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 MatMultTransposeAdd_MPIAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1126: {
1127: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1129: PetscFunctionBegin;
1130: /* do nondiagonal part */
1131: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
1132: /* do local part */
1133: PetscCall((*a->A->ops->multtransposeadd)(a->A, xx, yy, zz));
1134: /* add partial results together */
1135: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1136: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, zz, ADD_VALUES, SCATTER_REVERSE));
1137: PetscFunctionReturn(PETSC_SUCCESS);
1138: }
1140: /*
1141: This only works correctly for square matrices where the subblock A->A is the
1142: diagonal block
1143: */
1144: static PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A, Vec v)
1145: {
1146: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1148: PetscFunctionBegin;
1149: PetscCheck(A->rmap->N == A->cmap->N, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Supports only square matrix where A->A is diag block");
1150: PetscCheck(A->rmap->rstart == A->cmap->rstart && A->rmap->rend == A->cmap->rend, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "row partition must equal col partition");
1151: PetscCall(MatGetDiagonal(a->A, v));
1152: PetscFunctionReturn(PETSC_SUCCESS);
1153: }
1155: static PetscErrorCode MatScale_MPIAIJ(Mat A, PetscScalar aa)
1156: {
1157: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1159: PetscFunctionBegin;
1160: PetscCall(MatScale(a->A, aa));
1161: PetscCall(MatScale(a->B, aa));
1162: PetscFunctionReturn(PETSC_SUCCESS);
1163: }
1165: static PetscErrorCode MatView_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
1166: {
1167: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1168: Mat_SeqAIJ *A = (Mat_SeqAIJ *)aij->A->data;
1169: Mat_SeqAIJ *B = (Mat_SeqAIJ *)aij->B->data;
1170: const PetscInt *garray = aij->garray;
1171: const PetscScalar *aa, *ba;
1172: PetscInt header[4], M, N, m, rs, cs, cnt, i, ja, jb;
1173: PetscInt64 nz, hnz;
1174: PetscInt *rowlens;
1175: PetscInt *colidxs;
1176: PetscScalar *matvals;
1177: PetscMPIInt rank;
1179: PetscFunctionBegin;
1180: PetscCall(PetscViewerSetUp(viewer));
1182: M = mat->rmap->N;
1183: N = mat->cmap->N;
1184: m = mat->rmap->n;
1185: rs = mat->rmap->rstart;
1186: cs = mat->cmap->rstart;
1187: nz = A->nz + B->nz;
1189: /* write matrix header */
1190: header[0] = MAT_FILE_CLASSID;
1191: header[1] = M;
1192: header[2] = N;
1193: PetscCallMPI(MPI_Reduce(&nz, &hnz, 1, MPIU_INT64, MPI_SUM, 0, PetscObjectComm((PetscObject)mat)));
1194: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1195: if (rank == 0) {
1196: if (hnz > PETSC_MAX_INT) header[3] = PETSC_MAX_INT;
1197: else header[3] = (PetscInt)hnz;
1198: }
1199: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
1201: /* fill in and store row lengths */
1202: PetscCall(PetscMalloc1(m, &rowlens));
1203: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i] + B->i[i + 1] - B->i[i];
1204: PetscCall(PetscViewerBinaryWriteAll(viewer, rowlens, m, rs, M, PETSC_INT));
1205: PetscCall(PetscFree(rowlens));
1207: /* fill in and store column indices */
1208: PetscCall(PetscMalloc1(nz, &colidxs));
1209: for (cnt = 0, i = 0; i < m; i++) {
1210: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1211: if (garray[B->j[jb]] > cs) break;
1212: colidxs[cnt++] = garray[B->j[jb]];
1213: }
1214: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) colidxs[cnt++] = A->j[ja] + cs;
1215: for (; jb < B->i[i + 1]; jb++) colidxs[cnt++] = garray[B->j[jb]];
1216: }
1217: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1218: PetscCall(PetscViewerBinaryWriteAll(viewer, colidxs, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
1219: PetscCall(PetscFree(colidxs));
1221: /* fill in and store nonzero values */
1222: PetscCall(MatSeqAIJGetArrayRead(aij->A, &aa));
1223: PetscCall(MatSeqAIJGetArrayRead(aij->B, &ba));
1224: PetscCall(PetscMalloc1(nz, &matvals));
1225: for (cnt = 0, i = 0; i < m; i++) {
1226: for (jb = B->i[i]; jb < B->i[i + 1]; jb++) {
1227: if (garray[B->j[jb]] > cs) break;
1228: matvals[cnt++] = ba[jb];
1229: }
1230: for (ja = A->i[i]; ja < A->i[i + 1]; ja++) matvals[cnt++] = aa[ja];
1231: for (; jb < B->i[i + 1]; jb++) matvals[cnt++] = ba[jb];
1232: }
1233: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &aa));
1234: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &ba));
1235: PetscCheck(cnt == nz, PETSC_COMM_SELF, PETSC_ERR_LIB, "Internal PETSc error: cnt = %" PetscInt_FMT " nz = %" PetscInt64_FMT, cnt, nz);
1236: PetscCall(PetscViewerBinaryWriteAll(viewer, matvals, nz, PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
1237: PetscCall(PetscFree(matvals));
1239: /* write block size option to the viewer's .info file */
1240: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
1241: PetscFunctionReturn(PETSC_SUCCESS);
1242: }
1244: #include <petscdraw.h>
1245: static PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat, PetscViewer viewer)
1246: {
1247: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1248: PetscMPIInt rank = aij->rank, size = aij->size;
1249: PetscBool isdraw, iascii, isbinary;
1250: PetscViewer sviewer;
1251: PetscViewerFormat format;
1253: PetscFunctionBegin;
1254: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1255: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1256: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1257: if (iascii) {
1258: PetscCall(PetscViewerGetFormat(viewer, &format));
1259: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1260: PetscInt i, nmax = 0, nmin = PETSC_MAX_INT, navg = 0, *nz, nzlocal = ((Mat_SeqAIJ *)aij->A->data)->nz + ((Mat_SeqAIJ *)aij->B->data)->nz;
1261: PetscCall(PetscMalloc1(size, &nz));
1262: PetscCallMPI(MPI_Allgather(&nzlocal, 1, MPIU_INT, nz, 1, MPIU_INT, PetscObjectComm((PetscObject)mat)));
1263: for (i = 0; i < (PetscInt)size; i++) {
1264: nmax = PetscMax(nmax, nz[i]);
1265: nmin = PetscMin(nmin, nz[i]);
1266: navg += nz[i];
1267: }
1268: PetscCall(PetscFree(nz));
1269: navg = navg / size;
1270: PetscCall(PetscViewerASCIIPrintf(viewer, "Load Balance - Nonzeros: Min %" PetscInt_FMT " avg %" PetscInt_FMT " max %" PetscInt_FMT "\n", nmin, navg, nmax));
1271: PetscFunctionReturn(PETSC_SUCCESS);
1272: }
1273: PetscCall(PetscViewerGetFormat(viewer, &format));
1274: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1275: MatInfo info;
1276: PetscInt *inodes = NULL;
1278: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)mat), &rank));
1279: PetscCall(MatGetInfo(mat, MAT_LOCAL, &info));
1280: PetscCall(MatInodeGetInodeSizes(aij->A, NULL, &inodes, NULL));
1281: PetscCall(PetscViewerASCIIPushSynchronized(viewer));
1282: if (!inodes) {
1283: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " mem %g, not using I-node routines\n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
1284: (double)info.memory));
1285: } else {
1286: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Local rows %" PetscInt_FMT " nz %" PetscInt_FMT " nz alloced %" PetscInt_FMT " mem %g, using I-node routines\n", rank, mat->rmap->n, (PetscInt)info.nz_used, (PetscInt)info.nz_allocated,
1287: (double)info.memory));
1288: }
1289: PetscCall(MatGetInfo(aij->A, MAT_LOCAL, &info));
1290: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] on-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1291: PetscCall(MatGetInfo(aij->B, MAT_LOCAL, &info));
1292: PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] off-diagonal part: nz %" PetscInt_FMT " \n", rank, (PetscInt)info.nz_used));
1293: PetscCall(PetscViewerFlush(viewer));
1294: PetscCall(PetscViewerASCIIPopSynchronized(viewer));
1295: PetscCall(PetscViewerASCIIPrintf(viewer, "Information on VecScatter used in matrix-vector product: \n"));
1296: PetscCall(VecScatterView(aij->Mvctx, viewer));
1297: PetscFunctionReturn(PETSC_SUCCESS);
1298: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1299: PetscInt inodecount, inodelimit, *inodes;
1300: PetscCall(MatInodeGetInodeSizes(aij->A, &inodecount, &inodes, &inodelimit));
1301: if (inodes) {
1302: PetscCall(PetscViewerASCIIPrintf(viewer, "using I-node (on process 0) routines: found %" PetscInt_FMT " nodes, limit used is %" PetscInt_FMT "\n", inodecount, inodelimit));
1303: } else {
1304: PetscCall(PetscViewerASCIIPrintf(viewer, "not using I-node (on process 0) routines\n"));
1305: }
1306: PetscFunctionReturn(PETSC_SUCCESS);
1307: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1308: PetscFunctionReturn(PETSC_SUCCESS);
1309: }
1310: } else if (isbinary) {
1311: if (size == 1) {
1312: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1313: PetscCall(MatView(aij->A, viewer));
1314: } else {
1315: PetscCall(MatView_MPIAIJ_Binary(mat, viewer));
1316: }
1317: PetscFunctionReturn(PETSC_SUCCESS);
1318: } else if (iascii && size == 1) {
1319: PetscCall(PetscObjectSetName((PetscObject)aij->A, ((PetscObject)mat)->name));
1320: PetscCall(MatView(aij->A, viewer));
1321: PetscFunctionReturn(PETSC_SUCCESS);
1322: } else if (isdraw) {
1323: PetscDraw draw;
1324: PetscBool isnull;
1325: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1326: PetscCall(PetscDrawIsNull(draw, &isnull));
1327: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1328: }
1330: { /* assemble the entire matrix onto first processor */
1331: Mat A = NULL, Av;
1332: IS isrow, iscol;
1334: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->rmap->N : 0, 0, 1, &isrow));
1335: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat), rank == 0 ? mat->cmap->N : 0, 0, 1, &iscol));
1336: PetscCall(MatCreateSubMatrix(mat, isrow, iscol, MAT_INITIAL_MATRIX, &A));
1337: PetscCall(MatMPIAIJGetSeqAIJ(A, &Av, NULL, NULL));
1338: /* The commented code uses MatCreateSubMatrices instead */
1339: /*
1340: Mat *AA, A = NULL, Av;
1341: IS isrow,iscol;
1343: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->rmap->N : 0,0,1,&isrow));
1344: PetscCall(ISCreateStride(PetscObjectComm((PetscObject)mat),rank == 0 ? mat->cmap->N : 0,0,1,&iscol));
1345: PetscCall(MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA));
1346: if (rank == 0) {
1347: PetscCall(PetscObjectReference((PetscObject)AA[0]));
1348: A = AA[0];
1349: Av = AA[0];
1350: }
1351: PetscCall(MatDestroySubMatrices(1,&AA));
1352: */
1353: PetscCall(ISDestroy(&iscol));
1354: PetscCall(ISDestroy(&isrow));
1355: /*
1356: Everyone has to call to draw the matrix since the graphics waits are
1357: synchronized across all processors that share the PetscDraw object
1358: */
1359: PetscCall(PetscViewerGetSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1360: if (rank == 0) {
1361: if (((PetscObject)mat)->name) PetscCall(PetscObjectSetName((PetscObject)Av, ((PetscObject)mat)->name));
1362: PetscCall(MatView_SeqAIJ(Av, sviewer));
1363: }
1364: PetscCall(PetscViewerRestoreSubViewer(viewer, PETSC_COMM_SELF, &sviewer));
1365: PetscCall(MatDestroy(&A));
1366: }
1367: PetscFunctionReturn(PETSC_SUCCESS);
1368: }
1370: PetscErrorCode MatView_MPIAIJ(Mat mat, PetscViewer viewer)
1371: {
1372: PetscBool iascii, isdraw, issocket, isbinary;
1374: PetscFunctionBegin;
1375: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1376: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1377: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1378: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERSOCKET, &issocket));
1379: if (iascii || isdraw || isbinary || issocket) PetscCall(MatView_MPIAIJ_ASCIIorDraworSocket(mat, viewer));
1380: PetscFunctionReturn(PETSC_SUCCESS);
1381: }
1383: static PetscErrorCode MatSOR_MPIAIJ(Mat matin, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1384: {
1385: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1386: Vec bb1 = NULL;
1387: PetscBool hasop;
1389: PetscFunctionBegin;
1390: if (flag == SOR_APPLY_UPPER) {
1391: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1392: PetscFunctionReturn(PETSC_SUCCESS);
1393: }
1395: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) PetscCall(VecDuplicate(bb, &bb1));
1397: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1398: if (flag & SOR_ZERO_INITIAL_GUESS) {
1399: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1400: its--;
1401: }
1403: while (its--) {
1404: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1405: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1407: /* update rhs: bb1 = bb - B*x */
1408: PetscCall(VecScale(mat->lvec, -1.0));
1409: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1411: /* local sweep */
1412: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_SYMMETRIC_SWEEP, fshift, lits, 1, xx));
1413: }
1414: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1415: if (flag & SOR_ZERO_INITIAL_GUESS) {
1416: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1417: its--;
1418: }
1419: while (its--) {
1420: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1421: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1423: /* update rhs: bb1 = bb - B*x */
1424: PetscCall(VecScale(mat->lvec, -1.0));
1425: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1427: /* local sweep */
1428: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_FORWARD_SWEEP, fshift, lits, 1, xx));
1429: }
1430: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1431: if (flag & SOR_ZERO_INITIAL_GUESS) {
1432: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, flag, fshift, lits, 1, xx));
1433: its--;
1434: }
1435: while (its--) {
1436: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1437: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1439: /* update rhs: bb1 = bb - B*x */
1440: PetscCall(VecScale(mat->lvec, -1.0));
1441: PetscCall((*mat->B->ops->multadd)(mat->B, mat->lvec, bb, bb1));
1443: /* local sweep */
1444: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, SOR_BACKWARD_SWEEP, fshift, lits, 1, xx));
1445: }
1446: } else if (flag & SOR_EISENSTAT) {
1447: Vec xx1;
1449: PetscCall(VecDuplicate(bb, &xx1));
1450: PetscCall((*mat->A->ops->sor)(mat->A, bb, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP), fshift, lits, 1, xx));
1452: PetscCall(VecScatterBegin(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1453: PetscCall(VecScatterEnd(mat->Mvctx, xx, mat->lvec, INSERT_VALUES, SCATTER_FORWARD));
1454: if (!mat->diag) {
1455: PetscCall(MatCreateVecs(matin, &mat->diag, NULL));
1456: PetscCall(MatGetDiagonal(matin, mat->diag));
1457: }
1458: PetscCall(MatHasOperation(matin, MATOP_MULT_DIAGONAL_BLOCK, &hasop));
1459: if (hasop) {
1460: PetscCall(MatMultDiagonalBlock(matin, xx, bb1));
1461: } else {
1462: PetscCall(VecPointwiseMult(bb1, mat->diag, xx));
1463: }
1464: PetscCall(VecAYPX(bb1, (omega - 2.0) / omega, bb));
1466: PetscCall(MatMultAdd(mat->B, mat->lvec, bb1, bb1));
1468: /* local sweep */
1469: PetscCall((*mat->A->ops->sor)(mat->A, bb1, omega, (MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP), fshift, lits, 1, xx1));
1470: PetscCall(VecAXPY(xx, 1.0, xx1));
1471: PetscCall(VecDestroy(&xx1));
1472: } else SETERRQ(PetscObjectComm((PetscObject)matin), PETSC_ERR_SUP, "Parallel SOR not supported");
1474: PetscCall(VecDestroy(&bb1));
1476: matin->factorerrortype = mat->A->factorerrortype;
1477: PetscFunctionReturn(PETSC_SUCCESS);
1478: }
1480: static PetscErrorCode MatPermute_MPIAIJ(Mat A, IS rowp, IS colp, Mat *B)
1481: {
1482: Mat aA, aB, Aperm;
1483: const PetscInt *rwant, *cwant, *gcols, *ai, *bi, *aj, *bj;
1484: PetscScalar *aa, *ba;
1485: PetscInt i, j, m, n, ng, anz, bnz, *dnnz, *onnz, *tdnnz, *tonnz, *rdest, *cdest, *work, *gcdest;
1486: PetscSF rowsf, sf;
1487: IS parcolp = NULL;
1488: PetscBool done;
1490: PetscFunctionBegin;
1491: PetscCall(MatGetLocalSize(A, &m, &n));
1492: PetscCall(ISGetIndices(rowp, &rwant));
1493: PetscCall(ISGetIndices(colp, &cwant));
1494: PetscCall(PetscMalloc3(PetscMax(m, n), &work, m, &rdest, n, &cdest));
1496: /* Invert row permutation to find out where my rows should go */
1497: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &rowsf));
1498: PetscCall(PetscSFSetGraphLayout(rowsf, A->rmap, A->rmap->n, NULL, PETSC_OWN_POINTER, rwant));
1499: PetscCall(PetscSFSetFromOptions(rowsf));
1500: for (i = 0; i < m; i++) work[i] = A->rmap->rstart + i;
1501: PetscCall(PetscSFReduceBegin(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1502: PetscCall(PetscSFReduceEnd(rowsf, MPIU_INT, work, rdest, MPI_REPLACE));
1504: /* Invert column permutation to find out where my columns should go */
1505: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1506: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, A->cmap->n, NULL, PETSC_OWN_POINTER, cwant));
1507: PetscCall(PetscSFSetFromOptions(sf));
1508: for (i = 0; i < n; i++) work[i] = A->cmap->rstart + i;
1509: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1510: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, work, cdest, MPI_REPLACE));
1511: PetscCall(PetscSFDestroy(&sf));
1513: PetscCall(ISRestoreIndices(rowp, &rwant));
1514: PetscCall(ISRestoreIndices(colp, &cwant));
1515: PetscCall(MatMPIAIJGetSeqAIJ(A, &aA, &aB, &gcols));
1517: /* Find out where my gcols should go */
1518: PetscCall(MatGetSize(aB, NULL, &ng));
1519: PetscCall(PetscMalloc1(ng, &gcdest));
1520: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1521: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, ng, NULL, PETSC_OWN_POINTER, gcols));
1522: PetscCall(PetscSFSetFromOptions(sf));
1523: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1524: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, cdest, gcdest, MPI_REPLACE));
1525: PetscCall(PetscSFDestroy(&sf));
1527: PetscCall(PetscCalloc4(m, &dnnz, m, &onnz, m, &tdnnz, m, &tonnz));
1528: PetscCall(MatGetRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1529: PetscCall(MatGetRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1530: for (i = 0; i < m; i++) {
1531: PetscInt row = rdest[i];
1532: PetscMPIInt rowner;
1533: PetscCall(PetscLayoutFindOwner(A->rmap, row, &rowner));
1534: for (j = ai[i]; j < ai[i + 1]; j++) {
1535: PetscInt col = cdest[aj[j]];
1536: PetscMPIInt cowner;
1537: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner)); /* Could build an index for the columns to eliminate this search */
1538: if (rowner == cowner) dnnz[i]++;
1539: else onnz[i]++;
1540: }
1541: for (j = bi[i]; j < bi[i + 1]; j++) {
1542: PetscInt col = gcdest[bj[j]];
1543: PetscMPIInt cowner;
1544: PetscCall(PetscLayoutFindOwner(A->cmap, col, &cowner));
1545: if (rowner == cowner) dnnz[i]++;
1546: else onnz[i]++;
1547: }
1548: }
1549: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1550: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, dnnz, tdnnz, MPI_REPLACE));
1551: PetscCall(PetscSFBcastBegin(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1552: PetscCall(PetscSFBcastEnd(rowsf, MPIU_INT, onnz, tonnz, MPI_REPLACE));
1553: PetscCall(PetscSFDestroy(&rowsf));
1555: PetscCall(MatCreateAIJ(PetscObjectComm((PetscObject)A), A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N, 0, tdnnz, 0, tonnz, &Aperm));
1556: PetscCall(MatSeqAIJGetArray(aA, &aa));
1557: PetscCall(MatSeqAIJGetArray(aB, &ba));
1558: for (i = 0; i < m; i++) {
1559: PetscInt *acols = dnnz, *bcols = onnz; /* Repurpose now-unneeded arrays */
1560: PetscInt j0, rowlen;
1561: rowlen = ai[i + 1] - ai[i];
1562: for (j0 = j = 0; j < rowlen; j0 = j) { /* rowlen could be larger than number of rows m, so sum in batches */
1563: for (; j < PetscMin(rowlen, j0 + m); j++) acols[j - j0] = cdest[aj[ai[i] + j]];
1564: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, acols, aa + ai[i] + j0, INSERT_VALUES));
1565: }
1566: rowlen = bi[i + 1] - bi[i];
1567: for (j0 = j = 0; j < rowlen; j0 = j) {
1568: for (; j < PetscMin(rowlen, j0 + m); j++) bcols[j - j0] = gcdest[bj[bi[i] + j]];
1569: PetscCall(MatSetValues(Aperm, 1, &rdest[i], j - j0, bcols, ba + bi[i] + j0, INSERT_VALUES));
1570: }
1571: }
1572: PetscCall(MatAssemblyBegin(Aperm, MAT_FINAL_ASSEMBLY));
1573: PetscCall(MatAssemblyEnd(Aperm, MAT_FINAL_ASSEMBLY));
1574: PetscCall(MatRestoreRowIJ(aA, 0, PETSC_FALSE, PETSC_FALSE, &anz, &ai, &aj, &done));
1575: PetscCall(MatRestoreRowIJ(aB, 0, PETSC_FALSE, PETSC_FALSE, &bnz, &bi, &bj, &done));
1576: PetscCall(MatSeqAIJRestoreArray(aA, &aa));
1577: PetscCall(MatSeqAIJRestoreArray(aB, &ba));
1578: PetscCall(PetscFree4(dnnz, onnz, tdnnz, tonnz));
1579: PetscCall(PetscFree3(work, rdest, cdest));
1580: PetscCall(PetscFree(gcdest));
1581: if (parcolp) PetscCall(ISDestroy(&colp));
1582: *B = Aperm;
1583: PetscFunctionReturn(PETSC_SUCCESS);
1584: }
1586: static PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat, PetscInt *nghosts, const PetscInt *ghosts[])
1587: {
1588: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1590: PetscFunctionBegin;
1591: PetscCall(MatGetSize(aij->B, NULL, nghosts));
1592: if (ghosts) *ghosts = aij->garray;
1593: PetscFunctionReturn(PETSC_SUCCESS);
1594: }
1596: static PetscErrorCode MatGetInfo_MPIAIJ(Mat matin, MatInfoType flag, MatInfo *info)
1597: {
1598: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1599: Mat A = mat->A, B = mat->B;
1600: PetscLogDouble isend[5], irecv[5];
1602: PetscFunctionBegin;
1603: info->block_size = 1.0;
1604: PetscCall(MatGetInfo(A, MAT_LOCAL, info));
1606: isend[0] = info->nz_used;
1607: isend[1] = info->nz_allocated;
1608: isend[2] = info->nz_unneeded;
1609: isend[3] = info->memory;
1610: isend[4] = info->mallocs;
1612: PetscCall(MatGetInfo(B, MAT_LOCAL, info));
1614: isend[0] += info->nz_used;
1615: isend[1] += info->nz_allocated;
1616: isend[2] += info->nz_unneeded;
1617: isend[3] += info->memory;
1618: isend[4] += info->mallocs;
1619: if (flag == MAT_LOCAL) {
1620: info->nz_used = isend[0];
1621: info->nz_allocated = isend[1];
1622: info->nz_unneeded = isend[2];
1623: info->memory = isend[3];
1624: info->mallocs = isend[4];
1625: } else if (flag == MAT_GLOBAL_MAX) {
1626: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_MAX, PetscObjectComm((PetscObject)matin)));
1628: info->nz_used = irecv[0];
1629: info->nz_allocated = irecv[1];
1630: info->nz_unneeded = irecv[2];
1631: info->memory = irecv[3];
1632: info->mallocs = irecv[4];
1633: } else if (flag == MAT_GLOBAL_SUM) {
1634: PetscCall(MPIU_Allreduce(isend, irecv, 5, MPIU_PETSCLOGDOUBLE, MPI_SUM, PetscObjectComm((PetscObject)matin)));
1636: info->nz_used = irecv[0];
1637: info->nz_allocated = irecv[1];
1638: info->nz_unneeded = irecv[2];
1639: info->memory = irecv[3];
1640: info->mallocs = irecv[4];
1641: }
1642: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1643: info->fill_ratio_needed = 0;
1644: info->factor_mallocs = 0;
1645: PetscFunctionReturn(PETSC_SUCCESS);
1646: }
1648: PetscErrorCode MatSetOption_MPIAIJ(Mat A, MatOption op, PetscBool flg)
1649: {
1650: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
1652: PetscFunctionBegin;
1653: switch (op) {
1654: case MAT_NEW_NONZERO_LOCATIONS:
1655: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1656: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1657: case MAT_KEEP_NONZERO_PATTERN:
1658: case MAT_NEW_NONZERO_LOCATION_ERR:
1659: case MAT_USE_INODES:
1660: case MAT_IGNORE_ZERO_ENTRIES:
1661: case MAT_FORM_EXPLICIT_TRANSPOSE:
1662: MatCheckPreallocated(A, 1);
1663: PetscCall(MatSetOption(a->A, op, flg));
1664: PetscCall(MatSetOption(a->B, op, flg));
1665: break;
1666: case MAT_ROW_ORIENTED:
1667: MatCheckPreallocated(A, 1);
1668: a->roworiented = flg;
1670: PetscCall(MatSetOption(a->A, op, flg));
1671: PetscCall(MatSetOption(a->B, op, flg));
1672: break;
1673: case MAT_FORCE_DIAGONAL_ENTRIES:
1674: case MAT_SORTED_FULL:
1675: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1676: break;
1677: case MAT_IGNORE_OFF_PROC_ENTRIES:
1678: a->donotstash = flg;
1679: break;
1680: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1681: case MAT_SPD:
1682: case MAT_SYMMETRIC:
1683: case MAT_STRUCTURALLY_SYMMETRIC:
1684: case MAT_HERMITIAN:
1685: case MAT_SYMMETRY_ETERNAL:
1686: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1687: case MAT_SPD_ETERNAL:
1688: /* if the diagonal matrix is square it inherits some of the properties above */
1689: break;
1690: case MAT_SUBMAT_SINGLEIS:
1691: A->submat_singleis = flg;
1692: break;
1693: case MAT_STRUCTURE_ONLY:
1694: /* The option is handled directly by MatSetOption() */
1695: break;
1696: default:
1697: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1698: }
1699: PetscFunctionReturn(PETSC_SUCCESS);
1700: }
1702: PetscErrorCode MatGetRow_MPIAIJ(Mat matin, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1703: {
1704: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)matin->data;
1705: PetscScalar *vworkA, *vworkB, **pvA, **pvB, *v_p;
1706: PetscInt i, *cworkA, *cworkB, **pcA, **pcB, cstart = matin->cmap->rstart;
1707: PetscInt nztot, nzA, nzB, lrow, rstart = matin->rmap->rstart, rend = matin->rmap->rend;
1708: PetscInt *cmap, *idx_p;
1710: PetscFunctionBegin;
1711: PetscCheck(!mat->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Already active");
1712: mat->getrowactive = PETSC_TRUE;
1714: if (!mat->rowvalues && (idx || v)) {
1715: /*
1716: allocate enough space to hold information from the longest row.
1717: */
1718: Mat_SeqAIJ *Aa = (Mat_SeqAIJ *)mat->A->data, *Ba = (Mat_SeqAIJ *)mat->B->data;
1719: PetscInt max = 1, tmp;
1720: for (i = 0; i < matin->rmap->n; i++) {
1721: tmp = Aa->i[i + 1] - Aa->i[i] + Ba->i[i + 1] - Ba->i[i];
1722: if (max < tmp) max = tmp;
1723: }
1724: PetscCall(PetscMalloc2(max, &mat->rowvalues, max, &mat->rowindices));
1725: }
1727: PetscCheck(row >= rstart && row < rend, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Only local rows");
1728: lrow = row - rstart;
1730: pvA = &vworkA;
1731: pcA = &cworkA;
1732: pvB = &vworkB;
1733: pcB = &cworkB;
1734: if (!v) {
1735: pvA = NULL;
1736: pvB = NULL;
1737: }
1738: if (!idx) {
1739: pcA = NULL;
1740: if (!v) pcB = NULL;
1741: }
1742: PetscCall((*mat->A->ops->getrow)(mat->A, lrow, &nzA, pcA, pvA));
1743: PetscCall((*mat->B->ops->getrow)(mat->B, lrow, &nzB, pcB, pvB));
1744: nztot = nzA + nzB;
1746: cmap = mat->garray;
1747: if (v || idx) {
1748: if (nztot) {
1749: /* Sort by increasing column numbers, assuming A and B already sorted */
1750: PetscInt imark = -1;
1751: if (v) {
1752: *v = v_p = mat->rowvalues;
1753: for (i = 0; i < nzB; i++) {
1754: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1755: else break;
1756: }
1757: imark = i;
1758: for (i = 0; i < nzA; i++) v_p[imark + i] = vworkA[i];
1759: for (i = imark; i < nzB; i++) v_p[nzA + i] = vworkB[i];
1760: }
1761: if (idx) {
1762: *idx = idx_p = mat->rowindices;
1763: if (imark > -1) {
1764: for (i = 0; i < imark; i++) idx_p[i] = cmap[cworkB[i]];
1765: } else {
1766: for (i = 0; i < nzB; i++) {
1767: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1768: else break;
1769: }
1770: imark = i;
1771: }
1772: for (i = 0; i < nzA; i++) idx_p[imark + i] = cstart + cworkA[i];
1773: for (i = imark; i < nzB; i++) idx_p[nzA + i] = cmap[cworkB[i]];
1774: }
1775: } else {
1776: if (idx) *idx = NULL;
1777: if (v) *v = NULL;
1778: }
1779: }
1780: *nz = nztot;
1781: PetscCall((*mat->A->ops->restorerow)(mat->A, lrow, &nzA, pcA, pvA));
1782: PetscCall((*mat->B->ops->restorerow)(mat->B, lrow, &nzB, pcB, pvB));
1783: PetscFunctionReturn(PETSC_SUCCESS);
1784: }
1786: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
1787: {
1788: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1790: PetscFunctionBegin;
1791: PetscCheck(aij->getrowactive, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "MatGetRow() must be called first");
1792: aij->getrowactive = PETSC_FALSE;
1793: PetscFunctionReturn(PETSC_SUCCESS);
1794: }
1796: static PetscErrorCode MatNorm_MPIAIJ(Mat mat, NormType type, PetscReal *norm)
1797: {
1798: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1799: Mat_SeqAIJ *amat = (Mat_SeqAIJ *)aij->A->data, *bmat = (Mat_SeqAIJ *)aij->B->data;
1800: PetscInt i, j, cstart = mat->cmap->rstart;
1801: PetscReal sum = 0.0;
1802: const MatScalar *v, *amata, *bmata;
1804: PetscFunctionBegin;
1805: if (aij->size == 1) {
1806: PetscCall(MatNorm(aij->A, type, norm));
1807: } else {
1808: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1809: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1810: if (type == NORM_FROBENIUS) {
1811: v = amata;
1812: for (i = 0; i < amat->nz; i++) {
1813: sum += PetscRealPart(PetscConj(*v) * (*v));
1814: v++;
1815: }
1816: v = bmata;
1817: for (i = 0; i < bmat->nz; i++) {
1818: sum += PetscRealPart(PetscConj(*v) * (*v));
1819: v++;
1820: }
1821: PetscCall(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1822: *norm = PetscSqrtReal(*norm);
1823: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1824: } else if (type == NORM_1) { /* max column norm */
1825: PetscReal *tmp, *tmp2;
1826: PetscInt *jj, *garray = aij->garray;
1827: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1828: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1829: *norm = 0.0;
1830: v = amata;
1831: jj = amat->j;
1832: for (j = 0; j < amat->nz; j++) {
1833: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1834: v++;
1835: }
1836: v = bmata;
1837: jj = bmat->j;
1838: for (j = 0; j < bmat->nz; j++) {
1839: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1840: v++;
1841: }
1842: PetscCall(MPIU_Allreduce(tmp, tmp2, mat->cmap->N, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1843: for (j = 0; j < mat->cmap->N; j++) {
1844: if (tmp2[j] > *norm) *norm = tmp2[j];
1845: }
1846: PetscCall(PetscFree(tmp));
1847: PetscCall(PetscFree(tmp2));
1848: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1849: } else if (type == NORM_INFINITY) { /* max row norm */
1850: PetscReal ntemp = 0.0;
1851: for (j = 0; j < aij->A->rmap->n; j++) {
1852: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1853: sum = 0.0;
1854: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1855: sum += PetscAbsScalar(*v);
1856: v++;
1857: }
1858: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1859: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1860: sum += PetscAbsScalar(*v);
1861: v++;
1862: }
1863: if (sum > ntemp) ntemp = sum;
1864: }
1865: PetscCall(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1866: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1867: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1868: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1869: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1870: }
1871: PetscFunctionReturn(PETSC_SUCCESS);
1872: }
1874: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1875: {
1876: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1877: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1878: PetscInt M = A->rmap->N, N = A->cmap->N, ma, na, mb, nb, row, *cols, *cols_tmp, *B_diag_ilen, i, ncol, A_diag_ncol;
1879: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1880: Mat B, A_diag, *B_diag;
1881: const MatScalar *pbv, *bv;
1883: PetscFunctionBegin;
1884: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1885: ma = A->rmap->n;
1886: na = A->cmap->n;
1887: mb = a->B->rmap->n;
1888: nb = a->B->cmap->n;
1889: ai = Aloc->i;
1890: aj = Aloc->j;
1891: bi = Bloc->i;
1892: bj = Bloc->j;
1893: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1894: PetscInt *d_nnz, *g_nnz, *o_nnz;
1895: PetscSFNode *oloc;
1896: PETSC_UNUSED PetscSF sf;
1898: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1899: /* compute d_nnz for preallocation */
1900: PetscCall(PetscArrayzero(d_nnz, na));
1901: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1902: /* compute local off-diagonal contributions */
1903: PetscCall(PetscArrayzero(g_nnz, nb));
1904: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1905: /* map those to global */
1906: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1907: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1908: PetscCall(PetscSFSetFromOptions(sf));
1909: PetscCall(PetscArrayzero(o_nnz, na));
1910: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1911: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1912: PetscCall(PetscSFDestroy(&sf));
1914: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1915: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1916: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1917: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1918: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1919: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1920: } else {
1921: B = *matout;
1922: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1923: }
1925: b = (Mat_MPIAIJ *)B->data;
1926: A_diag = a->A;
1927: B_diag = &b->A;
1928: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1929: A_diag_ncol = A_diag->cmap->N;
1930: B_diag_ilen = sub_B_diag->ilen;
1931: B_diag_i = sub_B_diag->i;
1933: /* Set ilen for diagonal of B */
1934: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1936: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1937: very quickly (=without using MatSetValues), because all writes are local. */
1938: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1939: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1941: /* copy over the B part */
1942: PetscCall(PetscMalloc1(bi[mb], &cols));
1943: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1944: pbv = bv;
1945: row = A->rmap->rstart;
1946: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1947: cols_tmp = cols;
1948: for (i = 0; i < mb; i++) {
1949: ncol = bi[i + 1] - bi[i];
1950: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1951: row++;
1952: if (pbv) pbv += ncol;
1953: if (cols_tmp) cols_tmp += ncol;
1954: }
1955: PetscCall(PetscFree(cols));
1956: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1958: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1959: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1960: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1961: *matout = B;
1962: } else {
1963: PetscCall(MatHeaderMerge(A, &B));
1964: }
1965: PetscFunctionReturn(PETSC_SUCCESS);
1966: }
1968: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1969: {
1970: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1971: Mat a = aij->A, b = aij->B;
1972: PetscInt s1, s2, s3;
1974: PetscFunctionBegin;
1975: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1976: if (rr) {
1977: PetscCall(VecGetLocalSize(rr, &s1));
1978: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1979: /* Overlap communication with computation. */
1980: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1981: }
1982: if (ll) {
1983: PetscCall(VecGetLocalSize(ll, &s1));
1984: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1985: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1986: }
1987: /* scale the diagonal block */
1988: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1990: if (rr) {
1991: /* Do a scatter end and then right scale the off-diagonal block */
1992: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1993: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1994: }
1995: PetscFunctionReturn(PETSC_SUCCESS);
1996: }
1998: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1999: {
2000: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2002: PetscFunctionBegin;
2003: PetscCall(MatSetUnfactored(a->A));
2004: PetscFunctionReturn(PETSC_SUCCESS);
2005: }
2007: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2008: {
2009: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2010: Mat a, b, c, d;
2011: PetscBool flg;
2013: PetscFunctionBegin;
2014: a = matA->A;
2015: b = matA->B;
2016: c = matB->A;
2017: d = matB->B;
2019: PetscCall(MatEqual(a, c, &flg));
2020: if (flg) PetscCall(MatEqual(b, d, &flg));
2021: PetscCall(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2022: PetscFunctionReturn(PETSC_SUCCESS);
2023: }
2025: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2026: {
2027: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2028: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2030: PetscFunctionBegin;
2031: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2032: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2033: /* because of the column compression in the off-processor part of the matrix a->B,
2034: the number of columns in a->B and b->B may be different, hence we cannot call
2035: the MatCopy() directly on the two parts. If need be, we can provide a more
2036: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2037: then copying the submatrices */
2038: PetscCall(MatCopy_Basic(A, B, str));
2039: } else {
2040: PetscCall(MatCopy(a->A, b->A, str));
2041: PetscCall(MatCopy(a->B, b->B, str));
2042: }
2043: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2044: PetscFunctionReturn(PETSC_SUCCESS);
2045: }
2047: /*
2048: Computes the number of nonzeros per row needed for preallocation when X and Y
2049: have different nonzero structure.
2050: */
2051: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *xltog, const PetscInt *yi, const PetscInt *yj, const PetscInt *yltog, PetscInt *nnz)
2052: {
2053: PetscInt i, j, k, nzx, nzy;
2055: PetscFunctionBegin;
2056: /* Set the number of nonzeros in the new matrix */
2057: for (i = 0; i < m; i++) {
2058: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2059: nzx = xi[i + 1] - xi[i];
2060: nzy = yi[i + 1] - yi[i];
2061: nnz[i] = 0;
2062: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2063: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2064: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2065: nnz[i]++;
2066: }
2067: for (; k < nzy; k++) nnz[i]++;
2068: }
2069: PetscFunctionReturn(PETSC_SUCCESS);
2070: }
2072: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2073: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2074: {
2075: PetscInt m = Y->rmap->N;
2076: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2077: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2079: PetscFunctionBegin;
2080: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2081: PetscFunctionReturn(PETSC_SUCCESS);
2082: }
2084: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2085: {
2086: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2088: PetscFunctionBegin;
2089: if (str == SAME_NONZERO_PATTERN) {
2090: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2091: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2092: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2093: PetscCall(MatAXPY_Basic(Y, a, X, str));
2094: } else {
2095: Mat B;
2096: PetscInt *nnz_d, *nnz_o;
2098: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2099: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2100: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2101: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2102: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2103: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2104: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2105: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2106: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2107: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2108: PetscCall(MatHeaderMerge(Y, &B));
2109: PetscCall(PetscFree(nnz_d));
2110: PetscCall(PetscFree(nnz_o));
2111: }
2112: PetscFunctionReturn(PETSC_SUCCESS);
2113: }
2115: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2117: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2118: {
2119: PetscFunctionBegin;
2120: if (PetscDefined(USE_COMPLEX)) {
2121: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2123: PetscCall(MatConjugate_SeqAIJ(aij->A));
2124: PetscCall(MatConjugate_SeqAIJ(aij->B));
2125: }
2126: PetscFunctionReturn(PETSC_SUCCESS);
2127: }
2129: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2130: {
2131: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2133: PetscFunctionBegin;
2134: PetscCall(MatRealPart(a->A));
2135: PetscCall(MatRealPart(a->B));
2136: PetscFunctionReturn(PETSC_SUCCESS);
2137: }
2139: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2140: {
2141: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2143: PetscFunctionBegin;
2144: PetscCall(MatImaginaryPart(a->A));
2145: PetscCall(MatImaginaryPart(a->B));
2146: PetscFunctionReturn(PETSC_SUCCESS);
2147: }
2149: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2150: {
2151: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2152: PetscInt i, *idxb = NULL, m = A->rmap->n;
2153: PetscScalar *va, *vv;
2154: Vec vB, vA;
2155: const PetscScalar *vb;
2157: PetscFunctionBegin;
2158: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2159: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2161: PetscCall(VecGetArrayWrite(vA, &va));
2162: if (idx) {
2163: for (i = 0; i < m; i++) {
2164: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2165: }
2166: }
2168: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2169: PetscCall(PetscMalloc1(m, &idxb));
2170: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2172: PetscCall(VecGetArrayWrite(v, &vv));
2173: PetscCall(VecGetArrayRead(vB, &vb));
2174: for (i = 0; i < m; i++) {
2175: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2176: vv[i] = vb[i];
2177: if (idx) idx[i] = a->garray[idxb[i]];
2178: } else {
2179: vv[i] = va[i];
2180: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2181: }
2182: }
2183: PetscCall(VecRestoreArrayWrite(vA, &vv));
2184: PetscCall(VecRestoreArrayWrite(vA, &va));
2185: PetscCall(VecRestoreArrayRead(vB, &vb));
2186: PetscCall(PetscFree(idxb));
2187: PetscCall(VecDestroy(&vA));
2188: PetscCall(VecDestroy(&vB));
2189: PetscFunctionReturn(PETSC_SUCCESS);
2190: }
2192: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2193: {
2194: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2195: Vec vB, vA;
2197: PetscFunctionBegin;
2198: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2199: PetscCall(MatGetRowSumAbs(a->A, vA));
2200: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2201: PetscCall(MatGetRowSumAbs(a->B, vB));
2202: PetscCall(VecAXPY(vA, 1.0, vB));
2203: PetscCall(VecDestroy(&vB));
2204: PetscCall(VecCopy(vA, v));
2205: PetscCall(VecDestroy(&vA));
2206: PetscFunctionReturn(PETSC_SUCCESS);
2207: }
2209: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2210: {
2211: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2212: PetscInt m = A->rmap->n, n = A->cmap->n;
2213: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2214: PetscInt *cmap = mat->garray;
2215: PetscInt *diagIdx, *offdiagIdx;
2216: Vec diagV, offdiagV;
2217: PetscScalar *a, *diagA, *offdiagA;
2218: const PetscScalar *ba, *bav;
2219: PetscInt r, j, col, ncols, *bi, *bj;
2220: Mat B = mat->B;
2221: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2223: PetscFunctionBegin;
2224: /* When a process holds entire A and other processes have no entry */
2225: if (A->cmap->N == n) {
2226: PetscCall(VecGetArrayWrite(v, &diagA));
2227: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2228: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2229: PetscCall(VecDestroy(&diagV));
2230: PetscCall(VecRestoreArrayWrite(v, &diagA));
2231: PetscFunctionReturn(PETSC_SUCCESS);
2232: } else if (n == 0) {
2233: if (m) {
2234: PetscCall(VecGetArrayWrite(v, &a));
2235: for (r = 0; r < m; r++) {
2236: a[r] = 0.0;
2237: if (idx) idx[r] = -1;
2238: }
2239: PetscCall(VecRestoreArrayWrite(v, &a));
2240: }
2241: PetscFunctionReturn(PETSC_SUCCESS);
2242: }
2244: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2245: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2246: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2247: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2249: /* Get offdiagIdx[] for implicit 0.0 */
2250: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2251: ba = bav;
2252: bi = b->i;
2253: bj = b->j;
2254: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2255: for (r = 0; r < m; r++) {
2256: ncols = bi[r + 1] - bi[r];
2257: if (ncols == A->cmap->N - n) { /* Brow is dense */
2258: offdiagA[r] = *ba;
2259: offdiagIdx[r] = cmap[0];
2260: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2261: offdiagA[r] = 0.0;
2263: /* Find first hole in the cmap */
2264: for (j = 0; j < ncols; j++) {
2265: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2266: if (col > j && j < cstart) {
2267: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2268: break;
2269: } else if (col > j + n && j >= cstart) {
2270: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2271: break;
2272: }
2273: }
2274: if (j == ncols && ncols < A->cmap->N - n) {
2275: /* a hole is outside compressed Bcols */
2276: if (ncols == 0) {
2277: if (cstart) {
2278: offdiagIdx[r] = 0;
2279: } else offdiagIdx[r] = cend;
2280: } else { /* ncols > 0 */
2281: offdiagIdx[r] = cmap[ncols - 1] + 1;
2282: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2283: }
2284: }
2285: }
2287: for (j = 0; j < ncols; j++) {
2288: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2289: offdiagA[r] = *ba;
2290: offdiagIdx[r] = cmap[*bj];
2291: }
2292: ba++;
2293: bj++;
2294: }
2295: }
2297: PetscCall(VecGetArrayWrite(v, &a));
2298: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2299: for (r = 0; r < m; ++r) {
2300: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2301: a[r] = diagA[r];
2302: if (idx) idx[r] = cstart + diagIdx[r];
2303: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2304: a[r] = diagA[r];
2305: if (idx) {
2306: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2307: idx[r] = cstart + diagIdx[r];
2308: } else idx[r] = offdiagIdx[r];
2309: }
2310: } else {
2311: a[r] = offdiagA[r];
2312: if (idx) idx[r] = offdiagIdx[r];
2313: }
2314: }
2315: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2316: PetscCall(VecRestoreArrayWrite(v, &a));
2317: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2318: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2319: PetscCall(VecDestroy(&diagV));
2320: PetscCall(VecDestroy(&offdiagV));
2321: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2322: PetscFunctionReturn(PETSC_SUCCESS);
2323: }
2325: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2326: {
2327: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2328: PetscInt m = A->rmap->n, n = A->cmap->n;
2329: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2330: PetscInt *cmap = mat->garray;
2331: PetscInt *diagIdx, *offdiagIdx;
2332: Vec diagV, offdiagV;
2333: PetscScalar *a, *diagA, *offdiagA;
2334: const PetscScalar *ba, *bav;
2335: PetscInt r, j, col, ncols, *bi, *bj;
2336: Mat B = mat->B;
2337: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2339: PetscFunctionBegin;
2340: /* When a process holds entire A and other processes have no entry */
2341: if (A->cmap->N == n) {
2342: PetscCall(VecGetArrayWrite(v, &diagA));
2343: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2344: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2345: PetscCall(VecDestroy(&diagV));
2346: PetscCall(VecRestoreArrayWrite(v, &diagA));
2347: PetscFunctionReturn(PETSC_SUCCESS);
2348: } else if (n == 0) {
2349: if (m) {
2350: PetscCall(VecGetArrayWrite(v, &a));
2351: for (r = 0; r < m; r++) {
2352: a[r] = PETSC_MAX_REAL;
2353: if (idx) idx[r] = -1;
2354: }
2355: PetscCall(VecRestoreArrayWrite(v, &a));
2356: }
2357: PetscFunctionReturn(PETSC_SUCCESS);
2358: }
2360: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2361: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2362: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2363: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2365: /* Get offdiagIdx[] for implicit 0.0 */
2366: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2367: ba = bav;
2368: bi = b->i;
2369: bj = b->j;
2370: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2371: for (r = 0; r < m; r++) {
2372: ncols = bi[r + 1] - bi[r];
2373: if (ncols == A->cmap->N - n) { /* Brow is dense */
2374: offdiagA[r] = *ba;
2375: offdiagIdx[r] = cmap[0];
2376: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2377: offdiagA[r] = 0.0;
2379: /* Find first hole in the cmap */
2380: for (j = 0; j < ncols; j++) {
2381: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2382: if (col > j && j < cstart) {
2383: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2384: break;
2385: } else if (col > j + n && j >= cstart) {
2386: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2387: break;
2388: }
2389: }
2390: if (j == ncols && ncols < A->cmap->N - n) {
2391: /* a hole is outside compressed Bcols */
2392: if (ncols == 0) {
2393: if (cstart) {
2394: offdiagIdx[r] = 0;
2395: } else offdiagIdx[r] = cend;
2396: } else { /* ncols > 0 */
2397: offdiagIdx[r] = cmap[ncols - 1] + 1;
2398: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2399: }
2400: }
2401: }
2403: for (j = 0; j < ncols; j++) {
2404: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2405: offdiagA[r] = *ba;
2406: offdiagIdx[r] = cmap[*bj];
2407: }
2408: ba++;
2409: bj++;
2410: }
2411: }
2413: PetscCall(VecGetArrayWrite(v, &a));
2414: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2415: for (r = 0; r < m; ++r) {
2416: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2417: a[r] = diagA[r];
2418: if (idx) idx[r] = cstart + diagIdx[r];
2419: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2420: a[r] = diagA[r];
2421: if (idx) {
2422: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2423: idx[r] = cstart + diagIdx[r];
2424: } else idx[r] = offdiagIdx[r];
2425: }
2426: } else {
2427: a[r] = offdiagA[r];
2428: if (idx) idx[r] = offdiagIdx[r];
2429: }
2430: }
2431: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2432: PetscCall(VecRestoreArrayWrite(v, &a));
2433: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2434: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2435: PetscCall(VecDestroy(&diagV));
2436: PetscCall(VecDestroy(&offdiagV));
2437: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2438: PetscFunctionReturn(PETSC_SUCCESS);
2439: }
2441: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2442: {
2443: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2444: PetscInt m = A->rmap->n, n = A->cmap->n;
2445: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2446: PetscInt *cmap = mat->garray;
2447: PetscInt *diagIdx, *offdiagIdx;
2448: Vec diagV, offdiagV;
2449: PetscScalar *a, *diagA, *offdiagA;
2450: const PetscScalar *ba, *bav;
2451: PetscInt r, j, col, ncols, *bi, *bj;
2452: Mat B = mat->B;
2453: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2455: PetscFunctionBegin;
2456: /* When a process holds entire A and other processes have no entry */
2457: if (A->cmap->N == n) {
2458: PetscCall(VecGetArrayWrite(v, &diagA));
2459: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2460: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2461: PetscCall(VecDestroy(&diagV));
2462: PetscCall(VecRestoreArrayWrite(v, &diagA));
2463: PetscFunctionReturn(PETSC_SUCCESS);
2464: } else if (n == 0) {
2465: if (m) {
2466: PetscCall(VecGetArrayWrite(v, &a));
2467: for (r = 0; r < m; r++) {
2468: a[r] = PETSC_MIN_REAL;
2469: if (idx) idx[r] = -1;
2470: }
2471: PetscCall(VecRestoreArrayWrite(v, &a));
2472: }
2473: PetscFunctionReturn(PETSC_SUCCESS);
2474: }
2476: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2477: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2478: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2479: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2481: /* Get offdiagIdx[] for implicit 0.0 */
2482: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2483: ba = bav;
2484: bi = b->i;
2485: bj = b->j;
2486: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2487: for (r = 0; r < m; r++) {
2488: ncols = bi[r + 1] - bi[r];
2489: if (ncols == A->cmap->N - n) { /* Brow is dense */
2490: offdiagA[r] = *ba;
2491: offdiagIdx[r] = cmap[0];
2492: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2493: offdiagA[r] = 0.0;
2495: /* Find first hole in the cmap */
2496: for (j = 0; j < ncols; j++) {
2497: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2498: if (col > j && j < cstart) {
2499: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2500: break;
2501: } else if (col > j + n && j >= cstart) {
2502: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2503: break;
2504: }
2505: }
2506: if (j == ncols && ncols < A->cmap->N - n) {
2507: /* a hole is outside compressed Bcols */
2508: if (ncols == 0) {
2509: if (cstart) {
2510: offdiagIdx[r] = 0;
2511: } else offdiagIdx[r] = cend;
2512: } else { /* ncols > 0 */
2513: offdiagIdx[r] = cmap[ncols - 1] + 1;
2514: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2515: }
2516: }
2517: }
2519: for (j = 0; j < ncols; j++) {
2520: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2521: offdiagA[r] = *ba;
2522: offdiagIdx[r] = cmap[*bj];
2523: }
2524: ba++;
2525: bj++;
2526: }
2527: }
2529: PetscCall(VecGetArrayWrite(v, &a));
2530: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2531: for (r = 0; r < m; ++r) {
2532: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2533: a[r] = diagA[r];
2534: if (idx) idx[r] = cstart + diagIdx[r];
2535: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2536: a[r] = diagA[r];
2537: if (idx) {
2538: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2539: idx[r] = cstart + diagIdx[r];
2540: } else idx[r] = offdiagIdx[r];
2541: }
2542: } else {
2543: a[r] = offdiagA[r];
2544: if (idx) idx[r] = offdiagIdx[r];
2545: }
2546: }
2547: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2548: PetscCall(VecRestoreArrayWrite(v, &a));
2549: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2550: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2551: PetscCall(VecDestroy(&diagV));
2552: PetscCall(VecDestroy(&offdiagV));
2553: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2554: PetscFunctionReturn(PETSC_SUCCESS);
2555: }
2557: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2558: {
2559: Mat *dummy;
2561: PetscFunctionBegin;
2562: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2563: *newmat = *dummy;
2564: PetscCall(PetscFree(dummy));
2565: PetscFunctionReturn(PETSC_SUCCESS);
2566: }
2568: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2569: {
2570: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2572: PetscFunctionBegin;
2573: PetscCall(MatInvertBlockDiagonal(a->A, values));
2574: A->factorerrortype = a->A->factorerrortype;
2575: PetscFunctionReturn(PETSC_SUCCESS);
2576: }
2578: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2579: {
2580: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2582: PetscFunctionBegin;
2583: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2584: PetscCall(MatSetRandom(aij->A, rctx));
2585: if (x->assembled) {
2586: PetscCall(MatSetRandom(aij->B, rctx));
2587: } else {
2588: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2589: }
2590: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2591: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2592: PetscFunctionReturn(PETSC_SUCCESS);
2593: }
2595: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2596: {
2597: PetscFunctionBegin;
2598: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2599: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2600: PetscFunctionReturn(PETSC_SUCCESS);
2601: }
2603: /*@
2604: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2606: Not Collective
2608: Input Parameter:
2609: . A - the matrix
2611: Output Parameter:
2612: . nz - the number of nonzeros
2614: Level: advanced
2616: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2617: @*/
2618: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2619: {
2620: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2621: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2622: PetscBool isaij;
2624: PetscFunctionBegin;
2625: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2626: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2627: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2628: PetscFunctionReturn(PETSC_SUCCESS);
2629: }
2631: /*@
2632: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2634: Collective
2636: Input Parameters:
2637: + A - the matrix
2638: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2640: Level: advanced
2642: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2643: @*/
2644: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2645: {
2646: PetscFunctionBegin;
2647: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2648: PetscFunctionReturn(PETSC_SUCCESS);
2649: }
2651: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2652: {
2653: PetscBool sc = PETSC_FALSE, flg;
2655: PetscFunctionBegin;
2656: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2657: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2658: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2659: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2660: PetscOptionsHeadEnd();
2661: PetscFunctionReturn(PETSC_SUCCESS);
2662: }
2664: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2665: {
2666: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2667: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2669: PetscFunctionBegin;
2670: if (!Y->preallocated) {
2671: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2672: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2673: PetscInt nonew = aij->nonew;
2674: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2675: aij->nonew = nonew;
2676: }
2677: PetscCall(MatShift_Basic(Y, a));
2678: PetscFunctionReturn(PETSC_SUCCESS);
2679: }
2681: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2682: {
2683: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2685: PetscFunctionBegin;
2686: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2687: PetscCall(MatMissingDiagonal(a->A, missing, d));
2688: if (d) {
2689: PetscInt rstart;
2690: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2691: *d += rstart;
2692: }
2693: PetscFunctionReturn(PETSC_SUCCESS);
2694: }
2696: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2697: {
2698: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2700: PetscFunctionBegin;
2701: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2702: PetscFunctionReturn(PETSC_SUCCESS);
2703: }
2705: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2706: {
2707: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2709: PetscFunctionBegin;
2710: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2711: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2712: PetscFunctionReturn(PETSC_SUCCESS);
2713: }
2715: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2716: MatGetRow_MPIAIJ,
2717: MatRestoreRow_MPIAIJ,
2718: MatMult_MPIAIJ,
2719: /* 4*/ MatMultAdd_MPIAIJ,
2720: MatMultTranspose_MPIAIJ,
2721: MatMultTransposeAdd_MPIAIJ,
2722: NULL,
2723: NULL,
2724: NULL,
2725: /*10*/ NULL,
2726: NULL,
2727: NULL,
2728: MatSOR_MPIAIJ,
2729: MatTranspose_MPIAIJ,
2730: /*15*/ MatGetInfo_MPIAIJ,
2731: MatEqual_MPIAIJ,
2732: MatGetDiagonal_MPIAIJ,
2733: MatDiagonalScale_MPIAIJ,
2734: MatNorm_MPIAIJ,
2735: /*20*/ MatAssemblyBegin_MPIAIJ,
2736: MatAssemblyEnd_MPIAIJ,
2737: MatSetOption_MPIAIJ,
2738: MatZeroEntries_MPIAIJ,
2739: /*24*/ MatZeroRows_MPIAIJ,
2740: NULL,
2741: NULL,
2742: NULL,
2743: NULL,
2744: /*29*/ MatSetUp_MPI_Hash,
2745: NULL,
2746: NULL,
2747: MatGetDiagonalBlock_MPIAIJ,
2748: NULL,
2749: /*34*/ MatDuplicate_MPIAIJ,
2750: NULL,
2751: NULL,
2752: NULL,
2753: NULL,
2754: /*39*/ MatAXPY_MPIAIJ,
2755: MatCreateSubMatrices_MPIAIJ,
2756: MatIncreaseOverlap_MPIAIJ,
2757: MatGetValues_MPIAIJ,
2758: MatCopy_MPIAIJ,
2759: /*44*/ MatGetRowMax_MPIAIJ,
2760: MatScale_MPIAIJ,
2761: MatShift_MPIAIJ,
2762: MatDiagonalSet_MPIAIJ,
2763: MatZeroRowsColumns_MPIAIJ,
2764: /*49*/ MatSetRandom_MPIAIJ,
2765: MatGetRowIJ_MPIAIJ,
2766: MatRestoreRowIJ_MPIAIJ,
2767: NULL,
2768: NULL,
2769: /*54*/ MatFDColoringCreate_MPIXAIJ,
2770: NULL,
2771: MatSetUnfactored_MPIAIJ,
2772: MatPermute_MPIAIJ,
2773: NULL,
2774: /*59*/ MatCreateSubMatrix_MPIAIJ,
2775: MatDestroy_MPIAIJ,
2776: MatView_MPIAIJ,
2777: NULL,
2778: NULL,
2779: /*64*/ NULL,
2780: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2781: NULL,
2782: NULL,
2783: NULL,
2784: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2785: MatGetRowMinAbs_MPIAIJ,
2786: NULL,
2787: NULL,
2788: NULL,
2789: NULL,
2790: /*75*/ MatFDColoringApply_AIJ,
2791: MatSetFromOptions_MPIAIJ,
2792: NULL,
2793: NULL,
2794: MatFindZeroDiagonals_MPIAIJ,
2795: /*80*/ NULL,
2796: NULL,
2797: NULL,
2798: /*83*/ MatLoad_MPIAIJ,
2799: NULL,
2800: NULL,
2801: NULL,
2802: NULL,
2803: NULL,
2804: /*89*/ NULL,
2805: NULL,
2806: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2807: NULL,
2808: NULL,
2809: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2810: NULL,
2811: NULL,
2812: NULL,
2813: MatBindToCPU_MPIAIJ,
2814: /*99*/ MatProductSetFromOptions_MPIAIJ,
2815: NULL,
2816: NULL,
2817: MatConjugate_MPIAIJ,
2818: NULL,
2819: /*104*/ MatSetValuesRow_MPIAIJ,
2820: MatRealPart_MPIAIJ,
2821: MatImaginaryPart_MPIAIJ,
2822: NULL,
2823: NULL,
2824: /*109*/ NULL,
2825: NULL,
2826: MatGetRowMin_MPIAIJ,
2827: NULL,
2828: MatMissingDiagonal_MPIAIJ,
2829: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2830: NULL,
2831: MatGetGhosts_MPIAIJ,
2832: NULL,
2833: NULL,
2834: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2835: NULL,
2836: NULL,
2837: NULL,
2838: MatGetMultiProcBlock_MPIAIJ,
2839: /*124*/ MatFindNonzeroRows_MPIAIJ,
2840: MatGetColumnReductions_MPIAIJ,
2841: MatInvertBlockDiagonal_MPIAIJ,
2842: MatInvertVariableBlockDiagonal_MPIAIJ,
2843: MatCreateSubMatricesMPI_MPIAIJ,
2844: /*129*/ NULL,
2845: NULL,
2846: NULL,
2847: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2848: NULL,
2849: /*134*/ NULL,
2850: NULL,
2851: NULL,
2852: NULL,
2853: NULL,
2854: /*139*/ MatSetBlockSizes_MPIAIJ,
2855: NULL,
2856: NULL,
2857: MatFDColoringSetUp_MPIXAIJ,
2858: MatFindOffBlockDiagonalEntries_MPIAIJ,
2859: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2860: /*145*/ NULL,
2861: NULL,
2862: NULL,
2863: MatCreateGraph_Simple_AIJ,
2864: NULL,
2865: /*150*/ NULL,
2866: MatEliminateZeros_MPIAIJ,
2867: MatGetRowSumAbs_MPIAIJ};
2869: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2870: {
2871: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2873: PetscFunctionBegin;
2874: PetscCall(MatStoreValues(aij->A));
2875: PetscCall(MatStoreValues(aij->B));
2876: PetscFunctionReturn(PETSC_SUCCESS);
2877: }
2879: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2880: {
2881: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2883: PetscFunctionBegin;
2884: PetscCall(MatRetrieveValues(aij->A));
2885: PetscCall(MatRetrieveValues(aij->B));
2886: PetscFunctionReturn(PETSC_SUCCESS);
2887: }
2889: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2890: {
2891: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2892: PetscMPIInt size;
2894: PetscFunctionBegin;
2895: if (B->hash_active) {
2896: B->ops[0] = b->cops;
2897: B->hash_active = PETSC_FALSE;
2898: }
2899: PetscCall(PetscLayoutSetUp(B->rmap));
2900: PetscCall(PetscLayoutSetUp(B->cmap));
2902: #if defined(PETSC_USE_CTABLE)
2903: PetscCall(PetscHMapIDestroy(&b->colmap));
2904: #else
2905: PetscCall(PetscFree(b->colmap));
2906: #endif
2907: PetscCall(PetscFree(b->garray));
2908: PetscCall(VecDestroy(&b->lvec));
2909: PetscCall(VecScatterDestroy(&b->Mvctx));
2911: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2913: MatSeqXAIJGetOptions_Private(b->B);
2914: PetscCall(MatDestroy(&b->B));
2915: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2916: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2917: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2918: PetscCall(MatSetType(b->B, MATSEQAIJ));
2919: MatSeqXAIJRestoreOptions_Private(b->B);
2921: MatSeqXAIJGetOptions_Private(b->A);
2922: PetscCall(MatDestroy(&b->A));
2923: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2924: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2925: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2926: PetscCall(MatSetType(b->A, MATSEQAIJ));
2927: MatSeqXAIJRestoreOptions_Private(b->A);
2929: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2930: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2931: B->preallocated = PETSC_TRUE;
2932: B->was_assembled = PETSC_FALSE;
2933: B->assembled = PETSC_FALSE;
2934: PetscFunctionReturn(PETSC_SUCCESS);
2935: }
2937: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2938: {
2939: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2941: PetscFunctionBegin;
2943: PetscCall(PetscLayoutSetUp(B->rmap));
2944: PetscCall(PetscLayoutSetUp(B->cmap));
2946: #if defined(PETSC_USE_CTABLE)
2947: PetscCall(PetscHMapIDestroy(&b->colmap));
2948: #else
2949: PetscCall(PetscFree(b->colmap));
2950: #endif
2951: PetscCall(PetscFree(b->garray));
2952: PetscCall(VecDestroy(&b->lvec));
2953: PetscCall(VecScatterDestroy(&b->Mvctx));
2955: PetscCall(MatResetPreallocation(b->A));
2956: PetscCall(MatResetPreallocation(b->B));
2957: B->preallocated = PETSC_TRUE;
2958: B->was_assembled = PETSC_FALSE;
2959: B->assembled = PETSC_FALSE;
2960: PetscFunctionReturn(PETSC_SUCCESS);
2961: }
2963: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2964: {
2965: Mat mat;
2966: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2968: PetscFunctionBegin;
2969: *newmat = NULL;
2970: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2971: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2972: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2973: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2974: a = (Mat_MPIAIJ *)mat->data;
2976: mat->factortype = matin->factortype;
2977: mat->assembled = matin->assembled;
2978: mat->insertmode = NOT_SET_VALUES;
2980: a->size = oldmat->size;
2981: a->rank = oldmat->rank;
2982: a->donotstash = oldmat->donotstash;
2983: a->roworiented = oldmat->roworiented;
2984: a->rowindices = NULL;
2985: a->rowvalues = NULL;
2986: a->getrowactive = PETSC_FALSE;
2988: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2989: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2990: if (matin->hash_active) {
2991: PetscCall(MatSetUp(mat));
2992: } else {
2993: mat->preallocated = matin->preallocated;
2994: if (oldmat->colmap) {
2995: #if defined(PETSC_USE_CTABLE)
2996: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
2997: #else
2998: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
2999: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3000: #endif
3001: } else a->colmap = NULL;
3002: if (oldmat->garray) {
3003: PetscInt len;
3004: len = oldmat->B->cmap->n;
3005: PetscCall(PetscMalloc1(len + 1, &a->garray));
3006: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3007: } else a->garray = NULL;
3009: /* It may happen MatDuplicate is called with a non-assembled matrix
3010: In fact, MatDuplicate only requires the matrix to be preallocated
3011: This may happen inside a DMCreateMatrix_Shell */
3012: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3013: if (oldmat->Mvctx) PetscCall(VecScatterCopy(oldmat->Mvctx, &a->Mvctx));
3014: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3015: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3016: }
3017: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3018: *newmat = mat;
3019: PetscFunctionReturn(PETSC_SUCCESS);
3020: }
3022: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3023: {
3024: PetscBool isbinary, ishdf5;
3026: PetscFunctionBegin;
3029: /* force binary viewer to load .info file if it has not yet done so */
3030: PetscCall(PetscViewerSetUp(viewer));
3031: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3032: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3033: if (isbinary) {
3034: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3035: } else if (ishdf5) {
3036: #if defined(PETSC_HAVE_HDF5)
3037: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3038: #else
3039: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3040: #endif
3041: } else {
3042: 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);
3043: }
3044: PetscFunctionReturn(PETSC_SUCCESS);
3045: }
3047: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3048: {
3049: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3050: PetscInt *rowidxs, *colidxs;
3051: PetscScalar *matvals;
3053: PetscFunctionBegin;
3054: PetscCall(PetscViewerSetUp(viewer));
3056: /* read in matrix header */
3057: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3058: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3059: M = header[1];
3060: N = header[2];
3061: nz = header[3];
3062: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3063: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3064: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3066: /* set block sizes from the viewer's .info file */
3067: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3068: /* set global sizes if not set already */
3069: if (mat->rmap->N < 0) mat->rmap->N = M;
3070: if (mat->cmap->N < 0) mat->cmap->N = N;
3071: PetscCall(PetscLayoutSetUp(mat->rmap));
3072: PetscCall(PetscLayoutSetUp(mat->cmap));
3074: /* check if the matrix sizes are correct */
3075: PetscCall(MatGetSize(mat, &rows, &cols));
3076: 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);
3078: /* read in row lengths and build row indices */
3079: PetscCall(MatGetLocalSize(mat, &m, NULL));
3080: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3081: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3082: rowidxs[0] = 0;
3083: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3084: if (nz != PETSC_MAX_INT) {
3085: PetscCall(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3086: 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);
3087: }
3089: /* read in column indices and matrix values */
3090: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3091: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3092: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3093: /* store matrix indices and values */
3094: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3095: PetscCall(PetscFree(rowidxs));
3096: PetscCall(PetscFree2(colidxs, matvals));
3097: PetscFunctionReturn(PETSC_SUCCESS);
3098: }
3100: /* Not scalable because of ISAllGather() unless getting all columns. */
3101: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3102: {
3103: IS iscol_local;
3104: PetscBool isstride;
3105: PetscMPIInt lisstride = 0, gisstride;
3107: PetscFunctionBegin;
3108: /* check if we are grabbing all columns*/
3109: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3111: if (isstride) {
3112: PetscInt start, len, mstart, mlen;
3113: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3114: PetscCall(ISGetLocalSize(iscol, &len));
3115: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3116: if (mstart == start && mlen - mstart == len) lisstride = 1;
3117: }
3119: PetscCall(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3120: if (gisstride) {
3121: PetscInt N;
3122: PetscCall(MatGetSize(mat, NULL, &N));
3123: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3124: PetscCall(ISSetIdentity(iscol_local));
3125: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3126: } else {
3127: PetscInt cbs;
3128: PetscCall(ISGetBlockSize(iscol, &cbs));
3129: PetscCall(ISAllGather(iscol, &iscol_local));
3130: PetscCall(ISSetBlockSize(iscol_local, cbs));
3131: }
3133: *isseq = iscol_local;
3134: PetscFunctionReturn(PETSC_SUCCESS);
3135: }
3137: /*
3138: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3139: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3141: Input Parameters:
3142: + mat - matrix
3143: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3144: i.e., mat->rstart <= isrow[i] < mat->rend
3145: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3146: i.e., mat->cstart <= iscol[i] < mat->cend
3148: Output Parameters:
3149: + isrow_d - sequential row index set for retrieving mat->A
3150: . iscol_d - sequential column index set for retrieving mat->A
3151: . iscol_o - sequential column index set for retrieving mat->B
3152: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3153: */
3154: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3155: {
3156: Vec x, cmap;
3157: const PetscInt *is_idx;
3158: PetscScalar *xarray, *cmaparray;
3159: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3160: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3161: Mat B = a->B;
3162: Vec lvec = a->lvec, lcmap;
3163: PetscInt i, cstart, cend, Bn = B->cmap->N;
3164: MPI_Comm comm;
3165: VecScatter Mvctx = a->Mvctx;
3167: PetscFunctionBegin;
3168: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3169: PetscCall(ISGetLocalSize(iscol, &ncols));
3171: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3172: PetscCall(MatCreateVecs(mat, &x, NULL));
3173: PetscCall(VecSet(x, -1.0));
3174: PetscCall(VecDuplicate(x, &cmap));
3175: PetscCall(VecSet(cmap, -1.0));
3177: /* Get start indices */
3178: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3179: isstart -= ncols;
3180: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3182: PetscCall(ISGetIndices(iscol, &is_idx));
3183: PetscCall(VecGetArray(x, &xarray));
3184: PetscCall(VecGetArray(cmap, &cmaparray));
3185: PetscCall(PetscMalloc1(ncols, &idx));
3186: for (i = 0; i < ncols; i++) {
3187: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3188: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3189: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3190: }
3191: PetscCall(VecRestoreArray(x, &xarray));
3192: PetscCall(VecRestoreArray(cmap, &cmaparray));
3193: PetscCall(ISRestoreIndices(iscol, &is_idx));
3195: /* Get iscol_d */
3196: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3197: PetscCall(ISGetBlockSize(iscol, &i));
3198: PetscCall(ISSetBlockSize(*iscol_d, i));
3200: /* Get isrow_d */
3201: PetscCall(ISGetLocalSize(isrow, &m));
3202: rstart = mat->rmap->rstart;
3203: PetscCall(PetscMalloc1(m, &idx));
3204: PetscCall(ISGetIndices(isrow, &is_idx));
3205: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3206: PetscCall(ISRestoreIndices(isrow, &is_idx));
3208: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3209: PetscCall(ISGetBlockSize(isrow, &i));
3210: PetscCall(ISSetBlockSize(*isrow_d, i));
3212: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3213: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3214: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3216: PetscCall(VecDuplicate(lvec, &lcmap));
3218: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3219: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3221: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3222: /* off-process column indices */
3223: count = 0;
3224: PetscCall(PetscMalloc1(Bn, &idx));
3225: PetscCall(PetscMalloc1(Bn, &cmap1));
3227: PetscCall(VecGetArray(lvec, &xarray));
3228: PetscCall(VecGetArray(lcmap, &cmaparray));
3229: for (i = 0; i < Bn; i++) {
3230: if (PetscRealPart(xarray[i]) > -1.0) {
3231: idx[count] = i; /* local column index in off-diagonal part B */
3232: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3233: count++;
3234: }
3235: }
3236: PetscCall(VecRestoreArray(lvec, &xarray));
3237: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3239: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3240: /* cannot ensure iscol_o has same blocksize as iscol! */
3242: PetscCall(PetscFree(idx));
3243: *garray = cmap1;
3245: PetscCall(VecDestroy(&x));
3246: PetscCall(VecDestroy(&cmap));
3247: PetscCall(VecDestroy(&lcmap));
3248: PetscFunctionReturn(PETSC_SUCCESS);
3249: }
3251: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3252: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3253: {
3254: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3255: Mat M = NULL;
3256: MPI_Comm comm;
3257: IS iscol_d, isrow_d, iscol_o;
3258: Mat Asub = NULL, Bsub = NULL;
3259: PetscInt n;
3261: PetscFunctionBegin;
3262: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3264: if (call == MAT_REUSE_MATRIX) {
3265: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3266: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3267: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3269: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3270: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3272: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3273: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3275: /* Update diagonal and off-diagonal portions of submat */
3276: asub = (Mat_MPIAIJ *)(*submat)->data;
3277: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3278: PetscCall(ISGetLocalSize(iscol_o, &n));
3279: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3280: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3281: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3283: } else { /* call == MAT_INITIAL_MATRIX) */
3284: const PetscInt *garray;
3285: PetscInt BsubN;
3287: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3288: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3290: /* Create local submatrices Asub and Bsub */
3291: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3292: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3294: /* Create submatrix M */
3295: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3297: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3298: asub = (Mat_MPIAIJ *)M->data;
3300: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3301: n = asub->B->cmap->N;
3302: if (BsubN > n) {
3303: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3304: const PetscInt *idx;
3305: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3306: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3308: PetscCall(PetscMalloc1(n, &idx_new));
3309: j = 0;
3310: PetscCall(ISGetIndices(iscol_o, &idx));
3311: for (i = 0; i < n; i++) {
3312: if (j >= BsubN) break;
3313: while (subgarray[i] > garray[j]) j++;
3315: if (subgarray[i] == garray[j]) {
3316: idx_new[i] = idx[j++];
3317: } 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]);
3318: }
3319: PetscCall(ISRestoreIndices(iscol_o, &idx));
3321: PetscCall(ISDestroy(&iscol_o));
3322: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3324: } else if (BsubN < n) {
3325: 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);
3326: }
3328: PetscCall(PetscFree(garray));
3329: *submat = M;
3331: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3332: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3333: PetscCall(ISDestroy(&isrow_d));
3335: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3336: PetscCall(ISDestroy(&iscol_d));
3338: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3339: PetscCall(ISDestroy(&iscol_o));
3340: }
3341: PetscFunctionReturn(PETSC_SUCCESS);
3342: }
3344: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3345: {
3346: IS iscol_local = NULL, isrow_d;
3347: PetscInt csize;
3348: PetscInt n, i, j, start, end;
3349: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3350: MPI_Comm comm;
3352: PetscFunctionBegin;
3353: /* If isrow has same processor distribution as mat,
3354: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3355: if (call == MAT_REUSE_MATRIX) {
3356: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3357: if (isrow_d) {
3358: sameRowDist = PETSC_TRUE;
3359: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3360: } else {
3361: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3362: if (iscol_local) {
3363: sameRowDist = PETSC_TRUE;
3364: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3365: }
3366: }
3367: } else {
3368: /* Check if isrow has same processor distribution as mat */
3369: sameDist[0] = PETSC_FALSE;
3370: PetscCall(ISGetLocalSize(isrow, &n));
3371: if (!n) {
3372: sameDist[0] = PETSC_TRUE;
3373: } else {
3374: PetscCall(ISGetMinMax(isrow, &i, &j));
3375: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3376: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3377: }
3379: /* Check if iscol has same processor distribution as mat */
3380: sameDist[1] = PETSC_FALSE;
3381: PetscCall(ISGetLocalSize(iscol, &n));
3382: if (!n) {
3383: sameDist[1] = PETSC_TRUE;
3384: } else {
3385: PetscCall(ISGetMinMax(iscol, &i, &j));
3386: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3387: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3388: }
3390: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3391: PetscCall(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3392: sameRowDist = tsameDist[0];
3393: }
3395: if (sameRowDist) {
3396: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3397: /* isrow and iscol have same processor distribution as mat */
3398: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3399: PetscFunctionReturn(PETSC_SUCCESS);
3400: } else { /* sameRowDist */
3401: /* isrow has same processor distribution as mat */
3402: if (call == MAT_INITIAL_MATRIX) {
3403: PetscBool sorted;
3404: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3405: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3406: PetscCall(ISGetSize(iscol, &i));
3407: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3409: PetscCall(ISSorted(iscol_local, &sorted));
3410: if (sorted) {
3411: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3412: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3413: PetscFunctionReturn(PETSC_SUCCESS);
3414: }
3415: } else { /* call == MAT_REUSE_MATRIX */
3416: IS iscol_sub;
3417: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3418: if (iscol_sub) {
3419: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3420: PetscFunctionReturn(PETSC_SUCCESS);
3421: }
3422: }
3423: }
3424: }
3426: /* General case: iscol -> iscol_local which has global size of iscol */
3427: if (call == MAT_REUSE_MATRIX) {
3428: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3429: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3430: } else {
3431: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3432: }
3434: PetscCall(ISGetLocalSize(iscol, &csize));
3435: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3437: if (call == MAT_INITIAL_MATRIX) {
3438: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3439: PetscCall(ISDestroy(&iscol_local));
3440: }
3441: PetscFunctionReturn(PETSC_SUCCESS);
3442: }
3444: /*@C
3445: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3446: and "off-diagonal" part of the matrix in CSR format.
3448: Collective
3450: Input Parameters:
3451: + comm - MPI communicator
3452: . A - "diagonal" portion of matrix
3453: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3454: - garray - global index of `B` columns
3456: Output Parameter:
3457: . mat - the matrix, with input `A` as its local diagonal matrix
3459: Level: advanced
3461: Notes:
3462: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3464: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3466: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3467: @*/
3468: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3469: {
3470: Mat_MPIAIJ *maij;
3471: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3472: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3473: const PetscScalar *oa;
3474: Mat Bnew;
3475: PetscInt m, n, N;
3476: MatType mpi_mat_type;
3478: PetscFunctionBegin;
3479: PetscCall(MatCreate(comm, mat));
3480: PetscCall(MatGetSize(A, &m, &n));
3481: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3482: 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);
3483: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3484: /* 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); */
3486: /* Get global columns of mat */
3487: PetscCall(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3489: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3490: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3491: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3492: PetscCall(MatSetType(*mat, mpi_mat_type));
3494: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3495: maij = (Mat_MPIAIJ *)(*mat)->data;
3497: (*mat)->preallocated = PETSC_TRUE;
3499: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3500: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3502: /* Set A as diagonal portion of *mat */
3503: maij->A = A;
3505: nz = oi[m];
3506: for (i = 0; i < nz; i++) {
3507: col = oj[i];
3508: oj[i] = garray[col];
3509: }
3511: /* Set Bnew as off-diagonal portion of *mat */
3512: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3513: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3514: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3515: bnew = (Mat_SeqAIJ *)Bnew->data;
3516: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3517: maij->B = Bnew;
3519: 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);
3521: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3522: b->free_a = PETSC_FALSE;
3523: b->free_ij = PETSC_FALSE;
3524: PetscCall(MatDestroy(&B));
3526: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3527: bnew->free_a = PETSC_TRUE;
3528: bnew->free_ij = PETSC_TRUE;
3530: /* condense columns of maij->B */
3531: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3532: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3533: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3534: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3535: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3536: PetscFunctionReturn(PETSC_SUCCESS);
3537: }
3539: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3541: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3542: {
3543: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3544: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3545: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3546: Mat M, Msub, B = a->B;
3547: MatScalar *aa;
3548: Mat_SeqAIJ *aij;
3549: PetscInt *garray = a->garray, *colsub, Ncols;
3550: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3551: IS iscol_sub, iscmap;
3552: const PetscInt *is_idx, *cmap;
3553: PetscBool allcolumns = PETSC_FALSE;
3554: MPI_Comm comm;
3556: PetscFunctionBegin;
3557: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3558: if (call == MAT_REUSE_MATRIX) {
3559: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3560: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3561: PetscCall(ISGetLocalSize(iscol_sub, &count));
3563: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3564: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3566: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3567: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3569: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3571: } else { /* call == MAT_INITIAL_MATRIX) */
3572: PetscBool flg;
3574: PetscCall(ISGetLocalSize(iscol, &n));
3575: PetscCall(ISGetSize(iscol, &Ncols));
3577: /* (1) iscol -> nonscalable iscol_local */
3578: /* Check for special case: each processor gets entire matrix columns */
3579: PetscCall(ISIdentity(iscol_local, &flg));
3580: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3581: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3582: if (allcolumns) {
3583: iscol_sub = iscol_local;
3584: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3585: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3587: } else {
3588: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3589: PetscInt *idx, *cmap1, k;
3590: PetscCall(PetscMalloc1(Ncols, &idx));
3591: PetscCall(PetscMalloc1(Ncols, &cmap1));
3592: PetscCall(ISGetIndices(iscol_local, &is_idx));
3593: count = 0;
3594: k = 0;
3595: for (i = 0; i < Ncols; i++) {
3596: j = is_idx[i];
3597: if (j >= cstart && j < cend) {
3598: /* diagonal part of mat */
3599: idx[count] = j;
3600: cmap1[count++] = i; /* column index in submat */
3601: } else if (Bn) {
3602: /* off-diagonal part of mat */
3603: if (j == garray[k]) {
3604: idx[count] = j;
3605: cmap1[count++] = i; /* column index in submat */
3606: } else if (j > garray[k]) {
3607: while (j > garray[k] && k < Bn - 1) k++;
3608: if (j == garray[k]) {
3609: idx[count] = j;
3610: cmap1[count++] = i; /* column index in submat */
3611: }
3612: }
3613: }
3614: }
3615: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3617: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3618: PetscCall(ISGetBlockSize(iscol, &cbs));
3619: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3621: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3622: }
3624: /* (3) Create sequential Msub */
3625: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3626: }
3628: PetscCall(ISGetLocalSize(iscol_sub, &count));
3629: aij = (Mat_SeqAIJ *)(Msub)->data;
3630: ii = aij->i;
3631: PetscCall(ISGetIndices(iscmap, &cmap));
3633: /*
3634: m - number of local rows
3635: Ncols - number of columns (same on all processors)
3636: rstart - first row in new global matrix generated
3637: */
3638: PetscCall(MatGetSize(Msub, &m, NULL));
3640: if (call == MAT_INITIAL_MATRIX) {
3641: /* (4) Create parallel newmat */
3642: PetscMPIInt rank, size;
3643: PetscInt csize;
3645: PetscCallMPI(MPI_Comm_size(comm, &size));
3646: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3648: /*
3649: Determine the number of non-zeros in the diagonal and off-diagonal
3650: portions of the matrix in order to do correct preallocation
3651: */
3653: /* first get start and end of "diagonal" columns */
3654: PetscCall(ISGetLocalSize(iscol, &csize));
3655: if (csize == PETSC_DECIDE) {
3656: PetscCall(ISGetSize(isrow, &mglobal));
3657: if (mglobal == Ncols) { /* square matrix */
3658: nlocal = m;
3659: } else {
3660: nlocal = Ncols / size + ((Ncols % size) > rank);
3661: }
3662: } else {
3663: nlocal = csize;
3664: }
3665: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3666: rstart = rend - nlocal;
3667: PetscCheck(rank != size - 1 || rend == Ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, Ncols);
3669: /* next, compute all the lengths */
3670: jj = aij->j;
3671: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3672: olens = dlens + m;
3673: for (i = 0; i < m; i++) {
3674: jend = ii[i + 1] - ii[i];
3675: olen = 0;
3676: dlen = 0;
3677: for (j = 0; j < jend; j++) {
3678: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3679: else dlen++;
3680: jj++;
3681: }
3682: olens[i] = olen;
3683: dlens[i] = dlen;
3684: }
3686: PetscCall(ISGetBlockSize(isrow, &bs));
3687: PetscCall(ISGetBlockSize(iscol, &cbs));
3689: PetscCall(MatCreate(comm, &M));
3690: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3691: PetscCall(MatSetBlockSizes(M, bs, cbs));
3692: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3693: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3694: PetscCall(PetscFree(dlens));
3696: } else { /* call == MAT_REUSE_MATRIX */
3697: M = *newmat;
3698: PetscCall(MatGetLocalSize(M, &i, NULL));
3699: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3700: PetscCall(MatZeroEntries(M));
3701: /*
3702: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3703: rather than the slower MatSetValues().
3704: */
3705: M->was_assembled = PETSC_TRUE;
3706: M->assembled = PETSC_FALSE;
3707: }
3709: /* (5) Set values of Msub to *newmat */
3710: PetscCall(PetscMalloc1(count, &colsub));
3711: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3713: jj = aij->j;
3714: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3715: for (i = 0; i < m; i++) {
3716: row = rstart + i;
3717: nz = ii[i + 1] - ii[i];
3718: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3719: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3720: jj += nz;
3721: aa += nz;
3722: }
3723: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3724: PetscCall(ISRestoreIndices(iscmap, &cmap));
3726: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3727: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3729: PetscCall(PetscFree(colsub));
3731: /* save Msub, iscol_sub and iscmap used in processor for next request */
3732: if (call == MAT_INITIAL_MATRIX) {
3733: *newmat = M;
3734: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3735: PetscCall(MatDestroy(&Msub));
3737: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3738: PetscCall(ISDestroy(&iscol_sub));
3740: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3741: PetscCall(ISDestroy(&iscmap));
3743: if (iscol_local) {
3744: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3745: PetscCall(ISDestroy(&iscol_local));
3746: }
3747: }
3748: PetscFunctionReturn(PETSC_SUCCESS);
3749: }
3751: /*
3752: Not great since it makes two copies of the submatrix, first an SeqAIJ
3753: in local and then by concatenating the local matrices the end result.
3754: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3756: This requires a sequential iscol with all indices.
3757: */
3758: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3759: {
3760: PetscMPIInt rank, size;
3761: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3762: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3763: Mat M, Mreuse;
3764: MatScalar *aa, *vwork;
3765: MPI_Comm comm;
3766: Mat_SeqAIJ *aij;
3767: PetscBool colflag, allcolumns = PETSC_FALSE;
3769: PetscFunctionBegin;
3770: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3771: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3772: PetscCallMPI(MPI_Comm_size(comm, &size));
3774: /* Check for special case: each processor gets entire matrix columns */
3775: PetscCall(ISIdentity(iscol, &colflag));
3776: PetscCall(ISGetLocalSize(iscol, &n));
3777: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3778: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3780: if (call == MAT_REUSE_MATRIX) {
3781: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3782: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3783: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3784: } else {
3785: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3786: }
3788: /*
3789: m - number of local rows
3790: n - number of columns (same on all processors)
3791: rstart - first row in new global matrix generated
3792: */
3793: PetscCall(MatGetSize(Mreuse, &m, &n));
3794: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3795: if (call == MAT_INITIAL_MATRIX) {
3796: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3797: ii = aij->i;
3798: jj = aij->j;
3800: /*
3801: Determine the number of non-zeros in the diagonal and off-diagonal
3802: portions of the matrix in order to do correct preallocation
3803: */
3805: /* first get start and end of "diagonal" columns */
3806: if (csize == PETSC_DECIDE) {
3807: PetscCall(ISGetSize(isrow, &mglobal));
3808: if (mglobal == n) { /* square matrix */
3809: nlocal = m;
3810: } else {
3811: nlocal = n / size + ((n % size) > rank);
3812: }
3813: } else {
3814: nlocal = csize;
3815: }
3816: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3817: rstart = rend - nlocal;
3818: PetscCheck(rank != size - 1 || rend == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, n);
3820: /* next, compute all the lengths */
3821: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3822: olens = dlens + m;
3823: for (i = 0; i < m; i++) {
3824: jend = ii[i + 1] - ii[i];
3825: olen = 0;
3826: dlen = 0;
3827: for (j = 0; j < jend; j++) {
3828: if (*jj < rstart || *jj >= rend) olen++;
3829: else dlen++;
3830: jj++;
3831: }
3832: olens[i] = olen;
3833: dlens[i] = dlen;
3834: }
3835: PetscCall(MatCreate(comm, &M));
3836: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3837: PetscCall(MatSetBlockSizes(M, bs, cbs));
3838: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3839: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3840: PetscCall(PetscFree(dlens));
3841: } else {
3842: PetscInt ml, nl;
3844: M = *newmat;
3845: PetscCall(MatGetLocalSize(M, &ml, &nl));
3846: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3847: PetscCall(MatZeroEntries(M));
3848: /*
3849: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3850: rather than the slower MatSetValues().
3851: */
3852: M->was_assembled = PETSC_TRUE;
3853: M->assembled = PETSC_FALSE;
3854: }
3855: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3856: aij = (Mat_SeqAIJ *)(Mreuse)->data;
3857: ii = aij->i;
3858: jj = aij->j;
3860: /* trigger copy to CPU if needed */
3861: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3862: for (i = 0; i < m; i++) {
3863: row = rstart + i;
3864: nz = ii[i + 1] - ii[i];
3865: cwork = jj;
3866: jj = PetscSafePointerPlusOffset(jj, nz);
3867: vwork = aa;
3868: aa = PetscSafePointerPlusOffset(aa, nz);
3869: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3870: }
3871: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3873: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3874: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3875: *newmat = M;
3877: /* save submatrix used in processor for next request */
3878: if (call == MAT_INITIAL_MATRIX) {
3879: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3880: PetscCall(MatDestroy(&Mreuse));
3881: }
3882: PetscFunctionReturn(PETSC_SUCCESS);
3883: }
3885: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3886: {
3887: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3888: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii;
3889: const PetscInt *JJ;
3890: PetscBool nooffprocentries;
3891: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3893: PetscFunctionBegin;
3894: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
3896: PetscCall(PetscLayoutSetUp(B->rmap));
3897: PetscCall(PetscLayoutSetUp(B->cmap));
3898: m = B->rmap->n;
3899: cstart = B->cmap->rstart;
3900: cend = B->cmap->rend;
3901: rstart = B->rmap->rstart;
3903: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3905: if (PetscDefined(USE_DEBUG)) {
3906: for (i = 0; i < m; i++) {
3907: nnz = Ii[i + 1] - Ii[i];
3908: JJ = PetscSafePointerPlusOffset(J, Ii[i]);
3909: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3910: 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]);
3911: 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);
3912: }
3913: }
3915: for (i = 0; i < m; i++) {
3916: nnz = Ii[i + 1] - Ii[i];
3917: JJ = PetscSafePointerPlusOffset(J, Ii[i]);
3918: nnz_max = PetscMax(nnz_max, nnz);
3919: d = 0;
3920: for (j = 0; j < nnz; j++) {
3921: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3922: }
3923: d_nnz[i] = d;
3924: o_nnz[i] = nnz - d;
3925: }
3926: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3927: PetscCall(PetscFree2(d_nnz, o_nnz));
3929: for (i = 0; i < m; i++) {
3930: ii = i + rstart;
3931: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i]), PetscSafePointerPlusOffset(v, Ii[i]), INSERT_VALUES));
3932: }
3933: nooffprocentries = B->nooffprocentries;
3934: B->nooffprocentries = PETSC_TRUE;
3935: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3936: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3937: B->nooffprocentries = nooffprocentries;
3939: /* count number of entries below block diagonal */
3940: PetscCall(PetscFree(Aij->ld));
3941: PetscCall(PetscCalloc1(m, &ld));
3942: Aij->ld = ld;
3943: for (i = 0; i < m; i++) {
3944: nnz = Ii[i + 1] - Ii[i];
3945: j = 0;
3946: while (j < nnz && J[j] < cstart) j++;
3947: ld[i] = j;
3948: if (J) J += nnz;
3949: }
3951: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3952: PetscFunctionReturn(PETSC_SUCCESS);
3953: }
3955: /*@
3956: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3957: (the default parallel PETSc format).
3959: Collective
3961: Input Parameters:
3962: + B - the matrix
3963: . i - the indices into `j` for the start of each local row (indices start with zero)
3964: . j - the column indices for each local row (indices start with zero)
3965: - v - optional values in the matrix
3967: Level: developer
3969: Notes:
3970: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3971: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3972: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3974: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3976: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3978: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3980: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3981: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3983: The format which is used for the sparse matrix input, is equivalent to a
3984: row-major ordering.. i.e for the following matrix, the input data expected is
3985: as shown
3986: .vb
3987: 1 0 0
3988: 2 0 3 P0
3989: -------
3990: 4 5 6 P1
3992: Process0 [P0] rows_owned=[0,1]
3993: i = {0,1,3} [size = nrow+1 = 2+1]
3994: j = {0,0,2} [size = 3]
3995: v = {1,2,3} [size = 3]
3997: Process1 [P1] rows_owned=[2]
3998: i = {0,3} [size = nrow+1 = 1+1]
3999: j = {0,1,2} [size = 3]
4000: v = {4,5,6} [size = 3]
4001: .ve
4003: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4004: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4005: @*/
4006: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4007: {
4008: PetscFunctionBegin;
4009: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4010: PetscFunctionReturn(PETSC_SUCCESS);
4011: }
4013: /*@C
4014: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4015: (the default parallel PETSc format). For good matrix assembly performance
4016: the user should preallocate the matrix storage by setting the parameters
4017: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4019: Collective
4021: Input Parameters:
4022: + B - the matrix
4023: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4024: (same value is used for all local rows)
4025: . d_nnz - array containing the number of nonzeros in the various rows of the
4026: DIAGONAL portion of the local submatrix (possibly different for each row)
4027: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4028: The size of this array is equal to the number of local rows, i.e 'm'.
4029: For matrices that will be factored, you must leave room for (and set)
4030: the diagonal entry even if it is zero.
4031: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4032: submatrix (same value is used for all local rows).
4033: - o_nnz - array containing the number of nonzeros in the various rows of the
4034: OFF-DIAGONAL portion of the local submatrix (possibly different for
4035: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4036: structure. The size of this array is equal to the number
4037: of local rows, i.e 'm'.
4039: Example Usage:
4040: Consider the following 8x8 matrix with 34 non-zero values, that is
4041: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4042: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4043: as follows
4045: .vb
4046: 1 2 0 | 0 3 0 | 0 4
4047: Proc0 0 5 6 | 7 0 0 | 8 0
4048: 9 0 10 | 11 0 0 | 12 0
4049: -------------------------------------
4050: 13 0 14 | 15 16 17 | 0 0
4051: Proc1 0 18 0 | 19 20 21 | 0 0
4052: 0 0 0 | 22 23 0 | 24 0
4053: -------------------------------------
4054: Proc2 25 26 27 | 0 0 28 | 29 0
4055: 30 0 0 | 31 32 33 | 0 34
4056: .ve
4058: This can be represented as a collection of submatrices as
4059: .vb
4060: A B C
4061: D E F
4062: G H I
4063: .ve
4065: Where the submatrices A,B,C are owned by proc0, D,E,F are
4066: owned by proc1, G,H,I are owned by proc2.
4068: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4069: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4070: The 'M','N' parameters are 8,8, and have the same values on all procs.
4072: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4073: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4074: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4075: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4076: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4077: matrix, ans [DF] as another `MATSEQAIJ` matrix.
4079: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4080: allocated for every row of the local diagonal submatrix, and `o_nz`
4081: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4082: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4083: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4084: In this case, the values of `d_nz`, `o_nz` are
4085: .vb
4086: proc0 dnz = 2, o_nz = 2
4087: proc1 dnz = 3, o_nz = 2
4088: proc2 dnz = 1, o_nz = 4
4089: .ve
4090: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4091: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4092: for proc3. i.e we are using 12+15+10=37 storage locations to store
4093: 34 values.
4095: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4096: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4097: In the above case the values for `d_nnz`, `o_nnz` are
4098: .vb
4099: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4100: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4101: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4102: .ve
4103: Here the space allocated is sum of all the above values i.e 34, and
4104: hence pre-allocation is perfect.
4106: Level: intermediate
4108: Notes:
4109: If the *_nnz parameter is given then the *_nz parameter is ignored
4111: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4112: storage. The stored row and column indices begin with zero.
4113: See [Sparse Matrices](sec_matsparse) for details.
4115: The parallel matrix is partitioned such that the first m0 rows belong to
4116: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4117: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4119: The DIAGONAL portion of the local submatrix of a processor can be defined
4120: as the submatrix which is obtained by extraction the part corresponding to
4121: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4122: first row that belongs to the processor, r2 is the last row belonging to
4123: the this processor, and c1-c2 is range of indices of the local part of a
4124: vector suitable for applying the matrix to. This is an mxn matrix. In the
4125: common case of a square matrix, the row and column ranges are the same and
4126: the DIAGONAL part is also square. The remaining portion of the local
4127: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4129: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4131: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4132: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4133: You can also run with the option `-info` and look for messages with the string
4134: malloc in them to see if additional memory allocation was needed.
4136: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4137: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4138: @*/
4139: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4140: {
4141: PetscFunctionBegin;
4144: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4145: PetscFunctionReturn(PETSC_SUCCESS);
4146: }
4148: /*@
4149: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4150: CSR format for the local rows.
4152: Collective
4154: Input Parameters:
4155: + comm - MPI communicator
4156: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4157: . n - This value should be the same as the local size used in creating the
4158: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4159: calculated if `N` is given) For square matrices n is almost always `m`.
4160: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4161: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4162: . 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
4163: . j - global column indices
4164: - a - optional matrix values
4166: Output Parameter:
4167: . mat - the matrix
4169: Level: intermediate
4171: Notes:
4172: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4173: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4174: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4176: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4178: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4180: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4181: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4183: The format which is used for the sparse matrix input, is equivalent to a
4184: row-major ordering, i.e., for the following matrix, the input data expected is
4185: as shown
4186: .vb
4187: 1 0 0
4188: 2 0 3 P0
4189: -------
4190: 4 5 6 P1
4192: Process0 [P0] rows_owned=[0,1]
4193: i = {0,1,3} [size = nrow+1 = 2+1]
4194: j = {0,0,2} [size = 3]
4195: v = {1,2,3} [size = 3]
4197: Process1 [P1] rows_owned=[2]
4198: i = {0,3} [size = nrow+1 = 1+1]
4199: j = {0,1,2} [size = 3]
4200: v = {4,5,6} [size = 3]
4201: .ve
4203: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4204: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4205: @*/
4206: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4207: {
4208: PetscFunctionBegin;
4209: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4210: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4211: PetscCall(MatCreate(comm, mat));
4212: PetscCall(MatSetSizes(*mat, m, n, M, N));
4213: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4214: PetscCall(MatSetType(*mat, MATMPIAIJ));
4215: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4216: PetscFunctionReturn(PETSC_SUCCESS);
4217: }
4219: /*@
4220: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4221: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4222: from `MatCreateMPIAIJWithArrays()`
4224: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4226: Collective
4228: Input Parameters:
4229: + mat - the matrix
4230: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4231: . n - This value should be the same as the local size used in creating the
4232: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4233: calculated if N is given) For square matrices n is almost always m.
4234: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4235: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4236: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4237: . J - column indices
4238: - v - matrix values
4240: Level: deprecated
4242: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4243: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4244: @*/
4245: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4246: {
4247: PetscInt nnz, i;
4248: PetscBool nooffprocentries;
4249: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4250: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4251: PetscScalar *ad, *ao;
4252: PetscInt ldi, Iii, md;
4253: const PetscInt *Adi = Ad->i;
4254: PetscInt *ld = Aij->ld;
4256: PetscFunctionBegin;
4257: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4258: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4259: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4260: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4262: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4263: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4265: for (i = 0; i < m; i++) {
4266: if (PetscDefined(USE_DEBUG)) {
4267: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4268: PetscCheck(J[j] >= J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is not sorted", j - Ii[i], J[j], i);
4269: PetscCheck(J[j] != J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is identical to previous entry", j - Ii[i], J[j], i);
4270: }
4271: }
4272: nnz = Ii[i + 1] - Ii[i];
4273: Iii = Ii[i];
4274: ldi = ld[i];
4275: md = Adi[i + 1] - Adi[i];
4276: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4277: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4278: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4279: ad += md;
4280: ao += nnz - md;
4281: }
4282: nooffprocentries = mat->nooffprocentries;
4283: mat->nooffprocentries = PETSC_TRUE;
4284: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4285: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4286: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4287: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4288: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4289: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4290: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4291: mat->nooffprocentries = nooffprocentries;
4292: PetscFunctionReturn(PETSC_SUCCESS);
4293: }
4295: /*@
4296: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4298: Collective
4300: Input Parameters:
4301: + mat - the matrix
4302: - v - matrix values, stored by row
4304: Level: intermediate
4306: Notes:
4307: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4309: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4311: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4312: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4313: @*/
4314: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4315: {
4316: PetscInt nnz, i, m;
4317: PetscBool nooffprocentries;
4318: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4319: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4320: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4321: PetscScalar *ad, *ao;
4322: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4323: PetscInt ldi, Iii, md;
4324: PetscInt *ld = Aij->ld;
4326: PetscFunctionBegin;
4327: m = mat->rmap->n;
4329: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4330: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4331: Iii = 0;
4332: for (i = 0; i < m; i++) {
4333: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4334: ldi = ld[i];
4335: md = Adi[i + 1] - Adi[i];
4336: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4337: ad += md;
4338: if (ao) {
4339: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4340: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4341: ao += nnz - md;
4342: }
4343: Iii += nnz;
4344: }
4345: nooffprocentries = mat->nooffprocentries;
4346: mat->nooffprocentries = PETSC_TRUE;
4347: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4348: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4349: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4350: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4351: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4352: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4353: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4354: mat->nooffprocentries = nooffprocentries;
4355: PetscFunctionReturn(PETSC_SUCCESS);
4356: }
4358: /*@C
4359: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4360: (the default parallel PETSc format). For good matrix assembly performance
4361: the user should preallocate the matrix storage by setting the parameters
4362: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4364: Collective
4366: Input Parameters:
4367: + comm - MPI communicator
4368: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4369: This value should be the same as the local size used in creating the
4370: y vector for the matrix-vector product y = Ax.
4371: . n - This value should be the same as the local size used in creating the
4372: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4373: calculated if N is given) For square matrices n is almost always m.
4374: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4375: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4376: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4377: (same value is used for all local rows)
4378: . d_nnz - array containing the number of nonzeros in the various rows of the
4379: DIAGONAL portion of the local submatrix (possibly different for each row)
4380: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4381: The size of this array is equal to the number of local rows, i.e 'm'.
4382: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4383: submatrix (same value is used for all local rows).
4384: - o_nnz - array containing the number of nonzeros in the various rows of the
4385: OFF-DIAGONAL portion of the local submatrix (possibly different for
4386: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4387: structure. The size of this array is equal to the number
4388: of local rows, i.e 'm'.
4390: Output Parameter:
4391: . A - the matrix
4393: Options Database Keys:
4394: + -mat_no_inode - Do not use inodes
4395: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4396: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4397: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4398: to be viewed as a matrix. Entry (i,j) is the size of message (in bytes) rank i sends to rank j in one `MatMult()` call.
4400: Level: intermediate
4402: Notes:
4403: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4404: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4405: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4407: If the *_nnz parameter is given then the *_nz parameter is ignored
4409: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4410: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4411: storage requirements for this matrix.
4413: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4414: processor than it must be used on all processors that share the object for
4415: that argument.
4417: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4418: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4420: The user MUST specify either the local or global matrix dimensions
4421: (possibly both).
4423: The parallel matrix is partitioned across processors such that the
4424: first `m0` rows belong to process 0, the next `m1` rows belong to
4425: process 1, the next `m2` rows belong to process 2, etc., where
4426: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4427: values corresponding to [m x N] submatrix.
4429: The columns are logically partitioned with the n0 columns belonging
4430: to 0th partition, the next n1 columns belonging to the next
4431: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4433: The DIAGONAL portion of the local submatrix on any given processor
4434: is the submatrix corresponding to the rows and columns m,n
4435: corresponding to the given processor. i.e diagonal matrix on
4436: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4437: etc. The remaining portion of the local submatrix [m x (N-n)]
4438: constitute the OFF-DIAGONAL portion. The example below better
4439: illustrates this concept.
4441: For a square global matrix we define each processor's diagonal portion
4442: to be its local rows and the corresponding columns (a square submatrix);
4443: each processor's off-diagonal portion encompasses the remainder of the
4444: local matrix (a rectangular submatrix).
4446: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4448: When calling this routine with a single process communicator, a matrix of
4449: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4450: type of communicator, use the construction mechanism
4451: .vb
4452: MatCreate(..., &A);
4453: MatSetType(A, MATMPIAIJ);
4454: MatSetSizes(A, m, n, M, N);
4455: MatMPIAIJSetPreallocation(A, ...);
4456: .ve
4458: By default, this format uses inodes (identical nodes) when possible.
4459: We search for consecutive rows with the same nonzero structure, thereby
4460: reusing matrix information to achieve increased efficiency.
4462: Example Usage:
4463: Consider the following 8x8 matrix with 34 non-zero values, that is
4464: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4465: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4466: as follows
4468: .vb
4469: 1 2 0 | 0 3 0 | 0 4
4470: Proc0 0 5 6 | 7 0 0 | 8 0
4471: 9 0 10 | 11 0 0 | 12 0
4472: -------------------------------------
4473: 13 0 14 | 15 16 17 | 0 0
4474: Proc1 0 18 0 | 19 20 21 | 0 0
4475: 0 0 0 | 22 23 0 | 24 0
4476: -------------------------------------
4477: Proc2 25 26 27 | 0 0 28 | 29 0
4478: 30 0 0 | 31 32 33 | 0 34
4479: .ve
4481: This can be represented as a collection of submatrices as
4483: .vb
4484: A B C
4485: D E F
4486: G H I
4487: .ve
4489: Where the submatrices A,B,C are owned by proc0, D,E,F are
4490: owned by proc1, G,H,I are owned by proc2.
4492: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4493: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4494: The 'M','N' parameters are 8,8, and have the same values on all procs.
4496: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4497: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4498: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4499: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4500: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4501: matrix, ans [DF] as another SeqAIJ matrix.
4503: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4504: allocated for every row of the local diagonal submatrix, and `o_nz`
4505: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4506: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4507: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4508: In this case, the values of `d_nz`,`o_nz` are
4509: .vb
4510: proc0 dnz = 2, o_nz = 2
4511: proc1 dnz = 3, o_nz = 2
4512: proc2 dnz = 1, o_nz = 4
4513: .ve
4514: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4515: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4516: for proc3. i.e we are using 12+15+10=37 storage locations to store
4517: 34 values.
4519: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4520: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4521: In the above case the values for d_nnz,o_nnz are
4522: .vb
4523: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4524: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4525: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4526: .ve
4527: Here the space allocated is sum of all the above values i.e 34, and
4528: hence pre-allocation is perfect.
4530: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4531: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4532: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4533: @*/
4534: 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)
4535: {
4536: PetscMPIInt size;
4538: PetscFunctionBegin;
4539: PetscCall(MatCreate(comm, A));
4540: PetscCall(MatSetSizes(*A, m, n, M, N));
4541: PetscCallMPI(MPI_Comm_size(comm, &size));
4542: if (size > 1) {
4543: PetscCall(MatSetType(*A, MATMPIAIJ));
4544: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4545: } else {
4546: PetscCall(MatSetType(*A, MATSEQAIJ));
4547: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4548: }
4549: PetscFunctionReturn(PETSC_SUCCESS);
4550: }
4552: /*MC
4553: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4555: Synopsis:
4556: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4558: Not Collective
4560: Input Parameter:
4561: . A - the `MATMPIAIJ` matrix
4563: Output Parameters:
4564: + Ad - the diagonal portion of the matrix
4565: . Ao - the off-diagonal portion of the matrix
4566: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4567: - ierr - error code
4569: Level: advanced
4571: Note:
4572: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4574: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4575: M*/
4577: /*MC
4578: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4580: Synopsis:
4581: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4583: Not Collective
4585: Input Parameters:
4586: + A - the `MATMPIAIJ` matrix
4587: . Ad - the diagonal portion of the matrix
4588: . Ao - the off-diagonal portion of the matrix
4589: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4590: - ierr - error code
4592: Level: advanced
4594: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4595: M*/
4597: /*@C
4598: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4600: Not Collective
4602: Input Parameter:
4603: . A - The `MATMPIAIJ` matrix
4605: Output Parameters:
4606: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4607: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4608: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4610: Level: intermediate
4612: Note:
4613: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4614: 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
4615: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4616: local column numbers to global column numbers in the original matrix.
4618: Fortran Notes:
4619: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4621: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4622: @*/
4623: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4624: {
4625: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4626: PetscBool flg;
4628: PetscFunctionBegin;
4629: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4630: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4631: if (Ad) *Ad = a->A;
4632: if (Ao) *Ao = a->B;
4633: if (colmap) *colmap = a->garray;
4634: PetscFunctionReturn(PETSC_SUCCESS);
4635: }
4637: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4638: {
4639: PetscInt m, N, i, rstart, nnz, Ii;
4640: PetscInt *indx;
4641: PetscScalar *values;
4642: MatType rootType;
4644: PetscFunctionBegin;
4645: PetscCall(MatGetSize(inmat, &m, &N));
4646: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4647: PetscInt *dnz, *onz, sum, bs, cbs;
4649: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4650: /* Check sum(n) = N */
4651: PetscCall(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4652: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4654: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4655: rstart -= m;
4657: MatPreallocateBegin(comm, m, n, dnz, onz);
4658: for (i = 0; i < m; i++) {
4659: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4660: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4661: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4662: }
4664: PetscCall(MatCreate(comm, outmat));
4665: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4666: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4667: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4668: PetscCall(MatGetRootType_Private(inmat, &rootType));
4669: PetscCall(MatSetType(*outmat, rootType));
4670: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4671: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4672: MatPreallocateEnd(dnz, onz);
4673: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4674: }
4676: /* numeric phase */
4677: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4678: for (i = 0; i < m; i++) {
4679: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4680: Ii = i + rstart;
4681: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4682: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4683: }
4684: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4685: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4686: PetscFunctionReturn(PETSC_SUCCESS);
4687: }
4689: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4690: {
4691: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4693: PetscFunctionBegin;
4694: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4695: PetscCall(PetscFree(merge->id_r));
4696: PetscCall(PetscFree(merge->len_s));
4697: PetscCall(PetscFree(merge->len_r));
4698: PetscCall(PetscFree(merge->bi));
4699: PetscCall(PetscFree(merge->bj));
4700: PetscCall(PetscFree(merge->buf_ri[0]));
4701: PetscCall(PetscFree(merge->buf_ri));
4702: PetscCall(PetscFree(merge->buf_rj[0]));
4703: PetscCall(PetscFree(merge->buf_rj));
4704: PetscCall(PetscFree(merge->coi));
4705: PetscCall(PetscFree(merge->coj));
4706: PetscCall(PetscFree(merge->owners_co));
4707: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4708: PetscCall(PetscFree(merge));
4709: PetscFunctionReturn(PETSC_SUCCESS);
4710: }
4712: #include <../src/mat/utils/freespace.h>
4713: #include <petscbt.h>
4715: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4716: {
4717: MPI_Comm comm;
4718: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4719: PetscMPIInt size, rank, taga, *len_s;
4720: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj;
4721: PetscInt proc, m;
4722: PetscInt **buf_ri, **buf_rj;
4723: PetscInt k, anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4724: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4725: MPI_Request *s_waits, *r_waits;
4726: MPI_Status *status;
4727: const MatScalar *aa, *a_a;
4728: MatScalar **abuf_r, *ba_i;
4729: Mat_Merge_SeqsToMPI *merge;
4730: PetscContainer container;
4732: PetscFunctionBegin;
4733: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4734: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4736: PetscCallMPI(MPI_Comm_size(comm, &size));
4737: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4739: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4740: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4741: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4742: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4743: aa = a_a;
4745: bi = merge->bi;
4746: bj = merge->bj;
4747: buf_ri = merge->buf_ri;
4748: buf_rj = merge->buf_rj;
4750: PetscCall(PetscMalloc1(size, &status));
4751: owners = merge->rowmap->range;
4752: len_s = merge->len_s;
4754: /* send and recv matrix values */
4755: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4756: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4758: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4759: for (proc = 0, k = 0; proc < size; proc++) {
4760: if (!len_s[proc]) continue;
4761: i = owners[proc];
4762: PetscCallMPI(MPI_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4763: k++;
4764: }
4766: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4767: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4768: PetscCall(PetscFree(status));
4770: PetscCall(PetscFree(s_waits));
4771: PetscCall(PetscFree(r_waits));
4773: /* insert mat values of mpimat */
4774: PetscCall(PetscMalloc1(N, &ba_i));
4775: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4777: for (k = 0; k < merge->nrecv; k++) {
4778: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4779: nrows = *buf_ri_k[k];
4780: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4781: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4782: }
4784: /* set values of ba */
4785: m = merge->rowmap->n;
4786: for (i = 0; i < m; i++) {
4787: arow = owners[rank] + i;
4788: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4789: bnzi = bi[i + 1] - bi[i];
4790: PetscCall(PetscArrayzero(ba_i, bnzi));
4792: /* add local non-zero vals of this proc's seqmat into ba */
4793: anzi = ai[arow + 1] - ai[arow];
4794: aj = a->j + ai[arow];
4795: aa = a_a + ai[arow];
4796: nextaj = 0;
4797: for (j = 0; nextaj < anzi; j++) {
4798: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4799: ba_i[j] += aa[nextaj++];
4800: }
4801: }
4803: /* add received vals into ba */
4804: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4805: /* i-th row */
4806: if (i == *nextrow[k]) {
4807: anzi = *(nextai[k] + 1) - *nextai[k];
4808: aj = buf_rj[k] + *nextai[k];
4809: aa = abuf_r[k] + *nextai[k];
4810: nextaj = 0;
4811: for (j = 0; nextaj < anzi; j++) {
4812: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4813: ba_i[j] += aa[nextaj++];
4814: }
4815: }
4816: nextrow[k]++;
4817: nextai[k]++;
4818: }
4819: }
4820: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4821: }
4822: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4823: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4824: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4826: PetscCall(PetscFree(abuf_r[0]));
4827: PetscCall(PetscFree(abuf_r));
4828: PetscCall(PetscFree(ba_i));
4829: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4830: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4831: PetscFunctionReturn(PETSC_SUCCESS);
4832: }
4834: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4835: {
4836: Mat B_mpi;
4837: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4838: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4839: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4840: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4841: PetscInt len, proc, *dnz, *onz, bs, cbs;
4842: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4843: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4844: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4845: MPI_Status *status;
4846: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4847: PetscBT lnkbt;
4848: Mat_Merge_SeqsToMPI *merge;
4849: PetscContainer container;
4851: PetscFunctionBegin;
4852: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4854: /* make sure it is a PETSc comm */
4855: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4856: PetscCallMPI(MPI_Comm_size(comm, &size));
4857: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4859: PetscCall(PetscNew(&merge));
4860: PetscCall(PetscMalloc1(size, &status));
4862: /* determine row ownership */
4863: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4864: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4865: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4866: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4867: PetscCall(PetscLayoutSetUp(merge->rowmap));
4868: PetscCall(PetscMalloc1(size, &len_si));
4869: PetscCall(PetscMalloc1(size, &merge->len_s));
4871: m = merge->rowmap->n;
4872: owners = merge->rowmap->range;
4874: /* determine the number of messages to send, their lengths */
4875: len_s = merge->len_s;
4877: len = 0; /* length of buf_si[] */
4878: merge->nsend = 0;
4879: for (proc = 0; proc < size; proc++) {
4880: len_si[proc] = 0;
4881: if (proc == rank) {
4882: len_s[proc] = 0;
4883: } else {
4884: len_si[proc] = owners[proc + 1] - owners[proc] + 1;
4885: len_s[proc] = ai[owners[proc + 1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4886: }
4887: if (len_s[proc]) {
4888: merge->nsend++;
4889: nrows = 0;
4890: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4891: if (ai[i + 1] > ai[i]) nrows++;
4892: }
4893: len_si[proc] = 2 * (nrows + 1);
4894: len += len_si[proc];
4895: }
4896: }
4898: /* determine the number and length of messages to receive for ij-structure */
4899: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4900: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4902: /* post the Irecv of j-structure */
4903: PetscCall(PetscCommGetNewTag(comm, &tagj));
4904: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4906: /* post the Isend of j-structure */
4907: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4909: for (proc = 0, k = 0; proc < size; proc++) {
4910: if (!len_s[proc]) continue;
4911: i = owners[proc];
4912: PetscCallMPI(MPI_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4913: k++;
4914: }
4916: /* receives and sends of j-structure are complete */
4917: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4918: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4920: /* send and recv i-structure */
4921: PetscCall(PetscCommGetNewTag(comm, &tagi));
4922: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4924: PetscCall(PetscMalloc1(len + 1, &buf_s));
4925: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4926: for (proc = 0, k = 0; proc < size; proc++) {
4927: if (!len_s[proc]) continue;
4928: /* form outgoing message for i-structure:
4929: buf_si[0]: nrows to be sent
4930: [1:nrows]: row index (global)
4931: [nrows+1:2*nrows+1]: i-structure index
4932: */
4933: nrows = len_si[proc] / 2 - 1;
4934: buf_si_i = buf_si + nrows + 1;
4935: buf_si[0] = nrows;
4936: buf_si_i[0] = 0;
4937: nrows = 0;
4938: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4939: anzi = ai[i + 1] - ai[i];
4940: if (anzi) {
4941: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4942: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4943: nrows++;
4944: }
4945: }
4946: PetscCallMPI(MPI_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4947: k++;
4948: buf_si += len_si[proc];
4949: }
4951: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4952: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4954: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4955: 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]));
4957: PetscCall(PetscFree(len_si));
4958: PetscCall(PetscFree(len_ri));
4959: PetscCall(PetscFree(rj_waits));
4960: PetscCall(PetscFree2(si_waits, sj_waits));
4961: PetscCall(PetscFree(ri_waits));
4962: PetscCall(PetscFree(buf_s));
4963: PetscCall(PetscFree(status));
4965: /* compute a local seq matrix in each processor */
4966: /* allocate bi array and free space for accumulating nonzero column info */
4967: PetscCall(PetscMalloc1(m + 1, &bi));
4968: bi[0] = 0;
4970: /* create and initialize a linked list */
4971: nlnk = N + 1;
4972: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4974: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4975: len = ai[owners[rank + 1]] - ai[owners[rank]];
4976: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4978: current_space = free_space;
4980: /* determine symbolic info for each local row */
4981: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4983: for (k = 0; k < merge->nrecv; k++) {
4984: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4985: nrows = *buf_ri_k[k];
4986: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4987: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4988: }
4990: MatPreallocateBegin(comm, m, n, dnz, onz);
4991: len = 0;
4992: for (i = 0; i < m; i++) {
4993: bnzi = 0;
4994: /* add local non-zero cols of this proc's seqmat into lnk */
4995: arow = owners[rank] + i;
4996: anzi = ai[arow + 1] - ai[arow];
4997: aj = a->j + ai[arow];
4998: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
4999: bnzi += nlnk;
5000: /* add received col data into lnk */
5001: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5002: if (i == *nextrow[k]) { /* i-th row */
5003: anzi = *(nextai[k] + 1) - *nextai[k];
5004: aj = buf_rj[k] + *nextai[k];
5005: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5006: bnzi += nlnk;
5007: nextrow[k]++;
5008: nextai[k]++;
5009: }
5010: }
5011: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5013: /* if free space is not available, make more free space */
5014: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5015: /* copy data into free space, then initialize lnk */
5016: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5017: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5019: current_space->array += bnzi;
5020: current_space->local_used += bnzi;
5021: current_space->local_remaining -= bnzi;
5023: bi[i + 1] = bi[i] + bnzi;
5024: }
5026: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5028: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5029: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5030: PetscCall(PetscLLDestroy(lnk, lnkbt));
5032: /* create symbolic parallel matrix B_mpi */
5033: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5034: PetscCall(MatCreate(comm, &B_mpi));
5035: if (n == PETSC_DECIDE) {
5036: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5037: } else {
5038: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5039: }
5040: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5041: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5042: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5043: MatPreallocateEnd(dnz, onz);
5044: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5046: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5047: B_mpi->assembled = PETSC_FALSE;
5048: merge->bi = bi;
5049: merge->bj = bj;
5050: merge->buf_ri = buf_ri;
5051: merge->buf_rj = buf_rj;
5052: merge->coi = NULL;
5053: merge->coj = NULL;
5054: merge->owners_co = NULL;
5056: PetscCall(PetscCommDestroy(&comm));
5058: /* attach the supporting struct to B_mpi for reuse */
5059: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5060: PetscCall(PetscContainerSetPointer(container, merge));
5061: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5062: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5063: PetscCall(PetscContainerDestroy(&container));
5064: *mpimat = B_mpi;
5066: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5067: PetscFunctionReturn(PETSC_SUCCESS);
5068: }
5070: /*@
5071: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5072: matrices from each processor
5074: Collective
5076: Input Parameters:
5077: + comm - the communicators the parallel matrix will live on
5078: . seqmat - the input sequential matrices
5079: . m - number of local rows (or `PETSC_DECIDE`)
5080: . n - number of local columns (or `PETSC_DECIDE`)
5081: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5083: Output Parameter:
5084: . mpimat - the parallel matrix generated
5086: Level: advanced
5088: Note:
5089: The dimensions of the sequential matrix in each processor MUST be the same.
5090: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5091: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5093: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5094: @*/
5095: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5096: {
5097: PetscMPIInt size;
5099: PetscFunctionBegin;
5100: PetscCallMPI(MPI_Comm_size(comm, &size));
5101: if (size == 1) {
5102: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5103: if (scall == MAT_INITIAL_MATRIX) {
5104: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5105: } else {
5106: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5107: }
5108: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5109: PetscFunctionReturn(PETSC_SUCCESS);
5110: }
5111: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5112: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5113: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5114: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5115: PetscFunctionReturn(PETSC_SUCCESS);
5116: }
5118: /*@
5119: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5121: Not Collective
5123: Input Parameter:
5124: . A - the matrix
5126: Output Parameter:
5127: . A_loc - the local sequential matrix generated
5129: Level: developer
5131: Notes:
5132: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5133: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5134: `n` is the global column count obtained with `MatGetSize()`
5136: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5138: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5140: Destroy the matrix with `MatDestroy()`
5142: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5143: @*/
5144: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5145: {
5146: PetscBool mpi;
5148: PetscFunctionBegin;
5149: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5150: if (mpi) {
5151: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5152: } else {
5153: *A_loc = A;
5154: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5155: }
5156: PetscFunctionReturn(PETSC_SUCCESS);
5157: }
5159: /*@
5160: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5162: Not Collective
5164: Input Parameters:
5165: + A - the matrix
5166: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5168: Output Parameter:
5169: . A_loc - the local sequential matrix generated
5171: Level: developer
5173: Notes:
5174: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5175: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5176: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5178: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5180: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5181: 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
5182: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5183: 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.
5185: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5186: @*/
5187: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5188: {
5189: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5190: Mat_SeqAIJ *mat, *a, *b;
5191: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5192: const PetscScalar *aa, *ba, *aav, *bav;
5193: PetscScalar *ca, *cam;
5194: PetscMPIInt size;
5195: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5196: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5197: PetscBool match;
5199: PetscFunctionBegin;
5200: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5201: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5202: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5203: if (size == 1) {
5204: if (scall == MAT_INITIAL_MATRIX) {
5205: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5206: *A_loc = mpimat->A;
5207: } else if (scall == MAT_REUSE_MATRIX) {
5208: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5209: }
5210: PetscFunctionReturn(PETSC_SUCCESS);
5211: }
5213: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5214: a = (Mat_SeqAIJ *)mpimat->A->data;
5215: b = (Mat_SeqAIJ *)mpimat->B->data;
5216: ai = a->i;
5217: aj = a->j;
5218: bi = b->i;
5219: bj = b->j;
5220: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5221: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5222: aa = aav;
5223: ba = bav;
5224: if (scall == MAT_INITIAL_MATRIX) {
5225: PetscCall(PetscMalloc1(1 + am, &ci));
5226: ci[0] = 0;
5227: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5228: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5229: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5230: k = 0;
5231: for (i = 0; i < am; i++) {
5232: ncols_o = bi[i + 1] - bi[i];
5233: ncols_d = ai[i + 1] - ai[i];
5234: /* off-diagonal portion of A */
5235: for (jo = 0; jo < ncols_o; jo++) {
5236: col = cmap[*bj];
5237: if (col >= cstart) break;
5238: cj[k] = col;
5239: bj++;
5240: ca[k++] = *ba++;
5241: }
5242: /* diagonal portion of A */
5243: for (j = 0; j < ncols_d; j++) {
5244: cj[k] = cstart + *aj++;
5245: ca[k++] = *aa++;
5246: }
5247: /* off-diagonal portion of A */
5248: for (j = jo; j < ncols_o; j++) {
5249: cj[k] = cmap[*bj++];
5250: ca[k++] = *ba++;
5251: }
5252: }
5253: /* put together the new matrix */
5254: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5255: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5256: /* Since these are PETSc arrays, change flags to free them as necessary. */
5257: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5258: mat->free_a = PETSC_TRUE;
5259: mat->free_ij = PETSC_TRUE;
5260: mat->nonew = 0;
5261: } else if (scall == MAT_REUSE_MATRIX) {
5262: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5263: ci = mat->i;
5264: cj = mat->j;
5265: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5266: for (i = 0; i < am; i++) {
5267: /* off-diagonal portion of A */
5268: ncols_o = bi[i + 1] - bi[i];
5269: for (jo = 0; jo < ncols_o; jo++) {
5270: col = cmap[*bj];
5271: if (col >= cstart) break;
5272: *cam++ = *ba++;
5273: bj++;
5274: }
5275: /* diagonal portion of A */
5276: ncols_d = ai[i + 1] - ai[i];
5277: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5278: /* off-diagonal portion of A */
5279: for (j = jo; j < ncols_o; j++) {
5280: *cam++ = *ba++;
5281: bj++;
5282: }
5283: }
5284: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5285: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5286: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5287: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5288: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5289: PetscFunctionReturn(PETSC_SUCCESS);
5290: }
5292: /*@
5293: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5294: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5296: Not Collective
5298: Input Parameters:
5299: + A - the matrix
5300: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5302: Output Parameters:
5303: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5304: - A_loc - the local sequential matrix generated
5306: Level: developer
5308: Note:
5309: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5310: part, then those associated with the off-diagonal part (in its local ordering)
5312: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5313: @*/
5314: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5315: {
5316: Mat Ao, Ad;
5317: const PetscInt *cmap;
5318: PetscMPIInt size;
5319: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5321: PetscFunctionBegin;
5322: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5323: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5324: if (size == 1) {
5325: if (scall == MAT_INITIAL_MATRIX) {
5326: PetscCall(PetscObjectReference((PetscObject)Ad));
5327: *A_loc = Ad;
5328: } else if (scall == MAT_REUSE_MATRIX) {
5329: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5330: }
5331: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5332: PetscFunctionReturn(PETSC_SUCCESS);
5333: }
5334: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5335: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5336: if (f) {
5337: PetscCall((*f)(A, scall, glob, A_loc));
5338: } else {
5339: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5340: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5341: Mat_SeqAIJ *c;
5342: PetscInt *ai = a->i, *aj = a->j;
5343: PetscInt *bi = b->i, *bj = b->j;
5344: PetscInt *ci, *cj;
5345: const PetscScalar *aa, *ba;
5346: PetscScalar *ca;
5347: PetscInt i, j, am, dn, on;
5349: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5350: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5351: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5352: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5353: if (scall == MAT_INITIAL_MATRIX) {
5354: PetscInt k;
5355: PetscCall(PetscMalloc1(1 + am, &ci));
5356: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5357: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5358: ci[0] = 0;
5359: for (i = 0, k = 0; i < am; i++) {
5360: const PetscInt ncols_o = bi[i + 1] - bi[i];
5361: const PetscInt ncols_d = ai[i + 1] - ai[i];
5362: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5363: /* diagonal portion of A */
5364: for (j = 0; j < ncols_d; j++, k++) {
5365: cj[k] = *aj++;
5366: ca[k] = *aa++;
5367: }
5368: /* off-diagonal portion of A */
5369: for (j = 0; j < ncols_o; j++, k++) {
5370: cj[k] = dn + *bj++;
5371: ca[k] = *ba++;
5372: }
5373: }
5374: /* put together the new matrix */
5375: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5376: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5377: /* Since these are PETSc arrays, change flags to free them as necessary. */
5378: c = (Mat_SeqAIJ *)(*A_loc)->data;
5379: c->free_a = PETSC_TRUE;
5380: c->free_ij = PETSC_TRUE;
5381: c->nonew = 0;
5382: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5383: } else if (scall == MAT_REUSE_MATRIX) {
5384: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5385: for (i = 0; i < am; i++) {
5386: const PetscInt ncols_d = ai[i + 1] - ai[i];
5387: const PetscInt ncols_o = bi[i + 1] - bi[i];
5388: /* diagonal portion of A */
5389: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5390: /* off-diagonal portion of A */
5391: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5392: }
5393: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5394: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5395: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5396: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5397: if (glob) {
5398: PetscInt cst, *gidx;
5400: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5401: PetscCall(PetscMalloc1(dn + on, &gidx));
5402: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5403: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5404: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5405: }
5406: }
5407: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5408: PetscFunctionReturn(PETSC_SUCCESS);
5409: }
5411: /*@C
5412: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5414: Not Collective
5416: Input Parameters:
5417: + A - the matrix
5418: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5419: . row - index set of rows to extract (or `NULL`)
5420: - col - index set of columns to extract (or `NULL`)
5422: Output Parameter:
5423: . A_loc - the local sequential matrix generated
5425: Level: developer
5427: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5428: @*/
5429: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5430: {
5431: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5432: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5433: IS isrowa, iscola;
5434: Mat *aloc;
5435: PetscBool match;
5437: PetscFunctionBegin;
5438: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5439: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5440: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5441: if (!row) {
5442: start = A->rmap->rstart;
5443: end = A->rmap->rend;
5444: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5445: } else {
5446: isrowa = *row;
5447: }
5448: if (!col) {
5449: start = A->cmap->rstart;
5450: cmap = a->garray;
5451: nzA = a->A->cmap->n;
5452: nzB = a->B->cmap->n;
5453: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5454: ncols = 0;
5455: for (i = 0; i < nzB; i++) {
5456: if (cmap[i] < start) idx[ncols++] = cmap[i];
5457: else break;
5458: }
5459: imark = i;
5460: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5461: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5462: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5463: } else {
5464: iscola = *col;
5465: }
5466: if (scall != MAT_INITIAL_MATRIX) {
5467: PetscCall(PetscMalloc1(1, &aloc));
5468: aloc[0] = *A_loc;
5469: }
5470: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5471: if (!col) { /* attach global id of condensed columns */
5472: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5473: }
5474: *A_loc = aloc[0];
5475: PetscCall(PetscFree(aloc));
5476: if (!row) PetscCall(ISDestroy(&isrowa));
5477: if (!col) PetscCall(ISDestroy(&iscola));
5478: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5479: PetscFunctionReturn(PETSC_SUCCESS);
5480: }
5482: /*
5483: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5484: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5485: * on a global size.
5486: * */
5487: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5488: {
5489: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5490: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5491: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5492: PetscMPIInt owner;
5493: PetscSFNode *iremote, *oiremote;
5494: const PetscInt *lrowindices;
5495: PetscSF sf, osf;
5496: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5497: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5498: MPI_Comm comm;
5499: ISLocalToGlobalMapping mapping;
5500: const PetscScalar *pd_a, *po_a;
5502: PetscFunctionBegin;
5503: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5504: /* plocalsize is the number of roots
5505: * nrows is the number of leaves
5506: * */
5507: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5508: PetscCall(ISGetLocalSize(rows, &nrows));
5509: PetscCall(PetscCalloc1(nrows, &iremote));
5510: PetscCall(ISGetIndices(rows, &lrowindices));
5511: for (i = 0; i < nrows; i++) {
5512: /* Find a remote index and an owner for a row
5513: * The row could be local or remote
5514: * */
5515: owner = 0;
5516: lidx = 0;
5517: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5518: iremote[i].index = lidx;
5519: iremote[i].rank = owner;
5520: }
5521: /* Create SF to communicate how many nonzero columns for each row */
5522: PetscCall(PetscSFCreate(comm, &sf));
5523: /* SF will figure out the number of nonzero columns for each row, and their
5524: * offsets
5525: * */
5526: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5527: PetscCall(PetscSFSetFromOptions(sf));
5528: PetscCall(PetscSFSetUp(sf));
5530: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5531: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5532: PetscCall(PetscCalloc1(nrows, &pnnz));
5533: roffsets[0] = 0;
5534: roffsets[1] = 0;
5535: for (i = 0; i < plocalsize; i++) {
5536: /* diagonal */
5537: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5538: /* off-diagonal */
5539: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5540: /* compute offsets so that we relative location for each row */
5541: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5542: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5543: }
5544: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5545: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5546: /* 'r' means root, and 'l' means leaf */
5547: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5548: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5549: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5550: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5551: PetscCall(PetscSFDestroy(&sf));
5552: PetscCall(PetscFree(roffsets));
5553: PetscCall(PetscFree(nrcols));
5554: dntotalcols = 0;
5555: ontotalcols = 0;
5556: ncol = 0;
5557: for (i = 0; i < nrows; i++) {
5558: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5559: ncol = PetscMax(pnnz[i], ncol);
5560: /* diagonal */
5561: dntotalcols += nlcols[i * 2 + 0];
5562: /* off-diagonal */
5563: ontotalcols += nlcols[i * 2 + 1];
5564: }
5565: /* We do not need to figure the right number of columns
5566: * since all the calculations will be done by going through the raw data
5567: * */
5568: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5569: PetscCall(MatSetUp(*P_oth));
5570: PetscCall(PetscFree(pnnz));
5571: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5572: /* diagonal */
5573: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5574: /* off-diagonal */
5575: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5576: /* diagonal */
5577: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5578: /* off-diagonal */
5579: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5580: dntotalcols = 0;
5581: ontotalcols = 0;
5582: ntotalcols = 0;
5583: for (i = 0; i < nrows; i++) {
5584: owner = 0;
5585: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5586: /* Set iremote for diag matrix */
5587: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5588: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5589: iremote[dntotalcols].rank = owner;
5590: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5591: ilocal[dntotalcols++] = ntotalcols++;
5592: }
5593: /* off-diagonal */
5594: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5595: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5596: oiremote[ontotalcols].rank = owner;
5597: oilocal[ontotalcols++] = ntotalcols++;
5598: }
5599: }
5600: PetscCall(ISRestoreIndices(rows, &lrowindices));
5601: PetscCall(PetscFree(loffsets));
5602: PetscCall(PetscFree(nlcols));
5603: PetscCall(PetscSFCreate(comm, &sf));
5604: /* P serves as roots and P_oth is leaves
5605: * Diag matrix
5606: * */
5607: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5608: PetscCall(PetscSFSetFromOptions(sf));
5609: PetscCall(PetscSFSetUp(sf));
5611: PetscCall(PetscSFCreate(comm, &osf));
5612: /* off-diagonal */
5613: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5614: PetscCall(PetscSFSetFromOptions(osf));
5615: PetscCall(PetscSFSetUp(osf));
5616: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5617: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5618: /* operate on the matrix internal data to save memory */
5619: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5620: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5621: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5622: /* Convert to global indices for diag matrix */
5623: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5624: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5625: /* We want P_oth store global indices */
5626: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5627: /* Use memory scalable approach */
5628: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5629: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5630: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5631: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5632: /* Convert back to local indices */
5633: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5634: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5635: nout = 0;
5636: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5637: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5638: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5639: /* Exchange values */
5640: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5641: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5642: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5643: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5644: /* Stop PETSc from shrinking memory */
5645: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5646: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5647: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5648: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5649: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5650: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5651: PetscCall(PetscSFDestroy(&sf));
5652: PetscCall(PetscSFDestroy(&osf));
5653: PetscFunctionReturn(PETSC_SUCCESS);
5654: }
5656: /*
5657: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5658: * This supports MPIAIJ and MAIJ
5659: * */
5660: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5661: {
5662: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5663: Mat_SeqAIJ *p_oth;
5664: IS rows, map;
5665: PetscHMapI hamp;
5666: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5667: MPI_Comm comm;
5668: PetscSF sf, osf;
5669: PetscBool has;
5671: PetscFunctionBegin;
5672: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5673: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5674: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5675: * and then create a submatrix (that often is an overlapping matrix)
5676: * */
5677: if (reuse == MAT_INITIAL_MATRIX) {
5678: /* Use a hash table to figure out unique keys */
5679: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5680: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5681: count = 0;
5682: /* Assume that a->g is sorted, otherwise the following does not make sense */
5683: for (i = 0; i < a->B->cmap->n; i++) {
5684: key = a->garray[i] / dof;
5685: PetscCall(PetscHMapIHas(hamp, key, &has));
5686: if (!has) {
5687: mapping[i] = count;
5688: PetscCall(PetscHMapISet(hamp, key, count++));
5689: } else {
5690: /* Current 'i' has the same value the previous step */
5691: mapping[i] = count - 1;
5692: }
5693: }
5694: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5695: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5696: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5697: PetscCall(PetscCalloc1(htsize, &rowindices));
5698: off = 0;
5699: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5700: PetscCall(PetscHMapIDestroy(&hamp));
5701: PetscCall(PetscSortInt(htsize, rowindices));
5702: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5703: /* In case, the matrix was already created but users want to recreate the matrix */
5704: PetscCall(MatDestroy(P_oth));
5705: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5706: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5707: PetscCall(ISDestroy(&map));
5708: PetscCall(ISDestroy(&rows));
5709: } else if (reuse == MAT_REUSE_MATRIX) {
5710: /* If matrix was already created, we simply update values using SF objects
5711: * that as attached to the matrix earlier.
5712: */
5713: const PetscScalar *pd_a, *po_a;
5715: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5716: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5717: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5718: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5719: /* Update values in place */
5720: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5721: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5722: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5723: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5724: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5725: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5726: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5727: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5728: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5729: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5730: PetscFunctionReturn(PETSC_SUCCESS);
5731: }
5733: /*@C
5734: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5736: Collective
5738: Input Parameters:
5739: + A - the first matrix in `MATMPIAIJ` format
5740: . B - the second matrix in `MATMPIAIJ` format
5741: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5743: Output Parameters:
5744: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5745: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5746: - B_seq - the sequential matrix generated
5748: Level: developer
5750: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5751: @*/
5752: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5753: {
5754: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5755: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5756: IS isrowb, iscolb;
5757: Mat *bseq = NULL;
5759: PetscFunctionBegin;
5760: 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 ")",
5761: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5762: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5764: if (scall == MAT_INITIAL_MATRIX) {
5765: start = A->cmap->rstart;
5766: cmap = a->garray;
5767: nzA = a->A->cmap->n;
5768: nzB = a->B->cmap->n;
5769: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5770: ncols = 0;
5771: for (i = 0; i < nzB; i++) { /* row < local row index */
5772: if (cmap[i] < start) idx[ncols++] = cmap[i];
5773: else break;
5774: }
5775: imark = i;
5776: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5777: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5778: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5779: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5780: } else {
5781: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5782: isrowb = *rowb;
5783: iscolb = *colb;
5784: PetscCall(PetscMalloc1(1, &bseq));
5785: bseq[0] = *B_seq;
5786: }
5787: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5788: *B_seq = bseq[0];
5789: PetscCall(PetscFree(bseq));
5790: if (!rowb) {
5791: PetscCall(ISDestroy(&isrowb));
5792: } else {
5793: *rowb = isrowb;
5794: }
5795: if (!colb) {
5796: PetscCall(ISDestroy(&iscolb));
5797: } else {
5798: *colb = iscolb;
5799: }
5800: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5801: PetscFunctionReturn(PETSC_SUCCESS);
5802: }
5804: /*
5805: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5806: of the OFF-DIAGONAL portion of local A
5808: Collective
5810: Input Parameters:
5811: + A,B - the matrices in `MATMPIAIJ` format
5812: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5814: Output Parameter:
5815: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5816: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5817: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5818: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5820: Developer Note:
5821: This directly accesses information inside the VecScatter associated with the matrix-vector product
5822: for this matrix. This is not desirable..
5824: Level: developer
5826: */
5827: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5828: {
5829: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5830: Mat_SeqAIJ *b_oth;
5831: VecScatter ctx;
5832: MPI_Comm comm;
5833: const PetscMPIInt *rprocs, *sprocs;
5834: const PetscInt *srow, *rstarts, *sstarts;
5835: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5836: PetscInt i, j, k = 0, l, ll, nrecvs, nsends, nrows, *rstartsj = NULL, *sstartsj, len;
5837: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5838: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5839: PetscMPIInt size, tag, rank, nreqs;
5841: PetscFunctionBegin;
5842: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5843: PetscCallMPI(MPI_Comm_size(comm, &size));
5845: 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 ")",
5846: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5847: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5848: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5850: if (size == 1) {
5851: startsj_s = NULL;
5852: bufa_ptr = NULL;
5853: *B_oth = NULL;
5854: PetscFunctionReturn(PETSC_SUCCESS);
5855: }
5857: ctx = a->Mvctx;
5858: tag = ((PetscObject)ctx)->tag;
5860: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5861: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5862: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5863: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5864: PetscCall(PetscMalloc1(nreqs, &reqs));
5865: rwaits = reqs;
5866: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5868: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5869: if (scall == MAT_INITIAL_MATRIX) {
5870: /* i-array */
5871: /* post receives */
5872: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5873: for (i = 0; i < nrecvs; i++) {
5874: rowlen = rvalues + rstarts[i] * rbs;
5875: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5876: PetscCallMPI(MPI_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5877: }
5879: /* pack the outgoing message */
5880: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5882: sstartsj[0] = 0;
5883: rstartsj[0] = 0;
5884: len = 0; /* total length of j or a array to be sent */
5885: if (nsends) {
5886: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5887: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5888: }
5889: for (i = 0; i < nsends; i++) {
5890: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5891: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5892: for (j = 0; j < nrows; j++) {
5893: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5894: for (l = 0; l < sbs; l++) {
5895: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5897: rowlen[j * sbs + l] = ncols;
5899: len += ncols;
5900: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5901: }
5902: k++;
5903: }
5904: PetscCallMPI(MPI_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5906: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5907: }
5908: /* recvs and sends of i-array are completed */
5909: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5910: PetscCall(PetscFree(svalues));
5912: /* allocate buffers for sending j and a arrays */
5913: PetscCall(PetscMalloc1(len + 1, &bufj));
5914: PetscCall(PetscMalloc1(len + 1, &bufa));
5916: /* create i-array of B_oth */
5917: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5919: b_othi[0] = 0;
5920: len = 0; /* total length of j or a array to be received */
5921: k = 0;
5922: for (i = 0; i < nrecvs; i++) {
5923: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5924: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5925: for (j = 0; j < nrows; j++) {
5926: b_othi[k + 1] = b_othi[k] + rowlen[j];
5927: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5928: k++;
5929: }
5930: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5931: }
5932: PetscCall(PetscFree(rvalues));
5934: /* allocate space for j and a arrays of B_oth */
5935: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5936: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5938: /* j-array */
5939: /* post receives of j-array */
5940: for (i = 0; i < nrecvs; i++) {
5941: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5942: PetscCallMPI(MPI_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5943: }
5945: /* pack the outgoing message j-array */
5946: if (nsends) k = sstarts[0];
5947: for (i = 0; i < nsends; i++) {
5948: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5949: bufJ = bufj + sstartsj[i];
5950: for (j = 0; j < nrows; j++) {
5951: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5952: for (ll = 0; ll < sbs; ll++) {
5953: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5954: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5955: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5956: }
5957: }
5958: PetscCallMPI(MPI_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5959: }
5961: /* recvs and sends of j-array are completed */
5962: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5963: } else if (scall == MAT_REUSE_MATRIX) {
5964: sstartsj = *startsj_s;
5965: rstartsj = *startsj_r;
5966: bufa = *bufa_ptr;
5967: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
5968: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5969: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5971: /* a-array */
5972: /* post receives of a-array */
5973: for (i = 0; i < nrecvs; i++) {
5974: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5975: PetscCallMPI(MPI_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5976: }
5978: /* pack the outgoing message a-array */
5979: if (nsends) k = sstarts[0];
5980: for (i = 0; i < nsends; i++) {
5981: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5982: bufA = bufa + sstartsj[i];
5983: for (j = 0; j < nrows; j++) {
5984: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5985: for (ll = 0; ll < sbs; ll++) {
5986: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5987: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5988: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5989: }
5990: }
5991: PetscCallMPI(MPI_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5992: }
5993: /* recvs and sends of a-array are completed */
5994: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5995: PetscCall(PetscFree(reqs));
5997: if (scall == MAT_INITIAL_MATRIX) {
5998: /* put together the new matrix */
5999: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6001: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6002: /* Since these are PETSc arrays, change flags to free them as necessary. */
6003: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6004: b_oth->free_a = PETSC_TRUE;
6005: b_oth->free_ij = PETSC_TRUE;
6006: b_oth->nonew = 0;
6008: PetscCall(PetscFree(bufj));
6009: if (!startsj_s || !bufa_ptr) {
6010: PetscCall(PetscFree2(sstartsj, rstartsj));
6011: PetscCall(PetscFree(bufa_ptr));
6012: } else {
6013: *startsj_s = sstartsj;
6014: *startsj_r = rstartsj;
6015: *bufa_ptr = bufa;
6016: }
6017: } else if (scall == MAT_REUSE_MATRIX) {
6018: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6019: }
6021: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6022: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6023: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6024: PetscFunctionReturn(PETSC_SUCCESS);
6025: }
6027: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6028: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6029: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6030: #if defined(PETSC_HAVE_MKL_SPARSE)
6031: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6032: #endif
6033: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6034: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6035: #if defined(PETSC_HAVE_ELEMENTAL)
6036: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6037: #endif
6038: #if defined(PETSC_HAVE_SCALAPACK)
6039: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6040: #endif
6041: #if defined(PETSC_HAVE_HYPRE)
6042: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6043: #endif
6044: #if defined(PETSC_HAVE_CUDA)
6045: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6046: #endif
6047: #if defined(PETSC_HAVE_HIP)
6048: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6049: #endif
6050: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6051: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6052: #endif
6053: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6054: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6055: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6057: /*
6058: Computes (B'*A')' since computing B*A directly is untenable
6060: n p p
6061: [ ] [ ] [ ]
6062: m [ A ] * n [ B ] = m [ C ]
6063: [ ] [ ] [ ]
6065: */
6066: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6067: {
6068: Mat At, Bt, Ct;
6070: PetscFunctionBegin;
6071: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6072: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6073: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Ct));
6074: PetscCall(MatDestroy(&At));
6075: PetscCall(MatDestroy(&Bt));
6076: PetscCall(MatTransposeSetPrecursor(Ct, C));
6077: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6078: PetscCall(MatDestroy(&Ct));
6079: PetscFunctionReturn(PETSC_SUCCESS);
6080: }
6082: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6083: {
6084: PetscBool cisdense;
6086: PetscFunctionBegin;
6087: 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);
6088: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6089: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6090: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6091: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6092: PetscCall(MatSetUp(C));
6094: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6095: PetscFunctionReturn(PETSC_SUCCESS);
6096: }
6098: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6099: {
6100: Mat_Product *product = C->product;
6101: Mat A = product->A, B = product->B;
6103: PetscFunctionBegin;
6104: 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 ")",
6105: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6106: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6107: C->ops->productsymbolic = MatProductSymbolic_AB;
6108: PetscFunctionReturn(PETSC_SUCCESS);
6109: }
6111: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6112: {
6113: Mat_Product *product = C->product;
6115: PetscFunctionBegin;
6116: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6117: PetscFunctionReturn(PETSC_SUCCESS);
6118: }
6120: /*
6121: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6123: Input Parameters:
6125: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6126: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6128: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6130: For Set1, j1[] contains column indices of the nonzeros.
6131: 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
6132: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6133: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6135: Similar for Set2.
6137: This routine merges the two sets of nonzeros row by row and removes repeats.
6139: Output Parameters: (memory is allocated by the caller)
6141: i[],j[]: the CSR of the merged matrix, which has m rows.
6142: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6143: imap2[]: similar to imap1[], but for Set2.
6144: Note we order nonzeros row-by-row and from left to right.
6145: */
6146: 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[])
6147: {
6148: PetscInt r, m; /* Row index of mat */
6149: PetscCount t, t1, t2, b1, e1, b2, e2;
6151: PetscFunctionBegin;
6152: PetscCall(MatGetLocalSize(mat, &m, NULL));
6153: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6154: i[0] = 0;
6155: for (r = 0; r < m; r++) { /* Do row by row merging */
6156: b1 = rowBegin1[r];
6157: e1 = rowEnd1[r];
6158: b2 = rowBegin2[r];
6159: e2 = rowEnd2[r];
6160: while (b1 < e1 && b2 < e2) {
6161: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6162: j[t] = j1[b1];
6163: imap1[t1] = t;
6164: imap2[t2] = t;
6165: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6166: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6167: t1++;
6168: t2++;
6169: t++;
6170: } else if (j1[b1] < j2[b2]) {
6171: j[t] = j1[b1];
6172: imap1[t1] = t;
6173: b1 += jmap1[t1 + 1] - jmap1[t1];
6174: t1++;
6175: t++;
6176: } else {
6177: j[t] = j2[b2];
6178: imap2[t2] = t;
6179: b2 += jmap2[t2 + 1] - jmap2[t2];
6180: t2++;
6181: t++;
6182: }
6183: }
6184: /* Merge the remaining in either j1[] or j2[] */
6185: while (b1 < e1) {
6186: j[t] = j1[b1];
6187: imap1[t1] = t;
6188: b1 += jmap1[t1 + 1] - jmap1[t1];
6189: t1++;
6190: t++;
6191: }
6192: while (b2 < e2) {
6193: j[t] = j2[b2];
6194: imap2[t2] = t;
6195: b2 += jmap2[t2 + 1] - jmap2[t2];
6196: t2++;
6197: t++;
6198: }
6199: i[r + 1] = t;
6200: }
6201: PetscFunctionReturn(PETSC_SUCCESS);
6202: }
6204: /*
6205: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6207: Input Parameters:
6208: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6209: 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[]
6210: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6212: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6213: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6215: Output Parameters:
6216: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6217: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6218: 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,
6219: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6221: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6222: Atot: number of entries belonging to the diagonal block.
6223: Annz: number of unique nonzeros belonging to the diagonal block.
6224: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6225: repeats (i.e., same 'i,j' pair).
6226: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6227: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6229: Atot: number of entries belonging to the diagonal block
6230: Annz: number of unique nonzeros belonging to the diagonal block.
6232: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6234: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6235: */
6236: 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_)
6237: {
6238: PetscInt cstart, cend, rstart, rend, row, col;
6239: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6240: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6241: PetscCount k, m, p, q, r, s, mid;
6242: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6244: PetscFunctionBegin;
6245: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6246: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6247: m = rend - rstart;
6249: /* Skip negative rows */
6250: for (k = 0; k < n; k++)
6251: if (i[k] >= 0) break;
6253: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6254: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6255: */
6256: while (k < n) {
6257: row = i[k];
6258: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6259: for (s = k; s < n; s++)
6260: if (i[s] != row) break;
6262: /* Shift diag columns to range of [-PETSC_MAX_INT, -1] */
6263: for (p = k; p < s; p++) {
6264: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_MAX_INT;
6265: 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]);
6266: }
6267: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6268: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6269: rowBegin[row - rstart] = k;
6270: rowMid[row - rstart] = mid;
6271: rowEnd[row - rstart] = s;
6273: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6274: Atot += mid - k;
6275: Btot += s - mid;
6277: /* Count unique nonzeros of this diag row */
6278: for (p = k; p < mid;) {
6279: col = j[p];
6280: do {
6281: j[p] += PETSC_MAX_INT; /* Revert the modified diagonal indices */
6282: p++;
6283: } while (p < mid && j[p] == col);
6284: Annz++;
6285: }
6287: /* Count unique nonzeros of this offdiag row */
6288: for (p = mid; p < s;) {
6289: col = j[p];
6290: do {
6291: p++;
6292: } while (p < s && j[p] == col);
6293: Bnnz++;
6294: }
6295: k = s;
6296: }
6298: /* Allocation according to Atot, Btot, Annz, Bnnz */
6299: PetscCall(PetscMalloc1(Atot, &Aperm));
6300: PetscCall(PetscMalloc1(Btot, &Bperm));
6301: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6302: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6304: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6305: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6306: for (r = 0; r < m; r++) {
6307: k = rowBegin[r];
6308: mid = rowMid[r];
6309: s = rowEnd[r];
6310: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6311: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6312: Atot += mid - k;
6313: Btot += s - mid;
6315: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6316: for (p = k; p < mid;) {
6317: col = j[p];
6318: q = p;
6319: do {
6320: p++;
6321: } while (p < mid && j[p] == col);
6322: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6323: Annz++;
6324: }
6326: for (p = mid; p < s;) {
6327: col = j[p];
6328: q = p;
6329: do {
6330: p++;
6331: } while (p < s && j[p] == col);
6332: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6333: Bnnz++;
6334: }
6335: }
6336: /* Output */
6337: *Aperm_ = Aperm;
6338: *Annz_ = Annz;
6339: *Atot_ = Atot;
6340: *Ajmap_ = Ajmap;
6341: *Bperm_ = Bperm;
6342: *Bnnz_ = Bnnz;
6343: *Btot_ = Btot;
6344: *Bjmap_ = Bjmap;
6345: PetscFunctionReturn(PETSC_SUCCESS);
6346: }
6348: /*
6349: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6351: Input Parameters:
6352: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6353: nnz: number of unique nonzeros in the merged matrix
6354: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6355: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6357: Output Parameter: (memory is allocated by the caller)
6358: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6360: Example:
6361: nnz1 = 4
6362: nnz = 6
6363: imap = [1,3,4,5]
6364: jmap = [0,3,5,6,7]
6365: then,
6366: jmap_new = [0,0,3,3,5,6,7]
6367: */
6368: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6369: {
6370: PetscCount k, p;
6372: PetscFunctionBegin;
6373: jmap_new[0] = 0;
6374: p = nnz; /* p loops over jmap_new[] backwards */
6375: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6376: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6377: }
6378: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6379: PetscFunctionReturn(PETSC_SUCCESS);
6380: }
6382: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6383: {
6384: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6386: PetscFunctionBegin;
6387: PetscCall(PetscSFDestroy(&coo->sf));
6388: PetscCall(PetscFree(coo->Aperm1));
6389: PetscCall(PetscFree(coo->Bperm1));
6390: PetscCall(PetscFree(coo->Ajmap1));
6391: PetscCall(PetscFree(coo->Bjmap1));
6392: PetscCall(PetscFree(coo->Aimap2));
6393: PetscCall(PetscFree(coo->Bimap2));
6394: PetscCall(PetscFree(coo->Aperm2));
6395: PetscCall(PetscFree(coo->Bperm2));
6396: PetscCall(PetscFree(coo->Ajmap2));
6397: PetscCall(PetscFree(coo->Bjmap2));
6398: PetscCall(PetscFree(coo->Cperm1));
6399: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6400: PetscCall(PetscFree(coo));
6401: PetscFunctionReturn(PETSC_SUCCESS);
6402: }
6404: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6405: {
6406: MPI_Comm comm;
6407: PetscMPIInt rank, size;
6408: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6409: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6410: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6411: PetscContainer container;
6412: MatCOOStruct_MPIAIJ *coo;
6414: PetscFunctionBegin;
6415: PetscCall(PetscFree(mpiaij->garray));
6416: PetscCall(VecDestroy(&mpiaij->lvec));
6417: #if defined(PETSC_USE_CTABLE)
6418: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6419: #else
6420: PetscCall(PetscFree(mpiaij->colmap));
6421: #endif
6422: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6423: mat->assembled = PETSC_FALSE;
6424: mat->was_assembled = PETSC_FALSE;
6426: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6427: PetscCallMPI(MPI_Comm_size(comm, &size));
6428: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6429: PetscCall(PetscLayoutSetUp(mat->rmap));
6430: PetscCall(PetscLayoutSetUp(mat->cmap));
6431: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6432: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6433: PetscCall(MatGetLocalSize(mat, &m, &n));
6434: PetscCall(MatGetSize(mat, &M, &N));
6436: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6437: /* entries come first, then local rows, then remote rows. */
6438: PetscCount n1 = coo_n, *perm1;
6439: PetscInt *i1 = coo_i, *j1 = coo_j;
6441: PetscCall(PetscMalloc1(n1, &perm1));
6442: for (k = 0; k < n1; k++) perm1[k] = k;
6444: /* Manipulate indices so that entries with negative row or col indices will have smallest
6445: row indices, local entries will have greater but negative row indices, and remote entries
6446: will have positive row indices.
6447: */
6448: for (k = 0; k < n1; k++) {
6449: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_MIN_INT; /* e.g., -2^31, minimal to move them ahead */
6450: 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] */
6451: else {
6452: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6453: if (mpiaij->donotstash) i1[k] = PETSC_MIN_INT; /* Ignore offproc entries as if they had negative indices */
6454: }
6455: }
6457: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6458: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6460: /* Advance k to the first entry we need to take care of */
6461: for (k = 0; k < n1; k++)
6462: if (i1[k] > PETSC_MIN_INT) break;
6463: PetscInt i1start = k;
6465: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_MAX_INT, &rem)); /* rem is upper bound of the last local row */
6466: for (; k < rem; k++) i1[k] += PETSC_MAX_INT; /* Revert row indices of local rows*/
6468: /* Send remote rows to their owner */
6469: /* Find which rows should be sent to which remote ranks*/
6470: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6471: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6472: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6473: const PetscInt *ranges;
6474: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6476: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6477: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6478: for (k = rem; k < n1;) {
6479: PetscMPIInt owner;
6480: PetscInt firstRow, lastRow;
6482: /* Locate a row range */
6483: firstRow = i1[k]; /* first row of this owner */
6484: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6485: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6487: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6488: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6490: /* All entries in [k,p) belong to this remote owner */
6491: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6492: PetscMPIInt *sendto2;
6493: PetscInt *nentries2;
6494: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6496: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6497: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6498: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6499: PetscCall(PetscFree2(sendto, nentries2));
6500: sendto = sendto2;
6501: nentries = nentries2;
6502: maxNsend = maxNsend2;
6503: }
6504: sendto[nsend] = owner;
6505: nentries[nsend] = p - k;
6506: PetscCall(PetscCountCast(p - k, &nentries[nsend]));
6507: nsend++;
6508: k = p;
6509: }
6511: /* Build 1st SF to know offsets on remote to send data */
6512: PetscSF sf1;
6513: PetscInt nroots = 1, nroots2 = 0;
6514: PetscInt nleaves = nsend, nleaves2 = 0;
6515: PetscInt *offsets;
6516: PetscSFNode *iremote;
6518: PetscCall(PetscSFCreate(comm, &sf1));
6519: PetscCall(PetscMalloc1(nsend, &iremote));
6520: PetscCall(PetscMalloc1(nsend, &offsets));
6521: for (k = 0; k < nsend; k++) {
6522: iremote[k].rank = sendto[k];
6523: iremote[k].index = 0;
6524: nleaves2 += nentries[k];
6525: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6526: }
6527: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6528: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6529: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6530: PetscCall(PetscSFDestroy(&sf1));
6531: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6533: /* Build 2nd SF to send remote COOs to their owner */
6534: PetscSF sf2;
6535: nroots = nroots2;
6536: nleaves = nleaves2;
6537: PetscCall(PetscSFCreate(comm, &sf2));
6538: PetscCall(PetscSFSetFromOptions(sf2));
6539: PetscCall(PetscMalloc1(nleaves, &iremote));
6540: p = 0;
6541: for (k = 0; k < nsend; k++) {
6542: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6543: for (q = 0; q < nentries[k]; q++, p++) {
6544: iremote[p].rank = sendto[k];
6545: iremote[p].index = offsets[k] + q;
6546: }
6547: }
6548: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6550: /* Send the remote COOs to their owner */
6551: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6552: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6553: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6554: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6555: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6556: PetscInt *i1prem = i1 ? i1 + rem : NULL; /* silence ubsan warnings about pointer arithmetic on null pointer */
6557: PetscInt *j1prem = j1 ? j1 + rem : NULL;
6558: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6559: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6560: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6561: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, 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: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6573: PetscCount *perm1prem = perm1 ? perm1 + rem : NULL;
6574: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6575: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6577: /* Support for HYPRE matrices, kind of a hack.
6578: Swap min column with diagonal so that diagonal values will go first */
6579: PetscBool hypre;
6580: const char *name;
6581: PetscCall(PetscObjectGetName((PetscObject)mat, &name));
6582: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
6583: if (hypre) {
6584: PetscInt *minj;
6585: PetscBT hasdiag;
6587: PetscCall(PetscBTCreate(m, &hasdiag));
6588: PetscCall(PetscMalloc1(m, &minj));
6589: for (k = 0; k < m; k++) minj[k] = PETSC_MAX_INT;
6590: for (k = i1start; k < rem; k++) {
6591: if (j1[k] < cstart || j1[k] >= cend) continue;
6592: const PetscInt rindex = i1[k] - rstart;
6593: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6594: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6595: }
6596: for (k = 0; k < n2; k++) {
6597: if (j2[k] < cstart || j2[k] >= cend) continue;
6598: const PetscInt rindex = i2[k] - rstart;
6599: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6600: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6601: }
6602: for (k = i1start; k < rem; k++) {
6603: const PetscInt rindex = i1[k] - rstart;
6604: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6605: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6606: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6607: }
6608: for (k = 0; k < n2; k++) {
6609: const PetscInt rindex = i2[k] - rstart;
6610: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6611: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6612: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6613: }
6614: PetscCall(PetscBTDestroy(&hasdiag));
6615: PetscCall(PetscFree(minj));
6616: }
6618: /* Split local COOs and received COOs into diag/offdiag portions */
6619: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6620: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6621: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6622: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6623: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6624: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6626: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6627: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6628: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6629: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6631: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6632: PetscInt *Ai, *Bi;
6633: PetscInt *Aj, *Bj;
6635: PetscCall(PetscMalloc1(m + 1, &Ai));
6636: PetscCall(PetscMalloc1(m + 1, &Bi));
6637: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6638: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6640: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6641: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6642: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6643: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6644: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6646: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6647: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6649: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6650: /* expect nonzeros in A/B most likely have local contributing entries */
6651: PetscInt Annz = Ai[m];
6652: PetscInt Bnnz = Bi[m];
6653: PetscCount *Ajmap1_new, *Bjmap1_new;
6655: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6656: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6658: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6659: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6661: PetscCall(PetscFree(Aimap1));
6662: PetscCall(PetscFree(Ajmap1));
6663: PetscCall(PetscFree(Bimap1));
6664: PetscCall(PetscFree(Bjmap1));
6665: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6666: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6667: PetscCall(PetscFree(perm1));
6668: PetscCall(PetscFree3(i2, j2, perm2));
6670: Ajmap1 = Ajmap1_new;
6671: Bjmap1 = Bjmap1_new;
6673: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6674: if (Annz < Annz1 + Annz2) {
6675: PetscInt *Aj_new;
6676: PetscCall(PetscMalloc1(Annz, &Aj_new));
6677: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6678: PetscCall(PetscFree(Aj));
6679: Aj = Aj_new;
6680: }
6682: if (Bnnz < Bnnz1 + Bnnz2) {
6683: PetscInt *Bj_new;
6684: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6685: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6686: PetscCall(PetscFree(Bj));
6687: Bj = Bj_new;
6688: }
6690: /* Create new submatrices for on-process and off-process coupling */
6691: PetscScalar *Aa, *Ba;
6692: MatType rtype;
6693: Mat_SeqAIJ *a, *b;
6694: PetscObjectState state;
6695: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6696: PetscCall(PetscCalloc1(Bnnz, &Ba));
6697: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6698: if (cstart) {
6699: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6700: }
6702: PetscCall(MatGetRootType_Private(mat, &rtype));
6704: MatSeqXAIJGetOptions_Private(mpiaij->A);
6705: PetscCall(MatDestroy(&mpiaij->A));
6706: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6707: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6708: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6710: MatSeqXAIJGetOptions_Private(mpiaij->B);
6711: PetscCall(MatDestroy(&mpiaij->B));
6712: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6713: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6714: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6716: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6717: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6718: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6719: PetscCall(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6721: a = (Mat_SeqAIJ *)mpiaij->A->data;
6722: b = (Mat_SeqAIJ *)mpiaij->B->data;
6723: a->singlemalloc = b->singlemalloc = PETSC_FALSE; /* Let newmat own Ai,Aj,Aa,Bi,Bj,Ba */
6724: a->free_a = b->free_a = PETSC_TRUE;
6725: a->free_ij = b->free_ij = PETSC_TRUE;
6727: /* conversion must happen AFTER multiply setup */
6728: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6729: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6730: PetscCall(VecDestroy(&mpiaij->lvec));
6731: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6733: // Put the COO struct in a container and then attach that to the matrix
6734: PetscCall(PetscMalloc1(1, &coo));
6735: coo->n = coo_n;
6736: coo->sf = sf2;
6737: coo->sendlen = nleaves;
6738: coo->recvlen = nroots;
6739: coo->Annz = Annz;
6740: coo->Bnnz = Bnnz;
6741: coo->Annz2 = Annz2;
6742: coo->Bnnz2 = Bnnz2;
6743: coo->Atot1 = Atot1;
6744: coo->Atot2 = Atot2;
6745: coo->Btot1 = Btot1;
6746: coo->Btot2 = Btot2;
6747: coo->Ajmap1 = Ajmap1;
6748: coo->Aperm1 = Aperm1;
6749: coo->Bjmap1 = Bjmap1;
6750: coo->Bperm1 = Bperm1;
6751: coo->Aimap2 = Aimap2;
6752: coo->Ajmap2 = Ajmap2;
6753: coo->Aperm2 = Aperm2;
6754: coo->Bimap2 = Bimap2;
6755: coo->Bjmap2 = Bjmap2;
6756: coo->Bperm2 = Bperm2;
6757: coo->Cperm1 = Cperm1;
6758: // Allocate in preallocation. If not used, it has zero cost on host
6759: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6760: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6761: PetscCall(PetscContainerSetPointer(container, coo));
6762: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6763: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6764: PetscCall(PetscContainerDestroy(&container));
6765: PetscFunctionReturn(PETSC_SUCCESS);
6766: }
6768: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6769: {
6770: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6771: Mat A = mpiaij->A, B = mpiaij->B;
6772: PetscScalar *Aa, *Ba;
6773: PetscScalar *sendbuf, *recvbuf;
6774: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6775: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6776: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6777: const PetscCount *Cperm1;
6778: PetscContainer container;
6779: MatCOOStruct_MPIAIJ *coo;
6781: PetscFunctionBegin;
6782: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6783: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6784: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6785: sendbuf = coo->sendbuf;
6786: recvbuf = coo->recvbuf;
6787: Ajmap1 = coo->Ajmap1;
6788: Ajmap2 = coo->Ajmap2;
6789: Aimap2 = coo->Aimap2;
6790: Bjmap1 = coo->Bjmap1;
6791: Bjmap2 = coo->Bjmap2;
6792: Bimap2 = coo->Bimap2;
6793: Aperm1 = coo->Aperm1;
6794: Aperm2 = coo->Aperm2;
6795: Bperm1 = coo->Bperm1;
6796: Bperm2 = coo->Bperm2;
6797: Cperm1 = coo->Cperm1;
6799: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6800: PetscCall(MatSeqAIJGetArray(B, &Ba));
6802: /* Pack entries to be sent to remote */
6803: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6805: /* Send remote entries to their owner and overlap the communication with local computation */
6806: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6807: /* Add local entries to A and B */
6808: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6809: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6810: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6811: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6812: }
6813: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6814: PetscScalar sum = 0.0;
6815: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6816: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6817: }
6818: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6820: /* Add received remote entries to A and B */
6821: for (PetscCount i = 0; i < coo->Annz2; i++) {
6822: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6823: }
6824: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6825: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6826: }
6827: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6828: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6829: PetscFunctionReturn(PETSC_SUCCESS);
6830: }
6832: /*MC
6833: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6835: Options Database Keys:
6836: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6838: Level: beginner
6840: Notes:
6841: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6842: in this case the values associated with the rows and columns one passes in are set to zero
6843: in the matrix
6845: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6846: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6848: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6849: M*/
6850: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6851: {
6852: Mat_MPIAIJ *b;
6853: PetscMPIInt size;
6855: PetscFunctionBegin;
6856: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6858: PetscCall(PetscNew(&b));
6859: B->data = (void *)b;
6860: B->ops[0] = MatOps_Values;
6861: B->assembled = PETSC_FALSE;
6862: B->insertmode = NOT_SET_VALUES;
6863: b->size = size;
6865: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6867: /* build cache for off array entries formed */
6868: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6870: b->donotstash = PETSC_FALSE;
6871: b->colmap = NULL;
6872: b->garray = NULL;
6873: b->roworiented = PETSC_TRUE;
6875: /* stuff used for matrix vector multiply */
6876: b->lvec = NULL;
6877: b->Mvctx = NULL;
6879: /* stuff for MatGetRow() */
6880: b->rowindices = NULL;
6881: b->rowvalues = NULL;
6882: b->getrowactive = PETSC_FALSE;
6884: /* flexible pointer used in CUSPARSE classes */
6885: b->spptr = NULL;
6887: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6888: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6889: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6890: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6891: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6892: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6893: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6897: #if defined(PETSC_HAVE_CUDA)
6898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6899: #endif
6900: #if defined(PETSC_HAVE_HIP)
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6902: #endif
6903: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6905: #endif
6906: #if defined(PETSC_HAVE_MKL_SPARSE)
6907: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6908: #endif
6909: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6910: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6912: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6913: #if defined(PETSC_HAVE_ELEMENTAL)
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6915: #endif
6916: #if defined(PETSC_HAVE_SCALAPACK)
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6918: #endif
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6921: #if defined(PETSC_HAVE_HYPRE)
6922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6924: #endif
6925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6927: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6928: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6929: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6930: PetscFunctionReturn(PETSC_SUCCESS);
6931: }
6933: /*@C
6934: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6935: and "off-diagonal" part of the matrix in CSR format.
6937: Collective
6939: Input Parameters:
6940: + comm - MPI communicator
6941: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6942: . n - This value should be the same as the local size used in creating the
6943: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6944: calculated if `N` is given) For square matrices `n` is almost always `m`.
6945: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6946: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6947: . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
6948: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6949: . a - matrix values
6950: . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix
6951: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6952: - oa - matrix values
6954: Output Parameter:
6955: . mat - the matrix
6957: Level: advanced
6959: Notes:
6960: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
6961: must free the arrays once the matrix has been destroyed and not before.
6963: The `i` and `j` indices are 0 based
6965: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6967: This sets local rows and cannot be used to set off-processor values.
6969: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6970: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6971: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6972: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6973: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6974: communication if it is known that only local entries will be set.
6976: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6977: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6978: @*/
6979: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, PetscInt i[], PetscInt j[], PetscScalar a[], PetscInt oi[], PetscInt oj[], PetscScalar oa[], Mat *mat)
6980: {
6981: Mat_MPIAIJ *maij;
6983: PetscFunctionBegin;
6984: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6985: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6986: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6987: PetscCall(MatCreate(comm, mat));
6988: PetscCall(MatSetSizes(*mat, m, n, M, N));
6989: PetscCall(MatSetType(*mat, MATMPIAIJ));
6990: maij = (Mat_MPIAIJ *)(*mat)->data;
6992: (*mat)->preallocated = PETSC_TRUE;
6994: PetscCall(PetscLayoutSetUp((*mat)->rmap));
6995: PetscCall(PetscLayoutSetUp((*mat)->cmap));
6997: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
6998: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7000: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7001: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7002: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7003: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7004: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7005: PetscFunctionReturn(PETSC_SUCCESS);
7006: }
7008: typedef struct {
7009: Mat *mp; /* intermediate products */
7010: PetscBool *mptmp; /* is the intermediate product temporary ? */
7011: PetscInt cp; /* number of intermediate products */
7013: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7014: PetscInt *startsj_s, *startsj_r;
7015: PetscScalar *bufa;
7016: Mat P_oth;
7018: /* may take advantage of merging product->B */
7019: Mat Bloc; /* B-local by merging diag and off-diag */
7021: /* cusparse does not have support to split between symbolic and numeric phases.
7022: When api_user is true, we don't need to update the numerical values
7023: of the temporary storage */
7024: PetscBool reusesym;
7026: /* support for COO values insertion */
7027: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7028: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7029: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7030: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7031: PetscSF sf; /* used for non-local values insertion and memory malloc */
7032: PetscMemType mtype;
7034: /* customization */
7035: PetscBool abmerge;
7036: PetscBool P_oth_bind;
7037: } MatMatMPIAIJBACKEND;
7039: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7040: {
7041: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7042: PetscInt i;
7044: PetscFunctionBegin;
7045: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7046: PetscCall(PetscFree(mmdata->bufa));
7047: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7048: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7049: PetscCall(MatDestroy(&mmdata->P_oth));
7050: PetscCall(MatDestroy(&mmdata->Bloc));
7051: PetscCall(PetscSFDestroy(&mmdata->sf));
7052: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7053: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7054: PetscCall(PetscFree(mmdata->own[0]));
7055: PetscCall(PetscFree(mmdata->own));
7056: PetscCall(PetscFree(mmdata->off[0]));
7057: PetscCall(PetscFree(mmdata->off));
7058: PetscCall(PetscFree(mmdata));
7059: PetscFunctionReturn(PETSC_SUCCESS);
7060: }
7062: /* Copy selected n entries with indices in idx[] of A to v[].
7063: If idx is NULL, copy the whole data array of A to v[]
7064: */
7065: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7066: {
7067: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7069: PetscFunctionBegin;
7070: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7071: if (f) {
7072: PetscCall((*f)(A, n, idx, v));
7073: } else {
7074: const PetscScalar *vv;
7076: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7077: if (n && idx) {
7078: PetscScalar *w = v;
7079: const PetscInt *oi = idx;
7080: PetscInt j;
7082: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7083: } else {
7084: PetscCall(PetscArraycpy(v, vv, n));
7085: }
7086: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7087: }
7088: PetscFunctionReturn(PETSC_SUCCESS);
7089: }
7091: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7092: {
7093: MatMatMPIAIJBACKEND *mmdata;
7094: PetscInt i, n_d, n_o;
7096: PetscFunctionBegin;
7097: MatCheckProduct(C, 1);
7098: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7099: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7100: if (!mmdata->reusesym) { /* update temporary matrices */
7101: if (mmdata->P_oth) PetscCall(MatGetBrowsOfAoCols_MPIAIJ(C->product->A, C->product->B, MAT_REUSE_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7102: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7103: }
7104: mmdata->reusesym = PETSC_FALSE;
7106: for (i = 0; i < mmdata->cp; i++) {
7107: PetscCheck(mmdata->mp[i]->ops->productnumeric, PetscObjectComm((PetscObject)mmdata->mp[i]), PETSC_ERR_PLIB, "Missing numeric op for %s", MatProductTypes[mmdata->mp[i]->product->type]);
7108: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7109: }
7110: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7111: PetscInt noff = mmdata->off[i + 1] - mmdata->off[i];
7113: if (mmdata->mptmp[i]) continue;
7114: if (noff) {
7115: PetscInt nown = mmdata->own[i + 1] - mmdata->own[i];
7117: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7118: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7119: n_o += noff;
7120: n_d += nown;
7121: } else {
7122: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7124: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7125: n_d += mm->nz;
7126: }
7127: }
7128: if (mmdata->hasoffproc) { /* offprocess insertion */
7129: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7130: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7131: }
7132: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7133: PetscFunctionReturn(PETSC_SUCCESS);
7134: }
7136: /* Support for Pt * A, A * P, or Pt * A * P */
7137: #define MAX_NUMBER_INTERMEDIATE 4
7138: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7139: {
7140: Mat_Product *product = C->product;
7141: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7142: Mat_MPIAIJ *a, *p;
7143: MatMatMPIAIJBACKEND *mmdata;
7144: ISLocalToGlobalMapping P_oth_l2g = NULL;
7145: IS glob = NULL;
7146: const char *prefix;
7147: char pprefix[256];
7148: const PetscInt *globidx, *P_oth_idx;
7149: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7150: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7151: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7152: /* type-0: consecutive, start from 0; type-1: consecutive with */
7153: /* a base offset; type-2: sparse with a local to global map table */
7154: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7156: MatProductType ptype;
7157: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7158: PetscMPIInt size;
7160: PetscFunctionBegin;
7161: MatCheckProduct(C, 1);
7162: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7163: ptype = product->type;
7164: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7165: ptype = MATPRODUCT_AB;
7166: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7167: }
7168: switch (ptype) {
7169: case MATPRODUCT_AB:
7170: A = product->A;
7171: P = product->B;
7172: m = A->rmap->n;
7173: n = P->cmap->n;
7174: M = A->rmap->N;
7175: N = P->cmap->N;
7176: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7177: break;
7178: case MATPRODUCT_AtB:
7179: P = product->A;
7180: A = product->B;
7181: m = P->cmap->n;
7182: n = A->cmap->n;
7183: M = P->cmap->N;
7184: N = A->cmap->N;
7185: hasoffproc = PETSC_TRUE;
7186: break;
7187: case MATPRODUCT_PtAP:
7188: A = product->A;
7189: P = product->B;
7190: m = P->cmap->n;
7191: n = P->cmap->n;
7192: M = P->cmap->N;
7193: N = P->cmap->N;
7194: hasoffproc = PETSC_TRUE;
7195: break;
7196: default:
7197: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7198: }
7199: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7200: if (size == 1) hasoffproc = PETSC_FALSE;
7202: /* defaults */
7203: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7204: mp[i] = NULL;
7205: mptmp[i] = PETSC_FALSE;
7206: rmapt[i] = -1;
7207: cmapt[i] = -1;
7208: rmapa[i] = NULL;
7209: cmapa[i] = NULL;
7210: }
7212: /* customization */
7213: PetscCall(PetscNew(&mmdata));
7214: mmdata->reusesym = product->api_user;
7215: if (ptype == MATPRODUCT_AB) {
7216: if (product->api_user) {
7217: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7218: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7219: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7220: PetscOptionsEnd();
7221: } else {
7222: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7223: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7224: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7225: PetscOptionsEnd();
7226: }
7227: } else if (ptype == MATPRODUCT_PtAP) {
7228: if (product->api_user) {
7229: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7230: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7231: PetscOptionsEnd();
7232: } else {
7233: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7234: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7235: PetscOptionsEnd();
7236: }
7237: }
7238: a = (Mat_MPIAIJ *)A->data;
7239: p = (Mat_MPIAIJ *)P->data;
7240: PetscCall(MatSetSizes(C, m, n, M, N));
7241: PetscCall(PetscLayoutSetUp(C->rmap));
7242: PetscCall(PetscLayoutSetUp(C->cmap));
7243: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7244: PetscCall(MatGetOptionsPrefix(C, &prefix));
7246: cp = 0;
7247: switch (ptype) {
7248: case MATPRODUCT_AB: /* A * P */
7249: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7251: /* A_diag * P_local (merged or not) */
7252: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7253: /* P is product->B */
7254: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7255: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7256: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7257: PetscCall(MatProductSetFill(mp[cp], product->fill));
7258: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7259: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7260: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7261: mp[cp]->product->api_user = product->api_user;
7262: PetscCall(MatProductSetFromOptions(mp[cp]));
7263: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7264: PetscCall(ISGetIndices(glob, &globidx));
7265: rmapt[cp] = 1;
7266: cmapt[cp] = 2;
7267: cmapa[cp] = globidx;
7268: mptmp[cp] = PETSC_FALSE;
7269: cp++;
7270: } else { /* A_diag * P_diag and A_diag * P_off */
7271: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7272: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7273: PetscCall(MatProductSetFill(mp[cp], product->fill));
7274: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7275: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7276: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7277: mp[cp]->product->api_user = product->api_user;
7278: PetscCall(MatProductSetFromOptions(mp[cp]));
7279: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7280: rmapt[cp] = 1;
7281: cmapt[cp] = 1;
7282: mptmp[cp] = PETSC_FALSE;
7283: cp++;
7284: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7285: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7286: PetscCall(MatProductSetFill(mp[cp], product->fill));
7287: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7288: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7289: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7290: mp[cp]->product->api_user = product->api_user;
7291: PetscCall(MatProductSetFromOptions(mp[cp]));
7292: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7293: rmapt[cp] = 1;
7294: cmapt[cp] = 2;
7295: cmapa[cp] = p->garray;
7296: mptmp[cp] = PETSC_FALSE;
7297: cp++;
7298: }
7300: /* A_off * P_other */
7301: if (mmdata->P_oth) {
7302: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7303: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7304: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7305: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7306: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7307: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7308: PetscCall(MatProductSetFill(mp[cp], product->fill));
7309: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7310: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7311: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7312: mp[cp]->product->api_user = product->api_user;
7313: PetscCall(MatProductSetFromOptions(mp[cp]));
7314: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7315: rmapt[cp] = 1;
7316: cmapt[cp] = 2;
7317: cmapa[cp] = P_oth_idx;
7318: mptmp[cp] = PETSC_FALSE;
7319: cp++;
7320: }
7321: break;
7323: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7324: /* A is product->B */
7325: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7326: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7327: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7328: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7329: PetscCall(MatProductSetFill(mp[cp], product->fill));
7330: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7331: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7332: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7333: mp[cp]->product->api_user = product->api_user;
7334: PetscCall(MatProductSetFromOptions(mp[cp]));
7335: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7336: PetscCall(ISGetIndices(glob, &globidx));
7337: rmapt[cp] = 2;
7338: rmapa[cp] = globidx;
7339: cmapt[cp] = 2;
7340: cmapa[cp] = globidx;
7341: mptmp[cp] = PETSC_FALSE;
7342: cp++;
7343: } else {
7344: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7345: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7346: PetscCall(MatProductSetFill(mp[cp], product->fill));
7347: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7348: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7349: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7350: mp[cp]->product->api_user = product->api_user;
7351: PetscCall(MatProductSetFromOptions(mp[cp]));
7352: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7353: PetscCall(ISGetIndices(glob, &globidx));
7354: rmapt[cp] = 1;
7355: cmapt[cp] = 2;
7356: cmapa[cp] = globidx;
7357: mptmp[cp] = PETSC_FALSE;
7358: cp++;
7359: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7360: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7361: PetscCall(MatProductSetFill(mp[cp], product->fill));
7362: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7363: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7364: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7365: mp[cp]->product->api_user = product->api_user;
7366: PetscCall(MatProductSetFromOptions(mp[cp]));
7367: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7368: rmapt[cp] = 2;
7369: rmapa[cp] = p->garray;
7370: cmapt[cp] = 2;
7371: cmapa[cp] = globidx;
7372: mptmp[cp] = PETSC_FALSE;
7373: cp++;
7374: }
7375: break;
7376: case MATPRODUCT_PtAP:
7377: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7378: /* P is product->B */
7379: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7380: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7381: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7382: PetscCall(MatProductSetFill(mp[cp], product->fill));
7383: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7384: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7385: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7386: mp[cp]->product->api_user = product->api_user;
7387: PetscCall(MatProductSetFromOptions(mp[cp]));
7388: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7389: PetscCall(ISGetIndices(glob, &globidx));
7390: rmapt[cp] = 2;
7391: rmapa[cp] = globidx;
7392: cmapt[cp] = 2;
7393: cmapa[cp] = globidx;
7394: mptmp[cp] = PETSC_FALSE;
7395: cp++;
7396: if (mmdata->P_oth) {
7397: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7398: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7399: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7400: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7401: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7402: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7403: PetscCall(MatProductSetFill(mp[cp], product->fill));
7404: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7405: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7406: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7407: mp[cp]->product->api_user = product->api_user;
7408: PetscCall(MatProductSetFromOptions(mp[cp]));
7409: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7410: mptmp[cp] = PETSC_TRUE;
7411: cp++;
7412: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7413: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7414: PetscCall(MatProductSetFill(mp[cp], product->fill));
7415: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7416: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7417: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7418: mp[cp]->product->api_user = product->api_user;
7419: PetscCall(MatProductSetFromOptions(mp[cp]));
7420: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7421: rmapt[cp] = 2;
7422: rmapa[cp] = globidx;
7423: cmapt[cp] = 2;
7424: cmapa[cp] = P_oth_idx;
7425: mptmp[cp] = PETSC_FALSE;
7426: cp++;
7427: }
7428: break;
7429: default:
7430: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7431: }
7432: /* sanity check */
7433: if (size > 1)
7434: for (i = 0; i < cp; i++) PetscCheck(rmapt[i] != 2 || hasoffproc, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected offproc map type for product %" PetscInt_FMT, i);
7436: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7437: for (i = 0; i < cp; i++) {
7438: mmdata->mp[i] = mp[i];
7439: mmdata->mptmp[i] = mptmp[i];
7440: }
7441: mmdata->cp = cp;
7442: C->product->data = mmdata;
7443: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7444: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7446: /* memory type */
7447: mmdata->mtype = PETSC_MEMTYPE_HOST;
7448: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7449: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7450: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7451: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7452: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7453: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7455: /* prepare coo coordinates for values insertion */
7457: /* count total nonzeros of those intermediate seqaij Mats
7458: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7459: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7460: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7461: */
7462: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7463: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7464: if (mptmp[cp]) continue;
7465: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7466: const PetscInt *rmap = rmapa[cp];
7467: const PetscInt mr = mp[cp]->rmap->n;
7468: const PetscInt rs = C->rmap->rstart;
7469: const PetscInt re = C->rmap->rend;
7470: const PetscInt *ii = mm->i;
7471: for (i = 0; i < mr; i++) {
7472: const PetscInt gr = rmap[i];
7473: const PetscInt nz = ii[i + 1] - ii[i];
7474: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7475: else ncoo_oown += nz; /* this row is local */
7476: }
7477: } else ncoo_d += mm->nz;
7478: }
7480: /*
7481: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7483: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7485: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7487: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7488: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7489: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7491: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7492: Ex. coo_i[]: the beginning part (of size ncoo_d + ncoo_oown) stores i of local nonzeros, and the remaining part stores i of nonzeros I will receive.
7493: */
7494: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7495: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7497: /* gather (i,j) of nonzeros inserted by remote procs */
7498: if (hasoffproc) {
7499: PetscSF msf;
7500: PetscInt ncoo2, *coo_i2, *coo_j2;
7502: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7503: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7504: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7506: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7507: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7508: PetscInt *idxoff = mmdata->off[cp];
7509: PetscInt *idxown = mmdata->own[cp];
7510: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7511: const PetscInt *rmap = rmapa[cp];
7512: const PetscInt *cmap = cmapa[cp];
7513: const PetscInt *ii = mm->i;
7514: PetscInt *coi = coo_i + ncoo_o;
7515: PetscInt *coj = coo_j + ncoo_o;
7516: const PetscInt mr = mp[cp]->rmap->n;
7517: const PetscInt rs = C->rmap->rstart;
7518: const PetscInt re = C->rmap->rend;
7519: const PetscInt cs = C->cmap->rstart;
7520: for (i = 0; i < mr; i++) {
7521: const PetscInt *jj = mm->j + ii[i];
7522: const PetscInt gr = rmap[i];
7523: const PetscInt nz = ii[i + 1] - ii[i];
7524: if (gr < rs || gr >= re) { /* this is an offproc row */
7525: for (j = ii[i]; j < ii[i + 1]; j++) {
7526: *coi++ = gr;
7527: *idxoff++ = j;
7528: }
7529: if (!cmapt[cp]) { /* already global */
7530: for (j = 0; j < nz; j++) *coj++ = jj[j];
7531: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7532: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7533: } else { /* offdiag */
7534: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7535: }
7536: ncoo_o += nz;
7537: } else { /* this is a local row */
7538: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7539: }
7540: }
7541: }
7542: mmdata->off[cp + 1] = idxoff;
7543: mmdata->own[cp + 1] = idxown;
7544: }
7546: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7547: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, ncoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7548: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7549: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7550: ncoo = ncoo_d + ncoo_oown + ncoo2;
7551: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7552: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7553: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7554: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7555: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7556: PetscCall(PetscFree2(coo_i, coo_j));
7557: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7558: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7559: coo_i = coo_i2;
7560: coo_j = coo_j2;
7561: } else { /* no offproc values insertion */
7562: ncoo = ncoo_d;
7563: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7565: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7566: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7567: PetscCall(PetscSFSetUp(mmdata->sf));
7568: }
7569: mmdata->hasoffproc = hasoffproc;
7571: /* gather (i,j) of nonzeros inserted locally */
7572: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7573: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7574: PetscInt *coi = coo_i + ncoo_d;
7575: PetscInt *coj = coo_j + ncoo_d;
7576: const PetscInt *jj = mm->j;
7577: const PetscInt *ii = mm->i;
7578: const PetscInt *cmap = cmapa[cp];
7579: const PetscInt *rmap = rmapa[cp];
7580: const PetscInt mr = mp[cp]->rmap->n;
7581: const PetscInt rs = C->rmap->rstart;
7582: const PetscInt re = C->rmap->rend;
7583: const PetscInt cs = C->cmap->rstart;
7585: if (mptmp[cp]) continue;
7586: if (rmapt[cp] == 1) { /* consecutive rows */
7587: /* fill coo_i */
7588: for (i = 0; i < mr; i++) {
7589: const PetscInt gr = i + rs;
7590: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7591: }
7592: /* fill coo_j */
7593: if (!cmapt[cp]) { /* type-0, already global */
7594: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7595: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7596: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7597: } else { /* type-2, local to global for sparse columns */
7598: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7599: }
7600: ncoo_d += mm->nz;
7601: } else if (rmapt[cp] == 2) { /* sparse rows */
7602: for (i = 0; i < mr; i++) {
7603: const PetscInt *jj = mm->j + ii[i];
7604: const PetscInt gr = rmap[i];
7605: const PetscInt nz = ii[i + 1] - ii[i];
7606: if (gr >= rs && gr < re) { /* local rows */
7607: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7608: if (!cmapt[cp]) { /* type-0, already global */
7609: for (j = 0; j < nz; j++) *coj++ = jj[j];
7610: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7611: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7612: } else { /* type-2, local to global for sparse columns */
7613: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7614: }
7615: ncoo_d += nz;
7616: }
7617: }
7618: }
7619: }
7620: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7621: PetscCall(ISDestroy(&glob));
7622: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7623: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7624: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7625: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7627: /* preallocate with COO data */
7628: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7629: PetscCall(PetscFree2(coo_i, coo_j));
7630: PetscFunctionReturn(PETSC_SUCCESS);
7631: }
7633: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7634: {
7635: Mat_Product *product = mat->product;
7636: #if defined(PETSC_HAVE_DEVICE)
7637: PetscBool match = PETSC_FALSE;
7638: PetscBool usecpu = PETSC_FALSE;
7639: #else
7640: PetscBool match = PETSC_TRUE;
7641: #endif
7643: PetscFunctionBegin;
7644: MatCheckProduct(mat, 1);
7645: #if defined(PETSC_HAVE_DEVICE)
7646: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7647: if (match) { /* we can always fallback to the CPU if requested */
7648: switch (product->type) {
7649: case MATPRODUCT_AB:
7650: if (product->api_user) {
7651: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7652: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7653: PetscOptionsEnd();
7654: } else {
7655: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7656: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7657: PetscOptionsEnd();
7658: }
7659: break;
7660: case MATPRODUCT_AtB:
7661: if (product->api_user) {
7662: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7663: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7664: PetscOptionsEnd();
7665: } else {
7666: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7667: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7668: PetscOptionsEnd();
7669: }
7670: break;
7671: case MATPRODUCT_PtAP:
7672: if (product->api_user) {
7673: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7674: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7675: PetscOptionsEnd();
7676: } else {
7677: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7678: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7679: PetscOptionsEnd();
7680: }
7681: break;
7682: default:
7683: break;
7684: }
7685: match = (PetscBool)!usecpu;
7686: }
7687: #endif
7688: if (match) {
7689: switch (product->type) {
7690: case MATPRODUCT_AB:
7691: case MATPRODUCT_AtB:
7692: case MATPRODUCT_PtAP:
7693: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7694: break;
7695: default:
7696: break;
7697: }
7698: }
7699: /* fallback to MPIAIJ ops */
7700: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7701: PetscFunctionReturn(PETSC_SUCCESS);
7702: }
7704: /*
7705: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7707: n - the number of block indices in cc[]
7708: cc - the block indices (must be large enough to contain the indices)
7709: */
7710: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7711: {
7712: PetscInt cnt = -1, nidx, j;
7713: const PetscInt *idx;
7715: PetscFunctionBegin;
7716: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7717: if (nidx) {
7718: cnt = 0;
7719: cc[cnt] = idx[0] / bs;
7720: for (j = 1; j < nidx; j++) {
7721: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7722: }
7723: }
7724: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7725: *n = cnt + 1;
7726: PetscFunctionReturn(PETSC_SUCCESS);
7727: }
7729: /*
7730: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7732: ncollapsed - the number of block indices
7733: collapsed - the block indices (must be large enough to contain the indices)
7734: */
7735: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7736: {
7737: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7739: PetscFunctionBegin;
7740: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7741: for (i = start + 1; i < start + bs; i++) {
7742: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7743: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7744: cprevtmp = cprev;
7745: cprev = merged;
7746: merged = cprevtmp;
7747: }
7748: *ncollapsed = nprev;
7749: if (collapsed) *collapsed = cprev;
7750: PetscFunctionReturn(PETSC_SUCCESS);
7751: }
7753: /*
7754: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7756: Input Parameter:
7757: . Amat - matrix
7758: - symmetrize - make the result symmetric
7759: + scale - scale with diagonal
7761: Output Parameter:
7762: . a_Gmat - output scalar graph >= 0
7764: */
7765: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7766: {
7767: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7768: MPI_Comm comm;
7769: Mat Gmat;
7770: PetscBool ismpiaij, isseqaij;
7771: Mat a, b, c;
7772: MatType jtype;
7774: PetscFunctionBegin;
7775: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7776: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7777: PetscCall(MatGetSize(Amat, &MM, &NN));
7778: PetscCall(MatGetBlockSize(Amat, &bs));
7779: nloc = (Iend - Istart) / bs;
7781: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7782: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7783: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7785: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7786: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7787: implementation */
7788: if (bs > 1) {
7789: PetscCall(MatGetType(Amat, &jtype));
7790: PetscCall(MatCreate(comm, &Gmat));
7791: PetscCall(MatSetType(Gmat, jtype));
7792: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7793: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7794: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7795: PetscInt *d_nnz, *o_nnz;
7796: MatScalar *aa, val, *AA;
7797: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7798: if (isseqaij) {
7799: a = Amat;
7800: b = NULL;
7801: } else {
7802: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7803: a = d->A;
7804: b = d->B;
7805: }
7806: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7807: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7808: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7809: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7810: const PetscInt *cols1, *cols2;
7811: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7812: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7813: nnz[brow / bs] = nc2 / bs;
7814: if (nc2 % bs) ok = 0;
7815: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7816: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7817: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7818: if (nc1 != nc2) ok = 0;
7819: else {
7820: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7821: if (cols1[jj] != cols2[jj]) ok = 0;
7822: if (cols1[jj] % bs != jj % bs) ok = 0;
7823: }
7824: }
7825: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7826: }
7827: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7828: if (!ok) {
7829: PetscCall(PetscFree2(d_nnz, o_nnz));
7830: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7831: goto old_bs;
7832: }
7833: }
7834: }
7835: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7836: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7837: PetscCall(PetscFree2(d_nnz, o_nnz));
7838: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7839: // diag
7840: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7841: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7842: ai = aseq->i;
7843: n = ai[brow + 1] - ai[brow];
7844: aj = aseq->j + ai[brow];
7845: for (int k = 0; k < n; k += bs) { // block columns
7846: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7847: val = 0;
7848: if (index_size == 0) {
7849: for (int ii = 0; ii < bs; ii++) { // rows in block
7850: aa = aseq->a + ai[brow + ii] + k;
7851: for (int jj = 0; jj < bs; jj++) { // columns in block
7852: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7853: }
7854: }
7855: } else { // use (index,index) value if provided
7856: for (int iii = 0; iii < index_size; iii++) { // rows in block
7857: int ii = index[iii];
7858: aa = aseq->a + ai[brow + ii] + k;
7859: for (int jjj = 0; jjj < index_size; jjj++) { // columns in block
7860: int jj = index[jjj];
7861: val += PetscAbs(PetscRealPart(aa[jj]));
7862: }
7863: }
7864: }
7865: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7866: AA[k / bs] = val;
7867: }
7868: grow = Istart / bs + brow / bs;
7869: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, INSERT_VALUES));
7870: }
7871: // off-diag
7872: if (ismpiaij) {
7873: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7874: const PetscScalar *vals;
7875: const PetscInt *cols, *garray = aij->garray;
7876: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7877: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7878: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7879: for (int k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7880: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7881: AA[k / bs] = 0;
7882: AJ[cidx] = garray[cols[k]] / bs;
7883: }
7884: nc = ncols / bs;
7885: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7886: if (index_size == 0) {
7887: for (int ii = 0; ii < bs; ii++) { // rows in block
7888: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7889: for (int k = 0; k < ncols; k += bs) {
7890: for (int jj = 0; jj < bs; jj++) { // cols in block
7891: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7892: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7893: }
7894: }
7895: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7896: }
7897: } else { // use (index,index) value if provided
7898: for (int iii = 0; iii < index_size; iii++) { // rows in block
7899: int ii = index[iii];
7900: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7901: for (int k = 0; k < ncols; k += bs) {
7902: for (int jjj = 0; jjj < index_size; jjj++) { // cols in block
7903: int jj = index[jjj];
7904: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7905: }
7906: }
7907: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7908: }
7909: }
7910: grow = Istart / bs + brow / bs;
7911: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, INSERT_VALUES));
7912: }
7913: }
7914: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7915: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7916: PetscCall(PetscFree2(AA, AJ));
7917: } else {
7918: const PetscScalar *vals;
7919: const PetscInt *idx;
7920: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7921: old_bs:
7922: /*
7923: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7924: */
7925: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7926: PetscCall(PetscMalloc2(nloc, &d_nnz, isseqaij ? 0 : nloc, &o_nnz));
7927: if (isseqaij) {
7928: PetscInt max_d_nnz;
7929: /*
7930: Determine exact preallocation count for (sequential) scalar matrix
7931: */
7932: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7933: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7934: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7935: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7936: PetscCall(PetscFree3(w0, w1, w2));
7937: } else if (ismpiaij) {
7938: Mat Daij, Oaij;
7939: const PetscInt *garray;
7940: PetscInt max_d_nnz;
7941: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7942: /*
7943: Determine exact preallocation count for diagonal block portion of scalar matrix
7944: */
7945: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7946: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7947: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7948: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7949: PetscCall(PetscFree3(w0, w1, w2));
7950: /*
7951: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7952: */
7953: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7954: o_nnz[jj] = 0;
7955: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7956: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7957: o_nnz[jj] += ncols;
7958: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7959: }
7960: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7961: }
7962: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7963: /* get scalar copy (norms) of matrix */
7964: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7965: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7966: PetscCall(PetscFree2(d_nnz, o_nnz));
7967: for (Ii = Istart; Ii < Iend; Ii++) {
7968: PetscInt dest_row = Ii / bs;
7969: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7970: for (jj = 0; jj < ncols; jj++) {
7971: PetscInt dest_col = idx[jj] / bs;
7972: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7973: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7974: }
7975: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7976: }
7977: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7978: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7979: }
7980: } else {
7981: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
7982: else {
7983: Gmat = Amat;
7984: PetscCall(PetscObjectReference((PetscObject)Gmat));
7985: }
7986: if (isseqaij) {
7987: a = Gmat;
7988: b = NULL;
7989: } else {
7990: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
7991: a = d->A;
7992: b = d->B;
7993: }
7994: if (filter >= 0 || scale) {
7995: /* take absolute value of each entry */
7996: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7997: MatInfo info;
7998: PetscScalar *avals;
7999: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8000: PetscCall(MatSeqAIJGetArray(c, &avals));
8001: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8002: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8003: }
8004: }
8005: }
8006: if (symmetrize) {
8007: PetscBool isset, issym;
8008: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8009: if (!isset || !issym) {
8010: Mat matTrans;
8011: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8012: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8013: PetscCall(MatDestroy(&matTrans));
8014: }
8015: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8016: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8017: if (scale) {
8018: /* scale c for all diagonal values = 1 or -1 */
8019: Vec diag;
8020: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8021: PetscCall(MatGetDiagonal(Gmat, diag));
8022: PetscCall(VecReciprocal(diag));
8023: PetscCall(VecSqrtAbs(diag));
8024: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8025: PetscCall(VecDestroy(&diag));
8026: }
8027: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8029: if (filter >= 0) {
8030: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8031: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8032: }
8033: *a_Gmat = Gmat;
8034: PetscFunctionReturn(PETSC_SUCCESS);
8035: }
8037: /*
8038: Special version for direct calls from Fortran
8039: */
8040: #include <petsc/private/fortranimpl.h>
8042: /* Change these macros so can be used in void function */
8043: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8044: #undef PetscCall
8045: #define PetscCall(...) \
8046: do { \
8047: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8048: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8049: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8050: return; \
8051: } \
8052: } while (0)
8054: #undef SETERRQ
8055: #define SETERRQ(comm, ierr, ...) \
8056: do { \
8057: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8058: return; \
8059: } while (0)
8061: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8062: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8063: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8064: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8065: #else
8066: #endif
8067: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8068: {
8069: Mat mat = *mmat;
8070: PetscInt m = *mm, n = *mn;
8071: InsertMode addv = *maddv;
8072: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8073: PetscScalar value;
8075: MatCheckPreallocated(mat, 1);
8076: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8077: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8078: {
8079: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8080: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8081: PetscBool roworiented = aij->roworiented;
8083: /* Some Variables required in the macro */
8084: Mat A = aij->A;
8085: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8086: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8087: MatScalar *aa;
8088: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8089: Mat B = aij->B;
8090: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8091: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8092: MatScalar *ba;
8093: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8094: * cannot use "#if defined" inside a macro. */
8095: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8097: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8098: PetscInt nonew = a->nonew;
8099: MatScalar *ap1, *ap2;
8101: PetscFunctionBegin;
8102: PetscCall(MatSeqAIJGetArray(A, &aa));
8103: PetscCall(MatSeqAIJGetArray(B, &ba));
8104: for (i = 0; i < m; i++) {
8105: if (im[i] < 0) continue;
8106: 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);
8107: if (im[i] >= rstart && im[i] < rend) {
8108: row = im[i] - rstart;
8109: lastcol1 = -1;
8110: rp1 = aj + ai[row];
8111: ap1 = aa + ai[row];
8112: rmax1 = aimax[row];
8113: nrow1 = ailen[row];
8114: low1 = 0;
8115: high1 = nrow1;
8116: lastcol2 = -1;
8117: rp2 = bj + bi[row];
8118: ap2 = ba + bi[row];
8119: rmax2 = bimax[row];
8120: nrow2 = bilen[row];
8121: low2 = 0;
8122: high2 = nrow2;
8124: for (j = 0; j < n; j++) {
8125: if (roworiented) value = v[i * n + j];
8126: else value = v[i + j * m];
8127: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8128: if (in[j] >= cstart && in[j] < cend) {
8129: col = in[j] - cstart;
8130: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8131: } else if (in[j] < 0) continue;
8132: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8133: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8134: } else {
8135: if (mat->was_assembled) {
8136: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8137: #if defined(PETSC_USE_CTABLE)
8138: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8139: col--;
8140: #else
8141: col = aij->colmap[in[j]] - 1;
8142: #endif
8143: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8144: PetscCall(MatDisAssemble_MPIAIJ(mat));
8145: col = in[j];
8146: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8147: B = aij->B;
8148: b = (Mat_SeqAIJ *)B->data;
8149: bimax = b->imax;
8150: bi = b->i;
8151: bilen = b->ilen;
8152: bj = b->j;
8153: rp2 = bj + bi[row];
8154: ap2 = ba + bi[row];
8155: rmax2 = bimax[row];
8156: nrow2 = bilen[row];
8157: low2 = 0;
8158: high2 = nrow2;
8159: bm = aij->B->rmap->n;
8160: ba = b->a;
8161: inserted = PETSC_FALSE;
8162: }
8163: } else col = in[j];
8164: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8165: }
8166: }
8167: } else if (!aij->donotstash) {
8168: if (roworiented) {
8169: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8170: } else {
8171: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8172: }
8173: }
8174: }
8175: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8176: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8177: }
8178: PetscFunctionReturnVoid();
8179: }
8181: /* Undefining these here since they were redefined from their original definition above! No
8182: * other PETSc functions should be defined past this point, as it is impossible to recover the
8183: * original definitions */
8184: #undef PetscCall
8185: #undef SETERRQ