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: PetscCallMPI(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;
287: PetscMPIInt in;
289: PetscFunctionBegin;
290: PetscCall(MatGetSize(A, &m, &n));
291: PetscCall(PetscCalloc1(n, &work));
292: PetscCall(MatSeqAIJGetArrayRead(aij->A, &dummy));
293: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &dummy));
294: PetscCall(MatSeqAIJGetArrayRead(aij->B, &dummy));
295: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &dummy));
296: if (type == NORM_2) {
297: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i] * a_aij->a[i]);
298: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i] * b_aij->a[i]);
299: } else if (type == NORM_1) {
300: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
301: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
302: } else if (type == NORM_INFINITY) {
303: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
304: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]), work[garray[b_aij->j[i]]]);
305: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
306: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscRealPart(a_aij->a[i]);
307: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscRealPart(b_aij->a[i]);
308: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
309: for (i = 0; i < a_aij->i[aij->A->rmap->n]; i++) work[A->cmap->rstart + a_aij->j[i]] += PetscImaginaryPart(a_aij->a[i]);
310: for (i = 0; i < b_aij->i[aij->B->rmap->n]; i++) work[garray[b_aij->j[i]]] += PetscImaginaryPart(b_aij->a[i]);
311: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
312: PetscCall(PetscMPIIntCast(n, &in));
313: if (type == NORM_INFINITY) {
314: PetscCallMPI(MPIU_Allreduce(work, reductions, in, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)A)));
315: } else {
316: PetscCallMPI(MPIU_Allreduce(work, reductions, in, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)A)));
317: }
318: PetscCall(PetscFree(work));
319: if (type == NORM_2) {
320: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
321: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
322: for (i = 0; i < n; i++) reductions[i] /= m;
323: }
324: PetscFunctionReturn(PETSC_SUCCESS);
325: }
327: static PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A, IS *is)
328: {
329: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
330: IS sis, gis;
331: const PetscInt *isis, *igis;
332: PetscInt n, *iis, nsis, ngis, rstart, i;
334: PetscFunctionBegin;
335: PetscCall(MatFindOffBlockDiagonalEntries(a->A, &sis));
336: PetscCall(MatFindNonzeroRows(a->B, &gis));
337: PetscCall(ISGetSize(gis, &ngis));
338: PetscCall(ISGetSize(sis, &nsis));
339: PetscCall(ISGetIndices(sis, &isis));
340: PetscCall(ISGetIndices(gis, &igis));
342: PetscCall(PetscMalloc1(ngis + nsis, &iis));
343: PetscCall(PetscArraycpy(iis, igis, ngis));
344: PetscCall(PetscArraycpy(iis + ngis, isis, nsis));
345: n = ngis + nsis;
346: PetscCall(PetscSortRemoveDupsInt(&n, iis));
347: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
348: for (i = 0; i < n; i++) iis[i] += rstart;
349: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), n, iis, PETSC_OWN_POINTER, is));
351: PetscCall(ISRestoreIndices(sis, &isis));
352: PetscCall(ISRestoreIndices(gis, &igis));
353: PetscCall(ISDestroy(&sis));
354: PetscCall(ISDestroy(&gis));
355: PetscFunctionReturn(PETSC_SUCCESS);
356: }
358: /*
359: Local utility routine that creates a mapping from the global column
360: number to the local number in the off-diagonal part of the local
361: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
362: a slightly higher hash table cost; without it it is not scalable (each processor
363: has an order N integer array but is fast to access.
364: */
365: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
366: {
367: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
368: PetscInt n = aij->B->cmap->n, i;
370: PetscFunctionBegin;
371: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
372: #if defined(PETSC_USE_CTABLE)
373: PetscCall(PetscHMapICreateWithSize(n, &aij->colmap));
374: for (i = 0; i < n; i++) PetscCall(PetscHMapISet(aij->colmap, aij->garray[i] + 1, i + 1));
375: #else
376: PetscCall(PetscCalloc1(mat->cmap->N + 1, &aij->colmap));
377: for (i = 0; i < n; i++) aij->colmap[aij->garray[i]] = i + 1;
378: #endif
379: PetscFunctionReturn(PETSC_SUCCESS);
380: }
382: #define MatSetValues_SeqAIJ_A_Private(row, col, value, addv, orow, ocol) \
383: do { \
384: if (col <= lastcol1) low1 = 0; \
385: else high1 = nrow1; \
386: lastcol1 = col; \
387: while (high1 - low1 > 5) { \
388: t = (low1 + high1) / 2; \
389: if (rp1[t] > col) high1 = t; \
390: else low1 = t; \
391: } \
392: for (_i = low1; _i < high1; _i++) { \
393: if (rp1[_i] > col) break; \
394: if (rp1[_i] == col) { \
395: if (addv == ADD_VALUES) { \
396: ap1[_i] += value; \
397: /* Not sure LogFlops will slow dow the code or not */ \
398: (void)PetscLogFlops(1.0); \
399: } else ap1[_i] = value; \
400: goto a_noinsert; \
401: } \
402: } \
403: if (value == 0.0 && ignorezeroentries && row != col) { \
404: low1 = 0; \
405: high1 = nrow1; \
406: goto a_noinsert; \
407: } \
408: if (nonew == 1) { \
409: low1 = 0; \
410: high1 = nrow1; \
411: goto a_noinsert; \
412: } \
413: 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); \
414: MatSeqXAIJReallocateAIJ(A, am, 1, nrow1, row, col, rmax1, aa, ai, aj, rp1, ap1, aimax, nonew, MatScalar); \
415: N = nrow1++ - 1; \
416: a->nz++; \
417: high1++; \
418: /* shift up all the later entries in this row */ \
419: PetscCall(PetscArraymove(rp1 + _i + 1, rp1 + _i, N - _i + 1)); \
420: PetscCall(PetscArraymove(ap1 + _i + 1, ap1 + _i, N - _i + 1)); \
421: rp1[_i] = col; \
422: ap1[_i] = value; \
423: a_noinsert:; \
424: ailen[row] = nrow1; \
425: } while (0)
427: #define MatSetValues_SeqAIJ_B_Private(row, col, value, addv, orow, ocol) \
428: do { \
429: if (col <= lastcol2) low2 = 0; \
430: else high2 = nrow2; \
431: lastcol2 = col; \
432: while (high2 - low2 > 5) { \
433: t = (low2 + high2) / 2; \
434: if (rp2[t] > col) high2 = t; \
435: else low2 = t; \
436: } \
437: for (_i = low2; _i < high2; _i++) { \
438: if (rp2[_i] > col) break; \
439: if (rp2[_i] == col) { \
440: if (addv == ADD_VALUES) { \
441: ap2[_i] += value; \
442: (void)PetscLogFlops(1.0); \
443: } else ap2[_i] = value; \
444: goto b_noinsert; \
445: } \
446: } \
447: if (value == 0.0 && ignorezeroentries) { \
448: low2 = 0; \
449: high2 = nrow2; \
450: goto b_noinsert; \
451: } \
452: if (nonew == 1) { \
453: low2 = 0; \
454: high2 = nrow2; \
455: goto b_noinsert; \
456: } \
457: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \
458: MatSeqXAIJReallocateAIJ(B, bm, 1, nrow2, row, col, rmax2, ba, bi, bj, rp2, ap2, bimax, nonew, MatScalar); \
459: N = nrow2++ - 1; \
460: b->nz++; \
461: high2++; \
462: /* shift up all the later entries in this row */ \
463: PetscCall(PetscArraymove(rp2 + _i + 1, rp2 + _i, N - _i + 1)); \
464: PetscCall(PetscArraymove(ap2 + _i + 1, ap2 + _i, N - _i + 1)); \
465: rp2[_i] = col; \
466: ap2[_i] = value; \
467: b_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 = PetscSafePointerPlusOffset(bj, bi[row]);
584: ap2 = PetscSafePointerPlusOffset(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: PetscCallMPI(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: PetscCallMPI(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: PetscCallMPI(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: PetscInt 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: PetscCallMPI(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: PetscCallMPI(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_INT_MAX) header[3] = PETSC_INT_MAX;
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_INT_MAX, 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: PetscCallMPI(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: PetscCallMPI(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;
1803: PetscMPIInt iN;
1805: PetscFunctionBegin;
1806: if (aij->size == 1) {
1807: PetscCall(MatNorm(aij->A, type, norm));
1808: } else {
1809: PetscCall(MatSeqAIJGetArrayRead(aij->A, &amata));
1810: PetscCall(MatSeqAIJGetArrayRead(aij->B, &bmata));
1811: if (type == NORM_FROBENIUS) {
1812: v = amata;
1813: for (i = 0; i < amat->nz; i++) {
1814: sum += PetscRealPart(PetscConj(*v) * (*v));
1815: v++;
1816: }
1817: v = bmata;
1818: for (i = 0; i < bmat->nz; i++) {
1819: sum += PetscRealPart(PetscConj(*v) * (*v));
1820: v++;
1821: }
1822: PetscCallMPI(MPIU_Allreduce(&sum, norm, 1, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1823: *norm = PetscSqrtReal(*norm);
1824: PetscCall(PetscLogFlops(2.0 * amat->nz + 2.0 * bmat->nz));
1825: } else if (type == NORM_1) { /* max column norm */
1826: PetscReal *tmp, *tmp2;
1827: PetscInt *jj, *garray = aij->garray;
1828: PetscCall(PetscCalloc1(mat->cmap->N + 1, &tmp));
1829: PetscCall(PetscMalloc1(mat->cmap->N + 1, &tmp2));
1830: *norm = 0.0;
1831: v = amata;
1832: jj = amat->j;
1833: for (j = 0; j < amat->nz; j++) {
1834: tmp[cstart + *jj++] += PetscAbsScalar(*v);
1835: v++;
1836: }
1837: v = bmata;
1838: jj = bmat->j;
1839: for (j = 0; j < bmat->nz; j++) {
1840: tmp[garray[*jj++]] += PetscAbsScalar(*v);
1841: v++;
1842: }
1843: PetscCall(PetscMPIIntCast(mat->cmap->N, &iN));
1844: PetscCallMPI(MPIU_Allreduce(tmp, tmp2, iN, MPIU_REAL, MPIU_SUM, PetscObjectComm((PetscObject)mat)));
1845: for (j = 0; j < mat->cmap->N; j++) {
1846: if (tmp2[j] > *norm) *norm = tmp2[j];
1847: }
1848: PetscCall(PetscFree(tmp));
1849: PetscCall(PetscFree(tmp2));
1850: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1851: } else if (type == NORM_INFINITY) { /* max row norm */
1852: PetscReal ntemp = 0.0;
1853: for (j = 0; j < aij->A->rmap->n; j++) {
1854: v = PetscSafePointerPlusOffset(amata, amat->i[j]);
1855: sum = 0.0;
1856: for (i = 0; i < amat->i[j + 1] - amat->i[j]; i++) {
1857: sum += PetscAbsScalar(*v);
1858: v++;
1859: }
1860: v = PetscSafePointerPlusOffset(bmata, bmat->i[j]);
1861: for (i = 0; i < bmat->i[j + 1] - bmat->i[j]; i++) {
1862: sum += PetscAbsScalar(*v);
1863: v++;
1864: }
1865: if (sum > ntemp) ntemp = sum;
1866: }
1867: PetscCallMPI(MPIU_Allreduce(&ntemp, norm, 1, MPIU_REAL, MPIU_MAX, PetscObjectComm((PetscObject)mat)));
1868: PetscCall(PetscLogFlops(PetscMax(amat->nz + bmat->nz - 1, 0)));
1869: } else SETERRQ(PetscObjectComm((PetscObject)mat), PETSC_ERR_SUP, "No support for two norm");
1870: PetscCall(MatSeqAIJRestoreArrayRead(aij->A, &amata));
1871: PetscCall(MatSeqAIJRestoreArrayRead(aij->B, &bmata));
1872: }
1873: PetscFunctionReturn(PETSC_SUCCESS);
1874: }
1876: static PetscErrorCode MatTranspose_MPIAIJ(Mat A, MatReuse reuse, Mat *matout)
1877: {
1878: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *b;
1879: Mat_SeqAIJ *Aloc = (Mat_SeqAIJ *)a->A->data, *Bloc = (Mat_SeqAIJ *)a->B->data, *sub_B_diag;
1880: 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;
1881: const PetscInt *ai, *aj, *bi, *bj, *B_diag_i;
1882: Mat B, A_diag, *B_diag;
1883: const MatScalar *pbv, *bv;
1885: PetscFunctionBegin;
1886: if (reuse == MAT_REUSE_MATRIX) PetscCall(MatTransposeCheckNonzeroState_Private(A, *matout));
1887: ma = A->rmap->n;
1888: na = A->cmap->n;
1889: mb = a->B->rmap->n;
1890: nb = a->B->cmap->n;
1891: ai = Aloc->i;
1892: aj = Aloc->j;
1893: bi = Bloc->i;
1894: bj = Bloc->j;
1895: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1896: PetscInt *d_nnz, *g_nnz, *o_nnz;
1897: PetscSFNode *oloc;
1898: PETSC_UNUSED PetscSF sf;
1900: PetscCall(PetscMalloc4(na, &d_nnz, na, &o_nnz, nb, &g_nnz, nb, &oloc));
1901: /* compute d_nnz for preallocation */
1902: PetscCall(PetscArrayzero(d_nnz, na));
1903: for (i = 0; i < ai[ma]; i++) d_nnz[aj[i]]++;
1904: /* compute local off-diagonal contributions */
1905: PetscCall(PetscArrayzero(g_nnz, nb));
1906: for (i = 0; i < bi[ma]; i++) g_nnz[bj[i]]++;
1907: /* map those to global */
1908: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)A), &sf));
1909: PetscCall(PetscSFSetGraphLayout(sf, A->cmap, nb, NULL, PETSC_USE_POINTER, a->garray));
1910: PetscCall(PetscSFSetFromOptions(sf));
1911: PetscCall(PetscArrayzero(o_nnz, na));
1912: PetscCall(PetscSFReduceBegin(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1913: PetscCall(PetscSFReduceEnd(sf, MPIU_INT, g_nnz, o_nnz, MPI_SUM));
1914: PetscCall(PetscSFDestroy(&sf));
1916: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1917: PetscCall(MatSetSizes(B, A->cmap->n, A->rmap->n, N, M));
1918: PetscCall(MatSetBlockSizes(B, PetscAbs(A->cmap->bs), PetscAbs(A->rmap->bs)));
1919: PetscCall(MatSetType(B, ((PetscObject)A)->type_name));
1920: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
1921: PetscCall(PetscFree4(d_nnz, o_nnz, g_nnz, oloc));
1922: } else {
1923: B = *matout;
1924: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
1925: }
1927: b = (Mat_MPIAIJ *)B->data;
1928: A_diag = a->A;
1929: B_diag = &b->A;
1930: sub_B_diag = (Mat_SeqAIJ *)(*B_diag)->data;
1931: A_diag_ncol = A_diag->cmap->N;
1932: B_diag_ilen = sub_B_diag->ilen;
1933: B_diag_i = sub_B_diag->i;
1935: /* Set ilen for diagonal of B */
1936: for (i = 0; i < A_diag_ncol; i++) B_diag_ilen[i] = B_diag_i[i + 1] - B_diag_i[i];
1938: /* Transpose the diagonal part of the matrix. In contrast to the off-diagonal part, this can be done
1939: very quickly (=without using MatSetValues), because all writes are local. */
1940: PetscCall(MatTransposeSetPrecursor(A_diag, *B_diag));
1941: PetscCall(MatTranspose(A_diag, MAT_REUSE_MATRIX, B_diag));
1943: /* copy over the B part */
1944: PetscCall(PetscMalloc1(bi[mb], &cols));
1945: PetscCall(MatSeqAIJGetArrayRead(a->B, &bv));
1946: pbv = bv;
1947: row = A->rmap->rstart;
1948: for (i = 0; i < bi[mb]; i++) cols[i] = a->garray[bj[i]];
1949: cols_tmp = cols;
1950: for (i = 0; i < mb; i++) {
1951: ncol = bi[i + 1] - bi[i];
1952: PetscCall(MatSetValues(B, ncol, cols_tmp, 1, &row, pbv, INSERT_VALUES));
1953: row++;
1954: if (pbv) pbv += ncol;
1955: if (cols_tmp) cols_tmp += ncol;
1956: }
1957: PetscCall(PetscFree(cols));
1958: PetscCall(MatSeqAIJRestoreArrayRead(a->B, &bv));
1960: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
1961: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
1962: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
1963: *matout = B;
1964: } else {
1965: PetscCall(MatHeaderMerge(A, &B));
1966: }
1967: PetscFunctionReturn(PETSC_SUCCESS);
1968: }
1970: static PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat, Vec ll, Vec rr)
1971: {
1972: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
1973: Mat a = aij->A, b = aij->B;
1974: PetscInt s1, s2, s3;
1976: PetscFunctionBegin;
1977: PetscCall(MatGetLocalSize(mat, &s2, &s3));
1978: if (rr) {
1979: PetscCall(VecGetLocalSize(rr, &s1));
1980: PetscCheck(s1 == s3, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "right vector non-conforming local size");
1981: /* Overlap communication with computation. */
1982: PetscCall(VecScatterBegin(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1983: }
1984: if (ll) {
1985: PetscCall(VecGetLocalSize(ll, &s1));
1986: PetscCheck(s1 == s2, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "left vector non-conforming local size");
1987: PetscUseTypeMethod(b, diagonalscale, ll, NULL);
1988: }
1989: /* scale the diagonal block */
1990: PetscUseTypeMethod(a, diagonalscale, ll, rr);
1992: if (rr) {
1993: /* Do a scatter end and then right scale the off-diagonal block */
1994: PetscCall(VecScatterEnd(aij->Mvctx, rr, aij->lvec, INSERT_VALUES, SCATTER_FORWARD));
1995: PetscUseTypeMethod(b, diagonalscale, NULL, aij->lvec);
1996: }
1997: PetscFunctionReturn(PETSC_SUCCESS);
1998: }
2000: static PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2001: {
2002: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2004: PetscFunctionBegin;
2005: PetscCall(MatSetUnfactored(a->A));
2006: PetscFunctionReturn(PETSC_SUCCESS);
2007: }
2009: static PetscErrorCode MatEqual_MPIAIJ(Mat A, Mat B, PetscBool *flag)
2010: {
2011: Mat_MPIAIJ *matB = (Mat_MPIAIJ *)B->data, *matA = (Mat_MPIAIJ *)A->data;
2012: Mat a, b, c, d;
2013: PetscBool flg;
2015: PetscFunctionBegin;
2016: a = matA->A;
2017: b = matA->B;
2018: c = matB->A;
2019: d = matB->B;
2021: PetscCall(MatEqual(a, c, &flg));
2022: if (flg) PetscCall(MatEqual(b, d, &flg));
2023: PetscCallMPI(MPIU_Allreduce(&flg, flag, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)A)));
2024: PetscFunctionReturn(PETSC_SUCCESS);
2025: }
2027: static PetscErrorCode MatCopy_MPIAIJ(Mat A, Mat B, MatStructure str)
2028: {
2029: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2030: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2032: PetscFunctionBegin;
2033: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2034: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2035: /* because of the column compression in the off-processor part of the matrix a->B,
2036: the number of columns in a->B and b->B may be different, hence we cannot call
2037: the MatCopy() directly on the two parts. If need be, we can provide a more
2038: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2039: then copying the submatrices */
2040: PetscCall(MatCopy_Basic(A, B, str));
2041: } else {
2042: PetscCall(MatCopy(a->A, b->A, str));
2043: PetscCall(MatCopy(a->B, b->B, str));
2044: }
2045: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2046: PetscFunctionReturn(PETSC_SUCCESS);
2047: }
2049: /*
2050: Computes the number of nonzeros per row needed for preallocation when X and Y
2051: have different nonzero structure.
2052: */
2053: 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)
2054: {
2055: PetscInt i, j, k, nzx, nzy;
2057: PetscFunctionBegin;
2058: /* Set the number of nonzeros in the new matrix */
2059: for (i = 0; i < m; i++) {
2060: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2061: nzx = xi[i + 1] - xi[i];
2062: nzy = yi[i + 1] - yi[i];
2063: nnz[i] = 0;
2064: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2065: for (; k < nzy && yltog[yjj[k]] < xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2066: if (k < nzy && yltog[yjj[k]] == xltog[xjj[j]]) k++; /* Skip duplicate */
2067: nnz[i]++;
2068: }
2069: for (; k < nzy; k++) nnz[i]++;
2070: }
2071: PetscFunctionReturn(PETSC_SUCCESS);
2072: }
2074: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2075: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y, const PetscInt *yltog, Mat X, const PetscInt *xltog, PetscInt *nnz)
2076: {
2077: PetscInt m = Y->rmap->N;
2078: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2079: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2081: PetscFunctionBegin;
2082: PetscCall(MatAXPYGetPreallocation_MPIX_private(m, x->i, x->j, xltog, y->i, y->j, yltog, nnz));
2083: PetscFunctionReturn(PETSC_SUCCESS);
2084: }
2086: static PetscErrorCode MatAXPY_MPIAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2087: {
2088: Mat_MPIAIJ *xx = (Mat_MPIAIJ *)X->data, *yy = (Mat_MPIAIJ *)Y->data;
2090: PetscFunctionBegin;
2091: if (str == SAME_NONZERO_PATTERN) {
2092: PetscCall(MatAXPY(yy->A, a, xx->A, str));
2093: PetscCall(MatAXPY(yy->B, a, xx->B, str));
2094: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2095: PetscCall(MatAXPY_Basic(Y, a, X, str));
2096: } else {
2097: Mat B;
2098: PetscInt *nnz_d, *nnz_o;
2100: PetscCall(PetscMalloc1(yy->A->rmap->N, &nnz_d));
2101: PetscCall(PetscMalloc1(yy->B->rmap->N, &nnz_o));
2102: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2103: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2104: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2105: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2106: PetscCall(MatAXPYGetPreallocation_SeqAIJ(yy->A, xx->A, nnz_d));
2107: PetscCall(MatAXPYGetPreallocation_MPIAIJ(yy->B, yy->garray, xx->B, xx->garray, nnz_o));
2108: PetscCall(MatMPIAIJSetPreallocation(B, 0, nnz_d, 0, nnz_o));
2109: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2110: PetscCall(MatHeaderMerge(Y, &B));
2111: PetscCall(PetscFree(nnz_d));
2112: PetscCall(PetscFree(nnz_o));
2113: }
2114: PetscFunctionReturn(PETSC_SUCCESS);
2115: }
2117: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat);
2119: static PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2120: {
2121: PetscFunctionBegin;
2122: if (PetscDefined(USE_COMPLEX)) {
2123: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2125: PetscCall(MatConjugate_SeqAIJ(aij->A));
2126: PetscCall(MatConjugate_SeqAIJ(aij->B));
2127: }
2128: PetscFunctionReturn(PETSC_SUCCESS);
2129: }
2131: static PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2132: {
2133: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2135: PetscFunctionBegin;
2136: PetscCall(MatRealPart(a->A));
2137: PetscCall(MatRealPart(a->B));
2138: PetscFunctionReturn(PETSC_SUCCESS);
2139: }
2141: static PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2142: {
2143: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2145: PetscFunctionBegin;
2146: PetscCall(MatImaginaryPart(a->A));
2147: PetscCall(MatImaginaryPart(a->B));
2148: PetscFunctionReturn(PETSC_SUCCESS);
2149: }
2151: static PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2152: {
2153: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2154: PetscInt i, *idxb = NULL, m = A->rmap->n;
2155: PetscScalar *va, *vv;
2156: Vec vB, vA;
2157: const PetscScalar *vb;
2159: PetscFunctionBegin;
2160: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2161: PetscCall(MatGetRowMaxAbs(a->A, vA, idx));
2163: PetscCall(VecGetArrayWrite(vA, &va));
2164: if (idx) {
2165: for (i = 0; i < m; i++) {
2166: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2167: }
2168: }
2170: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2171: PetscCall(PetscMalloc1(m, &idxb));
2172: PetscCall(MatGetRowMaxAbs(a->B, vB, idxb));
2174: PetscCall(VecGetArrayWrite(v, &vv));
2175: PetscCall(VecGetArrayRead(vB, &vb));
2176: for (i = 0; i < m; i++) {
2177: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2178: vv[i] = vb[i];
2179: if (idx) idx[i] = a->garray[idxb[i]];
2180: } else {
2181: vv[i] = va[i];
2182: if (idx && PetscAbsScalar(va[i]) == PetscAbsScalar(vb[i]) && idxb[i] != -1 && idx[i] > a->garray[idxb[i]]) idx[i] = a->garray[idxb[i]];
2183: }
2184: }
2185: PetscCall(VecRestoreArrayWrite(vA, &vv));
2186: PetscCall(VecRestoreArrayWrite(vA, &va));
2187: PetscCall(VecRestoreArrayRead(vB, &vb));
2188: PetscCall(PetscFree(idxb));
2189: PetscCall(VecDestroy(&vA));
2190: PetscCall(VecDestroy(&vB));
2191: PetscFunctionReturn(PETSC_SUCCESS);
2192: }
2194: static PetscErrorCode MatGetRowSumAbs_MPIAIJ(Mat A, Vec v)
2195: {
2196: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2197: Vec vB, vA;
2199: PetscFunctionBegin;
2200: PetscCall(MatCreateVecs(a->A, NULL, &vA));
2201: PetscCall(MatGetRowSumAbs(a->A, vA));
2202: PetscCall(MatCreateVecs(a->B, NULL, &vB));
2203: PetscCall(MatGetRowSumAbs(a->B, vB));
2204: PetscCall(VecAXPY(vA, 1.0, vB));
2205: PetscCall(VecDestroy(&vB));
2206: PetscCall(VecCopy(vA, v));
2207: PetscCall(VecDestroy(&vA));
2208: PetscFunctionReturn(PETSC_SUCCESS);
2209: }
2211: static PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2212: {
2213: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2214: PetscInt m = A->rmap->n, n = A->cmap->n;
2215: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2216: PetscInt *cmap = mat->garray;
2217: PetscInt *diagIdx, *offdiagIdx;
2218: Vec diagV, offdiagV;
2219: PetscScalar *a, *diagA, *offdiagA;
2220: const PetscScalar *ba, *bav;
2221: PetscInt r, j, col, ncols, *bi, *bj;
2222: Mat B = mat->B;
2223: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2225: PetscFunctionBegin;
2226: /* When a process holds entire A and other processes have no entry */
2227: if (A->cmap->N == n) {
2228: PetscCall(VecGetArrayWrite(v, &diagA));
2229: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2230: PetscCall(MatGetRowMinAbs(mat->A, diagV, idx));
2231: PetscCall(VecDestroy(&diagV));
2232: PetscCall(VecRestoreArrayWrite(v, &diagA));
2233: PetscFunctionReturn(PETSC_SUCCESS);
2234: } else if (n == 0) {
2235: if (m) {
2236: PetscCall(VecGetArrayWrite(v, &a));
2237: for (r = 0; r < m; r++) {
2238: a[r] = 0.0;
2239: if (idx) idx[r] = -1;
2240: }
2241: PetscCall(VecRestoreArrayWrite(v, &a));
2242: }
2243: PetscFunctionReturn(PETSC_SUCCESS);
2244: }
2246: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2247: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2248: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2249: PetscCall(MatGetRowMinAbs(mat->A, diagV, diagIdx));
2251: /* Get offdiagIdx[] for implicit 0.0 */
2252: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2253: ba = bav;
2254: bi = b->i;
2255: bj = b->j;
2256: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2257: for (r = 0; r < m; r++) {
2258: ncols = bi[r + 1] - bi[r];
2259: if (ncols == A->cmap->N - n) { /* Brow is dense */
2260: offdiagA[r] = *ba;
2261: offdiagIdx[r] = cmap[0];
2262: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2263: offdiagA[r] = 0.0;
2265: /* Find first hole in the cmap */
2266: for (j = 0; j < ncols; j++) {
2267: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2268: if (col > j && j < cstart) {
2269: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2270: break;
2271: } else if (col > j + n && j >= cstart) {
2272: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2273: break;
2274: }
2275: }
2276: if (j == ncols && ncols < A->cmap->N - n) {
2277: /* a hole is outside compressed Bcols */
2278: if (ncols == 0) {
2279: if (cstart) {
2280: offdiagIdx[r] = 0;
2281: } else offdiagIdx[r] = cend;
2282: } else { /* ncols > 0 */
2283: offdiagIdx[r] = cmap[ncols - 1] + 1;
2284: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2285: }
2286: }
2287: }
2289: for (j = 0; j < ncols; j++) {
2290: if (PetscAbsScalar(offdiagA[r]) > PetscAbsScalar(*ba)) {
2291: offdiagA[r] = *ba;
2292: offdiagIdx[r] = cmap[*bj];
2293: }
2294: ba++;
2295: bj++;
2296: }
2297: }
2299: PetscCall(VecGetArrayWrite(v, &a));
2300: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2301: for (r = 0; r < m; ++r) {
2302: if (PetscAbsScalar(diagA[r]) < PetscAbsScalar(offdiagA[r])) {
2303: a[r] = diagA[r];
2304: if (idx) idx[r] = cstart + diagIdx[r];
2305: } else if (PetscAbsScalar(diagA[r]) == PetscAbsScalar(offdiagA[r])) {
2306: a[r] = diagA[r];
2307: if (idx) {
2308: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2309: idx[r] = cstart + diagIdx[r];
2310: } else idx[r] = offdiagIdx[r];
2311: }
2312: } else {
2313: a[r] = offdiagA[r];
2314: if (idx) idx[r] = offdiagIdx[r];
2315: }
2316: }
2317: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2318: PetscCall(VecRestoreArrayWrite(v, &a));
2319: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2320: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2321: PetscCall(VecDestroy(&diagV));
2322: PetscCall(VecDestroy(&offdiagV));
2323: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2324: PetscFunctionReturn(PETSC_SUCCESS);
2325: }
2327: static PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2328: {
2329: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2330: PetscInt m = A->rmap->n, n = A->cmap->n;
2331: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2332: PetscInt *cmap = mat->garray;
2333: PetscInt *diagIdx, *offdiagIdx;
2334: Vec diagV, offdiagV;
2335: PetscScalar *a, *diagA, *offdiagA;
2336: const PetscScalar *ba, *bav;
2337: PetscInt r, j, col, ncols, *bi, *bj;
2338: Mat B = mat->B;
2339: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2341: PetscFunctionBegin;
2342: /* When a process holds entire A and other processes have no entry */
2343: if (A->cmap->N == n) {
2344: PetscCall(VecGetArrayWrite(v, &diagA));
2345: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2346: PetscCall(MatGetRowMin(mat->A, diagV, idx));
2347: PetscCall(VecDestroy(&diagV));
2348: PetscCall(VecRestoreArrayWrite(v, &diagA));
2349: PetscFunctionReturn(PETSC_SUCCESS);
2350: } else if (n == 0) {
2351: if (m) {
2352: PetscCall(VecGetArrayWrite(v, &a));
2353: for (r = 0; r < m; r++) {
2354: a[r] = PETSC_MAX_REAL;
2355: if (idx) idx[r] = -1;
2356: }
2357: PetscCall(VecRestoreArrayWrite(v, &a));
2358: }
2359: PetscFunctionReturn(PETSC_SUCCESS);
2360: }
2362: PetscCall(PetscCalloc2(m, &diagIdx, m, &offdiagIdx));
2363: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2364: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2365: PetscCall(MatGetRowMin(mat->A, diagV, diagIdx));
2367: /* Get offdiagIdx[] for implicit 0.0 */
2368: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2369: ba = bav;
2370: bi = b->i;
2371: bj = b->j;
2372: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2373: for (r = 0; r < m; r++) {
2374: ncols = bi[r + 1] - bi[r];
2375: if (ncols == A->cmap->N - n) { /* Brow is dense */
2376: offdiagA[r] = *ba;
2377: offdiagIdx[r] = cmap[0];
2378: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2379: offdiagA[r] = 0.0;
2381: /* Find first hole in the cmap */
2382: for (j = 0; j < ncols; j++) {
2383: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2384: if (col > j && j < cstart) {
2385: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2386: break;
2387: } else if (col > j + n && j >= cstart) {
2388: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2389: break;
2390: }
2391: }
2392: if (j == ncols && ncols < A->cmap->N - n) {
2393: /* a hole is outside compressed Bcols */
2394: if (ncols == 0) {
2395: if (cstart) {
2396: offdiagIdx[r] = 0;
2397: } else offdiagIdx[r] = cend;
2398: } else { /* ncols > 0 */
2399: offdiagIdx[r] = cmap[ncols - 1] + 1;
2400: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2401: }
2402: }
2403: }
2405: for (j = 0; j < ncols; j++) {
2406: if (PetscRealPart(offdiagA[r]) > PetscRealPart(*ba)) {
2407: offdiagA[r] = *ba;
2408: offdiagIdx[r] = cmap[*bj];
2409: }
2410: ba++;
2411: bj++;
2412: }
2413: }
2415: PetscCall(VecGetArrayWrite(v, &a));
2416: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2417: for (r = 0; r < m; ++r) {
2418: if (PetscRealPart(diagA[r]) < PetscRealPart(offdiagA[r])) {
2419: a[r] = diagA[r];
2420: if (idx) idx[r] = cstart + diagIdx[r];
2421: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2422: a[r] = diagA[r];
2423: if (idx) {
2424: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2425: idx[r] = cstart + diagIdx[r];
2426: } else idx[r] = offdiagIdx[r];
2427: }
2428: } else {
2429: a[r] = offdiagA[r];
2430: if (idx) idx[r] = offdiagIdx[r];
2431: }
2432: }
2433: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2434: PetscCall(VecRestoreArrayWrite(v, &a));
2435: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2436: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2437: PetscCall(VecDestroy(&diagV));
2438: PetscCall(VecDestroy(&offdiagV));
2439: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2440: PetscFunctionReturn(PETSC_SUCCESS);
2441: }
2443: static PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2444: {
2445: Mat_MPIAIJ *mat = (Mat_MPIAIJ *)A->data;
2446: PetscInt m = A->rmap->n, n = A->cmap->n;
2447: PetscInt cstart = A->cmap->rstart, cend = A->cmap->rend;
2448: PetscInt *cmap = mat->garray;
2449: PetscInt *diagIdx, *offdiagIdx;
2450: Vec diagV, offdiagV;
2451: PetscScalar *a, *diagA, *offdiagA;
2452: const PetscScalar *ba, *bav;
2453: PetscInt r, j, col, ncols, *bi, *bj;
2454: Mat B = mat->B;
2455: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2457: PetscFunctionBegin;
2458: /* When a process holds entire A and other processes have no entry */
2459: if (A->cmap->N == n) {
2460: PetscCall(VecGetArrayWrite(v, &diagA));
2461: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, 1, m, diagA, &diagV));
2462: PetscCall(MatGetRowMax(mat->A, diagV, idx));
2463: PetscCall(VecDestroy(&diagV));
2464: PetscCall(VecRestoreArrayWrite(v, &diagA));
2465: PetscFunctionReturn(PETSC_SUCCESS);
2466: } else if (n == 0) {
2467: if (m) {
2468: PetscCall(VecGetArrayWrite(v, &a));
2469: for (r = 0; r < m; r++) {
2470: a[r] = PETSC_MIN_REAL;
2471: if (idx) idx[r] = -1;
2472: }
2473: PetscCall(VecRestoreArrayWrite(v, &a));
2474: }
2475: PetscFunctionReturn(PETSC_SUCCESS);
2476: }
2478: PetscCall(PetscMalloc2(m, &diagIdx, m, &offdiagIdx));
2479: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &diagV));
2480: PetscCall(VecCreateSeq(PETSC_COMM_SELF, m, &offdiagV));
2481: PetscCall(MatGetRowMax(mat->A, diagV, diagIdx));
2483: /* Get offdiagIdx[] for implicit 0.0 */
2484: PetscCall(MatSeqAIJGetArrayRead(B, &bav));
2485: ba = bav;
2486: bi = b->i;
2487: bj = b->j;
2488: PetscCall(VecGetArrayWrite(offdiagV, &offdiagA));
2489: for (r = 0; r < m; r++) {
2490: ncols = bi[r + 1] - bi[r];
2491: if (ncols == A->cmap->N - n) { /* Brow is dense */
2492: offdiagA[r] = *ba;
2493: offdiagIdx[r] = cmap[0];
2494: } else { /* Brow is sparse so already KNOW maximum is 0.0 or higher */
2495: offdiagA[r] = 0.0;
2497: /* Find first hole in the cmap */
2498: for (j = 0; j < ncols; j++) {
2499: col = cmap[bj[j]]; /* global column number = cmap[B column number] */
2500: if (col > j && j < cstart) {
2501: offdiagIdx[r] = j; /* global column number of first implicit 0.0 */
2502: break;
2503: } else if (col > j + n && j >= cstart) {
2504: offdiagIdx[r] = j + n; /* global column number of first implicit 0.0 */
2505: break;
2506: }
2507: }
2508: if (j == ncols && ncols < A->cmap->N - n) {
2509: /* a hole is outside compressed Bcols */
2510: if (ncols == 0) {
2511: if (cstart) {
2512: offdiagIdx[r] = 0;
2513: } else offdiagIdx[r] = cend;
2514: } else { /* ncols > 0 */
2515: offdiagIdx[r] = cmap[ncols - 1] + 1;
2516: if (offdiagIdx[r] == cstart) offdiagIdx[r] += n;
2517: }
2518: }
2519: }
2521: for (j = 0; j < ncols; j++) {
2522: if (PetscRealPart(offdiagA[r]) < PetscRealPart(*ba)) {
2523: offdiagA[r] = *ba;
2524: offdiagIdx[r] = cmap[*bj];
2525: }
2526: ba++;
2527: bj++;
2528: }
2529: }
2531: PetscCall(VecGetArrayWrite(v, &a));
2532: PetscCall(VecGetArrayRead(diagV, (const PetscScalar **)&diagA));
2533: for (r = 0; r < m; ++r) {
2534: if (PetscRealPart(diagA[r]) > PetscRealPart(offdiagA[r])) {
2535: a[r] = diagA[r];
2536: if (idx) idx[r] = cstart + diagIdx[r];
2537: } else if (PetscRealPart(diagA[r]) == PetscRealPart(offdiagA[r])) {
2538: a[r] = diagA[r];
2539: if (idx) {
2540: if (cstart + diagIdx[r] <= offdiagIdx[r]) {
2541: idx[r] = cstart + diagIdx[r];
2542: } else idx[r] = offdiagIdx[r];
2543: }
2544: } else {
2545: a[r] = offdiagA[r];
2546: if (idx) idx[r] = offdiagIdx[r];
2547: }
2548: }
2549: PetscCall(MatSeqAIJRestoreArrayRead(B, &bav));
2550: PetscCall(VecRestoreArrayWrite(v, &a));
2551: PetscCall(VecRestoreArrayRead(diagV, (const PetscScalar **)&diagA));
2552: PetscCall(VecRestoreArrayWrite(offdiagV, &offdiagA));
2553: PetscCall(VecDestroy(&diagV));
2554: PetscCall(VecDestroy(&offdiagV));
2555: PetscCall(PetscFree2(diagIdx, offdiagIdx));
2556: PetscFunctionReturn(PETSC_SUCCESS);
2557: }
2559: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat, Mat *newmat)
2560: {
2561: Mat *dummy;
2563: PetscFunctionBegin;
2564: PetscCall(MatCreateSubMatrix_MPIAIJ_All(mat, MAT_DO_NOT_GET_VALUES, MAT_INITIAL_MATRIX, &dummy));
2565: *newmat = *dummy;
2566: PetscCall(PetscFree(dummy));
2567: PetscFunctionReturn(PETSC_SUCCESS);
2568: }
2570: static PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A, const PetscScalar **values)
2571: {
2572: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2574: PetscFunctionBegin;
2575: PetscCall(MatInvertBlockDiagonal(a->A, values));
2576: A->factorerrortype = a->A->factorerrortype;
2577: PetscFunctionReturn(PETSC_SUCCESS);
2578: }
2580: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x, PetscRandom rctx)
2581: {
2582: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)x->data;
2584: PetscFunctionBegin;
2585: PetscCheck(x->assembled || x->preallocated, PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2586: PetscCall(MatSetRandom(aij->A, rctx));
2587: if (x->assembled) {
2588: PetscCall(MatSetRandom(aij->B, rctx));
2589: } else {
2590: PetscCall(MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B, x->cmap->rstart, x->cmap->rend, rctx));
2591: }
2592: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
2593: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
2594: PetscFunctionReturn(PETSC_SUCCESS);
2595: }
2597: static PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A, PetscBool sc)
2598: {
2599: PetscFunctionBegin;
2600: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2601: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2602: PetscFunctionReturn(PETSC_SUCCESS);
2603: }
2605: /*@
2606: MatMPIAIJGetNumberNonzeros - gets the number of nonzeros in the matrix on this MPI rank
2608: Not Collective
2610: Input Parameter:
2611: . A - the matrix
2613: Output Parameter:
2614: . nz - the number of nonzeros
2616: Level: advanced
2618: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2619: @*/
2620: PetscErrorCode MatMPIAIJGetNumberNonzeros(Mat A, PetscCount *nz)
2621: {
2622: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)A->data;
2623: Mat_SeqAIJ *aaij = (Mat_SeqAIJ *)maij->A->data, *baij = (Mat_SeqAIJ *)maij->B->data;
2624: PetscBool isaij;
2626: PetscFunctionBegin;
2627: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATMPIAIJ, &isaij));
2628: PetscCheck(isaij, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Not for type %s", ((PetscObject)A)->type_name);
2629: *nz = aaij->i[A->rmap->n] + baij->i[A->rmap->n];
2630: PetscFunctionReturn(PETSC_SUCCESS);
2631: }
2633: /*@
2634: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2636: Collective
2638: Input Parameters:
2639: + A - the matrix
2640: - sc - `PETSC_TRUE` indicates use the scalable algorithm (default is not to use the scalable algorithm)
2642: Level: advanced
2644: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`
2645: @*/
2646: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A, PetscBool sc)
2647: {
2648: PetscFunctionBegin;
2649: PetscTryMethod(A, "MatMPIAIJSetUseScalableIncreaseOverlap_C", (Mat, PetscBool), (A, sc));
2650: PetscFunctionReturn(PETSC_SUCCESS);
2651: }
2653: PetscErrorCode MatSetFromOptions_MPIAIJ(Mat A, PetscOptionItems *PetscOptionsObject)
2654: {
2655: PetscBool sc = PETSC_FALSE, flg;
2657: PetscFunctionBegin;
2658: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJ options");
2659: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2660: PetscCall(PetscOptionsBool("-mat_increase_overlap_scalable", "Use a scalable algorithm to compute the overlap", "MatIncreaseOverlap", sc, &sc, &flg));
2661: if (flg) PetscCall(MatMPIAIJSetUseScalableIncreaseOverlap(A, sc));
2662: PetscOptionsHeadEnd();
2663: PetscFunctionReturn(PETSC_SUCCESS);
2664: }
2666: static PetscErrorCode MatShift_MPIAIJ(Mat Y, PetscScalar a)
2667: {
2668: Mat_MPIAIJ *maij = (Mat_MPIAIJ *)Y->data;
2669: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)maij->A->data;
2671: PetscFunctionBegin;
2672: if (!Y->preallocated) {
2673: PetscCall(MatMPIAIJSetPreallocation(Y, 1, NULL, 0, NULL));
2674: } else if (!aij->nz) { /* It does not matter if diagonals of Y only partially lie in maij->A. We just need an estimated preallocation. */
2675: PetscInt nonew = aij->nonew;
2676: PetscCall(MatSeqAIJSetPreallocation(maij->A, 1, NULL));
2677: aij->nonew = nonew;
2678: }
2679: PetscCall(MatShift_Basic(Y, a));
2680: PetscFunctionReturn(PETSC_SUCCESS);
2681: }
2683: static PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A, PetscBool *missing, PetscInt *d)
2684: {
2685: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2687: PetscFunctionBegin;
2688: PetscCheck(A->rmap->n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only works for square matrices");
2689: PetscCall(MatMissingDiagonal(a->A, missing, d));
2690: if (d) {
2691: PetscInt rstart;
2692: PetscCall(MatGetOwnershipRange(A, &rstart, NULL));
2693: *d += rstart;
2694: }
2695: PetscFunctionReturn(PETSC_SUCCESS);
2696: }
2698: static PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
2699: {
2700: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2702: PetscFunctionBegin;
2703: PetscCall(MatInvertVariableBlockDiagonal(a->A, nblocks, bsizes, diag));
2704: PetscFunctionReturn(PETSC_SUCCESS);
2705: }
2707: static PetscErrorCode MatEliminateZeros_MPIAIJ(Mat A, PetscBool keep)
2708: {
2709: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
2711: PetscFunctionBegin;
2712: PetscCall(MatEliminateZeros_SeqAIJ(a->A, keep)); // possibly keep zero diagonal coefficients
2713: PetscCall(MatEliminateZeros_SeqAIJ(a->B, PETSC_FALSE)); // never keep zero diagonal coefficients
2714: PetscFunctionReturn(PETSC_SUCCESS);
2715: }
2717: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2718: MatGetRow_MPIAIJ,
2719: MatRestoreRow_MPIAIJ,
2720: MatMult_MPIAIJ,
2721: /* 4*/ MatMultAdd_MPIAIJ,
2722: MatMultTranspose_MPIAIJ,
2723: MatMultTransposeAdd_MPIAIJ,
2724: NULL,
2725: NULL,
2726: NULL,
2727: /*10*/ NULL,
2728: NULL,
2729: NULL,
2730: MatSOR_MPIAIJ,
2731: MatTranspose_MPIAIJ,
2732: /*15*/ MatGetInfo_MPIAIJ,
2733: MatEqual_MPIAIJ,
2734: MatGetDiagonal_MPIAIJ,
2735: MatDiagonalScale_MPIAIJ,
2736: MatNorm_MPIAIJ,
2737: /*20*/ MatAssemblyBegin_MPIAIJ,
2738: MatAssemblyEnd_MPIAIJ,
2739: MatSetOption_MPIAIJ,
2740: MatZeroEntries_MPIAIJ,
2741: /*24*/ MatZeroRows_MPIAIJ,
2742: NULL,
2743: NULL,
2744: NULL,
2745: NULL,
2746: /*29*/ MatSetUp_MPI_Hash,
2747: NULL,
2748: NULL,
2749: MatGetDiagonalBlock_MPIAIJ,
2750: NULL,
2751: /*34*/ MatDuplicate_MPIAIJ,
2752: NULL,
2753: NULL,
2754: NULL,
2755: NULL,
2756: /*39*/ MatAXPY_MPIAIJ,
2757: MatCreateSubMatrices_MPIAIJ,
2758: MatIncreaseOverlap_MPIAIJ,
2759: MatGetValues_MPIAIJ,
2760: MatCopy_MPIAIJ,
2761: /*44*/ MatGetRowMax_MPIAIJ,
2762: MatScale_MPIAIJ,
2763: MatShift_MPIAIJ,
2764: MatDiagonalSet_MPIAIJ,
2765: MatZeroRowsColumns_MPIAIJ,
2766: /*49*/ MatSetRandom_MPIAIJ,
2767: MatGetRowIJ_MPIAIJ,
2768: MatRestoreRowIJ_MPIAIJ,
2769: NULL,
2770: NULL,
2771: /*54*/ MatFDColoringCreate_MPIXAIJ,
2772: NULL,
2773: MatSetUnfactored_MPIAIJ,
2774: MatPermute_MPIAIJ,
2775: NULL,
2776: /*59*/ MatCreateSubMatrix_MPIAIJ,
2777: MatDestroy_MPIAIJ,
2778: MatView_MPIAIJ,
2779: NULL,
2780: NULL,
2781: /*64*/ NULL,
2782: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2783: NULL,
2784: NULL,
2785: NULL,
2786: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2787: MatGetRowMinAbs_MPIAIJ,
2788: NULL,
2789: NULL,
2790: NULL,
2791: NULL,
2792: /*75*/ MatFDColoringApply_AIJ,
2793: MatSetFromOptions_MPIAIJ,
2794: NULL,
2795: NULL,
2796: MatFindZeroDiagonals_MPIAIJ,
2797: /*80*/ NULL,
2798: NULL,
2799: NULL,
2800: /*83*/ MatLoad_MPIAIJ,
2801: NULL,
2802: NULL,
2803: NULL,
2804: NULL,
2805: NULL,
2806: /*89*/ NULL,
2807: NULL,
2808: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2809: NULL,
2810: NULL,
2811: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2812: NULL,
2813: NULL,
2814: NULL,
2815: MatBindToCPU_MPIAIJ,
2816: /*99*/ MatProductSetFromOptions_MPIAIJ,
2817: NULL,
2818: NULL,
2819: MatConjugate_MPIAIJ,
2820: NULL,
2821: /*104*/ MatSetValuesRow_MPIAIJ,
2822: MatRealPart_MPIAIJ,
2823: MatImaginaryPart_MPIAIJ,
2824: NULL,
2825: NULL,
2826: /*109*/ NULL,
2827: NULL,
2828: MatGetRowMin_MPIAIJ,
2829: NULL,
2830: MatMissingDiagonal_MPIAIJ,
2831: /*114*/ MatGetSeqNonzeroStructure_MPIAIJ,
2832: NULL,
2833: MatGetGhosts_MPIAIJ,
2834: NULL,
2835: NULL,
2836: /*119*/ MatMultDiagonalBlock_MPIAIJ,
2837: NULL,
2838: NULL,
2839: NULL,
2840: MatGetMultiProcBlock_MPIAIJ,
2841: /*124*/ MatFindNonzeroRows_MPIAIJ,
2842: MatGetColumnReductions_MPIAIJ,
2843: MatInvertBlockDiagonal_MPIAIJ,
2844: MatInvertVariableBlockDiagonal_MPIAIJ,
2845: MatCreateSubMatricesMPI_MPIAIJ,
2846: /*129*/ NULL,
2847: NULL,
2848: NULL,
2849: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2850: NULL,
2851: /*134*/ NULL,
2852: NULL,
2853: NULL,
2854: NULL,
2855: NULL,
2856: /*139*/ MatSetBlockSizes_MPIAIJ,
2857: NULL,
2858: NULL,
2859: MatFDColoringSetUp_MPIXAIJ,
2860: MatFindOffBlockDiagonalEntries_MPIAIJ,
2861: MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2862: /*145*/ NULL,
2863: NULL,
2864: NULL,
2865: MatCreateGraph_Simple_AIJ,
2866: NULL,
2867: /*150*/ NULL,
2868: MatEliminateZeros_MPIAIJ,
2869: MatGetRowSumAbs_MPIAIJ,
2870: NULL,
2871: NULL,
2872: NULL};
2874: static PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2875: {
2876: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2878: PetscFunctionBegin;
2879: PetscCall(MatStoreValues(aij->A));
2880: PetscCall(MatStoreValues(aij->B));
2881: PetscFunctionReturn(PETSC_SUCCESS);
2882: }
2884: static PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2885: {
2886: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
2888: PetscFunctionBegin;
2889: PetscCall(MatRetrieveValues(aij->A));
2890: PetscCall(MatRetrieveValues(aij->B));
2891: PetscFunctionReturn(PETSC_SUCCESS);
2892: }
2894: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
2895: {
2896: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2897: PetscMPIInt size;
2899: PetscFunctionBegin;
2900: if (B->hash_active) {
2901: B->ops[0] = b->cops;
2902: B->hash_active = PETSC_FALSE;
2903: }
2904: PetscCall(PetscLayoutSetUp(B->rmap));
2905: PetscCall(PetscLayoutSetUp(B->cmap));
2907: #if defined(PETSC_USE_CTABLE)
2908: PetscCall(PetscHMapIDestroy(&b->colmap));
2909: #else
2910: PetscCall(PetscFree(b->colmap));
2911: #endif
2912: PetscCall(PetscFree(b->garray));
2913: PetscCall(VecDestroy(&b->lvec));
2914: PetscCall(VecScatterDestroy(&b->Mvctx));
2916: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
2918: MatSeqXAIJGetOptions_Private(b->B);
2919: PetscCall(MatDestroy(&b->B));
2920: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
2921: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
2922: PetscCall(MatSetBlockSizesFromMats(b->B, B, B));
2923: PetscCall(MatSetType(b->B, MATSEQAIJ));
2924: MatSeqXAIJRestoreOptions_Private(b->B);
2926: MatSeqXAIJGetOptions_Private(b->A);
2927: PetscCall(MatDestroy(&b->A));
2928: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
2929: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
2930: PetscCall(MatSetBlockSizesFromMats(b->A, B, B));
2931: PetscCall(MatSetType(b->A, MATSEQAIJ));
2932: MatSeqXAIJRestoreOptions_Private(b->A);
2934: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
2935: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
2936: B->preallocated = PETSC_TRUE;
2937: B->was_assembled = PETSC_FALSE;
2938: B->assembled = PETSC_FALSE;
2939: PetscFunctionReturn(PETSC_SUCCESS);
2940: }
2942: static PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2943: {
2944: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
2946: PetscFunctionBegin;
2948: PetscCall(PetscLayoutSetUp(B->rmap));
2949: PetscCall(PetscLayoutSetUp(B->cmap));
2951: #if defined(PETSC_USE_CTABLE)
2952: PetscCall(PetscHMapIDestroy(&b->colmap));
2953: #else
2954: PetscCall(PetscFree(b->colmap));
2955: #endif
2956: PetscCall(PetscFree(b->garray));
2957: PetscCall(VecDestroy(&b->lvec));
2958: PetscCall(VecScatterDestroy(&b->Mvctx));
2960: PetscCall(MatResetPreallocation(b->A));
2961: PetscCall(MatResetPreallocation(b->B));
2962: B->preallocated = PETSC_TRUE;
2963: B->was_assembled = PETSC_FALSE;
2964: B->assembled = PETSC_FALSE;
2965: PetscFunctionReturn(PETSC_SUCCESS);
2966: }
2968: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2969: {
2970: Mat mat;
2971: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2973: PetscFunctionBegin;
2974: *newmat = NULL;
2975: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2976: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2977: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2978: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2979: a = (Mat_MPIAIJ *)mat->data;
2981: mat->factortype = matin->factortype;
2982: mat->assembled = matin->assembled;
2983: mat->insertmode = NOT_SET_VALUES;
2985: a->size = oldmat->size;
2986: a->rank = oldmat->rank;
2987: a->donotstash = oldmat->donotstash;
2988: a->roworiented = oldmat->roworiented;
2989: a->rowindices = NULL;
2990: a->rowvalues = NULL;
2991: a->getrowactive = PETSC_FALSE;
2993: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2994: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2995: if (matin->hash_active) {
2996: PetscCall(MatSetUp(mat));
2997: } else {
2998: mat->preallocated = matin->preallocated;
2999: if (oldmat->colmap) {
3000: #if defined(PETSC_USE_CTABLE)
3001: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3002: #else
3003: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3004: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3005: #endif
3006: } else a->colmap = NULL;
3007: if (oldmat->garray) {
3008: PetscInt len;
3009: len = oldmat->B->cmap->n;
3010: PetscCall(PetscMalloc1(len + 1, &a->garray));
3011: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3012: } else a->garray = NULL;
3014: /* It may happen MatDuplicate is called with a non-assembled matrix
3015: In fact, MatDuplicate only requires the matrix to be preallocated
3016: This may happen inside a DMCreateMatrix_Shell */
3017: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3018: if (oldmat->Mvctx) {
3019: a->Mvctx = oldmat->Mvctx;
3020: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3021: }
3022: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3023: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3024: }
3025: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3026: *newmat = mat;
3027: PetscFunctionReturn(PETSC_SUCCESS);
3028: }
3030: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3031: {
3032: PetscBool isbinary, ishdf5;
3034: PetscFunctionBegin;
3037: /* force binary viewer to load .info file if it has not yet done so */
3038: PetscCall(PetscViewerSetUp(viewer));
3039: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3040: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3041: if (isbinary) {
3042: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3043: } else if (ishdf5) {
3044: #if defined(PETSC_HAVE_HDF5)
3045: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3046: #else
3047: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3048: #endif
3049: } else {
3050: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "Viewer type %s not yet supported for reading %s matrices", ((PetscObject)viewer)->type_name, ((PetscObject)newMat)->type_name);
3051: }
3052: PetscFunctionReturn(PETSC_SUCCESS);
3053: }
3055: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3056: {
3057: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3058: PetscInt *rowidxs, *colidxs;
3059: PetscScalar *matvals;
3061: PetscFunctionBegin;
3062: PetscCall(PetscViewerSetUp(viewer));
3064: /* read in matrix header */
3065: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3066: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3067: M = header[1];
3068: N = header[2];
3069: nz = header[3];
3070: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3071: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3072: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3074: /* set block sizes from the viewer's .info file */
3075: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3076: /* set global sizes if not set already */
3077: if (mat->rmap->N < 0) mat->rmap->N = M;
3078: if (mat->cmap->N < 0) mat->cmap->N = N;
3079: PetscCall(PetscLayoutSetUp(mat->rmap));
3080: PetscCall(PetscLayoutSetUp(mat->cmap));
3082: /* check if the matrix sizes are correct */
3083: PetscCall(MatGetSize(mat, &rows, &cols));
3084: PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols);
3086: /* read in row lengths and build row indices */
3087: PetscCall(MatGetLocalSize(mat, &m, NULL));
3088: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3089: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3090: rowidxs[0] = 0;
3091: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3092: if (nz != PETSC_INT_MAX) {
3093: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3094: PetscCheck(sum == nz, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum);
3095: }
3097: /* read in column indices and matrix values */
3098: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3099: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3100: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3101: /* store matrix indices and values */
3102: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3103: PetscCall(PetscFree(rowidxs));
3104: PetscCall(PetscFree2(colidxs, matvals));
3105: PetscFunctionReturn(PETSC_SUCCESS);
3106: }
3108: /* Not scalable because of ISAllGather() unless getting all columns. */
3109: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3110: {
3111: IS iscol_local;
3112: PetscBool isstride;
3113: PetscMPIInt lisstride = 0, gisstride;
3115: PetscFunctionBegin;
3116: /* check if we are grabbing all columns*/
3117: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3119: if (isstride) {
3120: PetscInt start, len, mstart, mlen;
3121: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3122: PetscCall(ISGetLocalSize(iscol, &len));
3123: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3124: if (mstart == start && mlen - mstart == len) lisstride = 1;
3125: }
3127: PetscCallMPI(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3128: if (gisstride) {
3129: PetscInt N;
3130: PetscCall(MatGetSize(mat, NULL, &N));
3131: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3132: PetscCall(ISSetIdentity(iscol_local));
3133: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3134: } else {
3135: PetscInt cbs;
3136: PetscCall(ISGetBlockSize(iscol, &cbs));
3137: PetscCall(ISAllGather(iscol, &iscol_local));
3138: PetscCall(ISSetBlockSize(iscol_local, cbs));
3139: }
3141: *isseq = iscol_local;
3142: PetscFunctionReturn(PETSC_SUCCESS);
3143: }
3145: /*
3146: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3147: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3149: Input Parameters:
3150: + mat - matrix
3151: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3152: i.e., mat->rstart <= isrow[i] < mat->rend
3153: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3154: i.e., mat->cstart <= iscol[i] < mat->cend
3156: Output Parameters:
3157: + isrow_d - sequential row index set for retrieving mat->A
3158: . iscol_d - sequential column index set for retrieving mat->A
3159: . iscol_o - sequential column index set for retrieving mat->B
3160: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3161: */
3162: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3163: {
3164: Vec x, cmap;
3165: const PetscInt *is_idx;
3166: PetscScalar *xarray, *cmaparray;
3167: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3168: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3169: Mat B = a->B;
3170: Vec lvec = a->lvec, lcmap;
3171: PetscInt i, cstart, cend, Bn = B->cmap->N;
3172: MPI_Comm comm;
3173: VecScatter Mvctx = a->Mvctx;
3175: PetscFunctionBegin;
3176: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3177: PetscCall(ISGetLocalSize(iscol, &ncols));
3179: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3180: PetscCall(MatCreateVecs(mat, &x, NULL));
3181: PetscCall(VecSet(x, -1.0));
3182: PetscCall(VecDuplicate(x, &cmap));
3183: PetscCall(VecSet(cmap, -1.0));
3185: /* Get start indices */
3186: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3187: isstart -= ncols;
3188: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3190: PetscCall(ISGetIndices(iscol, &is_idx));
3191: PetscCall(VecGetArray(x, &xarray));
3192: PetscCall(VecGetArray(cmap, &cmaparray));
3193: PetscCall(PetscMalloc1(ncols, &idx));
3194: for (i = 0; i < ncols; i++) {
3195: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3196: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3197: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3198: }
3199: PetscCall(VecRestoreArray(x, &xarray));
3200: PetscCall(VecRestoreArray(cmap, &cmaparray));
3201: PetscCall(ISRestoreIndices(iscol, &is_idx));
3203: /* Get iscol_d */
3204: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3205: PetscCall(ISGetBlockSize(iscol, &i));
3206: PetscCall(ISSetBlockSize(*iscol_d, i));
3208: /* Get isrow_d */
3209: PetscCall(ISGetLocalSize(isrow, &m));
3210: rstart = mat->rmap->rstart;
3211: PetscCall(PetscMalloc1(m, &idx));
3212: PetscCall(ISGetIndices(isrow, &is_idx));
3213: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3214: PetscCall(ISRestoreIndices(isrow, &is_idx));
3216: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3217: PetscCall(ISGetBlockSize(isrow, &i));
3218: PetscCall(ISSetBlockSize(*isrow_d, i));
3220: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3221: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3222: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3224: PetscCall(VecDuplicate(lvec, &lcmap));
3226: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3227: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3229: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3230: /* off-process column indices */
3231: count = 0;
3232: PetscCall(PetscMalloc1(Bn, &idx));
3233: PetscCall(PetscMalloc1(Bn, &cmap1));
3235: PetscCall(VecGetArray(lvec, &xarray));
3236: PetscCall(VecGetArray(lcmap, &cmaparray));
3237: for (i = 0; i < Bn; i++) {
3238: if (PetscRealPart(xarray[i]) > -1.0) {
3239: idx[count] = i; /* local column index in off-diagonal part B */
3240: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3241: count++;
3242: }
3243: }
3244: PetscCall(VecRestoreArray(lvec, &xarray));
3245: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3247: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3248: /* cannot ensure iscol_o has same blocksize as iscol! */
3250: PetscCall(PetscFree(idx));
3251: *garray = cmap1;
3253: PetscCall(VecDestroy(&x));
3254: PetscCall(VecDestroy(&cmap));
3255: PetscCall(VecDestroy(&lcmap));
3256: PetscFunctionReturn(PETSC_SUCCESS);
3257: }
3259: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3260: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3261: {
3262: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3263: Mat M = NULL;
3264: MPI_Comm comm;
3265: IS iscol_d, isrow_d, iscol_o;
3266: Mat Asub = NULL, Bsub = NULL;
3267: PetscInt n;
3269: PetscFunctionBegin;
3270: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3272: if (call == MAT_REUSE_MATRIX) {
3273: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3274: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3275: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3277: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3278: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3280: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3281: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3283: /* Update diagonal and off-diagonal portions of submat */
3284: asub = (Mat_MPIAIJ *)(*submat)->data;
3285: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3286: PetscCall(ISGetLocalSize(iscol_o, &n));
3287: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3288: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3289: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3291: } else { /* call == MAT_INITIAL_MATRIX) */
3292: const PetscInt *garray;
3293: PetscInt BsubN;
3295: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3296: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3298: /* Create local submatrices Asub and Bsub */
3299: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3300: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3302: /* Create submatrix M */
3303: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3305: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3306: asub = (Mat_MPIAIJ *)M->data;
3308: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3309: n = asub->B->cmap->N;
3310: if (BsubN > n) {
3311: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3312: const PetscInt *idx;
3313: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3314: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3316: PetscCall(PetscMalloc1(n, &idx_new));
3317: j = 0;
3318: PetscCall(ISGetIndices(iscol_o, &idx));
3319: for (i = 0; i < n; i++) {
3320: if (j >= BsubN) break;
3321: while (subgarray[i] > garray[j]) j++;
3323: if (subgarray[i] == garray[j]) {
3324: idx_new[i] = idx[j++];
3325: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "subgarray[%" PetscInt_FMT "]=%" PetscInt_FMT " cannot < garray[%" PetscInt_FMT "]=%" PetscInt_FMT, i, subgarray[i], j, garray[j]);
3326: }
3327: PetscCall(ISRestoreIndices(iscol_o, &idx));
3329: PetscCall(ISDestroy(&iscol_o));
3330: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3332: } else if (BsubN < n) {
3333: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Columns of Bsub (%" PetscInt_FMT ") cannot be smaller than B's (%" PetscInt_FMT ")", BsubN, asub->B->cmap->N);
3334: }
3336: PetscCall(PetscFree(garray));
3337: *submat = M;
3339: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3340: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3341: PetscCall(ISDestroy(&isrow_d));
3343: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3344: PetscCall(ISDestroy(&iscol_d));
3346: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3347: PetscCall(ISDestroy(&iscol_o));
3348: }
3349: PetscFunctionReturn(PETSC_SUCCESS);
3350: }
3352: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3353: {
3354: IS iscol_local = NULL, isrow_d;
3355: PetscInt csize;
3356: PetscInt n, i, j, start, end;
3357: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3358: MPI_Comm comm;
3360: PetscFunctionBegin;
3361: /* If isrow has same processor distribution as mat,
3362: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3363: if (call == MAT_REUSE_MATRIX) {
3364: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3365: if (isrow_d) {
3366: sameRowDist = PETSC_TRUE;
3367: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3368: } else {
3369: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3370: if (iscol_local) {
3371: sameRowDist = PETSC_TRUE;
3372: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3373: }
3374: }
3375: } else {
3376: /* Check if isrow has same processor distribution as mat */
3377: sameDist[0] = PETSC_FALSE;
3378: PetscCall(ISGetLocalSize(isrow, &n));
3379: if (!n) {
3380: sameDist[0] = PETSC_TRUE;
3381: } else {
3382: PetscCall(ISGetMinMax(isrow, &i, &j));
3383: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3384: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3385: }
3387: /* Check if iscol has same processor distribution as mat */
3388: sameDist[1] = PETSC_FALSE;
3389: PetscCall(ISGetLocalSize(iscol, &n));
3390: if (!n) {
3391: sameDist[1] = PETSC_TRUE;
3392: } else {
3393: PetscCall(ISGetMinMax(iscol, &i, &j));
3394: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3395: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3396: }
3398: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3399: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3400: sameRowDist = tsameDist[0];
3401: }
3403: if (sameRowDist) {
3404: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3405: /* isrow and iscol have same processor distribution as mat */
3406: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3407: PetscFunctionReturn(PETSC_SUCCESS);
3408: } else { /* sameRowDist */
3409: /* isrow has same processor distribution as mat */
3410: if (call == MAT_INITIAL_MATRIX) {
3411: PetscBool sorted;
3412: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3413: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3414: PetscCall(ISGetSize(iscol, &i));
3415: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3417: PetscCall(ISSorted(iscol_local, &sorted));
3418: if (sorted) {
3419: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3420: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3421: PetscFunctionReturn(PETSC_SUCCESS);
3422: }
3423: } else { /* call == MAT_REUSE_MATRIX */
3424: IS iscol_sub;
3425: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3426: if (iscol_sub) {
3427: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3428: PetscFunctionReturn(PETSC_SUCCESS);
3429: }
3430: }
3431: }
3432: }
3434: /* General case: iscol -> iscol_local which has global size of iscol */
3435: if (call == MAT_REUSE_MATRIX) {
3436: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3437: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3438: } else {
3439: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3440: }
3442: PetscCall(ISGetLocalSize(iscol, &csize));
3443: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3445: if (call == MAT_INITIAL_MATRIX) {
3446: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3447: PetscCall(ISDestroy(&iscol_local));
3448: }
3449: PetscFunctionReturn(PETSC_SUCCESS);
3450: }
3452: /*@C
3453: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3454: and "off-diagonal" part of the matrix in CSR format.
3456: Collective
3458: Input Parameters:
3459: + comm - MPI communicator
3460: . A - "diagonal" portion of matrix
3461: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3462: - garray - global index of `B` columns
3464: Output Parameter:
3465: . mat - the matrix, with input `A` as its local diagonal matrix
3467: Level: advanced
3469: Notes:
3470: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3472: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3474: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3475: @*/
3476: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3477: {
3478: Mat_MPIAIJ *maij;
3479: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3480: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3481: const PetscScalar *oa;
3482: Mat Bnew;
3483: PetscInt m, n, N;
3484: MatType mpi_mat_type;
3486: PetscFunctionBegin;
3487: PetscCall(MatCreate(comm, mat));
3488: PetscCall(MatGetSize(A, &m, &n));
3489: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3490: PetscCheck(PetscAbs(A->rmap->bs) == PetscAbs(B->rmap->bs), PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "A row bs %" PetscInt_FMT " != B row bs %" PetscInt_FMT, A->rmap->bs, B->rmap->bs);
3491: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3492: /* PetscCheck(A->cmap->bs == B->cmap->bs,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %" PetscInt_FMT " != B column bs %" PetscInt_FMT,A->cmap->bs,B->cmap->bs); */
3494: /* Get global columns of mat */
3495: PetscCallMPI(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3497: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3498: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3499: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3500: PetscCall(MatSetType(*mat, mpi_mat_type));
3502: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3503: maij = (Mat_MPIAIJ *)(*mat)->data;
3505: (*mat)->preallocated = PETSC_TRUE;
3507: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3508: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3510: /* Set A as diagonal portion of *mat */
3511: maij->A = A;
3513: nz = oi[m];
3514: for (i = 0; i < nz; i++) {
3515: col = oj[i];
3516: oj[i] = garray[col];
3517: }
3519: /* Set Bnew as off-diagonal portion of *mat */
3520: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3521: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3522: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3523: bnew = (Mat_SeqAIJ *)Bnew->data;
3524: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3525: maij->B = Bnew;
3527: PetscCheck(B->rmap->N == Bnew->rmap->N, PETSC_COMM_SELF, PETSC_ERR_PLIB, "BN %" PetscInt_FMT " != BnewN %" PetscInt_FMT, B->rmap->N, Bnew->rmap->N);
3529: b->free_a = PETSC_FALSE;
3530: b->free_ij = PETSC_FALSE;
3531: PetscCall(MatDestroy(&B));
3533: bnew->free_a = PETSC_TRUE;
3534: bnew->free_ij = PETSC_TRUE;
3536: /* condense columns of maij->B */
3537: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3538: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3539: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3540: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3541: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3542: PetscFunctionReturn(PETSC_SUCCESS);
3543: }
3545: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3547: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3548: {
3549: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3550: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3551: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3552: Mat M, Msub, B = a->B;
3553: MatScalar *aa;
3554: Mat_SeqAIJ *aij;
3555: PetscInt *garray = a->garray, *colsub, Ncols;
3556: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3557: IS iscol_sub, iscmap;
3558: const PetscInt *is_idx, *cmap;
3559: PetscBool allcolumns = PETSC_FALSE;
3560: MPI_Comm comm;
3562: PetscFunctionBegin;
3563: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3564: if (call == MAT_REUSE_MATRIX) {
3565: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3566: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3567: PetscCall(ISGetLocalSize(iscol_sub, &count));
3569: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3570: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3572: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3573: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3575: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3577: } else { /* call == MAT_INITIAL_MATRIX) */
3578: PetscBool flg;
3580: PetscCall(ISGetLocalSize(iscol, &n));
3581: PetscCall(ISGetSize(iscol, &Ncols));
3583: /* (1) iscol -> nonscalable iscol_local */
3584: /* Check for special case: each processor gets entire matrix columns */
3585: PetscCall(ISIdentity(iscol_local, &flg));
3586: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3587: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3588: if (allcolumns) {
3589: iscol_sub = iscol_local;
3590: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3591: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3593: } else {
3594: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3595: PetscInt *idx, *cmap1, k;
3596: PetscCall(PetscMalloc1(Ncols, &idx));
3597: PetscCall(PetscMalloc1(Ncols, &cmap1));
3598: PetscCall(ISGetIndices(iscol_local, &is_idx));
3599: count = 0;
3600: k = 0;
3601: for (i = 0; i < Ncols; i++) {
3602: j = is_idx[i];
3603: if (j >= cstart && j < cend) {
3604: /* diagonal part of mat */
3605: idx[count] = j;
3606: cmap1[count++] = i; /* column index in submat */
3607: } else if (Bn) {
3608: /* off-diagonal part of mat */
3609: if (j == garray[k]) {
3610: idx[count] = j;
3611: cmap1[count++] = i; /* column index in submat */
3612: } else if (j > garray[k]) {
3613: while (j > garray[k] && k < Bn - 1) k++;
3614: if (j == garray[k]) {
3615: idx[count] = j;
3616: cmap1[count++] = i; /* column index in submat */
3617: }
3618: }
3619: }
3620: }
3621: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3623: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3624: PetscCall(ISGetBlockSize(iscol, &cbs));
3625: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3627: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3628: }
3630: /* (3) Create sequential Msub */
3631: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3632: }
3634: PetscCall(ISGetLocalSize(iscol_sub, &count));
3635: aij = (Mat_SeqAIJ *)Msub->data;
3636: ii = aij->i;
3637: PetscCall(ISGetIndices(iscmap, &cmap));
3639: /*
3640: m - number of local rows
3641: Ncols - number of columns (same on all processors)
3642: rstart - first row in new global matrix generated
3643: */
3644: PetscCall(MatGetSize(Msub, &m, NULL));
3646: if (call == MAT_INITIAL_MATRIX) {
3647: /* (4) Create parallel newmat */
3648: PetscMPIInt rank, size;
3649: PetscInt csize;
3651: PetscCallMPI(MPI_Comm_size(comm, &size));
3652: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3654: /*
3655: Determine the number of non-zeros in the diagonal and off-diagonal
3656: portions of the matrix in order to do correct preallocation
3657: */
3659: /* first get start and end of "diagonal" columns */
3660: PetscCall(ISGetLocalSize(iscol, &csize));
3661: if (csize == PETSC_DECIDE) {
3662: PetscCall(ISGetSize(isrow, &mglobal));
3663: if (mglobal == Ncols) { /* square matrix */
3664: nlocal = m;
3665: } else {
3666: nlocal = Ncols / size + ((Ncols % size) > rank);
3667: }
3668: } else {
3669: nlocal = csize;
3670: }
3671: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3672: rstart = rend - nlocal;
3673: 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);
3675: /* next, compute all the lengths */
3676: jj = aij->j;
3677: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3678: olens = dlens + m;
3679: for (i = 0; i < m; i++) {
3680: jend = ii[i + 1] - ii[i];
3681: olen = 0;
3682: dlen = 0;
3683: for (j = 0; j < jend; j++) {
3684: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3685: else dlen++;
3686: jj++;
3687: }
3688: olens[i] = olen;
3689: dlens[i] = dlen;
3690: }
3692: PetscCall(ISGetBlockSize(isrow, &bs));
3693: PetscCall(ISGetBlockSize(iscol, &cbs));
3695: PetscCall(MatCreate(comm, &M));
3696: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3697: PetscCall(MatSetBlockSizes(M, bs, cbs));
3698: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3699: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3700: PetscCall(PetscFree(dlens));
3702: } else { /* call == MAT_REUSE_MATRIX */
3703: M = *newmat;
3704: PetscCall(MatGetLocalSize(M, &i, NULL));
3705: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3706: PetscCall(MatZeroEntries(M));
3707: /*
3708: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3709: rather than the slower MatSetValues().
3710: */
3711: M->was_assembled = PETSC_TRUE;
3712: M->assembled = PETSC_FALSE;
3713: }
3715: /* (5) Set values of Msub to *newmat */
3716: PetscCall(PetscMalloc1(count, &colsub));
3717: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3719: jj = aij->j;
3720: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3721: for (i = 0; i < m; i++) {
3722: row = rstart + i;
3723: nz = ii[i + 1] - ii[i];
3724: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3725: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3726: jj += nz;
3727: aa += nz;
3728: }
3729: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3730: PetscCall(ISRestoreIndices(iscmap, &cmap));
3732: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3733: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3735: PetscCall(PetscFree(colsub));
3737: /* save Msub, iscol_sub and iscmap used in processor for next request */
3738: if (call == MAT_INITIAL_MATRIX) {
3739: *newmat = M;
3740: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3741: PetscCall(MatDestroy(&Msub));
3743: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3744: PetscCall(ISDestroy(&iscol_sub));
3746: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3747: PetscCall(ISDestroy(&iscmap));
3749: if (iscol_local) {
3750: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3751: PetscCall(ISDestroy(&iscol_local));
3752: }
3753: }
3754: PetscFunctionReturn(PETSC_SUCCESS);
3755: }
3757: /*
3758: Not great since it makes two copies of the submatrix, first an SeqAIJ
3759: in local and then by concatenating the local matrices the end result.
3760: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3762: This requires a sequential iscol with all indices.
3763: */
3764: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3765: {
3766: PetscMPIInt rank, size;
3767: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3768: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3769: Mat M, Mreuse;
3770: MatScalar *aa, *vwork;
3771: MPI_Comm comm;
3772: Mat_SeqAIJ *aij;
3773: PetscBool colflag, allcolumns = PETSC_FALSE;
3775: PetscFunctionBegin;
3776: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3777: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3778: PetscCallMPI(MPI_Comm_size(comm, &size));
3780: /* Check for special case: each processor gets entire matrix columns */
3781: PetscCall(ISIdentity(iscol, &colflag));
3782: PetscCall(ISGetLocalSize(iscol, &n));
3783: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3784: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3786: if (call == MAT_REUSE_MATRIX) {
3787: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3788: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3789: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3790: } else {
3791: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3792: }
3794: /*
3795: m - number of local rows
3796: n - number of columns (same on all processors)
3797: rstart - first row in new global matrix generated
3798: */
3799: PetscCall(MatGetSize(Mreuse, &m, &n));
3800: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3801: if (call == MAT_INITIAL_MATRIX) {
3802: aij = (Mat_SeqAIJ *)Mreuse->data;
3803: ii = aij->i;
3804: jj = aij->j;
3806: /*
3807: Determine the number of non-zeros in the diagonal and off-diagonal
3808: portions of the matrix in order to do correct preallocation
3809: */
3811: /* first get start and end of "diagonal" columns */
3812: if (csize == PETSC_DECIDE) {
3813: PetscCall(ISGetSize(isrow, &mglobal));
3814: if (mglobal == n) { /* square matrix */
3815: nlocal = m;
3816: } else {
3817: nlocal = n / size + ((n % size) > rank);
3818: }
3819: } else {
3820: nlocal = csize;
3821: }
3822: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3823: rstart = rend - nlocal;
3824: 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);
3826: /* next, compute all the lengths */
3827: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3828: olens = dlens + m;
3829: for (i = 0; i < m; i++) {
3830: jend = ii[i + 1] - ii[i];
3831: olen = 0;
3832: dlen = 0;
3833: for (j = 0; j < jend; j++) {
3834: if (*jj < rstart || *jj >= rend) olen++;
3835: else dlen++;
3836: jj++;
3837: }
3838: olens[i] = olen;
3839: dlens[i] = dlen;
3840: }
3841: PetscCall(MatCreate(comm, &M));
3842: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3843: PetscCall(MatSetBlockSizes(M, bs, cbs));
3844: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3845: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3846: PetscCall(PetscFree(dlens));
3847: } else {
3848: PetscInt ml, nl;
3850: M = *newmat;
3851: PetscCall(MatGetLocalSize(M, &ml, &nl));
3852: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3853: PetscCall(MatZeroEntries(M));
3854: /*
3855: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3856: rather than the slower MatSetValues().
3857: */
3858: M->was_assembled = PETSC_TRUE;
3859: M->assembled = PETSC_FALSE;
3860: }
3861: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3862: aij = (Mat_SeqAIJ *)Mreuse->data;
3863: ii = aij->i;
3864: jj = aij->j;
3866: /* trigger copy to CPU if needed */
3867: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3868: for (i = 0; i < m; i++) {
3869: row = rstart + i;
3870: nz = ii[i + 1] - ii[i];
3871: cwork = jj;
3872: jj = PetscSafePointerPlusOffset(jj, nz);
3873: vwork = aa;
3874: aa = PetscSafePointerPlusOffset(aa, nz);
3875: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3876: }
3877: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3879: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3880: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3881: *newmat = M;
3883: /* save submatrix used in processor for next request */
3884: if (call == MAT_INITIAL_MATRIX) {
3885: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3886: PetscCall(MatDestroy(&Mreuse));
3887: }
3888: PetscFunctionReturn(PETSC_SUCCESS);
3889: }
3891: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3892: {
3893: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3894: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3895: const PetscInt *JJ;
3896: PetscBool nooffprocentries;
3897: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3899: PetscFunctionBegin;
3900: PetscCall(PetscLayoutSetUp(B->rmap));
3901: PetscCall(PetscLayoutSetUp(B->cmap));
3902: m = B->rmap->n;
3903: cstart = B->cmap->rstart;
3904: cend = B->cmap->rend;
3905: rstart = B->rmap->rstart;
3906: irstart = Ii[0];
3908: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3910: if (PetscDefined(USE_DEBUG)) {
3911: for (i = 0; i < m; i++) {
3912: nnz = Ii[i + 1] - Ii[i];
3913: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3914: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3915: 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]);
3916: 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);
3917: }
3918: }
3920: for (i = 0; i < m; i++) {
3921: nnz = Ii[i + 1] - Ii[i];
3922: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3923: nnz_max = PetscMax(nnz_max, nnz);
3924: d = 0;
3925: for (j = 0; j < nnz; j++) {
3926: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3927: }
3928: d_nnz[i] = d;
3929: o_nnz[i] = nnz - d;
3930: }
3931: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3932: PetscCall(PetscFree2(d_nnz, o_nnz));
3934: for (i = 0; i < m; i++) {
3935: ii = i + rstart;
3936: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3937: }
3938: nooffprocentries = B->nooffprocentries;
3939: B->nooffprocentries = PETSC_TRUE;
3940: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3941: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3942: B->nooffprocentries = nooffprocentries;
3944: /* count number of entries below block diagonal */
3945: PetscCall(PetscFree(Aij->ld));
3946: PetscCall(PetscCalloc1(m, &ld));
3947: Aij->ld = ld;
3948: for (i = 0; i < m; i++) {
3949: nnz = Ii[i + 1] - Ii[i];
3950: j = 0;
3951: while (j < nnz && J[j] < cstart) j++;
3952: ld[i] = j;
3953: if (J) J += nnz;
3954: }
3956: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3957: PetscFunctionReturn(PETSC_SUCCESS);
3958: }
3960: /*@
3961: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3962: (the default parallel PETSc format).
3964: Collective
3966: Input Parameters:
3967: + B - the matrix
3968: . i - the indices into `j` for the start of each local row (indices start with zero)
3969: . j - the column indices for each local row (indices start with zero)
3970: - v - optional values in the matrix
3972: Level: developer
3974: Notes:
3975: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3976: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3977: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3979: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3981: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3983: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3985: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3986: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3988: The format which is used for the sparse matrix input, is equivalent to a
3989: row-major ordering.. i.e for the following matrix, the input data expected is
3990: as shown
3991: .vb
3992: 1 0 0
3993: 2 0 3 P0
3994: -------
3995: 4 5 6 P1
3997: Process0 [P0] rows_owned=[0,1]
3998: i = {0,1,3} [size = nrow+1 = 2+1]
3999: j = {0,0,2} [size = 3]
4000: v = {1,2,3} [size = 3]
4002: Process1 [P1] rows_owned=[2]
4003: i = {0,3} [size = nrow+1 = 1+1]
4004: j = {0,1,2} [size = 3]
4005: v = {4,5,6} [size = 3]
4006: .ve
4008: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4009: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4010: @*/
4011: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4012: {
4013: PetscFunctionBegin;
4014: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4015: PetscFunctionReturn(PETSC_SUCCESS);
4016: }
4018: /*@
4019: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4020: (the default parallel PETSc format). For good matrix assembly performance
4021: the user should preallocate the matrix storage by setting the parameters
4022: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4024: Collective
4026: Input Parameters:
4027: + B - the matrix
4028: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4029: (same value is used for all local rows)
4030: . d_nnz - array containing the number of nonzeros in the various rows of the
4031: DIAGONAL portion of the local submatrix (possibly different for each row)
4032: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4033: The size of this array is equal to the number of local rows, i.e 'm'.
4034: For matrices that will be factored, you must leave room for (and set)
4035: the diagonal entry even if it is zero.
4036: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4037: submatrix (same value is used for all local rows).
4038: - o_nnz - array containing the number of nonzeros in the various rows of the
4039: OFF-DIAGONAL portion of the local submatrix (possibly different for
4040: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4041: structure. The size of this array is equal to the number
4042: of local rows, i.e 'm'.
4044: Example Usage:
4045: Consider the following 8x8 matrix with 34 non-zero values, that is
4046: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4047: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4048: as follows
4050: .vb
4051: 1 2 0 | 0 3 0 | 0 4
4052: Proc0 0 5 6 | 7 0 0 | 8 0
4053: 9 0 10 | 11 0 0 | 12 0
4054: -------------------------------------
4055: 13 0 14 | 15 16 17 | 0 0
4056: Proc1 0 18 0 | 19 20 21 | 0 0
4057: 0 0 0 | 22 23 0 | 24 0
4058: -------------------------------------
4059: Proc2 25 26 27 | 0 0 28 | 29 0
4060: 30 0 0 | 31 32 33 | 0 34
4061: .ve
4063: This can be represented as a collection of submatrices as
4064: .vb
4065: A B C
4066: D E F
4067: G H I
4068: .ve
4070: Where the submatrices A,B,C are owned by proc0, D,E,F are
4071: owned by proc1, G,H,I are owned by proc2.
4073: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4074: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4075: The 'M','N' parameters are 8,8, and have the same values on all procs.
4077: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4078: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4079: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4080: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4081: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4082: matrix, ans [DF] as another `MATSEQAIJ` matrix.
4084: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4085: allocated for every row of the local diagonal submatrix, and `o_nz`
4086: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4087: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4088: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4089: In this case, the values of `d_nz`, `o_nz` are
4090: .vb
4091: proc0 dnz = 2, o_nz = 2
4092: proc1 dnz = 3, o_nz = 2
4093: proc2 dnz = 1, o_nz = 4
4094: .ve
4095: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4096: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4097: for proc3. i.e we are using 12+15+10=37 storage locations to store
4098: 34 values.
4100: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4101: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4102: In the above case the values for `d_nnz`, `o_nnz` are
4103: .vb
4104: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4105: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4106: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4107: .ve
4108: Here the space allocated is sum of all the above values i.e 34, and
4109: hence pre-allocation is perfect.
4111: Level: intermediate
4113: Notes:
4114: If the *_nnz parameter is given then the *_nz parameter is ignored
4116: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4117: storage. The stored row and column indices begin with zero.
4118: See [Sparse Matrices](sec_matsparse) for details.
4120: The parallel matrix is partitioned such that the first m0 rows belong to
4121: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4122: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4124: The DIAGONAL portion of the local submatrix of a processor can be defined
4125: as the submatrix which is obtained by extraction the part corresponding to
4126: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4127: first row that belongs to the processor, r2 is the last row belonging to
4128: the this processor, and c1-c2 is range of indices of the local part of a
4129: vector suitable for applying the matrix to. This is an mxn matrix. In the
4130: common case of a square matrix, the row and column ranges are the same and
4131: the DIAGONAL part is also square. The remaining portion of the local
4132: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4134: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4136: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4137: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4138: You can also run with the option `-info` and look for messages with the string
4139: malloc in them to see if additional memory allocation was needed.
4141: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4142: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4143: @*/
4144: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4145: {
4146: PetscFunctionBegin;
4149: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4150: PetscFunctionReturn(PETSC_SUCCESS);
4151: }
4153: /*@
4154: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4155: CSR format for the local rows.
4157: Collective
4159: Input Parameters:
4160: + comm - MPI communicator
4161: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4162: . n - This value should be the same as the local size used in creating the
4163: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4164: calculated if `N` is given) For square matrices n is almost always `m`.
4165: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4166: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4167: . 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
4168: . j - global column indices
4169: - a - optional matrix values
4171: Output Parameter:
4172: . mat - the matrix
4174: Level: intermediate
4176: Notes:
4177: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4178: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4179: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4181: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4183: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4185: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4186: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4188: The format which is used for the sparse matrix input, is equivalent to a
4189: row-major ordering, i.e., for the following matrix, the input data expected is
4190: as shown
4191: .vb
4192: 1 0 0
4193: 2 0 3 P0
4194: -------
4195: 4 5 6 P1
4197: Process0 [P0] rows_owned=[0,1]
4198: i = {0,1,3} [size = nrow+1 = 2+1]
4199: j = {0,0,2} [size = 3]
4200: v = {1,2,3} [size = 3]
4202: Process1 [P1] rows_owned=[2]
4203: i = {0,3} [size = nrow+1 = 1+1]
4204: j = {0,1,2} [size = 3]
4205: v = {4,5,6} [size = 3]
4206: .ve
4208: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4209: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4210: @*/
4211: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4212: {
4213: PetscFunctionBegin;
4214: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4215: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4216: PetscCall(MatCreate(comm, mat));
4217: PetscCall(MatSetSizes(*mat, m, n, M, N));
4218: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4219: PetscCall(MatSetType(*mat, MATMPIAIJ));
4220: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4221: PetscFunctionReturn(PETSC_SUCCESS);
4222: }
4224: /*@
4225: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4226: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4227: from `MatCreateMPIAIJWithArrays()`
4229: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4231: Collective
4233: Input Parameters:
4234: + mat - the matrix
4235: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4236: . n - This value should be the same as the local size used in creating the
4237: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4238: calculated if N is given) For square matrices n is almost always m.
4239: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4240: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4241: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4242: . J - column indices
4243: - v - matrix values
4245: Level: deprecated
4247: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4248: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4249: @*/
4250: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4251: {
4252: PetscInt nnz, i;
4253: PetscBool nooffprocentries;
4254: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4255: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4256: PetscScalar *ad, *ao;
4257: PetscInt ldi, Iii, md;
4258: const PetscInt *Adi = Ad->i;
4259: PetscInt *ld = Aij->ld;
4261: PetscFunctionBegin;
4262: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4263: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4264: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4265: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4267: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4268: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4270: for (i = 0; i < m; i++) {
4271: if (PetscDefined(USE_DEBUG)) {
4272: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4273: 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);
4274: 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);
4275: }
4276: }
4277: nnz = Ii[i + 1] - Ii[i];
4278: Iii = Ii[i];
4279: ldi = ld[i];
4280: md = Adi[i + 1] - Adi[i];
4281: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4282: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4283: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4284: ad += md;
4285: ao += nnz - md;
4286: }
4287: nooffprocentries = mat->nooffprocentries;
4288: mat->nooffprocentries = PETSC_TRUE;
4289: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4290: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4291: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4292: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4293: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4294: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4295: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4296: mat->nooffprocentries = nooffprocentries;
4297: PetscFunctionReturn(PETSC_SUCCESS);
4298: }
4300: /*@
4301: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4303: Collective
4305: Input Parameters:
4306: + mat - the matrix
4307: - v - matrix values, stored by row
4309: Level: intermediate
4311: Notes:
4312: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4314: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4316: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4317: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4318: @*/
4319: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4320: {
4321: PetscInt nnz, i, m;
4322: PetscBool nooffprocentries;
4323: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4324: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4325: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4326: PetscScalar *ad, *ao;
4327: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4328: PetscInt ldi, Iii, md;
4329: PetscInt *ld = Aij->ld;
4331: PetscFunctionBegin;
4332: m = mat->rmap->n;
4334: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4335: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4336: Iii = 0;
4337: for (i = 0; i < m; i++) {
4338: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4339: ldi = ld[i];
4340: md = Adi[i + 1] - Adi[i];
4341: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4342: ad += md;
4343: if (ao) {
4344: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4345: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4346: ao += nnz - md;
4347: }
4348: Iii += nnz;
4349: }
4350: nooffprocentries = mat->nooffprocentries;
4351: mat->nooffprocentries = PETSC_TRUE;
4352: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4353: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4354: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4355: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4356: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4357: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4358: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4359: mat->nooffprocentries = nooffprocentries;
4360: PetscFunctionReturn(PETSC_SUCCESS);
4361: }
4363: /*@
4364: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4365: (the default parallel PETSc format). For good matrix assembly performance
4366: the user should preallocate the matrix storage by setting the parameters
4367: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4369: Collective
4371: Input Parameters:
4372: + comm - MPI communicator
4373: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4374: This value should be the same as the local size used in creating the
4375: y vector for the matrix-vector product y = Ax.
4376: . n - This value should be the same as the local size used in creating the
4377: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4378: calculated if N is given) For square matrices n is almost always m.
4379: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4380: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4381: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4382: (same value is used for all local rows)
4383: . d_nnz - array containing the number of nonzeros in the various rows of the
4384: DIAGONAL portion of the local submatrix (possibly different for each row)
4385: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4386: The size of this array is equal to the number of local rows, i.e 'm'.
4387: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4388: submatrix (same value is used for all local rows).
4389: - o_nnz - array containing the number of nonzeros in the various rows of the
4390: OFF-DIAGONAL portion of the local submatrix (possibly different for
4391: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4392: structure. The size of this array is equal to the number
4393: of local rows, i.e 'm'.
4395: Output Parameter:
4396: . A - the matrix
4398: Options Database Keys:
4399: + -mat_no_inode - Do not use inodes
4400: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4401: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4402: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4403: 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.
4405: Level: intermediate
4407: Notes:
4408: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4409: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4410: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4412: If the *_nnz parameter is given then the *_nz parameter is ignored
4414: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4415: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4416: storage requirements for this matrix.
4418: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4419: processor than it must be used on all processors that share the object for
4420: that argument.
4422: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4423: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4425: The user MUST specify either the local or global matrix dimensions
4426: (possibly both).
4428: The parallel matrix is partitioned across processors such that the
4429: first `m0` rows belong to process 0, the next `m1` rows belong to
4430: process 1, the next `m2` rows belong to process 2, etc., where
4431: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4432: values corresponding to [m x N] submatrix.
4434: The columns are logically partitioned with the n0 columns belonging
4435: to 0th partition, the next n1 columns belonging to the next
4436: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4438: The DIAGONAL portion of the local submatrix on any given processor
4439: is the submatrix corresponding to the rows and columns m,n
4440: corresponding to the given processor. i.e diagonal matrix on
4441: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4442: etc. The remaining portion of the local submatrix [m x (N-n)]
4443: constitute the OFF-DIAGONAL portion. The example below better
4444: illustrates this concept.
4446: For a square global matrix we define each processor's diagonal portion
4447: to be its local rows and the corresponding columns (a square submatrix);
4448: each processor's off-diagonal portion encompasses the remainder of the
4449: local matrix (a rectangular submatrix).
4451: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4453: When calling this routine with a single process communicator, a matrix of
4454: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4455: type of communicator, use the construction mechanism
4456: .vb
4457: MatCreate(..., &A);
4458: MatSetType(A, MATMPIAIJ);
4459: MatSetSizes(A, m, n, M, N);
4460: MatMPIAIJSetPreallocation(A, ...);
4461: .ve
4463: By default, this format uses inodes (identical nodes) when possible.
4464: We search for consecutive rows with the same nonzero structure, thereby
4465: reusing matrix information to achieve increased efficiency.
4467: Example Usage:
4468: Consider the following 8x8 matrix with 34 non-zero values, that is
4469: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4470: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4471: as follows
4473: .vb
4474: 1 2 0 | 0 3 0 | 0 4
4475: Proc0 0 5 6 | 7 0 0 | 8 0
4476: 9 0 10 | 11 0 0 | 12 0
4477: -------------------------------------
4478: 13 0 14 | 15 16 17 | 0 0
4479: Proc1 0 18 0 | 19 20 21 | 0 0
4480: 0 0 0 | 22 23 0 | 24 0
4481: -------------------------------------
4482: Proc2 25 26 27 | 0 0 28 | 29 0
4483: 30 0 0 | 31 32 33 | 0 34
4484: .ve
4486: This can be represented as a collection of submatrices as
4488: .vb
4489: A B C
4490: D E F
4491: G H I
4492: .ve
4494: Where the submatrices A,B,C are owned by proc0, D,E,F are
4495: owned by proc1, G,H,I are owned by proc2.
4497: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4498: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4499: The 'M','N' parameters are 8,8, and have the same values on all procs.
4501: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4502: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4503: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4504: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4505: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4506: matrix, ans [DF] as another SeqAIJ matrix.
4508: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4509: allocated for every row of the local diagonal submatrix, and `o_nz`
4510: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4511: One way to choose `d_nz` and `o_nz` is to use the max nonzerors per local
4512: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4513: In this case, the values of `d_nz`,`o_nz` are
4514: .vb
4515: proc0 dnz = 2, o_nz = 2
4516: proc1 dnz = 3, o_nz = 2
4517: proc2 dnz = 1, o_nz = 4
4518: .ve
4519: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4520: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4521: for proc3. i.e we are using 12+15+10=37 storage locations to store
4522: 34 values.
4524: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4525: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4526: In the above case the values for d_nnz,o_nnz are
4527: .vb
4528: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4529: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4530: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4531: .ve
4532: Here the space allocated is sum of all the above values i.e 34, and
4533: hence pre-allocation is perfect.
4535: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4536: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4537: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4538: @*/
4539: 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)
4540: {
4541: PetscMPIInt size;
4543: PetscFunctionBegin;
4544: PetscCall(MatCreate(comm, A));
4545: PetscCall(MatSetSizes(*A, m, n, M, N));
4546: PetscCallMPI(MPI_Comm_size(comm, &size));
4547: if (size > 1) {
4548: PetscCall(MatSetType(*A, MATMPIAIJ));
4549: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4550: } else {
4551: PetscCall(MatSetType(*A, MATSEQAIJ));
4552: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4553: }
4554: PetscFunctionReturn(PETSC_SUCCESS);
4555: }
4557: /*MC
4558: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4560: Synopsis:
4561: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4563: Not Collective
4565: Input Parameter:
4566: . A - the `MATMPIAIJ` matrix
4568: Output Parameters:
4569: + Ad - the diagonal portion of the matrix
4570: . Ao - the off-diagonal portion of the matrix
4571: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4572: - ierr - error code
4574: Level: advanced
4576: Note:
4577: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4579: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4580: M*/
4582: /*MC
4583: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4585: Synopsis:
4586: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4588: Not Collective
4590: Input Parameters:
4591: + A - the `MATMPIAIJ` matrix
4592: . Ad - the diagonal portion of the matrix
4593: . Ao - the off-diagonal portion of the matrix
4594: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4595: - ierr - error code
4597: Level: advanced
4599: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4600: M*/
4602: /*@C
4603: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4605: Not Collective
4607: Input Parameter:
4608: . A - The `MATMPIAIJ` matrix
4610: Output Parameters:
4611: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4612: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4613: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4615: Level: intermediate
4617: Note:
4618: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4619: 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
4620: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4621: local column numbers to global column numbers in the original matrix.
4623: Fortran Notes:
4624: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4626: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4627: @*/
4628: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4629: {
4630: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4631: PetscBool flg;
4633: PetscFunctionBegin;
4634: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4635: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4636: if (Ad) *Ad = a->A;
4637: if (Ao) *Ao = a->B;
4638: if (colmap) *colmap = a->garray;
4639: PetscFunctionReturn(PETSC_SUCCESS);
4640: }
4642: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4643: {
4644: PetscInt m, N, i, rstart, nnz, Ii;
4645: PetscInt *indx;
4646: PetscScalar *values;
4647: MatType rootType;
4649: PetscFunctionBegin;
4650: PetscCall(MatGetSize(inmat, &m, &N));
4651: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4652: PetscInt *dnz, *onz, sum, bs, cbs;
4654: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4655: /* Check sum(n) = N */
4656: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4657: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4659: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4660: rstart -= m;
4662: MatPreallocateBegin(comm, m, n, dnz, onz);
4663: for (i = 0; i < m; i++) {
4664: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4665: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4666: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4667: }
4669: PetscCall(MatCreate(comm, outmat));
4670: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4671: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4672: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4673: PetscCall(MatGetRootType_Private(inmat, &rootType));
4674: PetscCall(MatSetType(*outmat, rootType));
4675: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4676: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4677: MatPreallocateEnd(dnz, onz);
4678: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4679: }
4681: /* numeric phase */
4682: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4683: for (i = 0; i < m; i++) {
4684: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4685: Ii = i + rstart;
4686: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4687: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4688: }
4689: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4690: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4691: PetscFunctionReturn(PETSC_SUCCESS);
4692: }
4694: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4695: {
4696: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4698: PetscFunctionBegin;
4699: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4700: PetscCall(PetscFree(merge->id_r));
4701: PetscCall(PetscFree(merge->len_s));
4702: PetscCall(PetscFree(merge->len_r));
4703: PetscCall(PetscFree(merge->bi));
4704: PetscCall(PetscFree(merge->bj));
4705: PetscCall(PetscFree(merge->buf_ri[0]));
4706: PetscCall(PetscFree(merge->buf_ri));
4707: PetscCall(PetscFree(merge->buf_rj[0]));
4708: PetscCall(PetscFree(merge->buf_rj));
4709: PetscCall(PetscFree(merge->coi));
4710: PetscCall(PetscFree(merge->coj));
4711: PetscCall(PetscFree(merge->owners_co));
4712: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4713: PetscCall(PetscFree(merge));
4714: PetscFunctionReturn(PETSC_SUCCESS);
4715: }
4717: #include <../src/mat/utils/freespace.h>
4718: #include <petscbt.h>
4720: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4721: {
4722: MPI_Comm comm;
4723: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4724: PetscMPIInt size, rank, taga, *len_s;
4725: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4726: PetscMPIInt proc, k;
4727: PetscInt **buf_ri, **buf_rj;
4728: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4729: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4730: MPI_Request *s_waits, *r_waits;
4731: MPI_Status *status;
4732: const MatScalar *aa, *a_a;
4733: MatScalar **abuf_r, *ba_i;
4734: Mat_Merge_SeqsToMPI *merge;
4735: PetscContainer container;
4737: PetscFunctionBegin;
4738: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4739: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4741: PetscCallMPI(MPI_Comm_size(comm, &size));
4742: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4744: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4745: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4746: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4747: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4748: aa = a_a;
4750: bi = merge->bi;
4751: bj = merge->bj;
4752: buf_ri = merge->buf_ri;
4753: buf_rj = merge->buf_rj;
4755: PetscCall(PetscMalloc1(size, &status));
4756: owners = merge->rowmap->range;
4757: len_s = merge->len_s;
4759: /* send and recv matrix values */
4760: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4761: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4763: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4764: for (proc = 0, k = 0; proc < size; proc++) {
4765: if (!len_s[proc]) continue;
4766: i = owners[proc];
4767: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4768: k++;
4769: }
4771: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4772: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4773: PetscCall(PetscFree(status));
4775: PetscCall(PetscFree(s_waits));
4776: PetscCall(PetscFree(r_waits));
4778: /* insert mat values of mpimat */
4779: PetscCall(PetscMalloc1(N, &ba_i));
4780: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4782: for (k = 0; k < merge->nrecv; k++) {
4783: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4784: nrows = *buf_ri_k[k];
4785: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4786: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4787: }
4789: /* set values of ba */
4790: m = merge->rowmap->n;
4791: for (i = 0; i < m; i++) {
4792: arow = owners[rank] + i;
4793: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4794: bnzi = bi[i + 1] - bi[i];
4795: PetscCall(PetscArrayzero(ba_i, bnzi));
4797: /* add local non-zero vals of this proc's seqmat into ba */
4798: anzi = ai[arow + 1] - ai[arow];
4799: aj = a->j + ai[arow];
4800: aa = a_a + ai[arow];
4801: nextaj = 0;
4802: for (j = 0; nextaj < anzi; j++) {
4803: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4804: ba_i[j] += aa[nextaj++];
4805: }
4806: }
4808: /* add received vals into ba */
4809: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4810: /* i-th row */
4811: if (i == *nextrow[k]) {
4812: anzi = *(nextai[k] + 1) - *nextai[k];
4813: aj = buf_rj[k] + *nextai[k];
4814: aa = abuf_r[k] + *nextai[k];
4815: nextaj = 0;
4816: for (j = 0; nextaj < anzi; j++) {
4817: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4818: ba_i[j] += aa[nextaj++];
4819: }
4820: }
4821: nextrow[k]++;
4822: nextai[k]++;
4823: }
4824: }
4825: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4826: }
4827: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4828: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4829: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4831: PetscCall(PetscFree(abuf_r[0]));
4832: PetscCall(PetscFree(abuf_r));
4833: PetscCall(PetscFree(ba_i));
4834: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4835: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4836: PetscFunctionReturn(PETSC_SUCCESS);
4837: }
4839: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4840: {
4841: Mat B_mpi;
4842: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4843: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4844: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4845: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4846: PetscInt len, *dnz, *onz, bs, cbs;
4847: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4848: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4849: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4850: MPI_Status *status;
4851: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4852: PetscBT lnkbt;
4853: Mat_Merge_SeqsToMPI *merge;
4854: PetscContainer container;
4856: PetscFunctionBegin;
4857: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4859: /* make sure it is a PETSc comm */
4860: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4861: PetscCallMPI(MPI_Comm_size(comm, &size));
4862: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4864: PetscCall(PetscNew(&merge));
4865: PetscCall(PetscMalloc1(size, &status));
4867: /* determine row ownership */
4868: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4869: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4870: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4871: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4872: PetscCall(PetscLayoutSetUp(merge->rowmap));
4873: PetscCall(PetscMalloc1(size, &len_si));
4874: PetscCall(PetscMalloc1(size, &merge->len_s));
4876: m = merge->rowmap->n;
4877: owners = merge->rowmap->range;
4879: /* determine the number of messages to send, their lengths */
4880: len_s = merge->len_s;
4882: len = 0; /* length of buf_si[] */
4883: merge->nsend = 0;
4884: for (PetscMPIInt proc = 0; proc < size; proc++) {
4885: len_si[proc] = 0;
4886: if (proc == rank) {
4887: len_s[proc] = 0;
4888: } else {
4889: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4890: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4891: }
4892: if (len_s[proc]) {
4893: merge->nsend++;
4894: nrows = 0;
4895: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4896: if (ai[i + 1] > ai[i]) nrows++;
4897: }
4898: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4899: len += len_si[proc];
4900: }
4901: }
4903: /* determine the number and length of messages to receive for ij-structure */
4904: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4905: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4907: /* post the Irecv of j-structure */
4908: PetscCall(PetscCommGetNewTag(comm, &tagj));
4909: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4911: /* post the Isend of j-structure */
4912: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4914: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4915: if (!len_s[proc]) continue;
4916: i = owners[proc];
4917: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4918: k++;
4919: }
4921: /* receives and sends of j-structure are complete */
4922: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4923: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4925: /* send and recv i-structure */
4926: PetscCall(PetscCommGetNewTag(comm, &tagi));
4927: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4929: PetscCall(PetscMalloc1(len + 1, &buf_s));
4930: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4931: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4932: if (!len_s[proc]) continue;
4933: /* form outgoing message for i-structure:
4934: buf_si[0]: nrows to be sent
4935: [1:nrows]: row index (global)
4936: [nrows+1:2*nrows+1]: i-structure index
4937: */
4938: nrows = len_si[proc] / 2 - 1;
4939: buf_si_i = buf_si + nrows + 1;
4940: buf_si[0] = nrows;
4941: buf_si_i[0] = 0;
4942: nrows = 0;
4943: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4944: anzi = ai[i + 1] - ai[i];
4945: if (anzi) {
4946: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4947: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4948: nrows++;
4949: }
4950: }
4951: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4952: k++;
4953: buf_si += len_si[proc];
4954: }
4956: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4957: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4959: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4960: 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]));
4962: PetscCall(PetscFree(len_si));
4963: PetscCall(PetscFree(len_ri));
4964: PetscCall(PetscFree(rj_waits));
4965: PetscCall(PetscFree2(si_waits, sj_waits));
4966: PetscCall(PetscFree(ri_waits));
4967: PetscCall(PetscFree(buf_s));
4968: PetscCall(PetscFree(status));
4970: /* compute a local seq matrix in each processor */
4971: /* allocate bi array and free space for accumulating nonzero column info */
4972: PetscCall(PetscMalloc1(m + 1, &bi));
4973: bi[0] = 0;
4975: /* create and initialize a linked list */
4976: nlnk = N + 1;
4977: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4979: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4980: len = ai[owners[rank + 1]] - ai[owners[rank]];
4981: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4983: current_space = free_space;
4985: /* determine symbolic info for each local row */
4986: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4988: for (k = 0; k < merge->nrecv; k++) {
4989: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4990: nrows = *buf_ri_k[k];
4991: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4992: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4993: }
4995: MatPreallocateBegin(comm, m, n, dnz, onz);
4996: len = 0;
4997: for (i = 0; i < m; i++) {
4998: bnzi = 0;
4999: /* add local non-zero cols of this proc's seqmat into lnk */
5000: arow = owners[rank] + i;
5001: anzi = ai[arow + 1] - ai[arow];
5002: aj = a->j + ai[arow];
5003: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5004: bnzi += nlnk;
5005: /* add received col data into lnk */
5006: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5007: if (i == *nextrow[k]) { /* i-th row */
5008: anzi = *(nextai[k] + 1) - *nextai[k];
5009: aj = buf_rj[k] + *nextai[k];
5010: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5011: bnzi += nlnk;
5012: nextrow[k]++;
5013: nextai[k]++;
5014: }
5015: }
5016: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5018: /* if free space is not available, make more free space */
5019: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5020: /* copy data into free space, then initialize lnk */
5021: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5022: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5024: current_space->array += bnzi;
5025: current_space->local_used += bnzi;
5026: current_space->local_remaining -= bnzi;
5028: bi[i + 1] = bi[i] + bnzi;
5029: }
5031: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5033: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5034: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5035: PetscCall(PetscLLDestroy(lnk, lnkbt));
5037: /* create symbolic parallel matrix B_mpi */
5038: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5039: PetscCall(MatCreate(comm, &B_mpi));
5040: if (n == PETSC_DECIDE) {
5041: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5042: } else {
5043: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5044: }
5045: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5046: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5047: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5048: MatPreallocateEnd(dnz, onz);
5049: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5051: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5052: B_mpi->assembled = PETSC_FALSE;
5053: merge->bi = bi;
5054: merge->bj = bj;
5055: merge->buf_ri = buf_ri;
5056: merge->buf_rj = buf_rj;
5057: merge->coi = NULL;
5058: merge->coj = NULL;
5059: merge->owners_co = NULL;
5061: PetscCall(PetscCommDestroy(&comm));
5063: /* attach the supporting struct to B_mpi for reuse */
5064: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5065: PetscCall(PetscContainerSetPointer(container, merge));
5066: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5067: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5068: PetscCall(PetscContainerDestroy(&container));
5069: *mpimat = B_mpi;
5071: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5072: PetscFunctionReturn(PETSC_SUCCESS);
5073: }
5075: /*@
5076: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5077: matrices from each processor
5079: Collective
5081: Input Parameters:
5082: + comm - the communicators the parallel matrix will live on
5083: . seqmat - the input sequential matrices
5084: . m - number of local rows (or `PETSC_DECIDE`)
5085: . n - number of local columns (or `PETSC_DECIDE`)
5086: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5088: Output Parameter:
5089: . mpimat - the parallel matrix generated
5091: Level: advanced
5093: Note:
5094: The dimensions of the sequential matrix in each processor MUST be the same.
5095: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5096: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5098: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5099: @*/
5100: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5101: {
5102: PetscMPIInt size;
5104: PetscFunctionBegin;
5105: PetscCallMPI(MPI_Comm_size(comm, &size));
5106: if (size == 1) {
5107: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5108: if (scall == MAT_INITIAL_MATRIX) {
5109: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5110: } else {
5111: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5112: }
5113: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5114: PetscFunctionReturn(PETSC_SUCCESS);
5115: }
5116: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5117: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5118: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5119: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5120: PetscFunctionReturn(PETSC_SUCCESS);
5121: }
5123: /*@
5124: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5126: Not Collective
5128: Input Parameter:
5129: . A - the matrix
5131: Output Parameter:
5132: . A_loc - the local sequential matrix generated
5134: Level: developer
5136: Notes:
5137: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5138: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5139: `n` is the global column count obtained with `MatGetSize()`
5141: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5143: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5145: Destroy the matrix with `MatDestroy()`
5147: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5148: @*/
5149: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5150: {
5151: PetscBool mpi;
5153: PetscFunctionBegin;
5154: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5155: if (mpi) {
5156: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5157: } else {
5158: *A_loc = A;
5159: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5160: }
5161: PetscFunctionReturn(PETSC_SUCCESS);
5162: }
5164: /*@
5165: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5167: Not Collective
5169: Input Parameters:
5170: + A - the matrix
5171: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5173: Output Parameter:
5174: . A_loc - the local sequential matrix generated
5176: Level: developer
5178: Notes:
5179: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5180: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5181: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5183: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5185: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5186: 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
5187: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5188: 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.
5190: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5191: @*/
5192: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5193: {
5194: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5195: Mat_SeqAIJ *mat, *a, *b;
5196: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5197: const PetscScalar *aa, *ba, *aav, *bav;
5198: PetscScalar *ca, *cam;
5199: PetscMPIInt size;
5200: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5201: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5202: PetscBool match;
5204: PetscFunctionBegin;
5205: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5206: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5207: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5208: if (size == 1) {
5209: if (scall == MAT_INITIAL_MATRIX) {
5210: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5211: *A_loc = mpimat->A;
5212: } else if (scall == MAT_REUSE_MATRIX) {
5213: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5214: }
5215: PetscFunctionReturn(PETSC_SUCCESS);
5216: }
5218: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5219: a = (Mat_SeqAIJ *)mpimat->A->data;
5220: b = (Mat_SeqAIJ *)mpimat->B->data;
5221: ai = a->i;
5222: aj = a->j;
5223: bi = b->i;
5224: bj = b->j;
5225: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5226: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5227: aa = aav;
5228: ba = bav;
5229: if (scall == MAT_INITIAL_MATRIX) {
5230: PetscCall(PetscMalloc1(1 + am, &ci));
5231: ci[0] = 0;
5232: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5233: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5234: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5235: k = 0;
5236: for (i = 0; i < am; i++) {
5237: ncols_o = bi[i + 1] - bi[i];
5238: ncols_d = ai[i + 1] - ai[i];
5239: /* off-diagonal portion of A */
5240: for (jo = 0; jo < ncols_o; jo++) {
5241: col = cmap[*bj];
5242: if (col >= cstart) break;
5243: cj[k] = col;
5244: bj++;
5245: ca[k++] = *ba++;
5246: }
5247: /* diagonal portion of A */
5248: for (j = 0; j < ncols_d; j++) {
5249: cj[k] = cstart + *aj++;
5250: ca[k++] = *aa++;
5251: }
5252: /* off-diagonal portion of A */
5253: for (j = jo; j < ncols_o; j++) {
5254: cj[k] = cmap[*bj++];
5255: ca[k++] = *ba++;
5256: }
5257: }
5258: /* put together the new matrix */
5259: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5260: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5261: /* Since these are PETSc arrays, change flags to free them as necessary. */
5262: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5263: mat->free_a = PETSC_TRUE;
5264: mat->free_ij = PETSC_TRUE;
5265: mat->nonew = 0;
5266: } else if (scall == MAT_REUSE_MATRIX) {
5267: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5268: ci = mat->i;
5269: cj = mat->j;
5270: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5271: for (i = 0; i < am; i++) {
5272: /* off-diagonal portion of A */
5273: ncols_o = bi[i + 1] - bi[i];
5274: for (jo = 0; jo < ncols_o; jo++) {
5275: col = cmap[*bj];
5276: if (col >= cstart) break;
5277: *cam++ = *ba++;
5278: bj++;
5279: }
5280: /* diagonal portion of A */
5281: ncols_d = ai[i + 1] - ai[i];
5282: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5283: /* off-diagonal portion of A */
5284: for (j = jo; j < ncols_o; j++) {
5285: *cam++ = *ba++;
5286: bj++;
5287: }
5288: }
5289: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5290: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5291: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5292: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5293: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5294: PetscFunctionReturn(PETSC_SUCCESS);
5295: }
5297: /*@
5298: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5299: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5301: Not Collective
5303: Input Parameters:
5304: + A - the matrix
5305: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5307: Output Parameters:
5308: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5309: - A_loc - the local sequential matrix generated
5311: Level: developer
5313: Note:
5314: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5315: part, then those associated with the off-diagonal part (in its local ordering)
5317: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5318: @*/
5319: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5320: {
5321: Mat Ao, Ad;
5322: const PetscInt *cmap;
5323: PetscMPIInt size;
5324: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5326: PetscFunctionBegin;
5327: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5328: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5329: if (size == 1) {
5330: if (scall == MAT_INITIAL_MATRIX) {
5331: PetscCall(PetscObjectReference((PetscObject)Ad));
5332: *A_loc = Ad;
5333: } else if (scall == MAT_REUSE_MATRIX) {
5334: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5335: }
5336: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5337: PetscFunctionReturn(PETSC_SUCCESS);
5338: }
5339: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5340: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5341: if (f) {
5342: PetscCall((*f)(A, scall, glob, A_loc));
5343: } else {
5344: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5345: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5346: Mat_SeqAIJ *c;
5347: PetscInt *ai = a->i, *aj = a->j;
5348: PetscInt *bi = b->i, *bj = b->j;
5349: PetscInt *ci, *cj;
5350: const PetscScalar *aa, *ba;
5351: PetscScalar *ca;
5352: PetscInt i, j, am, dn, on;
5354: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5355: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5356: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5357: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5358: if (scall == MAT_INITIAL_MATRIX) {
5359: PetscInt k;
5360: PetscCall(PetscMalloc1(1 + am, &ci));
5361: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5362: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5363: ci[0] = 0;
5364: for (i = 0, k = 0; i < am; i++) {
5365: const PetscInt ncols_o = bi[i + 1] - bi[i];
5366: const PetscInt ncols_d = ai[i + 1] - ai[i];
5367: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5368: /* diagonal portion of A */
5369: for (j = 0; j < ncols_d; j++, k++) {
5370: cj[k] = *aj++;
5371: ca[k] = *aa++;
5372: }
5373: /* off-diagonal portion of A */
5374: for (j = 0; j < ncols_o; j++, k++) {
5375: cj[k] = dn + *bj++;
5376: ca[k] = *ba++;
5377: }
5378: }
5379: /* put together the new matrix */
5380: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5381: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5382: /* Since these are PETSc arrays, change flags to free them as necessary. */
5383: c = (Mat_SeqAIJ *)(*A_loc)->data;
5384: c->free_a = PETSC_TRUE;
5385: c->free_ij = PETSC_TRUE;
5386: c->nonew = 0;
5387: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5388: } else if (scall == MAT_REUSE_MATRIX) {
5389: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5390: for (i = 0; i < am; i++) {
5391: const PetscInt ncols_d = ai[i + 1] - ai[i];
5392: const PetscInt ncols_o = bi[i + 1] - bi[i];
5393: /* diagonal portion of A */
5394: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5395: /* off-diagonal portion of A */
5396: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5397: }
5398: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5399: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5400: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5401: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5402: if (glob) {
5403: PetscInt cst, *gidx;
5405: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5406: PetscCall(PetscMalloc1(dn + on, &gidx));
5407: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5408: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5409: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5410: }
5411: }
5412: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5413: PetscFunctionReturn(PETSC_SUCCESS);
5414: }
5416: /*@C
5417: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5419: Not Collective
5421: Input Parameters:
5422: + A - the matrix
5423: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5424: . row - index set of rows to extract (or `NULL`)
5425: - col - index set of columns to extract (or `NULL`)
5427: Output Parameter:
5428: . A_loc - the local sequential matrix generated
5430: Level: developer
5432: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5433: @*/
5434: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5435: {
5436: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5437: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5438: IS isrowa, iscola;
5439: Mat *aloc;
5440: PetscBool match;
5442: PetscFunctionBegin;
5443: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5444: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5445: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5446: if (!row) {
5447: start = A->rmap->rstart;
5448: end = A->rmap->rend;
5449: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5450: } else {
5451: isrowa = *row;
5452: }
5453: if (!col) {
5454: start = A->cmap->rstart;
5455: cmap = a->garray;
5456: nzA = a->A->cmap->n;
5457: nzB = a->B->cmap->n;
5458: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5459: ncols = 0;
5460: for (i = 0; i < nzB; i++) {
5461: if (cmap[i] < start) idx[ncols++] = cmap[i];
5462: else break;
5463: }
5464: imark = i;
5465: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5466: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5467: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5468: } else {
5469: iscola = *col;
5470: }
5471: if (scall != MAT_INITIAL_MATRIX) {
5472: PetscCall(PetscMalloc1(1, &aloc));
5473: aloc[0] = *A_loc;
5474: }
5475: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5476: if (!col) { /* attach global id of condensed columns */
5477: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5478: }
5479: *A_loc = aloc[0];
5480: PetscCall(PetscFree(aloc));
5481: if (!row) PetscCall(ISDestroy(&isrowa));
5482: if (!col) PetscCall(ISDestroy(&iscola));
5483: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5484: PetscFunctionReturn(PETSC_SUCCESS);
5485: }
5487: /*
5488: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5489: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5490: * on a global size.
5491: * */
5492: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5493: {
5494: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5495: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5496: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5497: PetscMPIInt owner;
5498: PetscSFNode *iremote, *oiremote;
5499: const PetscInt *lrowindices;
5500: PetscSF sf, osf;
5501: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5502: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5503: MPI_Comm comm;
5504: ISLocalToGlobalMapping mapping;
5505: const PetscScalar *pd_a, *po_a;
5507: PetscFunctionBegin;
5508: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5509: /* plocalsize is the number of roots
5510: * nrows is the number of leaves
5511: * */
5512: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5513: PetscCall(ISGetLocalSize(rows, &nrows));
5514: PetscCall(PetscCalloc1(nrows, &iremote));
5515: PetscCall(ISGetIndices(rows, &lrowindices));
5516: for (i = 0; i < nrows; i++) {
5517: /* Find a remote index and an owner for a row
5518: * The row could be local or remote
5519: * */
5520: owner = 0;
5521: lidx = 0;
5522: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5523: iremote[i].index = lidx;
5524: iremote[i].rank = owner;
5525: }
5526: /* Create SF to communicate how many nonzero columns for each row */
5527: PetscCall(PetscSFCreate(comm, &sf));
5528: /* SF will figure out the number of nonzero columns for each row, and their
5529: * offsets
5530: * */
5531: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5532: PetscCall(PetscSFSetFromOptions(sf));
5533: PetscCall(PetscSFSetUp(sf));
5535: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5536: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5537: PetscCall(PetscCalloc1(nrows, &pnnz));
5538: roffsets[0] = 0;
5539: roffsets[1] = 0;
5540: for (i = 0; i < plocalsize; i++) {
5541: /* diagonal */
5542: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5543: /* off-diagonal */
5544: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5545: /* compute offsets so that we relative location for each row */
5546: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5547: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5548: }
5549: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5550: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5551: /* 'r' means root, and 'l' means leaf */
5552: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5553: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5554: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5555: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5556: PetscCall(PetscSFDestroy(&sf));
5557: PetscCall(PetscFree(roffsets));
5558: PetscCall(PetscFree(nrcols));
5559: dntotalcols = 0;
5560: ontotalcols = 0;
5561: ncol = 0;
5562: for (i = 0; i < nrows; i++) {
5563: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5564: ncol = PetscMax(pnnz[i], ncol);
5565: /* diagonal */
5566: dntotalcols += nlcols[i * 2 + 0];
5567: /* off-diagonal */
5568: ontotalcols += nlcols[i * 2 + 1];
5569: }
5570: /* We do not need to figure the right number of columns
5571: * since all the calculations will be done by going through the raw data
5572: * */
5573: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5574: PetscCall(MatSetUp(*P_oth));
5575: PetscCall(PetscFree(pnnz));
5576: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5577: /* diagonal */
5578: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5579: /* off-diagonal */
5580: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5581: /* diagonal */
5582: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5583: /* off-diagonal */
5584: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5585: dntotalcols = 0;
5586: ontotalcols = 0;
5587: ntotalcols = 0;
5588: for (i = 0; i < nrows; i++) {
5589: owner = 0;
5590: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5591: /* Set iremote for diag matrix */
5592: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5593: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5594: iremote[dntotalcols].rank = owner;
5595: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5596: ilocal[dntotalcols++] = ntotalcols++;
5597: }
5598: /* off-diagonal */
5599: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5600: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5601: oiremote[ontotalcols].rank = owner;
5602: oilocal[ontotalcols++] = ntotalcols++;
5603: }
5604: }
5605: PetscCall(ISRestoreIndices(rows, &lrowindices));
5606: PetscCall(PetscFree(loffsets));
5607: PetscCall(PetscFree(nlcols));
5608: PetscCall(PetscSFCreate(comm, &sf));
5609: /* P serves as roots and P_oth is leaves
5610: * Diag matrix
5611: * */
5612: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5613: PetscCall(PetscSFSetFromOptions(sf));
5614: PetscCall(PetscSFSetUp(sf));
5616: PetscCall(PetscSFCreate(comm, &osf));
5617: /* off-diagonal */
5618: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5619: PetscCall(PetscSFSetFromOptions(osf));
5620: PetscCall(PetscSFSetUp(osf));
5621: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5622: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5623: /* operate on the matrix internal data to save memory */
5624: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5625: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5626: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5627: /* Convert to global indices for diag matrix */
5628: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5629: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5630: /* We want P_oth store global indices */
5631: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5632: /* Use memory scalable approach */
5633: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5634: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5635: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5636: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5637: /* Convert back to local indices */
5638: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5639: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5640: nout = 0;
5641: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5642: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5643: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5644: /* Exchange values */
5645: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5646: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5647: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5648: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5649: /* Stop PETSc from shrinking memory */
5650: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5651: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5652: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5653: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5654: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5655: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5656: PetscCall(PetscSFDestroy(&sf));
5657: PetscCall(PetscSFDestroy(&osf));
5658: PetscFunctionReturn(PETSC_SUCCESS);
5659: }
5661: /*
5662: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5663: * This supports MPIAIJ and MAIJ
5664: * */
5665: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5666: {
5667: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5668: Mat_SeqAIJ *p_oth;
5669: IS rows, map;
5670: PetscHMapI hamp;
5671: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5672: MPI_Comm comm;
5673: PetscSF sf, osf;
5674: PetscBool has;
5676: PetscFunctionBegin;
5677: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5678: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5679: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5680: * and then create a submatrix (that often is an overlapping matrix)
5681: * */
5682: if (reuse == MAT_INITIAL_MATRIX) {
5683: /* Use a hash table to figure out unique keys */
5684: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5685: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5686: count = 0;
5687: /* Assume that a->g is sorted, otherwise the following does not make sense */
5688: for (i = 0; i < a->B->cmap->n; i++) {
5689: key = a->garray[i] / dof;
5690: PetscCall(PetscHMapIHas(hamp, key, &has));
5691: if (!has) {
5692: mapping[i] = count;
5693: PetscCall(PetscHMapISet(hamp, key, count++));
5694: } else {
5695: /* Current 'i' has the same value the previous step */
5696: mapping[i] = count - 1;
5697: }
5698: }
5699: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5700: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5701: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5702: PetscCall(PetscCalloc1(htsize, &rowindices));
5703: off = 0;
5704: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5705: PetscCall(PetscHMapIDestroy(&hamp));
5706: PetscCall(PetscSortInt(htsize, rowindices));
5707: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5708: /* In case, the matrix was already created but users want to recreate the matrix */
5709: PetscCall(MatDestroy(P_oth));
5710: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5711: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5712: PetscCall(ISDestroy(&map));
5713: PetscCall(ISDestroy(&rows));
5714: } else if (reuse == MAT_REUSE_MATRIX) {
5715: /* If matrix was already created, we simply update values using SF objects
5716: * that as attached to the matrix earlier.
5717: */
5718: const PetscScalar *pd_a, *po_a;
5720: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5721: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5722: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5723: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5724: /* Update values in place */
5725: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5726: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5727: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5728: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5729: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5730: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5731: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5732: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5733: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5734: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5735: PetscFunctionReturn(PETSC_SUCCESS);
5736: }
5738: /*@C
5739: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5741: Collective
5743: Input Parameters:
5744: + A - the first matrix in `MATMPIAIJ` format
5745: . B - the second matrix in `MATMPIAIJ` format
5746: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5748: Output Parameters:
5749: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5750: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5751: - B_seq - the sequential matrix generated
5753: Level: developer
5755: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5756: @*/
5757: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5758: {
5759: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5760: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5761: IS isrowb, iscolb;
5762: Mat *bseq = NULL;
5764: PetscFunctionBegin;
5765: 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 ")",
5766: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5767: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5769: if (scall == MAT_INITIAL_MATRIX) {
5770: start = A->cmap->rstart;
5771: cmap = a->garray;
5772: nzA = a->A->cmap->n;
5773: nzB = a->B->cmap->n;
5774: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5775: ncols = 0;
5776: for (i = 0; i < nzB; i++) { /* row < local row index */
5777: if (cmap[i] < start) idx[ncols++] = cmap[i];
5778: else break;
5779: }
5780: imark = i;
5781: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5782: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5783: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5784: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5785: } else {
5786: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5787: isrowb = *rowb;
5788: iscolb = *colb;
5789: PetscCall(PetscMalloc1(1, &bseq));
5790: bseq[0] = *B_seq;
5791: }
5792: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5793: *B_seq = bseq[0];
5794: PetscCall(PetscFree(bseq));
5795: if (!rowb) {
5796: PetscCall(ISDestroy(&isrowb));
5797: } else {
5798: *rowb = isrowb;
5799: }
5800: if (!colb) {
5801: PetscCall(ISDestroy(&iscolb));
5802: } else {
5803: *colb = iscolb;
5804: }
5805: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5806: PetscFunctionReturn(PETSC_SUCCESS);
5807: }
5809: /*
5810: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5811: of the OFF-DIAGONAL portion of local A
5813: Collective
5815: Input Parameters:
5816: + A,B - the matrices in `MATMPIAIJ` format
5817: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5819: Output Parameter:
5820: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5821: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5822: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5823: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5825: Developer Note:
5826: This directly accesses information inside the VecScatter associated with the matrix-vector product
5827: for this matrix. This is not desirable..
5829: Level: developer
5831: */
5833: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5834: {
5835: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5836: VecScatter ctx;
5837: MPI_Comm comm;
5838: const PetscMPIInt *rprocs, *sprocs;
5839: PetscMPIInt nrecvs, nsends;
5840: const PetscInt *srow, *rstarts, *sstarts;
5841: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5842: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5843: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5844: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5845: PetscMPIInt size, tag, rank, nreqs;
5847: PetscFunctionBegin;
5848: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5849: PetscCallMPI(MPI_Comm_size(comm, &size));
5851: 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 ")",
5852: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5853: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5854: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5856: if (size == 1) {
5857: startsj_s = NULL;
5858: bufa_ptr = NULL;
5859: *B_oth = NULL;
5860: PetscFunctionReturn(PETSC_SUCCESS);
5861: }
5863: ctx = a->Mvctx;
5864: tag = ((PetscObject)ctx)->tag;
5866: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5867: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5868: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5869: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5870: PetscCall(PetscMalloc1(nreqs, &reqs));
5871: rwaits = reqs;
5872: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5874: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5875: if (scall == MAT_INITIAL_MATRIX) {
5876: /* i-array */
5877: /* post receives */
5878: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5879: for (i = 0; i < nrecvs; i++) {
5880: rowlen = rvalues + rstarts[i] * rbs;
5881: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5882: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5883: }
5885: /* pack the outgoing message */
5886: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5888: sstartsj[0] = 0;
5889: rstartsj[0] = 0;
5890: len = 0; /* total length of j or a array to be sent */
5891: if (nsends) {
5892: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5893: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5894: }
5895: for (i = 0; i < nsends; i++) {
5896: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5897: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5898: for (j = 0; j < nrows; j++) {
5899: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5900: for (l = 0; l < sbs; l++) {
5901: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5903: rowlen[j * sbs + l] = ncols;
5905: len += ncols;
5906: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5907: }
5908: k++;
5909: }
5910: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5912: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5913: }
5914: /* recvs and sends of i-array are completed */
5915: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5916: PetscCall(PetscFree(svalues));
5918: /* allocate buffers for sending j and a arrays */
5919: PetscCall(PetscMalloc1(len + 1, &bufj));
5920: PetscCall(PetscMalloc1(len + 1, &bufa));
5922: /* create i-array of B_oth */
5923: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5925: b_othi[0] = 0;
5926: len = 0; /* total length of j or a array to be received */
5927: k = 0;
5928: for (i = 0; i < nrecvs; i++) {
5929: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5930: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5931: for (j = 0; j < nrows; j++) {
5932: b_othi[k + 1] = b_othi[k] + rowlen[j];
5933: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5934: k++;
5935: }
5936: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5937: }
5938: PetscCall(PetscFree(rvalues));
5940: /* allocate space for j and a arrays of B_oth */
5941: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5942: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5944: /* j-array */
5945: /* post receives of j-array */
5946: for (i = 0; i < nrecvs; i++) {
5947: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5948: PetscCallMPI(MPIU_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5949: }
5951: /* pack the outgoing message j-array */
5952: if (nsends) k = sstarts[0];
5953: for (i = 0; i < nsends; i++) {
5954: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5955: bufJ = bufj + sstartsj[i];
5956: for (j = 0; j < nrows; j++) {
5957: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5958: for (ll = 0; ll < sbs; ll++) {
5959: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5960: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5961: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5962: }
5963: }
5964: PetscCallMPI(MPIU_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5965: }
5967: /* recvs and sends of j-array are completed */
5968: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5969: } else if (scall == MAT_REUSE_MATRIX) {
5970: sstartsj = *startsj_s;
5971: rstartsj = *startsj_r;
5972: bufa = *bufa_ptr;
5973: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5974: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5976: /* a-array */
5977: /* post receives of a-array */
5978: for (i = 0; i < nrecvs; i++) {
5979: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5980: PetscCallMPI(MPIU_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5981: }
5983: /* pack the outgoing message a-array */
5984: if (nsends) k = sstarts[0];
5985: for (i = 0; i < nsends; i++) {
5986: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5987: bufA = bufa + sstartsj[i];
5988: for (j = 0; j < nrows; j++) {
5989: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5990: for (ll = 0; ll < sbs; ll++) {
5991: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5992: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5993: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5994: }
5995: }
5996: PetscCallMPI(MPIU_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
5997: }
5998: /* recvs and sends of a-array are completed */
5999: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
6000: PetscCall(PetscFree(reqs));
6002: if (scall == MAT_INITIAL_MATRIX) {
6003: Mat_SeqAIJ *b_oth;
6005: /* put together the new matrix */
6006: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6008: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6009: /* Since these are PETSc arrays, change flags to free them as necessary. */
6010: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6011: b_oth->free_a = PETSC_TRUE;
6012: b_oth->free_ij = PETSC_TRUE;
6013: b_oth->nonew = 0;
6015: PetscCall(PetscFree(bufj));
6016: if (!startsj_s || !bufa_ptr) {
6017: PetscCall(PetscFree2(sstartsj, rstartsj));
6018: PetscCall(PetscFree(bufa_ptr));
6019: } else {
6020: *startsj_s = sstartsj;
6021: *startsj_r = rstartsj;
6022: *bufa_ptr = bufa;
6023: }
6024: } else if (scall == MAT_REUSE_MATRIX) {
6025: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6026: }
6028: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6029: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6030: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6031: PetscFunctionReturn(PETSC_SUCCESS);
6032: }
6034: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6035: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6036: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6037: #if defined(PETSC_HAVE_MKL_SPARSE)
6038: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6039: #endif
6040: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6041: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6042: #if defined(PETSC_HAVE_ELEMENTAL)
6043: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6044: #endif
6045: #if defined(PETSC_HAVE_SCALAPACK)
6046: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6047: #endif
6048: #if defined(PETSC_HAVE_HYPRE)
6049: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6050: #endif
6051: #if defined(PETSC_HAVE_CUDA)
6052: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6053: #endif
6054: #if defined(PETSC_HAVE_HIP)
6055: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6058: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6059: #endif
6060: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6061: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6062: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6064: /*
6065: Computes (B'*A')' since computing B*A directly is untenable
6067: n p p
6068: [ ] [ ] [ ]
6069: m [ A ] * n [ B ] = m [ C ]
6070: [ ] [ ] [ ]
6072: */
6073: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6074: {
6075: Mat At, Bt, Ct;
6077: PetscFunctionBegin;
6078: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6079: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6080: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6081: PetscCall(MatDestroy(&At));
6082: PetscCall(MatDestroy(&Bt));
6083: PetscCall(MatTransposeSetPrecursor(Ct, C));
6084: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6085: PetscCall(MatDestroy(&Ct));
6086: PetscFunctionReturn(PETSC_SUCCESS);
6087: }
6089: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6090: {
6091: PetscBool cisdense;
6093: PetscFunctionBegin;
6094: 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);
6095: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6096: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6097: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6098: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6099: PetscCall(MatSetUp(C));
6101: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6102: PetscFunctionReturn(PETSC_SUCCESS);
6103: }
6105: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6106: {
6107: Mat_Product *product = C->product;
6108: Mat A = product->A, B = product->B;
6110: PetscFunctionBegin;
6111: 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 ")",
6112: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6113: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6114: C->ops->productsymbolic = MatProductSymbolic_AB;
6115: PetscFunctionReturn(PETSC_SUCCESS);
6116: }
6118: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6119: {
6120: Mat_Product *product = C->product;
6122: PetscFunctionBegin;
6123: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6124: PetscFunctionReturn(PETSC_SUCCESS);
6125: }
6127: /*
6128: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6130: Input Parameters:
6132: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6133: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6135: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6137: For Set1, j1[] contains column indices of the nonzeros.
6138: 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
6139: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6140: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6142: Similar for Set2.
6144: This routine merges the two sets of nonzeros row by row and removes repeats.
6146: Output Parameters: (memory is allocated by the caller)
6148: i[],j[]: the CSR of the merged matrix, which has m rows.
6149: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6150: imap2[]: similar to imap1[], but for Set2.
6151: Note we order nonzeros row-by-row and from left to right.
6152: */
6153: 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[])
6154: {
6155: PetscInt r, m; /* Row index of mat */
6156: PetscCount t, t1, t2, b1, e1, b2, e2;
6158: PetscFunctionBegin;
6159: PetscCall(MatGetLocalSize(mat, &m, NULL));
6160: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6161: i[0] = 0;
6162: for (r = 0; r < m; r++) { /* Do row by row merging */
6163: b1 = rowBegin1[r];
6164: e1 = rowEnd1[r];
6165: b2 = rowBegin2[r];
6166: e2 = rowEnd2[r];
6167: while (b1 < e1 && b2 < e2) {
6168: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6169: j[t] = j1[b1];
6170: imap1[t1] = t;
6171: imap2[t2] = t;
6172: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6173: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6174: t1++;
6175: t2++;
6176: t++;
6177: } else if (j1[b1] < j2[b2]) {
6178: j[t] = j1[b1];
6179: imap1[t1] = t;
6180: b1 += jmap1[t1 + 1] - jmap1[t1];
6181: t1++;
6182: t++;
6183: } else {
6184: j[t] = j2[b2];
6185: imap2[t2] = t;
6186: b2 += jmap2[t2 + 1] - jmap2[t2];
6187: t2++;
6188: t++;
6189: }
6190: }
6191: /* Merge the remaining in either j1[] or j2[] */
6192: while (b1 < e1) {
6193: j[t] = j1[b1];
6194: imap1[t1] = t;
6195: b1 += jmap1[t1 + 1] - jmap1[t1];
6196: t1++;
6197: t++;
6198: }
6199: while (b2 < e2) {
6200: j[t] = j2[b2];
6201: imap2[t2] = t;
6202: b2 += jmap2[t2 + 1] - jmap2[t2];
6203: t2++;
6204: t++;
6205: }
6206: PetscCall(PetscIntCast(t, i + r + 1));
6207: }
6208: PetscFunctionReturn(PETSC_SUCCESS);
6209: }
6211: /*
6212: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6214: Input Parameters:
6215: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6216: 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[]
6217: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6219: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6220: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6222: Output Parameters:
6223: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6224: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6225: 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,
6226: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6228: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6229: Atot: number of entries belonging to the diagonal block.
6230: Annz: number of unique nonzeros belonging to the diagonal block.
6231: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6232: repeats (i.e., same 'i,j' pair).
6233: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6234: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6236: Atot: number of entries belonging to the diagonal block
6237: Annz: number of unique nonzeros belonging to the diagonal block.
6239: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6241: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6242: */
6243: 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_)
6244: {
6245: PetscInt cstart, cend, rstart, rend, row, col;
6246: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6247: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6248: PetscCount k, m, p, q, r, s, mid;
6249: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6251: PetscFunctionBegin;
6252: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6253: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6254: m = rend - rstart;
6256: /* Skip negative rows */
6257: for (k = 0; k < n; k++)
6258: if (i[k] >= 0) break;
6260: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6261: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6262: */
6263: while (k < n) {
6264: row = i[k];
6265: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6266: for (s = k; s < n; s++)
6267: if (i[s] != row) break;
6269: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6270: for (p = k; p < s; p++) {
6271: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6272: 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]);
6273: }
6274: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6275: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6276: rowBegin[row - rstart] = k;
6277: rowMid[row - rstart] = mid;
6278: rowEnd[row - rstart] = s;
6280: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6281: Atot += mid - k;
6282: Btot += s - mid;
6284: /* Count unique nonzeros of this diag row */
6285: for (p = k; p < mid;) {
6286: col = j[p];
6287: do {
6288: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6289: p++;
6290: } while (p < mid && j[p] == col);
6291: Annz++;
6292: }
6294: /* Count unique nonzeros of this offdiag row */
6295: for (p = mid; p < s;) {
6296: col = j[p];
6297: do {
6298: p++;
6299: } while (p < s && j[p] == col);
6300: Bnnz++;
6301: }
6302: k = s;
6303: }
6305: /* Allocation according to Atot, Btot, Annz, Bnnz */
6306: PetscCall(PetscMalloc1(Atot, &Aperm));
6307: PetscCall(PetscMalloc1(Btot, &Bperm));
6308: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6309: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6311: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6312: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6313: for (r = 0; r < m; r++) {
6314: k = rowBegin[r];
6315: mid = rowMid[r];
6316: s = rowEnd[r];
6317: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6318: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6319: Atot += mid - k;
6320: Btot += s - mid;
6322: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6323: for (p = k; p < mid;) {
6324: col = j[p];
6325: q = p;
6326: do {
6327: p++;
6328: } while (p < mid && j[p] == col);
6329: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6330: Annz++;
6331: }
6333: for (p = mid; p < s;) {
6334: col = j[p];
6335: q = p;
6336: do {
6337: p++;
6338: } while (p < s && j[p] == col);
6339: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6340: Bnnz++;
6341: }
6342: }
6343: /* Output */
6344: *Aperm_ = Aperm;
6345: *Annz_ = Annz;
6346: *Atot_ = Atot;
6347: *Ajmap_ = Ajmap;
6348: *Bperm_ = Bperm;
6349: *Bnnz_ = Bnnz;
6350: *Btot_ = Btot;
6351: *Bjmap_ = Bjmap;
6352: PetscFunctionReturn(PETSC_SUCCESS);
6353: }
6355: /*
6356: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6358: Input Parameters:
6359: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6360: nnz: number of unique nonzeros in the merged matrix
6361: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6362: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6364: Output Parameter: (memory is allocated by the caller)
6365: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6367: Example:
6368: nnz1 = 4
6369: nnz = 6
6370: imap = [1,3,4,5]
6371: jmap = [0,3,5,6,7]
6372: then,
6373: jmap_new = [0,0,3,3,5,6,7]
6374: */
6375: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6376: {
6377: PetscCount k, p;
6379: PetscFunctionBegin;
6380: jmap_new[0] = 0;
6381: p = nnz; /* p loops over jmap_new[] backwards */
6382: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6383: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6384: }
6385: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6386: PetscFunctionReturn(PETSC_SUCCESS);
6387: }
6389: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6390: {
6391: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6393: PetscFunctionBegin;
6394: PetscCall(PetscSFDestroy(&coo->sf));
6395: PetscCall(PetscFree(coo->Aperm1));
6396: PetscCall(PetscFree(coo->Bperm1));
6397: PetscCall(PetscFree(coo->Ajmap1));
6398: PetscCall(PetscFree(coo->Bjmap1));
6399: PetscCall(PetscFree(coo->Aimap2));
6400: PetscCall(PetscFree(coo->Bimap2));
6401: PetscCall(PetscFree(coo->Aperm2));
6402: PetscCall(PetscFree(coo->Bperm2));
6403: PetscCall(PetscFree(coo->Ajmap2));
6404: PetscCall(PetscFree(coo->Bjmap2));
6405: PetscCall(PetscFree(coo->Cperm1));
6406: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6407: PetscCall(PetscFree(coo));
6408: PetscFunctionReturn(PETSC_SUCCESS);
6409: }
6411: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6412: {
6413: MPI_Comm comm;
6414: PetscMPIInt rank, size;
6415: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6416: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6417: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6418: PetscContainer container;
6419: MatCOOStruct_MPIAIJ *coo;
6421: PetscFunctionBegin;
6422: PetscCall(PetscFree(mpiaij->garray));
6423: PetscCall(VecDestroy(&mpiaij->lvec));
6424: #if defined(PETSC_USE_CTABLE)
6425: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6426: #else
6427: PetscCall(PetscFree(mpiaij->colmap));
6428: #endif
6429: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6430: mat->assembled = PETSC_FALSE;
6431: mat->was_assembled = PETSC_FALSE;
6433: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6434: PetscCallMPI(MPI_Comm_size(comm, &size));
6435: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6436: PetscCall(PetscLayoutSetUp(mat->rmap));
6437: PetscCall(PetscLayoutSetUp(mat->cmap));
6438: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6439: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6440: PetscCall(MatGetLocalSize(mat, &m, &n));
6441: PetscCall(MatGetSize(mat, &M, &N));
6443: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6444: /* entries come first, then local rows, then remote rows. */
6445: PetscCount n1 = coo_n, *perm1;
6446: PetscInt *i1 = coo_i, *j1 = coo_j;
6448: PetscCall(PetscMalloc1(n1, &perm1));
6449: for (k = 0; k < n1; k++) perm1[k] = k;
6451: /* Manipulate indices so that entries with negative row or col indices will have smallest
6452: row indices, local entries will have greater but negative row indices, and remote entries
6453: will have positive row indices.
6454: */
6455: for (k = 0; k < n1; k++) {
6456: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6457: else if (i1[k] >= rstart && i1[k] < rend) i1[k] -= PETSC_INT_MAX; /* e.g., minus 2^31-1 to shift local rows to range of [-PETSC_INT_MAX, -1] */
6458: else {
6459: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6460: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6461: }
6462: }
6464: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6465: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6467: /* Advance k to the first entry we need to take care of */
6468: for (k = 0; k < n1; k++)
6469: if (i1[k] > PETSC_INT_MIN) break;
6470: PetscCount i1start = k;
6472: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6473: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6475: /* Send remote rows to their owner */
6476: /* Find which rows should be sent to which remote ranks*/
6477: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6478: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6479: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6480: const PetscInt *ranges;
6481: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6483: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6484: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6485: for (k = rem; k < n1;) {
6486: PetscMPIInt owner;
6487: PetscInt firstRow, lastRow;
6489: /* Locate a row range */
6490: firstRow = i1[k]; /* first row of this owner */
6491: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6492: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6494: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6495: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6497: /* All entries in [k,p) belong to this remote owner */
6498: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6499: PetscMPIInt *sendto2;
6500: PetscInt *nentries2;
6501: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6503: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6504: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6505: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6506: PetscCall(PetscFree2(sendto, nentries2));
6507: sendto = sendto2;
6508: nentries = nentries2;
6509: maxNsend = maxNsend2;
6510: }
6511: sendto[nsend] = owner;
6512: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6513: nsend++;
6514: k = p;
6515: }
6517: /* Build 1st SF to know offsets on remote to send data */
6518: PetscSF sf1;
6519: PetscInt nroots = 1, nroots2 = 0;
6520: PetscInt nleaves = nsend, nleaves2 = 0;
6521: PetscInt *offsets;
6522: PetscSFNode *iremote;
6524: PetscCall(PetscSFCreate(comm, &sf1));
6525: PetscCall(PetscMalloc1(nsend, &iremote));
6526: PetscCall(PetscMalloc1(nsend, &offsets));
6527: for (k = 0; k < nsend; k++) {
6528: iremote[k].rank = sendto[k];
6529: iremote[k].index = 0;
6530: nleaves2 += nentries[k];
6531: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6532: }
6533: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6534: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6535: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6536: PetscCall(PetscSFDestroy(&sf1));
6537: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6539: /* Build 2nd SF to send remote COOs to their owner */
6540: PetscSF sf2;
6541: nroots = nroots2;
6542: nleaves = nleaves2;
6543: PetscCall(PetscSFCreate(comm, &sf2));
6544: PetscCall(PetscSFSetFromOptions(sf2));
6545: PetscCall(PetscMalloc1(nleaves, &iremote));
6546: p = 0;
6547: for (k = 0; k < nsend; k++) {
6548: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6549: for (q = 0; q < nentries[k]; q++, p++) {
6550: iremote[p].rank = sendto[k];
6551: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6552: }
6553: }
6554: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6556: /* Send the remote COOs to their owner */
6557: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6558: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6559: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6560: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6561: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6562: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6563: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6564: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6565: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6566: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6567: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6569: PetscCall(PetscFree(offsets));
6570: PetscCall(PetscFree2(sendto, nentries));
6572: /* Sort received COOs by row along with the permutation array */
6573: for (k = 0; k < n2; k++) perm2[k] = k;
6574: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6576: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6577: PetscCount *Cperm1;
6578: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6579: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6580: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6581: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6583: /* Support for HYPRE matrices, kind of a hack.
6584: Swap min column with diagonal so that diagonal values will go first */
6585: PetscBool hypre;
6586: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6587: if (hypre) {
6588: PetscInt *minj;
6589: PetscBT hasdiag;
6591: PetscCall(PetscBTCreate(m, &hasdiag));
6592: PetscCall(PetscMalloc1(m, &minj));
6593: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6594: for (k = i1start; k < rem; k++) {
6595: if (j1[k] < cstart || j1[k] >= cend) continue;
6596: const PetscInt rindex = i1[k] - rstart;
6597: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6598: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6599: }
6600: for (k = 0; k < n2; k++) {
6601: if (j2[k] < cstart || j2[k] >= cend) continue;
6602: const PetscInt rindex = i2[k] - rstart;
6603: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6604: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6605: }
6606: for (k = i1start; k < rem; k++) {
6607: const PetscInt rindex = i1[k] - rstart;
6608: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6609: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6610: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6611: }
6612: for (k = 0; k < n2; k++) {
6613: const PetscInt rindex = i2[k] - rstart;
6614: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6615: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6616: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6617: }
6618: PetscCall(PetscBTDestroy(&hasdiag));
6619: PetscCall(PetscFree(minj));
6620: }
6622: /* Split local COOs and received COOs into diag/offdiag portions */
6623: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6624: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6625: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6626: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6627: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6628: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6630: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6631: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6632: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6633: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6635: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6636: PetscInt *Ai, *Bi;
6637: PetscInt *Aj, *Bj;
6639: PetscCall(PetscMalloc1(m + 1, &Ai));
6640: PetscCall(PetscMalloc1(m + 1, &Bi));
6641: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6642: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6644: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6645: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6646: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6647: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6648: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6650: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6651: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6653: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6654: /* expect nonzeros in A/B most likely have local contributing entries */
6655: PetscInt Annz = Ai[m];
6656: PetscInt Bnnz = Bi[m];
6657: PetscCount *Ajmap1_new, *Bjmap1_new;
6659: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6660: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6662: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6663: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6665: PetscCall(PetscFree(Aimap1));
6666: PetscCall(PetscFree(Ajmap1));
6667: PetscCall(PetscFree(Bimap1));
6668: PetscCall(PetscFree(Bjmap1));
6669: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6670: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6671: PetscCall(PetscFree(perm1));
6672: PetscCall(PetscFree3(i2, j2, perm2));
6674: Ajmap1 = Ajmap1_new;
6675: Bjmap1 = Bjmap1_new;
6677: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6678: if (Annz < Annz1 + Annz2) {
6679: PetscInt *Aj_new;
6680: PetscCall(PetscMalloc1(Annz, &Aj_new));
6681: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6682: PetscCall(PetscFree(Aj));
6683: Aj = Aj_new;
6684: }
6686: if (Bnnz < Bnnz1 + Bnnz2) {
6687: PetscInt *Bj_new;
6688: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6689: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6690: PetscCall(PetscFree(Bj));
6691: Bj = Bj_new;
6692: }
6694: /* Create new submatrices for on-process and off-process coupling */
6695: PetscScalar *Aa, *Ba;
6696: MatType rtype;
6697: Mat_SeqAIJ *a, *b;
6698: PetscObjectState state;
6699: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6700: PetscCall(PetscCalloc1(Bnnz, &Ba));
6701: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6702: if (cstart) {
6703: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6704: }
6706: PetscCall(MatGetRootType_Private(mat, &rtype));
6708: MatSeqXAIJGetOptions_Private(mpiaij->A);
6709: PetscCall(MatDestroy(&mpiaij->A));
6710: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6711: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6712: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6714: MatSeqXAIJGetOptions_Private(mpiaij->B);
6715: PetscCall(MatDestroy(&mpiaij->B));
6716: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6717: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6718: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6720: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6721: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6722: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6723: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6725: a = (Mat_SeqAIJ *)mpiaij->A->data;
6726: b = (Mat_SeqAIJ *)mpiaij->B->data;
6727: a->free_a = PETSC_TRUE;
6728: a->free_ij = PETSC_TRUE;
6729: b->free_a = PETSC_TRUE;
6730: b->free_ij = PETSC_TRUE;
6731: a->maxnz = a->nz;
6732: b->maxnz = b->nz;
6734: /* conversion must happen AFTER multiply setup */
6735: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6736: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6737: PetscCall(VecDestroy(&mpiaij->lvec));
6738: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6740: // Put the COO struct in a container and then attach that to the matrix
6741: PetscCall(PetscMalloc1(1, &coo));
6742: coo->n = coo_n;
6743: coo->sf = sf2;
6744: coo->sendlen = nleaves;
6745: coo->recvlen = nroots;
6746: coo->Annz = Annz;
6747: coo->Bnnz = Bnnz;
6748: coo->Annz2 = Annz2;
6749: coo->Bnnz2 = Bnnz2;
6750: coo->Atot1 = Atot1;
6751: coo->Atot2 = Atot2;
6752: coo->Btot1 = Btot1;
6753: coo->Btot2 = Btot2;
6754: coo->Ajmap1 = Ajmap1;
6755: coo->Aperm1 = Aperm1;
6756: coo->Bjmap1 = Bjmap1;
6757: coo->Bperm1 = Bperm1;
6758: coo->Aimap2 = Aimap2;
6759: coo->Ajmap2 = Ajmap2;
6760: coo->Aperm2 = Aperm2;
6761: coo->Bimap2 = Bimap2;
6762: coo->Bjmap2 = Bjmap2;
6763: coo->Bperm2 = Bperm2;
6764: coo->Cperm1 = Cperm1;
6765: // Allocate in preallocation. If not used, it has zero cost on host
6766: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6767: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6768: PetscCall(PetscContainerSetPointer(container, coo));
6769: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6770: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6771: PetscCall(PetscContainerDestroy(&container));
6772: PetscFunctionReturn(PETSC_SUCCESS);
6773: }
6775: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6776: {
6777: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6778: Mat A = mpiaij->A, B = mpiaij->B;
6779: PetscScalar *Aa, *Ba;
6780: PetscScalar *sendbuf, *recvbuf;
6781: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6782: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6783: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6784: const PetscCount *Cperm1;
6785: PetscContainer container;
6786: MatCOOStruct_MPIAIJ *coo;
6788: PetscFunctionBegin;
6789: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6790: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6791: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6792: sendbuf = coo->sendbuf;
6793: recvbuf = coo->recvbuf;
6794: Ajmap1 = coo->Ajmap1;
6795: Ajmap2 = coo->Ajmap2;
6796: Aimap2 = coo->Aimap2;
6797: Bjmap1 = coo->Bjmap1;
6798: Bjmap2 = coo->Bjmap2;
6799: Bimap2 = coo->Bimap2;
6800: Aperm1 = coo->Aperm1;
6801: Aperm2 = coo->Aperm2;
6802: Bperm1 = coo->Bperm1;
6803: Bperm2 = coo->Bperm2;
6804: Cperm1 = coo->Cperm1;
6806: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6807: PetscCall(MatSeqAIJGetArray(B, &Ba));
6809: /* Pack entries to be sent to remote */
6810: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6812: /* Send remote entries to their owner and overlap the communication with local computation */
6813: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6814: /* Add local entries to A and B */
6815: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6816: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6817: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6818: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6819: }
6820: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6821: PetscScalar sum = 0.0;
6822: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6823: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6824: }
6825: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6827: /* Add received remote entries to A and B */
6828: for (PetscCount i = 0; i < coo->Annz2; i++) {
6829: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6830: }
6831: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6832: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6833: }
6834: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6835: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6836: PetscFunctionReturn(PETSC_SUCCESS);
6837: }
6839: /*MC
6840: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6842: Options Database Keys:
6843: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6845: Level: beginner
6847: Notes:
6848: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6849: in this case the values associated with the rows and columns one passes in are set to zero
6850: in the matrix
6852: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6853: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6855: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6856: M*/
6857: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6858: {
6859: Mat_MPIAIJ *b;
6860: PetscMPIInt size;
6862: PetscFunctionBegin;
6863: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6865: PetscCall(PetscNew(&b));
6866: B->data = (void *)b;
6867: B->ops[0] = MatOps_Values;
6868: B->assembled = PETSC_FALSE;
6869: B->insertmode = NOT_SET_VALUES;
6870: b->size = size;
6872: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6874: /* build cache for off array entries formed */
6875: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6877: b->donotstash = PETSC_FALSE;
6878: b->colmap = NULL;
6879: b->garray = NULL;
6880: b->roworiented = PETSC_TRUE;
6882: /* stuff used for matrix vector multiply */
6883: b->lvec = NULL;
6884: b->Mvctx = NULL;
6886: /* stuff for MatGetRow() */
6887: b->rowindices = NULL;
6888: b->rowvalues = NULL;
6889: b->getrowactive = PETSC_FALSE;
6891: /* flexible pointer used in CUSPARSE classes */
6892: b->spptr = NULL;
6894: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6895: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6896: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6897: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6899: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6903: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6904: #if defined(PETSC_HAVE_CUDA)
6905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6906: #endif
6907: #if defined(PETSC_HAVE_HIP)
6908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6909: #endif
6910: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6912: #endif
6913: #if defined(PETSC_HAVE_MKL_SPARSE)
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6915: #endif
6916: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6918: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6920: #if defined(PETSC_HAVE_ELEMENTAL)
6921: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6922: #endif
6923: #if defined(PETSC_HAVE_SCALAPACK)
6924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6925: #endif
6926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6927: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6928: #if defined(PETSC_HAVE_HYPRE)
6929: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6930: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6931: #endif
6932: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6933: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6934: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6935: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6936: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6937: PetscFunctionReturn(PETSC_SUCCESS);
6938: }
6940: /*@
6941: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6942: and "off-diagonal" part of the matrix in CSR format.
6944: Collective
6946: Input Parameters:
6947: + comm - MPI communicator
6948: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6949: . n - This value should be the same as the local size used in creating the
6950: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6951: calculated if `N` is given) For square matrices `n` is almost always `m`.
6952: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6953: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6954: . 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
6955: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6956: . a - matrix values
6957: . 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
6958: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6959: - oa - matrix values
6961: Output Parameter:
6962: . mat - the matrix
6964: Level: advanced
6966: Notes:
6967: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6968: must free the arrays once the matrix has been destroyed and not before.
6970: The `i` and `j` indices are 0 based
6972: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6974: This sets local rows and cannot be used to set off-processor values.
6976: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6977: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6978: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6979: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6980: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6981: communication if it is known that only local entries will be set.
6983: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6984: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6985: @*/
6986: 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)
6987: {
6988: Mat_MPIAIJ *maij;
6990: PetscFunctionBegin;
6991: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6992: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6993: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6994: PetscCall(MatCreate(comm, mat));
6995: PetscCall(MatSetSizes(*mat, m, n, M, N));
6996: PetscCall(MatSetType(*mat, MATMPIAIJ));
6997: maij = (Mat_MPIAIJ *)(*mat)->data;
6999: (*mat)->preallocated = PETSC_TRUE;
7001: PetscCall(PetscLayoutSetUp((*mat)->rmap));
7002: PetscCall(PetscLayoutSetUp((*mat)->cmap));
7004: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
7005: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7007: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7008: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7009: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7010: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7011: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7012: PetscFunctionReturn(PETSC_SUCCESS);
7013: }
7015: typedef struct {
7016: Mat *mp; /* intermediate products */
7017: PetscBool *mptmp; /* is the intermediate product temporary ? */
7018: PetscInt cp; /* number of intermediate products */
7020: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7021: PetscInt *startsj_s, *startsj_r;
7022: PetscScalar *bufa;
7023: Mat P_oth;
7025: /* may take advantage of merging product->B */
7026: Mat Bloc; /* B-local by merging diag and off-diag */
7028: /* cusparse does not have support to split between symbolic and numeric phases.
7029: When api_user is true, we don't need to update the numerical values
7030: of the temporary storage */
7031: PetscBool reusesym;
7033: /* support for COO values insertion */
7034: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7035: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7036: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7037: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7038: PetscSF sf; /* used for non-local values insertion and memory malloc */
7039: PetscMemType mtype;
7041: /* customization */
7042: PetscBool abmerge;
7043: PetscBool P_oth_bind;
7044: } MatMatMPIAIJBACKEND;
7046: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7047: {
7048: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7049: PetscInt i;
7051: PetscFunctionBegin;
7052: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7053: PetscCall(PetscFree(mmdata->bufa));
7054: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7055: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7056: PetscCall(MatDestroy(&mmdata->P_oth));
7057: PetscCall(MatDestroy(&mmdata->Bloc));
7058: PetscCall(PetscSFDestroy(&mmdata->sf));
7059: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7060: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7061: PetscCall(PetscFree(mmdata->own[0]));
7062: PetscCall(PetscFree(mmdata->own));
7063: PetscCall(PetscFree(mmdata->off[0]));
7064: PetscCall(PetscFree(mmdata->off));
7065: PetscCall(PetscFree(mmdata));
7066: PetscFunctionReturn(PETSC_SUCCESS);
7067: }
7069: /* Copy selected n entries with indices in idx[] of A to v[].
7070: If idx is NULL, copy the whole data array of A to v[]
7071: */
7072: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7073: {
7074: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7076: PetscFunctionBegin;
7077: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7078: if (f) {
7079: PetscCall((*f)(A, n, idx, v));
7080: } else {
7081: const PetscScalar *vv;
7083: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7084: if (n && idx) {
7085: PetscScalar *w = v;
7086: const PetscInt *oi = idx;
7087: PetscInt j;
7089: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7090: } else {
7091: PetscCall(PetscArraycpy(v, vv, n));
7092: }
7093: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7094: }
7095: PetscFunctionReturn(PETSC_SUCCESS);
7096: }
7098: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7099: {
7100: MatMatMPIAIJBACKEND *mmdata;
7101: PetscInt i, n_d, n_o;
7103: PetscFunctionBegin;
7104: MatCheckProduct(C, 1);
7105: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7106: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7107: if (!mmdata->reusesym) { /* update temporary matrices */
7108: 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));
7109: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7110: }
7111: mmdata->reusesym = PETSC_FALSE;
7113: for (i = 0; i < mmdata->cp; i++) {
7114: 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]);
7115: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7116: }
7117: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7118: PetscInt noff;
7120: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7121: if (mmdata->mptmp[i]) continue;
7122: if (noff) {
7123: PetscInt nown;
7125: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7126: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7127: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7128: n_o += noff;
7129: n_d += nown;
7130: } else {
7131: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7133: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7134: n_d += mm->nz;
7135: }
7136: }
7137: if (mmdata->hasoffproc) { /* offprocess insertion */
7138: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7139: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7140: }
7141: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7142: PetscFunctionReturn(PETSC_SUCCESS);
7143: }
7145: /* Support for Pt * A, A * P, or Pt * A * P */
7146: #define MAX_NUMBER_INTERMEDIATE 4
7147: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7148: {
7149: Mat_Product *product = C->product;
7150: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7151: Mat_MPIAIJ *a, *p;
7152: MatMatMPIAIJBACKEND *mmdata;
7153: ISLocalToGlobalMapping P_oth_l2g = NULL;
7154: IS glob = NULL;
7155: const char *prefix;
7156: char pprefix[256];
7157: const PetscInt *globidx, *P_oth_idx;
7158: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7159: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7160: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7161: /* type-0: consecutive, start from 0; type-1: consecutive with */
7162: /* a base offset; type-2: sparse with a local to global map table */
7163: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7165: MatProductType ptype;
7166: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7167: PetscMPIInt size;
7169: PetscFunctionBegin;
7170: MatCheckProduct(C, 1);
7171: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7172: ptype = product->type;
7173: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7174: ptype = MATPRODUCT_AB;
7175: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7176: }
7177: switch (ptype) {
7178: case MATPRODUCT_AB:
7179: A = product->A;
7180: P = product->B;
7181: m = A->rmap->n;
7182: n = P->cmap->n;
7183: M = A->rmap->N;
7184: N = P->cmap->N;
7185: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7186: break;
7187: case MATPRODUCT_AtB:
7188: P = product->A;
7189: A = product->B;
7190: m = P->cmap->n;
7191: n = A->cmap->n;
7192: M = P->cmap->N;
7193: N = A->cmap->N;
7194: hasoffproc = PETSC_TRUE;
7195: break;
7196: case MATPRODUCT_PtAP:
7197: A = product->A;
7198: P = product->B;
7199: m = P->cmap->n;
7200: n = P->cmap->n;
7201: M = P->cmap->N;
7202: N = P->cmap->N;
7203: hasoffproc = PETSC_TRUE;
7204: break;
7205: default:
7206: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7207: }
7208: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7209: if (size == 1) hasoffproc = PETSC_FALSE;
7211: /* defaults */
7212: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7213: mp[i] = NULL;
7214: mptmp[i] = PETSC_FALSE;
7215: rmapt[i] = -1;
7216: cmapt[i] = -1;
7217: rmapa[i] = NULL;
7218: cmapa[i] = NULL;
7219: }
7221: /* customization */
7222: PetscCall(PetscNew(&mmdata));
7223: mmdata->reusesym = product->api_user;
7224: if (ptype == MATPRODUCT_AB) {
7225: if (product->api_user) {
7226: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7227: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7228: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7229: PetscOptionsEnd();
7230: } else {
7231: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7232: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7233: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7234: PetscOptionsEnd();
7235: }
7236: } else if (ptype == MATPRODUCT_PtAP) {
7237: if (product->api_user) {
7238: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7239: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7240: PetscOptionsEnd();
7241: } else {
7242: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7243: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7244: PetscOptionsEnd();
7245: }
7246: }
7247: a = (Mat_MPIAIJ *)A->data;
7248: p = (Mat_MPIAIJ *)P->data;
7249: PetscCall(MatSetSizes(C, m, n, M, N));
7250: PetscCall(PetscLayoutSetUp(C->rmap));
7251: PetscCall(PetscLayoutSetUp(C->cmap));
7252: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7253: PetscCall(MatGetOptionsPrefix(C, &prefix));
7255: cp = 0;
7256: switch (ptype) {
7257: case MATPRODUCT_AB: /* A * P */
7258: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7260: /* A_diag * P_local (merged or not) */
7261: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7262: /* P is product->B */
7263: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7264: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7265: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7266: PetscCall(MatProductSetFill(mp[cp], product->fill));
7267: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7268: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7269: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7270: mp[cp]->product->api_user = product->api_user;
7271: PetscCall(MatProductSetFromOptions(mp[cp]));
7272: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7273: PetscCall(ISGetIndices(glob, &globidx));
7274: rmapt[cp] = 1;
7275: cmapt[cp] = 2;
7276: cmapa[cp] = globidx;
7277: mptmp[cp] = PETSC_FALSE;
7278: cp++;
7279: } else { /* A_diag * P_diag and A_diag * P_off */
7280: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7281: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7282: PetscCall(MatProductSetFill(mp[cp], product->fill));
7283: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7284: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7285: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7286: mp[cp]->product->api_user = product->api_user;
7287: PetscCall(MatProductSetFromOptions(mp[cp]));
7288: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7289: rmapt[cp] = 1;
7290: cmapt[cp] = 1;
7291: mptmp[cp] = PETSC_FALSE;
7292: cp++;
7293: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7294: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7295: PetscCall(MatProductSetFill(mp[cp], product->fill));
7296: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7297: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7298: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7299: mp[cp]->product->api_user = product->api_user;
7300: PetscCall(MatProductSetFromOptions(mp[cp]));
7301: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7302: rmapt[cp] = 1;
7303: cmapt[cp] = 2;
7304: cmapa[cp] = p->garray;
7305: mptmp[cp] = PETSC_FALSE;
7306: cp++;
7307: }
7309: /* A_off * P_other */
7310: if (mmdata->P_oth) {
7311: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7312: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7313: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7314: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7315: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7316: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7317: PetscCall(MatProductSetFill(mp[cp], product->fill));
7318: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7319: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7320: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7321: mp[cp]->product->api_user = product->api_user;
7322: PetscCall(MatProductSetFromOptions(mp[cp]));
7323: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7324: rmapt[cp] = 1;
7325: cmapt[cp] = 2;
7326: cmapa[cp] = P_oth_idx;
7327: mptmp[cp] = PETSC_FALSE;
7328: cp++;
7329: }
7330: break;
7332: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7333: /* A is product->B */
7334: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7335: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7336: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7337: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7338: PetscCall(MatProductSetFill(mp[cp], product->fill));
7339: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7340: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7341: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7342: mp[cp]->product->api_user = product->api_user;
7343: PetscCall(MatProductSetFromOptions(mp[cp]));
7344: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7345: PetscCall(ISGetIndices(glob, &globidx));
7346: rmapt[cp] = 2;
7347: rmapa[cp] = globidx;
7348: cmapt[cp] = 2;
7349: cmapa[cp] = globidx;
7350: mptmp[cp] = PETSC_FALSE;
7351: cp++;
7352: } else {
7353: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7354: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7355: PetscCall(MatProductSetFill(mp[cp], product->fill));
7356: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7357: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7358: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7359: mp[cp]->product->api_user = product->api_user;
7360: PetscCall(MatProductSetFromOptions(mp[cp]));
7361: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7362: PetscCall(ISGetIndices(glob, &globidx));
7363: rmapt[cp] = 1;
7364: cmapt[cp] = 2;
7365: cmapa[cp] = globidx;
7366: mptmp[cp] = PETSC_FALSE;
7367: cp++;
7368: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7369: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7370: PetscCall(MatProductSetFill(mp[cp], product->fill));
7371: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7372: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7373: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7374: mp[cp]->product->api_user = product->api_user;
7375: PetscCall(MatProductSetFromOptions(mp[cp]));
7376: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7377: rmapt[cp] = 2;
7378: rmapa[cp] = p->garray;
7379: cmapt[cp] = 2;
7380: cmapa[cp] = globidx;
7381: mptmp[cp] = PETSC_FALSE;
7382: cp++;
7383: }
7384: break;
7385: case MATPRODUCT_PtAP:
7386: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7387: /* P is product->B */
7388: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7389: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7390: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7391: PetscCall(MatProductSetFill(mp[cp], product->fill));
7392: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7393: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7394: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7395: mp[cp]->product->api_user = product->api_user;
7396: PetscCall(MatProductSetFromOptions(mp[cp]));
7397: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7398: PetscCall(ISGetIndices(glob, &globidx));
7399: rmapt[cp] = 2;
7400: rmapa[cp] = globidx;
7401: cmapt[cp] = 2;
7402: cmapa[cp] = globidx;
7403: mptmp[cp] = PETSC_FALSE;
7404: cp++;
7405: if (mmdata->P_oth) {
7406: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7407: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7408: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7409: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7410: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7411: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7412: PetscCall(MatProductSetFill(mp[cp], product->fill));
7413: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7414: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7415: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7416: mp[cp]->product->api_user = product->api_user;
7417: PetscCall(MatProductSetFromOptions(mp[cp]));
7418: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7419: mptmp[cp] = PETSC_TRUE;
7420: cp++;
7421: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7422: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7423: PetscCall(MatProductSetFill(mp[cp], product->fill));
7424: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7425: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7426: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7427: mp[cp]->product->api_user = product->api_user;
7428: PetscCall(MatProductSetFromOptions(mp[cp]));
7429: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7430: rmapt[cp] = 2;
7431: rmapa[cp] = globidx;
7432: cmapt[cp] = 2;
7433: cmapa[cp] = P_oth_idx;
7434: mptmp[cp] = PETSC_FALSE;
7435: cp++;
7436: }
7437: break;
7438: default:
7439: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7440: }
7441: /* sanity check */
7442: if (size > 1)
7443: 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);
7445: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7446: for (i = 0; i < cp; i++) {
7447: mmdata->mp[i] = mp[i];
7448: mmdata->mptmp[i] = mptmp[i];
7449: }
7450: mmdata->cp = cp;
7451: C->product->data = mmdata;
7452: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7453: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7455: /* memory type */
7456: mmdata->mtype = PETSC_MEMTYPE_HOST;
7457: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7458: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7459: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7460: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7461: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7462: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7464: /* prepare coo coordinates for values insertion */
7466: /* count total nonzeros of those intermediate seqaij Mats
7467: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7468: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7469: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7470: */
7471: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7472: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7473: if (mptmp[cp]) continue;
7474: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7475: const PetscInt *rmap = rmapa[cp];
7476: const PetscInt mr = mp[cp]->rmap->n;
7477: const PetscInt rs = C->rmap->rstart;
7478: const PetscInt re = C->rmap->rend;
7479: const PetscInt *ii = mm->i;
7480: for (i = 0; i < mr; i++) {
7481: const PetscInt gr = rmap[i];
7482: const PetscInt nz = ii[i + 1] - ii[i];
7483: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7484: else ncoo_oown += nz; /* this row is local */
7485: }
7486: } else ncoo_d += mm->nz;
7487: }
7489: /*
7490: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7492: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7494: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7496: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7497: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7498: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7500: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7501: 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.
7502: */
7503: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7504: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7506: /* gather (i,j) of nonzeros inserted by remote procs */
7507: if (hasoffproc) {
7508: PetscSF msf;
7509: PetscInt ncoo2, *coo_i2, *coo_j2;
7511: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7512: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7513: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7515: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7516: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7517: PetscInt *idxoff = mmdata->off[cp];
7518: PetscInt *idxown = mmdata->own[cp];
7519: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7520: const PetscInt *rmap = rmapa[cp];
7521: const PetscInt *cmap = cmapa[cp];
7522: const PetscInt *ii = mm->i;
7523: PetscInt *coi = coo_i + ncoo_o;
7524: PetscInt *coj = coo_j + ncoo_o;
7525: const PetscInt mr = mp[cp]->rmap->n;
7526: const PetscInt rs = C->rmap->rstart;
7527: const PetscInt re = C->rmap->rend;
7528: const PetscInt cs = C->cmap->rstart;
7529: for (i = 0; i < mr; i++) {
7530: const PetscInt *jj = mm->j + ii[i];
7531: const PetscInt gr = rmap[i];
7532: const PetscInt nz = ii[i + 1] - ii[i];
7533: if (gr < rs || gr >= re) { /* this is an offproc row */
7534: for (j = ii[i]; j < ii[i + 1]; j++) {
7535: *coi++ = gr;
7536: *idxoff++ = j;
7537: }
7538: if (!cmapt[cp]) { /* already global */
7539: for (j = 0; j < nz; j++) *coj++ = jj[j];
7540: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7541: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7542: } else { /* offdiag */
7543: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7544: }
7545: ncoo_o += nz;
7546: } else { /* this is a local row */
7547: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7548: }
7549: }
7550: }
7551: mmdata->off[cp + 1] = idxoff;
7552: mmdata->own[cp + 1] = idxown;
7553: }
7555: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7556: PetscInt incoo_o;
7557: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7558: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7559: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7560: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7561: ncoo = ncoo_d + ncoo_oown + ncoo2;
7562: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7563: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7564: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7565: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7566: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7567: PetscCall(PetscFree2(coo_i, coo_j));
7568: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7569: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7570: coo_i = coo_i2;
7571: coo_j = coo_j2;
7572: } else { /* no offproc values insertion */
7573: ncoo = ncoo_d;
7574: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7576: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7577: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7578: PetscCall(PetscSFSetUp(mmdata->sf));
7579: }
7580: mmdata->hasoffproc = hasoffproc;
7582: /* gather (i,j) of nonzeros inserted locally */
7583: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7584: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7585: PetscInt *coi = coo_i + ncoo_d;
7586: PetscInt *coj = coo_j + ncoo_d;
7587: const PetscInt *jj = mm->j;
7588: const PetscInt *ii = mm->i;
7589: const PetscInt *cmap = cmapa[cp];
7590: const PetscInt *rmap = rmapa[cp];
7591: const PetscInt mr = mp[cp]->rmap->n;
7592: const PetscInt rs = C->rmap->rstart;
7593: const PetscInt re = C->rmap->rend;
7594: const PetscInt cs = C->cmap->rstart;
7596: if (mptmp[cp]) continue;
7597: if (rmapt[cp] == 1) { /* consecutive rows */
7598: /* fill coo_i */
7599: for (i = 0; i < mr; i++) {
7600: const PetscInt gr = i + rs;
7601: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7602: }
7603: /* fill coo_j */
7604: if (!cmapt[cp]) { /* type-0, already global */
7605: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7606: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7607: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7608: } else { /* type-2, local to global for sparse columns */
7609: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7610: }
7611: ncoo_d += mm->nz;
7612: } else if (rmapt[cp] == 2) { /* sparse rows */
7613: for (i = 0; i < mr; i++) {
7614: const PetscInt *jj = mm->j + ii[i];
7615: const PetscInt gr = rmap[i];
7616: const PetscInt nz = ii[i + 1] - ii[i];
7617: if (gr >= rs && gr < re) { /* local rows */
7618: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7619: if (!cmapt[cp]) { /* type-0, already global */
7620: for (j = 0; j < nz; j++) *coj++ = jj[j];
7621: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7622: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7623: } else { /* type-2, local to global for sparse columns */
7624: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7625: }
7626: ncoo_d += nz;
7627: }
7628: }
7629: }
7630: }
7631: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7632: PetscCall(ISDestroy(&glob));
7633: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7634: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7635: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7636: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7638: /* preallocate with COO data */
7639: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7640: PetscCall(PetscFree2(coo_i, coo_j));
7641: PetscFunctionReturn(PETSC_SUCCESS);
7642: }
7644: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7645: {
7646: Mat_Product *product = mat->product;
7647: #if defined(PETSC_HAVE_DEVICE)
7648: PetscBool match = PETSC_FALSE;
7649: PetscBool usecpu = PETSC_FALSE;
7650: #else
7651: PetscBool match = PETSC_TRUE;
7652: #endif
7654: PetscFunctionBegin;
7655: MatCheckProduct(mat, 1);
7656: #if defined(PETSC_HAVE_DEVICE)
7657: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7658: if (match) { /* we can always fallback to the CPU if requested */
7659: switch (product->type) {
7660: case MATPRODUCT_AB:
7661: if (product->api_user) {
7662: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7663: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7664: PetscOptionsEnd();
7665: } else {
7666: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7667: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7668: PetscOptionsEnd();
7669: }
7670: break;
7671: case MATPRODUCT_AtB:
7672: if (product->api_user) {
7673: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7674: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7675: PetscOptionsEnd();
7676: } else {
7677: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7678: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7679: PetscOptionsEnd();
7680: }
7681: break;
7682: case MATPRODUCT_PtAP:
7683: if (product->api_user) {
7684: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7685: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7686: PetscOptionsEnd();
7687: } else {
7688: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7689: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7690: PetscOptionsEnd();
7691: }
7692: break;
7693: default:
7694: break;
7695: }
7696: match = (PetscBool)!usecpu;
7697: }
7698: #endif
7699: if (match) {
7700: switch (product->type) {
7701: case MATPRODUCT_AB:
7702: case MATPRODUCT_AtB:
7703: case MATPRODUCT_PtAP:
7704: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7705: break;
7706: default:
7707: break;
7708: }
7709: }
7710: /* fallback to MPIAIJ ops */
7711: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7712: PetscFunctionReturn(PETSC_SUCCESS);
7713: }
7715: /*
7716: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7718: n - the number of block indices in cc[]
7719: cc - the block indices (must be large enough to contain the indices)
7720: */
7721: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7722: {
7723: PetscInt cnt = -1, nidx, j;
7724: const PetscInt *idx;
7726: PetscFunctionBegin;
7727: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7728: if (nidx) {
7729: cnt = 0;
7730: cc[cnt] = idx[0] / bs;
7731: for (j = 1; j < nidx; j++) {
7732: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7733: }
7734: }
7735: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7736: *n = cnt + 1;
7737: PetscFunctionReturn(PETSC_SUCCESS);
7738: }
7740: /*
7741: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7743: ncollapsed - the number of block indices
7744: collapsed - the block indices (must be large enough to contain the indices)
7745: */
7746: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7747: {
7748: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7750: PetscFunctionBegin;
7751: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7752: for (i = start + 1; i < start + bs; i++) {
7753: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7754: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7755: cprevtmp = cprev;
7756: cprev = merged;
7757: merged = cprevtmp;
7758: }
7759: *ncollapsed = nprev;
7760: if (collapsed) *collapsed = cprev;
7761: PetscFunctionReturn(PETSC_SUCCESS);
7762: }
7764: /*
7765: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7767: Input Parameter:
7768: . Amat - matrix
7769: - symmetrize - make the result symmetric
7770: + scale - scale with diagonal
7772: Output Parameter:
7773: . a_Gmat - output scalar graph >= 0
7775: */
7776: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7777: {
7778: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7779: MPI_Comm comm;
7780: Mat Gmat;
7781: PetscBool ismpiaij, isseqaij;
7782: Mat a, b, c;
7783: MatType jtype;
7785: PetscFunctionBegin;
7786: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7787: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7788: PetscCall(MatGetSize(Amat, &MM, &NN));
7789: PetscCall(MatGetBlockSize(Amat, &bs));
7790: nloc = (Iend - Istart) / bs;
7792: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7793: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7794: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7796: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7797: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7798: implementation */
7799: if (bs > 1) {
7800: PetscCall(MatGetType(Amat, &jtype));
7801: PetscCall(MatCreate(comm, &Gmat));
7802: PetscCall(MatSetType(Gmat, jtype));
7803: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7804: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7805: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7806: PetscInt *d_nnz, *o_nnz;
7807: MatScalar *aa, val, *AA;
7808: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7810: if (isseqaij) {
7811: a = Amat;
7812: b = NULL;
7813: } else {
7814: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7815: a = d->A;
7816: b = d->B;
7817: }
7818: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7819: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7820: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7821: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7822: const PetscInt *cols1, *cols2;
7824: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7825: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7826: nnz[brow / bs] = nc2 / bs;
7827: if (nc2 % bs) ok = 0;
7828: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7829: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7830: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7831: if (nc1 != nc2) ok = 0;
7832: else {
7833: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7834: if (cols1[jj] != cols2[jj]) ok = 0;
7835: if (cols1[jj] % bs != jj % bs) ok = 0;
7836: }
7837: }
7838: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7839: }
7840: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7841: if (!ok) {
7842: PetscCall(PetscFree2(d_nnz, o_nnz));
7843: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7844: goto old_bs;
7845: }
7846: }
7847: }
7848: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7849: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7850: PetscCall(PetscFree2(d_nnz, o_nnz));
7851: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7852: // diag
7853: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7854: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7856: ai = aseq->i;
7857: n = ai[brow + 1] - ai[brow];
7858: aj = aseq->j + ai[brow];
7859: for (PetscInt k = 0; k < n; k += bs) { // block columns
7860: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7861: val = 0;
7862: if (index_size == 0) {
7863: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7864: aa = aseq->a + ai[brow + ii] + k;
7865: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7866: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7867: }
7868: }
7869: } else { // use (index,index) value if provided
7870: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7871: PetscInt ii = index[iii];
7872: aa = aseq->a + ai[brow + ii] + k;
7873: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7874: PetscInt jj = index[jjj];
7875: val += PetscAbs(PetscRealPart(aa[jj]));
7876: }
7877: }
7878: }
7879: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7880: AA[k / bs] = val;
7881: }
7882: grow = Istart / bs + brow / bs;
7883: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7884: }
7885: // off-diag
7886: if (ismpiaij) {
7887: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7888: const PetscScalar *vals;
7889: const PetscInt *cols, *garray = aij->garray;
7891: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7892: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7893: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7894: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7895: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7896: AA[k / bs] = 0;
7897: AJ[cidx] = garray[cols[k]] / bs;
7898: }
7899: nc = ncols / bs;
7900: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7901: if (index_size == 0) {
7902: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7903: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7904: for (PetscInt k = 0; k < ncols; k += bs) {
7905: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7906: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7907: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7908: }
7909: }
7910: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7911: }
7912: } else { // use (index,index) value if provided
7913: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7914: PetscInt ii = index[iii];
7915: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7916: for (PetscInt k = 0; k < ncols; k += bs) {
7917: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7918: PetscInt jj = index[jjj];
7919: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7920: }
7921: }
7922: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7923: }
7924: }
7925: grow = Istart / bs + brow / bs;
7926: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7927: }
7928: }
7929: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7930: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7931: PetscCall(PetscFree2(AA, AJ));
7932: } else {
7933: const PetscScalar *vals;
7934: const PetscInt *idx;
7935: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7936: old_bs:
7937: /*
7938: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7939: */
7940: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7941: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7942: if (isseqaij) {
7943: PetscInt max_d_nnz;
7945: /*
7946: Determine exact preallocation count for (sequential) scalar matrix
7947: */
7948: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7949: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7950: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7951: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7952: PetscCall(PetscFree3(w0, w1, w2));
7953: } else if (ismpiaij) {
7954: Mat Daij, Oaij;
7955: const PetscInt *garray;
7956: PetscInt max_d_nnz;
7958: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7959: /*
7960: Determine exact preallocation count for diagonal block portion of scalar matrix
7961: */
7962: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7963: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7964: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7965: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7966: PetscCall(PetscFree3(w0, w1, w2));
7967: /*
7968: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7969: */
7970: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7971: o_nnz[jj] = 0;
7972: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7973: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7974: o_nnz[jj] += ncols;
7975: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7976: }
7977: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7978: }
7979: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7980: /* get scalar copy (norms) of matrix */
7981: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7982: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7983: PetscCall(PetscFree2(d_nnz, o_nnz));
7984: for (Ii = Istart; Ii < Iend; Ii++) {
7985: PetscInt dest_row = Ii / bs;
7987: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7988: for (jj = 0; jj < ncols; jj++) {
7989: PetscInt dest_col = idx[jj] / bs;
7990: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7992: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7993: }
7994: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7995: }
7996: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7997: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7998: }
7999: } else {
8000: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8001: else {
8002: Gmat = Amat;
8003: PetscCall(PetscObjectReference((PetscObject)Gmat));
8004: }
8005: if (isseqaij) {
8006: a = Gmat;
8007: b = NULL;
8008: } else {
8009: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8010: a = d->A;
8011: b = d->B;
8012: }
8013: if (filter >= 0 || scale) {
8014: /* take absolute value of each entry */
8015: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8016: MatInfo info;
8017: PetscScalar *avals;
8019: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8020: PetscCall(MatSeqAIJGetArray(c, &avals));
8021: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8022: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8023: }
8024: }
8025: }
8026: if (symmetrize) {
8027: PetscBool isset, issym;
8029: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8030: if (!isset || !issym) {
8031: Mat matTrans;
8033: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8034: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8035: PetscCall(MatDestroy(&matTrans));
8036: }
8037: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8038: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8039: if (scale) {
8040: /* scale c for all diagonal values = 1 or -1 */
8041: Vec diag;
8043: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8044: PetscCall(MatGetDiagonal(Gmat, diag));
8045: PetscCall(VecReciprocal(diag));
8046: PetscCall(VecSqrtAbs(diag));
8047: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8048: PetscCall(VecDestroy(&diag));
8049: }
8050: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8051: if (filter >= 0) {
8052: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8053: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8054: }
8055: *a_Gmat = Gmat;
8056: PetscFunctionReturn(PETSC_SUCCESS);
8057: }
8059: /*
8060: Special version for direct calls from Fortran
8061: */
8063: /* Change these macros so can be used in void function */
8064: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8065: #undef PetscCall
8066: #define PetscCall(...) \
8067: do { \
8068: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8069: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8070: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8071: return; \
8072: } \
8073: } while (0)
8075: #undef SETERRQ
8076: #define SETERRQ(comm, ierr, ...) \
8077: do { \
8078: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8079: return; \
8080: } while (0)
8082: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8083: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8084: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8085: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8086: #else
8087: #endif
8088: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8089: {
8090: Mat mat = *mmat;
8091: PetscInt m = *mm, n = *mn;
8092: InsertMode addv = *maddv;
8093: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8094: PetscScalar value;
8096: MatCheckPreallocated(mat, 1);
8097: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8098: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8099: {
8100: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8101: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8102: PetscBool roworiented = aij->roworiented;
8104: /* Some Variables required in the macro */
8105: Mat A = aij->A;
8106: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8107: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8108: MatScalar *aa;
8109: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8110: Mat B = aij->B;
8111: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8112: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8113: MatScalar *ba;
8114: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8115: * cannot use "#if defined" inside a macro. */
8116: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8118: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8119: PetscInt nonew = a->nonew;
8120: MatScalar *ap1, *ap2;
8122: PetscFunctionBegin;
8123: PetscCall(MatSeqAIJGetArray(A, &aa));
8124: PetscCall(MatSeqAIJGetArray(B, &ba));
8125: for (i = 0; i < m; i++) {
8126: if (im[i] < 0) continue;
8127: 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);
8128: if (im[i] >= rstart && im[i] < rend) {
8129: row = im[i] - rstart;
8130: lastcol1 = -1;
8131: rp1 = aj + ai[row];
8132: ap1 = aa + ai[row];
8133: rmax1 = aimax[row];
8134: nrow1 = ailen[row];
8135: low1 = 0;
8136: high1 = nrow1;
8137: lastcol2 = -1;
8138: rp2 = bj + bi[row];
8139: ap2 = ba + bi[row];
8140: rmax2 = bimax[row];
8141: nrow2 = bilen[row];
8142: low2 = 0;
8143: high2 = nrow2;
8145: for (j = 0; j < n; j++) {
8146: if (roworiented) value = v[i * n + j];
8147: else value = v[i + j * m];
8148: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8149: if (in[j] >= cstart && in[j] < cend) {
8150: col = in[j] - cstart;
8151: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8152: } else if (in[j] < 0) continue;
8153: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8154: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8155: } else {
8156: if (mat->was_assembled) {
8157: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8158: #if defined(PETSC_USE_CTABLE)
8159: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8160: col--;
8161: #else
8162: col = aij->colmap[in[j]] - 1;
8163: #endif
8164: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8165: PetscCall(MatDisAssemble_MPIAIJ(mat));
8166: col = in[j];
8167: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8168: B = aij->B;
8169: b = (Mat_SeqAIJ *)B->data;
8170: bimax = b->imax;
8171: bi = b->i;
8172: bilen = b->ilen;
8173: bj = b->j;
8174: rp2 = bj + bi[row];
8175: ap2 = ba + bi[row];
8176: rmax2 = bimax[row];
8177: nrow2 = bilen[row];
8178: low2 = 0;
8179: high2 = nrow2;
8180: bm = aij->B->rmap->n;
8181: ba = b->a;
8182: inserted = PETSC_FALSE;
8183: }
8184: } else col = in[j];
8185: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8186: }
8187: }
8188: } else if (!aij->donotstash) {
8189: if (roworiented) {
8190: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8191: } else {
8192: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8193: }
8194: }
8195: }
8196: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8197: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8198: }
8199: PetscFunctionReturnVoid();
8200: }
8202: /* Undefining these here since they were redefined from their original definition above! No
8203: * other PETSc functions should be defined past this point, as it is impossible to recover the
8204: * original definitions */
8205: #undef PetscCall
8206: #undef SETERRQ