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, PETSC_FALSE)); /* 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, PETSC_FALSE));
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));
2950: if (B->assembled || B->was_assembled) PetscCall(MatDisAssemble_MPIAIJ(B, PETSC_TRUE));
2951: else {
2952: #if defined(PETSC_USE_CTABLE)
2953: PetscCall(PetscHMapIDestroy(&b->colmap));
2954: #else
2955: PetscCall(PetscFree(b->colmap));
2956: #endif
2957: PetscCall(PetscFree(b->garray));
2958: PetscCall(VecDestroy(&b->lvec));
2959: }
2960: PetscCall(VecScatterDestroy(&b->Mvctx));
2962: PetscCall(MatResetPreallocation(b->A));
2963: PetscCall(MatResetPreallocation(b->B));
2964: B->preallocated = PETSC_TRUE;
2965: B->was_assembled = PETSC_FALSE;
2966: B->assembled = PETSC_FALSE;
2967: PetscFunctionReturn(PETSC_SUCCESS);
2968: }
2970: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin, MatDuplicateOption cpvalues, Mat *newmat)
2971: {
2972: Mat mat;
2973: Mat_MPIAIJ *a, *oldmat = (Mat_MPIAIJ *)matin->data;
2975: PetscFunctionBegin;
2976: *newmat = NULL;
2977: PetscCall(MatCreate(PetscObjectComm((PetscObject)matin), &mat));
2978: PetscCall(MatSetSizes(mat, matin->rmap->n, matin->cmap->n, matin->rmap->N, matin->cmap->N));
2979: PetscCall(MatSetBlockSizesFromMats(mat, matin, matin));
2980: PetscCall(MatSetType(mat, ((PetscObject)matin)->type_name));
2981: a = (Mat_MPIAIJ *)mat->data;
2983: mat->factortype = matin->factortype;
2984: mat->assembled = matin->assembled;
2985: mat->insertmode = NOT_SET_VALUES;
2987: a->size = oldmat->size;
2988: a->rank = oldmat->rank;
2989: a->donotstash = oldmat->donotstash;
2990: a->roworiented = oldmat->roworiented;
2991: a->rowindices = NULL;
2992: a->rowvalues = NULL;
2993: a->getrowactive = PETSC_FALSE;
2995: PetscCall(PetscLayoutReference(matin->rmap, &mat->rmap));
2996: PetscCall(PetscLayoutReference(matin->cmap, &mat->cmap));
2997: if (matin->hash_active) {
2998: PetscCall(MatSetUp(mat));
2999: } else {
3000: mat->preallocated = matin->preallocated;
3001: if (oldmat->colmap) {
3002: #if defined(PETSC_USE_CTABLE)
3003: PetscCall(PetscHMapIDuplicate(oldmat->colmap, &a->colmap));
3004: #else
3005: PetscCall(PetscMalloc1(mat->cmap->N, &a->colmap));
3006: PetscCall(PetscArraycpy(a->colmap, oldmat->colmap, mat->cmap->N));
3007: #endif
3008: } else a->colmap = NULL;
3009: if (oldmat->garray) {
3010: PetscInt len;
3011: len = oldmat->B->cmap->n;
3012: PetscCall(PetscMalloc1(len + 1, &a->garray));
3013: if (len) PetscCall(PetscArraycpy(a->garray, oldmat->garray, len));
3014: } else a->garray = NULL;
3016: /* It may happen MatDuplicate is called with a non-assembled matrix
3017: In fact, MatDuplicate only requires the matrix to be preallocated
3018: This may happen inside a DMCreateMatrix_Shell */
3019: if (oldmat->lvec) PetscCall(VecDuplicate(oldmat->lvec, &a->lvec));
3020: if (oldmat->Mvctx) {
3021: a->Mvctx = oldmat->Mvctx;
3022: PetscCall(PetscObjectReference((PetscObject)oldmat->Mvctx));
3023: }
3024: PetscCall(MatDuplicate(oldmat->A, cpvalues, &a->A));
3025: PetscCall(MatDuplicate(oldmat->B, cpvalues, &a->B));
3026: }
3027: PetscCall(PetscFunctionListDuplicate(((PetscObject)matin)->qlist, &((PetscObject)mat)->qlist));
3028: *newmat = mat;
3029: PetscFunctionReturn(PETSC_SUCCESS);
3030: }
3032: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3033: {
3034: PetscBool isbinary, ishdf5;
3036: PetscFunctionBegin;
3039: /* force binary viewer to load .info file if it has not yet done so */
3040: PetscCall(PetscViewerSetUp(viewer));
3041: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
3042: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
3043: if (isbinary) {
3044: PetscCall(MatLoad_MPIAIJ_Binary(newMat, viewer));
3045: } else if (ishdf5) {
3046: #if defined(PETSC_HAVE_HDF5)
3047: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
3048: #else
3049: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
3050: #endif
3051: } else {
3052: 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);
3053: }
3054: PetscFunctionReturn(PETSC_SUCCESS);
3055: }
3057: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
3058: {
3059: PetscInt header[4], M, N, m, nz, rows, cols, sum, i;
3060: PetscInt *rowidxs, *colidxs;
3061: PetscScalar *matvals;
3063: PetscFunctionBegin;
3064: PetscCall(PetscViewerSetUp(viewer));
3066: /* read in matrix header */
3067: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
3068: PetscCheck(header[0] == MAT_FILE_CLASSID, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
3069: M = header[1];
3070: N = header[2];
3071: nz = header[3];
3072: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
3073: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
3074: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as MPIAIJ");
3076: /* set block sizes from the viewer's .info file */
3077: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
3078: /* set global sizes if not set already */
3079: if (mat->rmap->N < 0) mat->rmap->N = M;
3080: if (mat->cmap->N < 0) mat->cmap->N = N;
3081: PetscCall(PetscLayoutSetUp(mat->rmap));
3082: PetscCall(PetscLayoutSetUp(mat->cmap));
3084: /* check if the matrix sizes are correct */
3085: PetscCall(MatGetSize(mat, &rows, &cols));
3086: 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);
3088: /* read in row lengths and build row indices */
3089: PetscCall(MatGetLocalSize(mat, &m, NULL));
3090: PetscCall(PetscMalloc1(m + 1, &rowidxs));
3091: PetscCall(PetscViewerBinaryReadAll(viewer, rowidxs + 1, m, PETSC_DECIDE, M, PETSC_INT));
3092: rowidxs[0] = 0;
3093: for (i = 0; i < m; i++) rowidxs[i + 1] += rowidxs[i];
3094: if (nz != PETSC_INT_MAX) {
3095: PetscCallMPI(MPIU_Allreduce(&rowidxs[m], &sum, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject)viewer)));
3096: 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);
3097: }
3099: /* read in column indices and matrix values */
3100: PetscCall(PetscMalloc2(rowidxs[m], &colidxs, rowidxs[m], &matvals));
3101: PetscCall(PetscViewerBinaryReadAll(viewer, colidxs, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_INT));
3102: PetscCall(PetscViewerBinaryReadAll(viewer, matvals, rowidxs[m], PETSC_DETERMINE, PETSC_DETERMINE, PETSC_SCALAR));
3103: /* store matrix indices and values */
3104: PetscCall(MatMPIAIJSetPreallocationCSR(mat, rowidxs, colidxs, matvals));
3105: PetscCall(PetscFree(rowidxs));
3106: PetscCall(PetscFree2(colidxs, matvals));
3107: PetscFunctionReturn(PETSC_SUCCESS);
3108: }
3110: /* Not scalable because of ISAllGather() unless getting all columns. */
3111: static PetscErrorCode ISGetSeqIS_Private(Mat mat, IS iscol, IS *isseq)
3112: {
3113: IS iscol_local;
3114: PetscBool isstride;
3115: PetscMPIInt lisstride = 0, gisstride;
3117: PetscFunctionBegin;
3118: /* check if we are grabbing all columns*/
3119: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &isstride));
3121: if (isstride) {
3122: PetscInt start, len, mstart, mlen;
3123: PetscCall(ISStrideGetInfo(iscol, &start, NULL));
3124: PetscCall(ISGetLocalSize(iscol, &len));
3125: PetscCall(MatGetOwnershipRangeColumn(mat, &mstart, &mlen));
3126: if (mstart == start && mlen - mstart == len) lisstride = 1;
3127: }
3129: PetscCallMPI(MPIU_Allreduce(&lisstride, &gisstride, 1, MPI_INT, MPI_MIN, PetscObjectComm((PetscObject)mat)));
3130: if (gisstride) {
3131: PetscInt N;
3132: PetscCall(MatGetSize(mat, NULL, &N));
3133: PetscCall(ISCreateStride(PETSC_COMM_SELF, N, 0, 1, &iscol_local));
3134: PetscCall(ISSetIdentity(iscol_local));
3135: PetscCall(PetscInfo(mat, "Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n"));
3136: } else {
3137: PetscInt cbs;
3138: PetscCall(ISGetBlockSize(iscol, &cbs));
3139: PetscCall(ISAllGather(iscol, &iscol_local));
3140: PetscCall(ISSetBlockSize(iscol_local, cbs));
3141: }
3143: *isseq = iscol_local;
3144: PetscFunctionReturn(PETSC_SUCCESS);
3145: }
3147: /*
3148: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3149: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3151: Input Parameters:
3152: + mat - matrix
3153: . isrow - parallel row index set; its local indices are a subset of local columns of `mat`,
3154: i.e., mat->rstart <= isrow[i] < mat->rend
3155: - iscol - parallel column index set; its local indices are a subset of local columns of `mat`,
3156: i.e., mat->cstart <= iscol[i] < mat->cend
3158: Output Parameters:
3159: + isrow_d - sequential row index set for retrieving mat->A
3160: . iscol_d - sequential column index set for retrieving mat->A
3161: . iscol_o - sequential column index set for retrieving mat->B
3162: - garray - column map; garray[i] indicates global location of iscol_o[i] in `iscol`
3163: */
3164: static PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat, IS isrow, IS iscol, IS *isrow_d, IS *iscol_d, IS *iscol_o, const PetscInt *garray[])
3165: {
3166: Vec x, cmap;
3167: const PetscInt *is_idx;
3168: PetscScalar *xarray, *cmaparray;
3169: PetscInt ncols, isstart, *idx, m, rstart, *cmap1, count;
3170: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3171: Mat B = a->B;
3172: Vec lvec = a->lvec, lcmap;
3173: PetscInt i, cstart, cend, Bn = B->cmap->N;
3174: MPI_Comm comm;
3175: VecScatter Mvctx = a->Mvctx;
3177: PetscFunctionBegin;
3178: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3179: PetscCall(ISGetLocalSize(iscol, &ncols));
3181: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3182: PetscCall(MatCreateVecs(mat, &x, NULL));
3183: PetscCall(VecSet(x, -1.0));
3184: PetscCall(VecDuplicate(x, &cmap));
3185: PetscCall(VecSet(cmap, -1.0));
3187: /* Get start indices */
3188: PetscCallMPI(MPI_Scan(&ncols, &isstart, 1, MPIU_INT, MPI_SUM, comm));
3189: isstart -= ncols;
3190: PetscCall(MatGetOwnershipRangeColumn(mat, &cstart, &cend));
3192: PetscCall(ISGetIndices(iscol, &is_idx));
3193: PetscCall(VecGetArray(x, &xarray));
3194: PetscCall(VecGetArray(cmap, &cmaparray));
3195: PetscCall(PetscMalloc1(ncols, &idx));
3196: for (i = 0; i < ncols; i++) {
3197: xarray[is_idx[i] - cstart] = (PetscScalar)is_idx[i];
3198: cmaparray[is_idx[i] - cstart] = i + isstart; /* global index of iscol[i] */
3199: idx[i] = is_idx[i] - cstart; /* local index of iscol[i] */
3200: }
3201: PetscCall(VecRestoreArray(x, &xarray));
3202: PetscCall(VecRestoreArray(cmap, &cmaparray));
3203: PetscCall(ISRestoreIndices(iscol, &is_idx));
3205: /* Get iscol_d */
3206: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, iscol_d));
3207: PetscCall(ISGetBlockSize(iscol, &i));
3208: PetscCall(ISSetBlockSize(*iscol_d, i));
3210: /* Get isrow_d */
3211: PetscCall(ISGetLocalSize(isrow, &m));
3212: rstart = mat->rmap->rstart;
3213: PetscCall(PetscMalloc1(m, &idx));
3214: PetscCall(ISGetIndices(isrow, &is_idx));
3215: for (i = 0; i < m; i++) idx[i] = is_idx[i] - rstart;
3216: PetscCall(ISRestoreIndices(isrow, &is_idx));
3218: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, m, idx, PETSC_OWN_POINTER, isrow_d));
3219: PetscCall(ISGetBlockSize(isrow, &i));
3220: PetscCall(ISSetBlockSize(*isrow_d, i));
3222: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3223: PetscCall(VecScatterBegin(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3224: PetscCall(VecScatterEnd(Mvctx, x, lvec, INSERT_VALUES, SCATTER_FORWARD));
3226: PetscCall(VecDuplicate(lvec, &lcmap));
3228: PetscCall(VecScatterBegin(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3229: PetscCall(VecScatterEnd(Mvctx, cmap, lcmap, INSERT_VALUES, SCATTER_FORWARD));
3231: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3232: /* off-process column indices */
3233: count = 0;
3234: PetscCall(PetscMalloc1(Bn, &idx));
3235: PetscCall(PetscMalloc1(Bn, &cmap1));
3237: PetscCall(VecGetArray(lvec, &xarray));
3238: PetscCall(VecGetArray(lcmap, &cmaparray));
3239: for (i = 0; i < Bn; i++) {
3240: if (PetscRealPart(xarray[i]) > -1.0) {
3241: idx[count] = i; /* local column index in off-diagonal part B */
3242: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3243: count++;
3244: }
3245: }
3246: PetscCall(VecRestoreArray(lvec, &xarray));
3247: PetscCall(VecRestoreArray(lcmap, &cmaparray));
3249: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_COPY_VALUES, iscol_o));
3250: /* cannot ensure iscol_o has same blocksize as iscol! */
3252: PetscCall(PetscFree(idx));
3253: *garray = cmap1;
3255: PetscCall(VecDestroy(&x));
3256: PetscCall(VecDestroy(&cmap));
3257: PetscCall(VecDestroy(&lcmap));
3258: PetscFunctionReturn(PETSC_SUCCESS);
3259: }
3261: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3262: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *submat)
3263: {
3264: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data, *asub;
3265: Mat M = NULL;
3266: MPI_Comm comm;
3267: IS iscol_d, isrow_d, iscol_o;
3268: Mat Asub = NULL, Bsub = NULL;
3269: PetscInt n;
3271: PetscFunctionBegin;
3272: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3274: if (call == MAT_REUSE_MATRIX) {
3275: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3276: PetscCall(PetscObjectQuery((PetscObject)*submat, "isrow_d", (PetscObject *)&isrow_d));
3277: PetscCheck(isrow_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "isrow_d passed in was not used before, cannot reuse");
3279: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_d", (PetscObject *)&iscol_d));
3280: PetscCheck(iscol_d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_d passed in was not used before, cannot reuse");
3282: PetscCall(PetscObjectQuery((PetscObject)*submat, "iscol_o", (PetscObject *)&iscol_o));
3283: PetscCheck(iscol_o, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "iscol_o passed in was not used before, cannot reuse");
3285: /* Update diagonal and off-diagonal portions of submat */
3286: asub = (Mat_MPIAIJ *)(*submat)->data;
3287: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->A));
3288: PetscCall(ISGetLocalSize(iscol_o, &n));
3289: if (n) PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_REUSE_MATRIX, &asub->B));
3290: PetscCall(MatAssemblyBegin(*submat, MAT_FINAL_ASSEMBLY));
3291: PetscCall(MatAssemblyEnd(*submat, MAT_FINAL_ASSEMBLY));
3293: } else { /* call == MAT_INITIAL_MATRIX) */
3294: const PetscInt *garray;
3295: PetscInt BsubN;
3297: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3298: PetscCall(ISGetSeqIS_SameColDist_Private(mat, isrow, iscol, &isrow_d, &iscol_d, &iscol_o, &garray));
3300: /* Create local submatrices Asub and Bsub */
3301: PetscCall(MatCreateSubMatrix_SeqAIJ(a->A, isrow_d, iscol_d, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Asub));
3302: PetscCall(MatCreateSubMatrix_SeqAIJ(a->B, isrow_d, iscol_o, PETSC_DECIDE, MAT_INITIAL_MATRIX, &Bsub));
3304: /* Create submatrix M */
3305: PetscCall(MatCreateMPIAIJWithSeqAIJ(comm, Asub, Bsub, garray, &M));
3307: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3308: asub = (Mat_MPIAIJ *)M->data;
3310: PetscCall(ISGetLocalSize(iscol_o, &BsubN));
3311: n = asub->B->cmap->N;
3312: if (BsubN > n) {
3313: /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3314: const PetscInt *idx;
3315: PetscInt i, j, *idx_new, *subgarray = asub->garray;
3316: PetscCall(PetscInfo(M, "submatrix Bn %" PetscInt_FMT " != BsubN %" PetscInt_FMT ", update iscol_o\n", n, BsubN));
3318: PetscCall(PetscMalloc1(n, &idx_new));
3319: j = 0;
3320: PetscCall(ISGetIndices(iscol_o, &idx));
3321: for (i = 0; i < n; i++) {
3322: if (j >= BsubN) break;
3323: while (subgarray[i] > garray[j]) j++;
3325: if (subgarray[i] == garray[j]) {
3326: idx_new[i] = idx[j++];
3327: } 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]);
3328: }
3329: PetscCall(ISRestoreIndices(iscol_o, &idx));
3331: PetscCall(ISDestroy(&iscol_o));
3332: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, n, idx_new, PETSC_OWN_POINTER, &iscol_o));
3334: } else if (BsubN < n) {
3335: 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);
3336: }
3338: PetscCall(PetscFree(garray));
3339: *submat = M;
3341: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3342: PetscCall(PetscObjectCompose((PetscObject)M, "isrow_d", (PetscObject)isrow_d));
3343: PetscCall(ISDestroy(&isrow_d));
3345: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_d", (PetscObject)iscol_d));
3346: PetscCall(ISDestroy(&iscol_d));
3348: PetscCall(PetscObjectCompose((PetscObject)M, "iscol_o", (PetscObject)iscol_o));
3349: PetscCall(ISDestroy(&iscol_o));
3350: }
3351: PetscFunctionReturn(PETSC_SUCCESS);
3352: }
3354: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat, IS isrow, IS iscol, MatReuse call, Mat *newmat)
3355: {
3356: IS iscol_local = NULL, isrow_d;
3357: PetscInt csize;
3358: PetscInt n, i, j, start, end;
3359: PetscBool sameRowDist = PETSC_FALSE, sameDist[2], tsameDist[2];
3360: MPI_Comm comm;
3362: PetscFunctionBegin;
3363: /* If isrow has same processor distribution as mat,
3364: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3365: if (call == MAT_REUSE_MATRIX) {
3366: PetscCall(PetscObjectQuery((PetscObject)*newmat, "isrow_d", (PetscObject *)&isrow_d));
3367: if (isrow_d) {
3368: sameRowDist = PETSC_TRUE;
3369: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3370: } else {
3371: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_local));
3372: if (iscol_local) {
3373: sameRowDist = PETSC_TRUE;
3374: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3375: }
3376: }
3377: } else {
3378: /* Check if isrow has same processor distribution as mat */
3379: sameDist[0] = PETSC_FALSE;
3380: PetscCall(ISGetLocalSize(isrow, &n));
3381: if (!n) {
3382: sameDist[0] = PETSC_TRUE;
3383: } else {
3384: PetscCall(ISGetMinMax(isrow, &i, &j));
3385: PetscCall(MatGetOwnershipRange(mat, &start, &end));
3386: if (i >= start && j < end) sameDist[0] = PETSC_TRUE;
3387: }
3389: /* Check if iscol has same processor distribution as mat */
3390: sameDist[1] = PETSC_FALSE;
3391: PetscCall(ISGetLocalSize(iscol, &n));
3392: if (!n) {
3393: sameDist[1] = PETSC_TRUE;
3394: } else {
3395: PetscCall(ISGetMinMax(iscol, &i, &j));
3396: PetscCall(MatGetOwnershipRangeColumn(mat, &start, &end));
3397: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3398: }
3400: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3401: PetscCallMPI(MPIU_Allreduce(&sameDist, &tsameDist, 2, MPIU_BOOL, MPI_LAND, comm));
3402: sameRowDist = tsameDist[0];
3403: }
3405: if (sameRowDist) {
3406: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3407: /* isrow and iscol have same processor distribution as mat */
3408: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat, isrow, iscol, call, newmat));
3409: PetscFunctionReturn(PETSC_SUCCESS);
3410: } else { /* sameRowDist */
3411: /* isrow has same processor distribution as mat */
3412: if (call == MAT_INITIAL_MATRIX) {
3413: PetscBool sorted;
3414: PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3415: PetscCall(ISGetLocalSize(iscol_local, &n)); /* local size of iscol_local = global columns of newmat */
3416: PetscCall(ISGetSize(iscol, &i));
3417: PetscCheck(n == i, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "n %" PetscInt_FMT " != size of iscol %" PetscInt_FMT, n, i);
3419: PetscCall(ISSorted(iscol_local, &sorted));
3420: if (sorted) {
3421: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3422: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, iscol_local, MAT_INITIAL_MATRIX, newmat));
3423: PetscFunctionReturn(PETSC_SUCCESS);
3424: }
3425: } else { /* call == MAT_REUSE_MATRIX */
3426: IS iscol_sub;
3427: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3428: if (iscol_sub) {
3429: PetscCall(MatCreateSubMatrix_MPIAIJ_SameRowDist(mat, isrow, iscol, NULL, call, newmat));
3430: PetscFunctionReturn(PETSC_SUCCESS);
3431: }
3432: }
3433: }
3434: }
3436: /* General case: iscol -> iscol_local which has global size of iscol */
3437: if (call == MAT_REUSE_MATRIX) {
3438: PetscCall(PetscObjectQuery((PetscObject)*newmat, "ISAllGather", (PetscObject *)&iscol_local));
3439: PetscCheck(iscol_local, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3440: } else {
3441: if (!iscol_local) PetscCall(ISGetSeqIS_Private(mat, iscol, &iscol_local));
3442: }
3444: PetscCall(ISGetLocalSize(iscol, &csize));
3445: PetscCall(MatCreateSubMatrix_MPIAIJ_nonscalable(mat, isrow, iscol_local, csize, call, newmat));
3447: if (call == MAT_INITIAL_MATRIX) {
3448: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3449: PetscCall(ISDestroy(&iscol_local));
3450: }
3451: PetscFunctionReturn(PETSC_SUCCESS);
3452: }
3454: /*@C
3455: MatCreateMPIAIJWithSeqAIJ - creates a `MATMPIAIJ` matrix using `MATSEQAIJ` matrices that contain the "diagonal"
3456: and "off-diagonal" part of the matrix in CSR format.
3458: Collective
3460: Input Parameters:
3461: + comm - MPI communicator
3462: . A - "diagonal" portion of matrix
3463: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3464: - garray - global index of `B` columns
3466: Output Parameter:
3467: . mat - the matrix, with input `A` as its local diagonal matrix
3469: Level: advanced
3471: Notes:
3472: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3474: `A` becomes part of output mat, `B` is destroyed by this routine. The user cannot use `A` and `B` anymore.
3476: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MATSEQAIJ`, `MatCreateMPIAIJWithSplitArrays()`
3477: @*/
3478: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm, Mat A, Mat B, const PetscInt garray[], Mat *mat)
3479: {
3480: Mat_MPIAIJ *maij;
3481: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data, *bnew;
3482: PetscInt *oi = b->i, *oj = b->j, i, nz, col;
3483: const PetscScalar *oa;
3484: Mat Bnew;
3485: PetscInt m, n, N;
3486: MatType mpi_mat_type;
3488: PetscFunctionBegin;
3489: PetscCall(MatCreate(comm, mat));
3490: PetscCall(MatGetSize(A, &m, &n));
3491: PetscCheck(m == B->rmap->N, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Am %" PetscInt_FMT " != Bm %" PetscInt_FMT, m, B->rmap->N);
3492: 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);
3493: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3494: /* 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); */
3496: /* Get global columns of mat */
3497: PetscCallMPI(MPIU_Allreduce(&n, &N, 1, MPIU_INT, MPI_SUM, comm));
3499: PetscCall(MatSetSizes(*mat, m, n, PETSC_DECIDE, N));
3500: /* Determine the type of MPI matrix that should be created from the type of matrix A, which holds the "diagonal" portion. */
3501: PetscCall(MatGetMPIMatType_Private(A, &mpi_mat_type));
3502: PetscCall(MatSetType(*mat, mpi_mat_type));
3504: if (A->rmap->bs > 1 || A->cmap->bs > 1) PetscCall(MatSetBlockSizes(*mat, A->rmap->bs, A->cmap->bs));
3505: maij = (Mat_MPIAIJ *)(*mat)->data;
3507: (*mat)->preallocated = PETSC_TRUE;
3509: PetscCall(PetscLayoutSetUp((*mat)->rmap));
3510: PetscCall(PetscLayoutSetUp((*mat)->cmap));
3512: /* Set A as diagonal portion of *mat */
3513: maij->A = A;
3515: nz = oi[m];
3516: for (i = 0; i < nz; i++) {
3517: col = oj[i];
3518: oj[i] = garray[col];
3519: }
3521: /* Set Bnew as off-diagonal portion of *mat */
3522: PetscCall(MatSeqAIJGetArrayRead(B, &oa));
3523: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, N, oi, oj, (PetscScalar *)oa, &Bnew));
3524: PetscCall(MatSeqAIJRestoreArrayRead(B, &oa));
3525: bnew = (Mat_SeqAIJ *)Bnew->data;
3526: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3527: maij->B = Bnew;
3529: 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);
3531: b->free_a = PETSC_FALSE;
3532: b->free_ij = PETSC_FALSE;
3533: PetscCall(MatDestroy(&B));
3535: bnew->free_a = PETSC_TRUE;
3536: bnew->free_ij = PETSC_TRUE;
3538: /* condense columns of maij->B */
3539: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
3540: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
3541: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
3542: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
3543: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3544: PetscFunctionReturn(PETSC_SUCCESS);
3545: }
3547: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat, PetscInt, const IS[], const IS[], MatReuse, PetscBool, Mat *);
3549: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat, IS isrow, IS iscol, IS iscol_local, MatReuse call, Mat *newmat)
3550: {
3551: PetscInt i, m, n, rstart, row, rend, nz, j, bs, cbs;
3552: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3553: Mat_MPIAIJ *a = (Mat_MPIAIJ *)mat->data;
3554: Mat M, Msub, B = a->B;
3555: MatScalar *aa;
3556: Mat_SeqAIJ *aij;
3557: PetscInt *garray = a->garray, *colsub, Ncols;
3558: PetscInt count, Bn = B->cmap->N, cstart = mat->cmap->rstart, cend = mat->cmap->rend;
3559: IS iscol_sub, iscmap;
3560: const PetscInt *is_idx, *cmap;
3561: PetscBool allcolumns = PETSC_FALSE;
3562: MPI_Comm comm;
3564: PetscFunctionBegin;
3565: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3566: if (call == MAT_REUSE_MATRIX) {
3567: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubIScol", (PetscObject *)&iscol_sub));
3568: PetscCheck(iscol_sub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "SubIScol passed in was not used before, cannot reuse");
3569: PetscCall(ISGetLocalSize(iscol_sub, &count));
3571: PetscCall(PetscObjectQuery((PetscObject)*newmat, "Subcmap", (PetscObject *)&iscmap));
3572: PetscCheck(iscmap, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Subcmap passed in was not used before, cannot reuse");
3574: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Msub));
3575: PetscCheck(Msub, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3577: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_REUSE_MATRIX, PETSC_FALSE, &Msub));
3579: } else { /* call == MAT_INITIAL_MATRIX) */
3580: PetscBool flg;
3582: PetscCall(ISGetLocalSize(iscol, &n));
3583: PetscCall(ISGetSize(iscol, &Ncols));
3585: /* (1) iscol -> nonscalable iscol_local */
3586: /* Check for special case: each processor gets entire matrix columns */
3587: PetscCall(ISIdentity(iscol_local, &flg));
3588: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3589: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3590: if (allcolumns) {
3591: iscol_sub = iscol_local;
3592: PetscCall(PetscObjectReference((PetscObject)iscol_local));
3593: PetscCall(ISCreateStride(PETSC_COMM_SELF, n, 0, 1, &iscmap));
3595: } else {
3596: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3597: PetscInt *idx, *cmap1, k;
3598: PetscCall(PetscMalloc1(Ncols, &idx));
3599: PetscCall(PetscMalloc1(Ncols, &cmap1));
3600: PetscCall(ISGetIndices(iscol_local, &is_idx));
3601: count = 0;
3602: k = 0;
3603: for (i = 0; i < Ncols; i++) {
3604: j = is_idx[i];
3605: if (j >= cstart && j < cend) {
3606: /* diagonal part of mat */
3607: idx[count] = j;
3608: cmap1[count++] = i; /* column index in submat */
3609: } else if (Bn) {
3610: /* off-diagonal part of mat */
3611: if (j == garray[k]) {
3612: idx[count] = j;
3613: cmap1[count++] = i; /* column index in submat */
3614: } else if (j > garray[k]) {
3615: while (j > garray[k] && k < Bn - 1) k++;
3616: if (j == garray[k]) {
3617: idx[count] = j;
3618: cmap1[count++] = i; /* column index in submat */
3619: }
3620: }
3621: }
3622: }
3623: PetscCall(ISRestoreIndices(iscol_local, &is_idx));
3625: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, count, idx, PETSC_OWN_POINTER, &iscol_sub));
3626: PetscCall(ISGetBlockSize(iscol, &cbs));
3627: PetscCall(ISSetBlockSize(iscol_sub, cbs));
3629: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local), count, cmap1, PETSC_OWN_POINTER, &iscmap));
3630: }
3632: /* (3) Create sequential Msub */
3633: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol_sub, MAT_INITIAL_MATRIX, allcolumns, &Msub));
3634: }
3636: PetscCall(ISGetLocalSize(iscol_sub, &count));
3637: aij = (Mat_SeqAIJ *)Msub->data;
3638: ii = aij->i;
3639: PetscCall(ISGetIndices(iscmap, &cmap));
3641: /*
3642: m - number of local rows
3643: Ncols - number of columns (same on all processors)
3644: rstart - first row in new global matrix generated
3645: */
3646: PetscCall(MatGetSize(Msub, &m, NULL));
3648: if (call == MAT_INITIAL_MATRIX) {
3649: /* (4) Create parallel newmat */
3650: PetscMPIInt rank, size;
3651: PetscInt csize;
3653: PetscCallMPI(MPI_Comm_size(comm, &size));
3654: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3656: /*
3657: Determine the number of non-zeros in the diagonal and off-diagonal
3658: portions of the matrix in order to do correct preallocation
3659: */
3661: /* first get start and end of "diagonal" columns */
3662: PetscCall(ISGetLocalSize(iscol, &csize));
3663: if (csize == PETSC_DECIDE) {
3664: PetscCall(ISGetSize(isrow, &mglobal));
3665: if (mglobal == Ncols) { /* square matrix */
3666: nlocal = m;
3667: } else {
3668: nlocal = Ncols / size + ((Ncols % size) > rank);
3669: }
3670: } else {
3671: nlocal = csize;
3672: }
3673: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3674: rstart = rend - nlocal;
3675: PetscCheck(rank != size - 1 || rend == Ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, Ncols);
3677: /* next, compute all the lengths */
3678: jj = aij->j;
3679: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3680: olens = dlens + m;
3681: for (i = 0; i < m; i++) {
3682: jend = ii[i + 1] - ii[i];
3683: olen = 0;
3684: dlen = 0;
3685: for (j = 0; j < jend; j++) {
3686: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3687: else dlen++;
3688: jj++;
3689: }
3690: olens[i] = olen;
3691: dlens[i] = dlen;
3692: }
3694: PetscCall(ISGetBlockSize(isrow, &bs));
3695: PetscCall(ISGetBlockSize(iscol, &cbs));
3697: PetscCall(MatCreate(comm, &M));
3698: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, Ncols));
3699: PetscCall(MatSetBlockSizes(M, bs, cbs));
3700: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3701: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3702: PetscCall(PetscFree(dlens));
3704: } else { /* call == MAT_REUSE_MATRIX */
3705: M = *newmat;
3706: PetscCall(MatGetLocalSize(M, &i, NULL));
3707: PetscCheck(i == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3708: PetscCall(MatZeroEntries(M));
3709: /*
3710: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3711: rather than the slower MatSetValues().
3712: */
3713: M->was_assembled = PETSC_TRUE;
3714: M->assembled = PETSC_FALSE;
3715: }
3717: /* (5) Set values of Msub to *newmat */
3718: PetscCall(PetscMalloc1(count, &colsub));
3719: PetscCall(MatGetOwnershipRange(M, &rstart, NULL));
3721: jj = aij->j;
3722: PetscCall(MatSeqAIJGetArrayRead(Msub, (const PetscScalar **)&aa));
3723: for (i = 0; i < m; i++) {
3724: row = rstart + i;
3725: nz = ii[i + 1] - ii[i];
3726: for (j = 0; j < nz; j++) colsub[j] = cmap[jj[j]];
3727: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, colsub, aa, INSERT_VALUES));
3728: jj += nz;
3729: aa += nz;
3730: }
3731: PetscCall(MatSeqAIJRestoreArrayRead(Msub, (const PetscScalar **)&aa));
3732: PetscCall(ISRestoreIndices(iscmap, &cmap));
3734: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3735: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3737: PetscCall(PetscFree(colsub));
3739: /* save Msub, iscol_sub and iscmap used in processor for next request */
3740: if (call == MAT_INITIAL_MATRIX) {
3741: *newmat = M;
3742: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubMatrix", (PetscObject)Msub));
3743: PetscCall(MatDestroy(&Msub));
3745: PetscCall(PetscObjectCompose((PetscObject)*newmat, "SubIScol", (PetscObject)iscol_sub));
3746: PetscCall(ISDestroy(&iscol_sub));
3748: PetscCall(PetscObjectCompose((PetscObject)*newmat, "Subcmap", (PetscObject)iscmap));
3749: PetscCall(ISDestroy(&iscmap));
3751: if (iscol_local) {
3752: PetscCall(PetscObjectCompose((PetscObject)*newmat, "ISAllGather", (PetscObject)iscol_local));
3753: PetscCall(ISDestroy(&iscol_local));
3754: }
3755: }
3756: PetscFunctionReturn(PETSC_SUCCESS);
3757: }
3759: /*
3760: Not great since it makes two copies of the submatrix, first an SeqAIJ
3761: in local and then by concatenating the local matrices the end result.
3762: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3764: This requires a sequential iscol with all indices.
3765: */
3766: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat, IS isrow, IS iscol, PetscInt csize, MatReuse call, Mat *newmat)
3767: {
3768: PetscMPIInt rank, size;
3769: PetscInt i, m, n, rstart, row, rend, nz, *cwork, j, bs, cbs;
3770: PetscInt *ii, *jj, nlocal, *dlens, *olens, dlen, olen, jend, mglobal;
3771: Mat M, Mreuse;
3772: MatScalar *aa, *vwork;
3773: MPI_Comm comm;
3774: Mat_SeqAIJ *aij;
3775: PetscBool colflag, allcolumns = PETSC_FALSE;
3777: PetscFunctionBegin;
3778: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
3779: PetscCallMPI(MPI_Comm_rank(comm, &rank));
3780: PetscCallMPI(MPI_Comm_size(comm, &size));
3782: /* Check for special case: each processor gets entire matrix columns */
3783: PetscCall(ISIdentity(iscol, &colflag));
3784: PetscCall(ISGetLocalSize(iscol, &n));
3785: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3786: PetscCallMPI(MPIU_Allreduce(MPI_IN_PLACE, &allcolumns, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
3788: if (call == MAT_REUSE_MATRIX) {
3789: PetscCall(PetscObjectQuery((PetscObject)*newmat, "SubMatrix", (PetscObject *)&Mreuse));
3790: PetscCheck(Mreuse, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Submatrix passed in was not used before, cannot reuse");
3791: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_REUSE_MATRIX, allcolumns, &Mreuse));
3792: } else {
3793: PetscCall(MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat, 1, &isrow, &iscol, MAT_INITIAL_MATRIX, allcolumns, &Mreuse));
3794: }
3796: /*
3797: m - number of local rows
3798: n - number of columns (same on all processors)
3799: rstart - first row in new global matrix generated
3800: */
3801: PetscCall(MatGetSize(Mreuse, &m, &n));
3802: PetscCall(MatGetBlockSizes(Mreuse, &bs, &cbs));
3803: if (call == MAT_INITIAL_MATRIX) {
3804: aij = (Mat_SeqAIJ *)Mreuse->data;
3805: ii = aij->i;
3806: jj = aij->j;
3808: /*
3809: Determine the number of non-zeros in the diagonal and off-diagonal
3810: portions of the matrix in order to do correct preallocation
3811: */
3813: /* first get start and end of "diagonal" columns */
3814: if (csize == PETSC_DECIDE) {
3815: PetscCall(ISGetSize(isrow, &mglobal));
3816: if (mglobal == n) { /* square matrix */
3817: nlocal = m;
3818: } else {
3819: nlocal = n / size + ((n % size) > rank);
3820: }
3821: } else {
3822: nlocal = csize;
3823: }
3824: PetscCallMPI(MPI_Scan(&nlocal, &rend, 1, MPIU_INT, MPI_SUM, comm));
3825: rstart = rend - nlocal;
3826: PetscCheck(rank != size - 1 || rend == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Local column sizes %" PetscInt_FMT " do not add up to total number of columns %" PetscInt_FMT, rend, n);
3828: /* next, compute all the lengths */
3829: PetscCall(PetscMalloc1(2 * m + 1, &dlens));
3830: olens = dlens + m;
3831: for (i = 0; i < m; i++) {
3832: jend = ii[i + 1] - ii[i];
3833: olen = 0;
3834: dlen = 0;
3835: for (j = 0; j < jend; j++) {
3836: if (*jj < rstart || *jj >= rend) olen++;
3837: else dlen++;
3838: jj++;
3839: }
3840: olens[i] = olen;
3841: dlens[i] = dlen;
3842: }
3843: PetscCall(MatCreate(comm, &M));
3844: PetscCall(MatSetSizes(M, m, nlocal, PETSC_DECIDE, n));
3845: PetscCall(MatSetBlockSizes(M, bs, cbs));
3846: PetscCall(MatSetType(M, ((PetscObject)mat)->type_name));
3847: PetscCall(MatMPIAIJSetPreallocation(M, 0, dlens, 0, olens));
3848: PetscCall(PetscFree(dlens));
3849: } else {
3850: PetscInt ml, nl;
3852: M = *newmat;
3853: PetscCall(MatGetLocalSize(M, &ml, &nl));
3854: PetscCheck(ml == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Previous matrix must be same size/layout as request");
3855: PetscCall(MatZeroEntries(M));
3856: /*
3857: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3858: rather than the slower MatSetValues().
3859: */
3860: M->was_assembled = PETSC_TRUE;
3861: M->assembled = PETSC_FALSE;
3862: }
3863: PetscCall(MatGetOwnershipRange(M, &rstart, &rend));
3864: aij = (Mat_SeqAIJ *)Mreuse->data;
3865: ii = aij->i;
3866: jj = aij->j;
3868: /* trigger copy to CPU if needed */
3869: PetscCall(MatSeqAIJGetArrayRead(Mreuse, (const PetscScalar **)&aa));
3870: for (i = 0; i < m; i++) {
3871: row = rstart + i;
3872: nz = ii[i + 1] - ii[i];
3873: cwork = jj;
3874: jj = PetscSafePointerPlusOffset(jj, nz);
3875: vwork = aa;
3876: aa = PetscSafePointerPlusOffset(aa, nz);
3877: PetscCall(MatSetValues_MPIAIJ(M, 1, &row, nz, cwork, vwork, INSERT_VALUES));
3878: }
3879: PetscCall(MatSeqAIJRestoreArrayRead(Mreuse, (const PetscScalar **)&aa));
3881: PetscCall(MatAssemblyBegin(M, MAT_FINAL_ASSEMBLY));
3882: PetscCall(MatAssemblyEnd(M, MAT_FINAL_ASSEMBLY));
3883: *newmat = M;
3885: /* save submatrix used in processor for next request */
3886: if (call == MAT_INITIAL_MATRIX) {
3887: PetscCall(PetscObjectCompose((PetscObject)M, "SubMatrix", (PetscObject)Mreuse));
3888: PetscCall(MatDestroy(&Mreuse));
3889: }
3890: PetscFunctionReturn(PETSC_SUCCESS);
3891: }
3893: static PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
3894: {
3895: PetscInt m, cstart, cend, j, nnz, i, d, *ld;
3896: PetscInt *d_nnz, *o_nnz, nnz_max = 0, rstart, ii, irstart;
3897: const PetscInt *JJ;
3898: PetscBool nooffprocentries;
3899: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)B->data;
3901: PetscFunctionBegin;
3902: PetscCall(PetscLayoutSetUp(B->rmap));
3903: PetscCall(PetscLayoutSetUp(B->cmap));
3904: m = B->rmap->n;
3905: cstart = B->cmap->rstart;
3906: cend = B->cmap->rend;
3907: rstart = B->rmap->rstart;
3908: irstart = Ii[0];
3910: PetscCall(PetscCalloc2(m, &d_nnz, m, &o_nnz));
3912: if (PetscDefined(USE_DEBUG)) {
3913: for (i = 0; i < m; i++) {
3914: nnz = Ii[i + 1] - Ii[i];
3915: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3916: PetscCheck(nnz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative %" PetscInt_FMT " number of columns", i, nnz);
3917: 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]);
3918: 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);
3919: }
3920: }
3922: for (i = 0; i < m; i++) {
3923: nnz = Ii[i + 1] - Ii[i];
3924: JJ = PetscSafePointerPlusOffset(J, Ii[i] - irstart);
3925: nnz_max = PetscMax(nnz_max, nnz);
3926: d = 0;
3927: for (j = 0; j < nnz; j++) {
3928: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3929: }
3930: d_nnz[i] = d;
3931: o_nnz[i] = nnz - d;
3932: }
3933: PetscCall(MatMPIAIJSetPreallocation(B, 0, d_nnz, 0, o_nnz));
3934: PetscCall(PetscFree2(d_nnz, o_nnz));
3936: for (i = 0; i < m; i++) {
3937: ii = i + rstart;
3938: PetscCall(MatSetValues_MPIAIJ(B, 1, &ii, Ii[i + 1] - Ii[i], PetscSafePointerPlusOffset(J, Ii[i] - irstart), PetscSafePointerPlusOffset(v, Ii[i] - irstart), INSERT_VALUES));
3939: }
3940: nooffprocentries = B->nooffprocentries;
3941: B->nooffprocentries = PETSC_TRUE;
3942: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
3943: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
3944: B->nooffprocentries = nooffprocentries;
3946: /* count number of entries below block diagonal */
3947: PetscCall(PetscFree(Aij->ld));
3948: PetscCall(PetscCalloc1(m, &ld));
3949: Aij->ld = ld;
3950: for (i = 0; i < m; i++) {
3951: nnz = Ii[i + 1] - Ii[i];
3952: j = 0;
3953: while (j < nnz && J[j] < cstart) j++;
3954: ld[i] = j;
3955: if (J) J += nnz;
3956: }
3958: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
3959: PetscFunctionReturn(PETSC_SUCCESS);
3960: }
3962: /*@
3963: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in `MATAIJ` format
3964: (the default parallel PETSc format).
3966: Collective
3968: Input Parameters:
3969: + B - the matrix
3970: . i - the indices into `j` for the start of each local row (indices start with zero)
3971: . j - the column indices for each local row (indices start with zero)
3972: - v - optional values in the matrix
3974: Level: developer
3976: Notes:
3977: The `i`, `j`, and `v` arrays ARE copied by this routine into the internal format used by PETSc;
3978: thus you CANNOT change the matrix entries by changing the values of `v` after you have
3979: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
3981: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
3983: A convenience routine for this functionality is `MatCreateMPIAIJWithArrays()`.
3985: You can update the matrix with new numerical values using `MatUpdateMPIAIJWithArrays()` after this call if the column indices in `j` are sorted.
3987: If you do **not** use `MatUpdateMPIAIJWithArrays()`, the column indices in `j` do not need to be sorted. If you will use
3988: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
3990: The format which is used for the sparse matrix input, is equivalent to a
3991: row-major ordering.. i.e for the following matrix, the input data expected is
3992: as shown
3993: .vb
3994: 1 0 0
3995: 2 0 3 P0
3996: -------
3997: 4 5 6 P1
3999: Process0 [P0] rows_owned=[0,1]
4000: i = {0,1,3} [size = nrow+1 = 2+1]
4001: j = {0,0,2} [size = 3]
4002: v = {1,2,3} [size = 3]
4004: Process1 [P1] rows_owned=[2]
4005: i = {0,3} [size = nrow+1 = 1+1]
4006: j = {0,1,2} [size = 3]
4007: v = {4,5,6} [size = 3]
4008: .ve
4010: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatCreateAIJ()`,
4011: `MatCreateSeqAIJWithArrays()`, `MatCreateMPIAIJWithSplitArrays()`, `MatCreateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4012: @*/
4013: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4014: {
4015: PetscFunctionBegin;
4016: PetscTryMethod(B, "MatMPIAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4017: PetscFunctionReturn(PETSC_SUCCESS);
4018: }
4020: /*@
4021: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in `MATMPIAIJ` format
4022: (the default parallel PETSc format). For good matrix assembly performance
4023: the user should preallocate the matrix storage by setting the parameters
4024: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4026: Collective
4028: Input Parameters:
4029: + B - the matrix
4030: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4031: (same value is used for all local rows)
4032: . d_nnz - array containing the number of nonzeros in the various rows of the
4033: DIAGONAL portion of the local submatrix (possibly different for each row)
4034: or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `d_nz` is used to specify the nonzero structure.
4035: The size of this array is equal to the number of local rows, i.e 'm'.
4036: For matrices that will be factored, you must leave room for (and set)
4037: the diagonal entry even if it is zero.
4038: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4039: submatrix (same value is used for all local rows).
4040: - o_nnz - array containing the number of nonzeros in the various rows of the
4041: OFF-DIAGONAL portion of the local submatrix (possibly different for
4042: each row) or `NULL` (`PETSC_NULL_INTEGER` in Fortran), if `o_nz` is used to specify the nonzero
4043: structure. The size of this array is equal to the number
4044: of local rows, i.e 'm'.
4046: Example Usage:
4047: Consider the following 8x8 matrix with 34 non-zero values, that is
4048: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4049: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4050: as follows
4052: .vb
4053: 1 2 0 | 0 3 0 | 0 4
4054: Proc0 0 5 6 | 7 0 0 | 8 0
4055: 9 0 10 | 11 0 0 | 12 0
4056: -------------------------------------
4057: 13 0 14 | 15 16 17 | 0 0
4058: Proc1 0 18 0 | 19 20 21 | 0 0
4059: 0 0 0 | 22 23 0 | 24 0
4060: -------------------------------------
4061: Proc2 25 26 27 | 0 0 28 | 29 0
4062: 30 0 0 | 31 32 33 | 0 34
4063: .ve
4065: This can be represented as a collection of submatrices as
4066: .vb
4067: A B C
4068: D E F
4069: G H I
4070: .ve
4072: Where the submatrices A,B,C are owned by proc0, D,E,F are
4073: owned by proc1, G,H,I are owned by proc2.
4075: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4076: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4077: The 'M','N' parameters are 8,8, and have the same values on all procs.
4079: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4080: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4081: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4082: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4083: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4084: matrix, and [DF] as another `MATSEQAIJ` matrix.
4086: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4087: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4088: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4089: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4090: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4091: In this case, the values of `d_nz`, `o_nz` are
4092: .vb
4093: proc0 dnz = 2, o_nz = 2
4094: proc1 dnz = 3, o_nz = 2
4095: proc2 dnz = 1, o_nz = 4
4096: .ve
4097: We are allocating `m`*(`d_nz`+`o_nz`) storage locations for every proc. This
4098: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4099: for proc3. i.e we are using 12+15+10=37 storage locations to store
4100: 34 values.
4102: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4103: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4104: In the above case the values for `d_nnz`, `o_nnz` are
4105: .vb
4106: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4107: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4108: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4109: .ve
4110: Here the space allocated is sum of all the above values i.e 34, and
4111: hence pre-allocation is perfect.
4113: Level: intermediate
4115: Notes:
4116: If the *_nnz parameter is given then the *_nz parameter is ignored
4118: The `MATAIJ` format, also called compressed row storage (CSR), is compatible with standard Fortran
4119: storage. The stored row and column indices begin with zero.
4120: See [Sparse Matrices](sec_matsparse) for details.
4122: The parallel matrix is partitioned such that the first m0 rows belong to
4123: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4124: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4126: The DIAGONAL portion of the local submatrix of a processor can be defined
4127: as the submatrix which is obtained by extraction the part corresponding to
4128: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4129: first row that belongs to the processor, r2 is the last row belonging to
4130: the this processor, and c1-c2 is range of indices of the local part of a
4131: vector suitable for applying the matrix to. This is an mxn matrix. In the
4132: common case of a square matrix, the row and column ranges are the same and
4133: the DIAGONAL part is also square. The remaining portion of the local
4134: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4136: If `o_nnz` and `d_nnz` are specified, then `o_nz` and `d_nz` are ignored.
4138: You can call `MatGetInfo()` to get information on how effective the preallocation was;
4139: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4140: You can also run with the option `-info` and look for messages with the string
4141: malloc in them to see if additional memory allocation was needed.
4143: .seealso: [](ch_matrices), `Mat`, [Sparse Matrices](sec_matsparse), `MATMPIAIJ`, `MATAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatCreateAIJ()`, `MatMPIAIJSetPreallocationCSR()`,
4144: `MatGetInfo()`, `PetscSplitOwnership()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4145: @*/
4146: PetscErrorCode MatMPIAIJSetPreallocation(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
4147: {
4148: PetscFunctionBegin;
4151: PetscTryMethod(B, "MatMPIAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[]), (B, d_nz, d_nnz, o_nz, o_nnz));
4152: PetscFunctionReturn(PETSC_SUCCESS);
4153: }
4155: /*@
4156: MatCreateMPIAIJWithArrays - creates a `MATMPIAIJ` matrix using arrays that contain in standard
4157: CSR format for the local rows.
4159: Collective
4161: Input Parameters:
4162: + comm - MPI communicator
4163: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4164: . n - This value should be the same as the local size used in creating the
4165: x vector for the matrix-vector product $ y = Ax$. (or `PETSC_DECIDE` to have
4166: calculated if `N` is given) For square matrices n is almost always `m`.
4167: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
4168: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
4169: . 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
4170: . j - global column indices
4171: - a - optional matrix values
4173: Output Parameter:
4174: . mat - the matrix
4176: Level: intermediate
4178: Notes:
4179: The `i`, `j`, and `a` arrays ARE copied by this routine into the internal format used by PETSc;
4180: thus you CANNOT change the matrix entries by changing the values of `a[]` after you have
4181: called this routine. Use `MatCreateMPIAIJWithSplitArrays()` to avoid needing to copy the arrays.
4183: The `i` and `j` indices are 0 based, and `i` indices are indices corresponding to the local `j` array.
4185: Once you have created the matrix you can update it with new numerical values using `MatUpdateMPIAIJWithArray()`
4187: If you do **not** use `MatUpdateMPIAIJWithArray()`, the column indices in `j` do not need to be sorted. If you will use
4188: `MatUpdateMPIAIJWithArrays()`, the column indices **must** be sorted.
4190: The format which is used for the sparse matrix input, is equivalent to a
4191: row-major ordering, i.e., for the following matrix, the input data expected is
4192: as shown
4193: .vb
4194: 1 0 0
4195: 2 0 3 P0
4196: -------
4197: 4 5 6 P1
4199: Process0 [P0] rows_owned=[0,1]
4200: i = {0,1,3} [size = nrow+1 = 2+1]
4201: j = {0,0,2} [size = 3]
4202: v = {1,2,3} [size = 3]
4204: Process1 [P1] rows_owned=[2]
4205: i = {0,3} [size = nrow+1 = 1+1]
4206: j = {0,1,2} [size = 3]
4207: v = {4,5,6} [size = 3]
4208: .ve
4210: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4211: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4212: @*/
4213: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt i[], const PetscInt j[], const PetscScalar a[], Mat *mat)
4214: {
4215: PetscFunctionBegin;
4216: PetscCheck(!i || !i[0], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4217: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4218: PetscCall(MatCreate(comm, mat));
4219: PetscCall(MatSetSizes(*mat, m, n, M, N));
4220: /* PetscCall(MatSetBlockSizes(M,bs,cbs)); */
4221: PetscCall(MatSetType(*mat, MATMPIAIJ));
4222: PetscCall(MatMPIAIJSetPreallocationCSR(*mat, i, j, a));
4223: PetscFunctionReturn(PETSC_SUCCESS);
4224: }
4226: /*@
4227: MatUpdateMPIAIJWithArrays - updates a `MATMPIAIJ` matrix using arrays that contain in standard
4228: CSR format for the local rows. Only the numerical values are updated the other arrays must be identical to what was passed
4229: from `MatCreateMPIAIJWithArrays()`
4231: Deprecated: Use `MatUpdateMPIAIJWithArray()`
4233: Collective
4235: Input Parameters:
4236: + mat - the matrix
4237: . m - number of local rows (Cannot be `PETSC_DECIDE`)
4238: . n - This value should be the same as the local size used in creating the
4239: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4240: calculated if N is given) For square matrices n is almost always m.
4241: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4242: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4243: . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4244: . J - column indices
4245: - v - matrix values
4247: Level: deprecated
4249: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4250: `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArray()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4251: @*/
4252: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat, PetscInt m, PetscInt n, PetscInt M, PetscInt N, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4253: {
4254: PetscInt nnz, i;
4255: PetscBool nooffprocentries;
4256: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4257: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4258: PetscScalar *ad, *ao;
4259: PetscInt ldi, Iii, md;
4260: const PetscInt *Adi = Ad->i;
4261: PetscInt *ld = Aij->ld;
4263: PetscFunctionBegin;
4264: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
4265: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
4266: PetscCheck(m == mat->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4267: PetscCheck(n == mat->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");
4269: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4270: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4272: for (i = 0; i < m; i++) {
4273: if (PetscDefined(USE_DEBUG)) {
4274: for (PetscInt j = Ii[i] + 1; j < Ii[i + 1]; ++j) {
4275: 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);
4276: PetscCheck(J[j] != J[j - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is identical to previous entry", j - Ii[i], J[j], i);
4277: }
4278: }
4279: nnz = Ii[i + 1] - Ii[i];
4280: Iii = Ii[i];
4281: ldi = ld[i];
4282: md = Adi[i + 1] - Adi[i];
4283: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4284: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4285: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4286: ad += md;
4287: ao += nnz - md;
4288: }
4289: nooffprocentries = mat->nooffprocentries;
4290: mat->nooffprocentries = PETSC_TRUE;
4291: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4292: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4293: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4294: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4295: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4296: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4297: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4298: mat->nooffprocentries = nooffprocentries;
4299: PetscFunctionReturn(PETSC_SUCCESS);
4300: }
4302: /*@
4303: MatUpdateMPIAIJWithArray - updates an `MATMPIAIJ` matrix using an array that contains the nonzero values
4305: Collective
4307: Input Parameters:
4308: + mat - the matrix
4309: - v - matrix values, stored by row
4311: Level: intermediate
4313: Notes:
4314: The matrix must have been obtained with `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()`
4316: The column indices in the call to `MatCreateMPIAIJWithArrays()` or `MatMPIAIJSetPreallocationCSR()` must have been sorted for this call to work correctly
4318: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4319: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithSplitArrays()`, `MatUpdateMPIAIJWithArrays()`, `MatSetPreallocationCOO()`, `MatSetValuesCOO()`
4320: @*/
4321: PetscErrorCode MatUpdateMPIAIJWithArray(Mat mat, const PetscScalar v[])
4322: {
4323: PetscInt nnz, i, m;
4324: PetscBool nooffprocentries;
4325: Mat_MPIAIJ *Aij = (Mat_MPIAIJ *)mat->data;
4326: Mat_SeqAIJ *Ad = (Mat_SeqAIJ *)Aij->A->data;
4327: Mat_SeqAIJ *Ao = (Mat_SeqAIJ *)Aij->B->data;
4328: PetscScalar *ad, *ao;
4329: const PetscInt *Adi = Ad->i, *Adj = Ao->i;
4330: PetscInt ldi, Iii, md;
4331: PetscInt *ld = Aij->ld;
4333: PetscFunctionBegin;
4334: m = mat->rmap->n;
4336: PetscCall(MatSeqAIJGetArrayWrite(Aij->A, &ad));
4337: PetscCall(MatSeqAIJGetArrayWrite(Aij->B, &ao));
4338: Iii = 0;
4339: for (i = 0; i < m; i++) {
4340: nnz = Adi[i + 1] - Adi[i] + Adj[i + 1] - Adj[i];
4341: ldi = ld[i];
4342: md = Adi[i + 1] - Adi[i];
4343: PetscCall(PetscArraycpy(ad, v + Iii + ldi, md));
4344: ad += md;
4345: if (ao) {
4346: PetscCall(PetscArraycpy(ao, v + Iii, ldi));
4347: PetscCall(PetscArraycpy(ao + ldi, v + Iii + ldi + md, nnz - ldi - md));
4348: ao += nnz - md;
4349: }
4350: Iii += nnz;
4351: }
4352: nooffprocentries = mat->nooffprocentries;
4353: mat->nooffprocentries = PETSC_TRUE;
4354: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->A, &ad));
4355: PetscCall(MatSeqAIJRestoreArrayWrite(Aij->B, &ao));
4356: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->A));
4357: PetscCall(PetscObjectStateIncrease((PetscObject)Aij->B));
4358: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
4359: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
4360: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
4361: mat->nooffprocentries = nooffprocentries;
4362: PetscFunctionReturn(PETSC_SUCCESS);
4363: }
4365: /*@
4366: MatCreateAIJ - Creates a sparse parallel matrix in `MATAIJ` format
4367: (the default parallel PETSc format). For good matrix assembly performance
4368: the user should preallocate the matrix storage by setting the parameters
4369: `d_nz` (or `d_nnz`) and `o_nz` (or `o_nnz`).
4371: Collective
4373: Input Parameters:
4374: + comm - MPI communicator
4375: . m - number of local rows (or `PETSC_DECIDE` to have calculated if M is given)
4376: This value should be the same as the local size used in creating the
4377: y vector for the matrix-vector product y = Ax.
4378: . n - This value should be the same as the local size used in creating the
4379: x vector for the matrix-vector product y = Ax. (or `PETSC_DECIDE` to have
4380: calculated if N is given) For square matrices n is almost always m.
4381: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if m is given)
4382: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if n is given)
4383: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4384: (same value is used for all local rows)
4385: . d_nnz - array containing the number of nonzeros in the various rows of the
4386: DIAGONAL portion of the local submatrix (possibly different for each row)
4387: or `NULL`, if `d_nz` is used to specify the nonzero structure.
4388: The size of this array is equal to the number of local rows, i.e 'm'.
4389: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4390: submatrix (same value is used for all local rows).
4391: - o_nnz - array containing the number of nonzeros in the various rows of the
4392: OFF-DIAGONAL portion of the local submatrix (possibly different for
4393: each row) or `NULL`, if `o_nz` is used to specify the nonzero
4394: structure. The size of this array is equal to the number
4395: of local rows, i.e 'm'.
4397: Output Parameter:
4398: . A - the matrix
4400: Options Database Keys:
4401: + -mat_no_inode - Do not use inodes
4402: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4403: - -matmult_vecscatter_view <viewer> - View the vecscatter (i.e., communication pattern) used in `MatMult()` of sparse parallel matrices.
4404: See viewer types in manual of `MatView()`. Of them, ascii_matlab, draw or binary cause the `VecScatter`
4405: 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.
4407: Level: intermediate
4409: Notes:
4410: It is recommended that one use `MatCreateFromOptions()` or the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
4411: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4412: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
4414: If the *_nnz parameter is given then the *_nz parameter is ignored
4416: The `m`,`n`,`M`,`N` parameters specify the size of the matrix, and its partitioning across
4417: processors, while `d_nz`,`d_nnz`,`o_nz`,`o_nnz` parameters specify the approximate
4418: storage requirements for this matrix.
4420: If `PETSC_DECIDE` or `PETSC_DETERMINE` is used for a particular argument on one
4421: processor than it must be used on all processors that share the object for
4422: that argument.
4424: If `m` and `n` are not `PETSC_DECIDE`, then the values determine the `PetscLayout` of the matrix and the ranges returned by
4425: `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`, and `MatGetOwnershipRangesColumn()`.
4427: The user MUST specify either the local or global matrix dimensions
4428: (possibly both).
4430: The parallel matrix is partitioned across processors such that the
4431: first `m0` rows belong to process 0, the next `m1` rows belong to
4432: process 1, the next `m2` rows belong to process 2, etc., where
4433: `m0`, `m1`, `m2`... are the input parameter `m` on each MPI process. I.e., each MPI process stores
4434: values corresponding to [m x N] submatrix.
4436: The columns are logically partitioned with the n0 columns belonging
4437: to 0th partition, the next n1 columns belonging to the next
4438: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4440: The DIAGONAL portion of the local submatrix on any given processor
4441: is the submatrix corresponding to the rows and columns m,n
4442: corresponding to the given processor. i.e diagonal matrix on
4443: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4444: etc. The remaining portion of the local submatrix [m x (N-n)]
4445: constitute the OFF-DIAGONAL portion. The example below better
4446: illustrates this concept. The two matrices, the DIAGONAL portion and
4447: the OFF-DIAGONAL portion are each stored as `MATSEQAIJ` matrices.
4449: For a square global matrix we define each processor's diagonal portion
4450: to be its local rows and the corresponding columns (a square submatrix);
4451: each processor's off-diagonal portion encompasses the remainder of the
4452: local matrix (a rectangular submatrix).
4454: If `o_nnz`, `d_nnz` are specified, then `o_nz`, and `d_nz` are ignored.
4456: When calling this routine with a single process communicator, a matrix of
4457: type `MATSEQAIJ` is returned. If a matrix of type `MATMPIAIJ` is desired for this
4458: type of communicator, use the construction mechanism
4459: .vb
4460: MatCreate(..., &A);
4461: MatSetType(A, MATMPIAIJ);
4462: MatSetSizes(A, m, n, M, N);
4463: MatMPIAIJSetPreallocation(A, ...);
4464: .ve
4466: By default, this format uses inodes (identical nodes) when possible.
4467: We search for consecutive rows with the same nonzero structure, thereby
4468: reusing matrix information to achieve increased efficiency.
4470: Example Usage:
4471: Consider the following 8x8 matrix with 34 non-zero values, that is
4472: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4473: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4474: as follows
4476: .vb
4477: 1 2 0 | 0 3 0 | 0 4
4478: Proc0 0 5 6 | 7 0 0 | 8 0
4479: 9 0 10 | 11 0 0 | 12 0
4480: -------------------------------------
4481: 13 0 14 | 15 16 17 | 0 0
4482: Proc1 0 18 0 | 19 20 21 | 0 0
4483: 0 0 0 | 22 23 0 | 24 0
4484: -------------------------------------
4485: Proc2 25 26 27 | 0 0 28 | 29 0
4486: 30 0 0 | 31 32 33 | 0 34
4487: .ve
4489: This can be represented as a collection of submatrices as
4491: .vb
4492: A B C
4493: D E F
4494: G H I
4495: .ve
4497: Where the submatrices A,B,C are owned by proc0, D,E,F are
4498: owned by proc1, G,H,I are owned by proc2.
4500: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4501: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4502: The 'M','N' parameters are 8,8, and have the same values on all procs.
4504: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4505: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4506: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4507: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4508: part as `MATSEQAIJ` matrices. For example, proc1 will store [E] as a `MATSEQAIJ`
4509: matrix, and [DF] as another SeqAIJ matrix.
4511: When `d_nz`, `o_nz` parameters are specified, `d_nz` storage elements are
4512: allocated for every row of the local DIAGONAL submatrix, and `o_nz`
4513: storage locations are allocated for every row of the OFF-DIAGONAL submatrix.
4514: One way to choose `d_nz` and `o_nz` is to use the maximum number of nonzeros over
4515: the local rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4516: In this case, the values of `d_nz`,`o_nz` are
4517: .vb
4518: proc0 dnz = 2, o_nz = 2
4519: proc1 dnz = 3, o_nz = 2
4520: proc2 dnz = 1, o_nz = 4
4521: .ve
4522: We are allocating m*(`d_nz`+`o_nz`) storage locations for every proc. This
4523: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4524: for proc3. i.e we are using 12+15+10=37 storage locations to store
4525: 34 values.
4527: When `d_nnz`, `o_nnz` parameters are specified, the storage is specified
4528: for every row, corresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4529: In the above case the values for d_nnz,o_nnz are
4530: .vb
4531: proc0 d_nnz = [2,2,2] and o_nnz = [2,2,2]
4532: proc1 d_nnz = [3,3,2] and o_nnz = [2,1,1]
4533: proc2 d_nnz = [1,1] and o_nnz = [4,4]
4534: .ve
4535: Here the space allocated is sum of all the above values i.e 34, and
4536: hence pre-allocation is perfect.
4538: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
4539: `MATMPIAIJ`, `MatCreateMPIAIJWithArrays()`, `MatGetOwnershipRange()`, `MatGetOwnershipRanges()`, `MatGetOwnershipRangeColumn()`,
4540: `MatGetOwnershipRangesColumn()`, `PetscLayout`
4541: @*/
4542: 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)
4543: {
4544: PetscMPIInt size;
4546: PetscFunctionBegin;
4547: PetscCall(MatCreate(comm, A));
4548: PetscCall(MatSetSizes(*A, m, n, M, N));
4549: PetscCallMPI(MPI_Comm_size(comm, &size));
4550: if (size > 1) {
4551: PetscCall(MatSetType(*A, MATMPIAIJ));
4552: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
4553: } else {
4554: PetscCall(MatSetType(*A, MATSEQAIJ));
4555: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
4556: }
4557: PetscFunctionReturn(PETSC_SUCCESS);
4558: }
4560: /*MC
4561: MatMPIAIJGetSeqAIJF90 - Returns the local pieces of this distributed matrix
4563: Synopsis:
4564: MatMPIAIJGetSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4566: Not Collective
4568: Input Parameter:
4569: . A - the `MATMPIAIJ` matrix
4571: Output Parameters:
4572: + Ad - the diagonal portion of the matrix
4573: . Ao - the off-diagonal portion of the matrix
4574: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4575: - ierr - error code
4577: Level: advanced
4579: Note:
4580: Use `MatMPIAIJRestoreSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4582: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJRestoreSeqAIJF90()`
4583: M*/
4585: /*MC
4586: MatMPIAIJRestoreSeqAIJF90 - call after `MatMPIAIJGetSeqAIJF90()` when you no longer need access to the matrices and `colmap`
4588: Synopsis:
4589: MatMPIAIJRestoreSeqAIJF90(Mat A, Mat Ad, Mat Ao, {PetscInt, pointer :: colmap(:)},integer ierr)
4591: Not Collective
4593: Input Parameters:
4594: + A - the `MATMPIAIJ` matrix
4595: . Ad - the diagonal portion of the matrix
4596: . Ao - the off-diagonal portion of the matrix
4597: . colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4598: - ierr - error code
4600: Level: advanced
4602: .seealso: [](ch_matrices), `Mat`, [](sec_fortranarrays), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJ()`, `MatMPIAIJGetSeqAIJF90()`
4603: M*/
4605: /*@C
4606: MatMPIAIJGetSeqAIJ - Returns the local pieces of this distributed matrix
4608: Not Collective
4610: Input Parameter:
4611: . A - The `MATMPIAIJ` matrix
4613: Output Parameters:
4614: + Ad - The local diagonal block as a `MATSEQAIJ` matrix
4615: . Ao - The local off-diagonal block as a `MATSEQAIJ` matrix
4616: - colmap - An array mapping local column numbers of `Ao` to global column numbers of the parallel matrix
4618: Level: intermediate
4620: Note:
4621: The rows in `Ad` and `Ao` are in [0, Nr), where Nr is the number of local rows on this process. The columns
4622: 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
4623: the number of nonzero columns in the local off-diagonal piece of the matrix `A`. The array colmap maps these
4624: local column numbers to global column numbers in the original matrix.
4626: Fortran Notes:
4627: `MatMPIAIJGetSeqAIJ()` Fortran binding is deprecated (since PETSc 3.19), use `MatMPIAIJGetSeqAIJF90()`
4629: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatMPIAIJGetSeqAIJF90()`, `MatMPIAIJRestoreSeqAIJF90()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`, `MatCreateAIJ()`, `MATSEQAIJ`
4630: @*/
4631: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A, Mat *Ad, Mat *Ao, const PetscInt *colmap[])
4632: {
4633: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
4634: PetscBool flg;
4636: PetscFunctionBegin;
4637: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &flg));
4638: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "This function requires a MATMPIAIJ matrix as input");
4639: if (Ad) *Ad = a->A;
4640: if (Ao) *Ao = a->B;
4641: if (colmap) *colmap = a->garray;
4642: PetscFunctionReturn(PETSC_SUCCESS);
4643: }
4645: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
4646: {
4647: PetscInt m, N, i, rstart, nnz, Ii;
4648: PetscInt *indx;
4649: PetscScalar *values;
4650: MatType rootType;
4652: PetscFunctionBegin;
4653: PetscCall(MatGetSize(inmat, &m, &N));
4654: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4655: PetscInt *dnz, *onz, sum, bs, cbs;
4657: if (n == PETSC_DECIDE) PetscCall(PetscSplitOwnership(comm, &n, &N));
4658: /* Check sum(n) = N */
4659: PetscCallMPI(MPIU_Allreduce(&n, &sum, 1, MPIU_INT, MPI_SUM, comm));
4660: PetscCheck(sum == N, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Sum of local columns %" PetscInt_FMT " != global columns %" PetscInt_FMT, sum, N);
4662: PetscCallMPI(MPI_Scan(&m, &rstart, 1, MPIU_INT, MPI_SUM, comm));
4663: rstart -= m;
4665: MatPreallocateBegin(comm, m, n, dnz, onz);
4666: for (i = 0; i < m; i++) {
4667: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4668: PetscCall(MatPreallocateSet(i + rstart, nnz, indx, dnz, onz));
4669: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, NULL));
4670: }
4672: PetscCall(MatCreate(comm, outmat));
4673: PetscCall(MatSetSizes(*outmat, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
4674: PetscCall(MatGetBlockSizes(inmat, &bs, &cbs));
4675: PetscCall(MatSetBlockSizes(*outmat, bs, cbs));
4676: PetscCall(MatGetRootType_Private(inmat, &rootType));
4677: PetscCall(MatSetType(*outmat, rootType));
4678: PetscCall(MatSeqAIJSetPreallocation(*outmat, 0, dnz));
4679: PetscCall(MatMPIAIJSetPreallocation(*outmat, 0, dnz, 0, onz));
4680: MatPreallocateEnd(dnz, onz);
4681: PetscCall(MatSetOption(*outmat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
4682: }
4684: /* numeric phase */
4685: PetscCall(MatGetOwnershipRange(*outmat, &rstart, NULL));
4686: for (i = 0; i < m; i++) {
4687: PetscCall(MatGetRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4688: Ii = i + rstart;
4689: PetscCall(MatSetValues(*outmat, 1, &Ii, nnz, indx, values, INSERT_VALUES));
4690: PetscCall(MatRestoreRow_SeqAIJ(inmat, i, &nnz, &indx, &values));
4691: }
4692: PetscCall(MatAssemblyBegin(*outmat, MAT_FINAL_ASSEMBLY));
4693: PetscCall(MatAssemblyEnd(*outmat, MAT_FINAL_ASSEMBLY));
4694: PetscFunctionReturn(PETSC_SUCCESS);
4695: }
4697: static PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(void *data)
4698: {
4699: Mat_Merge_SeqsToMPI *merge = (Mat_Merge_SeqsToMPI *)data;
4701: PetscFunctionBegin;
4702: if (!merge) PetscFunctionReturn(PETSC_SUCCESS);
4703: PetscCall(PetscFree(merge->id_r));
4704: PetscCall(PetscFree(merge->len_s));
4705: PetscCall(PetscFree(merge->len_r));
4706: PetscCall(PetscFree(merge->bi));
4707: PetscCall(PetscFree(merge->bj));
4708: PetscCall(PetscFree(merge->buf_ri[0]));
4709: PetscCall(PetscFree(merge->buf_ri));
4710: PetscCall(PetscFree(merge->buf_rj[0]));
4711: PetscCall(PetscFree(merge->buf_rj));
4712: PetscCall(PetscFree(merge->coi));
4713: PetscCall(PetscFree(merge->coj));
4714: PetscCall(PetscFree(merge->owners_co));
4715: PetscCall(PetscLayoutDestroy(&merge->rowmap));
4716: PetscCall(PetscFree(merge));
4717: PetscFunctionReturn(PETSC_SUCCESS);
4718: }
4720: #include <../src/mat/utils/freespace.h>
4721: #include <petscbt.h>
4723: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat, Mat mpimat)
4724: {
4725: MPI_Comm comm;
4726: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4727: PetscMPIInt size, rank, taga, *len_s;
4728: PetscInt N = mpimat->cmap->N, i, j, *owners, *ai = a->i, *aj, m;
4729: PetscMPIInt proc, k;
4730: PetscInt **buf_ri, **buf_rj;
4731: PetscInt anzi, *bj_i, *bi, *bj, arow, bnzi, nextaj;
4732: PetscInt nrows, **buf_ri_k, **nextrow, **nextai;
4733: MPI_Request *s_waits, *r_waits;
4734: MPI_Status *status;
4735: const MatScalar *aa, *a_a;
4736: MatScalar **abuf_r, *ba_i;
4737: Mat_Merge_SeqsToMPI *merge;
4738: PetscContainer container;
4740: PetscFunctionBegin;
4741: PetscCall(PetscObjectGetComm((PetscObject)mpimat, &comm));
4742: PetscCall(PetscLogEventBegin(MAT_Seqstompinum, seqmat, 0, 0, 0));
4744: PetscCallMPI(MPI_Comm_size(comm, &size));
4745: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4747: PetscCall(PetscObjectQuery((PetscObject)mpimat, "MatMergeSeqsToMPI", (PetscObject *)&container));
4748: PetscCheck(container, PetscObjectComm((PetscObject)mpimat), PETSC_ERR_PLIB, "Mat not created from MatCreateMPIAIJSumSeqAIJSymbolic");
4749: PetscCall(PetscContainerGetPointer(container, (void **)&merge));
4750: PetscCall(MatSeqAIJGetArrayRead(seqmat, &a_a));
4751: aa = a_a;
4753: bi = merge->bi;
4754: bj = merge->bj;
4755: buf_ri = merge->buf_ri;
4756: buf_rj = merge->buf_rj;
4758: PetscCall(PetscMalloc1(size, &status));
4759: owners = merge->rowmap->range;
4760: len_s = merge->len_s;
4762: /* send and recv matrix values */
4763: PetscCall(PetscObjectGetNewTag((PetscObject)mpimat, &taga));
4764: PetscCall(PetscPostIrecvScalar(comm, taga, merge->nrecv, merge->id_r, merge->len_r, &abuf_r, &r_waits));
4766: PetscCall(PetscMalloc1(merge->nsend + 1, &s_waits));
4767: for (proc = 0, k = 0; proc < size; proc++) {
4768: if (!len_s[proc]) continue;
4769: i = owners[proc];
4770: PetscCallMPI(MPIU_Isend(aa + ai[i], len_s[proc], MPIU_MATSCALAR, proc, taga, comm, s_waits + k));
4771: k++;
4772: }
4774: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, r_waits, status));
4775: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, s_waits, status));
4776: PetscCall(PetscFree(status));
4778: PetscCall(PetscFree(s_waits));
4779: PetscCall(PetscFree(r_waits));
4781: /* insert mat values of mpimat */
4782: PetscCall(PetscMalloc1(N, &ba_i));
4783: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4785: for (k = 0; k < merge->nrecv; k++) {
4786: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4787: nrows = *buf_ri_k[k];
4788: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4789: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4790: }
4792: /* set values of ba */
4793: m = merge->rowmap->n;
4794: for (i = 0; i < m; i++) {
4795: arow = owners[rank] + i;
4796: bj_i = bj + bi[i]; /* col indices of the i-th row of mpimat */
4797: bnzi = bi[i + 1] - bi[i];
4798: PetscCall(PetscArrayzero(ba_i, bnzi));
4800: /* add local non-zero vals of this proc's seqmat into ba */
4801: anzi = ai[arow + 1] - ai[arow];
4802: aj = a->j + ai[arow];
4803: aa = a_a + ai[arow];
4804: nextaj = 0;
4805: for (j = 0; nextaj < anzi; j++) {
4806: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4807: ba_i[j] += aa[nextaj++];
4808: }
4809: }
4811: /* add received vals into ba */
4812: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
4813: /* i-th row */
4814: if (i == *nextrow[k]) {
4815: anzi = *(nextai[k] + 1) - *nextai[k];
4816: aj = buf_rj[k] + *nextai[k];
4817: aa = abuf_r[k] + *nextai[k];
4818: nextaj = 0;
4819: for (j = 0; nextaj < anzi; j++) {
4820: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4821: ba_i[j] += aa[nextaj++];
4822: }
4823: }
4824: nextrow[k]++;
4825: nextai[k]++;
4826: }
4827: }
4828: PetscCall(MatSetValues(mpimat, 1, &arow, bnzi, bj_i, ba_i, INSERT_VALUES));
4829: }
4830: PetscCall(MatSeqAIJRestoreArrayRead(seqmat, &a_a));
4831: PetscCall(MatAssemblyBegin(mpimat, MAT_FINAL_ASSEMBLY));
4832: PetscCall(MatAssemblyEnd(mpimat, MAT_FINAL_ASSEMBLY));
4834: PetscCall(PetscFree(abuf_r[0]));
4835: PetscCall(PetscFree(abuf_r));
4836: PetscCall(PetscFree(ba_i));
4837: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
4838: PetscCall(PetscLogEventEnd(MAT_Seqstompinum, seqmat, 0, 0, 0));
4839: PetscFunctionReturn(PETSC_SUCCESS);
4840: }
4842: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, Mat *mpimat)
4843: {
4844: Mat B_mpi;
4845: Mat_SeqAIJ *a = (Mat_SeqAIJ *)seqmat->data;
4846: PetscMPIInt size, rank, tagi, tagj, *len_s, *len_si, *len_ri;
4847: PetscInt **buf_rj, **buf_ri, **buf_ri_k;
4848: PetscInt M = seqmat->rmap->n, N = seqmat->cmap->n, i, *owners, *ai = a->i, *aj = a->j;
4849: PetscInt len, *dnz, *onz, bs, cbs;
4850: PetscInt k, anzi, *bi, *bj, *lnk, nlnk, arow, bnzi;
4851: PetscInt nrows, *buf_s, *buf_si, *buf_si_i, **nextrow, **nextai;
4852: MPI_Request *si_waits, *sj_waits, *ri_waits, *rj_waits;
4853: MPI_Status *status;
4854: PetscFreeSpaceList free_space = NULL, current_space = NULL;
4855: PetscBT lnkbt;
4856: Mat_Merge_SeqsToMPI *merge;
4857: PetscContainer container;
4859: PetscFunctionBegin;
4860: PetscCall(PetscLogEventBegin(MAT_Seqstompisym, seqmat, 0, 0, 0));
4862: /* make sure it is a PETSc comm */
4863: PetscCall(PetscCommDuplicate(comm, &comm, NULL));
4864: PetscCallMPI(MPI_Comm_size(comm, &size));
4865: PetscCallMPI(MPI_Comm_rank(comm, &rank));
4867: PetscCall(PetscNew(&merge));
4868: PetscCall(PetscMalloc1(size, &status));
4870: /* determine row ownership */
4871: PetscCall(PetscLayoutCreate(comm, &merge->rowmap));
4872: PetscCall(PetscLayoutSetLocalSize(merge->rowmap, m));
4873: PetscCall(PetscLayoutSetSize(merge->rowmap, M));
4874: PetscCall(PetscLayoutSetBlockSize(merge->rowmap, 1));
4875: PetscCall(PetscLayoutSetUp(merge->rowmap));
4876: PetscCall(PetscMalloc1(size, &len_si));
4877: PetscCall(PetscMalloc1(size, &merge->len_s));
4879: m = merge->rowmap->n;
4880: owners = merge->rowmap->range;
4882: /* determine the number of messages to send, their lengths */
4883: len_s = merge->len_s;
4885: len = 0; /* length of buf_si[] */
4886: merge->nsend = 0;
4887: for (PetscMPIInt proc = 0; proc < size; proc++) {
4888: len_si[proc] = 0;
4889: if (proc == rank) {
4890: len_s[proc] = 0;
4891: } else {
4892: PetscCall(PetscMPIIntCast(owners[proc + 1] - owners[proc] + 1, &len_si[proc]));
4893: PetscCall(PetscMPIIntCast(ai[owners[proc + 1]] - ai[owners[proc]], &len_s[proc])); /* num of rows to be sent to [proc] */
4894: }
4895: if (len_s[proc]) {
4896: merge->nsend++;
4897: nrows = 0;
4898: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4899: if (ai[i + 1] > ai[i]) nrows++;
4900: }
4901: PetscCall(PetscMPIIntCast(2 * (nrows + 1), &len_si[proc]));
4902: len += len_si[proc];
4903: }
4904: }
4906: /* determine the number and length of messages to receive for ij-structure */
4907: PetscCall(PetscGatherNumberOfMessages(comm, NULL, len_s, &merge->nrecv));
4908: PetscCall(PetscGatherMessageLengths2(comm, merge->nsend, merge->nrecv, len_s, len_si, &merge->id_r, &merge->len_r, &len_ri));
4910: /* post the Irecv of j-structure */
4911: PetscCall(PetscCommGetNewTag(comm, &tagj));
4912: PetscCall(PetscPostIrecvInt(comm, tagj, merge->nrecv, merge->id_r, merge->len_r, &buf_rj, &rj_waits));
4914: /* post the Isend of j-structure */
4915: PetscCall(PetscMalloc2(merge->nsend, &si_waits, merge->nsend, &sj_waits));
4917: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4918: if (!len_s[proc]) continue;
4919: i = owners[proc];
4920: PetscCallMPI(MPIU_Isend(aj + ai[i], len_s[proc], MPIU_INT, proc, tagj, comm, sj_waits + k));
4921: k++;
4922: }
4924: /* receives and sends of j-structure are complete */
4925: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, rj_waits, status));
4926: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, sj_waits, status));
4928: /* send and recv i-structure */
4929: PetscCall(PetscCommGetNewTag(comm, &tagi));
4930: PetscCall(PetscPostIrecvInt(comm, tagi, merge->nrecv, merge->id_r, len_ri, &buf_ri, &ri_waits));
4932: PetscCall(PetscMalloc1(len + 1, &buf_s));
4933: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4934: for (PetscMPIInt proc = 0, k = 0; proc < size; proc++) {
4935: if (!len_s[proc]) continue;
4936: /* form outgoing message for i-structure:
4937: buf_si[0]: nrows to be sent
4938: [1:nrows]: row index (global)
4939: [nrows+1:2*nrows+1]: i-structure index
4940: */
4941: nrows = len_si[proc] / 2 - 1;
4942: buf_si_i = buf_si + nrows + 1;
4943: buf_si[0] = nrows;
4944: buf_si_i[0] = 0;
4945: nrows = 0;
4946: for (i = owners[proc]; i < owners[proc + 1]; i++) {
4947: anzi = ai[i + 1] - ai[i];
4948: if (anzi) {
4949: buf_si_i[nrows + 1] = buf_si_i[nrows] + anzi; /* i-structure */
4950: buf_si[nrows + 1] = i - owners[proc]; /* local row index */
4951: nrows++;
4952: }
4953: }
4954: PetscCallMPI(MPIU_Isend(buf_si, len_si[proc], MPIU_INT, proc, tagi, comm, si_waits + k));
4955: k++;
4956: buf_si += len_si[proc];
4957: }
4959: if (merge->nrecv) PetscCallMPI(MPI_Waitall(merge->nrecv, ri_waits, status));
4960: if (merge->nsend) PetscCallMPI(MPI_Waitall(merge->nsend, si_waits, status));
4962: PetscCall(PetscInfo(seqmat, "nsend: %d, nrecv: %d\n", merge->nsend, merge->nrecv));
4963: 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]));
4965: PetscCall(PetscFree(len_si));
4966: PetscCall(PetscFree(len_ri));
4967: PetscCall(PetscFree(rj_waits));
4968: PetscCall(PetscFree2(si_waits, sj_waits));
4969: PetscCall(PetscFree(ri_waits));
4970: PetscCall(PetscFree(buf_s));
4971: PetscCall(PetscFree(status));
4973: /* compute a local seq matrix in each processor */
4974: /* allocate bi array and free space for accumulating nonzero column info */
4975: PetscCall(PetscMalloc1(m + 1, &bi));
4976: bi[0] = 0;
4978: /* create and initialize a linked list */
4979: nlnk = N + 1;
4980: PetscCall(PetscLLCreate(N, N, nlnk, lnk, lnkbt));
4982: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4983: len = ai[owners[rank + 1]] - ai[owners[rank]];
4984: PetscCall(PetscFreeSpaceGet(PetscIntMultTruncate(2, len) + 1, &free_space));
4986: current_space = free_space;
4988: /* determine symbolic info for each local row */
4989: PetscCall(PetscMalloc3(merge->nrecv, &buf_ri_k, merge->nrecv, &nextrow, merge->nrecv, &nextai));
4991: for (k = 0; k < merge->nrecv; k++) {
4992: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4993: nrows = *buf_ri_k[k];
4994: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4995: nextai[k] = buf_ri_k[k] + (nrows + 1); /* points to the next i-structure of k-th recved i-structure */
4996: }
4998: MatPreallocateBegin(comm, m, n, dnz, onz);
4999: len = 0;
5000: for (i = 0; i < m; i++) {
5001: bnzi = 0;
5002: /* add local non-zero cols of this proc's seqmat into lnk */
5003: arow = owners[rank] + i;
5004: anzi = ai[arow + 1] - ai[arow];
5005: aj = a->j + ai[arow];
5006: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5007: bnzi += nlnk;
5008: /* add received col data into lnk */
5009: for (k = 0; k < merge->nrecv; k++) { /* k-th received message */
5010: if (i == *nextrow[k]) { /* i-th row */
5011: anzi = *(nextai[k] + 1) - *nextai[k];
5012: aj = buf_rj[k] + *nextai[k];
5013: PetscCall(PetscLLAddSorted(anzi, aj, N, &nlnk, lnk, lnkbt));
5014: bnzi += nlnk;
5015: nextrow[k]++;
5016: nextai[k]++;
5017: }
5018: }
5019: if (len < bnzi) len = bnzi; /* =max(bnzi) */
5021: /* if free space is not available, make more free space */
5022: if (current_space->local_remaining < bnzi) PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(bnzi, current_space->total_array_size), ¤t_space));
5023: /* copy data into free space, then initialize lnk */
5024: PetscCall(PetscLLClean(N, N, bnzi, lnk, current_space->array, lnkbt));
5025: PetscCall(MatPreallocateSet(i + owners[rank], bnzi, current_space->array, dnz, onz));
5027: current_space->array += bnzi;
5028: current_space->local_used += bnzi;
5029: current_space->local_remaining -= bnzi;
5031: bi[i + 1] = bi[i] + bnzi;
5032: }
5034: PetscCall(PetscFree3(buf_ri_k, nextrow, nextai));
5036: PetscCall(PetscMalloc1(bi[m] + 1, &bj));
5037: PetscCall(PetscFreeSpaceContiguous(&free_space, bj));
5038: PetscCall(PetscLLDestroy(lnk, lnkbt));
5040: /* create symbolic parallel matrix B_mpi */
5041: PetscCall(MatGetBlockSizes(seqmat, &bs, &cbs));
5042: PetscCall(MatCreate(comm, &B_mpi));
5043: if (n == PETSC_DECIDE) {
5044: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, N));
5045: } else {
5046: PetscCall(MatSetSizes(B_mpi, m, n, PETSC_DETERMINE, PETSC_DETERMINE));
5047: }
5048: PetscCall(MatSetBlockSizes(B_mpi, bs, cbs));
5049: PetscCall(MatSetType(B_mpi, MATMPIAIJ));
5050: PetscCall(MatMPIAIJSetPreallocation(B_mpi, 0, dnz, 0, onz));
5051: MatPreallocateEnd(dnz, onz);
5052: PetscCall(MatSetOption(B_mpi, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE));
5054: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
5055: B_mpi->assembled = PETSC_FALSE;
5056: merge->bi = bi;
5057: merge->bj = bj;
5058: merge->buf_ri = buf_ri;
5059: merge->buf_rj = buf_rj;
5060: merge->coi = NULL;
5061: merge->coj = NULL;
5062: merge->owners_co = NULL;
5064: PetscCall(PetscCommDestroy(&comm));
5066: /* attach the supporting struct to B_mpi for reuse */
5067: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
5068: PetscCall(PetscContainerSetPointer(container, merge));
5069: PetscCall(PetscContainerSetUserDestroy(container, MatDestroy_MPIAIJ_SeqsToMPI));
5070: PetscCall(PetscObjectCompose((PetscObject)B_mpi, "MatMergeSeqsToMPI", (PetscObject)container));
5071: PetscCall(PetscContainerDestroy(&container));
5072: *mpimat = B_mpi;
5074: PetscCall(PetscLogEventEnd(MAT_Seqstompisym, seqmat, 0, 0, 0));
5075: PetscFunctionReturn(PETSC_SUCCESS);
5076: }
5078: /*@
5079: MatCreateMPIAIJSumSeqAIJ - Creates a `MATMPIAIJ` matrix by adding sequential
5080: matrices from each processor
5082: Collective
5084: Input Parameters:
5085: + comm - the communicators the parallel matrix will live on
5086: . seqmat - the input sequential matrices
5087: . m - number of local rows (or `PETSC_DECIDE`)
5088: . n - number of local columns (or `PETSC_DECIDE`)
5089: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5091: Output Parameter:
5092: . mpimat - the parallel matrix generated
5094: Level: advanced
5096: Note:
5097: The dimensions of the sequential matrix in each processor MUST be the same.
5098: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
5099: destroyed when `mpimat` is destroyed. Call `PetscObjectQuery()` to access `seqmat`.
5101: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`
5102: @*/
5103: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm, Mat seqmat, PetscInt m, PetscInt n, MatReuse scall, Mat *mpimat)
5104: {
5105: PetscMPIInt size;
5107: PetscFunctionBegin;
5108: PetscCallMPI(MPI_Comm_size(comm, &size));
5109: if (size == 1) {
5110: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5111: if (scall == MAT_INITIAL_MATRIX) {
5112: PetscCall(MatDuplicate(seqmat, MAT_COPY_VALUES, mpimat));
5113: } else {
5114: PetscCall(MatCopy(seqmat, *mpimat, SAME_NONZERO_PATTERN));
5115: }
5116: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5117: PetscFunctionReturn(PETSC_SUCCESS);
5118: }
5119: PetscCall(PetscLogEventBegin(MAT_Seqstompi, seqmat, 0, 0, 0));
5120: if (scall == MAT_INITIAL_MATRIX) PetscCall(MatCreateMPIAIJSumSeqAIJSymbolic(comm, seqmat, m, n, mpimat));
5121: PetscCall(MatCreateMPIAIJSumSeqAIJNumeric(seqmat, *mpimat));
5122: PetscCall(PetscLogEventEnd(MAT_Seqstompi, seqmat, 0, 0, 0));
5123: PetscFunctionReturn(PETSC_SUCCESS);
5124: }
5126: /*@
5127: MatAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATAIJ` matrix.
5129: Not Collective
5131: Input Parameter:
5132: . A - the matrix
5134: Output Parameter:
5135: . A_loc - the local sequential matrix generated
5137: Level: developer
5139: Notes:
5140: The matrix is created by taking `A`'s local rows and putting them into a sequential matrix
5141: with `mlocal` rows and `n` columns. Where `mlocal` is obtained with `MatGetLocalSize()` and
5142: `n` is the global column count obtained with `MatGetSize()`
5144: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5146: For parallel matrices this creates an entirely new matrix. If the matrix is sequential it merely increases the reference count.
5148: Destroy the matrix with `MatDestroy()`
5150: .seealso: [](ch_matrices), `Mat`, `MatMPIAIJGetLocalMat()`
5151: @*/
5152: PetscErrorCode MatAIJGetLocalMat(Mat A, Mat *A_loc)
5153: {
5154: PetscBool mpi;
5156: PetscFunctionBegin;
5157: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &mpi));
5158: if (mpi) {
5159: PetscCall(MatMPIAIJGetLocalMat(A, MAT_INITIAL_MATRIX, A_loc));
5160: } else {
5161: *A_loc = A;
5162: PetscCall(PetscObjectReference((PetscObject)*A_loc));
5163: }
5164: PetscFunctionReturn(PETSC_SUCCESS);
5165: }
5167: /*@
5168: MatMPIAIJGetLocalMat - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix.
5170: Not Collective
5172: Input Parameters:
5173: + A - the matrix
5174: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5176: Output Parameter:
5177: . A_loc - the local sequential matrix generated
5179: Level: developer
5181: Notes:
5182: The matrix is created by taking all `A`'s local rows and putting them into a sequential
5183: matrix with `mlocal` rows and `n` columns.`mlocal` is the row count obtained with
5184: `MatGetLocalSize()` and `n` is the global column count obtained with `MatGetSize()`.
5186: In other words combines the two parts of a parallel `MATMPIAIJ` matrix on each process to a single matrix.
5188: When `A` is sequential and `MAT_INITIAL_MATRIX` is requested, the matrix returned is the diagonal part of `A` (which contains the entire matrix),
5189: 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
5190: then `MatCopy`(Adiag,*`A_loc`,`SAME_NONZERO_PATTERN`) is called to fill `A_loc`. Thus one can preallocate the appropriate sequential matrix `A_loc`
5191: 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.
5193: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMatCondensed()`, `MatMPIAIJGetLocalMatMerge()`
5194: @*/
5195: PetscErrorCode MatMPIAIJGetLocalMat(Mat A, MatReuse scall, Mat *A_loc)
5196: {
5197: Mat_MPIAIJ *mpimat = (Mat_MPIAIJ *)A->data;
5198: Mat_SeqAIJ *mat, *a, *b;
5199: PetscInt *ai, *aj, *bi, *bj, *cmap = mpimat->garray;
5200: const PetscScalar *aa, *ba, *aav, *bav;
5201: PetscScalar *ca, *cam;
5202: PetscMPIInt size;
5203: PetscInt am = A->rmap->n, i, j, k, cstart = A->cmap->rstart;
5204: PetscInt *ci, *cj, col, ncols_d, ncols_o, jo;
5205: PetscBool match;
5207: PetscFunctionBegin;
5208: PetscCall(PetscStrbeginswith(((PetscObject)A)->type_name, MATMPIAIJ, &match));
5209: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5210: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5211: if (size == 1) {
5212: if (scall == MAT_INITIAL_MATRIX) {
5213: PetscCall(PetscObjectReference((PetscObject)mpimat->A));
5214: *A_loc = mpimat->A;
5215: } else if (scall == MAT_REUSE_MATRIX) {
5216: PetscCall(MatCopy(mpimat->A, *A_loc, SAME_NONZERO_PATTERN));
5217: }
5218: PetscFunctionReturn(PETSC_SUCCESS);
5219: }
5221: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5222: a = (Mat_SeqAIJ *)mpimat->A->data;
5223: b = (Mat_SeqAIJ *)mpimat->B->data;
5224: ai = a->i;
5225: aj = a->j;
5226: bi = b->i;
5227: bj = b->j;
5228: PetscCall(MatSeqAIJGetArrayRead(mpimat->A, &aav));
5229: PetscCall(MatSeqAIJGetArrayRead(mpimat->B, &bav));
5230: aa = aav;
5231: ba = bav;
5232: if (scall == MAT_INITIAL_MATRIX) {
5233: PetscCall(PetscMalloc1(1 + am, &ci));
5234: ci[0] = 0;
5235: for (i = 0; i < am; i++) ci[i + 1] = ci[i] + (ai[i + 1] - ai[i]) + (bi[i + 1] - bi[i]);
5236: PetscCall(PetscMalloc1(1 + ci[am], &cj));
5237: PetscCall(PetscMalloc1(1 + ci[am], &ca));
5238: k = 0;
5239: for (i = 0; i < am; i++) {
5240: ncols_o = bi[i + 1] - bi[i];
5241: ncols_d = ai[i + 1] - ai[i];
5242: /* off-diagonal portion of A */
5243: for (jo = 0; jo < ncols_o; jo++) {
5244: col = cmap[*bj];
5245: if (col >= cstart) break;
5246: cj[k] = col;
5247: bj++;
5248: ca[k++] = *ba++;
5249: }
5250: /* diagonal portion of A */
5251: for (j = 0; j < ncols_d; j++) {
5252: cj[k] = cstart + *aj++;
5253: ca[k++] = *aa++;
5254: }
5255: /* off-diagonal portion of A */
5256: for (j = jo; j < ncols_o; j++) {
5257: cj[k] = cmap[*bj++];
5258: ca[k++] = *ba++;
5259: }
5260: }
5261: /* put together the new matrix */
5262: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, A->cmap->N, ci, cj, ca, A_loc));
5263: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5264: /* Since these are PETSc arrays, change flags to free them as necessary. */
5265: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5266: mat->free_a = PETSC_TRUE;
5267: mat->free_ij = PETSC_TRUE;
5268: mat->nonew = 0;
5269: } else if (scall == MAT_REUSE_MATRIX) {
5270: mat = (Mat_SeqAIJ *)(*A_loc)->data;
5271: ci = mat->i;
5272: cj = mat->j;
5273: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &cam));
5274: for (i = 0; i < am; i++) {
5275: /* off-diagonal portion of A */
5276: ncols_o = bi[i + 1] - bi[i];
5277: for (jo = 0; jo < ncols_o; jo++) {
5278: col = cmap[*bj];
5279: if (col >= cstart) break;
5280: *cam++ = *ba++;
5281: bj++;
5282: }
5283: /* diagonal portion of A */
5284: ncols_d = ai[i + 1] - ai[i];
5285: for (j = 0; j < ncols_d; j++) *cam++ = *aa++;
5286: /* off-diagonal portion of A */
5287: for (j = jo; j < ncols_o; j++) {
5288: *cam++ = *ba++;
5289: bj++;
5290: }
5291: }
5292: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &cam));
5293: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5294: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->A, &aav));
5295: PetscCall(MatSeqAIJRestoreArrayRead(mpimat->B, &bav));
5296: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5297: PetscFunctionReturn(PETSC_SUCCESS);
5298: }
5300: /*@
5301: MatMPIAIJGetLocalMatMerge - Creates a `MATSEQAIJ` from a `MATMPIAIJ` matrix by taking all its local rows and putting them into a sequential matrix with
5302: mlocal rows and n columns. Where n is the sum of the number of columns of the diagonal and off-diagonal part
5304: Not Collective
5306: Input Parameters:
5307: + A - the matrix
5308: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5310: Output Parameters:
5311: + glob - sequential `IS` with global indices associated with the columns of the local sequential matrix generated (can be `NULL`)
5312: - A_loc - the local sequential matrix generated
5314: Level: developer
5316: Note:
5317: This is different from `MatMPIAIJGetLocalMat()` since the first columns in the returning matrix are those associated with the diagonal
5318: part, then those associated with the off-diagonal part (in its local ordering)
5320: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`, `MatMPIAIJGetLocalMatCondensed()`
5321: @*/
5322: PetscErrorCode MatMPIAIJGetLocalMatMerge(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
5323: {
5324: Mat Ao, Ad;
5325: const PetscInt *cmap;
5326: PetscMPIInt size;
5327: PetscErrorCode (*f)(Mat, MatReuse, IS *, Mat *);
5329: PetscFunctionBegin;
5330: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
5331: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
5332: if (size == 1) {
5333: if (scall == MAT_INITIAL_MATRIX) {
5334: PetscCall(PetscObjectReference((PetscObject)Ad));
5335: *A_loc = Ad;
5336: } else if (scall == MAT_REUSE_MATRIX) {
5337: PetscCall(MatCopy(Ad, *A_loc, SAME_NONZERO_PATTERN));
5338: }
5339: if (glob) PetscCall(ISCreateStride(PetscObjectComm((PetscObject)Ad), Ad->cmap->n, Ad->cmap->rstart, 1, glob));
5340: PetscFunctionReturn(PETSC_SUCCESS);
5341: }
5342: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", &f));
5343: PetscCall(PetscLogEventBegin(MAT_Getlocalmat, A, 0, 0, 0));
5344: if (f) {
5345: PetscCall((*f)(A, scall, glob, A_loc));
5346: } else {
5347: Mat_SeqAIJ *a = (Mat_SeqAIJ *)Ad->data;
5348: Mat_SeqAIJ *b = (Mat_SeqAIJ *)Ao->data;
5349: Mat_SeqAIJ *c;
5350: PetscInt *ai = a->i, *aj = a->j;
5351: PetscInt *bi = b->i, *bj = b->j;
5352: PetscInt *ci, *cj;
5353: const PetscScalar *aa, *ba;
5354: PetscScalar *ca;
5355: PetscInt i, j, am, dn, on;
5357: PetscCall(MatGetLocalSize(Ad, &am, &dn));
5358: PetscCall(MatGetLocalSize(Ao, NULL, &on));
5359: PetscCall(MatSeqAIJGetArrayRead(Ad, &aa));
5360: PetscCall(MatSeqAIJGetArrayRead(Ao, &ba));
5361: if (scall == MAT_INITIAL_MATRIX) {
5362: PetscInt k;
5363: PetscCall(PetscMalloc1(1 + am, &ci));
5364: PetscCall(PetscMalloc1(ai[am] + bi[am], &cj));
5365: PetscCall(PetscMalloc1(ai[am] + bi[am], &ca));
5366: ci[0] = 0;
5367: for (i = 0, k = 0; i < am; i++) {
5368: const PetscInt ncols_o = bi[i + 1] - bi[i];
5369: const PetscInt ncols_d = ai[i + 1] - ai[i];
5370: ci[i + 1] = ci[i] + ncols_o + ncols_d;
5371: /* diagonal portion of A */
5372: for (j = 0; j < ncols_d; j++, k++) {
5373: cj[k] = *aj++;
5374: ca[k] = *aa++;
5375: }
5376: /* off-diagonal portion of A */
5377: for (j = 0; j < ncols_o; j++, k++) {
5378: cj[k] = dn + *bj++;
5379: ca[k] = *ba++;
5380: }
5381: }
5382: /* put together the new matrix */
5383: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, am, dn + on, ci, cj, ca, A_loc));
5384: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5385: /* Since these are PETSc arrays, change flags to free them as necessary. */
5386: c = (Mat_SeqAIJ *)(*A_loc)->data;
5387: c->free_a = PETSC_TRUE;
5388: c->free_ij = PETSC_TRUE;
5389: c->nonew = 0;
5390: PetscCall(MatSetType(*A_loc, ((PetscObject)Ad)->type_name));
5391: } else if (scall == MAT_REUSE_MATRIX) {
5392: PetscCall(MatSeqAIJGetArrayWrite(*A_loc, &ca));
5393: for (i = 0; i < am; i++) {
5394: const PetscInt ncols_d = ai[i + 1] - ai[i];
5395: const PetscInt ncols_o = bi[i + 1] - bi[i];
5396: /* diagonal portion of A */
5397: for (j = 0; j < ncols_d; j++) *ca++ = *aa++;
5398: /* off-diagonal portion of A */
5399: for (j = 0; j < ncols_o; j++) *ca++ = *ba++;
5400: }
5401: PetscCall(MatSeqAIJRestoreArrayWrite(*A_loc, &ca));
5402: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid MatReuse %d", (int)scall);
5403: PetscCall(MatSeqAIJRestoreArrayRead(Ad, &aa));
5404: PetscCall(MatSeqAIJRestoreArrayRead(Ao, &aa));
5405: if (glob) {
5406: PetscInt cst, *gidx;
5408: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
5409: PetscCall(PetscMalloc1(dn + on, &gidx));
5410: for (i = 0; i < dn; i++) gidx[i] = cst + i;
5411: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
5412: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
5413: }
5414: }
5415: PetscCall(PetscLogEventEnd(MAT_Getlocalmat, A, 0, 0, 0));
5416: PetscFunctionReturn(PETSC_SUCCESS);
5417: }
5419: /*@C
5420: MatMPIAIJGetLocalMatCondensed - Creates a `MATSEQAIJ` matrix from an `MATMPIAIJ` matrix by taking all its local rows and NON-ZERO columns
5422: Not Collective
5424: Input Parameters:
5425: + A - the matrix
5426: . scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5427: . row - index set of rows to extract (or `NULL`)
5428: - col - index set of columns to extract (or `NULL`)
5430: Output Parameter:
5431: . A_loc - the local sequential matrix generated
5433: Level: developer
5435: .seealso: [](ch_matrices), `Mat`, `MATMPIAIJ`, `MatGetOwnershipRange()`, `MatMPIAIJGetLocalMat()`
5436: @*/
5437: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A, MatReuse scall, IS *row, IS *col, Mat *A_loc)
5438: {
5439: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5440: PetscInt i, start, end, ncols, nzA, nzB, *cmap, imark, *idx;
5441: IS isrowa, iscola;
5442: Mat *aloc;
5443: PetscBool match;
5445: PetscFunctionBegin;
5446: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATMPIAIJ, &match));
5447: PetscCheck(match, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Requires MATMPIAIJ matrix as input");
5448: PetscCall(PetscLogEventBegin(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5449: if (!row) {
5450: start = A->rmap->rstart;
5451: end = A->rmap->rend;
5452: PetscCall(ISCreateStride(PETSC_COMM_SELF, end - start, start, 1, &isrowa));
5453: } else {
5454: isrowa = *row;
5455: }
5456: if (!col) {
5457: start = A->cmap->rstart;
5458: cmap = a->garray;
5459: nzA = a->A->cmap->n;
5460: nzB = a->B->cmap->n;
5461: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5462: ncols = 0;
5463: for (i = 0; i < nzB; i++) {
5464: if (cmap[i] < start) idx[ncols++] = cmap[i];
5465: else break;
5466: }
5467: imark = i;
5468: for (i = 0; i < nzA; i++) idx[ncols++] = start + i;
5469: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i];
5470: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &iscola));
5471: } else {
5472: iscola = *col;
5473: }
5474: if (scall != MAT_INITIAL_MATRIX) {
5475: PetscCall(PetscMalloc1(1, &aloc));
5476: aloc[0] = *A_loc;
5477: }
5478: PetscCall(MatCreateSubMatrices(A, 1, &isrowa, &iscola, scall, &aloc));
5479: if (!col) { /* attach global id of condensed columns */
5480: PetscCall(PetscObjectCompose((PetscObject)aloc[0], "_petsc_GetLocalMatCondensed_iscol", (PetscObject)iscola));
5481: }
5482: *A_loc = aloc[0];
5483: PetscCall(PetscFree(aloc));
5484: if (!row) PetscCall(ISDestroy(&isrowa));
5485: if (!col) PetscCall(ISDestroy(&iscola));
5486: PetscCall(PetscLogEventEnd(MAT_Getlocalmatcondensed, A, 0, 0, 0));
5487: PetscFunctionReturn(PETSC_SUCCESS);
5488: }
5490: /*
5491: * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5492: * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5493: * on a global size.
5494: * */
5495: static PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P, IS rows, Mat *P_oth)
5496: {
5497: Mat_MPIAIJ *p = (Mat_MPIAIJ *)P->data;
5498: Mat_SeqAIJ *pd = (Mat_SeqAIJ *)p->A->data, *po = (Mat_SeqAIJ *)p->B->data, *p_oth;
5499: PetscInt plocalsize, nrows, *ilocal, *oilocal, i, lidx, *nrcols, *nlcols, ncol;
5500: PetscMPIInt owner;
5501: PetscSFNode *iremote, *oiremote;
5502: const PetscInt *lrowindices;
5503: PetscSF sf, osf;
5504: PetscInt pcstart, *roffsets, *loffsets, *pnnz, j;
5505: PetscInt ontotalcols, dntotalcols, ntotalcols, nout;
5506: MPI_Comm comm;
5507: ISLocalToGlobalMapping mapping;
5508: const PetscScalar *pd_a, *po_a;
5510: PetscFunctionBegin;
5511: PetscCall(PetscObjectGetComm((PetscObject)P, &comm));
5512: /* plocalsize is the number of roots
5513: * nrows is the number of leaves
5514: * */
5515: PetscCall(MatGetLocalSize(P, &plocalsize, NULL));
5516: PetscCall(ISGetLocalSize(rows, &nrows));
5517: PetscCall(PetscCalloc1(nrows, &iremote));
5518: PetscCall(ISGetIndices(rows, &lrowindices));
5519: for (i = 0; i < nrows; i++) {
5520: /* Find a remote index and an owner for a row
5521: * The row could be local or remote
5522: * */
5523: owner = 0;
5524: lidx = 0;
5525: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, &lidx));
5526: iremote[i].index = lidx;
5527: iremote[i].rank = owner;
5528: }
5529: /* Create SF to communicate how many nonzero columns for each row */
5530: PetscCall(PetscSFCreate(comm, &sf));
5531: /* SF will figure out the number of nonzero columns for each row, and their
5532: * offsets
5533: * */
5534: PetscCall(PetscSFSetGraph(sf, plocalsize, nrows, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5535: PetscCall(PetscSFSetFromOptions(sf));
5536: PetscCall(PetscSFSetUp(sf));
5538: PetscCall(PetscCalloc1(2 * (plocalsize + 1), &roffsets));
5539: PetscCall(PetscCalloc1(2 * plocalsize, &nrcols));
5540: PetscCall(PetscCalloc1(nrows, &pnnz));
5541: roffsets[0] = 0;
5542: roffsets[1] = 0;
5543: for (i = 0; i < plocalsize; i++) {
5544: /* diagonal */
5545: nrcols[i * 2 + 0] = pd->i[i + 1] - pd->i[i];
5546: /* off-diagonal */
5547: nrcols[i * 2 + 1] = po->i[i + 1] - po->i[i];
5548: /* compute offsets so that we relative location for each row */
5549: roffsets[(i + 1) * 2 + 0] = roffsets[i * 2 + 0] + nrcols[i * 2 + 0];
5550: roffsets[(i + 1) * 2 + 1] = roffsets[i * 2 + 1] + nrcols[i * 2 + 1];
5551: }
5552: PetscCall(PetscCalloc1(2 * nrows, &nlcols));
5553: PetscCall(PetscCalloc1(2 * nrows, &loffsets));
5554: /* 'r' means root, and 'l' means leaf */
5555: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5556: PetscCall(PetscSFBcastBegin(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5557: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, nrcols, nlcols, MPI_REPLACE));
5558: PetscCall(PetscSFBcastEnd(sf, MPIU_2INT, roffsets, loffsets, MPI_REPLACE));
5559: PetscCall(PetscSFDestroy(&sf));
5560: PetscCall(PetscFree(roffsets));
5561: PetscCall(PetscFree(nrcols));
5562: dntotalcols = 0;
5563: ontotalcols = 0;
5564: ncol = 0;
5565: for (i = 0; i < nrows; i++) {
5566: pnnz[i] = nlcols[i * 2 + 0] + nlcols[i * 2 + 1];
5567: ncol = PetscMax(pnnz[i], ncol);
5568: /* diagonal */
5569: dntotalcols += nlcols[i * 2 + 0];
5570: /* off-diagonal */
5571: ontotalcols += nlcols[i * 2 + 1];
5572: }
5573: /* We do not need to figure the right number of columns
5574: * since all the calculations will be done by going through the raw data
5575: * */
5576: PetscCall(MatCreateSeqAIJ(PETSC_COMM_SELF, nrows, ncol, 0, pnnz, P_oth));
5577: PetscCall(MatSetUp(*P_oth));
5578: PetscCall(PetscFree(pnnz));
5579: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5580: /* diagonal */
5581: PetscCall(PetscCalloc1(dntotalcols, &iremote));
5582: /* off-diagonal */
5583: PetscCall(PetscCalloc1(ontotalcols, &oiremote));
5584: /* diagonal */
5585: PetscCall(PetscCalloc1(dntotalcols, &ilocal));
5586: /* off-diagonal */
5587: PetscCall(PetscCalloc1(ontotalcols, &oilocal));
5588: dntotalcols = 0;
5589: ontotalcols = 0;
5590: ntotalcols = 0;
5591: for (i = 0; i < nrows; i++) {
5592: owner = 0;
5593: PetscCall(PetscLayoutFindOwnerIndex(P->rmap, lrowindices[i], &owner, NULL));
5594: /* Set iremote for diag matrix */
5595: for (j = 0; j < nlcols[i * 2 + 0]; j++) {
5596: iremote[dntotalcols].index = loffsets[i * 2 + 0] + j;
5597: iremote[dntotalcols].rank = owner;
5598: /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5599: ilocal[dntotalcols++] = ntotalcols++;
5600: }
5601: /* off-diagonal */
5602: for (j = 0; j < nlcols[i * 2 + 1]; j++) {
5603: oiremote[ontotalcols].index = loffsets[i * 2 + 1] + j;
5604: oiremote[ontotalcols].rank = owner;
5605: oilocal[ontotalcols++] = ntotalcols++;
5606: }
5607: }
5608: PetscCall(ISRestoreIndices(rows, &lrowindices));
5609: PetscCall(PetscFree(loffsets));
5610: PetscCall(PetscFree(nlcols));
5611: PetscCall(PetscSFCreate(comm, &sf));
5612: /* P serves as roots and P_oth is leaves
5613: * Diag matrix
5614: * */
5615: PetscCall(PetscSFSetGraph(sf, pd->i[plocalsize], dntotalcols, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
5616: PetscCall(PetscSFSetFromOptions(sf));
5617: PetscCall(PetscSFSetUp(sf));
5619: PetscCall(PetscSFCreate(comm, &osf));
5620: /* off-diagonal */
5621: PetscCall(PetscSFSetGraph(osf, po->i[plocalsize], ontotalcols, oilocal, PETSC_OWN_POINTER, oiremote, PETSC_OWN_POINTER));
5622: PetscCall(PetscSFSetFromOptions(osf));
5623: PetscCall(PetscSFSetUp(osf));
5624: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5625: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5626: /* operate on the matrix internal data to save memory */
5627: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5628: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5629: PetscCall(MatGetOwnershipRangeColumn(P, &pcstart, NULL));
5630: /* Convert to global indices for diag matrix */
5631: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] += pcstart;
5632: PetscCall(PetscSFBcastBegin(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5633: /* We want P_oth store global indices */
5634: PetscCall(ISLocalToGlobalMappingCreate(comm, 1, p->B->cmap->n, p->garray, PETSC_COPY_VALUES, &mapping));
5635: /* Use memory scalable approach */
5636: PetscCall(ISLocalToGlobalMappingSetType(mapping, ISLOCALTOGLOBALMAPPINGHASH));
5637: PetscCall(ISLocalToGlobalMappingApply(mapping, po->i[plocalsize], po->j, po->j));
5638: PetscCall(PetscSFBcastBegin(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5639: PetscCall(PetscSFBcastEnd(sf, MPIU_INT, pd->j, p_oth->j, MPI_REPLACE));
5640: /* Convert back to local indices */
5641: for (i = 0; i < pd->i[plocalsize]; i++) pd->j[i] -= pcstart;
5642: PetscCall(PetscSFBcastEnd(osf, MPIU_INT, po->j, p_oth->j, MPI_REPLACE));
5643: nout = 0;
5644: PetscCall(ISGlobalToLocalMappingApply(mapping, IS_GTOLM_DROP, po->i[plocalsize], po->j, &nout, po->j));
5645: PetscCheck(nout == po->i[plocalsize], comm, PETSC_ERR_ARG_INCOMP, "n %" PetscInt_FMT " does not equal to nout %" PetscInt_FMT " ", po->i[plocalsize], nout);
5646: PetscCall(ISLocalToGlobalMappingDestroy(&mapping));
5647: /* Exchange values */
5648: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5649: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5650: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5651: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5652: /* Stop PETSc from shrinking memory */
5653: for (i = 0; i < nrows; i++) p_oth->ilen[i] = p_oth->imax[i];
5654: PetscCall(MatAssemblyBegin(*P_oth, MAT_FINAL_ASSEMBLY));
5655: PetscCall(MatAssemblyEnd(*P_oth, MAT_FINAL_ASSEMBLY));
5656: /* Attach PetscSF objects to P_oth so that we can reuse it later */
5657: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "diagsf", (PetscObject)sf));
5658: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "offdiagsf", (PetscObject)osf));
5659: PetscCall(PetscSFDestroy(&sf));
5660: PetscCall(PetscSFDestroy(&osf));
5661: PetscFunctionReturn(PETSC_SUCCESS);
5662: }
5664: /*
5665: * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5666: * This supports MPIAIJ and MAIJ
5667: * */
5668: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A, Mat P, PetscInt dof, MatReuse reuse, Mat *P_oth)
5669: {
5670: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data, *p = (Mat_MPIAIJ *)P->data;
5671: Mat_SeqAIJ *p_oth;
5672: IS rows, map;
5673: PetscHMapI hamp;
5674: PetscInt i, htsize, *rowindices, off, *mapping, key, count;
5675: MPI_Comm comm;
5676: PetscSF sf, osf;
5677: PetscBool has;
5679: PetscFunctionBegin;
5680: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5681: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, P, 0, 0));
5682: /* If it is the first time, create an index set of off-diag nonzero columns of A,
5683: * and then create a submatrix (that often is an overlapping matrix)
5684: * */
5685: if (reuse == MAT_INITIAL_MATRIX) {
5686: /* Use a hash table to figure out unique keys */
5687: PetscCall(PetscHMapICreateWithSize(a->B->cmap->n, &hamp));
5688: PetscCall(PetscCalloc1(a->B->cmap->n, &mapping));
5689: count = 0;
5690: /* Assume that a->g is sorted, otherwise the following does not make sense */
5691: for (i = 0; i < a->B->cmap->n; i++) {
5692: key = a->garray[i] / dof;
5693: PetscCall(PetscHMapIHas(hamp, key, &has));
5694: if (!has) {
5695: mapping[i] = count;
5696: PetscCall(PetscHMapISet(hamp, key, count++));
5697: } else {
5698: /* Current 'i' has the same value the previous step */
5699: mapping[i] = count - 1;
5700: }
5701: }
5702: PetscCall(ISCreateGeneral(comm, a->B->cmap->n, mapping, PETSC_OWN_POINTER, &map));
5703: PetscCall(PetscHMapIGetSize(hamp, &htsize));
5704: PetscCheck(htsize == count, comm, PETSC_ERR_ARG_INCOMP, " Size of hash map %" PetscInt_FMT " is inconsistent with count %" PetscInt_FMT, htsize, count);
5705: PetscCall(PetscCalloc1(htsize, &rowindices));
5706: off = 0;
5707: PetscCall(PetscHMapIGetKeys(hamp, &off, rowindices));
5708: PetscCall(PetscHMapIDestroy(&hamp));
5709: PetscCall(PetscSortInt(htsize, rowindices));
5710: PetscCall(ISCreateGeneral(comm, htsize, rowindices, PETSC_OWN_POINTER, &rows));
5711: /* In case, the matrix was already created but users want to recreate the matrix */
5712: PetscCall(MatDestroy(P_oth));
5713: PetscCall(MatCreateSeqSubMatrixWithRows_Private(P, rows, P_oth));
5714: PetscCall(PetscObjectCompose((PetscObject)*P_oth, "aoffdiagtopothmapping", (PetscObject)map));
5715: PetscCall(ISDestroy(&map));
5716: PetscCall(ISDestroy(&rows));
5717: } else if (reuse == MAT_REUSE_MATRIX) {
5718: /* If matrix was already created, we simply update values using SF objects
5719: * that as attached to the matrix earlier.
5720: */
5721: const PetscScalar *pd_a, *po_a;
5723: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "diagsf", (PetscObject *)&sf));
5724: PetscCall(PetscObjectQuery((PetscObject)*P_oth, "offdiagsf", (PetscObject *)&osf));
5725: PetscCheck(sf && osf, comm, PETSC_ERR_ARG_NULL, "Matrix is not initialized yet");
5726: p_oth = (Mat_SeqAIJ *)(*P_oth)->data;
5727: /* Update values in place */
5728: PetscCall(MatSeqAIJGetArrayRead(p->A, &pd_a));
5729: PetscCall(MatSeqAIJGetArrayRead(p->B, &po_a));
5730: PetscCall(PetscSFBcastBegin(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5731: PetscCall(PetscSFBcastBegin(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5732: PetscCall(PetscSFBcastEnd(sf, MPIU_SCALAR, pd_a, p_oth->a, MPI_REPLACE));
5733: PetscCall(PetscSFBcastEnd(osf, MPIU_SCALAR, po_a, p_oth->a, MPI_REPLACE));
5734: PetscCall(MatSeqAIJRestoreArrayRead(p->A, &pd_a));
5735: PetscCall(MatSeqAIJRestoreArrayRead(p->B, &po_a));
5736: } else SETERRQ(comm, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown reuse type");
5737: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, P, 0, 0));
5738: PetscFunctionReturn(PETSC_SUCCESS);
5739: }
5741: /*@C
5742: MatGetBrowsOfAcols - Returns `IS` that contain rows of `B` that equal to nonzero columns of local `A`
5744: Collective
5746: Input Parameters:
5747: + A - the first matrix in `MATMPIAIJ` format
5748: . B - the second matrix in `MATMPIAIJ` format
5749: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5751: Output Parameters:
5752: + rowb - On input index sets of rows of B to extract (or `NULL`), modified on output
5753: . colb - On input index sets of columns of B to extract (or `NULL`), modified on output
5754: - B_seq - the sequential matrix generated
5756: Level: developer
5758: .seealso: `Mat`, `MATMPIAIJ`, `IS`, `MatReuse`
5759: @*/
5760: PetscErrorCode MatGetBrowsOfAcols(Mat A, Mat B, MatReuse scall, IS *rowb, IS *colb, Mat *B_seq)
5761: {
5762: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5763: PetscInt *idx, i, start, ncols, nzA, nzB, *cmap, imark;
5764: IS isrowb, iscolb;
5765: Mat *bseq = NULL;
5767: PetscFunctionBegin;
5768: 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 ")",
5769: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5770: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAcols, A, B, 0, 0));
5772: if (scall == MAT_INITIAL_MATRIX) {
5773: start = A->cmap->rstart;
5774: cmap = a->garray;
5775: nzA = a->A->cmap->n;
5776: nzB = a->B->cmap->n;
5777: PetscCall(PetscMalloc1(nzA + nzB, &idx));
5778: ncols = 0;
5779: for (i = 0; i < nzB; i++) { /* row < local row index */
5780: if (cmap[i] < start) idx[ncols++] = cmap[i];
5781: else break;
5782: }
5783: imark = i;
5784: for (i = 0; i < nzA; i++) idx[ncols++] = start + i; /* local rows */
5785: for (i = imark; i < nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5786: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, ncols, idx, PETSC_OWN_POINTER, &isrowb));
5787: PetscCall(ISCreateStride(PETSC_COMM_SELF, B->cmap->N, 0, 1, &iscolb));
5788: } else {
5789: PetscCheck(rowb && colb, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5790: isrowb = *rowb;
5791: iscolb = *colb;
5792: PetscCall(PetscMalloc1(1, &bseq));
5793: bseq[0] = *B_seq;
5794: }
5795: PetscCall(MatCreateSubMatrices(B, 1, &isrowb, &iscolb, scall, &bseq));
5796: *B_seq = bseq[0];
5797: PetscCall(PetscFree(bseq));
5798: if (!rowb) {
5799: PetscCall(ISDestroy(&isrowb));
5800: } else {
5801: *rowb = isrowb;
5802: }
5803: if (!colb) {
5804: PetscCall(ISDestroy(&iscolb));
5805: } else {
5806: *colb = iscolb;
5807: }
5808: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAcols, A, B, 0, 0));
5809: PetscFunctionReturn(PETSC_SUCCESS);
5810: }
5812: /*
5813: MatGetBrowsOfAoCols_MPIAIJ - Creates a `MATSEQAIJ` matrix by taking rows of B that equal to nonzero columns
5814: of the OFF-DIAGONAL portion of local A
5816: Collective
5818: Input Parameters:
5819: + A,B - the matrices in `MATMPIAIJ` format
5820: - scall - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
5822: Output Parameter:
5823: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5824: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5825: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5826: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5828: Developer Note:
5829: This directly accesses information inside the VecScatter associated with the matrix-vector product
5830: for this matrix. This is not desirable..
5832: Level: developer
5834: */
5836: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A, Mat B, MatReuse scall, PetscInt **startsj_s, PetscInt **startsj_r, MatScalar **bufa_ptr, Mat *B_oth)
5837: {
5838: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
5839: VecScatter ctx;
5840: MPI_Comm comm;
5841: const PetscMPIInt *rprocs, *sprocs;
5842: PetscMPIInt nrecvs, nsends;
5843: const PetscInt *srow, *rstarts, *sstarts;
5844: PetscInt *rowlen, *bufj, *bufJ, ncols = 0, aBn = a->B->cmap->n, row, *b_othi, *b_othj, *rvalues = NULL, *svalues = NULL, *cols, sbs, rbs;
5845: PetscInt i, j, k = 0, l, ll, nrows, *rstartsj = NULL, *sstartsj, len;
5846: PetscScalar *b_otha, *bufa, *bufA, *vals = NULL;
5847: MPI_Request *reqs = NULL, *rwaits = NULL, *swaits = NULL;
5848: PetscMPIInt size, tag, rank, nreqs;
5850: PetscFunctionBegin;
5851: PetscCall(PetscObjectGetComm((PetscObject)A, &comm));
5852: PetscCallMPI(MPI_Comm_size(comm, &size));
5854: 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 ")",
5855: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
5856: PetscCall(PetscLogEventBegin(MAT_GetBrowsOfAocols, A, B, 0, 0));
5857: PetscCallMPI(MPI_Comm_rank(comm, &rank));
5859: if (size == 1) {
5860: startsj_s = NULL;
5861: bufa_ptr = NULL;
5862: *B_oth = NULL;
5863: PetscFunctionReturn(PETSC_SUCCESS);
5864: }
5866: ctx = a->Mvctx;
5867: tag = ((PetscObject)ctx)->tag;
5869: PetscCall(VecScatterGetRemote_Private(ctx, PETSC_TRUE /*send*/, &nsends, &sstarts, &srow, &sprocs, &sbs));
5870: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5871: PetscCall(VecScatterGetRemoteOrdered_Private(ctx, PETSC_FALSE /*recv*/, &nrecvs, &rstarts, NULL /*indices not needed*/, &rprocs, &rbs));
5872: PetscCall(PetscMPIIntCast(nsends + nrecvs, &nreqs));
5873: PetscCall(PetscMalloc1(nreqs, &reqs));
5874: rwaits = reqs;
5875: swaits = PetscSafePointerPlusOffset(reqs, nrecvs);
5877: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5878: if (scall == MAT_INITIAL_MATRIX) {
5879: /* i-array */
5880: /* post receives */
5881: if (nrecvs) PetscCall(PetscMalloc1(rbs * (rstarts[nrecvs] - rstarts[0]), &rvalues)); /* rstarts can be NULL when nrecvs=0 */
5882: for (i = 0; i < nrecvs; i++) {
5883: rowlen = rvalues + rstarts[i] * rbs;
5884: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of indices to be received */
5885: PetscCallMPI(MPIU_Irecv(rowlen, nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5886: }
5888: /* pack the outgoing message */
5889: PetscCall(PetscMalloc2(nsends + 1, &sstartsj, nrecvs + 1, &rstartsj));
5891: sstartsj[0] = 0;
5892: rstartsj[0] = 0;
5893: len = 0; /* total length of j or a array to be sent */
5894: if (nsends) {
5895: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5896: PetscCall(PetscMalloc1(sbs * (sstarts[nsends] - sstarts[0]), &svalues));
5897: }
5898: for (i = 0; i < nsends; i++) {
5899: rowlen = svalues + (sstarts[i] - sstarts[0]) * sbs;
5900: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5901: for (j = 0; j < nrows; j++) {
5902: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5903: for (l = 0; l < sbs; l++) {
5904: PetscCall(MatGetRow_MPIAIJ(B, row + l, &ncols, NULL, NULL)); /* rowlength */
5906: rowlen[j * sbs + l] = ncols;
5908: len += ncols;
5909: PetscCall(MatRestoreRow_MPIAIJ(B, row + l, &ncols, NULL, NULL));
5910: }
5911: k++;
5912: }
5913: PetscCallMPI(MPIU_Isend(rowlen, nrows * sbs, MPIU_INT, sprocs[i], tag, comm, swaits + i));
5915: sstartsj[i + 1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5916: }
5917: /* recvs and sends of i-array are completed */
5918: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5919: PetscCall(PetscFree(svalues));
5921: /* allocate buffers for sending j and a arrays */
5922: PetscCall(PetscMalloc1(len + 1, &bufj));
5923: PetscCall(PetscMalloc1(len + 1, &bufa));
5925: /* create i-array of B_oth */
5926: PetscCall(PetscMalloc1(aBn + 2, &b_othi));
5928: b_othi[0] = 0;
5929: len = 0; /* total length of j or a array to be received */
5930: k = 0;
5931: for (i = 0; i < nrecvs; i++) {
5932: rowlen = rvalues + (rstarts[i] - rstarts[0]) * rbs;
5933: nrows = (rstarts[i + 1] - rstarts[i]) * rbs; /* num of rows to be received */
5934: for (j = 0; j < nrows; j++) {
5935: b_othi[k + 1] = b_othi[k] + rowlen[j];
5936: PetscCall(PetscIntSumError(rowlen[j], len, &len));
5937: k++;
5938: }
5939: rstartsj[i + 1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5940: }
5941: PetscCall(PetscFree(rvalues));
5943: /* allocate space for j and a arrays of B_oth */
5944: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_othj));
5945: PetscCall(PetscMalloc1(b_othi[aBn] + 1, &b_otha));
5947: /* j-array */
5948: /* post receives of j-array */
5949: for (i = 0; i < nrecvs; i++) {
5950: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5951: PetscCallMPI(MPIU_Irecv(b_othj + rstartsj[i], nrows, MPIU_INT, rprocs[i], tag, comm, rwaits + i));
5952: }
5954: /* pack the outgoing message j-array */
5955: if (nsends) k = sstarts[0];
5956: for (i = 0; i < nsends; i++) {
5957: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5958: bufJ = bufj + sstartsj[i];
5959: for (j = 0; j < nrows; j++) {
5960: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5961: for (ll = 0; ll < sbs; ll++) {
5962: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5963: for (l = 0; l < ncols; l++) *bufJ++ = cols[l];
5964: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, &cols, NULL));
5965: }
5966: }
5967: PetscCallMPI(MPIU_Isend(bufj + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_INT, sprocs[i], tag, comm, swaits + i));
5968: }
5970: /* recvs and sends of j-array are completed */
5971: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
5972: } else if (scall == MAT_REUSE_MATRIX) {
5973: sstartsj = *startsj_s;
5974: rstartsj = *startsj_r;
5975: bufa = *bufa_ptr;
5976: PetscCall(MatSeqAIJGetArrayWrite(*B_oth, &b_otha));
5977: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not possess an object container");
5979: /* a-array */
5980: /* post receives of a-array */
5981: for (i = 0; i < nrecvs; i++) {
5982: nrows = rstartsj[i + 1] - rstartsj[i]; /* length of the msg received */
5983: PetscCallMPI(MPIU_Irecv(b_otha + rstartsj[i], nrows, MPIU_SCALAR, rprocs[i], tag, comm, rwaits + i));
5984: }
5986: /* pack the outgoing message a-array */
5987: if (nsends) k = sstarts[0];
5988: for (i = 0; i < nsends; i++) {
5989: nrows = sstarts[i + 1] - sstarts[i]; /* num of block rows */
5990: bufA = bufa + sstartsj[i];
5991: for (j = 0; j < nrows; j++) {
5992: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5993: for (ll = 0; ll < sbs; ll++) {
5994: PetscCall(MatGetRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5995: for (l = 0; l < ncols; l++) *bufA++ = vals[l];
5996: PetscCall(MatRestoreRow_MPIAIJ(B, row + ll, &ncols, NULL, &vals));
5997: }
5998: }
5999: PetscCallMPI(MPIU_Isend(bufa + sstartsj[i], sstartsj[i + 1] - sstartsj[i], MPIU_SCALAR, sprocs[i], tag, comm, swaits + i));
6000: }
6001: /* recvs and sends of a-array are completed */
6002: if (nreqs) PetscCallMPI(MPI_Waitall(nreqs, reqs, MPI_STATUSES_IGNORE));
6003: PetscCall(PetscFree(reqs));
6005: if (scall == MAT_INITIAL_MATRIX) {
6006: Mat_SeqAIJ *b_oth;
6008: /* put together the new matrix */
6009: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, aBn, B->cmap->N, b_othi, b_othj, b_otha, B_oth));
6011: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
6012: /* Since these are PETSc arrays, change flags to free them as necessary. */
6013: b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
6014: b_oth->free_a = PETSC_TRUE;
6015: b_oth->free_ij = PETSC_TRUE;
6016: b_oth->nonew = 0;
6018: PetscCall(PetscFree(bufj));
6019: if (!startsj_s || !bufa_ptr) {
6020: PetscCall(PetscFree2(sstartsj, rstartsj));
6021: PetscCall(PetscFree(bufa_ptr));
6022: } else {
6023: *startsj_s = sstartsj;
6024: *startsj_r = rstartsj;
6025: *bufa_ptr = bufa;
6026: }
6027: } else if (scall == MAT_REUSE_MATRIX) {
6028: PetscCall(MatSeqAIJRestoreArrayWrite(*B_oth, &b_otha));
6029: }
6031: PetscCall(VecScatterRestoreRemote_Private(ctx, PETSC_TRUE, &nsends, &sstarts, &srow, &sprocs, &sbs));
6032: PetscCall(VecScatterRestoreRemoteOrdered_Private(ctx, PETSC_FALSE, &nrecvs, &rstarts, NULL, &rprocs, &rbs));
6033: PetscCall(PetscLogEventEnd(MAT_GetBrowsOfAocols, A, B, 0, 0));
6034: PetscFunctionReturn(PETSC_SUCCESS);
6035: }
6037: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat, MatType, MatReuse, Mat *);
6038: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat, MatType, MatReuse, Mat *);
6039: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat, MatType, MatReuse, Mat *);
6040: #if defined(PETSC_HAVE_MKL_SPARSE)
6041: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat, MatType, MatReuse, Mat *);
6042: #endif
6043: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat, MatType, MatReuse, Mat *);
6044: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat, MatType, MatReuse, Mat *);
6045: #if defined(PETSC_HAVE_ELEMENTAL)
6046: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
6047: #endif
6048: #if defined(PETSC_HAVE_SCALAPACK)
6049: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
6050: #endif
6051: #if defined(PETSC_HAVE_HYPRE)
6052: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat, MatType, MatReuse, Mat *);
6053: #endif
6054: #if defined(PETSC_HAVE_CUDA)
6055: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
6056: #endif
6057: #if defined(PETSC_HAVE_HIP)
6058: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
6059: #endif
6060: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6061: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJKokkos(Mat, MatType, MatReuse, Mat *);
6062: #endif
6063: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat, MatType, MatReuse, Mat *);
6064: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
6065: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
6067: /*
6068: Computes (B'*A')' since computing B*A directly is untenable
6070: n p p
6071: [ ] [ ] [ ]
6072: m [ A ] * n [ B ] = m [ C ]
6073: [ ] [ ] [ ]
6075: */
6076: static PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A, Mat B, Mat C)
6077: {
6078: Mat At, Bt, Ct;
6080: PetscFunctionBegin;
6081: PetscCall(MatTranspose(A, MAT_INITIAL_MATRIX, &At));
6082: PetscCall(MatTranspose(B, MAT_INITIAL_MATRIX, &Bt));
6083: PetscCall(MatMatMult(Bt, At, MAT_INITIAL_MATRIX, PETSC_CURRENT, &Ct));
6084: PetscCall(MatDestroy(&At));
6085: PetscCall(MatDestroy(&Bt));
6086: PetscCall(MatTransposeSetPrecursor(Ct, C));
6087: PetscCall(MatTranspose(Ct, MAT_REUSE_MATRIX, &C));
6088: PetscCall(MatDestroy(&Ct));
6089: PetscFunctionReturn(PETSC_SUCCESS);
6090: }
6092: static PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A, Mat B, PetscReal fill, Mat C)
6093: {
6094: PetscBool cisdense;
6096: PetscFunctionBegin;
6097: 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);
6098: PetscCall(MatSetSizes(C, A->rmap->n, B->cmap->n, A->rmap->N, B->cmap->N));
6099: PetscCall(MatSetBlockSizesFromMats(C, A, B));
6100: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATMPIDENSE, MATMPIDENSECUDA, MATMPIDENSEHIP, ""));
6101: if (!cisdense) PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
6102: PetscCall(MatSetUp(C));
6104: C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
6105: PetscFunctionReturn(PETSC_SUCCESS);
6106: }
6108: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
6109: {
6110: Mat_Product *product = C->product;
6111: Mat A = product->A, B = product->B;
6113: PetscFunctionBegin;
6114: 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 ")",
6115: A->cmap->rstart, A->cmap->rend, B->rmap->rstart, B->rmap->rend);
6116: C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
6117: C->ops->productsymbolic = MatProductSymbolic_AB;
6118: PetscFunctionReturn(PETSC_SUCCESS);
6119: }
6121: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
6122: {
6123: Mat_Product *product = C->product;
6125: PetscFunctionBegin;
6126: if (product->type == MATPRODUCT_AB) PetscCall(MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C));
6127: PetscFunctionReturn(PETSC_SUCCESS);
6128: }
6130: /*
6131: Merge two sets of sorted nonzeros and return a CSR for the merged (sequential) matrix
6133: Input Parameters:
6135: j1,rowBegin1,rowEnd1,jmap1: describe the first set of nonzeros (Set1)
6136: j2,rowBegin2,rowEnd2,jmap2: describe the second set of nonzeros (Set2)
6138: mat: both sets' nonzeros are on m rows, where m is the number of local rows of the matrix mat
6140: For Set1, j1[] contains column indices of the nonzeros.
6141: 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
6142: respectively (note rowEnd1[k] is not necessarily equal to rwoBegin1[k+1]). Indices in this range of j1[] are sorted,
6143: but might have repeats. jmap1[t+1] - jmap1[t] is the number of repeats for the t-th unique nonzero in Set1.
6145: Similar for Set2.
6147: This routine merges the two sets of nonzeros row by row and removes repeats.
6149: Output Parameters: (memory is allocated by the caller)
6151: i[],j[]: the CSR of the merged matrix, which has m rows.
6152: imap1[]: the k-th unique nonzero in Set1 (k=0,1,...) corresponds to imap1[k]-th unique nonzero in the merged matrix.
6153: imap2[]: similar to imap1[], but for Set2.
6154: Note we order nonzeros row-by-row and from left to right.
6155: */
6156: 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[])
6157: {
6158: PetscInt r, m; /* Row index of mat */
6159: PetscCount t, t1, t2, b1, e1, b2, e2;
6161: PetscFunctionBegin;
6162: PetscCall(MatGetLocalSize(mat, &m, NULL));
6163: t1 = t2 = t = 0; /* Count unique nonzeros of in Set1, Set1 and the merged respectively */
6164: i[0] = 0;
6165: for (r = 0; r < m; r++) { /* Do row by row merging */
6166: b1 = rowBegin1[r];
6167: e1 = rowEnd1[r];
6168: b2 = rowBegin2[r];
6169: e2 = rowEnd2[r];
6170: while (b1 < e1 && b2 < e2) {
6171: if (j1[b1] == j2[b2]) { /* Same column index and hence same nonzero */
6172: j[t] = j1[b1];
6173: imap1[t1] = t;
6174: imap2[t2] = t;
6175: b1 += jmap1[t1 + 1] - jmap1[t1]; /* Jump to next unique local nonzero */
6176: b2 += jmap2[t2 + 1] - jmap2[t2]; /* Jump to next unique remote nonzero */
6177: t1++;
6178: t2++;
6179: t++;
6180: } else if (j1[b1] < j2[b2]) {
6181: j[t] = j1[b1];
6182: imap1[t1] = t;
6183: b1 += jmap1[t1 + 1] - jmap1[t1];
6184: t1++;
6185: t++;
6186: } else {
6187: j[t] = j2[b2];
6188: imap2[t2] = t;
6189: b2 += jmap2[t2 + 1] - jmap2[t2];
6190: t2++;
6191: t++;
6192: }
6193: }
6194: /* Merge the remaining in either j1[] or j2[] */
6195: while (b1 < e1) {
6196: j[t] = j1[b1];
6197: imap1[t1] = t;
6198: b1 += jmap1[t1 + 1] - jmap1[t1];
6199: t1++;
6200: t++;
6201: }
6202: while (b2 < e2) {
6203: j[t] = j2[b2];
6204: imap2[t2] = t;
6205: b2 += jmap2[t2 + 1] - jmap2[t2];
6206: t2++;
6207: t++;
6208: }
6209: PetscCall(PetscIntCast(t, i + r + 1));
6210: }
6211: PetscFunctionReturn(PETSC_SUCCESS);
6212: }
6214: /*
6215: Split nonzeros in a block of local rows into two subsets: those in the diagonal block and those in the off-diagonal block
6217: Input Parameters:
6218: mat: an MPI matrix that provides row and column layout information for splitting. Let's say its number of local rows is m.
6219: 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[]
6220: respectively, along with a permutation array perm[]. Length of the i[],j[],perm[] arrays is n.
6222: i[] is already sorted, but within a row, j[] is not sorted and might have repeats.
6223: i[] might contain negative indices at the beginning, which means the corresponding entries should be ignored in the splitting.
6225: Output Parameters:
6226: j[],perm[]: the routine needs to sort j[] within each row along with perm[].
6227: rowBegin[],rowMid[],rowEnd[]: of length m, and the memory is preallocated and zeroed by the caller.
6228: 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,
6229: and [rowMid[r],rowEnd[r]) point to begin/end entries of row r of the off-diagonal block.
6231: Aperm[],Ajmap[],Atot,Annz: Arrays are allocated by this routine.
6232: Atot: number of entries belonging to the diagonal block.
6233: Annz: number of unique nonzeros belonging to the diagonal block.
6234: Aperm[Atot] stores values from perm[] for entries belonging to the diagonal block. Length of Aperm[] is Atot, though it may also count
6235: repeats (i.e., same 'i,j' pair).
6236: Ajmap[Annz+1] stores the number of repeats of each unique entry belonging to the diagonal block. More precisely, Ajmap[t+1] - Ajmap[t]
6237: is the number of repeats for the t-th unique entry in the diagonal block. Ajmap[0] is always 0.
6239: Atot: number of entries belonging to the diagonal block
6240: Annz: number of unique nonzeros belonging to the diagonal block.
6242: Bperm[], Bjmap[], Btot, Bnnz are similar but for the off-diagonal block.
6244: Aperm[],Bperm[],Ajmap[] and Bjmap[] are allocated separately by this routine with PetscMalloc1().
6245: */
6246: 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_)
6247: {
6248: PetscInt cstart, cend, rstart, rend, row, col;
6249: PetscCount Atot = 0, Btot = 0; /* Total number of nonzeros in the diagonal and off-diagonal blocks */
6250: PetscCount Annz = 0, Bnnz = 0; /* Number of unique nonzeros in the diagonal and off-diagonal blocks */
6251: PetscCount k, m, p, q, r, s, mid;
6252: PetscCount *Aperm, *Bperm, *Ajmap, *Bjmap;
6254: PetscFunctionBegin;
6255: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6256: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6257: m = rend - rstart;
6259: /* Skip negative rows */
6260: for (k = 0; k < n; k++)
6261: if (i[k] >= 0) break;
6263: /* Process [k,n): sort and partition each local row into diag and offdiag portions,
6264: fill rowBegin[], rowMid[], rowEnd[], and count Atot, Btot, Annz, Bnnz.
6265: */
6266: while (k < n) {
6267: row = i[k];
6268: /* Entries in [k,s) are in one row. Shift diagonal block col indices so that diag is ahead of offdiag after sorting the row */
6269: for (s = k; s < n; s++)
6270: if (i[s] != row) break;
6272: /* Shift diag columns to range of [-PETSC_INT_MAX, -1] */
6273: for (p = k; p < s; p++) {
6274: if (j[p] >= cstart && j[p] < cend) j[p] -= PETSC_INT_MAX;
6275: 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]);
6276: }
6277: PetscCall(PetscSortIntWithCountArray(s - k, j + k, perm + k));
6278: PetscCall(PetscSortedIntUpperBound(j, k, s, -1, &mid)); /* Separate [k,s) into [k,mid) for diag and [mid,s) for offdiag */
6279: rowBegin[row - rstart] = k;
6280: rowMid[row - rstart] = mid;
6281: rowEnd[row - rstart] = s;
6283: /* Count nonzeros of this diag/offdiag row, which might have repeats */
6284: Atot += mid - k;
6285: Btot += s - mid;
6287: /* Count unique nonzeros of this diag row */
6288: for (p = k; p < mid;) {
6289: col = j[p];
6290: do {
6291: j[p] += PETSC_INT_MAX; /* Revert the modified diagonal indices */
6292: p++;
6293: } while (p < mid && j[p] == col);
6294: Annz++;
6295: }
6297: /* Count unique nonzeros of this offdiag row */
6298: for (p = mid; p < s;) {
6299: col = j[p];
6300: do {
6301: p++;
6302: } while (p < s && j[p] == col);
6303: Bnnz++;
6304: }
6305: k = s;
6306: }
6308: /* Allocation according to Atot, Btot, Annz, Bnnz */
6309: PetscCall(PetscMalloc1(Atot, &Aperm));
6310: PetscCall(PetscMalloc1(Btot, &Bperm));
6311: PetscCall(PetscMalloc1(Annz + 1, &Ajmap));
6312: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap));
6314: /* Re-scan indices and copy diag/offdiag permutation indices to Aperm, Bperm and also fill Ajmap and Bjmap */
6315: Ajmap[0] = Bjmap[0] = Atot = Btot = Annz = Bnnz = 0;
6316: for (r = 0; r < m; r++) {
6317: k = rowBegin[r];
6318: mid = rowMid[r];
6319: s = rowEnd[r];
6320: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Aperm, Atot), PetscSafePointerPlusOffset(perm, k), mid - k));
6321: PetscCall(PetscArraycpy(PetscSafePointerPlusOffset(Bperm, Btot), PetscSafePointerPlusOffset(perm, mid), s - mid));
6322: Atot += mid - k;
6323: Btot += s - mid;
6325: /* Scan column indices in this row and find out how many repeats each unique nonzero has */
6326: for (p = k; p < mid;) {
6327: col = j[p];
6328: q = p;
6329: do {
6330: p++;
6331: } while (p < mid && j[p] == col);
6332: Ajmap[Annz + 1] = Ajmap[Annz] + (p - q);
6333: Annz++;
6334: }
6336: for (p = mid; p < s;) {
6337: col = j[p];
6338: q = p;
6339: do {
6340: p++;
6341: } while (p < s && j[p] == col);
6342: Bjmap[Bnnz + 1] = Bjmap[Bnnz] + (p - q);
6343: Bnnz++;
6344: }
6345: }
6346: /* Output */
6347: *Aperm_ = Aperm;
6348: *Annz_ = Annz;
6349: *Atot_ = Atot;
6350: *Ajmap_ = Ajmap;
6351: *Bperm_ = Bperm;
6352: *Bnnz_ = Bnnz;
6353: *Btot_ = Btot;
6354: *Bjmap_ = Bjmap;
6355: PetscFunctionReturn(PETSC_SUCCESS);
6356: }
6358: /*
6359: Expand the jmap[] array to make a new one in view of nonzeros in the merged matrix
6361: Input Parameters:
6362: nnz1: number of unique nonzeros in a set that was used to produce imap[], jmap[]
6363: nnz: number of unique nonzeros in the merged matrix
6364: imap[nnz1]: i-th nonzero in the set is the imap[i]-th nonzero in the merged matrix
6365: jmap[nnz1+1]: i-th nonzero in the set has jmap[i+1] - jmap[i] repeats in the set
6367: Output Parameter: (memory is allocated by the caller)
6368: jmap_new[nnz+1]: i-th nonzero in the merged matrix has jmap_new[i+1] - jmap_new[i] repeats in the set
6370: Example:
6371: nnz1 = 4
6372: nnz = 6
6373: imap = [1,3,4,5]
6374: jmap = [0,3,5,6,7]
6375: then,
6376: jmap_new = [0,0,3,3,5,6,7]
6377: */
6378: static PetscErrorCode ExpandJmap_Internal(PetscCount nnz1, PetscCount nnz, const PetscCount imap[], const PetscCount jmap[], PetscCount jmap_new[])
6379: {
6380: PetscCount k, p;
6382: PetscFunctionBegin;
6383: jmap_new[0] = 0;
6384: p = nnz; /* p loops over jmap_new[] backwards */
6385: for (k = nnz1 - 1; k >= 0; k--) { /* k loops over imap[] */
6386: for (; p > imap[k]; p--) jmap_new[p] = jmap[k + 1];
6387: }
6388: for (; p >= 0; p--) jmap_new[p] = jmap[0];
6389: PetscFunctionReturn(PETSC_SUCCESS);
6390: }
6392: static PetscErrorCode MatCOOStructDestroy_MPIAIJ(void *data)
6393: {
6394: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)data;
6396: PetscFunctionBegin;
6397: PetscCall(PetscSFDestroy(&coo->sf));
6398: PetscCall(PetscFree(coo->Aperm1));
6399: PetscCall(PetscFree(coo->Bperm1));
6400: PetscCall(PetscFree(coo->Ajmap1));
6401: PetscCall(PetscFree(coo->Bjmap1));
6402: PetscCall(PetscFree(coo->Aimap2));
6403: PetscCall(PetscFree(coo->Bimap2));
6404: PetscCall(PetscFree(coo->Aperm2));
6405: PetscCall(PetscFree(coo->Bperm2));
6406: PetscCall(PetscFree(coo->Ajmap2));
6407: PetscCall(PetscFree(coo->Bjmap2));
6408: PetscCall(PetscFree(coo->Cperm1));
6409: PetscCall(PetscFree2(coo->sendbuf, coo->recvbuf));
6410: PetscCall(PetscFree(coo));
6411: PetscFunctionReturn(PETSC_SUCCESS);
6412: }
6414: PetscErrorCode MatSetPreallocationCOO_MPIAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
6415: {
6416: MPI_Comm comm;
6417: PetscMPIInt rank, size;
6418: PetscInt m, n, M, N, rstart, rend, cstart, cend; /* Sizes, indices of row/col, therefore with type PetscInt */
6419: PetscCount k, p, q, rem; /* Loop variables over coo arrays */
6420: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6421: PetscContainer container;
6422: MatCOOStruct_MPIAIJ *coo;
6424: PetscFunctionBegin;
6425: PetscCall(PetscFree(mpiaij->garray));
6426: PetscCall(VecDestroy(&mpiaij->lvec));
6427: #if defined(PETSC_USE_CTABLE)
6428: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
6429: #else
6430: PetscCall(PetscFree(mpiaij->colmap));
6431: #endif
6432: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
6433: mat->assembled = PETSC_FALSE;
6434: mat->was_assembled = PETSC_FALSE;
6436: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
6437: PetscCallMPI(MPI_Comm_size(comm, &size));
6438: PetscCallMPI(MPI_Comm_rank(comm, &rank));
6439: PetscCall(PetscLayoutSetUp(mat->rmap));
6440: PetscCall(PetscLayoutSetUp(mat->cmap));
6441: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
6442: PetscCall(PetscLayoutGetRange(mat->cmap, &cstart, &cend));
6443: PetscCall(MatGetLocalSize(mat, &m, &n));
6444: PetscCall(MatGetSize(mat, &M, &N));
6446: /* Sort (i,j) by row along with a permutation array, so that the to-be-ignored */
6447: /* entries come first, then local rows, then remote rows. */
6448: PetscCount n1 = coo_n, *perm1;
6449: PetscInt *i1 = coo_i, *j1 = coo_j;
6451: PetscCall(PetscMalloc1(n1, &perm1));
6452: for (k = 0; k < n1; k++) perm1[k] = k;
6454: /* Manipulate indices so that entries with negative row or col indices will have smallest
6455: row indices, local entries will have greater but negative row indices, and remote entries
6456: will have positive row indices.
6457: */
6458: for (k = 0; k < n1; k++) {
6459: if (i1[k] < 0 || j1[k] < 0) i1[k] = PETSC_INT_MIN; /* e.g., -2^31, minimal to move them ahead */
6460: 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] */
6461: else {
6462: PetscCheck(!mat->nooffprocentries, PETSC_COMM_SELF, PETSC_ERR_USER_INPUT, "MAT_NO_OFF_PROC_ENTRIES is set but insert to remote rows");
6463: if (mpiaij->donotstash) i1[k] = PETSC_INT_MIN; /* Ignore offproc entries as if they had negative indices */
6464: }
6465: }
6467: /* Sort by row; after that, [0,k) have ignored entries, [k,rem) have local rows and [rem,n1) have remote rows */
6468: PetscCall(PetscSortIntWithIntCountArrayPair(n1, i1, j1, perm1));
6470: /* Advance k to the first entry we need to take care of */
6471: for (k = 0; k < n1; k++)
6472: if (i1[k] > PETSC_INT_MIN) break;
6473: PetscCount i1start = k;
6475: PetscCall(PetscSortedIntUpperBound(i1, k, n1, rend - 1 - PETSC_INT_MAX, &rem)); /* rem is upper bound of the last local row */
6476: for (; k < rem; k++) i1[k] += PETSC_INT_MAX; /* Revert row indices of local rows*/
6478: /* Send remote rows to their owner */
6479: /* Find which rows should be sent to which remote ranks*/
6480: PetscInt nsend = 0; /* Number of MPI ranks to send data to */
6481: PetscMPIInt *sendto; /* [nsend], storing remote ranks */
6482: PetscInt *nentries; /* [nsend], storing number of entries sent to remote ranks; Assume PetscInt is big enough for this count, and error if not */
6483: const PetscInt *ranges;
6484: PetscInt maxNsend = size >= 128 ? 128 : size; /* Assume max 128 neighbors; realloc when needed */
6486: PetscCall(PetscLayoutGetRanges(mat->rmap, &ranges));
6487: PetscCall(PetscMalloc2(maxNsend, &sendto, maxNsend, &nentries));
6488: for (k = rem; k < n1;) {
6489: PetscMPIInt owner;
6490: PetscInt firstRow, lastRow;
6492: /* Locate a row range */
6493: firstRow = i1[k]; /* first row of this owner */
6494: PetscCall(PetscLayoutFindOwner(mat->rmap, firstRow, &owner));
6495: lastRow = ranges[owner + 1] - 1; /* last row of this owner */
6497: /* Find the first index 'p' in [k,n) with i[p] belonging to next owner */
6498: PetscCall(PetscSortedIntUpperBound(i1, k, n1, lastRow, &p));
6500: /* All entries in [k,p) belong to this remote owner */
6501: if (nsend >= maxNsend) { /* Double the remote ranks arrays if not long enough */
6502: PetscMPIInt *sendto2;
6503: PetscInt *nentries2;
6504: PetscInt maxNsend2 = (maxNsend <= size / 2) ? maxNsend * 2 : size;
6506: PetscCall(PetscMalloc2(maxNsend2, &sendto2, maxNsend2, &nentries2));
6507: PetscCall(PetscArraycpy(sendto2, sendto, maxNsend));
6508: PetscCall(PetscArraycpy(nentries2, nentries2, maxNsend + 1));
6509: PetscCall(PetscFree2(sendto, nentries2));
6510: sendto = sendto2;
6511: nentries = nentries2;
6512: maxNsend = maxNsend2;
6513: }
6514: sendto[nsend] = owner;
6515: PetscCall(PetscIntCast(p - k, &nentries[nsend]));
6516: nsend++;
6517: k = p;
6518: }
6520: /* Build 1st SF to know offsets on remote to send data */
6521: PetscSF sf1;
6522: PetscInt nroots = 1, nroots2 = 0;
6523: PetscInt nleaves = nsend, nleaves2 = 0;
6524: PetscInt *offsets;
6525: PetscSFNode *iremote;
6527: PetscCall(PetscSFCreate(comm, &sf1));
6528: PetscCall(PetscMalloc1(nsend, &iremote));
6529: PetscCall(PetscMalloc1(nsend, &offsets));
6530: for (k = 0; k < nsend; k++) {
6531: iremote[k].rank = sendto[k];
6532: iremote[k].index = 0;
6533: nleaves2 += nentries[k];
6534: PetscCheck(nleaves2 >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF leaves is too large for PetscInt");
6535: }
6536: PetscCall(PetscSFSetGraph(sf1, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6537: PetscCall(PetscSFFetchAndOpWithMemTypeBegin(sf1, MPIU_INT, PETSC_MEMTYPE_HOST, &nroots2 /*rootdata*/, PETSC_MEMTYPE_HOST, nentries /*leafdata*/, PETSC_MEMTYPE_HOST, offsets /*leafupdate*/, MPI_SUM));
6538: PetscCall(PetscSFFetchAndOpEnd(sf1, MPIU_INT, &nroots2, nentries, offsets, MPI_SUM)); /* Would nroots2 overflow, we check offsets[] below */
6539: PetscCall(PetscSFDestroy(&sf1));
6540: PetscAssert(nleaves2 == n1 - rem, PETSC_COMM_SELF, PETSC_ERR_PLIB, "nleaves2 %" PetscInt_FMT " != number of remote entries %" PetscCount_FMT, nleaves2, n1 - rem);
6542: /* Build 2nd SF to send remote COOs to their owner */
6543: PetscSF sf2;
6544: nroots = nroots2;
6545: nleaves = nleaves2;
6546: PetscCall(PetscSFCreate(comm, &sf2));
6547: PetscCall(PetscSFSetFromOptions(sf2));
6548: PetscCall(PetscMalloc1(nleaves, &iremote));
6549: p = 0;
6550: for (k = 0; k < nsend; k++) {
6551: PetscCheck(offsets[k] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of SF roots is too large for PetscInt");
6552: for (q = 0; q < nentries[k]; q++, p++) {
6553: iremote[p].rank = sendto[k];
6554: PetscCall(PetscIntCast(offsets[k] + q, &iremote[p].index));
6555: }
6556: }
6557: PetscCall(PetscSFSetGraph(sf2, nroots, nleaves, NULL, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER));
6559: /* Send the remote COOs to their owner */
6560: PetscInt n2 = nroots, *i2, *j2; /* Buffers for received COOs from other ranks, along with a permutation array */
6561: PetscCount *perm2; /* Though PetscInt is enough for remote entries, we use PetscCount here as we want to reuse MatSplitEntries_Internal() */
6562: PetscCall(PetscMalloc3(n2, &i2, n2, &j2, n2, &perm2));
6563: PetscAssert(rem == 0 || i1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6564: PetscAssert(rem == 0 || j1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6565: PetscInt *i1prem = PetscSafePointerPlusOffset(i1, rem);
6566: PetscInt *j1prem = PetscSafePointerPlusOffset(j1, rem);
6567: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, i1prem, PETSC_MEMTYPE_HOST, i2, MPI_REPLACE));
6568: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, i1prem, i2, MPI_REPLACE));
6569: PetscCall(PetscSFReduceWithMemTypeBegin(sf2, MPIU_INT, PETSC_MEMTYPE_HOST, j1prem, PETSC_MEMTYPE_HOST, j2, MPI_REPLACE));
6570: PetscCall(PetscSFReduceEnd(sf2, MPIU_INT, j1prem, j2, MPI_REPLACE));
6572: PetscCall(PetscFree(offsets));
6573: PetscCall(PetscFree2(sendto, nentries));
6575: /* Sort received COOs by row along with the permutation array */
6576: for (k = 0; k < n2; k++) perm2[k] = k;
6577: PetscCall(PetscSortIntWithIntCountArrayPair(n2, i2, j2, perm2));
6579: /* sf2 only sends contiguous leafdata to contiguous rootdata. We record the permutation which will be used to fill leafdata */
6580: PetscCount *Cperm1;
6581: PetscAssert(rem == 0 || perm1 != NULL, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Cannot add nonzero offset to null");
6582: PetscCount *perm1prem = PetscSafePointerPlusOffset(perm1, rem);
6583: PetscCall(PetscMalloc1(nleaves, &Cperm1));
6584: PetscCall(PetscArraycpy(Cperm1, perm1prem, nleaves));
6586: /* Support for HYPRE matrices, kind of a hack.
6587: Swap min column with diagonal so that diagonal values will go first */
6588: PetscBool hypre;
6589: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
6590: if (hypre) {
6591: PetscInt *minj;
6592: PetscBT hasdiag;
6594: PetscCall(PetscBTCreate(m, &hasdiag));
6595: PetscCall(PetscMalloc1(m, &minj));
6596: for (k = 0; k < m; k++) minj[k] = PETSC_INT_MAX;
6597: for (k = i1start; k < rem; k++) {
6598: if (j1[k] < cstart || j1[k] >= cend) continue;
6599: const PetscInt rindex = i1[k] - rstart;
6600: if ((j1[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6601: minj[rindex] = PetscMin(minj[rindex], j1[k]);
6602: }
6603: for (k = 0; k < n2; k++) {
6604: if (j2[k] < cstart || j2[k] >= cend) continue;
6605: const PetscInt rindex = i2[k] - rstart;
6606: if ((j2[k] - cstart) == rindex) PetscCall(PetscBTSet(hasdiag, rindex));
6607: minj[rindex] = PetscMin(minj[rindex], j2[k]);
6608: }
6609: for (k = i1start; k < rem; k++) {
6610: const PetscInt rindex = i1[k] - rstart;
6611: if (j1[k] < cstart || j1[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6612: if (j1[k] == minj[rindex]) j1[k] = i1[k] + (cstart - rstart);
6613: else if ((j1[k] - cstart) == rindex) j1[k] = minj[rindex];
6614: }
6615: for (k = 0; k < n2; k++) {
6616: const PetscInt rindex = i2[k] - rstart;
6617: if (j2[k] < cstart || j2[k] >= cend || !PetscBTLookup(hasdiag, rindex)) continue;
6618: if (j2[k] == minj[rindex]) j2[k] = i2[k] + (cstart - rstart);
6619: else if ((j2[k] - cstart) == rindex) j2[k] = minj[rindex];
6620: }
6621: PetscCall(PetscBTDestroy(&hasdiag));
6622: PetscCall(PetscFree(minj));
6623: }
6625: /* Split local COOs and received COOs into diag/offdiag portions */
6626: PetscCount *rowBegin1, *rowMid1, *rowEnd1;
6627: PetscCount *Ajmap1, *Aperm1, *Bjmap1, *Bperm1;
6628: PetscCount Annz1, Bnnz1, Atot1, Btot1;
6629: PetscCount *rowBegin2, *rowMid2, *rowEnd2;
6630: PetscCount *Ajmap2, *Aperm2, *Bjmap2, *Bperm2;
6631: PetscCount Annz2, Bnnz2, Atot2, Btot2;
6633: PetscCall(PetscCalloc3(m, &rowBegin1, m, &rowMid1, m, &rowEnd1));
6634: PetscCall(PetscCalloc3(m, &rowBegin2, m, &rowMid2, m, &rowEnd2));
6635: PetscCall(MatSplitEntries_Internal(mat, rem, i1, j1, perm1, rowBegin1, rowMid1, rowEnd1, &Atot1, &Aperm1, &Annz1, &Ajmap1, &Btot1, &Bperm1, &Bnnz1, &Bjmap1));
6636: PetscCall(MatSplitEntries_Internal(mat, n2, i2, j2, perm2, rowBegin2, rowMid2, rowEnd2, &Atot2, &Aperm2, &Annz2, &Ajmap2, &Btot2, &Bperm2, &Bnnz2, &Bjmap2));
6638: /* Merge local COOs with received COOs: diag with diag, offdiag with offdiag */
6639: PetscInt *Ai, *Bi;
6640: PetscInt *Aj, *Bj;
6642: PetscCall(PetscMalloc1(m + 1, &Ai));
6643: PetscCall(PetscMalloc1(m + 1, &Bi));
6644: PetscCall(PetscMalloc1(Annz1 + Annz2, &Aj)); /* Since local and remote entries might have dups, we might allocate excess memory */
6645: PetscCall(PetscMalloc1(Bnnz1 + Bnnz2, &Bj));
6647: PetscCount *Aimap1, *Bimap1, *Aimap2, *Bimap2;
6648: PetscCall(PetscMalloc1(Annz1, &Aimap1));
6649: PetscCall(PetscMalloc1(Bnnz1, &Bimap1));
6650: PetscCall(PetscMalloc1(Annz2, &Aimap2));
6651: PetscCall(PetscMalloc1(Bnnz2, &Bimap2));
6653: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowBegin1, rowMid1, rowBegin2, rowMid2, Ajmap1, Ajmap2, Aimap1, Aimap2, Ai, Aj));
6654: PetscCall(MatMergeEntries_Internal(mat, j1, j2, rowMid1, rowEnd1, rowMid2, rowEnd2, Bjmap1, Bjmap2, Bimap1, Bimap2, Bi, Bj));
6656: /* Expand Ajmap1/Bjmap1 to make them based off nonzeros in A/B, since we */
6657: /* expect nonzeros in A/B most likely have local contributing entries */
6658: PetscInt Annz = Ai[m];
6659: PetscInt Bnnz = Bi[m];
6660: PetscCount *Ajmap1_new, *Bjmap1_new;
6662: PetscCall(PetscMalloc1(Annz + 1, &Ajmap1_new));
6663: PetscCall(PetscMalloc1(Bnnz + 1, &Bjmap1_new));
6665: PetscCall(ExpandJmap_Internal(Annz1, Annz, Aimap1, Ajmap1, Ajmap1_new));
6666: PetscCall(ExpandJmap_Internal(Bnnz1, Bnnz, Bimap1, Bjmap1, Bjmap1_new));
6668: PetscCall(PetscFree(Aimap1));
6669: PetscCall(PetscFree(Ajmap1));
6670: PetscCall(PetscFree(Bimap1));
6671: PetscCall(PetscFree(Bjmap1));
6672: PetscCall(PetscFree3(rowBegin1, rowMid1, rowEnd1));
6673: PetscCall(PetscFree3(rowBegin2, rowMid2, rowEnd2));
6674: PetscCall(PetscFree(perm1));
6675: PetscCall(PetscFree3(i2, j2, perm2));
6677: Ajmap1 = Ajmap1_new;
6678: Bjmap1 = Bjmap1_new;
6680: /* Reallocate Aj, Bj once we know actual numbers of unique nonzeros in A and B */
6681: if (Annz < Annz1 + Annz2) {
6682: PetscInt *Aj_new;
6683: PetscCall(PetscMalloc1(Annz, &Aj_new));
6684: PetscCall(PetscArraycpy(Aj_new, Aj, Annz));
6685: PetscCall(PetscFree(Aj));
6686: Aj = Aj_new;
6687: }
6689: if (Bnnz < Bnnz1 + Bnnz2) {
6690: PetscInt *Bj_new;
6691: PetscCall(PetscMalloc1(Bnnz, &Bj_new));
6692: PetscCall(PetscArraycpy(Bj_new, Bj, Bnnz));
6693: PetscCall(PetscFree(Bj));
6694: Bj = Bj_new;
6695: }
6697: /* Create new submatrices for on-process and off-process coupling */
6698: PetscScalar *Aa, *Ba;
6699: MatType rtype;
6700: Mat_SeqAIJ *a, *b;
6701: PetscObjectState state;
6702: PetscCall(PetscCalloc1(Annz, &Aa)); /* Zero matrix on device */
6703: PetscCall(PetscCalloc1(Bnnz, &Ba));
6704: /* make Aj[] local, i.e, based off the start column of the diagonal portion */
6705: if (cstart) {
6706: for (k = 0; k < Annz; k++) Aj[k] -= cstart;
6707: }
6709: PetscCall(MatGetRootType_Private(mat, &rtype));
6711: MatSeqXAIJGetOptions_Private(mpiaij->A);
6712: PetscCall(MatDestroy(&mpiaij->A));
6713: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, Ai, Aj, Aa, &mpiaij->A));
6714: PetscCall(MatSetBlockSizesFromMats(mpiaij->A, mat, mat));
6715: MatSeqXAIJRestoreOptions_Private(mpiaij->A);
6717: MatSeqXAIJGetOptions_Private(mpiaij->B);
6718: PetscCall(MatDestroy(&mpiaij->B));
6719: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, mat->cmap->N, Bi, Bj, Ba, &mpiaij->B));
6720: PetscCall(MatSetBlockSizesFromMats(mpiaij->B, mat, mat));
6721: MatSeqXAIJRestoreOptions_Private(mpiaij->B);
6723: PetscCall(MatSetUpMultiply_MPIAIJ(mat));
6724: mat->was_assembled = PETSC_TRUE; // was_assembled in effect means the Mvctx is built; doing so avoids redundant MatSetUpMultiply_MPIAIJ
6725: state = mpiaij->A->nonzerostate + mpiaij->B->nonzerostate;
6726: PetscCallMPI(MPIU_Allreduce(&state, &mat->nonzerostate, 1, MPIU_INT64, MPI_SUM, PetscObjectComm((PetscObject)mat)));
6728: a = (Mat_SeqAIJ *)mpiaij->A->data;
6729: b = (Mat_SeqAIJ *)mpiaij->B->data;
6730: a->free_a = PETSC_TRUE;
6731: a->free_ij = PETSC_TRUE;
6732: b->free_a = PETSC_TRUE;
6733: b->free_ij = PETSC_TRUE;
6734: a->maxnz = a->nz;
6735: b->maxnz = b->nz;
6737: /* conversion must happen AFTER multiply setup */
6738: PetscCall(MatConvert(mpiaij->A, rtype, MAT_INPLACE_MATRIX, &mpiaij->A));
6739: PetscCall(MatConvert(mpiaij->B, rtype, MAT_INPLACE_MATRIX, &mpiaij->B));
6740: PetscCall(VecDestroy(&mpiaij->lvec));
6741: PetscCall(MatCreateVecs(mpiaij->B, &mpiaij->lvec, NULL));
6743: // Put the COO struct in a container and then attach that to the matrix
6744: PetscCall(PetscMalloc1(1, &coo));
6745: coo->n = coo_n;
6746: coo->sf = sf2;
6747: coo->sendlen = nleaves;
6748: coo->recvlen = nroots;
6749: coo->Annz = Annz;
6750: coo->Bnnz = Bnnz;
6751: coo->Annz2 = Annz2;
6752: coo->Bnnz2 = Bnnz2;
6753: coo->Atot1 = Atot1;
6754: coo->Atot2 = Atot2;
6755: coo->Btot1 = Btot1;
6756: coo->Btot2 = Btot2;
6757: coo->Ajmap1 = Ajmap1;
6758: coo->Aperm1 = Aperm1;
6759: coo->Bjmap1 = Bjmap1;
6760: coo->Bperm1 = Bperm1;
6761: coo->Aimap2 = Aimap2;
6762: coo->Ajmap2 = Ajmap2;
6763: coo->Aperm2 = Aperm2;
6764: coo->Bimap2 = Bimap2;
6765: coo->Bjmap2 = Bjmap2;
6766: coo->Bperm2 = Bperm2;
6767: coo->Cperm1 = Cperm1;
6768: // Allocate in preallocation. If not used, it has zero cost on host
6769: PetscCall(PetscMalloc2(coo->sendlen, &coo->sendbuf, coo->recvlen, &coo->recvbuf));
6770: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
6771: PetscCall(PetscContainerSetPointer(container, coo));
6772: PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_MPIAIJ));
6773: PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container));
6774: PetscCall(PetscContainerDestroy(&container));
6775: PetscFunctionReturn(PETSC_SUCCESS);
6776: }
6778: static PetscErrorCode MatSetValuesCOO_MPIAIJ(Mat mat, const PetscScalar v[], InsertMode imode)
6779: {
6780: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
6781: Mat A = mpiaij->A, B = mpiaij->B;
6782: PetscScalar *Aa, *Ba;
6783: PetscScalar *sendbuf, *recvbuf;
6784: const PetscCount *Ajmap1, *Ajmap2, *Aimap2;
6785: const PetscCount *Bjmap1, *Bjmap2, *Bimap2;
6786: const PetscCount *Aperm1, *Aperm2, *Bperm1, *Bperm2;
6787: const PetscCount *Cperm1;
6788: PetscContainer container;
6789: MatCOOStruct_MPIAIJ *coo;
6791: PetscFunctionBegin;
6792: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
6793: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
6794: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
6795: sendbuf = coo->sendbuf;
6796: recvbuf = coo->recvbuf;
6797: Ajmap1 = coo->Ajmap1;
6798: Ajmap2 = coo->Ajmap2;
6799: Aimap2 = coo->Aimap2;
6800: Bjmap1 = coo->Bjmap1;
6801: Bjmap2 = coo->Bjmap2;
6802: Bimap2 = coo->Bimap2;
6803: Aperm1 = coo->Aperm1;
6804: Aperm2 = coo->Aperm2;
6805: Bperm1 = coo->Bperm1;
6806: Bperm2 = coo->Bperm2;
6807: Cperm1 = coo->Cperm1;
6809: PetscCall(MatSeqAIJGetArray(A, &Aa)); /* Might read and write matrix values */
6810: PetscCall(MatSeqAIJGetArray(B, &Ba));
6812: /* Pack entries to be sent to remote */
6813: for (PetscCount i = 0; i < coo->sendlen; i++) sendbuf[i] = v[Cperm1[i]];
6815: /* Send remote entries to their owner and overlap the communication with local computation */
6816: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_HOST, sendbuf, PETSC_MEMTYPE_HOST, recvbuf, MPI_REPLACE));
6817: /* Add local entries to A and B */
6818: for (PetscCount i = 0; i < coo->Annz; i++) { /* All nonzeros in A are either zero'ed or added with a value (i.e., initialized) */
6819: PetscScalar sum = 0.0; /* Do partial summation first to improve numerical stability */
6820: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += v[Aperm1[k]];
6821: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
6822: }
6823: for (PetscCount i = 0; i < coo->Bnnz; i++) {
6824: PetscScalar sum = 0.0;
6825: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += v[Bperm1[k]];
6826: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
6827: }
6828: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, sendbuf, recvbuf, MPI_REPLACE));
6830: /* Add received remote entries to A and B */
6831: for (PetscCount i = 0; i < coo->Annz2; i++) {
6832: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += recvbuf[Aperm2[k]];
6833: }
6834: for (PetscCount i = 0; i < coo->Bnnz2; i++) {
6835: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += recvbuf[Bperm2[k]];
6836: }
6837: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
6838: PetscCall(MatSeqAIJRestoreArray(B, &Ba));
6839: PetscFunctionReturn(PETSC_SUCCESS);
6840: }
6842: /*MC
6843: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
6845: Options Database Keys:
6846: . -mat_type mpiaij - sets the matrix type to `MATMPIAIJ` during a call to `MatSetFromOptions()`
6848: Level: beginner
6850: Notes:
6851: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
6852: in this case the values associated with the rows and columns one passes in are set to zero
6853: in the matrix
6855: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
6856: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
6858: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `MATAIJ`, `MatCreateAIJ()`
6859: M*/
6860: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
6861: {
6862: Mat_MPIAIJ *b;
6863: PetscMPIInt size;
6865: PetscFunctionBegin;
6866: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
6868: PetscCall(PetscNew(&b));
6869: B->data = (void *)b;
6870: B->ops[0] = MatOps_Values;
6871: B->assembled = PETSC_FALSE;
6872: B->insertmode = NOT_SET_VALUES;
6873: b->size = size;
6875: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)B), &b->rank));
6877: /* build cache for off array entries formed */
6878: PetscCall(MatStashCreate_Private(PetscObjectComm((PetscObject)B), 1, &B->stash));
6880: b->donotstash = PETSC_FALSE;
6881: b->colmap = NULL;
6882: b->garray = NULL;
6883: b->roworiented = PETSC_TRUE;
6885: /* stuff used for matrix vector multiply */
6886: b->lvec = NULL;
6887: b->Mvctx = NULL;
6889: /* stuff for MatGetRow() */
6890: b->rowindices = NULL;
6891: b->rowvalues = NULL;
6892: b->getrowactive = PETSC_FALSE;
6894: /* flexible pointer used in CUSPARSE classes */
6895: b->spptr = NULL;
6897: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetUseScalableIncreaseOverlap_C", MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ));
6898: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_MPIAIJ));
6899: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_MPIAIJ));
6900: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_MPIAIJ));
6901: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJ));
6902: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_MPIAIJ));
6903: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMPIAIJSetPreallocationCSR_C", MatMPIAIJSetPreallocationCSR_MPIAIJ));
6904: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatDiagonalScaleLocal_C", MatDiagonalScaleLocal_MPIAIJ));
6905: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijperm_C", MatConvert_MPIAIJ_MPIAIJPERM));
6906: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijsell_C", MatConvert_MPIAIJ_MPIAIJSELL));
6907: #if defined(PETSC_HAVE_CUDA)
6908: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcusparse_C", MatConvert_MPIAIJ_MPIAIJCUSPARSE));
6909: #endif
6910: #if defined(PETSC_HAVE_HIP)
6911: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijhipsparse_C", MatConvert_MPIAIJ_MPIAIJHIPSPARSE));
6912: #endif
6913: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
6914: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijkokkos_C", MatConvert_MPIAIJ_MPIAIJKokkos));
6915: #endif
6916: #if defined(PETSC_HAVE_MKL_SPARSE)
6917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijmkl_C", MatConvert_MPIAIJ_MPIAIJMKL));
6918: #endif
6919: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpiaijcrl_C", MatConvert_MPIAIJ_MPIAIJCRL));
6920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpibaij_C", MatConvert_MPIAIJ_MPIBAIJ));
6921: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisbaij_C", MatConvert_MPIAIJ_MPISBAIJ));
6922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpidense_C", MatConvert_MPIAIJ_MPIDense));
6923: #if defined(PETSC_HAVE_ELEMENTAL)
6924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_elemental_C", MatConvert_MPIAIJ_Elemental));
6925: #endif
6926: #if defined(PETSC_HAVE_SCALAPACK)
6927: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
6928: #endif
6929: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_is_C", MatConvert_XAIJ_IS));
6930: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_mpisell_C", MatConvert_MPIAIJ_MPISELL));
6931: #if defined(PETSC_HAVE_HYPRE)
6932: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_mpiaij_hypre_C", MatConvert_AIJ_HYPRE));
6933: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_mpiaij_mpiaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
6934: #endif
6935: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_mpiaij_C", MatProductSetFromOptions_IS_XAIJ));
6936: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_mpiaij_mpiaij_C", MatProductSetFromOptions_MPIAIJ));
6937: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJ));
6938: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJ));
6939: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATMPIAIJ));
6940: PetscFunctionReturn(PETSC_SUCCESS);
6941: }
6943: /*@
6944: MatCreateMPIAIJWithSplitArrays - creates a `MATMPIAIJ` matrix using arrays that contain the "diagonal"
6945: and "off-diagonal" part of the matrix in CSR format.
6947: Collective
6949: Input Parameters:
6950: + comm - MPI communicator
6951: . m - number of local rows (Cannot be `PETSC_DECIDE`)
6952: . n - This value should be the same as the local size used in creating the
6953: x vector for the matrix-vector product $y = Ax$. (or `PETSC_DECIDE` to have
6954: calculated if `N` is given) For square matrices `n` is almost always `m`.
6955: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
6956: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
6957: . 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
6958: . j - column indices, which must be local, i.e., based off the start column of the diagonal portion
6959: . a - matrix values
6960: . 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
6961: . oj - column indices, which must be global, representing global columns in the `MATMPIAIJ` matrix
6962: - oa - matrix values
6964: Output Parameter:
6965: . mat - the matrix
6967: Level: advanced
6969: Notes:
6970: The `i`, `j`, and `a` arrays ARE NOT copied by this routine into the internal format used by PETSc (even in Fortran). The user
6971: must free the arrays once the matrix has been destroyed and not before.
6973: The `i` and `j` indices are 0 based
6975: See `MatCreateAIJ()` for the definition of "diagonal" and "off-diagonal" portion of the matrix
6977: This sets local rows and cannot be used to set off-processor values.
6979: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
6980: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
6981: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
6982: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
6983: keep track of the underlying array. Use `MatSetOption`(A,`MAT_NO_OFF_PROC_ENTRIES`,`PETSC_TRUE`) to disable all
6984: communication if it is known that only local entries will be set.
6986: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatMPIAIJSetPreallocation()`, `MatMPIAIJSetPreallocationCSR()`,
6987: `MATMPIAIJ`, `MatCreateAIJ()`, `MatCreateMPIAIJWithArrays()`
6988: @*/
6989: 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)
6990: {
6991: Mat_MPIAIJ *maij;
6993: PetscFunctionBegin;
6994: PetscCheck(m >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "local number of rows (m) cannot be PETSC_DECIDE, or negative");
6995: PetscCheck(i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
6996: PetscCheck(oi[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "oi (row indices) must start with 0");
6997: PetscCall(MatCreate(comm, mat));
6998: PetscCall(MatSetSizes(*mat, m, n, M, N));
6999: PetscCall(MatSetType(*mat, MATMPIAIJ));
7000: maij = (Mat_MPIAIJ *)(*mat)->data;
7002: (*mat)->preallocated = PETSC_TRUE;
7004: PetscCall(PetscLayoutSetUp((*mat)->rmap));
7005: PetscCall(PetscLayoutSetUp((*mat)->cmap));
7007: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, n, i, j, a, &maij->A));
7008: PetscCall(MatCreateSeqAIJWithArrays(PETSC_COMM_SELF, m, (*mat)->cmap->N, oi, oj, oa, &maij->B));
7010: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE));
7011: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
7012: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
7013: PetscCall(MatSetOption(*mat, MAT_NO_OFF_PROC_ENTRIES, PETSC_FALSE));
7014: PetscCall(MatSetOption(*mat, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
7015: PetscFunctionReturn(PETSC_SUCCESS);
7016: }
7018: typedef struct {
7019: Mat *mp; /* intermediate products */
7020: PetscBool *mptmp; /* is the intermediate product temporary ? */
7021: PetscInt cp; /* number of intermediate products */
7023: /* support for MatGetBrowsOfAoCols_MPIAIJ for P_oth */
7024: PetscInt *startsj_s, *startsj_r;
7025: PetscScalar *bufa;
7026: Mat P_oth;
7028: /* may take advantage of merging product->B */
7029: Mat Bloc; /* B-local by merging diag and off-diag */
7031: /* cusparse does not have support to split between symbolic and numeric phases.
7032: When api_user is true, we don't need to update the numerical values
7033: of the temporary storage */
7034: PetscBool reusesym;
7036: /* support for COO values insertion */
7037: PetscScalar *coo_v, *coo_w; /* store on-process and off-process COO scalars, and used as MPI recv/send buffers respectively */
7038: PetscInt **own; /* own[i] points to address of on-process COO indices for Mat mp[i] */
7039: PetscInt **off; /* off[i] points to address of off-process COO indices for Mat mp[i] */
7040: PetscBool hasoffproc; /* if true, have off-process values insertion (i.e. AtB or PtAP) */
7041: PetscSF sf; /* used for non-local values insertion and memory malloc */
7042: PetscMemType mtype;
7044: /* customization */
7045: PetscBool abmerge;
7046: PetscBool P_oth_bind;
7047: } MatMatMPIAIJBACKEND;
7049: static PetscErrorCode MatDestroy_MatMatMPIAIJBACKEND(void *data)
7050: {
7051: MatMatMPIAIJBACKEND *mmdata = (MatMatMPIAIJBACKEND *)data;
7052: PetscInt i;
7054: PetscFunctionBegin;
7055: PetscCall(PetscFree2(mmdata->startsj_s, mmdata->startsj_r));
7056: PetscCall(PetscFree(mmdata->bufa));
7057: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_v));
7058: PetscCall(PetscSFFree(mmdata->sf, mmdata->mtype, mmdata->coo_w));
7059: PetscCall(MatDestroy(&mmdata->P_oth));
7060: PetscCall(MatDestroy(&mmdata->Bloc));
7061: PetscCall(PetscSFDestroy(&mmdata->sf));
7062: for (i = 0; i < mmdata->cp; i++) PetscCall(MatDestroy(&mmdata->mp[i]));
7063: PetscCall(PetscFree2(mmdata->mp, mmdata->mptmp));
7064: PetscCall(PetscFree(mmdata->own[0]));
7065: PetscCall(PetscFree(mmdata->own));
7066: PetscCall(PetscFree(mmdata->off[0]));
7067: PetscCall(PetscFree(mmdata->off));
7068: PetscCall(PetscFree(mmdata));
7069: PetscFunctionReturn(PETSC_SUCCESS);
7070: }
7072: /* Copy selected n entries with indices in idx[] of A to v[].
7073: If idx is NULL, copy the whole data array of A to v[]
7074: */
7075: static PetscErrorCode MatSeqAIJCopySubArray(Mat A, PetscInt n, const PetscInt idx[], PetscScalar v[])
7076: {
7077: PetscErrorCode (*f)(Mat, PetscInt, const PetscInt[], PetscScalar[]);
7079: PetscFunctionBegin;
7080: PetscCall(PetscObjectQueryFunction((PetscObject)A, "MatSeqAIJCopySubArray_C", &f));
7081: if (f) {
7082: PetscCall((*f)(A, n, idx, v));
7083: } else {
7084: const PetscScalar *vv;
7086: PetscCall(MatSeqAIJGetArrayRead(A, &vv));
7087: if (n && idx) {
7088: PetscScalar *w = v;
7089: const PetscInt *oi = idx;
7090: PetscInt j;
7092: for (j = 0; j < n; j++) *w++ = vv[*oi++];
7093: } else {
7094: PetscCall(PetscArraycpy(v, vv, n));
7095: }
7096: PetscCall(MatSeqAIJRestoreArrayRead(A, &vv));
7097: }
7098: PetscFunctionReturn(PETSC_SUCCESS);
7099: }
7101: static PetscErrorCode MatProductNumeric_MPIAIJBACKEND(Mat C)
7102: {
7103: MatMatMPIAIJBACKEND *mmdata;
7104: PetscInt i, n_d, n_o;
7106: PetscFunctionBegin;
7107: MatCheckProduct(C, 1);
7108: PetscCheck(C->product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data empty");
7109: mmdata = (MatMatMPIAIJBACKEND *)C->product->data;
7110: if (!mmdata->reusesym) { /* update temporary matrices */
7111: 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));
7112: if (mmdata->Bloc) PetscCall(MatMPIAIJGetLocalMatMerge(C->product->B, MAT_REUSE_MATRIX, NULL, &mmdata->Bloc));
7113: }
7114: mmdata->reusesym = PETSC_FALSE;
7116: for (i = 0; i < mmdata->cp; i++) {
7117: 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]);
7118: PetscCall((*mmdata->mp[i]->ops->productnumeric)(mmdata->mp[i]));
7119: }
7120: for (i = 0, n_d = 0, n_o = 0; i < mmdata->cp; i++) {
7121: PetscInt noff;
7123: PetscCall(PetscIntCast(mmdata->off[i + 1] - mmdata->off[i], &noff));
7124: if (mmdata->mptmp[i]) continue;
7125: if (noff) {
7126: PetscInt nown;
7128: PetscCall(PetscIntCast(mmdata->own[i + 1] - mmdata->own[i], &nown));
7129: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], noff, mmdata->off[i], mmdata->coo_w + n_o));
7130: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], nown, mmdata->own[i], mmdata->coo_v + n_d));
7131: n_o += noff;
7132: n_d += nown;
7133: } else {
7134: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mmdata->mp[i]->data;
7136: PetscCall(MatSeqAIJCopySubArray(mmdata->mp[i], mm->nz, NULL, mmdata->coo_v + n_d));
7137: n_d += mm->nz;
7138: }
7139: }
7140: if (mmdata->hasoffproc) { /* offprocess insertion */
7141: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7142: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_SCALAR, mmdata->coo_w, mmdata->coo_v + n_d));
7143: }
7144: PetscCall(MatSetValuesCOO(C, mmdata->coo_v, INSERT_VALUES));
7145: PetscFunctionReturn(PETSC_SUCCESS);
7146: }
7148: /* Support for Pt * A, A * P, or Pt * A * P */
7149: #define MAX_NUMBER_INTERMEDIATE 4
7150: PetscErrorCode MatProductSymbolic_MPIAIJBACKEND(Mat C)
7151: {
7152: Mat_Product *product = C->product;
7153: Mat A, P, mp[MAX_NUMBER_INTERMEDIATE]; /* A, P and a series of intermediate matrices */
7154: Mat_MPIAIJ *a, *p;
7155: MatMatMPIAIJBACKEND *mmdata;
7156: ISLocalToGlobalMapping P_oth_l2g = NULL;
7157: IS glob = NULL;
7158: const char *prefix;
7159: char pprefix[256];
7160: const PetscInt *globidx, *P_oth_idx;
7161: PetscInt i, j, cp, m, n, M, N, *coo_i, *coo_j;
7162: PetscCount ncoo, ncoo_d, ncoo_o, ncoo_oown;
7163: PetscInt cmapt[MAX_NUMBER_INTERMEDIATE], rmapt[MAX_NUMBER_INTERMEDIATE]; /* col/row map type for each Mat in mp[]. */
7164: /* type-0: consecutive, start from 0; type-1: consecutive with */
7165: /* a base offset; type-2: sparse with a local to global map table */
7166: const PetscInt *cmapa[MAX_NUMBER_INTERMEDIATE], *rmapa[MAX_NUMBER_INTERMEDIATE]; /* col/row local to global map array (table) for type-2 map type */
7168: MatProductType ptype;
7169: PetscBool mptmp[MAX_NUMBER_INTERMEDIATE], hasoffproc = PETSC_FALSE, iscuda, iship, iskokk;
7170: PetscMPIInt size;
7172: PetscFunctionBegin;
7173: MatCheckProduct(C, 1);
7174: PetscCheck(!product->data, PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Product data not empty");
7175: ptype = product->type;
7176: if (product->A->symmetric == PETSC_BOOL3_TRUE && ptype == MATPRODUCT_AtB) {
7177: ptype = MATPRODUCT_AB;
7178: product->symbolic_used_the_fact_A_is_symmetric = PETSC_TRUE;
7179: }
7180: switch (ptype) {
7181: case MATPRODUCT_AB:
7182: A = product->A;
7183: P = product->B;
7184: m = A->rmap->n;
7185: n = P->cmap->n;
7186: M = A->rmap->N;
7187: N = P->cmap->N;
7188: hasoffproc = PETSC_FALSE; /* will not scatter mat product values to other processes */
7189: break;
7190: case MATPRODUCT_AtB:
7191: P = product->A;
7192: A = product->B;
7193: m = P->cmap->n;
7194: n = A->cmap->n;
7195: M = P->cmap->N;
7196: N = A->cmap->N;
7197: hasoffproc = PETSC_TRUE;
7198: break;
7199: case MATPRODUCT_PtAP:
7200: A = product->A;
7201: P = product->B;
7202: m = P->cmap->n;
7203: n = P->cmap->n;
7204: M = P->cmap->N;
7205: N = P->cmap->N;
7206: hasoffproc = PETSC_TRUE;
7207: break;
7208: default:
7209: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7210: }
7211: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)C), &size));
7212: if (size == 1) hasoffproc = PETSC_FALSE;
7214: /* defaults */
7215: for (i = 0; i < MAX_NUMBER_INTERMEDIATE; i++) {
7216: mp[i] = NULL;
7217: mptmp[i] = PETSC_FALSE;
7218: rmapt[i] = -1;
7219: cmapt[i] = -1;
7220: rmapa[i] = NULL;
7221: cmapa[i] = NULL;
7222: }
7224: /* customization */
7225: PetscCall(PetscNew(&mmdata));
7226: mmdata->reusesym = product->api_user;
7227: if (ptype == MATPRODUCT_AB) {
7228: if (product->api_user) {
7229: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatMatMult", "Mat");
7230: PetscCall(PetscOptionsBool("-matmatmult_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7231: PetscCall(PetscOptionsBool("-matmatmult_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7232: PetscOptionsEnd();
7233: } else {
7234: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_AB", "Mat");
7235: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_mergeB", "Merge product->B local matrices", "MatMatMult", mmdata->abmerge, &mmdata->abmerge, NULL));
7236: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7237: PetscOptionsEnd();
7238: }
7239: } else if (ptype == MATPRODUCT_PtAP) {
7240: if (product->api_user) {
7241: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatPtAP", "Mat");
7242: PetscCall(PetscOptionsBool("-matptap_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7243: PetscOptionsEnd();
7244: } else {
7245: PetscOptionsBegin(PetscObjectComm((PetscObject)C), ((PetscObject)C)->prefix, "MatProduct_PtAP", "Mat");
7246: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_pothbind", "Bind P_oth to CPU", "MatBindToCPU", mmdata->P_oth_bind, &mmdata->P_oth_bind, NULL));
7247: PetscOptionsEnd();
7248: }
7249: }
7250: a = (Mat_MPIAIJ *)A->data;
7251: p = (Mat_MPIAIJ *)P->data;
7252: PetscCall(MatSetSizes(C, m, n, M, N));
7253: PetscCall(PetscLayoutSetUp(C->rmap));
7254: PetscCall(PetscLayoutSetUp(C->cmap));
7255: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
7256: PetscCall(MatGetOptionsPrefix(C, &prefix));
7258: cp = 0;
7259: switch (ptype) {
7260: case MATPRODUCT_AB: /* A * P */
7261: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7263: /* A_diag * P_local (merged or not) */
7264: if (mmdata->abmerge) { /* P's diagonal and off-diag blocks are merged to one matrix, then multiplied by A_diag */
7265: /* P is product->B */
7266: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7267: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7268: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7269: PetscCall(MatProductSetFill(mp[cp], product->fill));
7270: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7271: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7272: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7273: mp[cp]->product->api_user = product->api_user;
7274: PetscCall(MatProductSetFromOptions(mp[cp]));
7275: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7276: PetscCall(ISGetIndices(glob, &globidx));
7277: rmapt[cp] = 1;
7278: cmapt[cp] = 2;
7279: cmapa[cp] = globidx;
7280: mptmp[cp] = PETSC_FALSE;
7281: cp++;
7282: } else { /* A_diag * P_diag and A_diag * P_off */
7283: PetscCall(MatProductCreate(a->A, p->A, NULL, &mp[cp]));
7284: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7285: PetscCall(MatProductSetFill(mp[cp], product->fill));
7286: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7287: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7288: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7289: mp[cp]->product->api_user = product->api_user;
7290: PetscCall(MatProductSetFromOptions(mp[cp]));
7291: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7292: rmapt[cp] = 1;
7293: cmapt[cp] = 1;
7294: mptmp[cp] = PETSC_FALSE;
7295: cp++;
7296: PetscCall(MatProductCreate(a->A, p->B, NULL, &mp[cp]));
7297: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7298: PetscCall(MatProductSetFill(mp[cp], product->fill));
7299: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7300: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7301: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7302: mp[cp]->product->api_user = product->api_user;
7303: PetscCall(MatProductSetFromOptions(mp[cp]));
7304: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7305: rmapt[cp] = 1;
7306: cmapt[cp] = 2;
7307: cmapa[cp] = p->garray;
7308: mptmp[cp] = PETSC_FALSE;
7309: cp++;
7310: }
7312: /* A_off * P_other */
7313: if (mmdata->P_oth) {
7314: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g)); /* make P_oth use local col ids */
7315: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7316: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7317: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7318: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7319: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7320: PetscCall(MatProductSetFill(mp[cp], product->fill));
7321: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7322: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7323: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7324: mp[cp]->product->api_user = product->api_user;
7325: PetscCall(MatProductSetFromOptions(mp[cp]));
7326: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7327: rmapt[cp] = 1;
7328: cmapt[cp] = 2;
7329: cmapa[cp] = P_oth_idx;
7330: mptmp[cp] = PETSC_FALSE;
7331: cp++;
7332: }
7333: break;
7335: case MATPRODUCT_AtB: /* (P^t * A): P_diag * A_loc + P_off * A_loc */
7336: /* A is product->B */
7337: PetscCall(MatMPIAIJGetLocalMatMerge(A, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7338: if (A == P) { /* when A==P, we can take advantage of the already merged mmdata->Bloc */
7339: PetscCall(MatProductCreate(mmdata->Bloc, mmdata->Bloc, NULL, &mp[cp]));
7340: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7341: PetscCall(MatProductSetFill(mp[cp], product->fill));
7342: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7343: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7344: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7345: mp[cp]->product->api_user = product->api_user;
7346: PetscCall(MatProductSetFromOptions(mp[cp]));
7347: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7348: PetscCall(ISGetIndices(glob, &globidx));
7349: rmapt[cp] = 2;
7350: rmapa[cp] = globidx;
7351: cmapt[cp] = 2;
7352: cmapa[cp] = globidx;
7353: mptmp[cp] = PETSC_FALSE;
7354: cp++;
7355: } else {
7356: PetscCall(MatProductCreate(p->A, mmdata->Bloc, NULL, &mp[cp]));
7357: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7358: PetscCall(MatProductSetFill(mp[cp], product->fill));
7359: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7360: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7361: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7362: mp[cp]->product->api_user = product->api_user;
7363: PetscCall(MatProductSetFromOptions(mp[cp]));
7364: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7365: PetscCall(ISGetIndices(glob, &globidx));
7366: rmapt[cp] = 1;
7367: cmapt[cp] = 2;
7368: cmapa[cp] = globidx;
7369: mptmp[cp] = PETSC_FALSE;
7370: cp++;
7371: PetscCall(MatProductCreate(p->B, mmdata->Bloc, NULL, &mp[cp]));
7372: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7373: PetscCall(MatProductSetFill(mp[cp], product->fill));
7374: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7375: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7376: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7377: mp[cp]->product->api_user = product->api_user;
7378: PetscCall(MatProductSetFromOptions(mp[cp]));
7379: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7380: rmapt[cp] = 2;
7381: rmapa[cp] = p->garray;
7382: cmapt[cp] = 2;
7383: cmapa[cp] = globidx;
7384: mptmp[cp] = PETSC_FALSE;
7385: cp++;
7386: }
7387: break;
7388: case MATPRODUCT_PtAP:
7389: PetscCall(MatGetBrowsOfAoCols_MPIAIJ(A, P, MAT_INITIAL_MATRIX, &mmdata->startsj_s, &mmdata->startsj_r, &mmdata->bufa, &mmdata->P_oth));
7390: /* P is product->B */
7391: PetscCall(MatMPIAIJGetLocalMatMerge(P, MAT_INITIAL_MATRIX, &glob, &mmdata->Bloc));
7392: PetscCall(MatProductCreate(a->A, mmdata->Bloc, NULL, &mp[cp]));
7393: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_PtAP));
7394: PetscCall(MatProductSetFill(mp[cp], product->fill));
7395: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7396: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7397: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7398: mp[cp]->product->api_user = product->api_user;
7399: PetscCall(MatProductSetFromOptions(mp[cp]));
7400: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7401: PetscCall(ISGetIndices(glob, &globidx));
7402: rmapt[cp] = 2;
7403: rmapa[cp] = globidx;
7404: cmapt[cp] = 2;
7405: cmapa[cp] = globidx;
7406: mptmp[cp] = PETSC_FALSE;
7407: cp++;
7408: if (mmdata->P_oth) {
7409: PetscCall(MatSeqAIJCompactOutExtraColumns_SeqAIJ(mmdata->P_oth, &P_oth_l2g));
7410: PetscCall(ISLocalToGlobalMappingGetIndices(P_oth_l2g, &P_oth_idx));
7411: PetscCall(MatSetType(mmdata->P_oth, ((PetscObject)a->B)->type_name));
7412: PetscCall(MatBindToCPU(mmdata->P_oth, mmdata->P_oth_bind));
7413: PetscCall(MatProductCreate(a->B, mmdata->P_oth, NULL, &mp[cp]));
7414: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AB));
7415: PetscCall(MatProductSetFill(mp[cp], product->fill));
7416: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7417: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7418: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7419: mp[cp]->product->api_user = product->api_user;
7420: PetscCall(MatProductSetFromOptions(mp[cp]));
7421: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7422: mptmp[cp] = PETSC_TRUE;
7423: cp++;
7424: PetscCall(MatProductCreate(mmdata->Bloc, mp[1], NULL, &mp[cp]));
7425: PetscCall(MatProductSetType(mp[cp], MATPRODUCT_AtB));
7426: PetscCall(MatProductSetFill(mp[cp], product->fill));
7427: PetscCall(PetscSNPrintf(pprefix, sizeof(pprefix), "backend_p%" PetscInt_FMT "_", cp));
7428: PetscCall(MatSetOptionsPrefix(mp[cp], prefix));
7429: PetscCall(MatAppendOptionsPrefix(mp[cp], pprefix));
7430: mp[cp]->product->api_user = product->api_user;
7431: PetscCall(MatProductSetFromOptions(mp[cp]));
7432: PetscCall((*mp[cp]->ops->productsymbolic)(mp[cp]));
7433: rmapt[cp] = 2;
7434: rmapa[cp] = globidx;
7435: cmapt[cp] = 2;
7436: cmapa[cp] = P_oth_idx;
7437: mptmp[cp] = PETSC_FALSE;
7438: cp++;
7439: }
7440: break;
7441: default:
7442: SETERRQ(PetscObjectComm((PetscObject)C), PETSC_ERR_PLIB, "Not for product type %s", MatProductTypes[ptype]);
7443: }
7444: /* sanity check */
7445: if (size > 1)
7446: 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);
7448: PetscCall(PetscMalloc2(cp, &mmdata->mp, cp, &mmdata->mptmp));
7449: for (i = 0; i < cp; i++) {
7450: mmdata->mp[i] = mp[i];
7451: mmdata->mptmp[i] = mptmp[i];
7452: }
7453: mmdata->cp = cp;
7454: C->product->data = mmdata;
7455: C->product->destroy = MatDestroy_MatMatMPIAIJBACKEND;
7456: C->ops->productnumeric = MatProductNumeric_MPIAIJBACKEND;
7458: /* memory type */
7459: mmdata->mtype = PETSC_MEMTYPE_HOST;
7460: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iscuda, MATSEQAIJCUSPARSE, MATMPIAIJCUSPARSE, ""));
7461: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iship, MATSEQAIJHIPSPARSE, MATMPIAIJHIPSPARSE, ""));
7462: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &iskokk, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, ""));
7463: if (iscuda) mmdata->mtype = PETSC_MEMTYPE_CUDA;
7464: else if (iship) mmdata->mtype = PETSC_MEMTYPE_HIP;
7465: else if (iskokk) mmdata->mtype = PETSC_MEMTYPE_KOKKOS;
7467: /* prepare coo coordinates for values insertion */
7469: /* count total nonzeros of those intermediate seqaij Mats
7470: ncoo_d: # of nonzeros of matrices that do not have offproc entries
7471: ncoo_o: # of nonzeros (of matrices that might have offproc entries) that will be inserted to remote procs
7472: ncoo_oown: # of nonzeros (of matrices that might have offproc entries) that will be inserted locally
7473: */
7474: for (cp = 0, ncoo_d = 0, ncoo_o = 0, ncoo_oown = 0; cp < mmdata->cp; cp++) {
7475: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7476: if (mptmp[cp]) continue;
7477: if (rmapt[cp] == 2 && hasoffproc) { /* the rows need to be scatter to all processes (might include self) */
7478: const PetscInt *rmap = rmapa[cp];
7479: const PetscInt mr = mp[cp]->rmap->n;
7480: const PetscInt rs = C->rmap->rstart;
7481: const PetscInt re = C->rmap->rend;
7482: const PetscInt *ii = mm->i;
7483: for (i = 0; i < mr; i++) {
7484: const PetscInt gr = rmap[i];
7485: const PetscInt nz = ii[i + 1] - ii[i];
7486: if (gr < rs || gr >= re) ncoo_o += nz; /* this row is offproc */
7487: else ncoo_oown += nz; /* this row is local */
7488: }
7489: } else ncoo_d += mm->nz;
7490: }
7492: /*
7493: ncoo: total number of nonzeros (including those inserted by remote procs) belonging to this proc
7495: ncoo = ncoo_d + ncoo_oown + ncoo2, which ncoo2 is number of nonzeros inserted to me by other procs.
7497: off[0] points to a big index array, which is shared by off[1,2,...]. Similarly, for own[0].
7499: off[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert to others
7500: own[p]: points to the segment for matrix mp[p], storing location of nonzeros that mp[p] will insert locally
7501: so, off[p+1]-off[p] is the number of nonzeros that mp[p] will send to others.
7503: coo_i/j/v[]: [ncoo] row/col/val of nonzeros belonging to this proc.
7504: 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.
7505: */
7506: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->off)); /* +1 to make a csr-like data structure */
7507: PetscCall(PetscCalloc1(mmdata->cp + 1, &mmdata->own));
7509: /* gather (i,j) of nonzeros inserted by remote procs */
7510: if (hasoffproc) {
7511: PetscSF msf;
7512: PetscInt ncoo2, *coo_i2, *coo_j2;
7514: PetscCall(PetscMalloc1(ncoo_o, &mmdata->off[0]));
7515: PetscCall(PetscMalloc1(ncoo_oown, &mmdata->own[0]));
7516: PetscCall(PetscMalloc2(ncoo_o, &coo_i, ncoo_o, &coo_j)); /* to collect (i,j) of entries to be sent to others */
7518: for (cp = 0, ncoo_o = 0; cp < mmdata->cp; cp++) {
7519: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7520: PetscInt *idxoff = mmdata->off[cp];
7521: PetscInt *idxown = mmdata->own[cp];
7522: if (!mptmp[cp] && rmapt[cp] == 2) { /* row map is sparse */
7523: const PetscInt *rmap = rmapa[cp];
7524: const PetscInt *cmap = cmapa[cp];
7525: const PetscInt *ii = mm->i;
7526: PetscInt *coi = coo_i + ncoo_o;
7527: PetscInt *coj = coo_j + ncoo_o;
7528: const PetscInt mr = mp[cp]->rmap->n;
7529: const PetscInt rs = C->rmap->rstart;
7530: const PetscInt re = C->rmap->rend;
7531: const PetscInt cs = C->cmap->rstart;
7532: for (i = 0; i < mr; i++) {
7533: const PetscInt *jj = mm->j + ii[i];
7534: const PetscInt gr = rmap[i];
7535: const PetscInt nz = ii[i + 1] - ii[i];
7536: if (gr < rs || gr >= re) { /* this is an offproc row */
7537: for (j = ii[i]; j < ii[i + 1]; j++) {
7538: *coi++ = gr;
7539: *idxoff++ = j;
7540: }
7541: if (!cmapt[cp]) { /* already global */
7542: for (j = 0; j < nz; j++) *coj++ = jj[j];
7543: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7544: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7545: } else { /* offdiag */
7546: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7547: }
7548: ncoo_o += nz;
7549: } else { /* this is a local row */
7550: for (j = ii[i]; j < ii[i + 1]; j++) *idxown++ = j;
7551: }
7552: }
7553: }
7554: mmdata->off[cp + 1] = idxoff;
7555: mmdata->own[cp + 1] = idxown;
7556: }
7558: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7559: PetscInt incoo_o;
7560: PetscCall(PetscIntCast(ncoo_o, &incoo_o));
7561: PetscCall(PetscSFSetGraphLayout(mmdata->sf, C->rmap, incoo_o /*nleaves*/, NULL /*ilocal*/, PETSC_OWN_POINTER, coo_i));
7562: PetscCall(PetscSFGetMultiSF(mmdata->sf, &msf));
7563: PetscCall(PetscSFGetGraph(msf, &ncoo2 /*nroots*/, NULL, NULL, NULL));
7564: ncoo = ncoo_d + ncoo_oown + ncoo2;
7565: PetscCall(PetscMalloc2(ncoo, &coo_i2, ncoo, &coo_j2));
7566: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown)); /* put (i,j) of remote nonzeros at back */
7567: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_i, coo_i2 + ncoo_d + ncoo_oown));
7568: PetscCall(PetscSFGatherBegin(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7569: PetscCall(PetscSFGatherEnd(mmdata->sf, MPIU_INT, coo_j, coo_j2 + ncoo_d + ncoo_oown));
7570: PetscCall(PetscFree2(coo_i, coo_j));
7571: /* allocate MPI send buffer to collect nonzero values to be sent to remote procs */
7572: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo_o * sizeof(PetscScalar), (void **)&mmdata->coo_w));
7573: coo_i = coo_i2;
7574: coo_j = coo_j2;
7575: } else { /* no offproc values insertion */
7576: ncoo = ncoo_d;
7577: PetscCall(PetscMalloc2(ncoo, &coo_i, ncoo, &coo_j));
7579: PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)C), &mmdata->sf));
7580: PetscCall(PetscSFSetGraph(mmdata->sf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
7581: PetscCall(PetscSFSetUp(mmdata->sf));
7582: }
7583: mmdata->hasoffproc = hasoffproc;
7585: /* gather (i,j) of nonzeros inserted locally */
7586: for (cp = 0, ncoo_d = 0; cp < mmdata->cp; cp++) {
7587: Mat_SeqAIJ *mm = (Mat_SeqAIJ *)mp[cp]->data;
7588: PetscInt *coi = coo_i + ncoo_d;
7589: PetscInt *coj = coo_j + ncoo_d;
7590: const PetscInt *jj = mm->j;
7591: const PetscInt *ii = mm->i;
7592: const PetscInt *cmap = cmapa[cp];
7593: const PetscInt *rmap = rmapa[cp];
7594: const PetscInt mr = mp[cp]->rmap->n;
7595: const PetscInt rs = C->rmap->rstart;
7596: const PetscInt re = C->rmap->rend;
7597: const PetscInt cs = C->cmap->rstart;
7599: if (mptmp[cp]) continue;
7600: if (rmapt[cp] == 1) { /* consecutive rows */
7601: /* fill coo_i */
7602: for (i = 0; i < mr; i++) {
7603: const PetscInt gr = i + rs;
7604: for (j = ii[i]; j < ii[i + 1]; j++) coi[j] = gr;
7605: }
7606: /* fill coo_j */
7607: if (!cmapt[cp]) { /* type-0, already global */
7608: PetscCall(PetscArraycpy(coj, jj, mm->nz));
7609: } else if (cmapt[cp] == 1) { /* type-1, local to global for consecutive columns of C */
7610: for (j = 0; j < mm->nz; j++) coj[j] = jj[j] + cs; /* lid + col start */
7611: } else { /* type-2, local to global for sparse columns */
7612: for (j = 0; j < mm->nz; j++) coj[j] = cmap[jj[j]];
7613: }
7614: ncoo_d += mm->nz;
7615: } else if (rmapt[cp] == 2) { /* sparse rows */
7616: for (i = 0; i < mr; i++) {
7617: const PetscInt *jj = mm->j + ii[i];
7618: const PetscInt gr = rmap[i];
7619: const PetscInt nz = ii[i + 1] - ii[i];
7620: if (gr >= rs && gr < re) { /* local rows */
7621: for (j = ii[i]; j < ii[i + 1]; j++) *coi++ = gr;
7622: if (!cmapt[cp]) { /* type-0, already global */
7623: for (j = 0; j < nz; j++) *coj++ = jj[j];
7624: } else if (cmapt[cp] == 1) { /* local to global for owned columns of C */
7625: for (j = 0; j < nz; j++) *coj++ = jj[j] + cs;
7626: } else { /* type-2, local to global for sparse columns */
7627: for (j = 0; j < nz; j++) *coj++ = cmap[jj[j]];
7628: }
7629: ncoo_d += nz;
7630: }
7631: }
7632: }
7633: }
7634: if (glob) PetscCall(ISRestoreIndices(glob, &globidx));
7635: PetscCall(ISDestroy(&glob));
7636: if (P_oth_l2g) PetscCall(ISLocalToGlobalMappingRestoreIndices(P_oth_l2g, &P_oth_idx));
7637: PetscCall(ISLocalToGlobalMappingDestroy(&P_oth_l2g));
7638: /* allocate an array to store all nonzeros (inserted locally or remotely) belonging to this proc */
7639: PetscCall(PetscSFMalloc(mmdata->sf, mmdata->mtype, ncoo * sizeof(PetscScalar), (void **)&mmdata->coo_v));
7641: /* preallocate with COO data */
7642: PetscCall(MatSetPreallocationCOO(C, ncoo, coo_i, coo_j));
7643: PetscCall(PetscFree2(coo_i, coo_j));
7644: PetscFunctionReturn(PETSC_SUCCESS);
7645: }
7647: PetscErrorCode MatProductSetFromOptions_MPIAIJBACKEND(Mat mat)
7648: {
7649: Mat_Product *product = mat->product;
7650: #if defined(PETSC_HAVE_DEVICE)
7651: PetscBool match = PETSC_FALSE;
7652: PetscBool usecpu = PETSC_FALSE;
7653: #else
7654: PetscBool match = PETSC_TRUE;
7655: #endif
7657: PetscFunctionBegin;
7658: MatCheckProduct(mat, 1);
7659: #if defined(PETSC_HAVE_DEVICE)
7660: if (!product->A->boundtocpu && !product->B->boundtocpu) PetscCall(PetscObjectTypeCompare((PetscObject)product->B, ((PetscObject)product->A)->type_name, &match));
7661: if (match) { /* we can always fallback to the CPU if requested */
7662: switch (product->type) {
7663: case MATPRODUCT_AB:
7664: if (product->api_user) {
7665: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatMatMult", "Mat");
7666: PetscCall(PetscOptionsBool("-matmatmult_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7667: PetscOptionsEnd();
7668: } else {
7669: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AB", "Mat");
7670: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatMatMult", usecpu, &usecpu, NULL));
7671: PetscOptionsEnd();
7672: }
7673: break;
7674: case MATPRODUCT_AtB:
7675: if (product->api_user) {
7676: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatTransposeMatMult", "Mat");
7677: PetscCall(PetscOptionsBool("-mattransposematmult_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7678: PetscOptionsEnd();
7679: } else {
7680: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_AtB", "Mat");
7681: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatTransposeMatMult", usecpu, &usecpu, NULL));
7682: PetscOptionsEnd();
7683: }
7684: break;
7685: case MATPRODUCT_PtAP:
7686: if (product->api_user) {
7687: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatPtAP", "Mat");
7688: PetscCall(PetscOptionsBool("-matptap_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7689: PetscOptionsEnd();
7690: } else {
7691: PetscOptionsBegin(PetscObjectComm((PetscObject)mat), ((PetscObject)mat)->prefix, "MatProduct_PtAP", "Mat");
7692: PetscCall(PetscOptionsBool("-mat_product_algorithm_backend_cpu", "Use CPU code", "MatPtAP", usecpu, &usecpu, NULL));
7693: PetscOptionsEnd();
7694: }
7695: break;
7696: default:
7697: break;
7698: }
7699: match = (PetscBool)!usecpu;
7700: }
7701: #endif
7702: if (match) {
7703: switch (product->type) {
7704: case MATPRODUCT_AB:
7705: case MATPRODUCT_AtB:
7706: case MATPRODUCT_PtAP:
7707: mat->ops->productsymbolic = MatProductSymbolic_MPIAIJBACKEND;
7708: break;
7709: default:
7710: break;
7711: }
7712: }
7713: /* fallback to MPIAIJ ops */
7714: if (!mat->ops->productsymbolic) PetscCall(MatProductSetFromOptions_MPIAIJ(mat));
7715: PetscFunctionReturn(PETSC_SUCCESS);
7716: }
7718: /*
7719: Produces a set of block column indices of the matrix row, one for each block represented in the original row
7721: n - the number of block indices in cc[]
7722: cc - the block indices (must be large enough to contain the indices)
7723: */
7724: static inline PetscErrorCode MatCollapseRow(Mat Amat, PetscInt row, PetscInt bs, PetscInt *n, PetscInt *cc)
7725: {
7726: PetscInt cnt = -1, nidx, j;
7727: const PetscInt *idx;
7729: PetscFunctionBegin;
7730: PetscCall(MatGetRow(Amat, row, &nidx, &idx, NULL));
7731: if (nidx) {
7732: cnt = 0;
7733: cc[cnt] = idx[0] / bs;
7734: for (j = 1; j < nidx; j++) {
7735: if (cc[cnt] < idx[j] / bs) cc[++cnt] = idx[j] / bs;
7736: }
7737: }
7738: PetscCall(MatRestoreRow(Amat, row, &nidx, &idx, NULL));
7739: *n = cnt + 1;
7740: PetscFunctionReturn(PETSC_SUCCESS);
7741: }
7743: /*
7744: Produces a set of block column indices of the matrix block row, one for each block represented in the original set of rows
7746: ncollapsed - the number of block indices
7747: collapsed - the block indices (must be large enough to contain the indices)
7748: */
7749: static inline PetscErrorCode MatCollapseRows(Mat Amat, PetscInt start, PetscInt bs, PetscInt *w0, PetscInt *w1, PetscInt *w2, PetscInt *ncollapsed, PetscInt **collapsed)
7750: {
7751: PetscInt i, nprev, *cprev = w0, ncur = 0, *ccur = w1, *merged = w2, *cprevtmp;
7753: PetscFunctionBegin;
7754: PetscCall(MatCollapseRow(Amat, start, bs, &nprev, cprev));
7755: for (i = start + 1; i < start + bs; i++) {
7756: PetscCall(MatCollapseRow(Amat, i, bs, &ncur, ccur));
7757: PetscCall(PetscMergeIntArray(nprev, cprev, ncur, ccur, &nprev, &merged));
7758: cprevtmp = cprev;
7759: cprev = merged;
7760: merged = cprevtmp;
7761: }
7762: *ncollapsed = nprev;
7763: if (collapsed) *collapsed = cprev;
7764: PetscFunctionReturn(PETSC_SUCCESS);
7765: }
7767: /*
7768: MatCreateGraph_Simple_AIJ - create simple scalar matrix (graph) from potentially blocked matrix
7770: Input Parameter:
7771: . Amat - matrix
7772: - symmetrize - make the result symmetric
7773: + scale - scale with diagonal
7775: Output Parameter:
7776: . a_Gmat - output scalar graph >= 0
7778: */
7779: PETSC_INTERN PetscErrorCode MatCreateGraph_Simple_AIJ(Mat Amat, PetscBool symmetrize, PetscBool scale, PetscReal filter, PetscInt index_size, PetscInt index[], Mat *a_Gmat)
7780: {
7781: PetscInt Istart, Iend, Ii, jj, kk, ncols, nloc, NN, MM, bs;
7782: MPI_Comm comm;
7783: Mat Gmat;
7784: PetscBool ismpiaij, isseqaij;
7785: Mat a, b, c;
7786: MatType jtype;
7788: PetscFunctionBegin;
7789: PetscCall(PetscObjectGetComm((PetscObject)Amat, &comm));
7790: PetscCall(MatGetOwnershipRange(Amat, &Istart, &Iend));
7791: PetscCall(MatGetSize(Amat, &MM, &NN));
7792: PetscCall(MatGetBlockSize(Amat, &bs));
7793: nloc = (Iend - Istart) / bs;
7795: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATSEQAIJ, &isseqaij));
7796: PetscCall(PetscObjectBaseTypeCompare((PetscObject)Amat, MATMPIAIJ, &ismpiaij));
7797: PetscCheck(isseqaij || ismpiaij, comm, PETSC_ERR_USER, "Require (MPI)AIJ matrix type");
7799: /* TODO GPU: these calls are potentially expensive if matrices are large and we want to use the GPU */
7800: /* A solution consists in providing a new API, MatAIJGetCollapsedAIJ, and each class can provide a fast
7801: implementation */
7802: if (bs > 1) {
7803: PetscCall(MatGetType(Amat, &jtype));
7804: PetscCall(MatCreate(comm, &Gmat));
7805: PetscCall(MatSetType(Gmat, jtype));
7806: PetscCall(MatSetSizes(Gmat, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE));
7807: PetscCall(MatSetBlockSizes(Gmat, 1, 1));
7808: if (isseqaij || ((Mat_MPIAIJ *)Amat->data)->garray) {
7809: PetscInt *d_nnz, *o_nnz;
7810: MatScalar *aa, val, *AA;
7811: PetscInt *aj, *ai, *AJ, nc, nmax = 0;
7813: if (isseqaij) {
7814: a = Amat;
7815: b = NULL;
7816: } else {
7817: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Amat->data;
7818: a = d->A;
7819: b = d->B;
7820: }
7821: PetscCall(PetscInfo(Amat, "New bs>1 Graph. nloc=%" PetscInt_FMT "\n", nloc));
7822: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7823: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
7824: PetscInt *nnz = (c == a) ? d_nnz : o_nnz;
7825: const PetscInt *cols1, *cols2;
7827: for (PetscInt brow = 0, nc1, nc2, ok = 1; brow < nloc * bs; brow += bs) { // block rows
7828: PetscCall(MatGetRow(c, brow, &nc2, &cols2, NULL));
7829: nnz[brow / bs] = nc2 / bs;
7830: if (nc2 % bs) ok = 0;
7831: if (nnz[brow / bs] > nmax) nmax = nnz[brow / bs];
7832: for (PetscInt ii = 1; ii < bs; ii++) { // check for non-dense blocks
7833: PetscCall(MatGetRow(c, brow + ii, &nc1, &cols1, NULL));
7834: if (nc1 != nc2) ok = 0;
7835: else {
7836: for (PetscInt jj = 0; jj < nc1 && ok == 1; jj++) {
7837: if (cols1[jj] != cols2[jj]) ok = 0;
7838: if (cols1[jj] % bs != jj % bs) ok = 0;
7839: }
7840: }
7841: PetscCall(MatRestoreRow(c, brow + ii, &nc1, &cols1, NULL));
7842: }
7843: PetscCall(MatRestoreRow(c, brow, &nc2, &cols2, NULL));
7844: if (!ok) {
7845: PetscCall(PetscFree2(d_nnz, o_nnz));
7846: PetscCall(PetscInfo(Amat, "Found sparse blocks - revert to slow method\n"));
7847: goto old_bs;
7848: }
7849: }
7850: }
7851: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7852: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7853: PetscCall(PetscFree2(d_nnz, o_nnz));
7854: PetscCall(PetscMalloc2(nmax, &AA, nmax, &AJ));
7855: // diag
7856: for (PetscInt brow = 0, n, grow; brow < nloc * bs; brow += bs) { // block rows
7857: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)a->data;
7859: ai = aseq->i;
7860: n = ai[brow + 1] - ai[brow];
7861: aj = aseq->j + ai[brow];
7862: for (PetscInt k = 0; k < n; k += bs) { // block columns
7863: AJ[k / bs] = aj[k] / bs + Istart / bs; // diag starts at (Istart,Istart)
7864: val = 0;
7865: if (index_size == 0) {
7866: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7867: aa = aseq->a + ai[brow + ii] + k;
7868: for (PetscInt jj = 0; jj < bs; jj++) { // columns in block
7869: val += PetscAbs(PetscRealPart(aa[jj])); // a sort of norm
7870: }
7871: }
7872: } else { // use (index,index) value if provided
7873: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7874: PetscInt ii = index[iii];
7875: aa = aseq->a + ai[brow + ii] + k;
7876: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // columns in block
7877: PetscInt jj = index[jjj];
7878: val += PetscAbs(PetscRealPart(aa[jj]));
7879: }
7880: }
7881: }
7882: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7883: AA[k / bs] = val;
7884: }
7885: grow = Istart / bs + brow / bs;
7886: PetscCall(MatSetValues(Gmat, 1, &grow, n / bs, AJ, AA, ADD_VALUES));
7887: }
7888: // off-diag
7889: if (ismpiaij) {
7890: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)Amat->data;
7891: const PetscScalar *vals;
7892: const PetscInt *cols, *garray = aij->garray;
7894: PetscCheck(garray, PETSC_COMM_SELF, PETSC_ERR_USER, "No garray ?");
7895: for (PetscInt brow = 0, grow; brow < nloc * bs; brow += bs) { // block rows
7896: PetscCall(MatGetRow(b, brow, &ncols, &cols, NULL));
7897: for (PetscInt k = 0, cidx = 0; k < ncols; k += bs, cidx++) {
7898: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs >= nmax");
7899: AA[k / bs] = 0;
7900: AJ[cidx] = garray[cols[k]] / bs;
7901: }
7902: nc = ncols / bs;
7903: PetscCall(MatRestoreRow(b, brow, &ncols, &cols, NULL));
7904: if (index_size == 0) {
7905: for (PetscInt ii = 0; ii < bs; ii++) { // rows in block
7906: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7907: for (PetscInt k = 0; k < ncols; k += bs) {
7908: for (PetscInt jj = 0; jj < bs; jj++) { // cols in block
7909: PetscAssert(k / bs < nmax, comm, PETSC_ERR_USER, "k / bs (%d) >= nmax (%d)", (int)(k / bs), (int)nmax);
7910: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7911: }
7912: }
7913: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7914: }
7915: } else { // use (index,index) value if provided
7916: for (PetscInt iii = 0; iii < index_size; iii++) { // rows in block
7917: PetscInt ii = index[iii];
7918: PetscCall(MatGetRow(b, brow + ii, &ncols, &cols, &vals));
7919: for (PetscInt k = 0; k < ncols; k += bs) {
7920: for (PetscInt jjj = 0; jjj < index_size; jjj++) { // cols in block
7921: PetscInt jj = index[jjj];
7922: AA[k / bs] += PetscAbs(PetscRealPart(vals[k + jj]));
7923: }
7924: }
7925: PetscCall(MatRestoreRow(b, brow + ii, &ncols, &cols, &vals));
7926: }
7927: }
7928: grow = Istart / bs + brow / bs;
7929: PetscCall(MatSetValues(Gmat, 1, &grow, nc, AJ, AA, ADD_VALUES));
7930: }
7931: }
7932: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
7933: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
7934: PetscCall(PetscFree2(AA, AJ));
7935: } else {
7936: const PetscScalar *vals;
7937: const PetscInt *idx;
7938: PetscInt *d_nnz, *o_nnz, *w0, *w1, *w2;
7939: old_bs:
7940: /*
7941: Determine the preallocation needed for the scalar matrix derived from the vector matrix.
7942: */
7943: PetscCall(PetscInfo(Amat, "OLD bs>1 CreateGraph\n"));
7944: PetscCall(PetscMalloc2(nloc, &d_nnz, (isseqaij ? 0 : nloc), &o_nnz));
7945: if (isseqaij) {
7946: PetscInt max_d_nnz;
7948: /*
7949: Determine exact preallocation count for (sequential) scalar matrix
7950: */
7951: PetscCall(MatSeqAIJGetMaxRowNonzeros(Amat, &max_d_nnz));
7952: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7953: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7954: for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) PetscCall(MatCollapseRows(Amat, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7955: PetscCall(PetscFree3(w0, w1, w2));
7956: } else if (ismpiaij) {
7957: Mat Daij, Oaij;
7958: const PetscInt *garray;
7959: PetscInt max_d_nnz;
7961: PetscCall(MatMPIAIJGetSeqAIJ(Amat, &Daij, &Oaij, &garray));
7962: /*
7963: Determine exact preallocation count for diagonal block portion of scalar matrix
7964: */
7965: PetscCall(MatSeqAIJGetMaxRowNonzeros(Daij, &max_d_nnz));
7966: max_d_nnz = PetscMin(nloc, bs * max_d_nnz);
7967: PetscCall(PetscMalloc3(max_d_nnz, &w0, max_d_nnz, &w1, max_d_nnz, &w2));
7968: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) PetscCall(MatCollapseRows(Daij, Ii, bs, w0, w1, w2, &d_nnz[jj], NULL));
7969: PetscCall(PetscFree3(w0, w1, w2));
7970: /*
7971: Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
7972: */
7973: for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
7974: o_nnz[jj] = 0;
7975: for (kk = 0; kk < bs; kk++) { /* rows that get collapsed to a single row */
7976: PetscCall(MatGetRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7977: o_nnz[jj] += ncols;
7978: PetscCall(MatRestoreRow(Oaij, Ii + kk, &ncols, NULL, NULL));
7979: }
7980: if (o_nnz[jj] > (NN / bs - nloc)) o_nnz[jj] = NN / bs - nloc;
7981: }
7982: } else SETERRQ(comm, PETSC_ERR_USER, "Require AIJ matrix type");
7983: /* get scalar copy (norms) of matrix */
7984: PetscCall(MatSeqAIJSetPreallocation(Gmat, 0, d_nnz));
7985: PetscCall(MatMPIAIJSetPreallocation(Gmat, 0, d_nnz, 0, o_nnz));
7986: PetscCall(PetscFree2(d_nnz, o_nnz));
7987: for (Ii = Istart; Ii < Iend; Ii++) {
7988: PetscInt dest_row = Ii / bs;
7990: PetscCall(MatGetRow(Amat, Ii, &ncols, &idx, &vals));
7991: for (jj = 0; jj < ncols; jj++) {
7992: PetscInt dest_col = idx[jj] / bs;
7993: PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
7995: PetscCall(MatSetValues(Gmat, 1, &dest_row, 1, &dest_col, &sv, ADD_VALUES));
7996: }
7997: PetscCall(MatRestoreRow(Amat, Ii, &ncols, &idx, &vals));
7998: }
7999: PetscCall(MatAssemblyBegin(Gmat, MAT_FINAL_ASSEMBLY));
8000: PetscCall(MatAssemblyEnd(Gmat, MAT_FINAL_ASSEMBLY));
8001: }
8002: } else {
8003: if (symmetrize || filter >= 0 || scale) PetscCall(MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat));
8004: else {
8005: Gmat = Amat;
8006: PetscCall(PetscObjectReference((PetscObject)Gmat));
8007: }
8008: if (isseqaij) {
8009: a = Gmat;
8010: b = NULL;
8011: } else {
8012: Mat_MPIAIJ *d = (Mat_MPIAIJ *)Gmat->data;
8013: a = d->A;
8014: b = d->B;
8015: }
8016: if (filter >= 0 || scale) {
8017: /* take absolute value of each entry */
8018: for (c = a, kk = 0; c && kk < 2; c = b, kk++) {
8019: MatInfo info;
8020: PetscScalar *avals;
8022: PetscCall(MatGetInfo(c, MAT_LOCAL, &info));
8023: PetscCall(MatSeqAIJGetArray(c, &avals));
8024: for (int jj = 0; jj < info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
8025: PetscCall(MatSeqAIJRestoreArray(c, &avals));
8026: }
8027: }
8028: }
8029: if (symmetrize) {
8030: PetscBool isset, issym;
8032: PetscCall(MatIsSymmetricKnown(Amat, &isset, &issym));
8033: if (!isset || !issym) {
8034: Mat matTrans;
8036: PetscCall(MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans));
8037: PetscCall(MatAXPY(Gmat, 1.0, matTrans, Gmat->structurally_symmetric == PETSC_BOOL3_TRUE ? SAME_NONZERO_PATTERN : DIFFERENT_NONZERO_PATTERN));
8038: PetscCall(MatDestroy(&matTrans));
8039: }
8040: PetscCall(MatSetOption(Gmat, MAT_SYMMETRIC, PETSC_TRUE));
8041: } else if (Amat != Gmat) PetscCall(MatPropagateSymmetryOptions(Amat, Gmat));
8042: if (scale) {
8043: /* scale c for all diagonal values = 1 or -1 */
8044: Vec diag;
8046: PetscCall(MatCreateVecs(Gmat, &diag, NULL));
8047: PetscCall(MatGetDiagonal(Gmat, diag));
8048: PetscCall(VecReciprocal(diag));
8049: PetscCall(VecSqrtAbs(diag));
8050: PetscCall(MatDiagonalScale(Gmat, diag, diag));
8051: PetscCall(VecDestroy(&diag));
8052: }
8053: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_graph_view"));
8054: if (filter >= 0) {
8055: PetscCall(MatFilter(Gmat, filter, PETSC_TRUE, PETSC_TRUE));
8056: PetscCall(MatViewFromOptions(Gmat, NULL, "-mat_filter_graph_view"));
8057: }
8058: *a_Gmat = Gmat;
8059: PetscFunctionReturn(PETSC_SUCCESS);
8060: }
8062: /*
8063: Special version for direct calls from Fortran
8064: */
8066: /* Change these macros so can be used in void function */
8067: /* Identical to PetscCallVoid, except it assigns to *_ierr */
8068: #undef PetscCall
8069: #define PetscCall(...) \
8070: do { \
8071: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
8072: if (PetscUnlikely(ierr_msv_mpiaij)) { \
8073: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
8074: return; \
8075: } \
8076: } while (0)
8078: #undef SETERRQ
8079: #define SETERRQ(comm, ierr, ...) \
8080: do { \
8081: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
8082: return; \
8083: } while (0)
8085: #if defined(PETSC_HAVE_FORTRAN_CAPS)
8086: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
8087: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
8088: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
8089: #else
8090: #endif
8091: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat, PetscInt *mm, const PetscInt im[], PetscInt *mn, const PetscInt in[], const PetscScalar v[], InsertMode *maddv, PetscErrorCode *_ierr)
8092: {
8093: Mat mat = *mmat;
8094: PetscInt m = *mm, n = *mn;
8095: InsertMode addv = *maddv;
8096: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)mat->data;
8097: PetscScalar value;
8099: MatCheckPreallocated(mat, 1);
8100: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
8101: else PetscCheck(mat->insertmode == addv, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot mix add values and insert values");
8102: {
8103: PetscInt i, j, rstart = mat->rmap->rstart, rend = mat->rmap->rend;
8104: PetscInt cstart = mat->cmap->rstart, cend = mat->cmap->rend, row, col;
8105: PetscBool roworiented = aij->roworiented;
8107: /* Some Variables required in the macro */
8108: Mat A = aij->A;
8109: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
8110: PetscInt *aimax = a->imax, *ai = a->i, *ailen = a->ilen, *aj = a->j;
8111: MatScalar *aa;
8112: PetscBool ignorezeroentries = ((a->ignorezeroentries && (addv == ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
8113: Mat B = aij->B;
8114: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
8115: PetscInt *bimax = b->imax, *bi = b->i, *bilen = b->ilen, *bj = b->j, bm = aij->B->rmap->n, am = aij->A->rmap->n;
8116: MatScalar *ba;
8117: /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
8118: * cannot use "#if defined" inside a macro. */
8119: PETSC_UNUSED PetscBool inserted = PETSC_FALSE;
8121: PetscInt *rp1, *rp2, ii, nrow1, nrow2, _i, rmax1, rmax2, N, low1, high1, low2, high2, t, lastcol1, lastcol2;
8122: PetscInt nonew = a->nonew;
8123: MatScalar *ap1, *ap2;
8125: PetscFunctionBegin;
8126: PetscCall(MatSeqAIJGetArray(A, &aa));
8127: PetscCall(MatSeqAIJGetArray(B, &ba));
8128: for (i = 0; i < m; i++) {
8129: if (im[i] < 0) continue;
8130: 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);
8131: if (im[i] >= rstart && im[i] < rend) {
8132: row = im[i] - rstart;
8133: lastcol1 = -1;
8134: rp1 = aj + ai[row];
8135: ap1 = aa + ai[row];
8136: rmax1 = aimax[row];
8137: nrow1 = ailen[row];
8138: low1 = 0;
8139: high1 = nrow1;
8140: lastcol2 = -1;
8141: rp2 = bj + bi[row];
8142: ap2 = ba + bi[row];
8143: rmax2 = bimax[row];
8144: nrow2 = bilen[row];
8145: low2 = 0;
8146: high2 = nrow2;
8148: for (j = 0; j < n; j++) {
8149: if (roworiented) value = v[i * n + j];
8150: else value = v[i + j * m];
8151: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
8152: if (in[j] >= cstart && in[j] < cend) {
8153: col = in[j] - cstart;
8154: MatSetValues_SeqAIJ_A_Private(row, col, value, addv, im[i], in[j]);
8155: } else if (in[j] < 0) continue;
8156: else if (PetscUnlikelyDebug(in[j] >= mat->cmap->N)) {
8157: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[j], mat->cmap->N - 1);
8158: } else {
8159: if (mat->was_assembled) {
8160: if (!aij->colmap) PetscCall(MatCreateColmap_MPIAIJ_Private(mat));
8161: #if defined(PETSC_USE_CTABLE)
8162: PetscCall(PetscHMapIGetWithDefault(aij->colmap, in[j] + 1, 0, &col));
8163: col--;
8164: #else
8165: col = aij->colmap[in[j]] - 1;
8166: #endif
8167: if (col < 0 && !((Mat_SeqAIJ *)aij->A->data)->nonew) {
8168: PetscCall(MatDisAssemble_MPIAIJ(mat, PETSC_FALSE));
8169: col = in[j];
8170: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
8171: B = aij->B;
8172: b = (Mat_SeqAIJ *)B->data;
8173: bimax = b->imax;
8174: bi = b->i;
8175: bilen = b->ilen;
8176: bj = b->j;
8177: rp2 = bj + bi[row];
8178: ap2 = ba + bi[row];
8179: rmax2 = bimax[row];
8180: nrow2 = bilen[row];
8181: low2 = 0;
8182: high2 = nrow2;
8183: bm = aij->B->rmap->n;
8184: ba = b->a;
8185: inserted = PETSC_FALSE;
8186: }
8187: } else col = in[j];
8188: MatSetValues_SeqAIJ_B_Private(row, col, value, addv, im[i], in[j]);
8189: }
8190: }
8191: } else if (!aij->donotstash) {
8192: if (roworiented) {
8193: PetscCall(MatStashValuesRow_Private(&mat->stash, im[i], n, in, v + i * n, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8194: } else {
8195: PetscCall(MatStashValuesCol_Private(&mat->stash, im[i], n, in, v + i, m, (PetscBool)(ignorezeroentries && (addv == ADD_VALUES))));
8196: }
8197: }
8198: }
8199: PetscCall(MatSeqAIJRestoreArray(A, &aa));
8200: PetscCall(MatSeqAIJRestoreArray(B, &ba));
8201: }
8202: PetscFunctionReturnVoid();
8203: }
8205: /* Undefining these here since they were redefined from their original definition above! No
8206: * other PETSc functions should be defined past this point, as it is impossible to recover the
8207: * original definitions */
8208: #undef PetscCall
8209: #undef SETERRQ