: /* fill in and store row lengths */
1209: PetscCall(PetscMalloc1(m, &rowlens));
1210: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1211: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1212: PetscCall(PetscFree(rowlens));
1214: /* fill in and store column indices */
1215: PetscCall(PetscMalloc1(nz, &colidxs));
1216: for (cnt = 0, i = 0; i < m; i++) {
1217: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1218: if (garray[B->j[jb]] > cs) break;
1219: colidxs[cnt++] = garray[B->j[jb]];
1220: }
1221: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1222: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1223: }
1224: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1225: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1226: PetscCall(PetscFree(colidxs));
1228: /* fill in and store nonzero values */
1229: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1230: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1231: PetscCall(PetscMalloc1(nz, &matvals));
1232: for (cnt = 0, i = 0; i < m; i++) {
1233: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1234: if (garray[B->j[jb]] > cs) break;
1235: matvals[cnt++] = ba[jb];
1236: }
1237: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1238: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1239: }
1240: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1241: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1242: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1243: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1244: PetscCall(PetscFree(matvals));
1246: /* write block size option to the viewer's .info file */
1247: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1248: PetscFunctionReturn(PETSC_SUCCESS);
1249: }
1251: #include <petscdraw.h>
1252: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1253: {
1254: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1255: PetscMPIInt rank = aij->rank, size = aij->size;
1256: PetscBool isdraw, iascii, isbinary;
1257: PetscViewer sviewer;
1258: PetscViewerFormat format;
1260: PetscFunctionBegin;
1261: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1262: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1263: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1264: if (iascii) {
1265: PetscCall(PetscViewerGetFormat(viewer, &format));
1266: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1267: PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1268: PetscCall(PetscMalloc1(size, &nz));
1269: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1270: for (i = 0; i < (PetscInt)size; i++) {
1271: nmax = PetscMax(nmax, nz[i]);
1272: nmin = PetscMin(nmin, nz[i]);
1273: navg += nz[i];
1274: }
1275: PetscCall(PetscFree(nz));
1276: navg = navg / size;
1277: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1278: PetscFunctionReturn(PETSC_SUCCESS);
1279: }
1280: PetscCall(PetscViewerGetFormat(viewer, &format));
1281: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1282: MatInfo info;
1283: PetscInt *inodes = NULL;
1285: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1286: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1287: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1288: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1289: if (!inodes) {
1290: 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,
1291: (double)info.memory));
1292: } else {
1293: 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,
1294: (double)info.memory));
1295: }
1296: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1297: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1298: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1299: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1300: PetscCall(PetscViewerFlush(viewer));
1301: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1302: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1303: PetscCall(VecScatterView(aij->Mvctx, viewer));
1304: PetscFunctionReturn(PETSC_SUCCESS);
1305: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1306: PetscInt inodecount, inodelimit, *inodes;
1307: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1308: if (inodes) {
1309: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1310: } else {
1311: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1312: }
1313: PetscFunctionReturn(PETSC_SUCCESS);
1314: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1315: PetscFunctionReturn(PETSC_SUCCESS);
1316: }
1317: } else if (isbinary) {
1318: if (size == 1) {
1319: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1320: PetscCall(MatView(aij->A, viewer));
1321: } else {
1322: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1323: }
1324: PetscFunctionReturn(PETSC_SUCCESS);
1325: } else if (iascii && size == 1) {
1326: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1327: PetscCall(MatView(aij->A, viewer));
1328: PetscFunctionReturn(PETSC_SUCCESS);
1329: } else if (isdraw) {
1330: PetscDraw draw;
1331: PetscBool isnull;
1332: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1333: PetscCall(PetscDrawIsNull(draw, &isnull));
1334: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1335: }
1337: { /* assemble the entire matrix onto first processor */
1338: Mat 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(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1344: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1345: /* The commented code uses MatCreateSubMatrices instead */
1346: /*
1347: Mat *AA, A = NULL, Av;
1348: IS isrow,iscol;
1350: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1351: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1352: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1353: if (rank == 0) {
1354: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1355: A = AA[0];
1356: Av = AA[0];
1357: }
1358: PetscCall(MatDestroySubMatrices(1,&AA));
1359: */
1360: PetscCall(ISDestroy(&iscol));
1361: PetscCall(ISDestroy(&isrow));
1362: /*
1363: Everyone has to call to draw the matrix since the graphics waits are
1364: synchronized across all processors that share the PetscDraw object
1365: */
1366: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1367: if (rank == 0) {
1368: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1369: PetscCall(MatView_SeqAIJ(Av, sviewer));
1370: }
1371: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1372: PetscCall(MatDestroy(&A));
1373: }
1374: PetscFunctionReturn(PETSC_SUCCESS);
1375: }
1377: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1378: {
1379: PetscBool iascii, isdraw, issocket, isbinary;
1381: PetscFunctionBegin;
1382: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1383: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1384: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1385: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1386: if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1387: PetscFunctionReturn(PETSC_SUCCESS);
1388: }
1390: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1391: {
1392: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1393: Vec bb1 = NULL;
1394: PetscBool hasop;
1396: PetscFunctionBegin;
1397: if (flag == SOR_APPLY_UPPER) {
1398: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1399: PetscFunctionReturn(PETSC_SUCCESS);
1400: }
1402: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1404: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1405: if (flag & SOR_ZERO_INITIAL_GUESS) {
1406: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1407: its--;
1408: }
1410: while (its--) {
1411: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1412: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1414: /* update rhs: bb1 = bb - B*x */
1415: PetscCall(VecScale(mat->lvec, -1.0));
1416: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1418: /* local sweep */
1419: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1420: }
1421: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1422: if (flag & SOR_ZERO_INITIAL_GUESS) {
1423: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1424: its--;
1425: }
1426: while (its--) {
1427: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1428: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1430: /* update rhs: bb1 = bb - B*x */
1431: PetscCall(VecScale(mat->lvec, -1.0));
1432: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1434: /* local sweep */
1435: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1436: }
1437: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1438: if (flag & SOR_ZERO_INITIAL_GUESS) {
1439: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1440: its--;
1441: }
1442: while (its--) {
1443: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1444: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1446: /* update rhs: bb1 = bb - B*x */
1447: PetscCall(VecScale(mat->lvec, -1.0));
1448: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1450: /* local sweep */
1451: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1452: }
1453: } else if (flag & SOR_EISENSTAT) {
1454: Vec xx1;
1456: PetscCall(VecDuplicate(bb, &xx1));
1457: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1459: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1460: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1461: if (!mat->diag) {
1462: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1463: PetscCall(MatGetDiagonal(matin, mat->diag));
1464: }
1465: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1466: if (hasop) {
1467: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1468: } else {
1469: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1470: }
1471: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1473: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1475: /* local sweep */
1476: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1477: PetscCall(VecAXPY(xx, 1.0, xx1));
1478: PetscCall(VecDestroy(&xx1));
1479: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1481: PetscCall(VecDestroy(&bb1));
1483: matin->factorerrortype = mat->A->factorerrortype;
1484: PetscFunctionReturn(PETSC_SUCCESS);
1485: }
1487: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1488: {
1489: Mat aA, aB, Aperm;
1490: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1491: PetscScalar *aa, *ba;
1492: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1493: PetscSF rowsf, sf;
1494: IS parcolp = NULL;
1495: PetscBool done;
1497: PetscFunctionBegin;
1498: PetscCall(MatGetLocalSize(A, &m, &n));
1499: PetscCall(ISGetIndices(rowp, &rwant));
1500: PetscCall(ISGetIndices(colp, &cwant));
1501: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1503: /* Invert row permutation to find out where my rows should go */
1504: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1505: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1506: PetscCall(PetscSFSetFromOptions(rowsf));
1507: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1508: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1509: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1511: /* Invert column permutation to find out where my columns should go */
1512: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1513: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1514: PetscCall(PetscSFSetFromOptions(sf));
1515: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1516: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1517: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1518: PetscCall(PetscSFDestroy(&sf));
1520: PetscCall(ISRestoreIndices(rowp, &rwant));
1521: PetscCall(ISRestoreIndices(colp, &cwant));
1522: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1524: /* Find out where my gcols should go */
1525: PetscCall(MatGetSize(aB, NULL, &ng));
1526: PetscCall(PetscMalloc1(ng, &gcdest));
1527: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1528: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1529: PetscCall(PetscSFSetFromOptions(sf));
1530: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1531: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1532: PetscCall(PetscSFDestroy(&sf));
1534: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1535: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1536: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1537: for (i = 0; i < m; i++) {
1538: PetscInt row = rdest[i];
1539: PetscMPIInt rowner;
1540: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1541: for (j = ai[i]; j < ai[i + 1]; j++) {
1542: PetscInt col = cdest[aj[j]];
1543: PetscMPIInt cowner;
1544: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1545: if (rowner == cowner) dnnz[i]++;
1546: else onnz[i]++;
1547: }
1548: for (j = bi[i]; j < bi[i + 1]; j++) {
1549: PetscInt col = gcdest[bj[j]];
1550: PetscMPIInt cowner;
1551: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1552: if (rowner == cowner) dnnz[i]++;
1553: else onnz[i]++;
1554: }
1555: }
1556: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1557: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1558: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1559: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1560: PetscCall(PetscSFDestroy(&rowsf));
1562: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1563: PetscCall(MatSeqAIJGetArray(aA, &aa));
1564: PetscCall(MatSeqAIJGetArray(aB, &ba));
1565: for (i = 0; i < m; i++) {
1566: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1567: PetscInt j0, rowlen;
1568: rowlen = ai[i + 1] - ai[i];
1569: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1570: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1571: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1572: }
1573: rowlen = bi[i + 1] - bi[i];
1574: for (j0 = j = 0; j < rowlen; j0 = j) {
1575: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1576: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1577: }
1578: }
1579: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1580: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1581: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1582: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1583: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1584: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1585: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1586: PetscCall(PetscFree3(work, rdest, cdest));
1587: PetscCall(PetscFree(gcdest));
1588: if (parcolp) PetscCall(ISDestroy(&colp));
1589: *B = Aperm;
1590: PetscFunctionReturn(PETSC_SUCCESS);
1591: }
1593: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1594: {
1595: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1597: PetscFunctionBegin;
1598: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1599: if (ghosts) *ghosts = aij->garray;
1600: PetscFunctionReturn(PETSC_SUCCESS);
1601: }
1603: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1604: {
1605: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1606: Mat A = mat->A, B = mat->B;
1607: PetscLogDouble isend[5], irecv[5];
1609: PetscFunctionBegin;
1610: info->block_size = 1.0;
1611: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1613: isend[0] = info->nz_used;
1614: isend[1] = info->nz_allocated;
1615: isend[2] = info->nz_unneeded;
1616: isend[3] = info->memory;
1617: isend[4] = info->mallocs;
1619: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1621: isend[0] += info->nz_used;
1622: isend[1] += info->nz_allocated;
1623: isend[2] += info->nz_unneeded;
1624: isend[3] += info->memory;
1625: isend[4] += info->mallocs;
1626: if (flag == MAT_LOCAL) {
1627: info->nz_used = isend[0];
1628: info->nz_allocated = isend[1];
1629: info->nz_unneeded = isend[2];
1630: info->memory = isend[3];
1631: info->mallocs = isend[4];
1632: } else if (flag == MAT_GLOBAL_MAX) {
1633: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1635: info->nz_used = irecv[0];
1636: info->nz_allocated = irecv[1];
1637: info->nz_unneeded = irecv[2];
1638: info->memory = irecv[3];
1639: info->mallocs = irecv[4];
1640: } else if (flag == MAT_GLOBAL_SUM) {
1641: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1643: info->nz_used = irecv[0];
1644: info->nz_allocated = irecv[1];
1645: info->nz_unneeded = irecv[2];
1646: info->memory = irecv[3];
1647: info->mallocs = irecv[4];
1648: }
1649: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1650: info->fill_ratio_needed = 0;
1651: info->factor_mallocs = 0;
1652: PetscFunctionReturn(PETSC_SUCCESS);
1653: }
1655: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1656: {
1657: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1659: PetscFunctionBegin;
1660: switch (op) {
1661: case MAT_NEW_NONZERO_LOCATIONS:
1662: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1663: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1664: case MAT_KEEP_NONZERO_PATTERN:
1665: case MAT_NEW_NONZERO_LOCATION_ERR:
1666: case MAT_USE_INODES:
1667: case MAT_IGNORE_ZERO_ENTRIES:
1668: case MAT_FORM_EXPLICIT_TRANSPOSE:
1669: MatCheckPreallocated(A, 1);
1670: PetscCall(MatSetOption(a->A, op, flg));
1671: PetscCall(MatSetOption(a->B, op, flg));
1672: break;
1673: case MAT_ROW_ORIENTED:
1674: MatCheckPreallocated(A, 1);
1675: a->roworiented = flg;
1677: PetscCall(MatSetOption(a->A, op, flg));
1678: PetscCall(MatSetOption(a->B, op, flg));
1679: break;
1680: case MAT_FORCE_DIAGONAL_ENTRIES:
1681: case MAT_SORTED_FULL:
1682: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1683: break;
1684: case MAT_IGNORE_OFF_PROC_ENTRIES:
1685: a->donotstash = flg;
1686: break;
1687: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1688: case MAT_SPD:
1689: case MAT_SYMMETRIC:
1690: case MAT_STRUCTURALLY_SYMMETRIC:
1691: case MAT_HERMITIAN:
1692: case MAT_SYMMETRY_ETERNAL:
1693: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1694: case MAT_SPD_ETERNAL:
1695: /* if the diagonal matrix is square it inherits some of the properties above */
1696: break;
1697: case MAT_SUBMAT_SINGLEIS:
1698: A->submat_singleis = flg;
1699: break;
1700: case MAT_STRUCTURE_ONLY:
1701: /* The option is handled directly by MatSetOption() */
1702: break;
1703: default:
1704: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1705: }
1706: PetscFunctionReturn(PETSC_SUCCESS);
1707: }
1709: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1710: {
1711: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1712: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1713: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1714: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1715: PetscInt *cmap, *idx_p;
1717: PetscFunctionBegin;
1718: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1719: mat->getrowactive = PETSC_TRUE;
1721: if (!mat->rowvalues && (idx || v)) {
1722: /*
1723: allocate enough space to hold information from the longest row.
1724: */
1725: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1726: PetscInt max = 1, tmp;
1727: for (i = 0; i < matin->rmap->n; i++) {
1728: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1729: if (max < tmp) max = tmp;
1730: }
1731: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1732: }
1734: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1735: lrow = row - rstart;
1737: pvA = &vworkA;
1738: pcA = &cworkA;
1739: pvB = &vworkB;
1740: pcB = &cworkB;
1741: if (!v) {
1742: pvA = NULL;
1743: pvB = NULL;
1744: }
1745: if (!idx) {
1746: pcA = NULL;
1747: if (!v) pcB = NULL;
1748: }
1749: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1750: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1751: nztot = nzA + nzB;
1753: cmap = mat->garray;
1754: if (v || idx) {
1755: if (nztot) {
1756: /* Sort by increasing column numbers, assuming A and B already sorted */
1757: PetscInt imark = -1;
1758: if (v) {
1759: *v = v_p = mat->rowvalues;
1760: for (i = 0; i < nzB; i++) {
1761: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1762: else break;
1763: }
1764: imark = i;
1765: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1766: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1767: }
1768: if (idx) {
1769: *idx = idx_p = mat->rowindices;
1770: if (imark > -1) {
1771: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1772: } else {
1773: for (i = 0; i < nzB; i++) {
1774: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1775: else break;
1776: }
1777: imark = i;
1778: }
1779: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1780: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1781: }
1782: } else {
1783: if (idx) *idx = NULL;
1784: if (v) *v = NULL;
1785: }
1786: }
1787: *nz = nztot;
1788: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1789: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1790: PetscFunctionReturn(PETSC_SUCCESS);
1791: }
1793: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1794: {
1795: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1797: PetscFunctionBegin;
1798: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1799: aij->getrowactive = PETSC_FALSE;
1800: PetscFunctionReturn(PETSC_SUCCESS);
1801: }
1803: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1804: {
1805: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1806: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1807: PetscInt i, j, cstart = mat->cmap->rstart;
1808: PetscReal sum = 0.0;
1809: const MatScalar *v, *amata, *bmata;
1811: PetscFunctionBegin;
1812: if (aij->size == 1) {
1813: PetscCall(MatNorm(aij->A, type, norm));
1814: } else {
1815: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1816: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1817: if (type == NORM_FROBENIUS) {
1818: v = amata;
1819: for (i = 0; i < amat->nz; i++) {
1820: sum += PetscRealPart(PetscConj(*v) * (*v));
1821: v++;
1822: }
1823: v = bmata;
1824: for (i = 0; i < bmat->nz; i++) {
1825: sum += PetscRealPart(PetscConj(*v) * (*v));
1826: v++;
1827: }
1828: PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1829: *norm = PetscSqrtReal(*norm);
1830: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1831: } else if (type == NORM_1) { /* max column norm */
1832: PetscReal *tmp, *tmp2;
1833: PetscInt *jj, *garray = aij->garray;
1834: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1835: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1836: *norm = 0.0;
1837: v = amata;
1838: jj = amat->j;
1839: for (j = 0; j < amat->nz; j++) {
1840: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1841: v++;
1842: }
1843: v = bmata;
1844: jj = bmat->j;
1845: for (j = 0; j < bmat->nz; j++) {
1846: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1847: v++;
1848: }
1849: PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1850: for (j = 0; j < mat->cmap->N; j++) {
1851: if (tmp2[j] > *norm) *norm = tmp2[j];
1852: }
1853: PetscCall(PetscFree(tmp));
1854: PetscCall(PetscFree(tmp2));
1855: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1856: } else if (type == NORM_INFINITY) { /* max row norm */
1857: PetscReal ntemp = 0.0;
1858: for (j = 0; j < aij->A->rmap->n; j++) {
1859: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1860: sum = 0.0;
1861: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1862: sum += PetscAbsScalar(*v);
1863: v++;
1864: }
1865: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1866: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1867: sum += PetscAbsScalar(*v);
1868: v++;
1869: }
1870: if (sum > ntemp) ntemp = sum;
1871: }
1872: PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1873: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1874: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1875: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1876: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1877: }
1878: PetscFunctionReturn(PETSC_SUCCESS);
1879: }
1881: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1882: {
1883: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1884: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1885: 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;
1886: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1887: Mat B, A_diag, *B_diag;
1888: const MatScalar *pbv, *bv;
1890: PetscFunctionBegin;
1891: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1892: ma = A->rmap->n;
1893: na = A->cmap->n;
1894: mb = a->B->rmap->n;
1895: nb = a->B->cmap->n;
1896: ai = Aloc->i;
1897: aj = Aloc->j;
1898: bi = Bloc->i;
1899: bj = Bloc->j;
1900: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1901: PetscInt *d_nnz, *g_nnz, *o_nnz;
1902: PetscSFNode *oloc;
1903: PETSC_UNUSED PetscSF sf;
1905: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1906: /* compute d_nnz for preallocation */
1907: PetscCall(PetscArrayzero(d_nnz, na));
1908: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1909: /* compute local off-diagonal contributions */
1910: PetscCall(PetscArrayzero(g_nnz, nb));
1911: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1912: /* map those to global */
1913: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1914: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1915: PetscCall(PetscSFSetFromOptions(sf));
1916: PetscCall(PetscArrayzero(o_nnz, na));
1917: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1918: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1919: PetscCall(PetscSFDestroy(&sf));
1921: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1922: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1923: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1924: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1925: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1926: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1927: } else {
1928: B = *matout;
1929: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1930: }
1932: b = (Mat_MPIAIJ *)B->data;
1933: A_diag = a->A;
1934: B_diag = &b->A;
1935: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1936: A_diag_ncol = A_diag->cmap->N;
1937: B_diag_ilen = sub_B_diag->ilen;
1938: B_diag_i = sub_B_diag->i;
1940: /* Set ilen for diagonal of B */
1941: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1943: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1944: very quickly (=without using MatSetValues), because all writes are local. */
1945: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1946: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1948: /* copy over the B part */
1949: PetscCall(PetscMalloc1(bi[mb], &cols));
1950: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1951: pbv = bv;
1952: row = A->rmap->rstart;
1953: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1954: cols_tmp = cols;
1955: for (i = 0; i < mb; i++) {
1956: ncol = bi[i + 1] - bi[i];
1957: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1958: row++;
1959: if (pbv) pbv += ncol;
1960: if (cols_tmp) cols_tmp += ncol;
1961: }
1962: PetscCall(PetscFree(cols));
1963: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1965: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1966: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1967: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1968: *matout = B;
1969: } else {
1970: PetscCall(MatHeaderMerge(A, &B));
1971: }
1972: PetscFunctionReturn(PETSC_SUCCESS);
1973: }
1975: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1976: {
1977: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1978: Mat a = aij->A, b = aij->B;
1979: PetscInt s1, s2, s3;
1981: PetscFunctionBegin;
1982: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1983: if (rr) {
1984: PetscCall(VecGetLocalSize(rr, &s1));
1985: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1986: /* Overlap communication with computation. */
1987: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1988: }
1989: if (ll) {
1990: PetscCall(VecGetLocalSize(ll, &s1));
1991: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1992: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1993: }
1994: /* scale the diagonal block */
1995: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1997: if (rr) {
1998: /* Do a scatter end and then right scale the off-diagonal block */
1999: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
2000: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
2001: }
2002: PetscFunctionReturn(PETSC_SUCCESS);
2003: }
2005: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2006: {
2007: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2009: PetscFunctionBegin;
2010: PetscCall(MatSetUnfactored(a->A));
2011: PetscFunctionReturn(PETSC_SUCCESS);
2012: }
2014: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2015: {
2016: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2017: Mat a, b, c, d;
2018: PetscBool flg;
2020: PetscFunctionBegin;
2021: a = matA->A;
2022: b = matA->B;
2023: c = matB->A;
2024: d = matB->B;
2026: PetscCall(MatEqual(a, c, &flg));
2027: if (flg) PetscCall(MatEqual(b, d, &flg));
2028: PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2029: PetscFunctionReturn(PETSC_SUCCESS);
2030: }
2032: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2033: {
2034: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2035: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2037: PetscFunctionBegin;
2038: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2039: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2040: /* because of the column compression in the off-processor part of the matrix a->B,
2041: the number of columns in a->B and b->B may be different, hence we cannot call
2042: the MatCopy() directly on the two parts. If need be, we can provide a more
2043: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2044: then copying the submatrices */
2045: PetscCall(MatCopy_Basic(A, B, str));
2046: } else {
2047: PetscCall(MatCopy(a->A, b->A, str));
2048: PetscCall(MatCopy(a->B, b->B, str));
2049: }
2050: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2051: PetscFunctionReturn(PETSC_SUCCESS);
2052: }
2054: /*
2055: Computes the number of nonzeros per row needed for preallocation when X and Y
2056: have different nonzero structure.
2057: */
2058: 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)
2059: {
2060: PetscInt i, j, k, nzx, nzy;
2062: PetscFunctionBegin;
2063: /* Set the number of nonzeros in the new matrix */
2064: for (i = 0; i < m; i++) {
2065: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2066: nzx = xi[i + 1] - xi[i];
2067: nzy = yi[i + 1] - yi[i];
2068: nnz[i] = 0;
2069: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2070: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2071: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2072: nnz[i]++;
2073: }
2074: for (; k < nzy; k++) nnz[i]++;
2075: }
2076: PetscFunctionReturn(PETSC_SUCCESS);
2077: }
2079: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2080: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2081: {
2082: PetscInt m = Y->rmap->N;
2083: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2084: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2086: PetscFunctionBegin;
2087: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2088: PetscFunctionReturn(PETSC_SUCCESS);
2089: }
2091: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2092: {
2093: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2095: PetscFunctionBegin;
2096: if (str == SAME_NONZERO_PATTERN) {
2097: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2098: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2099: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2100: PetscCall(MatAXPY_Basic(Y, a, X, str));
2101: } else {
2102: Mat B;
2103: PetscInt *nnz_d, *nnz_o;
2105: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2106: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2107: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2108: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2109: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2110: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2111: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2112: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2113: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2114: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2115: PetscCall(MatHeaderMerge(Y, &B));
2116: PetscCall(PetscFree(nnz_d));
2117: PetscCall(PetscFree(nnz_o));
2118: }
2119: PetscFunctionReturn(PETSC_SUCCESS);
2120: }
2122: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2124: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2125: {
2126: PetscFunctionBegin;
2127: if (PetscDefined(USE_COMPLEX)) {
2128: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2130: PetscCall(MatConjugate_SeqAIJ(aij->A));
2131: PetscCall(MatConjugate_SeqAIJ(aij->B));
2132: }
2133: PetscFunctionReturn(PETSC_SUCCESS);
2134: }
2136: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2137: {
2138: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2140: PetscFunctionBegin;
2141: PetscCall(MatRealPart(a->A));
2142: PetscCall(MatRealPart(a->B));
2143: PetscFunctionReturn(PETSC_SUCCESS);
2144: }
2146: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2147: {
2148: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2150: PetscFunctionBegin;
2151: PetscCall(MatImaginaryPart(a->A));
2152: PetscCall(MatImaginaryPart(a->B));
2153: PetscFunctionReturn(PETSC_SUCCESS);
2154: }
2156: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2157: {
2158: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2159: PetscInt i, *idxb = NULL, m = A->rmap->n;
2160: PetscScalar *va, *vv;
2161: Vec vB, vA;
2162: const PetscScalar *vb;
2164: PetscFunctionBegin;
2165: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2166: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2168: PetscCall(VecGetArrayWrite(vA, &va));
2169: if (idx) {
2170: for (i = 0; i < m; i++) {
2171: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2172: }
2173: }
2175: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2176: PetscCall(PetscMalloc1(m, &idxb));
2177: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2179: PetscCall(VecGetArrayWrite(v, &vv));
2180: PetscCall(VecGetArrayRead(vB, &vb));
2181: for (i = 0; i < m; i++) {
2182: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2183: vv[i] = vb[i];
2184: if (idx) idx[i] = a->garray[idxb[i]];
2185: } else {
2186: vv[i] = va[i];
2187: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2188: }
2189: }
2190: PetscCall(VecRestoreArrayWrite(vA, &vv));
2191: PetscCall(VecRestoreArrayWrite(vA, &va));
2192: PetscCall(VecRestoreArrayRead(vB, &vb));
2193: PetscCall(PetscFree(idxb));
2194: PetscCall(VecDestroy(&vA));
2195: PetscCall(VecDestroy(&vB));
2196: PetscFunctionReturn(PETSC_SUCCESS);
2197: }
2199: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2200: {
2201: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2202: PetscInt m = A->rmap->n;
2203: Vec vB, vA;
2205: PetscFunctionBegin;
2206: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vA));
2207: PetscCall(MatGetRowSumAbs(a->A, vA));
2208: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &vB));
2209: PetscCall(MatGetRowSumAbs(a->B, vB));
2210: PetscCall(VecAXPY(vA, 1.0, vB));
2211: PetscCall(VecDestroy(&vB));
2212: PetscCall(VecCopy(vA, v));
2213: PetscCall(VecDestroy(&vA));
2214: PetscFunctionReturn(PETSC_SUCCESS);
2215: }
2217: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2218: {
2219: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2220: PetscInt m = A->rmap->n, n = A->cmap->n;
2221: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2222: PetscInt *cmap = mat->garray;
2223: PetscInt *diagIdx, *offdiagIdx;
2224: Vec diagV, offdiagV;
2225: PetscScalar *a, *diagA, *offdiagA;
2226: const PetscScalar *ba, *bav;
2227: PetscInt r, j, col, ncols, *bi, *bj;
2228: Mat B = mat->B;
2229: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2231: PetscFunctionBegin;
2232: /* When a process holds entire A and other processes have no entry */
2233: if (A->cmap->N == n) {
2234: PetscCall(VecGetArrayWrite(v, &diagA));
2235: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2236: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2237: PetscCall(VecDestroy(&diagV));
2238: PetscCall(VecRestoreArrayWrite(v, &diagA));
2239: PetscFunctionReturn(PETSC_SUCCESS);
2240: } else if (n == 0) {
2241: if (m) {
2242: PetscCall(VecGetArrayWrite(v, &a));
2243: for (r = 0; r < m; r++) {
2244: a[r] = 0.0;
2245: if (idx) idx[r] = -1;
2246: }
2247: PetscCall(VecRestoreArrayWrite(v, &a));
2248: }
2249: PetscFunctionReturn(PETSC_SUCCESS);
2250: }
2252: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2253: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2254: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2255: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2257: /* Get offdiagIdx[] for implicit 0.0 */
2258: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2259: ba = bav;
2260: bi = b->i;
2261: bj = b->j;
2262: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2263: for (r = 0; r < m; r++) {
2264: ncols = bi[r + 1] - bi[r];
2265: if (ncols == A->cmap->N - n) { /* Brow is dense */
2266: offdiagA[r] = *ba;
2267: offdiagIdx[r] = cmap[0];
2268: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2269: offdiagA[r] = 0.0;
2271: /* Find first hole in the cmap */
2272: for (j = 0; j < ncols; j++) {
2273: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2274: if (col > j && j < cstart) {
2275: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2276: break;
2277: } else if (col > j + n && j >= cstart) {
2278: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2279: break;
2280: }
2281: }
2282: if (j == ncols && ncols < A->cmap->N - n) {
2283: /* a hole is outside compressed Bcols */
2284: if (ncols == 0) {
2285: if (cstart) {
2286: offdiagIdx[r] = 0;
2287: } else offdiagIdx[r] = cend;
2288: } else { /* ncols > 0 */
2289: offdiagIdx[r] = cmap[ncols - 1] + 1;
2290: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2291: }
2292: }
2293: }
2295: for (j = 0; j < ncols; j++) {
2296: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2297: offdiagA[r] = *ba;
2298: offdiagIdx[r] = cmap[*bj];
2299: }
2300: ba++;
2301: bj++;
2302: }
2303: }
2305: PetscCall(VecGetArrayWrite(v, &a));
2306: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2307: for (r = 0; r < m; ++r) {
2308: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2309: a[r] = diagA[r];
2310: if (idx) idx[r] = cstart + diagIdx[r];
2311: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2312: a[r] = diagA[r];
2313: if (idx) {
2314: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2315: idx[r] = cstart + diagIdx[r];
2316: } else idx[r] = offdiagIdx[r];
2317: }
2318: } else {
2319: a[r] = offdiagA[r];
2320: if (idx) idx[r] = offdiagIdx[r];
2321: }
2322: }
2323: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2324: PetscCall(VecRestoreArrayWrite(v, &a));
2325: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2326: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2327: PetscCall(VecDestroy(&diagV));
2328: PetscCall(VecDestroy(&offdiagV));
2329: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2330: PetscFunctionReturn(PETSC_SUCCESS);
2331: }
2333: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2334: {
2335: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2336: PetscInt m = A->rmap->n, n = A->cmap->n;
2337: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2338: PetscInt *cmap = mat->garray;
2339: PetscInt *diagIdx, *offdiagIdx;
2340: Vec diagV, offdiagV;
2341: PetscScalar *a, *diagA, *offdiagA;
2342: const PetscScalar *ba, *bav;
2343: PetscInt r, j, col, ncols, *bi, *bj;
2344: Mat B = mat->B;
2345: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2347: PetscFunctionBegin;
2348: /* When a process holds entire A and other processes have no entry */
2349: if (A->cmap->N == n) {
2350: PetscCall(VecGetArrayWrite(v, &diagA));
2351: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2352: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2353: PetscCall(VecDestroy(&diagV));
2354: PetscCall(VecRestoreArrayWrite(v, &diagA));
2355: PetscFunctionReturn(PETSC_SUCCESS);
2356: } else if (n == 0) {
2357: if (m) {
2358: PetscCall(VecGetArrayWrite(v, &a));
2359: for (r = 0; r < m; r++) {
2360: a[r] = PETSC_MAX_REAL;
2361: if (idx) idx[r] = -1;
2362: }
2363: PetscCall(VecRestoreArrayWrite(v, &a));
2364: }
2365: PetscFunctionReturn(PETSC_SUCCESS);
2366: }
2368: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2369: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2370: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2371: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2373: /* Get offdiagIdx[] for implicit 0.0 */
2374: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2375: ba = bav;
2376: bi = b->i;
2377: bj = b->j;
2378: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2379: for (r = 0; r < m; r++) {
2380: ncols = bi[r + 1] - bi[r];
2381: if (ncols == A->cmap->N - n) { /* Brow is dense */
2382: offdiagA[r] = *ba;
2383: offdiagIdx[r] = cmap[0];
2384: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2385: offdiagA[r] = 0.0;
2387: /* Find first hole in the cmap */
2388: for (j = 0; j < ncols; j++) {
2389: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2390: if (col > j && j < cstart) {
2391: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2392: break;
2393: } else if (col > j + n && j >= cstart) {
2394: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2395: break;
2396: }
2397: }
2398: if (j == ncols && ncols < A->cmap->N - n) {
2399: /* a hole is outside compressed Bcols */
2400: if (ncols == 0) {
2401: if (cstart) {
2402: offdiagIdx[r] = 0;
2403: } else offdiagIdx[r] = cend;
2404: } else { /* ncols > 0 */
2405: offdiagIdx[r] = cmap[ncols - 1] + 1;
2406: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2407: }
2408: }
2409: }
2411: for (j = 0; j < ncols; j++) {
2412: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2413: offdiagA[r] = *ba;
2414: offdiagIdx[r] = cmap[*bj];
2415: }
2416: ba++;
2417: bj++;
2418: }
2419: }
2421: PetscCall(VecGetArrayWrite(v, &a));
2422: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2423: for (r = 0; r < m; ++r) {
2424: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2425: a[r] = diagA[r];
2426: if (idx) idx[r] = cstart + diagIdx[r];
2427: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2428: a[r] = diagA[r];
2429: if (idx) {
2430: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2431: idx[r] = cstart + diagIdx[r];
2432: } else idx[r] = offdiagIdx[r];
2433: }
2434: } else {
2435: a[r] = offdiagA[r];
2436: if (idx) idx[r] = offdiagIdx[r];
2437: }
2438: }
2439: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2440: PetscCall(VecRestoreArrayWrite(v, &a));
2441: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2442: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2443: PetscCall(VecDestroy(&diagV));
2444: PetscCall(VecDestroy(&offdiagV));
2445: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2446: PetscFunctionReturn(PETSC_SUCCESS);
2447: }
2449: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2450: {
2451: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2452: PetscInt m = A->rmap->n, n = A->cmap->n;
2453: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2454: PetscInt *cmap = mat->garray;
2455: PetscInt *diagIdx, *offdiagIdx;
2456: Vec diagV, offdiagV;
2457: PetscScalar *a, *diagA, *offdiagA;
2458: const PetscScalar *ba, *bav;
2459: PetscInt r, j, col, ncols, *bi, *bj;
2460: Mat B = mat->B;
2461: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2463: PetscFunctionBegin;
2464: /* When a process holds entire A and other processes have no entry */
2465: if (A->cmap->N == n) {
2466: PetscCall(VecGetArrayWrite(v, &diagA));
2467: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2468: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2469: PetscCall(VecDestroy(&diagV));
2470: PetscCall(VecRestoreArrayWrite(v, &diagA));
2471: PetscFunctionReturn(PETSC_SUCCESS);
2472: } else if (n == 0) {
2473: if (m) {
2474: PetscCall(VecGetArrayWrite(v, &a));
2475: for (r = 0; r < m; r++) {
2476: a[r] = PETSC_MIN_REAL;
2477: if (idx) idx[r] = -1;
2478: }
2479: PetscCall(VecRestoreArrayWrite(v, &a));
2480: }
2481: PetscFunctionReturn(PETSC_SUCCESS);
2482: }
2484: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2485: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2486: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2487: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2489: /* Get offdiagIdx[] for implicit 0.0 */
2490: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2491: ba = bav;
2492: bi = b->i;
2493: bj = b->j;
2494: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2495: for (r = 0; r < m; r++) {
2496: ncols = bi[r + 1] - bi[r];
2497: if (ncols == A->cmap->N - n) { /* Brow is dense */
2498: offdiagA[r] = *ba;
2499: offdiagIdx[r] = cmap[0];
2500: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2501: offdiagA[r] = 0.0;
2503: /* Find first hole in the cmap */
2504: for (j = 0; j < ncols; j++) {
2505: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2506: if (col > j && j < cstart) {
2507: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2508: break;
2509: } else if (col > j + n && j >= cstart) {
2510: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2511: break;
2512: }
2513: }
2514: if (j == ncols && ncols < A->cmap->N - n) {
2515: /* a hole is outside compressed Bcols */
2516: if (ncols == 0) {
2517: if (cstart) {
2518: offdiagIdx[r] = 0;
2519: } else offdiagIdx[r] = cend;
2520: } else { /* ncols > 0 */
2521: offdiagIdx[r] = cmap[ncols - 1] + 1;
2522: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2523: }
2524: }
2525: }
2527: for (j = 0; j < ncols; j++) {
2528: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2529: offdiagA[r] = *ba;
2530: offdiagIdx[r] = cmap[*bj];
2531: }
2532: ba++;
2533: bj++;
2534: }
2535: }
2537: PetscCall(VecGetArrayWrite(v, &a));
2538: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2539: for (r = 0; r < m; ++r) {
2540: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2541: a[r] = diagA[r];
2542: if (idx) idx[r] = cstart + diagIdx[r];
2543: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2544: a[r] = diagA[r];
2545: if (idx) {
2546: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2547: idx[r] = cstart + diagIdx[r];
2548: } else idx[r] = offdiagIdx[r];
2549: }
2550: } else {
2551: a[r] = offdiagA[r];
2552: if (idx) idx[r] = offdiagIdx[r];
2553: }
2554: }
2555: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2556: PetscCall(VecRestoreArrayWrite(v, &a));
2557: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2558: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2559: PetscCall(VecDestroy(&diagV));
2560: PetscCall(VecDestroy(&offdiagV));
2561: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2562: PetscFunctionReturn(PETSC_SUCCESS);
2563: }
2565: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2566: {
2567: Mat *dummy;
2569: PetscFunctionBegin;
2570: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2571: *newmat = *dummy;
2572: PetscCall(PetscFree(dummy));
2573: PetscFunctionReturn(PETSC_SUCCESS);
2574: }
2576: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2577: {
2578: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2580: PetscFunctionBegin;
2581: PetscCall(MatInvertBlockDiagonal(a->A, values));
2582: A->factorerrortype = a->A->factorerrortype;
2583: PetscFunctionReturn(PETSC_SUCCESS);
2584: }
2586: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2587: {
2588: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2590: PetscFunctionBegin;
2591: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2592: PetscCall(MatSetRandom(aij->A, rctx));
2593: if (x->assembled) {
2594: PetscCall(MatSetRandom(aij->B, rctx));
2595: } else {
2596: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2597: }
2598: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2599: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2600: PetscFunctionReturn(PETSC_SUCCESS);
2601: }
2603: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2604: {
2605: PetscFunctionBegin;
2606: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2607: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2608: PetscFunctionReturn(PETSC_SUCCESS);
2609: }
2611: /*@
2612: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2614: Not Collective
2616: Input Parameter:
2617: . A - the matrix
2619: Output Parameter:
2620: . nz - the number of nonzeros
2622: Level: advanced
2624: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2625: @*/
2626: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2627: {
2628: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2629: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2630: PetscBool isaij;
2632: PetscFunctionBegin;
2633: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2634: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2635: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2636: PetscFunctionReturn(PETSC_SUCCESS);
2637: }
2639: /*@
2640: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2642: Collective
2644: Input Parameters:
2645: + A - the matrix
2646: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2648: Level: advanced
2650: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2651: @*/
2652: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2653: {
2654: PetscFunctionBegin;
2655: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2656: PetscFunctionReturn(PETSC_SUCCESS);
2657: }
2659: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2660: {
2661: PetscBool sc = PETSC_FALSE, flg;
2663: PetscFunctionBegin;
2664: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2665: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2666: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2667: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2668: PetscOptionsHeadEnd();
2669: PetscFunctionReturn(PETSC_SUCCESS);
2670: }
2672: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2673: {
2674: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2675: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2677: PetscFunctionBegin;
2678: if (!Y->preallocated) {
2679: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2680: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2681: PetscInt nonew = aij->nonew;
2682: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2683: aij->nonew = nonew;
2684: }
2685: PetscCall(MatShift_Basic(Y, a));
2686: PetscFunctionReturn(PETSC_SUCCESS);
2687: }
2689: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2690: {
2691: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2693: PetscFunctionBegin;
2694: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2695: PetscCall(MatMissingDiagonal(a->A, missing, d));
2696: if (d) {
2697: PetscInt rstart;
2698: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2699: *d += rstart;
2700: }
2701: PetscFunctionReturn(PETSC_SUCCESS);
2702: }
2704: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2705: {
2706: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2708: PetscFunctionBegin;
2709: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2710: PetscFunctionReturn(PETSC_SUCCESS);
2711: }
2713: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2714: {
2715: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2717: PetscFunctionBegin;
2718: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2719: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2720: PetscFunctionReturn(PETSC_SUCCESS);
2721: }
2723: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2724: MatGetRow_MPIAIJ,
2725: MatRestoreRow_MPIAIJ,
2726: MatMult_MPIAIJ,
2727: /* 4*/ MatMultAdd_MPIAIJ,
2728: MatMultTranspose_MPIAIJ,
2729: MatMultTransposeAdd_MPIAIJ,
2730: NULL,
2731: NULL,
2732: NULL,
2733: /*10*/ NULL,
2734: NULL,
2735: NULL,
2736: MatSOR_MPIAIJ,
2737: MatTranspose_MPIAIJ,
2738: /*15*/ MatGetInfo_MPIAIJ,
2739: MatEqual_MPIAIJ,
2740: MatGetDiagonal_MPIAIJ,
2741: MatDiagonalScale_MPIAIJ,
2742: MatNorm_MPIAIJ,
2743: /*20*/ MatAssemblyBegin_MPIAIJ,
2744: MatAssemblyEnd_MPIAIJ,
2745: MatSetOption_MPIAIJ,
2746: MatZeroEntries_MPIAIJ,
2747: /*24*/ MatZeroRows_MPIAIJ,
2748: NULL,
2749: NULL,
2750: NULL,
2751: NULL,
2752: /*29*/ MatSetUp_MPI_Hash,
2753: NULL,
2754: NULL,
2755: MatGetDiagonalBlock_MPIAIJ,
2756: NULL,
2757: /*34*/ MatDuplicate_MPIAIJ,
2758: NULL,
2759: NULL,
2760: NULL,
2761: NULL,
2762: /*39*/ MatAXPY_MPIAIJ,
2763: MatCreateSubMatrices_MPIAIJ,
2764: MatIncreaseOverlap_MPIAIJ,
2765: MatGetValues_MPIAIJ,
2766: MatCopy_MPIAIJ,
2767: /*44*/ MatGetRowMax_MPIAIJ,
2768: MatScale_MPIAIJ,
2769: MatShift_MPIAIJ,
2770: MatDiagonalSet_MPIAIJ,
2771: MatZeroRowsColumns_MPIAIJ,
2772: /*49*/ MatSetRandom_MPIAIJ,
2773: MatGetRowIJ_MPIAIJ,
2774: MatRestoreRowIJ_MPIAIJ,
2775: NULL,
2776: NULL,
2777: /*54*/ MatFDColoringCreate_MPIXAIJ,
2778: NULL,
2779: MatSetUnfactored_MPIAIJ,
2780: MatPermute_MPIAIJ,
2781: NULL,
2782: /*59*/ MatCreateSubMatrix_MPIAIJ,
2783: MatDestroy_MPIAIJ,
2784: MatView_MPIAIJ,
2785: NULL,
2786: NULL,
2787: /*64*/ NULL,
2788: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2789: NULL,
2790: NULL,
2791: NULL,
2792: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2793: MatGetRowMinAbs_MPIAIJ,
2794: NULL,
2795: NULL,
2796: NULL,
2797: NULL,
2798: /*75*/ MatFDColoringApply_AIJ,
2799: MatSetFromOptions_MPIAIJ,
2800: NULL,
2801: NULL,
2802: MatFindZeroDiagonals_MPIAIJ,
2803: /*80*/ NULL,
2804: NULL,
2805: NULL,
2806: /*83*/ MatLoad_MPIAIJ,
2807: MatIsSymmetric_MPIAIJ,
2808: NULL,
2809: NULL,
2810: NULL,
2811: NULL,
2812: /*89*/ NULL,
2813: NULL,
2814: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2815: NULL,
2816: NULL,
2817: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2818: NULL,
2819: NULL,
2820: NULL,
2821: MatBindToCPU_MPIAIJ,
2822: /*99*/ MatProductSetFromOptions_MPIAIJ,
2823: NULL,
2824: NULL,
2825: MatConjugate_MPIAIJ,
2826: NULL,
2827: /*104*/ MatSetValuesRow_MPIAIJ,
2828: MatRealPart_MPIAIJ,
2829: MatImaginaryPart_MPIAIJ,
2830: NULL,
2831: NULL,
2832: /*109*/ NULL,
2833: NULL,
2834: MatGetRowMin_MPIAIJ,
2835: NULL,
2836: MatMissingDiagonal_MPIAIJ,
2837: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2838: NULL,
2839: MatGetGhosts_MPIAIJ,
2840: NULL,
2841: NULL,
2842: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2843: NULL,
2844: NULL,
2845: NULL,
2846: MatGetMultiProcBlock_MPIAIJ,
2847: /*124*/ MatFindNonzeroRows_MPIAIJ,
2848: MatGetColumnReductions_MPIAIJ,
2849: MatInvertBlockDiagonal_MPIAIJ,
2850: MatInvertVariableBlockDiagonal_MPIAIJ,
2851: MatCreateSubMatricesMPI_MPIAIJ,
2852: /*129*/ NULL,
2853: NULL,
2854: NULL,
2855: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2856: NULL,
2857: /*134*/ NULL,
2858: NULL,
2859: NULL,
2860: NULL,
2861: NULL,
2862: /*139*/ MatSetBlockSizes_MPIAIJ,
2863: NULL,
2864: NULL,
2865: MatFDColoringSetUp_MPIXAIJ,
2866: MatFindOffBlockDiagonalEntries_MPIAIJ,
2867: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2868: /*145*/ NULL,
2869: NULL,
2870: NULL,
2871: MatCreateGraph_Simple_AIJ,
2872: NULL,
2873: /*150*/ NULL,
2874: MatEliminateZeros_MPIAIJ,
2875: MatGetRowSumAbs_MPIAIJ};
2877: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2878: {
2879: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2881: PetscFunctionBegin;
2882: PetscCall(MatStoreValues(aij->A));
2883: PetscCall(MatStoreValues(aij->B));
2884: PetscFunctionReturn(PETSC_SUCCESS);
2885: }
2887: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2888: {
2889: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2891: PetscFunctionBegin;
2892: PetscCall(MatRetrieveValues(aij->A));
2893: PetscCall(MatRetrieveValues(aij->B));
2894: PetscFunctionReturn(PETSC_SUCCESS);
2895: }
2897: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2898: {
2899: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2900: PetscMPIInt size;
2902: PetscFunctionBegin;
2903: if (B->hash_active) {
2904: B->ops[0] = b->cops;
2905: B->hash_active = PETSC_FALSE;
2906: }
2907: PetscCall(PetscLayoutSetUp(B->rmap));
2908: PetscCall(PetscLayoutSetUp(B->cmap));
2910: #if defined(PETSC_USE_CTABLE)
2911: PetscCall(PetscHMapIDestroy(&b->colmap));
2912: #else
2913: PetscCall(PetscFree(b->colmap));
2914: #endif
2915: PetscCall(PetscFree(b->garray));
2916: PetscCall(VecDestroy(&b->lvec));
2917: PetscCall(VecScatterDestroy(&b->Mvctx));
2919: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2921: MatSeqXAIJGetOptions_Private(b->B);
2922: PetscCall(MatDestroy(&b->B));
2923: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2924: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2925: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2926: PetscCall(MatSetType(b->B, MATSEQAIJ));
2927: MatSeqXAIJRestoreOptions_Private(b->B);
2929: MatSeqXAIJGetOptions_Private(b->A);
2930: PetscCall(MatDestroy(&b->A));
2931: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2932: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2933: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2934: PetscCall(MatSetType(b->A, MATSEQAIJ));
2935: MatSeqXAIJRestoreOptions_Private(b->A);
2937: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2938: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2939: B->preallocated = PETSC_TRUE;
2940: B->was_assembled = PETSC_FALSE;
2941: B->assembled = PETSC_FALSE;
2942: PetscFunctionReturn(PETSC_SUCCESS);
2943: }
2945: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2946: {
2947: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2949: PetscFunctionBegin;
2951: PetscCall(PetscLayoutSetUp(B->rmap));
2952: PetscCall(PetscLayoutSetUp(B->cmap));
2954: #if defined(PETSC_USE_CTABLE)
2955: PetscCall(PetscHMapIDestroy(&b->colmap));
2956: #else
2957: PetscCall(PetscFree(b->colmap));
2958: #endif
2959: PetscCall(PetscFree(b->garray));
2960: PetscCall(VecDestroy(&b->lvec));
2961: PetscCall(VecScatterDestroy(&b->Mvctx));
2963: PetscCall(MatResetPreallocation(b->A));
2964: PetscCall(MatResetPreallocation(b->B));
2965: B->preallocated = PETSC_TRUE;
2966: B->was_assembled = PETSC_FALSE;
2967: B->assembled = PETSC_FALSE;
2968: PetscFunctionReturn(PETSC_SUCCESS);
2969: }
2971: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2972: {
2973: Mat mat;
2974: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2976: PetscFunctionBegin;
2977: *newmat = NULL;
2978: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2979: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2980: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2981: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2982: a = (Mat_MPIAIJ *)mat->data;
2984: mat->factortype = matin->factortype;
2985: mat->assembled = matin->assembled;
2986: mat->insertmode = NOT_SET_VALUES;
2988: a->size = oldmat->size;
2989: a->rank = oldmat->rank;
2990: a->donotstash = oldmat->donotstash;
2991: a->roworiented = oldmat->roworiented;
2992: a->rowindices = NULL;
2993: a->rowvalues = NULL;
2994: a->getrowactive = PETSC_FALSE;
2996: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2997: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2998: if (matin->hash_active) {
2999: PetscCall(MatSetUp(mat));
3000: } else {
3001: mat->preallocated = matin->preallocated;
3002: if (oldmat->colmap) {
3003: #if defined(PETSC_USE_CTABLE)
3004: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3005: #else
3006: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3007: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3008: #endif
3009: } else a->colmap = NULL;
3010: if (oldmat->garray) {
3011: PetscInt len;
3012: len = oldmat->B->cmap->n;
3013: PetscCall(PetscMalloc1(len + 1, &a->garray));
3014: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3015: } else a->garray = NULL;
3017: /* It may happen MatDuplicate is called with a non-assembled matrix
3018: In fact, MatDuplicate only requires the matrix to be preallocated
3019: This may happen inside a DMCreateMatrix_Shell */
3020: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3021: if (oldmat->Mvctx) PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3022: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3023: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3024: }
3025: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3026: *newmat = mat;
3027: PetscFunctionReturn(PETSC_SUCCESS);
3028: }
3030: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3031: {
3032: PetscBool isbinary, ishdf5;
3034: PetscFunctionBegin;
3037: /* force binary viewer to load .info file if it has not yet done so */
3038: PetscCall(PetscViewerSetUp(viewer));
3039: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3040: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3041: if (isbinary) {
3042: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3043: } else if (ishdf5) {
3044: #if defined(PETSC_HAVE_HDF5)
3045: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3046: #else
3047: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3048: #endif
3049: } else {
3050: 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);
3051: }
3052: PetscFunctionReturn(PETSC_SUCCESS);
3053: }
3055: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3056: {
3057: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3058: PetscInt *rowidxs, *colidxs;
3059: PetscScalar *matvals;
3061: PetscFunctionBegin;
3062: PetscCall(PetscViewerSetUp(viewer));
3064: /* read in matrix header */
3065: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3066: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3067: M = header[1];
3068: N = header[2];
3069: nz = header[3];
3070: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3071: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3072: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3074: /* set block sizes from the viewer's .info file */
3075: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3076: /* set global sizes if not set already */
3077: if (mat->rmap->N < 0) mat->rmap->N = M;
3078: if (mat->cmap->N < 0) mat->cmap->N = N;
3079: PetscCall(PetscLayoutSetUp(mat->rmap));
3080: PetscCall(PetscLayoutSetUp(mat->cmap));
3082: /* check if the matrix sizes are correct */
3083: PetscCall(MatGetSize(mat, &rows, &cols));
3084: 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);
3086: /* read in row lengths and build row indices */
3087: PetscCall(MatGetLocalSize(mat, &m, NULL));
3088: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3089: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3090: rowidxs[0] = 0;
3091: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3092: if (nz != PETSC_MAX_INT) {
3093: PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3094: 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);
3095: }
3097: /* read in column indices and matrix values */
3098: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3099: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3100: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3101: /* store matrix indices and values */
3102: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3103: PetscCall(PetscFree(rowidxs));
3104: PetscCall(PetscFree2(colidxs, matvals));
3105: PetscFunctionReturn(PETSC_SUCCESS);
3106: }
3108: /* Not scalable because of ISAllGather() unless getting all columns. */
3109: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3110: {
3111: IS iscol_local;
3112: PetscBool isstride;
3113: PetscMPIInt lisstride = 0, gisstride;
3115: PetscFunctionBegin;
3116: /* check if we are grabbing all columns*/
3117: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3119: if (isstride) {
3120: PetscInt start, len, mstart, mlen;
3121: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3122: PetscCall(ISGetLocalSize(iscol, &len));
3123: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3124: if (mstart == start && mlen - mstart == len) lisstride = 1;
3125: }
3127: PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3128: if (gisstride) {
3129: PetscInt N;
3130: PetscCall(MatGetSize(mat, NULL, &N));
3131: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3132: PetscCall(ISSetIdentity(iscol_local));
3133: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3134: } else {
3135: PetscInt cbs;
3136: PetscCall(ISGetBlockSize(iscol, &cbs));
3137: PetscCall(ISAllGather(iscol, &iscol_local));
3138: PetscCall(ISSetBlockSize(iscol_local, cbs));
3139: }
3141: *isseq = iscol_local;
3142: PetscFunctionReturn(PETSC_SUCCESS);
3143: }
3145: /*
3146: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3147: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3149: Input Parameters:
3150: + mat - matrix
3151: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3152: i.e., mat->rstart <= isrow[i] < mat->rend
3153: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3154: i.e., mat->cstart <= iscol[i] < mat->cend
3156: Output Parameters:
3157: + isrow_d - sequential row index set for retrieving mat->A
3158: . iscol_d - sequential column index set for retrieving mat->A
3159: . iscol_o - sequential column index set for retrieving mat->B
3160: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3161: */
3162: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3163: {
3164: Vec x, cmap;
3165: const PetscInt *is_idx;
3166: PetscScalar *xarray, *cmaparray;
3167: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3168: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3169: Mat B = a->B;
3170: Vec lvec = a->lvec, lcmap;
3171: PetscInt i, cstart, cend, Bn = B->cmap->N;
3172: MPI_Comm comm;
3173: VecScatter Mvctx = a->Mvctx;
3175: PetscFunctionBegin;
3176: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3177: PetscCall(ISGetLocalSize(iscol, &ncols));
3179: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3180: PetscCall(MatCreateVecs(mat, &x, NULL));
3181: PetscCall(VecSet(x, -1.0));
3182: PetscCall(VecDuplicate(x, &cmap));
3183: PetscCall(VecSet(cmap, -1.0));
3185: /* Get start indices */
3186: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3187: isstart -= ncols;
3188: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3190: PetscCall(ISGetIndices(iscol, &is_idx));
3191: PetscCall(VecGetArray(x, &xarray));
3192: PetscCall(VecGetArray(cmap, &cmaparray));
3193: PetscCall(PetscMalloc1(ncols, &idx));
3194: for (i = 0; i < ncols; i++) {
3195: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3196: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3197: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3198: }
3199: PetscCall(VecRestoreArray(x, &xarray));
3200: PetscCall(VecRestoreArray(cmap, &cmaparray));
3201: PetscCall(ISRestoreIndices(iscol, &is_idx));
3203: /* Get iscol_d */
3204: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3205: PetscCall(ISGetBlockSize(iscol, &i));
3206: PetscCall(ISSetBlockSize(*iscol_d, i));
3208: /* Get isrow_d */
3209: PetscCall(ISGetLocalSize(isrow, &m));
3210: rstart = mat->rmap->rstart;
3211: PetscCall(PetscMalloc1(m, &idx));
3212: PetscCall(ISGetIndices(isrow, &is_idx));
3213: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3214: PetscCall(ISRestoreIndices(isrow, &is_idx));
3216: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3217: PetscCall(ISGetBlockSize(isrow, &i));
3218: PetscCall(ISSetBlockSize(*isrow_d, i));
3220: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3221: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3222: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3224: PetscCall(VecDuplicate(lvec, &lcmap));
3226: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3227: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3229: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3230: /* off-process column indices */
3231: count = 0;
3232: PetscCall(PetscMalloc1(Bn, &idx));
3233: PetscCall(PetscMalloc1(Bn, &cmap1));
3235: PetscCall(VecGetArray(lvec, &xarray));
3236: PetscCall(VecGetArray(lcmap, &cmaparray));
3237: for (i = 0; i < Bn; i++) {
3238: if (PetscRealPart(xarray[i]) > -1.0) {
3239: idx[count] = i; /* local column index in off-diagonal part B */
3240: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3241: count++;
3242: }
3243: }
3244: PetscCall(VecRestoreArray(lvec, &xarray));
3245: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3247: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3248: /* cannot ensure iscol_o has same blocksize as iscol! */
3250: PetscCall(PetscFree(idx));
3251: *garray = cmap1;
3253: PetscCall(VecDestroy(&x));
3254: PetscCall(VecDestroy(&cmap));
3255: PetscCall(VecDestroy(&lcmap));
3256: PetscFunctionReturn(PETSC_SUCCESS);
3257: }
3259: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3260: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3261: {
3262: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3263: Mat M = NULL;
3264: MPI_Comm comm;
3265: IS iscol_d, isrow_d, iscol_o;
3266: Mat Asub = NULL, Bsub = NULL;
3267: PetscInt n;
3269: PetscFunctionBegin;
3270: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3272: if (call == MAT_REUSE_MATRIX) {
3273: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3274: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3275: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3277: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3278: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3280: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3281: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3283: /* Update diagonal and off-diagonal portions of submat */
3284: asub = (Mat_MPIAIJ *)(*submat)->data;
3285: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3286: PetscCall(ISGetLocalSize(iscol_o, &n));
3287: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3288: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3289: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3291: } else { /* call == MAT_INITIAL_MATRIX) */
3292: const PetscInt *garray;
3293: PetscInt BsubN;
3295: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3296: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3298: /* Create local submatrices Asub and Bsub */
3299: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3300: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3302: /* Create submatrix M */
3303: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3305: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3306: asub = (Mat_MPIAIJ *)M->data;
3308: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3309: n = asub->B->cmap->N;
3310: if (BsubN > n) {
3311: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3312: const PetscInt *idx;
3313: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3314: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3316: PetscCall(PetscMalloc1(n, &idx_new));
3317: j = 0;
3318: PetscCall(ISGetIndices(iscol_o, &idx));
3319: for (i = 0; i < n; i++) {
3320: if (j >= BsubN) break;
3321: while (subgarray[i] > garray[j]) j++;
3323: if (subgarray[i] == garray[j]) {
3324: idx_new[i] = idx[j++];
3325: } 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]);
3326: }
3327: PetscCall(ISRestoreIndices(iscol_o, &idx));
3329: PetscCall(ISDestroy(&iscol_o));
3330: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3332: } else if (BsubN < n) {
3333: 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);
3334: }
3336: PetscCall(PetscFree(garray));
3337: *submat = M;
3339: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3340: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3341: PetscCall(ISDestroy(&isrow_d));
3343: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3344: PetscCall(ISDestroy(&iscol_d));
3346: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3347: PetscCall(ISDestroy(&iscol_o));
3348: }
3349: PetscFunctionReturn(PETSC_SUCCESS);
3350: }
3352: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3353: {
3354: IS iscol_local = NULL, isrow_d;
3355: PetscInt csize;
3356: PetscInt n, i, j, start, end;
3357: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3358: MPI_Comm comm;
3360: PetscFunctionBegin;
3361: /* If isrow has same processor distribution as mat,
3362: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3363: if (call == MAT_REUSE_MATRIX) {
3364: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3365: if (isrow_d) {
3366: sameRowDist = PETSC_TRUE;
3367: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3368: } else {
3369: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3370: if (iscol_local) {
3371: sameRowDist = PETSC_TRUE;
3372: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3373: }
3374: }
3375: } else {
3376: /* Check if isrow has same processor distribution as mat */
3377: sameDist[0] = PETSC_FALSE;
3378: PetscCall(ISGetLocalSize(isrow, &n));
3379: if (!n) {
3380: sameDist[0] = PETSC_TRUE;
3381: } else {
3382: PetscCall(ISGetMinMax(isrow, &i, &j));
3383: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3384: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3385: }
3387: /* Check if iscol has same processor distribution as mat */
3388: sameDist[1] = PETSC_FALSE;
3389: PetscCall(ISGetLocalSize(iscol, &n));
3390: if (!n) {
3391: sameDist[1] = PETSC_TRUE;
3392: } else {
3393: PetscCall(ISGetMinMax(iscol, &i, &j));
3394: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3395: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3396: }
3398: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3399: PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3400: sameRowDist = tsameDist[0];
3401: }
3403: if (sameRowDist) {
3404: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3405: /* isrow and iscol have same processor distribution as mat */
3406: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3407: PetscFunctionReturn(PETSC_SUCCESS);
3408: } else { /* sameRowDist */
3409: /* isrow has same processor distribution as mat */
3410: if (call == MAT_INITIAL_MATRIX) {
3411: PetscBool sorted;
3412: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3413: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3414: PetscCall(ISGetSize(iscol, &i));
3415: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3417: PetscCall(ISSorted(iscol_local, &sorted));
3418: if (sorted) {
3419: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3420: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3421: PetscFunctionReturn(PETSC_SUCCESS);
3422: }
3423: } else { /* call == MAT_REUSE_MATRIX */
3424: IS iscol_sub;
3425: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3426: if (iscol_sub) {
3427: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3428: PetscFunctionReturn(PETSC_SUCCESS);
3429: }
3430: }
3431: }
3432: }
3434: /* General case: iscol -> iscol_local which has global size of iscol */
3435: if (call == MAT_REUSE_MATRIX) {
3436: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3437: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3438: } else {
3439: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3440: }
3442: PetscCall(ISGetLocalSize(iscol, &csize));
3443: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3445: if (call == MAT_INITIAL_MATRIX) {
3446: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3447: PetscCall(ISDestroy(&iscol_local));
3448: }
3449: PetscFunctionReturn(PETSC_SUCCESS);
3450: }
3452: /*@C
3453: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3454: and "off-diagonal" part of the matrix in CSR format.
3456: Collective
3458: Input Parameters:
3459: + comm - MPI communicator
3460: . A - "diagonal" portion of matrix
3461: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3462: - garray - global index of `B` columns
3464: Output Parameter:
3465: . mat - the matrix, with input `A` as its local diagonal matrix
3467: Level: advanced
3469: Notes:
3470: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3472: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3474: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3475: @*/
3476: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3477: {
3478: Mat_MPIAIJ *maij;
3479: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3480: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3481: const PetscScalar *oa;
3482: Mat Bnew;
3483: PetscInt m, n, N;
3484: MatType mpi_mat_type;
3486: PetscFunctionBegin;
3487: PetscCall(MatCreate(comm, mat));
3488: PetscCall(MatGetSize(A, &m, &n));
3489: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3490: 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);
3491: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3492: /* 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); */
3494: /* Get global columns of mat */
3495: PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3497: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3498: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3499: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3500: PetscCall(MatSetType(*mat, mpi_mat_type));
3502: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3503: maij = (Mat_MPIAIJ *)(*mat)->data;
3505: (*mat)->preallocated = PETSC_TRUE;
3507: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3508: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3510: /* Set A as diagonal portion of *mat */
3511: maij->A = A;
3513: nz = oi[m];
3514: for (i = 0; i < nz; i++) {
3515: col = oj[i];
3516: oj[i] = garray[col];
3517: }
3519: /* Set Bnew as off-diagonal portion of *mat */
3520: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3521: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3522: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3523: bnew = (Mat_SeqAIJ *)Bnew->data;
3524: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3525: maij->B = Bnew;
3527: 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);
3529: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3530: b->free_a = PETSC_FALSE;
3531: b->free_ij = PETSC_FALSE;
3532: PetscCall(MatDestroy(&B));
3534: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3535: bnew->free_a = PETSC_TRUE;
3536: bnew->free_ij = PETSC_TRUE;
3538: /* condense columns of maij->B */
3539: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3540: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3541: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3542: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3543: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3544: PetscFunctionReturn(PETSC_SUCCESS);
3545: }
3547: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3549: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3550: {
3551: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3552: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3553: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3554: Mat M, Msub, B = a->B;
3555: MatScalar *aa;
3556: Mat_SeqAIJ *aij;
3557: PetscInt *garray = a->garray, *colsub, Ncols;
3558: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3559: IS iscol_sub, iscmap;
3560: const PetscInt *is_idx, *cmap;
3561: PetscBool allcolumns = PETSC_FALSE;
3562: MPI_Comm comm;
3564: PetscFunctionBegin;
3565: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3566: if (call == MAT_REUSE_MATRIX) {
3567: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3568: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3569: PetscCall(ISGetLocalSize(iscol_sub, &count));
3571: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3572: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3574: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3575: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3577: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3579: } else { /* call == MAT_INITIAL_MATRIX) */
3580: PetscBool flg;
3582: PetscCall(ISGetLocalSize(iscol, &n));
3583: PetscCall(ISGetSize(iscol, &Ncols));
3585: /* (1) iscol -> nonscalable iscol_local */
3586: /* Check for special case: each processor gets entire matrix columns */
3587: PetscCall(ISIdentity(iscol_local, &flg));
3588: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3589: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3590: if (allcolumns) {
3591: iscol_sub = iscol_local;
3592: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3593: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3595: } else {
3596: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3597: PetscInt *idx, *cmap1, k;
3598: PetscCall(PetscMalloc1(Ncols, &idx));
3599: PetscCall(PetscMalloc1(Ncols, &cmap1));
3600: PetscCall(ISGetIndices(iscol_local, &is_idx));
3601: count = 0;
3602: k = 0;
3603: for (i = 0; i < Ncols; i++) {
3604: j = is_idx[i];
3605: if (j >= cstart && j < cend) {
3606: /* diagonal part of mat */
3607: idx[count] = j;
3608: cmap1[count++] = i; /* column index in submat */
3609: } else if (Bn) {
3610: /* off-diagonal part of mat */
3611: if (j == garray[k]) {
3612: idx[count] = j;
3613: cmap1[count++] = i; /* column index in submat */
3614: } else if (j > garray[k]) {
3615: while (j > garray[k] && k < Bn - 1) k++;
3616: if (j == garray[k]) {
3617: idx[count] = j;
3618: cmap1[count++] = i; /* column index in submat */
3619: }
3620: }
3621: }
3622: }
3623: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3625: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3626: PetscCall(ISGetBlockSize(iscol, &cbs));
3627: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3629: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3630: }
3632: /* (3) Create sequential Msub */
3633: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3634: }
3636: PetscCall(ISGetLocalSize(iscol_sub, &count));
3637: aij = (Mat_SeqAIJ *)(Msub)->data;
3638: ii = aij->i;
3639: PetscCall(ISGetIndices(iscmap, &cmap));
3641: /*
3642: m - number of local rows
3643: Ncols - number of columns (same on all processors)
3644: rstart - first row in new global matrix generated
3645: */
3646: PetscCall(MatGetSize(Msub, &m, NULL));
3648: if (call == MAT_INITIAL_MATRIX) {
3649: /* (4) Create parallel newmat */
3650: PetscMPIInt rank, size;
3651: PetscInt csize;
3653: PetscCallMPI(MPI_Comm_size(comm, &size));
3654: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3656: /*
3657: Determine the number of non-zeros in the diagonal and off-diagonal
3658: portions of the matrix in order to do correct preallocation
3659: */
3661: /* first get start and end of "diagonal" columns */
3662: PetscCall(ISGetLocalSize(iscol, &csize));
3663: if (csize == PETSC_DECIDE) {
3664: PetscCall(ISGetSize(isrow, &mglobal));
3665: if (mglobal == Ncols) { /* square matrix */
3666: nlocal = m;
3667: } else {
3668: nlocal = Ncols / size + ((Ncols % size) > rank);
3669: }
3670: } else {
3671: nlocal = csize;
3672: }
3673: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3674: rstart = rend - nlocal;
3675: 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);
3677: /* next, compute all the lengths */
3678: jj = aij->j;
3679: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3680: olens = dlens + m;
3681: for (i = 0; i < m; i++) {
3682: jend = ii[i + 1] - ii[i];
3683: olen = 0;
3684: dlen = 0;
3685: for (j = 0; j < jend; j++) {
3686: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3687: else dlen++;
3688: jj++;
3689: }
3690: olens[i] = olen;
3691: dlens[i] = dlen;
3692: }
3694: PetscCall(ISGetBlockSize(isrow, &bs));
3695: PetscCall(ISGetBlockSize(iscol, &cbs));
3697: PetscCall(MatCreate(comm, &M));
3698: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3699: PetscCall(MatSetBlockSizes(M, bs, cbs));
3700: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3701: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3702: PetscCall(PetscFree(dlens));
3704: } else { /* call == MAT_REUSE_MATRIX */
3705: M = *newmat;
3706: PetscCall(MatGetLocalSize(M, &i, NULL));
3707: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3708: PetscCall(MatZeroEntries(M));
3709: /*
3710: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3711: rather than the slower MatSetValues().
3712: */
3713: M->was_assembled = PETSC_TRUE;
3714: M->assembled = PETSC_FALSE;
3715: }
3717: /* (5) Set values of Msub to *newmat */
3718: PetscCall(PetscMalloc1(count, &colsub));
3719: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3721: jj = aij->j;
3722: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3723: for (i = 0; i < m; i++) {
3724: row = rstart + i;
3725: nz = ii[i + 1] - ii[i];
3726: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3727: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3728: jj += nz;
3729: aa += nz;
3730: }
3731: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3732: PetscCall(ISRestoreIndices(iscmap, &cmap));
3734: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3735: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3737: PetscCall(PetscFree(colsub));
3739: /* save Msub, iscol_sub and iscmap used in processor for next request */
3740: if (call == MAT_INITIAL_MATRIX) {
3741: *newmat = M;
3742: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3743: PetscCall(MatDestroy(&Msub));
3745: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3746: PetscCall(ISDestroy(&iscol_sub));
3748: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3749: PetscCall(ISDestroy(&iscmap));
3751: if (iscol_local) {
3752: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3753: PetscCall(ISDestroy(&iscol_local));
3754: }
3755: }
3756: PetscFunctionReturn(PETSC_SUCCESS);
3757: }
3759: /*
3760: Not great since it makes two copies of the submatrix, first an SeqAIJ
3761: in local and then by concatenating the local matrices the end result.
3762: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3764: This requires a sequential iscol with all indices.
3765: */
3766: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3767: {
3768: PetscMPIInt rank, size;
3769: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3770: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3771: Mat M, Mreuse;
3772: MatScalar *aa, *vwork;
3773: MPI_Comm comm;
3774: Mat_SeqAIJ *aij;
3775: PetscBool colflag, allcolumns = PETSC_FALSE;
3777: PetscFunctionBegin;
3778: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3779: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3780: PetscCallMPI(MPI_Comm_size(comm, &size));
3782: /* Check for special case: each processor gets entire matrix columns */
3783: PetscCall(ISIdentity(iscol, &colflag));
3784: PetscCall(ISGetLocalSize(iscol, &n));
3785: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3786: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3788: if (call == MAT_REUSE_MATRIX) {
3789: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3790: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3791: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3792: } else {
3793: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3794: }
3796: /*
3797: m - number of local rows
3798: n - number of columns (same on all processors)
3799: rstart - first row in new global matrix generated
3800: */
3801: PetscCall(MatGetSize(Mreuse, &m, &n));
3802: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3803: if (call == MAT_INITIAL_MATRIX) {
3804: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3805: ii = aij->i;
3806: jj = aij->j;
3808: /*
3809: Determine the number of non-zeros in the diagonal and off-diagonal
3810: portions of the matrix in order to do correct preallocation
3811: */
3813: /* first get start and end of "diagonal" columns */
3814: if (csize == PETSC_DECIDE) {
3815: PetscCall(ISGetSize(isrow, &mglobal));
3816: if (mglobal == n) { /* square matrix */
3817: nlocal = m;
3818: } else {
3819: nlocal = n / size + ((n % size) > rank);
3820: }
3821: } else {
3822: nlocal = csize;
3823: }
3824: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3825: rstart = rend - nlocal;
3826: 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);
3828: /* next, compute all the lengths */
3829: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3830: olens = dlens + m;
3831: for (i = 0; i < m; i++) {
3832: jend = ii[i + 1] - ii[i];
3833: olen = 0;
3834: dlen = 0;
3835: for (j = 0; j < jend; j++) {
3836: if (*jj < rstart || *jj >= rend) olen++;
3837: else dlen++;
3838: jj++;
3839: }
3840: olens[i] = olen;
3841: dlens[i] = dlen;
3842: }
3843: PetscCall(MatCreate(comm, &M));
3844: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3845: PetscCall(MatSetBlockSizes(M, bs, cbs));
3846: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3847: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3848: PetscCall(PetscFree(dlens));
3849: } else {
3850: PetscInt ml, nl;
3852: M = *newmat;
3853: PetscCall(MatGetLocalSize(M, &ml, &nl));
3854: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3855: PetscCall(MatZeroEntries(M));
3856: /*
3857: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3858: rather than the slower MatSetValues().
3859: */
3860: M->was_assembled = PETSC_TRUE;
3861: M->assembled = PETSC_FALSE;
3862: }
3863: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3864: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3865: ii = aij->i;
3866: jj = aij->j;
3868: /* trigger copy to CPU if needed */
3869: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3870: for (i = 0; i < m; i++) {
3871: row = rstart + i;
3872: nz = ii[i + 1] - ii[i];
3873: cwork = jj;
3874: jj = PetscSafePointerPlusOffset(jj, nz);
3875: vwork = aa;
3876: aa = PetscSafePointerPlusOffset(aa, nz);
3877: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3878: }
3879: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3881: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3882: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3883: *newmat = M;
3885: /* save submatrix used in processor for next request */
3886: if (call == MAT_INITIAL_MATRIX) {
3887: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3888: PetscCall(MatDestroy(&Mreuse));
3889: }
3890: PetscFunctionReturn(PETSC_SUCCESS);
3891: }
3893: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3894: {
3895: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3896: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii;
3897: const PetscInt *JJ;
3898: PetscBool nooffprocentries;
3899: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3901: PetscFunctionBegin;
3902: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
3904: PetscCall(PetscLayoutSetUp(B->rmap));
3905: PetscCall(PetscLayoutSetUp(B->cmap));
3906: m = B->rmap->n;
3907: cstart = B->cmap->rstart;
3908: cend = B->cmap->rend;
3909: rstart = B->rmap->rstart;
3911: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3913: if (PetscDefined(USE_DEBUG)) {
3914: for (i = 0; i < m; i++) {
3915: nnz = Ii[i + 1] - Ii[i];
3916: JJ = PetscSafePointerPlusOffset(J, Ii[i]);
3917: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3918: 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]);
3919: 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);
3920: }
3921: }
3923: for (i = 0; i < m; i++) {
3924: nnz = Ii[i + 1] - Ii[i];
3925: JJ = PetscSafePointerPlusOffset(J, Ii[i]);
3926: nnz_max = PetscMax(nnz_max, nnz);
3927: d = 0;
3928: for (j = 0; j < nnz; j++) {
3929: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3930: }
3931: d_nnz[i] = d;
3932: o_nnz[i] = nnz - d;
3933: }
3934: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3935: PetscCall(PetscFree2(d_nnz, o_nnz));
3937: for (i = 0; i < m; i++) {
3938: ii = i + rstart;
3939: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i]), PetscSafePointerPlusOffset(v, Ii[i]), INSERT_VALUES));
3940: }
3941: nooffprocentries = B->nooffprocentries;
3942: B->nooffprocentries = PETSC_TRUE;
3943: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3944: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3945: B->nooffprocentries = nooffprocentries;
3947: /* count number of entries below block diagonal */
3948: PetscCall(PetscFree(Aij->ld));
3949: PetscCall(PetscCalloc1(m, &ld));
3950: Aij->ld = ld;
3951: for (i = 0; i < m; i++) {
3952: nnz = Ii[i + 1] - Ii[i];
3953: j = 0;
3954: while (j < nnz && J[j] < cstart) j++;
3955: ld[i] = j;
3956: if (J) J += nnz;
3957: }
3959: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3960: PetscFunctionReturn(PETSC_SUCCESS);
3961: }
3963: /*@
3964: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3965: (the default parallel PETSc format).
3967: Collective
3969: Input Parameters:
3970: + B - the matrix
3971: . i - the indices into `j` for the start of each local row (indices start with zero)
3972: . j - the column indices for each local row (indices start with zero)
3973: - v - optional values in the matrix
3975: Level: developer
3977: Notes:
3978: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3979: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3980: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3982: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3984: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3986: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3988: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3989: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3991: The format which is used for the sparse matrix input, is equivalent to a
3992: row-major ordering.. i.e for the following matrix, the input data expected is
3993: as shown
3994: .vb
3995: 1 0 0
3996: 2 0 3 P0
3997: -------
3998: 4 5 6 P1
4000: Process0 [P0] rows_owned=[0,1]
4001: i = {0,1,3} [size = nrow+1 = 2+1]
4002: j = {0,0,2} [size = 3]
4003: v = {1,2,3} [size = 3]
4005: Process1 [P1] rows_owned=[2]
4006: i = {0,3} [size = nrow+1 = 1+1]
4007: j = {0,1,2} [size = 3]
4008: v = {4,5,6} [size = 3]
4009: .ve
4011: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4012: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4013: @*/
4014: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4015: {
4016: PetscFunctionBegin;
4017: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4018: PetscFunctionReturn(PETSC_SUCCESS);
4019: }
4021: /*@C
4022: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4023: (the default parallel PETSc format). For good matrix assembly performance
4024: the user should preallocate the matrix storage by setting the parameters
4025: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4027: Collective
4029: Input Parameters:
4030: + B - the matrix
4031: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4032: (same value is used for all local rows)
4033: . d_nnz - array containing the number of nonzeros in the various rows of the
4034: DIAGONAL portion of the local submatrix (possibly different for each row)
4035: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4036: The size of this array is equal to the number of local rows, i.e 'm'.
4037: For matrices that will be factored, you must leave room for (and set)
4038: the diagonal entry even if it is zero.
4039: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4040: submatrix (same value is used for all local rows).
4041: - o_nnz - array containing the number of nonzeros in the various rows of the
4042: OFF-DIAGONAL portion of the local submatrix (possibly different for
4043: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4044: structure. The size of this array is equal to the number
4045: of local rows, i.e 'm'.
4047: Example Usage:
4048: Consider the following 8x8 matrix with 34 non-zero values, that is
4049: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4050: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4051: as follows
4053: .vb
4054: 1 2 0 | 0 3 0 | 0 4
4055: Proc0 0 5 6 | 7 0 0 | 8 0
4056: 9 0 10 | 11 0 0 | 12 0
4057: -------------------------------------
4058: 13 0 14 | 15 16 17 | 0 0
4059: Proc1 0 18 0 | 19 20 21 | 0 0
4060: 0 0 0 | 22 23 0 | 24 0
4061: -------------------------------------
4062: Proc2 25 26 27 | 0 0 28 | 29 0
4063: 30 0 0 | 31 32 33 | 0 34
4064: .ve
4066: This can be represented as a collection of submatrices as
4067: .vb
4068: A B C
4069: D E F
4070: G H I
4071: .ve
4073: Where the submatrices A,B,C are owned by proc0, D,E,F are
4074: owned by proc1, G,H,I are owned by proc2.
4076: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4077: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4078: The 'M','N' parameters are 8,8, and have the same values on all procs.
4080: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4081: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4082: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4083: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4084: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4085: matrix, ans [DF] as another `MATSEQAIJ` matrix.
4087: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4088: allocated for every row of the local diagonal submatrix, and `o_nz`
4089: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4090: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4091: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4092: In this case, the values of `d_nz`, `o_nz` are
4093: .vb
4094: proc0 dnz = 2, o_nz = 2
4095: proc1 dnz = 3, o_nz = 2
4096: proc2 dnz = 1, o_nz = 4
4097: .ve
4098: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4099: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4100: for proc3. i.e we are using 12+15+10=37 storage locations to store
4101: 34 values.
4103: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4104: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4105: In the above case the values for `d_nnz`, `o_nnz` are
4106: .vb
4107: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4108: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4109: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4110: .ve
4111: Here the space allocated is sum of all the above values i.e 34, and
4112: hence pre-allocation is perfect.
4114: Level: intermediate
4116: Notes:
4117: If the *_nnz parameter is given then the *_nz parameter is ignored
4119: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4120: storage. The stored row and column indices begin with zero.
4121: See [Sparse Matrices](sec_matsparse) for details.
4123: The parallel matrix is partitioned such that the first m0 rows belong to
4124: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4125: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4127: The DIAGONAL portion of the local submatrix of a processor can be defined
4128: as the submatrix which is obtained by extraction the part corresponding to
4129: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4130: first row that belongs to the processor, r2 is the last row belonging to
4131: the this processor, and c1-c2 is range of indices of the local part of a
4132: vector suitable for applying the matrix to. This is an mxn matrix. In the
4133: common case of a square matrix, the row and column ranges are the same and
4134: the DIAGONAL part is also square. The remaining portion of the local
4135: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4137: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4139: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4140: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4141: You can also run with the option `-info` and look for messages with the string
4142: malloc in them to see if additional memory allocation was needed.
4144: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4145: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4146: @*/
4147: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4148: {
4149: PetscFunctionBegin;
4152: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4153: PetscFunctionReturn(PETSC_SUCCESS);
4154: }
4156: /*@
4157: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4158: CSR format for the local rows.
4160: Collective
4162: Input Parameters:
4163: + comm - MPI communicator
4164: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4165: . n - This value should be the same as the local size used in creating the
4166: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4167: calculated if `N` is given) For square matrices n is almost always `m`.
4168: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4169: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4170: . 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
4171: . j - global column indices
4172: - a - optional matrix values
4174: Output Parameter:
4175: . mat - the matrix
4177: Level: intermediate
4179: Notes:
4180: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4181: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4182: called this routine. Use `MatCreateMPIAIJWithSplitArray()` to avoid needing to copy the arrays.
4184: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4186: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4188: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4189: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4191: The format which is used for the sparse matrix input, is equivalent to a
4192: row-major ordering.. i.e for the following matrix, the input data expected is
4193: as shown
4194: .vb
4195: 1 0 0
4196: 2 0 3 P0
4197: -------
4198: 4 5 6 P1
4200: Process0 [P0] rows_owned=[0,1]
4201: i = {0,1,3} [size = nrow+1 = 2+1]
4202: j = {0,0,2} [size = 3]
4203: v = {1,2,3} [size = 3]
4205: Process1 [P1] rows_owned=[2]
4206: i = {0,3} [size = nrow+1 = 1+1]
4207: j = {0,1,2} [size = 3]
4208: v = {4,5,6} [size = 3]
4209: .ve
4211: .seealso: [](ch_matrices), `Mat`, `MATMPIAIK`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4212: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4213: @*/
4214: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4215: {
4216: PetscFunctionBegin;
4217: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4218: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4219: PetscCall(MatCreate(comm, mat));
4220: PetscCall(MatSetSizes(*mat, m, n, M, N));
4221: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4222: PetscCall(MatSetType(*mat, MATMPIAIJ));
4223: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4224: PetscFunctionReturn(PETSC_SUCCESS);
4225: }
4227: /*@
4228: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4229: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4230: from `MatCreateMPIAIJWithArrays()`
4232: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4234: Collective
4236: Input Parameters:
4237: + mat - the matrix
4238: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4239: . n - This value should be the same as the local size used in creating the
4240: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4241: calculated if N is given) For square matrices n is almost always m.
4242: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4243: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4244: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4245: . J - column indices
4246: - v - matrix values
4248: Level: deprecated
4250: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4251: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4252: @*/
4253: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4254: {
4255: PetscInt nnz, i;
4256: PetscBool nooffprocentries;
4257: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4258: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4259: PetscScalar *ad, *ao;
4260: PetscInt ldi, Iii, md;
4261: const PetscInt *Adi = Ad->i;
4262: PetscInt *ld = Aij->ld;
4264: PetscFunctionBegin;
4265: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4266: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4267: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4268: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4270: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4271: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4273: for (i = 0; i < m; i++) {
4274: if (PetscDefined(USE_DEBUG)) {
4275: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4276: 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);
4277: 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);
4278: }
4279: }
4280: nnz = Ii[i + 1] - Ii[i];
4281: Iii = Ii[i];
4282: ldi = ld[i];
4283: md = Adi[i + 1] - Adi[i];
4284: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4285: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4286: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4287: ad += md;
4288: ao += nnz - md;
4289: }
4290: nooffprocentries = mat->nooffprocentries;
4291: mat->nooffprocentries = PETSC_TRUE;
4292: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4293: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4294: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4295: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4296: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4297: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4298: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4299: mat->nooffprocentries = nooffprocentries;
4300: PetscFunctionReturn(PETSC_SUCCESS);
4301: }
4303: /*@
4304: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4306: Collective
4308: Input Parameters:
4309: + mat - the matrix
4310: - v - matrix values, stored by row
4312: Level: intermediate
4314: Notes:
4315: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4317: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4319: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4320: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4321: @*/
4322: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4323: {
4324: PetscInt nnz, i, m;
4325: PetscBool nooffprocentries;
4326: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4327: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4328: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4329: PetscScalar *ad, *ao;
4330: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4331: PetscInt ldi, Iii, md;
4332: PetscInt *ld = Aij->ld;
4334: PetscFunctionBegin;
4335: m = mat->rmap->n;
4337: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4338: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4339: Iii = 0;
4340: for (i = 0; i < m; i++) {
4341: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4342: ldi = ld[i];
4343: md = Adi[i + 1] - Adi[i];
4344: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4345: ad += md;
4346: if (ao) {
4347: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4348: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4349: ao += nnz - md;
4350: }
4351: Iii += nnz;
4352: }
4353: nooffprocentries = mat->nooffprocentries;
4354: mat->nooffprocentries = PETSC_TRUE;
4355: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4356: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4357: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4358: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4359: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4360: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4361: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4362: mat->nooffprocentries = nooffprocentries;
4363: PetscFunctionReturn(PETSC_SUCCESS);
4364: }
4366: /*@C
4367: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4368: (the default parallel PETSc format). For good matrix assembly performance
4369: the user should preallocate the matrix storage by setting the parameters
4370: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4372: Collective
4374: Input Parameters:
4375: + comm - MPI communicator
4376: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4377: This value should be the same as the local size used in creating the
4378: y vector for the matrix-vector product y = Ax.
4379: . n - This value should be the same as the local size used in creating the
4380: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4381: calculated if N is given) For square matrices n is almost always m.
4382: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4383: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4384: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4385: (same value is used for all local rows)
4386: . d_nnz - array containing the number of nonzeros in the various rows of the
4387: DIAGONAL portion of the local submatrix (possibly different for each row)
4388: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4389: The size of this array is equal to the number of local rows, i.e 'm'.
4390: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4391: submatrix (same value is used for all local rows).
4392: - o_nnz - array containing the number of nonzeros in the various rows of the
4393: OFF-DIAGONAL portion of the local submatrix (possibly different for
4394: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4395: structure. The size of this array is equal to the number
4396: of local rows, i.e 'm'.
4398: Output Parameter:
4399: . A - the matrix
4401: Options Database Keys:
4402: + -mat_no_inode - Do not use inodes
4403: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4404: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4405: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the vecscatter be viewed as a matrix.
4406: Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.
4408: Level: intermediate
4410: Notes:
4411: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4412: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4413: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4415: If the *_nnz parameter is given then the *_nz parameter is ignored
4417: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4418: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4419: storage requirements for this matrix.
4421: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4422: processor than it must be used on all processors that share the object for
4423: that argument.
4425: The user MUST specify either the local or global matrix dimensions
4426: (possibly both).
4428: The parallel matrix is partitioned across processors such that the
4429: first m0 rows belong to process 0, the next m1 rows belong to
4430: process 1, the next m2 rows belong to process 2 etc.. where
4431: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4432: values corresponding to [m x N] submatrix.
4434: The columns are logically partitioned with the n0 columns belonging
4435: to 0th partition, the next n1 columns belonging to the next
4436: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4438: The DIAGONAL portion of the local submatrix on any given processor
4439: is the submatrix corresponding to the rows and columns m,n
4440: corresponding to the given processor. i.e diagonal matrix on
4441: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4442: etc. The remaining portion of the local submatrix [m x (N-n)]
4443: constitute the OFF-DIAGONAL portion. The example below better
4444: illustrates this concept.
4446: For a square global matrix we define each processor's diagonal portion
4447: to be its local rows and the corresponding columns (a square submatrix);
4448: each processor's off-diagonal portion encompasses the remainder of the
4449: local matrix (a rectangular submatrix).
4451: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4453: When calling this routine with a single process communicator, a matrix of
4454: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4455: type of communicator, use the construction mechanism
4456: .vb
4457: MatCreate(..., &A);
4458: MatSetType(A, MATMPIAIJ);
4459: MatSetSizes(A, m, n, M, N);
4460: MatMPIAIJSetPreallocation(A, ...);
4461: .ve
4463: By default, this format uses inodes (identical nodes) when possible.
4464: We search for consecutive rows with the same nonzero structure, thereby
4465: reusing matrix information to achieve increased efficiency.
4467: Example Usage:
4468: Consider the following 8x8 matrix with 34 non-zero values, that is
4469: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4470: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4471: as follows
4473: .vb
4474: 1 2 0 | 0 3 0 | 0 4
4475: Proc0 0 5 6 | 7 0 0 | 8 0
4476: 9 0 10 | 11 0 0 | 12 0
4477: -------------------------------------
4478: 13 0 14 | 15 16 17 | 0 0
4479: Proc1 0 18 0 | 19 20 21 | 0 0
4480: 0 0 0 | 22 23 0 | 24 0
4481: -------------------------------------
4482: Proc2 25 26 27 | 0 0 28 | 29 0
4483: 30 0 0 | 31 32 33 | 0 34
4484: .ve
4486: This can be represented as a collection of submatrices as
4488: .vb
4489: A B C
4490: D E F
4491: G H I
4492: .ve
4494: Where the submatrices A,B,C are owned by proc0, D,E,F are
4495: owned by proc1, G,H,I are owned by proc2.
4497: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4498: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4499: The 'M','N' parameters are 8,8, and have the same values on all procs.
4501: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4502: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4503: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4504: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4505: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4506: matrix, ans [DF] as another SeqAIJ matrix.
4508: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4509: allocated for every row of the local diagonal submatrix, and `o_nz`
4510: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4511: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4512: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4513: In this case, the values of `d_nz`,`o_nz` are
4514: .vb
4515: proc0 dnz = 2, o_nz = 2
4516: proc1 dnz = 3, o_nz = 2
4517: proc2 dnz = 1, o_nz = 4
4518: .ve
4519: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4520: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4521: for proc3. i.e we are using 12+15+10=37 storage locations to store
4522: 34 values.
4524: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4525: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4526: In the above case the values for d_nnz,o_nnz are
4527: .vb
4528: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4529: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4530: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4531: .ve
4532: Here the space allocated is sum of all the above values i.e 34, and
4533: hence pre-allocation is perfect.
4535: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4536: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`
4537: @*/
4538: 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)
4539: {
4540: PetscMPIInt size;
4542: PetscFunctionBegin;
4543: PetscCall(MatCreate(comm, A));
4544: PetscCall(MatSetSizes(*A, m, n, M, N));
4545: PetscCallMPI(MPI_Comm_size(comm, &size));
4546: if (size > 1) {
4547: PetscCall(MatSetType(*A, MATMPIAIJ));
4548: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4549: } else {
4550: PetscCall(MatSetType(*A, MATSEQAIJ));
4551: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4552: }
4553: PetscFunctionReturn(PETSC_SUCCESS);
4554: }
4556: /*MC
4557: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4559: Synopsis:
4560: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4562: Not Collective
4564: Input Parameter:
4565: . A - the `MATMPIAIJ` matrix
4567: Output Parameters:
4568: + Ad - the diagonal portion of the matrix
4569: . Ao - the off-diagonal portion of the matrix
4570: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4571: - ierr - error code
4573: Level: advanced
4575: Note:
4576: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4578: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4579: M*/
4581: /*MC
4582: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4584: Synopsis:
4585: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4587: Not Collective
4589: Input Parameters:
4590: + A - the `MATMPIAIJ` matrix
4591: . Ad - the diagonal portion of the matrix
4592: . Ao - the off-diagonal portion of the matrix
4593: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4594: - ierr - error code
4596: Level: advanced
4598: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4599: M*/
4601: /*@C
4602: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4604: Not Collective
4606: Input Parameter:
4607: . A - The `MATMPIAIJ` matrix
4609: Output Parameters:
4610: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4611: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4612: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4614: Level: intermediate
4616: Note:
4617: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4618: 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
4619: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4620: local column numbers to global column numbers in the original matrix.
4622: Fortran Notes:
4623: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4625: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4626: @*/
4627: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4628: {
4629: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4630: PetscBool flg;
4632: PetscFunctionBegin;
4633: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4634: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4635: if (Ad) *Ad = a->A;
4636: if (Ao) *Ao = a->B;
4637: if (colmap) *colmap = a->garray;
4638: PetscFunctionReturn(PETSC_SUCCESS);
4639: }
4641: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4642: {
4643: PetscInt m, N, i, rstart, nnz, Ii;
4644: PetscInt *indx;
4645: PetscScalar *values;
4646: MatType rootType;
4648: PetscFunctionBegin;
4649: PetscCall(MatGetSize(inmat, &m, &N));
4650: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4651: PetscInt *dnz, *onz, sum, bs, cbs;
4653: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4654: /* Check sum(n) = N */
4655: PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4656: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4658: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4659: rstart -= m;
4661: MatPreallocateBegin(comm, m, n, dnz, onz);
4662: for (i = 0; i < m; i++) {
4663: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4664: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4665: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4666: }
4668: PetscCall(MatCreate(comm, outmat));
4669: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4670: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4671: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4672: PetscCall(MatGetRootType_Private(inmat, &rootType));
4673: PetscCall(MatSetType(*outmat, rootType));
4674: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4675: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4676: MatPreallocateEnd(dnz, onz);
4677: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4678: }
4680: /* numeric phase */
4681: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4682: for (i = 0; i < m; i++) {
4683: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4684: Ii = i + rstart;
4685: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4686: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4687: }
4688: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4689: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4690: PetscFunctionReturn(PETSC_SUCCESS);
4691: }
4693: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4694: {
4695: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4697: PetscFunctionBegin;
4698: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4699: PetscCall(PetscFree(merge->id_r));
4700: PetscCall(PetscFree(merge->len_s));
4701: PetscCall(PetscFree(merge->len_r));
4702: PetscCall(PetscFree(merge->bi));
4703: PetscCall(PetscFree(merge->bj));
4704: PetscCall(PetscFree(merge->buf_ri[0]));
4705: PetscCall(PetscFree(merge->buf_ri));
4706: PetscCall(PetscFree(merge->buf_rj[0]));
4707: PetscCall(PetscFree(merge->buf_rj));
4708: PetscCall(PetscFree(merge->coi));
4709: PetscCall(PetscFree(merge->coj));
4710: PetscCall(PetscFree(merge->owners_co));
4711: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4712: PetscCall(PetscFree(merge));
4713: PetscFunctionReturn(PETSC_SUCCESS);
4714: }
4716: #include <../src/mat/utils/freespace.h>
4717: #include <petscbt.h>
4719: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4720: {
4721: MPI_Comm comm;
4722: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4723: PetscMPIInt size, rank, taga, *len_s;
4724: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4725: PetscInt proc, m;
4726: PetscInt **buf_ri, **buf_rj;
4727: PetscInt k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4728: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4729: MPI_Request *s_waits, *r_waits;
4730: MPI_Status *status;
4731: const MatScalar *aa, *a_a;
4732: MatScalar **abuf_r, *ba_i;
4733: Mat_Merge_SeqsToMPI *merge;
4734: PetscContainer container;
4736: PetscFunctionBegin;
4737: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4738: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4740: PetscCallMPI(MPI_Comm_size(comm, &size));
4741: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4743: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4744: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4745: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4746: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4747: aa = a_a;
4749: bi = merge->bi;
4750: bj = merge->bj;
4751: buf_ri = merge->buf_ri;
4752: buf_rj = merge->buf_rj;
4754: PetscCall(PetscMalloc1(size, &status));
4755: owners = merge->rowmap->range;
4756: len_s = merge->len_s;
4758: /* send and recv matrix values */
4759: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4760: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4762: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4763: for (proc = 0, k = 0; proc < size; proc++) {
4764: if (!len_s[proc]) continue;
4765: i = owners[proc];
4766: PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4767: k++;
4768: }
4770: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4771: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4772: PetscCall(PetscFree(status));
4774: PetscCall(PetscFree(s_waits));
4775: PetscCall(PetscFree(r_waits));
4777: /* insert mat values of mpimat */
4778: PetscCall(PetscMalloc1(N, &ba_i));
4779: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4781: for (k = 0; k < merge->nrecv; k++) {
4782: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4783: nrows = *buf_ri_k[k];
4784: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4785: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4786: }
4788: /* set values of ba */
4789: m = merge->rowmap->n;
4790: for (i = 0; i < m; i++) {
4791: arow = owners[rank] + i;
4792: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4793: bnzi = bi[i + 1] - bi[i];
4794: PetscCall(PetscArrayzero(ba_i, bnzi));
4796: /* add local non-zero vals of this proc's seqmat into ba */
4797: anzi = ai[arow + 1] - ai[arow];
4798: aj = a->j + ai[arow];
4799: aa = a_a + ai[arow];
4800: nextaj = 0;
4801: for (j = 0; nextaj < anzi; j++) {
4802: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4803: ba_i[j] += aa[nextaj++];
4804: }
4805: }
4807: /* add received vals into ba */
4808: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4809: /* i-th row */
4810: if (i == *nextrow[k]) {
4811: anzi = *(nextai[k] + 1) - *nextai[k];
4812: aj = buf_rj[k] + *nextai[k];
4813: aa = abuf_r[k] + *nextai[k];
4814: nextaj = 0;
4815: for (j = 0; nextaj < anzi; j++) {
4816: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4817: ba_i[j] += aa[nextaj++];
4818: }
4819: }
4820: nextrow[k]++;
4821: nextai[k]++;
4822: }
4823: }
4824: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4825: }
4826: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4827: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4828: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4830: PetscCall(PetscFree(abuf_r[0]));
4831: PetscCall(PetscFree(abuf_r));
4832: PetscCall(PetscFree(ba_i));
4833: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4834: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4835: PetscFunctionReturn(PETSC_SUCCESS);
4836: }
4838: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4839: {
4840: Mat B_mpi;
4841: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4842: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4843: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4844: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4845: PetscInt len, proc, *dnz, *onz, bs, cbs;
4846: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4847: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4848: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4849: MPI_Status *status;
4850: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4851: PetscBT lnkbt;
4852: Mat_Merge_SeqsToMPI *merge;
4853: PetscContainer container;
4855: PetscFunctionBegin;
4856: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4858: /* make sure it is a PETSc comm */
4859: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4860: PetscCallMPI(MPI_Comm_size(comm, &size));
4861: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4863: PetscCall(PetscNew(&merge));
4864: PetscCall(PetscMalloc1(size, &status));
4866: /* determine row ownership */
4867: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4868: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4869: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4870: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4871: PetscCall(PetscLayoutSetUp(merge->rowmap));
4872: PetscCall(PetscMalloc1(size, &len_si));
4873: PetscCall(PetscMalloc1(size, &merge->len_s));
4875: m = merge->rowmap->n;
4876: owners = merge->rowmap->range;
4878: /* determine the number of messages to send, their lengths */
4879: len_s = merge->len_s;
4881: len = 0; /* length of buf_si[] */
4882: merge->nsend = 0;
4883: for (proc = 0; proc < size; proc++) {
4884: len_si[proc] = 0;
4885: if (proc == rank) {
4886: len_s[proc] = 0;
4887: } else {
4888: len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4889: len_s[proc] = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4890: }
4891: if (len_s[proc]) {
4892: merge->nsend++;
4893: nrows = 0;
4894: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4895: if (ai[i + 1] > ai[i]) nrows++;
4896: }
4897: len_si[proc] = 2 * (nrows + 1);
4898: len += len_si[proc];
4899: }
4900: }
4902: /* determine the number and length of messages to receive for ij-structure */
4903: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4904: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4906: /* post the Irecv of j-structure */
4907: PetscCall(PetscCommGetNewTag(comm, &tagj));
4908: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4910: /* post the Isend of j-structure */
4911: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4913: for (proc = 0, k = 0; proc < size; proc++) {
4914: if (!len_s[proc]) continue;
4915: i = owners[proc];
4916: PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4917: k++;
4918: }
4920: /* receives and sends of j-structure are complete */
4921: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4922: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4924: /* send and recv i-structure */
4925: PetscCall(PetscCommGetNewTag(comm, &tagi));
4926: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4928: PetscCall(PetscMalloc1(len + 1, &buf_s));
4929: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4930: for (proc = 0, k = 0; proc < size; proc++) {
4931: if (!len_s[proc]) continue;
4932: /* form outgoing message for i-structure:
4933: buf_si[0]: nrows to be sent
4934: [1:nrows]: row index (global)
4935: [nrows+1:2*nrows+1]: i-structure index
4936: */
4937: nrows = len_si[proc] / 2 - 1;
4938: buf_si_i = buf_si + nrows + 1;
4939: buf_si[0] = nrows;
4940: buf_si_i[0] = 0;
4941: nrows = 0;
4942: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4943: anzi = ai[i + 1] - ai[i];
4944: if (anzi) {
4945: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4946: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4947: nrows++;
4948: }
4949: }
4950: PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4951: k++;
4952: buf_si += len_si[proc];
4953: }
4955: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4956: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4958: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4959: 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]));
4961: PetscCall(PetscFree(len_si));
4962: PetscCall(PetscFree(len_ri));
4963: PetscCall(PetscFree(rj_waits));
4964: PetscCall(PetscFree2(si_waits, sj_waits));
4965: PetscCall(PetscFree(ri_waits));
4966: PetscCall(PetscFree(buf_s));
4967: PetscCall(PetscFree(status));
4969: /* compute a local seq matrix in each processor */
4970: /* allocate bi array and free space for accumulating nonzero column info */
4971: PetscCall(PetscMalloc1(m + 1, &bi));
4972: bi[0] = 0;
4974: /* create and initialize a linked list */
4975: nlnk = N + 1;
4976: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4978: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4979: len = ai[owners[rank + 1]] - ai[owners[rank]];
4980: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4982: current_space = free_space;
4984: /* determine symbolic info for each local row */
4985: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4987: for (k = 0; k < merge->nrecv; k++) {
4988: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4989: nrows = *buf_ri_k[k];
4990: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4991: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4992: }
4994: MatPreallocateBegin(comm, m, n, dnz, onz);
4995: len = 0;
4996: for (i = 0; i < m; i++) {
4997: bnzi = 0;
4998: /* add local non-zero cols of this proc's seqmat into lnk */
4999: arow = owners[rank] + i;
5000: anzi = ai[arow + 1] - ai[arow];
5001: aj = a->j + ai[arow];
5002: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5003: bnzi += nlnk;
5004: /* add received col data into lnk */
5005: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5006: if (i == *nextrow[k]) { /* i-th row */
5007: anzi = *(nextai[k] + 1) - *nextai[k];
5008: aj = buf_rj[k] + *nextai[k];
5009: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5010: bnzi += nlnk;
5011: nextrow[k]++;
5012: nextai[k]++;
5013: }
5014: }
5015: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5017: /* if free space is not available, make more free space */
5018: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5019: /* copy data into free space, then initialize lnk */
5020: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5021: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5023: current_space->array += bnzi;
5024: current_space->local_used += bnzi;
5025: current_space->local_remaining -= bnzi;
5027: bi[i + 1] = bi[i] + bnzi;
5028: }
5030: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5032: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5033: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5034: PetscCall(PetscLLDestroy(lnk, lnkbt));
5036: /* create symbolic parallel matrix B_mpi */
5037: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5038: PetscCall(MatCreate(comm, &B_mpi));
5039: if (n == PETSC_DECIDE) {
5040: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5041: } else {
5042: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5043: }
5044: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5045: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5046: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5047: MatPreallocateEnd(dnz, onz);
5048: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5050: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5051: B_mpi->assembled = PETSC_FALSE;
5052: merge->bi = bi;
5053: merge->bj = bj;
5054: merge->buf_ri = buf_ri;
5055: merge->buf_rj = buf_rj;
5056: merge->coi = NULL;
5057: merge->coj = NULL;
5058: merge->owners_co = NULL;
5060: PetscCall(PetscCommDestroy(&comm));
5062: /* attach the supporting struct to B_mpi for reuse */
5063: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5064: PetscCall(PetscContainerSetPointer(container, merge));
5065: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5066: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5067: PetscCall(PetscContainerDestroy(&container));
5068: *mpimat = B_mpi;
5070: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5071: PetscFunctionReturn(PETSC_SUCCESS);
5072: }
5074: /*@C
5075: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5076: matrices from each processor
5078: Collective
5080: Input Parameters:
5081: + comm - the communicators the parallel matrix will live on
5082: . seqmat - the input sequential matrices
5083: . m - number of local rows (or `PETSC_DECIDE`)
5084: . n - number of local columns (or `PETSC_DECIDE`)
5085: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5087: Output Parameter:
5088: . mpimat - the parallel matrix generated
5090: Level: advanced
5092: Note:
5093: The dimensions of the sequential matrix in each processor MUST be the same.
5094: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5095: destroyed when mpimat is destroyed. Call `PetscObjectQuery()` to access seqmat.
5097: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5098: @*/
5099: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5100: {
5101: PetscMPIInt size;
5103: PetscFunctionBegin;
5104: PetscCallMPI(MPI_Comm_size(comm, &size));
5105: if (size == 1) {
5106: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5107: if (scall == MAT_INITIAL_MATRIX) {
5108: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5109: } else {
5110: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5111: }
5112: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5113: PetscFunctionReturn(PETSC_SUCCESS);
5114: }
5115: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5116: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5117: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5118: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5119: PetscFunctionReturn(PETSC_SUCCESS);
5120: }
5122: /*@
5123: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5125: Not Collective
5127: Input Parameter:
5128: . A - the matrix
5130: Output Parameter:
5131: . A_loc - the local sequential matrix generated
5133: Level: developer
5135: Notes:
5136: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5137: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5138: `n` is the global column count obtained with `MatGetSize()`
5140: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5142: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5144: Destroy the matrix with `MatDestroy()`
5146: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5147: @*/
5148: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5149: {
5150: PetscBool mpi;
5152: PetscFunctionBegin;
5153: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5154: if (mpi) {
5155: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5156: } else {
5157: *A_loc = A;
5158: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5159: }
5160: PetscFunctionReturn(PETSC_SUCCESS);
5161: }
5163: /*@
5164: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5166: Not Collective
5168: Input Parameters:
5169: + A - the matrix
5170: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5172: Output Parameter:
5173: . A_loc - the local sequential matrix generated
5175: Level: developer
5177: Notes:
5178: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5179: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5180: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5182: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5184: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5185: 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
5186: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5187: 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.
5189: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5190: @*/
5191: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5192: {
5193: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5194: Mat_SeqAIJ *mat, *a, *b;
5195: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5196: const PetscScalar *aa, *ba, *aav, *bav;
5197: PetscScalar *ca, *cam;
5198: PetscMPIInt size;
5199: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5200: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5201: PetscBool match;
5203: PetscFunctionBegin;
5204: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5205: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5206: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5207: if (size == 1) {
5208: if (scall == MAT_INITIAL_MATRIX) {
5209: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5210: *A_loc = mpimat->A;
5211: } else if (scall == MAT_REUSE_MATRIX) {
5212: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5213: }
5214: PetscFunctionReturn(PETSC_SUCCESS);
5215: }
5217: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5218: a = (Mat_SeqAIJ *)mpimat->A->data;
5219: b = (Mat_SeqAIJ *)mpimat->B->data;
5220: ai = a->i;
5221: aj = a->j;
5222: bi = b->i;
5223: bj = b->j;
5224: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5225: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5226: aa = aav;
5227: ba = bav;
5228: if (scall == MAT_INITIAL_MATRIX) {
5229: PetscCall(PetscMalloc1(1 + am, &ci));
5230: ci[0] = 0;
5231: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5232: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5233: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5234: k = 0;
5235: for (i = 0; i < am; i++) {
5236: ncols_o = bi[i + 1] - bi[i];
5237: ncols_d = ai[i + 1] - ai[i];
5238: /* off-diagonal portion of A */
5239: for (jo = 0; jo < ncols_o; jo++) {
5240: col = cmap[*bj];
5241: if (col >= cstart) break;
5242: cj[k] = col;
5243: bj++;
5244: ca[k++] = *ba++;
5245: }
5246: /* diagonal portion of A */
5247: for (j = 0; j < ncols_d; j++) {
5248: cj[k] = cstart + *aj++;
5249: ca[k++] = *aa++;
5250: }
5251: /* off-diagonal portion of A */
5252: for (j = jo; j < ncols_o; j++) {
5253: cj[k] = cmap[*bj++];
5254: ca[k++] = *ba++;
5255: }
5256: }
5257: /* put together the new matrix */
5258: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5259: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5260: /* Since these are PETSc arrays, change flags to free them as necessary. */
5261: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5262: mat->free_a = PETSC_TRUE;
5263: mat->free_ij = PETSC_TRUE;
5264: mat->nonew = 0;
5265: } else if (scall == MAT_REUSE_MATRIX) {
5266: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5267: ci = mat->i;
5268: cj = mat->j;
5269: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5270: for (i = 0; i < am; i++) {
5271: /* off-diagonal portion of A */
5272: ncols_o = bi[i + 1] - bi[i];
5273: for (jo = 0; jo < ncols_o; jo++) {
5274: col = cmap[*bj];
5275: if (col >= cstart) break;
5276: *cam++ = *ba++;
5277: bj++;
5278: }
5279: /* diagonal portion of A */
5280: ncols_d = ai[i + 1] - ai[i];
5281: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5282: /* off-diagonal portion of A */
5283: for (j = jo; j < ncols_o; j++) {
5284: *cam++ = *ba++;
5285: bj++;
5286: }
5287: }
5288: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5289: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5290: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5291: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5292: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5293: PetscFunctionReturn(PETSC_SUCCESS);
5294: }
5296: /*@
5297: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5298: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5300: Not Collective
5302: Input Parameters:
5303: + A - the matrix
5304: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5306: Output Parameters:
5307: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5308: - A_loc - the local sequential matrix generated
5310: Level: developer
5312: Note:
5313: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5314: part, then those associated with the off-diagonal part (in its local ordering)
5316: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5317: @*/
5318: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5319: {
5320: Mat Ao, Ad;
5321: const PetscInt *cmap;
5322: PetscMPIInt size;
5323: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5325: PetscFunctionBegin;
5326: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5327: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5328: if (size == 1) {
5329: if (scall == MAT_INITIAL_MATRIX) {
5330: PetscCall(PetscObjectReference((PetscObject)Ad));
5331: *A_loc = Ad;
5332: } else if (scall == MAT_REUSE_MATRIX) {
5333: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5334: }
5335: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5336: PetscFunctionReturn(PETSC_SUCCESS);
5337: }
5338: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5339: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5340: if (f) {
5341: PetscCall((*f)(A, scall, glob, A_loc));
5342: } else {
5343: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5344: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5345: Mat_SeqAIJ *c;
5346: PetscInt *ai = a->i, *aj = a->j;
5347: PetscInt *bi = b->i, *bj = b->j;
5348: PetscInt *ci, *cj;
5349: const PetscScalar *aa, *ba;
5350: PetscScalar *ca;
5351: PetscInt i, j, am, dn, on;
5353: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5354: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5355: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5356: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5357: if (scall == MAT_INITIAL_MATRIX) {
5358: PetscInt k;
5359: PetscCall(PetscMalloc1(1 + am, &ci));
5360: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5361: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5362: ci[0] = 0;
5363: for (i = 0, k = 0; i < am; i++) {
5364: const PetscInt ncols_o = bi[i + 1] - bi[i];
5365: const PetscInt ncols_d = ai[i + 1] - ai[i];
5366: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5367: /* diagonal portion of A */
5368: for (j = 0; j < ncols_d; j++, k++) {
5369: cj[k] = *aj++;
5370: ca[k] = *aa++;
5371: }
5372: /* off-diagonal portion of A */
5373: for (j = 0; j < ncols_o; j++, k++) {
5374: cj[k] = dn + *bj++;
5375: ca[k] = *ba++;
5376: }
5377: }
5378: /* put together the new matrix */
5379: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5380: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5381: /* Since these are PETSc arrays, change flags to free them as necessary. */
5382: c = (Mat_SeqAIJ *)(*A_loc)->data;
5383: c->free_a = PETSC_TRUE;
5384: c->free_ij = PETSC_TRUE;
5385: c->nonew = 0;
5386: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5387: } else if (scall == MAT_REUSE_MATRIX) {
5388: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5389: for (i = 0; i < am; i++) {
5390: const PetscInt ncols_d = ai[i + 1] - ai[i];
5391: const PetscInt ncols_o = bi[i + 1] - bi[i];
5392: /* diagonal portion of A */
5393: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5394: /* off-diagonal portion of A */
5395: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5396: }
5397: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5398: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5399: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5400: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5401: if (glob) {
5402: PetscInt cst, *gidx;
5404: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5405: PetscCall(PetscMalloc1(dn + on, &gidx));
5406: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5407: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5408: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5409: }
5410: }
5411: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5412: PetscFunctionReturn(PETSC_SUCCESS);
5413: }
5415: /*@C
5416: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5418: Not Collective
5420: Input Parameters:
5421: + A - the matrix
5422: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5423: . row - index set of rows to extract (or `NULL`)
5424: - col - index set of columns to extract (or `NULL`)
5426: Output Parameter:
5427: . A_loc - the local sequential matrix generated
5429: Level: developer
5431: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5432: @*/
5433: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5434: {
5435: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5436: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5437: IS isrowa, iscola;
5438: Mat *aloc;
5439: PetscBool match;
5441: PetscFunctionBegin;
5442: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5443: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5444: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5445: if (!row) {
5446: start = A->rmap->rstart;
5447: end = A->rmap->rend;
5448: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5449: } else {
5450: isrowa = *row;
5451: }
5452: if (!col) {
5453: start = A->cmap->rstart;
5454: cmap = a->garray;
5455: nzA = a->A->cmap->n;
5456: nzB = a->B->cmap->n;
5457: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5458: ncols = 0;
5459: for (i = 0; i < nzB; i++) {
5460: if (cmap[i] < start) idx[ncols++] = cmap[i];
5461: else break;
5462: }
5463: imark = i;
5464: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5465: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5466: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5467: } else {
5468: iscola = *col;
5469: }
5470: if (scall != MAT_INITIAL_MATRIX) {
5471: PetscCall(PetscMalloc1(1, &aloc));
5472: aloc[0] = *A_loc;
5473: }
5474: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5475: if (!col) { /* attach global id of condensed columns */
5476: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5477: }
5478: *A_loc = aloc[0];
5479: PetscCall(PetscFree(aloc));
5480: if (!row) PetscCall(ISDestroy(&isrowa));
5481: if (!col) PetscCall(ISDestroy(&iscola));
5482: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5483: PetscFunctionReturn(PETSC_SUCCESS);
5484: }
5486: /*
5487: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5488: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5489: * on a global size.
5490: * */
5491: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5492: {
5493: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5494: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5495: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5496: PetscMPIInt owner;
5497: PetscSFNode *iremote, *oiremote;
5498: const PetscInt *lrowindices;
5499: PetscSF sf, osf;
5500: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5501: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5502: MPI_Comm comm;
5503: ISLocalToGlobalMapping mapping;
5504: const PetscScalar *pd_a, *po_a;
5506: PetscFunctionBegin;
5507: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5508: /* plocalsize is the number of roots
5509: * nrows is the number of leaves
5510: * */
5511: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5512: PetscCall(ISGetLocalSize(rows, &nrows));
5513: PetscCall(PetscCalloc1(nrows, &iremote));
5514: PetscCall(ISGetIndices(rows, &lrowindices));
5515: for (i = 0; i < nrows; i++) {
5516: /* Find a remote index and an owner for a row
5517: * The row could be local or remote
5518: * */
5519: owner = 0;
5520: lidx = 0;
5521: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5522: iremote[i].index = lidx;
5523: iremote[i].rank = owner;
5524: }
5525: /* Create SF to communicate how many nonzero columns for each row */
5526: PetscCall(PetscSFCreate(comm, &sf));
5527: /* SF will figure out the number of nonzero columns for each row, and their
5528: * offsets
5529: * */
5530: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5531: PetscCall(PetscSFSetFromOptions(sf));
5532: PetscCall(PetscSFSetUp(sf));
5534: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5535: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5536: PetscCall(PetscCalloc1(nrows, &pnnz));
5537: roffsets[0] = 0;
5538: roffsets[1] = 0;
5539: for (i = 0; i < plocalsize; i++) {
5540: /* diagonal */
5541: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5542: /* off-diagonal */
5543: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5544: /* compute offsets so that we relative location for each row */
5545: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5546: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5547: }
5548: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5549: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5550: /* 'r' means root, and 'l' means leaf */
5551: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5552: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5553: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5554: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5555: PetscCall(PetscSFDestroy(&sf));
5556: PetscCall(PetscFree(roffsets));
5557: PetscCall(PetscFree(nrcols));
5558: dntotalcols = 0;
5559: ontotalcols = 0;
5560: ncol = 0;
5561: for (i = 0; i < nrows; i++) {
5562: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5563: ncol = PetscMax(pnnz[i], ncol);
5564: /* diagonal */
5565: dntotalcols += nlcols[i * 2 + 0];
5566: /* off-diagonal */
5567: ontotalcols += nlcols[i * 2 + 1];
5568: }
5569: /* We do not need to figure the right number of columns
5570: * since all the calculations will be done by going through the raw data
5571: * */
5572: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5573: PetscCall(MatSetUp(*P_oth));
5574: PetscCall(PetscFree(pnnz));
5575: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5576: /* diagonal */
5577: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5578: /* off-diagonal */
5579: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5580: /* diagonal */
5581: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5582: /* off-diagonal */
5583: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5584: dntotalcols = 0;
5585: ontotalcols = 0;
5586: ntotalcols = 0;
5587: for (i = 0; i < nrows; i++) {
5588: owner = 0;
5589: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5590: /* Set iremote for diag matrix */
5591: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5592: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5593: iremote[dntotalcols].rank = owner;
5594: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5595: ilocal[dntotalcols++] = ntotalcols++;
5596: }
5597: /* off-diagonal */
5598: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5599: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5600: oiremote[ontotalcols].rank = owner;
5601: oilocal[ontotalcols++] = ntotalcols++;
5602: }
5603: }
5604: PetscCall(ISRestoreIndices(rows, &lrowindices));
5605: PetscCall(PetscFree(loffsets));
5606: PetscCall(PetscFree(nlcols));
5607: PetscCall(PetscSFCreate(comm, &sf));
5608: /* P serves as roots and P_oth is leaves
5609: * Diag matrix
5610: * */
5611: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5612: PetscCall(PetscSFSetFromOptions(sf));
5613: PetscCall(PetscSFSetUp(sf));
5615: PetscCall(PetscSFCreate(comm, &osf));
5616: /* off-diagonal */
5617: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5618: PetscCall(PetscSFSetFromOptions(osf));
5619: PetscCall(PetscSFSetUp(osf));
5620: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5621: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5622: /* operate on the matrix internal data to save memory */
5623: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5624: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5625: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5626: /* Convert to global indices for diag matrix */
5627: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5628: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5629: /* We want P_oth store global indices */
5630: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5631: /* Use memory scalable approach */
5632: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5633: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5634: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5635: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5636: /* Convert back to local indices */
5637: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5638: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5639: nout = 0;
5640: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5641: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5642: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5643: /* Exchange values */
5644: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5645: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5646: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5647: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5648: /* Stop PETSc from shrinking memory */
5649: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5650: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5651: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5652: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5653: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5654: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5655: PetscCall(PetscSFDestroy(&sf));
5656: PetscCall(PetscSFDestroy(&osf));
5657: PetscFunctionReturn(PETSC_SUCCESS);
5658: }
5660: /*
5661: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5662: * This supports MPIAIJ and MAIJ
5663: * */
5664: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5665: {
5666: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5667: Mat_SeqAIJ *p_oth;
5668: IS rows, map;
5669: PetscHMapI hamp;
5670: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5671: MPI_Comm comm;
5672: PetscSF sf, osf;
5673: PetscBool has;
5675: PetscFunctionBegin;
5676: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5677: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5678: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5679: * and then create a submatrix (that often is an overlapping matrix)
5680: * */
5681: if (reuse == MAT_INITIAL_MATRIX) {
5682: /* Use a hash table to figure out unique keys */
5683: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5684: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5685: count = 0;
5686: /* Assume that a->g is sorted, otherwise the following does not make sense */
5687: for (i = 0; i < a->B->cmap->n; i++) {
5688: key = a->garray[i] / dof;
5689: PetscCall(PetscHMapIHas(hamp, key, &has));
5690: if (!has) {
5691: mapping[i] = count;
5692: PetscCall(PetscHMapISet(hamp, key, count++));
5693: } else {
5694: /* Current 'i' has the same value the previous step */
5695: mapping[i] = count - 1;
5696: }
5697: }
5698: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5699: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5700: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5701: PetscCall(PetscCalloc1(htsize, &rowindices));
5702: off = 0;
5703: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5704: PetscCall(PetscHMapIDestroy(&hamp));
5705: PetscCall(PetscSortInt(htsize, rowindices));
5706: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5707: /* In case, the matrix was already created but users want to recreate the matrix */
5708: PetscCall(MatDestroy(P_oth));
5709: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5710: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5711: PetscCall(ISDestroy(&map));
5712: PetscCall(ISDestroy(&rows));
5713: } else if (reuse == MAT_REUSE_MATRIX) {
5714: /* If matrix was already created, we simply update values using SF objects
5715: * that as attached to the matrix earlier.
5716: */
5717: const PetscScalar *pd_a, *po_a;
5719: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5720: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5721: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5722: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5723: /* Update values in place */
5724: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5725: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5726: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5727: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5728: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5729: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5730: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5731: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5732: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5733: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5734: PetscFunctionReturn(PETSC_SUCCESS);
5735: }
5737: /*@C
5738: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5740: Collective
5742: Input Parameters:
5743: + A - the first matrix in `MATMPIAIJ` format
5744: . B - the second matrix in `MATMPIAIJ` format
5745: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5747: Output Parameters:
5748: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5749: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5750: - B_seq - the sequential matrix generated
5752: Level: developer
5754: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5755: @*/
5756: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5757: {
5758: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5759: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5760: IS isrowb, iscolb;
5761: Mat *bseq = NULL;
5763: PetscFunctionBegin;
5764: 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 ")",
5765: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5766: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5768: if (scall == MAT_INITIAL_MATRIX) {
5769: start = A->cmap->rstart;
5770: cmap = a->garray;
5771: nzA = a->A->cmap->n;
5772: nzB = a->B->cmap->n;
5773: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5774: ncols = 0;
5775: for (i = 0; i < nzB; i++) { /* row < local row index */
5776: if (cmap[i] < start) idx[ncols++] = cmap[i];
5777: else break;
5778: }
5779: imark = i;
5780: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5781: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5782: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5783: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5784: } else {
5785: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5786: isrowb = *rowb;
5787: iscolb = *colb;
5788: PetscCall(PetscMalloc1(1, &bseq));
5789: bseq[0] = *B_seq;
5790: }
5791: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5792: *B_seq = bseq[0];
5793: PetscCall(PetscFree(bseq));
5794: if (!rowb) {
5795: PetscCall(ISDestroy(&isrowb));
5796: } else {
5797: *rowb = isrowb;
5798: }
5799: if (!colb) {
5800: PetscCall(ISDestroy(&iscolb));
5801: } else {
5802: *colb = iscolb;
5803: }
5804: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5805: PetscFunctionReturn(PETSC_SUCCESS);
5806: }
5808: /*
5809: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5810: of the OFF-DIAGONAL portion of local A
5812: Collective
5814: Input Parameters:
5815: + A,B - the matrices in `MATMPIAIJ` format
5816: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5818: Output Parameter:
5819: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5820: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5821: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5822: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5824: Developer Note:
5825: This directly accesses information inside the VecScatter associated with the matrix-vector product
5826: for this matrix. This is not desirable..
5828: Level: developer
5830: */
5831: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5832: {
5833: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5834: Mat_SeqAIJ *b_oth;
5835: VecScatter ctx;
5836: MPI_Comm comm;
5837: const PetscMPIInt *rprocs, *sprocs;
5838: const PetscInt *srow, *rstarts, *sstarts;
5839: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5840: PetscInt i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5841: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5842: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5843: PetscMPIInt size, tag, rank, nreqs;
5845: PetscFunctionBegin;
5846: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5847: PetscCallMPI(MPI_Comm_size(comm, &size));
5849: 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 ")",
5850: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5851: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5852: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5854: if (size == 1) {
5855: startsj_s = NULL;
5856: bufa_ptr = NULL;
5857: *B_oth = NULL;
5858: PetscFunctionReturn(PETSC_SUCCESS);
5859: }
5861: ctx = a->Mvctx;
5862: tag = ((PetscObject)ctx)->tag;
5864: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5865: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5866: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5867: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5868: PetscCall(PetscMalloc1(nreqs, &reqs));
5869: rwaits = reqs;
5870: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5872: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5873: if (scall == MAT_INITIAL_MATRIX) {
5874: /* i-array */
5875: /* post receives */
5876: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5877: for (i = 0; i < nrecvs; i++) {
5878: rowlen = rvalues + rstarts[i] * rbs;
5879: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5880: PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5881: }
5883: /* pack the outgoing message */
5884: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5886: sstartsj[0] = 0;
5887: rstartsj[0] = 0;
5888: len = 0; /* total length of j or a array to be sent */
5889: if (nsends) {
5890: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5891: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5892: }
5893: for (i = 0; i < nsends; i++) {
5894: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5895: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5896: for (j = 0; j < nrows; j++) {
5897: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5898: for (l = 0; l < sbs; l++) {
5899: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5901: rowlen[j * sbs + l] = ncols;
5903: len += ncols;
5904: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5905: }
5906: k++;
5907: }
5908: PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5910: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5911: }
5912: /* recvs and sends of i-array are completed */
5913: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5914: PetscCall(PetscFree(svalues));
5916: /* allocate buffers for sending j and a arrays */
5917: PetscCall(PetscMalloc1(len + 1, &bufj));
5918: PetscCall(PetscMalloc1(len + 1, &bufa));
5920: /* create i-array of B_oth */
5921: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5923: b_othi[0] = 0;
5924: len = 0; /* total length of j or a array to be received */
5925: k = 0;
5926: for (i = 0; i < nrecvs; i++) {
5927: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5928: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5929: for (j = 0; j < nrows; j++) {
5930: b_othi[k + 1] = b_othi[k] + rowlen[j];
5931: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5932: k++;
5933: }
5934: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5935: }
5936: PetscCall(PetscFree(rvalues));
5938: /* allocate space for j and a arrays of B_oth */
5939: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5940: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5942: /* j-array */
5943: /* post receives of j-array */
5944: for (i = 0; i < nrecvs; i++) {
5945: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5946: PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5947: }
5949: /* pack the outgoing message j-array */
5950: if (nsends) k = sstarts[0];
5951: for (i = 0; i < nsends; i++) {
5952: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5953: bufJ = bufj + sstartsj[i];
5954: for (j = 0; j < nrows; j++) {
5955: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5956: for (ll = 0; ll < sbs; ll++) {
5957: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5958: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5959: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5960: }
5961: }
5962: PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5963: }
5965: /* recvs and sends of j-array are completed */
5966: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5967: } else if (scall == MAT_REUSE_MATRIX) {
5968: sstartsj = *startsj_s;
5969: rstartsj = *startsj_r;
5970: bufa = *bufa_ptr;
5971: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5972: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5973: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5975: /* a-array */
5976: /* post receives of a-array */
5977: for (i = 0; i < nrecvs; i++) {
5978: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5979: PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5980: }
5982: /* pack the outgoing message a-array */
5983: if (nsends) k = sstarts[0];
5984: for (i = 0; i < nsends; i++) {
5985: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5986: bufA = bufa + sstartsj[i];
5987: for (j = 0; j < nrows; j++) {
5988: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5989: for (ll = 0; ll < sbs; ll++) {
5990: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5991: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5992: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5993: }
5994: }
5995: PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5996: }
5997: /* recvs and sends of a-array are completed */
5998: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5999: PetscCall(PetscFree(reqs));
6001: if (scall == MAT_INITIAL_MATRIX) {
6002: /* put together the new matrix */
6003: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6005: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6006: /* Since these are PETSc arrays, change flags to free them as necessary. */
6007: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6008: b_oth->free_a = PETSC_TRUE;
6009: b_oth->free_ij = PETSC_TRUE;
6010: b_oth->nonew = 0;
6012: PetscCall(PetscFree(bufj));
6013: if (!startsj_s || !bufa_ptr) {
6014: PetscCall(PetscFree2(sstartsj, rstartsj));
6015: PetscCall(PetscFree(bufa_ptr));
6016: } else {
6017: *startsj_s = sstartsj;
6018: *startsj_r = rstartsj;
6019: *bufa_ptr = bufa;
6020: }
6021: } else if (scall == MAT_REUSE_MATRIX) {
6022: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6023: }
6025: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6026: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6027: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6028: PetscFunctionReturn(PETSC_SUCCESS);
6029: }
6031: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6032: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6033: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6034: #if defined(PETSC_HAVE_MKL_SPARSE)
6035: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6036: #endif
6037: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6038: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6039: #if defined(PETSC_HAVE_ELEMENTAL)
6040: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6041: #endif
6042: #if defined(PETSC_HAVE_SCALAPACK)
6043: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6044: #endif
6045: #if defined(PETSC_HAVE_HYPRE)
6046: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6047: #endif
6048: #if defined(PETSC_HAVE_CUDA)
6049: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6050: #endif
6051: #if defined(PETSC_HAVE_HIP)
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6053: #endif
6054: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6055: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6058: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6059: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6061: /*
6062: Computes (B'*A')' since computing B*A directly is untenable
6064: n p p
6065: [ ] [ ] [ ]
6066: m [ A ] * n [ B ] = m [ C ]
6067: [ ] [ ] [ ]
6069: */
6070: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6071: {
6072: Mat At, Bt, Ct;
6074: PetscFunctionBegin;
6075: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6076: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6077: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6078: PetscCall(MatDestroy(&At));
6079: PetscCall(MatDestroy(&Bt));
6080: PetscCall(MatTransposeSetPrecursor(Ct, C));
6081: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6082: PetscCall(MatDestroy(&Ct));
6083: PetscFunctionReturn(PETSC_SUCCESS);
6084: }
6086: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6087: {
6088: PetscBool cisdense;
6090: PetscFunctionBegin;
6091: 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);
6092: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6093: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6094: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6095: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6096: PetscCall(MatSetUp(C));
6098: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6099: PetscFunctionReturn(PETSC_SUCCESS);
6100: }
6102: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6103: {
6104: Mat_Product *product = C->product;
6105: Mat A = product->A, B = product->B;
6107: PetscFunctionBegin;
6108: 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 ")",
6109: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6110: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6111: C->ops->productsymbolic = MatProductSymbolic_AB;
6112: PetscFunctionReturn(PETSC_SUCCESS);
6113: }
6115: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6116: {
6117: Mat_Product *product = C->product;
6119: PetscFunctionBegin;
6120: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6121: PetscFunctionReturn(PETSC_SUCCESS);
6122: }
6124: /*
6125: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6127: Input Parameters:
6129: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6130: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6132: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6134: For Set1, j1[] contains column indices of the nonzeros.
6135: 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
6136: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6137: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6139: Similar for Set2.
6141: This routine merges the two sets of nonzeros row by row and removes repeats.
6143: Output Parameters: (memory is allocated by the caller)
6145: i[],j[]: the CSR of the merged matrix, which has m rows.
6146: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6147: imap2[]: similar to imap1[], but for Set2.
6148: Note we order nonzeros row-by-row and from left to right.
6149: */
6150: 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[])
6151: {
6152: PetscInt r, m; /* Row index of mat */
6153: PetscCount t, t1, t2, b1, e1, b2, e2;
6155: PetscFunctionBegin;
6156: PetscCall(MatGetLocalSize(mat, &m, NULL));
6157: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6158: i[0] = 0;
6159: for (r = 0; r < m; r++) { /* Do row by row merging */
6160: b1 = rowBegin1[r];
6161: e1 = rowEnd1[r];
6162: b2 = rowBegin2[r];
6163: e2 = rowEnd2[r];
6164: while (b1 < e1 && b2 < e2) {
6165: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6166: j[t] = j1[b1];
6167: imap1[t1] = t;
6168: imap2[t2] = t;
6169: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6170: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6171: t1++;
6172: t2++;
6173: t++;
6174: } else if (j1[b1] < j2[b2]) {
6175: j[t] = j1[b1];
6176: imap1[t1] = t;
6177: b1 += jmap1[t1 + 1] - jmap1[t1];
6178: t1++;
6179: t++;
6180: } else {
6181: j[t] = j2[b2];
6182: imap2[t2] = t;
6183: b2 += jmap2[t2 + 1] - jmap2[t2];
6184: t2++;
6185: t++;
6186: }
6187: }
6188: /* Merge the remaining in either j1[] or j2[] */
6189: while (b1 < e1) {
6190: j[t] = j1[b1];
6191: imap1[t1] = t;
6192: b1 += jmap1[t1 + 1] - jmap1[t1];
6193: t1++;
6194: t++;
6195: }
6196: while (b2 < e2) {
6197: j[t] = j2[b2];
6198: imap2[t2] = t;
6199: b2 += jmap2[t2 + 1] - jmap2[t2];
6200: t2++;
6201: t++;
6202: }
6203: i[r + 1] = t;
6204: }
6205: PetscFunctionReturn(PETSC_SUCCESS);
6206: }
6208: /*
6209: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6211: Input Parameters:
6212: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6213: 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[]
6214: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6216: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6217: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6219: Output Parameters:
6220: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6221: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6222: 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,
6223: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6225: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6226: Atot: number of entries belonging to the diagonal block.
6227: Annz: number of unique nonzeros belonging to the diagonal block.
6228: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6229: repeats (i.e., same 'i,j' pair).
6230: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6231: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6233: Atot: number of entries belonging to the diagonal block
6234: Annz: number of unique nonzeros belonging to the diagonal block.
6236: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6238: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6239: */
6240: 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_)
6241: {
6242: PetscInt cstart, cend, rstart, rend, row, col;
6243: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6244: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6245: PetscCount k, m, p, q, r, s, mid;
6246: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6248: PetscFunctionBegin;
6249: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6250: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6251: m = rend - rstart;
6253: /* Skip negative rows */
6254: for (k = 0; k < n; k++)
6255: if (i[k] >= 0) break;
6257: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6258: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6259: */
6260: while (k < n) {
6261: row = i[k];
6262: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6263: for (s = k; s < n; s++)
6264: if (i[s] != row) break;
6266: /* Shift diag columns to range of [-PETSC_MAX_INT, -1] */
6267: for (p = k; p < s; p++) {
6268: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT;
6269: 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]);
6270: }
6271: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6272: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6273: rowBegin[row - rstart] = k;
6274: rowMid[row - rstart] = mid;
6275: rowEnd[row - rstart] = s;
6277: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6278: Atot += mid - k;
6279: Btot += s - mid;
6281: /* Count unique nonzeros of this diag row */
6282: for (p = k; p < mid;) {
6283: col = j[p];
6284: do {
6285: j[p] += PETSC_MAX_INT; /* Revert the modified diagonal indices */
6286: p++;
6287: } while (p < mid && j[p] == col);
6288: Annz++;
6289: }
6291: /* Count unique nonzeros of this offdiag row */
6292: for (p = mid; p < s;) {
6293: col = j[p];
6294: do {
6295: p++;
6296: } while (p < s && j[p] == col);
6297: Bnnz++;
6298: }
6299: k = s;
6300: }
6302: /* Allocation according to Atot, Btot, Annz, Bnnz */
6303: PetscCall(PetscMalloc1(Atot, &Aperm));
6304: PetscCall(PetscMalloc1(Btot, &Bperm));
6305: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6306: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6308: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6309: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6310: for (r = 0; r < m; r++) {
6311: k = rowBegin[r];
6312: mid = rowMid[r];
6313: s = rowEnd[r];
6314: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6315: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6316: Atot += mid - k;
6317: Btot += s - mid;
6319: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6320: for (p = k; p < mid;) {
6321: col = j[p];
6322: q = p;
6323: do {
6324: p++;
6325: } while (p < mid && j[p] == col);
6326: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6327: Annz++;
6328: }
6330: for (p = mid; p < s;) {
6331: col = j[p];
6332: q = p;
6333: do {
6334: p++;
6335: } while (p < s && j[p] == col);
6336: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6337: Bnnz++;
6338: }
6339: }
6340: /* Output */
6341: *Aperm_ = Aperm;
6342: *Annz_ = Annz;
6343: *Atot_ = Atot;
6344: *Ajmap_ = Ajmap;
6345: *Bperm_ = Bperm;
6346: *Bnnz_ = Bnnz;
6347: *Btot_ = Btot;
6348: *Bjmap_ = Bjmap;
6349: PetscFunctionReturn(PETSC_SUCCESS);
6350: }
6352: /*
6353: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6355: Input Parameters:
6356: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6357: nnz: number of unique nonzeros in the merged matrix
6358: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6359: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6361: Output Parameter: (memory is allocated by the caller)
6362: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6364: Example:
6365: nnz1 = 4
6366: nnz = 6
6367: imap = [1,3,4,5]
6368: jmap = [0,3,5,6,7]
6369: then,
6370: jmap_new = [0,0,3,3,5,6,7]
6371: */
6372: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6373: {
6374: PetscCount k, p;
6376: PetscFunctionBegin;
6377: jmap_new[0] = 0;
6378: p = nnz; /* p loops over jmap_new[] backwards */
6379: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6380: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6381: }
6382: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6383: PetscFunctionReturn(PETSC_SUCCESS);
6384: }
6386: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6387: {
6388: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6390: PetscFunctionBegin;
6391: PetscCall(PetscSFDestroy(&coo->sf));
6392: PetscCall(PetscFree(coo->Aperm1));
6393: PetscCall(PetscFree(coo->Bperm1));
6394: PetscCall(PetscFree(coo->Ajmap1));
6395: PetscCall(PetscFree(coo->Bjmap1));
6396: PetscCall(PetscFree(coo->Aimap2));
6397: PetscCall(PetscFree(coo->Bimap2));
6398: PetscCall(PetscFree(coo->Aperm2));
6399: PetscCall(PetscFree(coo->Bperm2));
6400: PetscCall(PetscFree(coo->Ajmap2));
6401: PetscCall(PetscFree(coo->Bjmap2));
6402: PetscCall(PetscFree(coo->Cperm1));
6403: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6404: PetscCall(PetscFree(coo));
6405: PetscFunctionReturn(PETSC_SUCCESS);
6406: }
6408: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6409: {
6410: MPI_Comm comm;
6411: PetscMPIInt rank, size;
6412: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6413: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6414: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6415: PetscContainer container;
6416: MatCOOStruct_MPIAIJ *coo;
6418: PetscFunctionBegin;
6419: PetscCall(PetscFree(mpiaij->garray));
6420: PetscCall(VecDestroy(&mpiaij->lvec));
6421: #if defined(PETSC_USE_CTABLE)
6422: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6423: #else
6424: PetscCall(PetscFree(mpiaij->colmap));
6425: #endif
6426: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6427: mat->assembled = PETSC_FALSE;
6428: mat->was_assembled = PETSC_FALSE;
6430: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6431: PetscCallMPI(MPI_Comm_size(comm, &size));
6432: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6433: PetscCall(PetscLayoutSetUp(mat->rmap));
6434: PetscCall(PetscLayoutSetUp(mat->cmap));
6435: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6436: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6437: PetscCall(MatGetLocalSize(mat, &m, &n));
6438: PetscCall(MatGetSize(mat, &M, &N));
6440: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6441: /* entries come first, then local rows, then remote rows. */
6442: PetscCount n1 = coo_n, *perm1;
6443: PetscInt *i1 = coo_i, *j1 = coo_j;
6445: PetscCall(PetscMalloc1(n1, &perm1));
6446: for (k = 0; k < n1; k++) perm1[k] = k;
6448: /* Manipulate indices so that entries with negative row or col indices will have smallest
6449: row indices, local entries will have greater but negative row indices, and remote entries
6450: will have positive row indices.
6451: */
6452: for (k = 0; k < n1; k++) {
6453: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT; /* e.g., -2^31, minimal to move them ahead */
6454: 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] */
6455: else {
6456: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6457: if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6458: }
6459: }
6461: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6462: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6464: /* Advance k to the first entry we need to take care of */
6465: for (k = 0; k < n1; k++)
6466: if (i1[k] > PETSC_MIN_INT) break;
6467: PetscInt i1start = k;
6469: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6470: for (; k < rem; k++) i1[k] += PETSC_MAX_INT; /* Revert row indices of local rows*/
6472: /* Send remote rows to their owner */
6473: /* Find which rows should be sent to which remote ranks*/
6474: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6475: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6476: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6477: const PetscInt *ranges;
6478: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6480: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6481: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6482: for (k = rem; k < n1;) {
6483: PetscMPIInt owner;
6484: PetscInt firstRow, lastRow;
6486: /* Locate a row range */
6487: firstRow = i1[k]; /* first row of this owner */
6488: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6489: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6491: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6492: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6494: /* All entries in [k,p) belong to this remote owner */
6495: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6496: PetscMPIInt *sendto2;
6497: PetscInt *nentries2;
6498: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6500: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6501: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6502: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6503: PetscCall(PetscFree2(sendto, nentries2));
6504: sendto = sendto2;
6505: nentries = nentries2;
6506: maxNsend = maxNsend2;
6507: }
6508: sendto[nsend] = owner;
6509: nentries[nsend] = p - k;
6510: PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6511: nsend++;
6512: k = p;
6513: }
6515: /* Build 1st SF to know offsets on remote to send data */
6516: PetscSF sf1;
6517: PetscInt nroots = 1, nroots2 = 0;
6518: PetscInt nleaves = nsend, nleaves2 = 0;
6519: PetscInt *offsets;
6520: PetscSFNode *iremote;
6522: PetscCall(PetscSFCreate(comm, &sf1));
6523: PetscCall(PetscMalloc1(nsend, &iremote));
6524: PetscCall(PetscMalloc1(nsend, &offsets));
6525: for (k = 0; k < nsend; k++) {
6526: iremote[k].rank = sendto[k];
6527: iremote[k].index = 0;
6528: nleaves2 += nentries[k];
6529: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6530: }
6531: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6532: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6533: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6534: PetscCall(PetscSFDestroy(&sf1));
6535: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6537: /* Build 2nd SF to send remote COOs to their owner */
6538: PetscSF sf2;
6539: nroots = nroots2;
6540: nleaves = nleaves2;
6541: PetscCall(PetscSFCreate(comm, &sf2));
6542: PetscCall(PetscSFSetFromOptions(sf2));
6543: PetscCall(PetscMalloc1(nleaves, &iremote));
6544: p = 0;
6545: for (k = 0; k < nsend; k++) {
6546: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6547: for (q = 0; q < nentries[k]; q++, p++) {
6548: iremote[p].rank = sendto[k];
6549: iremote[p].index = offsets[k] + q;
6550: }
6551: }
6552: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6554: /* Send the remote COOs to their owner */
6555: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6556: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6557: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6558: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1 + rem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6559: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1 + rem, i2, MPI_REPLACE));
6560: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1 + rem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6561: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1 + rem, j2, MPI_REPLACE));
6563: PetscCall(PetscFree(offsets));
6564: PetscCall(PetscFree2(sendto, nentries));
6566: /* Sort received COOs by row along with the permutation array */
6567: for (k = 0; k < n2; k++) perm2[k] = k;
6568: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6570: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6571: PetscCount *Cperm1;
6572: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6573: PetscCall(PetscArraycpy(Cperm1, perm1 + rem, nleaves));
6575: /* Support for HYPRE matrices, kind of a hack.
6576: Swap min column with diagonal so that diagonal values will go first */
6577: PetscBool hypre;
6578: const char *name;
6579: PetscCall(PetscObjectGetName((PetscObject)mat, &name));
6580: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
6581: if (hypre) {
6582: PetscInt *minj;
6583: PetscBT hasdiag;
6585: PetscCall(PetscBTCreate(m, &hasdiag));
6586: PetscCall(PetscMalloc1(m, &minj));
6587: for (k = 0; k < m; k++) minj[k] = PETSC_MAX_INT;
6588: for (k = i1start; k < rem; k++) {
6589: if (j1[k] < cstart || j1[k] >= cend) continue;
6590: const PetscInt rindex = i1[k] - rstart;
6591: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6592: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6593: }
6594: for (k = 0; k < n2; k++) {
6595: if (j2[k] < cstart || j2[k] >= cend) continue;
6596: const PetscInt rindex = i2[k] - rstart;
6597: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6598: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6599: }
6600: for (k = i1start; k < rem; k++) {
6601: const PetscInt rindex = i1[k] - rstart;
6602: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6603: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6604: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6605: }
6606: for (k = 0; k < n2; k++) {
6607: const PetscInt rindex = i2[k] - rstart;
6608: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6609: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6610: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6611: }
6612: PetscCall(PetscBTDestroy(&hasdiag));
6613: PetscCall(PetscFree(minj));
6614: }
6616: /* Split local COOs and received COOs into diag/offdiag portions */
6617: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6618: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6619: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6620: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6621: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6622: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6624: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6625: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6626: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6627: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6629: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6630: PetscInt *Ai, *Bi;
6631: PetscInt *Aj, *Bj;
6633: PetscCall(PetscMalloc1(m + 1, &Ai));
6634: PetscCall(PetscMalloc1(m + 1, &Bi));
6635: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6636: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6638: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6639: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6640: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6641: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6642: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6644: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6645: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6647: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6648: /* expect nonzeros in A/B most likely have local contributing entries */
6649: PetscInt Annz = Ai[m];
6650: PetscInt Bnnz = Bi[m];
6651: PetscCount *Ajmap1_new, *Bjmap1_new;
6653: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6654: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6656: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6657: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6659: PetscCall(PetscFree(Aimap1));
6660: PetscCall(PetscFree(Ajmap1));
6661: PetscCall(PetscFree(Bimap1));
6662: PetscCall(PetscFree(Bjmap1));
6663: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6664: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6665: PetscCall(PetscFree(perm1));
6666: PetscCall(PetscFree3(i2, j2, perm2));
6668: Ajmap1 = Ajmap1_new;
6669: Bjmap1 = Bjmap1_new;
6671: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6672: if (Annz < Annz1 + Annz2) {
6673: PetscInt *Aj_new;
6674: PetscCall(PetscMalloc1(Annz, &Aj_new));
6675: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6676: PetscCall(PetscFree(Aj));
6677: Aj = Aj_new;
6678: }
6680: if (Bnnz < Bnnz1 + Bnnz2) {
6681: PetscInt *Bj_new;
6682: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6683: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6684: PetscCall(PetscFree(Bj));
6685: Bj = Bj_new;
6686: }
6688: /* Create new submatrices for on-process and off-process coupling */
6689: PetscScalar *Aa, *Ba;
6690: MatType rtype;
6691: Mat_SeqAIJ *a, *b;
6692: PetscObjectState state;
6693: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6694: PetscCall(PetscCalloc1(Bnnz, &Ba));
6695: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6696: if (cstart) {
6697: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6698: }
6700: PetscCall(MatGetRootType_Private(mat, &rtype));
6702: MatSeqXAIJGetOptions_Private(mpiaij->A);
6703: PetscCall(MatDestroy(&mpiaij->A));
6704: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6705: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6706: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6708: MatSeqXAIJGetOptions_Private(mpiaij->B);
6709: PetscCall(MatDestroy(&mpiaij->B));
6710: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6711: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6712: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6714: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6715: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6716: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6717: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6719: a = (Mat_SeqAIJ *)mpiaij->A->data;
6720: b = (Mat_SeqAIJ *)mpiaij->B->data;
6721: a->singlemalloc = b->singlemalloc = PETSC_FALSE; /* Let newmat own Ai,Aj,Aa,Bi,Bj,Ba */
6722: a->free_a = b->free_a = PETSC_TRUE;
6723: a->free_ij = b->free_ij = PETSC_TRUE;
6725: /* conversion must happen AFTER multiply setup */
6726: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6727: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6728: PetscCall(VecDestroy(&mpiaij->lvec));
6729: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6731: // Put the COO struct in a container and then attach that to the matrix
6732: PetscCall(PetscMalloc1(1, &coo));
6733: coo->n = coo_n;
6734: coo->sf = sf2;
6735: coo->sendlen = nleaves;
6736: coo->recvlen = nroots;
6737: coo->Annz = Annz;
6738: coo->Bnnz = Bnnz;
6739: coo->Annz2 = Annz2;
6740: coo->Bnnz2 = Bnnz2;
6741: coo->Atot1 = Atot1;
6742: coo->Atot2 = Atot2;
6743: coo->Btot1 = Btot1;
6744: coo->Btot2 = Btot2;
6745: coo->Ajmap1 = Ajmap1;
6746: coo->Aperm1 = Aperm1;
6747: coo->Bjmap1 = Bjmap1;
6748: coo->Bperm1 = Bperm1;
6749: coo->Aimap2 = Aimap2;
6750: coo->Ajmap2 = Ajmap2;
6751: coo->Aperm2 = Aperm2;
6752: coo->Bimap2 = Bimap2;
6753: coo->Bjmap2 = Bjmap2;
6754: coo->Bperm2 = Bperm2;
6755: coo->Cperm1 = Cperm1;
6756: // Allocate in preallocation. If not used, it has zero cost on host
6757: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6758: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6759: PetscCall(PetscContainerSetPointer(container, coo));
6760: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6761: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6762: PetscCall(PetscContainerDestroy(&container));
6763: PetscFunctionReturn(PETSC_SUCCESS);
6764: }
6766: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6767: {
6768: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6769: Mat A = mpiaij->A, B = mpiaij->B;
6770: PetscScalar *Aa, *Ba;
6771: PetscScalar *sendbuf, *recvbuf;
6772: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6773: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6774: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6775: const PetscCount *Cperm1;
6776: PetscContainer container;
6777: MatCOOStruct_MPIAIJ *coo;
6779: PetscFunctionBegin;
6780: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6781: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6782: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6783: sendbuf = coo->sendbuf;
6784: recvbuf = coo->recvbuf;
6785: Ajmap1 = coo->Ajmap1;
6786: Ajmap2 = coo->Ajmap2;
6787: Aimap2 = coo->Aimap2;
6788: Bjmap1 = coo->Bjmap1;
6789: Bjmap2 = coo->Bjmap2;
6790: Bimap2 = coo->Bimap2;
6791: Aperm1 = coo->Aperm1;
6792: Aperm2 = coo->Aperm2;
6793: Bperm1 = coo->Bperm1;
6794: Bperm2 = coo->Bperm2;
6795: Cperm1 = coo->Cperm1;
6797: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6798: PetscCall(MatSeqAIJGetArray(B, &Ba));
6800: /* Pack entries to be sent to remote */
6801: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6803: /* Send remote entries to their owner and overlap the communication with local computation */
6804: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6805: /* Add local entries to A and B */
6806: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6807: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6808: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6809: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6810: }
6811: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6812: PetscScalar sum = 0.0;
6813: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6814: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6815: }
6816: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6818: /* Add received remote entries to A and B */
6819: for (PetscCount i = 0; i < coo->Annz2; i++) {
6820: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6821: }
6822: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6823: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6824: }
6825: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6826: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6827: PetscFunctionReturn(PETSC_SUCCESS);
6828: }
6830: /*MC
6831: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6833: Options Database Keys:
6834: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6836: Level: beginner
6838: Notes:
6839: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6840: in this case the values associated with the rows and columns one passes in are set to zero
6841: in the matrix
6843: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6844: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6846: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6847: M*/
6848: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6849: {
6850: Mat_MPIAIJ *b;
6851: PetscMPIInt size;
6853: PetscFunctionBegin;
6854: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6856: PetscCall(PetscNew(&b));
6857: B->data = (void *)b;
6858: B->ops[0] = MatOps_Values;
6859: B->assembled = PETSC_FALSE;
6860: B->insertmode = NOT_SET_VALUES;
6861: b->size = size;
6863: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6865: /* build cache for off array entries formed */
6866: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6868: b->donotstash = PETSC_FALSE;
6869: b->colmap = NULL;
6870: b->garray = NULL;
6871: b->roworiented = PETSC_TRUE;
6873: /* stuff used for matrix vector multiply */
6874: b->lvec = NULL;
6875: b->Mvctx = NULL;
6877: /* stuff for MatGetRow() */
6878: b->rowindices = NULL;
6879: b->rowvalues = NULL;
6880: b->getrowactive = PETSC_FALSE;
6882: /* flexible pointer used in CUSPARSE classes */
6883: b->spptr = NULL;
6885: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6886: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6887: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6888: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6890: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6891: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6895: #if defined(PETSC_HAVE_CUDA)
6896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6897: #endif
6898: #if defined(PETSC_HAVE_HIP)
6899: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6900: #endif
6901: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6903: #endif
6904: #if defined(PETSC_HAVE_MKL_SPARSE)
6905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6906: #endif
6907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6910: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6911: #if defined(PETSC_HAVE_ELEMENTAL)
6912: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6913: #endif
6914: #if defined(PETSC_HAVE_SCALAPACK)
6915: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6916: #endif
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6918: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6919: #if defined(PETSC_HAVE_HYPRE)
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6921: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6922: #endif
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6927: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6928: PetscFunctionReturn(PETSC_SUCCESS);
6929: }
6931: /*@C
6932: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6933: and "off-diagonal" part of the matrix in CSR format.
6935: Collective
6937: Input Parameters:
6938: + comm - MPI communicator
6939: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6940: . n - This value should be the same as the local size used in creating the
6941: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6942: calculated if `N` is given) For square matrices `n` is almost always `m`.
6943: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6944: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6945: . 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
6946: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6947: . a - matrix values
6948: . 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
6949: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6950: - oa - matrix values
6952: Output Parameter:
6953: . mat - the matrix
6955: Level: advanced
6957: Notes:
6958: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6959: must free the arrays once the matrix has been destroyed and not before.
6961: The `i` and `j` indices are 0 based
6963: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6965: This sets local rows and cannot be used to set off-processor values.
6967: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6968: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6969: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6970: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6971: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6972: communication if it is known that only local entries will be set.
6974: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6975: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6976: @*/
6977: 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)
6978: {
6979: Mat_MPIAIJ *maij;
6981: PetscFunctionBegin;
6982: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6983: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6984: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6985: PetscCall(MatCreate(comm, mat));
6986: PetscCall(MatSetSizes(*mat, m, n, M, N));
6987: PetscCall(MatSetType(*mat, MATMPIAIJ));
6988: maij = (Mat_MPIAIJ *)(*mat)->data;
6990: (*mat)->preallocated = PETSC_TRUE;
6992: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6993: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6995: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6996: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
6998: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
6999: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7000: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7001: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7002: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7003: PetscFunctionReturn(PETSC_SUCCESS);
7004: }
7006: typedef struct {
7007: Mat *mp; /* intermediate products */
7008: PetscBool *mptmp; /* is the intermediate product temporary ? */
7009: PetscInt cp; /* number of intermediate products */
7011: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7012: PetscInt *startsj_s, *startsj_r;
7013: PetscScalar *bufa;
7014: Mat P_oth;
7016: /* may take advantage of merging product->B */
7017: Mat Bloc; /* B-local by merging diag and off-diag */
7019: /* cusparse does not have support to split between symbolic and numeric phases.
7020: When api_user is true, we don't need to update the numerical values
7021: of the temporary storage */
7022: PetscBool reusesym;
7024: /* support for COO values insertion */
7025: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7026: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7027: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7028: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7029: PetscSF sf; /* used for non-local values insertion and memory malloc */
7030: PetscMemType mtype;
7032: /* customization */
7033: PetscBool abmerge;
7034: PetscBool P_oth_bind;
7035: } MatMatMPIAIJBACKEND;
7037: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7038: {
7039: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7040: PetscInt i;
7042: PetscFunctionBegin;
7043: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7044: PetscCall(PetscFree(mmdata->bufa));
7045: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7046: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7047: PetscCall(MatDestroy(&mmdata->P_oth));
7048: PetscCall(MatDestroy(&mmdata->Bloc));
7049: PetscCall(PetscSFDestroy(&mmdata->sf));
7050: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7051: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7052: PetscCall(PetscFree(mmdata->own[0]));
7053: PetscCall(PetscFree(mmdata->own));
7054: PetscCall(PetscFree(mmdata->off[0]));
7055: PetscCall(PetscFree(mmdata->off));
7056: PetscCall(PetscFree(mmdata));
7057: PetscFunctionReturn(PETSC_SUCCESS);
7058: }
7060: /* Copy selected n entries with indices in idx[] of A to v[].
7061: If idx is NULL, copy the whole data array of A to v[]
7062: */
7063: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7064: {
7065: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7067: PetscFunctionBegin;
7068: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7069: if (f) {
7070: PetscCall((*f)(A, n, idx, v));
7071: } else {
7072: const PetscScalar *vv;
7074: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7075: if (n && idx) {
7076: PetscScalar *w = v;
7077: const PetscInt *oi = idx;
7078: PetscInt j;
7080: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7081: } else {
7082: PetscCall(PetscArraycpy(v, vv, n));
7083: }
7084: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7085: }
7086: PetscFunctionReturn(PETSC_SUCCESS);
7087: }
7089: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7090: {
7091: MatMatMPIAIJBACKEND *mmdata;
7092: PetscInt i, n_d, n_o;
7094: PetscFunctionBegin;
7095: MatCheckProduct(C, 1);
7096: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7097: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7098: if (!mmdata->reusesym) { /* update temporary matrices */
7099: 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));
7100: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7101: }
7102: mmdata->reusesym = PETSC_FALSE;
7104: for (i = 0; i < mmdata->cp; i++) {
7105: 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]);
7106: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7107: }
7108: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7109: PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];
7111: if (mmdata->mptmp[i]) continue;
7112: if (noff) {
7113: PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];
7115: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7116: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7117: n_o += noff;
7118: n_d += nown;
7119: } else {
7120: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7122: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7123: n_d += mm->nz;
7124: }
7125: }
7126: if (mmdata->hasoffproc) { /* offprocess insertion */
7127: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7128: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7129: }
7130: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7131: PetscFunctionReturn(PETSC_SUCCESS);
7132: }
7134: /* Support for Pt * A, A * P, or Pt * A * P */
7135: #define MAX_NUMBER_INTERMEDIATE 4
7136: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7137: {
7138: Mat_Product *product = C->product;
7139: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7140: Mat_MPIAIJ *a, *p;
7141: MatMatMPIAIJBACKEND *mmdata;
7142: ISLocalToGlobalMapping P_oth_l2g = NULL;
7143: IS glob = NULL;
7144: const char *prefix;
7145: char pprefix[256];
7146: const PetscInt *globidx, *P_oth_idx;
7147: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7148: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7149: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7150: /* type-0: consecutive, start from 0; type-1: consecutive with */
7151: /* a base offset; type-2: sparse with a local to global map table */
7152: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7154: MatProductType ptype;
7155: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7156: PetscMPIInt size;
7158: PetscFunctionBegin;
7159: MatCheckProduct(C, 1);
7160: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7161: ptype = product->type;
7162: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7163: ptype = MATPRODUCT_AB;
7164: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7165: }
7166: switch (ptype) {
7167: case MATPRODUCT_AB:
7168: A = product->A;
7169: P = product->B;
7170: m = A->rmap->n;
7171: n = P->cmap->n;
7172: M = A->rmap->N;
7173: N = P->cmap->N;
7174: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7175: break;
7176: case MATPRODUCT_AtB:
7177: P = product->A;
7178: A = product->B;
7179: m = P->cmap->n;
7180: n = A->cmap->n;
7181: M = P->cmap->N;
7182: N = A->cmap->N;
7183: hasoffproc = PETSC_TRUE;
7184: break;
7185: case MATPRODUCT_PtAP:
7186: A = product->A;
7187: P = product->B;
7188: m = P->cmap->n;
7189: n = P->cmap->n;
7190: M = P->cmap->N;
7191: N = P->cmap->N;
7192: hasoffproc = PETSC_TRUE;
7193: break;
7194: default:
7195: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7196: }
7197: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7198: if (size == 1) hasoffproc = PETSC_FALSE;
7200: /* defaults */
7201: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7202: mp[i] = NULL;
7203: mptmp[i] = PETSC_FALSE;
7204: rmapt[i] = -1;
7205: cmapt[i] = -1;
7206: rmapa[i] = NULL;
7207: cmapa[i] = NULL;
7208: }
7210: /* customization */
7211: PetscCall(PetscNew(&mmdata));
7212: mmdata->reusesym = product->api_user;
7213: if (ptype == MATPRODUCT_AB) {
7214: if (product->api_user) {
7215: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7216: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7217: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7218: PetscOptionsEnd();
7219: } else {
7220: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7221: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7222: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7223: PetscOptionsEnd();
7224: }
7225: } else if (ptype == MATPRODUCT_PtAP) {
7226: if (product->api_user) {
7227: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7228: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7229: PetscOptionsEnd();
7230: } else {
7231: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7232: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7233: PetscOptionsEnd();
7234: }
7235: }
7236: a = (Mat_MPIAIJ *)A->data;
7237: p = (Mat_MPIAIJ *)P->data;
7238: PetscCall(MatSetSizes(C, m, n, M, N));
7239: PetscCall(PetscLayoutSetUp(C->rmap));
7240: PetscCall(PetscLayoutSetUp(C->cmap));
7241: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7242: PetscCall(MatGetOptionsPrefix(C, &prefix));
7244: cp = 0;
7245: switch (ptype) {
7246: case MATPRODUCT_AB: /* A * P */
7247: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7249: /* A_diag * P_local (merged or not) */
7250: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7251: /* P is product->B */
7252: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7253: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7254: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7255: PetscCall(MatProductSetFill(mp[cp], product->fill));
7256: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7257: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7258: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7259: mp[cp]->product->api_user = product->api_user;
7260: PetscCall(MatProductSetFromOptions(mp[cp]));
7261: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7262: PetscCall(ISGetIndices(glob, &globidx));
7263: rmapt[cp] = 1;
7264: cmapt[cp] = 2;
7265: cmapa[cp] = globidx;
7266: mptmp[cp] = PETSC_FALSE;
7267: cp++;
7268: } else { /* A_diag * P_diag and A_diag * P_off */
7269: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7270: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7271: PetscCall(MatProductSetFill(mp[cp], product->fill));
7272: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7273: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7274: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7275: mp[cp]->product->api_user = product->api_user;
7276: PetscCall(MatProductSetFromOptions(mp[cp]));
7277: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7278: rmapt[cp] = 1;
7279: cmapt[cp] = 1;
7280: mptmp[cp] = PETSC_FALSE;
7281: cp++;
7282: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7283: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7284: PetscCall(MatProductSetFill(mp[cp], product->fill));
7285: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7286: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7287: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7288: mp[cp]->product->api_user = product->api_user;
7289: PetscCall(MatProductSetFromOptions(mp[cp]));
7290: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7291: rmapt[cp] = 1;
7292: cmapt[cp] = 2;
7293: cmapa[cp] = p->garray;
7294: mptmp[cp] = PETSC_FALSE;
7295: cp++;
7296: }
7298: /* A_off * P_other */
7299: if (mmdata->P_oth) {
7300: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7301: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7302: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7303: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7304: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7305: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7306: PetscCall(MatProductSetFill(mp[cp], product->fill));
7307: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7308: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7309: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7310: mp[cp]->product->api_user = product->api_user;
7311: PetscCall(MatProductSetFromOptions(mp[cp]));
7312: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7313: rmapt[cp] = 1;
7314: cmapt[cp] = 2;
7315: cmapa[cp] = P_oth_idx;
7316: mptmp[cp] = PETSC_FALSE;
7317: cp++;
7318: }
7319: break;
7321: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7322: /* A is product->B */
7323: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7324: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7325: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7326: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7327: PetscCall(MatProductSetFill(mp[cp], product->fill));
7328: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7329: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7330: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7331: mp[cp]->product->api_user = product->api_user;
7332: PetscCall(MatProductSetFromOptions(mp[cp]));
7333: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7334: PetscCall(ISGetIndices(glob, &globidx));
7335: rmapt[cp] = 2;
7336: rmapa[cp] = globidx;
7337: cmapt[cp] = 2;
7338: cmapa[cp] = globidx;
7339: mptmp[cp] = PETSC_FALSE;
7340: cp++;
7341: } else {
7342: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7343: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7344: PetscCall(MatProductSetFill(mp[cp], product->fill));
7345: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7346: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7347: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7348: mp[cp]->product->api_user = product->api_user;
7349: PetscCall(MatProductSetFromOptions(mp[cp]));
7350: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7351: PetscCall(ISGetIndices(glob, &globidx));
7352: rmapt[cp] = 1;
7353: cmapt[cp] = 2;
7354: cmapa[cp] = globidx;
7355: mptmp[cp] = PETSC_FALSE;
7356: cp++;
7357: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7358: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7359: PetscCall(MatProductSetFill(mp[cp], product->fill));
7360: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7361: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7362: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7363: mp[cp]->product->api_user = product->api_user;
7364: PetscCall(MatProductSetFromOptions(mp[cp]));
7365: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7366: rmapt[cp] = 2;
7367: rmapa[cp] = p->garray;
7368: cmapt[cp] = 2;
7369: cmapa[cp] = globidx;
7370: mptmp[cp] = PETSC_FALSE;
7371: cp++;
7372: }
7373: break;
7374: case MATPRODUCT_PtAP:
7375: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7376: /* P is product->B */
7377: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7378: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7379: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7380: PetscCall(MatProductSetFill(mp[cp], product->fill));
7381: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7382: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7383: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7384: mp[cp]->product->api_user = product->api_user;
7385: PetscCall(MatProductSetFromOptions(mp[cp]));
7386: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7387: PetscCall(ISGetIndices(glob, &globidx));
7388: rmapt[cp] = 2;
7389: rmapa[cp] = globidx;
7390: cmapt[cp] = 2;
7391: cmapa[cp] = globidx;
7392: mptmp[cp] = PETSC_FALSE;
7393: cp++;
7394: if (mmdata->P_oth) {
7395: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7396: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7397: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7398: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7399: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7400: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7401: PetscCall(MatProductSetFill(mp[cp], product->fill));
7402: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7403: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7404: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7405: mp[cp]->product->api_user = product->api_user;
7406: PetscCall(MatProductSetFromOptions(mp[cp]));
7407: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7408: mptmp[cp] = PETSC_TRUE;
7409: cp++;
7410: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7411: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7412: PetscCall(MatProductSetFill(mp[cp], product->fill));
7413: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7414: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7415: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7416: mp[cp]->product->api_user = product->api_user;
7417: PetscCall(MatProductSetFromOptions(mp[cp]));
7418: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7419: rmapt[cp] = 2;
7420: rmapa[cp] = globidx;
7421: cmapt[cp] = 2;
7422: cmapa[cp] = P_oth_idx;
7423: mptmp[cp] = PETSC_FALSE;
7424: cp++;
7425: }
7426: break;
7427: default:
7428: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7429: }
7430: /* sanity check */
7431: if (size > 1)
7432: 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);
7434: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7435: for (i = 0; i < cp; i++) {
7436: mmdata->mp[i] = mp[i];
7437: mmdata->mptmp[i] = mptmp[i];
7438: }
7439: mmdata->cp = cp;
7440: C->product->data = mmdata;
7441: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7442: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7444: /* memory type */
7445: mmdata->mtype = PETSC_MEMTYPE_HOST;
7446: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7447: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7448: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7449: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7450: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7451: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7453: /* prepare coo coordinates for values insertion */
7455: /* count total nonzeros of those intermediate seqaij Mats
7456: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7457: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7458: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7459: */
7460: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7461: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7462: if (mptmp[cp]) continue;
7463: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7464: const PetscInt *rmap = rmapa[cp];
7465: const PetscInt mr = mp[cp]->rmap->n;
7466: const PetscInt rs = C->rmap->rstart;
7467: const PetscInt re = C->rmap->rend;
7468: const PetscInt *ii = mm->i;
7469: for (i = 0; i < mr; i++) {
7470: const PetscInt gr = rmap[i];
7471: const PetscInt nz = ii[i + 1] - ii[i];
7472: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7473: else ncoo_oown += nz; /* this row is local */
7474: }
7475: } else ncoo_d += mm->nz;
7476: }
7478: /*
7479: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7481: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7483: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7485: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7486: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7487: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7489: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7490: 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.
7491: */
7492: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7493: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7495: /* gather (i,j) of nonzeros inserted by remote procs */
7496: if (hasoffproc) {
7497: PetscSF msf;
7498: PetscInt ncoo2, *coo_i2, *coo_j2;
7500: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7501: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7502: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7504: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7505: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7506: PetscInt *idxoff = mmdata->off[cp];
7507: PetscInt *idxown = mmdata->own[cp];
7508: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7509: const PetscInt *rmap = rmapa[cp];
7510: const PetscInt *cmap = cmapa[cp];
7511: const PetscInt *ii = mm->i;
7512: PetscInt *coi = coo_i + ncoo_o;
7513: PetscInt *coj = coo_j + ncoo_o;
7514: const PetscInt mr = mp[cp]->rmap->n;
7515: const PetscInt rs = C->rmap->rstart;
7516: const PetscInt re = C->rmap->rend;
7517: const PetscInt cs = C->cmap->rstart;
7518: for (i = 0; i < mr; i++) {
7519: const PetscInt *jj = mm->j + ii[i];
7520: const PetscInt gr = rmap[i];
7521: const PetscInt nz = ii[i + 1] - ii[i];
7522: if (gr < rs || gr >= re) { /* this is an offproc row */
7523: for (j = ii[i]; j < ii[i + 1]; j++) {
7524: *coi++ = gr;
7525: *idxoff++ = j;
7526: }
7527: if (!cmapt[cp]) { /* already global */
7528: for (j = 0; j < nz; j++) *coj++ = jj[j];
7529: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7530: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7531: } else { /* offdiag */
7532: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7533: }
7534: ncoo_o += nz;
7535: } else { /* this is a local row */
7536: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7537: }
7538: }
7539: }
7540: mmdata->off[cp + 1] = idxoff;
7541: mmdata->own[cp + 1] = idxown;
7542: }
7544: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7545: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7546: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7547: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7548: ncoo = ncoo_d + ncoo_oown + ncoo2;
7549: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7550: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7551: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7552: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7553: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7554: PetscCall(PetscFree2(coo_i, coo_j));
7555: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7556: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7557: coo_i = coo_i2;
7558: coo_j = coo_j2;
7559: } else { /* no offproc values insertion */
7560: ncoo = ncoo_d;
7561: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7563: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7564: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7565: PetscCall(PetscSFSetUp(mmdata->sf));
7566: }
7567: mmdata->hasoffproc = hasoffproc;
7569: /* gather (i,j) of nonzeros inserted locally */
7570: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7571: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7572: PetscInt *coi = coo_i + ncoo_d;
7573: PetscInt *coj = coo_j + ncoo_d;
7574: const PetscInt *jj = mm->j;
7575: const PetscInt *ii = mm->i;
7576: const PetscInt *cmap = cmapa[cp];
7577: const PetscInt *rmap = rmapa[cp];
7578: const PetscInt mr = mp[cp]->rmap->n;
7579: const PetscInt rs = C->rmap->rstart;
7580: const PetscInt re = C->rmap->rend;
7581: const PetscInt cs = C->cmap->rstart;
7583: if (mptmp[cp]) continue;
7584: if (rmapt[cp] == 1) { /* consecutive rows */
7585: /* fill coo_i */
7586: for (i = 0; i < mr; i++) {
7587: const PetscInt gr = i + rs;
7588: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7589: }
7590: /* fill coo_j */
7591: if (!cmapt[cp]) { /* type-0, already global */
7592: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7593: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7594: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7595: } else { /* type-2, local to global for sparse columns */
7596: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7597: }
7598: ncoo_d += mm->nz;
7599: } else if (rmapt[cp] == 2) { /* sparse rows */
7600: for (i = 0; i < mr; i++) {
7601: const PetscInt *jj = mm->j + ii[i];
7602: const PetscInt gr = rmap[i];
7603: const PetscInt nz = ii[i + 1] - ii[i];
7604: if (gr >= rs && gr < re) { /* local rows */
7605: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7606: if (!cmapt[cp]) { /* type-0, already global */
7607: for (j = 0; j < nz; j++) *coj++ = jj[j];
7608: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7609: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7610: } else { /* type-2, local to global for sparse columns */
7611: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7612: }
7613: ncoo_d += nz;
7614: }
7615: }
7616: }
7617: }
7618: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7619: PetscCall(ISDestroy(&glob));
7620: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7621: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7622: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7623: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7625: /* preallocate with COO data */
7626: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7627: PetscCall(PetscFree2(coo_i, coo_j));
7628: PetscFunctionReturn(PETSC_SUCCESS);
7629: }
7631: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7632: {
7633: Mat_Product *product = mat->product;
7634: #if defined(PETSC_HAVE_DEVICE)
7635: PetscBool match = PETSC_FALSE;
7636: PetscBool usecpu = PETSC_FALSE;
7637: #else
7638: PetscBool match = PETSC_TRUE;
7639: #endif
7641: PetscFunctionBegin;
7642: MatCheckProduct(mat, 1);
7643: #if defined(PETSC_HAVE_DEVICE)
7644: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7645: if (match) { /* we can always fallback to the CPU if requested */
7646: switch (product->type) {
7647: case MATPRODUCT_AB:
7648: if (product->api_user) {
7649: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7650: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7651: PetscOptionsEnd();
7652: } else {
7653: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7654: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7655: PetscOptionsEnd();
7656: }
7657: break;
7658: case MATPRODUCT_AtB:
7659: if (product->api_user) {
7660: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7661: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7662: PetscOptionsEnd();
7663: } else {
7664: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7665: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7666: PetscOptionsEnd();
7667: }
7668: break;
7669: case MATPRODUCT_PtAP:
7670: if (product->api_user) {
7671: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7672: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7673: PetscOptionsEnd();
7674: } else {
7675: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7676: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7677: PetscOptionsEnd();
7678: }
7679: break;
7680: default:
7681: break;
7682: }
7683: match = (PetscBool)!usecpu;
7684: }
7685: #endif
7686: if (match) {
7687: switch (product->type) {
7688: case MATPRODUCT_AB:
7689: case MATPRODUCT_AtB:
7690: case MATPRODUCT_PtAP:
7691: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7692: break;
7693: default:
7694: break;
7695: }
7696: }
7697: /* fallback to MPIAIJ ops */
7698: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7699: PetscFunctionReturn(PETSC_SUCCESS);
7700: }
7702: /*
7703: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7705: n - the number of block indices in cc[]
7706: cc - the block indices (must be large enough to contain the indices)
7707: */
7708: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7709: {
7710: PetscInt cnt = -1, nidx, j;
7711: const PetscInt *idx;
7713: PetscFunctionBegin;
7714: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7715: if (nidx) {
7716: cnt = 0;
7717: cc[cnt] = idx[0] / bs;
7718: for (j = 1; j < nidx; j++) {
7719: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7720: }
7721: }
7722: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7723: *n = cnt + 1;
7724: PetscFunctionReturn(PETSC_SUCCESS);
7725: }
7727: /*
7728: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7730: ncollapsed - the number of block indices
7731: collapsed - the block indices (must be large enough to contain the indices)
7732: */
7733: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7734: {
7735: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7737: PetscFunctionBegin;
7738: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7739: for (i = start + 1; i < start + bs; i++) {
7740: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7741: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7742: cprevtmp = cprev;
7743: cprev = merged;
7744: merged = cprevtmp;
7745: }
7746: *ncollapsed = nprev;
7747: if (collapsed) *collapsed = cprev;
7748: PetscFunctionReturn(PETSC_SUCCESS);
7749: }
7751: /*
7752: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7754: Input Parameter:
7755: . Amat - matrix
7756: - symmetrize - make the result symmetric
7757: + scale - scale with diagonal
7759: Output Parameter:
7760: . a_Gmat - output scalar graph >= 0
7762: */
7763: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7764: {
7765: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7766: MPI_Comm comm;
7767: Mat Gmat;
7768: PetscBool ismpiaij, isseqaij;
7769: Mat a, b, c;
7770: MatType jtype;
7772: PetscFunctionBegin;
7773: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7774: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7775: PetscCall(MatGetSize(Amat, &MM, &NN));
7776: PetscCall(MatGetBlockSize(Amat, &bs));
7777: nloc = (Iend - Istart) / bs;
7779: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7780: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7781: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7783: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7784: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7785: implementation */
7786: if (bs > 1) {
7787: PetscCall(MatGetType(Amat, &jtype));
7788: PetscCall(MatCreate(comm, &Gmat));
7789: PetscCall(MatSetType(Gmat, jtype));
7790: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7791: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7792: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7793: PetscInt *d_nnz, *o_nnz;
7794: MatScalar *aa, val, *AA;
7795: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7796: if (isseqaij) {
7797: a = Amat;
7798: b = NULL;
7799: } else {
7800: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7801: a = d->A;
7802: b = d->B;
7803: }
7804: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7805: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7806: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7807: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7808: const PetscInt *cols1, *cols2;
7809: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7810: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7811: nnz[brow / bs] = nc2 / bs;
7812: if (nc2 % bs) ok = 0;
7813: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7814: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7815: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7816: if (nc1 != nc2) ok = 0;
7817: else {
7818: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7819: if (cols1[jj] != cols2[jj]) ok = 0;
7820: if (cols1[jj] % bs != jj % bs) ok = 0;
7821: }
7822: }
7823: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7824: }
7825: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7826: if (!ok) {
7827: PetscCall(PetscFree2(d_nnz, o_nnz));
7828: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7829: goto old_bs;
7830: }
7831: }
7832: }
7833: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7834: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7835: PetscCall(PetscFree2(d_nnz, o_nnz));
7836: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7837: // diag
7838: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7839: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7840: ai = aseq->i;
7841: n = ai[brow + 1] - ai[brow];
7842: aj = aseq->j + ai[brow];
7843: for (int k = 0; k < n; k += bs) { // block columns
7844: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7845: val = 0;
7846: if (index_size == 0) {
7847: for (int ii = 0; ii < bs; ii++) { // rows in block
7848: aa = aseq->a + ai[brow + ii] + k;
7849: for (int jj = 0; jj < bs; jj++) { // columns in block
7850: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7851: }
7852: }
7853: } else { // use (index,index) value if provided
7854: for (int iii = 0; iii < index_size; iii++) { // rows in block
7855: int ii = index[iii];
7856: aa = aseq->a + ai[brow + ii] + k;
7857: for (int jjj = 0; jjj < index_size; jjj++) { // columns in block
7858: int jj = index[jjj];
7859: val += PetscAbs(PetscRealPart(aa[jj]));
7860: }
7861: }
7862: }
7863: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7864: AA[k / bs] = val;
7865: }
7866: grow = Istart / bs + brow / bs;
7867: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, INSERT_VALUES));
7868: }
7869: // off-diag
7870: if (ismpiaij) {
7871: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7872: const PetscScalar *vals;
7873: const PetscInt *cols, *garray = aij->garray;
7874: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7875: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7876: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7877: for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7878: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7879: AA[k / bs] = 0;
7880: AJ[cidx] = garray[cols[k]] / bs;
7881: }
7882: nc = ncols / bs;
7883: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7884: if (index_size == 0) {
7885: for (int ii = 0; ii < bs; ii++) { // rows in block
7886: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7887: for (int k = 0; k < ncols; k += bs) {
7888: for (int jj = 0; jj < bs; jj++) { // cols in block
7889: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7890: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7891: }
7892: }
7893: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7894: }
7895: } else { // use (index,index) value if provided
7896: for (int iii = 0; iii < index_size; iii++) { // rows in block
7897: int ii = index[iii];
7898: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7899: for (int k = 0; k < ncols; k += bs) {
7900: for (int jjj = 0; jjj < index_size; jjj++) { // cols in block
7901: int jj = index[jjj];
7902: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7903: }
7904: }
7905: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7906: }
7907: }
7908: grow = Istart / bs + brow / bs;
7909: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, INSERT_VALUES));
7910: }
7911: }
7912: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7913: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7914: PetscCall(PetscFree2(AA, AJ));
7915: } else {
7916: const PetscScalar *vals;
7917: const PetscInt *idx;
7918: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7919: old_bs:
7920: /*
7921: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7922: */
7923: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7924: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7925: if (isseqaij) {
7926: PetscInt max_d_nnz;
7927: /*
7928: Determine exact preallocation count for (sequential) scalar matrix
7929: */
7930: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7931: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7932: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7933: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7934: PetscCall(PetscFree3(w0, w1, w2));
7935: } else if (ismpiaij) {
7936: Mat Daij, Oaij;
7937: const PetscInt *garray;
7938: PetscInt max_d_nnz;
7939: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7940: /*
7941: Determine exact preallocation count for diagonal block portion of scalar matrix
7942: */
7943: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7944: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7945: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7946: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7947: PetscCall(PetscFree3(w0, w1, w2));
7948: /*
7949: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7950: */
7951: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7952: o_nnz[jj] = 0;
7953: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7954: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7955: o_nnz[jj] += ncols;
7956: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7957: }
7958: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7959: }
7960: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7961: /* get scalar copy (norms) of matrix */
7962: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7963: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7964: PetscCall(PetscFree2(d_nnz, o_nnz));
7965: for (Ii = Istart; Ii < Iend; Ii++) {
7966: PetscInt dest_row = Ii / bs;
7967: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7968: for (jj = 0; jj < ncols; jj++) {
7969: PetscInt dest_col = idx[jj] / bs;
7970: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7971: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7972: }
7973: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7974: }
7975: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7976: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7977: }
7978: } else {
7979: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7980: else {
7981: Gmat = Amat;
7982: PetscCall(PetscObjectReference((PetscObject)Gmat));
7983: }
7984: if (isseqaij) {
7985: a = Gmat;
7986: b = NULL;
7987: } else {
7988: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7989: a = d->A;
7990: b = d->B;
7991: }
7992: if (filter >= 0 || scale) {
7993: /* take absolute value of each entry */
7994: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7995: MatInfo info;
7996: PetscScalar *avals;
7997: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
7998: PetscCall(MatSeqAIJGetArray(c, &avals));
7999: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8000: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8001: }
8002: }
8003: }
8004: if (symmetrize) {
8005: PetscBool isset, issym;
8006: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8007: if (!isset || !issym) {
8008: Mat matTrans;
8009: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8010: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8011: PetscCall(MatDestroy(&matTrans));
8012: }
8013: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8014: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8015: if (scale) {
8016: /* scale c for all diagonal values = 1 or -1 */
8017: Vec diag;
8018: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8019: PetscCall(MatGetDiagonal(Gmat, diag));
8020: PetscCall(VecReciprocal(diag));
8021: PetscCall(VecSqrtAbs(diag));
8022: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8023: PetscCall(VecDestroy(&diag));
8024: }
8025: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8027: if (filter >= 0) {
8028: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8029: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8030: }
8031: *a_Gmat = Gmat;
8032: PetscFunctionReturn(PETSC_SUCCESS);
8033: }
8035: /*
8036: Special version for direct calls from Fortran
8037: */
8038: #include <petsc/private/fortranimpl.h>
8040: /* Change these macros so can be used in void function */
8041: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8042: #undef PetscCall
8043: #define PetscCall(...) \
8044: do { \
8045: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8046: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8047: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8048: return; \
8049: } \
8050: } while (0)
8052: #undef SETERRQ
8053: #define SETERRQ(comm, ierr, ...) \
8054: do { \
8055: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8056: return; \
8057: } while (0)
8059: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8060: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8061: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8062: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8063: #else
8064: #endif
8065: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8066: {
8067: Mat mat = *mmat;
8068: PetscInt m = *mm, n = *mn;
8069: InsertMode addv = *maddv;
8070: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8071: PetscScalar value;
8073: MatCheckPreallocated(mat, 1);
8074: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8075: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8076: {
8077: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8078: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8079: PetscBool roworiented = aij->roworiented;
8081: /* Some Variables required in the macro */
8082: Mat A = aij->A;
8083: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8084: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8085: MatScalar *aa;
8086: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8087: Mat B = aij->B;
8088: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8089: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8090: MatScalar *ba;
8091: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8092: * cannot use "#if defined" inside a macro. */
8093: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8095: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8096: PetscInt nonew = a->nonew;
8097: MatScalar *ap1, *ap2;
8099: PetscFunctionBegin;
8100: PetscCall(MatSeqAIJGetArray(A, &aa));
8101: PetscCall(MatSeqAIJGetArray(B, &ba));
8102: for (i = 0; i < m; i++) {
8103: if (im[i] < 0) continue;
8104: 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);
8105: if (im[i] >= rstart && im[i] < rend) {
8106: row = im[i] - rstart;
8107: lastcol1 = -1;
8108: rp1 = aj + ai[row];
8109: ap1 = aa + ai[row];
8110: rmax1 = aimax[row];
8111: nrow1 = ailen[row];
8112: low1 = 0;
8113: high1 = nrow1;
8114: lastcol2 = -1;
8115: rp2 = bj + bi[row];
8116: ap2 = ba + bi[row];
8117: rmax2 = bimax[row];
8118: nrow2 = bilen[row];
8119: low2 = 0;
8120: high2 = nrow2;
8122: for (j = 0; j < n; j++) {
8123: if (roworiented) value = v[i * n + j];
8124: else value = v[i + j * m];
8125: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8126: if (in[j] >= cstart && in[j] < cend) {
8127: col = in[j] - cstart;
8128: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8129: } else if (in[j] < 0) continue;
8130: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8131: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8132: } else {
8133: if (mat->was_assembled) {
8134: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8135: #if defined(PETSC_USE_CTABLE)
8136: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8137: col--;
8138: #else
8139: col = aij->colmap[in[j]] - 1;
8140: #endif
8141: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8142: PetscCall(MatDisAssemble_MPIAIJ(mat));
8143: col = in[j];
8144: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8145: B = aij->B;
8146: b = (Mat_SeqAIJ *)B->data;
8147: bimax = b->imax;
8148: bi = b->i;
8149: bilen = b->ilen;
8150: bj = b->j;
8151: rp2 = bj + bi[row];
8152: ap2 = ba + bi[row];
8153: rmax2 = bimax[row];
8154: nrow2 = bilen[row];
8155: low2 = 0;
8156: high2 = nrow2;
8157: bm = aij->B->rmap->n;
8158: ba = b->a;
8159: inserted = PETSC_FALSE;
8160: }
8161: } else col = in[j];
8162: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8163: }
8164: }
8165: } else if (!aij->donotstash) {
8166: if (roworiented) {
8167: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8168: } else {
8169: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8170: }
8171: }
8172: }
8173: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8174: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8175: }
8176: PetscFunctionReturnVoid();
8177: }
8179: /* Undefining these here since they were redefined from their original definition above! No
8180: * other PETSc functions should be defined past this point, as it is impossible to recover the
8181: * original definitions */
8182: #undef PetscCall
8183: #undef SETERRQ